Linux Audio

Check our new training course

Open-source upstreaming

Need help get the support for your hardware in upstream Linux?
Loading...
v6.13.7
    1// SPDX-License-Identifier: GPL-2.0-only
    2/*
    3 *  kernel/sched/core.c
    4 *
    5 *  Core kernel CPU scheduler code
    6 *
    7 *  Copyright (C) 1991-2002  Linus Torvalds
    8 *  Copyright (C) 1998-2024  Ingo Molnar, Red Hat
    9 */
   10#include <linux/highmem.h>
   11#include <linux/hrtimer_api.h>
   12#include <linux/ktime_api.h>
   13#include <linux/sched/signal.h>
   14#include <linux/syscalls_api.h>
   15#include <linux/debug_locks.h>
   16#include <linux/prefetch.h>
   17#include <linux/capability.h>
   18#include <linux/pgtable_api.h>
   19#include <linux/wait_bit.h>
   20#include <linux/jiffies.h>
   21#include <linux/spinlock_api.h>
   22#include <linux/cpumask_api.h>
   23#include <linux/lockdep_api.h>
   24#include <linux/hardirq.h>
   25#include <linux/softirq.h>
   26#include <linux/refcount_api.h>
   27#include <linux/topology.h>
   28#include <linux/sched/clock.h>
   29#include <linux/sched/cond_resched.h>
   30#include <linux/sched/cputime.h>
   31#include <linux/sched/debug.h>
   32#include <linux/sched/hotplug.h>
   33#include <linux/sched/init.h>
   34#include <linux/sched/isolation.h>
   35#include <linux/sched/loadavg.h>
   36#include <linux/sched/mm.h>
   37#include <linux/sched/nohz.h>
   38#include <linux/sched/rseq_api.h>
   39#include <linux/sched/rt.h>
   40
   41#include <linux/blkdev.h>
   42#include <linux/context_tracking.h>
   43#include <linux/cpuset.h>
   44#include <linux/delayacct.h>
   45#include <linux/init_task.h>
   46#include <linux/interrupt.h>
   47#include <linux/ioprio.h>
   48#include <linux/kallsyms.h>
   49#include <linux/kcov.h>
   50#include <linux/kprobes.h>
   51#include <linux/llist_api.h>
   52#include <linux/mmu_context.h>
   53#include <linux/mmzone.h>
   54#include <linux/mutex_api.h>
   55#include <linux/nmi.h>
   56#include <linux/nospec.h>
   57#include <linux/perf_event_api.h>
   58#include <linux/profile.h>
   59#include <linux/psi.h>
   60#include <linux/rcuwait_api.h>
   61#include <linux/rseq.h>
   62#include <linux/sched/wake_q.h>
   63#include <linux/scs.h>
   64#include <linux/slab.h>
   65#include <linux/syscalls.h>
   66#include <linux/vtime.h>
   67#include <linux/wait_api.h>
   68#include <linux/workqueue_api.h>
   69
   70#ifdef CONFIG_PREEMPT_DYNAMIC
   71# ifdef CONFIG_GENERIC_ENTRY
   72#  include <linux/entry-common.h>
   73# endif
   74#endif
   75
   76#include <uapi/linux/sched/types.h>
   77
   78#include <asm/irq_regs.h>
   79#include <asm/switch_to.h>
   80#include <asm/tlb.h>
   81
   82#define CREATE_TRACE_POINTS
   83#include <linux/sched/rseq_api.h>
   84#include <trace/events/sched.h>
   85#include <trace/events/ipi.h>
   86#undef CREATE_TRACE_POINTS
   87
   88#include "sched.h"
   89#include "stats.h"
   90
   91#include "autogroup.h"
   92#include "pelt.h"
   93#include "smp.h"
   94#include "stats.h"
 
 
 
   95
   96#include "../workqueue_internal.h"
   97#include "../../io_uring/io-wq.h"
   98#include "../smpboot.h"
   99
  100EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpu);
  101EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpumask);
  102
  103/*
  104 * Export tracepoints that act as a bare tracehook (ie: have no trace event
  105 * associated with them) to allow external modules to probe them.
  106 */
  107EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
  108EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
  109EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
  110EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
  111EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
  112EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_hw_tp);
  113EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
  114EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
  115EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
  116EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
  117EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
  118EXPORT_TRACEPOINT_SYMBOL_GPL(sched_compute_energy_tp);
  119
  120DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
  121
  122#ifdef CONFIG_SCHED_DEBUG
  123/*
  124 * Debugging: various feature bits
  125 *
  126 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
  127 * sysctl_sched_features, defined in sched.h, to allow constants propagation
  128 * at compile time and compiler optimization based on features default.
  129 */
  130#define SCHED_FEAT(name, enabled)	\
  131	(1UL << __SCHED_FEAT_##name) * enabled |
  132const_debug unsigned int sysctl_sched_features =
  133#include "features.h"
  134	0;
  135#undef SCHED_FEAT
  136
  137/*
  138 * Print a warning if need_resched is set for the given duration (if
  139 * LATENCY_WARN is enabled).
  140 *
  141 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
  142 * per boot.
  143 */
  144__read_mostly int sysctl_resched_latency_warn_ms = 100;
  145__read_mostly int sysctl_resched_latency_warn_once = 1;
  146#endif /* CONFIG_SCHED_DEBUG */
  147
  148/*
  149 * Number of tasks to iterate in a single balance run.
  150 * Limited because this is done with IRQs disabled.
  151 */
  152const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
 
 
 
 
 
 
  153
  154__read_mostly int scheduler_running;
  155
  156#ifdef CONFIG_SCHED_CORE
  157
  158DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
  159
  160/* kernel prio, less is more */
  161static inline int __task_prio(const struct task_struct *p)
  162{
  163	if (p->sched_class == &stop_sched_class) /* trumps deadline */
  164		return -2;
  165
  166	if (p->dl_server)
  167		return -1; /* deadline */
  168
  169	if (rt_or_dl_prio(p->prio))
  170		return p->prio; /* [-1, 99] */
  171
  172	if (p->sched_class == &idle_sched_class)
  173		return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
  174
  175	if (task_on_scx(p))
  176		return MAX_RT_PRIO + MAX_NICE + 1; /* 120, squash ext */
  177
  178	return MAX_RT_PRIO + MAX_NICE; /* 119, squash fair */
  179}
  180
  181/*
  182 * l(a,b)
  183 * le(a,b) := !l(b,a)
  184 * g(a,b)  := l(b,a)
  185 * ge(a,b) := !l(a,b)
  186 */
  187
  188/* real prio, less is less */
  189static inline bool prio_less(const struct task_struct *a,
  190			     const struct task_struct *b, bool in_fi)
  191{
  192
  193	int pa = __task_prio(a), pb = __task_prio(b);
  194
  195	if (-pa < -pb)
  196		return true;
  197
  198	if (-pb < -pa)
  199		return false;
  200
  201	if (pa == -1) { /* dl_prio() doesn't work because of stop_class above */
  202		const struct sched_dl_entity *a_dl, *b_dl;
  203
  204		a_dl = &a->dl;
  205		/*
  206		 * Since,'a' and 'b' can be CFS tasks served by DL server,
  207		 * __task_prio() can return -1 (for DL) even for those. In that
  208		 * case, get to the dl_server's DL entity.
  209		 */
  210		if (a->dl_server)
  211			a_dl = a->dl_server;
  212
  213		b_dl = &b->dl;
  214		if (b->dl_server)
  215			b_dl = b->dl_server;
  216
  217		return !dl_time_before(a_dl->deadline, b_dl->deadline);
  218	}
  219
  220	if (pa == MAX_RT_PRIO + MAX_NICE)	/* fair */
  221		return cfs_prio_less(a, b, in_fi);
  222
  223#ifdef CONFIG_SCHED_CLASS_EXT
  224	if (pa == MAX_RT_PRIO + MAX_NICE + 1)	/* ext */
  225		return scx_prio_less(a, b, in_fi);
  226#endif
  227
  228	return false;
  229}
  230
  231static inline bool __sched_core_less(const struct task_struct *a,
  232				     const struct task_struct *b)
  233{
  234	if (a->core_cookie < b->core_cookie)
  235		return true;
  236
  237	if (a->core_cookie > b->core_cookie)
  238		return false;
  239
  240	/* flip prio, so high prio is leftmost */
  241	if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
  242		return true;
  243
  244	return false;
  245}
  246
  247#define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
  248
  249static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
  250{
  251	return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
  252}
  253
  254static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
  255{
  256	const struct task_struct *p = __node_2_sc(node);
  257	unsigned long cookie = (unsigned long)key;
  258
  259	if (cookie < p->core_cookie)
  260		return -1;
  261
  262	if (cookie > p->core_cookie)
  263		return 1;
  264
  265	return 0;
  266}
  267
  268void sched_core_enqueue(struct rq *rq, struct task_struct *p)
  269{
  270	if (p->se.sched_delayed)
  271		return;
  272
  273	rq->core->core_task_seq++;
  274
  275	if (!p->core_cookie)
  276		return;
  277
  278	rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
  279}
  280
  281void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
  282{
  283	if (p->se.sched_delayed)
  284		return;
  285
  286	rq->core->core_task_seq++;
  287
  288	if (sched_core_enqueued(p)) {
  289		rb_erase(&p->core_node, &rq->core_tree);
  290		RB_CLEAR_NODE(&p->core_node);
  291	}
  292
  293	/*
  294	 * Migrating the last task off the cpu, with the cpu in forced idle
  295	 * state. Reschedule to create an accounting edge for forced idle,
  296	 * and re-examine whether the core is still in forced idle state.
  297	 */
  298	if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
  299	    rq->core->core_forceidle_count && rq->curr == rq->idle)
  300		resched_curr(rq);
  301}
  302
  303static int sched_task_is_throttled(struct task_struct *p, int cpu)
 
 
 
  304{
  305	if (p->sched_class->task_is_throttled)
  306		return p->sched_class->task_is_throttled(p, cpu);
  307
  308	return 0;
 
 
 
 
 
 
 
  309}
  310
  311static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
  312{
  313	struct rb_node *node = &p->core_node;
  314	int cpu = task_cpu(p);
  315
  316	do {
  317		node = rb_next(node);
  318		if (!node)
  319			return NULL;
  320
  321		p = __node_2_sc(node);
  322		if (p->core_cookie != cookie)
  323			return NULL;
  324
  325	} while (sched_task_is_throttled(p, cpu));
  326
  327	return p;
  328}
  329
  330/*
  331 * Find left-most (aka, highest priority) and unthrottled task matching @cookie.
  332 * If no suitable task is found, NULL will be returned.
  333 */
  334static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
  335{
  336	struct task_struct *p;
  337	struct rb_node *node;
  338
  339	node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
  340	if (!node)
  341		return NULL;
  342
  343	p = __node_2_sc(node);
  344	if (!sched_task_is_throttled(p, rq->cpu))
  345		return p;
  346
  347	return sched_core_next(p, cookie);
  348}
  349
  350/*
  351 * Magic required such that:
  352 *
  353 *	raw_spin_rq_lock(rq);
  354 *	...
  355 *	raw_spin_rq_unlock(rq);
  356 *
  357 * ends up locking and unlocking the _same_ lock, and all CPUs
  358 * always agree on what rq has what lock.
  359 *
  360 * XXX entirely possible to selectively enable cores, don't bother for now.
  361 */
  362
  363static DEFINE_MUTEX(sched_core_mutex);
  364static atomic_t sched_core_count;
  365static struct cpumask sched_core_mask;
  366
  367static void sched_core_lock(int cpu, unsigned long *flags)
  368{
  369	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
  370	int t, i = 0;
  371
  372	local_irq_save(*flags);
  373	for_each_cpu(t, smt_mask)
  374		raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
  375}
  376
  377static void sched_core_unlock(int cpu, unsigned long *flags)
  378{
  379	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
  380	int t;
  381
  382	for_each_cpu(t, smt_mask)
  383		raw_spin_unlock(&cpu_rq(t)->__lock);
  384	local_irq_restore(*flags);
  385}
  386
  387static void __sched_core_flip(bool enabled)
  388{
  389	unsigned long flags;
  390	int cpu, t;
  391
  392	cpus_read_lock();
  393
  394	/*
  395	 * Toggle the online cores, one by one.
  396	 */
  397	cpumask_copy(&sched_core_mask, cpu_online_mask);
  398	for_each_cpu(cpu, &sched_core_mask) {
  399		const struct cpumask *smt_mask = cpu_smt_mask(cpu);
  400
  401		sched_core_lock(cpu, &flags);
  402
  403		for_each_cpu(t, smt_mask)
  404			cpu_rq(t)->core_enabled = enabled;
  405
  406		cpu_rq(cpu)->core->core_forceidle_start = 0;
  407
  408		sched_core_unlock(cpu, &flags);
  409
  410		cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
  411	}
  412
  413	/*
  414	 * Toggle the offline CPUs.
  415	 */
  416	for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
 
 
 
  417		cpu_rq(cpu)->core_enabled = enabled;
  418
  419	cpus_read_unlock();
  420}
  421
  422static void sched_core_assert_empty(void)
  423{
  424	int cpu;
  425
  426	for_each_possible_cpu(cpu)
  427		WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
  428}
  429
  430static void __sched_core_enable(void)
  431{
  432	static_branch_enable(&__sched_core_enabled);
  433	/*
  434	 * Ensure all previous instances of raw_spin_rq_*lock() have finished
  435	 * and future ones will observe !sched_core_disabled().
  436	 */
  437	synchronize_rcu();
  438	__sched_core_flip(true);
  439	sched_core_assert_empty();
  440}
  441
  442static void __sched_core_disable(void)
  443{
  444	sched_core_assert_empty();
  445	__sched_core_flip(false);
  446	static_branch_disable(&__sched_core_enabled);
  447}
  448
  449void sched_core_get(void)
  450{
  451	if (atomic_inc_not_zero(&sched_core_count))
  452		return;
  453
  454	mutex_lock(&sched_core_mutex);
  455	if (!atomic_read(&sched_core_count))
  456		__sched_core_enable();
  457
  458	smp_mb__before_atomic();
  459	atomic_inc(&sched_core_count);
  460	mutex_unlock(&sched_core_mutex);
  461}
  462
  463static void __sched_core_put(struct work_struct *work)
  464{
  465	if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
  466		__sched_core_disable();
  467		mutex_unlock(&sched_core_mutex);
  468	}
  469}
  470
  471void sched_core_put(void)
  472{
  473	static DECLARE_WORK(_work, __sched_core_put);
  474
  475	/*
  476	 * "There can be only one"
  477	 *
  478	 * Either this is the last one, or we don't actually need to do any
  479	 * 'work'. If it is the last *again*, we rely on
  480	 * WORK_STRUCT_PENDING_BIT.
  481	 */
  482	if (!atomic_add_unless(&sched_core_count, -1, 1))
  483		schedule_work(&_work);
  484}
  485
  486#else /* !CONFIG_SCHED_CORE */
  487
  488static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
  489static inline void
  490sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
  491
  492#endif /* CONFIG_SCHED_CORE */
  493
  494/*
 
 
 
 
 
 
 
  495 * Serialization rules:
  496 *
  497 * Lock order:
  498 *
  499 *   p->pi_lock
  500 *     rq->lock
  501 *       hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
  502 *
  503 *  rq1->lock
  504 *    rq2->lock  where: rq1 < rq2
  505 *
  506 * Regular state:
  507 *
  508 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
  509 * local CPU's rq->lock, it optionally removes the task from the runqueue and
  510 * always looks at the local rq data structures to find the most eligible task
  511 * to run next.
  512 *
  513 * Task enqueue is also under rq->lock, possibly taken from another CPU.
  514 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
  515 * the local CPU to avoid bouncing the runqueue state around [ see
  516 * ttwu_queue_wakelist() ]
  517 *
  518 * Task wakeup, specifically wakeups that involve migration, are horribly
  519 * complicated to avoid having to take two rq->locks.
  520 *
  521 * Special state:
  522 *
  523 * System-calls and anything external will use task_rq_lock() which acquires
  524 * both p->pi_lock and rq->lock. As a consequence the state they change is
  525 * stable while holding either lock:
  526 *
  527 *  - sched_setaffinity()/
  528 *    set_cpus_allowed_ptr():	p->cpus_ptr, p->nr_cpus_allowed
  529 *  - set_user_nice():		p->se.load, p->*prio
  530 *  - __sched_setscheduler():	p->sched_class, p->policy, p->*prio,
  531 *				p->se.load, p->rt_priority,
  532 *				p->dl.dl_{runtime, deadline, period, flags, bw, density}
  533 *  - sched_setnuma():		p->numa_preferred_nid
  534 *  - sched_move_task():	p->sched_task_group
 
  535 *  - uclamp_update_active()	p->uclamp*
  536 *
  537 * p->state <- TASK_*:
  538 *
  539 *   is changed locklessly using set_current_state(), __set_current_state() or
  540 *   set_special_state(), see their respective comments, or by
  541 *   try_to_wake_up(). This latter uses p->pi_lock to serialize against
  542 *   concurrent self.
  543 *
  544 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
  545 *
  546 *   is set by activate_task() and cleared by deactivate_task(), under
  547 *   rq->lock. Non-zero indicates the task is runnable, the special
  548 *   ON_RQ_MIGRATING state is used for migration without holding both
  549 *   rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
  550 *
  551 *   Additionally it is possible to be ->on_rq but still be considered not
  552 *   runnable when p->se.sched_delayed is true. These tasks are on the runqueue
  553 *   but will be dequeued as soon as they get picked again. See the
  554 *   task_is_runnable() helper.
  555 *
  556 * p->on_cpu <- { 0, 1 }:
  557 *
  558 *   is set by prepare_task() and cleared by finish_task() such that it will be
  559 *   set before p is scheduled-in and cleared after p is scheduled-out, both
  560 *   under rq->lock. Non-zero indicates the task is running on its CPU.
  561 *
  562 *   [ The astute reader will observe that it is possible for two tasks on one
  563 *     CPU to have ->on_cpu = 1 at the same time. ]
  564 *
  565 * task_cpu(p): is changed by set_task_cpu(), the rules are:
  566 *
  567 *  - Don't call set_task_cpu() on a blocked task:
  568 *
  569 *    We don't care what CPU we're not running on, this simplifies hotplug,
  570 *    the CPU assignment of blocked tasks isn't required to be valid.
  571 *
  572 *  - for try_to_wake_up(), called under p->pi_lock:
  573 *
  574 *    This allows try_to_wake_up() to only take one rq->lock, see its comment.
  575 *
  576 *  - for migration called under rq->lock:
  577 *    [ see task_on_rq_migrating() in task_rq_lock() ]
  578 *
  579 *    o move_queued_task()
  580 *    o detach_task()
  581 *
  582 *  - for migration called under double_rq_lock():
  583 *
  584 *    o __migrate_swap_task()
  585 *    o push_rt_task() / pull_rt_task()
  586 *    o push_dl_task() / pull_dl_task()
  587 *    o dl_task_offline_migration()
  588 *
  589 */
  590
  591void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
  592{
  593	raw_spinlock_t *lock;
  594
  595	/* Matches synchronize_rcu() in __sched_core_enable() */
  596	preempt_disable();
  597	if (sched_core_disabled()) {
  598		raw_spin_lock_nested(&rq->__lock, subclass);
  599		/* preempt_count *MUST* be > 1 */
  600		preempt_enable_no_resched();
  601		return;
  602	}
  603
  604	for (;;) {
  605		lock = __rq_lockp(rq);
  606		raw_spin_lock_nested(lock, subclass);
  607		if (likely(lock == __rq_lockp(rq))) {
  608			/* preempt_count *MUST* be > 1 */
  609			preempt_enable_no_resched();
  610			return;
  611		}
  612		raw_spin_unlock(lock);
  613	}
  614}
  615
  616bool raw_spin_rq_trylock(struct rq *rq)
  617{
  618	raw_spinlock_t *lock;
  619	bool ret;
  620
  621	/* Matches synchronize_rcu() in __sched_core_enable() */
  622	preempt_disable();
  623	if (sched_core_disabled()) {
  624		ret = raw_spin_trylock(&rq->__lock);
  625		preempt_enable();
  626		return ret;
  627	}
  628
  629	for (;;) {
  630		lock = __rq_lockp(rq);
  631		ret = raw_spin_trylock(lock);
  632		if (!ret || (likely(lock == __rq_lockp(rq)))) {
  633			preempt_enable();
  634			return ret;
  635		}
  636		raw_spin_unlock(lock);
  637	}
  638}
  639
  640void raw_spin_rq_unlock(struct rq *rq)
  641{
  642	raw_spin_unlock(rq_lockp(rq));
  643}
  644
  645#ifdef CONFIG_SMP
  646/*
  647 * double_rq_lock - safely lock two runqueues
  648 */
  649void double_rq_lock(struct rq *rq1, struct rq *rq2)
  650{
  651	lockdep_assert_irqs_disabled();
  652
  653	if (rq_order_less(rq2, rq1))
  654		swap(rq1, rq2);
  655
  656	raw_spin_rq_lock(rq1);
  657	if (__rq_lockp(rq1) != __rq_lockp(rq2))
  658		raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
  659
  660	double_rq_clock_clear_update(rq1, rq2);
  661}
  662#endif
  663
  664/*
  665 * __task_rq_lock - lock the rq @p resides on.
  666 */
  667struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
  668	__acquires(rq->lock)
  669{
  670	struct rq *rq;
  671
  672	lockdep_assert_held(&p->pi_lock);
  673
  674	for (;;) {
  675		rq = task_rq(p);
  676		raw_spin_rq_lock(rq);
  677		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
  678			rq_pin_lock(rq, rf);
  679			return rq;
  680		}
  681		raw_spin_rq_unlock(rq);
  682
  683		while (unlikely(task_on_rq_migrating(p)))
  684			cpu_relax();
  685	}
  686}
  687
  688/*
  689 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
  690 */
  691struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
  692	__acquires(p->pi_lock)
  693	__acquires(rq->lock)
  694{
  695	struct rq *rq;
  696
  697	for (;;) {
  698		raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
  699		rq = task_rq(p);
  700		raw_spin_rq_lock(rq);
  701		/*
  702		 *	move_queued_task()		task_rq_lock()
  703		 *
  704		 *	ACQUIRE (rq->lock)
  705		 *	[S] ->on_rq = MIGRATING		[L] rq = task_rq()
  706		 *	WMB (__set_task_cpu())		ACQUIRE (rq->lock);
  707		 *	[S] ->cpu = new_cpu		[L] task_rq()
  708		 *					[L] ->on_rq
  709		 *	RELEASE (rq->lock)
  710		 *
  711		 * If we observe the old CPU in task_rq_lock(), the acquire of
  712		 * the old rq->lock will fully serialize against the stores.
  713		 *
  714		 * If we observe the new CPU in task_rq_lock(), the address
  715		 * dependency headed by '[L] rq = task_rq()' and the acquire
  716		 * will pair with the WMB to ensure we then also see migrating.
  717		 */
  718		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
  719			rq_pin_lock(rq, rf);
  720			return rq;
  721		}
  722		raw_spin_rq_unlock(rq);
  723		raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
  724
  725		while (unlikely(task_on_rq_migrating(p)))
  726			cpu_relax();
  727	}
  728}
  729
  730/*
  731 * RQ-clock updating methods:
  732 */
  733
  734static void update_rq_clock_task(struct rq *rq, s64 delta)
  735{
  736/*
  737 * In theory, the compile should just see 0 here, and optimize out the call
  738 * to sched_rt_avg_update. But I don't trust it...
  739 */
  740	s64 __maybe_unused steal = 0, irq_delta = 0;
  741
  742#ifdef CONFIG_IRQ_TIME_ACCOUNTING
  743	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
  744
  745	/*
  746	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
  747	 * this case when a previous update_rq_clock() happened inside a
  748	 * {soft,}IRQ region.
  749	 *
  750	 * When this happens, we stop ->clock_task and only update the
  751	 * prev_irq_time stamp to account for the part that fit, so that a next
  752	 * update will consume the rest. This ensures ->clock_task is
  753	 * monotonic.
  754	 *
  755	 * It does however cause some slight miss-attribution of {soft,}IRQ
  756	 * time, a more accurate solution would be to update the irq_time using
  757	 * the current rq->clock timestamp, except that would require using
  758	 * atomic ops.
  759	 */
  760	if (irq_delta > delta)
  761		irq_delta = delta;
  762
  763	rq->prev_irq_time += irq_delta;
  764	delta -= irq_delta;
  765	delayacct_irq(rq->curr, irq_delta);
  766#endif
  767#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
  768	if (static_key_false((&paravirt_steal_rq_enabled))) {
  769		u64 prev_steal;
  770
  771		steal = prev_steal = paravirt_steal_clock(cpu_of(rq));
  772		steal -= rq->prev_steal_time_rq;
  773
  774		if (unlikely(steal > delta))
  775			steal = delta;
  776
  777		rq->prev_steal_time_rq = prev_steal;
  778		delta -= steal;
  779	}
  780#endif
  781
  782	rq->clock_task += delta;
  783
  784#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
  785	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
  786		update_irq_load_avg(rq, irq_delta + steal);
  787#endif
  788	update_rq_clock_pelt(rq, delta);
  789}
  790
  791void update_rq_clock(struct rq *rq)
  792{
  793	s64 delta;
  794
  795	lockdep_assert_rq_held(rq);
  796
  797	if (rq->clock_update_flags & RQCF_ACT_SKIP)
  798		return;
  799
  800#ifdef CONFIG_SCHED_DEBUG
  801	if (sched_feat(WARN_DOUBLE_CLOCK))
  802		SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
  803	rq->clock_update_flags |= RQCF_UPDATED;
  804#endif
  805
  806	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
  807	if (delta < 0)
  808		return;
  809	rq->clock += delta;
  810	update_rq_clock_task(rq, delta);
  811}
  812
  813#ifdef CONFIG_SCHED_HRTICK
  814/*
  815 * Use HR-timers to deliver accurate preemption points.
  816 */
  817
  818static void hrtick_clear(struct rq *rq)
  819{
  820	if (hrtimer_active(&rq->hrtick_timer))
  821		hrtimer_cancel(&rq->hrtick_timer);
  822}
  823
  824/*
  825 * High-resolution timer tick.
  826 * Runs from hardirq context with interrupts disabled.
  827 */
  828static enum hrtimer_restart hrtick(struct hrtimer *timer)
  829{
  830	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
  831	struct rq_flags rf;
  832
  833	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
  834
  835	rq_lock(rq, &rf);
  836	update_rq_clock(rq);
  837	rq->donor->sched_class->task_tick(rq, rq->curr, 1);
  838	rq_unlock(rq, &rf);
  839
  840	return HRTIMER_NORESTART;
  841}
  842
  843#ifdef CONFIG_SMP
  844
  845static void __hrtick_restart(struct rq *rq)
  846{
  847	struct hrtimer *timer = &rq->hrtick_timer;
  848	ktime_t time = rq->hrtick_time;
  849
  850	hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
  851}
  852
  853/*
  854 * called from hardirq (IPI) context
  855 */
  856static void __hrtick_start(void *arg)
  857{
  858	struct rq *rq = arg;
  859	struct rq_flags rf;
  860
  861	rq_lock(rq, &rf);
  862	__hrtick_restart(rq);
  863	rq_unlock(rq, &rf);
  864}
  865
  866/*
  867 * Called to set the hrtick timer state.
  868 *
  869 * called with rq->lock held and IRQs disabled
  870 */
  871void hrtick_start(struct rq *rq, u64 delay)
  872{
  873	struct hrtimer *timer = &rq->hrtick_timer;
  874	s64 delta;
  875
  876	/*
  877	 * Don't schedule slices shorter than 10000ns, that just
  878	 * doesn't make sense and can cause timer DoS.
  879	 */
  880	delta = max_t(s64, delay, 10000LL);
  881	rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
  882
  883	if (rq == this_rq())
  884		__hrtick_restart(rq);
  885	else
  886		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
  887}
  888
  889#else
  890/*
  891 * Called to set the hrtick timer state.
  892 *
  893 * called with rq->lock held and IRQs disabled
  894 */
  895void hrtick_start(struct rq *rq, u64 delay)
  896{
  897	/*
  898	 * Don't schedule slices shorter than 10000ns, that just
  899	 * doesn't make sense. Rely on vruntime for fairness.
  900	 */
  901	delay = max_t(u64, delay, 10000LL);
  902	hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
  903		      HRTIMER_MODE_REL_PINNED_HARD);
  904}
  905
  906#endif /* CONFIG_SMP */
  907
  908static void hrtick_rq_init(struct rq *rq)
  909{
  910#ifdef CONFIG_SMP
  911	INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
  912#endif
  913	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
  914	rq->hrtick_timer.function = hrtick;
  915}
  916#else	/* CONFIG_SCHED_HRTICK */
  917static inline void hrtick_clear(struct rq *rq)
  918{
  919}
  920
  921static inline void hrtick_rq_init(struct rq *rq)
  922{
  923}
  924#endif	/* CONFIG_SCHED_HRTICK */
  925
  926/*
  927 * try_cmpxchg based fetch_or() macro so it works for different integer types:
  928 */
  929#define fetch_or(ptr, mask)						\
  930	({								\
  931		typeof(ptr) _ptr = (ptr);				\
  932		typeof(mask) _mask = (mask);				\
  933		typeof(*_ptr) _val = *_ptr;				\
  934									\
  935		do {							\
  936		} while (!try_cmpxchg(_ptr, &_val, _val | _mask));	\
  937	_val;								\
 
 
 
 
  938})
  939
  940#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
  941/*
  942 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
  943 * this avoids any races wrt polling state changes and thereby avoids
  944 * spurious IPIs.
  945 */
  946static inline bool set_nr_and_not_polling(struct thread_info *ti, int tif)
  947{
  948	return !(fetch_or(&ti->flags, 1 << tif) & _TIF_POLLING_NRFLAG);
 
  949}
  950
  951/*
  952 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
  953 *
  954 * If this returns true, then the idle task promises to call
  955 * sched_ttwu_pending() and reschedule soon.
  956 */
  957static bool set_nr_if_polling(struct task_struct *p)
  958{
  959	struct thread_info *ti = task_thread_info(p);
  960	typeof(ti->flags) val = READ_ONCE(ti->flags);
  961
  962	do {
  963		if (!(val & _TIF_POLLING_NRFLAG))
  964			return false;
  965		if (val & _TIF_NEED_RESCHED)
  966			return true;
  967	} while (!try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED));
  968
 
 
 
  969	return true;
  970}
  971
  972#else
  973static inline bool set_nr_and_not_polling(struct thread_info *ti, int tif)
  974{
  975	set_ti_thread_flag(ti, tif);
  976	return true;
  977}
  978
  979#ifdef CONFIG_SMP
  980static inline bool set_nr_if_polling(struct task_struct *p)
  981{
  982	return false;
  983}
  984#endif
  985#endif
  986
  987static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
  988{
  989	struct wake_q_node *node = &task->wake_q;
  990
  991	/*
  992	 * Atomically grab the task, if ->wake_q is !nil already it means
  993	 * it's already queued (either by us or someone else) and will get the
  994	 * wakeup due to that.
  995	 *
  996	 * In order to ensure that a pending wakeup will observe our pending
  997	 * state, even in the failed case, an explicit smp_mb() must be used.
  998	 */
  999	smp_mb__before_atomic();
 1000	if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
 1001		return false;
 1002
 1003	/*
 1004	 * The head is context local, there can be no concurrency.
 1005	 */
 1006	*head->lastp = node;
 1007	head->lastp = &node->next;
 1008	return true;
 1009}
 1010
 1011/**
 1012 * wake_q_add() - queue a wakeup for 'later' waking.
 1013 * @head: the wake_q_head to add @task to
 1014 * @task: the task to queue for 'later' wakeup
 1015 *
 1016 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
 1017 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
 1018 * instantly.
 1019 *
 1020 * This function must be used as-if it were wake_up_process(); IOW the task
 1021 * must be ready to be woken at this location.
 1022 */
 1023void wake_q_add(struct wake_q_head *head, struct task_struct *task)
 1024{
 1025	if (__wake_q_add(head, task))
 1026		get_task_struct(task);
 1027}
 1028
 1029/**
 1030 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
 1031 * @head: the wake_q_head to add @task to
 1032 * @task: the task to queue for 'later' wakeup
 1033 *
 1034 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
 1035 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
 1036 * instantly.
 1037 *
 1038 * This function must be used as-if it were wake_up_process(); IOW the task
 1039 * must be ready to be woken at this location.
 1040 *
 1041 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
 1042 * that already hold reference to @task can call the 'safe' version and trust
 1043 * wake_q to do the right thing depending whether or not the @task is already
 1044 * queued for wakeup.
 1045 */
 1046void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
 1047{
 1048	if (!__wake_q_add(head, task))
 1049		put_task_struct(task);
 1050}
 1051
 1052void wake_up_q(struct wake_q_head *head)
 1053{
 1054	struct wake_q_node *node = head->first;
 1055
 1056	while (node != WAKE_Q_TAIL) {
 1057		struct task_struct *task;
 1058
 1059		task = container_of(node, struct task_struct, wake_q);
 1060		/* Task can safely be re-inserted now: */
 1061		node = node->next;
 1062		task->wake_q.next = NULL;
 1063
 1064		/*
 1065		 * wake_up_process() executes a full barrier, which pairs with
 1066		 * the queueing in wake_q_add() so as not to miss wakeups.
 1067		 */
 1068		wake_up_process(task);
 1069		put_task_struct(task);
 1070	}
 1071}
 1072
 1073/*
 1074 * resched_curr - mark rq's current task 'to be rescheduled now'.
 1075 *
 1076 * On UP this means the setting of the need_resched flag, on SMP it
 1077 * might also involve a cross-CPU call to trigger the scheduler on
 1078 * the target CPU.
 1079 */
 1080static void __resched_curr(struct rq *rq, int tif)
 1081{
 1082	struct task_struct *curr = rq->curr;
 1083	struct thread_info *cti = task_thread_info(curr);
 1084	int cpu;
 1085
 1086	lockdep_assert_rq_held(rq);
 1087
 1088	/*
 1089	 * Always immediately preempt the idle task; no point in delaying doing
 1090	 * actual work.
 1091	 */
 1092	if (is_idle_task(curr) && tif == TIF_NEED_RESCHED_LAZY)
 1093		tif = TIF_NEED_RESCHED;
 1094
 1095	if (cti->flags & ((1 << tif) | _TIF_NEED_RESCHED))
 1096		return;
 1097
 1098	cpu = cpu_of(rq);
 1099
 1100	if (cpu == smp_processor_id()) {
 1101		set_ti_thread_flag(cti, tif);
 1102		if (tif == TIF_NEED_RESCHED)
 1103			set_preempt_need_resched();
 1104		return;
 1105	}
 1106
 1107	if (set_nr_and_not_polling(cti, tif)) {
 1108		if (tif == TIF_NEED_RESCHED)
 1109			smp_send_reschedule(cpu);
 1110	} else {
 1111		trace_sched_wake_idle_without_ipi(cpu);
 1112	}
 1113}
 1114
 1115void resched_curr(struct rq *rq)
 1116{
 1117	__resched_curr(rq, TIF_NEED_RESCHED);
 1118}
 1119
 1120#ifdef CONFIG_PREEMPT_DYNAMIC
 1121static DEFINE_STATIC_KEY_FALSE(sk_dynamic_preempt_lazy);
 1122static __always_inline bool dynamic_preempt_lazy(void)
 1123{
 1124	return static_branch_unlikely(&sk_dynamic_preempt_lazy);
 1125}
 1126#else
 1127static __always_inline bool dynamic_preempt_lazy(void)
 1128{
 1129	return IS_ENABLED(CONFIG_PREEMPT_LAZY);
 1130}
 1131#endif
 1132
 1133static __always_inline int get_lazy_tif_bit(void)
 1134{
 1135	if (dynamic_preempt_lazy())
 1136		return TIF_NEED_RESCHED_LAZY;
 1137
 1138	return TIF_NEED_RESCHED;
 1139}
 1140
 1141void resched_curr_lazy(struct rq *rq)
 1142{
 1143	__resched_curr(rq, get_lazy_tif_bit());
 1144}
 1145
 1146void resched_cpu(int cpu)
 1147{
 1148	struct rq *rq = cpu_rq(cpu);
 1149	unsigned long flags;
 1150
 1151	raw_spin_rq_lock_irqsave(rq, flags);
 1152	if (cpu_online(cpu) || cpu == smp_processor_id())
 1153		resched_curr(rq);
 1154	raw_spin_rq_unlock_irqrestore(rq, flags);
 1155}
 1156
 1157#ifdef CONFIG_SMP
 1158#ifdef CONFIG_NO_HZ_COMMON
 1159/*
 1160 * In the semi idle case, use the nearest busy CPU for migrating timers
 1161 * from an idle CPU.  This is good for power-savings.
 1162 *
 1163 * We don't do similar optimization for completely idle system, as
 1164 * selecting an idle CPU will add more delays to the timers than intended
 1165 * (as that CPU's timer base may not be up to date wrt jiffies etc).
 1166 */
 1167int get_nohz_timer_target(void)
 1168{
 1169	int i, cpu = smp_processor_id(), default_cpu = -1;
 1170	struct sched_domain *sd;
 1171	const struct cpumask *hk_mask;
 1172
 1173	if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
 1174		if (!idle_cpu(cpu))
 1175			return cpu;
 1176		default_cpu = cpu;
 1177	}
 1178
 1179	hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
 1180
 1181	guard(rcu)();
 1182
 1183	for_each_domain(cpu, sd) {
 1184		for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
 
 1185			if (cpu == i)
 1186				continue;
 1187
 1188			if (!idle_cpu(i))
 1189				return i;
 
 
 1190		}
 1191	}
 1192
 1193	if (default_cpu == -1)
 1194		default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
 1195
 1196	return default_cpu;
 
 
 1197}
 1198
 1199/*
 1200 * When add_timer_on() enqueues a timer into the timer wheel of an
 1201 * idle CPU then this timer might expire before the next timer event
 1202 * which is scheduled to wake up that CPU. In case of a completely
 1203 * idle system the next event might even be infinite time into the
 1204 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
 1205 * leaves the inner idle loop so the newly added timer is taken into
 1206 * account when the CPU goes back to idle and evaluates the timer
 1207 * wheel for the next timer event.
 1208 */
 1209static void wake_up_idle_cpu(int cpu)
 1210{
 1211	struct rq *rq = cpu_rq(cpu);
 1212
 1213	if (cpu == smp_processor_id())
 1214		return;
 1215
 1216	/*
 1217	 * Set TIF_NEED_RESCHED and send an IPI if in the non-polling
 1218	 * part of the idle loop. This forces an exit from the idle loop
 1219	 * and a round trip to schedule(). Now this could be optimized
 1220	 * because a simple new idle loop iteration is enough to
 1221	 * re-evaluate the next tick. Provided some re-ordering of tick
 1222	 * nohz functions that would need to follow TIF_NR_POLLING
 1223	 * clearing:
 1224	 *
 1225	 * - On most architectures, a simple fetch_or on ti::flags with a
 1226	 *   "0" value would be enough to know if an IPI needs to be sent.
 1227	 *
 1228	 * - x86 needs to perform a last need_resched() check between
 1229	 *   monitor and mwait which doesn't take timers into account.
 1230	 *   There a dedicated TIF_TIMER flag would be required to
 1231	 *   fetch_or here and be checked along with TIF_NEED_RESCHED
 1232	 *   before mwait().
 1233	 *
 1234	 * However, remote timer enqueue is not such a frequent event
 1235	 * and testing of the above solutions didn't appear to report
 1236	 * much benefits.
 1237	 */
 1238	if (set_nr_and_not_polling(task_thread_info(rq->idle), TIF_NEED_RESCHED))
 1239		smp_send_reschedule(cpu);
 1240	else
 1241		trace_sched_wake_idle_without_ipi(cpu);
 1242}
 1243
 1244static bool wake_up_full_nohz_cpu(int cpu)
 1245{
 1246	/*
 1247	 * We just need the target to call irq_exit() and re-evaluate
 1248	 * the next tick. The nohz full kick at least implies that.
 1249	 * If needed we can still optimize that later with an
 1250	 * empty IRQ.
 1251	 */
 1252	if (cpu_is_offline(cpu))
 1253		return true;  /* Don't try to wake offline CPUs. */
 1254	if (tick_nohz_full_cpu(cpu)) {
 1255		if (cpu != smp_processor_id() ||
 1256		    tick_nohz_tick_stopped())
 1257			tick_nohz_full_kick_cpu(cpu);
 1258		return true;
 1259	}
 1260
 1261	return false;
 1262}
 1263
 1264/*
 1265 * Wake up the specified CPU.  If the CPU is going offline, it is the
 1266 * caller's responsibility to deal with the lost wakeup, for example,
 1267 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
 1268 */
 1269void wake_up_nohz_cpu(int cpu)
 1270{
 1271	if (!wake_up_full_nohz_cpu(cpu))
 1272		wake_up_idle_cpu(cpu);
 1273}
 1274
 1275static void nohz_csd_func(void *info)
 1276{
 1277	struct rq *rq = info;
 1278	int cpu = cpu_of(rq);
 1279	unsigned int flags;
 1280
 1281	/*
 1282	 * Release the rq::nohz_csd.
 1283	 */
 1284	flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
 1285	WARN_ON(!(flags & NOHZ_KICK_MASK));
 1286
 1287	rq->idle_balance = idle_cpu(cpu);
 1288	if (rq->idle_balance) {
 1289		rq->nohz_idle_balance = flags;
 1290		__raise_softirq_irqoff(SCHED_SOFTIRQ);
 1291	}
 1292}
 1293
 1294#endif /* CONFIG_NO_HZ_COMMON */
 1295
 1296#ifdef CONFIG_NO_HZ_FULL
 1297static inline bool __need_bw_check(struct rq *rq, struct task_struct *p)
 1298{
 1299	if (rq->nr_running != 1)
 1300		return false;
 1301
 1302	if (p->sched_class != &fair_sched_class)
 1303		return false;
 1304
 1305	if (!task_on_rq_queued(p))
 1306		return false;
 1307
 1308	return true;
 1309}
 1310
 1311bool sched_can_stop_tick(struct rq *rq)
 1312{
 1313	int fifo_nr_running;
 1314
 1315	/* Deadline tasks, even if single, need the tick */
 1316	if (rq->dl.dl_nr_running)
 1317		return false;
 1318
 1319	/*
 1320	 * If there are more than one RR tasks, we need the tick to affect the
 1321	 * actual RR behaviour.
 1322	 */
 1323	if (rq->rt.rr_nr_running) {
 1324		if (rq->rt.rr_nr_running == 1)
 1325			return true;
 1326		else
 1327			return false;
 1328	}
 1329
 1330	/*
 1331	 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
 1332	 * forced preemption between FIFO tasks.
 1333	 */
 1334	fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
 1335	if (fifo_nr_running)
 1336		return true;
 1337
 1338	/*
 1339	 * If there are no DL,RR/FIFO tasks, there must only be CFS or SCX tasks
 1340	 * left. For CFS, if there's more than one we need the tick for
 1341	 * involuntary preemption. For SCX, ask.
 1342	 */
 1343	if (scx_enabled() && !scx_can_stop_tick(rq))
 1344		return false;
 1345
 1346	if (rq->cfs.h_nr_running > 1)
 1347		return false;
 1348
 1349	/*
 1350	 * If there is one task and it has CFS runtime bandwidth constraints
 1351	 * and it's on the cpu now we don't want to stop the tick.
 1352	 * This check prevents clearing the bit if a newly enqueued task here is
 1353	 * dequeued by migrating while the constrained task continues to run.
 1354	 * E.g. going from 2->1 without going through pick_next_task().
 1355	 */
 1356	if (__need_bw_check(rq, rq->curr)) {
 1357		if (cfs_task_bw_constrained(rq->curr))
 1358			return false;
 1359	}
 1360
 1361	return true;
 1362}
 1363#endif /* CONFIG_NO_HZ_FULL */
 1364#endif /* CONFIG_SMP */
 1365
 1366#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
 1367			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
 1368/*
 1369 * Iterate task_group tree rooted at *from, calling @down when first entering a
 1370 * node and @up when leaving it for the final time.
 1371 *
 1372 * Caller must hold rcu_lock or sufficient equivalent.
 1373 */
 1374int walk_tg_tree_from(struct task_group *from,
 1375			     tg_visitor down, tg_visitor up, void *data)
 1376{
 1377	struct task_group *parent, *child;
 1378	int ret;
 1379
 1380	parent = from;
 1381
 1382down:
 1383	ret = (*down)(parent, data);
 1384	if (ret)
 1385		goto out;
 1386	list_for_each_entry_rcu(child, &parent->children, siblings) {
 1387		parent = child;
 1388		goto down;
 1389
 1390up:
 1391		continue;
 1392	}
 1393	ret = (*up)(parent, data);
 1394	if (ret || parent == from)
 1395		goto out;
 1396
 1397	child = parent;
 1398	parent = parent->parent;
 1399	if (parent)
 1400		goto up;
 1401out:
 1402	return ret;
 1403}
 1404
 1405int tg_nop(struct task_group *tg, void *data)
 1406{
 1407	return 0;
 1408}
 1409#endif
 1410
 1411void set_load_weight(struct task_struct *p, bool update_load)
 1412{
 1413	int prio = p->static_prio - MAX_RT_PRIO;
 1414	struct load_weight lw;
 1415
 
 
 
 1416	if (task_has_idle_policy(p)) {
 1417		lw.weight = scale_load(WEIGHT_IDLEPRIO);
 1418		lw.inv_weight = WMULT_IDLEPRIO;
 1419	} else {
 1420		lw.weight = scale_load(sched_prio_to_weight[prio]);
 1421		lw.inv_weight = sched_prio_to_wmult[prio];
 1422	}
 1423
 1424	/*
 1425	 * SCHED_OTHER tasks have to update their load when changing their
 1426	 * weight
 1427	 */
 1428	if (update_load && p->sched_class->reweight_task)
 1429		p->sched_class->reweight_task(task_rq(p), p, &lw);
 1430	else
 1431		p->se.load = lw;
 
 
 1432}
 1433
 1434#ifdef CONFIG_UCLAMP_TASK
 1435/*
 1436 * Serializes updates of utilization clamp values
 1437 *
 1438 * The (slow-path) user-space triggers utilization clamp value updates which
 1439 * can require updates on (fast-path) scheduler's data structures used to
 1440 * support enqueue/dequeue operations.
 1441 * While the per-CPU rq lock protects fast-path update operations, user-space
 1442 * requests are serialized using a mutex to reduce the risk of conflicting
 1443 * updates or API abuses.
 1444 */
 1445static __maybe_unused DEFINE_MUTEX(uclamp_mutex);
 1446
 1447/* Max allowed minimum utilization */
 1448static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
 1449
 1450/* Max allowed maximum utilization */
 1451static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
 1452
 1453/*
 1454 * By default RT tasks run at the maximum performance point/capacity of the
 1455 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
 1456 * SCHED_CAPACITY_SCALE.
 1457 *
 1458 * This knob allows admins to change the default behavior when uclamp is being
 1459 * used. In battery powered devices, particularly, running at the maximum
 1460 * capacity and frequency will increase energy consumption and shorten the
 1461 * battery life.
 1462 *
 1463 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
 1464 *
 1465 * This knob will not override the system default sched_util_clamp_min defined
 1466 * above.
 1467 */
 1468unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
 1469
 1470/* All clamps are required to be less or equal than these values */
 1471static struct uclamp_se uclamp_default[UCLAMP_CNT];
 1472
 1473/*
 1474 * This static key is used to reduce the uclamp overhead in the fast path. It
 1475 * primarily disables the call to uclamp_rq_{inc, dec}() in
 1476 * enqueue/dequeue_task().
 1477 *
 1478 * This allows users to continue to enable uclamp in their kernel config with
 1479 * minimum uclamp overhead in the fast path.
 1480 *
 1481 * As soon as userspace modifies any of the uclamp knobs, the static key is
 1482 * enabled, since we have an actual users that make use of uclamp
 1483 * functionality.
 1484 *
 1485 * The knobs that would enable this static key are:
 1486 *
 1487 *   * A task modifying its uclamp value with sched_setattr().
 1488 *   * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
 1489 *   * An admin modifying the cgroup cpu.uclamp.{min, max}
 1490 */
 1491DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
 1492
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 1493static inline unsigned int
 1494uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
 1495		  unsigned int clamp_value)
 1496{
 1497	/*
 1498	 * Avoid blocked utilization pushing up the frequency when we go
 1499	 * idle (which drops the max-clamp) by retaining the last known
 1500	 * max-clamp.
 1501	 */
 1502	if (clamp_id == UCLAMP_MAX) {
 1503		rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
 1504		return clamp_value;
 1505	}
 1506
 1507	return uclamp_none(UCLAMP_MIN);
 1508}
 1509
 1510static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
 1511				     unsigned int clamp_value)
 1512{
 1513	/* Reset max-clamp retention only on idle exit */
 1514	if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
 1515		return;
 1516
 1517	uclamp_rq_set(rq, clamp_id, clamp_value);
 1518}
 1519
 1520static inline
 1521unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
 1522				   unsigned int clamp_value)
 1523{
 1524	struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
 1525	int bucket_id = UCLAMP_BUCKETS - 1;
 1526
 1527	/*
 1528	 * Since both min and max clamps are max aggregated, find the
 1529	 * top most bucket with tasks in.
 1530	 */
 1531	for ( ; bucket_id >= 0; bucket_id--) {
 1532		if (!bucket[bucket_id].tasks)
 1533			continue;
 1534		return bucket[bucket_id].value;
 1535	}
 1536
 1537	/* No tasks -- default clamp values */
 1538	return uclamp_idle_value(rq, clamp_id, clamp_value);
 1539}
 1540
 1541static void __uclamp_update_util_min_rt_default(struct task_struct *p)
 1542{
 1543	unsigned int default_util_min;
 1544	struct uclamp_se *uc_se;
 1545
 1546	lockdep_assert_held(&p->pi_lock);
 1547
 1548	uc_se = &p->uclamp_req[UCLAMP_MIN];
 1549
 1550	/* Only sync if user didn't override the default */
 1551	if (uc_se->user_defined)
 1552		return;
 1553
 1554	default_util_min = sysctl_sched_uclamp_util_min_rt_default;
 1555	uclamp_se_set(uc_se, default_util_min, false);
 1556}
 1557
 1558static void uclamp_update_util_min_rt_default(struct task_struct *p)
 1559{
 
 
 
 1560	if (!rt_task(p))
 1561		return;
 1562
 1563	/* Protect updates to p->uclamp_* */
 1564	guard(task_rq_lock)(p);
 1565	__uclamp_update_util_min_rt_default(p);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 1566}
 1567
 1568static inline struct uclamp_se
 1569uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
 1570{
 1571	/* Copy by value as we could modify it */
 1572	struct uclamp_se uc_req = p->uclamp_req[clamp_id];
 1573#ifdef CONFIG_UCLAMP_TASK_GROUP
 1574	unsigned int tg_min, tg_max, value;
 1575
 1576	/*
 1577	 * Tasks in autogroups or root task group will be
 1578	 * restricted by system defaults.
 1579	 */
 1580	if (task_group_is_autogroup(task_group(p)))
 1581		return uc_req;
 1582	if (task_group(p) == &root_task_group)
 1583		return uc_req;
 1584
 1585	tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
 1586	tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
 1587	value = uc_req.value;
 1588	value = clamp(value, tg_min, tg_max);
 1589	uclamp_se_set(&uc_req, value, false);
 1590#endif
 1591
 1592	return uc_req;
 1593}
 1594
 1595/*
 1596 * The effective clamp bucket index of a task depends on, by increasing
 1597 * priority:
 1598 * - the task specific clamp value, when explicitly requested from userspace
 1599 * - the task group effective clamp value, for tasks not either in the root
 1600 *   group or in an autogroup
 1601 * - the system default clamp value, defined by the sysadmin
 1602 */
 1603static inline struct uclamp_se
 1604uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
 1605{
 1606	struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
 1607	struct uclamp_se uc_max = uclamp_default[clamp_id];
 1608
 1609	/* System default restrictions always apply */
 1610	if (unlikely(uc_req.value > uc_max.value))
 1611		return uc_max;
 1612
 1613	return uc_req;
 1614}
 1615
 1616unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
 1617{
 1618	struct uclamp_se uc_eff;
 1619
 1620	/* Task currently refcounted: use back-annotated (effective) value */
 1621	if (p->uclamp[clamp_id].active)
 1622		return (unsigned long)p->uclamp[clamp_id].value;
 1623
 1624	uc_eff = uclamp_eff_get(p, clamp_id);
 1625
 1626	return (unsigned long)uc_eff.value;
 1627}
 1628
 1629/*
 1630 * When a task is enqueued on a rq, the clamp bucket currently defined by the
 1631 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
 1632 * updates the rq's clamp value if required.
 1633 *
 1634 * Tasks can have a task-specific value requested from user-space, track
 1635 * within each bucket the maximum value for tasks refcounted in it.
 1636 * This "local max aggregation" allows to track the exact "requested" value
 1637 * for each bucket when all its RUNNABLE tasks require the same clamp.
 1638 */
 1639static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
 1640				    enum uclamp_id clamp_id)
 1641{
 1642	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
 1643	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
 1644	struct uclamp_bucket *bucket;
 1645
 1646	lockdep_assert_rq_held(rq);
 1647
 1648	/* Update task effective clamp */
 1649	p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
 1650
 1651	bucket = &uc_rq->bucket[uc_se->bucket_id];
 1652	bucket->tasks++;
 1653	uc_se->active = true;
 1654
 1655	uclamp_idle_reset(rq, clamp_id, uc_se->value);
 1656
 1657	/*
 1658	 * Local max aggregation: rq buckets always track the max
 1659	 * "requested" clamp value of its RUNNABLE tasks.
 1660	 */
 1661	if (bucket->tasks == 1 || uc_se->value > bucket->value)
 1662		bucket->value = uc_se->value;
 1663
 1664	if (uc_se->value > uclamp_rq_get(rq, clamp_id))
 1665		uclamp_rq_set(rq, clamp_id, uc_se->value);
 1666}
 1667
 1668/*
 1669 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
 1670 * is released. If this is the last task reference counting the rq's max
 1671 * active clamp value, then the rq's clamp value is updated.
 1672 *
 1673 * Both refcounted tasks and rq's cached clamp values are expected to be
 1674 * always valid. If it's detected they are not, as defensive programming,
 1675 * enforce the expected state and warn.
 1676 */
 1677static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
 1678				    enum uclamp_id clamp_id)
 1679{
 1680	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
 1681	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
 1682	struct uclamp_bucket *bucket;
 1683	unsigned int bkt_clamp;
 1684	unsigned int rq_clamp;
 1685
 1686	lockdep_assert_rq_held(rq);
 1687
 1688	/*
 1689	 * If sched_uclamp_used was enabled after task @p was enqueued,
 1690	 * we could end up with unbalanced call to uclamp_rq_dec_id().
 1691	 *
 1692	 * In this case the uc_se->active flag should be false since no uclamp
 1693	 * accounting was performed at enqueue time and we can just return
 1694	 * here.
 1695	 *
 1696	 * Need to be careful of the following enqueue/dequeue ordering
 1697	 * problem too
 1698	 *
 1699	 *	enqueue(taskA)
 1700	 *	// sched_uclamp_used gets enabled
 1701	 *	enqueue(taskB)
 1702	 *	dequeue(taskA)
 1703	 *	// Must not decrement bucket->tasks here
 1704	 *	dequeue(taskB)
 1705	 *
 1706	 * where we could end up with stale data in uc_se and
 1707	 * bucket[uc_se->bucket_id].
 1708	 *
 1709	 * The following check here eliminates the possibility of such race.
 1710	 */
 1711	if (unlikely(!uc_se->active))
 1712		return;
 1713
 1714	bucket = &uc_rq->bucket[uc_se->bucket_id];
 1715
 1716	SCHED_WARN_ON(!bucket->tasks);
 1717	if (likely(bucket->tasks))
 1718		bucket->tasks--;
 1719
 1720	uc_se->active = false;
 1721
 1722	/*
 1723	 * Keep "local max aggregation" simple and accept to (possibly)
 1724	 * overboost some RUNNABLE tasks in the same bucket.
 1725	 * The rq clamp bucket value is reset to its base value whenever
 1726	 * there are no more RUNNABLE tasks refcounting it.
 1727	 */
 1728	if (likely(bucket->tasks))
 1729		return;
 1730
 1731	rq_clamp = uclamp_rq_get(rq, clamp_id);
 1732	/*
 1733	 * Defensive programming: this should never happen. If it happens,
 1734	 * e.g. due to future modification, warn and fix up the expected value.
 1735	 */
 1736	SCHED_WARN_ON(bucket->value > rq_clamp);
 1737	if (bucket->value >= rq_clamp) {
 1738		bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
 1739		uclamp_rq_set(rq, clamp_id, bkt_clamp);
 1740	}
 1741}
 1742
 1743static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
 1744{
 1745	enum uclamp_id clamp_id;
 1746
 1747	/*
 1748	 * Avoid any overhead until uclamp is actually used by the userspace.
 1749	 *
 1750	 * The condition is constructed such that a NOP is generated when
 1751	 * sched_uclamp_used is disabled.
 1752	 */
 1753	if (!static_branch_unlikely(&sched_uclamp_used))
 1754		return;
 1755
 1756	if (unlikely(!p->sched_class->uclamp_enabled))
 1757		return;
 1758
 1759	if (p->se.sched_delayed)
 1760		return;
 1761
 1762	for_each_clamp_id(clamp_id)
 1763		uclamp_rq_inc_id(rq, p, clamp_id);
 1764
 1765	/* Reset clamp idle holding when there is one RUNNABLE task */
 1766	if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
 1767		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
 1768}
 1769
 1770static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
 1771{
 1772	enum uclamp_id clamp_id;
 1773
 1774	/*
 1775	 * Avoid any overhead until uclamp is actually used by the userspace.
 1776	 *
 1777	 * The condition is constructed such that a NOP is generated when
 1778	 * sched_uclamp_used is disabled.
 1779	 */
 1780	if (!static_branch_unlikely(&sched_uclamp_used))
 1781		return;
 1782
 1783	if (unlikely(!p->sched_class->uclamp_enabled))
 1784		return;
 1785
 1786	if (p->se.sched_delayed)
 1787		return;
 1788
 1789	for_each_clamp_id(clamp_id)
 1790		uclamp_rq_dec_id(rq, p, clamp_id);
 1791}
 1792
 1793static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
 1794				      enum uclamp_id clamp_id)
 1795{
 1796	if (!p->uclamp[clamp_id].active)
 1797		return;
 1798
 1799	uclamp_rq_dec_id(rq, p, clamp_id);
 1800	uclamp_rq_inc_id(rq, p, clamp_id);
 1801
 1802	/*
 1803	 * Make sure to clear the idle flag if we've transiently reached 0
 1804	 * active tasks on rq.
 1805	 */
 1806	if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
 1807		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
 1808}
 1809
 1810static inline void
 1811uclamp_update_active(struct task_struct *p)
 1812{
 1813	enum uclamp_id clamp_id;
 1814	struct rq_flags rf;
 1815	struct rq *rq;
 1816
 1817	/*
 1818	 * Lock the task and the rq where the task is (or was) queued.
 1819	 *
 1820	 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
 1821	 * price to pay to safely serialize util_{min,max} updates with
 1822	 * enqueues, dequeues and migration operations.
 1823	 * This is the same locking schema used by __set_cpus_allowed_ptr().
 1824	 */
 1825	rq = task_rq_lock(p, &rf);
 1826
 1827	/*
 1828	 * Setting the clamp bucket is serialized by task_rq_lock().
 1829	 * If the task is not yet RUNNABLE and its task_struct is not
 1830	 * affecting a valid clamp bucket, the next time it's enqueued,
 1831	 * it will already see the updated clamp bucket value.
 1832	 */
 1833	for_each_clamp_id(clamp_id)
 1834		uclamp_rq_reinc_id(rq, p, clamp_id);
 1835
 1836	task_rq_unlock(rq, p, &rf);
 1837}
 1838
 1839#ifdef CONFIG_UCLAMP_TASK_GROUP
 1840static inline void
 1841uclamp_update_active_tasks(struct cgroup_subsys_state *css)
 1842{
 1843	struct css_task_iter it;
 1844	struct task_struct *p;
 1845
 1846	css_task_iter_start(css, 0, &it);
 1847	while ((p = css_task_iter_next(&it)))
 1848		uclamp_update_active(p);
 1849	css_task_iter_end(&it);
 1850}
 1851
 1852static void cpu_util_update_eff(struct cgroup_subsys_state *css);
 1853#endif
 1854
 1855#ifdef CONFIG_SYSCTL
 1856#ifdef CONFIG_UCLAMP_TASK_GROUP
 1857static void uclamp_update_root_tg(void)
 1858{
 1859	struct task_group *tg = &root_task_group;
 1860
 1861	uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
 1862		      sysctl_sched_uclamp_util_min, false);
 1863	uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
 1864		      sysctl_sched_uclamp_util_max, false);
 1865
 1866	guard(rcu)();
 1867	cpu_util_update_eff(&root_task_group.css);
 
 1868}
 1869#else
 1870static void uclamp_update_root_tg(void) { }
 1871#endif
 1872
 1873static void uclamp_sync_util_min_rt_default(void)
 1874{
 1875	struct task_struct *g, *p;
 1876
 1877	/*
 1878	 * copy_process()			sysctl_uclamp
 1879	 *					  uclamp_min_rt = X;
 1880	 *   write_lock(&tasklist_lock)		  read_lock(&tasklist_lock)
 1881	 *   // link thread			  smp_mb__after_spinlock()
 1882	 *   write_unlock(&tasklist_lock)	  read_unlock(&tasklist_lock);
 1883	 *   sched_post_fork()			  for_each_process_thread()
 1884	 *     __uclamp_sync_rt()		    __uclamp_sync_rt()
 1885	 *
 1886	 * Ensures that either sched_post_fork() will observe the new
 1887	 * uclamp_min_rt or for_each_process_thread() will observe the new
 1888	 * task.
 1889	 */
 1890	read_lock(&tasklist_lock);
 1891	smp_mb__after_spinlock();
 1892	read_unlock(&tasklist_lock);
 1893
 1894	guard(rcu)();
 1895	for_each_process_thread(g, p)
 1896		uclamp_update_util_min_rt_default(p);
 1897}
 1898
 1899static int sysctl_sched_uclamp_handler(const struct ctl_table *table, int write,
 1900				void *buffer, size_t *lenp, loff_t *ppos)
 1901{
 1902	bool update_root_tg = false;
 1903	int old_min, old_max, old_min_rt;
 1904	int result;
 1905
 1906	guard(mutex)(&uclamp_mutex);
 1907
 1908	old_min = sysctl_sched_uclamp_util_min;
 1909	old_max = sysctl_sched_uclamp_util_max;
 1910	old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
 1911
 1912	result = proc_dointvec(table, write, buffer, lenp, ppos);
 1913	if (result)
 1914		goto undo;
 1915	if (!write)
 1916		return 0;
 1917
 1918	if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
 1919	    sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE	||
 1920	    sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
 1921
 1922		result = -EINVAL;
 1923		goto undo;
 1924	}
 1925
 1926	if (old_min != sysctl_sched_uclamp_util_min) {
 1927		uclamp_se_set(&uclamp_default[UCLAMP_MIN],
 1928			      sysctl_sched_uclamp_util_min, false);
 1929		update_root_tg = true;
 1930	}
 1931	if (old_max != sysctl_sched_uclamp_util_max) {
 1932		uclamp_se_set(&uclamp_default[UCLAMP_MAX],
 1933			      sysctl_sched_uclamp_util_max, false);
 1934		update_root_tg = true;
 1935	}
 1936
 1937	if (update_root_tg) {
 1938		static_branch_enable(&sched_uclamp_used);
 1939		uclamp_update_root_tg();
 1940	}
 1941
 1942	if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
 1943		static_branch_enable(&sched_uclamp_used);
 1944		uclamp_sync_util_min_rt_default();
 1945	}
 1946
 1947	/*
 1948	 * We update all RUNNABLE tasks only when task groups are in use.
 1949	 * Otherwise, keep it simple and do just a lazy update at each next
 1950	 * task enqueue time.
 1951	 */
 1952	return 0;
 
 1953
 1954undo:
 1955	sysctl_sched_uclamp_util_min = old_min;
 1956	sysctl_sched_uclamp_util_max = old_max;
 1957	sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
 
 
 
 1958	return result;
 1959}
 1960#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 1961
 1962static void uclamp_fork(struct task_struct *p)
 1963{
 1964	enum uclamp_id clamp_id;
 1965
 1966	/*
 1967	 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
 1968	 * as the task is still at its early fork stages.
 1969	 */
 1970	for_each_clamp_id(clamp_id)
 1971		p->uclamp[clamp_id].active = false;
 1972
 1973	if (likely(!p->sched_reset_on_fork))
 1974		return;
 1975
 1976	for_each_clamp_id(clamp_id) {
 1977		uclamp_se_set(&p->uclamp_req[clamp_id],
 1978			      uclamp_none(clamp_id), false);
 1979	}
 1980}
 1981
 1982static void uclamp_post_fork(struct task_struct *p)
 1983{
 1984	uclamp_update_util_min_rt_default(p);
 1985}
 1986
 1987static void __init init_uclamp_rq(struct rq *rq)
 1988{
 1989	enum uclamp_id clamp_id;
 1990	struct uclamp_rq *uc_rq = rq->uclamp;
 1991
 1992	for_each_clamp_id(clamp_id) {
 1993		uc_rq[clamp_id] = (struct uclamp_rq) {
 1994			.value = uclamp_none(clamp_id)
 1995		};
 1996	}
 1997
 1998	rq->uclamp_flags = UCLAMP_FLAG_IDLE;
 1999}
 2000
 2001static void __init init_uclamp(void)
 2002{
 2003	struct uclamp_se uc_max = {};
 2004	enum uclamp_id clamp_id;
 2005	int cpu;
 2006
 2007	for_each_possible_cpu(cpu)
 2008		init_uclamp_rq(cpu_rq(cpu));
 2009
 2010	for_each_clamp_id(clamp_id) {
 2011		uclamp_se_set(&init_task.uclamp_req[clamp_id],
 2012			      uclamp_none(clamp_id), false);
 2013	}
 2014
 2015	/* System defaults allow max clamp values for both indexes */
 2016	uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
 2017	for_each_clamp_id(clamp_id) {
 2018		uclamp_default[clamp_id] = uc_max;
 2019#ifdef CONFIG_UCLAMP_TASK_GROUP
 2020		root_task_group.uclamp_req[clamp_id] = uc_max;
 2021		root_task_group.uclamp[clamp_id] = uc_max;
 2022#endif
 2023	}
 2024}
 2025
 2026#else /* !CONFIG_UCLAMP_TASK */
 2027static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
 2028static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
 
 
 
 
 
 
 
 2029static inline void uclamp_fork(struct task_struct *p) { }
 2030static inline void uclamp_post_fork(struct task_struct *p) { }
 2031static inline void init_uclamp(void) { }
 2032#endif /* CONFIG_UCLAMP_TASK */
 2033
 2034bool sched_task_on_rq(struct task_struct *p)
 2035{
 2036	return task_on_rq_queued(p);
 2037}
 2038
 2039unsigned long get_wchan(struct task_struct *p)
 2040{
 2041	unsigned long ip = 0;
 2042	unsigned int state;
 2043
 2044	if (!p || p == current)
 2045		return 0;
 2046
 2047	/* Only get wchan if task is blocked and we can keep it that way. */
 2048	raw_spin_lock_irq(&p->pi_lock);
 2049	state = READ_ONCE(p->__state);
 2050	smp_rmb(); /* see try_to_wake_up() */
 2051	if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
 2052		ip = __get_wchan(p);
 2053	raw_spin_unlock_irq(&p->pi_lock);
 2054
 2055	return ip;
 2056}
 2057
 2058void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
 2059{
 2060	if (!(flags & ENQUEUE_NOCLOCK))
 2061		update_rq_clock(rq);
 2062
 2063	p->sched_class->enqueue_task(rq, p, flags);
 2064	/*
 2065	 * Must be after ->enqueue_task() because ENQUEUE_DELAYED can clear
 2066	 * ->sched_delayed.
 2067	 */
 2068	uclamp_rq_inc(rq, p);
 2069
 2070	psi_enqueue(p, flags);
 2071
 2072	if (!(flags & ENQUEUE_RESTORE))
 2073		sched_info_enqueue(rq, p);
 
 
 
 
 
 2074
 2075	if (sched_core_enabled(rq))
 2076		sched_core_enqueue(rq, p);
 2077}
 2078
 2079/*
 2080 * Must only return false when DEQUEUE_SLEEP.
 2081 */
 2082inline bool dequeue_task(struct rq *rq, struct task_struct *p, int flags)
 2083{
 2084	if (sched_core_enabled(rq))
 2085		sched_core_dequeue(rq, p, flags);
 2086
 2087	if (!(flags & DEQUEUE_NOCLOCK))
 2088		update_rq_clock(rq);
 2089
 2090	if (!(flags & DEQUEUE_SAVE))
 2091		sched_info_dequeue(rq, p);
 
 
 2092
 2093	psi_dequeue(p, flags);
 2094
 2095	/*
 2096	 * Must be before ->dequeue_task() because ->dequeue_task() can 'fail'
 2097	 * and mark the task ->sched_delayed.
 2098	 */
 2099	uclamp_rq_dec(rq, p);
 2100	return p->sched_class->dequeue_task(rq, p, flags);
 2101}
 2102
 2103void activate_task(struct rq *rq, struct task_struct *p, int flags)
 2104{
 2105	if (task_on_rq_migrating(p))
 2106		flags |= ENQUEUE_MIGRATED;
 2107	if (flags & ENQUEUE_MIGRATED)
 2108		sched_mm_cid_migrate_to(rq, p);
 2109
 2110	enqueue_task(rq, p, flags);
 2111
 2112	WRITE_ONCE(p->on_rq, TASK_ON_RQ_QUEUED);
 2113	ASSERT_EXCLUSIVE_WRITER(p->on_rq);
 2114}
 2115
 2116void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
 2117{
 2118	SCHED_WARN_ON(flags & DEQUEUE_SLEEP);
 2119
 2120	WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
 2121	ASSERT_EXCLUSIVE_WRITER(p->on_rq);
 2122
 2123	/*
 2124	 * Code explicitly relies on TASK_ON_RQ_MIGRATING begin set *before*
 2125	 * dequeue_task() and cleared *after* enqueue_task().
 2126	 */
 2127
 2128	dequeue_task(rq, p, flags);
 
 
 
 
 
 
 
 2129}
 2130
 2131static void block_task(struct rq *rq, struct task_struct *p, int flags)
 
 
 
 
 
 
 
 2132{
 2133	if (dequeue_task(rq, p, DEQUEUE_SLEEP | flags))
 2134		__block_task(rq, p);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 2135}
 2136
 2137/**
 2138 * task_curr - is this task currently executing on a CPU?
 2139 * @p: the task in question.
 2140 *
 2141 * Return: 1 if the task is currently executing. 0 otherwise.
 2142 */
 2143inline int task_curr(const struct task_struct *p)
 2144{
 2145	return cpu_curr(task_cpu(p)) == p;
 2146}
 2147
 2148/*
 2149 * ->switching_to() is called with the pi_lock and rq_lock held and must not
 2150 * mess with locking.
 2151 */
 2152void check_class_changing(struct rq *rq, struct task_struct *p,
 2153			  const struct sched_class *prev_class)
 2154{
 2155	if (prev_class != p->sched_class && p->sched_class->switching_to)
 2156		p->sched_class->switching_to(rq, p);
 2157}
 2158
 2159/*
 2160 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
 2161 * use the balance_callback list if you want balancing.
 2162 *
 2163 * this means any call to check_class_changed() must be followed by a call to
 2164 * balance_callback().
 2165 */
 2166void check_class_changed(struct rq *rq, struct task_struct *p,
 2167			 const struct sched_class *prev_class,
 2168			 int oldprio)
 2169{
 2170	if (prev_class != p->sched_class) {
 2171		if (prev_class->switched_from)
 2172			prev_class->switched_from(rq, p);
 2173
 2174		p->sched_class->switched_to(rq, p);
 2175	} else if (oldprio != p->prio || dl_task(p))
 2176		p->sched_class->prio_changed(rq, p, oldprio);
 2177}
 2178
 2179void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags)
 2180{
 2181	struct task_struct *donor = rq->donor;
 2182
 2183	if (p->sched_class == donor->sched_class)
 2184		donor->sched_class->wakeup_preempt(rq, p, flags);
 2185	else if (sched_class_above(p->sched_class, donor->sched_class))
 2186		resched_curr(rq);
 2187
 2188	/*
 2189	 * A queue event has occurred, and we're going to schedule.  In
 2190	 * this case, we can save a useless back to back clock update.
 2191	 */
 2192	if (task_on_rq_queued(donor) && test_tsk_need_resched(rq->curr))
 2193		rq_clock_skip_update(rq);
 2194}
 2195
 2196static __always_inline
 2197int __task_state_match(struct task_struct *p, unsigned int state)
 2198{
 2199	if (READ_ONCE(p->__state) & state)
 2200		return 1;
 2201
 2202	if (READ_ONCE(p->saved_state) & state)
 2203		return -1;
 2204
 2205	return 0;
 2206}
 2207
 2208static __always_inline
 2209int task_state_match(struct task_struct *p, unsigned int state)
 2210{
 2211	/*
 2212	 * Serialize against current_save_and_set_rtlock_wait_state(),
 2213	 * current_restore_rtlock_saved_state(), and __refrigerator().
 2214	 */
 2215	guard(raw_spinlock_irq)(&p->pi_lock);
 2216	return __task_state_match(p, state);
 2217}
 2218
 2219/*
 2220 * wait_task_inactive - wait for a thread to unschedule.
 2221 *
 2222 * Wait for the thread to block in any of the states set in @match_state.
 2223 * If it changes, i.e. @p might have woken up, then return zero.  When we
 2224 * succeed in waiting for @p to be off its CPU, we return a positive number
 2225 * (its total switch count).  If a second call a short while later returns the
 2226 * same number, the caller can be sure that @p has remained unscheduled the
 2227 * whole time.
 2228 *
 2229 * The caller must ensure that the task *will* unschedule sometime soon,
 2230 * else this function might spin for a *long* time. This function can't
 2231 * be called with interrupts off, or it may introduce deadlock with
 2232 * smp_call_function() if an IPI is sent by the same process we are
 2233 * waiting to become inactive.
 2234 */
 2235unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
 2236{
 2237	int running, queued, match;
 2238	struct rq_flags rf;
 2239	unsigned long ncsw;
 2240	struct rq *rq;
 2241
 2242	for (;;) {
 2243		/*
 2244		 * We do the initial early heuristics without holding
 2245		 * any task-queue locks at all. We'll only try to get
 2246		 * the runqueue lock when things look like they will
 2247		 * work out!
 2248		 */
 2249		rq = task_rq(p);
 2250
 2251		/*
 2252		 * If the task is actively running on another CPU
 2253		 * still, just relax and busy-wait without holding
 2254		 * any locks.
 2255		 *
 2256		 * NOTE! Since we don't hold any locks, it's not
 2257		 * even sure that "rq" stays as the right runqueue!
 2258		 * But we don't care, since "task_on_cpu()" will
 2259		 * return false if the runqueue has changed and p
 2260		 * is actually now running somewhere else!
 2261		 */
 2262		while (task_on_cpu(rq, p)) {
 2263			if (!task_state_match(p, match_state))
 2264				return 0;
 2265			cpu_relax();
 2266		}
 2267
 2268		/*
 2269		 * Ok, time to look more closely! We need the rq
 2270		 * lock now, to be *sure*. If we're wrong, we'll
 2271		 * just go back and repeat.
 2272		 */
 2273		rq = task_rq_lock(p, &rf);
 2274		trace_sched_wait_task(p);
 2275		running = task_on_cpu(rq, p);
 2276		queued = task_on_rq_queued(p);
 2277		ncsw = 0;
 2278		if ((match = __task_state_match(p, match_state))) {
 2279			/*
 2280			 * When matching on p->saved_state, consider this task
 2281			 * still queued so it will wait.
 2282			 */
 2283			if (match < 0)
 2284				queued = 1;
 2285			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
 2286		}
 2287		task_rq_unlock(rq, p, &rf);
 2288
 2289		/*
 2290		 * If it changed from the expected state, bail out now.
 2291		 */
 2292		if (unlikely(!ncsw))
 2293			break;
 2294
 2295		/*
 2296		 * Was it really running after all now that we
 2297		 * checked with the proper locks actually held?
 2298		 *
 2299		 * Oops. Go back and try again..
 2300		 */
 2301		if (unlikely(running)) {
 2302			cpu_relax();
 2303			continue;
 2304		}
 2305
 2306		/*
 2307		 * It's not enough that it's not actively running,
 2308		 * it must be off the runqueue _entirely_, and not
 2309		 * preempted!
 2310		 *
 2311		 * So if it was still runnable (but just not actively
 2312		 * running right now), it's preempted, and we should
 2313		 * yield - it could be a while.
 2314		 */
 2315		if (unlikely(queued)) {
 2316			ktime_t to = NSEC_PER_SEC / HZ;
 2317
 2318			set_current_state(TASK_UNINTERRUPTIBLE);
 2319			schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
 2320			continue;
 2321		}
 2322
 2323		/*
 2324		 * Ahh, all good. It wasn't running, and it wasn't
 2325		 * runnable, which means that it will never become
 2326		 * running in the future either. We're all done!
 2327		 */
 2328		break;
 2329	}
 2330
 2331	return ncsw;
 2332}
 2333
 2334#ifdef CONFIG_SMP
 2335
 2336static void
 2337__do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
 
 
 
 
 2338
 2339static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
 2340{
 2341	struct affinity_context ac = {
 2342		.new_mask  = cpumask_of(rq->cpu),
 2343		.flags     = SCA_MIGRATE_DISABLE,
 2344	};
 2345
 2346	if (likely(!p->migration_disabled))
 2347		return;
 2348
 2349	if (p->cpus_ptr != &p->cpus_mask)
 2350		return;
 2351
 2352	/*
 2353	 * Violates locking rules! See comment in __do_set_cpus_allowed().
 2354	 */
 2355	__do_set_cpus_allowed(p, &ac);
 2356}
 2357
 2358void migrate_disable(void)
 2359{
 2360	struct task_struct *p = current;
 2361
 2362	if (p->migration_disabled) {
 2363#ifdef CONFIG_DEBUG_PREEMPT
 2364		/*
 2365		 *Warn about overflow half-way through the range.
 2366		 */
 2367		WARN_ON_ONCE((s16)p->migration_disabled < 0);
 2368#endif
 2369		p->migration_disabled++;
 2370		return;
 2371	}
 2372
 2373	guard(preempt)();
 2374	this_rq()->nr_pinned++;
 2375	p->migration_disabled = 1;
 
 2376}
 2377EXPORT_SYMBOL_GPL(migrate_disable);
 2378
 2379void migrate_enable(void)
 2380{
 2381	struct task_struct *p = current;
 2382	struct affinity_context ac = {
 2383		.new_mask  = &p->cpus_mask,
 2384		.flags     = SCA_MIGRATE_ENABLE,
 2385	};
 2386
 2387#ifdef CONFIG_DEBUG_PREEMPT
 2388	/*
 2389	 * Check both overflow from migrate_disable() and superfluous
 2390	 * migrate_enable().
 2391	 */
 2392	if (WARN_ON_ONCE((s16)p->migration_disabled <= 0))
 2393		return;
 2394#endif
 2395
 2396	if (p->migration_disabled > 1) {
 2397		p->migration_disabled--;
 2398		return;
 2399	}
 2400
 2401	/*
 2402	 * Ensure stop_task runs either before or after this, and that
 2403	 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
 2404	 */
 2405	guard(preempt)();
 2406	if (p->cpus_ptr != &p->cpus_mask)
 2407		__set_cpus_allowed_ptr(p, &ac);
 2408	/*
 2409	 * Mustn't clear migration_disabled() until cpus_ptr points back at the
 2410	 * regular cpus_mask, otherwise things that race (eg.
 2411	 * select_fallback_rq) get confused.
 2412	 */
 2413	barrier();
 2414	p->migration_disabled = 0;
 2415	this_rq()->nr_pinned--;
 
 2416}
 2417EXPORT_SYMBOL_GPL(migrate_enable);
 2418
 2419static inline bool rq_has_pinned_tasks(struct rq *rq)
 2420{
 2421	return rq->nr_pinned;
 2422}
 2423
 2424/*
 2425 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
 2426 * __set_cpus_allowed_ptr() and select_fallback_rq().
 2427 */
 2428static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
 2429{
 2430	/* When not in the task's cpumask, no point in looking further. */
 2431	if (!task_allowed_on_cpu(p, cpu))
 2432		return false;
 2433
 2434	/* migrate_disabled() must be allowed to finish. */
 2435	if (is_migration_disabled(p))
 2436		return cpu_online(cpu);
 2437
 2438	/* Non kernel threads are not allowed during either online or offline. */
 2439	if (!(p->flags & PF_KTHREAD))
 2440		return cpu_active(cpu);
 2441
 2442	/* KTHREAD_IS_PER_CPU is always allowed. */
 2443	if (kthread_is_per_cpu(p))
 2444		return cpu_online(cpu);
 2445
 2446	/* Regular kernel threads don't get to stay during offline. */
 2447	if (cpu_dying(cpu))
 2448		return false;
 2449
 2450	/* But are allowed during online. */
 2451	return cpu_online(cpu);
 2452}
 2453
 2454/*
 2455 * This is how migration works:
 2456 *
 2457 * 1) we invoke migration_cpu_stop() on the target CPU using
 2458 *    stop_one_cpu().
 2459 * 2) stopper starts to run (implicitly forcing the migrated thread
 2460 *    off the CPU)
 2461 * 3) it checks whether the migrated task is still in the wrong runqueue.
 2462 * 4) if it's in the wrong runqueue then the migration thread removes
 2463 *    it and puts it into the right queue.
 2464 * 5) stopper completes and stop_one_cpu() returns and the migration
 2465 *    is done.
 2466 */
 2467
 2468/*
 2469 * move_queued_task - move a queued task to new rq.
 2470 *
 2471 * Returns (locked) new rq. Old rq's lock is released.
 2472 */
 2473static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
 2474				   struct task_struct *p, int new_cpu)
 2475{
 2476	lockdep_assert_rq_held(rq);
 2477
 2478	deactivate_task(rq, p, DEQUEUE_NOCLOCK);
 2479	set_task_cpu(p, new_cpu);
 2480	rq_unlock(rq, rf);
 2481
 2482	rq = cpu_rq(new_cpu);
 2483
 2484	rq_lock(rq, rf);
 2485	WARN_ON_ONCE(task_cpu(p) != new_cpu);
 2486	activate_task(rq, p, 0);
 2487	wakeup_preempt(rq, p, 0);
 2488
 2489	return rq;
 2490}
 2491
 2492struct migration_arg {
 2493	struct task_struct		*task;
 2494	int				dest_cpu;
 2495	struct set_affinity_pending	*pending;
 2496};
 2497
 2498/*
 2499 * @refs: number of wait_for_completion()
 2500 * @stop_pending: is @stop_work in use
 2501 */
 2502struct set_affinity_pending {
 2503	refcount_t		refs;
 2504	unsigned int		stop_pending;
 2505	struct completion	done;
 2506	struct cpu_stop_work	stop_work;
 2507	struct migration_arg	arg;
 2508};
 2509
 2510/*
 2511 * Move (not current) task off this CPU, onto the destination CPU. We're doing
 2512 * this because either it can't run here any more (set_cpus_allowed()
 2513 * away from this CPU, or CPU going down), or because we're
 2514 * attempting to rebalance this task on exec (sched_exec).
 2515 *
 2516 * So we race with normal scheduler movements, but that's OK, as long
 2517 * as the task is no longer on this CPU.
 2518 */
 2519static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
 2520				 struct task_struct *p, int dest_cpu)
 2521{
 2522	/* Affinity changed (again). */
 2523	if (!is_cpu_allowed(p, dest_cpu))
 2524		return rq;
 2525
 
 2526	rq = move_queued_task(rq, rf, p, dest_cpu);
 2527
 2528	return rq;
 2529}
 2530
 2531/*
 2532 * migration_cpu_stop - this will be executed by a high-prio stopper thread
 2533 * and performs thread migration by bumping thread off CPU then
 2534 * 'pushing' onto another runqueue.
 2535 */
 2536static int migration_cpu_stop(void *data)
 2537{
 2538	struct migration_arg *arg = data;
 2539	struct set_affinity_pending *pending = arg->pending;
 2540	struct task_struct *p = arg->task;
 2541	struct rq *rq = this_rq();
 2542	bool complete = false;
 2543	struct rq_flags rf;
 2544
 2545	/*
 2546	 * The original target CPU might have gone down and we might
 2547	 * be on another CPU but it doesn't matter.
 2548	 */
 2549	local_irq_save(rf.flags);
 2550	/*
 2551	 * We need to explicitly wake pending tasks before running
 2552	 * __migrate_task() such that we will not miss enforcing cpus_ptr
 2553	 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
 2554	 */
 2555	flush_smp_call_function_queue();
 2556
 2557	raw_spin_lock(&p->pi_lock);
 2558	rq_lock(rq, &rf);
 2559
 2560	/*
 2561	 * If we were passed a pending, then ->stop_pending was set, thus
 2562	 * p->migration_pending must have remained stable.
 2563	 */
 2564	WARN_ON_ONCE(pending && pending != p->migration_pending);
 2565
 2566	/*
 2567	 * If task_rq(p) != rq, it cannot be migrated here, because we're
 2568	 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
 2569	 * we're holding p->pi_lock.
 2570	 */
 2571	if (task_rq(p) == rq) {
 2572		if (is_migration_disabled(p))
 2573			goto out;
 2574
 2575		if (pending) {
 2576			p->migration_pending = NULL;
 2577			complete = true;
 2578
 2579			if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
 2580				goto out;
 2581		}
 2582
 2583		if (task_on_rq_queued(p)) {
 2584			update_rq_clock(rq);
 2585			rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
 2586		} else {
 2587			p->wake_cpu = arg->dest_cpu;
 2588		}
 2589
 2590		/*
 2591		 * XXX __migrate_task() can fail, at which point we might end
 2592		 * up running on a dodgy CPU, AFAICT this can only happen
 2593		 * during CPU hotplug, at which point we'll get pushed out
 2594		 * anyway, so it's probably not a big deal.
 2595		 */
 2596
 2597	} else if (pending) {
 2598		/*
 2599		 * This happens when we get migrated between migrate_enable()'s
 2600		 * preempt_enable() and scheduling the stopper task. At that
 2601		 * point we're a regular task again and not current anymore.
 2602		 *
 2603		 * A !PREEMPT kernel has a giant hole here, which makes it far
 2604		 * more likely.
 2605		 */
 2606
 2607		/*
 2608		 * The task moved before the stopper got to run. We're holding
 2609		 * ->pi_lock, so the allowed mask is stable - if it got
 2610		 * somewhere allowed, we're done.
 2611		 */
 2612		if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
 2613			p->migration_pending = NULL;
 2614			complete = true;
 2615			goto out;
 2616		}
 2617
 2618		/*
 2619		 * When migrate_enable() hits a rq mis-match we can't reliably
 2620		 * determine is_migration_disabled() and so have to chase after
 2621		 * it.
 2622		 */
 2623		WARN_ON_ONCE(!pending->stop_pending);
 2624		preempt_disable();
 2625		task_rq_unlock(rq, p, &rf);
 2626		stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
 2627				    &pending->arg, &pending->stop_work);
 2628		preempt_enable();
 2629		return 0;
 2630	}
 2631out:
 2632	if (pending)
 2633		pending->stop_pending = false;
 2634	task_rq_unlock(rq, p, &rf);
 2635
 2636	if (complete)
 2637		complete_all(&pending->done);
 2638
 2639	return 0;
 2640}
 2641
 2642int push_cpu_stop(void *arg)
 2643{
 2644	struct rq *lowest_rq = NULL, *rq = this_rq();
 2645	struct task_struct *p = arg;
 2646
 2647	raw_spin_lock_irq(&p->pi_lock);
 2648	raw_spin_rq_lock(rq);
 2649
 2650	if (task_rq(p) != rq)
 2651		goto out_unlock;
 2652
 2653	if (is_migration_disabled(p)) {
 2654		p->migration_flags |= MDF_PUSH;
 2655		goto out_unlock;
 2656	}
 2657
 2658	p->migration_flags &= ~MDF_PUSH;
 2659
 2660	if (p->sched_class->find_lock_rq)
 2661		lowest_rq = p->sched_class->find_lock_rq(p, rq);
 2662
 2663	if (!lowest_rq)
 2664		goto out_unlock;
 2665
 2666	// XXX validate p is still the highest prio task
 2667	if (task_rq(p) == rq) {
 2668		move_queued_task_locked(rq, lowest_rq, p);
 
 
 2669		resched_curr(lowest_rq);
 2670	}
 2671
 2672	double_unlock_balance(rq, lowest_rq);
 2673
 2674out_unlock:
 2675	rq->push_busy = false;
 2676	raw_spin_rq_unlock(rq);
 2677	raw_spin_unlock_irq(&p->pi_lock);
 2678
 2679	put_task_struct(p);
 2680	return 0;
 2681}
 2682
 2683/*
 2684 * sched_class::set_cpus_allowed must do the below, but is not required to
 2685 * actually call this function.
 2686 */
 2687void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
 2688{
 2689	if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
 2690		p->cpus_ptr = ctx->new_mask;
 2691		return;
 2692	}
 2693
 2694	cpumask_copy(&p->cpus_mask, ctx->new_mask);
 2695	p->nr_cpus_allowed = cpumask_weight(ctx->new_mask);
 2696
 2697	/*
 2698	 * Swap in a new user_cpus_ptr if SCA_USER flag set
 2699	 */
 2700	if (ctx->flags & SCA_USER)
 2701		swap(p->user_cpus_ptr, ctx->user_mask);
 2702}
 2703
 2704static void
 2705__do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
 2706{
 2707	struct rq *rq = task_rq(p);
 2708	bool queued, running;
 2709
 2710	/*
 2711	 * This here violates the locking rules for affinity, since we're only
 2712	 * supposed to change these variables while holding both rq->lock and
 2713	 * p->pi_lock.
 2714	 *
 2715	 * HOWEVER, it magically works, because ttwu() is the only code that
 2716	 * accesses these variables under p->pi_lock and only does so after
 2717	 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
 2718	 * before finish_task().
 2719	 *
 2720	 * XXX do further audits, this smells like something putrid.
 2721	 */
 2722	if (ctx->flags & SCA_MIGRATE_DISABLE)
 2723		SCHED_WARN_ON(!p->on_cpu);
 2724	else
 2725		lockdep_assert_held(&p->pi_lock);
 2726
 2727	queued = task_on_rq_queued(p);
 2728	running = task_current_donor(rq, p);
 2729
 2730	if (queued) {
 2731		/*
 2732		 * Because __kthread_bind() calls this on blocked tasks without
 2733		 * holding rq->lock.
 2734		 */
 2735		lockdep_assert_rq_held(rq);
 2736		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
 2737	}
 2738	if (running)
 2739		put_prev_task(rq, p);
 2740
 2741	p->sched_class->set_cpus_allowed(p, ctx);
 2742	mm_set_cpus_allowed(p->mm, ctx->new_mask);
 2743
 2744	if (queued)
 2745		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
 2746	if (running)
 2747		set_next_task(rq, p);
 2748}
 2749
 2750/*
 2751 * Used for kthread_bind() and select_fallback_rq(), in both cases the user
 2752 * affinity (if any) should be destroyed too.
 2753 */
 2754void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
 2755{
 2756	struct affinity_context ac = {
 2757		.new_mask  = new_mask,
 2758		.user_mask = NULL,
 2759		.flags     = SCA_USER,	/* clear the user requested mask */
 2760	};
 2761	union cpumask_rcuhead {
 2762		cpumask_t cpumask;
 2763		struct rcu_head rcu;
 2764	};
 2765
 2766	__do_set_cpus_allowed(p, &ac);
 2767
 2768	/*
 2769	 * Because this is called with p->pi_lock held, it is not possible
 2770	 * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
 2771	 * kfree_rcu().
 2772	 */
 2773	kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
 2774}
 2775
 2776int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
 2777		      int node)
 2778{
 2779	cpumask_t *user_mask;
 2780	unsigned long flags;
 2781
 2782	/*
 2783	 * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
 2784	 * may differ by now due to racing.
 2785	 */
 2786	dst->user_cpus_ptr = NULL;
 2787
 2788	/*
 2789	 * This check is racy and losing the race is a valid situation.
 2790	 * It is not worth the extra overhead of taking the pi_lock on
 2791	 * every fork/clone.
 2792	 */
 2793	if (data_race(!src->user_cpus_ptr))
 2794		return 0;
 2795
 2796	user_mask = alloc_user_cpus_ptr(node);
 2797	if (!user_mask)
 2798		return -ENOMEM;
 2799
 2800	/*
 2801	 * Use pi_lock to protect content of user_cpus_ptr
 2802	 *
 2803	 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
 2804	 * do_set_cpus_allowed().
 2805	 */
 2806	raw_spin_lock_irqsave(&src->pi_lock, flags);
 2807	if (src->user_cpus_ptr) {
 2808		swap(dst->user_cpus_ptr, user_mask);
 2809		cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
 2810	}
 2811	raw_spin_unlock_irqrestore(&src->pi_lock, flags);
 2812
 2813	if (unlikely(user_mask))
 2814		kfree(user_mask);
 2815
 2816	return 0;
 2817}
 2818
 2819static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
 2820{
 2821	struct cpumask *user_mask = NULL;
 2822
 2823	swap(p->user_cpus_ptr, user_mask);
 2824
 2825	return user_mask;
 2826}
 2827
 2828void release_user_cpus_ptr(struct task_struct *p)
 2829{
 2830	kfree(clear_user_cpus_ptr(p));
 2831}
 2832
 2833/*
 2834 * This function is wildly self concurrent; here be dragons.
 2835 *
 2836 *
 2837 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
 2838 * designated task is enqueued on an allowed CPU. If that task is currently
 2839 * running, we have to kick it out using the CPU stopper.
 2840 *
 2841 * Migrate-Disable comes along and tramples all over our nice sandcastle.
 2842 * Consider:
 2843 *
 2844 *     Initial conditions: P0->cpus_mask = [0, 1]
 2845 *
 2846 *     P0@CPU0                  P1
 2847 *
 2848 *     migrate_disable();
 2849 *     <preempted>
 2850 *                              set_cpus_allowed_ptr(P0, [1]);
 2851 *
 2852 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
 2853 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
 2854 * This means we need the following scheme:
 2855 *
 2856 *     P0@CPU0                  P1
 2857 *
 2858 *     migrate_disable();
 2859 *     <preempted>
 2860 *                              set_cpus_allowed_ptr(P0, [1]);
 2861 *                                <blocks>
 2862 *     <resumes>
 2863 *     migrate_enable();
 2864 *       __set_cpus_allowed_ptr();
 2865 *       <wakes local stopper>
 2866 *                         `--> <woken on migration completion>
 2867 *
 2868 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
 2869 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
 2870 * task p are serialized by p->pi_lock, which we can leverage: the one that
 2871 * should come into effect at the end of the Migrate-Disable region is the last
 2872 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
 2873 * but we still need to properly signal those waiting tasks at the appropriate
 2874 * moment.
 2875 *
 2876 * This is implemented using struct set_affinity_pending. The first
 2877 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
 2878 * setup an instance of that struct and install it on the targeted task_struct.
 2879 * Any and all further callers will reuse that instance. Those then wait for
 2880 * a completion signaled at the tail of the CPU stopper callback (1), triggered
 2881 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
 2882 *
 2883 *
 2884 * (1) In the cases covered above. There is one more where the completion is
 2885 * signaled within affine_move_task() itself: when a subsequent affinity request
 2886 * occurs after the stopper bailed out due to the targeted task still being
 2887 * Migrate-Disable. Consider:
 2888 *
 2889 *     Initial conditions: P0->cpus_mask = [0, 1]
 2890 *
 2891 *     CPU0		  P1				P2
 2892 *     <P0>
 2893 *       migrate_disable();
 2894 *       <preempted>
 2895 *                        set_cpus_allowed_ptr(P0, [1]);
 2896 *                          <blocks>
 2897 *     <migration/0>
 2898 *       migration_cpu_stop()
 2899 *         is_migration_disabled()
 2900 *           <bails>
 2901 *                                                       set_cpus_allowed_ptr(P0, [0, 1]);
 2902 *                                                         <signal completion>
 2903 *                          <awakes>
 2904 *
 2905 * Note that the above is safe vs a concurrent migrate_enable(), as any
 2906 * pending affinity completion is preceded by an uninstallation of
 2907 * p->migration_pending done with p->pi_lock held.
 2908 */
 2909static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
 2910			    int dest_cpu, unsigned int flags)
 2911	__releases(rq->lock)
 2912	__releases(p->pi_lock)
 2913{
 2914	struct set_affinity_pending my_pending = { }, *pending = NULL;
 2915	bool stop_pending, complete = false;
 2916
 2917	/* Can the task run on the task's current CPU? If so, we're done */
 2918	if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
 2919		struct task_struct *push_task = NULL;
 2920
 2921		if ((flags & SCA_MIGRATE_ENABLE) &&
 2922		    (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
 2923			rq->push_busy = true;
 2924			push_task = get_task_struct(p);
 2925		}
 2926
 2927		/*
 2928		 * If there are pending waiters, but no pending stop_work,
 2929		 * then complete now.
 2930		 */
 2931		pending = p->migration_pending;
 2932		if (pending && !pending->stop_pending) {
 2933			p->migration_pending = NULL;
 2934			complete = true;
 2935		}
 2936
 2937		preempt_disable();
 2938		task_rq_unlock(rq, p, rf);
 
 2939		if (push_task) {
 2940			stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
 2941					    p, &rq->push_work);
 2942		}
 2943		preempt_enable();
 2944
 2945		if (complete)
 2946			complete_all(&pending->done);
 2947
 2948		return 0;
 2949	}
 2950
 2951	if (!(flags & SCA_MIGRATE_ENABLE)) {
 2952		/* serialized by p->pi_lock */
 2953		if (!p->migration_pending) {
 2954			/* Install the request */
 2955			refcount_set(&my_pending.refs, 1);
 2956			init_completion(&my_pending.done);
 2957			my_pending.arg = (struct migration_arg) {
 2958				.task = p,
 2959				.dest_cpu = dest_cpu,
 2960				.pending = &my_pending,
 2961			};
 2962
 2963			p->migration_pending = &my_pending;
 2964		} else {
 2965			pending = p->migration_pending;
 2966			refcount_inc(&pending->refs);
 2967			/*
 2968			 * Affinity has changed, but we've already installed a
 2969			 * pending. migration_cpu_stop() *must* see this, else
 2970			 * we risk a completion of the pending despite having a
 2971			 * task on a disallowed CPU.
 2972			 *
 2973			 * Serialized by p->pi_lock, so this is safe.
 2974			 */
 2975			pending->arg.dest_cpu = dest_cpu;
 2976		}
 2977	}
 2978	pending = p->migration_pending;
 2979	/*
 2980	 * - !MIGRATE_ENABLE:
 2981	 *   we'll have installed a pending if there wasn't one already.
 2982	 *
 2983	 * - MIGRATE_ENABLE:
 2984	 *   we're here because the current CPU isn't matching anymore,
 2985	 *   the only way that can happen is because of a concurrent
 2986	 *   set_cpus_allowed_ptr() call, which should then still be
 2987	 *   pending completion.
 2988	 *
 2989	 * Either way, we really should have a @pending here.
 2990	 */
 2991	if (WARN_ON_ONCE(!pending)) {
 2992		task_rq_unlock(rq, p, rf);
 2993		return -EINVAL;
 2994	}
 2995
 2996	if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
 2997		/*
 2998		 * MIGRATE_ENABLE gets here because 'p == current', but for
 2999		 * anything else we cannot do is_migration_disabled(), punt
 3000		 * and have the stopper function handle it all race-free.
 3001		 */
 3002		stop_pending = pending->stop_pending;
 3003		if (!stop_pending)
 3004			pending->stop_pending = true;
 3005
 3006		if (flags & SCA_MIGRATE_ENABLE)
 3007			p->migration_flags &= ~MDF_PUSH;
 3008
 3009		preempt_disable();
 3010		task_rq_unlock(rq, p, rf);
 
 3011		if (!stop_pending) {
 3012			stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
 3013					    &pending->arg, &pending->stop_work);
 3014		}
 3015		preempt_enable();
 3016
 3017		if (flags & SCA_MIGRATE_ENABLE)
 3018			return 0;
 3019	} else {
 3020
 3021		if (!is_migration_disabled(p)) {
 3022			if (task_on_rq_queued(p))
 3023				rq = move_queued_task(rq, rf, p, dest_cpu);
 3024
 3025			if (!pending->stop_pending) {
 3026				p->migration_pending = NULL;
 3027				complete = true;
 3028			}
 3029		}
 3030		task_rq_unlock(rq, p, rf);
 3031
 3032		if (complete)
 3033			complete_all(&pending->done);
 3034	}
 3035
 3036	wait_for_completion(&pending->done);
 3037
 3038	if (refcount_dec_and_test(&pending->refs))
 3039		wake_up_var(&pending->refs); /* No UaF, just an address */
 3040
 3041	/*
 3042	 * Block the original owner of &pending until all subsequent callers
 3043	 * have seen the completion and decremented the refcount
 3044	 */
 3045	wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
 3046
 3047	/* ARGH */
 3048	WARN_ON_ONCE(my_pending.stop_pending);
 3049
 3050	return 0;
 3051}
 3052
 3053/*
 3054 * Called with both p->pi_lock and rq->lock held; drops both before returning.
 
