Linux Audio

Check our new training course

Loading...
v4.17
 
   1/*
   2 *  kernel/sched/core.c
   3 *
   4 *  Core kernel scheduler code and related syscalls
   5 *
   6 *  Copyright (C) 1991-2002  Linus Torvalds
   7 */
   8#include "sched.h"
   9
  10#include <linux/kthread.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  11#include <linux/nospec.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  12
 
  13#include <asm/switch_to.h>
  14#include <asm/tlb.h>
  15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  16#include "../workqueue_internal.h"
 
  17#include "../smpboot.h"
  18
  19#define CREATE_TRACE_POINTS
  20#include <trace/events/sched.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  21
  22DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
  23
  24#if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
  25/*
  26 * Debugging: various feature bits
  27 *
  28 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
  29 * sysctl_sched_features, defined in sched.h, to allow constants propagation
  30 * at compile time and compiler optimization based on features default.
  31 */
  32#define SCHED_FEAT(name, enabled)	\
  33	(1UL << __SCHED_FEAT_##name) * enabled |
  34const_debug unsigned int sysctl_sched_features =
  35#include "features.h"
  36	0;
  37#undef SCHED_FEAT
  38#endif
 
 
 
 
 
 
 
 
 
 
  39
  40/*
  41 * Number of tasks to iterate in a single balance run.
  42 * Limited because this is done with IRQs disabled.
  43 */
  44const_debug unsigned int sysctl_sched_nr_migrate = 32;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  45
  46/*
  47 * period over which we average the RT time consumption, measured
  48 * in ms.
 
 
 
 
 
 
  49 *
  50 * default: 1s
  51 */
  52const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  53
  54/*
  55 * period over which we measure -rt task CPU usage in us.
  56 * default: 1s
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  57 */
  58unsigned int sysctl_sched_rt_period = 1000000;
  59
  60__read_mostly int scheduler_running;
 
 
  61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  62/*
  63 * part of the period that we allow rt tasks to run in us.
  64 * default: 0.95s
  65 */
  66int sysctl_sched_rt_runtime = 950000;
 
 
 
 
 
 
 
 
 
 
 
 
 
  67
  68/*
  69 * __task_rq_lock - lock the rq @p resides on.
  70 */
  71struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
  72	__acquires(rq->lock)
  73{
  74	struct rq *rq;
  75
  76	lockdep_assert_held(&p->pi_lock);
  77
  78	for (;;) {
  79		rq = task_rq(p);
  80		raw_spin_lock(&rq->lock);
  81		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
  82			rq_pin_lock(rq, rf);
  83			return rq;
  84		}
  85		raw_spin_unlock(&rq->lock);
  86
  87		while (unlikely(task_on_rq_migrating(p)))
  88			cpu_relax();
  89	}
  90}
  91
  92/*
  93 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
  94 */
  95struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
  96	__acquires(p->pi_lock)
  97	__acquires(rq->lock)
  98{
  99	struct rq *rq;
 100
 101	for (;;) {
 102		raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
 103		rq = task_rq(p);
 104		raw_spin_lock(&rq->lock);
 105		/*
 106		 *	move_queued_task()		task_rq_lock()
 107		 *
 108		 *	ACQUIRE (rq->lock)
 109		 *	[S] ->on_rq = MIGRATING		[L] rq = task_rq()
 110		 *	WMB (__set_task_cpu())		ACQUIRE (rq->lock);
 111		 *	[S] ->cpu = new_cpu		[L] task_rq()
 112		 *					[L] ->on_rq
 113		 *	RELEASE (rq->lock)
 114		 *
 115		 * If we observe the old CPU in task_rq_lock, the acquire of
 116		 * the old rq->lock will fully serialize against the stores.
 117		 *
 118		 * If we observe the new CPU in task_rq_lock, the acquire will
 119		 * pair with the WMB to ensure we must then also see migrating.
 
 120		 */
 121		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
 122			rq_pin_lock(rq, rf);
 123			return rq;
 124		}
 125		raw_spin_unlock(&rq->lock);
 126		raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
 127
 128		while (unlikely(task_on_rq_migrating(p)))
 129			cpu_relax();
 130	}
 131}
 132
 133/*
 134 * RQ-clock updating methods:
 135 */
 136
 137static void update_rq_clock_task(struct rq *rq, s64 delta)
 138{
 139/*
 140 * In theory, the compile should just see 0 here, and optimize out the call
 141 * to sched_rt_avg_update. But I don't trust it...
 142 */
 143#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
 144	s64 steal = 0, irq_delta = 0;
 145#endif
 146#ifdef CONFIG_IRQ_TIME_ACCOUNTING
 147	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
 148
 149	/*
 150	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
 151	 * this case when a previous update_rq_clock() happened inside a
 152	 * {soft,}irq region.
 153	 *
 154	 * When this happens, we stop ->clock_task and only update the
 155	 * prev_irq_time stamp to account for the part that fit, so that a next
 156	 * update will consume the rest. This ensures ->clock_task is
 157	 * monotonic.
 158	 *
 159	 * It does however cause some slight miss-attribution of {soft,}irq
 160	 * time, a more accurate solution would be to update the irq_time using
 161	 * the current rq->clock timestamp, except that would require using
 162	 * atomic ops.
 163	 */
 164	if (irq_delta > delta)
 165		irq_delta = delta;
 166
 167	rq->prev_irq_time += irq_delta;
 168	delta -= irq_delta;
 
