Loading...
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((¶virt_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, ¶m);
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(¬ifier->link, ¤t->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(¬ifier->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, "a);
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 */
1/*
2 * kernel/sched/core.c
3 *
4 * Kernel scheduler and related syscalls
5 *
6 * Copyright (C) 1991-2002 Linus Torvalds
7 *
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
11 * by Andrea Arcangeli
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
27 */
28
29#include <linux/mm.h>
30#include <linux/module.h>
31#include <linux/nmi.h>
32#include <linux/init.h>
33#include <linux/uaccess.h>
34#include <linux/highmem.h>
35#include <asm/mmu_context.h>
36#include <linux/interrupt.h>
37#include <linux/capability.h>
38#include <linux/completion.h>
39#include <linux/kernel_stat.h>
40#include <linux/debug_locks.h>
41#include <linux/perf_event.h>
42#include <linux/security.h>
43#include <linux/notifier.h>
44#include <linux/profile.h>
45#include <linux/freezer.h>
46#include <linux/vmalloc.h>
47#include <linux/blkdev.h>
48#include <linux/delay.h>
49#include <linux/pid_namespace.h>
50#include <linux/smp.h>
51#include <linux/threads.h>
52#include <linux/timer.h>
53#include <linux/rcupdate.h>
54#include <linux/cpu.h>
55#include <linux/cpuset.h>
56#include <linux/percpu.h>
57#include <linux/proc_fs.h>
58#include <linux/seq_file.h>
59#include <linux/sysctl.h>
60#include <linux/syscalls.h>
61#include <linux/times.h>
62#include <linux/tsacct_kern.h>
63#include <linux/kprobes.h>
64#include <linux/delayacct.h>
65#include <linux/unistd.h>
66#include <linux/pagemap.h>
67#include <linux/hrtimer.h>
68#include <linux/tick.h>
69#include <linux/debugfs.h>
70#include <linux/ctype.h>
71#include <linux/ftrace.h>
72#include <linux/slab.h>
73#include <linux/init_task.h>
74#include <linux/binfmts.h>
75
76#include <asm/switch_to.h>
77#include <asm/tlb.h>
78#include <asm/irq_regs.h>
79#include <asm/mutex.h>
80#ifdef CONFIG_PARAVIRT
81#include <asm/paravirt.h>
82#endif
83
84#include "sched.h"
85#include "../workqueue_sched.h"
86#include "../smpboot.h"
87
88#define CREATE_TRACE_POINTS
89#include <trace/events/sched.h>
90
91void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
92{
93 unsigned long delta;
94 ktime_t soft, hard, now;
95
96 for (;;) {
97 if (hrtimer_active(period_timer))
98 break;
99
100 now = hrtimer_cb_get_time(period_timer);
101 hrtimer_forward(period_timer, now, period);
102
103 soft = hrtimer_get_softexpires(period_timer);
104 hard = hrtimer_get_expires(period_timer);
105 delta = ktime_to_ns(ktime_sub(hard, soft));
106 __hrtimer_start_range_ns(period_timer, soft, delta,
107 HRTIMER_MODE_ABS_PINNED, 0);
108 }
109}
110
111DEFINE_MUTEX(sched_domains_mutex);
112DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
113
114static void update_rq_clock_task(struct rq *rq, s64 delta);
115
116void update_rq_clock(struct rq *rq)
117{
118 s64 delta;
119
120 if (rq->skip_clock_update > 0)
121 return;
122
123 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
124 rq->clock += delta;
125 update_rq_clock_task(rq, delta);
126}
127
128/*
129 * Debugging: various feature bits
130 */
131
132#define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
134
135const_debug unsigned int sysctl_sched_features =
136#include "features.h"
137 0;
138
139#undef SCHED_FEAT
140
141#ifdef CONFIG_SCHED_DEBUG
142#define SCHED_FEAT(name, enabled) \
143 #name ,
144
145static const char * const sched_feat_names[] = {
146#include "features.h"
147};
148
149#undef SCHED_FEAT
150
151static int sched_feat_show(struct seq_file *m, void *v)
152{
153 int i;
154
155 for (i = 0; i < __SCHED_FEAT_NR; i++) {
156 if (!(sysctl_sched_features & (1UL << i)))
157 seq_puts(m, "NO_");
158 seq_printf(m, "%s ", sched_feat_names[i]);
159 }
160 seq_puts(m, "\n");
161
162 return 0;
163}
164
165#ifdef HAVE_JUMP_LABEL
166
167#define jump_label_key__true STATIC_KEY_INIT_TRUE
168#define jump_label_key__false STATIC_KEY_INIT_FALSE
169
170#define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
172
173struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
174#include "features.h"
175};
176
177#undef SCHED_FEAT
178
179static void sched_feat_disable(int i)
180{
181 if (static_key_enabled(&sched_feat_keys[i]))
182 static_key_slow_dec(&sched_feat_keys[i]);
183}
184
185static void sched_feat_enable(int i)
186{
187 if (!static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_inc(&sched_feat_keys[i]);
189}
190#else
191static void sched_feat_disable(int i) { };
192static void sched_feat_enable(int i) { };
193#endif /* HAVE_JUMP_LABEL */
194
195static ssize_t
196sched_feat_write(struct file *filp, const char __user *ubuf,
197 size_t cnt, loff_t *ppos)
198{
199 char buf[64];
200 char *cmp;
201 int neg = 0;
202 int i;
203
204 if (cnt > 63)
205 cnt = 63;
206
207 if (copy_from_user(&buf, ubuf, cnt))
208 return -EFAULT;
209
210 buf[cnt] = 0;
211 cmp = strstrip(buf);
212
213 if (strncmp(cmp, "NO_", 3) == 0) {
214 neg = 1;
215 cmp += 3;
216 }
217
218 for (i = 0; i < __SCHED_FEAT_NR; i++) {
219 if (strcmp(cmp, sched_feat_names[i]) == 0) {
220 if (neg) {
221 sysctl_sched_features &= ~(1UL << i);
222 sched_feat_disable(i);
223 } else {
224 sysctl_sched_features |= (1UL << i);
225 sched_feat_enable(i);
226 }
227 break;
228 }
229 }
230
231 if (i == __SCHED_FEAT_NR)
232 return -EINVAL;
233
234 *ppos += cnt;
235
236 return cnt;
237}
238
239static int sched_feat_open(struct inode *inode, struct file *filp)
240{
241 return single_open(filp, sched_feat_show, NULL);
242}
243
244static const struct file_operations sched_feat_fops = {
245 .open = sched_feat_open,
246 .write = sched_feat_write,
247 .read = seq_read,
248 .llseek = seq_lseek,
249 .release = single_release,
250};
251
252static __init int sched_init_debug(void)
253{
254 debugfs_create_file("sched_features", 0644, NULL, NULL,
255 &sched_feat_fops);
256
257 return 0;
258}
259late_initcall(sched_init_debug);
260#endif /* CONFIG_SCHED_DEBUG */
261
262/*
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
265 */
266const_debug unsigned int sysctl_sched_nr_migrate = 32;
267
268/*
269 * period over which we average the RT time consumption, measured
270 * in ms.
271 *
272 * default: 1s
273 */
274const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
275
276/*
277 * period over which we measure -rt task cpu usage in us.
278 * default: 1s
279 */
280unsigned int sysctl_sched_rt_period = 1000000;
281
282__read_mostly int scheduler_running;
283
284/*
285 * part of the period that we allow rt tasks to run in us.
286 * default: 0.95s
287 */
288int sysctl_sched_rt_runtime = 950000;
289
290
291
292/*
293 * __task_rq_lock - lock the rq @p resides on.
294 */
295static inline struct rq *__task_rq_lock(struct task_struct *p)
296 __acquires(rq->lock)
297{
298 struct rq *rq;
299
300 lockdep_assert_held(&p->pi_lock);
301
302 for (;;) {
303 rq = task_rq(p);
304 raw_spin_lock(&rq->lock);
305 if (likely(rq == task_rq(p)))
306 return rq;
307 raw_spin_unlock(&rq->lock);
308 }
309}
310
311/*
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
313 */
314static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
315 __acquires(p->pi_lock)
316 __acquires(rq->lock)
317{
318 struct rq *rq;
319
320 for (;;) {
321 raw_spin_lock_irqsave(&p->pi_lock, *flags);
322 rq = task_rq(p);
323 raw_spin_lock(&rq->lock);
324 if (likely(rq == task_rq(p)))
325 return rq;
326 raw_spin_unlock(&rq->lock);
327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
328 }
329}
330
331static void __task_rq_unlock(struct rq *rq)
332 __releases(rq->lock)
333{
334 raw_spin_unlock(&rq->lock);
335}
336
337static inline void
338task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
339 __releases(rq->lock)
340 __releases(p->pi_lock)
341{
342 raw_spin_unlock(&rq->lock);
343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
344}
345
346/*
347 * this_rq_lock - lock this runqueue and disable interrupts.
348 */
349static struct rq *this_rq_lock(void)
350 __acquires(rq->lock)
351{
352 struct rq *rq;
353
354 local_irq_disable();
355 rq = this_rq();
356 raw_spin_lock(&rq->lock);
357
358 return rq;
359}
360
361#ifdef CONFIG_SCHED_HRTICK
362/*
363 * Use HR-timers to deliver accurate preemption points.
364 *
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
367 * reschedule event.
368 *
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
370 * rq->lock.
371 */
372
373static void hrtick_clear(struct rq *rq)
374{
375 if (hrtimer_active(&rq->hrtick_timer))
376 hrtimer_cancel(&rq->hrtick_timer);
377}
378
379/*
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
382 */
383static enum hrtimer_restart hrtick(struct hrtimer *timer)
384{
385 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
386
387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
388
389 raw_spin_lock(&rq->lock);
390 update_rq_clock(rq);
391 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
392 raw_spin_unlock(&rq->lock);
393
394 return HRTIMER_NORESTART;
395}
396
397#ifdef CONFIG_SMP
398/*
399 * called from hardirq (IPI) context
400 */
401static void __hrtick_start(void *arg)
402{
403 struct rq *rq = arg;
404
405 raw_spin_lock(&rq->lock);
406 hrtimer_restart(&rq->hrtick_timer);
407 rq->hrtick_csd_pending = 0;
408 raw_spin_unlock(&rq->lock);
409}
410
411/*
412 * Called to set the hrtick timer state.
413 *
414 * called with rq->lock held and irqs disabled
415 */
416void hrtick_start(struct rq *rq, u64 delay)
417{
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
420
421 hrtimer_set_expires(timer, time);
422
423 if (rq == this_rq()) {
424 hrtimer_restart(timer);
425 } else if (!rq->hrtick_csd_pending) {
426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
427 rq->hrtick_csd_pending = 1;
428 }
429}
430
431static int
432hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
433{
434 int cpu = (int)(long)hcpu;
435
436 switch (action) {
437 case CPU_UP_CANCELED:
438 case CPU_UP_CANCELED_FROZEN:
439 case CPU_DOWN_PREPARE:
440 case CPU_DOWN_PREPARE_FROZEN:
441 case CPU_DEAD:
442 case CPU_DEAD_FROZEN:
443 hrtick_clear(cpu_rq(cpu));
444 return NOTIFY_OK;
445 }
446
447 return NOTIFY_DONE;
448}
449
450static __init void init_hrtick(void)
451{
452 hotcpu_notifier(hotplug_hrtick, 0);
453}
454#else
455/*
456 * Called to set the hrtick timer state.
457 *
458 * called with rq->lock held and irqs disabled
459 */
460void hrtick_start(struct rq *rq, u64 delay)
461{
462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
463 HRTIMER_MODE_REL_PINNED, 0);
464}
465
466static inline void init_hrtick(void)
467{
468}
469#endif /* CONFIG_SMP */
470
471static void init_rq_hrtick(struct rq *rq)
472{
473#ifdef CONFIG_SMP
474 rq->hrtick_csd_pending = 0;
475
476 rq->hrtick_csd.flags = 0;
477 rq->hrtick_csd.func = __hrtick_start;
478 rq->hrtick_csd.info = rq;
479#endif
480
481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
482 rq->hrtick_timer.function = hrtick;
483}
484#else /* CONFIG_SCHED_HRTICK */
485static inline void hrtick_clear(struct rq *rq)
486{
487}
488
489static inline void init_rq_hrtick(struct rq *rq)
490{
491}
492
493static inline void init_hrtick(void)
494{
495}
496#endif /* CONFIG_SCHED_HRTICK */
497
498/*
499 * resched_task - mark a task 'to be rescheduled now'.
500 *
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
503 * the target CPU.
504 */
505#ifdef CONFIG_SMP
506
507#ifndef tsk_is_polling
508#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
509#endif
510
511void resched_task(struct task_struct *p)
512{
513 int cpu;
514
515 assert_raw_spin_locked(&task_rq(p)->lock);
516
517 if (test_tsk_need_resched(p))
518 return;
519
520 set_tsk_need_resched(p);
521
522 cpu = task_cpu(p);
523 if (cpu == smp_processor_id())
524 return;
525
526 /* NEED_RESCHED must be visible before we test polling */
527 smp_mb();
528 if (!tsk_is_polling(p))
529 smp_send_reschedule(cpu);
530}
531
532void resched_cpu(int cpu)
533{
534 struct rq *rq = cpu_rq(cpu);
535 unsigned long flags;
536
537 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
538 return;
539 resched_task(cpu_curr(cpu));
540 raw_spin_unlock_irqrestore(&rq->lock, flags);
541}
542
543#ifdef CONFIG_NO_HZ
544/*
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu. This is good for power-savings.
547 *
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
551 */
552int get_nohz_timer_target(void)
553{
554 int cpu = smp_processor_id();
555 int i;
556 struct sched_domain *sd;
557
558 rcu_read_lock();
559 for_each_domain(cpu, sd) {
560 for_each_cpu(i, sched_domain_span(sd)) {
561 if (!idle_cpu(i)) {
562 cpu = i;
563 goto unlock;
564 }
565 }
566 }
567unlock:
568 rcu_read_unlock();
569 return cpu;
570}
571/*
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
580 */
581void wake_up_idle_cpu(int cpu)
582{
583 struct rq *rq = cpu_rq(cpu);
584
585 if (cpu == smp_processor_id())
586 return;
587
588 /*
589 * This is safe, as this function is called with the timer
590 * wheel base lock of (cpu) held. When the CPU is on the way
591 * to idle and has not yet set rq->curr to idle then it will
592 * be serialized on the timer wheel base lock and take the new
593 * timer into account automatically.
594 */
595 if (rq->curr != rq->idle)
596 return;
597
598 /*
599 * We can set TIF_RESCHED on the idle task of the other CPU
600 * lockless. The worst case is that the other CPU runs the
601 * idle task through an additional NOOP schedule()
602 */
603 set_tsk_need_resched(rq->idle);
604
605 /* NEED_RESCHED must be visible before we test polling */
606 smp_mb();
607 if (!tsk_is_polling(rq->idle))
608 smp_send_reschedule(cpu);
609}
610
611static inline bool got_nohz_idle_kick(void)
612{
613 int cpu = smp_processor_id();
614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
615}
616
617#else /* CONFIG_NO_HZ */
618
619static inline bool got_nohz_idle_kick(void)
620{
621 return false;
622}
623
624#endif /* CONFIG_NO_HZ */
625
626void sched_avg_update(struct rq *rq)
627{
628 s64 period = sched_avg_period();
629
630 while ((s64)(rq->clock - rq->age_stamp) > period) {
631 /*
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
635 */
636 asm("" : "+rm" (rq->age_stamp));
637 rq->age_stamp += period;
638 rq->rt_avg /= 2;
639 }
640}
641
642#else /* !CONFIG_SMP */
643void resched_task(struct task_struct *p)
644{
645 assert_raw_spin_locked(&task_rq(p)->lock);
646 set_tsk_need_resched(p);
647}
648#endif /* CONFIG_SMP */
649
650#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
652/*
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
655 *
656 * Caller must hold rcu_lock or sufficient equivalent.
657 */
658int walk_tg_tree_from(struct task_group *from,
659 tg_visitor down, tg_visitor up, void *data)
660{
661 struct task_group *parent, *child;
662 int ret;
663
664 parent = from;
665
666down:
667 ret = (*down)(parent, data);
668 if (ret)
669 goto out;
670 list_for_each_entry_rcu(child, &parent->children, siblings) {
671 parent = child;
672 goto down;
673
674up:
675 continue;
676 }
677 ret = (*up)(parent, data);
678 if (ret || parent == from)
679 goto out;
680
681 child = parent;
682 parent = parent->parent;
683 if (parent)
684 goto up;
685out:
686 return ret;
687}
688
689int tg_nop(struct task_group *tg, void *data)
690{
691 return 0;
692}
693#endif
694
695static void set_load_weight(struct task_struct *p)
696{
697 int prio = p->static_prio - MAX_RT_PRIO;
698 struct load_weight *load = &p->se.load;
699
700 /*
701 * SCHED_IDLE tasks get minimal weight:
702 */
703 if (p->policy == SCHED_IDLE) {
704 load->weight = scale_load(WEIGHT_IDLEPRIO);
705 load->inv_weight = WMULT_IDLEPRIO;
706 return;
707 }
708
709 load->weight = scale_load(prio_to_weight[prio]);
710 load->inv_weight = prio_to_wmult[prio];
711}
712
713static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
714{
715 update_rq_clock(rq);
716 sched_info_queued(p);
717 p->sched_class->enqueue_task(rq, p, flags);
718}
719
720static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
721{
722 update_rq_clock(rq);
723 sched_info_dequeued(p);
724 p->sched_class->dequeue_task(rq, p, flags);
725}
726
727void activate_task(struct rq *rq, struct task_struct *p, int flags)
728{
729 if (task_contributes_to_load(p))
730 rq->nr_uninterruptible--;
731
732 enqueue_task(rq, p, flags);
733}
734
735void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
736{
737 if (task_contributes_to_load(p))
738 rq->nr_uninterruptible++;
739
740 dequeue_task(rq, p, flags);
741}
742
743#ifdef CONFIG_IRQ_TIME_ACCOUNTING
744
745/*
746 * There are no locks covering percpu hardirq/softirq time.
747 * They are only modified in account_system_vtime, on corresponding CPU
748 * with interrupts disabled. So, writes are safe.
749 * They are read and saved off onto struct rq in update_rq_clock().
750 * This may result in other CPU reading this CPU's irq time and can
751 * race with irq/account_system_vtime on this CPU. We would either get old
752 * or new value with a side effect of accounting a slice of irq time to wrong
753 * task when irq is in progress while we read rq->clock. That is a worthy
754 * compromise in place of having locks on each irq in account_system_time.
755 */
756static DEFINE_PER_CPU(u64, cpu_hardirq_time);
757static DEFINE_PER_CPU(u64, cpu_softirq_time);
758
759static DEFINE_PER_CPU(u64, irq_start_time);
760static int sched_clock_irqtime;
761
762void enable_sched_clock_irqtime(void)
763{
764 sched_clock_irqtime = 1;
765}
766
767void disable_sched_clock_irqtime(void)
768{
769 sched_clock_irqtime = 0;
770}
771
772#ifndef CONFIG_64BIT
773static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
774
775static inline void irq_time_write_begin(void)
776{
777 __this_cpu_inc(irq_time_seq.sequence);
778 smp_wmb();
779}
780
781static inline void irq_time_write_end(void)
782{
783 smp_wmb();
784 __this_cpu_inc(irq_time_seq.sequence);
785}
786
787static inline u64 irq_time_read(int cpu)
788{
789 u64 irq_time;
790 unsigned seq;
791
792 do {
793 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
794 irq_time = per_cpu(cpu_softirq_time, cpu) +
795 per_cpu(cpu_hardirq_time, cpu);
796 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
797
798 return irq_time;
799}
800#else /* CONFIG_64BIT */
801static inline void irq_time_write_begin(void)
802{
803}
804
805static inline void irq_time_write_end(void)
806{
807}
808
809static inline u64 irq_time_read(int cpu)
810{
811 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
812}
813#endif /* CONFIG_64BIT */
814
815/*
816 * Called before incrementing preempt_count on {soft,}irq_enter
817 * and before decrementing preempt_count on {soft,}irq_exit.
818 */
819void account_system_vtime(struct task_struct *curr)
820{
821 unsigned long flags;
822 s64 delta;
823 int cpu;
824
825 if (!sched_clock_irqtime)
826 return;
827
828 local_irq_save(flags);
829
830 cpu = smp_processor_id();
831 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
832 __this_cpu_add(irq_start_time, delta);
833
834 irq_time_write_begin();
835 /*
836 * We do not account for softirq time from ksoftirqd here.
837 * We want to continue accounting softirq time to ksoftirqd thread
838 * in that case, so as not to confuse scheduler with a special task
839 * that do not consume any time, but still wants to run.
840 */
841 if (hardirq_count())
842 __this_cpu_add(cpu_hardirq_time, delta);
843 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
844 __this_cpu_add(cpu_softirq_time, delta);
845
846 irq_time_write_end();
847 local_irq_restore(flags);
848}
849EXPORT_SYMBOL_GPL(account_system_vtime);
850
851#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
852
853#ifdef CONFIG_PARAVIRT
854static inline u64 steal_ticks(u64 steal)
855{
856 if (unlikely(steal > NSEC_PER_SEC))
857 return div_u64(steal, TICK_NSEC);
858
859 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
860}
861#endif
862
863static void update_rq_clock_task(struct rq *rq, s64 delta)
864{
865/*
866 * In theory, the compile should just see 0 here, and optimize out the call
867 * to sched_rt_avg_update. But I don't trust it...
868 */
869#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
870 s64 steal = 0, irq_delta = 0;
871#endif
872#ifdef CONFIG_IRQ_TIME_ACCOUNTING
873 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
874
875 /*
876 * Since irq_time is only updated on {soft,}irq_exit, we might run into
877 * this case when a previous update_rq_clock() happened inside a
878 * {soft,}irq region.
879 *
880 * When this happens, we stop ->clock_task and only update the
881 * prev_irq_time stamp to account for the part that fit, so that a next
882 * update will consume the rest. This ensures ->clock_task is
883 * monotonic.
884 *
885 * It does however cause some slight miss-attribution of {soft,}irq
886 * time, a more accurate solution would be to update the irq_time using
887 * the current rq->clock timestamp, except that would require using
888 * atomic ops.
889 */
890 if (irq_delta > delta)
891 irq_delta = delta;
892
893 rq->prev_irq_time += irq_delta;
894 delta -= irq_delta;
895#endif
896#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
897 if (static_key_false((¶virt_steal_rq_enabled))) {
898 u64 st;
899
900 steal = paravirt_steal_clock(cpu_of(rq));
901 steal -= rq->prev_steal_time_rq;
902
903 if (unlikely(steal > delta))
904 steal = delta;
905
906 st = steal_ticks(steal);
907 steal = st * TICK_NSEC;
908
909 rq->prev_steal_time_rq += steal;
910
911 delta -= steal;
912 }
913#endif
914
915 rq->clock_task += delta;
916
917#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
918 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
919 sched_rt_avg_update(rq, irq_delta + steal);
920#endif
921}
922
923#ifdef CONFIG_IRQ_TIME_ACCOUNTING
924static int irqtime_account_hi_update(void)
925{
926 u64 *cpustat = kcpustat_this_cpu->cpustat;
927 unsigned long flags;
928 u64 latest_ns;
929 int ret = 0;
930
931 local_irq_save(flags);
932 latest_ns = this_cpu_read(cpu_hardirq_time);
933 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
934 ret = 1;
935 local_irq_restore(flags);
936 return ret;
937}
938
939static int irqtime_account_si_update(void)
940{
941 u64 *cpustat = kcpustat_this_cpu->cpustat;
942 unsigned long flags;
943 u64 latest_ns;
944 int ret = 0;
945
946 local_irq_save(flags);
947 latest_ns = this_cpu_read(cpu_softirq_time);
948 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
949 ret = 1;
950 local_irq_restore(flags);
951 return ret;
952}
953
954#else /* CONFIG_IRQ_TIME_ACCOUNTING */
955
956#define sched_clock_irqtime (0)
957
958#endif
959
960void sched_set_stop_task(int cpu, struct task_struct *stop)
961{
962 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
963 struct task_struct *old_stop = cpu_rq(cpu)->stop;
964
965 if (stop) {
966 /*
967 * Make it appear like a SCHED_FIFO task, its something
968 * userspace knows about and won't get confused about.
969 *
970 * Also, it will make PI more or less work without too
971 * much confusion -- but then, stop work should not
972 * rely on PI working anyway.
973 */
974 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
975
976 stop->sched_class = &stop_sched_class;
977 }
978
979 cpu_rq(cpu)->stop = stop;
980
981 if (old_stop) {
982 /*
983 * Reset it back to a normal scheduling class so that
984 * it can die in pieces.
985 */
986 old_stop->sched_class = &rt_sched_class;
987 }
988}
989
990/*
991 * __normal_prio - return the priority that is based on the static prio
992 */
993static inline int __normal_prio(struct task_struct *p)
994{
995 return p->static_prio;
996}
997
998/*
999 * Calculate the expected normal priority: i.e. priority
1000 * without taking RT-inheritance into account. Might be
1001 * boosted by interactivity modifiers. Changes upon fork,
1002 * setprio syscalls, and whenever the interactivity
1003 * estimator recalculates.
1004 */
1005static inline int normal_prio(struct task_struct *p)
1006{
1007 int prio;
1008
1009 if (task_has_rt_policy(p))
1010 prio = MAX_RT_PRIO-1 - p->rt_priority;
1011 else
1012 prio = __normal_prio(p);
1013 return prio;
1014}
1015
1016/*
1017 * Calculate the current priority, i.e. the priority
1018 * taken into account by the scheduler. This value might
1019 * be boosted by RT tasks, or might be boosted by
1020 * interactivity modifiers. Will be RT if the task got
1021 * RT-boosted. If not then it returns p->normal_prio.
1022 */
1023static int effective_prio(struct task_struct *p)
1024{
1025 p->normal_prio = normal_prio(p);
1026 /*
1027 * If we are RT tasks or we were boosted to RT priority,
1028 * keep the priority unchanged. Otherwise, update priority
1029 * to the normal priority:
1030 */
1031 if (!rt_prio(p->prio))
1032 return p->normal_prio;
1033 return p->prio;
1034}
1035
1036/**
1037 * task_curr - is this task currently executing on a CPU?
1038 * @p: the task in question.
1039 */
1040inline int task_curr(const struct task_struct *p)
1041{
1042 return cpu_curr(task_cpu(p)) == p;
1043}
1044
1045static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1046 const struct sched_class *prev_class,
1047 int oldprio)
1048{
1049 if (prev_class != p->sched_class) {
1050 if (prev_class->switched_from)
1051 prev_class->switched_from(rq, p);
1052 p->sched_class->switched_to(rq, p);
1053 } else if (oldprio != p->prio)
1054 p->sched_class->prio_changed(rq, p, oldprio);
1055}
1056
1057void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1058{
1059 const struct sched_class *class;
1060
1061 if (p->sched_class == rq->curr->sched_class) {
1062 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1063 } else {
1064 for_each_class(class) {
1065 if (class == rq->curr->sched_class)
1066 break;
1067 if (class == p->sched_class) {
1068 resched_task(rq->curr);
1069 break;
1070 }
1071 }
1072 }
1073
1074 /*
1075 * A queue event has occurred, and we're going to schedule. In
1076 * this case, we can save a useless back to back clock update.
1077 */
1078 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1079 rq->skip_clock_update = 1;
1080}
1081
1082#ifdef CONFIG_SMP
1083void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1084{
1085#ifdef CONFIG_SCHED_DEBUG
1086 /*
1087 * We should never call set_task_cpu() on a blocked task,
1088 * ttwu() will sort out the placement.
1089 */
1090 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1091 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1092
1093#ifdef CONFIG_LOCKDEP
1094 /*
1095 * The caller should hold either p->pi_lock or rq->lock, when changing
1096 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1097 *
1098 * sched_move_task() holds both and thus holding either pins the cgroup,
1099 * see task_group().
1100 *
1101 * Furthermore, all task_rq users should acquire both locks, see
1102 * task_rq_lock().
1103 */
1104 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1105 lockdep_is_held(&task_rq(p)->lock)));
1106#endif
1107#endif
1108
1109 trace_sched_migrate_task(p, new_cpu);
1110
1111 if (task_cpu(p) != new_cpu) {
1112 p->se.nr_migrations++;
1113 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1114 }
1115
1116 __set_task_cpu(p, new_cpu);
1117}
1118
1119struct migration_arg {
1120 struct task_struct *task;
1121 int dest_cpu;
1122};
1123
1124static int migration_cpu_stop(void *data);
1125
1126/*
1127 * wait_task_inactive - wait for a thread to unschedule.
1128 *
1129 * If @match_state is nonzero, it's the @p->state value just checked and
1130 * not expected to change. If it changes, i.e. @p might have woken up,
1131 * then return zero. When we succeed in waiting for @p to be off its CPU,
1132 * we return a positive number (its total switch count). If a second call
1133 * a short while later returns the same number, the caller can be sure that
1134 * @p has remained unscheduled the whole time.
1135 *
1136 * The caller must ensure that the task *will* unschedule sometime soon,
1137 * else this function might spin for a *long* time. This function can't
1138 * be called with interrupts off, or it may introduce deadlock with
1139 * smp_call_function() if an IPI is sent by the same process we are
1140 * waiting to become inactive.
1141 */
1142unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1143{
1144 unsigned long flags;
1145 int running, on_rq;
1146 unsigned long ncsw;
1147 struct rq *rq;
1148
1149 for (;;) {
1150 /*
1151 * We do the initial early heuristics without holding
1152 * any task-queue locks at all. We'll only try to get
1153 * the runqueue lock when things look like they will
1154 * work out!
1155 */
1156 rq = task_rq(p);
1157
1158 /*
1159 * If the task is actively running on another CPU
1160 * still, just relax and busy-wait without holding
1161 * any locks.
1162 *
1163 * NOTE! Since we don't hold any locks, it's not
1164 * even sure that "rq" stays as the right runqueue!
1165 * But we don't care, since "task_running()" will
1166 * return false if the runqueue has changed and p
1167 * is actually now running somewhere else!