 
 
 
 
 
 3055 */
 3056static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
 3057					 struct affinity_context *ctx,
 3058					 struct rq *rq,
 3059					 struct rq_flags *rf)
 3060	__releases(rq->lock)
 3061	__releases(p->pi_lock)
 3062{
 3063	const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
 3064	const struct cpumask *cpu_valid_mask = cpu_active_mask;
 3065	bool kthread = p->flags & PF_KTHREAD;
 3066	unsigned int dest_cpu;
 
 
 3067	int ret = 0;
 3068
 
 3069	update_rq_clock(rq);
 3070
 3071	if (kthread || is_migration_disabled(p)) {
 3072		/*
 3073		 * Kernel threads are allowed on online && !active CPUs,
 3074		 * however, during cpu-hot-unplug, even these might get pushed
 3075		 * away if not KTHREAD_IS_PER_CPU.
 3076		 *
 3077		 * Specifically, migration_disabled() tasks must not fail the
 3078		 * cpumask_any_and_distribute() pick below, esp. so on
 3079		 * SCA_MIGRATE_ENABLE, otherwise we'll not call
 3080		 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
 3081		 */
 3082		cpu_valid_mask = cpu_online_mask;
 3083	}
 3084
 3085	if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) {
 3086		ret = -EINVAL;
 3087		goto out;
 3088	}
 3089
 3090	/*
 3091	 * Must re-check here, to close a race against __kthread_bind(),
 3092	 * sched_setaffinity() is not guaranteed to observe the flag.
 3093	 */
 3094	if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
 3095		ret = -EINVAL;
 3096		goto out;
 3097	}
 3098
 3099	if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
 3100		if (cpumask_equal(&p->cpus_mask, ctx->new_mask)) {
 3101			if (ctx->flags & SCA_USER)
 3102				swap(p->user_cpus_ptr, ctx->user_mask);
 3103			goto out;
 3104		}
 3105
 3106		if (WARN_ON_ONCE(p == current &&
 3107				 is_migration_disabled(p) &&
 3108				 !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
 3109			ret = -EBUSY;
 3110			goto out;
 3111		}
 3112	}
 3113
 3114	/*
 3115	 * Picking a ~random cpu helps in cases where we are changing affinity
 3116	 * for groups of tasks (ie. cpuset), so that load balancing is not
 3117	 * immediately required to distribute the tasks within their new mask.
 3118	 */
 3119	dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
 3120	if (dest_cpu >= nr_cpu_ids) {
 3121		ret = -EINVAL;
 3122		goto out;
 3123	}
 3124
 3125	__do_set_cpus_allowed(p, ctx);
 3126
 3127	return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
 3128
 3129out:
 3130	task_rq_unlock(rq, p, rf);
 3131
 3132	return ret;
 3133}
 3134
 3135/*
 3136 * Change a given task's CPU affinity. Migrate the thread to a
 3137 * proper CPU and schedule it away if the CPU it's executing on
 3138 * is removed from the allowed bitmask.
 3139 *
 3140 * NOTE: the caller must have a valid reference to the task, the
 3141 * task must not exit() & deallocate itself prematurely. The
 3142 * call is not atomic; no spinlocks may be held.
 3143 */
 3144int __set_cpus_allowed_ptr(struct task_struct *p, struct affinity_context *ctx)
 3145{
 3146	struct rq_flags rf;
 3147	struct rq *rq;
 3148
 3149	rq = task_rq_lock(p, &rf);
 3150	/*
 3151	 * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
 3152	 * flags are set.
 3153	 */
 3154	if (p->user_cpus_ptr &&
 3155	    !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
 3156	    cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr))
 3157		ctx->new_mask = rq->scratch_mask;
 3158
 3159	return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf);
 3160}
 3161
 3162int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
 3163{
 3164	struct affinity_context ac = {
 3165		.new_mask  = new_mask,
 3166		.flags     = 0,
 3167	};
 3168
 3169	return __set_cpus_allowed_ptr(p, &ac);
 3170}
 3171EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
 3172
 3173/*
 3174 * Change a given task's CPU affinity to the intersection of its current
 3175 * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
 3176 * If user_cpus_ptr is defined, use it as the basis for restricting CPU
 3177 * affinity or use cpu_online_mask instead.
 3178 *
 3179 * If the resulting mask is empty, leave the affinity unchanged and return
 3180 * -EINVAL.
 3181 */
 3182static int restrict_cpus_allowed_ptr(struct task_struct *p,
 3183				     struct cpumask *new_mask,
 3184				     const struct cpumask *subset_mask)
 3185{
 3186	struct affinity_context ac = {
 3187		.new_mask  = new_mask,
 3188		.flags     = 0,
 3189	};
 3190	struct rq_flags rf;
 3191	struct rq *rq;
 3192	int err;
 3193
 3194	rq = task_rq_lock(p, &rf);
 3195
 3196	/*
 3197	 * Forcefully restricting the affinity of a deadline task is
 3198	 * likely to cause problems, so fail and noisily override the
 3199	 * mask entirely.
 3200	 */
 3201	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
 3202		err = -EPERM;
 3203		goto err_unlock;
 3204	}
 3205
 3206	if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) {
 3207		err = -EINVAL;
 3208		goto err_unlock;
 3209	}
 3210
 3211	return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf);
 3212
 3213err_unlock:
 3214	task_rq_unlock(rq, p, &rf);
 3215	return err;
 3216}
 3217
 3218/*
 3219 * Restrict the CPU affinity of task @p so that it is a subset of
 3220 * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
 3221 * old affinity mask. If the resulting mask is empty, we warn and walk
 3222 * up the cpuset hierarchy until we find a suitable mask.
 3223 */
 3224void force_compatible_cpus_allowed_ptr(struct task_struct *p)
 3225{
 3226	cpumask_var_t new_mask;
 3227	const struct cpumask *override_mask = task_cpu_possible_mask(p);
 3228
 3229	alloc_cpumask_var(&new_mask, GFP_KERNEL);
 3230
 3231	/*
 3232	 * __migrate_task() can fail silently in the face of concurrent
 3233	 * offlining of the chosen destination CPU, so take the hotplug
 3234	 * lock to ensure that the migration succeeds.
 3235	 */
 3236	cpus_read_lock();
 3237	if (!cpumask_available(new_mask))
 3238		goto out_set_mask;
 3239
 3240	if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
 3241		goto out_free_mask;
 3242
 3243	/*
 3244	 * We failed to find a valid subset of the affinity mask for the
 3245	 * task, so override it based on its cpuset hierarchy.
 3246	 */
 3247	cpuset_cpus_allowed(p, new_mask);
 3248	override_mask = new_mask;
 3249
 3250out_set_mask:
 3251	if (printk_ratelimit()) {
 3252		printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
 3253				task_pid_nr(p), p->comm,
 3254				cpumask_pr_args(override_mask));
 3255	}
 3256
 3257	WARN_ON(set_cpus_allowed_ptr(p, override_mask));
 3258out_free_mask:
 3259	cpus_read_unlock();
 3260	free_cpumask_var(new_mask);
 3261}
 3262
 3263/*
 3264 * Restore the affinity of a task @p which was previously restricted by a
 3265 * call to force_compatible_cpus_allowed_ptr().
 3266 *
 3267 * It is the caller's responsibility to serialise this with any calls to
 3268 * force_compatible_cpus_allowed_ptr(@p).
 3269 */
 3270void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
 3271{
 3272	struct affinity_context ac = {
 3273		.new_mask  = task_user_cpus(p),
 3274		.flags     = 0,
 3275	};
 3276	int ret;
 3277
 3278	/*
 3279	 * Try to restore the old affinity mask with __sched_setaffinity().
 3280	 * Cpuset masking will be done there too.
 3281	 */
 3282	ret = __sched_setaffinity(p, &ac);
 3283	WARN_ON_ONCE(ret);
 3284}
 3285
 3286void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
 3287{
 3288#ifdef CONFIG_SCHED_DEBUG
 3289	unsigned int state = READ_ONCE(p->__state);
 3290
 3291	/*
 3292	 * We should never call set_task_cpu() on a blocked task,
 3293	 * ttwu() will sort out the placement.
 3294	 */
 3295	WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
 3296
 3297	/*
 3298	 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
 3299	 * because schedstat_wait_{start,end} rebase migrating task's wait_start
 3300	 * time relying on p->on_rq.
 3301	 */
 3302	WARN_ON_ONCE(state == TASK_RUNNING &&
 3303		     p->sched_class == &fair_sched_class &&
 3304		     (p->on_rq && !task_on_rq_migrating(p)));
 3305
 3306#ifdef CONFIG_LOCKDEP
 3307	/*
 3308	 * The caller should hold either p->pi_lock or rq->lock, when changing
 3309	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
 3310	 *
 3311	 * sched_move_task() holds both and thus holding either pins the cgroup,
 3312	 * see task_group().
 3313	 *
 3314	 * Furthermore, all task_rq users should acquire both locks, see
 3315	 * task_rq_lock().
 3316	 */
 3317	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
 3318				      lockdep_is_held(__rq_lockp(task_rq(p)))));
 3319#endif
 3320	/*
 3321	 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
 3322	 */
 3323	WARN_ON_ONCE(!cpu_online(new_cpu));
 3324
 3325	WARN_ON_ONCE(is_migration_disabled(p));
 3326#endif
 3327
 3328	trace_sched_migrate_task(p, new_cpu);
 3329
 3330	if (task_cpu(p) != new_cpu) {
 3331		if (p->sched_class->migrate_task_rq)
 3332			p->sched_class->migrate_task_rq(p, new_cpu);
 3333		p->se.nr_migrations++;
 3334		rseq_migrate(p);
 3335		sched_mm_cid_migrate_from(p);
 3336		perf_event_task_migrate(p);
 3337	}
 3338
 3339	__set_task_cpu(p, new_cpu);
 3340}
 3341
 3342#ifdef CONFIG_NUMA_BALANCING
 3343static void __migrate_swap_task(struct task_struct *p, int cpu)
 3344{
 3345	if (task_on_rq_queued(p)) {
 3346		struct rq *src_rq, *dst_rq;
 3347		struct rq_flags srf, drf;
 3348
 3349		src_rq = task_rq(p);
 3350		dst_rq = cpu_rq(cpu);
 3351
 3352		rq_pin_lock(src_rq, &srf);
 3353		rq_pin_lock(dst_rq, &drf);
 3354
 3355		move_queued_task_locked(src_rq, dst_rq, p);
 3356		wakeup_preempt(dst_rq, p, 0);
 
 
 3357
 3358		rq_unpin_lock(dst_rq, &drf);
 3359		rq_unpin_lock(src_rq, &srf);
 3360
 3361	} else {
 3362		/*
 3363		 * Task isn't running anymore; make it appear like we migrated
 3364		 * it before it went to sleep. This means on wakeup we make the
 3365		 * previous CPU our target instead of where it really is.
 3366		 */
 3367		p->wake_cpu = cpu;
 3368	}
 3369}
 3370
 3371struct migration_swap_arg {
 3372	struct task_struct *src_task, *dst_task;
 3373	int src_cpu, dst_cpu;
 3374};
 3375
 3376static int migrate_swap_stop(void *data)
 3377{
 3378	struct migration_swap_arg *arg = data;
 3379	struct rq *src_rq, *dst_rq;
 
 3380
 3381	if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
 3382		return -EAGAIN;
 3383
 3384	src_rq = cpu_rq(arg->src_cpu);
 3385	dst_rq = cpu_rq(arg->dst_cpu);
 3386
 3387	guard(double_raw_spinlock)(&arg->src_task->pi_lock, &arg->dst_task->pi_lock);
 3388	guard(double_rq_lock)(src_rq, dst_rq);
 
 3389
 3390	if (task_cpu(arg->dst_task) != arg->dst_cpu)
 3391		return -EAGAIN;
 3392
 3393	if (task_cpu(arg->src_task) != arg->src_cpu)
 3394		return -EAGAIN;
 3395
 3396	if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
 3397		return -EAGAIN;
 3398
 3399	if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
 3400		return -EAGAIN;
 3401
 3402	__migrate_swap_task(arg->src_task, arg->dst_cpu);
 3403	__migrate_swap_task(arg->dst_task, arg->src_cpu);
 3404
 3405	return 0;
 
 
 
 
 
 
 
 3406}
 3407
 3408/*
 3409 * Cross migrate two tasks
 3410 */
 3411int migrate_swap(struct task_struct *cur, struct task_struct *p,
 3412		int target_cpu, int curr_cpu)
 3413{
 3414	struct migration_swap_arg arg;
 3415	int ret = -EINVAL;
 3416
 3417	arg = (struct migration_swap_arg){
 3418		.src_task = cur,
 3419		.src_cpu = curr_cpu,
 3420		.dst_task = p,
 3421		.dst_cpu = target_cpu,
 3422	};
 3423
 3424	if (arg.src_cpu == arg.dst_cpu)
 3425		goto out;
 3426
 3427	/*
 3428	 * These three tests are all lockless; this is OK since all of them
 3429	 * will be re-checked with proper locks held further down the line.
 3430	 */
 3431	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
 3432		goto out;
 3433
 3434	if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
 3435		goto out;
 3436
 3437	if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
 3438		goto out;
 3439
 3440	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
 3441	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
 3442
 3443out:
 3444	return ret;
 3445}
 3446#endif /* CONFIG_NUMA_BALANCING */
 3447
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 3448/***
 3449 * kick_process - kick a running thread to enter/exit the kernel
 3450 * @p: the to-be-kicked thread
 3451 *
 3452 * Cause a process which is running on another CPU to enter
 3453 * kernel-mode, without any delay. (to get signals handled.)
 3454 *
 3455 * NOTE: this function doesn't have to take the runqueue lock,
 3456 * because all it wants to ensure is that the remote task enters
 3457 * the kernel. If the IPI races and the task has been migrated
 3458 * to another CPU then no harm is done and the purpose has been
 3459 * achieved as well.
 3460 */
 3461void kick_process(struct task_struct *p)
 3462{
 3463	guard(preempt)();
 3464	int cpu = task_cpu(p);
 3465
 
 
 3466	if ((cpu != smp_processor_id()) && task_curr(p))
 3467		smp_send_reschedule(cpu);
 
 3468}
 3469EXPORT_SYMBOL_GPL(kick_process);
 3470
 3471/*
 3472 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
 3473 *
 3474 * A few notes on cpu_active vs cpu_online:
 3475 *
 3476 *  - cpu_active must be a subset of cpu_online
 3477 *
 3478 *  - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
 3479 *    see __set_cpus_allowed_ptr(). At this point the newly online
 3480 *    CPU isn't yet part of the sched domains, and balancing will not
 3481 *    see it.
 3482 *
 3483 *  - on CPU-down we clear cpu_active() to mask the sched domains and
 3484 *    avoid the load balancer to place new tasks on the to be removed
 3485 *    CPU. Existing tasks will remain running there and will be taken
 3486 *    off.
 3487 *
 3488 * This means that fallback selection must not select !active CPUs.
 3489 * And can assume that any active CPU must be online. Conversely
 3490 * select_task_rq() below may allow selection of !active CPUs in order
 3491 * to satisfy the above rules.
 3492 */
 3493static int select_fallback_rq(int cpu, struct task_struct *p)
 3494{
 3495	int nid = cpu_to_node(cpu);
 3496	const struct cpumask *nodemask = NULL;
 3497	enum { cpuset, possible, fail } state = cpuset;
 3498	int dest_cpu;
 3499
 3500	/*
 3501	 * If the node that the CPU is on has been offlined, cpu_to_node()
 3502	 * will return -1. There is no CPU on the node, and we should
 3503	 * select the CPU on the other node.
 3504	 */
 3505	if (nid != -1) {
 3506		nodemask = cpumask_of_node(nid);
 3507
 3508		/* Look for allowed, online CPU in same node. */
 3509		for_each_cpu(dest_cpu, nodemask) {
 3510			if (is_cpu_allowed(p, dest_cpu))
 
 
 3511				return dest_cpu;
 3512		}
 3513	}
 3514
 3515	for (;;) {
 3516		/* Any allowed, online CPU? */
 3517		for_each_cpu(dest_cpu, p->cpus_ptr) {
 3518			if (!is_cpu_allowed(p, dest_cpu))
 3519				continue;
 3520
 3521			goto out;
 3522		}
 3523
 3524		/* No more Mr. Nice Guy. */
 3525		switch (state) {
 3526		case cpuset:
 3527			if (cpuset_cpus_allowed_fallback(p)) {
 
 3528				state = possible;
 3529				break;
 3530			}
 3531			fallthrough;
 3532		case possible:
 3533			/*
 3534			 * XXX When called from select_task_rq() we only
 3535			 * hold p->pi_lock and again violate locking order.
 3536			 *
 3537			 * More yuck to audit.
 3538			 */
 3539			do_set_cpus_allowed(p, task_cpu_possible_mask(p));
 3540			state = fail;
 3541			break;
 
 3542		case fail:
 3543			BUG();
 3544			break;
 3545		}
 3546	}
 3547
 3548out:
 3549	if (state != cpuset) {
 3550		/*
 3551		 * Don't tell them about moving exiting tasks or
 3552		 * kernel threads (both mm NULL), since they never
 3553		 * leave kernel.
 3554		 */
 3555		if (p->mm && printk_ratelimit()) {
 3556			printk_deferred("process %d (%s) no longer affine to cpu%d\n",
 3557					task_pid_nr(p), p->comm, cpu);
 3558		}
 3559	}
 3560
 3561	return dest_cpu;
 3562}
 3563
 3564/*
 3565 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
 3566 */
 3567static inline
 3568int select_task_rq(struct task_struct *p, int cpu, int *wake_flags)
 3569{
 3570	lockdep_assert_held(&p->pi_lock);
 3571
 3572	if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p)) {
 3573		cpu = p->sched_class->select_task_rq(p, cpu, *wake_flags);
 3574		*wake_flags |= WF_RQ_SELECTED;
 3575	} else {
 3576		cpu = cpumask_any(p->cpus_ptr);
 3577	}
 3578
 3579	/*
 3580	 * In order not to call set_task_cpu() on a blocking task we need
 3581	 * to rely on ttwu() to place the task on a valid ->cpus_ptr
 3582	 * CPU.
 3583	 *
 3584	 * Since this is common to all placement strategies, this lives here.
 3585	 *
 3586	 * [ this allows ->select_task() to simply return task_cpu(p) and
 3587	 *   not worry about this generic constraint ]
 3588	 */
 3589	if (unlikely(!is_cpu_allowed(p, cpu)))
 3590		cpu = select_fallback_rq(task_cpu(p), p);
 3591
 3592	return cpu;
 3593}
 3594
 3595void sched_set_stop_task(int cpu, struct task_struct *stop)
 3596{
 3597	static struct lock_class_key stop_pi_lock;
 3598	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
 3599	struct task_struct *old_stop = cpu_rq(cpu)->stop;
 3600
 3601	if (stop) {
 3602		/*
 3603		 * Make it appear like a SCHED_FIFO task, its something
 3604		 * userspace knows about and won't get confused about.
 3605		 *
 3606		 * Also, it will make PI more or less work without too
 3607		 * much confusion -- but then, stop work should not
 3608		 * rely on PI working anyway.
 3609		 */
 3610		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
 3611
 3612		stop->sched_class = &stop_sched_class;
 3613
 3614		/*
 3615		 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
 3616		 * adjust the effective priority of a task. As a result,
 3617		 * rt_mutex_setprio() can trigger (RT) balancing operations,
 3618		 * which can then trigger wakeups of the stop thread to push
 3619		 * around the current task.
 3620		 *
 3621		 * The stop task itself will never be part of the PI-chain, it
 3622		 * never blocks, therefore that ->pi_lock recursion is safe.
 3623		 * Tell lockdep about this by placing the stop->pi_lock in its
 3624		 * own class.
 3625		 */
 3626		lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
 3627	}
 3628
 3629	cpu_rq(cpu)->stop = stop;
 3630
 3631	if (old_stop) {
 3632		/*
 3633		 * Reset it back to a normal scheduling class so that
 3634		 * it can die in pieces.
 3635		 */
 3636		old_stop->sched_class = &rt_sched_class;
 3637	}
 3638}
 3639
 3640#else /* CONFIG_SMP */
 3641
 
 
 
 
 
 
 
 3642static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
 3643
 3644static inline bool rq_has_pinned_tasks(struct rq *rq)
 3645{
 3646	return false;
 3647}
 3648
 3649#endif /* !CONFIG_SMP */
 3650
 3651static void
 3652ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
 3653{
 3654	struct rq *rq;
 3655
 3656	if (!schedstat_enabled())
 3657		return;
 3658
 3659	rq = this_rq();
 3660
 3661#ifdef CONFIG_SMP
 3662	if (cpu == rq->cpu) {
 3663		__schedstat_inc(rq->ttwu_local);
 3664		__schedstat_inc(p->stats.nr_wakeups_local);
 3665	} else {
 3666		struct sched_domain *sd;
 3667
 3668		__schedstat_inc(p->stats.nr_wakeups_remote);
 3669
 3670		guard(rcu)();
 3671		for_each_domain(rq->cpu, sd) {
 3672			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
 3673				__schedstat_inc(sd->ttwu_wake_remote);
 3674				break;
 3675			}
 3676		}
 
 3677	}
 3678
 3679	if (wake_flags & WF_MIGRATED)
 3680		__schedstat_inc(p->stats.nr_wakeups_migrate);
 3681#endif /* CONFIG_SMP */
 3682
 3683	__schedstat_inc(rq->ttwu_count);
 3684	__schedstat_inc(p->stats.nr_wakeups);
 3685
 3686	if (wake_flags & WF_SYNC)
 3687		__schedstat_inc(p->stats.nr_wakeups_sync);
 3688}
 3689
 3690/*
 3691 * Mark the task runnable.
 3692 */
 3693static inline void ttwu_do_wakeup(struct task_struct *p)
 
 3694{
 
 3695	WRITE_ONCE(p->__state, TASK_RUNNING);
 3696	trace_sched_wakeup(p);
 3697}
 3698
 3699static void
 3700ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
 3701		 struct rq_flags *rf)
 3702{
 3703	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
 3704
 3705	lockdep_assert_rq_held(rq);
 3706
 3707	if (p->sched_contributes_to_load)
 3708		rq->nr_uninterruptible--;
 3709
 3710#ifdef CONFIG_SMP
 3711	if (wake_flags & WF_RQ_SELECTED)
 3712		en_flags |= ENQUEUE_RQ_SELECTED;
 3713	if (wake_flags & WF_MIGRATED)
 3714		en_flags |= ENQUEUE_MIGRATED;
 3715	else
 3716#endif
 3717	if (p->in_iowait) {
 3718		delayacct_blkio_end(p);
 3719		atomic_dec(&task_rq(p)->nr_iowait);
 3720	}
 3721
 3722	activate_task(rq, p, en_flags);
 3723	wakeup_preempt(rq, p, wake_flags);
 3724
 3725	ttwu_do_wakeup(p);
 3726
 3727#ifdef CONFIG_SMP
 3728	if (p->sched_class->task_woken) {
 3729		/*
 3730		 * Our task @p is fully woken up and running; so it's safe to
 3731		 * drop the rq->lock, hereafter rq is only used for statistics.
 3732		 */
 3733		rq_unpin_lock(rq, rf);
 3734		p->sched_class->task_woken(rq, p);
 3735		rq_repin_lock(rq, rf);
 3736	}
 3737
 3738	if (rq->idle_stamp) {
 3739		u64 delta = rq_clock(rq) - rq->idle_stamp;
 3740		u64 max = 2*rq->max_idle_balance_cost;
 3741
 3742		update_avg(&rq->avg_idle, delta);
 3743
 3744		if (rq->avg_idle > max)
 3745			rq->avg_idle = max;
 3746
 
 
 
 3747		rq->idle_stamp = 0;
 3748	}
 3749#endif
 3750}
 3751
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 3752/*
 3753 * Consider @p being inside a wait loop:
 3754 *
 3755 *   for (;;) {
 3756 *      set_current_state(TASK_UNINTERRUPTIBLE);
 3757 *
 3758 *      if (CONDITION)
 3759 *         break;
 3760 *
 3761 *      schedule();
 3762 *   }
 3763 *   __set_current_state(TASK_RUNNING);
 3764 *
 3765 * between set_current_state() and schedule(). In this case @p is still
 3766 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
 3767 * an atomic manner.
 3768 *
 3769 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
 3770 * then schedule() must still happen and p->state can be changed to
 3771 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
 3772 * need to do a full wakeup with enqueue.
 3773 *
 3774 * Returns: %true when the wakeup is done,
 3775 *          %false otherwise.
 3776 */
 3777static int ttwu_runnable(struct task_struct *p, int wake_flags)
 3778{
 3779	struct rq_flags rf;
 3780	struct rq *rq;
 3781	int ret = 0;
 3782
 3783	rq = __task_rq_lock(p, &rf);
 3784	if (task_on_rq_queued(p)) {
 
 3785		update_rq_clock(rq);
 3786		if (p->se.sched_delayed)
 3787			enqueue_task(rq, p, ENQUEUE_NOCLOCK | ENQUEUE_DELAYED);
 3788		if (!task_on_cpu(rq, p)) {
 3789			/*
 3790			 * When on_rq && !on_cpu the task is preempted, see if
 3791			 * it should preempt the task that is current now.
 3792			 */
 3793			wakeup_preempt(rq, p, wake_flags);
 3794		}
 3795		ttwu_do_wakeup(p);
 3796		ret = 1;
 3797	}
 3798	__task_rq_unlock(rq, &rf);
 3799
 3800	return ret;
 3801}
 3802
 3803#ifdef CONFIG_SMP
 3804void sched_ttwu_pending(void *arg)
 3805{
 3806	struct llist_node *llist = arg;
 3807	struct rq *rq = this_rq();
 3808	struct task_struct *p, *t;
 3809	struct rq_flags rf;
 3810
 3811	if (!llist)
 3812		return;
 3813
 
 
 
 
 
 
 
 3814	rq_lock_irqsave(rq, &rf);
 3815	update_rq_clock(rq);
 3816
 3817	llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
 3818		if (WARN_ON_ONCE(p->on_cpu))
 3819			smp_cond_load_acquire(&p->on_cpu, !VAL);
 3820
 3821		if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
 3822			set_task_cpu(p, cpu_of(rq));
 3823
 3824		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
 3825	}
 3826
 3827	/*
 3828	 * Must be after enqueueing at least once task such that
 3829	 * idle_cpu() does not observe a false-negative -- if it does,
 3830	 * it is possible for select_idle_siblings() to stack a number
 3831	 * of tasks on this CPU during that window.
 3832	 *
 3833	 * It is OK to clear ttwu_pending when another task pending.
 3834	 * We will receive IPI after local IRQ enabled and then enqueue it.
 3835	 * Since now nr_running > 0, idle_cpu() will always get correct result.
 3836	 */
 3837	WRITE_ONCE(rq->ttwu_pending, 0);
 3838	rq_unlock_irqrestore(rq, &rf);
 3839}
 3840
 3841/*
 3842 * Prepare the scene for sending an IPI for a remote smp_call
 3843 *
 3844 * Returns true if the caller can proceed with sending the IPI.
 3845 * Returns false otherwise.
 3846 */
 3847bool call_function_single_prep_ipi(int cpu)
 3848{
 3849	if (set_nr_if_polling(cpu_rq(cpu)->idle)) {
 3850		trace_sched_wake_idle_without_ipi(cpu);
 3851		return false;
 3852	}
 3853
 3854	return true;
 
 
 
 3855}
 3856
 3857/*
 3858 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
 3859 * necessary. The wakee CPU on receipt of the IPI will queue the task
 3860 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
 3861 * of the wakeup instead of the waker.
 3862 */
 3863static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
 3864{
 3865	struct rq *rq = cpu_rq(cpu);
 3866
 3867	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
 3868
 3869	WRITE_ONCE(rq->ttwu_pending, 1);
 3870	__smp_call_single_queue(cpu, &p->wake_entry.llist);
 3871}
 3872
 3873void wake_up_if_idle(int cpu)
 3874{
 3875	struct rq *rq = cpu_rq(cpu);
 
 3876
 3877	guard(rcu)();
 3878	if (is_idle_task(rcu_dereference(rq->curr))) {
 3879		guard(rq_lock_irqsave)(rq);
 3880		if (is_idle_task(rq->curr))
 3881			resched_curr(rq);
 3882	}
 3883}
 3884
 3885bool cpus_equal_capacity(int this_cpu, int that_cpu)
 3886{
 3887	if (!sched_asym_cpucap_active())
 3888		return true;
 3889
 3890	if (this_cpu == that_cpu)
 3891		return true;
 
 
 
 
 
 
 
 3892
 3893	return arch_scale_cpu_capacity(this_cpu) == arch_scale_cpu_capacity(that_cpu);
 
 3894}
 3895
 3896bool cpus_share_cache(int this_cpu, int that_cpu)
 3897{
 3898	if (this_cpu == that_cpu)
 3899		return true;
 3900
 3901	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
 3902}
 3903
 3904/*
 3905 * Whether CPUs are share cache resources, which means LLC on non-cluster
 3906 * machines and LLC tag or L2 on machines with clusters.
 3907 */
 3908bool cpus_share_resources(int this_cpu, int that_cpu)
 3909{
 3910	if (this_cpu == that_cpu)
 3911		return true;
 3912
 3913	return per_cpu(sd_share_id, this_cpu) == per_cpu(sd_share_id, that_cpu);
 3914}
 3915
 3916static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
 3917{
 3918	/*
 3919	 * The BPF scheduler may depend on select_task_rq() being invoked during
 3920	 * wakeups. In addition, @p may end up executing on a different CPU
 3921	 * regardless of what happens in the wakeup path making the ttwu_queue
 3922	 * optimization less meaningful. Skip if on SCX.
 3923	 */
 3924	if (task_on_scx(p))
 3925		return false;
 3926
 3927	/*
 3928	 * Do not complicate things with the async wake_list while the CPU is
 3929	 * in hotplug state.
 3930	 */
 3931	if (!cpu_active(cpu))
 3932		return false;
 3933
 3934	/* Ensure the task will still be allowed to run on the CPU. */
 3935	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
 3936		return false;
 3937
 3938	/*
 3939	 * If the CPU does not share cache, then queue the task on the
 3940	 * remote rqs wakelist to avoid accessing remote data.
 3941	 */
 3942	if (!cpus_share_cache(smp_processor_id(), cpu))
 3943		return true;
 3944
 3945	if (cpu == smp_processor_id())
 3946		return false;
 3947
 3948	/*
 3949	 * If the wakee cpu is idle, or the task is descheduling and the
 3950	 * only running task on the CPU, then use the wakelist to offload
 3951	 * the task activation to the idle (or soon-to-be-idle) CPU as
 3952	 * the current CPU is likely busy. nr_running is checked to
 3953	 * avoid unnecessary task stacking.
 3954	 *
 3955	 * Note that we can only get here with (wakee) p->on_rq=0,
 3956	 * p->on_cpu can be whatever, we've done the dequeue, so
 3957	 * the wakee has been accounted out of ->nr_running.
 3958	 */
 3959	if (!cpu_rq(cpu)->nr_running)
 3960		return true;
 3961
 3962	return false;
 3963}
 3964
 3965static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
 3966{
 3967	if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
 
 
 
 3968		sched_clock_cpu(cpu); /* Sync clocks across CPUs */
 3969		__ttwu_queue_wakelist(p, cpu, wake_flags);
 3970		return true;
 3971	}
 3972
 3973	return false;
 3974}
 3975
 3976#else /* !CONFIG_SMP */
 3977
 3978static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
 3979{
 3980	return false;
 3981}
 3982
 3983#endif /* CONFIG_SMP */
 3984
 3985static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
 3986{
 3987	struct rq *rq = cpu_rq(cpu);
 3988	struct rq_flags rf;
 3989
 3990	if (ttwu_queue_wakelist(p, cpu, wake_flags))
 3991		return;
 3992
 3993	rq_lock(rq, &rf);
 3994	update_rq_clock(rq);
 3995	ttwu_do_activate(rq, p, wake_flags, &rf);
 3996	rq_unlock(rq, &rf);
 3997}
 3998
 3999/*
 4000 * Invoked from try_to_wake_up() to check whether the task can be woken up.
 4001 *
 4002 * The caller holds p::pi_lock if p != current or has preemption
 4003 * disabled when p == current.
 4004 *
 4005 * The rules of saved_state:
 4006 *
 4007 *   The related locking code always holds p::pi_lock when updating
 4008 *   p::saved_state, which means the code is fully serialized in both cases.
 4009 *
 4010 *   For PREEMPT_RT, the lock wait and lock wakeups happen via TASK_RTLOCK_WAIT.
 4011 *   No other bits set. This allows to distinguish all wakeup scenarios.
 4012 *
 4013 *   For FREEZER, the wakeup happens via TASK_FROZEN. No other bits set. This
 4014 *   allows us to prevent early wakeup of tasks before they can be run on
 4015 *   asymmetric ISA architectures (eg ARMv9).
 4016 */
 4017static __always_inline
 4018bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
 4019{
 4020	int match;
 4021
 4022	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
 4023		WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
 4024			     state != TASK_RTLOCK_WAIT);
 4025	}
 4026
 4027	*success = !!(match = __task_state_match(p, state));
 4028
 4029	/*
 4030	 * Saved state preserves the task state across blocking on
 4031	 * an RT lock or TASK_FREEZABLE tasks.  If the state matches,
 4032	 * set p::saved_state to TASK_RUNNING, but do not wake the task
 4033	 * because it waits for a lock wakeup or __thaw_task(). Also
 4034	 * indicate success because from the regular waker's point of
 4035	 * view this has succeeded.
 4036	 *
 4037	 * After acquiring the lock the task will restore p::__state
 4038	 * from p::saved_state which ensures that the regular
 4039	 * wakeup is not lost. The restore will also set
 4040	 * p::saved_state to TASK_RUNNING so any further tests will
 4041	 * not result in false positives vs. @success
 4042	 */
 4043	if (match < 0)
 4044		p->saved_state = TASK_RUNNING;
 4045
 4046	return match > 0;
 4047}
 4048
 4049/*
 4050 * Notes on Program-Order guarantees on SMP systems.
 4051 *
 4052 *  MIGRATION
 4053 *
 4054 * The basic program-order guarantee on SMP systems is that when a task [t]
 4055 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
 4056 * execution on its new CPU [c1].
 4057 *
 4058 * For migration (of runnable tasks) this is provided by the following means:
 4059 *
 4060 *  A) UNLOCK of the rq(c0)->lock scheduling out task t
 4061 *  B) migration for t is required to synchronize *both* rq(c0)->lock and
 4062 *     rq(c1)->lock (if not at the same time, then in that order).
 4063 *  C) LOCK of the rq(c1)->lock scheduling in task
 4064 *
 4065 * Release/acquire chaining guarantees that B happens after A and C after B.
 4066 * Note: the CPU doing B need not be c0 or c1
 4067 *
 4068 * Example:
 4069 *
 4070 *   CPU0            CPU1            CPU2
 4071 *
 4072 *   LOCK rq(0)->lock
 4073 *   sched-out X
 4074 *   sched-in Y
 4075 *   UNLOCK rq(0)->lock
 4076 *
 4077 *                                   LOCK rq(0)->lock // orders against CPU0
 4078 *                                   dequeue X
 4079 *                                   UNLOCK rq(0)->lock
 4080 *
 4081 *                                   LOCK rq(1)->lock
 4082 *                                   enqueue X
 4083 *                                   UNLOCK rq(1)->lock
 4084 *
 4085 *                   LOCK rq(1)->lock // orders against CPU2
 4086 *                   sched-out Z
 4087 *                   sched-in X
 4088 *                   UNLOCK rq(1)->lock
 4089 *
 4090 *
 4091 *  BLOCKING -- aka. SLEEP + WAKEUP
 4092 *
 4093 * For blocking we (obviously) need to provide the same guarantee as for
 4094 * migration. However the means are completely different as there is no lock
 4095 * chain to provide order. Instead we do:
 4096 *
 4097 *   1) smp_store_release(X->on_cpu, 0)   -- finish_task()
 4098 *   2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
 4099 *
 4100 * Example:
 4101 *
 4102 *   CPU0 (schedule)  CPU1 (try_to_wake_up) CPU2 (schedule)
 4103 *
 4104 *   LOCK rq(0)->lock LOCK X->pi_lock
 4105 *   dequeue X
 4106 *   sched-out X
 4107 *   smp_store_release(X->on_cpu, 0);
 4108 *
 4109 *                    smp_cond_load_acquire(&X->on_cpu, !VAL);
 4110 *                    X->state = WAKING
 4111 *                    set_task_cpu(X,2)
 4112 *
 4113 *                    LOCK rq(2)->lock
 4114 *                    enqueue X
 4115 *                    X->state = RUNNING
 4116 *                    UNLOCK rq(2)->lock
 4117 *
 4118 *                                          LOCK rq(2)->lock // orders against CPU1
 4119 *                                          sched-out Z
 4120 *                                          sched-in X
 4121 *                                          UNLOCK rq(2)->lock
 4122 *
 4123 *                    UNLOCK X->pi_lock
 4124 *   UNLOCK rq(0)->lock
 4125 *
 4126 *
 4127 * However, for wakeups there is a second guarantee we must provide, namely we
 4128 * must ensure that CONDITION=1 done by the caller can not be reordered with
 4129 * accesses to the task state; see try_to_wake_up() and set_current_state().
 4130 */
 4131
 4132/**
 4133 * try_to_wake_up - wake up a thread
 4134 * @p: the thread to be awakened
 4135 * @state: the mask of task states that can be woken
 4136 * @wake_flags: wake modifier flags (WF_*)
 4137 *
 4138 * Conceptually does:
 4139 *
 4140 *   If (@state & @p->state) @p->state = TASK_RUNNING.
 4141 *
 4142 * If the task was not queued/runnable, also place it back on a runqueue.
 4143 *
 4144 * This function is atomic against schedule() which would dequeue the task.
 4145 *
 4146 * It issues a full memory barrier before accessing @p->state, see the comment
 4147 * with set_current_state().
 4148 *
 4149 * Uses p->pi_lock to serialize against concurrent wake-ups.
 4150 *
 4151 * Relies on p->pi_lock stabilizing:
 4152 *  - p->sched_class
 4153 *  - p->cpus_ptr
 4154 *  - p->sched_task_group
 4155 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
 4156 *
 4157 * Tries really hard to only take one task_rq(p)->lock for performance.
 4158 * Takes rq->lock in:
 4159 *  - ttwu_runnable()    -- old rq, unavoidable, see comment there;
 4160 *  - ttwu_queue()       -- new rq, for enqueue of the task;
 4161 *  - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
 4162 *
 4163 * As a consequence we race really badly with just about everything. See the
 4164 * many memory barriers and their comments for details.
 4165 *
 4166 * Return: %true if @p->state changes (an actual wakeup was done),
 4167 *	   %false otherwise.
 4168 */
 4169int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
 
 4170{
 4171	guard(preempt)();
 4172	int cpu, success = 0;
 4173
 4174	wake_flags |= WF_TTWU;
 4175
 4176	if (p == current) {
 4177		/*
 4178		 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
 4179		 * == smp_processor_id()'. Together this means we can special
 4180		 * case the whole 'p->on_rq && ttwu_runnable()' case below
 4181		 * without taking any locks.
 4182		 *
 4183		 * Specifically, given current runs ttwu() we must be before
 4184		 * schedule()'s block_task(), as such this must not observe
 4185		 * sched_delayed.
 4186		 *
 4187		 * In particular:
 4188		 *  - we rely on Program-Order guarantees for all the ordering,
 4189		 *  - we're serialized against set_special_state() by virtue of
 4190		 *    it disabling IRQs (this allows not taking ->pi_lock).
 4191		 */
 4192		SCHED_WARN_ON(p->se.sched_delayed);
 4193		if (!ttwu_state_match(p, state, &success))
 4194			goto out;
 4195
 
 4196		trace_sched_waking(p);
 4197		ttwu_do_wakeup(p);
 
 4198		goto out;
 4199	}
 4200
 4201	/*
 4202	 * If we are going to wake up a thread waiting for CONDITION we
 4203	 * need to ensure that CONDITION=1 done by the caller can not be
 4204	 * reordered with p->state check below. This pairs with smp_store_mb()
 4205	 * in set_current_state() that the waiting thread does.
 4206	 */
 4207	scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
 4208		smp_mb__after_spinlock();
 4209		if (!ttwu_state_match(p, state, &success))
 4210			break;
 4211
 4212		trace_sched_waking(p);
 4213
 4214		/*
 4215		 * Ensure we load p->on_rq _after_ p->state, otherwise it would
 4216		 * be possible to, falsely, observe p->on_rq == 0 and get stuck
 4217		 * in smp_cond_load_acquire() below.
 4218		 *
 4219		 * sched_ttwu_pending()			try_to_wake_up()
 4220		 *   STORE p->on_rq = 1			  LOAD p->state
 4221		 *   UNLOCK rq->lock
 4222		 *
 4223		 * __schedule() (switch to task 'p')
 4224		 *   LOCK rq->lock			  smp_rmb();
 4225		 *   smp_mb__after_spinlock();
 4226		 *   UNLOCK rq->lock
 4227		 *
 4228		 * [task p]
 4229		 *   STORE p->state = UNINTERRUPTIBLE	  LOAD p->on_rq
 4230		 *
 4231		 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
 4232		 * __schedule().  See the comment for smp_mb__after_spinlock().
 4233		 *
 4234		 * A similar smp_rmb() lives in __task_needs_rq_lock().
 4235		 */
 4236		smp_rmb();
 4237		if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
 4238			break;
 4239
 4240#ifdef CONFIG_SMP
 4241		/*
 4242		 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
 4243		 * possible to, falsely, observe p->on_cpu == 0.
 4244		 *
 4245		 * One must be running (->on_cpu == 1) in order to remove oneself
 4246		 * from the runqueue.
 4247		 *
 4248		 * __schedule() (switch to task 'p')	try_to_wake_up()
 4249		 *   STORE p->on_cpu = 1		  LOAD p->on_rq
 4250		 *   UNLOCK rq->lock
 4251		 *
 4252		 * __schedule() (put 'p' to sleep)
 4253		 *   LOCK rq->lock			  smp_rmb();
 4254		 *   smp_mb__after_spinlock();
 4255		 *   STORE p->on_rq = 0			  LOAD p->on_cpu
 4256		 *
 4257		 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
 4258		 * __schedule().  See the comment for smp_mb__after_spinlock().
 4259		 *
 4260		 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
 4261		 * schedule()'s deactivate_task() has 'happened' and p will no longer
 4262		 * care about it's own p->state. See the comment in __schedule().
 4263		 */
 4264		smp_acquire__after_ctrl_dep();
 4265
 4266		/*
 4267		 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
 4268		 * == 0), which means we need to do an enqueue, change p->state to
 4269		 * TASK_WAKING such that we can unlock p->pi_lock before doing the
 4270		 * enqueue, such as ttwu_queue_wakelist().
 4271		 */
 4272		WRITE_ONCE(p->__state, TASK_WAKING);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 4273
 4274		/*
 4275		 * If the owning (remote) CPU is still in the middle of schedule() with
 4276		 * this task as prev, considering queueing p on the remote CPUs wake_list
 4277		 * which potentially sends an IPI instead of spinning on p->on_cpu to
 4278		 * let the waker make forward progress. This is safe because IRQs are
 4279		 * disabled and the IPI will deliver after on_cpu is cleared.
 4280		 *
 4281		 * Ensure we load task_cpu(p) after p->on_cpu:
 4282		 *
 4283		 * set_task_cpu(p, cpu);
 4284		 *   STORE p->cpu = @cpu
 4285		 * __schedule() (switch to task 'p')
 4286		 *   LOCK rq->lock
 4287		 *   smp_mb__after_spin_lock()		smp_cond_load_acquire(&p->on_cpu)
 4288		 *   STORE p->on_cpu = 1		LOAD p->cpu
 4289		 *
 4290		 * to ensure we observe the correct CPU on which the task is currently
 4291		 * scheduling.
 4292		 */
 4293		if (smp_load_acquire(&p->on_cpu) &&
 4294		    ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
 4295			break;
 4296
 4297		/*
 4298		 * If the owning (remote) CPU is still in the middle of schedule() with
 4299		 * this task as prev, wait until it's done referencing the task.
 4300		 *
 4301		 * Pairs with the smp_store_release() in finish_task().
 4302		 *
 4303		 * This ensures that tasks getting woken will be fully ordered against
 4304		 * their previous state and preserve Program Order.
 4305		 */
 4306		smp_cond_load_acquire(&p->on_cpu, !VAL);
 
 
 
 
 
 
 
 
 
 
 
 
 4307
 4308		cpu = select_task_rq(p, p->wake_cpu, &wake_flags);
 4309		if (task_cpu(p) != cpu) {
 4310			if (p->in_iowait) {
 4311				delayacct_blkio_end(p);
 4312				atomic_dec(&task_rq(p)->nr_iowait);
 4313			}
 
 
 
 
 4314
 4315			wake_flags |= WF_MIGRATED;
 4316			psi_ttwu_dequeue(p);
 4317			set_task_cpu(p, cpu);
 
 
 4318		}
 
 
 
 
 
 4319#else
 4320		cpu = task_cpu(p);
 4321#endif /* CONFIG_SMP */
 4322
 4323		ttwu_queue(p, cpu, wake_flags);
 4324	}
 
 4325out:
 4326	if (success)
 4327		ttwu_stat(p, task_cpu(p), wake_flags);
 
 4328
 4329	return success;
 4330}
 4331
 4332static bool __task_needs_rq_lock(struct task_struct *p)
 4333{
 4334	unsigned int state = READ_ONCE(p->__state);
 4335
 4336	/*
 4337	 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
 4338	 * the task is blocked. Make sure to check @state since ttwu() can drop
 4339	 * locks at the end, see ttwu_queue_wakelist().
 4340	 */
 4341	if (state == TASK_RUNNING || state == TASK_WAKING)
 4342		return true;
 4343
 4344	/*
 4345	 * Ensure we load p->on_rq after p->__state, otherwise it would be
 4346	 * possible to, falsely, observe p->on_rq == 0.
 4347	 *
 4348	 * See try_to_wake_up() for a longer comment.
 4349	 */
 4350	smp_rmb();
 4351	if (p->on_rq)
 4352		return true;
 4353
 4354#ifdef CONFIG_SMP
 4355	/*
 4356	 * Ensure the task has finished __schedule() and will not be referenced
 4357	 * anymore. Again, see try_to_wake_up() for a longer comment.
 4358	 */
 4359	smp_rmb();
 4360	smp_cond_load_acquire(&p->on_cpu, !VAL);
 4361#endif
 4362
 4363	return false;
 4364}
 4365
 4366/**
 4367 * task_call_func - Invoke a function on task in fixed state
 4368 * @p: Process for which the function is to be invoked, can be @current.
 4369 * @func: Function to invoke.
 4370 * @arg: Argument to function.
 4371 *
 4372 * Fix the task in it's current state by avoiding wakeups and or rq operations
 4373 * and call @func(@arg) on it.  This function can use task_is_runnable() and
 4374 * task_curr() to work out what the state is, if required.  Given that @func
 4375 * can be invoked with a runqueue lock held, it had better be quite
 
 4376 * lightweight.
 4377 *
 4378 * Returns:
 4379 *   Whatever @func returns
 
 
 4380 */
 4381int task_call_func(struct task_struct *p, task_call_f func, void *arg)
 4382{
 4383	struct rq *rq = NULL;
 4384	struct rq_flags rf;
 4385	int ret;
 
 4386
 4387	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
 4388
 4389	if (__task_needs_rq_lock(p))
 4390		rq = __task_rq_lock(p, &rf);
 4391
 4392	/*
 4393	 * At this point the task is pinned; either:
 4394	 *  - blocked and we're holding off wakeups	 (pi->lock)
 4395	 *  - woken, and we're holding off enqueue	 (rq->lock)
 4396	 *  - queued, and we're holding off schedule	 (rq->lock)
 4397	 *  - running, and we're holding off de-schedule (rq->lock)
 4398	 *
 4399	 * The called function (@func) can use: task_curr(), p->on_rq and
 4400	 * p->__state to differentiate between these states.
 4401	 */
 4402	ret = func(p, arg);
 4403
 4404	if (rq)
 4405		rq_unlock(rq, &rf);
 4406
 
 
 
 
 
 
 
 
 
 
 4407	raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
 4408	return ret;
 4409}
 4410
 4411/**
 4412 * cpu_curr_snapshot - Return a snapshot of the currently running task
 4413 * @cpu: The CPU on which to snapshot the task.
 4414 *
 4415 * Returns the task_struct pointer of the task "currently" running on
 4416 * the specified CPU.
 4417 *
 4418 * If the specified CPU was offline, the return value is whatever it
 4419 * is, perhaps a pointer to the task_struct structure of that CPU's idle
 4420 * task, but there is no guarantee.  Callers wishing a useful return
 4421 * value must take some action to ensure that the specified CPU remains
 4422 * online throughout.
 4423 *
 4424 * This function executes full memory barriers before and after fetching
 4425 * the pointer, which permits the caller to confine this function's fetch
 4426 * with respect to the caller's accesses to other shared variables.
 4427 */
 4428struct task_struct *cpu_curr_snapshot(int cpu)
 4429{
 4430	struct rq *rq = cpu_rq(cpu);
 4431	struct task_struct *t;
 4432	struct rq_flags rf;
 4433
 4434	rq_lock_irqsave(rq, &rf);
 4435	smp_mb__after_spinlock(); /* Pairing determined by caller's synchronization design. */
 4436	t = rcu_dereference(cpu_curr(cpu));
 4437	rq_unlock_irqrestore(rq, &rf);
 4438	smp_mb(); /* Pairing determined by caller's synchronization design. */
 4439
 4440	return t;
 4441}
 4442
 4443/**
 4444 * wake_up_process - Wake up a specific process
 4445 * @p: The process to be woken up.
 4446 *
 4447 * Attempt to wake up the nominated process and move it to the set of runnable
 4448 * processes.
 4449 *
 4450 * Return: 1 if the process was woken up, 0 if it was already running.
 4451 *
 4452 * This function executes a full memory barrier before accessing the task state.
 4453 */
 4454int wake_up_process(struct task_struct *p)
 4455{
 4456	return try_to_wake_up(p, TASK_NORMAL, 0);
 4457}
 4458EXPORT_SYMBOL(wake_up_process);
 4459
 4460int wake_up_state(struct task_struct *p, unsigned int state)
 4461{
 4462	return try_to_wake_up(p, state, 0);
 4463}
 4464
 4465/*
 4466 * Perform scheduler related setup for a newly forked process p.
 4467 * p is forked by current.
 4468 *
 4469 * __sched_fork() is basic setup which is also used by sched_init() to
 4470 * initialize the boot CPU's idle task.
 4471 */
 4472static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
 4473{
 4474	p->on_rq			= 0;
 4475
 4476	p->se.on_rq			= 0;
 4477	p->se.exec_start		= 0;
 4478	p->se.sum_exec_runtime		= 0;
 4479	p->se.prev_sum_exec_runtime	= 0;
 4480	p->se.nr_migrations		= 0;
 4481	p->se.vruntime			= 0;
 4482	p->se.vlag			= 0;
 4483	INIT_LIST_HEAD(&p->se.group_node);
 4484
 4485	/* A delayed task cannot be in clone(). */
 4486	SCHED_WARN_ON(p->se.sched_delayed);
 4487
 4488#ifdef CONFIG_FAIR_GROUP_SCHED
 4489	p->se.cfs_rq			= NULL;
 4490#endif
 4491
 4492#ifdef CONFIG_SCHEDSTATS
 4493	/* Even if schedstat is disabled, there should not be garbage */
 4494	memset(&p->stats, 0, sizeof(p->stats));
 4495#endif
 4496
 4497	init_dl_entity(&p->dl);
 
 
 
 4498
 4499	INIT_LIST_HEAD(&p->rt.run_list);
 4500	p->rt.timeout		= 0;
 4501	p->rt.time_slice	= sched_rr_timeslice;
 4502	p->rt.on_rq		= 0;
 4503	p->rt.on_list		= 0;
 4504
 4505#ifdef CONFIG_SCHED_CLASS_EXT
 4506	init_scx_entity(&p->scx);
 4507#endif
 4508
 4509#ifdef CONFIG_PREEMPT_NOTIFIERS
 4510	INIT_HLIST_HEAD(&p->preempt_notifiers);
 4511#endif
 4512
 4513#ifdef CONFIG_COMPACTION
 4514	p->capture_control = NULL;
 4515#endif
 4516	init_numa_balancing(clone_flags, p);
 4517#ifdef CONFIG_SMP
 4518	p->wake_entry.u_flags = CSD_TYPE_TTWU;
 4519	p->migration_pending = NULL;
 4520#endif
 4521	init_sched_mm_cid(p);
 4522}
 4523
 4524DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
 4525
 4526#ifdef CONFIG_NUMA_BALANCING
 4527
 4528int sysctl_numa_balancing_mode;
 4529
 4530static void __set_numabalancing_state(bool enabled)
 4531{
 4532	if (enabled)
 4533		static_branch_enable(&sched_numa_balancing);
 4534	else
 4535		static_branch_disable(&sched_numa_balancing);
 4536}
 4537
 4538void set_numabalancing_state(bool enabled)
 4539{
 4540	if (enabled)
 4541		sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
 4542	else
 4543		sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
 4544	__set_numabalancing_state(enabled);
 4545}
 4546
 4547#ifdef CONFIG_PROC_SYSCTL
 4548static void reset_memory_tiering(void)
 4549{
 4550	struct pglist_data *pgdat;
 4551
 4552	for_each_online_pgdat(pgdat) {
 4553		pgdat->nbp_threshold = 0;
 4554		pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
 4555		pgdat->nbp_th_start = jiffies_to_msecs(jiffies);
 4556	}
 4557}
 4558
 4559static int sysctl_numa_balancing(const struct ctl_table *table, int write,
 4560			  void *buffer, size_t *lenp, loff_t *ppos)
 4561{
 4562	struct ctl_table t;
 4563	int err;
 4564	int state = sysctl_numa_balancing_mode;
 4565
 4566	if (write && !capable(CAP_SYS_ADMIN))
 4567		return -EPERM;
 4568
 4569	t = *table;
 4570	t.data = &state;
 4571	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
 4572	if (err < 0)
 4573		return err;
 4574	if (write) {
 4575		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
 4576		    (state & NUMA_BALANCING_MEMORY_TIERING))
 4577			reset_memory_tiering();
 4578		sysctl_numa_balancing_mode = state;
 4579		__set_numabalancing_state(state);
 4580	}
 4581	return err;
 4582}
 4583#endif
 4584#endif
 4585
 4586#ifdef CONFIG_SCHEDSTATS
 4587
 4588DEFINE_STATIC_KEY_FALSE(sched_schedstats);
 4589
 4590static void set_schedstats(bool enabled)
 4591{
 4592	if (enabled)
 4593		static_branch_enable(&sched_schedstats);
 4594	else
 4595		static_branch_disable(&sched_schedstats);
 4596}
 4597
 4598void force_schedstat_enabled(void)
 4599{
 4600	if (!schedstat_enabled()) {
 4601		pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
 4602		static_branch_enable(&sched_schedstats);
 4603	}
 4604}
 4605
 4606static int __init setup_schedstats(char *str)
 4607{
 4608	int ret = 0;
 4609	if (!str)
 4610		goto out;
 4611
 4612	if (!strcmp(str, "enable")) {
 4613		set_schedstats(true);
 4614		ret = 1;
 4615	} else if (!strcmp(str, "disable")) {
 4616		set_schedstats(false);
 4617		ret = 1;
 4618	}
 4619out:
 4620	if (!ret)
 4621		pr_warn("Unable to parse schedstats=\n");
 4622
 4623	return ret;
 4624}
 4625__setup("schedstats=", setup_schedstats);
 4626
 4627#ifdef CONFIG_PROC_SYSCTL
 4628static int sysctl_schedstats(const struct ctl_table *table, int write, void *buffer,
 4629		size_t *lenp, loff_t *ppos)
 4630{
 4631	struct ctl_table t;
 4632	int err;
 4633	int state = static_branch_likely(&sched_schedstats);
 4634
 4635	if (write && !capable(CAP_SYS_ADMIN))
 4636		return -EPERM;
 4637
 4638	t = *table;
 4639	t.data = &state;
 4640	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
 4641	if (err < 0)
 4642		return err;
 4643	if (write)
 4644		set_schedstats(state);
 4645	return err;
 4646}
 4647#endif /* CONFIG_PROC_SYSCTL */
 4648#endif /* CONFIG_SCHEDSTATS */
 4649
 4650#ifdef CONFIG_SYSCTL
 4651static struct ctl_table sched_core_sysctls[] = {
 4652#ifdef CONFIG_SCHEDSTATS
 4653	{
 4654		.procname       = "sched_schedstats",
 4655		.data           = NULL,
 4656		.maxlen         = sizeof(unsigned int),
 4657		.mode           = 0644,
 4658		.proc_handler   = sysctl_schedstats,
 4659		.extra1         = SYSCTL_ZERO,
 4660		.extra2         = SYSCTL_ONE,
 4661	},
 4662#endif /* CONFIG_SCHEDSTATS */
 4663#ifdef CONFIG_UCLAMP_TASK
 4664	{
 4665		.procname       = "sched_util_clamp_min",
 4666		.data           = &sysctl_sched_uclamp_util_min,
 4667		.maxlen         = sizeof(unsigned int),
 4668		.mode           = 0644,
 4669		.proc_handler   = sysctl_sched_uclamp_handler,
 4670	},
 4671	{
 4672		.procname       = "sched_util_clamp_max",
 4673		.data           = &sysctl_sched_uclamp_util_max,
 4674		.maxlen         = sizeof(unsigned int),
 4675		.mode           = 0644,
 4676		.proc_handler   = sysctl_sched_uclamp_handler,
 4677	},
 4678	{
 4679		.procname       = "sched_util_clamp_min_rt_default",
 4680		.data           = &sysctl_sched_uclamp_util_min_rt_default,
 4681		.maxlen         = sizeof(unsigned int),
 4682		.mode           = 0644,
 4683		.proc_handler   = sysctl_sched_uclamp_handler,
 4684	},
 4685#endif /* CONFIG_UCLAMP_TASK */
 4686#ifdef CONFIG_NUMA_BALANCING
 4687	{
 4688		.procname	= "numa_balancing",
 4689		.data		= NULL, /* filled in by handler */
 4690		.maxlen		= sizeof(unsigned int),
 4691		.mode		= 0644,
 4692		.proc_handler	= sysctl_numa_balancing,
 4693		.extra1		= SYSCTL_ZERO,
 4694		.extra2		= SYSCTL_FOUR,
 4695	},
 4696#endif /* CONFIG_NUMA_BALANCING */
 4697};
 4698static int __init sched_core_sysctl_init(void)
 4699{
 4700	register_sysctl_init("kernel", sched_core_sysctls);
 4701	return 0;
 4702}
 4703late_initcall(sched_core_sysctl_init);
 4704#endif /* CONFIG_SYSCTL */
 4705
 4706/*
 4707 * fork()/clone()-time setup:
 4708 */
 4709int sched_fork(unsigned long clone_flags, struct task_struct *p)
 4710{
 
 
 4711	__sched_fork(clone_flags, p);
 4712	/*
 4713	 * We mark the process as NEW here. This guarantees that
 4714	 * nobody will actually run it, and a signal or other external
 4715	 * event cannot wake it up and insert it on the runqueue either.
 4716	 */
 4717	p->__state = TASK_NEW;
 4718
 4719	/*
 4720	 * Make sure we do not leak PI boosting priority to the child.
 4721	 */
 4722	p->prio = current->normal_prio;
 4723
 4724	uclamp_fork(p);
 4725
 4726	/*
 4727	 * Revert to default priority/policy on fork if requested.
 4728	 */
 4729	if (unlikely(p->sched_reset_on_fork)) {
 4730		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
 4731			p->policy = SCHED_NORMAL;
 4732			p->static_prio = NICE_TO_PRIO(0);
 4733			p->rt_priority = 0;
 4734		} else if (PRIO_TO_NICE(p->static_prio) < 0)
 4735			p->static_prio = NICE_TO_PRIO(0);
 4736
 4737		p->prio = p->normal_prio = p->static_prio;
 4738		set_load_weight(p, false);
 4739		p->se.custom_slice = 0;
 4740		p->se.slice = sysctl_sched_base_slice;
 4741
 4742		/*
 4743		 * We don't need the reset flag anymore after the fork. It has
 4744		 * fulfilled its duty:
 4745		 */
 4746		p->sched_reset_on_fork = 0;
 4747	}
 4748
 4749	if (dl_prio(p->prio))
 4750		return -EAGAIN;
 4751
 4752	scx_pre_fork(p);
 4753
 4754	if (rt_prio(p->prio)) {
 4755		p->sched_class = &rt_sched_class;
 4756#ifdef CONFIG_SCHED_CLASS_EXT
 4757	} else if (task_should_scx(p->policy)) {
 4758		p->sched_class = &ext_sched_class;
 4759#endif
 4760	} else {
 4761		p->sched_class = &fair_sched_class;
 4762	}
 4763
 4764	init_entity_runnable_average(&p->se);
 4765
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 4766
 4767#ifdef CONFIG_SCHED_INFO
 4768	if (likely(sched_info_on()))
 4769		memset(&p->sched_info, 0, sizeof(p->sched_info));
 4770#endif
 4771#if defined(CONFIG_SMP)
 4772	p->on_cpu = 0;
 4773#endif
 4774	init_task_preempt_count(p);
 4775#ifdef CONFIG_SMP
 4776	plist_node_init(&p->pushable_tasks, MAX_PRIO);
 4777	RB_CLEAR_NODE(&p->pushable_dl_tasks);
 4778#endif
 4779	return 0;
 4780}
 4781
 4782int sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
 4783{
 4784	unsigned long flags;
 4785
 4786	/*
 4787	 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
 4788	 * required yet, but lockdep gets upset if rules are violated.
 4789	 */
 4790	raw_spin_lock_irqsave(&p->pi_lock, flags);
 4791#ifdef CONFIG_CGROUP_SCHED
 4792	if (1) {
 4793		struct task_group *tg;
 4794		tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
 4795				  struct task_group, css);
 4796		tg = autogroup_task_group(p, tg);
 4797		p->sched_task_group = tg;
 4798	}
 4799#endif
 4800	rseq_migrate(p);
 4801	/*
 4802	 * We're setting the CPU for the first time, we don't migrate,
 4803	 * so use __set_task_cpu().
 4804	 */
 4805	__set_task_cpu(p, smp_processor_id());
 4806	if (p->sched_class->task_fork)
 4807		p->sched_class->task_fork(p);
 4808	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
 4809
 4810	return scx_fork(p);
 4811}
 4812
 4813void sched_cancel_fork(struct task_struct *p)
 4814{
 4815	scx_cancel_fork(p);
 4816}
 4817
 4818void sched_post_fork(struct task_struct *p)
 4819{
 4820	uclamp_post_fork(p);
 4821	scx_post_fork(p);
 4822}
 4823
 4824unsigned long to_ratio(u64 period, u64 runtime)
 4825{
 4826	if (runtime == RUNTIME_INF)
 4827		return BW_UNIT;
 4828
 4829	/*
 4830	 * Doing this here saves a lot of checks in all
 4831	 * the calling paths, and returning zero seems
 4832	 * safe for them anyway.
 4833	 */
 4834	if (period == 0)
 4835		return 0;
 4836
 4837	return div64_u64(runtime << BW_SHIFT, period);
 4838}
 4839
 4840/*
 4841 * wake_up_new_task - wake up a newly created task for the first time.
 4842 *
 4843 * This function will do some initial scheduler statistics housekeeping
 4844 * that must be done for every newly created context, then puts the task
 4845 * on the runqueue and wakes it.
 4846 */
 4847void wake_up_new_task(struct task_struct *p)
 4848{
 4849	struct rq_flags rf;
 4850	struct rq *rq;
 4851	int wake_flags = WF_FORK;
 4852
 4853	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
 4854	WRITE_ONCE(p->__state, TASK_RUNNING);
 4855#ifdef CONFIG_SMP
 4856	/*
 4857	 * Fork balancing, do it here and not earlier because:
 4858	 *  - cpus_ptr can change in the fork path
 4859	 *  - any previously selected CPU might disappear through hotplug
 4860	 *
 4861	 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
 4862	 * as we're not fully set-up yet.
 4863	 */
 4864	p->recent_used_cpu = task_cpu(p);
 4865	rseq_migrate(p);
 4866	__set_task_cpu(p, select_task_rq(p, task_cpu(p), &wake_flags));
 4867#endif
 4868	rq = __task_rq_lock(p, &rf);
 4869	update_rq_clock(rq);
 4870	post_init_entity_util_avg(p);
 4871
 4872	activate_task(rq, p, ENQUEUE_NOCLOCK | ENQUEUE_INITIAL);
 4873	trace_sched_wakeup_new(p);
 4874	wakeup_preempt(rq, p, wake_flags);
 4875#ifdef CONFIG_SMP
 4876	if (p->sched_class->task_woken) {
 4877		/*
 4878		 * Nothing relies on rq->lock after this, so it's fine to
 4879		 * drop it.
 4880		 */
 4881		rq_unpin_lock(rq, &rf);
 4882		p->sched_class->task_woken(rq, p);
 4883		rq_repin_lock(rq, &rf);
 4884	}
 4885#endif
 4886	task_rq_unlock(rq, p, &rf);
 4887}
 4888
 4889#ifdef CONFIG_PREEMPT_NOTIFIERS
 4890
 4891static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
 4892
 4893void preempt_notifier_inc(void)
 4894{
 4895	static_branch_inc(&preempt_notifier_key);
 4896}
 4897EXPORT_SYMBOL_GPL(preempt_notifier_inc);
 4898
 4899void preempt_notifier_dec(void)
 4900{
 4901	static_branch_dec(&preempt_notifier_key);
 4902}
 4903EXPORT_SYMBOL_GPL(preempt_notifier_dec);
 4904
 4905/**
 4906 * preempt_notifier_register - tell me when current is being preempted & rescheduled
 4907 * @notifier: notifier struct to register
 4908 */
 4909void preempt_notifier_register(struct preempt_notifier *notifier)
 4910{
 4911	if (!static_branch_unlikely(&preempt_notifier_key))
 4912		WARN(1, "registering preempt_notifier while notifiers disabled\n");
 4913
 4914	hlist_add_head(&notifier->link, &current->preempt_notifiers);
 4915}
 4916EXPORT_SYMBOL_GPL(preempt_notifier_register);
 4917
 4918/**
 4919 * preempt_notifier_unregister - no longer interested in preemption notifications
 4920 * @notifier: notifier struct to unregister
 4921 *
 4922 * This is *not* safe to call from within a preemption notifier.
 4923 */
 4924void preempt_notifier_unregister(struct preempt_notifier *notifier)
 4925{
 4926	hlist_del(&notifier->link);
 4927}
 4928EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
 4929
 4930static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
 4931{
 4932	struct preempt_notifier *notifier;
 4933
 4934	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
 4935		notifier->ops->sched_in(notifier, raw_smp_processor_id());
 4936}
 4937
 4938static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
 4939{
 4940	if (static_branch_unlikely(&preempt_notifier_key))
 4941		__fire_sched_in_preempt_notifiers(curr);
 4942}
 4943
 4944static void
 4945__fire_sched_out_preempt_notifiers(struct task_struct *curr,
 4946				   struct task_struct *next)
 4947{
 4948	struct preempt_notifier *notifier;
 4949
 4950	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
 4951		notifier->ops->sched_out(notifier, next);
 4952}
 4953
 4954static __always_inline void
 4955fire_sched_out_preempt_notifiers(struct task_struct *curr,
 4956				 struct task_struct *next)
 4957{
 4958	if (static_branch_unlikely(&preempt_notifier_key))
 4959		__fire_sched_out_preempt_notifiers(curr, next);
 4960}
 4961
 4962#else /* !CONFIG_PREEMPT_NOTIFIERS */
 4963
 4964static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
 4965{
 4966}
 4967
 4968static inline void
 4969fire_sched_out_preempt_notifiers(struct task_struct *curr,
 4970				 struct task_struct *next)
 4971{
 4972}
 4973
 4974#endif /* CONFIG_PREEMPT_NOTIFIERS */
 4975
 4976static inline void prepare_task(struct task_struct *next)
 4977{
 4978#ifdef CONFIG_SMP
 4979	/*
 4980	 * Claim the task as running, we do this before switching to it
 4981	 * such that any running task will have this set.
 4982	 *
 4983	 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
 4984	 * its ordering comment.
 4985	 */
 4986	WRITE_ONCE(next->on_cpu, 1);
 4987#endif
 4988}
 4989
 4990static inline void finish_task(struct task_struct *prev)
 4991{
 4992#ifdef CONFIG_SMP
 4993	/*
 4994	 * This must be the very last reference to @prev from this CPU. After
 4995	 * p->on_cpu is cleared, the task can be moved to a different CPU. We
 4996	 * must ensure this doesn't happen until the switch is completely
 4997	 * finished.
 4998	 *
 4999	 * In particular, the load of prev->state in finish_task_switch() must
 5000	 * happen before this.
 5001	 *
 5002	 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
 5003	 */
 5004	smp_store_release(&prev->on_cpu, 0);
 5005#endif
 5006}
 5007
 5008#ifdef CONFIG_SMP
 5009
 5010static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
 5011{
 5012	void (*func)(struct rq *rq);
 5013	struct balance_callback *next;
 5014
 5015	lockdep_assert_rq_held(rq);
 5016
 5017	while (head) {
 5018		func = (void (*)(struct rq *))head->func;
 5019		next = head->next;
 5020		head->next = NULL;
 5021		head = next;
 5022
 5023		func(rq);
 5024	}
 5025}
 5026
 5027static void balance_push(struct rq *rq);
 5028
 5029/*
 5030 * balance_push_callback is a right abuse of the callback interface and plays
 5031 * by significantly different rules.
 5032 *
 5033 * Where the normal balance_callback's purpose is to be ran in the same context
 5034 * that queued it (only later, when it's safe to drop rq->lock again),
 5035 * balance_push_callback is specifically targeted at __schedule().
 5036 *
 5037 * This abuse is tolerated because it places all the unlikely/odd cases behind
 5038 * a single test, namely: rq->balance_callback == NULL.
 5039 */
 5040struct balance_callback balance_push_callback = {
 5041	.next = NULL,
 5042	.func = balance_push,
 5043};
 5044
 5045static inline struct balance_callback *
 5046__splice_balance_callbacks(struct rq *rq, bool split)
 5047{
 5048	struct balance_callback *head = rq->balance_callback;
 5049
 5050	if (likely(!head))
 5051		return NULL;
 5052
 5053	lockdep_assert_rq_held(rq);
 5054	/*
 5055	 * Must not take balance_push_callback off the list when
 5056	 * splice_balance_callbacks() and balance_callbacks() are not
 5057	 * in the same rq->lock section.
 5058	 *
 5059	 * In that case it would be possible for __schedule() to interleave
 5060	 * and observe the list empty.
 5061	 */
 5062	if (split && head == &balance_push_callback)
 5063		head = NULL;
 5064	else
 5065		rq->balance_callback = NULL;
 5066
 5067	return head;
 5068}
 5069
 5070struct balance_callback *splice_balance_callbacks(struct rq *rq)
 5071{
 5072	return __splice_balance_callbacks(rq, true);
 5073}
 5074
 5075static void __balance_callbacks(struct rq *rq)
 5076{
 5077	do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
 5078}
 5079
 5080void balance_callbacks(struct rq *rq, struct balance_callback *head)
 5081{
 5082	unsigned long flags;
 5083
 5084	if (unlikely(head)) {
 5085		raw_spin_rq_lock_irqsave(rq, flags);
 5086		do_balance_callbacks(rq, head);
 5087		raw_spin_rq_unlock_irqrestore(rq, flags);
 5088	}
 5089}
 5090
 5091#else
 5092
 5093static inline void __balance_callbacks(struct rq *rq)
 5094{
 5095}
 5096
 
 
 
 
 
 
 
 
 
 5097#endif
 5098
 5099static inline void
 5100prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
 5101{
 5102	/*
 5103	 * Since the runqueue lock will be released by the next
 5104	 * task (which is an invalid locking op but in the case
 5105	 * of the scheduler it's an obvious special-case), so we
 5106	 * do an early lockdep release here:
 5107	 */
 5108	rq_unpin_lock(rq, rf);
 5109	spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
 5110#ifdef CONFIG_DEBUG_SPINLOCK
 5111	/* this is a valid case when another task releases the spinlock */
 5112	rq_lockp(rq)->owner = next;
 5113#endif
 5114}
 5115
 5116static inline void finish_lock_switch(struct rq *rq)
 5117{
 5118	/*
 5119	 * If we are tracking spinlock dependencies then we have to
 5120	 * fix up the runqueue lock - which gets 'carried over' from
 5121	 * prev into current:
 5122	 */
 5123	spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
 5124	__balance_callbacks(rq);
 5125	raw_spin_rq_unlock_irq(rq);
 5126}
 5127
 5128/*
 5129 * NOP if the arch has not defined these:
 5130 */
 5131
 5132#ifndef prepare_arch_switch
 5133# define prepare_arch_switch(next)	do { } while (0)
 5134#endif
 5135
 5136#ifndef finish_arch_post_lock_switch
 5137# define finish_arch_post_lock_switch()	do { } while (0)
 5138#endif
 5139
 5140static inline void kmap_local_sched_out(void)
 5141{
 5142#ifdef CONFIG_KMAP_LOCAL
 5143	if (unlikely(current->kmap_ctrl.idx))
 5144		__kmap_local_sched_out();
 5145#endif
 5146}
 5147
 5148static inline void kmap_local_sched_in(void)
 5149{
 5150#ifdef CONFIG_KMAP_LOCAL
 5151	if (unlikely(current->kmap_ctrl.idx))
 5152		__kmap_local_sched_in();
 5153#endif
 5154}
 5155
 5156/**
 5157 * prepare_task_switch - prepare to switch tasks
 5158 * @rq: the runqueue preparing to switch
 5159 * @prev: the current task that is being switched out
 5160 * @next: the task we are going to switch to.
 5161 *
 5162 * This is called with the rq lock held and interrupts off. It must
 5163 * be paired with a subsequent finish_task_switch after the context
 5164 * switch.
 5165 *
 5166 * prepare_task_switch sets up locking and calls architecture specific
 5167 * hooks.
 5168 */
 5169static inline void
 5170prepare_task_switch(struct rq *rq, struct task_struct *prev,
 5171		    struct task_struct *next)
 5172{
 5173	kcov_prepare_switch(prev);
 5174	sched_info_switch(rq, prev, next);
 5175	perf_event_task_sched_out(prev, next);
 5176	rseq_preempt(prev);
 5177	fire_sched_out_preempt_notifiers(prev, next);
 5178	kmap_local_sched_out();
 5179	prepare_task(next);
 5180	prepare_arch_switch(next);
 5181}
 5182
 5183/**
 5184 * finish_task_switch - clean up after a task-switch
 5185 * @prev: the thread we just switched away from.
 5186 *
 5187 * finish_task_switch must be called after the context switch, paired
 5188 * with a prepare_task_switch call before the context switch.
 5189 * finish_task_switch will reconcile locking set up by prepare_task_switch,
 5190 * and do any other architecture-specific cleanup actions.
 5191 *
 5192 * Note that we may have delayed dropping an mm in context_switch(). If
 5193 * so, we finish that here outside of the runqueue lock. (Doing it
 5194 * with the lock held can cause deadlocks; see schedule() for
 5195 * details.)
 5196 *
 5197 * The context switch have flipped the stack from under us and restored the
 5198 * local variables which were saved when this task called schedule() in the
 5199 * past. 'prev == current' is still correct but we need to recalculate this_rq
 5200 * because prev may have moved to another CPU.
 5201 */
 5202static struct rq *finish_task_switch(struct task_struct *prev)
 5203	__releases(rq->lock)
 5204{
 5205	struct rq *rq = this_rq();
 5206	struct mm_struct *mm = rq->prev_mm;
 5207	unsigned int prev_state;
 5208
 5209	/*
 5210	 * The previous task will have left us with a preempt_count of 2
 5211	 * because it left us after:
 5212	 *
 5213	 *	schedule()
 5214	 *	  preempt_disable();			// 1
 5215	 *	  __schedule()
 5216	 *	    raw_spin_lock_irq(&rq->lock)	// 2
 5217	 *
 5218	 * Also, see FORK_PREEMPT_COUNT.
 5219	 */
 5220	if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
 5221		      "corrupted preempt_count: %s/%d/0x%x\n",
 5222		      current->comm, current->pid, preempt_count()))
 5223		preempt_count_set(FORK_PREEMPT_COUNT);
 5224
 5225	rq->prev_mm = NULL;
 5226
 5227	/*
 5228	 * A task struct has one reference for the use as "current".
 5229	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
 5230	 * schedule one last time. The schedule call will never return, and
 5231	 * the scheduled task must drop that reference.
 5232	 *
 5233	 * We must observe prev->state before clearing prev->on_cpu (in
 5234	 * finish_task), otherwise a concurrent wakeup can get prev
 5235	 * running on another CPU and we could rave with its RUNNING -> DEAD
 5236	 * transition, resulting in a double drop.
 5237	 */
 5238	prev_state = READ_ONCE(prev->__state);
 5239	vtime_task_switch(prev);
 5240	perf_event_task_sched_in(prev, current);
 5241	finish_task(prev);
 5242	tick_nohz_task_switch();
 5243	finish_lock_switch(rq);
 5244	finish_arch_post_lock_switch();
 5245	kcov_finish_switch(current);
 5246	/*
 5247	 * kmap_local_sched_out() is invoked with rq::lock held and
 5248	 * interrupts disabled. There is no requirement for that, but the
 5249	 * sched out code does not have an interrupt enabled section.
 5250	 * Restoring the maps on sched in does not require interrupts being
 5251	 * disabled either.
 5252	 */
 5253	kmap_local_sched_in();
 5254
 5255	fire_sched_in_preempt_notifiers(current);
 5256	/*
 5257	 * When switching through a kernel thread, the loop in
 5258	 * membarrier_{private,global}_expedited() may have observed that
 5259	 * kernel thread and not issued an IPI. It is therefore possible to
 5260	 * schedule between user->kernel->user threads without passing though
 5261	 * switch_mm(). Membarrier requires a barrier after storing to
 5262	 * rq->curr, before returning to userspace, so provide them here:
 5263	 *
 5264	 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
 5265	 *   provided by mmdrop_lazy_tlb(),
 5266	 * - a sync_core for SYNC_CORE.
 5267	 */
 5268	if (mm) {
 5269		membarrier_mm_sync_core_before_usermode(mm);
 5270		mmdrop_lazy_tlb_sched(mm);
 5271	}
 5272
 5273	if (unlikely(prev_state == TASK_DEAD)) {
 5274		if (prev->sched_class->task_dead)
 5275			prev->sched_class->task_dead(prev);
 5276
 