1168 */
1169 while (task_running(rq, p)) {
1170 if (match_state && unlikely(p->state != match_state))
1171 return 0;
1172 cpu_relax();
1173 }
1174
1175 /*
1176 * Ok, time to look more closely! We need the rq
1177 * lock now, to be *sure*. If we're wrong, we'll
1178 * just go back and repeat.
1179 */
1180 rq = task_rq_lock(p, &flags);
1181 trace_sched_wait_task(p);
1182 running = task_running(rq, p);
1183 on_rq = p->on_rq;
1184 ncsw = 0;
1185 if (!match_state || p->state == match_state)
1186 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1187 task_rq_unlock(rq, p, &flags);
1188
1189 /*
1190 * If it changed from the expected state, bail out now.
1191 */
1192 if (unlikely(!ncsw))
1193 break;
1194
1195 /*
1196 * Was it really running after all now that we
1197 * checked with the proper locks actually held?
1198 *
1199 * Oops. Go back and try again..
1200 */
1201 if (unlikely(running)) {
1202 cpu_relax();
1203 continue;
1204 }
1205
1206 /*
1207 * It's not enough that it's not actively running,
1208 * it must be off the runqueue _entirely_, and not
1209 * preempted!
1210 *
1211 * So if it was still runnable (but just not actively
1212 * running right now), it's preempted, and we should
1213 * yield - it could be a while.
1214 */
1215 if (unlikely(on_rq)) {
1216 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1217
1218 set_current_state(TASK_UNINTERRUPTIBLE);
1219 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1220 continue;
1221 }
1222
1223 /*
1224 * Ahh, all good. It wasn't running, and it wasn't
1225 * runnable, which means that it will never become
1226 * running in the future either. We're all done!
1227 */
1228 break;
1229 }
1230
1231 return ncsw;
1232}
1233
1234/***
1235 * kick_process - kick a running thread to enter/exit the kernel
1236 * @p: the to-be-kicked thread
1237 *
1238 * Cause a process which is running on another CPU to enter
1239 * kernel-mode, without any delay. (to get signals handled.)
1240 *
1241 * NOTE: this function doesn't have to take the runqueue lock,
1242 * because all it wants to ensure is that the remote task enters
1243 * the kernel. If the IPI races and the task has been migrated
1244 * to another CPU then no harm is done and the purpose has been
1245 * achieved as well.
1246 */
1247void kick_process(struct task_struct *p)
1248{
1249 int cpu;
1250
1251 preempt_disable();
1252 cpu = task_cpu(p);
1253 if ((cpu != smp_processor_id()) && task_curr(p))
1254 smp_send_reschedule(cpu);
1255 preempt_enable();
1256}
1257EXPORT_SYMBOL_GPL(kick_process);
1258#endif /* CONFIG_SMP */
1259
1260#ifdef CONFIG_SMP
1261/*
1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1263 */
1264static int select_fallback_rq(int cpu, struct task_struct *p)
1265{
1266 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1267 enum { cpuset, possible, fail } state = cpuset;
1268 int dest_cpu;
1269
1270 /* Look for allowed, online CPU in same node. */
1271 for_each_cpu(dest_cpu, nodemask) {
1272 if (!cpu_online(dest_cpu))
1273 continue;
1274 if (!cpu_active(dest_cpu))
1275 continue;
1276 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1277 return dest_cpu;
1278 }
1279
1280 for (;;) {
1281 /* Any allowed, online CPU? */
1282 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1283 if (!cpu_online(dest_cpu))
1284 continue;
1285 if (!cpu_active(dest_cpu))
1286 continue;
1287 goto out;
1288 }
1289
1290 switch (state) {
1291 case cpuset:
1292 /* No more Mr. Nice Guy. */
1293 cpuset_cpus_allowed_fallback(p);
1294 state = possible;
1295 break;
1296
1297 case possible:
1298 do_set_cpus_allowed(p, cpu_possible_mask);
1299 state = fail;
1300 break;
1301
1302 case fail:
1303 BUG();
1304 break;
1305 }
1306 }
1307
1308out:
1309 if (state != cpuset) {
1310 /*
1311 * Don't tell them about moving exiting tasks or
1312 * kernel threads (both mm NULL), since they never
1313 * leave kernel.
1314 */
1315 if (p->mm && printk_ratelimit()) {
1316 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1317 task_pid_nr(p), p->comm, cpu);
1318 }
1319 }
1320
1321 return dest_cpu;
1322}
1323
1324/*
1325 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1326 */
1327static inline
1328int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1329{
1330 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1331
1332 /*
1333 * In order not to call set_task_cpu() on a blocking task we need
1334 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1335 * cpu.
1336 *
1337 * Since this is common to all placement strategies, this lives here.
1338 *
1339 * [ this allows ->select_task() to simply return task_cpu(p) and
1340 * not worry about this generic constraint ]
1341 */
1342 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1343 !cpu_online(cpu)))
1344 cpu = select_fallback_rq(task_cpu(p), p);
1345
1346 return cpu;
1347}
1348
1349static void update_avg(u64 *avg, u64 sample)
1350{
1351 s64 diff = sample - *avg;
1352 *avg += diff >> 3;
1353}
1354#endif
1355
1356static void
1357ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1358{
1359#ifdef CONFIG_SCHEDSTATS
1360 struct rq *rq = this_rq();
1361
1362#ifdef CONFIG_SMP
1363 int this_cpu = smp_processor_id();
1364
1365 if (cpu == this_cpu) {
1366 schedstat_inc(rq, ttwu_local);
1367 schedstat_inc(p, se.statistics.nr_wakeups_local);
1368 } else {
1369 struct sched_domain *sd;
1370
1371 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1372 rcu_read_lock();
1373 for_each_domain(this_cpu, sd) {
1374 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1375 schedstat_inc(sd, ttwu_wake_remote);
1376 break;
1377 }
1378 }
1379 rcu_read_unlock();
1380 }
1381
1382 if (wake_flags & WF_MIGRATED)
1383 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1384
1385#endif /* CONFIG_SMP */
1386
1387 schedstat_inc(rq, ttwu_count);
1388 schedstat_inc(p, se.statistics.nr_wakeups);
1389
1390 if (wake_flags & WF_SYNC)
1391 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1392
1393#endif /* CONFIG_SCHEDSTATS */
1394}
1395
1396static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1397{
1398 activate_task(rq, p, en_flags);
1399 p->on_rq = 1;
1400
1401 /* if a worker is waking up, notify workqueue */
1402 if (p->flags & PF_WQ_WORKER)
1403 wq_worker_waking_up(p, cpu_of(rq));
1404}
1405
1406/*
1407 * Mark the task runnable and perform wakeup-preemption.
1408 */
1409static void
1410ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1411{
1412 trace_sched_wakeup(p, true);
1413 check_preempt_curr(rq, p, wake_flags);
1414
1415 p->state = TASK_RUNNING;
1416#ifdef CONFIG_SMP
1417 if (p->sched_class->task_woken)
1418 p->sched_class->task_woken(rq, p);
1419
1420 if (rq->idle_stamp) {
1421 u64 delta = rq->clock - rq->idle_stamp;
1422 u64 max = 2*sysctl_sched_migration_cost;
1423
1424 if (delta > max)
1425 rq->avg_idle = max;
1426 else
1427 update_avg(&rq->avg_idle, delta);
1428 rq->idle_stamp = 0;
1429 }
1430#endif
1431}
1432
1433static void
1434ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1435{
1436#ifdef CONFIG_SMP
1437 if (p->sched_contributes_to_load)
1438 rq->nr_uninterruptible--;
1439#endif
1440
1441 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1442 ttwu_do_wakeup(rq, p, wake_flags);
1443}
1444
1445/*
1446 * Called in case the task @p isn't fully descheduled from its runqueue,
1447 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1448 * since all we need to do is flip p->state to TASK_RUNNING, since
1449 * the task is still ->on_rq.
1450 */
1451static int ttwu_remote(struct task_struct *p, int wake_flags)
1452{
1453 struct rq *rq;
1454 int ret = 0;
1455
1456 rq = __task_rq_lock(p);
1457 if (p->on_rq) {
1458 ttwu_do_wakeup(rq, p, wake_flags);
1459 ret = 1;
1460 }
1461 __task_rq_unlock(rq);
1462
1463 return ret;
1464}
1465
1466#ifdef CONFIG_SMP
1467static void sched_ttwu_pending(void)
1468{
1469 struct rq *rq = this_rq();
1470 struct llist_node *llist = llist_del_all(&rq->wake_list);
1471 struct task_struct *p;
1472
1473 raw_spin_lock(&rq->lock);
1474
1475 while (llist) {
1476 p = llist_entry(llist, struct task_struct, wake_entry);
1477 llist = llist_next(llist);
1478 ttwu_do_activate(rq, p, 0);
1479 }
1480
1481 raw_spin_unlock(&rq->lock);
1482}
1483
1484void scheduler_ipi(void)
1485{
1486 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1487 return;
1488
1489 /*
1490 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1491 * traditionally all their work was done from the interrupt return
1492 * path. Now that we actually do some work, we need to make sure
1493 * we do call them.
1494 *
1495 * Some archs already do call them, luckily irq_enter/exit nest
1496 * properly.
1497 *
1498 * Arguably we should visit all archs and update all handlers,
1499 * however a fair share of IPIs are still resched only so this would
1500 * somewhat pessimize the simple resched case.
1501 */
1502 irq_enter();
1503 sched_ttwu_pending();
1504
1505 /*
1506 * Check if someone kicked us for doing the nohz idle load balance.
1507 */
1508 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1509 this_rq()->idle_balance = 1;
1510 raise_softirq_irqoff(SCHED_SOFTIRQ);
1511 }
1512 irq_exit();
1513}
1514
1515static void ttwu_queue_remote(struct task_struct *p, int cpu)
1516{
1517 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1518 smp_send_reschedule(cpu);
1519}
1520
1521#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1522static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1523{
1524 struct rq *rq;
1525 int ret = 0;
1526
1527 rq = __task_rq_lock(p);
1528 if (p->on_cpu) {
1529 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1530 ttwu_do_wakeup(rq, p, wake_flags);
1531 ret = 1;
1532 }
1533 __task_rq_unlock(rq);
1534
1535 return ret;
1536
1537}
1538#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1539
1540bool cpus_share_cache(int this_cpu, int that_cpu)
1541{
1542 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1543}
1544#endif /* CONFIG_SMP */
1545
1546static void ttwu_queue(struct task_struct *p, int cpu)
1547{
1548 struct rq *rq = cpu_rq(cpu);
1549
1550#if defined(CONFIG_SMP)
1551 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1552 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1553 ttwu_queue_remote(p, cpu);
1554 return;
1555 }
1556#endif
1557
1558 raw_spin_lock(&rq->lock);
1559 ttwu_do_activate(rq, p, 0);
1560 raw_spin_unlock(&rq->lock);
1561}
1562
1563/**
1564 * try_to_wake_up - wake up a thread
1565 * @p: the thread to be awakened
1566 * @state: the mask of task states that can be woken
1567 * @wake_flags: wake modifier flags (WF_*)
1568 *
1569 * Put it on the run-queue if it's not already there. The "current"
1570 * thread is always on the run-queue (except when the actual
1571 * re-schedule is in progress), and as such you're allowed to do
1572 * the simpler "current->state = TASK_RUNNING" to mark yourself
1573 * runnable without the overhead of this.
1574 *
1575 * Returns %true if @p was woken up, %false if it was already running
1576 * or @state didn't match @p's state.
1577 */
1578static int
1579try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1580{
1581 unsigned long flags;
1582 int cpu, success = 0;
1583
1584 smp_wmb();
1585 raw_spin_lock_irqsave(&p->pi_lock, flags);
1586 if (!(p->state & state))
1587 goto out;
1588
1589 success = 1; /* we're going to change ->state */
1590 cpu = task_cpu(p);
1591
1592 if (p->on_rq && ttwu_remote(p, wake_flags))
1593 goto stat;
1594
1595#ifdef CONFIG_SMP
1596 /*
1597 * If the owning (remote) cpu is still in the middle of schedule() with
1598 * this task as prev, wait until its done referencing the task.
1599 */
1600 while (p->on_cpu) {
1601#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1602 /*
1603 * In case the architecture enables interrupts in
1604 * context_switch(), we cannot busy wait, since that
1605 * would lead to deadlocks when an interrupt hits and
1606 * tries to wake up @prev. So bail and do a complete
1607 * remote wakeup.
1608 */
1609 if (ttwu_activate_remote(p, wake_flags))
1610 goto stat;
1611#else
1612 cpu_relax();
1613#endif
1614 }
1615 /*
1616 * Pairs with the smp_wmb() in finish_lock_switch().
1617 */
1618 smp_rmb();
1619
1620 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1621 p->state = TASK_WAKING;
1622
1623 if (p->sched_class->task_waking)
1624 p->sched_class->task_waking(p);
1625
1626 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1627 if (task_cpu(p) != cpu) {
1628 wake_flags |= WF_MIGRATED;
1629 set_task_cpu(p, cpu);
1630 }
1631#endif /* CONFIG_SMP */
1632
1633 ttwu_queue(p, cpu);
1634stat:
1635 ttwu_stat(p, cpu, wake_flags);
1636out:
1637 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1638
1639 return success;
1640}
1641
1642/**
1643 * try_to_wake_up_local - try to wake up a local task with rq lock held
1644 * @p: the thread to be awakened
1645 *
1646 * Put @p on the run-queue if it's not already there. The caller must
1647 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1648 * the current task.
1649 */
1650static void try_to_wake_up_local(struct task_struct *p)
1651{
1652 struct rq *rq = task_rq(p);
1653
1654 BUG_ON(rq != this_rq());
1655 BUG_ON(p == current);
1656 lockdep_assert_held(&rq->lock);
1657
1658 if (!raw_spin_trylock(&p->pi_lock)) {
1659 raw_spin_unlock(&rq->lock);
1660 raw_spin_lock(&p->pi_lock);
1661 raw_spin_lock(&rq->lock);
1662 }
1663
1664 if (!(p->state & TASK_NORMAL))
1665 goto out;
1666
1667 if (!p->on_rq)
1668 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1669
1670 ttwu_do_wakeup(rq, p, 0);
1671 ttwu_stat(p, smp_processor_id(), 0);
1672out:
1673 raw_spin_unlock(&p->pi_lock);
1674}
1675
1676/**
1677 * wake_up_process - Wake up a specific process
1678 * @p: The process to be woken up.
1679 *
1680 * Attempt to wake up the nominated process and move it to the set of runnable
1681 * processes. Returns 1 if the process was woken up, 0 if it was already
1682 * running.
1683 *
1684 * It may be assumed that this function implies a write memory barrier before
1685 * changing the task state if and only if any tasks are woken up.
1686 */
1687int wake_up_process(struct task_struct *p)
1688{
1689 return try_to_wake_up(p, TASK_ALL, 0);
1690}
1691EXPORT_SYMBOL(wake_up_process);
1692
1693int wake_up_state(struct task_struct *p, unsigned int state)
1694{
1695 return try_to_wake_up(p, state, 0);
1696}
1697
1698/*
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1701 *
1702 * __sched_fork() is basic setup used by init_idle() too:
1703 */
1704static void __sched_fork(struct task_struct *p)
1705{
1706 p->on_rq = 0;
1707
1708 p->se.on_rq = 0;
1709 p->se.exec_start = 0;
1710 p->se.sum_exec_runtime = 0;
1711 p->se.prev_sum_exec_runtime = 0;
1712 p->se.nr_migrations = 0;
1713 p->se.vruntime = 0;
1714 INIT_LIST_HEAD(&p->se.group_node);
1715
1716#ifdef CONFIG_SCHEDSTATS
1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1718#endif
1719
1720 INIT_LIST_HEAD(&p->rt.run_list);
1721
1722#ifdef CONFIG_PREEMPT_NOTIFIERS
1723 INIT_HLIST_HEAD(&p->preempt_notifiers);
1724#endif
1725}
1726
1727/*
1728 * fork()/clone()-time setup:
1729 */
1730void sched_fork(struct task_struct *p)
1731{
1732 unsigned long flags;
1733 int cpu = get_cpu();
1734
1735 __sched_fork(p);
1736 /*
1737 * We mark the process as running here. This guarantees that
1738 * nobody will actually run it, and a signal or other external
1739 * event cannot wake it up and insert it on the runqueue either.
1740 */
1741 p->state = TASK_RUNNING;
1742
1743 /*
1744 * Make sure we do not leak PI boosting priority to the child.
1745 */
1746 p->prio = current->normal_prio;
1747
1748 /*
1749 * Revert to default priority/policy on fork if requested.
1750 */
1751 if (unlikely(p->sched_reset_on_fork)) {
1752 if (task_has_rt_policy(p)) {
1753 p->policy = SCHED_NORMAL;
1754 p->static_prio = NICE_TO_PRIO(0);
1755 p->rt_priority = 0;
1756 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1757 p->static_prio = NICE_TO_PRIO(0);
1758
1759 p->prio = p->normal_prio = __normal_prio(p);
1760 set_load_weight(p);
1761
1762 /*
1763 * We don't need the reset flag anymore after the fork. It has
1764 * fulfilled its duty:
1765 */
1766 p->sched_reset_on_fork = 0;
1767 }
1768
1769 if (!rt_prio(p->prio))
1770 p->sched_class = &fair_sched_class;
1771
1772 if (p->sched_class->task_fork)
1773 p->sched_class->task_fork(p);
1774
1775 /*
1776 * The child is not yet in the pid-hash so no cgroup attach races,
1777 * and the cgroup is pinned to this child due to cgroup_fork()
1778 * is ran before sched_fork().
1779 *
1780 * Silence PROVE_RCU.
1781 */
1782 raw_spin_lock_irqsave(&p->pi_lock, flags);
1783 set_task_cpu(p, cpu);
1784 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1785
1786#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1787 if (likely(sched_info_on()))
1788 memset(&p->sched_info, 0, sizeof(p->sched_info));
1789#endif
1790#if defined(CONFIG_SMP)
1791 p->on_cpu = 0;
1792#endif
1793#ifdef CONFIG_PREEMPT_COUNT
1794 /* Want to start with kernel preemption disabled. */
1795 task_thread_info(p)->preempt_count = 1;
1796#endif
1797#ifdef CONFIG_SMP
1798 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1799#endif
1800
1801 put_cpu();
1802}
1803
1804/*
1805 * wake_up_new_task - wake up a newly created task for the first time.
1806 *
1807 * This function will do some initial scheduler statistics housekeeping
1808 * that must be done for every newly created context, then puts the task
1809 * on the runqueue and wakes it.
1810 */
1811void wake_up_new_task(struct task_struct *p)
1812{
1813 unsigned long flags;
1814 struct rq *rq;
1815
1816 raw_spin_lock_irqsave(&p->pi_lock, flags);
1817#ifdef CONFIG_SMP
1818 /*
1819 * Fork balancing, do it here and not earlier because:
1820 * - cpus_allowed can change in the fork path
1821 * - any previously selected cpu might disappear through hotplug
1822 */
1823 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1824#endif
1825
1826 rq = __task_rq_lock(p);
1827 activate_task(rq, p, 0);
1828 p->on_rq = 1;
1829 trace_sched_wakeup_new(p, true);
1830 check_preempt_curr(rq, p, WF_FORK);
1831#ifdef CONFIG_SMP
1832 if (p->sched_class->task_woken)
1833 p->sched_class->task_woken(rq, p);
1834#endif
1835 task_rq_unlock(rq, p, &flags);
1836}
1837
1838#ifdef CONFIG_PREEMPT_NOTIFIERS
1839
1840/**
1841 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1842 * @notifier: notifier struct to register
1843 */
1844void preempt_notifier_register(struct preempt_notifier *notifier)
1845{
1846 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1847}
1848EXPORT_SYMBOL_GPL(preempt_notifier_register);
1849
1850/**
1851 * preempt_notifier_unregister - no longer interested in preemption notifications
1852 * @notifier: notifier struct to unregister
1853 *
1854 * This is safe to call from within a preemption notifier.
1855 */
1856void preempt_notifier_unregister(struct preempt_notifier *notifier)
1857{
1858 hlist_del(¬ifier->link);
1859}
1860EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1861
1862static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1863{
1864 struct preempt_notifier *notifier;
1865 struct hlist_node *node;
1866
1867 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1868 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1869}
1870
1871static void
1872fire_sched_out_preempt_notifiers(struct task_struct *curr,
1873 struct task_struct *next)
1874{
1875 struct preempt_notifier *notifier;
1876 struct hlist_node *node;
1877
1878 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1879 notifier->ops->sched_out(notifier, next);
1880}
1881
1882#else /* !CONFIG_PREEMPT_NOTIFIERS */
1883
1884static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1885{
1886}
1887
1888static void
1889fire_sched_out_preempt_notifiers(struct task_struct *curr,
1890 struct task_struct *next)
1891{
1892}
1893
1894#endif /* CONFIG_PREEMPT_NOTIFIERS */
1895
1896/**
1897 * prepare_task_switch - prepare to switch tasks
1898 * @rq: the runqueue preparing to switch
1899 * @prev: the current task that is being switched out
1900 * @next: the task we are going to switch to.
1901 *
1902 * This is called with the rq lock held and interrupts off. It must
1903 * be paired with a subsequent finish_task_switch after the context
1904 * switch.
1905 *
1906 * prepare_task_switch sets up locking and calls architecture specific
1907 * hooks.
1908 */
1909static inline void
1910prepare_task_switch(struct rq *rq, struct task_struct *prev,
1911 struct task_struct *next)
1912{
1913 sched_info_switch(prev, next);
1914 perf_event_task_sched_out(prev, next);
1915 fire_sched_out_preempt_notifiers(prev, next);
1916 prepare_lock_switch(rq, next);
1917 prepare_arch_switch(next);
1918 trace_sched_switch(prev, next);
1919}
1920
1921/**
1922 * finish_task_switch - clean up after a task-switch
1923 * @rq: runqueue associated with task-switch
1924 * @prev: the thread we just switched away from.
1925 *
1926 * finish_task_switch must be called after the context switch, paired
1927 * with a prepare_task_switch call before the context switch.
1928 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1929 * and do any other architecture-specific cleanup actions.
1930 *
1931 * Note that we may have delayed dropping an mm in context_switch(). If
1932 * so, we finish that here outside of the runqueue lock. (Doing it
1933 * with the lock held can cause deadlocks; see schedule() for
1934 * details.)
1935 */
1936static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1937 __releases(rq->lock)
1938{
1939 struct mm_struct *mm = rq->prev_mm;
1940 long prev_state;
1941
1942 rq->prev_mm = NULL;
1943
1944 /*
1945 * A task struct has one reference for the use as "current".
1946 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1947 * schedule one last time. The schedule call will never return, and
1948 * the scheduled task must drop that reference.
1949 * The test for TASK_DEAD must occur while the runqueue locks are
1950 * still held, otherwise prev could be scheduled on another cpu, die
1951 * there before we look at prev->state, and then the reference would
1952 * be dropped twice.
1953 * Manfred Spraul <manfred@colorfullife.com>
1954 */
1955 prev_state = prev->state;
1956 finish_arch_switch(prev);
1957#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1958 local_irq_disable();
1959#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1960 perf_event_task_sched_in(prev, current);
1961#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1962 local_irq_enable();
1963#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1964 finish_lock_switch(rq, prev);
1965 finish_arch_post_lock_switch();
1966
1967 fire_sched_in_preempt_notifiers(current);
1968 if (mm)
1969 mmdrop(mm);
1970 if (unlikely(prev_state == TASK_DEAD)) {
1971 /*
1972 * Remove function-return probe instances associated with this
1973 * task and put them back on the free list.
1974 */
1975 kprobe_flush_task(prev);
1976 put_task_struct(prev);
1977 }
1978}
1979
1980#ifdef CONFIG_SMP
1981
1982/* assumes rq->lock is held */
1983static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1984{
1985 if (prev->sched_class->pre_schedule)
1986 prev->sched_class->pre_schedule(rq, prev);
1987}
1988
1989/* rq->lock is NOT held, but preemption is disabled */
1990static inline void post_schedule(struct rq *rq)
1991{
1992 if (rq->post_schedule) {
1993 unsigned long flags;
1994
1995 raw_spin_lock_irqsave(&rq->lock, flags);
1996 if (rq->curr->sched_class->post_schedule)
1997 rq->curr->sched_class->post_schedule(rq);
1998 raw_spin_unlock_irqrestore(&rq->lock, flags);
1999
2000 rq->post_schedule = 0;
2001 }
2002}
2003
2004#else
2005
2006static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2007{
2008}
2009
2010static inline void post_schedule(struct rq *rq)
2011{
2012}
2013
2014#endif
2015
2016/**
2017 * schedule_tail - first thing a freshly forked thread must call.
2018 * @prev: the thread we just switched away from.
2019 */
2020asmlinkage void schedule_tail(struct task_struct *prev)
2021 __releases(rq->lock)
2022{
2023 struct rq *rq = this_rq();
2024
2025 finish_task_switch(rq, prev);
2026
2027 /*
2028 * FIXME: do we need to worry about rq being invalidated by the
2029 * task_switch?
2030 */
2031 post_schedule(rq);
2032
2033#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2034 /* In this case, finish_task_switch does not reenable preemption */
2035 preempt_enable();
2036#endif
2037 if (current->set_child_tid)
2038 put_user(task_pid_vnr(current), current->set_child_tid);
2039}
2040
2041/*
2042 * context_switch - switch to the new MM and the new
2043 * thread's register state.
2044 */
2045static inline void
2046context_switch(struct rq *rq, struct task_struct *prev,
2047 struct task_struct *next)
2048{
2049 struct mm_struct *mm, *oldmm;
2050
2051 prepare_task_switch(rq, prev, next);
2052
2053 mm = next->mm;
2054 oldmm = prev->active_mm;
2055 /*
2056 * For paravirt, this is coupled with an exit in switch_to to
2057 * combine the page table reload and the switch backend into
2058 * one hypercall.
2059 */
2060 arch_start_context_switch(prev);
2061
2062 if (!mm) {
2063 next->active_mm = oldmm;
2064 atomic_inc(&oldmm->mm_count);
2065 enter_lazy_tlb(oldmm, next);
2066 } else
2067 switch_mm(oldmm, mm, next);
2068
2069 if (!prev->mm) {
2070 prev->active_mm = NULL;
2071 rq->prev_mm = oldmm;
2072 }
2073 /*
2074 * Since the runqueue lock will be released by the next
2075 * task (which is an invalid locking op but in the case
2076 * of the scheduler it's an obvious special-case), so we
2077 * do an early lockdep release here:
2078 */
2079#ifndef __ARCH_WANT_UNLOCKED_CTXSW
2080 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2081#endif
2082
2083 /* Here we just switch the register state and the stack. */
2084 switch_to(prev, next, prev);
2085
2086 barrier();
2087 /*
2088 * this_rq must be evaluated again because prev may have moved
2089 * CPUs since it called schedule(), thus the 'rq' on its stack
2090 * frame will be invalid.
2091 */
2092 finish_task_switch(this_rq(), prev);
2093}
2094
2095/*
2096 * nr_running, nr_uninterruptible and nr_context_switches:
2097 *
2098 * externally visible scheduler statistics: current number of runnable
2099 * threads, current number of uninterruptible-sleeping threads, total
2100 * number of context switches performed since bootup.
2101 */
2102unsigned long nr_running(void)
2103{
2104 unsigned long i, sum = 0;
2105
2106 for_each_online_cpu(i)
2107 sum += cpu_rq(i)->nr_running;
2108
2109 return sum;
2110}
2111
2112unsigned long nr_uninterruptible(void)
2113{
2114 unsigned long i, sum = 0;
2115
2116 for_each_possible_cpu(i)
2117 sum += cpu_rq(i)->nr_uninterruptible;
2118
2119 /*
2120 * Since we read the counters lockless, it might be slightly
2121 * inaccurate. Do not allow it to go below zero though:
2122 */
2123 if (unlikely((long)sum < 0))
2124 sum = 0;
2125
2126 return sum;
2127}
2128
2129unsigned long long nr_context_switches(void)
2130{
2131 int i;
2132 unsigned long long sum = 0;
2133
2134 for_each_possible_cpu(i)
2135 sum += cpu_rq(i)->nr_switches;
2136
2137 return sum;
2138}
2139
2140unsigned long nr_iowait(void)
2141{
2142 unsigned long i, sum = 0;
2143
2144 for_each_possible_cpu(i)
2145 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2146
2147 return sum;
2148}
2149
2150unsigned long nr_iowait_cpu(int cpu)
2151{
2152 struct rq *this = cpu_rq(cpu);
2153 return atomic_read(&this->nr_iowait);
2154}
2155
2156unsigned long this_cpu_load(void)
2157{
2158 struct rq *this = this_rq();
2159 return this->cpu_load[0];
2160}
2161
2162
2163/*
2164 * Global load-average calculations
2165 *
2166 * We take a distributed and async approach to calculating the global load-avg
2167 * in order to minimize overhead.
2168 *
2169 * The global load average is an exponentially decaying average of nr_running +
2170 * nr_uninterruptible.
2171 *
2172 * Once every LOAD_FREQ:
2173 *
2174 * nr_active = 0;
2175 * for_each_possible_cpu(cpu)
2176 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2177 *
2178 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2179 *
2180 * Due to a number of reasons the above turns in the mess below:
2181 *
2182 * - for_each_possible_cpu() is prohibitively expensive on machines with
2183 * serious number of cpus, therefore we need to take a distributed approach
2184 * to calculating nr_active.
2185 *
2186 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2187 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2188 *
2189 * So assuming nr_active := 0 when we start out -- true per definition, we
2190 * can simply take per-cpu deltas and fold those into a global accumulate
2191 * to obtain the same result. See calc_load_fold_active().
2192 *
2193 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2194 * across the machine, we assume 10 ticks is sufficient time for every
2195 * cpu to have completed this task.
2196 *
2197 * This places an upper-bound on the IRQ-off latency of the machine. Then
2198 * again, being late doesn't loose the delta, just wrecks the sample.
2199 *
2200 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2201 * this would add another cross-cpu cacheline miss and atomic operation
2202 * to the wakeup path. Instead we increment on whatever cpu the task ran
2203 * when it went into uninterruptible state and decrement on whatever cpu
2204 * did the wakeup. This means that only the sum of nr_uninterruptible over
2205 * all cpus yields the correct result.
2206 *
2207 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2208 */
2209
2210/* Variables and functions for calc_load */
2211static atomic_long_t calc_load_tasks;
2212static unsigned long calc_load_update;
2213unsigned long avenrun[3];
2214EXPORT_SYMBOL(avenrun); /* should be removed */
2215
2216/**
2217 * get_avenrun - get the load average array
2218 * @loads: pointer to dest load array
2219 * @offset: offset to add
2220 * @shift: shift count to shift the result left
2221 *
2222 * These values are estimates at best, so no need for locking.
2223 */
2224void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2225{
2226 loads[0] = (avenrun[0] + offset) << shift;
2227 loads[1] = (avenrun[1] + offset) << shift;
2228 loads[2] = (avenrun[2] + offset) << shift;
2229}
2230
2231static long calc_load_fold_active(struct rq *this_rq)
2232{
2233 long nr_active, delta = 0;
2234
2235 nr_active = this_rq->nr_running;
2236 nr_active += (long) this_rq->nr_uninterruptible;
2237
2238 if (nr_active != this_rq->calc_load_active) {
2239 delta = nr_active - this_rq->calc_load_active;
2240 this_rq->calc_load_active = nr_active;
2241 }
2242
2243 return delta;
2244}
2245
2246/*
2247 * a1 = a0 * e + a * (1 - e)
2248 */
2249static unsigned long
2250calc_load(unsigned long load, unsigned long exp, unsigned long active)
2251{
2252 load *= exp;
2253 load += active * (FIXED_1 - exp);
2254 load += 1UL << (FSHIFT - 1);
2255 return load >> FSHIFT;
2256}
2257
2258#ifdef CONFIG_NO_HZ
2259/*
2260 * Handle NO_HZ for the global load-average.
2261 *
2262 * Since the above described distributed algorithm to compute the global
2263 * load-average relies on per-cpu sampling from the tick, it is affected by
2264 * NO_HZ.
2265 *
2266 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2267 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2268 * when we read the global state.
2269 *
2270 * Obviously reality has to ruin such a delightfully simple scheme:
2271 *
2272 * - When we go NO_HZ idle during the window, we can negate our sample
2273 * contribution, causing under-accounting.
2274 *
2275 * We avoid this by keeping two idle-delta counters and flipping them
2276 * when the window starts, thus separating old and new NO_HZ load.
2277 *
2278 * The only trick is the slight shift in index flip for read vs write.
2279 *
2280 * 0s 5s 10s 15s
2281 * +10 +10 +10 +10
2282 * |-|-----------|-|-----------|-|-----------|-|
2283 * r:0 0 1 1 0 0 1 1 0
2284 * w:0 1 1 0 0 1 1 0 0
2285 *
2286 * This ensures we'll fold the old idle contribution in this window while
2287 * accumlating the new one.
2288 *
2289 * - When we wake up from NO_HZ idle during the window, we push up our
2290 * contribution, since we effectively move our sample point to a known
2291 * busy state.
2292 *
2293 * This is solved by pushing the window forward, and thus skipping the
2294 * sample, for this cpu (effectively using the idle-delta for this cpu which
2295 * was in effect at the time the window opened). This also solves the issue
2296 * of having to deal with a cpu having been in NOHZ idle for multiple
2297 * LOAD_FREQ intervals.
2298 *
2299 * When making the ILB scale, we should try to pull this in as well.
2300 */
2301static atomic_long_t calc_load_idle[2];
2302static int calc_load_idx;
2303
2304static inline int calc_load_write_idx(void)
2305{
2306 int idx = calc_load_idx;
2307
2308 /*
2309 * See calc_global_nohz(), if we observe the new index, we also
2310 * need to observe the new update time.
2311 */
2312 smp_rmb();
2313
2314 /*
2315 * If the folding window started, make sure we start writing in the
2316 * next idle-delta.
2317 */
2318 if (!time_before(jiffies, calc_load_update))
2319 idx++;
2320
2321 return idx & 1;
2322}
2323
2324static inline int calc_load_read_idx(void)
2325{
2326 return calc_load_idx & 1;
2327}
2328
2329void calc_load_enter_idle(void)
2330{
2331 struct rq *this_rq = this_rq();
2332 long delta;
2333
2334 /*
2335 * We're going into NOHZ mode, if there's any pending delta, fold it
2336 * into the pending idle delta.
2337 */
2338 delta = calc_load_fold_active(this_rq);
2339 if (delta) {
2340 int idx = calc_load_write_idx();
2341 atomic_long_add(delta, &calc_load_idle[idx]);
2342 }
2343}
2344
2345void calc_load_exit_idle(void)
2346{
2347 struct rq *this_rq = this_rq();
2348
2349 /*
2350 * If we're still before the sample window, we're done.
2351 */
2352 if (time_before(jiffies, this_rq->calc_load_update))
2353 return;
2354
2355 /*
2356 * We woke inside or after the sample window, this means we're already
2357 * accounted through the nohz accounting, so skip the entire deal and
2358 * sync up for the next window.
2359 */
2360 this_rq->calc_load_update = calc_load_update;
2361 if (time_before(jiffies, this_rq->calc_load_update + 10))
2362 this_rq->calc_load_update += LOAD_FREQ;
2363}
2364
2365static long calc_load_fold_idle(void)
2366{
2367 int idx = calc_load_read_idx();
2368 long delta = 0;
2369
2370 if (atomic_long_read(&calc_load_idle[idx]))
2371 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2372
2373 return delta;
2374}
2375
2376/**
2377 * fixed_power_int - compute: x^n, in O(log n) time
2378 *
2379 * @x: base of the power
2380 * @frac_bits: fractional bits of @x
2381 * @n: power to raise @x to.
2382 *
2383 * By exploiting the relation between the definition of the natural power
2384 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2385 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2386 * (where: n_i \elem {0, 1}, the binary vector representing n),
2387 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2388 * of course trivially computable in O(log_2 n), the length of our binary
2389 * vector.
2390 */
2391static unsigned long
2392fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2393{
2394 unsigned long result = 1UL << frac_bits;
2395
2396 if (n) for (;;) {
2397 if (n & 1) {
2398 result *= x;
2399 result += 1UL << (frac_bits - 1);
2400 result >>= frac_bits;
2401 }
2402 n >>= 1;
2403 if (!n)
2404 break;
2405 x *= x;
2406 x += 1UL << (frac_bits - 1);
2407 x >>= frac_bits;
2408 }
2409
2410 return result;
2411}
2412
2413/*
2414 * a1 = a0 * e + a * (1 - e)
2415 *
2416 * a2 = a1 * e + a * (1 - e)
2417 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2418 * = a0 * e^2 + a * (1 - e) * (1 + e)
2419 *
2420 * a3 = a2 * e + a * (1 - e)
2421 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2422 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2423 *
2424 * ...
2425 *
2426 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2427 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2428 * = a0 * e^n + a * (1 - e^n)
2429 *
2430 * [1] application of the geometric series:
2431 *
2432 * n 1 - x^(n+1)
2433 * S_n := \Sum x^i = -------------
2434 * i=0 1 - x
2435 */
2436static unsigned long
2437calc_load_n(unsigned long load, unsigned long exp,
2438 unsigned long active, unsigned int n)
2439{
2440
2441 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2442}
2443
2444/*
2445 * NO_HZ can leave us missing all per-cpu ticks calling
2446 * calc_load_account_active(), but since an idle CPU folds its delta into
2447 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2448 * in the pending idle delta if our idle period crossed a load cycle boundary.
2449 *
2450 * Once we've updated the global active value, we need to apply the exponential
2451 * weights adjusted to the number of cycles missed.
2452 */
2453static void calc_global_nohz(void)
2454{
2455 long delta, active, n;
2456
2457 if (!time_before(jiffies, calc_load_update + 10)) {
2458 /*
2459 * Catch-up, fold however many we are behind still
2460 */
2461 delta = jiffies - calc_load_update - 10;
2462 n = 1 + (delta / LOAD_FREQ);
2463
2464 active = atomic_long_read(&calc_load_tasks);
2465 active = active > 0 ? active * FIXED_1 : 0;
2466
2467 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2468 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2469 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2470
2471 calc_load_update += n * LOAD_FREQ;
2472 }
2473
2474 /*
2475 * Flip the idle index...
2476 *
2477 * Make sure we first write the new time then flip the index, so that
2478 * calc_load_write_idx() will see the new time when it reads the new
2479 * index, this avoids a double flip messing things up.
2480 */
2481 smp_wmb();
2482 calc_load_idx++;
2483}
2484#else /* !CONFIG_NO_HZ */
2485
2486static inline long calc_load_fold_idle(void) { return 0; }
2487static inline void calc_global_nohz(void) { }
2488
2489#endif /* CONFIG_NO_HZ */
2490
2491/*
2492 * calc_load - update the avenrun load estimates 10 ticks after the
2493 * CPUs have updated calc_load_tasks.
2494 */
2495void calc_global_load(unsigned long ticks)
2496{
2497 long active, delta;
2498
2499 if (time_before(jiffies, calc_load_update + 10))
2500 return;
2501
2502 /*
2503 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2504 */
2505 delta = calc_load_fold_idle();
2506 if (delta)
2507 atomic_long_add(delta, &calc_load_tasks);
2508
2509 active = atomic_long_read(&calc_load_tasks);
2510 active = active > 0 ? active * FIXED_1 : 0;
2511
2512 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2513 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2514 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2515
2516 calc_load_update += LOAD_FREQ;
2517
2518 /*
2519 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2520 */
2521 calc_global_nohz();
2522}
2523
2524/*
2525 * Called from update_cpu_load() to periodically update this CPU's
2526 * active count.
2527 */
2528static void calc_load_account_active(struct rq *this_rq)
2529{
2530 long delta;
2531
2532 if (time_before(jiffies, this_rq->calc_load_update))
2533 return;
2534
2535 delta = calc_load_fold_active(this_rq);
2536 if (delta)
2537 atomic_long_add(delta, &calc_load_tasks);
2538
2539 this_rq->calc_load_update += LOAD_FREQ;
2540}
2541
2542/*
2543 * End of global load-average stuff
2544 */
2545
2546/*
2547 * The exact cpuload at various idx values, calculated at every tick would be
2548 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2549 *
2550 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2551 * on nth tick when cpu may be busy, then we have:
2552 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2553 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2554 *
2555 * decay_load_missed() below does efficient calculation of
2556 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2557 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2558 *
2559 * The calculation is approximated on a 128 point scale.
2560 * degrade_zero_ticks is the number of ticks after which load at any
2561 * particular idx is approximated to be zero.
2562 * degrade_factor is a precomputed table, a row for each load idx.
2563 * Each column corresponds to degradation factor for a power of two ticks,
2564 * based on 128 point scale.
2565 * Example:
2566 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2567 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2568 *
2569 * With this power of 2 load factors, we can degrade the load n times
2570 * by looking at 1 bits in n and doing as many mult/shift instead of
2571 * n mult/shifts needed by the exact degradation.
2572 */
2573#define DEGRADE_SHIFT 7
2574static const unsigned char
2575 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2576static const unsigned char
2577 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2578 {0, 0, 0, 0, 0, 0, 0, 0},
2579 {64, 32, 8, 0, 0, 0, 0, 0},
2580 {96, 72, 40, 12, 1, 0, 0},
2581 {112, 98, 75, 43, 15, 1, 0},
2582 {120, 112, 98, 76, 45, 16, 2} };
2583
2584/*
2585 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2586 * would be when CPU is idle and so we just decay the old load without
2587 * adding any new load.
2588 */
2589static unsigned long
2590decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2591{
2592 int j = 0;
2593
2594 if (!missed_updates)
2595 return load;
2596
2597 if (missed_updates >= degrade_zero_ticks[idx])
2598 return 0;
2599
2600 if (idx == 1)
2601 return load >> missed_updates;
2602
2603 while (missed_updates) {
2604 if (missed_updates % 2)
2605 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2606
2607 missed_updates >>= 1;
2608 j++;
2609 }
2610 return load;
2611}
2612
2613/*
2614 * Update rq->cpu_load[] statistics. This function is usually called every
2615 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2616 * every tick. We fix it up based on jiffies.
2617 */
2618static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2619 unsigned long pending_updates)
2620{
2621 int i, scale;
2622
2623 this_rq->nr_load_updates++;
2624
2625 /* Update our load: */
2626 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2627 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2628 unsigned long old_load, new_load;
2629
2630 /* scale is effectively 1 << i now, and >> i divides by scale */
2631
2632 old_load = this_rq->cpu_load[i];
2633 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2634 new_load = this_load;
2635 /*
2636 * Round up the averaging division if load is increasing. This
2637 * prevents us from getting stuck on 9 if the load is 10, for
2638 * example.
2639 */
2640 if (new_load > old_load)
2641 new_load += scale - 1;
2642
2643 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2644 }
2645
2646 sched_avg_update(this_rq);
2647}
2648
2649#ifdef CONFIG_NO_HZ
2650/*
2651 * There is no sane way to deal with nohz on smp when using jiffies because the
2652 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2653 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2654 *
2655 * Therefore we cannot use the delta approach from the regular tick since that
2656 * would seriously skew the load calculation. However we'll make do for those
2657 * updates happening while idle (nohz_idle_balance) or coming out of idle
2658 * (tick_nohz_idle_exit).
2659 *
2660 * This means we might still be one tick off for nohz periods.
2661 */
2662
2663/*
2664 * Called from nohz_idle_balance() to update the load ratings before doing the
2665 * idle balance.
2666 */
2667void update_idle_cpu_load(struct rq *this_rq)
2668{
2669 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2670 unsigned long load = this_rq->load.weight;
2671 unsigned long pending_updates;
2672
2673 /*
2674 * bail if there's load or we're actually up-to-date.
2675 */
2676 if (load || curr_jiffies == this_rq->last_load_update_tick)
2677 return;
2678
2679 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2680 this_rq->last_load_update_tick = curr_jiffies;
2681
2682 __update_cpu_load(this_rq, load, pending_updates);
2683}
2684
2685/*
2686 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2687 */
2688void update_cpu_load_nohz(void)
2689{
2690 struct rq *this_rq = this_rq();
2691 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2692 unsigned long pending_updates;
2693
2694 if (curr_jiffies == this_rq->last_load_update_tick)
2695 return;
2696
2697 raw_spin_lock(&this_rq->lock);
2698 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2699 if (pending_updates) {
2700 this_rq->last_load_update_tick = curr_jiffies;
2701 /*
2702 * We were idle, this means load 0, the current load might be
2703 * !0 due to remote wakeups and the sort.
2704 */
2705 __update_cpu_load(this_rq, 0, pending_updates);
2706 }
2707 raw_spin_unlock(&this_rq->lock);
2708}
2709#endif /* CONFIG_NO_HZ */
2710
2711/*
2712 * Called from scheduler_tick()
2713 */
2714static void update_cpu_load_active(struct rq *this_rq)
2715{
2716 /*
2717 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2718 */
2719 this_rq->last_load_update_tick = jiffies;
2720 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2721
2722 calc_load_account_active(this_rq);
2723}
2724
2725#ifdef CONFIG_SMP
2726
2727/*
2728 * sched_exec - execve() is a valuable balancing opportunity, because at
2729 * this point the task has the smallest effective memory and cache footprint.
2730 */
2731void sched_exec(void)
2732{
2733 struct task_struct *p = current;
2734 unsigned long flags;
2735 int dest_cpu;
2736
2737 raw_spin_lock_irqsave(&p->pi_lock, flags);
2738 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2739 if (dest_cpu == smp_processor_id())
2740 goto unlock;
2741
2742 if (likely(cpu_active(dest_cpu))) {
2743 struct migration_arg arg = { p, dest_cpu };
2744
2745 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2746 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2747 return;
2748 }
2749unlock:
2750 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2751}
2752
2753#endif
2754
2755DEFINE_PER_CPU(struct kernel_stat, kstat);
2756DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2757
2758EXPORT_PER_CPU_SYMBOL(kstat);
2759EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2760
2761/*
2762 * Return any ns on the sched_clock that have not yet been accounted in
2763 * @p in case that task is currently running.
2764 *
2765 * Called with task_rq_lock() held on @rq.
2766 */
2767static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2768{
2769 u64 ns = 0;
2770
2771 if (task_current(rq, p)) {
2772 update_rq_clock(rq);
2773 ns = rq->clock_task - p->se.exec_start;
2774 if ((s64)ns < 0)
2775 ns = 0;
2776 }
2777
2778 return ns;
2779}
2780
2781unsigned long long task_delta_exec(struct task_struct *p)
2782{
2783 unsigned long flags;
2784 struct rq *rq;
2785 u64 ns = 0;
2786
2787 rq = task_rq_lock(p, &flags);
2788 ns = do_task_delta_exec(p, rq);
2789 task_rq_unlock(rq, p, &flags);
2790
2791 return ns;
2792}
2793
2794/*
2795 * Return accounted runtime for the task.
2796 * In case the task is currently running, return the runtime plus current's
2797 * pending runtime that have not been accounted yet.
2798 */
2799unsigned long long task_sched_runtime(struct task_struct *p)
2800{
2801 unsigned long flags;
2802 struct rq *rq;
2803 u64 ns = 0;
2804
2805 rq = task_rq_lock(p, &flags);
2806 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2807 task_rq_unlock(rq, p, &flags);
2808
2809 return ns;
2810}
2811
2812#ifdef CONFIG_CGROUP_CPUACCT
2813struct cgroup_subsys cpuacct_subsys;
2814struct cpuacct root_cpuacct;
2815#endif
2816
2817static inline void task_group_account_field(struct task_struct *p, int index,
2818 u64 tmp)
2819{
2820#ifdef CONFIG_CGROUP_CPUACCT
2821 struct kernel_cpustat *kcpustat;
2822 struct cpuacct *ca;
2823#endif
2824 /*
2825 * Since all updates are sure to touch the root cgroup, we
2826 * get ourselves ahead and touch it first. If the root cgroup
2827 * is the only cgroup, then nothing else should be necessary.
2828 *
2829 */
2830 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2831
2832#ifdef CONFIG_CGROUP_CPUACCT
2833 if (unlikely(!cpuacct_subsys.active))
2834 return;
2835
2836 rcu_read_lock();
2837 ca = task_ca(p);
2838 while (ca && (ca != &root_cpuacct)) {
2839 kcpustat = this_cpu_ptr(ca->cpustat);
2840 kcpustat->cpustat[index] += tmp;
2841 ca = parent_ca(ca);
2842 }
2843 rcu_read_unlock();
2844#endif
2845}
2846
2847
2848/*
2849 * Account user cpu time to a process.
2850 * @p: the process that the cpu time gets accounted to
2851 * @cputime: the cpu time spent in user space since the last update
2852 * @cputime_scaled: cputime scaled by cpu frequency
2853 */
2854void account_user_time(struct task_struct *p, cputime_t cputime,
2855 cputime_t cputime_scaled)
2856{
2857 int index;
2858
2859 /* Add user time to process. */
2860 p->utime += cputime;
2861 p->utimescaled += cputime_scaled;
2862 account_group_user_time(p, cputime);
2863
2864 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2865
2866 /* Add user time to cpustat. */
2867 task_group_account_field(p, index, (__force u64) cputime);
2868
2869 /* Account for user time used */
2870 acct_update_integrals(p);
2871}
2872
2873/*
2874 * Account guest cpu time to a process.
2875 * @p: the process that the cpu time gets accounted to
2876 * @cputime: the cpu time spent in virtual machine since the last update
2877 * @cputime_scaled: cputime scaled by cpu frequency
2878 */
2879static void account_guest_time(struct task_struct *p, cputime_t cputime,
2880 cputime_t cputime_scaled)
2881{
2882 u64 *cpustat = kcpustat_this_cpu->cpustat;
2883
2884 /* Add guest time to process. */
2885 p->utime += cputime;
2886 p->utimescaled += cputime_scaled;
2887 account_group_user_time(p, cputime);
2888 p->gtime += cputime;
2889
2890 /* Add guest time to cpustat. */
2891 if (TASK_NICE(p) > 0) {
2892 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2893 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2894 } else {
2895 cpustat[CPUTIME_USER] += (__force u64) cputime;
2896 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2897 }
2898}
2899
2900/*
2901 * Account system cpu time to a process and desired cpustat field
2902 * @p: the process that the cpu time gets accounted to
2903 * @cputime: the cpu time spent in kernel space since the last update
2904 * @cputime_scaled: cputime scaled by cpu frequency
2905 * @target_cputime64: pointer to cpustat field that has to be updated
2906 */
2907static inline
2908void __account_system_time(struct task_struct *p, cputime_t cputime,
2909 cputime_t cputime_scaled, int index)
2910{
2911 /* Add system time to process. */
2912 p->stime += cputime;
2913 p->stimescaled += cputime_scaled;
2914 account_group_system_time(p, cputime);
2915
2916 /* Add system time to cpustat. */
2917 task_group_account_field(p, index, (__force u64) cputime);
2918
2919 /* Account for system time used */
2920 acct_update_integrals(p);
2921}
2922
2923/*
2924 * Account system cpu time to a process.
2925 * @p: the process that the cpu time gets accounted to
2926 * @hardirq_offset: the offset to subtract from hardirq_count()
2927 * @cputime: the cpu time spent in kernel space since the last update
2928 * @cputime_scaled: cputime scaled by cpu frequency
2929 */
2930void account_system_time(struct task_struct *p, int hardirq_offset,
2931 cputime_t cputime, cputime_t cputime_scaled)
2932{
2933 int index;
2934
2935 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2936 account_guest_time(p, cputime, cputime_scaled);
2937 return;
2938 }
2939
2940 if (hardirq_count() - hardirq_offset)
2941 index = CPUTIME_IRQ;
2942 else if (in_serving_softirq())
2943 index = CPUTIME_SOFTIRQ;
2944 else
2945 index = CPUTIME_SYSTEM;
2946
2947 __account_system_time(p, cputime, cputime_scaled, index);
2948}
2949
2950/*
2951 * Account for involuntary wait time.
2952 * @cputime: the cpu time spent in involuntary wait
2953 */
2954void account_steal_time(cputime_t cputime)
2955{
2956 u64 *cpustat = kcpustat_this_cpu->cpustat;
2957
2958 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2959}
2960
2961/*
2962 * Account for idle time.
2963 * @cputime: the cpu time spent in idle wait
2964 */
2965void account_idle_time(cputime_t cputime)
2966{
2967 u64 *cpustat = kcpustat_this_cpu->cpustat;
2968 struct rq *rq = this_rq();
2969
2970 if (atomic_read(&rq->nr_iowait) > 0)
2971 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2972 else
2973 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2974}
2975
2976static __always_inline bool steal_account_process_tick(void)
2977{
2978#ifdef CONFIG_PARAVIRT
2979 if (static_key_false(¶virt_steal_enabled)) {
2980 u64 steal, st = 0;
2981
2982 steal = paravirt_steal_clock(smp_processor_id());
2983 steal -= this_rq()->prev_steal_time;
2984
2985 st = steal_ticks(steal);
2986 this_rq()->prev_steal_time += st * TICK_NSEC;
2987
2988 account_steal_time(st);
2989 return st;
2990 }
2991#endif
2992 return false;
2993}
2994
2995#ifndef CONFIG_VIRT_CPU_ACCOUNTING
2996
2997#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2998/*
2999 * Account a tick to a process and cpustat
3000 * @p: the process that the cpu time gets accounted to
3001 * @user_tick: is the tick from userspace
3002 * @rq: the pointer to rq
3003 *
3004 * Tick demultiplexing follows the order
3005 * - pending hardirq update
3006 * - pending softirq update
3007 * - user_time
3008 * - idle_time
3009 * - system time
3010 * - check for guest_time
3011 * - else account as system_time
3012 *
3013 * Check for hardirq is done both for system and user time as there is
3014 * no timer going off while we are on hardirq and hence we may never get an
3015 * opportunity to update it solely in system time.
3016 * p->stime and friends are only updated on system time and not on irq
3017 * softirq as those do not count in task exec_runtime any more.
3018 */
3019static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3020 struct rq *rq)
3021{
3022 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3023 u64 *cpustat = kcpustat_this_cpu->cpustat;
3024
3025 if (steal_account_process_tick())
3026 return;
3027
3028 if (irqtime_account_hi_update()) {
3029 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
3030 } else if (irqtime_account_si_update()) {
3031 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
3032 } else if (this_cpu_ksoftirqd() == p) {
3033 /*
3034 * ksoftirqd time do not get accounted in cpu_softirq_time.
3035 * So, we have to handle it separately here.
3036 * Also, p->stime needs to be updated for ksoftirqd.
3037 */
3038 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3039 CPUTIME_SOFTIRQ);
3040 } else if (user_tick) {
3041 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3042 } else if (p == rq->idle) {
3043 account_idle_time(cputime_one_jiffy);
3044 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3045 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3046 } else {
3047 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3048 CPUTIME_SYSTEM);
3049 }
3050}
3051
3052static void irqtime_account_idle_ticks(int ticks)
3053{
3054 int i;
3055 struct rq *rq = this_rq();
3056
3057 for (i = 0; i < ticks; i++)
3058 irqtime_account_process_tick(current, 0, rq);
3059}
3060#else /* CONFIG_IRQ_TIME_ACCOUNTING */
3061static void irqtime_account_idle_ticks(int ticks) {}
3062static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3063 struct rq *rq) {}
3064#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3065
3066/*
3067 * Account a single tick of cpu time.
3068 * @p: the process that the cpu time gets accounted to
3069 * @user_tick: indicates if the tick is a user or a system tick
3070 */
3071void account_process_tick(struct task_struct *p, int user_tick)
3072{
3073 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3074 struct rq *rq = this_rq();
3075
3076 if (sched_clock_irqtime) {
3077 irqtime_account_process_tick(p, user_tick, rq);
3078 return;
3079 }
3080
3081 if (steal_account_process_tick())
3082 return;
3083
3084 if (user_tick)
3085 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3086 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3087 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3088 one_jiffy_scaled);
3089 else
3090 account_idle_time(cputime_one_jiffy);
3091}
3092
3093/*
3094 * Account multiple ticks of steal time.
3095 * @p: the process from which the cpu time has been stolen
3096 * @ticks: number of stolen ticks
3097 */
3098void account_steal_ticks(unsigned long ticks)
3099{
3100 account_steal_time(jiffies_to_cputime(ticks));
3101}
3102
3103/*
3104 * Account multiple ticks of idle time.
3105 * @ticks: number of stolen ticks
3106 */
3107void account_idle_ticks(unsigned long ticks)
3108{
3109
3110 if (sched_clock_irqtime) {
3111 irqtime_account_idle_ticks(ticks);
3112 return;
3113 }
3114
3115 account_idle_time(jiffies_to_cputime(ticks));
3116}
3117
3118#endif
3119
3120/*
3121 * Use precise platform statistics if available:
3122 */
3123#ifdef CONFIG_VIRT_CPU_ACCOUNTING
3124void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3125{
3126 *ut = p->utime;
3127 *st = p->stime;
3128}
3129
3130void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3131{
3132 struct task_cputime cputime;
3133
3134 thread_group_cputime(p, &cputime);
3135
3136 *ut = cputime.utime;
3137 *st = cputime.stime;
3138}
3139#else
3140
3141#ifndef nsecs_to_cputime
3142# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3143#endif
3144
3145static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total)
3146{
3147 u64 temp = (__force u64) rtime;
3148
3149 temp *= (__force u64) utime;
3150
3151 if (sizeof(cputime_t) == 4)
3152 temp = div_u64(temp, (__force u32) total);
3153 else
3154 temp = div64_u64(temp, (__force u64) total);
3155
3156 return (__force cputime_t) temp;
3157}
3158
3159void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3160{
3161 cputime_t rtime, utime = p->utime, total = utime + p->stime;
3162
3163 /*
3164 * Use CFS's precise accounting:
3165 */
3166 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3167
3168 if (total)
3169 utime = scale_utime(utime, rtime, total);
3170 else
3171 utime = rtime;
3172
3173 /*
3174 * Compare with previous values, to keep monotonicity:
3175 */
3176 p->prev_utime = max(p->prev_utime, utime);
3177 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
3178
3179 *ut = p->prev_utime;
3180 *st = p->prev_stime;
3181}
3182
3183/*
3184 * Must be called with siglock held.
3185 */
3186void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3187{
3188 struct signal_struct *sig = p->signal;
3189 struct task_cputime cputime;
3190 cputime_t rtime, utime, total;
3191
3192 thread_group_cputime(p, &cputime);
3193
3194 total = cputime.utime + cputime.stime;
3195 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3196
3197 if (total)
3198 utime = scale_utime(cputime.utime, rtime, total);
3199 else
3200 utime = rtime;
3201
3202 sig->prev_utime = max(sig->prev_utime, utime);
3203 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3204
3205 *ut = sig->prev_utime;
3206 *st = sig->prev_stime;
3207}
3208#endif
3209
3210/*
3211 * This function gets called by the timer code, with HZ frequency.
3212 * We call it with interrupts disabled.
3213 */
3214void scheduler_tick(void)
3215{
3216 int cpu = smp_processor_id();
3217 struct rq *rq = cpu_rq(cpu);
3218 struct task_struct *curr = rq->curr;
3219
3220 sched_clock_tick();
3221
3222 raw_spin_lock(&rq->lock);
3223 update_rq_clock(rq);
3224 update_cpu_load_active(rq);
3225 curr->sched_class->task_tick(rq, curr, 0);
3226 raw_spin_unlock(&rq->lock);
3227
3228 perf_event_task_tick();
3229
3230#ifdef CONFIG_SMP
3231 rq->idle_balance = idle_cpu(cpu);
3232 trigger_load_balance(rq, cpu);
3233#endif
3234}
3235
3236notrace unsigned long get_parent_ip(unsigned long addr)
3237{
3238 if (in_lock_functions(addr)) {
3239 addr = CALLER_ADDR2;
3240 if (in_lock_functions(addr))
3241 addr = CALLER_ADDR3;
3242 }
3243 return addr;
3244}
3245
3246#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3247 defined(CONFIG_PREEMPT_TRACER))
3248
3249void __kprobes add_preempt_count(int val)
3250{
3251#ifdef CONFIG_DEBUG_PREEMPT
3252 /*
3253 * Underflow?