 
 
 
 
 
 5277		/* Task is done with its stack. */
 5278		put_task_stack(prev);
 5279
 5280		put_task_struct_rcu_user(prev);
 5281	}
 5282
 5283	return rq;
 5284}
 5285
 5286/**
 5287 * schedule_tail - first thing a freshly forked thread must call.
 5288 * @prev: the thread we just switched away from.
 5289 */
 5290asmlinkage __visible void schedule_tail(struct task_struct *prev)
 5291	__releases(rq->lock)
 5292{
 5293	/*
 5294	 * New tasks start with FORK_PREEMPT_COUNT, see there and
 5295	 * finish_task_switch() for details.
 5296	 *
 5297	 * finish_task_switch() will drop rq->lock() and lower preempt_count
 5298	 * and the preempt_enable() will end up enabling preemption (on
 5299	 * PREEMPT_COUNT kernels).
 5300	 */
 5301
 5302	finish_task_switch(prev);
 5303	preempt_enable();
 5304
 5305	if (current->set_child_tid)
 5306		put_user(task_pid_vnr(current), current->set_child_tid);
 5307
 5308	calculate_sigpending();
 5309}
 5310
 5311/*
 5312 * context_switch - switch to the new MM and the new thread's register state.
 5313 */
 5314static __always_inline struct rq *
 5315context_switch(struct rq *rq, struct task_struct *prev,
 5316	       struct task_struct *next, struct rq_flags *rf)
 5317{
 5318	prepare_task_switch(rq, prev, next);
 5319
 5320	/*
 5321	 * For paravirt, this is coupled with an exit in switch_to to
 5322	 * combine the page table reload and the switch backend into
 5323	 * one hypercall.
 5324	 */
 5325	arch_start_context_switch(prev);
 5326
 5327	/*
 5328	 * kernel -> kernel   lazy + transfer active
 5329	 *   user -> kernel   lazy + mmgrab_lazy_tlb() active
 5330	 *
 5331	 * kernel ->   user   switch + mmdrop_lazy_tlb() active
 5332	 *   user ->   user   switch
 5333	 *
 5334	 * switch_mm_cid() needs to be updated if the barriers provided
 5335	 * by context_switch() are modified.
 5336	 */
 5337	if (!next->mm) {                                // to kernel
 5338		enter_lazy_tlb(prev->active_mm, next);
 5339
 5340		next->active_mm = prev->active_mm;
 5341		if (prev->mm)                           // from user
 5342			mmgrab_lazy_tlb(prev->active_mm);
 5343		else
 5344			prev->active_mm = NULL;
 5345	} else {                                        // to user
 5346		membarrier_switch_mm(rq, prev->active_mm, next->mm);
 5347		/*
 5348		 * sys_membarrier() requires an smp_mb() between setting
 5349		 * rq->curr / membarrier_switch_mm() and returning to userspace.
 5350		 *
 5351		 * The below provides this either through switch_mm(), or in
 5352		 * case 'prev->active_mm == next->mm' through
 5353		 * finish_task_switch()'s mmdrop().
 5354		 */
 5355		switch_mm_irqs_off(prev->active_mm, next->mm, next);
 5356		lru_gen_use_mm(next->mm);
 5357
 5358		if (!prev->mm) {                        // from kernel
 5359			/* will mmdrop_lazy_tlb() in finish_task_switch(). */
 5360			rq->prev_mm = prev->active_mm;
 5361			prev->active_mm = NULL;
 5362		}
 5363	}
 5364
 5365	/* switch_mm_cid() requires the memory barriers above. */
 5366	switch_mm_cid(rq, prev, next);
 5367
 5368	prepare_lock_switch(rq, next, rf);
 5369
 5370	/* Here we just switch the register state and the stack. */
 5371	switch_to(prev, next, prev);
 5372	barrier();
 5373
 5374	return finish_task_switch(prev);
 5375}
 5376
 5377/*
 5378 * nr_running and nr_context_switches:
 5379 *
 5380 * externally visible scheduler statistics: current number of runnable
 5381 * threads, total number of context switches performed since bootup.
 5382 */
 5383unsigned int nr_running(void)
 5384{
 5385	unsigned int i, sum = 0;
 5386
 5387	for_each_online_cpu(i)
 5388		sum += cpu_rq(i)->nr_running;
 5389
 5390	return sum;
 5391}
 5392
 5393/*
 5394 * Check if only the current task is running on the CPU.
 5395 *
 5396 * Caution: this function does not check that the caller has disabled
 5397 * preemption, thus the result might have a time-of-check-to-time-of-use
 5398 * race.  The caller is responsible to use it correctly, for example:
 5399 *
 5400 * - from a non-preemptible section (of course)
 5401 *
 5402 * - from a thread that is bound to a single CPU
 5403 *
 5404 * - in a loop with very short iterations (e.g. a polling loop)
 5405 */
 5406bool single_task_running(void)
 5407{
 5408	return raw_rq()->nr_running == 1;
 5409}
 5410EXPORT_SYMBOL(single_task_running);
 5411
 5412unsigned long long nr_context_switches_cpu(int cpu)
 5413{
 5414	return cpu_rq(cpu)->nr_switches;
 5415}
 5416
 5417unsigned long long nr_context_switches(void)
 5418{
 5419	int i;
 5420	unsigned long long sum = 0;
 5421
 5422	for_each_possible_cpu(i)
 5423		sum += cpu_rq(i)->nr_switches;
 5424
 5425	return sum;
 5426}
 5427
 5428/*
 5429 * Consumers of these two interfaces, like for example the cpuidle menu
 5430 * governor, are using nonsensical data. Preferring shallow idle state selection
 5431 * for a CPU that has IO-wait which might not even end up running the task when
 5432 * it does become runnable.
 5433 */
 5434
 5435unsigned int nr_iowait_cpu(int cpu)
 5436{
 5437	return atomic_read(&cpu_rq(cpu)->nr_iowait);
 5438}
 5439
 5440/*
 5441 * IO-wait accounting, and how it's mostly bollocks (on SMP).
 5442 *
 5443 * The idea behind IO-wait account is to account the idle time that we could
 5444 * have spend running if it were not for IO. That is, if we were to improve the
 5445 * storage performance, we'd have a proportional reduction in IO-wait time.
 5446 *
 5447 * This all works nicely on UP, where, when a task blocks on IO, we account
 5448 * idle time as IO-wait, because if the storage were faster, it could've been
 5449 * running and we'd not be idle.
 5450 *
 5451 * This has been extended to SMP, by doing the same for each CPU. This however
 5452 * is broken.
 5453 *
 5454 * Imagine for instance the case where two tasks block on one CPU, only the one
 5455 * CPU will have IO-wait accounted, while the other has regular idle. Even
 5456 * though, if the storage were faster, both could've ran at the same time,
 5457 * utilising both CPUs.
 5458 *
 5459 * This means, that when looking globally, the current IO-wait accounting on
 5460 * SMP is a lower bound, by reason of under accounting.
 5461 *
 5462 * Worse, since the numbers are provided per CPU, they are sometimes
 5463 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
 5464 * associated with any one particular CPU, it can wake to another CPU than it
 5465 * blocked on. This means the per CPU IO-wait number is meaningless.
 5466 *
 5467 * Task CPU affinities can make all that even more 'interesting'.
 5468 */
 5469
 5470unsigned int nr_iowait(void)
 5471{
 5472	unsigned int i, sum = 0;
 5473
 5474	for_each_possible_cpu(i)
 5475		sum += nr_iowait_cpu(i);
 5476
 5477	return sum;
 5478}
 5479
 5480#ifdef CONFIG_SMP
 5481
 5482/*
 5483 * sched_exec - execve() is a valuable balancing opportunity, because at
 5484 * this point the task has the smallest effective memory and cache footprint.
 5485 */
 5486void sched_exec(void)
 5487{
 5488	struct task_struct *p = current;
 5489	struct migration_arg arg;
 5490	int dest_cpu;
 5491
 5492	scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
 5493		dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
 5494		if (dest_cpu == smp_processor_id())
 5495			return;
 5496
 5497		if (unlikely(!cpu_active(dest_cpu)))
 5498			return;
 5499
 5500		arg = (struct migration_arg){ p, dest_cpu };
 
 
 5501	}
 5502	stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
 
 5503}
 5504
 5505#endif
 5506
 5507DEFINE_PER_CPU(struct kernel_stat, kstat);
 5508DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
 5509
 5510EXPORT_PER_CPU_SYMBOL(kstat);
 5511EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
 5512
 5513/*
 5514 * The function fair_sched_class.update_curr accesses the struct curr
 5515 * and its field curr->exec_start; when called from task_sched_runtime(),
 5516 * we observe a high rate of cache misses in practice.
 5517 * Prefetching this data results in improved performance.
 5518 */
 5519static inline void prefetch_curr_exec_start(struct task_struct *p)
 5520{
 5521#ifdef CONFIG_FAIR_GROUP_SCHED
 5522	struct sched_entity *curr = p->se.cfs_rq->curr;
 5523#else
 5524	struct sched_entity *curr = task_rq(p)->cfs.curr;
 5525#endif
 5526	prefetch(curr);
 5527	prefetch(&curr->exec_start);
 5528}
 5529
 5530/*
 5531 * Return accounted runtime for the task.
 5532 * In case the task is currently running, return the runtime plus current's
 5533 * pending runtime that have not been accounted yet.
 5534 */
 5535unsigned long long task_sched_runtime(struct task_struct *p)
 5536{
 5537	struct rq_flags rf;
 5538	struct rq *rq;
 5539	u64 ns;
 5540
 5541#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
 5542	/*
 5543	 * 64-bit doesn't need locks to atomically read a 64-bit value.
 5544	 * So we have a optimization chance when the task's delta_exec is 0.
 5545	 * Reading ->on_cpu is racy, but this is OK.
 5546	 *
 5547	 * If we race with it leaving CPU, we'll take a lock. So we're correct.
 5548	 * If we race with it entering CPU, unaccounted time is 0. This is
 5549	 * indistinguishable from the read occurring a few cycles earlier.
 5550	 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
 5551	 * been accounted, so we're correct here as well.
 5552	 */
 5553	if (!p->on_cpu || !task_on_rq_queued(p))
 5554		return p->se.sum_exec_runtime;
 5555#endif
 5556
 5557	rq = task_rq_lock(p, &rf);
 5558	/*
 5559	 * Must be ->curr _and_ ->on_rq.  If dequeued, we would
 5560	 * project cycles that may never be accounted to this
 5561	 * thread, breaking clock_gettime().
 5562	 */
 5563	if (task_current_donor(rq, p) && task_on_rq_queued(p)) {
 5564		prefetch_curr_exec_start(p);
 5565		update_rq_clock(rq);
 5566		p->sched_class->update_curr(rq);
 5567	}
 5568	ns = p->se.sum_exec_runtime;
 5569	task_rq_unlock(rq, p, &rf);
 5570
 5571	return ns;
 5572}
 5573
 5574#ifdef CONFIG_SCHED_DEBUG
 5575static u64 cpu_resched_latency(struct rq *rq)
 5576{
 5577	int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
 5578	u64 resched_latency, now = rq_clock(rq);
 5579	static bool warned_once;
 5580
 5581	if (sysctl_resched_latency_warn_once && warned_once)
 5582		return 0;
 5583
 5584	if (!need_resched() || !latency_warn_ms)
 5585		return 0;
 5586
 5587	if (system_state == SYSTEM_BOOTING)
 5588		return 0;
 5589
 5590	if (!rq->last_seen_need_resched_ns) {
 5591		rq->last_seen_need_resched_ns = now;
 5592		rq->ticks_without_resched = 0;
 5593		return 0;
 5594	}
 5595
 5596	rq->ticks_without_resched++;
 5597	resched_latency = now - rq->last_seen_need_resched_ns;
 5598	if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
 5599		return 0;
 5600
 5601	warned_once = true;
 5602
 5603	return resched_latency;
 5604}
 5605
 5606static int __init setup_resched_latency_warn_ms(char *str)
 5607{
 5608	long val;
 5609
 5610	if ((kstrtol(str, 0, &val))) {
 5611		pr_warn("Unable to set resched_latency_warn_ms\n");
 5612		return 1;
 5613	}
 5614
 5615	sysctl_resched_latency_warn_ms = val;
 5616	return 1;
 5617}
 5618__setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
 5619#else
 5620static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
 5621#endif /* CONFIG_SCHED_DEBUG */
 5622
 5623/*
 5624 * This function gets called by the timer code, with HZ frequency.
 5625 * We call it with interrupts disabled.
 5626 */
 5627void sched_tick(void)
 5628{
 5629	int cpu = smp_processor_id();
 5630	struct rq *rq = cpu_rq(cpu);
 5631	/* accounting goes to the donor task */
 5632	struct task_struct *donor;
 5633	struct rq_flags rf;
 5634	unsigned long hw_pressure;
 5635	u64 resched_latency;
 5636
 5637	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
 5638		arch_scale_freq_tick();
 5639
 5640	sched_clock_tick();
 5641
 5642	rq_lock(rq, &rf);
 5643	donor = rq->donor;
 5644
 5645	psi_account_irqtime(rq, donor, NULL);
 5646
 5647	update_rq_clock(rq);
 5648	hw_pressure = arch_scale_hw_pressure(cpu_of(rq));
 5649	update_hw_load_avg(rq_clock_task(rq), rq, hw_pressure);
 5650
 5651	if (dynamic_preempt_lazy() && tif_test_bit(TIF_NEED_RESCHED_LAZY))
 5652		resched_curr(rq);
 5653
 5654	donor->sched_class->task_tick(rq, donor, 0);
 5655	if (sched_feat(LATENCY_WARN))
 5656		resched_latency = cpu_resched_latency(rq);
 5657	calc_global_load_tick(rq);
 5658	sched_core_tick(rq);
 5659	task_tick_mm_cid(rq, donor);
 5660	scx_tick(rq);
 5661
 5662	rq_unlock(rq, &rf);
 5663
 5664	if (sched_feat(LATENCY_WARN) && resched_latency)
 5665		resched_latency_warn(cpu, resched_latency);
 5666
 5667	perf_event_task_tick();
 5668
 5669	if (donor->flags & PF_WQ_WORKER)
 5670		wq_worker_tick(donor);
 5671
 5672#ifdef CONFIG_SMP
 5673	if (!scx_switched_all()) {
 5674		rq->idle_balance = idle_cpu(cpu);
 5675		sched_balance_trigger(rq);
 5676	}
 5677#endif
 5678}
 5679
 5680#ifdef CONFIG_NO_HZ_FULL
 5681
 5682struct tick_work {
 5683	int			cpu;
 5684	atomic_t		state;
 5685	struct delayed_work	work;
 5686};
 5687/* Values for ->state, see diagram below. */
 5688#define TICK_SCHED_REMOTE_OFFLINE	0
 5689#define TICK_SCHED_REMOTE_OFFLINING	1
 5690#define TICK_SCHED_REMOTE_RUNNING	2
 5691
 5692/*
 5693 * State diagram for ->state:
 5694 *
 5695 *
 5696 *          TICK_SCHED_REMOTE_OFFLINE
 5697 *                    |   ^
 5698 *                    |   |
 5699 *                    |   | sched_tick_remote()
 5700 *                    |   |
 5701 *                    |   |
 5702 *                    +--TICK_SCHED_REMOTE_OFFLINING
 5703 *                    |   ^
 5704 *                    |   |
 5705 * sched_tick_start() |   | sched_tick_stop()
 5706 *                    |   |
 5707 *                    V   |
 5708 *          TICK_SCHED_REMOTE_RUNNING
 5709 *
 5710 *
 5711 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
 5712 * and sched_tick_start() are happy to leave the state in RUNNING.
 5713 */
 5714
 5715static struct tick_work __percpu *tick_work_cpu;
 5716
 5717static void sched_tick_remote(struct work_struct *work)
 5718{
 5719	struct delayed_work *dwork = to_delayed_work(work);
 5720	struct tick_work *twork = container_of(dwork, struct tick_work, work);
 5721	int cpu = twork->cpu;
 5722	struct rq *rq = cpu_rq(cpu);
 
 
 
 5723	int os;
 5724
 5725	/*
 5726	 * Handle the tick only if it appears the remote CPU is running in full
 5727	 * dynticks mode. The check is racy by nature, but missing a tick or
 5728	 * having one too much is no big deal because the scheduler tick updates
 5729	 * statistics and checks timeslices in a time-independent way, regardless
 5730	 * of when exactly it is running.
 5731	 */
 5732	if (tick_nohz_tick_stopped_cpu(cpu)) {
 5733		guard(rq_lock_irq)(rq);
 5734		struct task_struct *curr = rq->curr;
 5735
 5736		if (cpu_online(cpu)) {
 5737			/*
 5738			 * Since this is a remote tick for full dynticks mode,
 5739			 * we are always sure that there is no proxy (only a
 5740			 * single task is running).
 5741			 */
 5742			SCHED_WARN_ON(rq->curr != rq->donor);
 5743			update_rq_clock(rq);
 5744
 5745			if (!is_idle_task(curr)) {
 5746				/*
 5747				 * Make sure the next tick runs within a
 5748				 * reasonable amount of time.
 5749				 */
 5750				u64 delta = rq_clock_task(rq) - curr->se.exec_start;
 5751				WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
 5752			}
 5753			curr->sched_class->task_tick(rq, curr, 0);
 5754
 5755			calc_load_nohz_remote(rq);
 5756		}
 
 
 
 
 
 5757	}
 
 
 
 
 
 
 5758
 5759	/*
 5760	 * Run the remote tick once per second (1Hz). This arbitrary
 5761	 * frequency is large enough to avoid overload but short enough
 5762	 * to keep scheduler internal stats reasonably up to date.  But
 5763	 * first update state to reflect hotplug activity if required.
 5764	 */
 5765	os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
 5766	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
 5767	if (os == TICK_SCHED_REMOTE_RUNNING)
 5768		queue_delayed_work(system_unbound_wq, dwork, HZ);
 5769}
 5770
 5771static void sched_tick_start(int cpu)
 5772{
 5773	int os;
 5774	struct tick_work *twork;
 5775
 5776	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
 5777		return;
 5778
 5779	WARN_ON_ONCE(!tick_work_cpu);
 5780
 5781	twork = per_cpu_ptr(tick_work_cpu, cpu);
 5782	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
 5783	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
 5784	if (os == TICK_SCHED_REMOTE_OFFLINE) {
 5785		twork->cpu = cpu;
 5786		INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
 5787		queue_delayed_work(system_unbound_wq, &twork->work, HZ);
 5788	}
 5789}
 5790
 5791#ifdef CONFIG_HOTPLUG_CPU
 5792static void sched_tick_stop(int cpu)
 5793{
 5794	struct tick_work *twork;
 5795	int os;
 5796
 5797	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
 5798		return;
 5799
 5800	WARN_ON_ONCE(!tick_work_cpu);
 5801
 5802	twork = per_cpu_ptr(tick_work_cpu, cpu);
 5803	/* There cannot be competing actions, but don't rely on stop-machine. */
 5804	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
 5805	WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
 5806	/* Don't cancel, as this would mess up the state machine. */
 5807}
 5808#endif /* CONFIG_HOTPLUG_CPU */
 5809
 5810int __init sched_tick_offload_init(void)
 5811{
 5812	tick_work_cpu = alloc_percpu(struct tick_work);
 5813	BUG_ON(!tick_work_cpu);
 5814	return 0;
 5815}
 5816
 5817#else /* !CONFIG_NO_HZ_FULL */
 5818static inline void sched_tick_start(int cpu) { }
 5819static inline void sched_tick_stop(int cpu) { }
 5820#endif
 5821
 5822#if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
 5823				defined(CONFIG_TRACE_PREEMPT_TOGGLE))
 5824/*
 5825 * If the value passed in is equal to the current preempt count
 5826 * then we just disabled preemption. Start timing the latency.
 5827 */
 5828static inline void preempt_latency_start(int val)
 5829{
 5830	if (preempt_count() == val) {
 5831		unsigned long ip = get_lock_parent_ip();
 5832#ifdef CONFIG_DEBUG_PREEMPT
 5833		current->preempt_disable_ip = ip;
 5834#endif
 5835		trace_preempt_off(CALLER_ADDR0, ip);
 5836	}
 5837}
 5838
 5839void preempt_count_add(int val)
 5840{
 5841#ifdef CONFIG_DEBUG_PREEMPT
 5842	/*
 5843	 * Underflow?
 5844	 */
 5845	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
 5846		return;
 5847#endif
 5848	__preempt_count_add(val);
 5849#ifdef CONFIG_DEBUG_PREEMPT
 5850	/*
 5851	 * Spinlock count overflowing soon?
 5852	 */
 5853	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
 5854				PREEMPT_MASK - 10);
 5855#endif
 5856	preempt_latency_start(val);
 5857}
 5858EXPORT_SYMBOL(preempt_count_add);
 5859NOKPROBE_SYMBOL(preempt_count_add);
 5860
 5861/*
 5862 * If the value passed in equals to the current preempt count
 5863 * then we just enabled preemption. Stop timing the latency.
 5864 */
 5865static inline void preempt_latency_stop(int val)
 5866{
 5867	if (preempt_count() == val)
 5868		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
 5869}
 5870
 5871void preempt_count_sub(int val)
 5872{
 5873#ifdef CONFIG_DEBUG_PREEMPT
 5874	/*
 5875	 * Underflow?
 5876	 */
 5877	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
 5878		return;
 5879	/*
 5880	 * Is the spinlock portion underflowing?
 5881	 */
 5882	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
 5883			!(preempt_count() & PREEMPT_MASK)))
 5884		return;
 5885#endif
 5886
 5887	preempt_latency_stop(val);
 5888	__preempt_count_sub(val);
 5889}
 5890EXPORT_SYMBOL(preempt_count_sub);
 5891NOKPROBE_SYMBOL(preempt_count_sub);
 5892
 5893#else
 5894static inline void preempt_latency_start(int val) { }
 5895static inline void preempt_latency_stop(int val) { }
 5896#endif
 5897
 5898static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
 5899{
 5900#ifdef CONFIG_DEBUG_PREEMPT
 5901	return p->preempt_disable_ip;
 5902#else
 5903	return 0;
 5904#endif
 5905}
 5906
 5907/*
 5908 * Print scheduling while atomic bug:
 5909 */
 5910static noinline void __schedule_bug(struct task_struct *prev)
 5911{
 5912	/* Save this before calling printk(), since that will clobber it */
 5913	unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
 5914
 5915	if (oops_in_progress)
 5916		return;
 5917
 5918	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
 5919		prev->comm, prev->pid, preempt_count());
 5920
 5921	debug_show_held_locks(prev);
 5922	print_modules();
 5923	if (irqs_disabled())
 5924		print_irqtrace_events(prev);
 5925	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
 
 5926		pr_err("Preemption disabled at:");
 5927		print_ip_sym(KERN_ERR, preempt_disable_ip);
 5928	}
 5929	check_panic_on_warn("scheduling while atomic");
 
 5930
 5931	dump_stack();
 5932	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 5933}
 5934
 5935/*
 5936 * Various schedule()-time debugging checks and statistics:
 5937 */
 5938static inline void schedule_debug(struct task_struct *prev, bool preempt)
 5939{
 5940#ifdef CONFIG_SCHED_STACK_END_CHECK
 5941	if (task_stack_end_corrupted(prev))
 5942		panic("corrupted stack end detected inside scheduler\n");
 5943
 5944	if (task_scs_end_corrupted(prev))
 5945		panic("corrupted shadow stack detected inside scheduler\n");
 5946#endif
 5947
 5948#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
 5949	if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
 5950		printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
 5951			prev->comm, prev->pid, prev->non_block_count);
 5952		dump_stack();
 5953		add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 5954	}
 5955#endif
 5956
 5957	if (unlikely(in_atomic_preempt_off())) {
 5958		__schedule_bug(prev);
 5959		preempt_count_set(PREEMPT_DISABLED);
 5960	}
 5961	rcu_sleep_check();
 5962	SCHED_WARN_ON(ct_state() == CT_STATE_USER);
 5963
 5964	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
 5965
 5966	schedstat_inc(this_rq()->sched_count);
 5967}
 5968
 5969static void prev_balance(struct rq *rq, struct task_struct *prev,
 5970			 struct rq_flags *rf)
 5971{
 5972	const struct sched_class *start_class = prev->sched_class;
 5973	const struct sched_class *class;
 5974
 5975#ifdef CONFIG_SCHED_CLASS_EXT
 5976	/*
 5977	 * SCX requires a balance() call before every pick_task() including when
 5978	 * waking up from SCHED_IDLE. If @start_class is below SCX, start from
 5979	 * SCX instead. Also, set a flag to detect missing balance() call.
 5980	 */
 5981	if (scx_enabled()) {
 5982		rq->scx.flags |= SCX_RQ_BAL_PENDING;
 5983		if (sched_class_above(&ext_sched_class, start_class))
 5984			start_class = &ext_sched_class;
 5985	}
 5986#endif
 5987
 5988	/*
 5989	 * We must do the balancing pass before put_prev_task(), such
 5990	 * that when we release the rq->lock the task is in the same
 5991	 * state as before we took rq->lock.
 5992	 *
 5993	 * We can terminate the balance pass as soon as we know there is
 5994	 * a runnable task of @class priority or higher.
 5995	 */
 5996	for_active_class_range(class, start_class, &idle_sched_class) {
 5997		if (class->balance && class->balance(rq, prev, rf))
 5998			break;
 5999	}
 
 
 
 6000}
 6001
 6002/*
 6003 * Pick up the highest-prio task:
 6004 */
 6005static inline struct task_struct *
 6006__pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
 6007{
 6008	const struct sched_class *class;
 6009	struct task_struct *p;
 6010
 6011	rq->dl_server = NULL;
 6012
 6013	if (scx_enabled())
 6014		goto restart;
 6015
 6016	/*
 6017	 * Optimization: we know that if all tasks are in the fair class we can
 6018	 * call that function directly, but only if the @prev task wasn't of a
 6019	 * higher scheduling class, because otherwise those lose the
 6020	 * opportunity to pull in more work from other CPUs.
 6021	 */
 6022	if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
 6023		   rq->nr_running == rq->cfs.h_nr_running)) {
 6024
 6025		p = pick_next_task_fair(rq, prev, rf);
 6026		if (unlikely(p == RETRY_TASK))
 6027			goto restart;
 6028
 6029		/* Assume the next prioritized class is idle_sched_class */
 6030		if (!p) {
 6031			p = pick_task_idle(rq);
 6032			put_prev_set_next_task(rq, prev, p);
 6033		}
 6034
 6035		return p;
 6036	}
 6037
 6038restart:
 6039	prev_balance(rq, prev, rf);
 6040
 6041	for_each_active_class(class) {
 6042		if (class->pick_next_task) {
 6043			p = class->pick_next_task(rq, prev);
 6044			if (p)
 6045				return p;
 6046		} else {
 6047			p = class->pick_task(rq);
 6048			if (p) {
 6049				put_prev_set_next_task(rq, prev, p);
 6050				return p;
 6051			}
 6052		}
 6053	}
 6054
 6055	BUG(); /* The idle class should always have a runnable task. */
 
 6056}
 6057
 6058#ifdef CONFIG_SCHED_CORE
 6059static inline bool is_task_rq_idle(struct task_struct *t)
 6060{
 6061	return (task_rq(t)->idle == t);
 6062}
 6063
 6064static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
 6065{
 6066	return is_task_rq_idle(a) || (a->core_cookie == cookie);
 6067}
 6068
 6069static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
 6070{
 6071	if (is_task_rq_idle(a) || is_task_rq_idle(b))
 6072		return true;
 6073
 6074	return a->core_cookie == b->core_cookie;
 6075}
 6076
 6077static inline struct task_struct *pick_task(struct rq *rq)
 
 
 
 
 
 
 
 
 
 6078{
 6079	const struct sched_class *class;
 6080	struct task_struct *p;
 6081
 6082	rq->dl_server = NULL;
 
 
 6083
 6084	for_each_active_class(class) {
 6085		p = class->pick_task(rq);
 6086		if (p)
 6087			return p;
 
 
 
 
 
 
 6088	}
 6089
 6090	BUG(); /* The idle class should always have a runnable task. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 6091}
 6092
 6093extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
 6094
 6095static void queue_core_balance(struct rq *rq);
 6096
 6097static struct task_struct *
 6098pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
 6099{
 6100	struct task_struct *next, *p, *max = NULL;
 
 6101	const struct cpumask *smt_mask;
 6102	bool fi_before = false;
 6103	bool core_clock_updated = (rq == rq->core);
 6104	unsigned long cookie;
 6105	int i, cpu, occ = 0;
 6106	struct rq *rq_i;
 6107	bool need_sync;
 6108
 6109	if (!sched_core_enabled(rq))
 6110		return __pick_next_task(rq, prev, rf);
 6111
 6112	cpu = cpu_of(rq);
 6113
 6114	/* Stopper task is switching into idle, no need core-wide selection. */
 6115	if (cpu_is_offline(cpu)) {
 6116		/*
 6117		 * Reset core_pick so that we don't enter the fastpath when
 6118		 * coming online. core_pick would already be migrated to
 6119		 * another cpu during offline.
 6120		 */
 6121		rq->core_pick = NULL;
 6122		rq->core_dl_server = NULL;
 6123		return __pick_next_task(rq, prev, rf);
 6124	}
 6125
 6126	/*
 6127	 * If there were no {en,de}queues since we picked (IOW, the task
 6128	 * pointers are all still valid), and we haven't scheduled the last
 6129	 * pick yet, do so now.
 6130	 *
 6131	 * rq->core_pick can be NULL if no selection was made for a CPU because
 6132	 * it was either offline or went offline during a sibling's core-wide
 6133	 * selection. In this case, do a core-wide selection.
 6134	 */
 6135	if (rq->core->core_pick_seq == rq->core->core_task_seq &&
 6136	    rq->core->core_pick_seq != rq->core_sched_seq &&
 6137	    rq->core_pick) {
 6138		WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
 6139
 6140		next = rq->core_pick;
 6141		rq->dl_server = rq->core_dl_server;
 
 
 
 
 6142		rq->core_pick = NULL;
 6143		rq->core_dl_server = NULL;
 6144		goto out_set_next;
 6145	}
 6146
 6147	prev_balance(rq, prev, rf);
 6148
 6149	smt_mask = cpu_smt_mask(cpu);
 6150	need_sync = !!rq->core->core_cookie;
 6151
 6152	/* reset state */
 6153	rq->core->core_cookie = 0UL;
 6154	if (rq->core->core_forceidle_count) {
 6155		if (!core_clock_updated) {
 6156			update_rq_clock(rq->core);
 6157			core_clock_updated = true;
 6158		}
 6159		sched_core_account_forceidle(rq);
 6160		/* reset after accounting force idle */
 6161		rq->core->core_forceidle_start = 0;
 6162		rq->core->core_forceidle_count = 0;
 6163		rq->core->core_forceidle_occupation = 0;
 6164		need_sync = true;
 6165		fi_before = true;
 
 6166	}
 6167
 6168	/*
 6169	 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
 6170	 *
 6171	 * @task_seq guards the task state ({en,de}queues)
 6172	 * @pick_seq is the @task_seq we did a selection on
 6173	 * @sched_seq is the @pick_seq we scheduled
 6174	 *
 6175	 * However, preemptions can cause multiple picks on the same task set.
 6176	 * 'Fix' this by also increasing @task_seq for every pick.
 6177	 */
 6178	rq->core->core_task_seq++;
 6179
 6180	/*
 6181	 * Optimize for common case where this CPU has no cookies
 6182	 * and there are no cookied tasks running on siblings.
 6183	 */
 6184	if (!need_sync) {
 6185		next = pick_task(rq);
 
 
 
 
 
 6186		if (!next->core_cookie) {
 6187			rq->core_pick = NULL;
 6188			rq->core_dl_server = NULL;
 6189			/*
 6190			 * For robustness, update the min_vruntime_fi for
 6191			 * unconstrained picks as well.
 6192			 */
 6193			WARN_ON_ONCE(fi_before);
 6194			task_vruntime_update(rq, next, false);
 6195			goto out_set_next;
 6196		}
 6197	}
 6198
 6199	/*
 6200	 * For each thread: do the regular task pick and find the max prio task
 6201	 * amongst them.
 6202	 *
 6203	 * Tie-break prio towards the current CPU
 6204	 */
 6205	for_each_cpu_wrap(i, smt_mask, cpu) {
 6206		rq_i = cpu_rq(i);
 6207
 6208		/*
 6209		 * Current cpu always has its clock updated on entrance to
 6210		 * pick_next_task(). If the current cpu is not the core,
 6211		 * the core may also have been updated above.
 6212		 */
 6213		if (i != cpu && (rq_i != rq->core || !core_clock_updated))
 6214			update_rq_clock(rq_i);
 6215
 6216		rq_i->core_pick = p = pick_task(rq_i);
 6217		rq_i->core_dl_server = rq_i->dl_server;
 6218
 6219		if (!max || prio_less(max, p, fi_before))
 6220			max = p;
 6221	}
 6222
 6223	cookie = rq->core->core_cookie = max->core_cookie;
 6224
 6225	/*
 6226	 * For each thread: try and find a runnable task that matches @max or
 6227	 * force idle.
 6228	 */
 6229	for_each_cpu(i, smt_mask) {
 6230		rq_i = cpu_rq(i);
 6231		p = rq_i->core_pick;
 
 
 
 
 
 6232
 6233		if (!cookie_equals(p, cookie)) {
 6234			p = NULL;
 6235			if (cookie)
 6236				p = sched_core_find(rq_i, cookie);
 
 
 
 6237			if (!p)
 6238				p = idle_sched_class.pick_task(rq_i);
 6239		}
 6240
 6241		rq_i->core_pick = p;
 6242		rq_i->core_dl_server = NULL;
 6243
 6244		if (p == rq_i->idle) {
 6245			if (rq_i->nr_running) {
 6246				rq->core->core_forceidle_count++;
 6247				if (!fi_before)
 6248					rq->core->core_forceidle_seq++;
 6249			}
 6250		} else {
 6251			occ++;
 6252		}
 6253	}
 6254
 6255	if (schedstat_enabled() && rq->core->core_forceidle_count) {
 6256		rq->core->core_forceidle_start = rq_clock(rq->core);
 6257		rq->core->core_forceidle_occupation = occ;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 6258	}
 6259
 6260	rq->core->core_pick_seq = rq->core->core_task_seq;
 6261	next = rq->core_pick;
 6262	rq->core_sched_seq = rq->core->core_pick_seq;
 6263
 6264	/* Something should have been selected for current CPU */
 6265	WARN_ON_ONCE(!next);
 6266
 6267	/*
 6268	 * Reschedule siblings
 6269	 *
 6270	 * NOTE: L1TF -- at this point we're no longer running the old task and
 6271	 * sending an IPI (below) ensures the sibling will no longer be running
 6272	 * their task. This ensures there is no inter-sibling overlap between
 6273	 * non-matching user state.
 6274	 */
 6275	for_each_cpu(i, smt_mask) {
 6276		rq_i = cpu_rq(i);
 6277
 6278		/*
 6279		 * An online sibling might have gone offline before a task
 6280		 * could be picked for it, or it might be offline but later
 6281		 * happen to come online, but its too late and nothing was
 6282		 * picked for it.  That's Ok - it will pick tasks for itself,
 6283		 * so ignore it.
 6284		 */
 6285		if (!rq_i->core_pick)
 6286			continue;
 6287
 6288		/*
 6289		 * Update for new !FI->FI transitions, or if continuing to be in !FI:
 6290		 * fi_before     fi      update?
 6291		 *  0            0       1
 6292		 *  0            1       1
 6293		 *  1            0       1
 6294		 *  1            1       0
 6295		 */
 6296		if (!(fi_before && rq->core->core_forceidle_count))
 6297			task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
 6298
 6299		rq_i->core_pick->core_occupation = occ;
 6300
 6301		if (i == cpu) {
 6302			rq_i->core_pick = NULL;
 6303			rq_i->core_dl_server = NULL;
 6304			continue;
 6305		}
 6306
 6307		/* Did we break L1TF mitigation requirements? */
 6308		WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
 6309
 6310		if (rq_i->curr == rq_i->core_pick) {
 6311			rq_i->core_pick = NULL;
 6312			rq_i->core_dl_server = NULL;
 6313			continue;
 6314		}
 6315
 6316		resched_curr(rq_i);
 6317	}
 6318
 6319out_set_next:
 6320	put_prev_set_next_task(rq, prev, next);
 6321	if (rq->core->core_forceidle_count && next == rq->idle)
 6322		queue_core_balance(rq);
 6323
 6324	return next;
 6325}
 6326
 6327static bool try_steal_cookie(int this, int that)
 6328{
 6329	struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
 6330	struct task_struct *p;
 6331	unsigned long cookie;
 6332	bool success = false;
 6333
 6334	guard(irq)();
 6335	guard(double_rq_lock)(dst, src);
 6336
 6337	cookie = dst->core->core_cookie;
 6338	if (!cookie)
 6339		return false;
 6340
 6341	if (dst->curr != dst->idle)
 6342		return false;
 6343
 6344	p = sched_core_find(src, cookie);
 6345	if (!p)
 6346		return false;
 6347
 6348	do {
 6349		if (p == src->core_pick || p == src->curr)
 6350			goto next;
 6351
 6352		if (!is_cpu_allowed(p, this))
 6353			goto next;
 6354
 6355		if (p->core_occupation > dst->idle->core_occupation)
 6356			goto next;
 6357		/*
 6358		 * sched_core_find() and sched_core_next() will ensure
 6359		 * that task @p is not throttled now, we also need to
 6360		 * check whether the runqueue of the destination CPU is
 6361		 * being throttled.
 6362		 */
 6363		if (sched_task_is_throttled(p, this))
 6364			goto next;
 6365
 6366		move_queued_task_locked(src, dst, p);
 
 
 
 
 
 6367		resched_curr(dst);
 6368
 6369		success = true;
 6370		break;
 6371
 6372next:
 6373		p = sched_core_next(p, cookie);
 6374	} while (p);
 6375
 
 
 
 
 6376	return success;
 6377}
 6378
 6379static bool steal_cookie_task(int cpu, struct sched_domain *sd)
 6380{
 6381	int i;
 6382
 6383	for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) {
 6384		if (i == cpu)
 6385			continue;
 6386
 6387		if (need_resched())
 6388			break;
 6389
 6390		if (try_steal_cookie(cpu, i))
 6391			return true;
 6392	}
 6393
 6394	return false;
 6395}
 6396
 6397static void sched_core_balance(struct rq *rq)
 6398{
 6399	struct sched_domain *sd;
 6400	int cpu = cpu_of(rq);
 6401
 6402	guard(preempt)();
 6403	guard(rcu)();
 6404
 6405	raw_spin_rq_unlock_irq(rq);
 6406	for_each_domain(cpu, sd) {
 6407		if (need_resched())
 6408			break;
 6409
 6410		if (steal_cookie_task(cpu, sd))
 6411			break;
 6412	}
 6413	raw_spin_rq_lock_irq(rq);
 
 
 6414}
 6415
 6416static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
 6417
 6418static void queue_core_balance(struct rq *rq)
 6419{
 6420	if (!sched_core_enabled(rq))
 6421		return;
 6422
 6423	if (!rq->core->core_cookie)
 6424		return;
 6425
 6426	if (!rq->nr_running) /* not forced idle */
 6427		return;
 6428
 6429	queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
 6430}
 6431
 6432DEFINE_LOCK_GUARD_1(core_lock, int,
 6433		    sched_core_lock(*_T->lock, &_T->flags),
 6434		    sched_core_unlock(*_T->lock, &_T->flags),
 6435		    unsigned long flags)
 6436
 6437static void sched_core_cpu_starting(unsigned int cpu)
 6438{
 6439	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
 6440	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
 
 6441	int t;
 6442
 6443	guard(core_lock)(&cpu);
 6444
 6445	WARN_ON_ONCE(rq->core != rq);
 6446
 6447	/* if we're the first, we'll be our own leader */
 6448	if (cpumask_weight(smt_mask) == 1)
 6449		return;
 6450
 6451	/* find the leader */
 6452	for_each_cpu(t, smt_mask) {
 6453		if (t == cpu)
 6454			continue;
 6455		rq = cpu_rq(t);
 6456		if (rq->core == rq) {
 6457			core_rq = rq;
 6458			break;
 6459		}
 6460	}
 6461
 6462	if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
 6463		return;
 6464
 6465	/* install and validate core_rq */
 6466	for_each_cpu(t, smt_mask) {
 6467		rq = cpu_rq(t);
 6468
 6469		if (t == cpu)
 6470			rq->core = core_rq;
 6471
 6472		WARN_ON_ONCE(rq->core != core_rq);
 6473	}
 
 
 
 6474}
 6475
 6476static void sched_core_cpu_deactivate(unsigned int cpu)
 6477{
 6478	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
 6479	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
 
 6480	int t;
 6481
 6482	guard(core_lock)(&cpu);
 6483
 6484	/* if we're the last man standing, nothing to do */
 6485	if (cpumask_weight(smt_mask) == 1) {
 6486		WARN_ON_ONCE(rq->core != rq);
 6487		return;
 6488	}
 6489
 6490	/* if we're not the leader, nothing to do */
 6491	if (rq->core != rq)
 6492		return;
 6493
 6494	/* find a new leader */
 6495	for_each_cpu(t, smt_mask) {
 6496		if (t == cpu)
 6497			continue;
 6498		core_rq = cpu_rq(t);
 6499		break;
 6500	}
 6501
 6502	if (WARN_ON_ONCE(!core_rq)) /* impossible */
 6503		return;
 6504
 6505	/* copy the shared state to the new leader */
 6506	core_rq->core_task_seq             = rq->core_task_seq;
 6507	core_rq->core_pick_seq             = rq->core_pick_seq;
 6508	core_rq->core_cookie               = rq->core_cookie;
 6509	core_rq->core_forceidle_count      = rq->core_forceidle_count;
 6510	core_rq->core_forceidle_seq        = rq->core_forceidle_seq;
 6511	core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
 6512
 6513	/*
 6514	 * Accounting edge for forced idle is handled in pick_next_task().
 6515	 * Don't need another one here, since the hotplug thread shouldn't
 6516	 * have a cookie.
 6517	 */
 6518	core_rq->core_forceidle_start = 0;
 6519
 6520	/* install new leader */
 6521	for_each_cpu(t, smt_mask) {
 6522		rq = cpu_rq(t);
 6523		rq->core = core_rq;
 6524	}
 
 
 
 6525}
 6526
 6527static inline void sched_core_cpu_dying(unsigned int cpu)
 6528{
 6529	struct rq *rq = cpu_rq(cpu);
 6530
 6531	if (rq->core != rq)
 6532		rq->core = rq;
 6533}
 6534
 6535#else /* !CONFIG_SCHED_CORE */
 6536
 6537static inline void sched_core_cpu_starting(unsigned int cpu) {}
 6538static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
 6539static inline void sched_core_cpu_dying(unsigned int cpu) {}
 6540
 6541static struct task_struct *
 6542pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
 6543{
 6544	return __pick_next_task(rq, prev, rf);
 6545}
 6546
 6547#endif /* CONFIG_SCHED_CORE */
 6548
 6549/*
 6550 * Constants for the sched_mode argument of __schedule().
 6551 *
 6552 * The mode argument allows RT enabled kernels to differentiate a
 6553 * preemption from blocking on an 'sleeping' spin/rwlock.
 6554 */
 6555#define SM_IDLE			(-1)
 6556#define SM_NONE			0
 6557#define SM_PREEMPT		1
 6558#define SM_RTLOCK_WAIT		2
 6559
 6560/*
 6561 * Helper function for __schedule()
 6562 *
 6563 * If a task does not have signals pending, deactivate it
 6564 * Otherwise marks the task's __state as RUNNING
 6565 */
 6566static bool try_to_block_task(struct rq *rq, struct task_struct *p,
 6567			      unsigned long task_state)
 6568{
 6569	int flags = DEQUEUE_NOCLOCK;
 6570
 6571	if (signal_pending_state(task_state, p)) {
 6572		WRITE_ONCE(p->__state, TASK_RUNNING);
 6573		return false;
 6574	}
 6575
 6576	p->sched_contributes_to_load =
 6577		(task_state & TASK_UNINTERRUPTIBLE) &&
 6578		!(task_state & TASK_NOLOAD) &&
 6579		!(task_state & TASK_FROZEN);
 6580
 6581	if (unlikely(is_special_task_state(task_state)))
 6582		flags |= DEQUEUE_SPECIAL;
 6583
 6584	/*
 6585	 * __schedule()			ttwu()
 6586	 *   prev_state = prev->state;    if (p->on_rq && ...)
 6587	 *   if (prev_state)		    goto out;
 6588	 *     p->on_rq = 0;		  smp_acquire__after_ctrl_dep();
 6589	 *				  p->state = TASK_WAKING
 6590	 *
 6591	 * Where __schedule() and ttwu() have matching control dependencies.
 6592	 *
 6593	 * After this, schedule() must not care about p->state any more.
 6594	 */
 6595	block_task(rq, p, flags);
 6596	return true;
 6597}
 6598
 6599/*
 6600 * __schedule() is the main scheduler function.
 6601 *
 6602 * The main means of driving the scheduler and thus entering this function are:
 6603 *
 6604 *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
 6605 *
 6606 *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
 6607 *      paths. For example, see arch/x86/entry_64.S.
 6608 *
 6609 *      To drive preemption between tasks, the scheduler sets the flag in timer
 6610 *      interrupt handler sched_tick().
 6611 *
 6612 *   3. Wakeups don't really cause entry into schedule(). They add a
 6613 *      task to the run-queue and that's it.
 6614 *
 6615 *      Now, if the new task added to the run-queue preempts the current
 6616 *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
 6617 *      called on the nearest possible occasion:
 6618 *
 6619 *       - If the kernel is preemptible (CONFIG_PREEMPTION=y):
 6620 *
 6621 *         - in syscall or exception context, at the next outmost
 6622 *           preempt_enable(). (this might be as soon as the wake_up()'s
 6623 *           spin_unlock()!)
 6624 *
 6625 *         - in IRQ context, return from interrupt-handler to
 6626 *           preemptible context
 6627 *
 6628 *       - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
 6629 *         then at the next:
 6630 *
 6631 *          - cond_resched() call
 6632 *          - explicit schedule() call
 6633 *          - return from syscall or exception to user-space
 6634 *          - return from interrupt-handler to user-space
 6635 *
 6636 * WARNING: must be called with preemption disabled!
 6637 */
 6638static void __sched notrace __schedule(int sched_mode)
 6639{
 6640	struct task_struct *prev, *next;
 6641	/*
 6642	 * On PREEMPT_RT kernel, SM_RTLOCK_WAIT is noted
 6643	 * as a preemption by schedule_debug() and RCU.
 6644	 */
 6645	bool preempt = sched_mode > SM_NONE;
 6646	unsigned long *switch_count;
 6647	unsigned long prev_state;
 6648	struct rq_flags rf;
 6649	struct rq *rq;
 6650	int cpu;
 6651
 6652	cpu = smp_processor_id();
 6653	rq = cpu_rq(cpu);
 6654	prev = rq->curr;
 6655
 6656	schedule_debug(prev, preempt);
 6657
 6658	if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
 6659		hrtick_clear(rq);
 6660
 6661	local_irq_disable();
 6662	rcu_note_context_switch(preempt);
 6663
 6664	/*
 6665	 * Make sure that signal_pending_state()->signal_pending() below
 6666	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
 6667	 * done by the caller to avoid the race with signal_wake_up():
 6668	 *
 6669	 * __set_current_state(@state)		signal_wake_up()
 6670	 * schedule()				  set_tsk_thread_flag(p, TIF_SIGPENDING)
 6671	 *					  wake_up_state(p, state)
 6672	 *   LOCK rq->lock			    LOCK p->pi_state
 6673	 *   smp_mb__after_spinlock()		    smp_mb__after_spinlock()
 6674	 *     if (signal_pending_state())	    if (p->state & @state)
 6675	 *
 6676	 * Also, the membarrier system call requires a full memory barrier
 6677	 * after coming from user-space, before storing to rq->curr; this
 6678	 * barrier matches a full barrier in the proximity of the membarrier
 6679	 * system call exit.
 6680	 */
 6681	rq_lock(rq, &rf);
 6682	smp_mb__after_spinlock();
 6683
 6684	/* Promote REQ to ACT */
 6685	rq->clock_update_flags <<= 1;
 6686	update_rq_clock(rq);
 6687	rq->clock_update_flags = RQCF_UPDATED;
 6688
 6689	switch_count = &prev->nivcsw;
 6690
 6691	/* Task state changes only considers SM_PREEMPT as preemption */
 6692	preempt = sched_mode == SM_PREEMPT;
 6693
 6694	/*
 6695	 * We must load prev->state once (task_struct::state is volatile), such
 6696	 * that we form a control dependency vs deactivate_task() below.
 
 
 
 6697	 */
 6698	prev_state = READ_ONCE(prev->__state);
 6699	if (sched_mode == SM_IDLE) {
 6700		/* SCX must consult the BPF scheduler to tell if rq is empty */
 6701		if (!rq->nr_running && !scx_enabled()) {
 6702			next = prev;
 6703			goto picked;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 6704		}
 6705	} else if (!preempt && prev_state) {
 6706		try_to_block_task(rq, prev, prev_state);
 6707		switch_count = &prev->nvcsw;
 6708	}
 6709
 6710	next = pick_next_task(rq, prev, &rf);
 6711	rq_set_donor(rq, next);
 6712picked:
 6713	clear_tsk_need_resched(prev);
 6714	clear_preempt_need_resched();
 6715#ifdef CONFIG_SCHED_DEBUG
 6716	rq->last_seen_need_resched_ns = 0;
 6717#endif
 6718
 6719	if (likely(prev != next)) {
 6720		rq->nr_switches++;
 6721		/*
 6722		 * RCU users of rcu_dereference(rq->curr) may not see
 6723		 * changes to task_struct made by pick_next_task().
 6724		 */
 6725		RCU_INIT_POINTER(rq->curr, next);
 6726		/*
 6727		 * The membarrier system call requires each architecture
 6728		 * to have a full memory barrier after updating
 6729		 * rq->curr, before returning to user-space.
 6730		 *
 6731		 * Here are the schemes providing that barrier on the
 6732		 * various architectures:
 6733		 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC,
 6734		 *   RISC-V.  switch_mm() relies on membarrier_arch_switch_mm()
 6735		 *   on PowerPC and on RISC-V.
 6736		 * - finish_lock_switch() for weakly-ordered
 6737		 *   architectures where spin_unlock is a full barrier,
 6738		 * - switch_to() for arm64 (weakly-ordered, spin_unlock
 6739		 *   is a RELEASE barrier),
 6740		 *
 6741		 * The barrier matches a full barrier in the proximity of
 6742		 * the membarrier system call entry.
 6743		 *
 6744		 * On RISC-V, this barrier pairing is also needed for the
 6745		 * SYNC_CORE command when switching between processes, cf.
 6746		 * the inline comments in membarrier_arch_switch_mm().
 6747		 */
 6748		++*switch_count;
 6749
 6750		migrate_disable_switch(rq, prev);
 6751		psi_account_irqtime(rq, prev, next);
 6752		psi_sched_switch(prev, next, !task_on_rq_queued(prev) ||
 6753					     prev->se.sched_delayed);
 6754
 6755		trace_sched_switch(preempt, prev, next, prev_state);
 6756
 6757		/* Also unlocks the rq: */
 6758		rq = context_switch(rq, prev, next, &rf);
 6759	} else {
 
 
 6760		rq_unpin_lock(rq, &rf);
 6761		__balance_callbacks(rq);
 6762		raw_spin_rq_unlock_irq(rq);
 6763	}
 6764}
 6765
 6766void __noreturn do_task_dead(void)
 6767{
 6768	/* Causes final put_task_struct in finish_task_switch(): */
 6769	set_special_state(TASK_DEAD);
 6770
 6771	/* Tell freezer to ignore us: */
 6772	current->flags |= PF_NOFREEZE;
 6773
 6774	__schedule(SM_NONE);
 6775	BUG();
 6776
 6777	/* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
 6778	for (;;)
 6779		cpu_relax();
 6780}
 6781
 6782static inline void sched_submit_work(struct task_struct *tsk)
 6783{
 6784	static DEFINE_WAIT_OVERRIDE_MAP(sched_map, LD_WAIT_CONFIG);
 6785	unsigned int task_flags;
 6786
 6787	/*
 6788	 * Establish LD_WAIT_CONFIG context to ensure none of the code called
 6789	 * will use a blocking primitive -- which would lead to recursion.
 6790	 */
 6791	lock_map_acquire_try(&sched_map);
 6792
 6793	task_flags = tsk->flags;
 6794	/*
 6795	 * If a worker goes to sleep, notify and ask workqueue whether it
 6796	 * wants to wake up a task to maintain concurrency.
 
 
 
 
 6797	 */
 6798	if (task_flags & PF_WQ_WORKER)
 6799		wq_worker_sleeping(tsk);
 6800	else if (task_flags & PF_IO_WORKER)
 6801		io_wq_worker_sleeping(tsk);
 
 
 
 
 6802
 6803	/*
 6804	 * spinlock and rwlock must not flush block requests.  This will
 6805	 * deadlock if the callback attempts to acquire a lock which is
 6806	 * already acquired.
 6807	 */
 6808	SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
 6809
 6810	/*
 6811	 * If we are going to sleep and we have plugged IO queued,
 6812	 * make sure to submit it to avoid deadlocks.
 6813	 */
 6814	blk_flush_plug(tsk->plug, true);
 6815
 6816	lock_map_release(&sched_map);
 6817}
 6818
 6819static void sched_update_worker(struct task_struct *tsk)
 6820{
 6821	if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER | PF_BLOCK_TS)) {
 6822		if (tsk->flags & PF_BLOCK_TS)
 6823			blk_plug_invalidate_ts(tsk);
 6824		if (tsk->flags & PF_WQ_WORKER)
 6825			wq_worker_running(tsk);
 6826		else if (tsk->flags & PF_IO_WORKER)
 6827			io_wq_worker_running(tsk);
 6828	}
 6829}
 6830
 6831static __always_inline void __schedule_loop(int sched_mode)
 6832{
 
 
 
 6833	do {
 6834		preempt_disable();
 6835		__schedule(sched_mode);
 6836		sched_preempt_enable_no_resched();
 6837	} while (need_resched());
 6838}
 6839
 6840asmlinkage __visible void __sched schedule(void)
 6841{
 6842	struct task_struct *tsk = current;
 6843
 6844#ifdef CONFIG_RT_MUTEXES
 6845	lockdep_assert(!tsk->sched_rt_mutex);
 6846#endif
 6847
 6848	if (!task_is_running(tsk))
 6849		sched_submit_work(tsk);
 6850	__schedule_loop(SM_NONE);
 6851	sched_update_worker(tsk);
 6852}
 6853EXPORT_SYMBOL(schedule);
 6854
 6855/*
 6856 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
 6857 * state (have scheduled out non-voluntarily) by making sure that all
 6858 * tasks have either left the run queue or have gone into user space.
 6859 * As idle tasks do not do either, they must not ever be preempted
 6860 * (schedule out non-voluntarily).
 6861 *
 6862 * schedule_idle() is similar to schedule_preempt_disable() except that it
 6863 * never enables preemption because it does not call sched_submit_work().
 6864 */
 6865void __sched schedule_idle(void)
 6866{
 6867	/*
 6868	 * As this skips calling sched_submit_work(), which the idle task does
 6869	 * regardless because that function is a NOP when the task is in a
 6870	 * TASK_RUNNING state, make sure this isn't used someplace that the
 6871	 * current task can be in any other state. Note, idle is always in the
 6872	 * TASK_RUNNING state.
 6873	 */
 6874	WARN_ON_ONCE(current->__state);
 6875	do {
 6876		__schedule(SM_IDLE);
 6877	} while (need_resched());
 6878}
 6879
 6880#if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
 6881asmlinkage __visible void __sched schedule_user(void)
 6882{
 6883	/*
 6884	 * If we come here after a random call to set_need_resched(),
 6885	 * or we have been woken up remotely but the IPI has not yet arrived,
 6886	 * we haven't yet exited the RCU idle mode. Do it here manually until
 6887	 * we find a better solution.
 6888	 *
 6889	 * NB: There are buggy callers of this function.  Ideally we
 6890	 * should warn if prev_state != CT_STATE_USER, but that will trigger
 6891	 * too frequently to make sense yet.
 6892	 */
 6893	enum ctx_state prev_state = exception_enter();
 6894	schedule();
 6895	exception_exit(prev_state);
 6896}
 6897#endif
 6898
 6899/**
 6900 * schedule_preempt_disabled - called with preemption disabled
 6901 *
 6902 * Returns with preemption disabled. Note: preempt_count must be 1
 6903 */
 6904void __sched schedule_preempt_disabled(void)
 6905{
 6906	sched_preempt_enable_no_resched();
 6907	schedule();
 6908	preempt_disable();
 6909}
 6910
 6911#ifdef CONFIG_PREEMPT_RT
 6912void __sched notrace schedule_rtlock(void)
 6913{
 6914	__schedule_loop(SM_RTLOCK_WAIT);
 6915}
 6916NOKPROBE_SYMBOL(schedule_rtlock);
 6917#endif
 6918
 6919static void __sched notrace preempt_schedule_common(void)
 6920{
 6921	do {
 6922		/*
 6923		 * Because the function tracer can trace preempt_count_sub()
 6924		 * and it also uses preempt_enable/disable_notrace(), if
 6925		 * NEED_RESCHED is set, the preempt_enable_notrace() called
 6926		 * by the function tracer will call this function again and
 6927		 * cause infinite recursion.
 6928		 *
 6929		 * Preemption must be disabled here before the function
 6930		 * tracer can trace. Break up preempt_disable() into two
 6931		 * calls. One to disable preemption without fear of being
 6932		 * traced. The other to still record the preemption latency,
 6933		 * which can also be traced by the function tracer.
 6934		 */
 6935		preempt_disable_notrace();
 6936		preempt_latency_start(1);
 6937		__schedule(SM_PREEMPT);
 6938		preempt_latency_stop(1);
 6939		preempt_enable_no_resched_notrace();
 6940
 6941		/*
 6942		 * Check again in case we missed a preemption opportunity
 6943		 * between schedule and now.
 6944		 */
 6945	} while (need_resched());
 6946}
 6947
 6948#ifdef CONFIG_PREEMPTION
 6949/*
 6950 * This is the entry point to schedule() from in-kernel preemption
 6951 * off of preempt_enable.
 6952 */
 6953asmlinkage __visible void __sched notrace preempt_schedule(void)
 6954{
 6955	/*
 6956	 * If there is a non-zero preempt_count or interrupts are disabled,
 6957	 * we do not want to preempt the current task. Just return..
 6958	 */
 6959	if (likely(!preemptible()))
 6960		return;
 
 6961	preempt_schedule_common();
 6962}
 6963NOKPROBE_SYMBOL(preempt_schedule);
 6964EXPORT_SYMBOL(preempt_schedule);
 6965
 6966#ifdef CONFIG_PREEMPT_DYNAMIC
 6967#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 6968#ifndef preempt_schedule_dynamic_enabled
 6969#define preempt_schedule_dynamic_enabled	preempt_schedule
 6970#define preempt_schedule_dynamic_disabled	NULL
 6971#endif
 6972DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
 6973EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
 6974#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 6975static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
 6976void __sched notrace dynamic_preempt_schedule(void)
 6977{
 6978	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
 6979		return;
 6980	preempt_schedule();
 6981}
 6982NOKPROBE_SYMBOL(dynamic_preempt_schedule);
 6983EXPORT_SYMBOL(dynamic_preempt_schedule);
 6984#endif
 6985#endif
 6986
 
 6987/**
 6988 * preempt_schedule_notrace - preempt_schedule called by tracing
 6989 *
 6990 * The tracing infrastructure uses preempt_enable_notrace to prevent
 6991 * recursion and tracing preempt enabling caused by the tracing
 6992 * infrastructure itself. But as tracing can happen in areas coming
 6993 * from userspace or just about to enter userspace, a preempt enable
 6994 * can occur before user_exit() is called. This will cause the scheduler
 6995 * to be called when the system is still in usermode.
 6996 *
 6997 * To prevent this, the preempt_enable_notrace will use this function
 6998 * instead of preempt_schedule() to exit user context if needed before
 6999 * calling the scheduler.
 7000 */
 7001asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
 7002{
 7003	enum ctx_state prev_ctx;
 7004
 7005	if (likely(!preemptible()))
 7006		return;
 7007
 7008	do {
 7009		/*
 7010		 * Because the function tracer can trace preempt_count_sub()
 7011		 * and it also uses preempt_enable/disable_notrace(), if
 7012		 * NEED_RESCHED is set, the preempt_enable_notrace() called
 7013		 * by the function tracer will call this function again and
 7014		 * cause infinite recursion.
 7015		 *
 7016		 * Preemption must be disabled here before the function
 7017		 * tracer can trace. Break up preempt_disable() into two
 7018		 * calls. One to disable preemption without fear of being
 7019		 * traced. The other to still record the preemption latency,
 7020		 * which can also be traced by the function tracer.
 7021		 */
 7022		preempt_disable_notrace();
 7023		preempt_latency_start(1);
 7024		/*
 7025		 * Needs preempt disabled in case user_exit() is traced
 7026		 * and the tracer calls preempt_enable_notrace() causing
 7027		 * an infinite recursion.
 7028		 */
 7029		prev_ctx = exception_enter();
 7030		__schedule(SM_PREEMPT);
 7031		exception_exit(prev_ctx);
 7032
 7033		preempt_latency_stop(1);
 7034		preempt_enable_no_resched_notrace();
 7035	} while (need_resched());
 7036}
 7037EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
 7038
 7039#ifdef CONFIG_PREEMPT_DYNAMIC
 7040#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 7041#ifndef preempt_schedule_notrace_dynamic_enabled
 7042#define preempt_schedule_notrace_dynamic_enabled	preempt_schedule_notrace
 7043#define preempt_schedule_notrace_dynamic_disabled	NULL
 7044#endif
 7045DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
 7046EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
 7047#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 7048static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
 7049void __sched notrace dynamic_preempt_schedule_notrace(void)
 7050{
 7051	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
 7052		return;
 7053	preempt_schedule_notrace();
 7054}
 7055NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
 7056EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
 7057#endif
 7058#endif
 7059
 7060#endif /* CONFIG_PREEMPTION */
 7061
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 7062/*
 7063 * This is the entry point to schedule() from kernel preemption
 7064 * off of IRQ context.
 7065 * Note, that this is called and return with IRQs disabled. This will
 7066 * protect us against recursive calling from IRQ contexts.
 7067 */
 7068asmlinkage __visible void __sched preempt_schedule_irq(void)
 7069{
 7070	enum ctx_state prev_state;
 7071
 7072	/* Catch callers which need to be fixed */
 7073	BUG_ON(preempt_count() || !irqs_disabled());
 7074
 7075	prev_state = exception_enter();
 7076
 7077	do {
 7078		preempt_disable();
 7079		local_irq_enable();
 7080		__schedule(SM_PREEMPT);
 7081		local_irq_disable();
 7082		sched_preempt_enable_no_resched();
 7083	} while (need_resched());
 7084
 7085	exception_exit(prev_state);
 7086}
 7087
 7088int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
 7089			  void *key)
 7090{
 7091	WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~(WF_SYNC|WF_CURRENT_CPU));
 7092	return try_to_wake_up(curr->private, mode, wake_flags);
 7093}
 7094EXPORT_SYMBOL(default_wake_function);
 7095
 7096const struct sched_class *__setscheduler_class(int policy, int prio)
 7097{
 7098	if (dl_prio(prio))
 7099		return &dl_sched_class;
 7100
 7101	if (rt_prio(prio))
 7102		return &rt_sched_class;
 7103
 7104#ifdef CONFIG_SCHED_CLASS_EXT
 7105	if (task_should_scx(policy))
 7106		return &ext_sched_class;
 7107#endif
 7108
 7109	return &fair_sched_class;
 7110}
 7111
 7112#ifdef CONFIG_RT_MUTEXES
 7113
 7114/*
 7115 * Would be more useful with typeof()/auto_type but they don't mix with
 7116 * bit-fields. Since it's a local thing, use int. Keep the generic sounding
 7117 * name such that if someone were to implement this function we get to compare
 7118 * notes.
 7119 */
 7120#define fetch_and_set(x, v) ({ int _x = (x); (x) = (v); _x; })
 7121
 7122void rt_mutex_pre_schedule(void)
 7123{
 7124	lockdep_assert(!fetch_and_set(current->sched_rt_mutex, 1));
 7125	sched_submit_work(current);
 7126}
 7127
 7128void rt_mutex_schedule(void)
 7129{
 7130	lockdep_assert(current->sched_rt_mutex);
 7131	__schedule_loop(SM_NONE);
 7132}
 7133
 7134void rt_mutex_post_schedule(void)
 7135{
 7136	sched_update_worker(current);
 7137	lockdep_assert(fetch_and_set(current->sched_rt_mutex, 0));
 
 7138}
 7139
 7140/*
 7141 * rt_mutex_setprio - set the current priority of a task
 7142 * @p: task to boost
 7143 * @pi_task: donor task
 7144 *
 7145 * This function changes the 'effective' priority of a task. It does
 7146 * not touch ->normal_prio like __setscheduler().
 7147 *
 7148 * Used by the rt_mutex code to implement priority inheritance
 7149 * logic. Call site only calls if the priority of the task changed.
 7150 */
 7151void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
 7152{
 7153	int prio, oldprio, queued, running, queue_flag =
 7154		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
 7155	const struct sched_class *prev_class, *next_class;
 7156	struct rq_flags rf;
 7157	struct rq *rq;
 7158
 7159	/* XXX used to be waiter->prio, not waiter->task->prio */
 7160	prio = __rt_effective_prio(pi_task, p->normal_prio);
 7161
 7162	/*
 7163	 * If nothing changed; bail early.
 7164	 */
 7165	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
 7166		return;
 7167
 7168	rq = __task_rq_lock(p, &rf);
 7169	update_rq_clock(rq);
 7170	/*
 7171	 * Set under pi_lock && rq->lock, such that the value can be used under
 7172	 * either lock.
 7173	 *
 7174	 * Note that there is loads of tricky to make this pointer cache work
 7175	 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
 7176	 * ensure a task is de-boosted (pi_task is set to NULL) before the
 7177	 * task is allowed to run again (and can exit). This ensures the pointer
 7178	 * points to a blocked task -- which guarantees the task is present.
 7179	 */
 7180	p->pi_top_task = pi_task;
 7181
 7182	/*
 7183	 * For FIFO/RR we only need to set prio, if that matches we're done.
 7184	 */
 7185	if (prio == p->prio && !dl_prio(prio))
 7186		goto out_unlock;
 7187
 7188	/*
 7189	 * Idle task boosting is a no-no in general. There is one
 7190	 * exception, when PREEMPT_RT and NOHZ is active:
 7191	 *
 7192	 * The idle task calls get_next_timer_interrupt() and holds
 7193	 * the timer wheel base->lock on the CPU and another CPU wants
 7194	 * to access the timer (probably to cancel it). We can safely
 7195	 * ignore the boosting request, as the idle CPU runs this code
 7196	 * with interrupts disabled and will complete the lock
 7197	 * protected section without being interrupted. So there is no
 7198	 * real need to boost.
 7199	 */
 7200	if (unlikely(p == rq->idle)) {
 7201		WARN_ON(p != rq->curr);
 7202		WARN_ON(p->pi_blocked_on);
 7203		goto out_unlock;
 7204	}
 7205
 7206	trace_sched_pi_setprio(p, pi_task);
 7207	oldprio = p->prio;
 7208
 7209	if (oldprio == prio)
 7210		queue_flag &= ~DEQUEUE_MOVE;
 7211
 7212	prev_class = p->sched_class;
 7213	next_class = __setscheduler_class(p->policy, prio);
 7214
 7215	if (prev_class != next_class && p->se.sched_delayed)
 7216		dequeue_task(rq, p, DEQUEUE_SLEEP | DEQUEUE_DELAYED | DEQUEUE_NOCLOCK);
 7217
 7218	queued = task_on_rq_queued(p);
 7219	running = task_current_donor(rq, p);
 7220	if (queued)
 7221		dequeue_task(rq, p, queue_flag);
 7222	if (running)
 7223		put_prev_task(rq, p);
 7224
 7225	/*
 7226	 * Boosting condition are:
 7227	 * 1. -rt task is running and holds mutex A
 7228	 *      --> -dl task blocks on mutex A
 7229	 *
 7230	 * 2. -dl task is running and holds mutex A
 7231	 *      --> -dl task blocks on mutex A and could preempt the
 7232	 *          running task
 7233	 */
 7234	if (dl_prio(prio)) {
 7235		if (!dl_prio(p->normal_prio) ||
 7236		    (pi_task && dl_prio(pi_task->prio) &&
 7237		     dl_entity_preempt(&pi_task->dl, &p->dl))) {
 7238			p->dl.pi_se = pi_task->dl.pi_se;
 7239			queue_flag |= ENQUEUE_REPLENISH;
 7240		} else {
 7241			p->dl.pi_se = &p->dl;
 7242		}
 7243	} else if (rt_prio(prio)) {
 7244		if (dl_prio(oldprio))
 7245			p->dl.pi_se = &p->dl;
 7246		if (oldprio < prio)
 7247			queue_flag |= ENQUEUE_HEAD;
 7248	} else {
 7249		if (dl_prio(oldprio))
 7250			p->dl.pi_se = &p->dl;
 7251		if (rt_prio(oldprio))
 7252			p->rt.timeout = 0;
 7253	}
 7254
 7255	p->sched_class = next_class;
 7256	p->prio = prio;
 7257
 7258	check_class_changing(rq, p, prev_class);
 7259
 7260	if (queued)
 7261		enqueue_task(rq, p, queue_flag);
 7262	if (running)
 7263		set_next_task(rq, p);
 7264
 7265	check_class_changed(rq, p, prev_class, oldprio);
 7266out_unlock:
 7267	/* Avoid rq from going away on us: */
 7268	preempt_disable();
 7269
 7270	rq_unpin_lock(rq, &rf);
 7271	__balance_callbacks(rq);
 7272	raw_spin_rq_unlock(rq);
 7273
 7274	preempt_enable();
 7275}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 7276#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 7277
 7278#if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
 7279int __sched __cond_resched(void)
 7280{
 7281	if (should_resched(0) && !irqs_disabled()) {
 7282		preempt_schedule_common();
 7283		return 1;
 7284	}
 7285	/*
 7286	 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
 7287	 * whether the current CPU is in an RCU read-side critical section,
 7288	 * so the tick can report quiescent states even for CPUs looping
 7289	 * in kernel context.  In contrast, in non-preemptible kernels,
 7290	 * RCU readers leave no in-memory hints, which means that CPU-bound
 7291	 * processes executing in kernel context might never report an
 7292	 * RCU quiescent state.  Therefore, the following code causes
 7293	 * cond_resched() to report a quiescent state, but only when RCU
 7294	 * is in urgent need of one.
 7295	 */
 7296#ifndef CONFIG_PREEMPT_RCU
 7297	rcu_all_qs();
 7298#endif
 7299	return 0;
 7300}
 7301EXPORT_SYMBOL(__cond_resched);
 7302#endif
 7303
 7304#ifdef CONFIG_PREEMPT_DYNAMIC
 7305#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 7306#define cond_resched_dynamic_enabled	__cond_resched
 7307#define cond_resched_dynamic_disabled	((void *)&__static_call_return0)
 7308DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
 7309EXPORT_STATIC_CALL_TRAMP(cond_resched);
 7310
 7311#define might_resched_dynamic_enabled	__cond_resched
 7312#define might_resched_dynamic_disabled	((void *)&__static_call_return0)
 7313DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
 7314EXPORT_STATIC_CALL_TRAMP(might_resched);
 7315#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 7316static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
 7317int __sched dynamic_cond_resched(void)
 7318{
 7319	klp_sched_try_switch();
 7320	if (!static_branch_unlikely(&sk_dynamic_cond_resched))
 7321		return 0;
 7322	return __cond_resched();
 7323}
 7324EXPORT_SYMBOL(dynamic_cond_resched);
 7325
 7326static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
 7327int __sched dynamic_might_resched(void)
 7328{
 7329	if (!static_branch_unlikely(&sk_dynamic_might_resched))
 7330		return 0;
 7331	return __cond_resched();
 7332}
 7333EXPORT_SYMBOL(dynamic_might_resched);
 7334#endif
 7335#endif
 7336
 7337/*
 7338 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
 7339 * call schedule, and on return reacquire the lock.
 7340 *
 7341 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
 7342 * operations here to prevent schedule() from being called twice (once via
 7343 * spin_unlock(), once by hand).
 7344 */
 7345int __cond_resched_lock(spinlock_t *lock)
 7346{
 7347	int resched = should_resched(PREEMPT_LOCK_OFFSET);
 7348	int ret = 0;
 7349
 7350	lockdep_assert_held(lock);
 7351
 7352	if (spin_needbreak(lock) || resched) {
 7353		spin_unlock(lock);
 7354		if (!_cond_resched())
 
 
 7355			cpu_relax();
 7356		ret = 1;
 7357		spin_lock(lock);
 7358	}
 7359	return ret;
 7360}
 7361EXPORT_SYMBOL(__cond_resched_lock);
 7362
 7363int __cond_resched_rwlock_read(rwlock_t *lock)
 7364{
 7365	int resched = should_resched(PREEMPT_LOCK_OFFSET);
 7366	int ret = 0;
 7367
 7368	lockdep_assert_held_read(lock);
 7369
 7370	if (rwlock_needbreak(lock) || resched) {
 7371		read_unlock(lock);
 7372		if (!_cond_resched())
 
 
 7373			cpu_relax();
 7374		ret = 1;
 7375		read_lock(lock);
 7376	}
 7377	return ret;
 7378}
 7379EXPORT_SYMBOL(__cond_resched_rwlock_read);
 7380
 7381int __cond_resched_rwlock_write(rwlock_t *lock)
 7382{
 7383	int resched = should_resched(PREEMPT_LOCK_OFFSET);
 7384	int ret = 0;
 7385
 7386	lockdep_assert_held_write(lock);
 7387
 7388	if (rwlock_needbreak(lock) || resched) {
 7389		write_unlock(lock);
 7390		if (!_cond_resched())
 
 
 7391			cpu_relax();
 7392		ret = 1;
 7393		write_lock(lock);
 7394	}
 7395	return ret;
 7396}
 7397EXPORT_SYMBOL(__cond_resched_rwlock_write);
 7398
 7399#ifdef CONFIG_PREEMPT_DYNAMIC
 7400
 7401#ifdef CONFIG_GENERIC_ENTRY
 7402#include <linux/entry-common.h>
 7403#endif
 7404
 7405/*
 7406 * SC:cond_resched
 7407 * SC:might_resched
 7408 * SC:preempt_schedule
 7409 * SC:preempt_schedule_notrace
 7410 * SC:irqentry_exit_cond_resched
 7411 *
 
 7412 *
 7413 * NONE:
 7414 *   cond_resched               <- __cond_resched
 7415 *   might_resched              <- RET0
 7416 *   preempt_schedule           <- NOP
 7417 *   preempt_schedule_notrace   <- NOP
 7418 *   irqentry_exit_cond_resched <- NOP
 7419 *   dynamic_preempt_lazy       <- false
 7420 *
 7421 * VOLUNTARY:
 7422 *   cond_resched               <- __cond_resched
 7423 *   might_resched              <- __cond_resched
 7424 *   preempt_schedule           <- NOP
 7425 *   preempt_schedule_notrace   <- NOP
 7426 *   irqentry_exit_cond_resched <- NOP
 7427 *   dynamic_preempt_lazy       <- false
 7428 *
 7429 * FULL:
 7430 *   cond_resched               <- RET0
 7431 *   might_resched              <- RET0
 7432 *   preempt_schedule           <- preempt_schedule
 7433 *   preempt_schedule_notrace   <- preempt_schedule_notrace
 7434 *   irqentry_exit_cond_resched <- irqentry_exit_cond_resched
 7435 *   dynamic_preempt_lazy       <- false
 7436 *
 7437 * LAZY:
 7438 *   cond_resched               <- RET0
 7439 *   might_resched              <- RET0
 7440 *   preempt_schedule           <- preempt_schedule
 7441 *   preempt_schedule_notrace   <- preempt_schedule_notrace
 7442 *   irqentry_exit_cond_resched <- irqentry_exit_cond_resched
 7443 *   dynamic_preempt_lazy       <- true
 7444 */
 7445
 7446enum {
 7447	preempt_dynamic_undefined = -1,
 7448	preempt_dynamic_none,
 7449	preempt_dynamic_voluntary,
 7450	preempt_dynamic_full,
 7451	preempt_dynamic_lazy,
 7452};
 7453
 7454int preempt_dynamic_mode = preempt_dynamic_undefined;
 7455
 7456int sched_dynamic_mode(const char *str)
 7457{
 7458#ifndef CONFIG_PREEMPT_RT
 7459	if (!strcmp(str, "none"))
 7460		return preempt_dynamic_none;
 7461
 7462	if (!strcmp(str, "voluntary"))
 7463		return preempt_dynamic_voluntary;
 7464#endif
 7465
 7466	if (!strcmp(str, "full"))
 7467		return preempt_dynamic_full;
 7468
 7469#ifdef CONFIG_ARCH_HAS_PREEMPT_LAZY
 7470	if (!strcmp(str, "lazy"))
 7471		return preempt_dynamic_lazy;
 7472#endif
 7473
 7474	return -EINVAL;
 7475}
 
 7476
 7477#define preempt_dynamic_key_enable(f)	static_key_enable(&sk_dynamic_##f.key)
 7478#define preempt_dynamic_key_disable(f)	static_key_disable(&sk_dynamic_##f.key)
 7479
 7480#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 7481#define preempt_dynamic_enable(f)	static_call_update(f, f##_dynamic_enabled)
 7482#define preempt_dynamic_disable(f)	static_call_update(f, f##_dynamic_disabled)
 7483#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 7484#define preempt_dynamic_enable(f)	preempt_dynamic_key_enable(f)
 7485#define preempt_dynamic_disable(f)	preempt_dynamic_key_disable(f)
 7486#else
 7487#error "Unsupported PREEMPT_DYNAMIC mechanism"
 7488#endif
 
 
 
 
 
 
 
 
 
 7489
 7490static DEFINE_MUTEX(sched_dynamic_mutex);
 7491static bool klp_override;
 7492
 7493static void __sched_dynamic_update(int mode)
 7494{
 7495	/*
 7496	 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
 7497	 * the ZERO state, which is invalid.
 7498	 */
 7499	if (!klp_override)
 7500		preempt_dynamic_enable(cond_resched);
 7501	preempt_dynamic_enable(might_resched);
 7502	preempt_dynamic_enable(preempt_schedule);
 7503	preempt_dynamic_enable(preempt_schedule_notrace);
 7504	preempt_dynamic_enable(irqentry_exit_cond_resched);
 7505	preempt_dynamic_key_disable(preempt_lazy);
 7506
 7507	switch (mode) {
 7508	case preempt_dynamic_none:
 7509		if (!klp_override)
 7510			preempt_dynamic_enable(cond_resched);
 7511		preempt_dynamic_disable(might_resched);
 7512		preempt_dynamic_disable(preempt_schedule);
 7513		preempt_dynamic_disable(preempt_schedule_notrace);
 7514		preempt_dynamic_disable(irqentry_exit_cond_resched);
 7515		preempt_dynamic_key_disable(preempt_lazy);
 7516		if (mode != preempt_dynamic_mode)
 7517			pr_info("Dynamic Preempt: none\n");
 7518		break;
 7519
 7520	case preempt_dynamic_voluntary:
 7521		if (!klp_override)
 7522			preempt_dynamic_enable(cond_resched);
 7523		preempt_dynamic_enable(might_resched);
 7524		preempt_dynamic_disable(preempt_schedule);
 7525		preempt_dynamic_disable(preempt_schedule_notrace);
 7526		preempt_dynamic_disable(irqentry_exit_cond_resched);
 7527		preempt_dynamic_key_disable(preempt_lazy);
 7528		if (mode != preempt_dynamic_mode)
 7529			pr_info("Dynamic Preempt: voluntary\n");
 7530		break;
 7531
 7532	case preempt_dynamic_full:
 7533		if (!klp_override)
 7534			preempt_dynamic_disable(cond_resched);
 7535		preempt_dynamic_disable(might_resched);
 7536		preempt_dynamic_enable(preempt_schedule);
 7537		preempt_dynamic_enable(preempt_schedule_notrace);
 7538		preempt_dynamic_enable(irqentry_exit_cond_resched);
 7539		preempt_dynamic_key_disable(preempt_lazy);
 7540		if (mode != preempt_dynamic_mode)
 7541			pr_info("Dynamic Preempt: full\n");
 7542		break;
 7543
 7544	case preempt_dynamic_lazy:
 7545		if (!klp_override)
 7546			preempt_dynamic_disable(cond_resched);
 7547		preempt_dynamic_disable(might_resched);
 7548		preempt_dynamic_enable(preempt_schedule);
 7549		preempt_dynamic_enable(preempt_schedule_notrace);
 7550		preempt_dynamic_enable(irqentry_exit_cond_resched);
 7551		preempt_dynamic_key_enable(preempt_lazy);
 7552		if (mode != preempt_dynamic_mode)
 7553			pr_info("Dynamic Preempt: lazy\n");
 7554		break;
 7555	}
 7556
 7557	preempt_dynamic_mode = mode;
 7558}
 7559
 7560void sched_dynamic_update(int mode)
 7561{
 7562	mutex_lock(&sched_dynamic_mutex);
 7563	__sched_dynamic_update(mode);
 7564	mutex_unlock(&sched_dynamic_mutex);
 7565}
 7566
 7567#ifdef CONFIG_HAVE_PREEMPT_DYNAMIC_CALL
 7568
 7569static int klp_cond_resched(void)
 7570{
 7571	__klp_sched_try_switch();
 7572	return __cond_resched();
 7573}
 7574
 7575void sched_dynamic_klp_enable(void)
 7576{
 7577	mutex_lock(&sched_dynamic_mutex);
 7578
 7579	klp_override = true;
 7580	static_call_update(cond_resched, klp_cond_resched);
 7581
 7582	mutex_unlock(&sched_dynamic_mutex);
 7583}
 7584
 7585void sched_dynamic_klp_disable(void)
 7586{
 7587	mutex_lock(&sched_dynamic_mutex);
 7588
 7589	klp_override = false;
 7590	__sched_dynamic_update(preempt_dynamic_mode);
 7591
 7592	mutex_unlock(&sched_dynamic_mutex);
 7593}
 7594
 7595#endif /* CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
 7596
 7597static int __init setup_preempt_mode(char *str)
 7598{
 7599	int mode = sched_dynamic_mode(str);
 7600	if (mode < 0) {
 7601		pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
 7602		return 0;
 7603	}
 7604
 7605	sched_dynamic_update(mode);
 7606	return 1;
 7607}
 7608__setup("preempt=", setup_preempt_mode);
 7609
 7610static void __init preempt_dynamic_init(void)
 7611{
 7612	if (preempt_dynamic_mode == preempt_dynamic_undefined) {
 7613		if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
 7614			sched_dynamic_update(preempt_dynamic_none);
 7615		} else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
 7616			sched_dynamic_update(preempt_dynamic_voluntary);
 7617		} else if (IS_ENABLED(CONFIG_PREEMPT_LAZY)) {
 7618			sched_dynamic_update(preempt_dynamic_lazy);
 7619		} else {
 7620			/* Default static call setting, nothing to do */
 7621			WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
 7622			preempt_dynamic_mode = preempt_dynamic_full;
 7623			pr_info("Dynamic Preempt: full\n");
 7624		}
 7625	}
 7626}
 7627
 7628#define PREEMPT_MODEL_ACCESSOR(mode) \
 7629	bool preempt_model_##mode(void)						 \
 7630	{									 \
 7631		WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
 7632		return preempt_dynamic_mode == preempt_dynamic_##mode;		 \
 7633	}									 \
 7634	EXPORT_SYMBOL_GPL(preempt_model_##mode)
 7635
 7636PREEMPT_MODEL_ACCESSOR(none);
 7637PREEMPT_MODEL_ACCESSOR(voluntary);
 7638PREEMPT_MODEL_ACCESSOR(full);
 7639PREEMPT_MODEL_ACCESSOR(lazy);
 
 
 
 
 
 
 7640
 7641#else /* !CONFIG_PREEMPT_DYNAMIC: */
 
 
 
 7642
 7643static inline void preempt_dynamic_init(void) { }
 
 7644
 7645#endif /* CONFIG_PREEMPT_DYNAMIC */
 
 
 7646
 7647int io_schedule_prepare(void)
 7648{
 7649	int old_iowait = current->in_iowait;
 7650
 7651	current->in_iowait = 1;
 7652	blk_flush_plug(current->plug, true);
 
 7653	return old_iowait;
 7654}
 7655
 7656void io_schedule_finish(int token)
 7657{
 7658	current->in_iowait = token;
 7659}
 7660
 7661/*
 7662 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
 7663 * that process accounting knows that this is a task in IO wait state.
 7664 */
 7665long __sched io_schedule_timeout(long timeout)
 7666{
 7667	int token;
 7668	long ret;
 7669
 7670	token = io_schedule_prepare();
 7671	ret = schedule_timeout(timeout);
 7672	io_schedule_finish(token);
 7673
 7674	return ret;
 7675}
 7676EXPORT_SYMBOL(io_schedule_timeout);
 7677
 7678void __sched io_schedule(void)
 7679{
 7680	int token;
 7681
 7682	token = io_schedule_prepare();
 7683	schedule();
 7684	io_schedule_finish(token);
 7685}
 7686EXPORT_SYMBOL(io_schedule);
 7687
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 7688void sched_show_task(struct task_struct *p)
 7689{
 7690	unsigned long free;
 7691	int ppid;
 7692
 7693	if (!try_get_task_stack(p))
 7694		return;
 7695
 7696	pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
 7697
 7698	if (task_is_running(p))
 7699		pr_cont("  running task    ");
 
 7700	free = stack_not_used(p);
 
 7701	ppid = 0;
 7702	rcu_read_lock();
 7703	if (pid_alive(p))
 7704		ppid = task_pid_nr(rcu_dereference(p->real_parent));
 7705	rcu_read_unlock();
 7706	pr_cont(" stack:%-5lu pid:%-5d tgid:%-5d ppid:%-6d flags:0x%08lx\n",
 7707		free, task_pid_nr(p), task_tgid_nr(p),
 7708		ppid, read_task_thread_flags(p));
 7709
 7710	print_worker_info(KERN_INFO, p);
 7711	print_stop_info(KERN_INFO, p);
 7712	print_scx_info(KERN_INFO, p);
 7713	show_stack(p, NULL, KERN_INFO);
 7714	put_task_stack(p);
 7715}
 7716EXPORT_SYMBOL_GPL(sched_show_task);
 7717
 7718static inline bool
 7719state_filter_match(unsigned long state_filter, struct task_struct *p)
 7720{
 7721	unsigned int state = READ_ONCE(p->__state);
 7722
 7723	/* no filter, everything matches */
 7724	if (!state_filter)
 7725		return true;
 7726
 7727	/* filter, but doesn't match */
 7728	if (!(state & state_filter))
 7729		return false;
 7730
 7731	/*
 7732	 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
 7733	 * TASK_KILLABLE).
 7734	 */
 7735	if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
 7736		return false;
 7737
 7738	return true;
 7739}
 7740
 7741
 7742void show_state_filter(unsigned int state_filter)
 7743{
 7744	struct task_struct *g, *p;
 7745
 7746	rcu_read_lock();
 7747	for_each_process_thread(g, p) {
 7748		/*
 7749		 * reset the NMI-timeout, listing all files on a slow
 7750		 * console might take a lot of time:
 7751		 * Also, reset softlockup watchdogs on all CPUs, because
 7752		 * another CPU might be blocked waiting for us to process
 7753		 * an IPI.
 7754		 */
 7755		touch_nmi_watchdog();
 7756		touch_all_softlockup_watchdogs();
 7757		if (state_filter_match(state_filter, p))
 7758			sched_show_task(p);
 7759	}
 7760
 7761#ifdef CONFIG_SCHED_DEBUG
 7762	if (!state_filter)
 7763		sysrq_sched_debug_show();
 7764#endif
 7765	rcu_read_unlock();
 7766	/*
 7767	 * Only show locks if all tasks are dumped:
 7768	 */
 7769	if (!state_filter)
 7770		debug_show_all_locks();
 7771}
 7772
 7773/**
 7774 * init_idle - set up an idle thread for a given CPU
 7775 * @idle: task in question
 7776 * @cpu: CPU the idle task belongs to
 7777 *
 7778 * NOTE: this function does not set the idle thread's NEED_RESCHED
 7779 * flag, to make booting more robust.
 7780 */
 7781void __init init_idle(struct task_struct *idle, int cpu)
 7782{
 7783#ifdef CONFIG_SMP
 7784	struct affinity_context ac = (struct affinity_context) {
 7785		.new_mask  = cpumask_of(cpu),
 7786		.flags     = 0,
 7787	};
 7788#endif
 7789	struct rq *rq = cpu_rq(cpu);
 7790	unsigned long flags;
 7791
 
 
 
 
 
 
 
 
 
 
 7792	raw_spin_lock_irqsave(&idle->pi_lock, flags);
 7793	raw_spin_rq_lock(rq);
 7794
 7795	idle->__state = TASK_RUNNING;
 7796	idle->se.exec_start = sched_clock();
 7797	/*
 7798	 * PF_KTHREAD should already be set at this point; regardless, make it
 7799	 * look like a proper per-CPU kthread.
 7800	 */
 7801	idle->flags |= PF_KTHREAD | PF_NO_SETAFFINITY;
 7802	kthread_set_per_cpu(idle, cpu);
 7803
 
 
 
 7804#ifdef CONFIG_SMP
 7805	/*
 7806	 * No validation and serialization required at boot time and for
 7807	 * setting up the idle tasks of not yet online CPUs.
 