3254 */
3255 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3256 return;
3257#endif
3258 preempt_count() += val;
3259#ifdef CONFIG_DEBUG_PREEMPT
3260 /*
3261 * Spinlock count overflowing soon?
3262 */
3263 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3264 PREEMPT_MASK - 10);
3265#endif
3266 if (preempt_count() == val)
3267 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3268}
3269EXPORT_SYMBOL(add_preempt_count);
3270
3271void __kprobes sub_preempt_count(int val)
3272{
3273#ifdef CONFIG_DEBUG_PREEMPT
3274 /*
3275 * Underflow?
3276 */
3277 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3278 return;
3279 /*
3280 * Is the spinlock portion underflowing?
3281 */
3282 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3283 !(preempt_count() & PREEMPT_MASK)))
3284 return;
3285#endif
3286
3287 if (preempt_count() == val)
3288 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3289 preempt_count() -= val;
3290}
3291EXPORT_SYMBOL(sub_preempt_count);
3292
3293#endif
3294
3295/*
3296 * Print scheduling while atomic bug:
3297 */
3298static noinline void __schedule_bug(struct task_struct *prev)
3299{
3300 if (oops_in_progress)
3301 return;
3302
3303 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3304 prev->comm, prev->pid, preempt_count());
3305
3306 debug_show_held_locks(prev);
3307 print_modules();
3308 if (irqs_disabled())
3309 print_irqtrace_events(prev);
3310 dump_stack();
3311 add_taint(TAINT_WARN);
3312}
3313
3314/*
3315 * Various schedule()-time debugging checks and statistics:
3316 */
3317static inline void schedule_debug(struct task_struct *prev)
3318{
3319 /*
3320 * Test if we are atomic. Since do_exit() needs to call into
3321 * schedule() atomically, we ignore that path for now.
3322 * Otherwise, whine if we are scheduling when we should not be.
3323 */
3324 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3325 __schedule_bug(prev);
3326 rcu_sleep_check();
3327
3328 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3329
3330 schedstat_inc(this_rq(), sched_count);
3331}
3332
3333static void put_prev_task(struct rq *rq, struct task_struct *prev)
3334{
3335 if (prev->on_rq || rq->skip_clock_update < 0)
3336 update_rq_clock(rq);
3337 prev->sched_class->put_prev_task(rq, prev);
3338}
3339
3340/*
3341 * Pick up the highest-prio task:
3342 */
3343static inline struct task_struct *
3344pick_next_task(struct rq *rq)
3345{
3346 const struct sched_class *class;
3347 struct task_struct *p;
3348
3349 /*
3350 * Optimization: we know that if all tasks are in
3351 * the fair class we can call that function directly:
3352 */
3353 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3354 p = fair_sched_class.pick_next_task(rq);
3355 if (likely(p))
3356 return p;
3357 }
3358
3359 for_each_class(class) {
3360 p = class->pick_next_task(rq);
3361 if (p)
3362 return p;
3363 }
3364
3365 BUG(); /* the idle class will always have a runnable task */
3366}
3367
3368/*
3369 * __schedule() is the main scheduler function.
3370 */
3371static void __sched __schedule(void)
3372{
3373 struct task_struct *prev, *next;
3374 unsigned long *switch_count;
3375 struct rq *rq;
3376 int cpu;
3377
3378need_resched:
3379 preempt_disable();
3380 cpu = smp_processor_id();
3381 rq = cpu_rq(cpu);
3382 rcu_note_context_switch(cpu);
3383 prev = rq->curr;
3384
3385 schedule_debug(prev);
3386
3387 if (sched_feat(HRTICK))
3388 hrtick_clear(rq);
3389
3390 raw_spin_lock_irq(&rq->lock);
3391
3392 switch_count = &prev->nivcsw;
3393 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3394 if (unlikely(signal_pending_state(prev->state, prev))) {
3395 prev->state = TASK_RUNNING;
3396 } else {
3397 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3398 prev->on_rq = 0;
3399
3400 /*
3401 * If a worker went to sleep, notify and ask workqueue
3402 * whether it wants to wake up a task to maintain
3403 * concurrency.
3404 */
3405 if (prev->flags & PF_WQ_WORKER) {
3406 struct task_struct *to_wakeup;
3407
3408 to_wakeup = wq_worker_sleeping(prev, cpu);
3409 if (to_wakeup)
3410 try_to_wake_up_local(to_wakeup);
3411 }
3412 }
3413 switch_count = &prev->nvcsw;
3414 }
3415
3416 pre_schedule(rq, prev);
3417
3418 if (unlikely(!rq->nr_running))
3419 idle_balance(cpu, rq);
3420
3421 put_prev_task(rq, prev);
3422 next = pick_next_task(rq);
3423 clear_tsk_need_resched(prev);
3424 rq->skip_clock_update = 0;
3425
3426 if (likely(prev != next)) {
3427 rq->nr_switches++;
3428 rq->curr = next;
3429 ++*switch_count;
3430
3431 context_switch(rq, prev, next); /* unlocks the rq */
3432 /*
3433 * The context switch have flipped the stack from under us
3434 * and restored the local variables which were saved when
3435 * this task called schedule() in the past. prev == current
3436 * is still correct, but it can be moved to another cpu/rq.
3437 */
3438 cpu = smp_processor_id();
3439 rq = cpu_rq(cpu);
3440 } else
3441 raw_spin_unlock_irq(&rq->lock);
3442
3443 post_schedule(rq);
3444
3445 sched_preempt_enable_no_resched();
3446 if (need_resched())
3447 goto need_resched;
3448}
3449
3450static inline void sched_submit_work(struct task_struct *tsk)
3451{
3452 if (!tsk->state || tsk_is_pi_blocked(tsk))
3453 return;
3454 /*
3455 * If we are going to sleep and we have plugged IO queued,
3456 * make sure to submit it to avoid deadlocks.
3457 */
3458 if (blk_needs_flush_plug(tsk))
3459 blk_schedule_flush_plug(tsk);
3460}
3461
3462asmlinkage void __sched schedule(void)
3463{
3464 struct task_struct *tsk = current;
3465
3466 sched_submit_work(tsk);
3467 __schedule();
3468}
3469EXPORT_SYMBOL(schedule);
3470
3471/**
3472 * schedule_preempt_disabled - called with preemption disabled
3473 *
3474 * Returns with preemption disabled. Note: preempt_count must be 1
3475 */
3476void __sched schedule_preempt_disabled(void)
3477{
3478 sched_preempt_enable_no_resched();
3479 schedule();
3480 preempt_disable();
3481}
3482
3483#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3484
3485static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3486{
3487 if (lock->owner != owner)
3488 return false;
3489
3490 /*
3491 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3492 * lock->owner still matches owner, if that fails, owner might
3493 * point to free()d memory, if it still matches, the rcu_read_lock()
3494 * ensures the memory stays valid.
3495 */
3496 barrier();
3497
3498 return owner->on_cpu;
3499}
3500
3501/*
3502 * Look out! "owner" is an entirely speculative pointer
3503 * access and not reliable.
3504 */
3505int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3506{
3507 if (!sched_feat(OWNER_SPIN))
3508 return 0;
3509
3510 rcu_read_lock();
3511 while (owner_running(lock, owner)) {
3512 if (need_resched())
3513 break;
3514
3515 arch_mutex_cpu_relax();
3516 }
3517 rcu_read_unlock();
3518
3519 /*
3520 * We break out the loop above on need_resched() and when the
3521 * owner changed, which is a sign for heavy contention. Return
3522 * success only when lock->owner is NULL.
3523 */
3524 return lock->owner == NULL;
3525}
3526#endif
3527
3528#ifdef CONFIG_PREEMPT
3529/*
3530 * this is the entry point to schedule() from in-kernel preemption
3531 * off of preempt_enable. Kernel preemptions off return from interrupt
3532 * occur there and call schedule directly.
3533 */
3534asmlinkage void __sched notrace preempt_schedule(void)
3535{
3536 struct thread_info *ti = current_thread_info();
3537
3538 /*
3539 * If there is a non-zero preempt_count or interrupts are disabled,
3540 * we do not want to preempt the current task. Just return..
3541 */
3542 if (likely(ti->preempt_count || irqs_disabled()))
3543 return;
3544
3545 do {
3546 add_preempt_count_notrace(PREEMPT_ACTIVE);
3547 __schedule();
3548 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3549
3550 /*
3551 * Check again in case we missed a preemption opportunity
3552 * between schedule and now.
3553 */
3554 barrier();
3555 } while (need_resched());
3556}
3557EXPORT_SYMBOL(preempt_schedule);
3558
3559/*
3560 * this is the entry point to schedule() from kernel preemption
3561 * off of irq context.
3562 * Note, that this is called and return with irqs disabled. This will
3563 * protect us against recursive calling from irq.
3564 */
3565asmlinkage void __sched preempt_schedule_irq(void)
3566{
3567 struct thread_info *ti = current_thread_info();
3568
3569 /* Catch callers which need to be fixed */
3570 BUG_ON(ti->preempt_count || !irqs_disabled());
3571
3572 do {
3573 add_preempt_count(PREEMPT_ACTIVE);
3574 local_irq_enable();
3575 __schedule();
3576 local_irq_disable();
3577 sub_preempt_count(PREEMPT_ACTIVE);
3578
3579 /*
3580 * Check again in case we missed a preemption opportunity
3581 * between schedule and now.
3582 */
3583 barrier();
3584 } while (need_resched());
3585}
3586
3587#endif /* CONFIG_PREEMPT */
3588
3589int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3590 void *key)
3591{
3592 return try_to_wake_up(curr->private, mode, wake_flags);
3593}
3594EXPORT_SYMBOL(default_wake_function);
3595
3596/*
3597 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3598 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3599 * number) then we wake all the non-exclusive tasks and one exclusive task.
3600 *
3601 * There are circumstances in which we can try to wake a task which has already
3602 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3603 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3604 */
3605static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3606 int nr_exclusive, int wake_flags, void *key)
3607{
3608 wait_queue_t *curr, *next;
3609
3610 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3611 unsigned flags = curr->flags;
3612
3613 if (curr->func(curr, mode, wake_flags, key) &&
3614 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3615 break;
3616 }
3617}
3618
3619/**
3620 * __wake_up - wake up threads blocked on a waitqueue.
3621 * @q: the waitqueue
3622 * @mode: which threads
3623 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3624 * @key: is directly passed to the wakeup function
3625 *
3626 * It may be assumed that this function implies a write memory barrier before
3627 * changing the task state if and only if any tasks are woken up.
3628 */
3629void __wake_up(wait_queue_head_t *q, unsigned int mode,
3630 int nr_exclusive, void *key)
3631{
3632 unsigned long flags;
3633
3634 spin_lock_irqsave(&q->lock, flags);
3635 __wake_up_common(q, mode, nr_exclusive, 0, key);
3636 spin_unlock_irqrestore(&q->lock, flags);
3637}
3638EXPORT_SYMBOL(__wake_up);
3639
3640/*
3641 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3642 */
3643void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3644{
3645 __wake_up_common(q, mode, nr, 0, NULL);
3646}
3647EXPORT_SYMBOL_GPL(__wake_up_locked);
3648
3649void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3650{
3651 __wake_up_common(q, mode, 1, 0, key);
3652}
3653EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3654
3655/**
3656 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3657 * @q: the waitqueue
3658 * @mode: which threads
3659 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3660 * @key: opaque value to be passed to wakeup targets
3661 *
3662 * The sync wakeup differs that the waker knows that it will schedule
3663 * away soon, so while the target thread will be woken up, it will not
3664 * be migrated to another CPU - ie. the two threads are 'synchronized'
3665 * with each other. This can prevent needless bouncing between CPUs.
3666 *
3667 * On UP it can prevent extra preemption.
3668 *
3669 * It may be assumed that this function implies a write memory barrier before
3670 * changing the task state if and only if any tasks are woken up.
3671 */
3672void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3673 int nr_exclusive, void *key)
3674{
3675 unsigned long flags;
3676 int wake_flags = WF_SYNC;
3677
3678 if (unlikely(!q))
3679 return;
3680
3681 if (unlikely(!nr_exclusive))
3682 wake_flags = 0;
3683
3684 spin_lock_irqsave(&q->lock, flags);
3685 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3686 spin_unlock_irqrestore(&q->lock, flags);
3687}
3688EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3689
3690/*
3691 * __wake_up_sync - see __wake_up_sync_key()
3692 */
3693void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3694{
3695 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3696}
3697EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3698
3699/**
3700 * complete: - signals a single thread waiting on this completion
3701 * @x: holds the state of this particular completion
3702 *
3703 * This will wake up a single thread waiting on this completion. Threads will be
3704 * awakened in the same order in which they were queued.
3705 *
3706 * See also complete_all(), wait_for_completion() and related routines.
3707 *
3708 * It may be assumed that this function implies a write memory barrier before
3709 * changing the task state if and only if any tasks are woken up.
3710 */
3711void complete(struct completion *x)
3712{
3713 unsigned long flags;
3714
3715 spin_lock_irqsave(&x->wait.lock, flags);
3716 x->done++;
3717 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3718 spin_unlock_irqrestore(&x->wait.lock, flags);
3719}
3720EXPORT_SYMBOL(complete);
3721
3722/**
3723 * complete_all: - signals all threads waiting on this completion
3724 * @x: holds the state of this particular completion
3725 *
3726 * This will wake up all threads waiting on this particular completion event.
3727 *
3728 * It may be assumed that this function implies a write memory barrier before
3729 * changing the task state if and only if any tasks are woken up.
3730 */
3731void complete_all(struct completion *x)
3732{
3733 unsigned long flags;
3734
3735 spin_lock_irqsave(&x->wait.lock, flags);
3736 x->done += UINT_MAX/2;
3737 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3738 spin_unlock_irqrestore(&x->wait.lock, flags);
3739}
3740EXPORT_SYMBOL(complete_all);
3741
3742static inline long __sched
3743do_wait_for_common(struct completion *x, long timeout, int state)
3744{
3745 if (!x->done) {
3746 DECLARE_WAITQUEUE(wait, current);
3747
3748 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3749 do {
3750 if (signal_pending_state(state, current)) {
3751 timeout = -ERESTARTSYS;
3752 break;
3753 }
3754 __set_current_state(state);
3755 spin_unlock_irq(&x->wait.lock);
3756 timeout = schedule_timeout(timeout);
3757 spin_lock_irq(&x->wait.lock);
3758 } while (!x->done && timeout);
3759 __remove_wait_queue(&x->wait, &wait);
3760 if (!x->done)
3761 return timeout;
3762 }
3763 x->done--;
3764 return timeout ?: 1;
3765}
3766
3767static long __sched
3768wait_for_common(struct completion *x, long timeout, int state)
3769{
3770 might_sleep();
3771
3772 spin_lock_irq(&x->wait.lock);
3773 timeout = do_wait_for_common(x, timeout, state);
3774 spin_unlock_irq(&x->wait.lock);
3775 return timeout;
3776}
3777
3778/**
3779 * wait_for_completion: - waits for completion of a task
3780 * @x: holds the state of this particular completion
3781 *
3782 * This waits to be signaled for completion of a specific task. It is NOT
3783 * interruptible and there is no timeout.
3784 *
3785 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3786 * and interrupt capability. Also see complete().
3787 */
3788void __sched wait_for_completion(struct completion *x)
3789{
3790 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3791}
3792EXPORT_SYMBOL(wait_for_completion);
3793
3794/**
3795 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3796 * @x: holds the state of this particular completion
3797 * @timeout: timeout value in jiffies
3798 *
3799 * This waits for either a completion of a specific task to be signaled or for a
3800 * specified timeout to expire. The timeout is in jiffies. It is not
3801 * interruptible.
3802 *
3803 * The return value is 0 if timed out, and positive (at least 1, or number of
3804 * jiffies left till timeout) if completed.
3805 */
3806unsigned long __sched
3807wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3808{
3809 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3810}
3811EXPORT_SYMBOL(wait_for_completion_timeout);
3812
3813/**
3814 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3815 * @x: holds the state of this particular completion
3816 *
3817 * This waits for completion of a specific task to be signaled. It is
3818 * interruptible.
3819 *
3820 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3821 */
3822int __sched wait_for_completion_interruptible(struct completion *x)
3823{
3824 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3825 if (t == -ERESTARTSYS)
3826 return t;
3827 return 0;
3828}
3829EXPORT_SYMBOL(wait_for_completion_interruptible);
3830
3831/**
3832 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3833 * @x: holds the state of this particular completion
3834 * @timeout: timeout value in jiffies
3835 *
3836 * This waits for either a completion of a specific task to be signaled or for a
3837 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3838 *
3839 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3840 * positive (at least 1, or number of jiffies left till timeout) if completed.
3841 */
3842long __sched
3843wait_for_completion_interruptible_timeout(struct completion *x,
3844 unsigned long timeout)
3845{
3846 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3847}
3848EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3849
3850/**
3851 * wait_for_completion_killable: - waits for completion of a task (killable)
3852 * @x: holds the state of this particular completion
3853 *
3854 * This waits to be signaled for completion of a specific task. It can be
3855 * interrupted by a kill signal.
3856 *
3857 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3858 */
3859int __sched wait_for_completion_killable(struct completion *x)
3860{
3861 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3862 if (t == -ERESTARTSYS)
3863 return t;
3864 return 0;
3865}
3866EXPORT_SYMBOL(wait_for_completion_killable);
3867
3868/**
3869 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3870 * @x: holds the state of this particular completion
3871 * @timeout: timeout value in jiffies
3872 *
3873 * This waits for either a completion of a specific task to be
3874 * signaled or for a specified timeout to expire. It can be
3875 * interrupted by a kill signal. The timeout is in jiffies.
3876 *
3877 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3878 * positive (at least 1, or number of jiffies left till timeout) if completed.
3879 */
3880long __sched
3881wait_for_completion_killable_timeout(struct completion *x,
3882 unsigned long timeout)
3883{
3884 return wait_for_common(x, timeout, TASK_KILLABLE);
3885}
3886EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3887
3888/**
3889 * try_wait_for_completion - try to decrement a completion without blocking
3890 * @x: completion structure
3891 *
3892 * Returns: 0 if a decrement cannot be done without blocking
3893 * 1 if a decrement succeeded.
3894 *
3895 * If a completion is being used as a counting completion,
3896 * attempt to decrement the counter without blocking. This
3897 * enables us to avoid waiting if the resource the completion
3898 * is protecting is not available.
3899 */
3900bool try_wait_for_completion(struct completion *x)
3901{
3902 unsigned long flags;
3903 int ret = 1;
3904
3905 spin_lock_irqsave(&x->wait.lock, flags);
3906 if (!x->done)
3907 ret = 0;
3908 else
3909 x->done--;
3910 spin_unlock_irqrestore(&x->wait.lock, flags);
3911 return ret;
3912}
3913EXPORT_SYMBOL(try_wait_for_completion);
3914
3915/**
3916 * completion_done - Test to see if a completion has any waiters
3917 * @x: completion structure
3918 *
3919 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3920 * 1 if there are no waiters.
3921 *
3922 */
3923bool completion_done(struct completion *x)
3924{
3925 unsigned long flags;
3926 int ret = 1;
3927
3928 spin_lock_irqsave(&x->wait.lock, flags);
3929 if (!x->done)
3930 ret = 0;
3931 spin_unlock_irqrestore(&x->wait.lock, flags);
3932 return ret;
3933}
3934EXPORT_SYMBOL(completion_done);
3935
3936static long __sched
3937sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3938{
3939 unsigned long flags;
3940 wait_queue_t wait;
3941
3942 init_waitqueue_entry(&wait, current);
3943
3944 __set_current_state(state);
3945
3946 spin_lock_irqsave(&q->lock, flags);
3947 __add_wait_queue(q, &wait);
3948 spin_unlock(&q->lock);
3949 timeout = schedule_timeout(timeout);
3950 spin_lock_irq(&q->lock);
3951 __remove_wait_queue(q, &wait);
3952 spin_unlock_irqrestore(&q->lock, flags);
3953
3954 return timeout;
3955}
3956
3957void __sched interruptible_sleep_on(wait_queue_head_t *q)
3958{
3959 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3960}
3961EXPORT_SYMBOL(interruptible_sleep_on);
3962
3963long __sched
3964interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3965{
3966 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3967}
3968EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3969
3970void __sched sleep_on(wait_queue_head_t *q)
3971{
3972 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3973}
3974EXPORT_SYMBOL(sleep_on);
3975
3976long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3977{
3978 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3979}
3980EXPORT_SYMBOL(sleep_on_timeout);
3981
3982#ifdef CONFIG_RT_MUTEXES
3983
3984/*
3985 * rt_mutex_setprio - set the current priority of a task
3986 * @p: task
3987 * @prio: prio value (kernel-internal form)
3988 *
3989 * This function changes the 'effective' priority of a task. It does
3990 * not touch ->normal_prio like __setscheduler().
3991 *
3992 * Used by the rt_mutex code to implement priority inheritance logic.
3993 */
3994void rt_mutex_setprio(struct task_struct *p, int prio)
3995{
3996 int oldprio, on_rq, running;
3997 struct rq *rq;
3998 const struct sched_class *prev_class;
3999
4000 BUG_ON(prio < 0 || prio > MAX_PRIO);
4001
4002 rq = __task_rq_lock(p);
4003
4004 /*
4005 * Idle task boosting is a nono in general. There is one
4006 * exception, when PREEMPT_RT and NOHZ is active:
4007 *
4008 * The idle task calls get_next_timer_interrupt() and holds
4009 * the timer wheel base->lock on the CPU and another CPU wants
4010 * to access the timer (probably to cancel it). We can safely
4011 * ignore the boosting request, as the idle CPU runs this code
4012 * with interrupts disabled and will complete the lock
4013 * protected section without being interrupted. So there is no
4014 * real need to boost.
4015 */
4016 if (unlikely(p == rq->idle)) {
4017 WARN_ON(p != rq->curr);
4018 WARN_ON(p->pi_blocked_on);
4019 goto out_unlock;
4020 }
4021
4022 trace_sched_pi_setprio(p, prio);
4023 oldprio = p->prio;
4024 prev_class = p->sched_class;
4025 on_rq = p->on_rq;
4026 running = task_current(rq, p);
4027 if (on_rq)
4028 dequeue_task(rq, p, 0);
4029 if (running)
4030 p->sched_class->put_prev_task(rq, p);
4031
4032 if (rt_prio(prio))
4033 p->sched_class = &rt_sched_class;
4034 else
4035 p->sched_class = &fair_sched_class;
4036
4037 p->prio = prio;
4038
4039 if (running)
4040 p->sched_class->set_curr_task(rq);
4041 if (on_rq)
4042 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4043
4044 check_class_changed(rq, p, prev_class, oldprio);
4045out_unlock:
4046 __task_rq_unlock(rq);
4047}
4048#endif
4049void set_user_nice(struct task_struct *p, long nice)
4050{
4051 int old_prio, delta, on_rq;
4052 unsigned long flags;
4053 struct rq *rq;
4054
4055 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4056 return;
4057 /*
4058 * We have to be careful, if called from sys_setpriority(),
4059 * the task might be in the middle of scheduling on another CPU.
4060 */
4061 rq = task_rq_lock(p, &flags);
4062 /*
4063 * The RT priorities are set via sched_setscheduler(), but we still
4064 * allow the 'normal' nice value to be set - but as expected
4065 * it wont have any effect on scheduling until the task is
4066 * SCHED_FIFO/SCHED_RR:
4067 */
4068 if (task_has_rt_policy(p)) {
4069 p->static_prio = NICE_TO_PRIO(nice);
4070 goto out_unlock;
4071 }
4072 on_rq = p->on_rq;
4073 if (on_rq)
4074 dequeue_task(rq, p, 0);
4075
4076 p->static_prio = NICE_TO_PRIO(nice);
4077 set_load_weight(p);
4078 old_prio = p->prio;
4079 p->prio = effective_prio(p);
4080 delta = p->prio - old_prio;
4081
4082 if (on_rq) {
4083 enqueue_task(rq, p, 0);
4084 /*
4085 * If the task increased its priority or is running and
4086 * lowered its priority, then reschedule its CPU:
4087 */
4088 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4089 resched_task(rq->curr);
4090 }
4091out_unlock:
4092 task_rq_unlock(rq, p, &flags);
4093}
4094EXPORT_SYMBOL(set_user_nice);
4095
4096/*
4097 * can_nice - check if a task can reduce its nice value
4098 * @p: task
4099 * @nice: nice value
4100 */
4101int can_nice(const struct task_struct *p, const int nice)
4102{
4103 /* convert nice value [19,-20] to rlimit style value [1,40] */
4104 int nice_rlim = 20 - nice;
4105
4106 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4107 capable(CAP_SYS_NICE));
4108}
4109
4110#ifdef __ARCH_WANT_SYS_NICE
4111
4112/*
4113 * sys_nice - change the priority of the current process.
4114 * @increment: priority increment
4115 *
4116 * sys_setpriority is a more generic, but much slower function that
4117 * does similar things.
4118 */
4119SYSCALL_DEFINE1(nice, int, increment)
4120{
4121 long nice, retval;
4122
4123 /*
4124 * Setpriority might change our priority at the same moment.
4125 * We don't have to worry. Conceptually one call occurs first
4126 * and we have a single winner.
4127 */
4128 if (increment < -40)
4129 increment = -40;
4130 if (increment > 40)
4131 increment = 40;
4132
4133 nice = TASK_NICE(current) + increment;
4134 if (nice < -20)
4135 nice = -20;
4136 if (nice > 19)
4137 nice = 19;
4138
4139 if (increment < 0 && !can_nice(current, nice))
4140 return -EPERM;
4141
4142 retval = security_task_setnice(current, nice);
4143 if (retval)
4144 return retval;
4145
4146 set_user_nice(current, nice);
4147 return 0;
4148}
4149
4150#endif
4151
4152/**
4153 * task_prio - return the priority value of a given task.
4154 * @p: the task in question.
4155 *
4156 * This is the priority value as seen by users in /proc.
4157 * RT tasks are offset by -200. Normal tasks are centered
4158 * around 0, value goes from -16 to +15.
4159 */
4160int task_prio(const struct task_struct *p)
4161{
4162 return p->prio - MAX_RT_PRIO;
4163}
4164
4165/**
4166 * task_nice - return the nice value of a given task.
4167 * @p: the task in question.
4168 */
4169int task_nice(const struct task_struct *p)
4170{
4171 return TASK_NICE(p);
4172}
4173EXPORT_SYMBOL(task_nice);
4174
4175/**
4176 * idle_cpu - is a given cpu idle currently?
4177 * @cpu: the processor in question.
4178 */
4179int idle_cpu(int cpu)
4180{
4181 struct rq *rq = cpu_rq(cpu);
4182
4183 if (rq->curr != rq->idle)
4184 return 0;
4185
4186 if (rq->nr_running)
4187 return 0;
4188
4189#ifdef CONFIG_SMP
4190 if (!llist_empty(&rq->wake_list))
4191 return 0;
4192#endif
4193
4194 return 1;
4195}
4196
4197/**
4198 * idle_task - return the idle task for a given cpu.
4199 * @cpu: the processor in question.
4200 */
4201struct task_struct *idle_task(int cpu)
4202{
4203 return cpu_rq(cpu)->idle;
4204}
4205
4206/**
4207 * find_process_by_pid - find a process with a matching PID value.
4208 * @pid: the pid in question.
4209 */
4210static struct task_struct *find_process_by_pid(pid_t pid)
4211{
4212 return pid ? find_task_by_vpid(pid) : current;
4213}
4214
4215/* Actually do priority change: must hold rq lock. */
4216static void
4217__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4218{
4219 p->policy = policy;
4220 p->rt_priority = prio;
4221 p->normal_prio = normal_prio(p);
4222 /* we are holding p->pi_lock already */
4223 p->prio = rt_mutex_getprio(p);
4224 if (rt_prio(p->prio))
4225 p->sched_class = &rt_sched_class;
4226 else
4227 p->sched_class = &fair_sched_class;
4228 set_load_weight(p);
4229}
4230
4231/*
4232 * check the target process has a UID that matches the current process's
4233 */
4234static bool check_same_owner(struct task_struct *p)
4235{
4236 const struct cred *cred = current_cred(), *pcred;
4237 bool match;
4238
4239 rcu_read_lock();
4240 pcred = __task_cred(p);
4241 match = (uid_eq(cred->euid, pcred->euid) ||
4242 uid_eq(cred->euid, pcred->uid));
4243 rcu_read_unlock();
4244 return match;
4245}
4246
4247static int __sched_setscheduler(struct task_struct *p, int policy,
4248 const struct sched_param *param, bool user)
4249{
4250 int retval, oldprio, oldpolicy = -1, on_rq, running;
4251 unsigned long flags;
4252 const struct sched_class *prev_class;
4253 struct rq *rq;
4254 int reset_on_fork;
4255
4256 /* may grab non-irq protected spin_locks */
4257 BUG_ON(in_interrupt());
4258recheck:
4259 /* double check policy once rq lock held */
4260 if (policy < 0) {
4261 reset_on_fork = p->sched_reset_on_fork;
4262 policy = oldpolicy = p->policy;
4263 } else {
4264 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4265 policy &= ~SCHED_RESET_ON_FORK;
4266
4267 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4268 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4269 policy != SCHED_IDLE)
4270 return -EINVAL;
4271 }
4272
4273 /*
4274 * Valid priorities for SCHED_FIFO and SCHED_RR are
4275 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4276 * SCHED_BATCH and SCHED_IDLE is 0.