 
 7808	 */
 7809	set_cpus_allowed_common(idle, &ac);
 7810#endif
 7811	/*
 7812	 * We're having a chicken and egg problem, even though we are
 7813	 * holding rq->lock, the CPU isn't yet set to this CPU so the
 7814	 * lockdep check in task_group() will fail.
 7815	 *
 7816	 * Similar case to sched_fork(). / Alternatively we could
 7817	 * use task_rq_lock() here and obtain the other rq->lock.
 7818	 *
 7819	 * Silence PROVE_RCU
 7820	 */
 7821	rcu_read_lock();
 7822	__set_task_cpu(idle, cpu);
 7823	rcu_read_unlock();
 7824
 7825	rq->idle = idle;
 7826	rq_set_donor(rq, idle);
 7827	rcu_assign_pointer(rq->curr, idle);
 7828	idle->on_rq = TASK_ON_RQ_QUEUED;
 7829#ifdef CONFIG_SMP
 7830	idle->on_cpu = 1;
 7831#endif
 7832	raw_spin_rq_unlock(rq);
 7833	raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
 7834
 7835	/* Set the preempt count _outside_ the spinlocks! */
 7836	init_idle_preempt_count(idle, cpu);
 7837
 7838	/*
 7839	 * The idle tasks have their own, simple scheduling class:
 7840	 */
 7841	idle->sched_class = &idle_sched_class;
 7842	ftrace_graph_init_idle_task(idle, cpu);
 7843	vtime_init_idle(idle, cpu);
 7844#ifdef CONFIG_SMP
 7845	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
 7846#endif
 7847}
 7848
 7849#ifdef CONFIG_SMP
 7850
 7851int cpuset_cpumask_can_shrink(const struct cpumask *cur,
 7852			      const struct cpumask *trial)
 7853{
 7854	int ret = 1;
 7855
 7856	if (cpumask_empty(cur))
 7857		return ret;
 7858
 7859	ret = dl_cpuset_cpumask_can_shrink(cur, trial);
 7860
 7861	return ret;
 7862}
 7863
 7864int task_can_attach(struct task_struct *p)
 
 7865{
 7866	int ret = 0;
 7867
 7868	/*
 7869	 * Kthreads which disallow setaffinity shouldn't be moved
 7870	 * to a new cpuset; we don't want to change their CPU
 7871	 * affinity and isolating such threads by their set of
 7872	 * allowed nodes is unnecessary.  Thus, cpusets are not
 7873	 * applicable for such threads.  This prevents checking for
 7874	 * success of set_cpus_allowed_ptr() on all attached tasks
 7875	 * before cpus_mask may be changed.
 7876	 */
 7877	if (p->flags & PF_NO_SETAFFINITY)
 7878		ret = -EINVAL;
 
 
 
 
 
 
 7879
 
 7880	return ret;
 7881}
 7882
 7883bool sched_smp_initialized __read_mostly;
 7884
 7885#ifdef CONFIG_NUMA_BALANCING
 7886/* Migrate current task p to target_cpu */
 7887int migrate_task_to(struct task_struct *p, int target_cpu)
 7888{
 7889	struct migration_arg arg = { p, target_cpu };
 7890	int curr_cpu = task_cpu(p);
 7891
 7892	if (curr_cpu == target_cpu)
 7893		return 0;
 7894
 7895	if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
 7896		return -EINVAL;
 7897
 7898	/* TODO: This is not properly updating schedstats */
 7899
 7900	trace_sched_move_numa(p, curr_cpu, target_cpu);
 7901	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
 7902}
 7903
 7904/*
 7905 * Requeue a task on a given node and accurately track the number of NUMA
 7906 * tasks on the runqueues
 7907 */
 7908void sched_setnuma(struct task_struct *p, int nid)
 7909{
 7910	bool queued, running;
 7911	struct rq_flags rf;
 7912	struct rq *rq;
 7913
 7914	rq = task_rq_lock(p, &rf);
 7915	queued = task_on_rq_queued(p);
 7916	running = task_current_donor(rq, p);
 7917
 7918	if (queued)
 7919		dequeue_task(rq, p, DEQUEUE_SAVE);
 7920	if (running)
 7921		put_prev_task(rq, p);
 7922
 7923	p->numa_preferred_nid = nid;
 7924
 7925	if (queued)
 7926		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
 7927	if (running)
 7928		set_next_task(rq, p);
 7929	task_rq_unlock(rq, p, &rf);
 7930}
 7931#endif /* CONFIG_NUMA_BALANCING */
 7932
 7933#ifdef CONFIG_HOTPLUG_CPU
 7934/*
 7935 * Ensure that the idle task is using init_mm right before its CPU goes
 7936 * offline.
 7937 */
 7938void idle_task_exit(void)
 7939{
 7940	struct mm_struct *mm = current->active_mm;
 7941
 7942	BUG_ON(cpu_online(smp_processor_id()));
 7943	BUG_ON(current != this_rq()->idle);
 7944
 7945	if (mm != &init_mm) {
 7946		switch_mm(mm, &init_mm, current);
 7947		finish_arch_post_lock_switch();
 7948	}
 7949
 
 7950	/* finish_cpu(), as ran on the BP, will clean up the active_mm state */
 7951}
 7952
 7953static int __balance_push_cpu_stop(void *arg)
 7954{
 7955	struct task_struct *p = arg;
 7956	struct rq *rq = this_rq();
 7957	struct rq_flags rf;
 7958	int cpu;
 7959
 7960	raw_spin_lock_irq(&p->pi_lock);
 7961	rq_lock(rq, &rf);
 7962
 7963	update_rq_clock(rq);
 7964
 7965	if (task_rq(p) == rq && task_on_rq_queued(p)) {
 7966		cpu = select_fallback_rq(rq->cpu, p);
 7967		rq = __migrate_task(rq, &rf, p, cpu);
 7968	}
 7969
 7970	rq_unlock(rq, &rf);
 7971	raw_spin_unlock_irq(&p->pi_lock);
 7972
 7973	put_task_struct(p);
 7974
 7975	return 0;
 7976}
 7977
 7978static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
 7979
 7980/*
 7981 * Ensure we only run per-cpu kthreads once the CPU goes !active.
 7982 *
 7983 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
 7984 * effective when the hotplug motion is down.
 7985 */
 7986static void balance_push(struct rq *rq)
 7987{
 7988	struct task_struct *push_task = rq->curr;
 7989
 7990	lockdep_assert_rq_held(rq);
 7991
 7992	/*
 7993	 * Ensure the thing is persistent until balance_push_set(.on = false);
 7994	 */
 7995	rq->balance_callback = &balance_push_callback;
 7996
 7997	/*
 7998	 * Only active while going offline and when invoked on the outgoing
 7999	 * CPU.
 8000	 */
 8001	if (!cpu_dying(rq->cpu) || rq != this_rq())
 8002		return;
 8003
 8004	/*
 8005	 * Both the cpu-hotplug and stop task are in this case and are
 8006	 * required to complete the hotplug process.
 8007	 */
 8008	if (kthread_is_per_cpu(push_task) ||
 8009	    is_migration_disabled(push_task)) {
 8010
 8011		/*
 8012		 * If this is the idle task on the outgoing CPU try to wake
 8013		 * up the hotplug control thread which might wait for the
 8014		 * last task to vanish. The rcuwait_active() check is
 8015		 * accurate here because the waiter is pinned on this CPU
 8016		 * and can't obviously be running in parallel.
 8017		 *
 8018		 * On RT kernels this also has to check whether there are
 8019		 * pinned and scheduled out tasks on the runqueue. They
 8020		 * need to leave the migrate disabled section first.
 8021		 */
 8022		if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
 8023		    rcuwait_active(&rq->hotplug_wait)) {
 8024			raw_spin_rq_unlock(rq);
 8025			rcuwait_wake_up(&rq->hotplug_wait);
 8026			raw_spin_rq_lock(rq);
 8027		}
 8028		return;
 8029	}
 8030
 8031	get_task_struct(push_task);
 8032	/*
 8033	 * Temporarily drop rq->lock such that we can wake-up the stop task.
 8034	 * Both preemption and IRQs are still disabled.
 8035	 */
 8036	preempt_disable();
 8037	raw_spin_rq_unlock(rq);
 8038	stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
 8039			    this_cpu_ptr(&push_work));
 8040	preempt_enable();
 8041	/*
 8042	 * At this point need_resched() is true and we'll take the loop in
 8043	 * schedule(). The next pick is obviously going to be the stop task
 8044	 * which kthread_is_per_cpu() and will push this task away.
 8045	 */
 8046	raw_spin_rq_lock(rq);
 8047}
 8048
 8049static void balance_push_set(int cpu, bool on)
 8050{
 8051	struct rq *rq = cpu_rq(cpu);
 8052	struct rq_flags rf;
 8053
 8054	rq_lock_irqsave(rq, &rf);
 8055	if (on) {
 8056		WARN_ON_ONCE(rq->balance_callback);
 8057		rq->balance_callback = &balance_push_callback;
 8058	} else if (rq->balance_callback == &balance_push_callback) {
 8059		rq->balance_callback = NULL;
 8060	}
 8061	rq_unlock_irqrestore(rq, &rf);
 8062}
 8063
 8064/*
 8065 * Invoked from a CPUs hotplug control thread after the CPU has been marked
 8066 * inactive. All tasks which are not per CPU kernel threads are either
 8067 * pushed off this CPU now via balance_push() or placed on a different CPU
 8068 * during wakeup. Wait until the CPU is quiescent.
 8069 */
 8070static void balance_hotplug_wait(void)
 8071{
 8072	struct rq *rq = this_rq();
 8073
 8074	rcuwait_wait_event(&rq->hotplug_wait,
 8075			   rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
 8076			   TASK_UNINTERRUPTIBLE);
 8077}
 8078
 8079#else
 8080
 8081static inline void balance_push(struct rq *rq)
 8082{
 8083}
 8084
 8085static inline void balance_push_set(int cpu, bool on)
 8086{
 8087}
 8088
 8089static inline void balance_hotplug_wait(void)
 8090{
 8091}
 8092
 8093#endif /* CONFIG_HOTPLUG_CPU */
 8094
 8095void set_rq_online(struct rq *rq)
 8096{
 8097	if (!rq->online) {
 8098		const struct sched_class *class;
 8099
 8100		cpumask_set_cpu(rq->cpu, rq->rd->online);
 8101		rq->online = 1;
 8102
 8103		for_each_class(class) {
 8104			if (class->rq_online)
 8105				class->rq_online(rq);
 8106		}
 8107	}
 8108}
 8109
 8110void set_rq_offline(struct rq *rq)
 8111{
 8112	if (rq->online) {
 8113		const struct sched_class *class;
 8114
 8115		update_rq_clock(rq);
 8116		for_each_class(class) {
 8117			if (class->rq_offline)
 8118				class->rq_offline(rq);
 8119		}
 8120
 8121		cpumask_clear_cpu(rq->cpu, rq->rd->online);
 8122		rq->online = 0;
 8123	}
 8124}
 8125
 8126static inline void sched_set_rq_online(struct rq *rq, int cpu)
 8127{
 8128	struct rq_flags rf;
 8129
 8130	rq_lock_irqsave(rq, &rf);
 8131	if (rq->rd) {
 8132		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
 8133		set_rq_online(rq);
 8134	}
 8135	rq_unlock_irqrestore(rq, &rf);
 8136}
 8137
 8138static inline void sched_set_rq_offline(struct rq *rq, int cpu)
 8139{
 8140	struct rq_flags rf;
 8141
 8142	rq_lock_irqsave(rq, &rf);
 8143	if (rq->rd) {
 8144		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
 8145		set_rq_offline(rq);
 8146	}
 8147	rq_unlock_irqrestore(rq, &rf);
 8148}
 8149
 8150/*
 8151 * used to mark begin/end of suspend/resume:
 8152 */
 8153static int num_cpus_frozen;
 8154
 8155/*
 8156 * Update cpusets according to cpu_active mask.  If cpusets are
 8157 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
 8158 * around partition_sched_domains().
 8159 *
 8160 * If we come here as part of a suspend/resume, don't touch cpusets because we
 8161 * want to restore it back to its original state upon resume anyway.
 8162 */
 8163static void cpuset_cpu_active(void)
 8164{
 8165	if (cpuhp_tasks_frozen) {
 8166		/*
 8167		 * num_cpus_frozen tracks how many CPUs are involved in suspend
 8168		 * resume sequence. As long as this is not the last online
 8169		 * operation in the resume sequence, just build a single sched
 8170		 * domain, ignoring cpusets.
 8171		 */
 8172		partition_sched_domains(1, NULL, NULL);
 8173		if (--num_cpus_frozen)
 8174			return;
 8175		/*
 8176		 * This is the last CPU online operation. So fall through and
 8177		 * restore the original sched domains by considering the
 8178		 * cpuset configurations.
 8179		 */
 8180		cpuset_force_rebuild();
 8181	}
 8182	cpuset_update_active_cpus();
 8183}
 8184
 8185static int cpuset_cpu_inactive(unsigned int cpu)
 8186{
 8187	if (!cpuhp_tasks_frozen) {
 8188		int ret = dl_bw_check_overflow(cpu);
 8189
 8190		if (ret)
 8191			return ret;
 8192		cpuset_update_active_cpus();
 8193	} else {
 8194		num_cpus_frozen++;
 8195		partition_sched_domains(1, NULL, NULL);
 8196	}
 8197	return 0;
 8198}
 8199
 8200static inline void sched_smt_present_inc(int cpu)
 8201{
 8202#ifdef CONFIG_SCHED_SMT
 8203	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
 8204		static_branch_inc_cpuslocked(&sched_smt_present);
 8205#endif
 8206}
 8207
 8208static inline void sched_smt_present_dec(int cpu)
 8209{
 8210#ifdef CONFIG_SCHED_SMT
 8211	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
 8212		static_branch_dec_cpuslocked(&sched_smt_present);
 8213#endif
 8214}
 8215
 8216int sched_cpu_activate(unsigned int cpu)
 8217{
 8218	struct rq *rq = cpu_rq(cpu);
 
 8219
 8220	/*
 8221	 * Clear the balance_push callback and prepare to schedule
 8222	 * regular tasks.
 8223	 */
 8224	balance_push_set(cpu, false);
 8225
 
 8226	/*
 8227	 * When going up, increment the number of cores with SMT present.
 8228	 */
 8229	sched_smt_present_inc(cpu);
 
 
 8230	set_cpu_active(cpu, true);
 8231
 8232	if (sched_smp_initialized) {
 8233		sched_update_numa(cpu, true);
 8234		sched_domains_numa_masks_set(cpu);
 8235		cpuset_cpu_active();
 8236	}
 8237
 8238	scx_rq_activate(rq);
 8239
 8240	/*
 8241	 * Put the rq online, if not already. This happens:
 8242	 *
 8243	 * 1) In the early boot process, because we build the real domains
 8244	 *    after all CPUs have been brought up.
 8245	 *
 8246	 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
 8247	 *    domains.
 8248	 */
 8249	sched_set_rq_online(rq, cpu);
 
 
 
 
 
 8250
 8251	return 0;
 8252}
 8253
 8254int sched_cpu_deactivate(unsigned int cpu)
 8255{
 8256	struct rq *rq = cpu_rq(cpu);
 
 8257	int ret;
 8258
 8259	/*
 8260	 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
 8261	 * load balancing when not active
 8262	 */
 8263	nohz_balance_exit_idle(rq);
 8264
 8265	set_cpu_active(cpu, false);
 8266
 8267	/*
 8268	 * From this point forward, this CPU will refuse to run any task that
 8269	 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
 8270	 * push those tasks away until this gets cleared, see
 8271	 * sched_cpu_dying().
 8272	 */
 8273	balance_push_set(cpu, true);
 8274
 8275	/*
 8276	 * We've cleared cpu_active_mask / set balance_push, wait for all
 8277	 * preempt-disabled and RCU users of this state to go away such that
 8278	 * all new such users will observe it.
 8279	 *
 8280	 * Specifically, we rely on ttwu to no longer target this CPU, see
 8281	 * ttwu_queue_cond() and is_cpu_allowed().
 8282	 *
 8283	 * Do sync before park smpboot threads to take care the RCU boost case.
 8284	 */
 8285	synchronize_rcu();
 8286
 8287	sched_set_rq_offline(rq, cpu);
 8288
 8289	scx_rq_deactivate(rq);
 
 
 
 
 8290
 
 8291	/*
 8292	 * When going down, decrement the number of cores with SMT present.
 8293	 */
 8294	sched_smt_present_dec(cpu);
 
 8295
 8296#ifdef CONFIG_SCHED_SMT
 8297	sched_core_cpu_deactivate(cpu);
 8298#endif
 8299
 8300	if (!sched_smp_initialized)
 8301		return 0;
 8302
 8303	sched_update_numa(cpu, false);
 8304	ret = cpuset_cpu_inactive(cpu);
 8305	if (ret) {
 8306		sched_smt_present_inc(cpu);
 8307		sched_set_rq_online(rq, cpu);
 8308		balance_push_set(cpu, false);
 8309		set_cpu_active(cpu, true);
 8310		sched_update_numa(cpu, true);
 8311		return ret;
 8312	}
 8313	sched_domains_numa_masks_clear(cpu);
 8314	return 0;
 8315}
 8316
 8317static void sched_rq_cpu_starting(unsigned int cpu)
 8318{
 8319	struct rq *rq = cpu_rq(cpu);
 8320
 8321	rq->calc_load_update = calc_load_update;
 8322	update_max_interval();
 8323}
 8324
 8325int sched_cpu_starting(unsigned int cpu)
 8326{
 8327	sched_core_cpu_starting(cpu);
 8328	sched_rq_cpu_starting(cpu);
 8329	sched_tick_start(cpu);
 8330	return 0;
 8331}
 8332
 8333#ifdef CONFIG_HOTPLUG_CPU
 8334
 8335/*
 8336 * Invoked immediately before the stopper thread is invoked to bring the
 8337 * CPU down completely. At this point all per CPU kthreads except the
 8338 * hotplug thread (current) and the stopper thread (inactive) have been
 8339 * either parked or have been unbound from the outgoing CPU. Ensure that
 8340 * any of those which might be on the way out are gone.
 8341 *
 8342 * If after this point a bound task is being woken on this CPU then the
 8343 * responsible hotplug callback has failed to do it's job.
 8344 * sched_cpu_dying() will catch it with the appropriate fireworks.
 8345 */
 8346int sched_cpu_wait_empty(unsigned int cpu)
 8347{
 8348	balance_hotplug_wait();
 8349	return 0;
 8350}
 8351
 8352/*
 8353 * Since this CPU is going 'away' for a while, fold any nr_active delta we
 8354 * might have. Called from the CPU stopper task after ensuring that the
 8355 * stopper is the last running task on the CPU, so nr_active count is
 8356 * stable. We need to take the tear-down thread which is calling this into
 8357 * account, so we hand in adjust = 1 to the load calculation.
 8358 *
 8359 * Also see the comment "Global load-average calculations".
 8360 */
 8361static void calc_load_migrate(struct rq *rq)
 8362{
 8363	long delta = calc_load_fold_active(rq, 1);
 8364
 8365	if (delta)
 8366		atomic_long_add(delta, &calc_load_tasks);
 8367}
 8368
 8369static void dump_rq_tasks(struct rq *rq, const char *loglvl)
 8370{
 8371	struct task_struct *g, *p;
 8372	int cpu = cpu_of(rq);
 8373
 8374	lockdep_assert_rq_held(rq);
 8375
 8376	printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
 8377	for_each_process_thread(g, p) {
 8378		if (task_cpu(p) != cpu)
 8379			continue;
 8380
 8381		if (!task_on_rq_queued(p))
 8382			continue;
 8383
 8384		printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
 8385	}
 8386}
 8387
 8388int sched_cpu_dying(unsigned int cpu)
 8389{
 8390	struct rq *rq = cpu_rq(cpu);
 8391	struct rq_flags rf;
 8392
 8393	/* Handle pending wakeups and then migrate everything off */
 8394	sched_tick_stop(cpu);
 8395
 8396	rq_lock_irqsave(rq, &rf);
 8397	if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
 8398		WARN(true, "Dying CPU not properly vacated!");
 8399		dump_rq_tasks(rq, KERN_WARNING);
 8400	}
 8401	rq_unlock_irqrestore(rq, &rf);
 8402
 8403	calc_load_migrate(rq);
 8404	update_max_interval();
 8405	hrtick_clear(rq);
 8406	sched_core_cpu_dying(cpu);
 8407	return 0;
 8408}
 8409#endif
 8410
 8411void __init sched_init_smp(void)
 8412{
 8413	sched_init_numa(NUMA_NO_NODE);
 8414
 8415	/*
 8416	 * There's no userspace yet to cause hotplug operations; hence all the
 8417	 * CPU masks are stable and all blatant races in the below code cannot
 8418	 * happen.
 8419	 */
 8420	mutex_lock(&sched_domains_mutex);
 8421	sched_init_domains(cpu_active_mask);
 8422	mutex_unlock(&sched_domains_mutex);
 8423
 8424	/* Move init over to a non-isolated CPU */
 8425	if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
 8426		BUG();
 8427	current->flags &= ~PF_NO_SETAFFINITY;
 8428	sched_init_granularity();
 8429
 8430	init_sched_rt_class();
 8431	init_sched_dl_class();
 8432
 8433	sched_smp_initialized = true;
 8434}
 8435
 8436static int __init migration_init(void)
 8437{
 8438	sched_cpu_starting(smp_processor_id());
 8439	return 0;
 8440}
 8441early_initcall(migration_init);
 8442
 8443#else
 8444void __init sched_init_smp(void)
 8445{
 8446	sched_init_granularity();
 8447}
 8448#endif /* CONFIG_SMP */
 8449
 8450int in_sched_functions(unsigned long addr)
 8451{
 8452	return in_lock_functions(addr) ||
 8453		(addr >= (unsigned long)__sched_text_start
 8454		&& addr < (unsigned long)__sched_text_end);
 8455}
 8456
 8457#ifdef CONFIG_CGROUP_SCHED
 8458/*
 8459 * Default task group.
 8460 * Every task in system belongs to this group at bootup.
 8461 */
 8462struct task_group root_task_group;
 8463LIST_HEAD(task_groups);
 8464
 8465/* Cacheline aligned slab cache for task_group */
 8466static struct kmem_cache *task_group_cache __ro_after_init;
 8467#endif
 8468
 
 
 
 8469void __init sched_init(void)
 8470{
 8471	unsigned long ptr = 0;
 8472	int i;
 8473
 8474	/* Make sure the linker didn't screw up */
 
 
 
 8475#ifdef CONFIG_SMP
 8476	BUG_ON(!sched_class_above(&stop_sched_class, &dl_sched_class));
 8477#endif
 8478	BUG_ON(!sched_class_above(&dl_sched_class, &rt_sched_class));
 8479	BUG_ON(!sched_class_above(&rt_sched_class, &fair_sched_class));
 8480	BUG_ON(!sched_class_above(&fair_sched_class, &idle_sched_class));
 8481#ifdef CONFIG_SCHED_CLASS_EXT
 8482	BUG_ON(!sched_class_above(&fair_sched_class, &ext_sched_class));
 8483	BUG_ON(!sched_class_above(&ext_sched_class, &idle_sched_class));
 8484#endif
 8485
 8486	wait_bit_init();
 8487
 8488#ifdef CONFIG_FAIR_GROUP_SCHED
 8489	ptr += 2 * nr_cpu_ids * sizeof(void **);
 8490#endif
 8491#ifdef CONFIG_RT_GROUP_SCHED
 8492	ptr += 2 * nr_cpu_ids * sizeof(void **);
 8493#endif
 8494	if (ptr) {
 8495		ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
 8496
 8497#ifdef CONFIG_FAIR_GROUP_SCHED
 8498		root_task_group.se = (struct sched_entity **)ptr;
 8499		ptr += nr_cpu_ids * sizeof(void **);
 8500
 8501		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
 8502		ptr += nr_cpu_ids * sizeof(void **);
 8503
 8504		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
 8505		init_cfs_bandwidth(&root_task_group.cfs_bandwidth, NULL);
 8506#endif /* CONFIG_FAIR_GROUP_SCHED */
 8507#ifdef CONFIG_EXT_GROUP_SCHED
 8508		root_task_group.scx_weight = CGROUP_WEIGHT_DFL;
 8509#endif /* CONFIG_EXT_GROUP_SCHED */
 8510#ifdef CONFIG_RT_GROUP_SCHED
 8511		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
 8512		ptr += nr_cpu_ids * sizeof(void **);
 8513
 8514		root_task_group.rt_rq = (struct rt_rq **)ptr;
 8515		ptr += nr_cpu_ids * sizeof(void **);
 8516
 8517#endif /* CONFIG_RT_GROUP_SCHED */
 8518	}
 
 
 
 
 
 
 
 
 
 
 
 8519
 8520#ifdef CONFIG_SMP
 8521	init_defrootdomain();
 8522#endif
 8523
 8524#ifdef CONFIG_RT_GROUP_SCHED
 8525	init_rt_bandwidth(&root_task_group.rt_bandwidth,
 8526			global_rt_period(), global_rt_runtime());
 8527#endif /* CONFIG_RT_GROUP_SCHED */
 8528
 8529#ifdef CONFIG_CGROUP_SCHED
 8530	task_group_cache = KMEM_CACHE(task_group, 0);
 8531
 8532	list_add(&root_task_group.list, &task_groups);
 8533	INIT_LIST_HEAD(&root_task_group.children);
 8534	INIT_LIST_HEAD(&root_task_group.siblings);
 8535	autogroup_init(&init_task);
 8536#endif /* CONFIG_CGROUP_SCHED */
 8537
 8538	for_each_possible_cpu(i) {
 8539		struct rq *rq;
 8540
 8541		rq = cpu_rq(i);
 8542		raw_spin_lock_init(&rq->__lock);
 8543		rq->nr_running = 0;
 8544		rq->calc_load_active = 0;
 8545		rq->calc_load_update = jiffies + LOAD_FREQ;
 8546		init_cfs_rq(&rq->cfs);
 8547		init_rt_rq(&rq->rt);
 8548		init_dl_rq(&rq->dl);
 8549#ifdef CONFIG_FAIR_GROUP_SCHED
 8550		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
 8551		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
 8552		/*
 8553		 * How much CPU bandwidth does root_task_group get?
 8554		 *
 8555		 * In case of task-groups formed through the cgroup filesystem, it
 8556		 * gets 100% of the CPU resources in the system. This overall
 8557		 * system CPU resource is divided among the tasks of
 8558		 * root_task_group and its child task-groups in a fair manner,
 8559		 * based on each entity's (task or task-group's) weight
 8560		 * (se->load.weight).
 8561		 *
 8562		 * In other words, if root_task_group has 10 tasks of weight
 8563		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
 8564		 * then A0's share of the CPU resource is:
 8565		 *
 8566		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
 8567		 *
 8568		 * We achieve this by letting root_task_group's tasks sit
 8569		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
 8570		 */
 8571		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
 8572#endif /* CONFIG_FAIR_GROUP_SCHED */
 8573
 
 8574#ifdef CONFIG_RT_GROUP_SCHED
 8575		/*
 8576		 * This is required for init cpu because rt.c:__enable_runtime()
 8577		 * starts working after scheduler_running, which is not the case
 8578		 * yet.
 8579		 */
 8580		rq->rt.rt_runtime = global_rt_runtime();
 8581		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
 8582#endif
 8583#ifdef CONFIG_SMP
 8584		rq->sd = NULL;
 8585		rq->rd = NULL;
 8586		rq->cpu_capacity = SCHED_CAPACITY_SCALE;
 8587		rq->balance_callback = &balance_push_callback;
 8588		rq->active_balance = 0;
 8589		rq->next_balance = jiffies;
 8590		rq->push_cpu = 0;
 8591		rq->cpu = i;
 8592		rq->online = 0;
 8593		rq->idle_stamp = 0;
 8594		rq->avg_idle = 2*sysctl_sched_migration_cost;
 
 
 8595		rq->max_idle_balance_cost = sysctl_sched_migration_cost;
 8596
 8597		INIT_LIST_HEAD(&rq->cfs_tasks);
 8598
 8599		rq_attach_root(rq, &def_root_domain);
 8600#ifdef CONFIG_NO_HZ_COMMON
 8601		rq->last_blocked_load_update_tick = jiffies;
 8602		atomic_set(&rq->nohz_flags, 0);
 8603
 8604		INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
 8605#endif
 8606#ifdef CONFIG_HOTPLUG_CPU
 8607		rcuwait_init(&rq->hotplug_wait);
 8608#endif
 8609#endif /* CONFIG_SMP */
 8610		hrtick_rq_init(rq);
 8611		atomic_set(&rq->nr_iowait, 0);
 8612		fair_server_init(rq);
 8613
 8614#ifdef CONFIG_SCHED_CORE
 8615		rq->core = rq;
 8616		rq->core_pick = NULL;
 8617		rq->core_dl_server = NULL;
 8618		rq->core_enabled = 0;
 8619		rq->core_tree = RB_ROOT;
 8620		rq->core_forceidle_count = 0;
 8621		rq->core_forceidle_occupation = 0;
 8622		rq->core_forceidle_start = 0;
 8623
 8624		rq->core_cookie = 0UL;
 8625#endif
 8626		zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i));
 8627	}
 8628
 8629	set_load_weight(&init_task, false);
 8630	init_task.se.slice = sysctl_sched_base_slice,
 8631
 8632	/*
 8633	 * The boot idle thread does lazy MMU switching as well:
 8634	 */
 8635	mmgrab_lazy_tlb(&init_mm);
 8636	enter_lazy_tlb(&init_mm, current);
 8637
 8638	/*
 8639	 * The idle task doesn't need the kthread struct to function, but it
 8640	 * is dressed up as a per-CPU kthread and thus needs to play the part
 8641	 * if we want to avoid special-casing it in code that deals with per-CPU
 8642	 * kthreads.
 8643	 */
 8644	WARN_ON(!set_kthread_struct(current));
 8645
 8646	/*
 8647	 * Make us the idle thread. Technically, schedule() should not be
 8648	 * called from this thread, however somewhere below it might be,
 8649	 * but because we are the idle thread, we just pick up running again
 8650	 * when this runqueue becomes "idle".
 8651	 */
 8652	__sched_fork(0, current);
 8653	init_idle(current, smp_processor_id());
 8654
 8655	calc_load_update = jiffies + LOAD_FREQ;
 8656
 8657#ifdef CONFIG_SMP
 8658	idle_thread_set_boot_cpu();
 8659	balance_push_set(smp_processor_id(), false);
 8660#endif
 8661	init_sched_fair_class();
 8662	init_sched_ext_class();
 8663
 8664	psi_init();
 8665
 8666	init_uclamp();
 8667
 8668	preempt_dynamic_init();
 8669
 8670	scheduler_running = 1;
 8671}
 8672
 8673#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
 
 
 
 8674
 8675void __might_sleep(const char *file, int line)
 
 
 
 8676{
 8677	unsigned int state = get_current_state();
 8678	/*
 8679	 * Blocking primitives will set (and therefore destroy) current->state,
 8680	 * since we will exit with TASK_RUNNING make sure we enter with it,
 8681	 * otherwise we will destroy state.
 8682	 */
 8683	WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
 8684			"do not call blocking ops when !TASK_RUNNING; "
 8685			"state=%x set at [<%p>] %pS\n", state,
 8686			(void *)current->task_state_change,
 8687			(void *)current->task_state_change);
 8688
 8689	__might_resched(file, line, 0);
 8690}
 8691EXPORT_SYMBOL(__might_sleep);
 8692
 8693static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
 8694{
 8695	if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
 8696		return;
 8697
 8698	if (preempt_count() == preempt_offset)
 8699		return;
 8700
 8701	pr_err("Preemption disabled at:");
 8702	print_ip_sym(KERN_ERR, ip);
 8703}
 8704
 8705static inline bool resched_offsets_ok(unsigned int offsets)
 8706{
 8707	unsigned int nested = preempt_count();
 8708
 8709	nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
 8710
 8711	return nested == offsets;
 8712}
 8713
 8714void __might_resched(const char *file, int line, unsigned int offsets)
 8715{
 8716	/* Ratelimiting timestamp: */
 8717	static unsigned long prev_jiffy;
 8718
 8719	unsigned long preempt_disable_ip;
 8720
 8721	/* WARN_ON_ONCE() by default, no rate limit required: */
 8722	rcu_sleep_check();
 8723
 8724	if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
 8725	     !is_idle_task(current) && !current->non_block_count) ||
 8726	    system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
 8727	    oops_in_progress)
 8728		return;
 8729
 8730	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
 8731		return;
 8732	prev_jiffy = jiffies;
 8733
 8734	/* Save this before calling printk(), since that will clobber it: */
 8735	preempt_disable_ip = get_preempt_disable_ip(current);
 8736
 8737	pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
 8738	       file, line);
 8739	pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
 8740	       in_atomic(), irqs_disabled(), current->non_block_count,
 8741	       current->pid, current->comm);
 8742	pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
 8743	       offsets & MIGHT_RESCHED_PREEMPT_MASK);
 8744
 8745	if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
 8746		pr_err("RCU nest depth: %d, expected: %u\n",
 8747		       rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
 8748	}
 8749
 8750	if (task_stack_end_corrupted(current))
 8751		pr_emerg("Thread overran stack, or stack corrupted\n");
 8752
 8753	debug_show_held_locks(current);
 8754	if (irqs_disabled())
 8755		print_irqtrace_events(current);
 8756
 8757	print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
 8758				 preempt_disable_ip);
 8759
 
 8760	dump_stack();
 8761	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 8762}
 8763EXPORT_SYMBOL(__might_resched);
 8764
 8765void __cant_sleep(const char *file, int line, int preempt_offset)
 8766{
 8767	static unsigned long prev_jiffy;
 8768
 8769	if (irqs_disabled())
 8770		return;
 8771
 8772	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
 8773		return;
 8774
 8775	if (preempt_count() > preempt_offset)
 8776		return;
 8777
 8778	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
 8779		return;
 8780	prev_jiffy = jiffies;
 8781
 8782	printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
 8783	printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
 8784			in_atomic(), irqs_disabled(),
 8785			current->pid, current->comm);
 8786
 8787	debug_show_held_locks(current);
 8788	dump_stack();
 8789	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 8790}
 8791EXPORT_SYMBOL_GPL(__cant_sleep);
 8792
 8793#ifdef CONFIG_SMP
 8794void __cant_migrate(const char *file, int line)
 8795{
 8796	static unsigned long prev_jiffy;
 8797
 8798	if (irqs_disabled())
 8799		return;
 8800
 8801	if (is_migration_disabled(current))
 8802		return;
 8803
 8804	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
 8805		return;
 8806
 8807	if (preempt_count() > 0)
 8808		return;
 8809
 8810	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
 8811		return;
 8812	prev_jiffy = jiffies;
 8813
 8814	pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
 8815	pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
 8816	       in_atomic(), irqs_disabled(), is_migration_disabled(current),
 8817	       current->pid, current->comm);
 8818
 8819	debug_show_held_locks(current);
 8820	dump_stack();
 8821	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 8822}
 8823EXPORT_SYMBOL_GPL(__cant_migrate);
 8824#endif
 8825#endif
 8826
 8827#ifdef CONFIG_MAGIC_SYSRQ
 8828void normalize_rt_tasks(void)
 8829{
 8830	struct task_struct *g, *p;
 8831	struct sched_attr attr = {
 8832		.sched_policy = SCHED_NORMAL,
 8833	};
 8834
 8835	read_lock(&tasklist_lock);
 8836	for_each_process_thread(g, p) {
 8837		/*
 8838		 * Only normalize user tasks:
 8839		 */
 8840		if (p->flags & PF_KTHREAD)
 8841			continue;
 8842
 8843		p->se.exec_start = 0;
 8844		schedstat_set(p->stats.wait_start,  0);
 8845		schedstat_set(p->stats.sleep_start, 0);
 8846		schedstat_set(p->stats.block_start, 0);
 8847
 8848		if (!rt_or_dl_task(p)) {
 8849			/*
 8850			 * Renice negative nice level userspace
 8851			 * tasks back to 0:
 8852			 */
 8853			if (task_nice(p) < 0)
 8854				set_user_nice(p, 0);
 8855			continue;
 8856		}
 8857
 8858		__sched_setscheduler(p, &attr, false, false);
 8859	}
 8860	read_unlock(&tasklist_lock);
 8861}
 8862
 8863#endif /* CONFIG_MAGIC_SYSRQ */
 8864
 8865#if defined(CONFIG_KGDB_KDB)
 8866/*
 8867 * These functions are only useful for KDB.
 8868 *
 8869 * They can only be called when the whole system has been
 8870 * stopped - every CPU needs to be quiescent, and no scheduling
 8871 * activity can take place. Using them for anything else would
 8872 * be a serious bug, and as a result, they aren't even visible
 8873 * under any other configuration.
 8874 */
 8875
 8876/**
 8877 * curr_task - return the current task for a given CPU.
 8878 * @cpu: the processor in question.
 8879 *
 8880 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
 8881 *
 8882 * Return: The current task for @cpu.
 8883 */
 8884struct task_struct *curr_task(int cpu)
 8885{
 8886	return cpu_curr(cpu);
 8887}
 8888
 8889#endif /* defined(CONFIG_KGDB_KDB) */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 8890
 8891#ifdef CONFIG_CGROUP_SCHED
 8892/* task_group_lock serializes the addition/removal of task groups */
 8893static DEFINE_SPINLOCK(task_group_lock);
 8894
 8895static inline void alloc_uclamp_sched_group(struct task_group *tg,
 8896					    struct task_group *parent)
 8897{
 8898#ifdef CONFIG_UCLAMP_TASK_GROUP
 8899	enum uclamp_id clamp_id;
 8900
 8901	for_each_clamp_id(clamp_id) {
 8902		uclamp_se_set(&tg->uclamp_req[clamp_id],
 8903			      uclamp_none(clamp_id), false);
 8904		tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
 8905	}
 8906#endif
 8907}
 8908
 8909static void sched_free_group(struct task_group *tg)
 8910{
 8911	free_fair_sched_group(tg);
 8912	free_rt_sched_group(tg);
 8913	autogroup_free(tg);
 8914	kmem_cache_free(task_group_cache, tg);
 8915}
 8916
 8917static void sched_free_group_rcu(struct rcu_head *rcu)
 8918{
 8919	sched_free_group(container_of(rcu, struct task_group, rcu));
 8920}
 8921
 8922static void sched_unregister_group(struct task_group *tg)
 8923{
 8924	unregister_fair_sched_group(tg);
 8925	unregister_rt_sched_group(tg);
 8926	/*
 8927	 * We have to wait for yet another RCU grace period to expire, as
 8928	 * print_cfs_stats() might run concurrently.
 8929	 */
 8930	call_rcu(&tg->rcu, sched_free_group_rcu);
 8931}
 8932
 8933/* allocate runqueue etc for a new task group */
 8934struct task_group *sched_create_group(struct task_group *parent)
 8935{
 8936	struct task_group *tg;
 8937
 8938	tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
 8939	if (!tg)
 8940		return ERR_PTR(-ENOMEM);
 8941
 8942	if (!alloc_fair_sched_group(tg, parent))
 8943		goto err;
 8944
 8945	if (!alloc_rt_sched_group(tg, parent))
 8946		goto err;
 8947
 8948	scx_group_set_weight(tg, CGROUP_WEIGHT_DFL);
 8949	alloc_uclamp_sched_group(tg, parent);
 8950
 8951	return tg;
 8952
 8953err:
 8954	sched_free_group(tg);
 8955	return ERR_PTR(-ENOMEM);
 8956}
 8957
 8958void sched_online_group(struct task_group *tg, struct task_group *parent)
 8959{
 8960	unsigned long flags;
 8961
 8962	spin_lock_irqsave(&task_group_lock, flags);
 8963	list_add_rcu(&tg->list, &task_groups);
 8964
 8965	/* Root should already exist: */
 8966	WARN_ON(!parent);
 8967
 8968	tg->parent = parent;
 8969	INIT_LIST_HEAD(&tg->children);
 8970	list_add_rcu(&tg->siblings, &parent->children);
 8971	spin_unlock_irqrestore(&task_group_lock, flags);
 8972
 8973	online_fair_sched_group(tg);
 8974}
 8975
 8976/* RCU callback to free various structures associated with a task group */
 8977static void sched_unregister_group_rcu(struct rcu_head *rhp)
 8978{
 8979	/* Now it should be safe to free those cfs_rqs: */
 8980	sched_unregister_group(container_of(rhp, struct task_group, rcu));
 8981}
 8982
 8983void sched_destroy_group(struct task_group *tg)
 8984{
 8985	/* Wait for possible concurrent references to cfs_rqs complete: */
 8986	call_rcu(&tg->rcu, sched_unregister_group_rcu);
 8987}
 8988
 8989void sched_release_group(struct task_group *tg)
 8990{
 8991	unsigned long flags;
 8992
 8993	/*
 8994	 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
 8995	 * sched_cfs_period_timer()).
 8996	 *
 8997	 * For this to be effective, we have to wait for all pending users of
 8998	 * this task group to leave their RCU critical section to ensure no new
 8999	 * user will see our dying task group any more. Specifically ensure
 9000	 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
 9001	 *
 9002	 * We therefore defer calling unregister_fair_sched_group() to
 9003	 * sched_unregister_group() which is guarantied to get called only after the
 9004	 * current RCU grace period has expired.
 9005	 */
 9006	spin_lock_irqsave(&task_group_lock, flags);
 9007	list_del_rcu(&tg->list);
 9008	list_del_rcu(&tg->siblings);
 9009	spin_unlock_irqrestore(&task_group_lock, flags);
 9010}
 9011
 9012static struct task_group *sched_get_task_group(struct task_struct *tsk)
 9013{
 9014	struct task_group *tg;
 9015
 9016	/*
 9017	 * All callers are synchronized by task_rq_lock(); we do not use RCU
 9018	 * which is pointless here. Thus, we pass "true" to task_css_check()
 9019	 * to prevent lockdep warnings.
 9020	 */
 9021	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
 9022			  struct task_group, css);
 9023	tg = autogroup_task_group(tsk, tg);
 9024
 9025	return tg;
 9026}
 9027
 9028static void sched_change_group(struct task_struct *tsk, struct task_group *group)
 9029{
 9030	tsk->sched_task_group = group;
 9031
 9032#ifdef CONFIG_FAIR_GROUP_SCHED
 9033	if (tsk->sched_class->task_change_group)
 9034		tsk->sched_class->task_change_group(tsk);
 9035	else
 9036#endif
 9037		set_task_rq(tsk, task_cpu(tsk));
 9038}
 9039
 9040/*
 9041 * Change task's runqueue when it moves between groups.
 9042 *
 9043 * The caller of this function should have put the task in its new group by
 9044 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
 9045 * its new group.
 9046 */
 9047void sched_move_task(struct task_struct *tsk, bool for_autogroup)
 9048{
 9049	int queued, running, queue_flags =
 9050		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
 9051	struct task_group *group;
 9052	struct rq *rq;
 9053
 9054	CLASS(task_rq_lock, rq_guard)(tsk);
 9055	rq = rq_guard.rq;
 9056
 9057	/*
 9058	 * Esp. with SCHED_AUTOGROUP enabled it is possible to get superfluous
 9059	 * group changes.
 9060	 */
 9061	group = sched_get_task_group(tsk);
 9062	if (group == tsk->sched_task_group)
 9063		return;
 9064
 9065	update_rq_clock(rq);
 9066
 9067	running = task_current_donor(rq, tsk);
 9068	queued = task_on_rq_queued(tsk);
 9069
 9070	if (queued)
 9071		dequeue_task(rq, tsk, queue_flags);
 9072	if (running)
 9073		put_prev_task(rq, tsk);
 9074
 9075	sched_change_group(tsk, group);
 9076	if (!for_autogroup)
 9077		scx_cgroup_move_task(tsk);
 9078
 9079	if (queued)
 9080		enqueue_task(rq, tsk, queue_flags);
 9081	if (running) {
 9082		set_next_task(rq, tsk);
 9083		/*
 9084		 * After changing group, the running task may have joined a
 9085		 * throttled one but it's still the running task. Trigger a
 9086		 * resched to make sure that task can still run.
 9087		 */
 9088		resched_curr(rq);
 9089	}
 
 
 
 
 
 
 
 9090}
 9091
 9092static struct cgroup_subsys_state *
 9093cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
 9094{
 9095	struct task_group *parent = css_tg(parent_css);
 9096	struct task_group *tg;
 9097
 9098	if (!parent) {
 9099		/* This is early initialization for the top cgroup */
 9100		return &root_task_group.css;
 9101	}
 9102
 9103	tg = sched_create_group(parent);
 9104	if (IS_ERR(tg))
 9105		return ERR_PTR(-ENOMEM);
 9106
 9107	return &tg->css;
 9108}
 9109
 9110/* Expose task group only after completing cgroup initialization */
 9111static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
 9112{
 9113	struct task_group *tg = css_tg(css);
 9114	struct task_group *parent = css_tg(css->parent);
 9115	int ret;
 9116
 9117	ret = scx_tg_online(tg);
 9118	if (ret)
 9119		return ret;
 9120
 9121	if (parent)
 9122		sched_online_group(tg, parent);
 9123
 9124#ifdef CONFIG_UCLAMP_TASK_GROUP
 9125	/* Propagate the effective uclamp value for the new group */
 9126	guard(mutex)(&uclamp_mutex);
 9127	guard(rcu)();
 9128	cpu_util_update_eff(css);
 
 
 9129#endif
 9130
 9131	return 0;
 9132}
 9133
 9134static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
 9135{
 9136	struct task_group *tg = css_tg(css);
 9137
 9138	scx_tg_offline(tg);
 9139}
 9140
 9141static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
 9142{
 9143	struct task_group *tg = css_tg(css);
 9144
 9145	sched_release_group(tg);
 9146}
 9147
 9148static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
 9149{
 9150	struct task_group *tg = css_tg(css);
 9151
 9152	/*
 9153	 * Relies on the RCU grace period between css_released() and this.
 9154	 */
 9155	sched_unregister_group(tg);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 9156}
 9157
 9158static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
 9159{
 9160#ifdef CONFIG_RT_GROUP_SCHED
 9161	struct task_struct *task;
 9162	struct cgroup_subsys_state *css;
 
 9163
 9164	cgroup_taskset_for_each(task, css, tset) {
 
 9165		if (!sched_rt_can_attach(css_tg(css), task))
 9166			return -EINVAL;
 9167	}
 9168#endif
 9169	return scx_cgroup_can_attach(tset);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 9170}
 9171
 9172static void cpu_cgroup_attach(struct cgroup_taskset *tset)
 9173{
 9174	struct task_struct *task;
 9175	struct cgroup_subsys_state *css;
 9176
 9177	cgroup_taskset_for_each(task, css, tset)
 9178		sched_move_task(task, false);
 9179
 9180	scx_cgroup_finish_attach();
 9181}
 9182
 9183static void cpu_cgroup_cancel_attach(struct cgroup_taskset *tset)
 9184{
 9185	scx_cgroup_cancel_attach(tset);
 9186}
 9187
 9188#ifdef CONFIG_UCLAMP_TASK_GROUP
 9189static void cpu_util_update_eff(struct cgroup_subsys_state *css)
 9190{
 9191	struct cgroup_subsys_state *top_css = css;
 9192	struct uclamp_se *uc_parent = NULL;
 9193	struct uclamp_se *uc_se = NULL;
 9194	unsigned int eff[UCLAMP_CNT];
 9195	enum uclamp_id clamp_id;
 9196	unsigned int clamps;
 9197
 9198	lockdep_assert_held(&uclamp_mutex);
 9199	SCHED_WARN_ON(!rcu_read_lock_held());
 9200
 9201	css_for_each_descendant_pre(css, top_css) {
 9202		uc_parent = css_tg(css)->parent
 9203			? css_tg(css)->parent->uclamp : NULL;
 9204
 9205		for_each_clamp_id(clamp_id) {
 9206			/* Assume effective clamps matches requested clamps */
 9207			eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
 9208			/* Cap effective clamps with parent's effective clamps */
 9209			if (uc_parent &&
 9210			    eff[clamp_id] > uc_parent[clamp_id].value) {
 9211				eff[clamp_id] = uc_parent[clamp_id].value;
 9212			}
 9213		}
 9214		/* Ensure protection is always capped by limit */
 9215		eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
 9216
 9217		/* Propagate most restrictive effective clamps */
 9218		clamps = 0x0;
 9219		uc_se = css_tg(css)->uclamp;
 9220		for_each_clamp_id(clamp_id) {
 9221			if (eff[clamp_id] == uc_se[clamp_id].value)
 9222				continue;
 9223			uc_se[clamp_id].value = eff[clamp_id];
 9224			uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
 9225			clamps |= (0x1 << clamp_id);
 9226		}
 9227		if (!clamps) {
 9228			css = css_rightmost_descendant(css);
 9229			continue;
 9230		}
 9231
 9232		/* Immediately update descendants RUNNABLE tasks */
 9233		uclamp_update_active_tasks(css);
 9234	}
 9235}
 9236
 9237/*
 9238 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
 9239 * C expression. Since there is no way to convert a macro argument (N) into a
 9240 * character constant, use two levels of macros.
 9241 */
 9242#define _POW10(exp) ((unsigned int)1e##exp)
 9243#define POW10(exp) _POW10(exp)
 9244
 9245struct uclamp_request {
 9246#define UCLAMP_PERCENT_SHIFT	2
 9247#define UCLAMP_PERCENT_SCALE	(100 * POW10(UCLAMP_PERCENT_SHIFT))
 9248	s64 percent;
 9249	u64 util;
 9250	int ret;
 9251};
 9252
 9253static inline struct uclamp_request
 9254capacity_from_percent(char *buf)
 9255{
 9256	struct uclamp_request req = {
 9257		.percent = UCLAMP_PERCENT_SCALE,
 9258		.util = SCHED_CAPACITY_SCALE,
 9259		.ret = 0,
 9260	};
 9261
 9262	buf = strim(buf);
 9263	if (strcmp(buf, "max")) {
 9264		req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
 9265					     &req.percent);
 9266		if (req.ret)
 9267			return req;
 9268		if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
 9269			req.ret = -ERANGE;
 9270			return req;
 9271		}
 9272
 9273		req.util = req.percent << SCHED_CAPACITY_SHIFT;
 9274		req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
 9275	}
 9276
 9277	return req;
 9278}
 9279
 9280static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
 9281				size_t nbytes, loff_t off,
 9282				enum uclamp_id clamp_id)
 9283{
 9284	struct uclamp_request req;
 9285	struct task_group *tg;
 9286
 9287	req = capacity_from_percent(buf);
 9288	if (req.ret)
 9289		return req.ret;
 9290
 9291	static_branch_enable(&sched_uclamp_used);
 9292
 9293	guard(mutex)(&uclamp_mutex);
 9294	guard(rcu)();
 9295
 9296	tg = css_tg(of_css(of));
 9297	if (tg->uclamp_req[clamp_id].value != req.util)
 9298		uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
 9299
 9300	/*
 9301	 * Because of not recoverable conversion rounding we keep track of the
 9302	 * exact requested value
 9303	 */
 9304	tg->uclamp_pct[clamp_id] = req.percent;
 9305
 9306	/* Update effective clamps to track the most restrictive value */
 9307	cpu_util_update_eff(of_css(of));
 9308
 
 
 
 9309	return nbytes;
 9310}
 9311
 9312static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
 9313				    char *buf, size_t nbytes,
 9314				    loff_t off)
 9315{
 9316	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
 9317}
 9318
 9319static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
 9320				    char *buf, size_t nbytes,
 9321				    loff_t off)
 9322{
 9323	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
 9324}
 9325
 9326static inline void cpu_uclamp_print(struct seq_file *sf,
 9327				    enum uclamp_id clamp_id)
 9328{
 9329	struct task_group *tg;
 9330	u64 util_clamp;
 9331	u64 percent;
 9332	u32 rem;
 9333
 9334	scoped_guard (rcu) {
 9335		tg = css_tg(seq_css(sf));
 9336		util_clamp = tg->uclamp_req[clamp_id].value;
 9337	}
 9338
 9339	if (util_clamp == SCHED_CAPACITY_SCALE) {
 9340		seq_puts(sf, "max\n");
 9341		return;
 9342	}
 9343
 9344	percent = tg->uclamp_pct[clamp_id];
 9345	percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
 9346	seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
 9347}
 9348
 9349static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
 9350{
 9351	cpu_uclamp_print(sf, UCLAMP_MIN);
 9352	return 0;
 9353}
 9354
 9355static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
 9356{
 9357	cpu_uclamp_print(sf, UCLAMP_MAX);
 9358	return 0;
 9359}
 9360#endif /* CONFIG_UCLAMP_TASK_GROUP */
 9361
 9362#ifdef CONFIG_GROUP_SCHED_WEIGHT
 9363static unsigned long tg_weight(struct task_group *tg)
 9364{
 9365#ifdef CONFIG_FAIR_GROUP_SCHED
 9366	return scale_load_down(tg->shares);
 9367#else
 9368	return sched_weight_from_cgroup(tg->scx_weight);
 9369#endif
 9370}
 9371
 9372static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
 9373				struct cftype *cftype, u64 shareval)
 9374{
 9375	int ret;
 9376
 9377	if (shareval > scale_load_down(ULONG_MAX))
 9378		shareval = MAX_SHARES;
 9379	ret = sched_group_set_shares(css_tg(css), scale_load(shareval));
 9380	if (!ret)
 9381		scx_group_set_weight(css_tg(css),
 9382				     sched_weight_to_cgroup(shareval));
 9383	return ret;
 9384}
 9385
 9386static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
 9387			       struct cftype *cft)
 9388{
 9389	return tg_weight(css_tg(css));
 
 
 9390}
 9391#endif /* CONFIG_GROUP_SCHED_WEIGHT */
 9392
 9393#ifdef CONFIG_CFS_BANDWIDTH
 9394static DEFINE_MUTEX(cfs_constraints_mutex);
 9395
 9396const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
 9397static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
 9398/* More than 203 days if BW_SHIFT equals 20. */
 9399static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
 9400
 9401static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
 9402
 9403static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
 9404				u64 burst)
 9405{
 9406	int i, ret = 0, runtime_enabled, runtime_was_enabled;
 9407	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
 9408
 9409	if (tg == &root_task_group)
 9410		return -EINVAL;
 9411
 9412	/*
 9413	 * Ensure we have at some amount of bandwidth every period.  This is
 9414	 * to prevent reaching a state of large arrears when throttled via
 9415	 * entity_tick() resulting in prolonged exit starvation.
 9416	 */
 9417	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
 9418		return -EINVAL;
 9419
 9420	/*
 9421	 * Likewise, bound things on the other side by preventing insane quota
 9422	 * periods.  This also allows us to normalize in computing quota
 9423	 * feasibility.
 9424	 */
 9425	if (period > max_cfs_quota_period)
 9426		return -EINVAL;
 9427
 9428	/*
 9429	 * Bound quota to defend quota against overflow during bandwidth shift.
 9430	 */
 9431	if (quota != RUNTIME_INF && quota > max_cfs_runtime)
 9432		return -EINVAL;
 9433
 9434	if (quota != RUNTIME_INF && (burst > quota ||
 9435				     burst + quota > max_cfs_runtime))
 9436		return -EINVAL;
 9437
 9438	/*
 9439	 * Prevent race between setting of cfs_rq->runtime_enabled and
 9440	 * unthrottle_offline_cfs_rqs().
 9441	 */
 9442	guard(cpus_read_lock)();
 9443	guard(mutex)(&cfs_constraints_mutex);
 9444
 9445	ret = __cfs_schedulable(tg, period, quota);
 9446	if (ret)
 9447		return ret;
 9448
 9449	runtime_enabled = quota != RUNTIME_INF;
 9450	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
 9451	/*
 9452	 * If we need to toggle cfs_bandwidth_used, off->on must occur
 9453	 * before making related changes, and on->off must occur afterwards
 9454	 */
 9455	if (runtime_enabled && !runtime_was_enabled)
 9456		cfs_bandwidth_usage_inc();
 
 
 
 
 
 
 
 
 
 
 9457
 9458	scoped_guard (raw_spinlock_irq, &cfs_b->lock) {
 9459		cfs_b->period = ns_to_ktime(period);
 9460		cfs_b->quota = quota;
 9461		cfs_b->burst = burst;
 9462
 9463		__refill_cfs_bandwidth_runtime(cfs_b);
 9464
 9465		/*
 9466		 * Restart the period timer (if active) to handle new
 9467		 * period expiry:
 9468		 */
 9469		if (runtime_enabled)
 9470			start_cfs_bandwidth(cfs_b);
 9471	}
 9472
 9473	for_each_online_cpu(i) {
 9474		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
 9475		struct rq *rq = cfs_rq->rq;
 
 9476
 9477		guard(rq_lock_irq)(rq);
 9478		cfs_rq->runtime_enabled = runtime_enabled;
 9479		cfs_rq->runtime_remaining = 0;
 9480
 9481		if (cfs_rq->throttled)
 9482			unthrottle_cfs_rq(cfs_rq);
 
 9483	}
 9484
 9485	if (runtime_was_enabled && !runtime_enabled)
 9486		cfs_bandwidth_usage_dec();
 
 
 
 9487
 9488	return 0;
 9489}
 9490
 9491static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
 9492{
 9493	u64 quota, period, burst;
 9494
 9495	period = ktime_to_ns(tg->cfs_bandwidth.period);
 9496	burst = tg->cfs_bandwidth.burst;
 9497	if (cfs_quota_us < 0)
 9498		quota = RUNTIME_INF;
 9499	else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
 9500		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
 9501	else
 9502		return -EINVAL;
 9503
 9504	return tg_set_cfs_bandwidth(tg, period, quota, burst);
 9505}
 9506
 9507static long tg_get_cfs_quota(struct task_group *tg)
 9508{
 9509	u64 quota_us;
 9510
 9511	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
 9512		return -1;
 9513
 9514	quota_us = tg->cfs_bandwidth.quota;
 9515	do_div(quota_us, NSEC_PER_USEC);
 9516
 9517	return quota_us;
 9518}
 9519
 9520static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
 9521{
 9522	u64 quota, period, burst;
 9523
 9524	if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
 9525		return -EINVAL;
 9526
 9527	period = (u64)cfs_period_us * NSEC_PER_USEC;
 9528	quota = tg->cfs_bandwidth.quota;
 9529	burst = tg->cfs_bandwidth.burst;
 9530
 9531	return tg_set_cfs_bandwidth(tg, period, quota, burst);
 9532}
 9533
 9534static long tg_get_cfs_period(struct task_group *tg)
 9535{
 9536	u64 cfs_period_us;
 9537
 9538	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
 9539	do_div(cfs_period_us, NSEC_PER_USEC);
 9540
 9541	return cfs_period_us;
 9542}
 9543
 9544static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
 9545{
 9546	u64 quota, period, burst;
 9547
 9548	if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
 9549		return -EINVAL;
 9550
 9551	burst = (u64)cfs_burst_us * NSEC_PER_USEC;
 9552	period = ktime_to_ns(tg->cfs_bandwidth.period);
 9553	quota = tg->cfs_bandwidth.quota;
 9554
 9555	return tg_set_cfs_bandwidth(tg, period, quota, burst);
 9556}
 9557
 9558static long tg_get_cfs_burst(struct task_group *tg)
 9559{
 9560	u64 burst_us;
 9561
 9562	burst_us = tg->cfs_bandwidth.burst;
 9563	do_div(burst_us, NSEC_PER_USEC);
 9564
 9565	return burst_us;
 9566}
 9567
 9568static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
 9569				  struct cftype *cft)
 9570{
 9571	return tg_get_cfs_quota(css_tg(css));
 9572}
 9573
 9574static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
 9575				   struct cftype *cftype, s64 cfs_quota_us)
 9576{
 9577	return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
 9578}
 9579
 9580static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
 9581				   struct cftype *cft)
 9582{
 9583	return tg_get_cfs_period(css_tg(css));
 9584}
 9585
 9586static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
 9587				    struct cftype *cftype, u64 cfs_period_us)
 9588{
 9589	return tg_set_cfs_period(css_tg(css), cfs_period_us);
 9590}
 9591
 9592static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
 9593				  struct cftype *cft)
 9594{
 9595	return tg_get_cfs_burst(css_tg(css));
 9596}
 9597
 9598static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
 9599				   struct cftype *cftype, u64 cfs_burst_us)
 9600{
 9601	return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
 9602}
 9603
 9604struct cfs_schedulable_data {
 9605	struct task_group *tg;
 9606	u64 period, quota;
 9607};
 9608
 9609/*
 9610 * normalize group quota/period to be quota/max_period
 9611 * note: units are usecs
 9612 */
 9613static u64 normalize_cfs_quota(struct task_group *tg,
 9614			       struct cfs_schedulable_data *d)
 9615{
 9616	u64 quota, period;
 9617
 9618	if (tg == d->tg) {
 9619		period = d->period;
 9620		quota = d->quota;
 9621	} else {
 9622		period = tg_get_cfs_period(tg);
 9623		quota = tg_get_cfs_quota(tg);
 9624	}
 9625
 9626	/* note: these should typically be equivalent */
 9627	if (quota == RUNTIME_INF || quota == -1)
 9628		return RUNTIME_INF;
 9629
 9630	return to_ratio(period, quota);
 9631}
 9632
 9633static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
 9634{
 9635	struct cfs_schedulable_data *d = data;
 9636	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
 9637	s64 quota = 0, parent_quota = -1;
 9638
 9639	if (!tg->parent) {
 9640		quota = RUNTIME_INF;
 9641	} else {
 9642		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
 9643
 9644		quota = normalize_cfs_quota(tg, d);
 9645		parent_quota = parent_b->hierarchical_quota;
 9646
 9647		/*
 9648		 * Ensure max(child_quota) <= parent_quota.  On cgroup2,
 9649		 * always take the non-RUNTIME_INF min.  On cgroup1, only
 9650		 * inherit when no limit is set. In both cases this is used
 9651		 * by the scheduler to determine if a given CFS task has a
 9652		 * bandwidth constraint at some higher level.
 9653		 */
 9654		if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
 9655			if (quota == RUNTIME_INF)
 9656				quota = parent_quota;
 9657			else if (parent_quota != RUNTIME_INF)
 9658				quota = min(quota, parent_quota);
 9659		} else {
 9660			if (quota == RUNTIME_INF)
 9661				quota = parent_quota;
 9662			else if (parent_quota != RUNTIME_INF && quota > parent_quota)
 9663				return -EINVAL;
 9664		}
 9665	}
 9666	cfs_b->hierarchical_quota = quota;
 9667
 9668	return 0;
 9669}
 9670
 9671static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
 9672{
 
 9673	struct cfs_schedulable_data data = {
 9674		.tg = tg,
 9675		.period = period,
 9676		.quota = quota,
 9677	};
 9678
 9679	if (quota != RUNTIME_INF) {
 9680		do_div(data.period, NSEC_PER_USEC);
 9681		do_div(data.quota, NSEC_PER_USEC);
 9682	}
 9683
 9684	guard(rcu)();
 9685	return walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
 
 
 
 9686}
 9687
 9688static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
 9689{
 9690	struct task_group *tg = css_tg(seq_css(sf));
 9691	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
 9692
 9693	seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
 9694	seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
 9695	seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
 9696
 9697	if (schedstat_enabled() && tg != &root_task_group) {
 9698		struct sched_statistics *stats;
 9699		u64 ws = 0;
 9700		int i;
 9701
 9702		for_each_possible_cpu(i) {
 9703			stats = __schedstats_from_se(tg->se[i]);
 9704			ws += schedstat_val(stats->wait_sum);
 9705		}
 9706
 9707		seq_printf(sf, "wait_sum %llu\n", ws);
 9708	}
 9709
 9710	seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
 9711	seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
 9712
 9713	return 0;
 9714}
 9715
 9716static u64 throttled_time_self(struct task_group *tg)
 9717{
 9718	int i;
 9719	u64 total = 0;
 9720
 9721	for_each_possible_cpu(i) {
 9722		total += READ_ONCE(tg->cfs_rq[i]->throttled_clock_self_time);
 9723	}
 9724
 9725	return total;
 9726}
 9727
 9728static int cpu_cfs_local_stat_show(struct seq_file *sf, void *v)
 9729{
 9730	struct task_group *tg = css_tg(seq_css(sf));
 9731
 9732	seq_printf(sf, "throttled_time %llu\n", throttled_time_self(tg));
 9733
 9734	return 0;
 9735}
 9736#endif /* CONFIG_CFS_BANDWIDTH */
 
 9737
 9738#ifdef CONFIG_RT_GROUP_SCHED
 9739static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
 9740				struct cftype *cft, s64 val)
 9741{
 9742	return sched_group_set_rt_runtime(css_tg(css), val);
 9743}
 9744
 9745static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
 9746			       struct cftype *cft)
 9747{
 9748	return sched_group_rt_runtime(css_tg(css));
 9749}
 9750
 9751static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
 9752				    struct cftype *cftype, u64 rt_period_us)
 9753{
 9754	return sched_group_set_rt_period(css_tg(css), rt_period_us);
 9755}
 9756
 9757static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
 9758				   struct cftype *cft)
 9759{
 9760	return sched_group_rt_period(css_tg(css));
 9761}
 9762#endif /* CONFIG_RT_GROUP_SCHED */
 9763
 9764#ifdef CONFIG_GROUP_SCHED_WEIGHT
 9765static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
 9766			       struct cftype *cft)
 9767{
 9768	return css_tg(css)->idle;
 9769}
 9770
 9771static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
 9772				struct cftype *cft, s64 idle)
 9773{
 9774	int ret;
 9775
 9776	ret = sched_group_set_idle(css_tg(css), idle);
 9777	if (!ret)
 9778		scx_group_set_idle(css_tg(css), idle);
 9779	return ret;
 9780}
 9781#endif
 9782
 9783static struct cftype cpu_legacy_files[] = {
 9784#ifdef CONFIG_GROUP_SCHED_WEIGHT
 9785	{
 9786		.name = "shares",
 9787		.read_u64 = cpu_shares_read_u64,
 9788		.write_u64 = cpu_shares_write_u64,
 9789	},
 9790	{
 9791		.name = "idle",
 9792		.read_s64 = cpu_idle_read_s64,
 9793		.write_s64 = cpu_idle_write_s64,
 9794	},
 9795#endif
 9796#ifdef CONFIG_CFS_BANDWIDTH
 9797	{
 9798		.name = "cfs_quota_us",
 9799		.read_s64 = cpu_cfs_quota_read_s64,
 9800		.write_s64 = cpu_cfs_quota_write_s64,
 9801	},
 9802	{
 9803		.name = "cfs_period_us",
 9804		.read_u64 = cpu_cfs_period_read_u64,
 9805		.write_u64 = cpu_cfs_period_write_u64,
 9806	},
 9807	{
 9808		.name = "cfs_burst_us",
 9809		.read_u64 = cpu_cfs_burst_read_u64,
 9810		.write_u64 = cpu_cfs_burst_write_u64,
 9811	},
 9812	{
 9813		.name = "stat",
 9814		.seq_show = cpu_cfs_stat_show,
 9815	},
 9816	{
 9817		.name = "stat.local",
 9818		.seq_show = cpu_cfs_local_stat_show,
 9819	},
 9820#endif
 9821#ifdef CONFIG_RT_GROUP_SCHED
 9822	{
 9823		.name = "rt_runtime_us",
 9824		.read_s64 = cpu_rt_runtime_read,
 9825		.write_s64 = cpu_rt_runtime_write,
 9826	},
 9827	{
 9828		.name = "rt_period_us",
 9829		.read_u64 = cpu_rt_period_read_uint,
 9830		.write_u64 = cpu_rt_period_write_uint,
 9831	},
 9832#endif
 9833#ifdef CONFIG_UCLAMP_TASK_GROUP
 9834	{
 9835		.name = "uclamp.min",
 9836		.flags = CFTYPE_NOT_ON_ROOT,
 9837		.seq_show = cpu_uclamp_min_show,
 9838		.write = cpu_uclamp_min_write,
 9839	},
 9840	{
 9841		.name = "uclamp.max",
 9842		.flags = CFTYPE_NOT_ON_ROOT,
 9843		.seq_show = cpu_uclamp_max_show,
 9844		.write = cpu_uclamp_max_write,
 9845	},
 9846#endif
 9847	{ }	/* Terminate */
 9848};
 9849
 9850static int cpu_extra_stat_show(struct seq_file *sf,
 9851			       struct cgroup_subsys_state *css)
 9852{
 9853#ifdef CONFIG_CFS_BANDWIDTH
 9854	{
 9855		struct task_group *tg = css_tg(css);
 9856		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
 9857		u64 throttled_usec, burst_usec;
 9858
 9859		throttled_usec = cfs_b->throttled_time;
 9860		do_div(throttled_usec, NSEC_PER_USEC);
 9861		burst_usec = cfs_b->burst_time;
 9862		do_div(burst_usec, NSEC_PER_USEC);
 9863
 9864		seq_printf(sf, "nr_periods %d\n"
 9865			   "nr_throttled %d\n"
 9866			   "throttled_usec %llu\n"
 9867			   "nr_bursts %d\n"
 9868			   "burst_usec %llu\n",
 9869			   cfs_b->nr_periods, cfs_b->nr_throttled,
 9870			   throttled_usec, cfs_b->nr_burst, burst_usec);
 9871	}
 9872#endif
 9873	return 0;
 9874}
 9875
 9876static int cpu_local_stat_show(struct seq_file *sf,
 9877			       struct cgroup_subsys_state *css)
 9878{
 9879#ifdef CONFIG_CFS_BANDWIDTH
 9880	{
 9881		struct task_group *tg = css_tg(css);
 9882		u64 throttled_self_usec;
 9883
 9884		throttled_self_usec = throttled_time_self(tg);
 9885		do_div(throttled_self_usec, NSEC_PER_USEC);
 9886
 9887		seq_printf(sf, "throttled_usec %llu\n",
 9888			   throttled_self_usec);
 9889	}
 9890#endif
 9891	return 0;
 9892}
 9893
 9894#ifdef CONFIG_GROUP_SCHED_WEIGHT
 9895
 9896static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
 9897			       struct cftype *cft)
 9898{
 9899	return sched_weight_to_cgroup(tg_weight(css_tg(css)));
 
 
 
 9900}
 9901
 9902static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
 9903				struct cftype *cft, u64 cgrp_weight)
 9904{
 9905	unsigned long weight;
 9906	int ret;
 9907
 9908	if (cgrp_weight < CGROUP_WEIGHT_MIN || cgrp_weight > CGROUP_WEIGHT_MAX)
 
 
 
 
 9909		return -ERANGE;
 9910
 9911	weight = sched_weight_from_cgroup(cgrp_weight);
 9912
 9913	ret = sched_group_set_shares(css_tg(css), scale_load(weight));
 9914	if (!ret)
 9915		scx_group_set_weight(css_tg(css), cgrp_weight);
 9916	return ret;
 9917}
 9918
 9919static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
 9920				    struct cftype *cft)
 9921{
 9922	unsigned long weight = tg_weight(css_tg(css));
 9923	int last_delta = INT_MAX;
 9924	int prio, delta;
 9925
 9926	/* find the closest nice value to the current weight */
 9927	for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
 9928		delta = abs(sched_prio_to_weight[prio] - weight);
 9929		if (delta >= last_delta)
 9930			break;
 9931		last_delta = delta;
 9932	}
 9933
 9934	return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
 9935}
 9936
 9937static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
 9938				     struct cftype *cft, s64 nice)
 9939{
 9940	unsigned long weight;
 9941	int idx, ret;
 9942
 9943	if (nice < MIN_NICE || nice > MAX_NICE)
 9944		return -ERANGE;
 9945
 9946	idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
 9947	idx = array_index_nospec(idx, 40);
 9948	weight = sched_prio_to_weight[idx];
 9949
 9950	ret = sched_group_set_shares(css_tg(css), scale_load(weight));
 9951	if (!ret)
 9952		scx_group_set_weight(css_tg(css),
 9953				     sched_weight_to_cgroup(weight));
 9954	return ret;
 9955}
 9956#endif /* CONFIG_GROUP_SCHED_WEIGHT */
 9957
 9958static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
 9959						  long period, long quota)
 9960{
 9961	if (quota < 0)
 9962		seq_puts(sf, "max");
 9963	else
 9964		seq_printf(sf, "%ld", quota);
 9965
 9966	seq_printf(sf, " %ld\n", period);
 9967}
 9968
 9969/* caller should put the current value in *@periodp before calling */
 9970static int __maybe_unused cpu_period_quota_parse(char *buf,
 9971						 u64 *periodp, u64 *quotap)
 9972{
 9973	char tok[21];	/* U64_MAX */
 9974
 9975	if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
 9976		return -EINVAL;
 9977
 9978	*periodp *= NSEC_PER_USEC;
 9979
 9980	if (sscanf(tok, "%llu", quotap))
 9981		*quotap *= NSEC_PER_USEC;
 9982	else if (!strcmp(tok, "max"))
 9983		*quotap = RUNTIME_INF;
 9984	else
 9985		return -EINVAL;
 9986
 9987	return 0;
 9988}
 9989
 9990#ifdef CONFIG_CFS_BANDWIDTH
 9991static int cpu_max_show(struct seq_file *sf, void *v)
 9992{
 9993	struct task_group *tg = css_tg(seq_css(sf));
 9994
 9995	cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
 9996	return 0;
 9997}
 9998
 9999static ssize_t cpu_max_write(struct kernfs_open_file *of,
10000			     char *buf, size_t nbytes, loff_t off)
10001{
10002	struct task_group *tg = css_tg(of_css(of));
10003	u64 period = tg_get_cfs_period(tg);
10004	u64 burst = tg->cfs_bandwidth.burst;
10005	u64 quota;
10006	int ret;
10007
10008	ret = cpu_period_quota_parse(buf, &period, &quota);
10009	if (!ret)
10010		ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10011	return ret ?: nbytes;
10012}
10013#endif
10014
10015static struct cftype cpu_files[] = {
10016#ifdef CONFIG_GROUP_SCHED_WEIGHT
10017	{
10018		.name = "weight",
10019		.flags = CFTYPE_NOT_ON_ROOT,
10020		.read_u64 = cpu_weight_read_u64,
10021		.write_u64 = cpu_weight_write_u64,
10022	},
10023	{
10024		.name = "weight.nice",
10025		.flags = CFTYPE_NOT_ON_ROOT,
10026		.read_s64 = cpu_weight_nice_read_s64,
10027		.write_s64 = cpu_weight_nice_write_s64,
10028	},
10029	{
10030		.name = "idle",
10031		.flags = CFTYPE_NOT_ON_ROOT,
10032		.read_s64 = cpu_idle_read_s64,
10033		.write_s64 = cpu_idle_write_s64,
10034	},
10035#endif
10036#ifdef CONFIG_CFS_BANDWIDTH
10037	{
10038		.name = "max",
10039		.flags = CFTYPE_NOT_ON_ROOT,
10040		.seq_show = cpu_max_show,
10041		.write = cpu_max_write,
10042	},
10043	{
10044		.name = "max.burst",
10045		.flags = CFTYPE_NOT_ON_ROOT,
10046		.read_u64 = cpu_cfs_burst_read_u64,
10047		.write_u64 = cpu_cfs_burst_write_u64,
10048	},
10049#endif
10050#ifdef CONFIG_UCLAMP_TASK_GROUP
10051	{
10052		.name = "uclamp.min",
10053		.flags = CFTYPE_NOT_ON_ROOT,
10054		.seq_show = cpu_uclamp_min_show,
10055		.write = cpu_uclamp_min_write,
10056	},
10057	{
10058		.name = "uclamp.max",
10059		.flags = CFTYPE_NOT_ON_ROOT,
10060		.seq_show = cpu_uclamp_max_show,
10061		.write = cpu_uclamp_max_write,
10062	},
10063#endif
10064	{ }	/* terminate */
10065};
10066
10067struct cgroup_subsys cpu_cgrp_subsys = {
10068	.css_alloc	= cpu_cgroup_css_alloc,
10069	.css_online	= cpu_cgroup_css_online,
10070	.css_offline	= cpu_cgroup_css_offline,
10071	.css_released	= cpu_cgroup_css_released,
10072	.css_free	= cpu_cgroup_css_free,
10073	.css_extra_stat_show = cpu_extra_stat_show,
10074	.css_local_stat_show = cpu_local_stat_show,
10075	.can_attach	= cpu_cgroup_can_attach,
10076	.attach		= cpu_cgroup_attach,
10077	.cancel_attach	= cpu_cgroup_cancel_attach,
10078	.legacy_cftypes	= cpu_legacy_files,
10079	.dfl_cftypes	= cpu_files,
10080	.early_init	= true,
10081	.threaded	= true,
10082};
10083
10084#endif	/* CONFIG_CGROUP_SCHED */
10085
10086void dump_cpu_task(int cpu)
10087{
10088	if (in_hardirq() && cpu == smp_processor_id()) {
10089		struct pt_regs *regs;
10090
10091		regs = get_irq_regs();
10092		if (regs) {
10093			show_regs(regs);
10094			return;
10095		}
10096	}
10097
10098	if (trigger_single_cpu_backtrace(cpu))
10099		return;
10100
10101	pr_info("Task dump for CPU %d:\n", cpu);
10102	sched_show_task(cpu_curr(cpu));
10103}
10104
10105/*
10106 * Nice levels are multiplicative, with a gentle 10% change for every
10107 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10108 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10109 * that remained on nice 0.
10110 *
10111 * The "10% effect" is relative and cumulative: from _any_ nice level,
10112 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10113 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10114 * If a task goes up by ~10% and another task goes down by ~10% then
10115 * the relative distance between them is ~25%.)
10116 */
10117const int sched_prio_to_weight[40] = {
10118 /* -20 */     88761,     71755,     56483,     46273,     36291,
10119 /* -15 */     29154,     23254,     18705,     14949,     11916,
10120 /* -10 */      9548,      7620,      6100,      4904,      3906,
10121 /*  -5 */      3121,      2501,      1991,      1586,      1277,
10122 /*   0 */      1024,       820,       655,       526,       423,
10123 /*   5 */       335,       272,       215,       172,       137,
10124 /*  10 */       110,        87,        70,        56,        45,
10125 /*  15 */        36,        29,        23,        18,        15,
10126};
10127
10128/*
10129 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, pre-calculated.
10130 *
10131 * In cases where the weight does not change often, we can use the
10132 * pre-calculated inverse to speed up arithmetics by turning divisions
10133 * into multiplications:
10134 */
10135const u32 sched_prio_to_wmult[40] = {
10136 /* -20 */     48388,     59856,     76040,     92818,    118348,
10137 /* -15 */    147320,    184698,    229616,    287308,    360437,
10138 /* -10 */    449829,    563644,    704093,    875809,   1099582,
10139 /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
10140 /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
10141 /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
10142 /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
10143 /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10144};
10145
10146void call_trace_sched_update_nr_running(struct rq *rq, int count)
10147{
10148        trace_sched_update_nr_running_tp(rq, count);
10149}
10150
10151#ifdef CONFIG_SCHED_MM_CID
10152
10153/*
10154 * @cid_lock: Guarantee forward-progress of cid allocation.
10155 *
10156 * Concurrency ID allocation within a bitmap is mostly lock-free. The cid_lock
10157 * is only used when contention is detected by the lock-free allocation so
10158 * forward progress can be guaranteed.
10159 */
10160DEFINE_RAW_SPINLOCK(cid_lock);
10161
10162/*
10163 * @use_cid_lock: Select cid allocation behavior: lock-free vs spinlock.
10164 *
10165 * When @use_cid_lock is 0, the cid allocation is lock-free. When contention is
10166 * detected, it is set to 1 to ensure that all newly coming allocations are
10167 * serialized by @cid_lock until the allocation which detected contention
10168 * completes and sets @use_cid_lock back to 0. This guarantees forward progress
10169 * of a cid allocation.
10170 */
10171int use_cid_lock;
10172
10173/*
10174 * mm_cid remote-clear implements a lock-free algorithm to clear per-mm/cpu cid
10175 * concurrently with respect to the execution of the source runqueue context
10176 * switch.
10177 *
10178 * There is one basic properties we want to guarantee here:
10179 *
10180 * (1) Remote-clear should _never_ mark a per-cpu cid UNSET when it is actively
10181 * used by a task. That would lead to concurrent allocation of the cid and
10182 * userspace corruption.
10183 *
10184 * Provide this guarantee by introducing a Dekker memory ordering to guarantee
10185 * that a pair of loads observe at least one of a pair of stores, which can be
10186 * shown as:
10187 *
10188 *      X = Y = 0
10189 *
10190 *      w[X]=1          w[Y]=1
10191 *      MB              MB
10192 *      r[Y]=y          r[X]=x
10193 *
10194 * Which guarantees that x==0 && y==0 is impossible. But rather than using
10195 * values 0 and 1, this algorithm cares about specific state transitions of the
10196 * runqueue current task (as updated by the scheduler context switch), and the
10197 * per-mm/cpu cid value.
10198 *
10199 * Let's introduce task (Y) which has task->mm == mm and task (N) which has
10200 * task->mm != mm for the rest of the discussion. There are two scheduler state
10201 * transitions on context switch we care about:
10202 *
10203 * (TSA) Store to rq->curr with transition from (N) to (Y)
10204 *
10205 * (TSB) Store to rq->curr with transition from (Y) to (N)
10206 *
10207 * On the remote-clear side, there is one transition we care about:
10208 *
10209 * (TMA) cmpxchg to *pcpu_cid to set the LAZY flag
10210 *
10211 * There is also a transition to UNSET state which can be performed from all
10212 * sides (scheduler, remote-clear). It is always performed with a cmpxchg which
10213 * guarantees that only a single thread will succeed:
10214 *
10215 * (TMB) cmpxchg to *pcpu_cid to mark UNSET
10216 *
10217 * Just to be clear, what we do _not_ want to happen is a transition to UNSET
10218 * when a thread is actively using the cid (property (1)).
10219 *
10220 * Let's looks at the relevant combinations of TSA/TSB, and TMA transitions.
10221 *
10222 * Scenario A) (TSA)+(TMA) (from next task perspective)
10223 *
10224 * CPU0                                      CPU1
10225 *
10226 * Context switch CS-1                       Remote-clear
10227 *   - store to rq->curr: (N)->(Y) (TSA)     - cmpxchg to *pcpu_id to LAZY (TMA)
10228 *                                             (implied barrier after cmpxchg)
10229 *   - switch_mm_cid()
10230 *     - memory barrier (see switch_mm_cid()
10231 *       comment explaining how this barrier
10232 *       is combined with other scheduler
10233 *       barriers)
10234 *     - mm_cid_get (next)
10235 *       - READ_ONCE(*pcpu_cid)              - rcu_dereference(src_rq->curr)
10236 *
10237 * This Dekker ensures that either task (Y) is observed by the
10238 * rcu_dereference() or the LAZY flag is observed by READ_ONCE(), or both are
10239 * observed.
10240 *
10241 * If task (Y) store is observed by rcu_dereference(), it means that there is
10242 * still an active task on the cpu. Remote-clear will therefore not transition
10243 * to UNSET, which fulfills property (1).
10244 *
10245 * If task (Y) is not observed, but the lazy flag is observed by READ_ONCE(),
10246 * it will move its state to UNSET, which clears the percpu cid perhaps
10247 * uselessly (which is not an issue for correctness). Because task (Y) is not
10248 * observed, CPU1 can move ahead to set the state to UNSET. Because moving
10249 * state to UNSET is done with a cmpxchg expecting that the old state has the
10250 * LAZY flag set, only one thread will successfully UNSET.
10251 *
10252 * If both states (LAZY flag and task (Y)) are observed, the thread on CPU0
10253 * will observe the LAZY flag and transition to UNSET (perhaps uselessly), and
10254 * CPU1 will observe task (Y) and do nothing more, which is fine.
10255 *
10256 * What we are effectively preventing with this Dekker is a scenario where
10257 * neither LAZY flag nor store (Y) are observed, which would fail property (1)
10258 * because this would UNSET a cid which is actively used.
10259 */
10260
10261void sched_mm_cid_migrate_from(struct task_struct *t)
10262{
10263	t->migrate_from_cpu = task_cpu(t);
10264}
10265
10266static
10267int __sched_mm_cid_migrate_from_fetch_cid(struct rq *src_rq,
10268					  struct task_struct *t,
10269					  struct mm_cid *src_pcpu_cid)
10270{
10271	struct mm_struct *mm = t->mm;
10272	struct task_struct *src_task;
10273	int src_cid, last_mm_cid;
10274
10275	if (!mm)
10276		return -1;
10277
10278	last_mm_cid = t->last_mm_cid;
10279	/*
10280	 * If the migrated task has no last cid, or if the current
10281	 * task on src rq uses the cid, it means the source cid does not need
10282	 * to be moved to the destination cpu.
10283	 */
10284	if (last_mm_cid == -1)
10285		return -1;
10286	src_cid = READ_ONCE(src_pcpu_cid->cid);
10287	if (!mm_cid_is_valid(src_cid) || last_mm_cid != src_cid)
10288		return -1;
10289
10290	/*
10291	 * If we observe an active task using the mm on this rq, it means we
10292	 * are not the last task to be migrated from this cpu for this mm, so
10293	 * there is no need to move src_cid to the destination cpu.
10294	 */
10295	guard(rcu)();
10296	src_task = rcu_dereference(src_rq->curr);
10297	if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
10298		t->last_mm_cid = -1;
10299		return -1;
10300	}
10301
10302	return src_cid;
10303}
10304
10305static
10306int __sched_mm_cid_migrate_from_try_steal_cid(struct rq *src_rq,
10307					      struct task_struct *t,
10308					      struct mm_cid *src_pcpu_cid,
10309					      int src_cid)
10310{
10311	struct task_struct *src_task;
10312	struct mm_struct *mm = t->mm;
10313	int lazy_cid;
10314
10315	if (src_cid == -1)
10316		return -1;
10317
10318	/*
10319	 * Attempt to clear the source cpu cid to move it to the destination
10320	 * cpu.
10321	 */
10322	lazy_cid = mm_cid_set_lazy_put(src_cid);
10323	if (!try_cmpxchg(&src_pcpu_cid->cid, &src_cid, lazy_cid))
10324		return -1;
10325
10326	/*
10327	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
10328	 * rq->curr->mm matches the scheduler barrier in context_switch()
10329	 * between store to rq->curr and load of prev and next task's
10330	 * per-mm/cpu cid.
10331	 *
10332	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
10333	 * rq->curr->mm_cid_active matches the barrier in
10334	 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
10335	 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
10336	 * load of per-mm/cpu cid.
10337	 */
10338
10339	/*
10340	 * If we observe an active task using the mm on this rq after setting
10341	 * the lazy-put flag, this task will be responsible for transitioning
10342	 * from lazy-put flag set to MM_CID_UNSET.
10343	 */
10344	scoped_guard (rcu) {
10345		src_task = rcu_dereference(src_rq->curr);
10346		if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
10347			/*
10348			 * We observed an active task for this mm, there is therefore
10349			 * no point in moving this cid to the destination cpu.
10350			 */
10351			t->last_mm_cid = -1;
10352			return -1;
10353		}
10354	}
10355
10356	/*
10357	 * The src_cid is unused, so it can be unset.
10358	 */
10359	if (!try_cmpxchg(&src_pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
10360		return -1;
10361	WRITE_ONCE(src_pcpu_cid->recent_cid, MM_CID_UNSET);
10362	return src_cid;
10363}
10364
10365/*
10366 * Migration to dst cpu. Called with dst_rq lock held.
10367 * Interrupts are disabled, which keeps the window of cid ownership without the
10368 * source rq lock held small.
10369 */
10370void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t)
10371{
10372	struct mm_cid *src_pcpu_cid, *dst_pcpu_cid;
10373	struct mm_struct *mm = t->mm;
10374	int src_cid, src_cpu;
10375	bool dst_cid_is_set;
10376	struct rq *src_rq;
10377
10378	lockdep_assert_rq_held(dst_rq);
10379
10380	if (!mm)
10381		return;
10382	src_cpu = t->migrate_from_cpu;
10383	if (src_cpu == -1) {
10384		t->last_mm_cid = -1;
10385		return;
10386	}
10387	/*
10388	 * Move the src cid if the dst cid is unset. This keeps id
10389	 * allocation closest to 0 in cases where few threads migrate around
10390	 * many CPUs.
10391	 *
10392	 * If destination cid or recent cid is already set, we may have
10393	 * to just clear the src cid to ensure compactness in frequent
10394	 * migrations scenarios.
10395	 *
10396	 * It is not useful to clear the src cid when the number of threads is
10397	 * greater or equal to the number of allowed CPUs, because user-space
10398	 * can expect that the number of allowed cids can reach the number of
10399	 * allowed CPUs.
10400	 */
10401	dst_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(dst_rq));
10402	dst_cid_is_set = !mm_cid_is_unset(READ_ONCE(dst_pcpu_cid->cid)) ||
10403			 !mm_cid_is_unset(READ_ONCE(dst_pcpu_cid->recent_cid));
10404	if (dst_cid_is_set && atomic_read(&mm->mm_users) >= READ_ONCE(mm->nr_cpus_allowed))
10405		return;
10406	src_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, src_cpu);
10407	src_rq = cpu_rq(src_cpu);
10408	src_cid = __sched_mm_cid_migrate_from_fetch_cid(src_rq, t, src_pcpu_cid);
10409	if (src_cid == -1)
10410		return;
10411	src_cid = __sched_mm_cid_migrate_from_try_steal_cid(src_rq, t, src_pcpu_cid,
10412							    src_cid);
10413	if (src_cid == -1)
10414		return;
10415	if (dst_cid_is_set) {
10416		__mm_cid_put(mm, src_cid);
10417		return;
10418	}
10419	/* Move src_cid to dst cpu. */
10420	mm_cid_snapshot_time(dst_rq, mm);
10421	WRITE_ONCE(dst_pcpu_cid->cid, src_cid);
10422	WRITE_ONCE(dst_pcpu_cid->recent_cid, src_cid);
10423}
10424
10425static void sched_mm_cid_remote_clear(struct mm_struct *mm, struct mm_cid *pcpu_cid,
10426				      int cpu)
10427{
10428	struct rq *rq = cpu_rq(cpu);
10429	struct task_struct *t;
10430	int cid, lazy_cid;
10431
10432	cid = READ_ONCE(pcpu_cid->cid);
10433	if (!mm_cid_is_valid(cid))
10434		return;
10435
10436	/*
10437	 * Clear the cpu cid if it is set to keep cid allocation compact.  If
10438	 * there happens to be other tasks left on the source cpu using this
10439	 * mm, the next task using this mm will reallocate its cid on context
10440	 * switch.
10441	 */
10442	lazy_cid = mm_cid_set_lazy_put(cid);
10443	if (!try_cmpxchg(&pcpu_cid->cid, &cid, lazy_cid))
10444		return;
10445
10446	/*
10447	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
10448	 * rq->curr->mm matches the scheduler barrier in context_switch()
10449	 * between store to rq->curr and load of prev and next task's
10450	 * per-mm/cpu cid.
10451	 *
10452	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
10453	 * rq->curr->mm_cid_active matches the barrier in
10454	 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
10455	 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
10456	 * load of per-mm/cpu cid.
10457	 */
10458
10459	/*
10460	 * If we observe an active task using the mm on this rq after setting
10461	 * the lazy-put flag, that task will be responsible for transitioning
10462	 * from lazy-put flag set to MM_CID_UNSET.
10463	 */
10464	scoped_guard (rcu) {
10465		t = rcu_dereference(rq->curr);
10466		if (READ_ONCE(t->mm_cid_active) && t->mm == mm)
10467			return;
10468	}
10469
10470	/*
10471	 * The cid is unused, so it can be unset.
10472	 * Disable interrupts to keep the window of cid ownership without rq
10473	 * lock small.
10474	 */
10475	scoped_guard (irqsave) {
10476		if (try_cmpxchg(&pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
10477			__mm_cid_put(mm, cid);
10478	}
10479}
10480
10481static void sched_mm_cid_remote_clear_old(struct mm_struct *mm, int cpu)
10482{
10483	struct rq *rq = cpu_rq(cpu);
10484	struct mm_cid *pcpu_cid;
10485	struct task_struct *curr;
10486	u64 rq_clock;
10487
10488	/*
10489	 * rq->clock load is racy on 32-bit but one spurious clear once in a
10490	 * while is irrelevant.
10491	 */
10492	rq_clock = READ_ONCE(rq->clock);
10493	pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
10494
10495	/*
10496	 * In order to take care of infrequently scheduled tasks, bump the time
10497	 * snapshot associated with this cid if an active task using the mm is
10498	 * observed on this rq.
10499	 */
10500	scoped_guard (rcu) {
10501		curr = rcu_dereference(rq->curr);
10502		if (READ_ONCE(curr->mm_cid_active) && curr->mm == mm) {
10503			WRITE_ONCE(pcpu_cid->time, rq_clock);
10504			return;
10505		}
10506	}
10507
10508	if (rq_clock < pcpu_cid->time + SCHED_MM_CID_PERIOD_NS)
10509		return;
10510	sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
10511}
10512
10513static void sched_mm_cid_remote_clear_weight(struct mm_struct *mm, int cpu,
10514					     int weight)
10515{
10516	struct mm_cid *pcpu_cid;
10517	int cid;
10518
10519	pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
10520	cid = READ_ONCE(pcpu_cid->cid);
10521	if (!mm_cid_is_valid(cid) || cid < weight)
10522		return;
10523	sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
10524}
10525
10526static void task_mm_cid_work(struct callback_head *work)
10527{
10528	unsigned long now = jiffies, old_scan, next_scan;
10529	struct task_struct *t = current;
10530	struct cpumask *cidmask;
10531	struct mm_struct *mm;
10532	int weight, cpu;
10533
10534	SCHED_WARN_ON(t != container_of(work, struct task_struct, cid_work));
10535
10536	work->next = work;	/* Prevent double-add */
10537	if (t->flags & PF_EXITING)
10538		return;
10539	mm = t->mm;
10540	if (!mm)
10541		return;
10542	old_scan = READ_ONCE(mm->mm_cid_next_scan);
10543	next_scan = now + msecs_to_jiffies(MM_CID_SCAN_DELAY);
10544	if (!old_scan) {
10545		unsigned long res;
10546
10547		res = cmpxchg(&mm->mm_cid_next_scan, old_scan, next_scan);
10548		if (res != old_scan)
10549			old_scan = res;
10550		else
10551			old_scan = next_scan;
10552	}
10553	if (time_before(now, old_scan))
10554		return;
10555	if (!try_cmpxchg(&mm->mm_cid_next_scan, &old_scan, next_scan))
10556		return;
10557	cidmask = mm_cidmask(mm);
10558	/* Clear cids that were not recently used. */
10559	for_each_possible_cpu(cpu)
10560		sched_mm_cid_remote_clear_old(mm, cpu);
10561	weight = cpumask_weight(cidmask);
10562	/*
10563	 * Clear cids that are greater or equal to the cidmask weight to
10564	 * recompact it.
10565	 */
10566	for_each_possible_cpu(cpu)
10567		sched_mm_cid_remote_clear_weight(mm, cpu, weight);
10568}
10569
10570void init_sched_mm_cid(struct task_struct *t)
10571{
10572	struct mm_struct *mm = t->mm;
10573	int mm_users = 0;
10574
10575	if (mm) {
10576		mm_users = atomic_read(&mm->mm_users);
10577		if (mm_users == 1)
10578			mm->mm_cid_next_scan = jiffies + msecs_to_jiffies(MM_CID_SCAN_DELAY);
10579	}
10580	t->cid_work.next = &t->cid_work;	/* Protect against double add */
10581	init_task_work(&t->cid_work, task_mm_cid_work);
10582}
10583
10584void task_tick_mm_cid(struct rq *rq, struct task_struct *curr)
10585{
10586	struct callback_head *work = &curr->cid_work;
10587	unsigned long now = jiffies;
10588
10589	if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) ||
10590	    work->next != work)
10591		return;
10592	if (time_before(now, READ_ONCE(curr->mm->mm_cid_next_scan)))
10593		return;
10594
10595	/* No page allocation under rq lock */
10596	task_work_add(curr, work, TWA_RESUME | TWAF_NO_ALLOC);
10597}
10598
10599void sched_mm_cid_exit_signals(struct task_struct *t)
10600{
10601	struct mm_struct *mm = t->mm;
10602	struct rq *rq;
10603
10604	if (!mm)
10605		return;
10606
10607	preempt_disable();
10608	rq = this_rq();
10609	guard(rq_lock_irqsave)(rq);
10610	preempt_enable_no_resched();	/* holding spinlock */
10611	WRITE_ONCE(t->mm_cid_active, 0);
10612	/*
10613	 * Store t->mm_cid_active before loading per-mm/cpu cid.
10614	 * Matches barrier in sched_mm_cid_remote_clear_old().
10615	 */
10616	smp_mb();
10617	mm_cid_put(mm);
10618	t->last_mm_cid = t->mm_cid = -1;
10619}
10620
10621void sched_mm_cid_before_execve(struct task_struct *t)
10622{
10623	struct mm_struct *mm = t->mm;
10624	struct rq *rq;
10625
10626	if (!mm)
10627		return;
10628
10629	preempt_disable();
10630	rq = this_rq();
10631	guard(rq_lock_irqsave)(rq);
10632	preempt_enable_no_resched();	/* holding spinlock */
10633	WRITE_ONCE(t->mm_cid_active, 0);
10634	/*
10635	 * Store t->mm_cid_active before loading per-mm/cpu cid.
10636	 * Matches barrier in sched_mm_cid_remote_clear_old().
10637	 */
10638	smp_mb();
10639	mm_cid_put(mm);
10640	t->last_mm_cid = t->mm_cid = -1;
10641}
10642
10643void sched_mm_cid_after_execve(struct task_struct *t)
10644{
10645	struct mm_struct *mm = t->mm;
10646	struct rq *rq;
10647
10648	if (!mm)
10649		return;
10650
10651	preempt_disable();
10652	rq = this_rq();
10653	scoped_guard (rq_lock_irqsave, rq) {
10654		preempt_enable_no_resched();	/* holding spinlock */
10655		WRITE_ONCE(t->mm_cid_active, 1);
10656		/*
10657		 * Store t->mm_cid_active before loading per-mm/cpu cid.
10658		 * Matches barrier in sched_mm_cid_remote_clear_old().
10659		 */
10660		smp_mb();
10661		t->last_mm_cid = t->mm_cid = mm_cid_get(rq, t, mm);
10662	}
10663	rseq_set_notify_resume(t);
10664}
10665
10666void sched_mm_cid_fork(struct task_struct *t)
10667{
10668	WARN_ON_ONCE(!t->mm || t->mm_cid != -1);
10669	t->mm_cid_active = 1;
10670}
10671#endif
10672
10673#ifdef CONFIG_SCHED_CLASS_EXT
10674void sched_deq_and_put_task(struct task_struct *p, int queue_flags,
10675			    struct sched_enq_and_set_ctx *ctx)
10676{
10677	struct rq *rq = task_rq(p);
10678
10679	lockdep_assert_rq_held(rq);
10680
10681	*ctx = (struct sched_enq_and_set_ctx){
10682		.p = p,
10683		.queue_flags = queue_flags,
10684		.queued = task_on_rq_queued(p),
10685		.running = task_current(rq, p),
10686	};
10687
10688	update_rq_clock(rq);
10689	if (ctx->queued)
10690		dequeue_task(rq, p, queue_flags | DEQUEUE_NOCLOCK);
10691	if (ctx->running)
10692		put_prev_task(rq, p);
10693}
10694
10695void sched_enq_and_set_task(struct sched_enq_and_set_ctx *ctx)
10696{
10697	struct rq *rq = task_rq(ctx->p);
10698
10699	lockdep_assert_rq_held(rq);
10700
10701	if (ctx->queued)
10702		enqueue_task(rq, ctx->p, ctx->queue_flags | ENQUEUE_NOCLOCK);
10703	if (ctx->running)
10704		set_next_task(rq, ctx->p);
10705}
10706#endif	/* CONFIG_SCHED_CLASS_EXT */
v5.14.15
    1// SPDX-License-Identifier: GPL-2.0-only
    2/*
    3 *  kernel/sched/core.c
    4 *
    5 *  Core kernel scheduler code and related syscalls
    6 *
    7 *  Copyright (C) 1991-2002  Linus Torvalds
 
    8 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
    9#define CREATE_TRACE_POINTS
 
   10#include <trace/events/sched.h>
 
   11#undef CREATE_TRACE_POINTS
   12
   13#include "sched.h"
 
   14
   15#include <linux/nospec.h>
   16
   17#include <linux/kcov.h>
   18#include <linux/scs.h>
   19
   20#include <asm/switch_to.h>
   21#include <asm/tlb.h>
   22
   23#include "../workqueue_internal.h"
   24#include "../../fs/io-wq.h"
   25#include "../smpboot.h"
   26
   27#include "pelt.h"
   28#include "smp.h"
   29
   30/*
   31 * Export tracepoints that act as a bare tracehook (ie: have no trace event
   32 * associated with them) to allow external modules to probe them.
   33 */
   34EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
   35EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
   36EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
   37EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
   38EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
 
   39EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
   40EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
   41EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
   42EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
   43EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
 
   44
   45DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
   46
   47#ifdef CONFIG_SCHED_DEBUG
   48/*
   49 * Debugging: various feature bits
   50 *
   51 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
   52 * sysctl_sched_features, defined in sched.h, to allow constants propagation
   53 * at compile time and compiler optimization based on features default.
   54 */
   55#define SCHED_FEAT(name, enabled)	\
   56	(1UL << __SCHED_FEAT_##name) * enabled |
   57const_debug unsigned int sysctl_sched_features =
   58#include "features.h"
   59	0;
   60#undef SCHED_FEAT
   61
   62/*
   63 * Print a warning if need_resched is set for the given duration (if
   64 * LATENCY_WARN is enabled).
   65 *
   66 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
   67 * per boot.
   68 */
   69__read_mostly int sysctl_resched_latency_warn_ms = 100;
   70__read_mostly int sysctl_resched_latency_warn_once = 1;
   71#endif /* CONFIG_SCHED_DEBUG */
   72
   73/*
   74 * Number of tasks to iterate in a single balance run.
   75 * Limited because this is done with IRQs disabled.
   76 */
   77const_debug unsigned int sysctl_sched_nr_migrate = 32;
   78
   79/*
   80 * period over which we measure -rt task CPU usage in us.
   81 * default: 1s
   82 */
   83unsigned int sysctl_sched_rt_period = 1000000;
   84
   85__read_mostly int scheduler_running;
   86
   87#ifdef CONFIG_SCHED_CORE
   88
   89DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
   90
   91/* kernel prio, less is more */
   92static inline int __task_prio(struct task_struct *p)
   93{
   94	if (p->sched_class == &stop_sched_class) /* trumps deadline */
   95		return -2;
   96
   97	if (rt_prio(p->prio)) /* includes deadline */
 
 
 
   98		return p->prio; /* [-1, 99] */
   99
  100	if (p->sched_class == &idle_sched_class)
  101		return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
  102
  103	return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
 
 
 
  104}
  105
  106/*
  107 * l(a,b)
  108 * le(a,b) := !l(b,a)
  109 * g(a,b)  := l(b,a)
  110 * ge(a,b) := !l(a,b)
  111 */
  112
  113/* real prio, less is less */
  114static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
 
  115{
  116
  117	int pa = __task_prio(a), pb = __task_prio(b);
  118
  119	if (-pa < -pb)
  120		return true;
  121
  122	if (-pb < -pa)
  123		return false;
  124
  125	if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
  126		return !dl_time_before(a->dl.deadline, b->dl.deadline);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  127
  128	if (pa == MAX_RT_PRIO + MAX_NICE)	/* fair */
  129		return cfs_prio_less(a, b, in_fi);
  130
 
 
 
 
 
  131	return false;
  132}
  133
  134static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
 
  135{
  136	if (a->core_cookie < b->core_cookie)
  137		return true;
  138
  139	if (a->core_cookie > b->core_cookie)
  140		return false;
  141
  142	/* flip prio, so high prio is leftmost */
  143	if (prio_less(b, a, task_rq(a)->core->core_forceidle))
  144		return true;
  145
  146	return false;
  147}
  148
  149#define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
  150
  151static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
  152{
  153	return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
  154}
  155
  156static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
  157{
  158	const struct task_struct *p = __node_2_sc(node);
  159	unsigned long cookie = (unsigned long)key;
  160
  161	if (cookie < p->core_cookie)
  162		return -1;
  163
  164	if (cookie > p->core_cookie)
  165		return 1;
  166
  167	return 0;
  168}
  169
  170void sched_core_enqueue(struct rq *rq, struct task_struct *p)
  171{
 
 
 
  172	rq->core->core_task_seq++;
  173
  174	if (!p->core_cookie)
  175		return;
  176
  177	rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
  178}
  179
  180void sched_core_dequeue(struct rq *rq, struct task_struct *p)
  181{
 
 
 
  182	rq->core->core_task_seq++;
  183
  184	if (!sched_core_enqueued(p))
  185		return;
 
 
  186
  187	rb_erase(&p->core_node, &rq->core_tree);
  188	RB_CLEAR_NODE(&p->core_node);
 
 
 
 
 
 
  189}
  190
  191/*
  192 * Find left-most (aka, highest priority) task matching @cookie.
  193 */
  194static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
  195{
  196	struct rb_node *node;
 
  197
  198	node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
  199	/*
  200	 * The idle task always matches any cookie!
  201	 */
  202	if (!node)
  203		return idle_sched_class.pick_task(rq);
  204
  205	return __node_2_sc(node);
  206}
  207
  208static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
  209{
  210	struct rb_node *node = &p->core_node;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  211
  212	node = rb_next(node);
  213	if (!node)
  214		return NULL;
  215
  216	p = container_of(node, struct task_struct, core_node);
  217	if (p->core_cookie != cookie)
  218		return NULL;
  219
  220	return p;
  221}
  222
  223/*
  224 * Magic required such that:
  225 *
  226 *	raw_spin_rq_lock(rq);
  227 *	...
  228 *	raw_spin_rq_unlock(rq);
  229 *
  230 * ends up locking and unlocking the _same_ lock, and all CPUs
  231 * always agree on what rq has what lock.
  232 *
  233 * XXX entirely possible to selectively enable cores, don't bother for now.
  234 */
  235
  236static DEFINE_MUTEX(sched_core_mutex);
  237static atomic_t sched_core_count;
  238static struct cpumask sched_core_mask;
  239
  240static void sched_core_lock(int cpu, unsigned long *flags)
  241{
  242	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
  243	int t, i = 0;
  244
  245	local_irq_save(*flags);
  246	for_each_cpu(t, smt_mask)
  247		raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
  248}
  249
  250static void sched_core_unlock(int cpu, unsigned long *flags)
  251{
  252	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
  253	int t;
  254
  255	for_each_cpu(t, smt_mask)
  256		raw_spin_unlock(&cpu_rq(t)->__lock);
  257	local_irq_restore(*flags);
  258}
  259
  260static void __sched_core_flip(bool enabled)
  261{
  262	unsigned long flags;
  263	int cpu, t;
  264
  265	cpus_read_lock();
  266
  267	/*
  268	 * Toggle the online cores, one by one.
  269	 */
  270	cpumask_copy(&sched_core_mask, cpu_online_mask);
  271	for_each_cpu(cpu, &sched_core_mask) {
  272		const struct cpumask *smt_mask = cpu_smt_mask(cpu);
  273
  274		sched_core_lock(cpu, &flags);
  275
  276		for_each_cpu(t, smt_mask)
  277			cpu_rq(t)->core_enabled = enabled;
  278
 
 
  279		sched_core_unlock(cpu, &flags);
  280
  281		cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
  282	}
  283
  284	/*
  285	 * Toggle the offline CPUs.
  286	 */
  287	cpumask_copy(&sched_core_mask, cpu_possible_mask);
  288	cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
  289
  290	for_each_cpu(cpu, &sched_core_mask)
  291		cpu_rq(cpu)->core_enabled = enabled;
  292
  293	cpus_read_unlock();
  294}
  295
  296static void sched_core_assert_empty(void)
  297{
  298	int cpu;
  299
  300	for_each_possible_cpu(cpu)
  301		WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
  302}
  303
  304static void __sched_core_enable(void)
  305{
  306	static_branch_enable(&__sched_core_enabled);
  307	/*
  308	 * Ensure all previous instances of raw_spin_rq_*lock() have finished
  309	 * and future ones will observe !sched_core_disabled().
  310	 */
  311	synchronize_rcu();
  312	__sched_core_flip(true);
  313	sched_core_assert_empty();
  314}
  315
  316static void __sched_core_disable(void)
  317{
  318	sched_core_assert_empty();
  319	__sched_core_flip(false);
  320	static_branch_disable(&__sched_core_enabled);
  321}
  322
  323void sched_core_get(void)
  324{
  325	if (atomic_inc_not_zero(&sched_core_count))
  326		return;
  327
  328	mutex_lock(&sched_core_mutex);
  329	if (!atomic_read(&sched_core_count))
  330		__sched_core_enable();
  331
  332	smp_mb__before_atomic();
  333	atomic_inc(&sched_core_count);
  334	mutex_unlock(&sched_core_mutex);
  335}
  336
  337static void __sched_core_put(struct work_struct *work)
  338{
  339	if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
  340		__sched_core_disable();
  341		mutex_unlock(&sched_core_mutex);
  342	}
  343}
  344
  345void sched_core_put(void)
  346{
  347	static DECLARE_WORK(_work, __sched_core_put);
  348
  349	/*
  350	 * "There can be only one"
  351	 *
  352	 * Either this is the last one, or we don't actually need to do any
  353	 * 'work'. If it is the last *again*, we rely on
  354	 * WORK_STRUCT_PENDING_BIT.
  355	 */
  356	if (!atomic_add_unless(&sched_core_count, -1, 1))
  357		schedule_work(&_work);
  358}
  359
  360#else /* !CONFIG_SCHED_CORE */
  361
  362static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
  363static inline void sched_core_dequeue(struct rq *rq, struct task_struct *p) { }
 
  364
  365#endif /* CONFIG_SCHED_CORE */
  366
  367/*
  368 * part of the period that we allow rt tasks to run in us.
  369 * default: 0.95s
  370 */
  371int sysctl_sched_rt_runtime = 950000;
  372
  373
  374/*
  375 * Serialization rules:
  376 *
  377 * Lock order:
  378 *
  379 *   p->pi_lock
  380 *     rq->lock
  381 *       hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
  382 *
  383 *  rq1->lock
  384 *    rq2->lock  where: rq1 < rq2
  385 *
  386 * Regular state:
  387 *
  388 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
  389 * local CPU's rq->lock, it optionally removes the task from the runqueue and
  390 * always looks at the local rq data structures to find the most eligible task
  391 * to run next.
  392 *
  393 * Task enqueue is also under rq->lock, possibly taken from another CPU.
  394 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
  395 * the local CPU to avoid bouncing the runqueue state around [ see
  396 * ttwu_queue_wakelist() ]
  397 *
  398 * Task wakeup, specifically wakeups that involve migration, are horribly
  399 * complicated to avoid having to take two rq->locks.
  400 *
  401 * Special state:
  402 *
  403 * System-calls and anything external will use task_rq_lock() which acquires
  404 * both p->pi_lock and rq->lock. As a consequence the state they change is
  405 * stable while holding either lock:
  406 *
  407 *  - sched_setaffinity()/
  408 *    set_cpus_allowed_ptr():	p->cpus_ptr, p->nr_cpus_allowed
  409 *  - set_user_nice():		p->se.load, p->*prio
  410 *  - __sched_setscheduler():	p->sched_class, p->policy, p->*prio,
  411 *				p->se.load, p->rt_priority,
  412 *				p->dl.dl_{runtime, deadline, period, flags, bw, density}
  413 *  - sched_setnuma():		p->numa_preferred_nid
  414 *  - sched_move_task()/
  415 *    cpu_cgroup_fork():	p->sched_task_group
  416 *  - uclamp_update_active()	p->uclamp*
  417 *
  418 * p->state <- TASK_*:
  419 *
  420 *   is changed locklessly using set_current_state(), __set_current_state() or
  421 *   set_special_state(), see their respective comments, or by
  422 *   try_to_wake_up(). This latter uses p->pi_lock to serialize against
  423 *   concurrent self.
  424 *
  425 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
  426 *
  427 *   is set by activate_task() and cleared by deactivate_task(), under
  428 *   rq->lock. Non-zero indicates the task is runnable, the special
  429 *   ON_RQ_MIGRATING state is used for migration without holding both
  430 *   rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
  431 *
 
 
 
 
 
  432 * p->on_cpu <- { 0, 1 }:
  433 *
  434 *   is set by prepare_task() and cleared by finish_task() such that it will be
  435 *   set before p is scheduled-in and cleared after p is scheduled-out, both
  436 *   under rq->lock. Non-zero indicates the task is running on its CPU.
  437 *
  438 *   [ The astute reader will observe that it is possible for two tasks on one
  439 *     CPU to have ->on_cpu = 1 at the same time. ]
  440 *
  441 * task_cpu(p): is changed by set_task_cpu(), the rules are:
  442 *
  443 *  - Don't call set_task_cpu() on a blocked task:
  444 *
  445 *    We don't care what CPU we're not running on, this simplifies hotplug,
  446 *    the CPU assignment of blocked tasks isn't required to be valid.
  447 *
  448 *  - for try_to_wake_up(), called under p->pi_lock:
  449 *
  450 *    This allows try_to_wake_up() to only take one rq->lock, see its comment.
  451 *
  452 *  - for migration called under rq->lock:
  453 *    [ see task_on_rq_migrating() in task_rq_lock() ]
  454 *
  455 *    o move_queued_task()
  456 *    o detach_task()
  457 *
  458 *  - for migration called under double_rq_lock():
  459 *
  460 *    o __migrate_swap_task()
  461 *    o push_rt_task() / pull_rt_task()
  462 *    o push_dl_task() / pull_dl_task()
  463 *    o dl_task_offline_migration()
  464 *
  465 */
  466
  467void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
  468{
  469	raw_spinlock_t *lock;
  470
  471	/* Matches synchronize_rcu() in __sched_core_enable() */
  472	preempt_disable();
  473	if (sched_core_disabled()) {
  474		raw_spin_lock_nested(&rq->__lock, subclass);
  475		/* preempt_count *MUST* be > 1 */
  476		preempt_enable_no_resched();
  477		return;
  478	}
  479
  480	for (;;) {
  481		lock = __rq_lockp(rq);
  482		raw_spin_lock_nested(lock, subclass);
  483		if (likely(lock == __rq_lockp(rq))) {
  484			/* preempt_count *MUST* be > 1 */
  485			preempt_enable_no_resched();
  486			return;
  487		}
  488		raw_spin_unlock(lock);
  489	}
  490}
  491
  492bool raw_spin_rq_trylock(struct rq *rq)
  493{
  494	raw_spinlock_t *lock;
  495	bool ret;
  496
  497	/* Matches synchronize_rcu() in __sched_core_enable() */
  498	preempt_disable();
  499	if (sched_core_disabled()) {
  500		ret = raw_spin_trylock(&rq->__lock);
  501		preempt_enable();
  502		return ret;
  503	}
  504
  505	for (;;) {
  506		lock = __rq_lockp(rq);
  507		ret = raw_spin_trylock(lock);
  508		if (!ret || (likely(lock == __rq_lockp(rq)))) {
  509			preempt_enable();
  510			return ret;
  511		}
  512		raw_spin_unlock(lock);
  513	}
  514}
  515
  516void raw_spin_rq_unlock(struct rq *rq)
  517{
  518	raw_spin_unlock(rq_lockp(rq));
  519}
  520
  521#ifdef CONFIG_SMP
  522/*
  523 * double_rq_lock - safely lock two runqueues
  524 */
  525void double_rq_lock(struct rq *rq1, struct rq *rq2)
  526{
  527	lockdep_assert_irqs_disabled();
  528
  529	if (rq_order_less(rq2, rq1))
  530		swap(rq1, rq2);
  531
  532	raw_spin_rq_lock(rq1);
  533	if (__rq_lockp(rq1) == __rq_lockp(rq2))
  534		return;
  535
  536	raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
  537}
  538#endif
  539
  540/*
  541 * __task_rq_lock - lock the rq @p resides on.
  542 */
  543struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
  544	__acquires(rq->lock)
  545{
  546	struct rq *rq;
  547
  548	lockdep_assert_held(&p->pi_lock);
  549
  550	for (;;) {
  551		rq = task_rq(p);
  552		raw_spin_rq_lock(rq);
  553		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
  554			rq_pin_lock(rq, rf);
  555			return rq;
  556		}
  557		raw_spin_rq_unlock(rq);
  558
  559		while (unlikely(task_on_rq_migrating(p)))
  560			cpu_relax();
  561	}
  562}
  563
  564/*
  565 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
  566 */
  567struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
  568	__acquires(p->pi_lock)
  569	__acquires(rq->lock)
  570{
  571	struct rq *rq;
  572
  573	for (;;) {
  574		raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
  575		rq = task_rq(p);
  576		raw_spin_rq_lock(rq);
  577		/*
  578		 *	move_queued_task()		task_rq_lock()
  579		 *
  580		 *	ACQUIRE (rq->lock)
  581		 *	[S] ->on_rq = MIGRATING		[L] rq = task_rq()
  582		 *	WMB (__set_task_cpu())		ACQUIRE (rq->lock);
  583		 *	[S] ->cpu = new_cpu		[L] task_rq()
  584		 *					[L] ->on_rq
  585		 *	RELEASE (rq->lock)
  586		 *
  587		 * If we observe the old CPU in task_rq_lock(), the acquire of
  588		 * the old rq->lock will fully serialize against the stores.
  589		 *
  590		 * If we observe the new CPU in task_rq_lock(), the address
  591		 * dependency headed by '[L] rq = task_rq()' and the acquire
  592		 * will pair with the WMB to ensure we then also see migrating.
  593		 */
  594		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
  595			rq_pin_lock(rq, rf);
  596			return rq;
  597		}
  598		raw_spin_rq_unlock(rq);
  599		raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
  600
  601		while (unlikely(task_on_rq_migrating(p)))
  602			cpu_relax();
  603	}
  604}
  605
  606/*
  607 * RQ-clock updating methods:
  608 */
  609
  610static void update_rq_clock_task(struct rq *rq, s64 delta)
  611{
  612/*
  613 * In theory, the compile should just see 0 here, and optimize out the call
  614 * to sched_rt_avg_update. But I don't trust it...
  615 */
  616	s64 __maybe_unused steal = 0, irq_delta = 0;
  617
  618#ifdef CONFIG_IRQ_TIME_ACCOUNTING
  619	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
  620
  621	/*
  622	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
  623	 * this case when a previous update_rq_clock() happened inside a
  624	 * {soft,}irq region.
  625	 *
  626	 * When this happens, we stop ->clock_task and only update the
  627	 * prev_irq_time stamp to account for the part that fit, so that a next
  628	 * update will consume the rest. This ensures ->clock_task is
  629	 * monotonic.
  630	 *
  631	 * It does however cause some slight miss-attribution of {soft,}irq
  632	 * time, a more accurate solution would be to update the irq_time using
  633	 * the current rq->clock timestamp, except that would require using
  634	 * atomic ops.
  635	 */
  636	if (irq_delta > delta)
  637		irq_delta = delta;
  638
  639	rq->prev_irq_time += irq_delta;
  640	delta -= irq_delta;
 
  641#endif
  642#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
  643	if (static_key_false((&paravirt_steal_rq_enabled))) {
  644		steal = paravirt_steal_clock(cpu_of(rq));
 
 
  645		steal -= rq->prev_steal_time_rq;
  646
  647		if (unlikely(steal > delta))
  648			steal = delta;
  649
  650		rq->prev_steal_time_rq += steal;
  651		delta -= steal;
  652	}
  653#endif
  654
  655	rq->clock_task += delta;
  656
  657#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
  658	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
  659		update_irq_load_avg(rq, irq_delta + steal);
  660#endif
  661	update_rq_clock_pelt(rq, delta);
  662}
  663
  664void update_rq_clock(struct rq *rq)
  665{
  666	s64 delta;
  667
  668	lockdep_assert_rq_held(rq);
  669
  670	if (rq->clock_update_flags & RQCF_ACT_SKIP)
  671		return;
  672
  673#ifdef CONFIG_SCHED_DEBUG
  674	if (sched_feat(WARN_DOUBLE_CLOCK))
  675		SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
  676	rq->clock_update_flags |= RQCF_UPDATED;
  677#endif
  678
  679	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
  680	if (delta < 0)
  681		return;
  682	rq->clock += delta;
  683	update_rq_clock_task(rq, delta);
  684}
  685
  686#ifdef CONFIG_SCHED_HRTICK
  687/*
  688 * Use HR-timers to deliver accurate preemption points.
  689 */
  690
  691static void hrtick_clear(struct rq *rq)
  692{
  693	if (hrtimer_active(&rq->hrtick_timer))
  694		hrtimer_cancel(&rq->hrtick_timer);
  695}
  696
  697/*
  698 * High-resolution timer tick.
  699 * Runs from hardirq context with interrupts disabled.
  700 */
  701static enum hrtimer_restart hrtick(struct hrtimer *timer)
  702{
  703	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
  704	struct rq_flags rf;
  705
  706	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
  707
  708	rq_lock(rq, &rf);
  709	update_rq_clock(rq);
  710	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
  711	rq_unlock(rq, &rf);
  712
  713	return HRTIMER_NORESTART;
  714}
  715
  716#ifdef CONFIG_SMP
  717
  718static void __hrtick_restart(struct rq *rq)
  719{
  720	struct hrtimer *timer = &rq->hrtick_timer;
  721	ktime_t time = rq->hrtick_time;
  722
  723	hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
  724}
  725
  726/*
  727 * called from hardirq (IPI) context
  728 */
  729static void __hrtick_start(void *arg)
  730{
  731	struct rq *rq = arg;
  732	struct rq_flags rf;
  733
  734	rq_lock(rq, &rf);
  735	__hrtick_restart(rq);
  736	rq_unlock(rq, &rf);
  737}
  738
  739/*
  740 * Called to set the hrtick timer state.
  741 *
  742 * called with rq->lock held and irqs disabled
  743 */
  744void hrtick_start(struct rq *rq, u64 delay)
  745{
  746	struct hrtimer *timer = &rq->hrtick_timer;
  747	s64 delta;
  748
  749	/*
  750	 * Don't schedule slices shorter than 10000ns, that just
  751	 * doesn't make sense and can cause timer DoS.
  752	 */
  753	delta = max_t(s64, delay, 10000LL);
  754	rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
  755
  756	if (rq == this_rq())
  757		__hrtick_restart(rq);
  758	else
  759		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
  760}
  761
  762#else
  763/*
  764 * Called to set the hrtick timer state.
  765 *
  766 * called with rq->lock held and irqs disabled
  767 */
  768void hrtick_start(struct rq *rq, u64 delay)
  769{
  770	/*
  771	 * Don't schedule slices shorter than 10000ns, that just
  772	 * doesn't make sense. Rely on vruntime for fairness.
  773	 */
  774	delay = max_t(u64, delay, 10000LL);
  775	hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
  776		      HRTIMER_MODE_REL_PINNED_HARD);
  777}
  778
  779#endif /* CONFIG_SMP */
  780
  781static void hrtick_rq_init(struct rq *rq)
  782{
  783#ifdef CONFIG_SMP
  784	INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
  785#endif
  786	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
  787	rq->hrtick_timer.function = hrtick;
  788}
  789#else	/* CONFIG_SCHED_HRTICK */
  790static inline void hrtick_clear(struct rq *rq)
  791{
  792}
  793
  794static inline void hrtick_rq_init(struct rq *rq)
  795{
  796}
  797#endif	/* CONFIG_SCHED_HRTICK */
  798
  799/*
  800 * cmpxchg based fetch_or, macro so it works for different integer types
  801 */
  802#define fetch_or(ptr, mask)						\
  803	({								\
  804		typeof(ptr) _ptr = (ptr);				\
  805		typeof(mask) _mask = (mask);				\
  806		typeof(*_ptr) _old, _val = *_ptr;			\
  807									\
  808		for (;;) {						\
  809			_old = cmpxchg(_ptr, _val, _val | _mask);	\
  810			if (_old == _val)				\
  811				break;					\
  812			_val = _old;					\
  813		}							\
  814	_old;								\
  815})
  816
  817#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
  818/*
  819 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
  820 * this avoids any races wrt polling state changes and thereby avoids
  821 * spurious IPIs.
  822 */
  823static bool set_nr_and_not_polling(struct task_struct *p)
  824{
  825	struct thread_info *ti = task_thread_info(p);
  826	return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
  827}
  828
  829/*
  830 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
  831 *
  832 * If this returns true, then the idle task promises to call
  833 * sched_ttwu_pending() and reschedule soon.
  834 */
  835static bool set_nr_if_polling(struct task_struct *p)
  836{
  837	struct thread_info *ti = task_thread_info(p);
  838	typeof(ti->flags) old, val = READ_ONCE(ti->flags);
  839
  840	for (;;) {
  841		if (!(val & _TIF_POLLING_NRFLAG))
  842			return false;
  843		if (val & _TIF_NEED_RESCHED)
  844			return true;
  845		old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
  846		if (old == val)
  847			break;
  848		val = old;
  849	}
  850	return true;
  851}
  852
  853#else
  854static bool set_nr_and_not_polling(struct task_struct *p)
  855{
  856	set_tsk_need_resched(p);
  857	return true;
  858}
  859
  860#ifdef CONFIG_SMP
  861static bool set_nr_if_polling(struct task_struct *p)
  862{
  863	return false;
  864}
  865#endif
  866#endif
  867
  868static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
  869{
  870	struct wake_q_node *node = &task->wake_q;
  871
  872	/*
  873	 * Atomically grab the task, if ->wake_q is !nil already it means
  874	 * it's already queued (either by us or someone else) and will get the
  875	 * wakeup due to that.
  876	 *
  877	 * In order to ensure that a pending wakeup will observe our pending
  878	 * state, even in the failed case, an explicit smp_mb() must be used.
  879	 */
  880	smp_mb__before_atomic();
  881	if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
  882		return false;
  883
  884	/*
  885	 * The head is context local, there can be no concurrency.
  886	 */
  887	*head->lastp = node;
  888	head->lastp = &node->next;
  889	return true;
  890}
  891
  892/**
  893 * wake_q_add() - queue a wakeup for 'later' waking.
  894 * @head: the wake_q_head to add @task to
  895 * @task: the task to queue for 'later' wakeup
  896 *
  897 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
  898 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
  899 * instantly.
  900 *
  901 * This function must be used as-if it were wake_up_process(); IOW the task
  902 * must be ready to be woken at this location.
  903 */
  904void wake_q_add(struct wake_q_head *head, struct task_struct *task)
  905{
  906	if (__wake_q_add(head, task))
  907		get_task_struct(task);
  908}
  909
  910/**
  911 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
  912 * @head: the wake_q_head to add @task to
  913 * @task: the task to queue for 'later' wakeup
  914 *
  915 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
  916 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
  917 * instantly.
  918 *
  919 * This function must be used as-if it were wake_up_process(); IOW the task
  920 * must be ready to be woken at this location.
  921 *
  922 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
  923 * that already hold reference to @task can call the 'safe' version and trust
  924 * wake_q to do the right thing depending whether or not the @task is already
  925 * queued for wakeup.
  926 */
  927void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
  928{
  929	if (!__wake_q_add(head, task))
  930		put_task_struct(task);
  931}
  932
  933void wake_up_q(struct wake_q_head *head)
  934{
  935	struct wake_q_node *node = head->first;
  936
  937	while (node != WAKE_Q_TAIL) {
  938		struct task_struct *task;
  939
  940		task = container_of(node, struct task_struct, wake_q);
  941		/* Task can safely be re-inserted now: */
  942		node = node->next;
  943		task->wake_q.next = NULL;
  944
  945		/*
  946		 * wake_up_process() executes a full barrier, which pairs with
  947		 * the queueing in wake_q_add() so as not to miss wakeups.
  948		 */
  949		wake_up_process(task);
  950		put_task_struct(task);
  951	}
  952}
  953
  954/*
  955 * resched_curr - mark rq's current task 'to be rescheduled now'.
  956 *
  957 * On UP this means the setting of the need_resched flag, on SMP it
  958 * might also involve a cross-CPU call to trigger the scheduler on
  959 * the target CPU.
  960 */
  961void resched_curr(struct rq *rq)
  962{
  963	struct task_struct *curr = rq->curr;
 
  964	int cpu;
  965
  966	lockdep_assert_rq_held(rq);
  967
  968	if (test_tsk_need_resched(curr))
 
 
 
 
 
 
 
  969		return;
  970
  971	cpu = cpu_of(rq);
  972
  973	if (cpu == smp_processor_id()) {
  974		set_tsk_need_resched(curr);
  975		set_preempt_need_resched();
 
  976		return;
  977	}
  978
  979	if (set_nr_and_not_polling(curr))
  980		smp_send_reschedule(cpu);
  981	else
 
  982		trace_sched_wake_idle_without_ipi(cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  983}
  984
  985void resched_cpu(int cpu)
  986{
  987	struct rq *rq = cpu_rq(cpu);
  988	unsigned long flags;
  989
  990	raw_spin_rq_lock_irqsave(rq, flags);
  991	if (cpu_online(cpu) || cpu == smp_processor_id())
  992		resched_curr(rq);
  993	raw_spin_rq_unlock_irqrestore(rq, flags);
  994}
  995
  996#ifdef CONFIG_SMP
  997#ifdef CONFIG_NO_HZ_COMMON
  998/*
  999 * In the semi idle case, use the nearest busy CPU for migrating timers
 1000 * from an idle CPU.  This is good for power-savings.
 1001 *
 1002 * We don't do similar optimization for completely idle system, as
 1003 * selecting an idle CPU will add more delays to the timers than intended
 1004 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
 1005 */
 1006int get_nohz_timer_target(void)
 1007{
 1008	int i, cpu = smp_processor_id(), default_cpu = -1;
 1009	struct sched_domain *sd;
 
 1010
 1011	if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
 1012		if (!idle_cpu(cpu))
 1013			return cpu;
 1014		default_cpu = cpu;
 1015	}
 1016
 1017	rcu_read_lock();
 
 
 
 1018	for_each_domain(cpu, sd) {
 1019		for_each_cpu_and(i, sched_domain_span(sd),
 1020			housekeeping_cpumask(HK_FLAG_TIMER)) {
 1021			if (cpu == i)
 1022				continue;
 1023
 1024			if (!idle_cpu(i)) {
 1025				cpu = i;
 1026				goto unlock;
 1027			}
 1028		}
 1029	}
 1030
 1031	if (default_cpu == -1)
 1032		default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
 1033	cpu = default_cpu;
 1034unlock:
 1035	rcu_read_unlock();
 1036	return cpu;
 1037}
 1038
 1039/*
 1040 * When add_timer_on() enqueues a timer into the timer wheel of an
 1041 * idle CPU then this timer might expire before the next timer event
 1042 * which is scheduled to wake up that CPU. In case of a completely
 1043 * idle system the next event might even be infinite time into the
 1044 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
 1045 * leaves the inner idle loop so the newly added timer is taken into
 1046 * account when the CPU goes back to idle and evaluates the timer
 1047 * wheel for the next timer event.
 1048 */
 1049static void wake_up_idle_cpu(int cpu)
 1050{
 1051	struct rq *rq = cpu_rq(cpu);
 1052
 1053	if (cpu == smp_processor_id())
 1054		return;
 1055
 1056	if (set_nr_and_not_polling(rq->idle))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 1057		smp_send_reschedule(cpu);
 1058	else
 1059		trace_sched_wake_idle_without_ipi(cpu);
 1060}
 1061
 1062static bool wake_up_full_nohz_cpu(int cpu)
 1063{
 1064	/*
 1065	 * We just need the target to call irq_exit() and re-evaluate
 1066	 * the next tick. The nohz full kick at least implies that.
 1067	 * If needed we can still optimize that later with an
 1068	 * empty IRQ.
 1069	 */
 1070	if (cpu_is_offline(cpu))
 1071		return true;  /* Don't try to wake offline CPUs. */
 1072	if (tick_nohz_full_cpu(cpu)) {
 1073		if (cpu != smp_processor_id() ||
 1074		    tick_nohz_tick_stopped())
 1075			tick_nohz_full_kick_cpu(cpu);
 1076		return true;
 1077	}
 1078
 1079	return false;
 1080}
 1081
 1082/*
 1083 * Wake up the specified CPU.  If the CPU is going offline, it is the
 1084 * caller's responsibility to deal with the lost wakeup, for example,
 1085 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
 1086 */
 1087void wake_up_nohz_cpu(int cpu)
 1088{
 1089	if (!wake_up_full_nohz_cpu(cpu))
 1090		wake_up_idle_cpu(cpu);
 1091}
 1092
 1093static void nohz_csd_func(void *info)
 1094{
 1095	struct rq *rq = info;
 1096	int cpu = cpu_of(rq);
 1097	unsigned int flags;
 1098
 1099	/*
 1100	 * Release the rq::nohz_csd.
 1101	 */
 1102	flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
 1103	WARN_ON(!(flags & NOHZ_KICK_MASK));
 1104
 1105	rq->idle_balance = idle_cpu(cpu);
 1106	if (rq->idle_balance && !need_resched()) {
 1107		rq->nohz_idle_balance = flags;
 1108		raise_softirq_irqoff(SCHED_SOFTIRQ);
 1109	}
 1110}
 1111
 1112#endif /* CONFIG_NO_HZ_COMMON */
 1113
 1114#ifdef CONFIG_NO_HZ_FULL
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 1115bool sched_can_stop_tick(struct rq *rq)
 1116{
 1117	int fifo_nr_running;
 1118
 1119	/* Deadline tasks, even if single, need the tick */
 1120	if (rq->dl.dl_nr_running)
 1121		return false;
 1122
 1123	/*
 1124	 * If there are more than one RR tasks, we need the tick to affect the
 1125	 * actual RR behaviour.
 1126	 */
 1127	if (rq->rt.rr_nr_running) {
 1128		if (rq->rt.rr_nr_running == 1)
 1129			return true;
 1130		else
 1131			return false;
 1132	}
 1133
 1134	/*
 1135	 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
 1136	 * forced preemption between FIFO tasks.
 1137	 */
 1138	fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
 1139	if (fifo_nr_running)
 1140		return true;
 1141
 1142	/*
 1143	 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
 1144	 * if there's more than one we need the tick for involuntary
 1145	 * preemption.
 1146	 */
 1147	if (rq->nr_running > 1)
 
 
 
 1148		return false;
 1149
 
 
 
 
 
 
 
 
 
 
 
 
 1150	return true;
 1151}
 1152#endif /* CONFIG_NO_HZ_FULL */
 1153#endif /* CONFIG_SMP */
 1154
 1155#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
 1156			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
 1157/*
 1158 * Iterate task_group tree rooted at *from, calling @down when first entering a
 1159 * node and @up when leaving it for the final time.
 1160 *
 1161 * Caller must hold rcu_lock or sufficient equivalent.
 1162 */
 1163int walk_tg_tree_from(struct task_group *from,
 1164			     tg_visitor down, tg_visitor up, void *data)
 1165{
 1166	struct task_group *parent, *child;
 1167	int ret;
 1168
 1169	parent = from;
 1170
 1171down:
 1172	ret = (*down)(parent, data);
 1173	if (ret)
 1174		goto out;
 1175	list_for_each_entry_rcu(child, &parent->children, siblings) {
 1176		parent = child;
 1177		goto down;
 1178
 1179up:
 1180		continue;
 1181	}
 1182	ret = (*up)(parent, data);
 1183	if (ret || parent == from)
 1184		goto out;
 1185
 1186	child = parent;
 1187	parent = parent->parent;
 1188	if (parent)
 1189		goto up;
 1190out:
 1191	return ret;
 1192}
 1193
 1194int tg_nop(struct task_group *tg, void *data)
 1195{
 1196	return 0;
 1197}
 1198#endif
 1199
 1200static void set_load_weight(struct task_struct *p, bool update_load)
 1201{
 1202	int prio = p->static_prio - MAX_RT_PRIO;
 1203	struct load_weight *load = &p->se.load;
 1204
 1205	/*
 1206	 * SCHED_IDLE tasks get minimal weight:
 1207	 */
 1208	if (task_has_idle_policy(p)) {
 1209		load->weight = scale_load(WEIGHT_IDLEPRIO);
 1210		load->inv_weight = WMULT_IDLEPRIO;
 1211		return;
 
 
 1212	}
 1213
 1214	/*
 1215	 * SCHED_OTHER tasks have to update their load when changing their
 1216	 * weight
 1217	 */
 1218	if (update_load && p->sched_class == &fair_sched_class) {
 1219		reweight_task(p, prio);
 1220	} else {
 1221		load->weight = scale_load(sched_prio_to_weight[prio]);
 1222		load->inv_weight = sched_prio_to_wmult[prio];
 1223	}
 1224}
 1225
 1226#ifdef CONFIG_UCLAMP_TASK
 1227/*
 1228 * Serializes updates of utilization clamp values
 1229 *
 1230 * The (slow-path) user-space triggers utilization clamp value updates which
 1231 * can require updates on (fast-path) scheduler's data structures used to
 1232 * support enqueue/dequeue operations.
 1233 * While the per-CPU rq lock protects fast-path update operations, user-space
 1234 * requests are serialized using a mutex to reduce the risk of conflicting
 1235 * updates or API abuses.
 1236 */
 1237static DEFINE_MUTEX(uclamp_mutex);
 1238
 1239/* Max allowed minimum utilization */
 1240unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
 1241
 1242/* Max allowed maximum utilization */
 1243unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
 1244
 1245/*
 1246 * By default RT tasks run at the maximum performance point/capacity of the
 1247 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
 1248 * SCHED_CAPACITY_SCALE.
 1249 *
 1250 * This knob allows admins to change the default behavior when uclamp is being
 1251 * used. In battery powered devices, particularly, running at the maximum
 1252 * capacity and frequency will increase energy consumption and shorten the
 1253 * battery life.
 1254 *
 1255 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
 1256 *
 1257 * This knob will not override the system default sched_util_clamp_min defined
 1258 * above.
 1259 */
 1260unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
 1261
 1262/* All clamps are required to be less or equal than these values */
 1263static struct uclamp_se uclamp_default[UCLAMP_CNT];
 1264
 1265/*
 1266 * This static key is used to reduce the uclamp overhead in the fast path. It
 1267 * primarily disables the call to uclamp_rq_{inc, dec}() in
 1268 * enqueue/dequeue_task().
 1269 *
 1270 * This allows users to continue to enable uclamp in their kernel config with
 1271 * minimum uclamp overhead in the fast path.
 1272 *
 1273 * As soon as userspace modifies any of the uclamp knobs, the static key is
 1274 * enabled, since we have an actual users that make use of uclamp
 1275 * functionality.
 1276 *
 1277 * The knobs that would enable this static key are:
 1278 *
 1279 *   * A task modifying its uclamp value with sched_setattr().
 1280 *   * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
 1281 *   * An admin modifying the cgroup cpu.uclamp.{min, max}
 1282 */
 1283DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
 1284
 1285/* Integer rounded range for each bucket */
 1286#define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
 1287
 1288#define for_each_clamp_id(clamp_id) \
 1289	for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
 1290
 1291static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
 1292{
 1293	return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
 1294}
 1295
 1296static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
 1297{
 1298	if (clamp_id == UCLAMP_MIN)
 1299		return 0;
 1300	return SCHED_CAPACITY_SCALE;
 1301}
 1302
 1303static inline void uclamp_se_set(struct uclamp_se *uc_se,
 1304				 unsigned int value, bool user_defined)
 1305{
 1306	uc_se->value = value;
 1307	uc_se->bucket_id = uclamp_bucket_id(value);
 1308	uc_se->user_defined = user_defined;
 1309}
 1310
 1311static inline unsigned int
 1312uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
 1313		  unsigned int clamp_value)
 1314{
 1315	/*
 1316	 * Avoid blocked utilization pushing up the frequency when we go
 1317	 * idle (which drops the max-clamp) by retaining the last known
 1318	 * max-clamp.
 1319	 */
 1320	if (clamp_id == UCLAMP_MAX) {
 1321		rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
 1322		return clamp_value;
 1323	}
 1324
 1325	return uclamp_none(UCLAMP_MIN);
 1326}
 1327
 1328static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
 1329				     unsigned int clamp_value)
 1330{
 1331	/* Reset max-clamp retention only on idle exit */
 1332	if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
 1333		return;
 1334
 1335	WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
 1336}
 1337
 1338static inline
 1339unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
 1340				   unsigned int clamp_value)
 1341{
 1342	struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
 1343	int bucket_id = UCLAMP_BUCKETS - 1;
 1344
 1345	/*
 1346	 * Since both min and max clamps are max aggregated, find the
 1347	 * top most bucket with tasks in.
 1348	 */
 1349	for ( ; bucket_id >= 0; bucket_id--) {
 1350		if (!bucket[bucket_id].tasks)
 1351			continue;
 1352		return bucket[bucket_id].value;
 1353	}
 1354
 1355	/* No tasks -- default clamp values */
 1356	return uclamp_idle_value(rq, clamp_id, clamp_value);
 1357}
 1358
 1359static void __uclamp_update_util_min_rt_default(struct task_struct *p)
 1360{
 1361	unsigned int default_util_min;
 1362	struct uclamp_se *uc_se;
 1363
 1364	lockdep_assert_held(&p->pi_lock);
 1365
 1366	uc_se = &p->uclamp_req[UCLAMP_MIN];
 1367
 1368	/* Only sync if user didn't override the default */
 1369	if (uc_se->user_defined)
 1370		return;
 1371
 1372	default_util_min = sysctl_sched_uclamp_util_min_rt_default;
 1373	uclamp_se_set(uc_se, default_util_min, false);
 1374}
 1375
 1376static void uclamp_update_util_min_rt_default(struct task_struct *p)
 1377{
 1378	struct rq_flags rf;
 1379	struct rq *rq;
 1380
 1381	if (!rt_task(p))
 1382		return;
 1383
 1384	/* Protect updates to p->uclamp_* */
 1385	rq = task_rq_lock(p, &rf);
 1386	__uclamp_update_util_min_rt_default(p);
 1387	task_rq_unlock(rq, p, &rf);
 1388}
 1389
 1390static void uclamp_sync_util_min_rt_default(void)
 1391{
 1392	struct task_struct *g, *p;
 1393
 1394	/*
 1395	 * copy_process()			sysctl_uclamp
 1396	 *					  uclamp_min_rt = X;
 1397	 *   write_lock(&tasklist_lock)		  read_lock(&tasklist_lock)
 1398	 *   // link thread			  smp_mb__after_spinlock()
 1399	 *   write_unlock(&tasklist_lock)	  read_unlock(&tasklist_lock);
 1400	 *   sched_post_fork()			  for_each_process_thread()
 1401	 *     __uclamp_sync_rt()		    __uclamp_sync_rt()
 1402	 *
 1403	 * Ensures that either sched_post_fork() will observe the new
 1404	 * uclamp_min_rt or for_each_process_thread() will observe the new
 1405	 * task.
 1406	 */
 1407	read_lock(&tasklist_lock);
 1408	smp_mb__after_spinlock();
 1409	read_unlock(&tasklist_lock);
 1410
 1411	rcu_read_lock();
 1412	for_each_process_thread(g, p)
 1413		uclamp_update_util_min_rt_default(p);
 1414	rcu_read_unlock();
 1415}
 1416
 1417static inline struct uclamp_se
 1418uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
 1419{
 1420	/* Copy by value as we could modify it */
 1421	struct uclamp_se uc_req = p->uclamp_req[clamp_id];
 1422#ifdef CONFIG_UCLAMP_TASK_GROUP
 1423	unsigned int tg_min, tg_max, value;
 1424
 1425	/*
 1426	 * Tasks in autogroups or root task group will be
 1427	 * restricted by system defaults.
 1428	 */
 1429	if (task_group_is_autogroup(task_group(p)))
 1430		return uc_req;
 1431	if (task_group(p) == &root_task_group)
 1432		return uc_req;
 1433
 1434	tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
 1435	tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
 1436	value = uc_req.value;
 1437	value = clamp(value, tg_min, tg_max);
 1438	uclamp_se_set(&uc_req, value, false);
 1439#endif
 1440
 1441	return uc_req;
 1442}
 1443
 1444/*
 1445 * The effective clamp bucket index of a task depends on, by increasing
 1446 * priority:
 1447 * - the task specific clamp value, when explicitly requested from userspace
 1448 * - the task group effective clamp value, for tasks not either in the root
 1449 *   group or in an autogroup
 1450 * - the system default clamp value, defined by the sysadmin
 1451 */
 1452static inline struct uclamp_se
 1453uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
 1454{
 1455	struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
 1456	struct uclamp_se uc_max = uclamp_default[clamp_id];
 1457
 1458	/* System default restrictions always apply */
 1459	if (unlikely(uc_req.value > uc_max.value))
 1460		return uc_max;
 1461
 1462	return uc_req;
 1463}
 1464
 1465unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
 1466{
 1467	struct uclamp_se uc_eff;
 1468
 1469	/* Task currently refcounted: use back-annotated (effective) value */
 1470	if (p->uclamp[clamp_id].active)
 1471		return (unsigned long)p->uclamp[clamp_id].value;
 1472
 1473	uc_eff = uclamp_eff_get(p, clamp_id);
 1474
 1475	return (unsigned long)uc_eff.value;
 1476}
 1477
 1478/*
 1479 * When a task is enqueued on a rq, the clamp bucket currently defined by the
 1480 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
 1481 * updates the rq's clamp value if required.
 1482 *
 1483 * Tasks can have a task-specific value requested from user-space, track
 1484 * within each bucket the maximum value for tasks refcounted in it.
 1485 * This "local max aggregation" allows to track the exact "requested" value
 1486 * for each bucket when all its RUNNABLE tasks require the same clamp.
 1487 */
 1488static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
 1489				    enum uclamp_id clamp_id)
 1490{
 1491	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
 1492	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
 1493	struct uclamp_bucket *bucket;
 1494
 1495	lockdep_assert_rq_held(rq);
 1496
 1497	/* Update task effective clamp */
 1498	p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
 1499
 1500	bucket = &uc_rq->bucket[uc_se->bucket_id];
 1501	bucket->tasks++;
 1502	uc_se->active = true;
 1503
 1504	uclamp_idle_reset(rq, clamp_id, uc_se->value);
 1505
 1506	/*
 1507	 * Local max aggregation: rq buckets always track the max
 1508	 * "requested" clamp value of its RUNNABLE tasks.
 1509	 */
 1510	if (bucket->tasks == 1 || uc_se->value > bucket->value)
 1511		bucket->value = uc_se->value;
 1512
 1513	if (uc_se->value > READ_ONCE(uc_rq->value))
 1514		WRITE_ONCE(uc_rq->value, uc_se->value);
 1515}
 1516
 1517/*
 1518 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
 1519 * is released. If this is the last task reference counting the rq's max
 1520 * active clamp value, then the rq's clamp value is updated.
 1521 *
 1522 * Both refcounted tasks and rq's cached clamp values are expected to be
 1523 * always valid. If it's detected they are not, as defensive programming,
 1524 * enforce the expected state and warn.
 1525 */
 1526static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
 1527				    enum uclamp_id clamp_id)
 1528{
 1529	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
 1530	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
 1531	struct uclamp_bucket *bucket;
 1532	unsigned int bkt_clamp;
 1533	unsigned int rq_clamp;
 1534
 1535	lockdep_assert_rq_held(rq);
 1536
 1537	/*
 1538	 * If sched_uclamp_used was enabled after task @p was enqueued,
 1539	 * we could end up with unbalanced call to uclamp_rq_dec_id().
 1540	 *
 1541	 * In this case the uc_se->active flag should be false since no uclamp
 1542	 * accounting was performed at enqueue time and we can just return
 1543	 * here.
 1544	 *
 1545	 * Need to be careful of the following enqueue/dequeue ordering
 1546	 * problem too
 1547	 *
 1548	 *	enqueue(taskA)
 1549	 *	// sched_uclamp_used gets enabled
 1550	 *	enqueue(taskB)
 1551	 *	dequeue(taskA)
 1552	 *	// Must not decrement bucket->tasks here
 1553	 *	dequeue(taskB)
 1554	 *
 1555	 * where we could end up with stale data in uc_se and
 1556	 * bucket[uc_se->bucket_id].
 1557	 *
 1558	 * The following check here eliminates the possibility of such race.
 1559	 */
 1560	if (unlikely(!uc_se->active))
 1561		return;
 1562
 1563	bucket = &uc_rq->bucket[uc_se->bucket_id];
 1564
 1565	SCHED_WARN_ON(!bucket->tasks);
 1566	if (likely(bucket->tasks))
 1567		bucket->tasks--;
 1568
 1569	uc_se->active = false;
 1570
 1571	/*
 1572	 * Keep "local max aggregation" simple and accept to (possibly)
 1573	 * overboost some RUNNABLE tasks in the same bucket.
 1574	 * The rq clamp bucket value is reset to its base value whenever
 1575	 * there are no more RUNNABLE tasks refcounting it.
 1576	 */
 1577	if (likely(bucket->tasks))
 1578		return;
 1579
 1580	rq_clamp = READ_ONCE(uc_rq->value);
 1581	/*
 1582	 * Defensive programming: this should never happen. If it happens,
 1583	 * e.g. due to future modification, warn and fixup the expected value.
 1584	 */
 1585	SCHED_WARN_ON(bucket->value > rq_clamp);
 1586	if (bucket->value >= rq_clamp) {
 1587		bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
 1588		WRITE_ONCE(uc_rq->value, bkt_clamp);
 1589	}
 1590}
 1591
 1592static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
 1593{
 1594	enum uclamp_id clamp_id;
 1595
 1596	/*
 1597	 * Avoid any overhead until uclamp is actually used by the userspace.
 1598	 *
 1599	 * The condition is constructed such that a NOP is generated when
 1600	 * sched_uclamp_used is disabled.
 1601	 */
 1602	if (!static_branch_unlikely(&sched_uclamp_used))
 1603		return;
 1604
 1605	if (unlikely(!p->sched_class->uclamp_enabled))
 1606		return;
 1607
 
 
 
 1608	for_each_clamp_id(clamp_id)
 1609		uclamp_rq_inc_id(rq, p, clamp_id);
 1610
 1611	/* Reset clamp idle holding when there is one RUNNABLE task */
 1612	if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
 1613		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
 1614}
 1615
 1616static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
 1617{
 1618	enum uclamp_id clamp_id;
 1619
 1620	/*
 1621	 * Avoid any overhead until uclamp is actually used by the userspace.
 1622	 *
 1623	 * The condition is constructed such that a NOP is generated when
 1624	 * sched_uclamp_used is disabled.
 1625	 */
 1626	if (!static_branch_unlikely(&sched_uclamp_used))
 1627		return;
 1628
 1629	if (unlikely(!p->sched_class->uclamp_enabled))
 1630		return;
 1631
 
 
 
 1632	for_each_clamp_id(clamp_id)
 1633		uclamp_rq_dec_id(rq, p, clamp_id);
 1634}
 1635
 1636static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
 1637				      enum uclamp_id clamp_id)
 1638{
 1639	if (!p->uclamp[clamp_id].active)
 1640		return;
 1641
 1642	uclamp_rq_dec_id(rq, p, clamp_id);
 1643	uclamp_rq_inc_id(rq, p, clamp_id);
 1644
 1645	/*
 1646	 * Make sure to clear the idle flag if we've transiently reached 0
 1647	 * active tasks on rq.
 1648	 */
 1649	if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
 1650		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
 1651}
 1652
 1653static inline void
 1654uclamp_update_active(struct task_struct *p)
 1655{
 1656	enum uclamp_id clamp_id;
 1657	struct rq_flags rf;
 1658	struct rq *rq;
 1659
 1660	/*
 1661	 * Lock the task and the rq where the task is (or was) queued.
 1662	 *
 1663	 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
 1664	 * price to pay to safely serialize util_{min,max} updates with
 1665	 * enqueues, dequeues and migration operations.
 1666	 * This is the same locking schema used by __set_cpus_allowed_ptr().
 1667	 */
 1668	rq = task_rq_lock(p, &rf);
 1669
 1670	/*
 1671	 * Setting the clamp bucket is serialized by task_rq_lock().
 1672	 * If the task is not yet RUNNABLE and its task_struct is not
 1673	 * affecting a valid clamp bucket, the next time it's enqueued,
 1674	 * it will already see the updated clamp bucket value.
 1675	 */
 1676	for_each_clamp_id(clamp_id)
 1677		uclamp_rq_reinc_id(rq, p, clamp_id);
 1678
 1679	task_rq_unlock(rq, p, &rf);
 1680}
 1681
 1682#ifdef CONFIG_UCLAMP_TASK_GROUP
 1683static inline void
 1684uclamp_update_active_tasks(struct cgroup_subsys_state *css)
 1685{
 1686	struct css_task_iter it;
 1687	struct task_struct *p;
 1688
 1689	css_task_iter_start(css, 0, &it);
 1690	while ((p = css_task_iter_next(&it)))
 1691		uclamp_update_active(p);
 1692	css_task_iter_end(&it);
 1693}
 1694
 1695static void cpu_util_update_eff(struct cgroup_subsys_state *css);
 
 
 
 
 1696static void uclamp_update_root_tg(void)
 1697{
 1698	struct task_group *tg = &root_task_group;
 1699
 1700	uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
 1701		      sysctl_sched_uclamp_util_min, false);
 1702	uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
 1703		      sysctl_sched_uclamp_util_max, false);
 1704
 1705	rcu_read_lock();
 1706	cpu_util_update_eff(&root_task_group.css);
 1707	rcu_read_unlock();
 1708}
 1709#else
 1710static void uclamp_update_root_tg(void) { }
 1711#endif
 1712
 1713int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 1714				void *buffer, size_t *lenp, loff_t *ppos)
 1715{
 1716	bool update_root_tg = false;
 1717	int old_min, old_max, old_min_rt;
 1718	int result;
 1719
 1720	mutex_lock(&uclamp_mutex);
 
 1721	old_min = sysctl_sched_uclamp_util_min;
 1722	old_max = sysctl_sched_uclamp_util_max;
 1723	old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
 1724
 1725	result = proc_dointvec(table, write, buffer, lenp, ppos);
 1726	if (result)
 1727		goto undo;
 1728	if (!write)
 1729		goto done;
 1730
 1731	if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
 1732	    sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE	||
 1733	    sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
 1734
 1735		result = -EINVAL;
 1736		goto undo;
 1737	}
 1738
 1739	if (old_min != sysctl_sched_uclamp_util_min) {
 1740		uclamp_se_set(&uclamp_default[UCLAMP_MIN],
 1741			      sysctl_sched_uclamp_util_min, false);
 1742		update_root_tg = true;
 1743	}
 1744	if (old_max != sysctl_sched_uclamp_util_max) {
 1745		uclamp_se_set(&uclamp_default[UCLAMP_MAX],
 1746			      sysctl_sched_uclamp_util_max, false);
 1747		update_root_tg = true;
 1748	}
 1749
 1750	if (update_root_tg) {
 1751		static_branch_enable(&sched_uclamp_used);
 1752		uclamp_update_root_tg();
 1753	}
 1754
 1755	if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
 1756		static_branch_enable(&sched_uclamp_used);
 1757		uclamp_sync_util_min_rt_default();
 1758	}
 1759
 1760	/*
 1761	 * We update all RUNNABLE tasks only when task groups are in use.
 1762	 * Otherwise, keep it simple and do just a lazy update at each next
 1763	 * task enqueue time.
 1764	 */
 1765
 1766	goto done;
 1767
 1768undo:
 1769	sysctl_sched_uclamp_util_min = old_min;
 1770	sysctl_sched_uclamp_util_max = old_max;
 1771	sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
 1772done:
 1773	mutex_unlock(&uclamp_mutex);
 1774
 1775	return result;
 1776}
 1777
 1778static int uclamp_validate(struct task_struct *p,
 1779			   const struct sched_attr *attr)
 1780{
 1781	int util_min = p->uclamp_req[UCLAMP_MIN].value;
 1782	int util_max = p->uclamp_req[UCLAMP_MAX].value;
 1783
 1784	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
 1785		util_min = attr->sched_util_min;
 1786
 1787		if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
 1788			return -EINVAL;
 1789	}
 1790
 1791	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
 1792		util_max = attr->sched_util_max;
 1793
 1794		if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
 1795			return -EINVAL;
 1796	}
 1797
 1798	if (util_min != -1 && util_max != -1 && util_min > util_max)
 1799		return -EINVAL;
 1800
 1801	/*
 1802	 * We have valid uclamp attributes; make sure uclamp is enabled.
 1803	 *
 1804	 * We need to do that here, because enabling static branches is a
 1805	 * blocking operation which obviously cannot be done while holding
 1806	 * scheduler locks.
 1807	 */
 1808	static_branch_enable(&sched_uclamp_used);
 1809
 1810	return 0;
 1811}
 1812
 1813static bool uclamp_reset(const struct sched_attr *attr,
 1814			 enum uclamp_id clamp_id,
 1815			 struct uclamp_se *uc_se)
 1816{
 1817	/* Reset on sched class change for a non user-defined clamp value. */
 1818	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
 1819	    !uc_se->user_defined)
 1820		return true;
 1821
 1822	/* Reset on sched_util_{min,max} == -1. */
 1823	if (clamp_id == UCLAMP_MIN &&
 1824	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
 1825	    attr->sched_util_min == -1) {
 1826		return true;
 1827	}
 1828
 1829	if (clamp_id == UCLAMP_MAX &&
 1830	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
 1831	    attr->sched_util_max == -1) {
 1832		return true;
 1833	}
 1834
 1835	return false;
 1836}
 1837
 1838static void __setscheduler_uclamp(struct task_struct *p,
 1839				  const struct sched_attr *attr)
 1840{
 1841	enum uclamp_id clamp_id;
 1842
 1843	for_each_clamp_id(clamp_id) {
 1844		struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
 1845		unsigned int value;
 1846
 1847		if (!uclamp_reset(attr, clamp_id, uc_se))
 1848			continue;
 1849
 1850		/*
 1851		 * RT by default have a 100% boost value that could be modified
 1852		 * at runtime.
 1853		 */
 1854		if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
 1855			value = sysctl_sched_uclamp_util_min_rt_default;
 1856		else
 1857			value = uclamp_none(clamp_id);
 1858
 1859		uclamp_se_set(uc_se, value, false);
 1860
 1861	}
 1862
 1863	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
 1864		return;
 1865
 1866	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
 1867	    attr->sched_util_min != -1) {
 1868		uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
 1869			      attr->sched_util_min, true);
 1870	}
 1871
 1872	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
 1873	    attr->sched_util_max != -1) {
 1874		uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
 1875			      attr->sched_util_max, true);
 1876	}
 1877}
 1878
 1879static void uclamp_fork(struct task_struct *p)
 1880{
 1881	enum uclamp_id clamp_id;
 1882
 1883	/*
 1884	 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
 1885	 * as the task is still at its early fork stages.
 1886	 */
 1887	for_each_clamp_id(clamp_id)
 1888		p->uclamp[clamp_id].active = false;
 1889
 1890	if (likely(!p->sched_reset_on_fork))
 1891		return;
 1892
 1893	for_each_clamp_id(clamp_id) {
 1894		uclamp_se_set(&p->uclamp_req[clamp_id],
 1895			      uclamp_none(clamp_id), false);
 1896	}
 1897}
 1898
 1899static void uclamp_post_fork(struct task_struct *p)
 1900{
 1901	uclamp_update_util_min_rt_default(p);
 1902}
 1903
 1904static void __init init_uclamp_rq(struct rq *rq)
 1905{
 1906	enum uclamp_id clamp_id;
 1907	struct uclamp_rq *uc_rq = rq->uclamp;
 1908
 1909	for_each_clamp_id(clamp_id) {
 1910		uc_rq[clamp_id] = (struct uclamp_rq) {
 1911			.value = uclamp_none(clamp_id)
 1912		};
 1913	}
 1914
 1915	rq->uclamp_flags = 0;
 1916}
 1917
 1918static void __init init_uclamp(void)
 1919{
 1920	struct uclamp_se uc_max = {};
 1921	enum uclamp_id clamp_id;
 1922	int cpu;
 1923
 1924	for_each_possible_cpu(cpu)
 1925		init_uclamp_rq(cpu_rq(cpu));
 1926
 1927	for_each_clamp_id(clamp_id) {
 1928		uclamp_se_set(&init_task.uclamp_req[clamp_id],
 1929			      uclamp_none(clamp_id), false);
 1930	}
 1931
 1932	/* System defaults allow max clamp values for both indexes */
 1933	uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
 1934	for_each_clamp_id(clamp_id) {
 1935		uclamp_default[clamp_id] = uc_max;
 1936#ifdef CONFIG_UCLAMP_TASK_GROUP
 1937		root_task_group.uclamp_req[clamp_id] = uc_max;
 1938		root_task_group.uclamp[clamp_id] = uc_max;
 1939#endif
 1940	}
 1941}
 1942
 1943#else /* CONFIG_UCLAMP_TASK */
 1944static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
 1945static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
 1946static inline int uclamp_validate(struct task_struct *p,
 1947				  const struct sched_attr *attr)
 1948{
 1949	return -EOPNOTSUPP;
 1950}
 1951static void __setscheduler_uclamp(struct task_struct *p,
 1952				  const struct sched_attr *attr) { }
 1953static inline void uclamp_fork(struct task_struct *p) { }
 1954static inline void uclamp_post_fork(struct task_struct *p) { }
 1955static inline void init_uclamp(void) { }
 1956#endif /* CONFIG_UCLAMP_TASK */
 1957
 1958bool sched_task_on_rq(struct task_struct *p)
 1959{
 1960	return task_on_rq_queued(p);
 1961}
 1962
 1963static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 1964{
 1965	if (!(flags & ENQUEUE_NOCLOCK))
 1966		update_rq_clock(rq);
 1967
 1968	if (!(flags & ENQUEUE_RESTORE)) {
 
 
 
 
 
 
 
 
 
 1969		sched_info_enqueue(rq, p);
 1970		psi_enqueue(p, flags & ENQUEUE_WAKEUP);
 1971	}
 1972
 1973	uclamp_rq_inc(rq, p);
 1974	p->sched_class->enqueue_task(rq, p, flags);
 1975
 1976	if (sched_core_enabled(rq))
 1977		sched_core_enqueue(rq, p);
 1978}
 1979
 1980static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
 
 
 
 1981{
 1982	if (sched_core_enabled(rq))
 1983		sched_core_dequeue(rq, p);
 1984
 1985	if (!(flags & DEQUEUE_NOCLOCK))
 1986		update_rq_clock(rq);
 1987
 1988	if (!(flags & DEQUEUE_SAVE)) {
 1989		sched_info_dequeue(rq, p);
 1990		psi_dequeue(p, flags & DEQUEUE_SLEEP);
 1991	}
 1992
 
 
 
 
 
 
 1993	uclamp_rq_dec(rq, p);
 1994	p->sched_class->dequeue_task(rq, p, flags);
 1995}
 1996
 1997void activate_task(struct rq *rq, struct task_struct *p, int flags)
 1998{
 
 
 
 
 
 1999	enqueue_task(rq, p, flags);
 2000
 2001	p->on_rq = TASK_ON_RQ_QUEUED;
 
 2002}
 2003
 2004void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
 2005{
 2006	p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
 2007
 2008	dequeue_task(rq, p, flags);
 2009}
 2010
 2011static inline int __normal_prio(int policy, int rt_prio, int nice)
 2012{
 2013	int prio;
 
 2014
 2015	if (dl_policy(policy))
 2016		prio = MAX_DL_PRIO - 1;
 2017	else if (rt_policy(policy))
 2018		prio = MAX_RT_PRIO - 1 - rt_prio;
 2019	else
 2020		prio = NICE_TO_PRIO(nice);
 2021
 2022	return prio;
 2023}
 2024
 2025/*
 2026 * Calculate the expected normal priority: i.e. priority
 2027 * without taking RT-inheritance into account. Might be
 2028 * boosted by interactivity modifiers. Changes upon fork,
 2029 * setprio syscalls, and whenever the interactivity
 2030 * estimator recalculates.
 2031 */
 2032static inline int normal_prio(struct task_struct *p)
 2033{
 2034	return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
 2035}
 2036
 2037/*
 2038 * Calculate the current priority, i.e. the priority
 2039 * taken into account by the scheduler. This value might
 2040 * be boosted by RT tasks, or might be boosted by
 2041 * interactivity modifiers. Will be RT if the task got
 2042 * RT-boosted. If not then it returns p->normal_prio.
 2043 */
 2044static int effective_prio(struct task_struct *p)
 2045{
 2046	p->normal_prio = normal_prio(p);
 2047	/*
 2048	 * If we are RT tasks or we were boosted to RT priority,
 2049	 * keep the priority unchanged. Otherwise, update priority
 2050	 * to the normal priority:
 2051	 */
 2052	if (!rt_prio(p->prio))
 2053		return p->normal_prio;
 2054	return p->prio;
 2055}
 2056
 2057/**
 2058 * task_curr - is this task currently executing on a CPU?
 2059 * @p: the task in question.
 2060 *
 2061 * Return: 1 if the task is currently executing. 0 otherwise.
 2062 */
 2063inline int task_curr(const struct task_struct *p)
 2064{
 2065	return cpu_curr(task_cpu(p)) == p;
 2066}
 2067
 2068/*
 
 
 
 
 
 
 
 
 
 
 
 2069 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
 2070 * use the balance_callback list if you want balancing.
 2071 *
 2072 * this means any call to check_class_changed() must be followed by a call to
 2073 * balance_callback().
 2074 */
 2075static inline void check_class_changed(struct rq *rq, struct task_struct *p,
 2076				       const struct sched_class *prev_class,
 2077				       int oldprio)
 2078{
 2079	if (prev_class != p->sched_class) {
 2080		if (prev_class->switched_from)
 2081			prev_class->switched_from(rq, p);
 2082
 2083		p->sched_class->switched_to(rq, p);
 2084	} else if (oldprio != p->prio || dl_task(p))
 2085		p->sched_class->prio_changed(rq, p, oldprio);
 2086}
 2087
 2088void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
 2089{
 2090	if (p->sched_class == rq->curr->sched_class)
 2091		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
 2092	else if (p->sched_class > rq->curr->sched_class)
 
 
 2093		resched_curr(rq);
 2094
 2095	/*
 2096	 * A queue event has occurred, and we're going to schedule.  In
 2097	 * this case, we can save a useless back to back clock update.
 2098	 */
 2099	if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
 2100		rq_clock_skip_update(rq);
 2101}
 2102
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 2103#ifdef CONFIG_SMP
 2104
 2105static void
 2106__do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
 2107
 2108static int __set_cpus_allowed_ptr(struct task_struct *p,
 2109				  const struct cpumask *new_mask,
 2110				  u32 flags);
 2111
 2112static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
 2113{
 
 
 
 
 
 2114	if (likely(!p->migration_disabled))
 2115		return;
 2116
 2117	if (p->cpus_ptr != &p->cpus_mask)
 2118		return;
 2119
 2120	/*
 2121	 * Violates locking rules! see comment in __do_set_cpus_allowed().
 2122	 */
 2123	__do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
 2124}
 2125
 2126void migrate_disable(void)
 2127{
 2128	struct task_struct *p = current;
 2129
 2130	if (p->migration_disabled) {
 
 
 
 
 
 
 2131		p->migration_disabled++;
 2132		return;
 2133	}
 2134
 2135	preempt_disable();
 2136	this_rq()->nr_pinned++;
 2137	p->migration_disabled = 1;
 2138	preempt_enable();
 2139}
 2140EXPORT_SYMBOL_GPL(migrate_disable);
 2141
 2142void migrate_enable(void)
 2143{
 2144	struct task_struct *p = current;
 
 
 
 
 
 
 
 
 
 
 
 
 
 2145
 2146	if (p->migration_disabled > 1) {
 2147		p->migration_disabled--;
 2148		return;
 2149	}
 2150
 2151	/*
 2152	 * Ensure stop_task runs either before or after this, and that
 2153	 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
 2154	 */
 2155	preempt_disable();
 2156	if (p->cpus_ptr != &p->cpus_mask)
 2157		__set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
 2158	/*
 2159	 * Mustn't clear migration_disabled() until cpus_ptr points back at the
 2160	 * regular cpus_mask, otherwise things that race (eg.
 2161	 * select_fallback_rq) get confused.
 2162	 */
 2163	barrier();
 2164	p->migration_disabled = 0;
 2165	this_rq()->nr_pinned--;
 2166	preempt_enable();
 2167}
 2168EXPORT_SYMBOL_GPL(migrate_enable);
 2169
 2170static inline bool rq_has_pinned_tasks(struct rq *rq)
 2171{
 2172	return rq->nr_pinned;
 2173}
 2174
 2175/*
 2176 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
 2177 * __set_cpus_allowed_ptr() and select_fallback_rq().
 2178 */
 2179static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
 2180{
 2181	/* When not in the task's cpumask, no point in looking further. */
 2182	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
 2183		return false;
 2184
 2185	/* migrate_disabled() must be allowed to finish. */
 2186	if (is_migration_disabled(p))
 2187		return cpu_online(cpu);
 2188
 2189	/* Non kernel threads are not allowed during either online or offline. */
 2190	if (!(p->flags & PF_KTHREAD))
 2191		return cpu_active(cpu);
 2192
 2193	/* KTHREAD_IS_PER_CPU is always allowed. */
 2194	if (kthread_is_per_cpu(p))
 2195		return cpu_online(cpu);
 2196
 2197	/* Regular kernel threads don't get to stay during offline. */
 2198	if (cpu_dying(cpu))
 2199		return false;
 2200
 2201	/* But are allowed during online. */
 2202	return cpu_online(cpu);
 2203}
 2204
 2205/*
 2206 * This is how migration works:
 2207 *
 2208 * 1) we invoke migration_cpu_stop() on the target CPU using
 2209 *    stop_one_cpu().
 2210 * 2) stopper starts to run (implicitly forcing the migrated thread
 2211 *    off the CPU)
 2212 * 3) it checks whether the migrated task is still in the wrong runqueue.
 2213 * 4) if it's in the wrong runqueue then the migration thread removes
 2214 *    it and puts it into the right queue.
 2215 * 5) stopper completes and stop_one_cpu() returns and the migration
 2216 *    is done.
 2217 */
 2218
 2219/*
 2220 * move_queued_task - move a queued task to new rq.
 2221 *
 2222 * Returns (locked) new rq. Old rq's lock is released.
 2223 */
 2224static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
 2225				   struct task_struct *p, int new_cpu)
 2226{
 2227	lockdep_assert_rq_held(rq);
 2228
 2229	deactivate_task(rq, p, DEQUEUE_NOCLOCK);
 2230	set_task_cpu(p, new_cpu);
 2231	rq_unlock(rq, rf);
 2232
 2233	rq = cpu_rq(new_cpu);
 2234
 2235	rq_lock(rq, rf);
 2236	BUG_ON(task_cpu(p) != new_cpu);
 2237	activate_task(rq, p, 0);
 2238	check_preempt_curr(rq, p, 0);
 2239
 2240	return rq;
 2241}
 2242
 2243struct migration_arg {
 2244	struct task_struct		*task;
 2245	int				dest_cpu;
 2246	struct set_affinity_pending	*pending;
 2247};
 2248
 2249/*
 2250 * @refs: number of wait_for_completion()
 2251 * @stop_pending: is @stop_work in use
 2252 */
 2253struct set_affinity_pending {
 2254	refcount_t		refs;
 2255	unsigned int		stop_pending;
 2256	struct completion	done;
 2257	struct cpu_stop_work	stop_work;
 2258	struct migration_arg	arg;
 2259};
 2260
 2261/*
 2262 * Move (not current) task off this CPU, onto the destination CPU. We're doing
 2263 * this because either it can't run here any more (set_cpus_allowed()
 2264 * away from this CPU, or CPU going down), or because we're
 2265 * attempting to rebalance this task on exec (sched_exec).
 2266 *
 2267 * So we race with normal scheduler movements, but that's OK, as long
 2268 * as the task is no longer on this CPU.
 2269 */
 2270static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
 2271				 struct task_struct *p, int dest_cpu)
 2272{
 2273	/* Affinity changed (again). */
 2274	if (!is_cpu_allowed(p, dest_cpu))
 2275		return rq;
 2276
 2277	update_rq_clock(rq);
 2278	rq = move_queued_task(rq, rf, p, dest_cpu);
 2279
 2280	return rq;
 2281}
 2282
 2283/*
 2284 * migration_cpu_stop - this will be executed by a highprio stopper thread
 2285 * and performs thread migration by bumping thread off CPU then
 2286 * 'pushing' onto another runqueue.
 2287 */
 2288static int migration_cpu_stop(void *data)
 2289{
 2290	struct migration_arg *arg = data;
 2291	struct set_affinity_pending *pending = arg->pending;
 2292	struct task_struct *p = arg->task;
 2293	struct rq *rq = this_rq();
 2294	bool complete = false;
 2295	struct rq_flags rf;
 2296
 2297	/*
 2298	 * The original target CPU might have gone down and we might
 2299	 * be on another CPU but it doesn't matter.
 2300	 */
 2301	local_irq_save(rf.flags);
 2302	/*
 2303	 * We need to explicitly wake pending tasks before running
 2304	 * __migrate_task() such that we will not miss enforcing cpus_ptr
 2305	 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
 2306	 */
 2307	flush_smp_call_function_from_idle();
 2308
 2309	raw_spin_lock(&p->pi_lock);
 2310	rq_lock(rq, &rf);
 2311
 2312	/*
 2313	 * If we were passed a pending, then ->stop_pending was set, thus
 2314	 * p->migration_pending must have remained stable.
 2315	 */
 2316	WARN_ON_ONCE(pending && pending != p->migration_pending);
 2317
 2318	/*
 2319	 * If task_rq(p) != rq, it cannot be migrated here, because we're
 2320	 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
 2321	 * we're holding p->pi_lock.
 2322	 */
 2323	if (task_rq(p) == rq) {
 2324		if (is_migration_disabled(p))
 2325			goto out;
 2326
 2327		if (pending) {
 2328			p->migration_pending = NULL;
 2329			complete = true;
 2330
 2331			if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
 2332				goto out;
 2333		}
 2334
 2335		if (task_on_rq_queued(p))
 
 2336			rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
 2337		else
 2338			p->wake_cpu = arg->dest_cpu;
 
 2339
 2340		/*
 2341		 * XXX __migrate_task() can fail, at which point we might end
 2342		 * up running on a dodgy CPU, AFAICT this can only happen
 2343		 * during CPU hotplug, at which point we'll get pushed out
 2344		 * anyway, so it's probably not a big deal.
 2345		 */
 2346
 2347	} else if (pending) {
 2348		/*
 2349		 * This happens when we get migrated between migrate_enable()'s
 2350		 * preempt_enable() and scheduling the stopper task. At that
 2351		 * point we're a regular task again and not current anymore.
 2352		 *
 2353		 * A !PREEMPT kernel has a giant hole here, which makes it far
 2354		 * more likely.
 2355		 */
 2356
 2357		/*
 2358		 * The task moved before the stopper got to run. We're holding
 2359		 * ->pi_lock, so the allowed mask is stable - if it got
 2360		 * somewhere allowed, we're done.
 2361		 */
 2362		if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
 2363			p->migration_pending = NULL;
 2364			complete = true;
 2365			goto out;
 2366		}
 2367
 2368		/*
 2369		 * When migrate_enable() hits a rq mis-match we can't reliably
 2370		 * determine is_migration_disabled() and so have to chase after
 2371		 * it.
 2372		 */
 2373		WARN_ON_ONCE(!pending->stop_pending);
 
 2374		task_rq_unlock(rq, p, &rf);
 2375		stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
 2376				    &pending->arg, &pending->stop_work);
 
 2377		return 0;
 2378	}
 2379out:
 2380	if (pending)
 2381		pending->stop_pending = false;
 2382	task_rq_unlock(rq, p, &rf);
 2383
 2384	if (complete)
 2385		complete_all(&pending->done);
 2386
 2387	return 0;
 2388}
 2389
 2390int push_cpu_stop(void *arg)
 2391{
 2392	struct rq *lowest_rq = NULL, *rq = this_rq();
 2393	struct task_struct *p = arg;
 2394
 2395	raw_spin_lock_irq(&p->pi_lock);
 2396	raw_spin_rq_lock(rq);
 2397
 2398	if (task_rq(p) != rq)
 2399		goto out_unlock;
 2400
 2401	if (is_migration_disabled(p)) {
 2402		p->migration_flags |= MDF_PUSH;
 2403		goto out_unlock;
 2404	}
 2405
 2406	p->migration_flags &= ~MDF_PUSH;
 2407
 2408	if (p->sched_class->find_lock_rq)
 2409		lowest_rq = p->sched_class->find_lock_rq(p, rq);
 2410
 2411	if (!lowest_rq)
 2412		goto out_unlock;
 2413
 2414	// XXX validate p is still the highest prio task
 2415	if (task_rq(p) == rq) {
 2416		deactivate_task(rq, p, 0);
 2417		set_task_cpu(p, lowest_rq->cpu);
 2418		activate_task(lowest_rq, p, 0);
 2419		resched_curr(lowest_rq);
 2420	}
 2421
 2422	double_unlock_balance(rq, lowest_rq);
 2423
 2424out_unlock:
 2425	rq->push_busy = false;
 2426	raw_spin_rq_unlock(rq);
 2427	raw_spin_unlock_irq(&p->pi_lock);
 2428
 2429	put_task_struct(p);
 2430	return 0;
 2431}
 2432
 2433/*
 2434 * sched_class::set_cpus_allowed must do the below, but is not required to
 2435 * actually call this function.
 2436 */
 2437void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
 2438{
 2439	if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
 2440		p->cpus_ptr = new_mask;
 2441		return;
 2442	}
 2443
 2444	cpumask_copy(&p->cpus_mask, new_mask);
 2445	p->nr_cpus_allowed = cpumask_weight(new_mask);
 
 
 
 
 
 
 2446}
 2447
 2448static void
 2449__do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
 2450{
 2451	struct rq *rq = task_rq(p);
 2452	bool queued, running;
 2453
 2454	/*
 2455	 * This here violates the locking rules for affinity, since we're only
 2456	 * supposed to change these variables while holding both rq->lock and
 2457	 * p->pi_lock.
 2458	 *
 2459	 * HOWEVER, it magically works, because ttwu() is the only code that
 2460	 * accesses these variables under p->pi_lock and only does so after
 2461	 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
 2462	 * before finish_task().
 2463	 *
 2464	 * XXX do further audits, this smells like something putrid.
 2465	 */
 2466	if (flags & SCA_MIGRATE_DISABLE)
 2467		SCHED_WARN_ON(!p->on_cpu);
 2468	else
 2469		lockdep_assert_held(&p->pi_lock);
 2470
 2471	queued = task_on_rq_queued(p);
 2472	running = task_current(rq, p);
 2473
 2474	if (queued) {
 2475		/*
 2476		 * Because __kthread_bind() calls this on blocked tasks without
 2477		 * holding rq->lock.
 2478		 */
 2479		lockdep_assert_rq_held(rq);
 2480		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
 2481	}
 2482	if (running)
 2483		put_prev_task(rq, p);
 2484
 2485	p->sched_class->set_cpus_allowed(p, new_mask, flags);
 
 2486
 2487	if (queued)
 2488		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
 2489	if (running)
 2490		set_next_task(rq, p);
 2491}
 2492
 
 
 