4277 */
4278 if (param->sched_priority < 0 ||
4279 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4280 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4281 return -EINVAL;
4282 if (rt_policy(policy) != (param->sched_priority != 0))
4283 return -EINVAL;
4284
4285 /*
4286 * Allow unprivileged RT tasks to decrease priority:
4287 */
4288 if (user && !capable(CAP_SYS_NICE)) {
4289 if (rt_policy(policy)) {
4290 unsigned long rlim_rtprio =
4291 task_rlimit(p, RLIMIT_RTPRIO);
4292
4293 /* can't set/change the rt policy */
4294 if (policy != p->policy && !rlim_rtprio)
4295 return -EPERM;
4296
4297 /* can't increase priority */
4298 if (param->sched_priority > p->rt_priority &&
4299 param->sched_priority > rlim_rtprio)
4300 return -EPERM;
4301 }
4302
4303 /*
4304 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4305 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4306 */
4307 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4308 if (!can_nice(p, TASK_NICE(p)))
4309 return -EPERM;
4310 }
4311
4312 /* can't change other user's priorities */
4313 if (!check_same_owner(p))
4314 return -EPERM;
4315
4316 /* Normal users shall not reset the sched_reset_on_fork flag */
4317 if (p->sched_reset_on_fork && !reset_on_fork)
4318 return -EPERM;
4319 }
4320
4321 if (user) {
4322 retval = security_task_setscheduler(p);
4323 if (retval)
4324 return retval;
4325 }
4326
4327 /*
4328 * make sure no PI-waiters arrive (or leave) while we are
4329 * changing the priority of the task:
4330 *
4331 * To be able to change p->policy safely, the appropriate
4332 * runqueue lock must be held.
4333 */
4334 rq = task_rq_lock(p, &flags);
4335
4336 /*
4337 * Changing the policy of the stop threads its a very bad idea
4338 */
4339 if (p == rq->stop) {
4340 task_rq_unlock(rq, p, &flags);
4341 return -EINVAL;
4342 }
4343
4344 /*
4345 * If not changing anything there's no need to proceed further:
4346 */
4347 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4348 param->sched_priority == p->rt_priority))) {
4349
4350 __task_rq_unlock(rq);
4351 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4352 return 0;
4353 }
4354
4355#ifdef CONFIG_RT_GROUP_SCHED
4356 if (user) {
4357 /*
4358 * Do not allow realtime tasks into groups that have no runtime
4359 * assigned.
4360 */
4361 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4362 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4363 !task_group_is_autogroup(task_group(p))) {
4364 task_rq_unlock(rq, p, &flags);
4365 return -EPERM;
4366 }
4367 }
4368#endif
4369
4370 /* recheck policy now with rq lock held */
4371 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4372 policy = oldpolicy = -1;
4373 task_rq_unlock(rq, p, &flags);
4374 goto recheck;
4375 }
4376 on_rq = p->on_rq;
4377 running = task_current(rq, p);
4378 if (on_rq)
4379 dequeue_task(rq, p, 0);
4380 if (running)
4381 p->sched_class->put_prev_task(rq, p);
4382
4383 p->sched_reset_on_fork = reset_on_fork;
4384
4385 oldprio = p->prio;
4386 prev_class = p->sched_class;
4387 __setscheduler(rq, p, policy, param->sched_priority);
4388
4389 if (running)
4390 p->sched_class->set_curr_task(rq);
4391 if (on_rq)
4392 enqueue_task(rq, p, 0);
4393
4394 check_class_changed(rq, p, prev_class, oldprio);
4395 task_rq_unlock(rq, p, &flags);
4396
4397 rt_mutex_adjust_pi(p);
4398
4399 return 0;
4400}
4401
4402/**
4403 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4404 * @p: the task in question.
4405 * @policy: new policy.
4406 * @param: structure containing the new RT priority.
4407 *
4408 * NOTE that the task may be already dead.
4409 */
4410int sched_setscheduler(struct task_struct *p, int policy,
4411 const struct sched_param *param)
4412{
4413 return __sched_setscheduler(p, policy, param, true);
4414}
4415EXPORT_SYMBOL_GPL(sched_setscheduler);
4416
4417/**
4418 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4419 * @p: the task in question.
4420 * @policy: new policy.
4421 * @param: structure containing the new RT priority.
4422 *
4423 * Just like sched_setscheduler, only don't bother checking if the
4424 * current context has permission. For example, this is needed in
4425 * stop_machine(): we create temporary high priority worker threads,
4426 * but our caller might not have that capability.
4427 */
4428int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4429 const struct sched_param *param)
4430{
4431 return __sched_setscheduler(p, policy, param, false);
4432}
4433
4434static int
4435do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4436{
4437 struct sched_param lparam;
4438 struct task_struct *p;
4439 int retval;
4440
4441 if (!param || pid < 0)
4442 return -EINVAL;
4443 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4444 return -EFAULT;
4445
4446 rcu_read_lock();
4447 retval = -ESRCH;
4448 p = find_process_by_pid(pid);
4449 if (p != NULL)
4450 retval = sched_setscheduler(p, policy, &lparam);
4451 rcu_read_unlock();
4452
4453 return retval;
4454}
4455
4456/**
4457 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4458 * @pid: the pid in question.
4459 * @policy: new policy.
4460 * @param: structure containing the new RT priority.
4461 */
4462SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4463 struct sched_param __user *, param)
4464{
4465 /* negative values for policy are not valid */
4466 if (policy < 0)
4467 return -EINVAL;
4468
4469 return do_sched_setscheduler(pid, policy, param);
4470}
4471
4472/**
4473 * sys_sched_setparam - set/change the RT priority of a thread
4474 * @pid: the pid in question.
4475 * @param: structure containing the new RT priority.
4476 */
4477SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4478{
4479 return do_sched_setscheduler(pid, -1, param);
4480}
4481
4482/**
4483 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4484 * @pid: the pid in question.
4485 */
4486SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4487{
4488 struct task_struct *p;
4489 int retval;
4490
4491 if (pid < 0)
4492 return -EINVAL;
4493
4494 retval = -ESRCH;
4495 rcu_read_lock();
4496 p = find_process_by_pid(pid);
4497 if (p) {
4498 retval = security_task_getscheduler(p);
4499 if (!retval)
4500 retval = p->policy
4501 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4502 }
4503 rcu_read_unlock();
4504 return retval;
4505}
4506
4507/**
4508 * sys_sched_getparam - get the RT priority of a thread
4509 * @pid: the pid in question.
4510 * @param: structure containing the RT priority.
4511 */
4512SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4513{
4514 struct sched_param lp;
4515 struct task_struct *p;
4516 int retval;
4517
4518 if (!param || pid < 0)
4519 return -EINVAL;
4520
4521 rcu_read_lock();
4522 p = find_process_by_pid(pid);
4523 retval = -ESRCH;
4524 if (!p)
4525 goto out_unlock;
4526
4527 retval = security_task_getscheduler(p);
4528 if (retval)
4529 goto out_unlock;
4530
4531 lp.sched_priority = p->rt_priority;
4532 rcu_read_unlock();
4533
4534 /*
4535 * This one might sleep, we cannot do it with a spinlock held ...
4536 */
4537 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4538
4539 return retval;
4540
4541out_unlock:
4542 rcu_read_unlock();
4543 return retval;
4544}
4545
4546long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4547{
4548 cpumask_var_t cpus_allowed, new_mask;
4549 struct task_struct *p;
4550 int retval;
4551
4552 get_online_cpus();
4553 rcu_read_lock();
4554
4555 p = find_process_by_pid(pid);
4556 if (!p) {
4557 rcu_read_unlock();
4558 put_online_cpus();
4559 return -ESRCH;
4560 }
4561
4562 /* Prevent p going away */
4563 get_task_struct(p);
4564 rcu_read_unlock();
4565
4566 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4567 retval = -ENOMEM;
4568 goto out_put_task;
4569 }
4570 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4571 retval = -ENOMEM;
4572 goto out_free_cpus_allowed;
4573 }
4574 retval = -EPERM;
4575 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4576 goto out_unlock;
4577
4578 retval = security_task_setscheduler(p);
4579 if (retval)
4580 goto out_unlock;
4581
4582 cpuset_cpus_allowed(p, cpus_allowed);
4583 cpumask_and(new_mask, in_mask, cpus_allowed);
4584again:
4585 retval = set_cpus_allowed_ptr(p, new_mask);
4586
4587 if (!retval) {
4588 cpuset_cpus_allowed(p, cpus_allowed);
4589 if (!cpumask_subset(new_mask, cpus_allowed)) {
4590 /*
4591 * We must have raced with a concurrent cpuset
4592 * update. Just reset the cpus_allowed to the
4593 * cpuset's cpus_allowed
4594 */
4595 cpumask_copy(new_mask, cpus_allowed);
4596 goto again;
4597 }
4598 }
4599out_unlock:
4600 free_cpumask_var(new_mask);
4601out_free_cpus_allowed:
4602 free_cpumask_var(cpus_allowed);
4603out_put_task:
4604 put_task_struct(p);
4605 put_online_cpus();
4606 return retval;
4607}
4608
4609static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4610 struct cpumask *new_mask)
4611{
4612 if (len < cpumask_size())
4613 cpumask_clear(new_mask);
4614 else if (len > cpumask_size())
4615 len = cpumask_size();
4616
4617 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4618}
4619
4620/**
4621 * sys_sched_setaffinity - set the cpu affinity of a process
4622 * @pid: pid of the process
4623 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4624 * @user_mask_ptr: user-space pointer to the new cpu mask
4625 */
4626SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4627 unsigned long __user *, user_mask_ptr)
4628{
4629 cpumask_var_t new_mask;
4630 int retval;
4631
4632 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4633 return -ENOMEM;
4634
4635 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4636 if (retval == 0)
4637 retval = sched_setaffinity(pid, new_mask);
4638 free_cpumask_var(new_mask);
4639 return retval;
4640}
4641
4642long sched_getaffinity(pid_t pid, struct cpumask *mask)
4643{
4644 struct task_struct *p;
4645 unsigned long flags;
4646 int retval;
4647
4648 get_online_cpus();
4649 rcu_read_lock();
4650
4651 retval = -ESRCH;
4652 p = find_process_by_pid(pid);
4653 if (!p)
4654 goto out_unlock;
4655
4656 retval = security_task_getscheduler(p);
4657 if (retval)
4658 goto out_unlock;
4659
4660 raw_spin_lock_irqsave(&p->pi_lock, flags);
4661 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4662 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4663
4664out_unlock:
4665 rcu_read_unlock();
4666 put_online_cpus();
4667
4668 return retval;
4669}
4670
4671/**
4672 * sys_sched_getaffinity - get the cpu affinity of a process
4673 * @pid: pid of the process
4674 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4675 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4676 */
4677SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4678 unsigned long __user *, user_mask_ptr)
4679{
4680 int ret;
4681 cpumask_var_t mask;
4682
4683 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4684 return -EINVAL;
4685 if (len & (sizeof(unsigned long)-1))
4686 return -EINVAL;
4687
4688 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4689 return -ENOMEM;
4690
4691 ret = sched_getaffinity(pid, mask);
4692 if (ret == 0) {
4693 size_t retlen = min_t(size_t, len, cpumask_size());
4694
4695 if (copy_to_user(user_mask_ptr, mask, retlen))
4696 ret = -EFAULT;
4697 else
4698 ret = retlen;
4699 }
4700 free_cpumask_var(mask);
4701
4702 return ret;
4703}
4704
4705/**
4706 * sys_sched_yield - yield the current processor to other threads.
4707 *
4708 * This function yields the current CPU to other tasks. If there are no
4709 * other threads running on this CPU then this function will return.
4710 */
4711SYSCALL_DEFINE0(sched_yield)
4712{
4713 struct rq *rq = this_rq_lock();
4714
4715 schedstat_inc(rq, yld_count);
4716 current->sched_class->yield_task(rq);
4717
4718 /*
4719 * Since we are going to call schedule() anyway, there's
4720 * no need to preempt or enable interrupts:
4721 */
4722 __release(rq->lock);
4723 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4724 do_raw_spin_unlock(&rq->lock);
4725 sched_preempt_enable_no_resched();
4726
4727 schedule();
4728
4729 return 0;
4730}
4731
4732static inline int should_resched(void)
4733{
4734 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4735}
4736
4737static void __cond_resched(void)
4738{
4739 add_preempt_count(PREEMPT_ACTIVE);
4740 __schedule();
4741 sub_preempt_count(PREEMPT_ACTIVE);
4742}
4743
4744int __sched _cond_resched(void)
4745{
4746 if (should_resched()) {
4747 __cond_resched();
4748 return 1;
4749 }
4750 return 0;
4751}
4752EXPORT_SYMBOL(_cond_resched);
4753
4754/*
4755 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4756 * call schedule, and on return reacquire the lock.
4757 *
4758 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4759 * operations here to prevent schedule() from being called twice (once via
4760 * spin_unlock(), once by hand).
4761 */
4762int __cond_resched_lock(spinlock_t *lock)
4763{
4764 int resched = should_resched();
4765 int ret = 0;
4766
4767 lockdep_assert_held(lock);
4768
4769 if (spin_needbreak(lock) || resched) {
4770 spin_unlock(lock);
4771 if (resched)
4772 __cond_resched();
4773 else
4774 cpu_relax();
4775 ret = 1;
4776 spin_lock(lock);
4777 }
4778 return ret;
4779}
4780EXPORT_SYMBOL(__cond_resched_lock);
4781
4782int __sched __cond_resched_softirq(void)
4783{
4784 BUG_ON(!in_softirq());
4785
4786 if (should_resched()) {
4787 local_bh_enable();
4788 __cond_resched();
4789 local_bh_disable();
4790 return 1;
4791 }
4792 return 0;
4793}
4794EXPORT_SYMBOL(__cond_resched_softirq);
4795
4796/**
4797 * yield - yield the current processor to other threads.
4798 *
4799 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4800 *
4801 * The scheduler is at all times free to pick the calling task as the most
4802 * eligible task to run, if removing the yield() call from your code breaks
4803 * it, its already broken.
4804 *
4805 * Typical broken usage is:
4806 *
4807 * while (!event)
4808 * yield();
4809 *
4810 * where one assumes that yield() will let 'the other' process run that will
4811 * make event true. If the current task is a SCHED_FIFO task that will never
4812 * happen. Never use yield() as a progress guarantee!!
4813 *
4814 * If you want to use yield() to wait for something, use wait_event().
4815 * If you want to use yield() to be 'nice' for others, use cond_resched().
4816 * If you still want to use yield(), do not!
4817 */
4818void __sched yield(void)
4819{
4820 set_current_state(TASK_RUNNING);
4821 sys_sched_yield();
4822}
4823EXPORT_SYMBOL(yield);
4824
4825/**
4826 * yield_to - yield the current processor to another thread in
4827 * your thread group, or accelerate that thread toward the
4828 * processor it's on.
4829 * @p: target task
4830 * @preempt: whether task preemption is allowed or not
4831 *
4832 * It's the caller's job to ensure that the target task struct
4833 * can't go away on us before we can do any checks.
4834 *
4835 * Returns true if we indeed boosted the target task.
4836 */
4837bool __sched yield_to(struct task_struct *p, bool preempt)
4838{
4839 struct task_struct *curr = current;
4840 struct rq *rq, *p_rq;
4841 unsigned long flags;
4842 bool yielded = 0;
4843
4844 local_irq_save(flags);
4845 rq = this_rq();
4846
4847again:
4848 p_rq = task_rq(p);
4849 double_rq_lock(rq, p_rq);
4850 while (task_rq(p) != p_rq) {
4851 double_rq_unlock(rq, p_rq);
4852 goto again;
4853 }
4854
4855 if (!curr->sched_class->yield_to_task)
4856 goto out;
4857
4858 if (curr->sched_class != p->sched_class)
4859 goto out;
4860
4861 if (task_running(p_rq, p) || p->state)
4862 goto out;
4863
4864 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4865 if (yielded) {
4866 schedstat_inc(rq, yld_count);
4867 /*
4868 * Make p's CPU reschedule; pick_next_entity takes care of
4869 * fairness.
4870 */
4871 if (preempt && rq != p_rq)
4872 resched_task(p_rq->curr);
4873 } else {
4874 /*
4875 * We might have set it in task_yield_fair(), but are
4876 * not going to schedule(), so don't want to skip
4877 * the next update.
4878 */
4879 rq->skip_clock_update = 0;
4880 }
4881
4882out:
4883 double_rq_unlock(rq, p_rq);
4884 local_irq_restore(flags);
4885
4886 if (yielded)
4887 schedule();
4888
4889 return yielded;
4890}
4891EXPORT_SYMBOL_GPL(yield_to);
4892
4893/*
4894 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4895 * that process accounting knows that this is a task in IO wait state.
4896 */
4897void __sched io_schedule(void)
4898{
4899 struct rq *rq = raw_rq();
4900
4901 delayacct_blkio_start();
4902 atomic_inc(&rq->nr_iowait);
4903 blk_flush_plug(current);
4904 current->in_iowait = 1;
4905 schedule();
4906 current->in_iowait = 0;
4907 atomic_dec(&rq->nr_iowait);
4908 delayacct_blkio_end();
4909}
4910EXPORT_SYMBOL(io_schedule);
4911
4912long __sched io_schedule_timeout(long timeout)
4913{
4914 struct rq *rq = raw_rq();
4915 long ret;
4916
4917 delayacct_blkio_start();
4918 atomic_inc(&rq->nr_iowait);
4919 blk_flush_plug(current);
4920 current->in_iowait = 1;
4921 ret = schedule_timeout(timeout);
4922 current->in_iowait = 0;
4923 atomic_dec(&rq->nr_iowait);
4924 delayacct_blkio_end();
4925 return ret;
4926}
4927
4928/**
4929 * sys_sched_get_priority_max - return maximum RT priority.
4930 * @policy: scheduling class.
4931 *
4932 * this syscall returns the maximum rt_priority that can be used
4933 * by a given scheduling class.
4934 */
4935SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4936{
4937 int ret = -EINVAL;
4938
4939 switch (policy) {
4940 case SCHED_FIFO:
4941 case SCHED_RR:
4942 ret = MAX_USER_RT_PRIO-1;
4943 break;
4944 case SCHED_NORMAL:
4945 case SCHED_BATCH:
4946 case SCHED_IDLE:
4947 ret = 0;
4948 break;
4949 }
4950 return ret;
4951}
4952
4953/**
4954 * sys_sched_get_priority_min - return minimum RT priority.
4955 * @policy: scheduling class.
4956 *
4957 * this syscall returns the minimum rt_priority that can be used
4958 * by a given scheduling class.
4959 */
4960SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4961{
4962 int ret = -EINVAL;
4963
4964 switch (policy) {
4965 case SCHED_FIFO:
4966 case SCHED_RR:
4967 ret = 1;
4968 break;
4969 case SCHED_NORMAL:
4970 case SCHED_BATCH:
4971 case SCHED_IDLE:
4972 ret = 0;
4973 }
4974 return ret;
4975}
4976
4977/**
4978 * sys_sched_rr_get_interval - return the default timeslice of a process.
4979 * @pid: pid of the process.
4980 * @interval: userspace pointer to the timeslice value.
4981 *
4982 * this syscall writes the default timeslice value of a given process
4983 * into the user-space timespec buffer. A value of '0' means infinity.
4984 */
4985SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4986 struct timespec __user *, interval)
4987{
4988 struct task_struct *p;
4989 unsigned int time_slice;
4990 unsigned long flags;
4991 struct rq *rq;
4992 int retval;
4993 struct timespec t;
4994
4995 if (pid < 0)
4996 return -EINVAL;
4997
4998 retval = -ESRCH;
4999 rcu_read_lock();
5000 p = find_process_by_pid(pid);
5001 if (!p)
5002 goto out_unlock;
5003
5004 retval = security_task_getscheduler(p);
5005 if (retval)
5006 goto out_unlock;
5007
5008 rq = task_rq_lock(p, &flags);
5009 time_slice = p->sched_class->get_rr_interval(rq, p);
5010 task_rq_unlock(rq, p, &flags);
5011
5012 rcu_read_unlock();
5013 jiffies_to_timespec(time_slice, &t);
5014 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5015 return retval;
5016
5017out_unlock:
5018 rcu_read_unlock();
5019 return retval;
5020}
5021
5022static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5023
5024void sched_show_task(struct task_struct *p)
5025{
5026 unsigned long free = 0;
5027 unsigned state;
5028
5029 state = p->state ? __ffs(p->state) + 1 : 0;
5030 printk(KERN_INFO "%-15.15s %c", p->comm,
5031 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5032#if BITS_PER_LONG == 32
5033 if (state == TASK_RUNNING)
5034 printk(KERN_CONT " running ");
5035 else
5036 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5037#else
5038 if (state == TASK_RUNNING)
5039 printk(KERN_CONT " running task ");
5040 else
5041 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5042#endif
5043#ifdef CONFIG_DEBUG_STACK_USAGE
5044 free = stack_not_used(p);
5045#endif
5046 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5047 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
5048 (unsigned long)task_thread_info(p)->flags);
5049
5050 show_stack(p, NULL);
5051}
5052
5053void show_state_filter(unsigned long state_filter)
5054{
5055 struct task_struct *g, *p;
5056
5057#if BITS_PER_LONG == 32
5058 printk(KERN_INFO
5059 " task PC stack pid father\n");
5060#else
5061 printk(KERN_INFO
5062 " task PC stack pid father\n");
5063#endif
5064 rcu_read_lock();
5065 do_each_thread(g, p) {
5066 /*
5067 * reset the NMI-timeout, listing all files on a slow
5068 * console might take a lot of time:
5069 */
5070 touch_nmi_watchdog();
5071 if (!state_filter || (p->state & state_filter))
5072 sched_show_task(p);
5073 } while_each_thread(g, p);
5074
5075 touch_all_softlockup_watchdogs();
5076
5077#ifdef CONFIG_SCHED_DEBUG
5078 sysrq_sched_debug_show();
5079#endif
5080 rcu_read_unlock();
5081 /*
5082 * Only show locks if all tasks are dumped:
5083 */
5084 if (!state_filter)
5085 debug_show_all_locks();
5086}
5087
5088void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5089{
5090 idle->sched_class = &idle_sched_class;
5091}
5092
5093/**
5094 * init_idle - set up an idle thread for a given CPU
5095 * @idle: task in question
5096 * @cpu: cpu the idle task belongs to
5097 *
5098 * NOTE: this function does not set the idle thread's NEED_RESCHED
5099 * flag, to make booting more robust.
5100 */
5101void __cpuinit init_idle(struct task_struct *idle, int cpu)
5102{
5103 struct rq *rq = cpu_rq(cpu);
5104 unsigned long flags;
5105
5106 raw_spin_lock_irqsave(&rq->lock, flags);
5107
5108 __sched_fork(idle);
5109 idle->state = TASK_RUNNING;
5110 idle->se.exec_start = sched_clock();
5111
5112 do_set_cpus_allowed(idle, cpumask_of(cpu));
5113 /*
5114 * We're having a chicken and egg problem, even though we are
5115 * holding rq->lock, the cpu isn't yet set to this cpu so the
5116 * lockdep check in task_group() will fail.
5117 *
5118 * Similar case to sched_fork(). / Alternatively we could
5119 * use task_rq_lock() here and obtain the other rq->lock.
5120 *
5121 * Silence PROVE_RCU
5122 */
5123 rcu_read_lock();
5124 __set_task_cpu(idle, cpu);
5125 rcu_read_unlock();
5126
5127 rq->curr = rq->idle = idle;
5128#if defined(CONFIG_SMP)
5129 idle->on_cpu = 1;
5130#endif
5131 raw_spin_unlock_irqrestore(&rq->lock, flags);
5132
5133 /* Set the preempt count _outside_ the spinlocks! */
5134 task_thread_info(idle)->preempt_count = 0;
5135
5136 /*
5137 * The idle tasks have their own, simple scheduling class:
5138 */
5139 idle->sched_class = &idle_sched_class;
5140 ftrace_graph_init_idle_task(idle, cpu);
5141#if defined(CONFIG_SMP)
5142 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5143#endif
5144}
5145
5146#ifdef CONFIG_SMP
5147void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5148{
5149 if (p->sched_class && p->sched_class->set_cpus_allowed)
5150 p->sched_class->set_cpus_allowed(p, new_mask);
5151
5152 cpumask_copy(&p->cpus_allowed, new_mask);
5153 p->nr_cpus_allowed = cpumask_weight(new_mask);
5154}
5155
5156/*
5157 * This is how migration works:
5158 *
5159 * 1) we invoke migration_cpu_stop() on the target CPU using
5160 * stop_one_cpu().
5161 * 2) stopper starts to run (implicitly forcing the migrated thread
5162 * off the CPU)
5163 * 3) it checks whether the migrated task is still in the wrong runqueue.
5164 * 4) if it's in the wrong runqueue then the migration thread removes
5165 * it and puts it into the right queue.
5166 * 5) stopper completes and stop_one_cpu() returns and the migration
5167 * is done.
5168 */
5169
5170/*
5171 * Change a given task's CPU affinity. Migrate the thread to a
5172 * proper CPU and schedule it away if the CPU it's executing on
5173 * is removed from the allowed bitmask.
5174 *
5175 * NOTE: the caller must have a valid reference to the task, the
5176 * task must not exit() & deallocate itself prematurely. The
5177 * call is not atomic; no spinlocks may be held.
5178 */
5179int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5180{
5181 unsigned long flags;
5182 struct rq *rq;
5183 unsigned int dest_cpu;
5184 int ret = 0;
5185
5186 rq = task_rq_lock(p, &flags);
5187
5188 if (cpumask_equal(&p->cpus_allowed, new_mask))
5189 goto out;
5190
5191 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5192 ret = -EINVAL;
5193 goto out;
5194 }
5195
5196 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5197 ret = -EINVAL;
5198 goto out;
5199 }
5200
5201 do_set_cpus_allowed(p, new_mask);
5202
5203 /* Can the task run on the task's current CPU? If so, we're done */
5204 if (cpumask_test_cpu(task_cpu(p), new_mask))
5205 goto out;
5206
5207 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5208 if (p->on_rq) {
5209 struct migration_arg arg = { p, dest_cpu };
5210 /* Need help from migration thread: drop lock and wait. */
5211 task_rq_unlock(rq, p, &flags);
5212 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5213 tlb_migrate_finish(p->mm);
5214 return 0;
5215 }
5216out:
5217 task_rq_unlock(rq, p, &flags);
5218
5219 return ret;
5220}
5221EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5222
5223/*
5224 * Move (not current) task off this cpu, onto dest cpu. We're doing
5225 * this because either it can't run here any more (set_cpus_allowed()
5226 * away from this CPU, or CPU going down), or because we're
5227 * attempting to rebalance this task on exec (sched_exec).
5228 *
5229 * So we race with normal scheduler movements, but that's OK, as long
5230 * as the task is no longer on this CPU.
5231 *
5232 * Returns non-zero if task was successfully migrated.
5233 */
5234static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5235{
5236 struct rq *rq_dest, *rq_src;
5237 int ret = 0;
5238
5239 if (unlikely(!cpu_active(dest_cpu)))
5240 return ret;
5241
5242 rq_src = cpu_rq(src_cpu);
5243 rq_dest = cpu_rq(dest_cpu);
5244
5245 raw_spin_lock(&p->pi_lock);
5246 double_rq_lock(rq_src, rq_dest);
5247 /* Already moved. */
5248 if (task_cpu(p) != src_cpu)
5249 goto done;
5250 /* Affinity changed (again). */
5251 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5252 goto fail;
5253
5254 /*
5255 * If we're not on a rq, the next wake-up will ensure we're
5256 * placed properly.
5257 */
5258 if (p->on_rq) {
5259 dequeue_task(rq_src, p, 0);
5260 set_task_cpu(p, dest_cpu);
5261 enqueue_task(rq_dest, p, 0);
5262 check_preempt_curr(rq_dest, p, 0);
5263 }
5264done:
5265 ret = 1;
5266fail:
5267 double_rq_unlock(rq_src, rq_dest);
5268 raw_spin_unlock(&p->pi_lock);
5269 return ret;
5270}
5271
5272/*
5273 * migration_cpu_stop - this will be executed by a highprio stopper thread
5274 * and performs thread migration by bumping thread off CPU then
5275 * 'pushing' onto another runqueue.
5276 */
5277static int migration_cpu_stop(void *data)
5278{
5279 struct migration_arg *arg = data;
5280
5281 /*
5282 * The original target cpu might have gone down and we might
5283 * be on another cpu but it doesn't matter.
5284 */
5285 local_irq_disable();
5286 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5287 local_irq_enable();
5288 return 0;
5289}
5290
5291#ifdef CONFIG_HOTPLUG_CPU
5292
5293/*
5294 * Ensures that the idle task is using init_mm right before its cpu goes
5295 * offline.
5296 */
5297void idle_task_exit(void)
5298{
5299 struct mm_struct *mm = current->active_mm;
5300
5301 BUG_ON(cpu_online(smp_processor_id()));
5302
5303 if (mm != &init_mm)
5304 switch_mm(mm, &init_mm, current);
5305 mmdrop(mm);
5306}
5307
5308/*
5309 * While a dead CPU has no uninterruptible tasks queued at this point,
5310 * it might still have a nonzero ->nr_uninterruptible counter, because
5311 * for performance reasons the counter is not stricly tracking tasks to
5312 * their home CPUs. So we just add the counter to another CPU's counter,
5313 * to keep the global sum constant after CPU-down:
5314 */
5315static void migrate_nr_uninterruptible(struct rq *rq_src)
5316{
5317 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5318
5319 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5320 rq_src->nr_uninterruptible = 0;
5321}
5322
5323/*
5324 * remove the tasks which were accounted by rq from calc_load_tasks.
5325 */
5326static void calc_global_load_remove(struct rq *rq)
5327{
5328 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5329 rq->calc_load_active = 0;
5330}
5331
5332/*
5333 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5334 * try_to_wake_up()->select_task_rq().
5335 *
5336 * Called with rq->lock held even though we'er in stop_machine() and
5337 * there's no concurrency possible, we hold the required locks anyway
5338 * because of lock validation efforts.
5339 */
5340static void migrate_tasks(unsigned int dead_cpu)
5341{
5342 struct rq *rq = cpu_rq(dead_cpu);
5343 struct task_struct *next, *stop = rq->stop;
5344 int dest_cpu;
5345
5346 /*
5347 * Fudge the rq selection such that the below task selection loop
5348 * doesn't get stuck on the currently eligible stop task.
5349 *
5350 * We're currently inside stop_machine() and the rq is either stuck
5351 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5352 * either way we should never end up calling schedule() until we're
5353 * done here.