 
 2493void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
 2494{
 2495	__do_set_cpus_allowed(p, new_mask, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 2496}
 2497
 2498/*
 2499 * This function is wildly self concurrent; here be dragons.
 2500 *
 2501 *
 2502 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
 2503 * designated task is enqueued on an allowed CPU. If that task is currently
 2504 * running, we have to kick it out using the CPU stopper.
 2505 *
 2506 * Migrate-Disable comes along and tramples all over our nice sandcastle.
 2507 * Consider:
 2508 *
 2509 *     Initial conditions: P0->cpus_mask = [0, 1]
 2510 *
 2511 *     P0@CPU0                  P1
 2512 *
 2513 *     migrate_disable();
 2514 *     <preempted>
 2515 *                              set_cpus_allowed_ptr(P0, [1]);
 2516 *
 2517 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
 2518 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
 2519 * This means we need the following scheme:
 2520 *
 2521 *     P0@CPU0                  P1
 2522 *
 2523 *     migrate_disable();
 2524 *     <preempted>
 2525 *                              set_cpus_allowed_ptr(P0, [1]);
 2526 *                                <blocks>
 2527 *     <resumes>
 2528 *     migrate_enable();
 2529 *       __set_cpus_allowed_ptr();
 2530 *       <wakes local stopper>
 2531 *                         `--> <woken on migration completion>
 2532 *
 2533 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
 2534 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
 2535 * task p are serialized by p->pi_lock, which we can leverage: the one that
 2536 * should come into effect at the end of the Migrate-Disable region is the last
 2537 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
 2538 * but we still need to properly signal those waiting tasks at the appropriate
 2539 * moment.
 2540 *
 2541 * This is implemented using struct set_affinity_pending. The first
 2542 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
 2543 * setup an instance of that struct and install it on the targeted task_struct.
 2544 * Any and all further callers will reuse that instance. Those then wait for
 2545 * a completion signaled at the tail of the CPU stopper callback (1), triggered
 2546 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
 2547 *
 2548 *
 2549 * (1) In the cases covered above. There is one more where the completion is
 2550 * signaled within affine_move_task() itself: when a subsequent affinity request
 2551 * occurs after the stopper bailed out due to the targeted task still being
 2552 * Migrate-Disable. Consider:
 2553 *
 2554 *     Initial conditions: P0->cpus_mask = [0, 1]
 2555 *
 2556 *     CPU0		  P1				P2
 2557 *     <P0>
 2558 *       migrate_disable();
 2559 *       <preempted>
 2560 *                        set_cpus_allowed_ptr(P0, [1]);
 2561 *                          <blocks>
 2562 *     <migration/0>
 2563 *       migration_cpu_stop()
 2564 *         is_migration_disabled()
 2565 *           <bails>
 2566 *                                                       set_cpus_allowed_ptr(P0, [0, 1]);
 2567 *                                                         <signal completion>
 2568 *                          <awakes>
 2569 *
 2570 * Note that the above is safe vs a concurrent migrate_enable(), as any
 2571 * pending affinity completion is preceded by an uninstallation of
 2572 * p->migration_pending done with p->pi_lock held.
 2573 */
 2574static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
 2575			    int dest_cpu, unsigned int flags)
 
 
 2576{
 2577	struct set_affinity_pending my_pending = { }, *pending = NULL;
 2578	bool stop_pending, complete = false;
 2579
 2580	/* Can the task run on the task's current CPU? If so, we're done */
 2581	if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
 2582		struct task_struct *push_task = NULL;
 2583
 2584		if ((flags & SCA_MIGRATE_ENABLE) &&
 2585		    (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
 2586			rq->push_busy = true;
 2587			push_task = get_task_struct(p);
 2588		}
 2589
 2590		/*
 2591		 * If there are pending waiters, but no pending stop_work,
 2592		 * then complete now.
 2593		 */
 2594		pending = p->migration_pending;
 2595		if (pending && !pending->stop_pending) {
 2596			p->migration_pending = NULL;
 2597			complete = true;
 2598		}
 2599
 
 2600		task_rq_unlock(rq, p, rf);
 2601
 2602		if (push_task) {
 2603			stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
 2604					    p, &rq->push_work);
 2605		}
 
 2606
 2607		if (complete)
 2608			complete_all(&pending->done);
 2609
 2610		return 0;
 2611	}
 2612
 2613	if (!(flags & SCA_MIGRATE_ENABLE)) {
 2614		/* serialized by p->pi_lock */
 2615		if (!p->migration_pending) {
 2616			/* Install the request */
 2617			refcount_set(&my_pending.refs, 1);
 2618			init_completion(&my_pending.done);
 2619			my_pending.arg = (struct migration_arg) {
 2620				.task = p,
 2621				.dest_cpu = dest_cpu,
 2622				.pending = &my_pending,
 2623			};
 2624
 2625			p->migration_pending = &my_pending;
 2626		} else {
 2627			pending = p->migration_pending;
 2628			refcount_inc(&pending->refs);
 2629			/*
 2630			 * Affinity has changed, but we've already installed a
 2631			 * pending. migration_cpu_stop() *must* see this, else
 2632			 * we risk a completion of the pending despite having a
 2633			 * task on a disallowed CPU.
 2634			 *
 2635			 * Serialized by p->pi_lock, so this is safe.
 2636			 */
 2637			pending->arg.dest_cpu = dest_cpu;
 2638		}
 2639	}
 2640	pending = p->migration_pending;
 2641	/*
 2642	 * - !MIGRATE_ENABLE:
 2643	 *   we'll have installed a pending if there wasn't one already.
 2644	 *
 2645	 * - MIGRATE_ENABLE:
 2646	 *   we're here because the current CPU isn't matching anymore,
 2647	 *   the only way that can happen is because of a concurrent
 2648	 *   set_cpus_allowed_ptr() call, which should then still be
 2649	 *   pending completion.
 2650	 *
 2651	 * Either way, we really should have a @pending here.
 2652	 */
 2653	if (WARN_ON_ONCE(!pending)) {
 2654		task_rq_unlock(rq, p, rf);
 2655		return -EINVAL;
 2656	}
 2657
 2658	if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
 2659		/*
 2660		 * MIGRATE_ENABLE gets here because 'p == current', but for
 2661		 * anything else we cannot do is_migration_disabled(), punt
 2662		 * and have the stopper function handle it all race-free.
 2663		 */
 2664		stop_pending = pending->stop_pending;
 2665		if (!stop_pending)
 2666			pending->stop_pending = true;
 2667
 2668		if (flags & SCA_MIGRATE_ENABLE)
 2669			p->migration_flags &= ~MDF_PUSH;
 2670
 
 2671		task_rq_unlock(rq, p, rf);
 2672
 2673		if (!stop_pending) {
 2674			stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
 2675					    &pending->arg, &pending->stop_work);
 2676		}
 
 2677
 2678		if (flags & SCA_MIGRATE_ENABLE)
 2679			return 0;
 2680	} else {
 2681
 2682		if (!is_migration_disabled(p)) {
 2683			if (task_on_rq_queued(p))
 2684				rq = move_queued_task(rq, rf, p, dest_cpu);
 2685
 2686			if (!pending->stop_pending) {
 2687				p->migration_pending = NULL;
 2688				complete = true;
 2689			}
 2690		}
 2691		task_rq_unlock(rq, p, rf);
 2692
 2693		if (complete)
 2694			complete_all(&pending->done);
 2695	}
 2696
 2697	wait_for_completion(&pending->done);
 2698
 2699	if (refcount_dec_and_test(&pending->refs))
 2700		wake_up_var(&pending->refs); /* No UaF, just an address */
 2701
 2702	/*
 2703	 * Block the original owner of &pending until all subsequent callers
 2704	 * have seen the completion and decremented the refcount
 2705	 */
 2706	wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
 2707
 2708	/* ARGH */
 2709	WARN_ON_ONCE(my_pending.stop_pending);
 2710
 2711	return 0;
 2712}
 2713
 2714/*
 2715 * Change a given task's CPU affinity. Migrate the thread to a
 2716 * proper CPU and schedule it away if the CPU it's executing on
 2717 * is removed from the allowed bitmask.
 2718 *
 2719 * NOTE: the caller must have a valid reference to the task, the
 2720 * task must not exit() & deallocate itself prematurely. The
 2721 * call is not atomic; no spinlocks may be held.
 2722 */
 2723static int __set_cpus_allowed_ptr(struct task_struct *p,
 2724				  const struct cpumask *new_mask,
 2725				  u32 flags)
 
 
 
 2726{
 
 2727	const struct cpumask *cpu_valid_mask = cpu_active_mask;
 
 2728	unsigned int dest_cpu;
 2729	struct rq_flags rf;
 2730	struct rq *rq;
 2731	int ret = 0;
 2732
 2733	rq = task_rq_lock(p, &rf);
 2734	update_rq_clock(rq);
 2735
 2736	if (p->flags & PF_KTHREAD || is_migration_disabled(p)) {
 2737		/*
 2738		 * Kernel threads are allowed on online && !active CPUs,
 2739		 * however, during cpu-hot-unplug, even these might get pushed
 2740		 * away if not KTHREAD_IS_PER_CPU.
 2741		 *
 2742		 * Specifically, migration_disabled() tasks must not fail the
 2743		 * cpumask_any_and_distribute() pick below, esp. so on
 2744		 * SCA_MIGRATE_ENABLE, otherwise we'll not call
 2745		 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
 2746		 */
 2747		cpu_valid_mask = cpu_online_mask;
 2748	}
 2749
 
 
 
 
 
 2750	/*
 2751	 * Must re-check here, to close a race against __kthread_bind(),
 2752	 * sched_setaffinity() is not guaranteed to observe the flag.
 2753	 */
 2754	if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
 2755		ret = -EINVAL;
 2756		goto out;
 2757	}
 2758
 2759	if (!(flags & SCA_MIGRATE_ENABLE)) {
 2760		if (cpumask_equal(&p->cpus_mask, new_mask))
 
 
 2761			goto out;
 
 2762
 2763		if (WARN_ON_ONCE(p == current &&
 2764				 is_migration_disabled(p) &&
 2765				 !cpumask_test_cpu(task_cpu(p), new_mask))) {
 2766			ret = -EBUSY;
 2767			goto out;
 2768		}
 2769	}
 2770
 2771	/*
 2772	 * Picking a ~random cpu helps in cases where we are changing affinity
 2773	 * for groups of tasks (ie. cpuset), so that load balancing is not
 2774	 * immediately required to distribute the tasks within their new mask.
 2775	 */
 2776	dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
 2777	if (dest_cpu >= nr_cpu_ids) {
 2778		ret = -EINVAL;
 2779		goto out;
 2780	}
 2781
 2782	__do_set_cpus_allowed(p, new_mask, flags);
 2783
 2784	return affine_move_task(rq, p, &rf, dest_cpu, flags);
 2785
 2786out:
 2787	task_rq_unlock(rq, p, &rf);
 2788
 2789	return ret;
 2790}
 2791
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 2792int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
 2793{
 2794	return __set_cpus_allowed_ptr(p, new_mask, 0);
 
 
 
 
 
 2795}
 2796EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
 2797
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 2798void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
 2799{
 2800#ifdef CONFIG_SCHED_DEBUG
 2801	unsigned int state = READ_ONCE(p->__state);
 2802
 2803	/*
 2804	 * We should never call set_task_cpu() on a blocked task,
 2805	 * ttwu() will sort out the placement.
 2806	 */
 2807	WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
 2808
 2809	/*
 2810	 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
 2811	 * because schedstat_wait_{start,end} rebase migrating task's wait_start
 2812	 * time relying on p->on_rq.
 2813	 */
 2814	WARN_ON_ONCE(state == TASK_RUNNING &&
 2815		     p->sched_class == &fair_sched_class &&
 2816		     (p->on_rq && !task_on_rq_migrating(p)));
 2817
 2818#ifdef CONFIG_LOCKDEP
 2819	/*
 2820	 * The caller should hold either p->pi_lock or rq->lock, when changing
 2821	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
 2822	 *
 2823	 * sched_move_task() holds both and thus holding either pins the cgroup,
 2824	 * see task_group().
 2825	 *
 2826	 * Furthermore, all task_rq users should acquire both locks, see
 2827	 * task_rq_lock().
 2828	 */
 2829	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
 2830				      lockdep_is_held(__rq_lockp(task_rq(p)))));
 2831#endif
 2832	/*
 2833	 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
 2834	 */
 2835	WARN_ON_ONCE(!cpu_online(new_cpu));
 2836
 2837	WARN_ON_ONCE(is_migration_disabled(p));
 2838#endif
 2839
 2840	trace_sched_migrate_task(p, new_cpu);
 2841
 2842	if (task_cpu(p) != new_cpu) {
 2843		if (p->sched_class->migrate_task_rq)
 2844			p->sched_class->migrate_task_rq(p, new_cpu);
 2845		p->se.nr_migrations++;
 2846		rseq_migrate(p);
 
 2847		perf_event_task_migrate(p);
 2848	}
 2849
 2850	__set_task_cpu(p, new_cpu);
 2851}
 2852
 2853#ifdef CONFIG_NUMA_BALANCING
 2854static void __migrate_swap_task(struct task_struct *p, int cpu)
 2855{
 2856	if (task_on_rq_queued(p)) {
 2857		struct rq *src_rq, *dst_rq;
 2858		struct rq_flags srf, drf;
 2859
 2860		src_rq = task_rq(p);
 2861		dst_rq = cpu_rq(cpu);
 2862
 2863		rq_pin_lock(src_rq, &srf);
 2864		rq_pin_lock(dst_rq, &drf);
 2865
 2866		deactivate_task(src_rq, p, 0);
 2867		set_task_cpu(p, cpu);
 2868		activate_task(dst_rq, p, 0);
 2869		check_preempt_curr(dst_rq, p, 0);
 2870
 2871		rq_unpin_lock(dst_rq, &drf);
 2872		rq_unpin_lock(src_rq, &srf);
 2873
 2874	} else {
 2875		/*
 2876		 * Task isn't running anymore; make it appear like we migrated
 2877		 * it before it went to sleep. This means on wakeup we make the
 2878		 * previous CPU our target instead of where it really is.
 2879		 */
 2880		p->wake_cpu = cpu;
 2881	}
 2882}
 2883
 2884struct migration_swap_arg {
 2885	struct task_struct *src_task, *dst_task;
 2886	int src_cpu, dst_cpu;
 2887};
 2888
 2889static int migrate_swap_stop(void *data)
 2890{
 2891	struct migration_swap_arg *arg = data;
 2892	struct rq *src_rq, *dst_rq;
 2893	int ret = -EAGAIN;
 2894
 2895	if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
 2896		return -EAGAIN;
 2897
 2898	src_rq = cpu_rq(arg->src_cpu);
 2899	dst_rq = cpu_rq(arg->dst_cpu);
 2900
 2901	double_raw_lock(&arg->src_task->pi_lock,
 2902			&arg->dst_task->pi_lock);
 2903	double_rq_lock(src_rq, dst_rq);
 2904
 2905	if (task_cpu(arg->dst_task) != arg->dst_cpu)
 2906		goto unlock;
 2907
 2908	if (task_cpu(arg->src_task) != arg->src_cpu)
 2909		goto unlock;
 2910
 2911	if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
 2912		goto unlock;
 2913
 2914	if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
 2915		goto unlock;
 2916
 2917	__migrate_swap_task(arg->src_task, arg->dst_cpu);
 2918	__migrate_swap_task(arg->dst_task, arg->src_cpu);
 2919
 2920	ret = 0;
 2921
 2922unlock:
 2923	double_rq_unlock(src_rq, dst_rq);
 2924	raw_spin_unlock(&arg->dst_task->pi_lock);
 2925	raw_spin_unlock(&arg->src_task->pi_lock);
 2926
 2927	return ret;
 2928}
 2929
 2930/*
 2931 * Cross migrate two tasks
 2932 */
 2933int migrate_swap(struct task_struct *cur, struct task_struct *p,
 2934		int target_cpu, int curr_cpu)
 2935{
 2936	struct migration_swap_arg arg;
 2937	int ret = -EINVAL;
 2938
 2939	arg = (struct migration_swap_arg){
 2940		.src_task = cur,
 2941		.src_cpu = curr_cpu,
 2942		.dst_task = p,
 2943		.dst_cpu = target_cpu,
 2944	};
 2945
 2946	if (arg.src_cpu == arg.dst_cpu)
 2947		goto out;
 2948
 2949	/*
 2950	 * These three tests are all lockless; this is OK since all of them
 2951	 * will be re-checked with proper locks held further down the line.
 2952	 */
 2953	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
 2954		goto out;
 2955
 2956	if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
 2957		goto out;
 2958
 2959	if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
 2960		goto out;
 2961
 2962	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
 2963	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
 2964
 2965out:
 2966	return ret;
 2967}
 2968#endif /* CONFIG_NUMA_BALANCING */
 2969
 2970/*
 2971 * wait_task_inactive - wait for a thread to unschedule.
 2972 *
 2973 * If @match_state is nonzero, it's the @p->state value just checked and
 2974 * not expected to change.  If it changes, i.e. @p might have woken up,
 2975 * then return zero.  When we succeed in waiting for @p to be off its CPU,
 2976 * we return a positive number (its total switch count).  If a second call
 2977 * a short while later returns the same number, the caller can be sure that
 2978 * @p has remained unscheduled the whole time.
 2979 *
 2980 * The caller must ensure that the task *will* unschedule sometime soon,
 2981 * else this function might spin for a *long* time. This function can't
 2982 * be called with interrupts off, or it may introduce deadlock with
 2983 * smp_call_function() if an IPI is sent by the same process we are
 2984 * waiting to become inactive.
 2985 */
 2986unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
 2987{
 2988	int running, queued;
 2989	struct rq_flags rf;
 2990	unsigned long ncsw;
 2991	struct rq *rq;
 2992
 2993	for (;;) {
 2994		/*
 2995		 * We do the initial early heuristics without holding
 2996		 * any task-queue locks at all. We'll only try to get
 2997		 * the runqueue lock when things look like they will
 2998		 * work out!
 2999		 */
 3000		rq = task_rq(p);
 3001
 3002		/*
 3003		 * If the task is actively running on another CPU
 3004		 * still, just relax and busy-wait without holding
 3005		 * any locks.
 3006		 *
 3007		 * NOTE! Since we don't hold any locks, it's not
 3008		 * even sure that "rq" stays as the right runqueue!
 3009		 * But we don't care, since "task_running()" will
 3010		 * return false if the runqueue has changed and p
 3011		 * is actually now running somewhere else!
 3012		 */
 3013		while (task_running(rq, p)) {
 3014			if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
 3015				return 0;
 3016			cpu_relax();
 3017		}
 3018
 3019		/*
 3020		 * Ok, time to look more closely! We need the rq
 3021		 * lock now, to be *sure*. If we're wrong, we'll
 3022		 * just go back and repeat.
 3023		 */
 3024		rq = task_rq_lock(p, &rf);
 3025		trace_sched_wait_task(p);
 3026		running = task_running(rq, p);
 3027		queued = task_on_rq_queued(p);
 3028		ncsw = 0;
 3029		if (!match_state || READ_ONCE(p->__state) == match_state)
 3030			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
 3031		task_rq_unlock(rq, p, &rf);
 3032
 3033		/*
 3034		 * If it changed from the expected state, bail out now.
 3035		 */
 3036		if (unlikely(!ncsw))
 3037			break;
 3038
 3039		/*
 3040		 * Was it really running after all now that we
 3041		 * checked with the proper locks actually held?
 3042		 *
 3043		 * Oops. Go back and try again..
 3044		 */
 3045		if (unlikely(running)) {
 3046			cpu_relax();
 3047			continue;
 3048		}
 3049
 3050		/*
 3051		 * It's not enough that it's not actively running,
 3052		 * it must be off the runqueue _entirely_, and not
 3053		 * preempted!
 3054		 *
 3055		 * So if it was still runnable (but just not actively
 3056		 * running right now), it's preempted, and we should
 3057		 * yield - it could be a while.
 3058		 */
 3059		if (unlikely(queued)) {
 3060			ktime_t to = NSEC_PER_SEC / HZ;
 3061
 3062			set_current_state(TASK_UNINTERRUPTIBLE);
 3063			schedule_hrtimeout(&to, HRTIMER_MODE_REL);
 3064			continue;
 3065		}
 3066
 3067		/*
 3068		 * Ahh, all good. It wasn't running, and it wasn't
 3069		 * runnable, which means that it will never become
 3070		 * running in the future either. We're all done!
 3071		 */
 3072		break;
 3073	}
 3074
 3075	return ncsw;
 3076}
 3077
 3078/***
 3079 * kick_process - kick a running thread to enter/exit the kernel
 3080 * @p: the to-be-kicked thread
 3081 *
 3082 * Cause a process which is running on another CPU to enter
 3083 * kernel-mode, without any delay. (to get signals handled.)
 3084 *
 3085 * NOTE: this function doesn't have to take the runqueue lock,
 3086 * because all it wants to ensure is that the remote task enters
 3087 * the kernel. If the IPI races and the task has been migrated
 3088 * to another CPU then no harm is done and the purpose has been
 3089 * achieved as well.
 3090 */
 3091void kick_process(struct task_struct *p)
 3092{
 3093	int cpu;
 
 3094
 3095	preempt_disable();
 3096	cpu = task_cpu(p);
 3097	if ((cpu != smp_processor_id()) && task_curr(p))
 3098		smp_send_reschedule(cpu);
 3099	preempt_enable();
 3100}
 3101EXPORT_SYMBOL_GPL(kick_process);
 3102
 3103/*
 3104 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
 3105 *
 3106 * A few notes on cpu_active vs cpu_online:
 3107 *
 3108 *  - cpu_active must be a subset of cpu_online
 3109 *
 3110 *  - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
 3111 *    see __set_cpus_allowed_ptr(). At this point the newly online
 3112 *    CPU isn't yet part of the sched domains, and balancing will not
 3113 *    see it.
 3114 *
 3115 *  - on CPU-down we clear cpu_active() to mask the sched domains and
 3116 *    avoid the load balancer to place new tasks on the to be removed
 3117 *    CPU. Existing tasks will remain running there and will be taken
 3118 *    off.
 3119 *
 3120 * This means that fallback selection must not select !active CPUs.
 3121 * And can assume that any active CPU must be online. Conversely
 3122 * select_task_rq() below may allow selection of !active CPUs in order
 3123 * to satisfy the above rules.
 3124 */
 3125static int select_fallback_rq(int cpu, struct task_struct *p)
 3126{
 3127	int nid = cpu_to_node(cpu);
 3128	const struct cpumask *nodemask = NULL;
 3129	enum { cpuset, possible, fail } state = cpuset;
 3130	int dest_cpu;
 3131
 3132	/*
 3133	 * If the node that the CPU is on has been offlined, cpu_to_node()
 3134	 * will return -1. There is no CPU on the node, and we should
 3135	 * select the CPU on the other node.
 3136	 */
 3137	if (nid != -1) {
 3138		nodemask = cpumask_of_node(nid);
 3139
 3140		/* Look for allowed, online CPU in same node. */
 3141		for_each_cpu(dest_cpu, nodemask) {
 3142			if (!cpu_active(dest_cpu))
 3143				continue;
 3144			if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
 3145				return dest_cpu;
 3146		}
 3147	}
 3148
 3149	for (;;) {
 3150		/* Any allowed, online CPU? */
 3151		for_each_cpu(dest_cpu, p->cpus_ptr) {
 3152			if (!is_cpu_allowed(p, dest_cpu))
 3153				continue;
 3154
 3155			goto out;
 3156		}
 3157
 3158		/* No more Mr. Nice Guy. */
 3159		switch (state) {
 3160		case cpuset:
 3161			if (IS_ENABLED(CONFIG_CPUSETS)) {
 3162				cpuset_cpus_allowed_fallback(p);
 3163				state = possible;
 3164				break;
 3165			}
 3166			fallthrough;
 3167		case possible:
 3168			/*
 3169			 * XXX When called from select_task_rq() we only
 3170			 * hold p->pi_lock and again violate locking order.
 3171			 *
 3172			 * More yuck to audit.
 3173			 */
 3174			do_set_cpus_allowed(p, cpu_possible_mask);
 3175			state = fail;
 3176			break;
 3177
 3178		case fail:
 3179			BUG();
 3180			break;
 3181		}
 3182	}
 3183
 3184out:
 3185	if (state != cpuset) {
 3186		/*
 3187		 * Don't tell them about moving exiting tasks or
 3188		 * kernel threads (both mm NULL), since they never
 3189		 * leave kernel.
 3190		 */
 3191		if (p->mm && printk_ratelimit()) {
 3192			printk_deferred("process %d (%s) no longer affine to cpu%d\n",
 3193					task_pid_nr(p), p->comm, cpu);
 3194		}
 3195	}
 3196
 3197	return dest_cpu;
 3198}
 3199
 3200/*
 3201 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
 3202 */
 3203static inline
 3204int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
 3205{
 3206	lockdep_assert_held(&p->pi_lock);
 3207
 3208	if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
 3209		cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
 3210	else
 
 3211		cpu = cpumask_any(p->cpus_ptr);
 
 3212
 3213	/*
 3214	 * In order not to call set_task_cpu() on a blocking task we need
 3215	 * to rely on ttwu() to place the task on a valid ->cpus_ptr
 3216	 * CPU.
 3217	 *
 3218	 * Since this is common to all placement strategies, this lives here.
 3219	 *
 3220	 * [ this allows ->select_task() to simply return task_cpu(p) and
 3221	 *   not worry about this generic constraint ]
 3222	 */
 3223	if (unlikely(!is_cpu_allowed(p, cpu)))
 3224		cpu = select_fallback_rq(task_cpu(p), p);
 3225
 3226	return cpu;
 3227}
 3228
 3229void sched_set_stop_task(int cpu, struct task_struct *stop)
 3230{
 3231	static struct lock_class_key stop_pi_lock;
 3232	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
 3233	struct task_struct *old_stop = cpu_rq(cpu)->stop;
 3234
 3235	if (stop) {
 3236		/*
 3237		 * Make it appear like a SCHED_FIFO task, its something
 3238		 * userspace knows about and won't get confused about.
 3239		 *
 3240		 * Also, it will make PI more or less work without too
 3241		 * much confusion -- but then, stop work should not
 3242		 * rely on PI working anyway.
 3243		 */
 3244		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
 3245
 3246		stop->sched_class = &stop_sched_class;
 3247
 3248		/*
 3249		 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
 3250		 * adjust the effective priority of a task. As a result,
 3251		 * rt_mutex_setprio() can trigger (RT) balancing operations,
 3252		 * which can then trigger wakeups of the stop thread to push
 3253		 * around the current task.
 3254		 *
 3255		 * The stop task itself will never be part of the PI-chain, it
 3256		 * never blocks, therefore that ->pi_lock recursion is safe.
 3257		 * Tell lockdep about this by placing the stop->pi_lock in its
 3258		 * own class.
 3259		 */
 3260		lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
 3261	}
 3262
 3263	cpu_rq(cpu)->stop = stop;
 3264
 3265	if (old_stop) {
 3266		/*
 3267		 * Reset it back to a normal scheduling class so that
 3268		 * it can die in pieces.
 3269		 */
 3270		old_stop->sched_class = &rt_sched_class;
 3271	}
 3272}
 3273
 3274#else /* CONFIG_SMP */
 3275
 3276static inline int __set_cpus_allowed_ptr(struct task_struct *p,
 3277					 const struct cpumask *new_mask,
 3278					 u32 flags)
 3279{
 3280	return set_cpus_allowed_ptr(p, new_mask);
 3281}
 3282
 3283static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
 3284
 3285static inline bool rq_has_pinned_tasks(struct rq *rq)
 3286{
 3287	return false;
 3288}
 3289
 3290#endif /* !CONFIG_SMP */
 3291
 3292static void
 3293ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
 3294{
 3295	struct rq *rq;
 3296
 3297	if (!schedstat_enabled())
 3298		return;
 3299
 3300	rq = this_rq();
 3301
 3302#ifdef CONFIG_SMP
 3303	if (cpu == rq->cpu) {
 3304		__schedstat_inc(rq->ttwu_local);
 3305		__schedstat_inc(p->se.statistics.nr_wakeups_local);
 3306	} else {
 3307		struct sched_domain *sd;
 3308
 3309		__schedstat_inc(p->se.statistics.nr_wakeups_remote);
 3310		rcu_read_lock();
 
 3311		for_each_domain(rq->cpu, sd) {
 3312			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
 3313				__schedstat_inc(sd->ttwu_wake_remote);
 3314				break;
 3315			}
 3316		}
 3317		rcu_read_unlock();
 3318	}
 3319
 3320	if (wake_flags & WF_MIGRATED)
 3321		__schedstat_inc(p->se.statistics.nr_wakeups_migrate);
 3322#endif /* CONFIG_SMP */
 3323
 3324	__schedstat_inc(rq->ttwu_count);
 3325	__schedstat_inc(p->se.statistics.nr_wakeups);
 3326
 3327	if (wake_flags & WF_SYNC)
 3328		__schedstat_inc(p->se.statistics.nr_wakeups_sync);
 3329}
 3330
 3331/*
 3332 * Mark the task runnable and perform wakeup-preemption.
 3333 */
 3334static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
 3335			   struct rq_flags *rf)
 3336{
 3337	check_preempt_curr(rq, p, wake_flags);
 3338	WRITE_ONCE(p->__state, TASK_RUNNING);
 3339	trace_sched_wakeup(p);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 3340
 3341#ifdef CONFIG_SMP
 3342	if (p->sched_class->task_woken) {
 3343		/*
 3344		 * Our task @p is fully woken up and running; so it's safe to
 3345		 * drop the rq->lock, hereafter rq is only used for statistics.
 3346		 */
 3347		rq_unpin_lock(rq, rf);
 3348		p->sched_class->task_woken(rq, p);
 3349		rq_repin_lock(rq, rf);
 3350	}
 3351
 3352	if (rq->idle_stamp) {
 3353		u64 delta = rq_clock(rq) - rq->idle_stamp;
 3354		u64 max = 2*rq->max_idle_balance_cost;
 3355
 3356		update_avg(&rq->avg_idle, delta);
 3357
 3358		if (rq->avg_idle > max)
 3359			rq->avg_idle = max;
 3360
 3361		rq->wake_stamp = jiffies;
 3362		rq->wake_avg_idle = rq->avg_idle / 2;
 3363
 3364		rq->idle_stamp = 0;
 3365	}
 3366#endif
 3367}
 3368
 3369static void
 3370ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
 3371		 struct rq_flags *rf)
 3372{
 3373	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
 3374
 3375	lockdep_assert_rq_held(rq);
 3376
 3377	if (p->sched_contributes_to_load)
 3378		rq->nr_uninterruptible--;
 3379
 3380#ifdef CONFIG_SMP
 3381	if (wake_flags & WF_MIGRATED)
 3382		en_flags |= ENQUEUE_MIGRATED;
 3383	else
 3384#endif
 3385	if (p->in_iowait) {
 3386		delayacct_blkio_end(p);
 3387		atomic_dec(&task_rq(p)->nr_iowait);
 3388	}
 3389
 3390	activate_task(rq, p, en_flags);
 3391	ttwu_do_wakeup(rq, p, wake_flags, rf);
 3392}
 3393
 3394/*
 3395 * Consider @p being inside a wait loop:
 3396 *
 3397 *   for (;;) {
 3398 *      set_current_state(TASK_UNINTERRUPTIBLE);
 3399 *
 3400 *      if (CONDITION)
 3401 *         break;
 3402 *
 3403 *      schedule();
 3404 *   }
 3405 *   __set_current_state(TASK_RUNNING);
 3406 *
 3407 * between set_current_state() and schedule(). In this case @p is still
 3408 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
 3409 * an atomic manner.
 3410 *
 3411 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
 3412 * then schedule() must still happen and p->state can be changed to
 3413 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
 3414 * need to do a full wakeup with enqueue.
 3415 *
 3416 * Returns: %true when the wakeup is done,
 3417 *          %false otherwise.
 3418 */
 3419static int ttwu_runnable(struct task_struct *p, int wake_flags)
 3420{
 3421	struct rq_flags rf;
 3422	struct rq *rq;
 3423	int ret = 0;
 3424
 3425	rq = __task_rq_lock(p, &rf);
 3426	if (task_on_rq_queued(p)) {
 3427		/* check_preempt_curr() may use rq clock */
 3428		update_rq_clock(rq);
 3429		ttwu_do_wakeup(rq, p, wake_flags, &rf);
 
 
 
 
 
 
 
 
 
 3430		ret = 1;
 3431	}
 3432	__task_rq_unlock(rq, &rf);
 3433
 3434	return ret;
 3435}
 3436
 3437#ifdef CONFIG_SMP
 3438void sched_ttwu_pending(void *arg)
 3439{
 3440	struct llist_node *llist = arg;
 3441	struct rq *rq = this_rq();
 3442	struct task_struct *p, *t;
 3443	struct rq_flags rf;
 3444
 3445	if (!llist)
 3446		return;
 3447
 3448	/*
 3449	 * rq::ttwu_pending racy indication of out-standing wakeups.
 3450	 * Races such that false-negatives are possible, since they
 3451	 * are shorter lived that false-positives would be.
 3452	 */
 3453	WRITE_ONCE(rq->ttwu_pending, 0);
 3454
 3455	rq_lock_irqsave(rq, &rf);
 3456	update_rq_clock(rq);
 3457
 3458	llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
 3459		if (WARN_ON_ONCE(p->on_cpu))
 3460			smp_cond_load_acquire(&p->on_cpu, !VAL);
 3461
 3462		if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
 3463			set_task_cpu(p, cpu_of(rq));
 3464
 3465		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
 3466	}
 3467
 
 
 
 
 
 
 
 
 
 
 
 3468	rq_unlock_irqrestore(rq, &rf);
 3469}
 3470
 3471void send_call_function_single_ipi(int cpu)
 
 
 
 
 
 
 3472{
 3473	struct rq *rq = cpu_rq(cpu);
 
 
 
 3474
 3475	if (!set_nr_if_polling(rq->idle))
 3476		arch_send_call_function_single_ipi(cpu);
 3477	else
 3478		trace_sched_wake_idle_without_ipi(cpu);
 3479}
 3480
 3481/*
 3482 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
 3483 * necessary. The wakee CPU on receipt of the IPI will queue the task
 3484 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
 3485 * of the wakeup instead of the waker.
 3486 */
 3487static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
 3488{
 3489	struct rq *rq = cpu_rq(cpu);
 3490
 3491	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
 3492
 3493	WRITE_ONCE(rq->ttwu_pending, 1);
 3494	__smp_call_single_queue(cpu, &p->wake_entry.llist);
 3495}
 3496
 3497void wake_up_if_idle(int cpu)
 3498{
 3499	struct rq *rq = cpu_rq(cpu);
 3500	struct rq_flags rf;
 3501
 3502	rcu_read_lock();
 
 
 
 
 
 
 3503
 3504	if (!is_idle_task(rcu_dereference(rq->curr)))
 3505		goto out;
 
 
 3506
 3507	if (set_nr_if_polling(rq->idle)) {
 3508		trace_sched_wake_idle_without_ipi(cpu);
 3509	} else {
 3510		rq_lock_irqsave(rq, &rf);
 3511		if (is_idle_task(rq->curr))
 3512			smp_send_reschedule(cpu);
 3513		/* Else CPU is not idle, do nothing here: */
 3514		rq_unlock_irqrestore(rq, &rf);
 3515	}
 3516
 3517out:
 3518	rcu_read_unlock();
 3519}
 3520
 3521bool cpus_share_cache(int this_cpu, int that_cpu)
 3522{
 
 
 
 3523	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
 3524}
 3525
 3526static inline bool ttwu_queue_cond(int cpu, int wake_flags)
 
 
 
 
 
 
 
 
 
 
 
 
 3527{
 3528	/*
 
 
 
 
 
 
 
 
 
 3529	 * Do not complicate things with the async wake_list while the CPU is
 3530	 * in hotplug state.
 3531	 */
 3532	if (!cpu_active(cpu))
 3533		return false;
 3534
 
 
 
 
 3535	/*
 3536	 * If the CPU does not share cache, then queue the task on the
 3537	 * remote rqs wakelist to avoid accessing remote data.
 3538	 */
 3539	if (!cpus_share_cache(smp_processor_id(), cpu))
 3540		return true;
 3541
 
 
 
 3542	/*
 3543	 * If the task is descheduling and the only running task on the
 3544	 * CPU then use the wakelist to offload the task activation to
 3545	 * the soon-to-be-idle CPU as the current CPU is likely busy.
 3546	 * nr_running is checked to avoid unnecessary task stacking.
 
 
 
 
 
 3547	 */
 3548	if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
 3549		return true;
 3550
 3551	return false;
 3552}
 3553
 3554static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
 3555{
 3556	if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
 3557		if (WARN_ON_ONCE(cpu == smp_processor_id()))
 3558			return false;
 3559
 3560		sched_clock_cpu(cpu); /* Sync clocks across CPUs */
 3561		__ttwu_queue_wakelist(p, cpu, wake_flags);
 3562		return true;
 3563	}
 3564
 3565	return false;
 3566}
 3567
 3568#else /* !CONFIG_SMP */
 3569
 3570static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
 3571{
 3572	return false;
 3573}
 3574
 3575#endif /* CONFIG_SMP */
 3576
 3577static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
 3578{
 3579	struct rq *rq = cpu_rq(cpu);
 3580	struct rq_flags rf;
 3581
 3582	if (ttwu_queue_wakelist(p, cpu, wake_flags))
 3583		return;
 3584
 3585	rq_lock(rq, &rf);
 3586	update_rq_clock(rq);
 3587	ttwu_do_activate(rq, p, wake_flags, &rf);
 3588	rq_unlock(rq, &rf);
 3589}
 3590
 3591/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 3592 * Notes on Program-Order guarantees on SMP systems.
 3593 *
 3594 *  MIGRATION
 3595 *
 3596 * The basic program-order guarantee on SMP systems is that when a task [t]
 3597 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
 3598 * execution on its new CPU [c1].
 3599 *
 3600 * For migration (of runnable tasks) this is provided by the following means:
 3601 *
 3602 *  A) UNLOCK of the rq(c0)->lock scheduling out task t
 3603 *  B) migration for t is required to synchronize *both* rq(c0)->lock and
 3604 *     rq(c1)->lock (if not at the same time, then in that order).
 3605 *  C) LOCK of the rq(c1)->lock scheduling in task
 3606 *
 3607 * Release/acquire chaining guarantees that B happens after A and C after B.
 3608 * Note: the CPU doing B need not be c0 or c1
 3609 *
 3610 * Example:
 3611 *
 3612 *   CPU0            CPU1            CPU2
 3613 *
 3614 *   LOCK rq(0)->lock
 3615 *   sched-out X
 3616 *   sched-in Y
 3617 *   UNLOCK rq(0)->lock
 3618 *
 3619 *                                   LOCK rq(0)->lock // orders against CPU0
 3620 *                                   dequeue X
 3621 *                                   UNLOCK rq(0)->lock
 3622 *
 3623 *                                   LOCK rq(1)->lock
 3624 *                                   enqueue X
 3625 *                                   UNLOCK rq(1)->lock
 3626 *
 3627 *                   LOCK rq(1)->lock // orders against CPU2
 3628 *                   sched-out Z
 3629 *                   sched-in X
 3630 *                   UNLOCK rq(1)->lock
 3631 *
 3632 *
 3633 *  BLOCKING -- aka. SLEEP + WAKEUP
 3634 *
 3635 * For blocking we (obviously) need to provide the same guarantee as for
 3636 * migration. However the means are completely different as there is no lock
 3637 * chain to provide order. Instead we do:
 3638 *
 3639 *   1) smp_store_release(X->on_cpu, 0)   -- finish_task()
 3640 *   2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
 3641 *
 3642 * Example:
 3643 *
 3644 *   CPU0 (schedule)  CPU1 (try_to_wake_up) CPU2 (schedule)
 3645 *
 3646 *   LOCK rq(0)->lock LOCK X->pi_lock
 3647 *   dequeue X
 3648 *   sched-out X
 3649 *   smp_store_release(X->on_cpu, 0);
 3650 *
 3651 *                    smp_cond_load_acquire(&X->on_cpu, !VAL);
 3652 *                    X->state = WAKING
 3653 *                    set_task_cpu(X,2)
 3654 *
 3655 *                    LOCK rq(2)->lock
 3656 *                    enqueue X
 3657 *                    X->state = RUNNING
 3658 *                    UNLOCK rq(2)->lock
 3659 *
 3660 *                                          LOCK rq(2)->lock // orders against CPU1
 3661 *                                          sched-out Z
 3662 *                                          sched-in X
 3663 *                                          UNLOCK rq(2)->lock
 3664 *
 3665 *                    UNLOCK X->pi_lock
 3666 *   UNLOCK rq(0)->lock
 3667 *
 3668 *
 3669 * However, for wakeups there is a second guarantee we must provide, namely we
 3670 * must ensure that CONDITION=1 done by the caller can not be reordered with
 3671 * accesses to the task state; see try_to_wake_up() and set_current_state().
 3672 */
 3673
 3674/**
 3675 * try_to_wake_up - wake up a thread
 3676 * @p: the thread to be awakened
 3677 * @state: the mask of task states that can be woken
 3678 * @wake_flags: wake modifier flags (WF_*)
 3679 *
 3680 * Conceptually does:
 3681 *
 3682 *   If (@state & @p->state) @p->state = TASK_RUNNING.
 3683 *
 3684 * If the task was not queued/runnable, also place it back on a runqueue.
 3685 *
 3686 * This function is atomic against schedule() which would dequeue the task.
 3687 *
 3688 * It issues a full memory barrier before accessing @p->state, see the comment
 3689 * with set_current_state().
 3690 *
 3691 * Uses p->pi_lock to serialize against concurrent wake-ups.
 3692 *
 3693 * Relies on p->pi_lock stabilizing:
 3694 *  - p->sched_class
 3695 *  - p->cpus_ptr
 3696 *  - p->sched_task_group
 3697 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
 3698 *
 3699 * Tries really hard to only take one task_rq(p)->lock for performance.
 3700 * Takes rq->lock in:
 3701 *  - ttwu_runnable()    -- old rq, unavoidable, see comment there;
 3702 *  - ttwu_queue()       -- new rq, for enqueue of the task;
 3703 *  - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
 3704 *
 3705 * As a consequence we race really badly with just about everything. See the
 3706 * many memory barriers and their comments for details.
 3707 *
 3708 * Return: %true if @p->state changes (an actual wakeup was done),
 3709 *	   %false otherwise.
 3710 */
 3711static int
 3712try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
 3713{
 3714	unsigned long flags;
 3715	int cpu, success = 0;
 3716
 3717	preempt_disable();
 
 3718	if (p == current) {
 3719		/*
 3720		 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
 3721		 * == smp_processor_id()'. Together this means we can special
 3722		 * case the whole 'p->on_rq && ttwu_runnable()' case below
 3723		 * without taking any locks.
 3724		 *
 
 
 
 
 3725		 * In particular:
 3726		 *  - we rely on Program-Order guarantees for all the ordering,
 3727		 *  - we're serialized against set_special_state() by virtue of
 3728		 *    it disabling IRQs (this allows not taking ->pi_lock).
 3729		 */
 3730		if (!(READ_ONCE(p->__state) & state))
 
 3731			goto out;
 3732
 3733		success = 1;
 3734		trace_sched_waking(p);
 3735		WRITE_ONCE(p->__state, TASK_RUNNING);
 3736		trace_sched_wakeup(p);
 3737		goto out;
 3738	}
 3739
 3740	/*
 3741	 * If we are going to wake up a thread waiting for CONDITION we
 3742	 * need to ensure that CONDITION=1 done by the caller can not be
 3743	 * reordered with p->state check below. This pairs with smp_store_mb()
 3744	 * in set_current_state() that the waiting thread does.
 3745	 */
 3746	raw_spin_lock_irqsave(&p->pi_lock, flags);
 3747	smp_mb__after_spinlock();
 3748	if (!(READ_ONCE(p->__state) & state))
 3749		goto unlock;
 3750
 3751	trace_sched_waking(p);
 3752
 3753	/* We're going to change ->state: */
 3754	success = 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 3755
 3756	/*
 3757	 * Ensure we load p->on_rq _after_ p->state, otherwise it would
 3758	 * be possible to, falsely, observe p->on_rq == 0 and get stuck
 3759	 * in smp_cond_load_acquire() below.
 3760	 *
 3761	 * sched_ttwu_pending()			try_to_wake_up()
 3762	 *   STORE p->on_rq = 1			  LOAD p->state
 3763	 *   UNLOCK rq->lock
 3764	 *
 3765	 * __schedule() (switch to task 'p')
 3766	 *   LOCK rq->lock			  smp_rmb();
 3767	 *   smp_mb__after_spinlock();
 3768	 *   UNLOCK rq->lock
 3769	 *
 3770	 * [task p]
 3771	 *   STORE p->state = UNINTERRUPTIBLE	  LOAD p->on_rq
 3772	 *
 3773	 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
 3774	 * __schedule().  See the comment for smp_mb__after_spinlock().
 3775	 *
 3776	 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
 3777	 */
 3778	smp_rmb();
 3779	if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
 3780		goto unlock;
 3781
 3782#ifdef CONFIG_SMP
 3783	/*
 3784	 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
 3785	 * possible to, falsely, observe p->on_cpu == 0.
 3786	 *
 3787	 * One must be running (->on_cpu == 1) in order to remove oneself
 3788	 * from the runqueue.
 3789	 *
 3790	 * __schedule() (switch to task 'p')	try_to_wake_up()
 3791	 *   STORE p->on_cpu = 1		  LOAD p->on_rq
 3792	 *   UNLOCK rq->lock
 3793	 *
 3794	 * __schedule() (put 'p' to sleep)
 3795	 *   LOCK rq->lock			  smp_rmb();
 3796	 *   smp_mb__after_spinlock();
 3797	 *   STORE p->on_rq = 0			  LOAD p->on_cpu
 3798	 *
 3799	 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
 3800	 * __schedule().  See the comment for smp_mb__after_spinlock().
 3801	 *
 3802	 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
 3803	 * schedule()'s deactivate_task() has 'happened' and p will no longer
 3804	 * care about it's own p->state. See the comment in __schedule().
 3805	 */
 3806	smp_acquire__after_ctrl_dep();
 3807
 3808	/*
 3809	 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
 3810	 * == 0), which means we need to do an enqueue, change p->state to
 3811	 * TASK_WAKING such that we can unlock p->pi_lock before doing the
 3812	 * enqueue, such as ttwu_queue_wakelist().
 3813	 */
 3814	WRITE_ONCE(p->__state, TASK_WAKING);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 3815
 3816	/*
 3817	 * If the owning (remote) CPU is still in the middle of schedule() with
 3818	 * this task as prev, considering queueing p on the remote CPUs wake_list
 3819	 * which potentially sends an IPI instead of spinning on p->on_cpu to
 3820	 * let the waker make forward progress. This is safe because IRQs are
 3821	 * disabled and the IPI will deliver after on_cpu is cleared.
 3822	 *
 3823	 * Ensure we load task_cpu(p) after p->on_cpu:
 3824	 *
 3825	 * set_task_cpu(p, cpu);
 3826	 *   STORE p->cpu = @cpu
 3827	 * __schedule() (switch to task 'p')
 3828	 *   LOCK rq->lock
 3829	 *   smp_mb__after_spin_lock()		smp_cond_load_acquire(&p->on_cpu)
 3830	 *   STORE p->on_cpu = 1		LOAD p->cpu
 3831	 *
 3832	 * to ensure we observe the correct CPU on which the task is currently
 3833	 * scheduling.
 3834	 */
 3835	if (smp_load_acquire(&p->on_cpu) &&
 3836	    ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
 3837		goto unlock;
 3838
 3839	/*
 3840	 * If the owning (remote) CPU is still in the middle of schedule() with
 3841	 * this task as prev, wait until it's done referencing the task.
 3842	 *
 3843	 * Pairs with the smp_store_release() in finish_task().
 3844	 *
 3845	 * This ensures that tasks getting woken will be fully ordered against
 3846	 * their previous state and preserve Program Order.
 3847	 */
 3848	smp_cond_load_acquire(&p->on_cpu, !VAL);
 3849
 3850	cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
 3851	if (task_cpu(p) != cpu) {
 3852		if (p->in_iowait) {
 3853			delayacct_blkio_end(p);
 3854			atomic_dec(&task_rq(p)->nr_iowait);
 3855		}
 3856
 3857		wake_flags |= WF_MIGRATED;
 3858		psi_ttwu_dequeue(p);
 3859		set_task_cpu(p, cpu);
 3860	}
 3861#else
 3862	cpu = task_cpu(p);
 3863#endif /* CONFIG_SMP */
 3864
 3865	ttwu_queue(p, cpu, wake_flags);
 3866unlock:
 3867	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
 3868out:
 3869	if (success)
 3870		ttwu_stat(p, task_cpu(p), wake_flags);
 3871	preempt_enable();
 3872
 3873	return success;
 3874}
 3875
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 3876/**
 3877 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
 3878 * @p: Process for which the function is to be invoked, can be @current.
 3879 * @func: Function to invoke.
 3880 * @arg: Argument to function.
 3881 *
 3882 * If the specified task can be quickly locked into a definite state
 3883 * (either sleeping or on a given runqueue), arrange to keep it in that
 3884 * state while invoking @func(@arg).  This function can use ->on_rq and
 3885 * task_curr() to work out what the state is, if required.  Given that
 3886 * @func can be invoked with a runqueue lock held, it had better be quite
 3887 * lightweight.
 3888 *
 3889 * Returns:
 3890 *	@false if the task slipped out from under the locks.
 3891 *	@true if the task was locked onto a runqueue or is sleeping.
 3892 *		However, @func can override this by returning @false.
 3893 */
 3894bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
 3895{
 
 3896	struct rq_flags rf;
 3897	bool ret = false;
 3898	struct rq *rq;
 3899
 3900	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
 3901	if (p->on_rq) {
 
 3902		rq = __task_rq_lock(p, &rf);
 3903		if (task_rq(p) == rq)
 3904			ret = func(p, arg);
 
 
 
 
 
 
 
 
 
 
 
 
 3905		rq_unlock(rq, &rf);
 3906	} else {
 3907		switch (READ_ONCE(p->__state)) {
 3908		case TASK_RUNNING:
 3909		case TASK_WAKING:
 3910			break;
 3911		default:
 3912			smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
 3913			if (!p->on_rq)
 3914				ret = func(p, arg);
 3915		}
 3916	}
 3917	raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
 3918	return ret;
 3919}
 3920
 3921/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 3922 * wake_up_process - Wake up a specific process
 3923 * @p: The process to be woken up.
 3924 *
 3925 * Attempt to wake up the nominated process and move it to the set of runnable
 3926 * processes.
 3927 *
 3928 * Return: 1 if the process was woken up, 0 if it was already running.
 3929 *
 3930 * This function executes a full memory barrier before accessing the task state.
 3931 */
 3932int wake_up_process(struct task_struct *p)
 3933{
 3934	return try_to_wake_up(p, TASK_NORMAL, 0);
 3935}
 3936EXPORT_SYMBOL(wake_up_process);
 3937
 3938int wake_up_state(struct task_struct *p, unsigned int state)
 3939{
 3940	return try_to_wake_up(p, state, 0);
 3941}
 3942
 3943/*
 3944 * Perform scheduler related setup for a newly forked process p.
 3945 * p is forked by current.
 3946 *
 3947 * __sched_fork() is basic setup used by init_idle() too:
 
 3948 */
 3949static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
 3950{
 3951	p->on_rq			= 0;
 3952
 3953	p->se.on_rq			= 0;
 3954	p->se.exec_start		= 0;
 3955	p->se.sum_exec_runtime		= 0;
 3956	p->se.prev_sum_exec_runtime	= 0;
 3957	p->se.nr_migrations		= 0;
 3958	p->se.vruntime			= 0;
 
 3959	INIT_LIST_HEAD(&p->se.group_node);
 3960
 
 
 
 3961#ifdef CONFIG_FAIR_GROUP_SCHED
 3962	p->se.cfs_rq			= NULL;
 3963#endif
 3964
 3965#ifdef CONFIG_SCHEDSTATS
 3966	/* Even if schedstat is disabled, there should not be garbage */
 3967	memset(&p->se.statistics, 0, sizeof(p->se.statistics));
 3968#endif
 3969
 3970	RB_CLEAR_NODE(&p->dl.rb_node);
 3971	init_dl_task_timer(&p->dl);
 3972	init_dl_inactive_task_timer(&p->dl);
 3973	__dl_clear_params(p);
 3974
 3975	INIT_LIST_HEAD(&p->rt.run_list);
 3976	p->rt.timeout		= 0;
 3977	p->rt.time_slice	= sched_rr_timeslice;
 3978	p->rt.on_rq		= 0;
 3979	p->rt.on_list		= 0;
 3980
 
 
 
 
 3981#ifdef CONFIG_PREEMPT_NOTIFIERS
 3982	INIT_HLIST_HEAD(&p->preempt_notifiers);
 3983#endif
 3984
 3985#ifdef CONFIG_COMPACTION
 3986	p->capture_control = NULL;
 3987#endif
 3988	init_numa_balancing(clone_flags, p);
 3989#ifdef CONFIG_SMP
 3990	p->wake_entry.u_flags = CSD_TYPE_TTWU;
 3991	p->migration_pending = NULL;
 3992#endif
 
 3993}
 3994
 3995DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
 3996
 3997#ifdef CONFIG_NUMA_BALANCING
 3998
 3999void set_numabalancing_state(bool enabled)
 
 
 4000{
 4001	if (enabled)
 4002		static_branch_enable(&sched_numa_balancing);
 4003	else
 4004		static_branch_disable(&sched_numa_balancing);
 4005}
 4006
 
 
 
 
 
 
 
 
 
 4007#ifdef CONFIG_PROC_SYSCTL
 4008int sysctl_numa_balancing(struct ctl_table *table, int write,
 
 
 
 
 
 
 
 
 
 
 
 4009			  void *buffer, size_t *lenp, loff_t *ppos)
 4010{
 4011	struct ctl_table t;
 4012	int err;
 4013	int state = static_branch_likely(&sched_numa_balancing);
 4014
 4015	if (write && !capable(CAP_SYS_ADMIN))
 4016		return -EPERM;
 4017
 4018	t = *table;
 4019	t.data = &state;
 4020	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
 4021	if (err < 0)
 4022		return err;
 4023	if (write)
 4024		set_numabalancing_state(state);
 
 
 
 
 
 4025	return err;
 4026}
 4027#endif
 4028#endif
 4029
 4030#ifdef CONFIG_SCHEDSTATS
 4031
 4032DEFINE_STATIC_KEY_FALSE(sched_schedstats);
 4033
 4034static void set_schedstats(bool enabled)
 4035{
 4036	if (enabled)
 4037		static_branch_enable(&sched_schedstats);
 4038	else
 4039		static_branch_disable(&sched_schedstats);
 4040}
 4041
 4042void force_schedstat_enabled(void)
 4043{
 4044	if (!schedstat_enabled()) {
 4045		pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
 4046		static_branch_enable(&sched_schedstats);
 4047	}
 4048}
 4049
 4050static int __init setup_schedstats(char *str)
 4051{
 4052	int ret = 0;
 4053	if (!str)
 4054		goto out;
 4055
 4056	if (!strcmp(str, "enable")) {
 4057		set_schedstats(true);
 4058		ret = 1;
 4059	} else if (!strcmp(str, "disable")) {
 4060		set_schedstats(false);
 4061		ret = 1;
 4062	}
 4063out:
 4064	if (!ret)
 4065		pr_warn("Unable to parse schedstats=\n");
 4066
 4067	return ret;
 4068}
 4069__setup("schedstats=", setup_schedstats);
 4070
 4071#ifdef CONFIG_PROC_SYSCTL
 4072int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
 4073		size_t *lenp, loff_t *ppos)
 4074{
 4075	struct ctl_table t;
 4076	int err;
 4077	int state = static_branch_likely(&sched_schedstats);
 4078
 4079	if (write && !capable(CAP_SYS_ADMIN))
 4080		return -EPERM;
 4081
 4082	t = *table;
 4083	t.data = &state;
 4084	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
 4085	if (err < 0)
 4086		return err;
 4087	if (write)
 4088		set_schedstats(state);
 4089	return err;
 4090}
 4091#endif /* CONFIG_PROC_SYSCTL */
 4092#endif /* CONFIG_SCHEDSTATS */
 4093
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 4094/*
 4095 * fork()/clone()-time setup:
 4096 */
 4097int sched_fork(unsigned long clone_flags, struct task_struct *p)
 4098{
 4099	unsigned long flags;
 4100
 4101	__sched_fork(clone_flags, p);
 4102	/*
 4103	 * We mark the process as NEW here. This guarantees that
 4104	 * nobody will actually run it, and a signal or other external
 4105	 * event cannot wake it up and insert it on the runqueue either.
 4106	 */
 4107	p->__state = TASK_NEW;
 4108
 4109	/*
 4110	 * Make sure we do not leak PI boosting priority to the child.
 4111	 */
 4112	p->prio = current->normal_prio;
 4113
 4114	uclamp_fork(p);
 4115
 4116	/*
 4117	 * Revert to default priority/policy on fork if requested.
 4118	 */
 4119	if (unlikely(p->sched_reset_on_fork)) {
 4120		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
 4121			p->policy = SCHED_NORMAL;
 4122			p->static_prio = NICE_TO_PRIO(0);
 4123			p->rt_priority = 0;
 4124		} else if (PRIO_TO_NICE(p->static_prio) < 0)
 4125			p->static_prio = NICE_TO_PRIO(0);
 4126
 4127		p->prio = p->normal_prio = p->static_prio;
 4128		set_load_weight(p, false);
 
 
 4129
 4130		/*
 4131		 * We don't need the reset flag anymore after the fork. It has
 4132		 * fulfilled its duty:
 4133		 */
 4134		p->sched_reset_on_fork = 0;
 4135	}
 4136
 4137	if (dl_prio(p->prio))
 4138		return -EAGAIN;
 4139	else if (rt_prio(p->prio))
 
 
 
 4140		p->sched_class = &rt_sched_class;
 4141	else
 
 
 
 
 4142		p->sched_class = &fair_sched_class;
 
 4143
 4144	init_entity_runnable_average(&p->se);
 4145
 4146	/*
 4147	 * The child is not yet in the pid-hash so no cgroup attach races,
 4148	 * and the cgroup is pinned to this child due to cgroup_fork()
 4149	 * is ran before sched_fork().
 4150	 *
 4151	 * Silence PROVE_RCU.
 4152	 */
 4153	raw_spin_lock_irqsave(&p->pi_lock, flags);
 4154	rseq_migrate(p);
 4155	/*
 4156	 * We're setting the CPU for the first time, we don't migrate,
 4157	 * so use __set_task_cpu().
 4158	 */
 4159	__set_task_cpu(p, smp_processor_id());
 4160	if (p->sched_class->task_fork)
 4161		p->sched_class->task_fork(p);
 4162	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
 4163
 4164#ifdef CONFIG_SCHED_INFO
 4165	if (likely(sched_info_on()))
 4166		memset(&p->sched_info, 0, sizeof(p->sched_info));
 4167#endif
 4168#if defined(CONFIG_SMP)
 4169	p->on_cpu = 0;
 4170#endif
 4171	init_task_preempt_count(p);
 4172#ifdef CONFIG_SMP
 4173	plist_node_init(&p->pushable_tasks, MAX_PRIO);
 4174	RB_CLEAR_NODE(&p->pushable_dl_tasks);
 4175#endif
 4176	return 0;
 4177}
 4178
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 4179void sched_post_fork(struct task_struct *p)
 4180{
 4181	uclamp_post_fork(p);
 
 4182}
 4183
 4184unsigned long to_ratio(u64 period, u64 runtime)
 4185{
 4186	if (runtime == RUNTIME_INF)
 4187		return BW_UNIT;
 4188
 4189	/*
 4190	 * Doing this here saves a lot of checks in all
 4191	 * the calling paths, and returning zero seems
 4192	 * safe for them anyway.
 4193	 */
 4194	if (period == 0)
 4195		return 0;
 4196
 4197	return div64_u64(runtime << BW_SHIFT, period);
 4198}
 4199
 4200/*
 4201 * wake_up_new_task - wake up a newly created task for the first time.
 4202 *
 4203 * This function will do some initial scheduler statistics housekeeping
 4204 * that must be done for every newly created context, then puts the task
 4205 * on the runqueue and wakes it.
 4206 */
 4207void wake_up_new_task(struct task_struct *p)
 4208{
 4209	struct rq_flags rf;
 4210	struct rq *rq;
 
 4211
 4212	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
 4213	WRITE_ONCE(p->__state, TASK_RUNNING);
 4214#ifdef CONFIG_SMP
 4215	/*
 4216	 * Fork balancing, do it here and not earlier because:
 4217	 *  - cpus_ptr can change in the fork path
 4218	 *  - any previously selected CPU might disappear through hotplug
 4219	 *
 4220	 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
 4221	 * as we're not fully set-up yet.
 4222	 */
 4223	p->recent_used_cpu = task_cpu(p);
 4224	rseq_migrate(p);
 4225	__set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
 4226#endif
 4227	rq = __task_rq_lock(p, &rf);
 4228	update_rq_clock(rq);
 4229	post_init_entity_util_avg(p);
 4230
 4231	activate_task(rq, p, ENQUEUE_NOCLOCK);
 4232	trace_sched_wakeup_new(p);
 4233	check_preempt_curr(rq, p, WF_FORK);
 4234#ifdef CONFIG_SMP
 4235	if (p->sched_class->task_woken) {
 4236		/*
 4237		 * Nothing relies on rq->lock after this, so it's fine to
 4238		 * drop it.
 4239		 */
 4240		rq_unpin_lock(rq, &rf);
 4241		p->sched_class->task_woken(rq, p);
 4242		rq_repin_lock(rq, &rf);
 4243	}
 4244#endif
 4245	task_rq_unlock(rq, p, &rf);
 4246}
 4247
 4248#ifdef CONFIG_PREEMPT_NOTIFIERS
 4249
 4250static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
 4251
 4252void preempt_notifier_inc(void)
 4253{
 4254	static_branch_inc(&preempt_notifier_key);
 4255}
 4256EXPORT_SYMBOL_GPL(preempt_notifier_inc);
 4257
 4258void preempt_notifier_dec(void)
 4259{
 4260	static_branch_dec(&preempt_notifier_key);
 4261}
 4262EXPORT_SYMBOL_GPL(preempt_notifier_dec);
 4263
 4264/**
 4265 * preempt_notifier_register - tell me when current is being preempted & rescheduled
 4266 * @notifier: notifier struct to register
 4267 */
 4268void preempt_notifier_register(struct preempt_notifier *notifier)
 4269{
 4270	if (!static_branch_unlikely(&preempt_notifier_key))
 4271		WARN(1, "registering preempt_notifier while notifiers disabled\n");
 4272
 4273	hlist_add_head(&notifier->link, &current->preempt_notifiers);
 4274}
 4275EXPORT_SYMBOL_GPL(preempt_notifier_register);
 4276
 4277/**
 4278 * preempt_notifier_unregister - no longer interested in preemption notifications
 4279 * @notifier: notifier struct to unregister
 4280 *
 4281 * This is *not* safe to call from within a preemption notifier.
 4282 */
 4283void preempt_notifier_unregister(struct preempt_notifier *notifier)
 4284{
 4285	hlist_del(&notifier->link);
 4286}
 4287EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
 4288
 4289static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
 4290{
 4291	struct preempt_notifier *notifier;
 4292
 4293	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
 4294		notifier->ops->sched_in(notifier, raw_smp_processor_id());
 4295}
 4296
 4297static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
 4298{
 4299	if (static_branch_unlikely(&preempt_notifier_key))
 4300		__fire_sched_in_preempt_notifiers(curr);
 4301}
 4302
 4303static void
 4304__fire_sched_out_preempt_notifiers(struct task_struct *curr,
 4305				   struct task_struct *next)
 4306{
 4307	struct preempt_notifier *notifier;
 4308
 4309	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
 4310		notifier->ops->sched_out(notifier, next);
 4311}
 4312
 4313static __always_inline void
 4314fire_sched_out_preempt_notifiers(struct task_struct *curr,
 4315				 struct task_struct *next)
 4316{
 4317	if (static_branch_unlikely(&preempt_notifier_key))
 4318		__fire_sched_out_preempt_notifiers(curr, next);
 4319}
 4320
 4321#else /* !CONFIG_PREEMPT_NOTIFIERS */
 4322
 4323static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
 4324{
 4325}
 4326
 4327static inline void
 4328fire_sched_out_preempt_notifiers(struct task_struct *curr,
 4329				 struct task_struct *next)
 4330{
 4331}
 4332
 4333#endif /* CONFIG_PREEMPT_NOTIFIERS */
 4334
 4335static inline void prepare_task(struct task_struct *next)
 4336{
 4337#ifdef CONFIG_SMP
 4338	/*
 4339	 * Claim the task as running, we do this before switching to it
 4340	 * such that any running task will have this set.
 4341	 *
 4342	 * See the ttwu() WF_ON_CPU case and its ordering comment.
 
 4343	 */
 4344	WRITE_ONCE(next->on_cpu, 1);
 4345#endif
 4346}
 4347
 4348static inline void finish_task(struct task_struct *prev)
 4349{
 4350#ifdef CONFIG_SMP
 4351	/*
 4352	 * This must be the very last reference to @prev from this CPU. After
 4353	 * p->on_cpu is cleared, the task can be moved to a different CPU. We
 4354	 * must ensure this doesn't happen until the switch is completely
 4355	 * finished.
 4356	 *
 4357	 * In particular, the load of prev->state in finish_task_switch() must
 4358	 * happen before this.
 4359	 *
 4360	 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
 4361	 */
 4362	smp_store_release(&prev->on_cpu, 0);
 4363#endif
 4364}
 4365
 4366#ifdef CONFIG_SMP
 4367
 4368static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
 4369{
 4370	void (*func)(struct rq *rq);
 4371	struct callback_head *next;
 4372
 4373	lockdep_assert_rq_held(rq);
 4374
 4375	while (head) {
 4376		func = (void (*)(struct rq *))head->func;
 4377		next = head->next;
 4378		head->next = NULL;
 4379		head = next;
 4380
 4381		func(rq);
 4382	}
 4383}
 4384
 4385static void balance_push(struct rq *rq);
 4386
 4387struct callback_head balance_push_callback = {
 
 
 
 
 
 
 
 
 
 
 
 4388	.next = NULL,
 4389	.func = (void (*)(struct callback_head *))balance_push,
 4390};
 4391
 4392static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
 
 4393{
 4394	struct callback_head *head = rq->balance_callback;
 
 
 
 4395
 4396	lockdep_assert_rq_held(rq);
 4397	if (head)
 
 
 
 
 
 
 
 
 
 
 4398		rq->balance_callback = NULL;
 4399
 4400	return head;
 4401}
 4402
 
 
 
 
 
 4403static void __balance_callbacks(struct rq *rq)
 4404{
 4405	do_balance_callbacks(rq, splice_balance_callbacks(rq));
 4406}
 4407
 4408static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
 4409{
 4410	unsigned long flags;
 4411
 4412	if (unlikely(head)) {
 4413		raw_spin_rq_lock_irqsave(rq, flags);
 4414		do_balance_callbacks(rq, head);
 4415		raw_spin_rq_unlock_irqrestore(rq, flags);
 4416	}
 4417}
 4418
 4419#else
 4420
 4421static inline void __balance_callbacks(struct rq *rq)
 4422{
 4423}
 4424
 4425static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
 4426{
 4427	return NULL;
 4428}
 4429
 4430static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
 4431{
 4432}
 4433
 4434#endif
 4435
 4436static inline void
 4437prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
 4438{
 4439	/*
 4440	 * Since the runqueue lock will be released by the next
 4441	 * task (which is an invalid locking op but in the case
 4442	 * of the scheduler it's an obvious special-case), so we
 4443	 * do an early lockdep release here:
 4444	 */
 4445	rq_unpin_lock(rq, rf);
 4446	spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
 4447#ifdef CONFIG_DEBUG_SPINLOCK
 4448	/* this is a valid case when another task releases the spinlock */
 4449	rq_lockp(rq)->owner = next;
 4450#endif
 4451}
 4452
 4453static inline void finish_lock_switch(struct rq *rq)
 4454{
 4455	/*
 4456	 * If we are tracking spinlock dependencies then we have to
 4457	 * fix up the runqueue lock - which gets 'carried over' from
 4458	 * prev into current:
 4459	 */
 4460	spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
 4461	__balance_callbacks(rq);
 4462	raw_spin_rq_unlock_irq(rq);
 4463}
 4464
 4465/*
 4466 * NOP if the arch has not defined these:
 4467 */
 4468
 4469#ifndef prepare_arch_switch
 4470# define prepare_arch_switch(next)	do { } while (0)
 4471#endif
 4472
 4473#ifndef finish_arch_post_lock_switch
 4474# define finish_arch_post_lock_switch()	do { } while (0)
 4475#endif
 4476
 4477static inline void kmap_local_sched_out(void)
 4478{
 4479#ifdef CONFIG_KMAP_LOCAL
 4480	if (unlikely(current->kmap_ctrl.idx))
 4481		__kmap_local_sched_out();
 4482#endif
 4483}
 4484
 4485static inline void kmap_local_sched_in(void)
 4486{
 4487#ifdef CONFIG_KMAP_LOCAL
 4488	if (unlikely(current->kmap_ctrl.idx))
 4489		__kmap_local_sched_in();
 4490#endif
 4491}
 4492
 4493/**
 4494 * prepare_task_switch - prepare to switch tasks
 4495 * @rq: the runqueue preparing to switch
 4496 * @prev: the current task that is being switched out
 4497 * @next: the task we are going to switch to.
 4498 *
 4499 * This is called with the rq lock held and interrupts off. It must
 4500 * be paired with a subsequent finish_task_switch after the context
 4501 * switch.
 4502 *
 4503 * prepare_task_switch sets up locking and calls architecture specific
 4504 * hooks.
 4505 */
 4506static inline void
 4507prepare_task_switch(struct rq *rq, struct task_struct *prev,
 4508		    struct task_struct *next)
 4509{
 4510	kcov_prepare_switch(prev);
 4511	sched_info_switch(rq, prev, next);
 4512	perf_event_task_sched_out(prev, next);
 4513	rseq_preempt(prev);
 4514	fire_sched_out_preempt_notifiers(prev, next);
 4515	kmap_local_sched_out();
 4516	prepare_task(next);
 4517	prepare_arch_switch(next);
 4518}
 4519
 4520/**
 4521 * finish_task_switch - clean up after a task-switch
 4522 * @prev: the thread we just switched away from.
 4523 *
 4524 * finish_task_switch must be called after the context switch, paired
 4525 * with a prepare_task_switch call before the context switch.
 4526 * finish_task_switch will reconcile locking set up by prepare_task_switch,
 4527 * and do any other architecture-specific cleanup actions.
 4528 *
 4529 * Note that we may have delayed dropping an mm in context_switch(). If
 4530 * so, we finish that here outside of the runqueue lock. (Doing it
 4531 * with the lock held can cause deadlocks; see schedule() for
 4532 * details.)
 4533 *
 4534 * The context switch have flipped the stack from under us and restored the
 4535 * local variables which were saved when this task called schedule() in the
 4536 * past. prev == current is still correct but we need to recalculate this_rq
 4537 * because prev may have moved to another CPU.
 4538 */
 4539static struct rq *finish_task_switch(struct task_struct *prev)
 4540	__releases(rq->lock)
 4541{
 4542	struct rq *rq = this_rq();
 4543	struct mm_struct *mm = rq->prev_mm;
 4544	long prev_state;
 4545
 4546	/*
 4547	 * The previous task will have left us with a preempt_count of 2
 4548	 * because it left us after:
 4549	 *
 4550	 *	schedule()
 4551	 *	  preempt_disable();			// 1
 4552	 *	  __schedule()
 4553	 *	    raw_spin_lock_irq(&rq->lock)	// 2
 4554	 *
 4555	 * Also, see FORK_PREEMPT_COUNT.
 4556	 */
 4557	if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
 4558		      "corrupted preempt_count: %s/%d/0x%x\n",
 4559		      current->comm, current->pid, preempt_count()))
 4560		preempt_count_set(FORK_PREEMPT_COUNT);
 4561
 4562	rq->prev_mm = NULL;
 4563
 4564	/*
 4565	 * A task struct has one reference for the use as "current".
 4566	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
 4567	 * schedule one last time. The schedule call will never return, and
 4568	 * the scheduled task must drop that reference.
 4569	 *
 4570	 * We must observe prev->state before clearing prev->on_cpu (in
 4571	 * finish_task), otherwise a concurrent wakeup can get prev
 4572	 * running on another CPU and we could rave with its RUNNING -> DEAD
 4573	 * transition, resulting in a double drop.
 4574	 */
 4575	prev_state = READ_ONCE(prev->__state);
 4576	vtime_task_switch(prev);
 4577	perf_event_task_sched_in(prev, current);
 4578	finish_task(prev);
 4579	tick_nohz_task_switch();
 4580	finish_lock_switch(rq);
 4581	finish_arch_post_lock_switch();
 4582	kcov_finish_switch(current);
 4583	/*
 4584	 * kmap_local_sched_out() is invoked with rq::lock held and
 4585	 * interrupts disabled. There is no requirement for that, but the
 4586	 * sched out code does not have an interrupt enabled section.
 4587	 * Restoring the maps on sched in does not require interrupts being
 4588	 * disabled either.
 4589	 */
 4590	kmap_local_sched_in();
 4591
 4592	fire_sched_in_preempt_notifiers(current);
 4593	/*
 4594	 * When switching through a kernel thread, the loop in
 4595	 * membarrier_{private,global}_expedited() may have observed that
 4596	 * kernel thread and not issued an IPI. It is therefore possible to
 4597	 * schedule between user->kernel->user threads without passing though
 4598	 * switch_mm(). Membarrier requires a barrier after storing to
 4599	 * rq->curr, before returning to userspace, so provide them here:
 4600	 *
 4601	 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
 4602	 *   provided by mmdrop(),
 4603	 * - a sync_core for SYNC_CORE.
 4604	 */
 4605	if (mm) {
 4606		membarrier_mm_sync_core_before_usermode(mm);
 4607		mmdrop(mm);
 4608	}
 
 4609	if (unlikely(prev_state == TASK_DEAD)) {
 4610		if (prev->sched_class->task_dead)
 4611			prev->sched_class->task_dead(prev);
 4612
 4613		/*
 4614		 * Remove function-return probe instances associated with this
 4615		 * task and put them back on the free list.
 4616		 */
 4617		kprobe_flush_task(prev);
 4618
 4619		/* Task is done with its stack. */
 4620		put_task_stack(prev);
 4621
 4622		put_task_struct_rcu_user(prev);
 4623	}
 4624
 4625	return rq;
 4626}
 4627
 4628/**
 4629 * schedule_tail - first thing a freshly forked thread must call.
 4630 * @prev: the thread we just switched away from.
 4631 */
 4632asmlinkage __visible void schedule_tail(struct task_struct *prev)
 4633	__releases(rq->lock)
 4634{
 4635	/*
 4636	 * New tasks start with FORK_PREEMPT_COUNT, see there and
 4637	 * finish_task_switch() for details.
 4638	 *
 4639	 * finish_task_switch() will drop rq->lock() and lower preempt_count
 4640	 * and the preempt_enable() will end up enabling preemption (on
 4641	 * PREEMPT_COUNT kernels).
 4642	 */
 4643
 4644	finish_task_switch(prev);
 4645	preempt_enable();
 4646
 4647	if (current->set_child_tid)
 4648		put_user(task_pid_vnr(current), current->set_child_tid);
 4649
 4650	calculate_sigpending();
 4651}
 4652
 4653/*
 4654 * context_switch - switch to the new MM and the new thread's register state.
 4655 */
 4656static __always_inline struct rq *
 4657context_switch(struct rq *rq, struct task_struct *prev,
 4658	       struct task_struct *next, struct rq_flags *rf)
 4659{
 4660	prepare_task_switch(rq, prev, next);
 4661
 4662	/*
 4663	 * For paravirt, this is coupled with an exit in switch_to to
 4664	 * combine the page table reload and the switch backend into
 4665	 * one hypercall.
 4666	 */
 4667	arch_start_context_switch(prev);
 4668
 4669	/*
 4670	 * kernel -> kernel   lazy + transfer active
 4671	 *   user -> kernel   lazy + mmgrab() active
 4672	 *
 4673	 * kernel ->   user   switch + mmdrop() active
 4674	 *   user ->   user   switch
 
 
 
 4675	 */
 4676	if (!next->mm) {                                // to kernel
 4677		enter_lazy_tlb(prev->active_mm, next);
 4678
 4679		next->active_mm = prev->active_mm;
 4680		if (prev->mm)                           // from user
 4681			mmgrab(prev->active_mm);
 4682		else
 4683			prev->active_mm = NULL;
 4684	} else {                                        // to user
 4685		membarrier_switch_mm(rq, prev->active_mm, next->mm);
 4686		/*
 4687		 * sys_membarrier() requires an smp_mb() between setting
 4688		 * rq->curr / membarrier_switch_mm() and returning to userspace.
 4689		 *
 4690		 * The below provides this either through switch_mm(), or in
 4691		 * case 'prev->active_mm == next->mm' through
 4692		 * finish_task_switch()'s mmdrop().
 4693		 */
 4694		switch_mm_irqs_off(prev->active_mm, next->mm, next);
 
 4695
 4696		if (!prev->mm) {                        // from kernel
 4697			/* will mmdrop() in finish_task_switch(). */
 4698			rq->prev_mm = prev->active_mm;
 4699			prev->active_mm = NULL;
 4700		}
 4701	}
 4702
 4703	rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
 
 4704
 4705	prepare_lock_switch(rq, next, rf);
 4706
 4707	/* Here we just switch the register state and the stack. */
 4708	switch_to(prev, next, prev);
 4709	barrier();
 4710
 4711	return finish_task_switch(prev);
 4712}
 4713
 4714/*
 4715 * nr_running and nr_context_switches:
 4716 *
 4717 * externally visible scheduler statistics: current number of runnable
 4718 * threads, total number of context switches performed since bootup.
 4719 */
 4720unsigned int nr_running(void)
 4721{
 4722	unsigned int i, sum = 0;
 4723
 4724	for_each_online_cpu(i)
 4725		sum += cpu_rq(i)->nr_running;
 4726
 4727	return sum;
 4728}
 4729
 4730/*
 4731 * Check if only the current task is running on the CPU.
 4732 *
 4733 * Caution: this function does not check that the caller has disabled
 4734 * preemption, thus the result might have a time-of-check-to-time-of-use
 4735 * race.  The caller is responsible to use it correctly, for example:
 4736 *
 4737 * - from a non-preemptible section (of course)
 4738 *
 4739 * - from a thread that is bound to a single CPU
 4740 *
 4741 * - in a loop with very short iterations (e.g. a polling loop)
 4742 */
 4743bool single_task_running(void)
 4744{
 4745	return raw_rq()->nr_running == 1;
 4746}
 4747EXPORT_SYMBOL(single_task_running);
 4748
 
 
 
 
 
 4749unsigned long long nr_context_switches(void)
 4750{
 4751	int i;
 4752	unsigned long long sum = 0;
 4753
 4754	for_each_possible_cpu(i)
 4755		sum += cpu_rq(i)->nr_switches;
 4756
 4757	return sum;
 4758}
 4759
 4760/*
 4761 * Consumers of these two interfaces, like for example the cpuidle menu
 4762 * governor, are using nonsensical data. Preferring shallow idle state selection
 4763 * for a CPU that has IO-wait which might not even end up running the task when
 4764 * it does become runnable.
 4765 */
 4766
 4767unsigned int nr_iowait_cpu(int cpu)
 4768{
 4769	return atomic_read(&cpu_rq(cpu)->nr_iowait);
 4770}
 4771
 4772/*
 4773 * IO-wait accounting, and how it's mostly bollocks (on SMP).
 4774 *
 4775 * The idea behind IO-wait account is to account the idle time that we could
 4776 * have spend running if it were not for IO. That is, if we were to improve the
 4777 * storage performance, we'd have a proportional reduction in IO-wait time.
 4778 *
 4779 * This all works nicely on UP, where, when a task blocks on IO, we account
 4780 * idle time as IO-wait, because if the storage were faster, it could've been
 4781 * running and we'd not be idle.
 4782 *
 4783 * This has been extended to SMP, by doing the same for each CPU. This however
 4784 * is broken.
 4785 *
 4786 * Imagine for instance the case where two tasks block on one CPU, only the one
 4787 * CPU will have IO-wait accounted, while the other has regular idle. Even
 4788 * though, if the storage were faster, both could've ran at the same time,
 4789 * utilising both CPUs.
 4790 *
 4791 * This means, that when looking globally, the current IO-wait accounting on
 4792 * SMP is a lower bound, by reason of under accounting.
 4793 *
 4794 * Worse, since the numbers are provided per CPU, they are sometimes
 4795 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
 4796 * associated with any one particular CPU, it can wake to another CPU than it
 4797 * blocked on. This means the per CPU IO-wait number is meaningless.
 4798 *
 4799 * Task CPU affinities can make all that even more 'interesting'.
 4800 */
 4801
 4802unsigned int nr_iowait(void)
 4803{
 4804	unsigned int i, sum = 0;
 4805
 4806	for_each_possible_cpu(i)
 4807		sum += nr_iowait_cpu(i);
 4808
 4809	return sum;
 4810}
 4811
 4812#ifdef CONFIG_SMP
 4813
 4814/*
 4815 * sched_exec - execve() is a valuable balancing opportunity, because at
 4816 * this point the task has the smallest effective memory and cache footprint.
 4817 */
 4818void sched_exec(void)
 4819{
 4820	struct task_struct *p = current;
 4821	unsigned long flags;
 4822	int dest_cpu;
 4823
 4824	raw_spin_lock_irqsave(&p->pi_lock, flags);
 4825	dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
 4826	if (dest_cpu == smp_processor_id())
 4827		goto unlock;
 4828
 4829	if (likely(cpu_active(dest_cpu))) {
 4830		struct migration_arg arg = { p, dest_cpu };
 4831
 4832		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
 4833		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
 4834		return;
 4835	}
 4836unlock:
 4837	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
 4838}
 4839
 4840#endif
 4841
 4842DEFINE_PER_CPU(struct kernel_stat, kstat);
 4843DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
 4844
 4845EXPORT_PER_CPU_SYMBOL(kstat);
 4846EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
 4847
 4848/*
 4849 * The function fair_sched_class.update_curr accesses the struct curr
 4850 * and its field curr->exec_start; when called from task_sched_runtime(),
 4851 * we observe a high rate of cache misses in practice.
 4852 * Prefetching this data results in improved performance.
 4853 */
 4854static inline void prefetch_curr_exec_start(struct task_struct *p)
 4855{
 4856#ifdef CONFIG_FAIR_GROUP_SCHED
 4857	struct sched_entity *curr = (&p->se)->cfs_rq->curr;
 4858#else
 4859	struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
 4860#endif
 4861	prefetch(curr);
 4862	prefetch(&curr->exec_start);
 4863}
 4864
 4865/*
 4866 * Return accounted runtime for the task.
 4867 * In case the task is currently running, return the runtime plus current's
 4868 * pending runtime that have not been accounted yet.
 4869 */
 4870unsigned long long task_sched_runtime(struct task_struct *p)
 4871{
 4872	struct rq_flags rf;
 4873	struct rq *rq;
 4874	u64 ns;
 4875
 4876#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
 4877	/*
 4878	 * 64-bit doesn't need locks to atomically read a 64-bit value.
 4879	 * So we have a optimization chance when the task's delta_exec is 0.
 4880	 * Reading ->on_cpu is racy, but this is ok.
 4881	 *
 4882	 * If we race with it leaving CPU, we'll take a lock. So we're correct.
 4883	 * If we race with it entering CPU, unaccounted time is 0. This is
 4884	 * indistinguishable from the read occurring a few cycles earlier.
 4885	 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
 4886	 * been accounted, so we're correct here as well.
 4887	 */
 4888	if (!p->on_cpu || !task_on_rq_queued(p))
 4889		return p->se.sum_exec_runtime;
 4890#endif
 4891
 4892	rq = task_rq_lock(p, &rf);
 4893	/*
 4894	 * Must be ->curr _and_ ->on_rq.  If dequeued, we would
 4895	 * project cycles that may never be accounted to this
 4896	 * thread, breaking clock_gettime().
 4897	 */
 4898	if (task_current(rq, p) && task_on_rq_queued(p)) {
 4899		prefetch_curr_exec_start(p);
 4900		update_rq_clock(rq);
 4901		p->sched_class->update_curr(rq);
 4902	}
 4903	ns = p->se.sum_exec_runtime;
 4904	task_rq_unlock(rq, p, &rf);
 4905
 4906	return ns;
 4907}
 4908
 4909#ifdef CONFIG_SCHED_DEBUG
 4910static u64 cpu_resched_latency(struct rq *rq)
 4911{
 4912	int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
 4913	u64 resched_latency, now = rq_clock(rq);
 4914	static bool warned_once;
 4915
 4916	if (sysctl_resched_latency_warn_once && warned_once)
 4917		return 0;
 4918
 4919	if (!need_resched() || !latency_warn_ms)
 4920		return 0;
 4921
 4922	if (system_state == SYSTEM_BOOTING)
 4923		return 0;
 4924
 4925	if (!rq->last_seen_need_resched_ns) {
 4926		rq->last_seen_need_resched_ns = now;
 4927		rq->ticks_without_resched = 0;
 4928		return 0;
 4929	}
 4930
 4931	rq->ticks_without_resched++;
 4932	resched_latency = now - rq->last_seen_need_resched_ns;
 4933	if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
 4934		return 0;
 4935
 4936	warned_once = true;
 4937
 4938	return resched_latency;
 4939}
 4940
 4941static int __init setup_resched_latency_warn_ms(char *str)
 4942{
 4943	long val;
 4944
 4945	if ((kstrtol(str, 0, &val))) {
 4946		pr_warn("Unable to set resched_latency_warn_ms\n");
 4947		return 1;
 4948	}
 4949
 4950	sysctl_resched_latency_warn_ms = val;
 4951	return 1;
 4952}
 4953__setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
 4954#else
 4955static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
 4956#endif /* CONFIG_SCHED_DEBUG */
 4957
 4958/*
 4959 * This function gets called by the timer code, with HZ frequency.
 4960 * We call it with interrupts disabled.
 4961 */
 4962void scheduler_tick(void)
 4963{
 4964	int cpu = smp_processor_id();
 4965	struct rq *rq = cpu_rq(cpu);
 4966	struct task_struct *curr = rq->curr;
 
 4967	struct rq_flags rf;
 4968	unsigned long thermal_pressure;
 4969	u64 resched_latency;
 4970
 4971	arch_scale_freq_tick();
 
 
 4972	sched_clock_tick();
 4973
 4974	rq_lock(rq, &rf);
 
 
 
 4975
 4976	update_rq_clock(rq);
 4977	thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
 4978	update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
 4979	curr->sched_class->task_tick(rq, curr, 0);
 
 
 
 
 4980	if (sched_feat(LATENCY_WARN))
 4981		resched_latency = cpu_resched_latency(rq);
 4982	calc_global_load_tick(rq);
 
 
 
 4983
 4984	rq_unlock(rq, &rf);
 4985
 4986	if (sched_feat(LATENCY_WARN) && resched_latency)
 4987		resched_latency_warn(cpu, resched_latency);
 4988
 4989	perf_event_task_tick();
 4990
 
 
 
 4991#ifdef CONFIG_SMP
 4992	rq->idle_balance = idle_cpu(cpu);
 4993	trigger_load_balance(rq);
 
 
 4994#endif
 4995}
 4996
 4997#ifdef CONFIG_NO_HZ_FULL
 4998
 4999struct tick_work {
 5000	int			cpu;
 5001	atomic_t		state;
 5002	struct delayed_work	work;
 5003};
 5004/* Values for ->state, see diagram below. */
 5005#define TICK_SCHED_REMOTE_OFFLINE	0
 5006#define TICK_SCHED_REMOTE_OFFLINING	1
 5007#define TICK_SCHED_REMOTE_RUNNING	2
 5008
 5009/*
 5010 * State diagram for ->state:
 5011 *
 5012 *
 5013 *          TICK_SCHED_REMOTE_OFFLINE
 5014 *                    |   ^
 5015 *                    |   |
 5016 *                    |   | sched_tick_remote()
 5017 *                    |   |
 5018 *                    |   |
 5019 *                    +--TICK_SCHED_REMOTE_OFFLINING
 5020 *                    |   ^
 5021 *                    |   |
 5022 * sched_tick_start() |   | sched_tick_stop()
 5023 *                    |   |
 5024 *                    V   |
 5025 *          TICK_SCHED_REMOTE_RUNNING
 5026 *
 5027 *
 5028 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
 5029 * and sched_tick_start() are happy to leave the state in RUNNING.
 5030 */
 5031
 5032static struct tick_work __percpu *tick_work_cpu;
 5033
 5034static void sched_tick_remote(struct work_struct *work)
 5035{
 5036	struct delayed_work *dwork = to_delayed_work(work);
 5037	struct tick_work *twork = container_of(dwork, struct tick_work, work);
 5038	int cpu = twork->cpu;
 5039	struct rq *rq = cpu_rq(cpu);
 5040	struct task_struct *curr;
 5041	struct rq_flags rf;
 5042	u64 delta;
 5043	int os;
 5044
 5045	/*
 5046	 * Handle the tick only if it appears the remote CPU is running in full
 5047	 * dynticks mode. The check is racy by nature, but missing a tick or
 5048	 * having one too much is no big deal because the scheduler tick updates
 5049	 * statistics and checks timeslices in a time-independent way, regardless
 5050	 * of when exactly it is running.
 5051	 */
 5052	if (!tick_nohz_tick_stopped_cpu(cpu))
 5053		goto out_requeue;
 
 5054
 5055	rq_lock_irq(rq, &rf);
 5056	curr = rq->curr;
 5057	if (cpu_is_offline(cpu))
 5058		goto out_unlock;
 
 
 
 
 5059
 5060	update_rq_clock(rq);
 
 
 
 
 
 
 