5354 */
5355 rq->stop = NULL;
5356
5357 /* Ensure any throttled groups are reachable by pick_next_task */
5358 unthrottle_offline_cfs_rqs(rq);
5359
5360 for ( ; ; ) {
5361 /*
5362 * There's this thread running, bail when that's the only
5363 * remaining thread.
5364 */
5365 if (rq->nr_running == 1)
5366 break;
5367
5368 next = pick_next_task(rq);
5369 BUG_ON(!next);
5370 next->sched_class->put_prev_task(rq, next);
5371
5372 /* Find suitable destination for @next, with force if needed. */
5373 dest_cpu = select_fallback_rq(dead_cpu, next);
5374 raw_spin_unlock(&rq->lock);
5375
5376 __migrate_task(next, dead_cpu, dest_cpu);
5377
5378 raw_spin_lock(&rq->lock);
5379 }
5380
5381 rq->stop = stop;
5382}
5383
5384#endif /* CONFIG_HOTPLUG_CPU */
5385
5386#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5387
5388static struct ctl_table sd_ctl_dir[] = {
5389 {
5390 .procname = "sched_domain",
5391 .mode = 0555,
5392 },
5393 {}
5394};
5395
5396static struct ctl_table sd_ctl_root[] = {
5397 {
5398 .procname = "kernel",
5399 .mode = 0555,
5400 .child = sd_ctl_dir,
5401 },
5402 {}
5403};
5404
5405static struct ctl_table *sd_alloc_ctl_entry(int n)
5406{
5407 struct ctl_table *entry =
5408 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5409
5410 return entry;
5411}
5412
5413static void sd_free_ctl_entry(struct ctl_table **tablep)
5414{
5415 struct ctl_table *entry;
5416
5417 /*
5418 * In the intermediate directories, both the child directory and
5419 * procname are dynamically allocated and could fail but the mode
5420 * will always be set. In the lowest directory the names are
5421 * static strings and all have proc handlers.
5422 */
5423 for (entry = *tablep; entry->mode; entry++) {
5424 if (entry->child)
5425 sd_free_ctl_entry(&entry->child);
5426 if (entry->proc_handler == NULL)
5427 kfree(entry->procname);
5428 }
5429
5430 kfree(*tablep);
5431 *tablep = NULL;
5432}
5433
5434static void
5435set_table_entry(struct ctl_table *entry,
5436 const char *procname, void *data, int maxlen,
5437 umode_t mode, proc_handler *proc_handler)
5438{
5439 entry->procname = procname;
5440 entry->data = data;
5441 entry->maxlen = maxlen;
5442 entry->mode = mode;
5443 entry->proc_handler = proc_handler;
5444}
5445
5446static struct ctl_table *
5447sd_alloc_ctl_domain_table(struct sched_domain *sd)
5448{
5449 struct ctl_table *table = sd_alloc_ctl_entry(13);
5450
5451 if (table == NULL)
5452 return NULL;
5453
5454 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5455 sizeof(long), 0644, proc_doulongvec_minmax);
5456 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5457 sizeof(long), 0644, proc_doulongvec_minmax);
5458 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5459 sizeof(int), 0644, proc_dointvec_minmax);
5460 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5461 sizeof(int), 0644, proc_dointvec_minmax);
5462 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5463 sizeof(int), 0644, proc_dointvec_minmax);
5464 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5465 sizeof(int), 0644, proc_dointvec_minmax);
5466 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5467 sizeof(int), 0644, proc_dointvec_minmax);
5468 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5471 sizeof(int), 0644, proc_dointvec_minmax);
5472 set_table_entry(&table[9], "cache_nice_tries",
5473 &sd->cache_nice_tries,
5474 sizeof(int), 0644, proc_dointvec_minmax);
5475 set_table_entry(&table[10], "flags", &sd->flags,
5476 sizeof(int), 0644, proc_dointvec_minmax);
5477 set_table_entry(&table[11], "name", sd->name,
5478 CORENAME_MAX_SIZE, 0444, proc_dostring);
5479 /* &table[12] is terminator */
5480
5481 return table;
5482}
5483
5484static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5485{
5486 struct ctl_table *entry, *table;
5487 struct sched_domain *sd;
5488 int domain_num = 0, i;
5489 char buf[32];
5490
5491 for_each_domain(cpu, sd)
5492 domain_num++;
5493 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5494 if (table == NULL)
5495 return NULL;
5496
5497 i = 0;
5498 for_each_domain(cpu, sd) {
5499 snprintf(buf, 32, "domain%d", i);
5500 entry->procname = kstrdup(buf, GFP_KERNEL);
5501 entry->mode = 0555;
5502 entry->child = sd_alloc_ctl_domain_table(sd);
5503 entry++;
5504 i++;
5505 }
5506 return table;
5507}
5508
5509static struct ctl_table_header *sd_sysctl_header;
5510static void register_sched_domain_sysctl(void)
5511{
5512 int i, cpu_num = num_possible_cpus();
5513 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5514 char buf[32];
5515
5516 WARN_ON(sd_ctl_dir[0].child);
5517 sd_ctl_dir[0].child = entry;
5518
5519 if (entry == NULL)
5520 return;
5521
5522 for_each_possible_cpu(i) {
5523 snprintf(buf, 32, "cpu%d", i);
5524 entry->procname = kstrdup(buf, GFP_KERNEL);
5525 entry->mode = 0555;
5526 entry->child = sd_alloc_ctl_cpu_table(i);
5527 entry++;
5528 }
5529
5530 WARN_ON(sd_sysctl_header);
5531 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5532}
5533
5534/* may be called multiple times per register */
5535static void unregister_sched_domain_sysctl(void)
5536{
5537 if (sd_sysctl_header)
5538 unregister_sysctl_table(sd_sysctl_header);
5539 sd_sysctl_header = NULL;
5540 if (sd_ctl_dir[0].child)
5541 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5542}
5543#else
5544static void register_sched_domain_sysctl(void)
5545{
5546}
5547static void unregister_sched_domain_sysctl(void)
5548{
5549}
5550#endif
5551
5552static void set_rq_online(struct rq *rq)
5553{
5554 if (!rq->online) {
5555 const struct sched_class *class;
5556
5557 cpumask_set_cpu(rq->cpu, rq->rd->online);
5558 rq->online = 1;
5559
5560 for_each_class(class) {
5561 if (class->rq_online)
5562 class->rq_online(rq);
5563 }
5564 }
5565}
5566
5567static void set_rq_offline(struct rq *rq)
5568{
5569 if (rq->online) {
5570 const struct sched_class *class;
5571
5572 for_each_class(class) {
5573 if (class->rq_offline)
5574 class->rq_offline(rq);
5575 }
5576
5577 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5578 rq->online = 0;
5579 }
5580}
5581
5582/*
5583 * migration_call - callback that gets triggered when a CPU is added.
5584 * Here we can start up the necessary migration thread for the new CPU.
5585 */
5586static int __cpuinit
5587migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5588{
5589 int cpu = (long)hcpu;
5590 unsigned long flags;
5591 struct rq *rq = cpu_rq(cpu);
5592
5593 switch (action & ~CPU_TASKS_FROZEN) {
5594
5595 case CPU_UP_PREPARE:
5596 rq->calc_load_update = calc_load_update;
5597 break;
5598
5599 case CPU_ONLINE:
5600 /* Update our root-domain */
5601 raw_spin_lock_irqsave(&rq->lock, flags);
5602 if (rq->rd) {
5603 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5604
5605 set_rq_online(rq);
5606 }
5607 raw_spin_unlock_irqrestore(&rq->lock, flags);
5608 break;
5609
5610#ifdef CONFIG_HOTPLUG_CPU
5611 case CPU_DYING:
5612 sched_ttwu_pending();
5613 /* Update our root-domain */
5614 raw_spin_lock_irqsave(&rq->lock, flags);
5615 if (rq->rd) {
5616 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5617 set_rq_offline(rq);
5618 }
5619 migrate_tasks(cpu);
5620 BUG_ON(rq->nr_running != 1); /* the migration thread */
5621 raw_spin_unlock_irqrestore(&rq->lock, flags);
5622
5623 migrate_nr_uninterruptible(rq);
5624 calc_global_load_remove(rq);
5625 break;
5626#endif
5627 }
5628
5629 update_max_interval();
5630
5631 return NOTIFY_OK;
5632}
5633
5634/*
5635 * Register at high priority so that task migration (migrate_all_tasks)
5636 * happens before everything else. This has to be lower priority than
5637 * the notifier in the perf_event subsystem, though.
5638 */
5639static struct notifier_block __cpuinitdata migration_notifier = {
5640 .notifier_call = migration_call,
5641 .priority = CPU_PRI_MIGRATION,
5642};
5643
5644static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5645 unsigned long action, void *hcpu)
5646{
5647 switch (action & ~CPU_TASKS_FROZEN) {
5648 case CPU_STARTING:
5649 case CPU_DOWN_FAILED:
5650 set_cpu_active((long)hcpu, true);
5651 return NOTIFY_OK;
5652 default:
5653 return NOTIFY_DONE;
5654 }
5655}
5656
5657static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5658 unsigned long action, void *hcpu)
5659{
5660 switch (action & ~CPU_TASKS_FROZEN) {
5661 case CPU_DOWN_PREPARE:
5662 set_cpu_active((long)hcpu, false);
5663 return NOTIFY_OK;
5664 default:
5665 return NOTIFY_DONE;
5666 }
5667}
5668
5669static int __init migration_init(void)
5670{
5671 void *cpu = (void *)(long)smp_processor_id();
5672 int err;
5673
5674 /* Initialize migration for the boot CPU */
5675 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5676 BUG_ON(err == NOTIFY_BAD);
5677 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5678 register_cpu_notifier(&migration_notifier);
5679
5680 /* Register cpu active notifiers */
5681 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5682 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5683
5684 return 0;
5685}
5686early_initcall(migration_init);
5687#endif
5688
5689#ifdef CONFIG_SMP
5690
5691static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5692
5693#ifdef CONFIG_SCHED_DEBUG
5694
5695static __read_mostly int sched_debug_enabled;
5696
5697static int __init sched_debug_setup(char *str)
5698{
5699 sched_debug_enabled = 1;
5700
5701 return 0;
5702}
5703early_param("sched_debug", sched_debug_setup);
5704
5705static inline bool sched_debug(void)
5706{
5707 return sched_debug_enabled;
5708}
5709
5710static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5711 struct cpumask *groupmask)
5712{
5713 struct sched_group *group = sd->groups;
5714 char str[256];
5715
5716 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5717 cpumask_clear(groupmask);
5718
5719 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5720
5721 if (!(sd->flags & SD_LOAD_BALANCE)) {
5722 printk("does not load-balance\n");
5723 if (sd->parent)
5724 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5725 " has parent");
5726 return -1;
5727 }
5728
5729 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5730
5731 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5732 printk(KERN_ERR "ERROR: domain->span does not contain "
5733 "CPU%d\n", cpu);
5734 }
5735 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5736 printk(KERN_ERR "ERROR: domain->groups does not contain"
5737 " CPU%d\n", cpu);
5738 }
5739
5740 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5741 do {
5742 if (!group) {
5743 printk("\n");
5744 printk(KERN_ERR "ERROR: group is NULL\n");
5745 break;
5746 }
5747
5748 /*
5749 * Even though we initialize ->power to something semi-sane,
5750 * we leave power_orig unset. This allows us to detect if
5751 * domain iteration is still funny without causing /0 traps.
5752 */
5753 if (!group->sgp->power_orig) {
5754 printk(KERN_CONT "\n");
5755 printk(KERN_ERR "ERROR: domain->cpu_power not "
5756 "set\n");
5757 break;
5758 }
5759
5760 if (!cpumask_weight(sched_group_cpus(group))) {
5761 printk(KERN_CONT "\n");
5762 printk(KERN_ERR "ERROR: empty group\n");
5763 break;
5764 }
5765
5766 if (!(sd->flags & SD_OVERLAP) &&
5767 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5768 printk(KERN_CONT "\n");
5769 printk(KERN_ERR "ERROR: repeated CPUs\n");
5770 break;
5771 }
5772
5773 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5774
5775 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5776
5777 printk(KERN_CONT " %s", str);
5778 if (group->sgp->power != SCHED_POWER_SCALE) {
5779 printk(KERN_CONT " (cpu_power = %d)",
5780 group->sgp->power);
5781 }
5782
5783 group = group->next;
5784 } while (group != sd->groups);
5785 printk(KERN_CONT "\n");
5786
5787 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5788 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5789
5790 if (sd->parent &&
5791 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5792 printk(KERN_ERR "ERROR: parent span is not a superset "
5793 "of domain->span\n");
5794 return 0;
5795}
5796
5797static void sched_domain_debug(struct sched_domain *sd, int cpu)
5798{
5799 int level = 0;
5800
5801 if (!sched_debug_enabled)
5802 return;
5803
5804 if (!sd) {
5805 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5806 return;
5807 }
5808
5809 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5810
5811 for (;;) {
5812 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5813 break;
5814 level++;
5815 sd = sd->parent;
5816 if (!sd)
5817 break;
5818 }
5819}
5820#else /* !CONFIG_SCHED_DEBUG */
5821# define sched_domain_debug(sd, cpu) do { } while (0)
5822static inline bool sched_debug(void)
5823{
5824 return false;
5825}
5826#endif /* CONFIG_SCHED_DEBUG */
5827
5828static int sd_degenerate(struct sched_domain *sd)
5829{
5830 if (cpumask_weight(sched_domain_span(sd)) == 1)
5831 return 1;
5832
5833 /* Following flags need at least 2 groups */
5834 if (sd->flags & (SD_LOAD_BALANCE |
5835 SD_BALANCE_NEWIDLE |
5836 SD_BALANCE_FORK |
5837 SD_BALANCE_EXEC |
5838 SD_SHARE_CPUPOWER |
5839 SD_SHARE_PKG_RESOURCES)) {
5840 if (sd->groups != sd->groups->next)
5841 return 0;
5842 }
5843
5844 /* Following flags don't use groups */
5845 if (sd->flags & (SD_WAKE_AFFINE))
5846 return 0;
5847
5848 return 1;
5849}
5850
5851static int
5852sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5853{
5854 unsigned long cflags = sd->flags, pflags = parent->flags;
5855
5856 if (sd_degenerate(parent))
5857 return 1;
5858
5859 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5860 return 0;
5861
5862 /* Flags needing groups don't count if only 1 group in parent */
5863 if (parent->groups == parent->groups->next) {
5864 pflags &= ~(SD_LOAD_BALANCE |
5865 SD_BALANCE_NEWIDLE |
5866 SD_BALANCE_FORK |
5867 SD_BALANCE_EXEC |
5868 SD_SHARE_CPUPOWER |
5869 SD_SHARE_PKG_RESOURCES);
5870 if (nr_node_ids == 1)
5871 pflags &= ~SD_SERIALIZE;
5872 }
5873 if (~cflags & pflags)
5874 return 0;
5875
5876 return 1;
5877}
5878
5879static void free_rootdomain(struct rcu_head *rcu)
5880{
5881 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5882
5883 cpupri_cleanup(&rd->cpupri);
5884 free_cpumask_var(rd->rto_mask);
5885 free_cpumask_var(rd->online);
5886 free_cpumask_var(rd->span);
5887 kfree(rd);
5888}
5889
5890static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5891{
5892 struct root_domain *old_rd = NULL;
5893 unsigned long flags;
5894
5895 raw_spin_lock_irqsave(&rq->lock, flags);
5896
5897 if (rq->rd) {
5898 old_rd = rq->rd;
5899
5900 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5901 set_rq_offline(rq);
5902
5903 cpumask_clear_cpu(rq->cpu, old_rd->span);
5904
5905 /*
5906 * If we dont want to free the old_rt yet then
5907 * set old_rd to NULL to skip the freeing later
5908 * in this function:
5909 */
5910 if (!atomic_dec_and_test(&old_rd->refcount))
5911 old_rd = NULL;
5912 }
5913
5914 atomic_inc(&rd->refcount);
5915 rq->rd = rd;
5916
5917 cpumask_set_cpu(rq->cpu, rd->span);
5918 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5919 set_rq_online(rq);
5920
5921 raw_spin_unlock_irqrestore(&rq->lock, flags);
5922
5923 if (old_rd)
5924 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5925}
5926
5927static int init_rootdomain(struct root_domain *rd)
5928{
5929 memset(rd, 0, sizeof(*rd));
5930
5931 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5932 goto out;
5933 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5934 goto free_span;
5935 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5936 goto free_online;
5937
5938 if (cpupri_init(&rd->cpupri) != 0)
5939 goto free_rto_mask;
5940 return 0;
5941
5942free_rto_mask:
5943 free_cpumask_var(rd->rto_mask);
5944free_online:
5945 free_cpumask_var(rd->online);
5946free_span:
5947 free_cpumask_var(rd->span);
5948out:
5949 return -ENOMEM;
5950}
5951
5952/*
5953 * By default the system creates a single root-domain with all cpus as
5954 * members (mimicking the global state we have today).
5955 */
5956struct root_domain def_root_domain;
5957
5958static void init_defrootdomain(void)
5959{
5960 init_rootdomain(&def_root_domain);
5961
5962 atomic_set(&def_root_domain.refcount, 1);
5963}
5964
5965static struct root_domain *alloc_rootdomain(void)
5966{
5967 struct root_domain *rd;
5968
5969 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5970 if (!rd)
5971 return NULL;
5972
5973 if (init_rootdomain(rd) != 0) {
5974 kfree(rd);
5975 return NULL;
5976 }
5977
5978 return rd;
5979}
5980
5981static void free_sched_groups(struct sched_group *sg, int free_sgp)
5982{
5983 struct sched_group *tmp, *first;
5984
5985 if (!sg)
5986 return;
5987
5988 first = sg;
5989 do {
5990 tmp = sg->next;
5991
5992 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5993 kfree(sg->sgp);
5994
5995 kfree(sg);
5996 sg = tmp;
5997 } while (sg != first);
5998}
5999
6000static void free_sched_domain(struct rcu_head *rcu)
6001{
6002 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6003
6004 /*
6005 * If its an overlapping domain it has private groups, iterate and
6006 * nuke them all.
6007 */
6008 if (sd->flags & SD_OVERLAP) {
6009 free_sched_groups(sd->groups, 1);
6010 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6011 kfree(sd->groups->sgp);
6012 kfree(sd->groups);
6013 }
6014 kfree(sd);
6015}
6016
6017static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6018{
6019 call_rcu(&sd->rcu, free_sched_domain);
6020}
6021
6022static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6023{
6024 for (; sd; sd = sd->parent)
6025 destroy_sched_domain(sd, cpu);
6026}
6027
6028/*
6029 * Keep a special pointer to the highest sched_domain that has
6030 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6031 * allows us to avoid some pointer chasing select_idle_sibling().
6032 *
6033 * Also keep a unique ID per domain (we use the first cpu number in
6034 * the cpumask of the domain), this allows us to quickly tell if
6035 * two cpus are in the same cache domain, see cpus_share_cache().
6036 */
6037DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6038DEFINE_PER_CPU(int, sd_llc_id);
6039
6040static void update_top_cache_domain(int cpu)
6041{
6042 struct sched_domain *sd;
6043 int id = cpu;
6044
6045 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6046 if (sd)
6047 id = cpumask_first(sched_domain_span(sd));
6048
6049 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6050 per_cpu(sd_llc_id, cpu) = id;
6051}
6052
6053/*
6054 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6055 * hold the hotplug lock.
6056 */
6057static void
6058cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6059{
6060 struct rq *rq = cpu_rq(cpu);
6061 struct sched_domain *tmp;
6062
6063 /* Remove the sched domains which do not contribute to scheduling. */
6064 for (tmp = sd; tmp; ) {
6065 struct sched_domain *parent = tmp->parent;
6066 if (!parent)
6067 break;
6068
6069 if (sd_parent_degenerate(tmp, parent)) {
6070 tmp->parent = parent->parent;
6071 if (parent->parent)
6072 parent->parent->child = tmp;
6073 destroy_sched_domain(parent, cpu);
6074 } else
6075 tmp = tmp->parent;
6076 }
6077
6078 if (sd && sd_degenerate(sd)) {
6079 tmp = sd;
6080 sd = sd->parent;
6081 destroy_sched_domain(tmp, cpu);
6082 if (sd)
6083 sd->child = NULL;
6084 }
6085
6086 sched_domain_debug(sd, cpu);
6087
6088 rq_attach_root(rq, rd);
6089 tmp = rq->sd;
6090 rcu_assign_pointer(rq->sd, sd);
6091 destroy_sched_domains(tmp, cpu);
6092
6093 update_top_cache_domain(cpu);
6094}
6095
6096/* cpus with isolated domains */
6097static cpumask_var_t cpu_isolated_map;
6098
6099/* Setup the mask of cpus configured for isolated domains */
6100static int __init isolated_cpu_setup(char *str)
6101{
6102 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6103 cpulist_parse(str, cpu_isolated_map);
6104 return 1;
6105}
6106
6107__setup("isolcpus=", isolated_cpu_setup);
6108
6109static const struct cpumask *cpu_cpu_mask(int cpu)
6110{
6111 return cpumask_of_node(cpu_to_node(cpu));
6112}
6113
6114struct sd_data {
6115 struct sched_domain **__percpu sd;
6116 struct sched_group **__percpu sg;
6117 struct sched_group_power **__percpu sgp;
6118};
6119
6120struct s_data {
6121 struct sched_domain ** __percpu sd;
6122 struct root_domain *rd;
6123};
6124
6125enum s_alloc {
6126 sa_rootdomain,
6127 sa_sd,
6128 sa_sd_storage,
6129 sa_none,
6130};
6131
6132struct sched_domain_topology_level;
6133
6134typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6135typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6136
6137#define SDTL_OVERLAP 0x01
6138
6139struct sched_domain_topology_level {
6140 sched_domain_init_f init;
6141 sched_domain_mask_f mask;
6142 int flags;
6143 int numa_level;
6144 struct sd_data data;
6145};
6146
6147/*
6148 * Build an iteration mask that can exclude certain CPUs from the upwards
6149 * domain traversal.
6150 *
6151 * Asymmetric node setups can result in situations where the domain tree is of
6152 * unequal depth, make sure to skip domains that already cover the entire
6153 * range.
6154 *
6155 * In that case build_sched_domains() will have terminated the iteration early
6156 * and our sibling sd spans will be empty. Domains should always include the
6157 * cpu they're built on, so check that.
6158 *
6159 */
6160static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6161{
6162 const struct cpumask *span = sched_domain_span(sd);
6163 struct sd_data *sdd = sd->private;
6164 struct sched_domain *sibling;
6165 int i;
6166
6167 for_each_cpu(i, span) {
6168 sibling = *per_cpu_ptr(sdd->sd, i);
6169 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6170 continue;
6171
6172 cpumask_set_cpu(i, sched_group_mask(sg));
6173 }
6174}
6175
6176/*
6177 * Return the canonical balance cpu for this group, this is the first cpu
6178 * of this group that's also in the iteration mask.
6179 */
6180int group_balance_cpu(struct sched_group *sg)
6181{
6182 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6183}
6184
6185static int
6186build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6187{
6188 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6189 const struct cpumask *span = sched_domain_span(sd);
6190 struct cpumask *covered = sched_domains_tmpmask;
6191 struct sd_data *sdd = sd->private;
6192 struct sched_domain *child;
6193 int i;
6194
6195 cpumask_clear(covered);
6196
6197 for_each_cpu(i, span) {
6198 struct cpumask *sg_span;
6199
6200 if (cpumask_test_cpu(i, covered))
6201 continue;
6202
6203 child = *per_cpu_ptr(sdd->sd, i);
6204
6205 /* See the comment near build_group_mask(). */
6206 if (!cpumask_test_cpu(i, sched_domain_span(child)))
6207 continue;
6208
6209 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6210 GFP_KERNEL, cpu_to_node(cpu));
6211
6212 if (!sg)
6213 goto fail;
6214
6215 sg_span = sched_group_cpus(sg);
6216 if (child->child) {
6217 child = child->child;
6218 cpumask_copy(sg_span, sched_domain_span(child));
6219 } else
6220 cpumask_set_cpu(i, sg_span);
6221
6222 cpumask_or(covered, covered, sg_span);
6223
6224 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
6225 if (atomic_inc_return(&sg->sgp->ref) == 1)
6226 build_group_mask(sd, sg);
6227
6228 /*
6229 * Initialize sgp->power such that even if we mess up the
6230 * domains and no possible iteration will get us here, we won't
6231 * die on a /0 trap.
6232 */
6233 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
6234
6235 /*
6236 * Make sure the first group of this domain contains the
6237 * canonical balance cpu. Otherwise the sched_domain iteration
6238 * breaks. See update_sg_lb_stats().
6239 */
6240 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6241 group_balance_cpu(sg) == cpu)
6242 groups = sg;
6243
6244 if (!first)
6245 first = sg;
6246 if (last)
6247 last->next = sg;
6248 last = sg;
6249 last->next = first;
6250 }
6251 sd->groups = groups;
6252
6253 return 0;
6254
6255fail:
6256 free_sched_groups(first, 0);
6257
6258 return -ENOMEM;
6259}
6260
6261static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6262{
6263 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6264 struct sched_domain *child = sd->child;
6265
6266 if (child)
6267 cpu = cpumask_first(sched_domain_span(child));
6268
6269 if (sg) {
6270 *sg = *per_cpu_ptr(sdd->sg, cpu);
6271 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6272 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6273 }
6274
6275 return cpu;
6276}
6277
6278/*
6279 * build_sched_groups will build a circular linked list of the groups
6280 * covered by the given span, and will set each group's ->cpumask correctly,
6281 * and ->cpu_power to 0.
6282 *
6283 * Assumes the sched_domain tree is fully constructed
6284 */
6285static int
6286build_sched_groups(struct sched_domain *sd, int cpu)
6287{
6288 struct sched_group *first = NULL, *last = NULL;
6289 struct sd_data *sdd = sd->private;
6290 const struct cpumask *span = sched_domain_span(sd);
6291 struct cpumask *covered;
6292 int i;
6293
6294 get_group(cpu, sdd, &sd->groups);
6295 atomic_inc(&sd->groups->ref);
6296
6297 if (cpu != cpumask_first(sched_domain_span(sd)))
6298 return 0;
6299
6300 lockdep_assert_held(&sched_domains_mutex);
6301 covered = sched_domains_tmpmask;
6302
6303 cpumask_clear(covered);
6304
6305 for_each_cpu(i, span) {
6306 struct sched_group *sg;
6307 int group = get_group(i, sdd, &sg);
6308 int j;
6309
6310 if (cpumask_test_cpu(i, covered))
6311 continue;
6312
6313 cpumask_clear(sched_group_cpus(sg));
6314 sg->sgp->power = 0;
6315 cpumask_setall(sched_group_mask(sg));
6316
6317 for_each_cpu(j, span) {
6318 if (get_group(j, sdd, NULL) != group)
6319 continue;
6320
6321 cpumask_set_cpu(j, covered);
6322 cpumask_set_cpu(j, sched_group_cpus(sg));
6323 }
6324
6325 if (!first)
6326 first = sg;
6327 if (last)
6328 last->next = sg;
6329 last = sg;
6330 }
6331 last->next = first;
6332
6333 return 0;
6334}
6335
6336/*
6337 * Initialize sched groups cpu_power.
6338 *
6339 * cpu_power indicates the capacity of sched group, which is used while
6340 * distributing the load between different sched groups in a sched domain.
6341 * Typically cpu_power for all the groups in a sched domain will be same unless
6342 * there are asymmetries in the topology. If there are asymmetries, group
6343 * having more cpu_power will pickup more load compared to the group having
6344 * less cpu_power.
6345 */
6346static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6347{
6348 struct sched_group *sg = sd->groups;
6349
6350 WARN_ON(!sd || !sg);
6351
6352 do {
6353 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6354 sg = sg->next;
6355 } while (sg != sd->groups);
6356
6357 if (cpu != group_balance_cpu(sg))
6358 return;
6359
6360 update_group_power(sd, cpu);
6361 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6362}
6363
6364int __weak arch_sd_sibling_asym_packing(void)
6365{
6366 return 0*SD_ASYM_PACKING;
6367}
6368
6369/*
6370 * Initializers for schedule domains
6371 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6372 */
6373
6374#ifdef CONFIG_SCHED_DEBUG
6375# define SD_INIT_NAME(sd, type) sd->name = #type
6376#else
6377# define SD_INIT_NAME(sd, type) do { } while (0)
6378#endif
6379
6380#define SD_INIT_FUNC(type) \
6381static noinline struct sched_domain * \
6382sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6383{ \
6384 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6385 *sd = SD_##type##_INIT; \
6386 SD_INIT_NAME(sd, type); \
6387 sd->private = &tl->data; \
6388 return sd; \
6389}
6390
6391SD_INIT_FUNC(CPU)
6392#ifdef CONFIG_SCHED_SMT
6393 SD_INIT_FUNC(SIBLING)
6394#endif
6395#ifdef CONFIG_SCHED_MC
6396 SD_INIT_FUNC(MC)
6397#endif
6398#ifdef CONFIG_SCHED_BOOK
6399 SD_INIT_FUNC(BOOK)
6400#endif
6401
6402static int default_relax_domain_level = -1;
6403int sched_domain_level_max;
6404
6405static int __init setup_relax_domain_level(char *str)
6406{
6407 if (kstrtoint(str, 0, &default_relax_domain_level))
6408 pr_warn("Unable to set relax_domain_level\n");
6409
6410 return 1;
6411}
6412__setup("relax_domain_level=", setup_relax_domain_level);
6413
6414static void set_domain_attribute(struct sched_domain *sd,
6415 struct sched_domain_attr *attr)
6416{
6417 int request;
6418
6419 if (!attr || attr->relax_domain_level < 0) {
6420 if (default_relax_domain_level < 0)
6421 return;
6422 else
6423 request = default_relax_domain_level;
6424 } else
6425 request = attr->relax_domain_level;
6426 if (request < sd->level) {
6427 /* turn off idle balance on this domain */
6428 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6429 } else {
6430 /* turn on idle balance on this domain */
6431 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6432 }
6433}
6434
6435static void __sdt_free(const struct cpumask *cpu_map);
6436static int __sdt_alloc(const struct cpumask *cpu_map);
6437
6438static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6439 const struct cpumask *cpu_map)
6440{
6441 switch (what) {
6442 case sa_rootdomain:
6443 if (!atomic_read(&d->rd->refcount))
6444 free_rootdomain(&d->rd->rcu); /* fall through */
6445 case sa_sd:
6446 free_percpu(d->sd); /* fall through */
6447 case sa_sd_storage:
6448 __sdt_free(cpu_map); /* fall through */
6449 case sa_none:
6450 break;
6451 }
6452}
6453
6454static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6455 const struct cpumask *cpu_map)
6456{
6457 memset(d, 0, sizeof(*d));
6458
6459 if (__sdt_alloc(cpu_map))
6460 return sa_sd_storage;
6461 d->sd = alloc_percpu(struct sched_domain *);
6462 if (!d->sd)
6463 return sa_sd_storage;
6464 d->rd = alloc_rootdomain();
6465 if (!d->rd)
6466 return sa_sd;
6467 return sa_rootdomain;
6468}
6469
6470/*
6471 * NULL the sd_data elements we've used to build the sched_domain and
6472 * sched_group structure so that the subsequent __free_domain_allocs()
6473 * will not free the data we're using.