 
 5061
 5062	if (!is_idle_task(curr)) {
 5063		/*
 5064		 * Make sure the next tick runs within a reasonable
 5065		 * amount of time.
 5066		 */
 5067		delta = rq_clock_task(rq) - curr->se.exec_start;
 5068		WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
 5069	}
 5070	curr->sched_class->task_tick(rq, curr, 0);
 5071
 5072	calc_load_nohz_remote(rq);
 5073out_unlock:
 5074	rq_unlock_irq(rq, &rf);
 5075out_requeue:
 5076
 5077	/*
 5078	 * Run the remote tick once per second (1Hz). This arbitrary
 5079	 * frequency is large enough to avoid overload but short enough
 5080	 * to keep scheduler internal stats reasonably up to date.  But
 5081	 * first update state to reflect hotplug activity if required.
 5082	 */
 5083	os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
 5084	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
 5085	if (os == TICK_SCHED_REMOTE_RUNNING)
 5086		queue_delayed_work(system_unbound_wq, dwork, HZ);
 5087}
 5088
 5089static void sched_tick_start(int cpu)
 5090{
 5091	int os;
 5092	struct tick_work *twork;
 5093
 5094	if (housekeeping_cpu(cpu, HK_FLAG_TICK))
 5095		return;
 5096
 5097	WARN_ON_ONCE(!tick_work_cpu);
 5098
 5099	twork = per_cpu_ptr(tick_work_cpu, cpu);
 5100	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
 5101	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
 5102	if (os == TICK_SCHED_REMOTE_OFFLINE) {
 5103		twork->cpu = cpu;
 5104		INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
 5105		queue_delayed_work(system_unbound_wq, &twork->work, HZ);
 5106	}
 5107}
 5108
 5109#ifdef CONFIG_HOTPLUG_CPU
 5110static void sched_tick_stop(int cpu)
 5111{
 5112	struct tick_work *twork;
 5113	int os;
 5114
 5115	if (housekeeping_cpu(cpu, HK_FLAG_TICK))
 5116		return;
 5117
 5118	WARN_ON_ONCE(!tick_work_cpu);
 5119
 5120	twork = per_cpu_ptr(tick_work_cpu, cpu);
 5121	/* There cannot be competing actions, but don't rely on stop-machine. */
 5122	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
 5123	WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
 5124	/* Don't cancel, as this would mess up the state machine. */
 5125}
 5126#endif /* CONFIG_HOTPLUG_CPU */
 5127
 5128int __init sched_tick_offload_init(void)
 5129{
 5130	tick_work_cpu = alloc_percpu(struct tick_work);
 5131	BUG_ON(!tick_work_cpu);
 5132	return 0;
 5133}
 5134
 5135#else /* !CONFIG_NO_HZ_FULL */
 5136static inline void sched_tick_start(int cpu) { }
 5137static inline void sched_tick_stop(int cpu) { }
 5138#endif
 5139
 5140#if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
 5141				defined(CONFIG_TRACE_PREEMPT_TOGGLE))
 5142/*
 5143 * If the value passed in is equal to the current preempt count
 5144 * then we just disabled preemption. Start timing the latency.
 5145 */
 5146static inline void preempt_latency_start(int val)
 5147{
 5148	if (preempt_count() == val) {
 5149		unsigned long ip = get_lock_parent_ip();
 5150#ifdef CONFIG_DEBUG_PREEMPT
 5151		current->preempt_disable_ip = ip;
 5152#endif
 5153		trace_preempt_off(CALLER_ADDR0, ip);
 5154	}
 5155}
 5156
 5157void preempt_count_add(int val)
 5158{
 5159#ifdef CONFIG_DEBUG_PREEMPT
 5160	/*
 5161	 * Underflow?
 5162	 */
 5163	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
 5164		return;
 5165#endif
 5166	__preempt_count_add(val);
 5167#ifdef CONFIG_DEBUG_PREEMPT
 5168	/*
 5169	 * Spinlock count overflowing soon?
 5170	 */
 5171	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
 5172				PREEMPT_MASK - 10);
 5173#endif
 5174	preempt_latency_start(val);
 5175}
 5176EXPORT_SYMBOL(preempt_count_add);
 5177NOKPROBE_SYMBOL(preempt_count_add);
 5178
 5179/*
 5180 * If the value passed in equals to the current preempt count
 5181 * then we just enabled preemption. Stop timing the latency.
 5182 */
 5183static inline void preempt_latency_stop(int val)
 5184{
 5185	if (preempt_count() == val)
 5186		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
 5187}
 5188
 5189void preempt_count_sub(int val)
 5190{
 5191#ifdef CONFIG_DEBUG_PREEMPT
 5192	/*
 5193	 * Underflow?
 5194	 */
 5195	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
 5196		return;
 5197	/*
 5198	 * Is the spinlock portion underflowing?
 5199	 */
 5200	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
 5201			!(preempt_count() & PREEMPT_MASK)))
 5202		return;
 5203#endif
 5204
 5205	preempt_latency_stop(val);
 5206	__preempt_count_sub(val);
 5207}
 5208EXPORT_SYMBOL(preempt_count_sub);
 5209NOKPROBE_SYMBOL(preempt_count_sub);
 5210
 5211#else
 5212static inline void preempt_latency_start(int val) { }
 5213static inline void preempt_latency_stop(int val) { }
 5214#endif
 5215
 5216static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
 5217{
 5218#ifdef CONFIG_DEBUG_PREEMPT
 5219	return p->preempt_disable_ip;
 5220#else
 5221	return 0;
 5222#endif
 5223}
 5224
 5225/*
 5226 * Print scheduling while atomic bug:
 5227 */
 5228static noinline void __schedule_bug(struct task_struct *prev)
 5229{
 5230	/* Save this before calling printk(), since that will clobber it */
 5231	unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
 5232
 5233	if (oops_in_progress)
 5234		return;
 5235
 5236	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
 5237		prev->comm, prev->pid, preempt_count());
 5238
 5239	debug_show_held_locks(prev);
 5240	print_modules();
 5241	if (irqs_disabled())
 5242		print_irqtrace_events(prev);
 5243	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
 5244	    && in_atomic_preempt_off()) {
 5245		pr_err("Preemption disabled at:");
 5246		print_ip_sym(KERN_ERR, preempt_disable_ip);
 5247	}
 5248	if (panic_on_warn)
 5249		panic("scheduling while atomic\n");
 5250
 5251	dump_stack();
 5252	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 5253}
 5254
 5255/*
 5256 * Various schedule()-time debugging checks and statistics:
 5257 */
 5258static inline void schedule_debug(struct task_struct *prev, bool preempt)
 5259{
 5260#ifdef CONFIG_SCHED_STACK_END_CHECK
 5261	if (task_stack_end_corrupted(prev))
 5262		panic("corrupted stack end detected inside scheduler\n");
 5263
 5264	if (task_scs_end_corrupted(prev))
 5265		panic("corrupted shadow stack detected inside scheduler\n");
 5266#endif
 5267
 5268#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
 5269	if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
 5270		printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
 5271			prev->comm, prev->pid, prev->non_block_count);
 5272		dump_stack();
 5273		add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 5274	}
 5275#endif
 5276
 5277	if (unlikely(in_atomic_preempt_off())) {
 5278		__schedule_bug(prev);
 5279		preempt_count_set(PREEMPT_DISABLED);
 5280	}
 5281	rcu_sleep_check();
 5282	SCHED_WARN_ON(ct_state() == CONTEXT_USER);
 5283
 5284	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
 5285
 5286	schedstat_inc(this_rq()->sched_count);
 5287}
 5288
 5289static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
 5290				  struct rq_flags *rf)
 5291{
 5292#ifdef CONFIG_SMP
 5293	const struct sched_class *class;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 5294	/*
 5295	 * We must do the balancing pass before put_prev_task(), such
 5296	 * that when we release the rq->lock the task is in the same
 5297	 * state as before we took rq->lock.
 5298	 *
 5299	 * We can terminate the balance pass as soon as we know there is
 5300	 * a runnable task of @class priority or higher.
 5301	 */
 5302	for_class_range(class, prev->sched_class, &idle_sched_class) {
 5303		if (class->balance(rq, prev, rf))
 5304			break;
 5305	}
 5306#endif
 5307
 5308	put_prev_task(rq, prev);
 5309}
 5310
 5311/*
 5312 * Pick up the highest-prio task:
 5313 */
 5314static inline struct task_struct *
 5315__pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
 5316{
 5317	const struct sched_class *class;
 5318	struct task_struct *p;
 5319
 
 
 
 
 
 5320	/*
 5321	 * Optimization: we know that if all tasks are in the fair class we can
 5322	 * call that function directly, but only if the @prev task wasn't of a
 5323	 * higher scheduling class, because otherwise those lose the
 5324	 * opportunity to pull in more work from other CPUs.
 5325	 */
 5326	if (likely(prev->sched_class <= &fair_sched_class &&
 5327		   rq->nr_running == rq->cfs.h_nr_running)) {
 5328
 5329		p = pick_next_task_fair(rq, prev, rf);
 5330		if (unlikely(p == RETRY_TASK))
 5331			goto restart;
 5332
 5333		/* Assume the next prioritized class is idle_sched_class */
 5334		if (!p) {
 5335			put_prev_task(rq, prev);
 5336			p = pick_next_task_idle(rq);
 5337		}
 5338
 5339		return p;
 5340	}
 5341
 5342restart:
 5343	put_prev_task_balance(rq, prev, rf);
 5344
 5345	for_each_class(class) {
 5346		p = class->pick_next_task(rq);
 5347		if (p)
 5348			return p;
 
 
 
 
 
 
 
 
 5349	}
 5350
 5351	/* The idle class should always have a runnable task: */
 5352	BUG();
 5353}
 5354
 5355#ifdef CONFIG_SCHED_CORE
 5356static inline bool is_task_rq_idle(struct task_struct *t)
 5357{
 5358	return (task_rq(t)->idle == t);
 5359}
 5360
 5361static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
 5362{
 5363	return is_task_rq_idle(a) || (a->core_cookie == cookie);
 5364}
 5365
 5366static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
 5367{
 5368	if (is_task_rq_idle(a) || is_task_rq_idle(b))
 5369		return true;
 5370
 5371	return a->core_cookie == b->core_cookie;
 5372}
 5373
 5374// XXX fairness/fwd progress conditions
 5375/*
 5376 * Returns
 5377 * - NULL if there is no runnable task for this class.
 5378 * - the highest priority task for this runqueue if it matches
 5379 *   rq->core->core_cookie or its priority is greater than max.
 5380 * - Else returns idle_task.
 5381 */
 5382static struct task_struct *
 5383pick_task(struct rq *rq, const struct sched_class *class, struct task_struct *max, bool in_fi)
 5384{
 5385	struct task_struct *class_pick, *cookie_pick;
 5386	unsigned long cookie = rq->core->core_cookie;
 5387
 5388	class_pick = class->pick_task(rq);
 5389	if (!class_pick)
 5390		return NULL;
 5391
 5392	if (!cookie) {
 5393		/*
 5394		 * If class_pick is tagged, return it only if it has
 5395		 * higher priority than max.
 5396		 */
 5397		if (max && class_pick->core_cookie &&
 5398		    prio_less(class_pick, max, in_fi))
 5399			return idle_sched_class.pick_task(rq);
 5400
 5401		return class_pick;
 5402	}
 5403
 5404	/*
 5405	 * If class_pick is idle or matches cookie, return early.
 5406	 */
 5407	if (cookie_equals(class_pick, cookie))
 5408		return class_pick;
 5409
 5410	cookie_pick = sched_core_find(rq, cookie);
 5411
 5412	/*
 5413	 * If class > max && class > cookie, it is the highest priority task on
 5414	 * the core (so far) and it must be selected, otherwise we must go with
 5415	 * the cookie pick in order to satisfy the constraint.
 5416	 */
 5417	if (prio_less(cookie_pick, class_pick, in_fi) &&
 5418	    (!max || prio_less(max, class_pick, in_fi)))
 5419		return class_pick;
 5420
 5421	return cookie_pick;
 5422}
 5423
 5424extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
 5425
 
 
 5426static struct task_struct *
 5427pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
 5428{
 5429	struct task_struct *next, *max = NULL;
 5430	const struct sched_class *class;
 5431	const struct cpumask *smt_mask;
 5432	bool fi_before = false;
 5433	int i, j, cpu, occ = 0;
 
 
 
 5434	bool need_sync;
 5435
 5436	if (!sched_core_enabled(rq))
 5437		return __pick_next_task(rq, prev, rf);
 5438
 5439	cpu = cpu_of(rq);
 5440
 5441	/* Stopper task is switching into idle, no need core-wide selection. */
 5442	if (cpu_is_offline(cpu)) {
 5443		/*
 5444		 * Reset core_pick so that we don't enter the fastpath when
 5445		 * coming online. core_pick would already be migrated to
 5446		 * another cpu during offline.
 5447		 */
 5448		rq->core_pick = NULL;
 
 5449		return __pick_next_task(rq, prev, rf);
 5450	}
 5451
 5452	/*
 5453	 * If there were no {en,de}queues since we picked (IOW, the task
 5454	 * pointers are all still valid), and we haven't scheduled the last
 5455	 * pick yet, do so now.
 5456	 *
 5457	 * rq->core_pick can be NULL if no selection was made for a CPU because
 5458	 * it was either offline or went offline during a sibling's core-wide
 5459	 * selection. In this case, do a core-wide selection.
 5460	 */
 5461	if (rq->core->core_pick_seq == rq->core->core_task_seq &&
 5462	    rq->core->core_pick_seq != rq->core_sched_seq &&
 5463	    rq->core_pick) {
 5464		WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
 5465
 5466		next = rq->core_pick;
 5467		if (next != prev) {
 5468			put_prev_task(rq, prev);
 5469			set_next_task(rq, next);
 5470		}
 5471
 5472		rq->core_pick = NULL;
 5473		return next;
 
 5474	}
 5475
 5476	put_prev_task_balance(rq, prev, rf);
 5477
 5478	smt_mask = cpu_smt_mask(cpu);
 5479	need_sync = !!rq->core->core_cookie;
 5480
 5481	/* reset state */
 5482	rq->core->core_cookie = 0UL;
 5483	if (rq->core->core_forceidle) {
 
 
 
 
 
 
 
 
 
 5484		need_sync = true;
 5485		fi_before = true;
 5486		rq->core->core_forceidle = false;
 5487	}
 5488
 5489	/*
 5490	 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
 5491	 *
 5492	 * @task_seq guards the task state ({en,de}queues)
 5493	 * @pick_seq is the @task_seq we did a selection on
 5494	 * @sched_seq is the @pick_seq we scheduled
 5495	 *
 5496	 * However, preemptions can cause multiple picks on the same task set.
 5497	 * 'Fix' this by also increasing @task_seq for every pick.
 5498	 */
 5499	rq->core->core_task_seq++;
 5500
 5501	/*
 5502	 * Optimize for common case where this CPU has no cookies
 5503	 * and there are no cookied tasks running on siblings.
 5504	 */
 5505	if (!need_sync) {
 5506		for_each_class(class) {
 5507			next = class->pick_task(rq);
 5508			if (next)
 5509				break;
 5510		}
 5511
 5512		if (!next->core_cookie) {
 5513			rq->core_pick = NULL;
 
 5514			/*
 5515			 * For robustness, update the min_vruntime_fi for
 5516			 * unconstrained picks as well.
 5517			 */
 5518			WARN_ON_ONCE(fi_before);
 5519			task_vruntime_update(rq, next, false);
 5520			goto done;
 5521		}
 5522	}
 5523
 5524	for_each_cpu(i, smt_mask) {
 5525		struct rq *rq_i = cpu_rq(i);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 5526
 5527		rq_i->core_pick = NULL;
 
 5528
 5529		if (i != cpu)
 5530			update_rq_clock(rq_i);
 5531	}
 5532
 
 
 5533	/*
 5534	 * Try and select tasks for each sibling in descending sched_class
 5535	 * order.
 5536	 */
 5537	for_each_class(class) {
 5538again:
 5539		for_each_cpu_wrap(i, smt_mask, cpu) {
 5540			struct rq *rq_i = cpu_rq(i);
 5541			struct task_struct *p;
 5542
 5543			if (rq_i->core_pick)
 5544				continue;
 5545
 5546			/*
 5547			 * If this sibling doesn't yet have a suitable task to
 5548			 * run; ask for the most eligible task, given the
 5549			 * highest priority task already selected for this
 5550			 * core.
 5551			 */
 5552			p = pick_task(rq_i, class, max, fi_before);
 5553			if (!p)
 5554				continue;
 
 5555
 5556			if (!is_task_rq_idle(p))
 5557				occ++;
 5558
 5559			rq_i->core_pick = p;
 5560			if (rq_i->idle == p && rq_i->nr_running) {
 5561				rq->core->core_forceidle = true;
 5562				if (!fi_before)
 5563					rq->core->core_forceidle_seq++;
 5564			}
 
 
 
 
 5565
 5566			/*
 5567			 * If this new candidate is of higher priority than the
 5568			 * previous; and they're incompatible; we need to wipe
 5569			 * the slate and start over. pick_task makes sure that
 5570			 * p's priority is more than max if it doesn't match
 5571			 * max's cookie.
 5572			 *
 5573			 * NOTE: this is a linear max-filter and is thus bounded
 5574			 * in execution time.
 5575			 */
 5576			if (!max || !cookie_match(max, p)) {
 5577				struct task_struct *old_max = max;
 5578
 5579				rq->core->core_cookie = p->core_cookie;
 5580				max = p;
 5581
 5582				if (old_max) {
 5583					rq->core->core_forceidle = false;
 5584					for_each_cpu(j, smt_mask) {
 5585						if (j == i)
 5586							continue;
 5587
 5588						cpu_rq(j)->core_pick = NULL;
 5589					}
 5590					occ = 1;
 5591					goto again;
 5592				}
 5593			}
 5594		}
 5595	}
 5596
 5597	rq->core->core_pick_seq = rq->core->core_task_seq;
 5598	next = rq->core_pick;
 5599	rq->core_sched_seq = rq->core->core_pick_seq;
 5600
 5601	/* Something should have been selected for current CPU */
 5602	WARN_ON_ONCE(!next);
 5603
 5604	/*
 5605	 * Reschedule siblings
 5606	 *
 5607	 * NOTE: L1TF -- at this point we're no longer running the old task and
 5608	 * sending an IPI (below) ensures the sibling will no longer be running
 5609	 * their task. This ensures there is no inter-sibling overlap between
 5610	 * non-matching user state.
 5611	 */
 5612	for_each_cpu(i, smt_mask) {
 5613		struct rq *rq_i = cpu_rq(i);
 5614
 5615		/*
 5616		 * An online sibling might have gone offline before a task
 5617		 * could be picked for it, or it might be offline but later
 5618		 * happen to come online, but its too late and nothing was
 5619		 * picked for it.  That's Ok - it will pick tasks for itself,
 5620		 * so ignore it.
 5621		 */
 5622		if (!rq_i->core_pick)
 5623			continue;
 5624
 5625		/*
 5626		 * Update for new !FI->FI transitions, or if continuing to be in !FI:
 5627		 * fi_before     fi      update?
 5628		 *  0            0       1
 5629		 *  0            1       1
 5630		 *  1            0       1
 5631		 *  1            1       0
 5632		 */
 5633		if (!(fi_before && rq->core->core_forceidle))
 5634			task_vruntime_update(rq_i, rq_i->core_pick, rq->core->core_forceidle);
 5635
 5636		rq_i->core_pick->core_occupation = occ;
 5637
 5638		if (i == cpu) {
 5639			rq_i->core_pick = NULL;
 
 5640			continue;
 5641		}
 5642
 5643		/* Did we break L1TF mitigation requirements? */
 5644		WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
 5645
 5646		if (rq_i->curr == rq_i->core_pick) {
 5647			rq_i->core_pick = NULL;
 
 5648			continue;
 5649		}
 5650
 5651		resched_curr(rq_i);
 5652	}
 5653
 5654done:
 5655	set_next_task(rq, next);
 
 
 
 5656	return next;
 5657}
 5658
 5659static bool try_steal_cookie(int this, int that)
 5660{
 5661	struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
 5662	struct task_struct *p;
 5663	unsigned long cookie;
 5664	bool success = false;
 5665
 5666	local_irq_disable();
 5667	double_rq_lock(dst, src);
 5668
 5669	cookie = dst->core->core_cookie;
 5670	if (!cookie)
 5671		goto unlock;
 5672
 5673	if (dst->curr != dst->idle)
 5674		goto unlock;
 5675
 5676	p = sched_core_find(src, cookie);
 5677	if (p == src->idle)
 5678		goto unlock;
 5679
 5680	do {
 5681		if (p == src->core_pick || p == src->curr)
 5682			goto next;
 5683
 5684		if (!cpumask_test_cpu(this, &p->cpus_mask))
 5685			goto next;
 5686
 5687		if (p->core_occupation > dst->idle->core_occupation)
 5688			goto next;
 
 
 
 
 
 
 
 
 5689
 5690		p->on_rq = TASK_ON_RQ_MIGRATING;
 5691		deactivate_task(src, p, 0);
 5692		set_task_cpu(p, this);
 5693		activate_task(dst, p, 0);
 5694		p->on_rq = TASK_ON_RQ_QUEUED;
 5695
 5696		resched_curr(dst);
 5697
 5698		success = true;
 5699		break;
 5700
 5701next:
 5702		p = sched_core_next(p, cookie);
 5703	} while (p);
 5704
 5705unlock:
 5706	double_rq_unlock(dst, src);
 5707	local_irq_enable();
 5708
 5709	return success;
 5710}
 5711
 5712static bool steal_cookie_task(int cpu, struct sched_domain *sd)
 5713{
 5714	int i;
 5715
 5716	for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
 5717		if (i == cpu)
 5718			continue;
 5719
 5720		if (need_resched())
 5721			break;
 5722
 5723		if (try_steal_cookie(cpu, i))
 5724			return true;
 5725	}
 5726
 5727	return false;
 5728}
 5729
 5730static void sched_core_balance(struct rq *rq)
 5731{
 5732	struct sched_domain *sd;
 5733	int cpu = cpu_of(rq);
 5734
 5735	preempt_disable();
 5736	rcu_read_lock();
 
 5737	raw_spin_rq_unlock_irq(rq);
 5738	for_each_domain(cpu, sd) {
 5739		if (need_resched())
 5740			break;
 5741
 5742		if (steal_cookie_task(cpu, sd))
 5743			break;
 5744	}
 5745	raw_spin_rq_lock_irq(rq);
 5746	rcu_read_unlock();
 5747	preempt_enable();
 5748}
 5749
 5750static DEFINE_PER_CPU(struct callback_head, core_balance_head);
 5751
 5752void queue_core_balance(struct rq *rq)
 5753{
 5754	if (!sched_core_enabled(rq))
 5755		return;
 5756
 5757	if (!rq->core->core_cookie)
 5758		return;
 5759
 5760	if (!rq->nr_running) /* not forced idle */
 5761		return;
 5762
 5763	queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
 5764}
 5765
 
 
 
 
 
 5766static void sched_core_cpu_starting(unsigned int cpu)
 5767{
 5768	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
 5769	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
 5770	unsigned long flags;
 5771	int t;
 5772
 5773	sched_core_lock(cpu, &flags);
 5774
 5775	WARN_ON_ONCE(rq->core != rq);
 5776
 5777	/* if we're the first, we'll be our own leader */
 5778	if (cpumask_weight(smt_mask) == 1)
 5779		goto unlock;
 5780
 5781	/* find the leader */
 5782	for_each_cpu(t, smt_mask) {
 5783		if (t == cpu)
 5784			continue;
 5785		rq = cpu_rq(t);
 5786		if (rq->core == rq) {
 5787			core_rq = rq;
 5788			break;
 5789		}
 5790	}
 5791
 5792	if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
 5793		goto unlock;
 5794
 5795	/* install and validate core_rq */
 5796	for_each_cpu(t, smt_mask) {
 5797		rq = cpu_rq(t);
 5798
 5799		if (t == cpu)
 5800			rq->core = core_rq;
 5801
 5802		WARN_ON_ONCE(rq->core != core_rq);
 5803	}
 5804
 5805unlock:
 5806	sched_core_unlock(cpu, &flags);
 5807}
 5808
 5809static void sched_core_cpu_deactivate(unsigned int cpu)
 5810{
 5811	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
 5812	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
 5813	unsigned long flags;
 5814	int t;
 5815
 5816	sched_core_lock(cpu, &flags);
 5817
 5818	/* if we're the last man standing, nothing to do */
 5819	if (cpumask_weight(smt_mask) == 1) {
 5820		WARN_ON_ONCE(rq->core != rq);
 5821		goto unlock;
 5822	}
 5823
 5824	/* if we're not the leader, nothing to do */
 5825	if (rq->core != rq)
 5826		goto unlock;
 5827
 5828	/* find a new leader */
 5829	for_each_cpu(t, smt_mask) {
 5830		if (t == cpu)
 5831			continue;
 5832		core_rq = cpu_rq(t);
 5833		break;
 5834	}
 5835
 5836	if (WARN_ON_ONCE(!core_rq)) /* impossible */
 5837		goto unlock;
 5838
 5839	/* copy the shared state to the new leader */
 5840	core_rq->core_task_seq      = rq->core_task_seq;
 5841	core_rq->core_pick_seq      = rq->core_pick_seq;
 5842	core_rq->core_cookie        = rq->core_cookie;
 5843	core_rq->core_forceidle     = rq->core_forceidle;
 5844	core_rq->core_forceidle_seq = rq->core_forceidle_seq;
 
 
 
 
 
 
 
 
 5845
 5846	/* install new leader */
 5847	for_each_cpu(t, smt_mask) {
 5848		rq = cpu_rq(t);
 5849		rq->core = core_rq;
 5850	}
 5851
 5852unlock:
 5853	sched_core_unlock(cpu, &flags);
 5854}
 5855
 5856static inline void sched_core_cpu_dying(unsigned int cpu)
 5857{
 5858	struct rq *rq = cpu_rq(cpu);
 5859
 5860	if (rq->core != rq)
 5861		rq->core = rq;
 5862}
 5863
 5864#else /* !CONFIG_SCHED_CORE */
 5865
 5866static inline void sched_core_cpu_starting(unsigned int cpu) {}
 5867static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
 5868static inline void sched_core_cpu_dying(unsigned int cpu) {}
 5869
 5870static struct task_struct *
 5871pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
 5872{
 5873	return __pick_next_task(rq, prev, rf);
 5874}
 5875
 5876#endif /* CONFIG_SCHED_CORE */
 5877
 5878/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 5879 * __schedule() is the main scheduler function.
 5880 *
 5881 * The main means of driving the scheduler and thus entering this function are:
 5882 *
 5883 *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
 5884 *
 5885 *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
 5886 *      paths. For example, see arch/x86/entry_64.S.
 5887 *
 5888 *      To drive preemption between tasks, the scheduler sets the flag in timer
 5889 *      interrupt handler scheduler_tick().
 5890 *
 5891 *   3. Wakeups don't really cause entry into schedule(). They add a
 5892 *      task to the run-queue and that's it.
 5893 *
 5894 *      Now, if the new task added to the run-queue preempts the current
 5895 *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
 5896 *      called on the nearest possible occasion:
 5897 *
 5898 *       - If the kernel is preemptible (CONFIG_PREEMPTION=y):
 5899 *
 5900 *         - in syscall or exception context, at the next outmost
 5901 *           preempt_enable(). (this might be as soon as the wake_up()'s
 5902 *           spin_unlock()!)
 5903 *
 5904 *         - in IRQ context, return from interrupt-handler to
 5905 *           preemptible context
 5906 *
 5907 *       - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
 5908 *         then at the next:
 5909 *
 5910 *          - cond_resched() call
 5911 *          - explicit schedule() call
 5912 *          - return from syscall or exception to user-space
 5913 *          - return from interrupt-handler to user-space
 5914 *
 5915 * WARNING: must be called with preemption disabled!
 5916 */
 5917static void __sched notrace __schedule(bool preempt)
 5918{
 5919	struct task_struct *prev, *next;
 
 
 
 
 
 5920	unsigned long *switch_count;
 5921	unsigned long prev_state;
 5922	struct rq_flags rf;
 5923	struct rq *rq;
 5924	int cpu;
 5925
 5926	cpu = smp_processor_id();
 5927	rq = cpu_rq(cpu);
 5928	prev = rq->curr;
 5929
 5930	schedule_debug(prev, preempt);
 5931
 5932	if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
 5933		hrtick_clear(rq);
 5934
 5935	local_irq_disable();
 5936	rcu_note_context_switch(preempt);
 5937
 5938	/*
 5939	 * Make sure that signal_pending_state()->signal_pending() below
 5940	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
 5941	 * done by the caller to avoid the race with signal_wake_up():
 5942	 *
 5943	 * __set_current_state(@state)		signal_wake_up()
 5944	 * schedule()				  set_tsk_thread_flag(p, TIF_SIGPENDING)
 5945	 *					  wake_up_state(p, state)
 5946	 *   LOCK rq->lock			    LOCK p->pi_state
 5947	 *   smp_mb__after_spinlock()		    smp_mb__after_spinlock()
 5948	 *     if (signal_pending_state())	    if (p->state & @state)
 5949	 *
 5950	 * Also, the membarrier system call requires a full memory barrier
 5951	 * after coming from user-space, before storing to rq->curr.
 
 
 5952	 */
 5953	rq_lock(rq, &rf);
 5954	smp_mb__after_spinlock();
 5955
 5956	/* Promote REQ to ACT */
 5957	rq->clock_update_flags <<= 1;
 5958	update_rq_clock(rq);
 
 5959
 5960	switch_count = &prev->nivcsw;
 5961
 
 
 
 5962	/*
 5963	 * We must load prev->state once (task_struct::state is volatile), such
 5964	 * that:
 5965	 *
 5966	 *  - we form a control dependency vs deactivate_task() below.
 5967	 *  - ptrace_{,un}freeze_traced() can change ->state underneath us.
 5968	 */
 5969	prev_state = READ_ONCE(prev->__state);
 5970	if (!preempt && prev_state) {
 5971		if (signal_pending_state(prev_state, prev)) {
 5972			WRITE_ONCE(prev->__state, TASK_RUNNING);
 5973		} else {
 5974			prev->sched_contributes_to_load =
 5975				(prev_state & TASK_UNINTERRUPTIBLE) &&
 5976				!(prev_state & TASK_NOLOAD) &&
 5977				!(prev->flags & PF_FROZEN);
 5978
 5979			if (prev->sched_contributes_to_load)
 5980				rq->nr_uninterruptible++;
 5981
 5982			/*
 5983			 * __schedule()			ttwu()
 5984			 *   prev_state = prev->state;    if (p->on_rq && ...)
 5985			 *   if (prev_state)		    goto out;
 5986			 *     p->on_rq = 0;		  smp_acquire__after_ctrl_dep();
 5987			 *				  p->state = TASK_WAKING
 5988			 *
 5989			 * Where __schedule() and ttwu() have matching control dependencies.
 5990			 *
 5991			 * After this, schedule() must not care about p->state any more.
 5992			 */
 5993			deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
 5994
 5995			if (prev->in_iowait) {
 5996				atomic_inc(&rq->nr_iowait);
 5997				delayacct_blkio_start();
 5998			}
 5999		}
 
 
 6000		switch_count = &prev->nvcsw;
 6001	}
 6002
 6003	next = pick_next_task(rq, prev, &rf);
 
 
 6004	clear_tsk_need_resched(prev);
 6005	clear_preempt_need_resched();
 6006#ifdef CONFIG_SCHED_DEBUG
 6007	rq->last_seen_need_resched_ns = 0;
 6008#endif
 6009
 6010	if (likely(prev != next)) {
 6011		rq->nr_switches++;
 6012		/*
 6013		 * RCU users of rcu_dereference(rq->curr) may not see
 6014		 * changes to task_struct made by pick_next_task().
 6015		 */
 6016		RCU_INIT_POINTER(rq->curr, next);
 6017		/*
 6018		 * The membarrier system call requires each architecture
 6019		 * to have a full memory barrier after updating
 6020		 * rq->curr, before returning to user-space.
 6021		 *
 6022		 * Here are the schemes providing that barrier on the
 6023		 * various architectures:
 6024		 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
 6025		 *   switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
 
 6026		 * - finish_lock_switch() for weakly-ordered
 6027		 *   architectures where spin_unlock is a full barrier,
 6028		 * - switch_to() for arm64 (weakly-ordered, spin_unlock
 6029		 *   is a RELEASE barrier),
 
 
 
 
 
 
 
 6030		 */
 6031		++*switch_count;
 6032
 6033		migrate_disable_switch(rq, prev);
 6034		psi_sched_switch(prev, next, !task_on_rq_queued(prev));
 
 
 6035
 6036		trace_sched_switch(preempt, prev, next);
 6037
 6038		/* Also unlocks the rq: */
 6039		rq = context_switch(rq, prev, next, &rf);
 6040	} else {
 6041		rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
 6042
 6043		rq_unpin_lock(rq, &rf);
 6044		__balance_callbacks(rq);
 6045		raw_spin_rq_unlock_irq(rq);
 6046	}
 6047}
 6048
 6049void __noreturn do_task_dead(void)
 6050{
 6051	/* Causes final put_task_struct in finish_task_switch(): */
 6052	set_special_state(TASK_DEAD);
 6053
 6054	/* Tell freezer to ignore us: */
 6055	current->flags |= PF_NOFREEZE;
 6056
 6057	__schedule(false);
 6058	BUG();
 6059
 6060	/* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
 6061	for (;;)
 6062		cpu_relax();
 6063}
 6064
 6065static inline void sched_submit_work(struct task_struct *tsk)
 6066{
 
 6067	unsigned int task_flags;
 6068
 6069	if (task_is_running(tsk))
 6070		return;
 
 
 
 6071
 6072	task_flags = tsk->flags;
 6073	/*
 6074	 * If a worker went to sleep, notify and ask workqueue whether
 6075	 * it wants to wake up a task to maintain concurrency.
 6076	 * As this function is called inside the schedule() context,
 6077	 * we disable preemption to avoid it calling schedule() again
 6078	 * in the possible wakeup of a kworker and because wq_worker_sleeping()
 6079	 * requires it.
 6080	 */
 6081	if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
 6082		preempt_disable();
 6083		if (task_flags & PF_WQ_WORKER)
 6084			wq_worker_sleeping(tsk);
 6085		else
 6086			io_wq_worker_sleeping(tsk);
 6087		preempt_enable_no_resched();
 6088	}
 6089
 6090	if (tsk_is_pi_blocked(tsk))
 6091		return;
 
 
 
 
 6092
 6093	/*
 6094	 * If we are going to sleep and we have plugged IO queued,
 6095	 * make sure to submit it to avoid deadlocks.
 6096	 */
 6097	if (blk_needs_flush_plug(tsk))
 6098		blk_schedule_flush_plug(tsk);
 
 6099}
 6100
 6101static void sched_update_worker(struct task_struct *tsk)
 6102{
 6103	if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
 
 
 6104		if (tsk->flags & PF_WQ_WORKER)
 6105			wq_worker_running(tsk);
 6106		else
 6107			io_wq_worker_running(tsk);
 6108	}
 6109}
 6110
 6111asmlinkage __visible void __sched schedule(void)
 6112{
 6113	struct task_struct *tsk = current;
 6114
 6115	sched_submit_work(tsk);
 6116	do {
 6117		preempt_disable();
 6118		__schedule(false);
 6119		sched_preempt_enable_no_resched();
 6120	} while (need_resched());
 
 
 
 
 
 
 
 
 
 
 
 
 
 6121	sched_update_worker(tsk);
 6122}
 6123EXPORT_SYMBOL(schedule);
 6124
 6125/*
 6126 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
 6127 * state (have scheduled out non-voluntarily) by making sure that all
 6128 * tasks have either left the run queue or have gone into user space.
 6129 * As idle tasks do not do either, they must not ever be preempted
 6130 * (schedule out non-voluntarily).
 6131 *
 6132 * schedule_idle() is similar to schedule_preempt_disable() except that it
 6133 * never enables preemption because it does not call sched_submit_work().
 6134 */
 6135void __sched schedule_idle(void)
 6136{
 6137	/*
 6138	 * As this skips calling sched_submit_work(), which the idle task does
 6139	 * regardless because that function is a nop when the task is in a
 6140	 * TASK_RUNNING state, make sure this isn't used someplace that the
 6141	 * current task can be in any other state. Note, idle is always in the
 6142	 * TASK_RUNNING state.
 6143	 */
 6144	WARN_ON_ONCE(current->__state);
 6145	do {
 6146		__schedule(false);
 6147	} while (need_resched());
 6148}
 6149
 6150#if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
 6151asmlinkage __visible void __sched schedule_user(void)
 6152{
 6153	/*
 6154	 * If we come here after a random call to set_need_resched(),
 6155	 * or we have been woken up remotely but the IPI has not yet arrived,
 6156	 * we haven't yet exited the RCU idle mode. Do it here manually until
 6157	 * we find a better solution.
 6158	 *
 6159	 * NB: There are buggy callers of this function.  Ideally we
 6160	 * should warn if prev_state != CONTEXT_USER, but that will trigger
 6161	 * too frequently to make sense yet.
 6162	 */
 6163	enum ctx_state prev_state = exception_enter();
 6164	schedule();
 6165	exception_exit(prev_state);
 6166}
 6167#endif
 6168
 6169/**
 6170 * schedule_preempt_disabled - called with preemption disabled
 6171 *
 6172 * Returns with preemption disabled. Note: preempt_count must be 1
 6173 */
 6174void __sched schedule_preempt_disabled(void)
 6175{
 6176	sched_preempt_enable_no_resched();
 6177	schedule();
 6178	preempt_disable();
 6179}
 6180
 
 
 
 
 
 
 
 
 6181static void __sched notrace preempt_schedule_common(void)
 6182{
 6183	do {
 6184		/*
 6185		 * Because the function tracer can trace preempt_count_sub()
 6186		 * and it also uses preempt_enable/disable_notrace(), if
 6187		 * NEED_RESCHED is set, the preempt_enable_notrace() called
 6188		 * by the function tracer will call this function again and
 6189		 * cause infinite recursion.
 6190		 *
 6191		 * Preemption must be disabled here before the function
 6192		 * tracer can trace. Break up preempt_disable() into two
 6193		 * calls. One to disable preemption without fear of being
 6194		 * traced. The other to still record the preemption latency,
 6195		 * which can also be traced by the function tracer.
 6196		 */
 6197		preempt_disable_notrace();
 6198		preempt_latency_start(1);
 6199		__schedule(true);
 6200		preempt_latency_stop(1);
 6201		preempt_enable_no_resched_notrace();
 6202
 6203		/*
 6204		 * Check again in case we missed a preemption opportunity
 6205		 * between schedule and now.
 6206		 */
 6207	} while (need_resched());
 6208}
 6209
 6210#ifdef CONFIG_PREEMPTION
 6211/*
 6212 * This is the entry point to schedule() from in-kernel preemption
 6213 * off of preempt_enable.
 6214 */
 6215asmlinkage __visible void __sched notrace preempt_schedule(void)
 6216{
 6217	/*
 6218	 * If there is a non-zero preempt_count or interrupts are disabled,
 6219	 * we do not want to preempt the current task. Just return..
 6220	 */
 6221	if (likely(!preemptible()))
 6222		return;
 6223
 6224	preempt_schedule_common();
 6225}
 6226NOKPROBE_SYMBOL(preempt_schedule);
 6227EXPORT_SYMBOL(preempt_schedule);
 6228
 6229#ifdef CONFIG_PREEMPT_DYNAMIC
 6230DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
 
 
 
 
 
 6231EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
 
 
 
 
 
 
 
 
 
 
 
 6232#endif
 6233
 6234
 6235/**
 6236 * preempt_schedule_notrace - preempt_schedule called by tracing
 6237 *
 6238 * The tracing infrastructure uses preempt_enable_notrace to prevent
 6239 * recursion and tracing preempt enabling caused by the tracing
 6240 * infrastructure itself. But as tracing can happen in areas coming
 6241 * from userspace or just about to enter userspace, a preempt enable
 6242 * can occur before user_exit() is called. This will cause the scheduler
 6243 * to be called when the system is still in usermode.
 6244 *
 6245 * To prevent this, the preempt_enable_notrace will use this function
 6246 * instead of preempt_schedule() to exit user context if needed before
 6247 * calling the scheduler.
 6248 */
 6249asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
 6250{
 6251	enum ctx_state prev_ctx;
 6252
 6253	if (likely(!preemptible()))
 6254		return;
 6255
 6256	do {
 6257		/*
 6258		 * Because the function tracer can trace preempt_count_sub()
 6259		 * and it also uses preempt_enable/disable_notrace(), if
 6260		 * NEED_RESCHED is set, the preempt_enable_notrace() called
 6261		 * by the function tracer will call this function again and
 6262		 * cause infinite recursion.
 6263		 *
 6264		 * Preemption must be disabled here before the function
 6265		 * tracer can trace. Break up preempt_disable() into two
 6266		 * calls. One to disable preemption without fear of being
 6267		 * traced. The other to still record the preemption latency,
 6268		 * which can also be traced by the function tracer.
 6269		 */
 6270		preempt_disable_notrace();
 6271		preempt_latency_start(1);
 6272		/*
 6273		 * Needs preempt disabled in case user_exit() is traced
 6274		 * and the tracer calls preempt_enable_notrace() causing
 6275		 * an infinite recursion.
 6276		 */
 6277		prev_ctx = exception_enter();
 6278		__schedule(true);
 6279		exception_exit(prev_ctx);
 6280
 6281		preempt_latency_stop(1);
 6282		preempt_enable_no_resched_notrace();
 6283	} while (need_resched());
 6284}
 6285EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
 6286
 6287#ifdef CONFIG_PREEMPT_DYNAMIC
 6288DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
 
 
 
 
 
 6289EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
 
 
 
 
 
 
 
 
 
 
 
 6290#endif
 6291
 6292#endif /* CONFIG_PREEMPTION */
 6293
 6294#ifdef CONFIG_PREEMPT_DYNAMIC
 6295
 6296#include <linux/entry-common.h>
 6297
 6298/*
 6299 * SC:cond_resched
 6300 * SC:might_resched
 6301 * SC:preempt_schedule
 6302 * SC:preempt_schedule_notrace
 6303 * SC:irqentry_exit_cond_resched
 6304 *
 6305 *
 6306 * NONE:
 6307 *   cond_resched               <- __cond_resched
 6308 *   might_resched              <- RET0
 6309 *   preempt_schedule           <- NOP
 6310 *   preempt_schedule_notrace   <- NOP
 6311 *   irqentry_exit_cond_resched <- NOP
 6312 *
 6313 * VOLUNTARY:
 6314 *   cond_resched               <- __cond_resched
 6315 *   might_resched              <- __cond_resched
 6316 *   preempt_schedule           <- NOP
 6317 *   preempt_schedule_notrace   <- NOP
 6318 *   irqentry_exit_cond_resched <- NOP
 6319 *
 6320 * FULL:
 6321 *   cond_resched               <- RET0
 6322 *   might_resched              <- RET0
 6323 *   preempt_schedule           <- preempt_schedule
 6324 *   preempt_schedule_notrace   <- preempt_schedule_notrace
 6325 *   irqentry_exit_cond_resched <- irqentry_exit_cond_resched
 6326 */
 6327
 6328enum {
 6329	preempt_dynamic_none = 0,
 6330	preempt_dynamic_voluntary,
 6331	preempt_dynamic_full,
 6332};
 6333
 6334int preempt_dynamic_mode = preempt_dynamic_full;
 6335
 6336int sched_dynamic_mode(const char *str)
 6337{
 6338	if (!strcmp(str, "none"))
 6339		return preempt_dynamic_none;
 6340
 6341	if (!strcmp(str, "voluntary"))
 6342		return preempt_dynamic_voluntary;
 6343
 6344	if (!strcmp(str, "full"))
 6345		return preempt_dynamic_full;
 6346
 6347	return -EINVAL;
 6348}
 6349
 6350void sched_dynamic_update(int mode)
 6351{
 6352	/*
 6353	 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
 6354	 * the ZERO state, which is invalid.
 6355	 */
 6356	static_call_update(cond_resched, __cond_resched);
 6357	static_call_update(might_resched, __cond_resched);
 6358	static_call_update(preempt_schedule, __preempt_schedule_func);
 6359	static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
 6360	static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
 6361
 6362	switch (mode) {
 6363	case preempt_dynamic_none:
 6364		static_call_update(cond_resched, __cond_resched);
 6365		static_call_update(might_resched, (void *)&__static_call_return0);
 6366		static_call_update(preempt_schedule, NULL);
 6367		static_call_update(preempt_schedule_notrace, NULL);
 6368		static_call_update(irqentry_exit_cond_resched, NULL);
 6369		pr_info("Dynamic Preempt: none\n");
 6370		break;
 6371
 6372	case preempt_dynamic_voluntary:
 6373		static_call_update(cond_resched, __cond_resched);
 6374		static_call_update(might_resched, __cond_resched);
 6375		static_call_update(preempt_schedule, NULL);
 6376		static_call_update(preempt_schedule_notrace, NULL);
 6377		static_call_update(irqentry_exit_cond_resched, NULL);
 6378		pr_info("Dynamic Preempt: voluntary\n");
 6379		break;
 6380
 6381	case preempt_dynamic_full:
 6382		static_call_update(cond_resched, (void *)&__static_call_return0);
 6383		static_call_update(might_resched, (void *)&__static_call_return0);
 6384		static_call_update(preempt_schedule, __preempt_schedule_func);
 6385		static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
 6386		static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
 6387		pr_info("Dynamic Preempt: full\n");
 6388		break;
 6389	}
 6390
 6391	preempt_dynamic_mode = mode;
 6392}
 6393
 6394static int __init setup_preempt_mode(char *str)
 6395{
 6396	int mode = sched_dynamic_mode(str);
 6397	if (mode < 0) {
 6398		pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
 6399		return 1;
 6400	}
 6401
 6402	sched_dynamic_update(mode);
 6403	return 0;
 6404}
 6405__setup("preempt=", setup_preempt_mode);
 6406
 6407#endif /* CONFIG_PREEMPT_DYNAMIC */
 6408
 6409/*
 6410 * This is the entry point to schedule() from kernel preemption
 6411 * off of irq context.
 6412 * Note, that this is called and return with irqs disabled. This will
 6413 * protect us against recursive calling from irq.
 6414 */
 6415asmlinkage __visible void __sched preempt_schedule_irq(void)
 6416{
 6417	enum ctx_state prev_state;
 6418
 6419	/* Catch callers which need to be fixed */
 6420	BUG_ON(preempt_count() || !irqs_disabled());
 6421
 6422	prev_state = exception_enter();
 6423
 6424	do {
 6425		preempt_disable();
 6426		local_irq_enable();
 6427		__schedule(true);
 6428		local_irq_disable();
 6429		sched_preempt_enable_no_resched();
 6430	} while (need_resched());
 6431
 6432	exception_exit(prev_state);
 6433}
 6434
 6435int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
 6436			  void *key)
 6437{
 6438	WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
 6439	return try_to_wake_up(curr->private, mode, wake_flags);
 6440}
 6441EXPORT_SYMBOL(default_wake_function);
 6442
 6443static void __setscheduler_prio(struct task_struct *p, int prio)
 6444{
 6445	if (dl_prio(prio))
 6446		p->sched_class = &dl_sched_class;
 6447	else if (rt_prio(prio))
 6448		p->sched_class = &rt_sched_class;
 6449	else
 6450		p->sched_class = &fair_sched_class;
 
 
 
 
 6451
 6452	p->prio = prio;
 6453}
 6454
 6455#ifdef CONFIG_RT_MUTEXES
 6456
 6457static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
 
 
 
 
 
 
 
 
 6458{
 6459	if (pi_task)
 6460		prio = min(prio, pi_task->prio);
 
 6461
 6462	return prio;
 
 
 
 6463}
 6464
 6465static inline int rt_effective_prio(struct task_struct *p, int prio)
 6466{
 6467	struct task_struct *pi_task = rt_mutex_get_top_task(p);
 6468
 6469	return __rt_effective_prio(pi_task, prio);
 6470}
 6471
 6472/*
 6473 * rt_mutex_setprio - set the current priority of a task
 6474 * @p: task to boost
 6475 * @pi_task: donor task
 6476 *
 6477 * This function changes the 'effective' priority of a task. It does
 6478 * not touch ->normal_prio like __setscheduler().
 6479 *
 6480 * Used by the rt_mutex code to implement priority inheritance
 6481 * logic. Call site only calls if the priority of the task changed.
 6482 */
 6483void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
 6484{
 6485	int prio, oldprio, queued, running, queue_flag =
 6486		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
 6487	const struct sched_class *prev_class;
 6488	struct rq_flags rf;
 6489	struct rq *rq;
 6490
 6491	/* XXX used to be waiter->prio, not waiter->task->prio */
 6492	prio = __rt_effective_prio(pi_task, p->normal_prio);
 6493
 6494	/*
 6495	 * If nothing changed; bail early.
 6496	 */
 6497	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
 6498		return;
 6499
 6500	rq = __task_rq_lock(p, &rf);
 6501	update_rq_clock(rq);
 6502	/*
 6503	 * Set under pi_lock && rq->lock, such that the value can be used under
 6504	 * either lock.
 6505	 *
 6506	 * Note that there is loads of tricky to make this pointer cache work
 6507	 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
 6508	 * ensure a task is de-boosted (pi_task is set to NULL) before the
 6509	 * task is allowed to run again (and can exit). This ensures the pointer
 6510	 * points to a blocked task -- which guarantees the task is present.
 6511	 */
 6512	p->pi_top_task = pi_task;
 6513
 6514	/*
 6515	 * For FIFO/RR we only need to set prio, if that matches we're done.
 6516	 */
 6517	if (prio == p->prio && !dl_prio(prio))
 6518		goto out_unlock;
 6519
 6520	/*
 6521	 * Idle task boosting is a nono in general. There is one
 6522	 * exception, when PREEMPT_RT and NOHZ is active:
 6523	 *
 6524	 * The idle task calls get_next_timer_interrupt() and holds
 6525	 * the timer wheel base->lock on the CPU and another CPU wants
 6526	 * to access the timer (probably to cancel it). We can safely
 6527	 * ignore the boosting request, as the idle CPU runs this code
 6528	 * with interrupts disabled and will complete the lock
 6529	 * protected section without being interrupted. So there is no
 6530	 * real need to boost.
 6531	 */
 6532	if (unlikely(p == rq->idle)) {
 6533		WARN_ON(p != rq->curr);
 6534		WARN_ON(p->pi_blocked_on);
 6535		goto out_unlock;
 6536	}
 6537
 6538	trace_sched_pi_setprio(p, pi_task);
 6539	oldprio = p->prio;
 6540
 6541	if (oldprio == prio)
 6542		queue_flag &= ~DEQUEUE_MOVE;
 6543
 6544	prev_class = p->sched_class;
 
 
 
 
 
 6545	queued = task_on_rq_queued(p);
 6546	running = task_current(rq, p);
 6547	if (queued)
 6548		dequeue_task(rq, p, queue_flag);
 6549	if (running)
 6550		put_prev_task(rq, p);
 6551
 6552	/*
 6553	 * Boosting condition are:
 6554	 * 1. -rt task is running and holds mutex A
 6555	 *      --> -dl task blocks on mutex A
 6556	 *
 6557	 * 2. -dl task is running and holds mutex A
 6558	 *      --> -dl task blocks on mutex A and could preempt the
 6559	 *          running task
 6560	 */
 6561	if (dl_prio(prio)) {
 6562		if (!dl_prio(p->normal_prio) ||
 6563		    (pi_task && dl_prio(pi_task->prio) &&
 6564		     dl_entity_preempt(&pi_task->dl, &p->dl))) {
 6565			p->dl.pi_se = pi_task->dl.pi_se;
 6566			queue_flag |= ENQUEUE_REPLENISH;
 6567		} else {
 6568			p->dl.pi_se = &p->dl;
 6569		}
 6570	} else if (rt_prio(prio)) {
 6571		if (dl_prio(oldprio))
 6572			p->dl.pi_se = &p->dl;
 6573		if (oldprio < prio)
 6574			queue_flag |= ENQUEUE_HEAD;
 6575	} else {
 6576		if (dl_prio(oldprio))
 6577			p->dl.pi_se = &p->dl;
 6578		if (rt_prio(oldprio))
 6579			p->rt.timeout = 0;
 6580	}
 6581
 6582	__setscheduler_prio(p, prio);
 
 
 
 6583
 6584	if (queued)
 6585		enqueue_task(rq, p, queue_flag);
 6586	if (running)
 6587		set_next_task(rq, p);
 6588
 6589	check_class_changed(rq, p, prev_class, oldprio);
 6590out_unlock:
 6591	/* Avoid rq from going away on us: */
 6592	preempt_disable();
 6593
 6594	rq_unpin_lock(rq, &rf);
 6595	__balance_callbacks(rq);
 6596	raw_spin_rq_unlock(rq);
 6597
 6598	preempt_enable();
 6599}
 6600#else
 6601static inline int rt_effective_prio(struct task_struct *p, int prio)
 6602{
 6603	return prio;
 6604}
 6605#endif
 6606
 6607void set_user_nice(struct task_struct *p, long nice)
 6608{
 6609	bool queued, running;
 6610	int old_prio;
 6611	struct rq_flags rf;
 6612	struct rq *rq;
 6613
 6614	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
 6615		return;
 6616	/*
 6617	 * We have to be careful, if called from sys_setpriority(),
 6618	 * the task might be in the middle of scheduling on another CPU.
 6619	 */
 6620	rq = task_rq_lock(p, &rf);
 6621	update_rq_clock(rq);
 6622
 6623	/*
 6624	 * The RT priorities are set via sched_setscheduler(), but we still
 6625	 * allow the 'normal' nice value to be set - but as expected
 6626	 * it won't have any effect on scheduling until the task is
 6627	 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
 6628	 */
 6629	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
 6630		p->static_prio = NICE_TO_PRIO(nice);
 6631		goto out_unlock;
 6632	}
 6633	queued = task_on_rq_queued(p);
 6634	running = task_current(rq, p);
 6635	if (queued)
 6636		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
 6637	if (running)
 6638		put_prev_task(rq, p);
 6639
 6640	p->static_prio = NICE_TO_PRIO(nice);
 6641	set_load_weight(p, true);
 6642	old_prio = p->prio;
 6643	p->prio = effective_prio(p);
 6644
 6645	if (queued)
 6646		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
 6647	if (running)
 6648		set_next_task(rq, p);
 6649
 6650	/*
 6651	 * If the task increased its priority or is running and
 6652	 * lowered its priority, then reschedule its CPU:
 6653	 */
 6654	p->sched_class->prio_changed(rq, p, old_prio);
 6655
 6656out_unlock:
 6657	task_rq_unlock(rq, p, &rf);
 6658}
 6659EXPORT_SYMBOL(set_user_nice);
 6660
 6661/*
 6662 * can_nice - check if a task can reduce its nice value
 6663 * @p: task
 6664 * @nice: nice value
 6665 */
 6666int can_nice(const struct task_struct *p, const int nice)
 6667{
 6668	/* Convert nice value [19,-20] to rlimit style value [1,40]: */
 6669	int nice_rlim = nice_to_rlimit(nice);
 6670
 6671	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
 6672		capable(CAP_SYS_NICE));
 6673}
 6674
 6675#ifdef __ARCH_WANT_SYS_NICE
 6676
 6677/*
 6678 * sys_nice - change the priority of the current process.
 6679 * @increment: priority increment
 6680 *
 6681 * sys_setpriority is a more generic, but much slower function that
 6682 * does similar things.
 6683 */
 6684SYSCALL_DEFINE1(nice, int, increment)
 6685{
 6686	long nice, retval;
 6687
 6688	/*
 6689	 * Setpriority might change our priority at the same moment.
 6690	 * We don't have to worry. Conceptually one call occurs first
 6691	 * and we have a single winner.
 6692	 */
 6693	increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
 6694	nice = task_nice(current) + increment;
 6695
 6696	nice = clamp_val(nice, MIN_NICE, MAX_NICE);
 6697	if (increment < 0 && !can_nice(current, nice))
 6698		return -EPERM;
 6699
 6700	retval = security_task_setnice(current, nice);
 6701	if (retval)
 6702		return retval;
 6703
 6704	set_user_nice(current, nice);
 6705	return 0;
 6706}
 6707
 6708#endif
 6709
 6710/**
 6711 * task_prio - return the priority value of a given task.
 6712 * @p: the task in question.
 6713 *
 6714 * Return: The priority value as seen by users in /proc.
 6715 *
 6716 * sched policy         return value   kernel prio    user prio/nice
 6717 *
 6718 * normal, batch, idle     [0 ... 39]  [100 ... 139]          0/[-20 ... 19]
 6719 * fifo, rr             [-2 ... -100]     [98 ... 0]  [1 ... 99]
 6720 * deadline                     -101             -1           0
 6721 */
 6722int task_prio(const struct task_struct *p)
 6723{
 6724	return p->prio - MAX_RT_PRIO;
 6725}
 6726
 6727/**
 6728 * idle_cpu - is a given CPU idle currently?
 6729 * @cpu: the processor in question.
 6730 *
 6731 * Return: 1 if the CPU is currently idle. 0 otherwise.
 6732 */
 6733int idle_cpu(int cpu)
 6734{
 6735	struct rq *rq = cpu_rq(cpu);
 6736
 6737	if (rq->curr != rq->idle)
 6738		return 0;
 6739
 6740	if (rq->nr_running)
 6741		return 0;
 6742
 6743#ifdef CONFIG_SMP
 6744	if (rq->ttwu_pending)
 6745		return 0;
 6746#endif
 6747
 6748	return 1;
 6749}
 6750
 6751/**
 6752 * available_idle_cpu - is a given CPU idle for enqueuing work.
 6753 * @cpu: the CPU in question.
 6754 *
 6755 * Return: 1 if the CPU is currently idle. 0 otherwise.
 6756 */
 6757int available_idle_cpu(int cpu)
 6758{
 6759	if (!idle_cpu(cpu))
 6760		return 0;
 6761
 6762	if (vcpu_is_preempted(cpu))
 6763		return 0;
 6764
 6765	return 1;
 6766}
 6767
 6768/**
 6769 * idle_task - return the idle task for a given CPU.
 6770 * @cpu: the processor in question.
 6771 *
 6772 * Return: The idle task for the CPU @cpu.
 6773 */
 6774struct task_struct *idle_task(int cpu)
 6775{
 6776	return cpu_rq(cpu)->idle;
 6777}
 6778
 6779#ifdef CONFIG_SMP
 6780/*
 6781 * This function computes an effective utilization for the given CPU, to be
 6782 * used for frequency selection given the linear relation: f = u * f_max.
 6783 *
 6784 * The scheduler tracks the following metrics:
 6785 *
 6786 *   cpu_util_{cfs,rt,dl,irq}()
 6787 *   cpu_bw_dl()
 6788 *
 6789 * Where the cfs,rt and dl util numbers are tracked with the same metric and
 6790 * synchronized windows and are thus directly comparable.
 6791 *
 6792 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
 6793 * which excludes things like IRQ and steal-time. These latter are then accrued
 6794 * in the irq utilization.
 6795 *
 6796 * The DL bandwidth number otoh is not a measured metric but a value computed
 6797 * based on the task model parameters and gives the minimal utilization
 6798 * required to meet deadlines.
 6799 */
 6800unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
 6801				 unsigned long max, enum cpu_util_type type,
 6802				 struct task_struct *p)
 6803{
 6804	unsigned long dl_util, util, irq;
 6805	struct rq *rq = cpu_rq(cpu);
 6806
 6807	if (!uclamp_is_used() &&
 6808	    type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
 6809		return max;
 6810	}
 6811
 6812	/*
 6813	 * Early check to see if IRQ/steal time saturates the CPU, can be
 6814	 * because of inaccuracies in how we track these -- see
 6815	 * update_irq_load_avg().
 6816	 */
 6817	irq = cpu_util_irq(rq);
 6818	if (unlikely(irq >= max))
 6819		return max;
 6820
 6821	/*
 6822	 * Because the time spend on RT/DL tasks is visible as 'lost' time to
 6823	 * CFS tasks and we use the same metric to track the effective
 6824	 * utilization (PELT windows are synchronized) we can directly add them
 6825	 * to obtain the CPU's actual utilization.
 6826	 *
 6827	 * CFS and RT utilization can be boosted or capped, depending on
 6828	 * utilization clamp constraints requested by currently RUNNABLE
 6829	 * tasks.
 6830	 * When there are no CFS RUNNABLE tasks, clamps are released and
 6831	 * frequency will be gracefully reduced with the utilization decay.
 6832	 */
 6833	util = util_cfs + cpu_util_rt(rq);
 6834	if (type == FREQUENCY_UTIL)
 6835		util = uclamp_rq_util_with(rq, util, p);
 6836
 6837	dl_util = cpu_util_dl(rq);
 6838
 6839	/*
 6840	 * For frequency selection we do not make cpu_util_dl() a permanent part
 6841	 * of this sum because we want to use cpu_bw_dl() later on, but we need
 6842	 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
 6843	 * that we select f_max when there is no idle time.
 6844	 *
 6845	 * NOTE: numerical errors or stop class might cause us to not quite hit
 6846	 * saturation when we should -- something for later.
 6847	 */
 6848	if (util + dl_util >= max)
 6849		return max;
 6850
 6851	/*
 6852	 * OTOH, for energy computation we need the estimated running time, so
 6853	 * include util_dl and ignore dl_bw.
 6854	 */
 6855	if (type == ENERGY_UTIL)
 6856		util += dl_util;
 6857
 6858	/*
 6859	 * There is still idle time; further improve the number by using the
 6860	 * irq metric. Because IRQ/steal time is hidden from the task clock we
 6861	 * need to scale the task numbers:
 6862	 *
 6863	 *              max - irq
 6864	 *   U' = irq + --------- * U
 6865	 *                 max
 6866	 */
 6867	util = scale_irq_capacity(util, irq, max);
 6868	util += irq;
 6869
 6870	/*
 6871	 * Bandwidth required by DEADLINE must always be granted while, for
 6872	 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
 6873	 * to gracefully reduce the frequency when no tasks show up for longer
 6874	 * periods of time.
 6875	 *
 6876	 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
 6877	 * bw_dl as requested freq. However, cpufreq is not yet ready for such
 6878	 * an interface. So, we only do the latter for now.
 6879	 */
 6880	if (type == FREQUENCY_UTIL)
 6881		util += cpu_bw_dl(rq);
 6882
 6883	return min(max, util);
 6884}
 6885
 6886unsigned long sched_cpu_util(int cpu, unsigned long max)
 6887{
 6888	return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max,
 6889				  ENERGY_UTIL, NULL);
 6890}
 6891#endif /* CONFIG_SMP */
 6892
 6893/**
 6894 * find_process_by_pid - find a process with a matching PID value.
 6895 * @pid: the pid in question.
 6896 *
 6897 * The task of @pid, if found. %NULL otherwise.
 6898 */
 6899static struct task_struct *find_process_by_pid(pid_t pid)
 6900{
 6901	return pid ? find_task_by_vpid(pid) : current;
 6902}
 6903
 6904/*
 6905 * sched_setparam() passes in -1 for its policy, to let the functions
 6906 * it calls know not to change it.
 6907 */
 6908#define SETPARAM_POLICY	-1
 6909
 6910static void __setscheduler_params(struct task_struct *p,
 6911		const struct sched_attr *attr)
 6912{
 6913	int policy = attr->sched_policy;
 6914
 6915	if (policy == SETPARAM_POLICY)
 6916		policy = p->policy;
 6917
 6918	p->policy = policy;
 6919
 6920	if (dl_policy(policy))
 6921		__setparam_dl(p, attr);
 6922	else if (fair_policy(policy))
 6923		p->static_prio = NICE_TO_PRIO(attr->sched_nice);
 6924
 6925	/*
 6926	 * __sched_setscheduler() ensures attr->sched_priority == 0 when
 6927	 * !rt_policy. Always setting this ensures that things like
 6928	 * getparam()/getattr() don't report silly values for !rt tasks.
 6929	 */
 6930	p->rt_priority = attr->sched_priority;
 6931	p->normal_prio = normal_prio(p);
 6932	set_load_weight(p, true);
 6933}
 6934
 6935/*
 6936 * Check the target process has a UID that matches the current process's:
 6937 */
 6938static bool check_same_owner(struct task_struct *p)
 6939{
 6940	const struct cred *cred = current_cred(), *pcred;
 6941	bool match;
 6942
 6943	rcu_read_lock();
 6944	pcred = __task_cred(p);
 6945	match = (uid_eq(cred->euid, pcred->euid) ||
 6946		 uid_eq(cred->euid, pcred->uid));
 6947	rcu_read_unlock();
 6948	return match;
 6949}
 6950
 6951static int __sched_setscheduler(struct task_struct *p,
 6952				const struct sched_attr *attr,
 6953				bool user, bool pi)
 6954{
 6955	int oldpolicy = -1, policy = attr->sched_policy;
 6956	int retval, oldprio, newprio, queued, running;
 6957	const struct sched_class *prev_class;
 6958	struct callback_head *head;
 6959	struct rq_flags rf;
 6960	int reset_on_fork;
 6961	int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
 6962	struct rq *rq;
 6963
 6964	/* The pi code expects interrupts enabled */
 6965	BUG_ON(pi && in_interrupt());
 6966recheck:
 6967	/* Double check policy once rq lock held: */
 6968	if (policy < 0) {
 6969		reset_on_fork = p->sched_reset_on_fork;
 6970		policy = oldpolicy = p->policy;
 6971	} else {
 6972		reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
 6973
 6974		if (!valid_policy(policy))
 6975			return -EINVAL;
 6976	}
 6977
 6978	if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
 6979		return -EINVAL;
 6980
 6981	/*
 6982	 * Valid priorities for SCHED_FIFO and SCHED_RR are
 6983	 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
 6984	 * SCHED_BATCH and SCHED_IDLE is 0.
 6985	 */
 6986	if (attr->sched_priority > MAX_RT_PRIO-1)
 6987		return -EINVAL;
 6988	if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
 6989	    (rt_policy(policy) != (attr->sched_priority != 0)))
 6990		return -EINVAL;
 6991
 6992	/*
 6993	 * Allow unprivileged RT tasks to decrease priority:
 6994	 */
 6995	if (user && !capable(CAP_SYS_NICE)) {
 6996		if (fair_policy(policy)) {
 6997			if (attr->sched_nice < task_nice(p) &&
 6998			    !can_nice(p, attr->sched_nice))
 6999				return -EPERM;
 7000		}
 7001
 7002		if (rt_policy(policy)) {
 7003			unsigned long rlim_rtprio =
 7004					task_rlimit(p, RLIMIT_RTPRIO);
 7005
 7006			/* Can't set/change the rt policy: */
 7007			if (policy != p->policy && !rlim_rtprio)
 7008				return -EPERM;
 7009
 7010			/* Can't increase priority: */
 7011			if (attr->sched_priority > p->rt_priority &&
 7012			    attr->sched_priority > rlim_rtprio)
 7013				return -EPERM;
 7014		}
 7015
 7016		 /*
 7017		  * Can't set/change SCHED_DEADLINE policy at all for now
 7018		  * (safest behavior); in the future we would like to allow
 7019		  * unprivileged DL tasks to increase their relative deadline
 7020		  * or reduce their runtime (both ways reducing utilization)
 7021		  */
 7022		if (dl_policy(policy))
 7023			return -EPERM;
 7024
 7025		/*
 7026		 * Treat SCHED_IDLE as nice 20. Only allow a switch to
 7027		 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
 7028		 */
 7029		if (task_has_idle_policy(p) && !idle_policy(policy)) {
 7030			if (!can_nice(p, task_nice(p)))
 7031				return -EPERM;
 7032		}
 7033
 7034		/* Can't change other user's priorities: */
 7035		if (!check_same_owner(p))
 7036			return -EPERM;
 7037
 7038		/* Normal users shall not reset the sched_reset_on_fork flag: */
 7039		if (p->sched_reset_on_fork && !reset_on_fork)
 7040			return -EPERM;
 7041	}
 7042
 7043	if (user) {
 7044		if (attr->sched_flags & SCHED_FLAG_SUGOV)
 7045			return -EINVAL;
 7046
 7047		retval = security_task_setscheduler(p);
 7048		if (retval)
 7049			return retval;
 7050	}
 7051
 7052	/* Update task specific "requested" clamps */
 7053	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
 7054		retval = uclamp_validate(p, attr);
 7055		if (retval)
 7056			return retval;
 7057	}
 7058
 7059	if (pi)
 7060		cpuset_read_lock();
 7061
 7062	/*
 7063	 * Make sure no PI-waiters arrive (or leave) while we are
 7064	 * changing the priority of the task:
 7065	 *
 7066	 * To be able to change p->policy safely, the appropriate
 7067	 * runqueue lock must be held.
 7068	 */
 7069	rq = task_rq_lock(p, &rf);
 7070	update_rq_clock(rq);
 7071
 7072	/*
 7073	 * Changing the policy of the stop threads its a very bad idea:
 7074	 */
 7075	if (p == rq->stop) {
 7076		retval = -EINVAL;
 7077		goto unlock;
 7078	}
 7079
 7080	/*
 7081	 * If not changing anything there's no need to proceed further,
 7082	 * but store a possible modification of reset_on_fork.
 7083	 */
 7084	if (unlikely(policy == p->policy)) {
 7085		if (fair_policy(policy) && attr->sched_nice != task_nice(p))
 7086			goto change;
 7087		if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
 7088			goto change;
 7089		if (dl_policy(policy) && dl_param_changed(p, attr))
 7090			goto change;
 7091		if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
 7092			goto change;
 7093
 7094		p->sched_reset_on_fork = reset_on_fork;
 7095		retval = 0;
 7096		goto unlock;
 7097	}
 7098change:
 7099
 7100	if (user) {
 7101#ifdef CONFIG_RT_GROUP_SCHED
 7102		/*
 7103		 * Do not allow realtime tasks into groups that have no runtime
 7104		 * assigned.
 7105		 */
 7106		if (rt_bandwidth_enabled() && rt_policy(policy) &&
 7107				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
 7108				!task_group_is_autogroup(task_group(p))) {
 7109			retval = -EPERM;
 7110			goto unlock;
 7111		}
 7112#endif
 7113#ifdef CONFIG_SMP
 7114		if (dl_bandwidth_enabled() && dl_policy(policy) &&
 7115				!(attr->sched_flags & SCHED_FLAG_SUGOV)) {
 7116			cpumask_t *span = rq->rd->span;
 7117
 7118			/*
 7119			 * Don't allow tasks with an affinity mask smaller than
 7120			 * the entire root_domain to become SCHED_DEADLINE. We
 7121			 * will also fail if there's no bandwidth available.
 7122			 */
 7123			if (!cpumask_subset(span, p->cpus_ptr) ||
 7124			    rq->rd->dl_bw.bw == 0) {
 7125				retval = -EPERM;
 7126				goto unlock;
 7127			}
 7128		}
 7129#endif
 7130	}
 7131
 7132	/* Re-check policy now with rq lock held: */
 7133	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
 7134		policy = oldpolicy = -1;
 7135		task_rq_unlock(rq, p, &rf);
 7136		if (pi)
 7137			cpuset_read_unlock();
 7138		goto recheck;
 7139	}
 7140
 7141	/*
 7142	 * If setscheduling to SCHED_DEADLINE (or changing the parameters
 7143	 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
 7144	 * is available.
 7145	 */
 7146	if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
 7147		retval = -EBUSY;
 7148		goto unlock;
 7149	}
 7150
 7151	p->sched_reset_on_fork = reset_on_fork;
 7152	oldprio = p->prio;
 7153
 7154	newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
 7155	if (pi) {
 7156		/*
 7157		 * Take priority boosted tasks into account. If the new
 7158		 * effective priority is unchanged, we just store the new
 7159		 * normal parameters and do not touch the scheduler class and
 7160		 * the runqueue. This will be done when the task deboost
 7161		 * itself.
 7162		 */
 7163		newprio = rt_effective_prio(p, newprio);
 7164		if (newprio == oldprio)
 7165			queue_flags &= ~DEQUEUE_MOVE;
 7166	}
 7167
 7168	queued = task_on_rq_queued(p);
 7169	running = task_current(rq, p);
 7170	if (queued)
 7171		dequeue_task(rq, p, queue_flags);
 7172	if (running)
 7173		put_prev_task(rq, p);
 7174
 7175	prev_class = p->sched_class;
 7176
 7177	if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
 7178		__setscheduler_params(p, attr);
 7179		__setscheduler_prio(p, newprio);
 7180	}
 7181	__setscheduler_uclamp(p, attr);
 7182
 7183	if (queued) {
 7184		/*
 7185		 * We enqueue to tail when the priority of a task is
 7186		 * increased (user space view).
 7187		 */
 7188		if (oldprio < p->prio)
 7189			queue_flags |= ENQUEUE_HEAD;
 7190
 7191		enqueue_task(rq, p, queue_flags);
 7192	}
 7193	if (running)
 7194		set_next_task(rq, p);
 7195
 7196	check_class_changed(rq, p, prev_class, oldprio);
 7197
 7198	/* Avoid rq from going away on us: */
 7199	preempt_disable();
 7200	head = splice_balance_callbacks(rq);
 7201	task_rq_unlock(rq, p, &rf);
 7202
 7203	if (pi) {
 7204		cpuset_read_unlock();
 7205		rt_mutex_adjust_pi(p);
 7206	}
 7207
 7208	/* Run balance callbacks after we've adjusted the PI chain: */
 7209	balance_callbacks(rq, head);
 7210	preempt_enable();
 7211
 7212	return 0;
 7213
 7214unlock:
 7215	task_rq_unlock(rq, p, &rf);
 7216	if (pi)
 7217		cpuset_read_unlock();
 7218	return retval;
 7219}
 7220
 7221static int _sched_setscheduler(struct task_struct *p, int policy,
 7222			       const struct sched_param *param, bool check)
 7223{
 7224	struct sched_attr attr = {
 7225		.sched_policy   = policy,
 7226		.sched_priority = param->sched_priority,
 7227		.sched_nice	= PRIO_TO_NICE(p->static_prio),
 7228	};
 7229
 7230	/* Fixup the legacy SCHED_RESET_ON_FORK hack. */
 7231	if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
 7232		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
 7233		policy &= ~SCHED_RESET_ON_FORK;
 7234		attr.sched_policy = policy;
 7235	}
 7236
 7237	return __sched_setscheduler(p, &attr, check, true);
 7238}
 7239/**
 7240 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
 7241 * @p: the task in question.
 7242 * @policy: new policy.
 7243 * @param: structure containing the new RT priority.
 7244 *
 7245 * Use sched_set_fifo(), read its comment.
 7246 *
 7247 * Return: 0 on success. An error code otherwise.
 7248 *
 7249 * NOTE that the task may be already dead.
 7250 */
 7251int sched_setscheduler(struct task_struct *p, int policy,
 7252		       const struct sched_param *param)
 7253{
 7254	return _sched_setscheduler(p, policy, param, true);
 7255}
 7256
 7257int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
 7258{
 7259	return __sched_setscheduler(p, attr, true, true);
 7260}
 7261
 7262int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
 7263{
 7264	return __sched_setscheduler(p, attr, false, true);
 7265}
 7266EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
 7267
 7268/**
 7269 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
 7270 * @p: the task in question.
 7271 * @policy: new policy.
 7272 * @param: structure containing the new RT priority.
 7273 *
 7274 * Just like sched_setscheduler, only don't bother checking if the
 7275 * current context has permission.  For example, this is needed in
 7276 * stop_machine(): we create temporary high priority worker threads,
 7277 * but our caller might not have that capability.
 7278 *
 7279 * Return: 0 on success. An error code otherwise.
 7280 */
 7281int sched_setscheduler_nocheck(struct task_struct *p, int policy,
 7282			       const struct sched_param *param)
 7283{
 7284	return _sched_setscheduler(p, policy, param, false);
 7285}
 7286
 7287/*
 7288 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
 7289 * incapable of resource management, which is the one thing an OS really should
 7290 * be doing.
 7291 *
 7292 * This is of course the reason it is limited to privileged users only.
 7293 *
 7294 * Worse still; it is fundamentally impossible to compose static priority
 7295 * workloads. You cannot take two correctly working static prio workloads
 7296 * and smash them together and still expect them to work.
 7297 *
 7298 * For this reason 'all' FIFO tasks the kernel creates are basically at:
 7299 *
 7300 *   MAX_RT_PRIO / 2
 7301 *
 7302 * The administrator _MUST_ configure the system, the kernel simply doesn't
 7303 * know enough information to make a sensible choice.
 7304 */
 7305void sched_set_fifo(struct task_struct *p)
 7306{
 7307	struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
 7308	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
 7309}
 7310EXPORT_SYMBOL_GPL(sched_set_fifo);
 7311
 7312/*
 7313 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
 7314 */
 7315void sched_set_fifo_low(struct task_struct *p)
 7316{
 7317	struct sched_param sp = { .sched_priority = 1 };
 7318	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
 7319}
 7320EXPORT_SYMBOL_GPL(sched_set_fifo_low);
 7321
 7322void sched_set_normal(struct task_struct *p, int nice)
 7323{
 7324	struct sched_attr attr = {
 7325		.sched_policy = SCHED_NORMAL,
 7326		.sched_nice = nice,
 7327	};
 7328	WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
 7329}
 7330EXPORT_SYMBOL_GPL(sched_set_normal);
 7331
 7332static int
 7333do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
 7334{
 7335	struct sched_param lparam;
 7336	struct task_struct *p;
 7337	int retval;
 7338
 7339	if (!param || pid < 0)
 7340		return -EINVAL;
 7341	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
 7342		return -EFAULT;
 7343
 7344	rcu_read_lock();
 7345	retval = -ESRCH;
 7346	p = find_process_by_pid(pid);
 7347	if (likely(p))
 7348		get_task_struct(p);
 7349	rcu_read_unlock();
 7350
 7351	if (likely(p)) {
 7352		retval = sched_setscheduler(p, policy, &lparam);
 7353		put_task_struct(p);
 7354	}
 7355
 7356	return retval;
 7357}
 7358
 7359/*
 7360 * Mimics kernel/events/core.c perf_copy_attr().
 7361 */
 7362static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
 7363{
 7364	u32 size;
 7365	int ret;
 7366
 7367	/* Zero the full structure, so that a short copy will be nice: */
 7368	memset(attr, 0, sizeof(*attr));
 7369
 7370	ret = get_user(size, &uattr->size);
 7371	if (ret)
 7372		return ret;
 7373
 7374	/* ABI compatibility quirk: */
 7375	if (!size)
 7376		size = SCHED_ATTR_SIZE_VER0;
 7377	if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
 7378		goto err_size;
 7379
 7380	ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
 7381	if (ret) {
 7382		if (ret == -E2BIG)
 7383			goto err_size;
 7384		return ret;
 7385	}
 7386
 7387	if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
 7388	    size < SCHED_ATTR_SIZE_VER1)
 7389		return -EINVAL;
 7390
 7391	/*
 7392	 * XXX: Do we want to be lenient like existing syscalls; or do we want
 7393	 * to be strict and return an error on out-of-bounds values?
 7394	 */
 7395	attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
 7396
 7397	return 0;
 7398
 7399err_size:
 7400	put_user(sizeof(*attr), &uattr->size);
 7401	return -E2BIG;
 7402}
 7403
 7404/**
 7405 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
 7406 * @pid: the pid in question.
 7407 * @policy: new policy.
 7408 * @param: structure containing the new RT priority.
 7409 *
 7410 * Return: 0 on success. An error code otherwise.
 7411 */
 7412SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
 7413{
 7414	if (policy < 0)
 7415		return -EINVAL;
 7416
 7417	return do_sched_setscheduler(pid, policy, param);
 7418}
 7419
 7420/**
 7421 * sys_sched_setparam - set/change the RT priority of a thread
 7422 * @pid: the pid in question.
 7423 * @param: structure containing the new RT priority.
 7424 *
 7425 * Return: 0 on success. An error code otherwise.
 7426 */
 7427SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
 7428{
 7429	return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
 7430}
 7431
 7432/**
 7433 * sys_sched_setattr - same as above, but with extended sched_attr
 7434 * @pid: the pid in question.
 7435 * @uattr: structure containing the extended parameters.
 7436 * @flags: for future extension.
 7437 */
 7438SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
 7439			       unsigned int, flags)
 7440{
 7441	struct sched_attr attr;
 7442	struct task_struct *p;
 7443	int retval;
 7444
 7445	if (!uattr || pid < 0 || flags)
 7446		return -EINVAL;
 7447
 7448	retval = sched_copy_attr(uattr, &attr);
 7449	if (retval)
 7450		return retval;
 7451
 7452	if ((int)attr.sched_policy < 0)
 7453		return -EINVAL;
 7454	if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
 7455		attr.sched_policy = SETPARAM_POLICY;
 7456
 7457	rcu_read_lock();
 7458	retval = -ESRCH;
 7459	p = find_process_by_pid(pid);
 7460	if (likely(p))
 7461		get_task_struct(p);
 7462	rcu_read_unlock();
 7463
 7464	if (likely(p)) {
 7465		retval = sched_setattr(p, &attr);
 7466		put_task_struct(p);
 7467	}
 7468
 7469	return retval;
 7470}
 7471
 7472/**
 7473 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
 7474 * @pid: the pid in question.
 7475 *
 7476 * Return: On success, the policy of the thread. Otherwise, a negative error
 7477 * code.
 7478 */
 7479SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
 7480{
 7481	struct task_struct *p;
 7482	int retval;
 7483
 7484	if (pid < 0)
 7485		return -EINVAL;
 7486
 7487	retval = -ESRCH;
 7488	rcu_read_lock();
 7489	p = find_process_by_pid(pid);
 7490	if (p) {
 7491		retval = security_task_getscheduler(p);
 7492		if (!retval)
 7493			retval = p->policy
 7494				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
 7495	}
 7496	rcu_read_unlock();
 7497	return retval;
 7498}
 7499
 7500/**
 7501 * sys_sched_getparam - get the RT priority of a thread
 7502 * @pid: the pid in question.
 7503 * @param: structure containing the RT priority.
 7504 *
 7505 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
 7506 * code.
 7507 */
 7508SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
 7509{
 7510	struct sched_param lp = { .sched_priority = 0 };
 7511	struct task_struct *p;
 7512	int retval;
 7513
 7514	if (!param || pid < 0)
 7515		return -EINVAL;
 7516
 7517	rcu_read_lock();
 7518	p = find_process_by_pid(pid);
 7519	retval = -ESRCH;
 7520	if (!p)
 7521		goto out_unlock;
 7522
 7523	retval = security_task_getscheduler(p);
 7524	if (retval)
 7525		goto out_unlock;
 7526
 7527	if (task_has_rt_policy(p))
 7528		lp.sched_priority = p->rt_priority;
 7529	rcu_read_unlock();
 7530
 7531	/*
 7532	 * This one might sleep, we cannot do it with a spinlock held ...
 7533	 */
 7534	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
 7535
 7536	return retval;
 7537
 7538out_unlock:
 7539	rcu_read_unlock();
 7540	return retval;
 7541}
 7542
 7543/*
 7544 * Copy the kernel size attribute structure (which might be larger
 7545 * than what user-space knows about) to user-space.
 7546 *
 7547 * Note that all cases are valid: user-space buffer can be larger or
 7548 * smaller than the kernel-space buffer. The usual case is that both
 7549 * have the same size.
 7550 */
 7551static int
 7552sched_attr_copy_to_user(struct sched_attr __user *uattr,
 7553			struct sched_attr *kattr,
 7554			unsigned int usize)
 7555{
 7556	unsigned int ksize = sizeof(*kattr);
 7557
 7558	if (!access_ok(uattr, usize))
 7559		return -EFAULT;
 7560
 7561	/*
 7562	 * sched_getattr() ABI forwards and backwards compatibility:
 7563	 *
 7564	 * If usize == ksize then we just copy everything to user-space and all is good.
 7565	 *
 7566	 * If usize < ksize then we only copy as much as user-space has space for,
 7567	 * this keeps ABI compatibility as well. We skip the rest.
 7568	 *
 7569	 * If usize > ksize then user-space is using a newer version of the ABI,
 7570	 * which part the kernel doesn't know about. Just ignore it - tooling can
 7571	 * detect the kernel's knowledge of attributes from the attr->size value
 7572	 * which is set to ksize in this case.
 7573	 */
 7574	kattr->size = min(usize, ksize);
 7575
 7576	if (copy_to_user(uattr, kattr, kattr->size))
 7577		return -EFAULT;
 7578
 7579	return 0;
 7580}
 7581
 7582/**
 7583 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
 7584 * @pid: the pid in question.
 7585 * @uattr: structure containing the extended parameters.
 7586 * @usize: sizeof(attr) for fwd/bwd comp.
 7587 * @flags: for future extension.
 7588 */
 7589SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
 7590		unsigned int, usize, unsigned int, flags)
 7591{
 7592	struct sched_attr kattr = { };
 7593	struct task_struct *p;
 7594	int retval;
 7595
 7596	if (!uattr || pid < 0 || usize > PAGE_SIZE ||
 7597	    usize < SCHED_ATTR_SIZE_VER0 || flags)
 7598		return -EINVAL;
 7599
 7600	rcu_read_lock();
 7601	p = find_process_by_pid(pid);
 7602	retval = -ESRCH;
 7603	if (!p)
 7604		goto out_unlock;
 7605
 7606	retval = security_task_getscheduler(p);
 7607	if (retval)
 7608		goto out_unlock;
 7609
 7610	kattr.sched_policy = p->policy;
 7611	if (p->sched_reset_on_fork)
 7612		kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
 7613	if (task_has_dl_policy(p))
 7614		__getparam_dl(p, &kattr);
 7615	else if (task_has_rt_policy(p))
 7616		kattr.sched_priority = p->rt_priority;
 7617	else
 7618		kattr.sched_nice = task_nice(p);
 7619
 7620#ifdef CONFIG_UCLAMP_TASK
 7621	/*
 7622	 * This could race with another potential updater, but this is fine
 7623	 * because it'll correctly read the old or the new value. We don't need
 7624	 * to guarantee who wins the race as long as it doesn't return garbage.
 7625	 */
 7626	kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
 7627	kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
 7628#endif
 7629
 7630	rcu_read_unlock();
 7631
 7632	return sched_attr_copy_to_user(uattr, &kattr, usize);
 7633
 7634out_unlock:
 7635	rcu_read_unlock();
 7636	return retval;
 7637}
 7638
 7639long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
 7640{
 7641	cpumask_var_t cpus_allowed, new_mask;
 7642	struct task_struct *p;
 7643	int retval;
 7644
 7645	rcu_read_lock();
 7646
 7647	p = find_process_by_pid(pid);
 7648	if (!p) {
 7649		rcu_read_unlock();
 7650		return -ESRCH;
 7651	}
 7652
 7653	/* Prevent p going away */
 7654	get_task_struct(p);
 7655	rcu_read_unlock();
 7656
 7657	if (p->flags & PF_NO_SETAFFINITY) {
 7658		retval = -EINVAL;
 7659		goto out_put_task;
 7660	}
 7661	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
 7662		retval = -ENOMEM;
 7663		goto out_put_task;
 7664	}
 7665	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
 7666		retval = -ENOMEM;
 7667		goto out_free_cpus_allowed;
 7668	}
 7669	retval = -EPERM;
 7670	if (!check_same_owner(p)) {
 7671		rcu_read_lock();
 7672		if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
 7673			rcu_read_unlock();
 7674			goto out_free_new_mask;
 7675		}
 7676		rcu_read_unlock();
 7677	}
 7678
 7679	retval = security_task_setscheduler(p);
 7680	if (retval)
 7681		goto out_free_new_mask;
 7682
 7683
 7684	cpuset_cpus_allowed(p, cpus_allowed);
 7685	cpumask_and(new_mask, in_mask, cpus_allowed);
 7686
 7687	/*
 7688	 * Since bandwidth control happens on root_domain basis,
 7689	 * if admission test is enabled, we only admit -deadline
 7690	 * tasks allowed to run on all the CPUs in the task's
 7691	 * root_domain.
 7692	 */
 7693#ifdef CONFIG_SMP
 7694	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
 7695		rcu_read_lock();
 7696		if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
 7697			retval = -EBUSY;
 7698			rcu_read_unlock();
 7699			goto out_free_new_mask;
 7700		}
 7701		rcu_read_unlock();
 7702	}
 7703#endif
 7704again:
 7705	retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK);
 7706
 7707	if (!retval) {
 7708		cpuset_cpus_allowed(p, cpus_allowed);
 7709		if (!cpumask_subset(new_mask, cpus_allowed)) {
 7710			/*
 7711			 * We must have raced with a concurrent cpuset
 7712			 * update. Just reset the cpus_allowed to the
 7713			 * cpuset's cpus_allowed
 7714			 */
 7715			cpumask_copy(new_mask, cpus_allowed);
 7716			goto again;
 7717		}
 7718	}
 7719out_free_new_mask:
 7720	free_cpumask_var(new_mask);
 7721out_free_cpus_allowed:
 7722	free_cpumask_var(cpus_allowed);
 7723out_put_task:
 7724	put_task_struct(p);
 7725	return retval;
 7726}
 7727
 7728static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
 7729			     struct cpumask *new_mask)
 7730{
 7731	if (len < cpumask_size())
 7732		cpumask_clear(new_mask);
 7733	else if (len > cpumask_size())
 7734		len = cpumask_size();
 7735
 7736	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
 7737}
 7738
 7739/**
 7740 * sys_sched_setaffinity - set the CPU affinity of a process
 7741 * @pid: pid of the process
 7742 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
 7743 * @user_mask_ptr: user-space pointer to the new CPU mask
 7744 *
 7745 * Return: 0 on success. An error code otherwise.
 7746 */
 7747SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
 7748		unsigned long __user *, user_mask_ptr)
 7749{
 7750	cpumask_var_t new_mask;
 7751	int retval;
 7752
 7753	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
 7754		return -ENOMEM;
 7755
 7756	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
 7757	if (retval == 0)
 7758		retval = sched_setaffinity(pid, new_mask);
 7759	free_cpumask_var(new_mask);
 7760	return retval;
 7761}
 7762
 7763long sched_getaffinity(pid_t pid, struct cpumask *mask)
 7764{
 7765	struct task_struct *p;
 7766	unsigned long flags;
 7767	int retval;
 7768
 7769	rcu_read_lock();
 7770
 7771	retval = -ESRCH;
 7772	p = find_process_by_pid(pid);
 7773	if (!p)
 7774		goto out_unlock;
 7775
 7776	retval = security_task_getscheduler(p);
 7777	if (retval)
 7778		goto out_unlock;
 7779
 7780	raw_spin_lock_irqsave(&p->pi_lock, flags);
 7781	cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
 7782	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
 7783
 7784out_unlock:
 7785	rcu_read_unlock();
 7786
 7787	return retval;
 7788}
 7789
 7790/**
 7791 * sys_sched_getaffinity - get the CPU affinity of a process
 7792 * @pid: pid of the process
 7793 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
 7794 * @user_mask_ptr: user-space pointer to hold the current CPU mask
 7795 *
 7796 * Return: size of CPU mask copied to user_mask_ptr on success. An
 7797 * error code otherwise.
 7798 */
 7799SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
 7800		unsigned long __user *, user_mask_ptr)
 7801{
 7802	int ret;
 7803	cpumask_var_t mask;
 7804
 7805	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
 7806		return -EINVAL;
 7807	if (len & (sizeof(unsigned long)-1))
 7808		return -EINVAL;
 7809
 7810	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
 7811		return -ENOMEM;
 7812
 7813	ret = sched_getaffinity(pid, mask);
 7814	if (ret == 0) {
 7815		unsigned int retlen = min(len, cpumask_size());
 7816
 7817		if (copy_to_user(user_mask_ptr, mask, retlen))
 7818			ret = -EFAULT;
 7819		else
 7820			ret = retlen;
 7821	}
 7822	free_cpumask_var(mask);
 7823
 7824	return ret;
 7825}
 7826
 7827static void do_sched_yield(void)
 7828{
 7829	struct rq_flags rf;
 7830	struct rq *rq;
 7831
 7832	rq = this_rq_lock_irq(&rf);
 7833
 7834	schedstat_inc(rq->yld_count);
 7835	current->sched_class->yield_task(rq);
 7836
 7837	preempt_disable();
 7838	rq_unlock_irq(rq, &rf);
 7839	sched_preempt_enable_no_resched();
 7840
 7841	schedule();
 7842}
 7843
 7844/**
 7845 * sys_sched_yield - yield the current processor to other threads.
 7846 *
 7847 * This function yields the current CPU to other tasks. If there are no
 7848 * other threads running on this CPU then this function will return.
 7849 *
 7850 * Return: 0.
 7851 */
 7852SYSCALL_DEFINE0(sched_yield)
 7853{
 7854	do_sched_yield();
 7855	return 0;
 7856}
 7857
 7858#if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
 7859int __sched __cond_resched(void)
 7860{
 7861	if (should_resched(0)) {
 7862		preempt_schedule_common();
 7863		return 1;
 7864	}
 
 
 
 
 
 
 
 
 
 
 
 7865#ifndef CONFIG_PREEMPT_RCU
 7866	rcu_all_qs();
 7867#endif
 7868	return 0;
 7869}
 7870EXPORT_SYMBOL(__cond_resched);
 7871#endif
 7872
 7873#ifdef CONFIG_PREEMPT_DYNAMIC
 
 
 
 7874DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
 7875EXPORT_STATIC_CALL_TRAMP(cond_resched);
 7876
 
 
 7877DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
 7878EXPORT_STATIC_CALL_TRAMP(might_resched);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 7879#endif
 7880
 7881/*
 7882 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
 7883 * call schedule, and on return reacquire the lock.
 7884 *
 7885 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
 7886 * operations here to prevent schedule() from being called twice (once via
 7887 * spin_unlock(), once by hand).
 7888 */
 7889int __cond_resched_lock(spinlock_t *lock)
 7890{
 7891	int resched = should_resched(PREEMPT_LOCK_OFFSET);
 7892	int ret = 0;
 7893
 7894	lockdep_assert_held(lock);
 7895
 7896	if (spin_needbreak(lock) || resched) {
 7897		spin_unlock(lock);
 7898		if (resched)
 7899			preempt_schedule_common();
 7900		else
 7901			cpu_relax();
 7902		ret = 1;
 7903		spin_lock(lock);
 7904	}
 7905	return ret;
 7906}
 7907EXPORT_SYMBOL(__cond_resched_lock);
 7908
 7909int __cond_resched_rwlock_read(rwlock_t *lock)
 7910{
 7911	int resched = should_resched(PREEMPT_LOCK_OFFSET);
 7912	int ret = 0;
 7913
 7914	lockdep_assert_held_read(lock);
 7915
 7916	if (rwlock_needbreak(lock) || resched) {
 7917		read_unlock(lock);
 7918		if (resched)
 7919			preempt_schedule_common();
 7920		else
 7921			cpu_relax();
 7922		ret = 1;
 7923		read_lock(lock);
 7924	}
 7925	return ret;
 7926}
 7927EXPORT_SYMBOL(__cond_resched_rwlock_read);
 7928
 7929int __cond_resched_rwlock_write(rwlock_t *lock)
 7930{
 7931	int resched = should_resched(PREEMPT_LOCK_OFFSET);
 7932	int ret = 0;
 7933
 7934	lockdep_assert_held_write(lock);
 7935
 7936	if (rwlock_needbreak(lock) || resched) {
 7937		write_unlock(lock);
 7938		if (resched)
 7939			preempt_schedule_common();
 7940		else
 7941			cpu_relax();
 7942		ret = 1;
 7943		write_lock(lock);
 7944	}
 7945	return ret;
 7946}
 7947EXPORT_SYMBOL(__cond_resched_rwlock_write);
 7948
 7949/**
 7950 * yield - yield the current processor to other threads.
 