6474 */
6475static void claim_allocations(int cpu, struct sched_domain *sd)
6476{
6477 struct sd_data *sdd = sd->private;
6478
6479 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6480 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6481
6482 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6483 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6484
6485 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6486 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6487}
6488
6489#ifdef CONFIG_SCHED_SMT
6490static const struct cpumask *cpu_smt_mask(int cpu)
6491{
6492 return topology_thread_cpumask(cpu);
6493}
6494#endif
6495
6496/*
6497 * Topology list, bottom-up.
6498 */
6499static struct sched_domain_topology_level default_topology[] = {
6500#ifdef CONFIG_SCHED_SMT
6501 { sd_init_SIBLING, cpu_smt_mask, },
6502#endif
6503#ifdef CONFIG_SCHED_MC
6504 { sd_init_MC, cpu_coregroup_mask, },
6505#endif
6506#ifdef CONFIG_SCHED_BOOK
6507 { sd_init_BOOK, cpu_book_mask, },
6508#endif
6509 { sd_init_CPU, cpu_cpu_mask, },
6510 { NULL, },
6511};
6512
6513static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6514
6515#ifdef CONFIG_NUMA
6516
6517static int sched_domains_numa_levels;
6518static int *sched_domains_numa_distance;
6519static struct cpumask ***sched_domains_numa_masks;
6520static int sched_domains_curr_level;
6521
6522static inline int sd_local_flags(int level)
6523{
6524 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6525 return 0;
6526
6527 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6528}
6529
6530static struct sched_domain *
6531sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6532{
6533 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6534 int level = tl->numa_level;
6535 int sd_weight = cpumask_weight(
6536 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6537
6538 *sd = (struct sched_domain){
6539 .min_interval = sd_weight,
6540 .max_interval = 2*sd_weight,
6541 .busy_factor = 32,
6542 .imbalance_pct = 125,
6543 .cache_nice_tries = 2,
6544 .busy_idx = 3,
6545 .idle_idx = 2,
6546 .newidle_idx = 0,
6547 .wake_idx = 0,
6548 .forkexec_idx = 0,
6549
6550 .flags = 1*SD_LOAD_BALANCE
6551 | 1*SD_BALANCE_NEWIDLE
6552 | 0*SD_BALANCE_EXEC
6553 | 0*SD_BALANCE_FORK
6554 | 0*SD_BALANCE_WAKE
6555 | 0*SD_WAKE_AFFINE
6556 | 0*SD_PREFER_LOCAL
6557 | 0*SD_SHARE_CPUPOWER
6558 | 0*SD_SHARE_PKG_RESOURCES
6559 | 1*SD_SERIALIZE
6560 | 0*SD_PREFER_SIBLING
6561 | sd_local_flags(level)
6562 ,
6563 .last_balance = jiffies,
6564 .balance_interval = sd_weight,
6565 };
6566 SD_INIT_NAME(sd, NUMA);
6567 sd->private = &tl->data;
6568
6569 /*
6570 * Ugly hack to pass state to sd_numa_mask()...
6571 */
6572 sched_domains_curr_level = tl->numa_level;
6573
6574 return sd;
6575}
6576
6577static const struct cpumask *sd_numa_mask(int cpu)
6578{
6579 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6580}
6581
6582static void sched_numa_warn(const char *str)
6583{
6584 static int done = false;
6585 int i,j;
6586
6587 if (done)
6588 return;
6589
6590 done = true;
6591
6592 printk(KERN_WARNING "ERROR: %s\n\n", str);
6593
6594 for (i = 0; i < nr_node_ids; i++) {
6595 printk(KERN_WARNING " ");
6596 for (j = 0; j < nr_node_ids; j++)
6597 printk(KERN_CONT "%02d ", node_distance(i,j));
6598 printk(KERN_CONT "\n");
6599 }
6600 printk(KERN_WARNING "\n");
6601}
6602
6603static bool find_numa_distance(int distance)
6604{
6605 int i;
6606
6607 if (distance == node_distance(0, 0))
6608 return true;
6609
6610 for (i = 0; i < sched_domains_numa_levels; i++) {
6611 if (sched_domains_numa_distance[i] == distance)
6612 return true;
6613 }
6614
6615 return false;
6616}
6617
6618static void sched_init_numa(void)
6619{
6620 int next_distance, curr_distance = node_distance(0, 0);
6621 struct sched_domain_topology_level *tl;
6622 int level = 0;
6623 int i, j, k;
6624
6625 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6626 if (!sched_domains_numa_distance)
6627 return;
6628
6629 /*
6630 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6631 * unique distances in the node_distance() table.
6632 *
6633 * Assumes node_distance(0,j) includes all distances in
6634 * node_distance(i,j) in order to avoid cubic time.
6635 */
6636 next_distance = curr_distance;
6637 for (i = 0; i < nr_node_ids; i++) {
6638 for (j = 0; j < nr_node_ids; j++) {
6639 for (k = 0; k < nr_node_ids; k++) {
6640 int distance = node_distance(i, k);
6641
6642 if (distance > curr_distance &&
6643 (distance < next_distance ||
6644 next_distance == curr_distance))
6645 next_distance = distance;
6646
6647 /*
6648 * While not a strong assumption it would be nice to know
6649 * about cases where if node A is connected to B, B is not
6650 * equally connected to A.
6651 */
6652 if (sched_debug() && node_distance(k, i) != distance)
6653 sched_numa_warn("Node-distance not symmetric");
6654
6655 if (sched_debug() && i && !find_numa_distance(distance))
6656 sched_numa_warn("Node-0 not representative");
6657 }
6658 if (next_distance != curr_distance) {
6659 sched_domains_numa_distance[level++] = next_distance;
6660 sched_domains_numa_levels = level;
6661 curr_distance = next_distance;
6662 } else break;
6663 }
6664
6665 /*
6666 * In case of sched_debug() we verify the above assumption.
6667 */
6668 if (!sched_debug())
6669 break;
6670 }
6671 /*
6672 * 'level' contains the number of unique distances, excluding the
6673 * identity distance node_distance(i,i).
6674 *
6675 * The sched_domains_nume_distance[] array includes the actual distance
6676 * numbers.
6677 */
6678
6679 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6680 if (!sched_domains_numa_masks)
6681 return;
6682
6683 /*
6684 * Now for each level, construct a mask per node which contains all
6685 * cpus of nodes that are that many hops away from us.
6686 */
6687 for (i = 0; i < level; i++) {
6688 sched_domains_numa_masks[i] =
6689 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6690 if (!sched_domains_numa_masks[i])
6691 return;
6692
6693 for (j = 0; j < nr_node_ids; j++) {
6694 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6695 if (!mask)
6696 return;
6697
6698 sched_domains_numa_masks[i][j] = mask;
6699
6700 for (k = 0; k < nr_node_ids; k++) {
6701 if (node_distance(j, k) > sched_domains_numa_distance[i])
6702 continue;
6703
6704 cpumask_or(mask, mask, cpumask_of_node(k));
6705 }
6706 }
6707 }
6708
6709 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6710 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6711 if (!tl)
6712 return;
6713
6714 /*
6715 * Copy the default topology bits..
6716 */
6717 for (i = 0; default_topology[i].init; i++)
6718 tl[i] = default_topology[i];
6719
6720 /*
6721 * .. and append 'j' levels of NUMA goodness.
6722 */
6723 for (j = 0; j < level; i++, j++) {
6724 tl[i] = (struct sched_domain_topology_level){
6725 .init = sd_numa_init,
6726 .mask = sd_numa_mask,
6727 .flags = SDTL_OVERLAP,
6728 .numa_level = j,
6729 };
6730 }
6731
6732 sched_domain_topology = tl;
6733}
6734#else
6735static inline void sched_init_numa(void)
6736{
6737}
6738#endif /* CONFIG_NUMA */
6739
6740static int __sdt_alloc(const struct cpumask *cpu_map)
6741{
6742 struct sched_domain_topology_level *tl;
6743 int j;
6744
6745 for (tl = sched_domain_topology; tl->init; tl++) {
6746 struct sd_data *sdd = &tl->data;
6747
6748 sdd->sd = alloc_percpu(struct sched_domain *);
6749 if (!sdd->sd)
6750 return -ENOMEM;
6751
6752 sdd->sg = alloc_percpu(struct sched_group *);
6753 if (!sdd->sg)
6754 return -ENOMEM;
6755
6756 sdd->sgp = alloc_percpu(struct sched_group_power *);
6757 if (!sdd->sgp)
6758 return -ENOMEM;
6759
6760 for_each_cpu(j, cpu_map) {
6761 struct sched_domain *sd;
6762 struct sched_group *sg;
6763 struct sched_group_power *sgp;
6764
6765 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6766 GFP_KERNEL, cpu_to_node(j));
6767 if (!sd)
6768 return -ENOMEM;
6769
6770 *per_cpu_ptr(sdd->sd, j) = sd;
6771
6772 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6773 GFP_KERNEL, cpu_to_node(j));
6774 if (!sg)
6775 return -ENOMEM;
6776
6777 sg->next = sg;
6778
6779 *per_cpu_ptr(sdd->sg, j) = sg;
6780
6781 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6782 GFP_KERNEL, cpu_to_node(j));
6783 if (!sgp)
6784 return -ENOMEM;
6785
6786 *per_cpu_ptr(sdd->sgp, j) = sgp;
6787 }
6788 }
6789
6790 return 0;
6791}
6792
6793static void __sdt_free(const struct cpumask *cpu_map)
6794{
6795 struct sched_domain_topology_level *tl;
6796 int j;
6797
6798 for (tl = sched_domain_topology; tl->init; tl++) {
6799 struct sd_data *sdd = &tl->data;
6800
6801 for_each_cpu(j, cpu_map) {
6802 struct sched_domain *sd;
6803
6804 if (sdd->sd) {
6805 sd = *per_cpu_ptr(sdd->sd, j);
6806 if (sd && (sd->flags & SD_OVERLAP))
6807 free_sched_groups(sd->groups, 0);
6808 kfree(*per_cpu_ptr(sdd->sd, j));
6809 }
6810
6811 if (sdd->sg)
6812 kfree(*per_cpu_ptr(sdd->sg, j));
6813 if (sdd->sgp)
6814 kfree(*per_cpu_ptr(sdd->sgp, j));
6815 }
6816 free_percpu(sdd->sd);
6817 sdd->sd = NULL;
6818 free_percpu(sdd->sg);
6819 sdd->sg = NULL;
6820 free_percpu(sdd->sgp);
6821 sdd->sgp = NULL;
6822 }
6823}
6824
6825struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6826 struct s_data *d, const struct cpumask *cpu_map,
6827 struct sched_domain_attr *attr, struct sched_domain *child,
6828 int cpu)
6829{
6830 struct sched_domain *sd = tl->init(tl, cpu);
6831 if (!sd)
6832 return child;
6833
6834 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6835 if (child) {
6836 sd->level = child->level + 1;
6837 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6838 child->parent = sd;
6839 }
6840 sd->child = child;
6841 set_domain_attribute(sd, attr);
6842
6843 return sd;
6844}
6845
6846/*
6847 * Build sched domains for a given set of cpus and attach the sched domains
6848 * to the individual cpus
6849 */
6850static int build_sched_domains(const struct cpumask *cpu_map,
6851 struct sched_domain_attr *attr)
6852{
6853 enum s_alloc alloc_state = sa_none;
6854 struct sched_domain *sd;
6855 struct s_data d;
6856 int i, ret = -ENOMEM;
6857
6858 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6859 if (alloc_state != sa_rootdomain)
6860 goto error;
6861
6862 /* Set up domains for cpus specified by the cpu_map. */
6863 for_each_cpu(i, cpu_map) {
6864 struct sched_domain_topology_level *tl;
6865
6866 sd = NULL;
6867 for (tl = sched_domain_topology; tl->init; tl++) {
6868 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6869 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6870 sd->flags |= SD_OVERLAP;
6871 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6872 break;
6873 }
6874
6875 while (sd->child)
6876 sd = sd->child;
6877
6878 *per_cpu_ptr(d.sd, i) = sd;
6879 }
6880
6881 /* Build the groups for the domains */
6882 for_each_cpu(i, cpu_map) {
6883 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6884 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6885 if (sd->flags & SD_OVERLAP) {
6886 if (build_overlap_sched_groups(sd, i))
6887 goto error;
6888 } else {
6889 if (build_sched_groups(sd, i))
6890 goto error;
6891 }
6892 }
6893 }
6894
6895 /* Calculate CPU power for physical packages and nodes */
6896 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6897 if (!cpumask_test_cpu(i, cpu_map))
6898 continue;
6899
6900 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6901 claim_allocations(i, sd);
6902 init_sched_groups_power(i, sd);
6903 }
6904 }
6905
6906 /* Attach the domains */
6907 rcu_read_lock();
6908 for_each_cpu(i, cpu_map) {
6909 sd = *per_cpu_ptr(d.sd, i);
6910 cpu_attach_domain(sd, d.rd, i);
6911 }
6912 rcu_read_unlock();
6913
6914 ret = 0;
6915error:
6916 __free_domain_allocs(&d, alloc_state, cpu_map);
6917 return ret;
6918}
6919
6920static cpumask_var_t *doms_cur; /* current sched domains */
6921static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6922static struct sched_domain_attr *dattr_cur;
6923 /* attribues of custom domains in 'doms_cur' */
6924
6925/*
6926 * Special case: If a kmalloc of a doms_cur partition (array of
6927 * cpumask) fails, then fallback to a single sched domain,
6928 * as determined by the single cpumask fallback_doms.
6929 */
6930static cpumask_var_t fallback_doms;
6931
6932/*
6933 * arch_update_cpu_topology lets virtualized architectures update the
6934 * cpu core maps. It is supposed to return 1 if the topology changed
6935 * or 0 if it stayed the same.
6936 */
6937int __attribute__((weak)) arch_update_cpu_topology(void)
6938{
6939 return 0;
6940}
6941
6942cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6943{
6944 int i;
6945 cpumask_var_t *doms;
6946
6947 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6948 if (!doms)
6949 return NULL;
6950 for (i = 0; i < ndoms; i++) {
6951 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6952 free_sched_domains(doms, i);
6953 return NULL;
6954 }
6955 }
6956 return doms;
6957}
6958
6959void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6960{
6961 unsigned int i;
6962 for (i = 0; i < ndoms; i++)
6963 free_cpumask_var(doms[i]);
6964 kfree(doms);
6965}
6966
6967/*
6968 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6969 * For now this just excludes isolated cpus, but could be used to
6970 * exclude other special cases in the future.
6971 */
6972static int init_sched_domains(const struct cpumask *cpu_map)
6973{
6974 int err;
6975
6976 arch_update_cpu_topology();
6977 ndoms_cur = 1;
6978 doms_cur = alloc_sched_domains(ndoms_cur);
6979 if (!doms_cur)
6980 doms_cur = &fallback_doms;
6981 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6982 err = build_sched_domains(doms_cur[0], NULL);
6983 register_sched_domain_sysctl();
6984
6985 return err;
6986}
6987
6988/*
6989 * Detach sched domains from a group of cpus specified in cpu_map
6990 * These cpus will now be attached to the NULL domain
6991 */
6992static void detach_destroy_domains(const struct cpumask *cpu_map)
6993{
6994 int i;
6995
6996 rcu_read_lock();
6997 for_each_cpu(i, cpu_map)
6998 cpu_attach_domain(NULL, &def_root_domain, i);
6999 rcu_read_unlock();
7000}
7001
7002/* handle null as "default" */
7003static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7004 struct sched_domain_attr *new, int idx_new)
7005{
7006 struct sched_domain_attr tmp;
7007
7008 /* fast path */
7009 if (!new && !cur)
7010 return 1;
7011
7012 tmp = SD_ATTR_INIT;
7013 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7014 new ? (new + idx_new) : &tmp,
7015 sizeof(struct sched_domain_attr));
7016}
7017
7018/*
7019 * Partition sched domains as specified by the 'ndoms_new'
7020 * cpumasks in the array doms_new[] of cpumasks. This compares
7021 * doms_new[] to the current sched domain partitioning, doms_cur[].
7022 * It destroys each deleted domain and builds each new domain.
7023 *
7024 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7025 * The masks don't intersect (don't overlap.) We should setup one
7026 * sched domain for each mask. CPUs not in any of the cpumasks will
7027 * not be load balanced. If the same cpumask appears both in the
7028 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7029 * it as it is.
7030 *
7031 * The passed in 'doms_new' should be allocated using
7032 * alloc_sched_domains. This routine takes ownership of it and will
7033 * free_sched_domains it when done with it. If the caller failed the
7034 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7035 * and partition_sched_domains() will fallback to the single partition
7036 * 'fallback_doms', it also forces the domains to be rebuilt.
7037 *
7038 * If doms_new == NULL it will be replaced with cpu_online_mask.
7039 * ndoms_new == 0 is a special case for destroying existing domains,
7040 * and it will not create the default domain.
7041 *
7042 * Call with hotplug lock held
7043 */
7044void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7045 struct sched_domain_attr *dattr_new)
7046{
7047 int i, j, n;
7048 int new_topology;
7049
7050 mutex_lock(&sched_domains_mutex);
7051
7052 /* always unregister in case we don't destroy any domains */
7053 unregister_sched_domain_sysctl();
7054
7055 /* Let architecture update cpu core mappings. */
7056 new_topology = arch_update_cpu_topology();
7057
7058 n = doms_new ? ndoms_new : 0;
7059
7060 /* Destroy deleted domains */
7061 for (i = 0; i < ndoms_cur; i++) {
7062 for (j = 0; j < n && !new_topology; j++) {
7063 if (cpumask_equal(doms_cur[i], doms_new[j])
7064 && dattrs_equal(dattr_cur, i, dattr_new, j))
7065 goto match1;
7066 }
7067 /* no match - a current sched domain not in new doms_new[] */
7068 detach_destroy_domains(doms_cur[i]);
7069match1:
7070 ;
7071 }
7072
7073 if (doms_new == NULL) {
7074 ndoms_cur = 0;
7075 doms_new = &fallback_doms;
7076 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7077 WARN_ON_ONCE(dattr_new);
7078 }
7079
7080 /* Build new domains */
7081 for (i = 0; i < ndoms_new; i++) {
7082 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7083 if (cpumask_equal(doms_new[i], doms_cur[j])
7084 && dattrs_equal(dattr_new, i, dattr_cur, j))
7085 goto match2;
7086 }
7087 /* no match - add a new doms_new */
7088 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7089match2:
7090 ;
7091 }
7092
7093 /* Remember the new sched domains */
7094 if (doms_cur != &fallback_doms)
7095 free_sched_domains(doms_cur, ndoms_cur);
7096 kfree(dattr_cur); /* kfree(NULL) is safe */
7097 doms_cur = doms_new;
7098 dattr_cur = dattr_new;
7099 ndoms_cur = ndoms_new;
7100
7101 register_sched_domain_sysctl();
7102
7103 mutex_unlock(&sched_domains_mutex);
7104}
7105
7106/*
7107 * Update cpusets according to cpu_active mask. If cpusets are
7108 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7109 * around partition_sched_domains().
7110 */
7111static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7112 void *hcpu)
7113{
7114 switch (action & ~CPU_TASKS_FROZEN) {
7115 case CPU_ONLINE:
7116 case CPU_DOWN_FAILED:
7117 cpuset_update_active_cpus();
7118 return NOTIFY_OK;
7119 default:
7120 return NOTIFY_DONE;
7121 }
7122}
7123
7124static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7125 void *hcpu)
7126{
7127 switch (action & ~CPU_TASKS_FROZEN) {
7128 case CPU_DOWN_PREPARE:
7129 cpuset_update_active_cpus();
7130 return NOTIFY_OK;
7131 default:
7132 return NOTIFY_DONE;
7133 }
7134}
7135
7136void __init sched_init_smp(void)
7137{
7138 cpumask_var_t non_isolated_cpus;
7139
7140 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7141 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7142
7143 sched_init_numa();
7144
7145 get_online_cpus();
7146 mutex_lock(&sched_domains_mutex);
7147 init_sched_domains(cpu_active_mask);
7148 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7149 if (cpumask_empty(non_isolated_cpus))
7150 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7151 mutex_unlock(&sched_domains_mutex);
7152 put_online_cpus();
7153
7154 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7155 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7156
7157 /* RT runtime code needs to handle some hotplug events */
7158 hotcpu_notifier(update_runtime, 0);
7159
7160 init_hrtick();
7161
7162 /* Move init over to a non-isolated CPU */
7163 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7164 BUG();
7165 sched_init_granularity();
7166 free_cpumask_var(non_isolated_cpus);
7167
7168 init_sched_rt_class();
7169}
7170#else
7171void __init sched_init_smp(void)
7172{
7173 sched_init_granularity();
7174}
7175#endif /* CONFIG_SMP */
7176
7177const_debug unsigned int sysctl_timer_migration = 1;
7178
7179int in_sched_functions(unsigned long addr)
7180{
7181 return in_lock_functions(addr) ||
7182 (addr >= (unsigned long)__sched_text_start
7183 && addr < (unsigned long)__sched_text_end);
7184}
7185
7186#ifdef CONFIG_CGROUP_SCHED
7187struct task_group root_task_group;
7188LIST_HEAD(task_groups);
7189#endif
7190
7191DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
7192
7193void __init sched_init(void)
7194{
7195 int i, j;
7196 unsigned long alloc_size = 0, ptr;
7197
7198#ifdef CONFIG_FAIR_GROUP_SCHED
7199 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7200#endif
7201#ifdef CONFIG_RT_GROUP_SCHED
7202 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7203#endif
7204#ifdef CONFIG_CPUMASK_OFFSTACK
7205 alloc_size += num_possible_cpus() * cpumask_size();
7206#endif
7207 if (alloc_size) {
7208 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7209
7210#ifdef CONFIG_FAIR_GROUP_SCHED
7211 root_task_group.se = (struct sched_entity **)ptr;
7212 ptr += nr_cpu_ids * sizeof(void **);
7213
7214 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7215 ptr += nr_cpu_ids * sizeof(void **);
7216
7217#endif /* CONFIG_FAIR_GROUP_SCHED */
7218#ifdef CONFIG_RT_GROUP_SCHED
7219 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7220 ptr += nr_cpu_ids * sizeof(void **);
7221
7222 root_task_group.rt_rq = (struct rt_rq **)ptr;
7223 ptr += nr_cpu_ids * sizeof(void **);
7224
7225#endif /* CONFIG_RT_GROUP_SCHED */
7226#ifdef CONFIG_CPUMASK_OFFSTACK
7227 for_each_possible_cpu(i) {
7228 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7229 ptr += cpumask_size();
7230 }
7231#endif /* CONFIG_CPUMASK_OFFSTACK */
7232 }
7233
7234#ifdef CONFIG_SMP
7235 init_defrootdomain();
7236#endif
7237
7238 init_rt_bandwidth(&def_rt_bandwidth,
7239 global_rt_period(), global_rt_runtime());
7240
7241#ifdef CONFIG_RT_GROUP_SCHED
7242 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7243 global_rt_period(), global_rt_runtime());
7244#endif /* CONFIG_RT_GROUP_SCHED */
7245
7246#ifdef CONFIG_CGROUP_SCHED
7247 list_add(&root_task_group.list, &task_groups);
7248 INIT_LIST_HEAD(&root_task_group.children);
7249 INIT_LIST_HEAD(&root_task_group.siblings);
7250 autogroup_init(&init_task);
7251
7252#endif /* CONFIG_CGROUP_SCHED */
7253
7254#ifdef CONFIG_CGROUP_CPUACCT
7255 root_cpuacct.cpustat = &kernel_cpustat;
7256 root_cpuacct.cpuusage = alloc_percpu(u64);
7257 /* Too early, not expected to fail */
7258 BUG_ON(!root_cpuacct.cpuusage);
7259#endif
7260 for_each_possible_cpu(i) {
7261 struct rq *rq;
7262
7263 rq = cpu_rq(i);
7264 raw_spin_lock_init(&rq->lock);
7265 rq->nr_running = 0;
7266 rq->calc_load_active = 0;
7267 rq->calc_load_update = jiffies + LOAD_FREQ;
7268 init_cfs_rq(&rq->cfs);
7269 init_rt_rq(&rq->rt, rq);
7270#ifdef CONFIG_FAIR_GROUP_SCHED
7271 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7272 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7273 /*
7274 * How much cpu bandwidth does root_task_group get?
7275 *
7276 * In case of task-groups formed thr' the cgroup filesystem, it
7277 * gets 100% of the cpu resources in the system. This overall
7278 * system cpu resource is divided among the tasks of
7279 * root_task_group and its child task-groups in a fair manner,
7280 * based on each entity's (task or task-group's) weight
7281 * (se->load.weight).
7282 *
7283 * In other words, if root_task_group has 10 tasks of weight
7284 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7285 * then A0's share of the cpu resource is:
7286 *
7287 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7288 *
7289 * We achieve this by letting root_task_group's tasks sit
7290 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7291 */
7292 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7293 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7294#endif /* CONFIG_FAIR_GROUP_SCHED */
7295
7296 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7297#ifdef CONFIG_RT_GROUP_SCHED
7298 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7299 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7300#endif
7301
7302 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7303 rq->cpu_load[j] = 0;
7304
7305 rq->last_load_update_tick = jiffies;
7306
7307#ifdef CONFIG_SMP
7308 rq->sd = NULL;
7309 rq->rd = NULL;
7310 rq->cpu_power = SCHED_POWER_SCALE;
7311 rq->post_schedule = 0;
7312 rq->active_balance = 0;
7313 rq->next_balance = jiffies;
7314 rq->push_cpu = 0;
7315 rq->cpu = i;
7316 rq->online = 0;
7317 rq->idle_stamp = 0;
7318 rq->avg_idle = 2*sysctl_sched_migration_cost;
7319
7320 INIT_LIST_HEAD(&rq->cfs_tasks);
7321
7322 rq_attach_root(rq, &def_root_domain);
7323#ifdef CONFIG_NO_HZ
7324 rq->nohz_flags = 0;
7325#endif
7326#endif
7327 init_rq_hrtick(rq);
7328 atomic_set(&rq->nr_iowait, 0);
7329 }
7330
7331 set_load_weight(&init_task);
7332
7333#ifdef CONFIG_PREEMPT_NOTIFIERS
7334 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7335#endif
7336
7337#ifdef CONFIG_RT_MUTEXES
7338 plist_head_init(&init_task.pi_waiters);
7339#endif
7340
7341 /*
7342 * The boot idle thread does lazy MMU switching as well:
7343 */
7344 atomic_inc(&init_mm.mm_count);
7345 enter_lazy_tlb(&init_mm, current);
7346
7347 /*
7348 * Make us the idle thread. Technically, schedule() should not be
7349 * called from this thread, however somewhere below it might be,
7350 * but because we are the idle thread, we just pick up running again
7351 * when this runqueue becomes "idle".
7352 */
7353 init_idle(current, smp_processor_id());
7354
7355 calc_load_update = jiffies + LOAD_FREQ;
7356
7357 /*
7358 * During early bootup we pretend to be a normal task:
7359 */
7360 current->sched_class = &fair_sched_class;
7361
7362#ifdef CONFIG_SMP
7363 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7364 /* May be allocated at isolcpus cmdline parse time */
7365 if (cpu_isolated_map == NULL)
7366 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7367 idle_thread_set_boot_cpu();
7368#endif
7369 init_sched_fair_class();
7370
7371 scheduler_running = 1;
7372}
7373
7374#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7375static inline int preempt_count_equals(int preempt_offset)
7376{
7377 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7378
7379 return (nested == preempt_offset);
7380}
7381
7382void __might_sleep(const char *file, int line, int preempt_offset)
7383{
7384 static unsigned long prev_jiffy; /* ratelimiting */
7385
7386 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7387 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7388 system_state != SYSTEM_RUNNING || oops_in_progress)
7389 return;
7390 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7391 return;
7392 prev_jiffy = jiffies;
7393
7394 printk(KERN_ERR
7395 "BUG: sleeping function called from invalid context at %s:%d\n",
7396 file, line);
7397 printk(KERN_ERR
7398 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7399 in_atomic(), irqs_disabled(),
7400 current->pid, current->comm);
7401
7402 debug_show_held_locks(current);
7403 if (irqs_disabled())
7404 print_irqtrace_events(current);
7405 dump_stack();
7406}
7407EXPORT_SYMBOL(__might_sleep);
7408#endif
7409
7410#ifdef CONFIG_MAGIC_SYSRQ
7411static void normalize_task(struct rq *rq, struct task_struct *p)
7412{
7413 const struct sched_class *prev_class = p->sched_class;
7414 int old_prio = p->prio;
7415 int on_rq;
7416
7417 on_rq = p->on_rq;
7418 if (on_rq)
7419 dequeue_task(rq, p, 0);
7420 __setscheduler(rq, p, SCHED_NORMAL, 0);
7421 if (on_rq) {
7422 enqueue_task(rq, p, 0);
7423 resched_task(rq->curr);
7424 }
7425
7426 check_class_changed(rq, p, prev_class, old_prio);
7427}
7428
7429void normalize_rt_tasks(void)
7430{
7431 struct task_struct *g, *p;
7432 unsigned long flags;
7433 struct rq *rq;
7434
7435 read_lock_irqsave(&tasklist_lock, flags);
7436 do_each_thread(g, p) {
7437 /*
7438 * Only normalize user tasks:
7439 */
7440 if (!p->mm)
7441 continue;
7442
7443 p->se.exec_start = 0;
7444#ifdef CONFIG_SCHEDSTATS
7445 p->se.statistics.wait_start = 0;
7446 p->se.statistics.sleep_start = 0;
7447 p->se.statistics.block_start = 0;
7448#endif
7449
7450 if (!rt_task(p)) {
7451 /*
7452 * Renice negative nice level userspace
7453 * tasks back to 0:
7454 */
7455 if (TASK_NICE(p) < 0 && p->mm)
7456 set_user_nice(p, 0);
7457 continue;
7458 }
7459
7460 raw_spin_lock(&p->pi_lock);
7461 rq = __task_rq_lock(p);
7462
7463 normalize_task(rq, p);
7464
7465 __task_rq_unlock(rq);
7466 raw_spin_unlock(&p->pi_lock);
7467 } while_each_thread(g, p);
7468
7469 read_unlock_irqrestore(&tasklist_lock, flags);
7470}
7471
7472#endif /* CONFIG_MAGIC_SYSRQ */
7473
7474#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7475/*
7476 * These functions are only useful for the IA64 MCA handling, or kdb.
7477 *
7478 * They can only be called when the whole system has been
7479 * stopped - every CPU needs to be quiescent, and no scheduling
7480 * activity can take place. Using them for anything else would
7481 * be a serious bug, and as a result, they aren't even visible
7482 * under any other configuration.
7483 */
7484
7485/**
7486 * curr_task - return the current task for a given cpu.
7487 * @cpu: the processor in question.
7488 *
7489 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7490 */
7491struct task_struct *curr_task(int cpu)
7492{
7493 return cpu_curr(cpu);
7494}
7495
7496#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7497
7498#ifdef CONFIG_IA64
7499/**
7500 * set_curr_task - set the current task for a given cpu.
7501 * @cpu: the processor in question.
7502 * @p: the task pointer to set.
7503 *
7504 * Description: This function must only be used when non-maskable interrupts
7505 * are serviced on a separate stack. It allows the architecture to switch the
7506 * notion of the current task on a cpu in a non-blocking manner. This function
7507 * must be called with all CPU's synchronized, and interrupts disabled, the
7508 * and caller must save the original value of the current task (see
7509 * curr_task() above) and restore that value before reenabling interrupts and
7510 * re-starting the system.
7511 *
7512 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7513 */
7514void set_curr_task(int cpu, struct task_struct *p)
7515{
7516 cpu_curr(cpu) = p;
7517}
7518
7519#endif
7520
7521#ifdef CONFIG_CGROUP_SCHED
7522/* task_group_lock serializes the addition/removal of task groups */
7523static DEFINE_SPINLOCK(task_group_lock);
7524
7525static void free_sched_group(struct task_group *tg)
7526{
7527 free_fair_sched_group(tg);
7528 free_rt_sched_group(tg);
7529 autogroup_free(tg);
7530 kfree(tg);
7531}
7532
7533/* allocate runqueue etc for a new task group */
7534struct task_group *sched_create_group(struct task_group *parent)
7535{
7536 struct task_group *tg;
7537 unsigned long flags;
7538
7539 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7540 if (!tg)
7541 return ERR_PTR(-ENOMEM);
7542
7543 if (!alloc_fair_sched_group(tg, parent))
7544 goto err;
7545
7546 if (!alloc_rt_sched_group(tg, parent))
7547 goto err;
7548
7549 spin_lock_irqsave(&task_group_lock, flags);
7550 list_add_rcu(&tg->list, &task_groups);
7551
7552 WARN_ON(!parent); /* root should already exist */
7553
7554 tg->parent = parent;
7555 INIT_LIST_HEAD(&tg->children);
7556 list_add_rcu(&tg->siblings, &parent->children);
7557 spin_unlock_irqrestore(&task_group_lock, flags);
7558
7559 return tg;
7560
7561err:
7562 free_sched_group(tg);
7563 return ERR_PTR(-ENOMEM);
7564}
7565
7566/* rcu callback to free various structures associated with a task group */
7567static void free_sched_group_rcu(struct rcu_head *rhp)
7568{
7569 /* now it should be safe to free those cfs_rqs */
7570 free_sched_group(container_of(rhp, struct task_group, rcu));
7571}
7572
7573/* Destroy runqueue etc associated with a task group */
7574void sched_destroy_group(struct task_group *tg)
7575{
7576 unsigned long flags;
7577 int i;
7578
7579 /* end participation in shares distribution */
7580 for_each_possible_cpu(i)
7581 unregister_fair_sched_group(tg, i);
7582
7583 spin_lock_irqsave(&task_group_lock, flags);
7584 list_del_rcu(&tg->list);
7585 list_del_rcu(&tg->siblings);
7586 spin_unlock_irqrestore(&task_group_lock, flags);
7587
7588 /* wait for possible concurrent references to cfs_rqs complete */
7589 call_rcu(&tg->rcu, free_sched_group_rcu);
7590}
7591
7592/* change task's runqueue when it moves between groups.
7593 * The caller of this function should have put the task in its new group
7594 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7595 * reflect its new group.
7596 */
7597void sched_move_task(struct task_struct *tsk)
7598{
7599 struct task_group *tg;
7600 int on_rq, running;
7601 unsigned long flags;
7602 struct rq *rq;
7603
7604 rq = task_rq_lock(tsk, &flags);
7605
7606 running = task_current(rq, tsk);
7607 on_rq = tsk->on_rq;
7608
7609 if (on_rq)
7610 dequeue_task(rq, tsk, 0);
7611 if (unlikely(running))
7612 tsk->sched_class->put_prev_task(rq, tsk);
7613
7614 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7615 lockdep_is_held(&tsk->sighand->siglock)),
7616 struct task_group, css);
7617 tg = autogroup_task_group(tsk, tg);
7618 tsk->sched_task_group = tg;
7619
7620#ifdef CONFIG_FAIR_GROUP_SCHED
7621 if (tsk->sched_class->task_move_group)
7622 tsk->sched_class->task_move_group(tsk, on_rq);
7623 else
7624#endif
7625 set_task_rq(tsk, task_cpu(tsk));
7626
7627 if (unlikely(running))
7628 tsk->sched_class->set_curr_task(rq);
7629 if (on_rq)
7630 enqueue_task(rq, tsk, 0);
7631
7632 task_rq_unlock(rq, tsk, &flags);
7633}
7634#endif /* CONFIG_CGROUP_SCHED */
7635
7636#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7637static unsigned long to_ratio(u64 period, u64 runtime)
7638{
7639 if (runtime == RUNTIME_INF)
7640 return 1ULL << 20;
7641
7642 return div64_u64(runtime << 20, period);
7643}
7644#endif
7645
7646#ifdef CONFIG_RT_GROUP_SCHED
7647/*
7648 * Ensure that the real time constraints are schedulable.
7649 */
7650static DEFINE_MUTEX(rt_constraints_mutex);
7651
7652/* Must be called with tasklist_lock held */
7653static inline int tg_has_rt_tasks(struct task_group *tg)
7654{
7655 struct task_struct *g, *p;
7656
7657 do_each_thread(g, p) {
7658 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7659 return 1;
7660 } while_each_thread(g, p);
7661
7662 return 0;
7663}
7664
7665struct rt_schedulable_data {
7666 struct task_group *tg;
7667 u64 rt_period;
7668 u64 rt_runtime;
7669};
7670
7671static int tg_rt_schedulable(struct task_group *tg, void *data)
7672{
7673 struct rt_schedulable_data *d = data;
7674 struct task_group *child;
7675 unsigned long total, sum = 0;
7676 u64 period, runtime;
7677
7678 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7679 runtime = tg->rt_bandwidth.rt_runtime;
7680
7681 if (tg == d->tg) {
7682 period = d->rt_period;
7683 runtime = d->rt_runtime;
7684 }
7685
7686 /*
7687 * Cannot have more runtime than the period.
7688 */
7689 if (runtime > period && runtime != RUNTIME_INF)
7690 return -EINVAL;
7691
7692 /*
7693 * Ensure we don't starve existing RT tasks.
7694 */
7695 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7696 return -EBUSY;
7697
7698 total = to_ratio(period, runtime);
7699
7700 /*
7701 * Nobody can have more than the global setting allows.
7702 */
7703 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7704 return -EINVAL;
7705
7706 /*
7707 * The sum of our children's runtime should not exceed our own.
7708 */
7709 list_for_each_entry_rcu(child, &tg->children, siblings) {
7710 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7711 runtime = child->rt_bandwidth.rt_runtime;
7712
7713 if (child == d->tg) {
7714 period = d->rt_period;
7715 runtime = d->rt_runtime;
7716 }
7717
7718 sum += to_ratio(period, runtime);
7719 }
7720
7721 if (sum > total)
7722 return -EINVAL;
7723
7724 return 0;
7725}
7726
7727static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7728{
7729 int ret;
7730
7731 struct rt_schedulable_data data = {
7732 .tg = tg,
7733 .rt_period = period,
7734 .rt_runtime = runtime,
7735 };
7736
7737 rcu_read_lock();
7738 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7739 rcu_read_unlock();
7740
7741 return ret;
7742}
7743
7744static int tg_set_rt_bandwidth(struct task_group *tg,
7745 u64 rt_period, u64 rt_runtime)
7746{
7747 int i, err = 0;
7748
7749 mutex_lock(&rt_constraints_mutex);
7750 read_lock(&tasklist_lock);
7751 err = __rt_schedulable(tg, rt_period, rt_runtime);
7752 if (err)
7753 goto unlock;
7754
7755 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7756 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7757 tg->rt_bandwidth.rt_runtime = rt_runtime;
7758
7759 for_each_possible_cpu(i) {
7760 struct rt_rq *rt_rq = tg->rt_rq[i];
7761
7762 raw_spin_lock(&rt_rq->rt_runtime_lock);
7763 rt_rq->rt_runtime = rt_runtime;
7764 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7765 }
7766 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7767unlock:
7768 read_unlock(&tasklist_lock);
7769 mutex_unlock(&rt_constraints_mutex);
7770
7771 return err;
7772}
7773
7774int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7775{
7776 u64 rt_runtime, rt_period;
7777
7778 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7779 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7780 if (rt_runtime_us < 0)
7781 rt_runtime = RUNTIME_INF;
7782
7783 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7784}
7785
7786long sched_group_rt_runtime(struct task_group *tg)
7787{
7788 u64 rt_runtime_us;
7789
7790 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7791 return -1;
7792
7793 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7794 do_div(rt_runtime_us, NSEC_PER_USEC);
7795 return rt_runtime_us;
7796}
7797
7798int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7799{
7800 u64 rt_runtime, rt_period;
7801
7802 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7803 rt_runtime = tg->rt_bandwidth.rt_runtime;
7804
7805 if (rt_period == 0)
7806 return -EINVAL;
7807
7808 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7809}
7810
7811long sched_group_rt_period(struct task_group *tg)
7812{
7813 u64 rt_period_us;
7814
7815 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7816 do_div(rt_period_us, NSEC_PER_USEC);
7817 return rt_period_us;
7818}
7819
7820static int sched_rt_global_constraints(void)
7821{
7822 u64 runtime, period;
7823 int ret = 0;
7824
7825 if (sysctl_sched_rt_period <= 0)
7826 return -EINVAL;
7827
7828 runtime = global_rt_runtime();
7829 period = global_rt_period();
7830
7831 /*
7832 * Sanity check on the sysctl variables.
7833 */
7834 if (runtime > period && runtime != RUNTIME_INF)
7835 return -EINVAL;
7836
7837 mutex_lock(&rt_constraints_mutex);
7838 read_lock(&tasklist_lock);
7839 ret = __rt_schedulable(NULL, 0, 0);
7840 read_unlock(&tasklist_lock);
7841 mutex_unlock(&rt_constraints_mutex);
7842
7843 return ret;
7844}
7845
7846int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7847{
7848 /* Don't accept realtime tasks when there is no way for them to run */
7849 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7850 return 0;
7851
7852 return 1;
7853}
7854
7855#else /* !CONFIG_RT_GROUP_SCHED */
7856static int sched_rt_global_constraints(void)
7857{
7858 unsigned long flags;
7859 int i;
7860
7861 if (sysctl_sched_rt_period <= 0)
7862 return -EINVAL;
7863
7864 /*
7865 * There's always some RT tasks in the root group
7866 * -- migration, kstopmachine etc..
7867 */
7868 if (sysctl_sched_rt_runtime == 0)
7869 return -EBUSY;
7870
7871 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7872 for_each_possible_cpu(i) {
7873 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7874
7875 raw_spin_lock(&rt_rq->rt_runtime_lock);
7876 rt_rq->rt_runtime = global_rt_runtime();
7877 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7878 }
7879 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7880
7881 return 0;
7882}
7883#endif /* CONFIG_RT_GROUP_SCHED */
7884
7885int sched_rt_handler(struct ctl_table *table, int write,
7886 void __user *buffer, size_t *lenp,
7887 loff_t *ppos)
7888{
7889 int ret;
7890 int old_period, old_runtime;
7891 static DEFINE_MUTEX(mutex);
7892
7893 mutex_lock(&mutex);
7894 old_period = sysctl_sched_rt_period;
7895 old_runtime = sysctl_sched_rt_runtime;
7896
7897 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7898
7899 if (!ret && write) {
7900 ret = sched_rt_global_constraints();
7901 if (ret) {
7902 sysctl_sched_rt_period = old_period;
7903 sysctl_sched_rt_runtime = old_runtime;
7904 } else {
7905 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7906 def_rt_bandwidth.rt_period =
7907 ns_to_ktime(global_rt_period());
7908 }
7909 }
7910 mutex_unlock(&mutex);
7911
7912 return ret;
7913}
7914
7915#ifdef CONFIG_CGROUP_SCHED
7916
7917/* return corresponding task_group object of a cgroup */
7918static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7919{
7920 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7921 struct task_group, css);
7922}
7923
7924static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7925{
7926 struct task_group *tg, *parent;
7927
7928 if (!cgrp->parent) {
7929 /* This is early initialization for the top cgroup */
7930 return &root_task_group.css;
7931 }
7932
7933 parent = cgroup_tg(cgrp->parent);
7934 tg = sched_create_group(parent);
7935 if (IS_ERR(tg))
7936 return ERR_PTR(-ENOMEM);
7937
7938 return &tg->css;
7939}
7940
7941static void cpu_cgroup_destroy(struct cgroup *cgrp)
7942{
7943 struct task_group *tg = cgroup_tg(cgrp);
7944
7945 sched_destroy_group(tg);
7946}
7947
7948static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7949 struct cgroup_taskset *tset)
7950{
7951 struct task_struct *task;
7952
7953 cgroup_taskset_for_each(task, cgrp, tset) {
7954#ifdef CONFIG_RT_GROUP_SCHED
7955 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7956 return -EINVAL;
7957#else
7958 /* We don't support RT-tasks being in separate groups */
7959 if (task->sched_class != &fair_sched_class)
7960 return -EINVAL;
7961#endif
7962 }
7963 return 0;
7964}
7965
7966static void cpu_cgroup_attach(struct cgroup *cgrp,
7967 struct cgroup_taskset *tset)
7968{
7969 struct task_struct *task;
7970
7971 cgroup_taskset_for_each(task, cgrp, tset)
7972 sched_move_task(task);
7973}
7974
7975static void
7976cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7977 struct task_struct *task)
7978{
7979 /*
7980 * cgroup_exit() is called in the copy_process() failure path.
7981 * Ignore this case since the task hasn't ran yet, this avoids
7982 * trying to poke a half freed task state from generic code.
7983 */
7984 if (!(task->flags & PF_EXITING))
7985 return;
7986
7987 sched_move_task(task);
7988}
7989
7990#ifdef CONFIG_FAIR_GROUP_SCHED
7991static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7992 u64 shareval)
7993{
7994 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7995}
7996
7997static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7998{
7999 struct task_group *tg = cgroup_tg(cgrp);
8000
8001 return (u64) scale_load_down(tg->shares);
8002}
8003
8004#ifdef CONFIG_CFS_BANDWIDTH
8005static DEFINE_MUTEX(cfs_constraints_mutex);
8006
8007const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8008const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8009
8010static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8011
8012static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8013{
8014 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8015 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8016
8017 if (tg == &root_task_group)
8018 return -EINVAL;
8019
8020 /*
8021 * Ensure we have at some amount of bandwidth every period. This is
8022 * to prevent reaching a state of large arrears when throttled via
8023 * entity_tick() resulting in prolonged exit starvation.
8024 */
8025 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8026 return -EINVAL;
8027
8028 /*
8029 * Likewise, bound things on the otherside by preventing insane quota
8030 * periods. This also allows us to normalize in computing quota
8031 * feasibility.
8032 */
8033 if (period > max_cfs_quota_period)
8034 return -EINVAL;
8035
8036 mutex_lock(&cfs_constraints_mutex);
8037 ret = __cfs_schedulable(tg, period, quota);
8038 if (ret)
8039 goto out_unlock;
8040
8041 runtime_enabled = quota != RUNTIME_INF;
8042 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8043 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
8044 raw_spin_lock_irq(&cfs_b->lock);
8045 cfs_b->period = ns_to_ktime(period);
8046 cfs_b->quota = quota;
8047
8048 __refill_cfs_bandwidth_runtime(cfs_b);
8049 /* restart the period timer (if active) to handle new period expiry */
8050 if (runtime_enabled && cfs_b->timer_active) {
8051 /* force a reprogram */
8052 cfs_b->timer_active = 0;
8053 __start_cfs_bandwidth(cfs_b);
8054 }
8055 raw_spin_unlock_irq(&cfs_b->lock);
8056
8057 for_each_possible_cpu(i) {
8058 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8059 struct rq *rq = cfs_rq->rq;
8060
8061 raw_spin_lock_irq(&rq->lock);
8062 cfs_rq->runtime_enabled = runtime_enabled;
8063 cfs_rq->runtime_remaining = 0;
8064
8065 if (cfs_rq->throttled)
8066 unthrottle_cfs_rq(cfs_rq);
8067 raw_spin_unlock_irq(&rq->lock);
8068 }
8069out_unlock:
8070 mutex_unlock(&cfs_constraints_mutex);
8071
8072 return ret;
8073}
8074
8075int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8076{
8077 u64 quota, period;
8078
8079 period = ktime_to_ns(tg->cfs_bandwidth.period);
8080 if (cfs_quota_us < 0)
8081 quota = RUNTIME_INF;
8082 else
8083 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8084
8085 return tg_set_cfs_bandwidth(tg, period, quota);
8086}
8087
8088long tg_get_cfs_quota(struct task_group *tg)
8089{
8090 u64 quota_us;
8091
8092 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8093 return -1;
8094
8095 quota_us = tg->cfs_bandwidth.quota;
8096 do_div(quota_us, NSEC_PER_USEC);
8097
8098 return quota_us;
8099}
8100
8101int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8102{
8103 u64 quota, period;
8104
8105 period = (u64)cfs_period_us * NSEC_PER_USEC;
8106 quota = tg->cfs_bandwidth.quota;
8107
8108 return tg_set_cfs_bandwidth(tg, period, quota);
8109}
8110
8111long tg_get_cfs_period(struct task_group *tg)
8112{
8113 u64 cfs_period_us;
8114
8115 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8116 do_div(cfs_period_us, NSEC_PER_USEC);
8117
8118 return cfs_period_us;
8119}
8120
8121static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
8122{
8123 return tg_get_cfs_quota(cgroup_tg(cgrp));
8124}
8125
8126static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
8127 s64 cfs_quota_us)
8128{
8129 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
8130}
8131
8132static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
8133{
8134 return tg_get_cfs_period(cgroup_tg(cgrp));
8135}
8136
8137static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8138 u64 cfs_period_us)
8139{
8140 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
8141}
8142
8143struct cfs_schedulable_data {
8144 struct task_group *tg;
8145 u64 period, quota;
8146};
8147
8148/*
8149 * normalize group quota/period to be quota/max_period
8150 * note: units are usecs
8151 */
8152static u64 normalize_cfs_quota(struct task_group *tg,
8153 struct cfs_schedulable_data *d)
8154{
8155 u64 quota, period;
8156
8157 if (tg == d->tg) {
8158 period = d->period;
8159 quota = d->quota;
8160 } else {
8161 period = tg_get_cfs_period(tg);
8162 quota = tg_get_cfs_quota(tg);
8163 }
8164
8165 /* note: these should typically be equivalent */
8166 if (quota == RUNTIME_INF || quota == -1)
8167 return RUNTIME_INF;
8168
8169 return to_ratio(period, quota);
8170}
8171
8172static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8173{
8174 struct cfs_schedulable_data *d = data;
8175 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8176 s64 quota = 0, parent_quota = -1;
8177
8178 if (!tg->parent) {
8179 quota = RUNTIME_INF;
8180 } else {
8181 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8182
8183 quota = normalize_cfs_quota(tg, d);
8184 parent_quota = parent_b->hierarchal_quota;
8185
8186 /*
8187 * ensure max(child_quota) <= parent_quota, inherit when no
8188 * limit is set
8189 */
8190 if (quota == RUNTIME_INF)
8191 quota = parent_quota;
8192 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8193 return -EINVAL;
8194 }
8195 cfs_b->hierarchal_quota = quota;
8196
8197 return 0;
8198}
8199
8200static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8201{
8202 int ret;
8203 struct cfs_schedulable_data data = {
8204 .tg = tg,
8205 .period = period,
8206 .quota = quota,
8207 };
8208
8209 if (quota != RUNTIME_INF) {
8210 do_div(data.period, NSEC_PER_USEC);
8211 do_div(data.quota, NSEC_PER_USEC);
8212 }
8213
8214 rcu_read_lock();
8215 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8216 rcu_read_unlock();
8217
8218 return ret;
8219}
8220
8221static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
8222 struct cgroup_map_cb *cb)
8223{
8224 struct task_group *tg = cgroup_tg(cgrp);
8225 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8226
8227 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
8228 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
8229 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
8230
8231 return 0;
8232}
8233#endif /* CONFIG_CFS_BANDWIDTH */
8234#endif /* CONFIG_FAIR_GROUP_SCHED */
8235
8236#ifdef CONFIG_RT_GROUP_SCHED
8237static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8238 s64 val)
8239{
8240 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8241}
8242
8243static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8244{
8245 return sched_group_rt_runtime(cgroup_tg(cgrp));
8246}
8247
8248static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8249 u64 rt_period_us)
8250{
8251 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8252}
8253
8254static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8255{
8256 return sched_group_rt_period(cgroup_tg(cgrp));
8257}
8258#endif /* CONFIG_RT_GROUP_SCHED */
8259
8260static struct cftype cpu_files[] = {
8261#ifdef CONFIG_FAIR_GROUP_SCHED
8262 {
8263 .name = "shares",
8264 .read_u64 = cpu_shares_read_u64,
8265 .write_u64 = cpu_shares_write_u64,
8266 },
8267#endif
8268#ifdef CONFIG_CFS_BANDWIDTH
8269 {
8270 .name = "cfs_quota_us",
8271 .read_s64 = cpu_cfs_quota_read_s64,
8272 .write_s64 = cpu_cfs_quota_write_s64,
8273 },
8274 {
8275 .name = "cfs_period_us",
8276 .read_u64 = cpu_cfs_period_read_u64,
8277 .write_u64 = cpu_cfs_period_write_u64,
8278 },
8279 {
8280 .name = "stat",
8281 .read_map = cpu_stats_show,
8282 },
8283#endif
8284#ifdef CONFIG_RT_GROUP_SCHED
8285 {
8286 .name = "rt_runtime_us",
8287 .read_s64 = cpu_rt_runtime_read,
8288 .write_s64 = cpu_rt_runtime_write,
8289 },
8290 {
8291 .name = "rt_period_us",
8292 .read_u64 = cpu_rt_period_read_uint,
8293 .write_u64 = cpu_rt_period_write_uint,
8294 },
8295#endif
8296 { } /* terminate */
8297};
8298
8299struct cgroup_subsys cpu_cgroup_subsys = {
8300 .name = "cpu",
8301 .create = cpu_cgroup_create,
8302 .destroy = cpu_cgroup_destroy,
8303 .can_attach = cpu_cgroup_can_attach,
8304 .attach = cpu_cgroup_attach,
8305 .exit = cpu_cgroup_exit,
8306 .subsys_id = cpu_cgroup_subsys_id,
8307 .base_cftypes = cpu_files,
8308 .early_init = 1,
8309};
8310
8311#endif /* CONFIG_CGROUP_SCHED */
8312
8313#ifdef CONFIG_CGROUP_CPUACCT
8314
8315/*
8316 * CPU accounting code for task groups.
8317 *
8318 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8319 * (balbir@in.ibm.com).
8320 */
8321
8322/* create a new cpu accounting group */
8323static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
8324{
8325 struct cpuacct *ca;
8326
8327 if (!cgrp->parent)
8328 return &root_cpuacct.css;
8329
8330 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8331 if (!ca)
8332 goto out;
8333
8334 ca->cpuusage = alloc_percpu(u64);
8335 if (!ca->cpuusage)
8336 goto out_free_ca;
8337
8338 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8339 if (!ca->cpustat)
8340 goto out_free_cpuusage;
8341
8342 return &ca->css;
8343
8344out_free_cpuusage:
8345 free_percpu(ca->cpuusage);
8346out_free_ca:
8347 kfree(ca);
8348out:
8349 return ERR_PTR(-ENOMEM);
8350}
8351
8352/* destroy an existing cpu accounting group */
8353static void cpuacct_destroy(struct cgroup *cgrp)
8354{
8355 struct cpuacct *ca = cgroup_ca(cgrp);
8356
8357 free_percpu(ca->cpustat);
8358 free_percpu(ca->cpuusage);
8359 kfree(ca);
8360}
8361
8362static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8363{
8364 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8365 u64 data;
8366
8367#ifndef CONFIG_64BIT
8368 /*
8369 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8370 */
8371 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8372 data = *cpuusage;
8373 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8374#else
8375 data = *cpuusage;
8376#endif
8377
8378 return data;
8379}
8380
8381static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8382{
8383 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8384
8385#ifndef CONFIG_64BIT
8386 /*
8387 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8388 */
8389 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8390 *cpuusage = val;
8391 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8392#else
8393 *cpuusage = val;
8394#endif
8395}
8396
8397/* return total cpu usage (in nanoseconds) of a group */
8398static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8399{
8400 struct cpuacct *ca = cgroup_ca(cgrp);
8401 u64 totalcpuusage = 0;
8402 int i;
8403
8404 for_each_present_cpu(i)
8405 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8406
8407 return totalcpuusage;
8408}
8409
8410static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8411 u64 reset)
8412{
8413 struct cpuacct *ca = cgroup_ca(cgrp);
8414 int err = 0;
8415 int i;
8416
8417 if (reset) {
8418 err = -EINVAL;
8419 goto out;
8420 }
8421
8422 for_each_present_cpu(i)
8423 cpuacct_cpuusage_write(ca, i, 0);
8424
8425out:
8426 return err;
8427}
8428
8429static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8430 struct seq_file *m)
8431{
8432 struct cpuacct *ca = cgroup_ca(cgroup);
8433 u64 percpu;
8434 int i;
8435
8436 for_each_present_cpu(i) {
8437 percpu = cpuacct_cpuusage_read(ca, i);
8438 seq_printf(m, "%llu ", (unsigned long long) percpu);
8439 }
8440 seq_printf(m, "\n");
8441 return 0;
8442}
8443
8444static const char *cpuacct_stat_desc[] = {
8445 [CPUACCT_STAT_USER] = "user",
8446 [CPUACCT_STAT_SYSTEM] = "system",
8447};
8448
8449static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8450 struct cgroup_map_cb *cb)
8451{
8452 struct cpuacct *ca = cgroup_ca(cgrp);
8453 int cpu;
8454 s64 val = 0;
8455
8456 for_each_online_cpu(cpu) {
8457 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8458 val += kcpustat->cpustat[CPUTIME_USER];
8459 val += kcpustat->cpustat[CPUTIME_NICE];
8460 }
8461 val = cputime64_to_clock_t(val);
8462 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8463
8464 val = 0;
8465 for_each_online_cpu(cpu) {
8466 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8467 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8468 val += kcpustat->cpustat[CPUTIME_IRQ];
8469 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8470 }
8471
8472 val = cputime64_to_clock_t(val);
8473 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8474
8475 return 0;
8476}
8477
8478static struct cftype files[] = {
8479 {
8480 .name = "usage",
8481 .read_u64 = cpuusage_read,
8482 .write_u64 = cpuusage_write,
8483 },
8484 {
8485 .name = "usage_percpu",
8486 .read_seq_string = cpuacct_percpu_seq_read,
8487 },
8488 {
8489 .name = "stat",
8490 .read_map = cpuacct_stats_show,
8491 },
8492 { } /* terminate */
8493};
8494
8495/*
8496 * charge this task's execution time to its accounting group.
8497 *
8498 * called with rq->lock held.
8499 */
8500void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8501{
8502 struct cpuacct *ca;
8503 int cpu;
8504
8505 if (unlikely(!cpuacct_subsys.active))
8506 return;
8507
8508 cpu = task_cpu(tsk);
8509
8510 rcu_read_lock();
8511
8512 ca = task_ca(tsk);
8513
8514 for (; ca; ca = parent_ca(ca)) {
8515 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8516 *cpuusage += cputime;
8517 }
8518
8519 rcu_read_unlock();
8520}
8521
8522struct cgroup_subsys cpuacct_subsys = {
8523 .name = "cpuacct",
8524 .create = cpuacct_create,
8525 .destroy = cpuacct_destroy,
8526 .subsys_id = cpuacct_subsys_id,
8527 .base_cftypes = files,
8528};
8529#endif /* CONFIG_CGROUP_CPUACCT */