 
 
 
 
 
 
 
 
 
 7951 *
 7952 * Do not ever use this function, there's a 99% chance you're doing it wrong.
 7953 *
 7954 * The scheduler is at all times free to pick the calling task as the most
 7955 * eligible task to run, if removing the yield() call from your code breaks
 7956 * it, it's already broken.
 
 
 
 
 7957 *
 7958 * Typical broken usage is:
 
 
 
 
 
 
 7959 *
 7960 * while (!event)
 7961 *	yield();
 
 
 
 
 
 7962 *
 7963 * where one assumes that yield() will let 'the other' process run that will
 7964 * make event true. If the current task is a SCHED_FIFO task that will never
 7965 * happen. Never use yield() as a progress guarantee!!
 7966 *
 7967 * If you want to use yield() to wait for something, use wait_event().
 7968 * If you want to use yield() to be 'nice' for others, use cond_resched().
 7969 * If you still want to use yield(), do not!
 7970 */
 7971void __sched yield(void)
 
 
 
 
 
 
 
 
 
 
 
 7972{
 7973	set_current_state(TASK_RUNNING);
 7974	do_sched_yield();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 7975}
 7976EXPORT_SYMBOL(yield);
 7977
 7978/**
 7979 * yield_to - yield the current processor to another thread in
 7980 * your thread group, or accelerate that thread toward the
 7981 * processor it's on.
 7982 * @p: target task
 7983 * @preempt: whether task preemption is allowed or not
 7984 *
 7985 * It's the caller's job to ensure that the target task struct
 7986 * can't go away on us before we can do any checks.
 7987 *
 7988 * Return:
 7989 *	true (>0) if we indeed boosted the target task.
 7990 *	false (0) if we failed to boost the target.
 7991 *	-ESRCH if there's no task to yield to.
 7992 */
 7993int __sched yield_to(struct task_struct *p, bool preempt)
 7994{
 7995	struct task_struct *curr = current;
 7996	struct rq *rq, *p_rq;
 7997	unsigned long flags;
 7998	int yielded = 0;
 7999
 8000	local_irq_save(flags);
 8001	rq = this_rq();
 8002
 8003again:
 8004	p_rq = task_rq(p);
 8005	/*
 8006	 * If we're the only runnable task on the rq and target rq also
 8007	 * has only one task, there's absolutely no point in yielding.
 8008	 */
 8009	if (rq->nr_running == 1 && p_rq->nr_running == 1) {
 8010		yielded = -ESRCH;
 8011		goto out_irq;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 8012	}
 8013
 8014	double_rq_lock(rq, p_rq);
 8015	if (task_rq(p) != p_rq) {
 8016		double_rq_unlock(rq, p_rq);
 8017		goto again;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 8018	}
 8019
 8020	if (!curr->sched_class->yield_to_task)
 8021		goto out_unlock;
 
 
 8022
 8023	if (curr->sched_class != p->sched_class)
 8024		goto out_unlock;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 8025
 8026	if (task_running(p_rq, p) || !task_is_running(p))
 8027		goto out_unlock;
 
 
 
 
 
 8028
 8029	yielded = curr->sched_class->yield_to_task(rq, p);
 8030	if (yielded) {
 8031		schedstat_inc(rq->yld_count);
 8032		/*
 8033		 * Make p's CPU reschedule; pick_next_entity takes care of
 8034		 * fairness.
 8035		 */
 8036		if (preempt && rq != p_rq)
 8037			resched_curr(p_rq);
 8038	}
 8039
 8040out_unlock:
 8041	double_rq_unlock(rq, p_rq);
 8042out_irq:
 8043	local_irq_restore(flags);
 8044
 8045	if (yielded > 0)
 8046		schedule();
 8047
 8048	return yielded;
 8049}
 8050EXPORT_SYMBOL_GPL(yield_to);
 8051
 8052int io_schedule_prepare(void)
 8053{
 8054	int old_iowait = current->in_iowait;
 8055
 8056	current->in_iowait = 1;
 8057	blk_schedule_flush_plug(current);
 8058
 8059	return old_iowait;
 8060}
 8061
 8062void io_schedule_finish(int token)
 8063{
 8064	current->in_iowait = token;
 8065}
 8066
 8067/*
 8068 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
 8069 * that process accounting knows that this is a task in IO wait state.
 8070 */
 8071long __sched io_schedule_timeout(long timeout)
 8072{
 8073	int token;
 8074	long ret;
 8075
 8076	token = io_schedule_prepare();
 8077	ret = schedule_timeout(timeout);
 8078	io_schedule_finish(token);
 8079
 8080	return ret;
 8081}
 8082EXPORT_SYMBOL(io_schedule_timeout);
 8083
 8084void __sched io_schedule(void)
 8085{
 8086	int token;
 8087
 8088	token = io_schedule_prepare();
 8089	schedule();
 8090	io_schedule_finish(token);
 8091}
 8092EXPORT_SYMBOL(io_schedule);
 8093
 8094/**
 8095 * sys_sched_get_priority_max - return maximum RT priority.
 8096 * @policy: scheduling class.
 8097 *
 8098 * Return: On success, this syscall returns the maximum
 8099 * rt_priority that can be used by a given scheduling class.
 8100 * On failure, a negative error code is returned.
 8101 */
 8102SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
 8103{
 8104	int ret = -EINVAL;
 8105
 8106	switch (policy) {
 8107	case SCHED_FIFO:
 8108	case SCHED_RR:
 8109		ret = MAX_RT_PRIO-1;
 8110		break;
 8111	case SCHED_DEADLINE:
 8112	case SCHED_NORMAL:
 8113	case SCHED_BATCH:
 8114	case SCHED_IDLE:
 8115		ret = 0;
 8116		break;
 8117	}
 8118	return ret;
 8119}
 8120
 8121/**
 8122 * sys_sched_get_priority_min - return minimum RT priority.
 8123 * @policy: scheduling class.
 8124 *
 8125 * Return: On success, this syscall returns the minimum
 8126 * rt_priority that can be used by a given scheduling class.
 8127 * On failure, a negative error code is returned.
 8128 */
 8129SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
 8130{
 8131	int ret = -EINVAL;
 8132
 8133	switch (policy) {
 8134	case SCHED_FIFO:
 8135	case SCHED_RR:
 8136		ret = 1;
 8137		break;
 8138	case SCHED_DEADLINE:
 8139	case SCHED_NORMAL:
 8140	case SCHED_BATCH:
 8141	case SCHED_IDLE:
 8142		ret = 0;
 8143	}
 8144	return ret;
 8145}
 8146
 8147static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
 8148{
 8149	struct task_struct *p;
 8150	unsigned int time_slice;
 8151	struct rq_flags rf;
 8152	struct rq *rq;
 8153	int retval;
 8154
 8155	if (pid < 0)
 8156		return -EINVAL;
 8157
 8158	retval = -ESRCH;
 8159	rcu_read_lock();
 8160	p = find_process_by_pid(pid);
 8161	if (!p)
 8162		goto out_unlock;
 8163
 8164	retval = security_task_getscheduler(p);
 8165	if (retval)
 8166		goto out_unlock;
 8167
 8168	rq = task_rq_lock(p, &rf);
 8169	time_slice = 0;
 8170	if (p->sched_class->get_rr_interval)
 8171		time_slice = p->sched_class->get_rr_interval(rq, p);
 8172	task_rq_unlock(rq, p, &rf);
 8173
 8174	rcu_read_unlock();
 8175	jiffies_to_timespec64(time_slice, t);
 8176	return 0;
 8177
 8178out_unlock:
 8179	rcu_read_unlock();
 8180	return retval;
 8181}
 8182
 8183/**
 8184 * sys_sched_rr_get_interval - return the default timeslice of a process.
 8185 * @pid: pid of the process.
 8186 * @interval: userspace pointer to the timeslice value.
 8187 *
 8188 * this syscall writes the default timeslice value of a given process
 8189 * into the user-space timespec buffer. A value of '0' means infinity.
 8190 *
 8191 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
 8192 * an error code.
 8193 */
 8194SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
 8195		struct __kernel_timespec __user *, interval)
 8196{
 8197	struct timespec64 t;
 8198	int retval = sched_rr_get_interval(pid, &t);
 8199
 8200	if (retval == 0)
 8201		retval = put_timespec64(&t, interval);
 8202
 8203	return retval;
 8204}
 8205
 8206#ifdef CONFIG_COMPAT_32BIT_TIME
 8207SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
 8208		struct old_timespec32 __user *, interval)
 8209{
 8210	struct timespec64 t;
 8211	int retval = sched_rr_get_interval(pid, &t);
 8212
 8213	if (retval == 0)
 8214		retval = put_old_timespec32(&t, interval);
 8215	return retval;
 8216}
 8217#endif
 8218
 8219void sched_show_task(struct task_struct *p)
 8220{
 8221	unsigned long free = 0;
 8222	int ppid;
 8223
 8224	if (!try_get_task_stack(p))
 8225		return;
 8226
 8227	pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
 8228
 8229	if (task_is_running(p))
 8230		pr_cont("  running task    ");
 8231#ifdef CONFIG_DEBUG_STACK_USAGE
 8232	free = stack_not_used(p);
 8233#endif
 8234	ppid = 0;
 8235	rcu_read_lock();
 8236	if (pid_alive(p))
 8237		ppid = task_pid_nr(rcu_dereference(p->real_parent));
 8238	rcu_read_unlock();
 8239	pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
 8240		free, task_pid_nr(p), ppid,
 8241		(unsigned long)task_thread_info(p)->flags);
 8242
 8243	print_worker_info(KERN_INFO, p);
 8244	print_stop_info(KERN_INFO, p);
 
 8245	show_stack(p, NULL, KERN_INFO);
 8246	put_task_stack(p);
 8247}
 8248EXPORT_SYMBOL_GPL(sched_show_task);
 8249
 8250static inline bool
 8251state_filter_match(unsigned long state_filter, struct task_struct *p)
 8252{
 8253	unsigned int state = READ_ONCE(p->__state);
 8254
 8255	/* no filter, everything matches */
 8256	if (!state_filter)
 8257		return true;
 8258
 8259	/* filter, but doesn't match */
 8260	if (!(state & state_filter))
 8261		return false;
 8262
 8263	/*
 8264	 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
 8265	 * TASK_KILLABLE).
 8266	 */
 8267	if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
 8268		return false;
 8269
 8270	return true;
 8271}
 8272
 8273
 8274void show_state_filter(unsigned int state_filter)
 8275{
 8276	struct task_struct *g, *p;
 8277
 8278	rcu_read_lock();
 8279	for_each_process_thread(g, p) {
 8280		/*
 8281		 * reset the NMI-timeout, listing all files on a slow
 8282		 * console might take a lot of time:
 8283		 * Also, reset softlockup watchdogs on all CPUs, because
 8284		 * another CPU might be blocked waiting for us to process
 8285		 * an IPI.
 8286		 */
 8287		touch_nmi_watchdog();
 8288		touch_all_softlockup_watchdogs();
 8289		if (state_filter_match(state_filter, p))
 8290			sched_show_task(p);
 8291	}
 8292
 8293#ifdef CONFIG_SCHED_DEBUG
 8294	if (!state_filter)
 8295		sysrq_sched_debug_show();
 8296#endif
 8297	rcu_read_unlock();
 8298	/*
 8299	 * Only show locks if all tasks are dumped:
 8300	 */
 8301	if (!state_filter)
 8302		debug_show_all_locks();
 8303}
 8304
 8305/**
 8306 * init_idle - set up an idle thread for a given CPU
 8307 * @idle: task in question
 8308 * @cpu: CPU the idle task belongs to
 8309 *
 8310 * NOTE: this function does not set the idle thread's NEED_RESCHED
 8311 * flag, to make booting more robust.
 8312 */
 8313void __init init_idle(struct task_struct *idle, int cpu)
 8314{
 
 
 
 
 
 
 8315	struct rq *rq = cpu_rq(cpu);
 8316	unsigned long flags;
 8317
 8318	__sched_fork(0, idle);
 8319
 8320	/*
 8321	 * The idle task doesn't need the kthread struct to function, but it
 8322	 * is dressed up as a per-CPU kthread and thus needs to play the part
 8323	 * if we want to avoid special-casing it in code that deals with per-CPU
 8324	 * kthreads.
 8325	 */
 8326	set_kthread_struct(idle);
 8327
 8328	raw_spin_lock_irqsave(&idle->pi_lock, flags);
 8329	raw_spin_rq_lock(rq);
 8330
 8331	idle->__state = TASK_RUNNING;
 8332	idle->se.exec_start = sched_clock();
 8333	/*
 8334	 * PF_KTHREAD should already be set at this point; regardless, make it
 8335	 * look like a proper per-CPU kthread.
 8336	 */
 8337	idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
 8338	kthread_set_per_cpu(idle, cpu);
 8339
 8340	scs_task_reset(idle);
 8341	kasan_unpoison_task_stack(idle);
 8342
 8343#ifdef CONFIG_SMP
 8344	/*
 8345	 * It's possible that init_idle() gets called multiple times on a task,
 8346	 * in that case do_set_cpus_allowed() will not do the right thing.
 8347	 *
 8348	 * And since this is boot we can forgo the serialization.
 8349	 */
 8350	set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
 8351#endif
 8352	/*
 8353	 * We're having a chicken and egg problem, even though we are
 8354	 * holding rq->lock, the CPU isn't yet set to this CPU so the
 8355	 * lockdep check in task_group() will fail.
 8356	 *
 8357	 * Similar case to sched_fork(). / Alternatively we could
 8358	 * use task_rq_lock() here and obtain the other rq->lock.
 8359	 *
 8360	 * Silence PROVE_RCU
 8361	 */
 8362	rcu_read_lock();
 8363	__set_task_cpu(idle, cpu);
 8364	rcu_read_unlock();
 8365
 8366	rq->idle = idle;
 
 8367	rcu_assign_pointer(rq->curr, idle);
 8368	idle->on_rq = TASK_ON_RQ_QUEUED;
 8369#ifdef CONFIG_SMP
 8370	idle->on_cpu = 1;
 8371#endif
 8372	raw_spin_rq_unlock(rq);
 8373	raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
 8374
 8375	/* Set the preempt count _outside_ the spinlocks! */
 8376	init_idle_preempt_count(idle, cpu);
 8377
 8378	/*
 8379	 * The idle tasks have their own, simple scheduling class:
 8380	 */
 8381	idle->sched_class = &idle_sched_class;
 8382	ftrace_graph_init_idle_task(idle, cpu);
 8383	vtime_init_idle(idle, cpu);
 8384#ifdef CONFIG_SMP
 8385	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
 8386#endif
 8387}
 8388
 8389#ifdef CONFIG_SMP
 8390
 8391int cpuset_cpumask_can_shrink(const struct cpumask *cur,
 8392			      const struct cpumask *trial)
 8393{
 8394	int ret = 1;
 8395
 8396	if (!cpumask_weight(cur))
 8397		return ret;
 8398
 8399	ret = dl_cpuset_cpumask_can_shrink(cur, trial);
 8400
 8401	return ret;
 8402}
 8403
 8404int task_can_attach(struct task_struct *p,
 8405		    const struct cpumask *cs_cpus_allowed)
 8406{
 8407	int ret = 0;
 8408
 8409	/*
 8410	 * Kthreads which disallow setaffinity shouldn't be moved
 8411	 * to a new cpuset; we don't want to change their CPU
 8412	 * affinity and isolating such threads by their set of
 8413	 * allowed nodes is unnecessary.  Thus, cpusets are not
 8414	 * applicable for such threads.  This prevents checking for
 8415	 * success of set_cpus_allowed_ptr() on all attached tasks
 8416	 * before cpus_mask may be changed.
 8417	 */
 8418	if (p->flags & PF_NO_SETAFFINITY) {
 8419		ret = -EINVAL;
 8420		goto out;
 8421	}
 8422
 8423	if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
 8424					      cs_cpus_allowed))
 8425		ret = dl_task_can_attach(p, cs_cpus_allowed);
 8426
 8427out:
 8428	return ret;
 8429}
 8430
 8431bool sched_smp_initialized __read_mostly;
 8432
 8433#ifdef CONFIG_NUMA_BALANCING
 8434/* Migrate current task p to target_cpu */
 8435int migrate_task_to(struct task_struct *p, int target_cpu)
 8436{
 8437	struct migration_arg arg = { p, target_cpu };
 8438	int curr_cpu = task_cpu(p);
 8439
 8440	if (curr_cpu == target_cpu)
 8441		return 0;
 8442
 8443	if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
 8444		return -EINVAL;
 8445
 8446	/* TODO: This is not properly updating schedstats */
 8447
 8448	trace_sched_move_numa(p, curr_cpu, target_cpu);
 8449	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
 8450}
 8451
 8452/*
 8453 * Requeue a task on a given node and accurately track the number of NUMA
 8454 * tasks on the runqueues
 8455 */
 8456void sched_setnuma(struct task_struct *p, int nid)
 8457{
 8458	bool queued, running;
 8459	struct rq_flags rf;
 8460	struct rq *rq;
 8461
 8462	rq = task_rq_lock(p, &rf);
 8463	queued = task_on_rq_queued(p);
 8464	running = task_current(rq, p);
 8465
 8466	if (queued)
 8467		dequeue_task(rq, p, DEQUEUE_SAVE);
 8468	if (running)
 8469		put_prev_task(rq, p);
 8470
 8471	p->numa_preferred_nid = nid;
 8472
 8473	if (queued)
 8474		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
 8475	if (running)
 8476		set_next_task(rq, p);
 8477	task_rq_unlock(rq, p, &rf);
 8478}
 8479#endif /* CONFIG_NUMA_BALANCING */
 8480
 8481#ifdef CONFIG_HOTPLUG_CPU
 8482/*
 8483 * Ensure that the idle task is using init_mm right before its CPU goes
 8484 * offline.
 8485 */
 8486void idle_task_exit(void)
 8487{
 8488	struct mm_struct *mm = current->active_mm;
 8489
 8490	BUG_ON(cpu_online(smp_processor_id()));
 8491	BUG_ON(current != this_rq()->idle);
 8492
 8493	if (mm != &init_mm) {
 8494		switch_mm(mm, &init_mm, current);
 8495		finish_arch_post_lock_switch();
 8496	}
 8497
 8498	scs_task_reset(current);
 8499	/* finish_cpu(), as ran on the BP, will clean up the active_mm state */
 8500}
 8501
 8502static int __balance_push_cpu_stop(void *arg)
 8503{
 8504	struct task_struct *p = arg;
 8505	struct rq *rq = this_rq();
 8506	struct rq_flags rf;
 8507	int cpu;
 8508
 8509	raw_spin_lock_irq(&p->pi_lock);
 8510	rq_lock(rq, &rf);
 8511
 8512	update_rq_clock(rq);
 8513
 8514	if (task_rq(p) == rq && task_on_rq_queued(p)) {
 8515		cpu = select_fallback_rq(rq->cpu, p);
 8516		rq = __migrate_task(rq, &rf, p, cpu);
 8517	}
 8518
 8519	rq_unlock(rq, &rf);
 8520	raw_spin_unlock_irq(&p->pi_lock);
 8521
 8522	put_task_struct(p);
 8523
 8524	return 0;
 8525}
 8526
 8527static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
 8528
 8529/*
 8530 * Ensure we only run per-cpu kthreads once the CPU goes !active.
 8531 *
 8532 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
 8533 * effective when the hotplug motion is down.
 8534 */
 8535static void balance_push(struct rq *rq)
 8536{
 8537	struct task_struct *push_task = rq->curr;
 8538
 8539	lockdep_assert_rq_held(rq);
 8540
 8541	/*
 8542	 * Ensure the thing is persistent until balance_push_set(.on = false);
 8543	 */
 8544	rq->balance_callback = &balance_push_callback;
 8545
 8546	/*
 8547	 * Only active while going offline and when invoked on the outgoing
 8548	 * CPU.
 8549	 */
 8550	if (!cpu_dying(rq->cpu) || rq != this_rq())
 8551		return;
 8552
 8553	/*
 8554	 * Both the cpu-hotplug and stop task are in this case and are
 8555	 * required to complete the hotplug process.
 8556	 */
 8557	if (kthread_is_per_cpu(push_task) ||
 8558	    is_migration_disabled(push_task)) {
 8559
 8560		/*
 8561		 * If this is the idle task on the outgoing CPU try to wake
 8562		 * up the hotplug control thread which might wait for the
 8563		 * last task to vanish. The rcuwait_active() check is
 8564		 * accurate here because the waiter is pinned on this CPU
 8565		 * and can't obviously be running in parallel.
 8566		 *
 8567		 * On RT kernels this also has to check whether there are
 8568		 * pinned and scheduled out tasks on the runqueue. They
 8569		 * need to leave the migrate disabled section first.
 8570		 */
 8571		if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
 8572		    rcuwait_active(&rq->hotplug_wait)) {
 8573			raw_spin_rq_unlock(rq);
 8574			rcuwait_wake_up(&rq->hotplug_wait);
 8575			raw_spin_rq_lock(rq);
 8576		}
 8577		return;
 8578	}
 8579
 8580	get_task_struct(push_task);
 8581	/*
 8582	 * Temporarily drop rq->lock such that we can wake-up the stop task.
 8583	 * Both preemption and IRQs are still disabled.
 8584	 */
 
 8585	raw_spin_rq_unlock(rq);
 8586	stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
 8587			    this_cpu_ptr(&push_work));
 
 8588	/*
 8589	 * At this point need_resched() is true and we'll take the loop in
 8590	 * schedule(). The next pick is obviously going to be the stop task
 8591	 * which kthread_is_per_cpu() and will push this task away.
 8592	 */
 8593	raw_spin_rq_lock(rq);
 8594}
 8595
 8596static void balance_push_set(int cpu, bool on)
 8597{
 8598	struct rq *rq = cpu_rq(cpu);
 8599	struct rq_flags rf;
 8600
 8601	rq_lock_irqsave(rq, &rf);
 8602	if (on) {
 8603		WARN_ON_ONCE(rq->balance_callback);
 8604		rq->balance_callback = &balance_push_callback;
 8605	} else if (rq->balance_callback == &balance_push_callback) {
 8606		rq->balance_callback = NULL;
 8607	}
 8608	rq_unlock_irqrestore(rq, &rf);
 8609}
 8610
 8611/*
 8612 * Invoked from a CPUs hotplug control thread after the CPU has been marked
 8613 * inactive. All tasks which are not per CPU kernel threads are either
 8614 * pushed off this CPU now via balance_push() or placed on a different CPU
 8615 * during wakeup. Wait until the CPU is quiescent.
 8616 */
 8617static void balance_hotplug_wait(void)
 8618{
 8619	struct rq *rq = this_rq();
 8620
 8621	rcuwait_wait_event(&rq->hotplug_wait,
 8622			   rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
 8623			   TASK_UNINTERRUPTIBLE);
 8624}
 8625
 8626#else
 8627
 8628static inline void balance_push(struct rq *rq)
 8629{
 8630}
 8631
 8632static inline void balance_push_set(int cpu, bool on)
 8633{
 8634}
 8635
 8636static inline void balance_hotplug_wait(void)
 8637{
 8638}
 8639
 8640#endif /* CONFIG_HOTPLUG_CPU */
 8641
 8642void set_rq_online(struct rq *rq)
 8643{
 8644	if (!rq->online) {
 8645		const struct sched_class *class;
 8646
 8647		cpumask_set_cpu(rq->cpu, rq->rd->online);
 8648		rq->online = 1;
 8649
 8650		for_each_class(class) {
 8651			if (class->rq_online)
 8652				class->rq_online(rq);
 8653		}
 8654	}
 8655}
 8656
 8657void set_rq_offline(struct rq *rq)
 8658{
 8659	if (rq->online) {
 8660		const struct sched_class *class;
 8661
 
 8662		for_each_class(class) {
 8663			if (class->rq_offline)
 8664				class->rq_offline(rq);
 8665		}
 8666
 8667		cpumask_clear_cpu(rq->cpu, rq->rd->online);
 8668		rq->online = 0;
 8669	}
 8670}
 8671
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 8672/*
 8673 * used to mark begin/end of suspend/resume:
 8674 */
 8675static int num_cpus_frozen;
 8676
 8677/*
 8678 * Update cpusets according to cpu_active mask.  If cpusets are
 8679 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
 8680 * around partition_sched_domains().
 8681 *
 8682 * If we come here as part of a suspend/resume, don't touch cpusets because we
 8683 * want to restore it back to its original state upon resume anyway.
 8684 */
 8685static void cpuset_cpu_active(void)
 8686{
 8687	if (cpuhp_tasks_frozen) {
 8688		/*
 8689		 * num_cpus_frozen tracks how many CPUs are involved in suspend
 8690		 * resume sequence. As long as this is not the last online
 8691		 * operation in the resume sequence, just build a single sched
 8692		 * domain, ignoring cpusets.
 8693		 */
 8694		partition_sched_domains(1, NULL, NULL);
 8695		if (--num_cpus_frozen)
 8696			return;
 8697		/*
 8698		 * This is the last CPU online operation. So fall through and
 8699		 * restore the original sched domains by considering the
 8700		 * cpuset configurations.
 8701		 */
 8702		cpuset_force_rebuild();
 8703	}
 8704	cpuset_update_active_cpus();
 8705}
 8706
 8707static int cpuset_cpu_inactive(unsigned int cpu)
 8708{
 8709	if (!cpuhp_tasks_frozen) {
 8710		if (dl_cpu_busy(cpu))
 8711			return -EBUSY;
 
 
 8712		cpuset_update_active_cpus();
 8713	} else {
 8714		num_cpus_frozen++;
 8715		partition_sched_domains(1, NULL, NULL);
 8716	}
 8717	return 0;
 8718}
 8719
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 8720int sched_cpu_activate(unsigned int cpu)
 8721{
 8722	struct rq *rq = cpu_rq(cpu);
 8723	struct rq_flags rf;
 8724
 8725	/*
 8726	 * Clear the balance_push callback and prepare to schedule
 8727	 * regular tasks.
 8728	 */
 8729	balance_push_set(cpu, false);
 8730
 8731#ifdef CONFIG_SCHED_SMT
 8732	/*
 8733	 * When going up, increment the number of cores with SMT present.
 8734	 */
 8735	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
 8736		static_branch_inc_cpuslocked(&sched_smt_present);
 8737#endif
 8738	set_cpu_active(cpu, true);
 8739
 8740	if (sched_smp_initialized) {
 
 8741		sched_domains_numa_masks_set(cpu);
 8742		cpuset_cpu_active();
 8743	}
 8744
 
 
 8745	/*
 8746	 * Put the rq online, if not already. This happens:
 8747	 *
 8748	 * 1) In the early boot process, because we build the real domains
 8749	 *    after all CPUs have been brought up.
 8750	 *
 8751	 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
 8752	 *    domains.
 8753	 */
 8754	rq_lock_irqsave(rq, &rf);
 8755	if (rq->rd) {
 8756		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
 8757		set_rq_online(rq);
 8758	}
 8759	rq_unlock_irqrestore(rq, &rf);
 8760
 8761	return 0;
 8762}
 8763
 8764int sched_cpu_deactivate(unsigned int cpu)
 8765{
 8766	struct rq *rq = cpu_rq(cpu);
 8767	struct rq_flags rf;
 8768	int ret;
 8769
 8770	/*
 8771	 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
 8772	 * load balancing when not active
 8773	 */
 8774	nohz_balance_exit_idle(rq);
 8775
 8776	set_cpu_active(cpu, false);
 8777
 8778	/*
 8779	 * From this point forward, this CPU will refuse to run any task that
 8780	 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
 8781	 * push those tasks away until this gets cleared, see
 8782	 * sched_cpu_dying().
 8783	 */
 8784	balance_push_set(cpu, true);
 8785
 8786	/*
 8787	 * We've cleared cpu_active_mask / set balance_push, wait for all
 8788	 * preempt-disabled and RCU users of this state to go away such that
 8789	 * all new such users will observe it.
 8790	 *
 8791	 * Specifically, we rely on ttwu to no longer target this CPU, see
 8792	 * ttwu_queue_cond() and is_cpu_allowed().
 8793	 *
 8794	 * Do sync before park smpboot threads to take care the rcu boost case.
 8795	 */
 8796	synchronize_rcu();
 8797
 8798	rq_lock_irqsave(rq, &rf);
 8799	if (rq->rd) {
 8800		update_rq_clock(rq);
 8801		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
 8802		set_rq_offline(rq);
 8803	}
 8804	rq_unlock_irqrestore(rq, &rf);
 8805
 8806#ifdef CONFIG_SCHED_SMT
 8807	/*
 8808	 * When going down, decrement the number of cores with SMT present.
 8809	 */
 8810	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
 8811		static_branch_dec_cpuslocked(&sched_smt_present);
 8812
 
 8813	sched_core_cpu_deactivate(cpu);
 8814#endif
 8815
 8816	if (!sched_smp_initialized)
 8817		return 0;
 8818
 
 8819	ret = cpuset_cpu_inactive(cpu);
 8820	if (ret) {
 
 
 8821		balance_push_set(cpu, false);
 8822		set_cpu_active(cpu, true);
 
 8823		return ret;
 8824	}
 8825	sched_domains_numa_masks_clear(cpu);
 8826	return 0;
 8827}
 8828
 8829static void sched_rq_cpu_starting(unsigned int cpu)
 8830{
 8831	struct rq *rq = cpu_rq(cpu);
 8832
 8833	rq->calc_load_update = calc_load_update;
 8834	update_max_interval();
 8835}
 8836
 8837int sched_cpu_starting(unsigned int cpu)
 8838{
 8839	sched_core_cpu_starting(cpu);
 8840	sched_rq_cpu_starting(cpu);
 8841	sched_tick_start(cpu);
 8842	return 0;
 8843}
 8844
 8845#ifdef CONFIG_HOTPLUG_CPU
 8846
 8847/*
 8848 * Invoked immediately before the stopper thread is invoked to bring the
 8849 * CPU down completely. At this point all per CPU kthreads except the
 8850 * hotplug thread (current) and the stopper thread (inactive) have been
 8851 * either parked or have been unbound from the outgoing CPU. Ensure that
 8852 * any of those which might be on the way out are gone.
 8853 *
 8854 * If after this point a bound task is being woken on this CPU then the
 8855 * responsible hotplug callback has failed to do it's job.
 8856 * sched_cpu_dying() will catch it with the appropriate fireworks.
 8857 */
 8858int sched_cpu_wait_empty(unsigned int cpu)
 8859{
 8860	balance_hotplug_wait();
 8861	return 0;
 8862}
 8863
 8864/*
 8865 * Since this CPU is going 'away' for a while, fold any nr_active delta we
 8866 * might have. Called from the CPU stopper task after ensuring that the
 8867 * stopper is the last running task on the CPU, so nr_active count is
 8868 * stable. We need to take the teardown thread which is calling this into
 8869 * account, so we hand in adjust = 1 to the load calculation.
 8870 *
 8871 * Also see the comment "Global load-average calculations".
 8872 */
 8873static void calc_load_migrate(struct rq *rq)
 8874{
 8875	long delta = calc_load_fold_active(rq, 1);
 8876
 8877	if (delta)
 8878		atomic_long_add(delta, &calc_load_tasks);
 8879}
 8880
 8881static void dump_rq_tasks(struct rq *rq, const char *loglvl)
 8882{
 8883	struct task_struct *g, *p;
 8884	int cpu = cpu_of(rq);
 8885
 8886	lockdep_assert_rq_held(rq);
 8887
 8888	printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
 8889	for_each_process_thread(g, p) {
 8890		if (task_cpu(p) != cpu)
 8891			continue;
 8892
 8893		if (!task_on_rq_queued(p))
 8894			continue;
 8895
 8896		printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
 8897	}
 8898}
 8899
 8900int sched_cpu_dying(unsigned int cpu)
 8901{
 8902	struct rq *rq = cpu_rq(cpu);
 8903	struct rq_flags rf;
 8904
 8905	/* Handle pending wakeups and then migrate everything off */
 8906	sched_tick_stop(cpu);
 8907
 8908	rq_lock_irqsave(rq, &rf);
 8909	if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
 8910		WARN(true, "Dying CPU not properly vacated!");
 8911		dump_rq_tasks(rq, KERN_WARNING);
 8912	}
 8913	rq_unlock_irqrestore(rq, &rf);
 8914
 8915	calc_load_migrate(rq);
 8916	update_max_interval();
 8917	hrtick_clear(rq);
 8918	sched_core_cpu_dying(cpu);
 8919	return 0;
 8920}
 8921#endif
 8922
 8923void __init sched_init_smp(void)
 8924{
 8925	sched_init_numa();
 8926
 8927	/*
 8928	 * There's no userspace yet to cause hotplug operations; hence all the
 8929	 * CPU masks are stable and all blatant races in the below code cannot
 8930	 * happen.
 8931	 */
 8932	mutex_lock(&sched_domains_mutex);
 8933	sched_init_domains(cpu_active_mask);
 8934	mutex_unlock(&sched_domains_mutex);
 8935
 8936	/* Move init over to a non-isolated CPU */
 8937	if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
 8938		BUG();
 8939	current->flags &= ~PF_NO_SETAFFINITY;
 8940	sched_init_granularity();
 8941
 8942	init_sched_rt_class();
 8943	init_sched_dl_class();
 8944
 8945	sched_smp_initialized = true;
 8946}
 8947
 8948static int __init migration_init(void)
 8949{
 8950	sched_cpu_starting(smp_processor_id());
 8951	return 0;
 8952}
 8953early_initcall(migration_init);
 8954
 8955#else
 8956void __init sched_init_smp(void)
 8957{
 8958	sched_init_granularity();
 8959}
 8960#endif /* CONFIG_SMP */
 8961
 8962int in_sched_functions(unsigned long addr)
 8963{
 8964	return in_lock_functions(addr) ||
 8965		(addr >= (unsigned long)__sched_text_start
 8966		&& addr < (unsigned long)__sched_text_end);
 8967}
 8968
 8969#ifdef CONFIG_CGROUP_SCHED
 8970/*
 8971 * Default task group.
 8972 * Every task in system belongs to this group at bootup.
 8973 */
 8974struct task_group root_task_group;
 8975LIST_HEAD(task_groups);
 8976
 8977/* Cacheline aligned slab cache for task_group */
 8978static struct kmem_cache *task_group_cache __read_mostly;
 8979#endif
 8980
 8981DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
 8982DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
 8983
 8984void __init sched_init(void)
 8985{
 8986	unsigned long ptr = 0;
 8987	int i;
 8988
 8989	/* Make sure the linker didn't screw up */
 8990	BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
 8991	       &fair_sched_class + 1 != &rt_sched_class ||
 8992	       &rt_sched_class + 1   != &dl_sched_class);
 8993#ifdef CONFIG_SMP
 8994	BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
 
 
 
 
 
 
 
 8995#endif
 8996
 8997	wait_bit_init();
 8998
 8999#ifdef CONFIG_FAIR_GROUP_SCHED
 9000	ptr += 2 * nr_cpu_ids * sizeof(void **);
 9001#endif
 9002#ifdef CONFIG_RT_GROUP_SCHED
 9003	ptr += 2 * nr_cpu_ids * sizeof(void **);
 9004#endif
 9005	if (ptr) {
 9006		ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
 9007
 9008#ifdef CONFIG_FAIR_GROUP_SCHED
 9009		root_task_group.se = (struct sched_entity **)ptr;
 9010		ptr += nr_cpu_ids * sizeof(void **);
 9011
 9012		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
 9013		ptr += nr_cpu_ids * sizeof(void **);
 9014
 9015		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
 9016		init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
 9017#endif /* CONFIG_FAIR_GROUP_SCHED */
 
 
 
 9018#ifdef CONFIG_RT_GROUP_SCHED
 9019		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
 9020		ptr += nr_cpu_ids * sizeof(void **);
 9021
 9022		root_task_group.rt_rq = (struct rt_rq **)ptr;
 9023		ptr += nr_cpu_ids * sizeof(void **);
 9024
 9025#endif /* CONFIG_RT_GROUP_SCHED */
 9026	}
 9027#ifdef CONFIG_CPUMASK_OFFSTACK
 9028	for_each_possible_cpu(i) {
 9029		per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
 9030			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
 9031		per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
 9032			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
 9033	}
 9034#endif /* CONFIG_CPUMASK_OFFSTACK */
 9035
 9036	init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
 9037	init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
 9038
 9039#ifdef CONFIG_SMP
 9040	init_defrootdomain();
 9041#endif
 9042
 9043#ifdef CONFIG_RT_GROUP_SCHED
 9044	init_rt_bandwidth(&root_task_group.rt_bandwidth,
 9045			global_rt_period(), global_rt_runtime());
 9046#endif /* CONFIG_RT_GROUP_SCHED */
 9047
 9048#ifdef CONFIG_CGROUP_SCHED
 9049	task_group_cache = KMEM_CACHE(task_group, 0);
 9050
 9051	list_add(&root_task_group.list, &task_groups);
 9052	INIT_LIST_HEAD(&root_task_group.children);
 9053	INIT_LIST_HEAD(&root_task_group.siblings);
 9054	autogroup_init(&init_task);
 9055#endif /* CONFIG_CGROUP_SCHED */
 9056
 9057	for_each_possible_cpu(i) {
 9058		struct rq *rq;
 9059
 9060		rq = cpu_rq(i);
 9061		raw_spin_lock_init(&rq->__lock);
 9062		rq->nr_running = 0;
 9063		rq->calc_load_active = 0;
 9064		rq->calc_load_update = jiffies + LOAD_FREQ;
 9065		init_cfs_rq(&rq->cfs);
 9066		init_rt_rq(&rq->rt);
 9067		init_dl_rq(&rq->dl);
 9068#ifdef CONFIG_FAIR_GROUP_SCHED
 9069		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
 9070		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
 9071		/*
 9072		 * How much CPU bandwidth does root_task_group get?
 9073		 *
 9074		 * In case of task-groups formed thr' the cgroup filesystem, it
 9075		 * gets 100% of the CPU resources in the system. This overall
 9076		 * system CPU resource is divided among the tasks of
 9077		 * root_task_group and its child task-groups in a fair manner,
 9078		 * based on each entity's (task or task-group's) weight
 9079		 * (se->load.weight).
 9080		 *
 9081		 * In other words, if root_task_group has 10 tasks of weight
 9082		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
 9083		 * then A0's share of the CPU resource is:
 9084		 *
 9085		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
 9086		 *
 9087		 * We achieve this by letting root_task_group's tasks sit
 9088		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
 9089		 */
 9090		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
 9091#endif /* CONFIG_FAIR_GROUP_SCHED */
 9092
 9093		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
 9094#ifdef CONFIG_RT_GROUP_SCHED
 
 
 
 
 
 
 9095		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
 9096#endif
 9097#ifdef CONFIG_SMP
 9098		rq->sd = NULL;
 9099		rq->rd = NULL;
 9100		rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
 9101		rq->balance_callback = &balance_push_callback;
 9102		rq->active_balance = 0;
 9103		rq->next_balance = jiffies;
 9104		rq->push_cpu = 0;
 9105		rq->cpu = i;
 9106		rq->online = 0;
 9107		rq->idle_stamp = 0;
 9108		rq->avg_idle = 2*sysctl_sched_migration_cost;
 9109		rq->wake_stamp = jiffies;
 9110		rq->wake_avg_idle = rq->avg_idle;
 9111		rq->max_idle_balance_cost = sysctl_sched_migration_cost;
 9112
 9113		INIT_LIST_HEAD(&rq->cfs_tasks);
 9114
 9115		rq_attach_root(rq, &def_root_domain);
 9116#ifdef CONFIG_NO_HZ_COMMON
 9117		rq->last_blocked_load_update_tick = jiffies;
 9118		atomic_set(&rq->nohz_flags, 0);
 9119
 9120		INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
 9121#endif
 9122#ifdef CONFIG_HOTPLUG_CPU
 9123		rcuwait_init(&rq->hotplug_wait);
 9124#endif
 9125#endif /* CONFIG_SMP */
 9126		hrtick_rq_init(rq);
 9127		atomic_set(&rq->nr_iowait, 0);
 
 9128
 9129#ifdef CONFIG_SCHED_CORE
 9130		rq->core = rq;
 9131		rq->core_pick = NULL;
 
 9132		rq->core_enabled = 0;
 9133		rq->core_tree = RB_ROOT;
 9134		rq->core_forceidle = false;
 
 
 9135
 9136		rq->core_cookie = 0UL;
 9137#endif
 
 9138	}
 9139
 9140	set_load_weight(&init_task, false);
 
 9141
 9142	/*
 9143	 * The boot idle thread does lazy MMU switching as well:
 9144	 */
 9145	mmgrab(&init_mm);
 9146	enter_lazy_tlb(&init_mm, current);
 9147
 9148	/*
 
 
 
 
 
 
 
 
 9149	 * Make us the idle thread. Technically, schedule() should not be
 9150	 * called from this thread, however somewhere below it might be,
 9151	 * but because we are the idle thread, we just pick up running again
 9152	 * when this runqueue becomes "idle".
 9153	 */
 
 9154	init_idle(current, smp_processor_id());
 9155
 9156	calc_load_update = jiffies + LOAD_FREQ;
 9157
 9158#ifdef CONFIG_SMP
 9159	idle_thread_set_boot_cpu();
 9160	balance_push_set(smp_processor_id(), false);
 9161#endif
 9162	init_sched_fair_class();
 
 9163
 9164	psi_init();
 9165
 9166	init_uclamp();
 9167
 
 
 9168	scheduler_running = 1;
 9169}
 9170
 9171#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
 9172static inline int preempt_count_equals(int preempt_offset)
 9173{
 9174	int nested = preempt_count() + rcu_preempt_depth();
 9175
 9176	return (nested == preempt_offset);
 9177}
 9178
 9179void __might_sleep(const char *file, int line, int preempt_offset)
 9180{
 9181	unsigned int state = get_current_state();
 9182	/*
 9183	 * Blocking primitives will set (and therefore destroy) current->state,
 9184	 * since we will exit with TASK_RUNNING make sure we enter with it,
 9185	 * otherwise we will destroy state.
 9186	 */
 9187	WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
 9188			"do not call blocking ops when !TASK_RUNNING; "
 9189			"state=%x set at [<%p>] %pS\n", state,
 9190			(void *)current->task_state_change,
 9191			(void *)current->task_state_change);
 9192
 9193	___might_sleep(file, line, preempt_offset);
 9194}
 9195EXPORT_SYMBOL(__might_sleep);
 9196
 9197void ___might_sleep(const char *file, int line, int preempt_offset)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 9198{
 9199	/* Ratelimiting timestamp: */
 9200	static unsigned long prev_jiffy;
 9201
 9202	unsigned long preempt_disable_ip;
 9203
 9204	/* WARN_ON_ONCE() by default, no rate limit required: */
 9205	rcu_sleep_check();
 9206
 9207	if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
 9208	     !is_idle_task(current) && !current->non_block_count) ||
 9209	    system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
 9210	    oops_in_progress)
 9211		return;
 9212
 9213	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
 9214		return;
 9215	prev_jiffy = jiffies;
 9216
 9217	/* Save this before calling printk(), since that will clobber it: */
 9218	preempt_disable_ip = get_preempt_disable_ip(current);
 9219
 9220	printk(KERN_ERR
 9221		"BUG: sleeping function called from invalid context at %s:%d\n",
 9222			file, line);
 9223	printk(KERN_ERR
 9224		"in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
 9225			in_atomic(), irqs_disabled(), current->non_block_count,
 9226			current->pid, current->comm);
 
 
 
 
 
 9227
 9228	if (task_stack_end_corrupted(current))
 9229		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
 9230
 9231	debug_show_held_locks(current);
 9232	if (irqs_disabled())
 9233		print_irqtrace_events(current);
 9234	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
 9235	    && !preempt_count_equals(preempt_offset)) {
 9236		pr_err("Preemption disabled at:");
 9237		print_ip_sym(KERN_ERR, preempt_disable_ip);
 9238	}
 9239	dump_stack();
 9240	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 9241}
 9242EXPORT_SYMBOL(___might_sleep);
 9243
 9244void __cant_sleep(const char *file, int line, int preempt_offset)
 9245{
 9246	static unsigned long prev_jiffy;
 9247
 9248	if (irqs_disabled())
 9249		return;
 9250
 9251	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
 9252		return;
 9253
 9254	if (preempt_count() > preempt_offset)
 9255		return;
 9256
 9257	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
 9258		return;
 9259	prev_jiffy = jiffies;
 9260
 9261	printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
 9262	printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
 9263			in_atomic(), irqs_disabled(),
 9264			current->pid, current->comm);
 9265
 9266	debug_show_held_locks(current);
 9267	dump_stack();
 9268	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 9269}
 9270EXPORT_SYMBOL_GPL(__cant_sleep);
 9271
 9272#ifdef CONFIG_SMP
 9273void __cant_migrate(const char *file, int line)
 9274{
 9275	static unsigned long prev_jiffy;
 9276
 9277	if (irqs_disabled())
 9278		return;
 9279
 9280	if (is_migration_disabled(current))
 9281		return;
 9282
 9283	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
 9284		return;
 9285
 9286	if (preempt_count() > 0)
 9287		return;
 9288
 9289	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
 9290		return;
 9291	prev_jiffy = jiffies;
 9292
 9293	pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
 9294	pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
 9295	       in_atomic(), irqs_disabled(), is_migration_disabled(current),
 9296	       current->pid, current->comm);
 9297
 9298	debug_show_held_locks(current);
 9299	dump_stack();
 9300	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 9301}
 9302EXPORT_SYMBOL_GPL(__cant_migrate);
 9303#endif
 9304#endif
 9305
 9306#ifdef CONFIG_MAGIC_SYSRQ
 9307void normalize_rt_tasks(void)
 9308{
 9309	struct task_struct *g, *p;
 9310	struct sched_attr attr = {
 9311		.sched_policy = SCHED_NORMAL,
 9312	};
 9313
 9314	read_lock(&tasklist_lock);
 9315	for_each_process_thread(g, p) {
 9316		/*
 9317		 * Only normalize user tasks:
 9318		 */
 9319		if (p->flags & PF_KTHREAD)
 9320			continue;
 9321
 9322		p->se.exec_start = 0;
 9323		schedstat_set(p->se.statistics.wait_start,  0);
 9324		schedstat_set(p->se.statistics.sleep_start, 0);
 9325		schedstat_set(p->se.statistics.block_start, 0);
 9326
 9327		if (!dl_task(p) && !rt_task(p)) {
 9328			/*
 9329			 * Renice negative nice level userspace
 9330			 * tasks back to 0:
 9331			 */
 9332			if (task_nice(p) < 0)
 9333				set_user_nice(p, 0);
 9334			continue;
 9335		}
 9336
 9337		__sched_setscheduler(p, &attr, false, false);
 9338	}
 9339	read_unlock(&tasklist_lock);
 9340}
 9341
 9342#endif /* CONFIG_MAGIC_SYSRQ */
 9343
 9344#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
 9345/*
 9346 * These functions are only useful for the IA64 MCA handling, or kdb.
 9347 *
 9348 * They can only be called when the whole system has been
 9349 * stopped - every CPU needs to be quiescent, and no scheduling
 9350 * activity can take place. Using them for anything else would
 9351 * be a serious bug, and as a result, they aren't even visible
 9352 * under any other configuration.
 9353 */
 9354
 9355/**
 9356 * curr_task - return the current task for a given CPU.
 9357 * @cpu: the processor in question.
 9358 *
 9359 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
 9360 *
 9361 * Return: The current task for @cpu.
 9362 */
 9363struct task_struct *curr_task(int cpu)
 9364{
 9365	return cpu_curr(cpu);
 9366}
 9367
 9368#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
 9369
 9370#ifdef CONFIG_IA64
 9371/**
 9372 * ia64_set_curr_task - set the current task for a given CPU.
 9373 * @cpu: the processor in question.
 9374 * @p: the task pointer to set.
 9375 *
 9376 * Description: This function must only be used when non-maskable interrupts
 9377 * are serviced on a separate stack. It allows the architecture to switch the
 9378 * notion of the current task on a CPU in a non-blocking manner. This function
 9379 * must be called with all CPU's synchronized, and interrupts disabled, the
 9380 * and caller must save the original value of the current task (see
 9381 * curr_task() above) and restore that value before reenabling interrupts and
 9382 * re-starting the system.
 9383 *
 9384 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
 9385 */
 9386void ia64_set_curr_task(int cpu, struct task_struct *p)
 9387{
 9388	cpu_curr(cpu) = p;
 9389}
 9390
 9391#endif
 9392
 9393#ifdef CONFIG_CGROUP_SCHED
 9394/* task_group_lock serializes the addition/removal of task groups */
 9395static DEFINE_SPINLOCK(task_group_lock);
 9396
 9397static inline void alloc_uclamp_sched_group(struct task_group *tg,
 9398					    struct task_group *parent)
 9399{
 9400#ifdef CONFIG_UCLAMP_TASK_GROUP
 9401	enum uclamp_id clamp_id;
 9402
 9403	for_each_clamp_id(clamp_id) {
 9404		uclamp_se_set(&tg->uclamp_req[clamp_id],
 9405			      uclamp_none(clamp_id), false);
 9406		tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
 9407	}
 9408#endif
 9409}
 9410
 9411static void sched_free_group(struct task_group *tg)
 9412{
 9413	free_fair_sched_group(tg);
 9414	free_rt_sched_group(tg);
 9415	autogroup_free(tg);
 9416	kmem_cache_free(task_group_cache, tg);
 9417}
 9418
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 9419/* allocate runqueue etc for a new task group */
 9420struct task_group *sched_create_group(struct task_group *parent)
 9421{
 9422	struct task_group *tg;
 9423
 9424	tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
 9425	if (!tg)
 9426		return ERR_PTR(-ENOMEM);
 9427
 9428	if (!alloc_fair_sched_group(tg, parent))
 9429		goto err;
 9430
 9431	if (!alloc_rt_sched_group(tg, parent))
 9432		goto err;
 9433
 
 9434	alloc_uclamp_sched_group(tg, parent);
 9435
 9436	return tg;
 9437
 9438err:
 9439	sched_free_group(tg);
 9440	return ERR_PTR(-ENOMEM);
 9441}
 9442
 9443void sched_online_group(struct task_group *tg, struct task_group *parent)
 9444{
 9445	unsigned long flags;
 9446
 9447	spin_lock_irqsave(&task_group_lock, flags);
 9448	list_add_rcu(&tg->list, &task_groups);
 9449
 9450	/* Root should already exist: */
 9451	WARN_ON(!parent);
 9452
 9453	tg->parent = parent;
 9454	INIT_LIST_HEAD(&tg->children);
 9455	list_add_rcu(&tg->siblings, &parent->children);
 9456	spin_unlock_irqrestore(&task_group_lock, flags);
 9457
 9458	online_fair_sched_group(tg);
 9459}
 9460
 9461/* rcu callback to free various structures associated with a task group */
 9462static void sched_free_group_rcu(struct rcu_head *rhp)
 9463{
 9464	/* Now it should be safe to free those cfs_rqs: */
 9465	sched_free_group(container_of(rhp, struct task_group, rcu));
 9466}
 9467
 9468void sched_destroy_group(struct task_group *tg)
 9469{
 9470	/* Wait for possible concurrent references to cfs_rqs complete: */
 9471	call_rcu(&tg->rcu, sched_free_group_rcu);
 9472}
 9473
 9474void sched_offline_group(struct task_group *tg)
 9475{
 9476	unsigned long flags;
 9477
 9478	/* End participation in shares distribution: */
 9479	unregister_fair_sched_group(tg);
 9480
 
 
 
 
 
 
 
 
 
 
 9481	spin_lock_irqsave(&task_group_lock, flags);
 9482	list_del_rcu(&tg->list);
 9483	list_del_rcu(&tg->siblings);
 9484	spin_unlock_irqrestore(&task_group_lock, flags);
 9485}
 9486
 9487static void sched_change_group(struct task_struct *tsk, int type)
 9488{
 9489	struct task_group *tg;
 9490
 9491	/*
 9492	 * All callers are synchronized by task_rq_lock(); we do not use RCU
 9493	 * which is pointless here. Thus, we pass "true" to task_css_check()
 9494	 * to prevent lockdep warnings.
 9495	 */
 9496	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
 9497			  struct task_group, css);
 9498	tg = autogroup_task_group(tsk, tg);
 9499	tsk->sched_task_group = tg;
 
 
 
 
 
 
 9500
 9501#ifdef CONFIG_FAIR_GROUP_SCHED
 9502	if (tsk->sched_class->task_change_group)
 9503		tsk->sched_class->task_change_group(tsk, type);
 9504	else
 9505#endif
 9506		set_task_rq(tsk, task_cpu(tsk));
 9507}
 9508
 9509/*
 9510 * Change task's runqueue when it moves between groups.
 9511 *
 9512 * The caller of this function should have put the task in its new group by
 9513 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
 9514 * its new group.
 9515 */
 9516void sched_move_task(struct task_struct *tsk)
 9517{
 9518	int queued, running, queue_flags =
 9519		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
 9520	struct rq_flags rf;
 9521	struct rq *rq;
 9522
 9523	rq = task_rq_lock(tsk, &rf);
 
 
 
 
 
 
 
 
 
 
 9524	update_rq_clock(rq);
 9525
 9526	running = task_current(rq, tsk);
 9527	queued = task_on_rq_queued(tsk);
 9528
 9529	if (queued)
 9530		dequeue_task(rq, tsk, queue_flags);
 9531	if (running)
 9532		put_prev_task(rq, tsk);
 9533
 9534	sched_change_group(tsk, TASK_MOVE_GROUP);
 
 
 9535
 9536	if (queued)
 9537		enqueue_task(rq, tsk, queue_flags);
 9538	if (running) {
 9539		set_next_task(rq, tsk);
 9540		/*
 9541		 * After changing group, the running task may have joined a
 9542		 * throttled one but it's still the running task. Trigger a
 9543		 * resched to make sure that task can still run.
 9544		 */
 9545		resched_curr(rq);
 9546	}
 9547
 9548	task_rq_unlock(rq, tsk, &rf);
 9549}
 9550
 9551static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
 9552{
 9553	return css ? container_of(css, struct task_group, css) : NULL;
 9554}
 9555
 9556static struct cgroup_subsys_state *
 9557cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
 9558{
 9559	struct task_group *parent = css_tg(parent_css);
 9560	struct task_group *tg;
 9561
 9562	if (!parent) {
 9563		/* This is early initialization for the top cgroup */
 9564		return &root_task_group.css;
 9565	}
 9566
 9567	tg = sched_create_group(parent);
 9568	if (IS_ERR(tg))
 9569		return ERR_PTR(-ENOMEM);
 9570
 9571	return &tg->css;
 9572}
 9573
 9574/* Expose task group only after completing cgroup initialization */
 9575static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
 9576{
 9577	struct task_group *tg = css_tg(css);
 9578	struct task_group *parent = css_tg(css->parent);
 
 
 
 
 
 9579
 9580	if (parent)
 9581		sched_online_group(tg, parent);
 9582
 9583#ifdef CONFIG_UCLAMP_TASK_GROUP
 9584	/* Propagate the effective uclamp value for the new group */
 9585	mutex_lock(&uclamp_mutex);
 9586	rcu_read_lock();
 9587	cpu_util_update_eff(css);
 9588	rcu_read_unlock();
 9589	mutex_unlock(&uclamp_mutex);
 9590#endif
 9591
 9592	return 0;
 9593}
 9594
 
 
 
 
 
 
 
 9595static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
 9596{
 9597	struct task_group *tg = css_tg(css);
 9598
 9599	sched_offline_group(tg);
 9600}
 9601
 9602static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
 9603{
 9604	struct task_group *tg = css_tg(css);
 9605
 9606	/*
 9607	 * Relies on the RCU grace period between css_released() and this.
 9608	 */
 9609	sched_free_group(tg);
 9610}
 9611
 9612/*
 9613 * This is called before wake_up_new_task(), therefore we really only
 9614 * have to set its group bits, all the other stuff does not apply.
 9615 */
 9616static void cpu_cgroup_fork(struct task_struct *task)
 9617{
 9618	struct rq_flags rf;
 9619	struct rq *rq;
 9620
 9621	rq = task_rq_lock(task, &rf);
 9622
 9623	update_rq_clock(rq);
 9624	sched_change_group(task, TASK_SET_GROUP);
 9625
 9626	task_rq_unlock(rq, task, &rf);
 9627}
 9628
 9629static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
 9630{
 
 9631	struct task_struct *task;
 9632	struct cgroup_subsys_state *css;
 9633	int ret = 0;
 9634
 9635	cgroup_taskset_for_each(task, css, tset) {
 9636#ifdef CONFIG_RT_GROUP_SCHED
 9637		if (!sched_rt_can_attach(css_tg(css), task))
 9638			return -EINVAL;
 
 9639#endif
 9640		/*
 9641		 * Serialize against wake_up_new_task() such that if it's
 9642		 * running, we're sure to observe its full state.
 9643		 */
 9644		raw_spin_lock_irq(&task->pi_lock);
 9645		/*
 9646		 * Avoid calling sched_move_task() before wake_up_new_task()
 9647		 * has happened. This would lead to problems with PELT, due to
 9648		 * move wanting to detach+attach while we're not attached yet.
 9649		 */
 9650		if (READ_ONCE(task->__state) == TASK_NEW)
 9651			ret = -EINVAL;
 9652		raw_spin_unlock_irq(&task->pi_lock);
 9653
 9654		if (ret)
 9655			break;
 9656	}
 9657	return ret;
 9658}
 9659
 9660static void cpu_cgroup_attach(struct cgroup_taskset *tset)
 9661{
 9662	struct task_struct *task;
 9663	struct cgroup_subsys_state *css;
 9664
 9665	cgroup_taskset_for_each(task, css, tset)
 9666		sched_move_task(task);
 
 
 
 
 
 
 
 9667}
 9668
 9669#ifdef CONFIG_UCLAMP_TASK_GROUP
 9670static void cpu_util_update_eff(struct cgroup_subsys_state *css)
 9671{
 9672	struct cgroup_subsys_state *top_css = css;
 9673	struct uclamp_se *uc_parent = NULL;
 9674	struct uclamp_se *uc_se = NULL;
 9675	unsigned int eff[UCLAMP_CNT];
 9676	enum uclamp_id clamp_id;
 9677	unsigned int clamps;
 9678
 9679	lockdep_assert_held(&uclamp_mutex);
 9680	SCHED_WARN_ON(!rcu_read_lock_held());
 9681
 9682	css_for_each_descendant_pre(css, top_css) {
 9683		uc_parent = css_tg(css)->parent
 9684			? css_tg(css)->parent->uclamp : NULL;
 9685
 9686		for_each_clamp_id(clamp_id) {
 9687			/* Assume effective clamps matches requested clamps */
 9688			eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
 9689			/* Cap effective clamps with parent's effective clamps */
 9690			if (uc_parent &&
 9691			    eff[clamp_id] > uc_parent[clamp_id].value) {
 9692				eff[clamp_id] = uc_parent[clamp_id].value;
 9693			}
 9694		}
 9695		/* Ensure protection is always capped by limit */
 9696		eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
 9697
 9698		/* Propagate most restrictive effective clamps */
 9699		clamps = 0x0;
 9700		uc_se = css_tg(css)->uclamp;
 9701		for_each_clamp_id(clamp_id) {
 9702			if (eff[clamp_id] == uc_se[clamp_id].value)
 9703				continue;
 9704			uc_se[clamp_id].value = eff[clamp_id];
 9705			uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
 9706			clamps |= (0x1 << clamp_id);
 9707		}
 9708		if (!clamps) {
 9709			css = css_rightmost_descendant(css);
 9710			continue;
 9711		}
 9712
 9713		/* Immediately update descendants RUNNABLE tasks */
 9714		uclamp_update_active_tasks(css);
 9715	}
 9716}
 9717
 9718/*
 9719 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
 9720 * C expression. Since there is no way to convert a macro argument (N) into a
 9721 * character constant, use two levels of macros.
 9722 */
 9723#define _POW10(exp) ((unsigned int)1e##exp)
 9724#define POW10(exp) _POW10(exp)
 9725
 9726struct uclamp_request {
 9727#define UCLAMP_PERCENT_SHIFT	2
 9728#define UCLAMP_PERCENT_SCALE	(100 * POW10(UCLAMP_PERCENT_SHIFT))
 9729	s64 percent;
 9730	u64 util;
 9731	int ret;
 9732};
 9733
 9734static inline struct uclamp_request
 9735capacity_from_percent(char *buf)
 9736{
 9737	struct uclamp_request req = {
 9738		.percent = UCLAMP_PERCENT_SCALE,
 9739		.util = SCHED_CAPACITY_SCALE,
 9740		.ret = 0,
 9741	};
 9742
 9743	buf = strim(buf);
 9744	if (strcmp(buf, "max")) {
 9745		req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
 9746					     &req.percent);
 9747		if (req.ret)
 9748			return req;
 9749		if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
 9750			req.ret = -ERANGE;
 9751			return req;
 9752		}
 9753
 9754		req.util = req.percent << SCHED_CAPACITY_SHIFT;
 9755		req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
 9756	}
 9757
 9758	return req;
 9759}
 9760
 9761static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
 9762				size_t nbytes, loff_t off,
 9763				enum uclamp_id clamp_id)
 9764{
 9765	struct uclamp_request req;
 9766	struct task_group *tg;
 9767
 9768	req = capacity_from_percent(buf);
 9769	if (req.ret)
 9770		return req.ret;
 9771
 9772	static_branch_enable(&sched_uclamp_used);
 9773
 9774	mutex_lock(&uclamp_mutex);
 9775	rcu_read_lock();
 9776
 9777	tg = css_tg(of_css(of));
 9778	if (tg->uclamp_req[clamp_id].value != req.util)
 9779		uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
 9780
 9781	/*
 9782	 * Because of not recoverable conversion rounding we keep track of the
 9783	 * exact requested value
 9784	 */
 9785	tg->uclamp_pct[clamp_id] = req.percent;
 9786
 9787	/* Update effective clamps to track the most restrictive value */
 9788	cpu_util_update_eff(of_css(of));
 9789
 9790	rcu_read_unlock();
 9791	mutex_unlock(&uclamp_mutex);
 9792
 9793	return nbytes;
 9794}
 9795
 9796static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
 9797				    char *buf, size_t nbytes,
 9798				    loff_t off)
 9799{
 9800	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
 9801}
 9802
 9803static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
 9804				    char *buf, size_t nbytes,
 9805				    loff_t off)
 9806{
 9807	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
 9808}
 9809
 9810static inline void cpu_uclamp_print(struct seq_file *sf,
 9811				    enum uclamp_id clamp_id)
 9812{
 9813	struct task_group *tg;
 9814	u64 util_clamp;
 9815	u64 percent;
 9816	u32 rem;
 9817
 9818	rcu_read_lock();
 9819	tg = css_tg(seq_css(sf));
 9820	util_clamp = tg->uclamp_req[clamp_id].value;
 9821	rcu_read_unlock();
 9822
 9823	if (util_clamp == SCHED_CAPACITY_SCALE) {
 9824		seq_puts(sf, "max\n");
 9825		return;
 9826	}
 9827
 9828	percent = tg->uclamp_pct[clamp_id];
 9829	percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
 9830	seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
 9831}
 9832
 9833static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
 9834{
 9835	cpu_uclamp_print(sf, UCLAMP_MIN);
 9836	return 0;
 9837}
 9838
 9839static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
 9840{
 9841	cpu_uclamp_print(sf, UCLAMP_MAX);
 9842	return 0;
 9843}
 9844#endif /* CONFIG_UCLAMP_TASK_GROUP */
 9845
 
 
 
 9846#ifdef CONFIG_FAIR_GROUP_SCHED
 
 
 
 
 
 
 9847static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
 9848				struct cftype *cftype, u64 shareval)
 9849{
 
 
 9850	if (shareval > scale_load_down(ULONG_MAX))
 9851		shareval = MAX_SHARES;
 9852	return sched_group_set_shares(css_tg(css), scale_load(shareval));
 
 
 
 
 9853}
 9854
 9855static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
 9856			       struct cftype *cft)
 9857{
 9858	struct task_group *tg = css_tg(css);
 9859
 9860	return (u64) scale_load_down(tg->shares);
 9861}
 
 9862
 9863#ifdef CONFIG_CFS_BANDWIDTH
 9864static DEFINE_MUTEX(cfs_constraints_mutex);
 9865
 9866const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
 9867static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
 9868/* More than 203 days if BW_SHIFT equals 20. */
 9869static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
 9870
 9871static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
 9872
 9873static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
 9874				u64 burst)
 9875{
 9876	int i, ret = 0, runtime_enabled, runtime_was_enabled;
 9877	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
 9878
 9879	if (tg == &root_task_group)
 9880		return -EINVAL;
 9881
 9882	/*
 9883	 * Ensure we have at some amount of bandwidth every period.  This is
 9884	 * to prevent reaching a state of large arrears when throttled via
 9885	 * entity_tick() resulting in prolonged exit starvation.
 9886	 */
 9887	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
 9888		return -EINVAL;
 9889
 9890	/*
 9891	 * Likewise, bound things on the other side by preventing insane quota
 9892	 * periods.  This also allows us to normalize in computing quota
 9893	 * feasibility.
 9894	 */
 9895	if (period > max_cfs_quota_period)
 9896		return -EINVAL;
 9897
 9898	/*
 9899	 * Bound quota to defend quota against overflow during bandwidth shift.
 9900	 */
 9901	if (quota != RUNTIME_INF && quota > max_cfs_runtime)
 9902		return -EINVAL;
 9903
 9904	if (quota != RUNTIME_INF && (burst > quota ||
 9905				     burst + quota > max_cfs_runtime))
 9906		return -EINVAL;
 9907
 9908	/*
 9909	 * Prevent race between setting of cfs_rq->runtime_enabled and
 9910	 * unthrottle_offline_cfs_rqs().
 9911	 */
 9912	get_online_cpus();
 9913	mutex_lock(&cfs_constraints_mutex);
 
 9914	ret = __cfs_schedulable(tg, period, quota);
 9915	if (ret)
 9916		goto out_unlock;
 9917
 9918	runtime_enabled = quota != RUNTIME_INF;
 9919	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
 9920	/*
 9921	 * If we need to toggle cfs_bandwidth_used, off->on must occur
 9922	 * before making related changes, and on->off must occur afterwards
 9923	 */
 9924	if (runtime_enabled && !runtime_was_enabled)
 9925		cfs_bandwidth_usage_inc();
 9926	raw_spin_lock_irq(&cfs_b->lock);
 9927	cfs_b->period = ns_to_ktime(period);
 9928	cfs_b->quota = quota;
 9929	cfs_b->burst = burst;
 9930
 9931	__refill_cfs_bandwidth_runtime(cfs_b);
 9932
 9933	/* Restart the period timer (if active) to handle new period expiry: */
 9934	if (runtime_enabled)
 9935		start_cfs_bandwidth(cfs_b);
 9936
 9937	raw_spin_unlock_irq(&cfs_b->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 9938
 9939	for_each_online_cpu(i) {
 9940		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
 9941		struct rq *rq = cfs_rq->rq;
 9942		struct rq_flags rf;
 9943
 9944		rq_lock_irq(rq, &rf);
 9945		cfs_rq->runtime_enabled = runtime_enabled;
 9946		cfs_rq->runtime_remaining = 0;
 9947
 9948		if (cfs_rq->throttled)
 9949			unthrottle_cfs_rq(cfs_rq);
 9950		rq_unlock_irq(rq, &rf);
 9951	}
 
 9952	if (runtime_was_enabled && !runtime_enabled)
 9953		cfs_bandwidth_usage_dec();
 9954out_unlock:
 9955	mutex_unlock(&cfs_constraints_mutex);
 9956	put_online_cpus();
 9957
 9958	return ret;
 9959}
 9960
 9961static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
 9962{
 9963	u64 quota, period, burst;
 9964
 9965	period = ktime_to_ns(tg->cfs_bandwidth.period);
 9966	burst = tg->cfs_bandwidth.burst;
 9967	if (cfs_quota_us < 0)
 9968		quota = RUNTIME_INF;
 9969	else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
 9970		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
 9971	else
 9972		return -EINVAL;
 9973
 9974	return tg_set_cfs_bandwidth(tg, period, quota, burst);
 9975}
 9976
 9977static long tg_get_cfs_quota(struct task_group *tg)
 9978{
 9979	u64 quota_us;
 9980
 9981	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
 9982		return -1;
 9983
 9984	quota_us = tg->cfs_bandwidth.quota;
 9985	do_div(quota_us, NSEC_PER_USEC);
 9986
 9987	return quota_us;
 9988}
 9989
 9990static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
 9991{
 9992	u64 quota, period, burst;
 9993
 9994	if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
 9995		return -EINVAL;
 9996
 9997	period = (u64)cfs_period_us * NSEC_PER_USEC;
 9998	quota = tg->cfs_bandwidth.quota;
 9999	burst = tg->cfs_bandwidth.burst;
10000
10001	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10002}
10003
10004static long tg_get_cfs_period(struct task_group *tg)
10005{
10006	u64 cfs_period_us;
10007
10008	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10009	do_div(cfs_period_us, NSEC_PER_USEC);
10010
10011	return cfs_period_us;
10012}
10013
10014static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10015{
10016	u64 quota, period, burst;
10017
10018	if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10019		return -EINVAL;
10020
10021	burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10022	period = ktime_to_ns(tg->cfs_bandwidth.period);
10023	quota = tg->cfs_bandwidth.quota;
10024
10025	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10026}
10027
10028static long tg_get_cfs_burst(struct task_group *tg)
10029{
10030	u64 burst_us;
10031
10032	burst_us = tg->cfs_bandwidth.burst;
10033	do_div(burst_us, NSEC_PER_USEC);
10034
10035	return burst_us;
10036}
10037
10038static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10039				  struct cftype *cft)
10040{
10041	return tg_get_cfs_quota(css_tg(css));
10042}
10043
10044static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10045				   struct cftype *cftype, s64 cfs_quota_us)
10046{
10047	return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10048}
10049
10050static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10051				   struct cftype *cft)
10052{
10053	return tg_get_cfs_period(css_tg(css));
10054}
10055
10056static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10057				    struct cftype *cftype, u64 cfs_period_us)
10058{
10059	return tg_set_cfs_period(css_tg(css), cfs_period_us);
10060}
10061
10062static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10063				  struct cftype *cft)
10064{
10065	return tg_get_cfs_burst(css_tg(css));
10066}
10067
10068static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10069				   struct cftype *cftype, u64 cfs_burst_us)
10070{
10071	return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10072}
10073
10074struct cfs_schedulable_data {
10075	struct task_group *tg;
10076	u64 period, quota;
10077};
10078
10079/*
10080 * normalize group quota/period to be quota/max_period
10081 * note: units are usecs
10082 */
10083static u64 normalize_cfs_quota(struct task_group *tg,
10084			       struct cfs_schedulable_data *d)
10085{
10086	u64 quota, period;
10087
10088	if (tg == d->tg) {
10089		period = d->period;
10090		quota = d->quota;
10091	} else {
10092		period = tg_get_cfs_period(tg);
10093		quota = tg_get_cfs_quota(tg);
10094	}
10095
10096	/* note: these should typically be equivalent */
10097	if (quota == RUNTIME_INF || quota == -1)
10098		return RUNTIME_INF;
10099
10100	return to_ratio(period, quota);
10101}
10102
10103static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10104{
10105	struct cfs_schedulable_data *d = data;
10106	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10107	s64 quota = 0, parent_quota = -1;
10108
10109	if (!tg->parent) {
10110		quota = RUNTIME_INF;
10111	} else {
10112		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10113
10114		quota = normalize_cfs_quota(tg, d);
10115		parent_quota = parent_b->hierarchical_quota;
10116
10117		/*
10118		 * Ensure max(child_quota) <= parent_quota.  On cgroup2,
10119		 * always take the min.  On cgroup1, only inherit when no
10120		 * limit is set:
 
 
10121		 */
10122		if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10123			quota = min(quota, parent_quota);
 
 
 
10124		} else {
10125			if (quota == RUNTIME_INF)
10126				quota = parent_quota;
10127			else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10128				return -EINVAL;
10129		}
10130	}
10131	cfs_b->hierarchical_quota = quota;
10132
10133	return 0;
10134}
10135
10136static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10137{
10138	int ret;
10139	struct cfs_schedulable_data data = {
10140		.tg = tg,
10141		.period = period,
10142		.quota = quota,
10143	};
10144
10145	if (quota != RUNTIME_INF) {
10146		do_div(data.period, NSEC_PER_USEC);
10147		do_div(data.quota, NSEC_PER_USEC);
10148	}
10149
10150	rcu_read_lock();
10151	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10152	rcu_read_unlock();
10153
10154	return ret;
10155}
10156
10157static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10158{
10159	struct task_group *tg = css_tg(seq_css(sf));
10160	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10161
10162	seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10163	seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10164	seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10165
10166	if (schedstat_enabled() && tg != &root_task_group) {
 
10167		u64 ws = 0;
10168		int i;
10169
10170		for_each_possible_cpu(i)
10171			ws += schedstat_val(tg->se[i]->statistics.wait_sum);
 
 
10172
10173		seq_printf(sf, "wait_sum %llu\n", ws);
10174	}
10175
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
10176	return 0;
10177}
10178#endif /* CONFIG_CFS_BANDWIDTH */
10179#endif /* CONFIG_FAIR_GROUP_SCHED */
10180
10181#ifdef CONFIG_RT_GROUP_SCHED
10182static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10183				struct cftype *cft, s64 val)
10184{
10185	return sched_group_set_rt_runtime(css_tg(css), val);
10186}
10187
10188static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10189			       struct cftype *cft)
10190{
10191	return sched_group_rt_runtime(css_tg(css));
10192}
10193
10194static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10195				    struct cftype *cftype, u64 rt_period_us)
10196{
10197	return sched_group_set_rt_period(css_tg(css), rt_period_us);
10198}
10199
10200static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10201				   struct cftype *cft)
10202{
10203	return sched_group_rt_period(css_tg(css));
10204}
10205#endif /* CONFIG_RT_GROUP_SCHED */
10206
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
10207static struct cftype cpu_legacy_files[] = {
10208#ifdef CONFIG_FAIR_GROUP_SCHED
10209	{
10210		.name = "shares",
10211		.read_u64 = cpu_shares_read_u64,
10212		.write_u64 = cpu_shares_write_u64,
10213	},
 
 
 
 
 
10214#endif
10215#ifdef CONFIG_CFS_BANDWIDTH
10216	{
10217		.name = "cfs_quota_us",
10218		.read_s64 = cpu_cfs_quota_read_s64,
10219		.write_s64 = cpu_cfs_quota_write_s64,
10220	},
10221	{
10222		.name = "cfs_period_us",
10223		.read_u64 = cpu_cfs_period_read_u64,
10224		.write_u64 = cpu_cfs_period_write_u64,
10225	},
10226	{
10227		.name = "cfs_burst_us",
10228		.read_u64 = cpu_cfs_burst_read_u64,
10229		.write_u64 = cpu_cfs_burst_write_u64,
10230	},
10231	{
10232		.name = "stat",
10233		.seq_show = cpu_cfs_stat_show,
10234	},
 
 
 
 
10235#endif
10236#ifdef CONFIG_RT_GROUP_SCHED
10237	{
10238		.name = "rt_runtime_us",
10239		.read_s64 = cpu_rt_runtime_read,
10240		.write_s64 = cpu_rt_runtime_write,
10241	},
10242	{
10243		.name = "rt_period_us",
10244		.read_u64 = cpu_rt_period_read_uint,
10245		.write_u64 = cpu_rt_period_write_uint,
10246	},
10247#endif
10248#ifdef CONFIG_UCLAMP_TASK_GROUP
10249	{
10250		.name = "uclamp.min",
10251		.flags = CFTYPE_NOT_ON_ROOT,
10252		.seq_show = cpu_uclamp_min_show,
10253		.write = cpu_uclamp_min_write,
10254	},
10255	{
10256		.name = "uclamp.max",
10257		.flags = CFTYPE_NOT_ON_ROOT,
10258		.seq_show = cpu_uclamp_max_show,
10259		.write = cpu_uclamp_max_write,
10260	},
10261#endif
10262	{ }	/* Terminate */
10263};
10264
10265static int cpu_extra_stat_show(struct seq_file *sf,
10266			       struct cgroup_subsys_state *css)
10267{
10268#ifdef CONFIG_CFS_BANDWIDTH
10269	{
10270		struct task_group *tg = css_tg(css);
10271		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10272		u64 throttled_usec;
10273
10274		throttled_usec = cfs_b->throttled_time;
10275		do_div(throttled_usec, NSEC_PER_USEC);
 
 
10276
10277		seq_printf(sf, "nr_periods %d\n"
10278			   "nr_throttled %d\n"
10279			   "throttled_usec %llu\n",
 
 
10280			   cfs_b->nr_periods, cfs_b->nr_throttled,
10281			   throttled_usec);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
10282	}
10283#endif
10284	return 0;
10285}
10286
10287#ifdef CONFIG_FAIR_GROUP_SCHED
 
10288static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10289			       struct cftype *cft)
10290{
10291	struct task_group *tg = css_tg(css);
10292	u64 weight = scale_load_down(tg->shares);
10293
10294	return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10295}
10296
10297static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10298				struct cftype *cft, u64 weight)
10299{
10300	/*
10301	 * cgroup weight knobs should use the common MIN, DFL and MAX
10302	 * values which are 1, 100 and 10000 respectively.  While it loses
10303	 * a bit of range on both ends, it maps pretty well onto the shares
10304	 * value used by scheduler and the round-trip conversions preserve
10305	 * the original value over the entire range.
10306	 */
10307	if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10308		return -ERANGE;
10309
10310	weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10311
10312	return sched_group_set_shares(css_tg(css), scale_load(weight));
 
 
 
10313}
10314
10315static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10316				    struct cftype *cft)
10317{
10318	unsigned long weight = scale_load_down(css_tg(css)->shares);
10319	int last_delta = INT_MAX;
10320	int prio, delta;
10321
10322	/* find the closest nice value to the current weight */
10323	for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10324		delta = abs(sched_prio_to_weight[prio] - weight);
10325		if (delta >= last_delta)
10326			break;
10327		last_delta = delta;
10328	}
10329
10330	return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10331}
10332
10333static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10334				     struct cftype *cft, s64 nice)
10335{
10336	unsigned long weight;
10337	int idx;
10338
10339	if (nice < MIN_NICE || nice > MAX_NICE)
10340		return -ERANGE;
10341
10342	idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10343	idx = array_index_nospec(idx, 40);
10344	weight = sched_prio_to_weight[idx];
10345
10346	return sched_group_set_shares(css_tg(css), scale_load(weight));
 
 
 
 
10347}
10348#endif
10349
10350static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10351						  long period, long quota)
10352{
10353	if (quota < 0)
10354		seq_puts(sf, "max");
10355	else
10356		seq_printf(sf, "%ld", quota);
10357
10358	seq_printf(sf, " %ld\n", period);
10359}
10360
10361/* caller should put the current value in *@periodp before calling */
10362static int __maybe_unused cpu_period_quota_parse(char *buf,
10363						 u64 *periodp, u64 *quotap)
10364{
10365	char tok[21];	/* U64_MAX */
10366
10367	if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10368		return -EINVAL;
10369
10370	*periodp *= NSEC_PER_USEC;
10371
10372	if (sscanf(tok, "%llu", quotap))
10373		*quotap *= NSEC_PER_USEC;
10374	else if (!strcmp(tok, "max"))
10375		*quotap = RUNTIME_INF;
10376	else
10377		return -EINVAL;
10378
10379	return 0;
10380}
10381
10382#ifdef CONFIG_CFS_BANDWIDTH
10383static int cpu_max_show(struct seq_file *sf, void *v)
10384{
10385	struct task_group *tg = css_tg(seq_css(sf));
10386
10387	cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10388	return 0;
10389}
10390
10391static ssize_t cpu_max_write(struct kernfs_open_file *of,
10392			     char *buf, size_t nbytes, loff_t off)
10393{
10394	struct task_group *tg = css_tg(of_css(of));
10395	u64 period = tg_get_cfs_period(tg);
10396	u64 burst = tg_get_cfs_burst(tg);
10397	u64 quota;
10398	int ret;
10399
10400	ret = cpu_period_quota_parse(buf, &period, &quota);
10401	if (!ret)
10402		ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10403	return ret ?: nbytes;
10404}
10405#endif
10406
10407static struct cftype cpu_files[] = {
10408#ifdef CONFIG_FAIR_GROUP_SCHED
10409	{
10410		.name = "weight",
10411		.flags = CFTYPE_NOT_ON_ROOT,
10412		.read_u64 = cpu_weight_read_u64,
10413		.write_u64 = cpu_weight_write_u64,
10414	},
10415	{
10416		.name = "weight.nice",
10417		.flags = CFTYPE_NOT_ON_ROOT,
10418		.read_s64 = cpu_weight_nice_read_s64,
10419		.write_s64 = cpu_weight_nice_write_s64,
10420	},
 
 
 
 
 
 
10421#endif
10422#ifdef CONFIG_CFS_BANDWIDTH
10423	{
10424		.name = "max",
10425		.flags = CFTYPE_NOT_ON_ROOT,
10426		.seq_show = cpu_max_show,
10427		.write = cpu_max_write,
10428	},
10429	{
10430		.name = "max.burst",
10431		.flags = CFTYPE_NOT_ON_ROOT,
10432		.read_u64 = cpu_cfs_burst_read_u64,
10433		.write_u64 = cpu_cfs_burst_write_u64,
10434	},
10435#endif
10436#ifdef CONFIG_UCLAMP_TASK_GROUP
10437	{
10438		.name = "uclamp.min",
10439		.flags = CFTYPE_NOT_ON_ROOT,
10440		.seq_show = cpu_uclamp_min_show,
10441		.write = cpu_uclamp_min_write,
10442	},
10443	{
10444		.name = "uclamp.max",
10445		.flags = CFTYPE_NOT_ON_ROOT,
10446		.seq_show = cpu_uclamp_max_show,
10447		.write = cpu_uclamp_max_write,
10448	},
10449#endif
10450	{ }	/* terminate */
10451};
10452
10453struct cgroup_subsys cpu_cgrp_subsys = {
10454	.css_alloc	= cpu_cgroup_css_alloc,
10455	.css_online	= cpu_cgroup_css_online,
 
10456	.css_released	= cpu_cgroup_css_released,
10457	.css_free	= cpu_cgroup_css_free,
10458	.css_extra_stat_show = cpu_extra_stat_show,
10459	.fork		= cpu_cgroup_fork,
10460	.can_attach	= cpu_cgroup_can_attach,
10461	.attach		= cpu_cgroup_attach,
 
10462	.legacy_cftypes	= cpu_legacy_files,
10463	.dfl_cftypes	= cpu_files,
10464	.early_init	= true,
10465	.threaded	= true,
10466};
10467
10468#endif	/* CONFIG_CGROUP_SCHED */
10469
10470void dump_cpu_task(int cpu)
10471{
 
 
 
 
 
 
 
 
 
 
 
 
 
10472	pr_info("Task dump for CPU %d:\n", cpu);
10473	sched_show_task(cpu_curr(cpu));
10474}
10475
10476/*
10477 * Nice levels are multiplicative, with a gentle 10% change for every
10478 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10479 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10480 * that remained on nice 0.
10481 *
10482 * The "10% effect" is relative and cumulative: from _any_ nice level,
10483 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10484 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10485 * If a task goes up by ~10% and another task goes down by ~10% then
10486 * the relative distance between them is ~25%.)
10487 */
10488const int sched_prio_to_weight[40] = {
10489 /* -20 */     88761,     71755,     56483,     46273,     36291,
10490 /* -15 */     29154,     23254,     18705,     14949,     11916,
10491 /* -10 */      9548,      7620,      6100,      4904,      3906,
10492 /*  -5 */      3121,      2501,      1991,      1586,      1277,
10493 /*   0 */      1024,       820,       655,       526,       423,
10494 /*   5 */       335,       272,       215,       172,       137,
10495 /*  10 */       110,        87,        70,        56,        45,
10496 /*  15 */        36,        29,        23,        18,        15,
10497};
10498
10499/*
10500 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10501 *
10502 * In cases where the weight does not change often, we can use the
10503 * precalculated inverse to speed up arithmetics by turning divisions
10504 * into multiplications:
10505 */
10506const u32 sched_prio_to_wmult[40] = {
10507 /* -20 */     48388,     59856,     76040,     92818,    118348,
10508 /* -15 */    147320,    184698,    229616,    287308,    360437,
10509 /* -10 */    449829,    563644,    704093,    875809,   1099582,
10510 /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
10511 /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
10512 /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
10513 /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
10514 /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10515};
10516
10517void call_trace_sched_update_nr_running(struct rq *rq, int count)
10518{
10519        trace_sched_update_nr_running_tp(rq, count);
10520}