 
 169#endif
 170#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
 171	if (static_key_false((&paravirt_steal_rq_enabled))) {
 172		steal = paravirt_steal_clock(cpu_of(rq));
 173		steal -= rq->prev_steal_time_rq;
 174
 175		if (unlikely(steal > delta))
 176			steal = delta;
 177
 178		rq->prev_steal_time_rq += steal;
 179		delta -= steal;
 180	}
 181#endif
 182
 183	rq->clock_task += delta;
 184
 185#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
 186	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
 187		sched_rt_avg_update(rq, irq_delta + steal);
 188#endif
 
 189}
 190
 191void update_rq_clock(struct rq *rq)
 192{
 193	s64 delta;
 194
 195	lockdep_assert_held(&rq->lock);
 196
 197	if (rq->clock_update_flags & RQCF_ACT_SKIP)
 198		return;
 199
 200#ifdef CONFIG_SCHED_DEBUG
 201	if (sched_feat(WARN_DOUBLE_CLOCK))
 202		SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
 203	rq->clock_update_flags |= RQCF_UPDATED;
 204#endif
 205
 206	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
 207	if (delta < 0)
 208		return;
 209	rq->clock += delta;
 210	update_rq_clock_task(rq, delta);
 211}
 212
 213
 214#ifdef CONFIG_SCHED_HRTICK
 215/*
 216 * Use HR-timers to deliver accurate preemption points.
 217 */
 218
 219static void hrtick_clear(struct rq *rq)
 220{
 221	if (hrtimer_active(&rq->hrtick_timer))
 222		hrtimer_cancel(&rq->hrtick_timer);
 223}
 224
 225/*
 226 * High-resolution timer tick.
 227 * Runs from hardirq context with interrupts disabled.
 228 */
 229static enum hrtimer_restart hrtick(struct hrtimer *timer)
 230{
 231	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
 232	struct rq_flags rf;
 233
 234	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
 235
 236	rq_lock(rq, &rf);
 237	update_rq_clock(rq);
 238	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
 239	rq_unlock(rq, &rf);
 240
 241	return HRTIMER_NORESTART;
 242}
 243
 244#ifdef CONFIG_SMP
 245
 246static void __hrtick_restart(struct rq *rq)
 247{
 248	struct hrtimer *timer = &rq->hrtick_timer;
 
 249
 250	hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
 251}
 252
 253/*
 254 * called from hardirq (IPI) context
 255 */
 256static void __hrtick_start(void *arg)
 257{
 258	struct rq *rq = arg;
 259	struct rq_flags rf;
 260
 261	rq_lock(rq, &rf);
 262	__hrtick_restart(rq);
 263	rq->hrtick_csd_pending = 0;
 264	rq_unlock(rq, &rf);
 265}
 266
 267/*
 268 * Called to set the hrtick timer state.
 269 *
 270 * called with rq->lock held and irqs disabled
 271 */
 272void hrtick_start(struct rq *rq, u64 delay)
 273{
 274	struct hrtimer *timer = &rq->hrtick_timer;
 275	ktime_t time;
 276	s64 delta;
 277
 278	/*
 279	 * Don't schedule slices shorter than 10000ns, that just
 280	 * doesn't make sense and can cause timer DoS.
 281	 */
 282	delta = max_t(s64, delay, 10000LL);
 283	time = ktime_add_ns(timer->base->get_time(), delta);
 284
 285	hrtimer_set_expires(timer, time);
 286
 287	if (rq == this_rq()) {
 288		__hrtick_restart(rq);
 289	} else if (!rq->hrtick_csd_pending) {
 290		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
 291		rq->hrtick_csd_pending = 1;
 292	}
 293}
 294
 295#else
 296/*
 297 * Called to set the hrtick timer state.
 298 *
 299 * called with rq->lock held and irqs disabled
 300 */
 301void hrtick_start(struct rq *rq, u64 delay)
 302{
 303	/*
 304	 * Don't schedule slices shorter than 10000ns, that just
 305	 * doesn't make sense. Rely on vruntime for fairness.
 306	 */
 307	delay = max_t(u64, delay, 10000LL);
 308	hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
 309		      HRTIMER_MODE_REL_PINNED);
 310}
 
 311#endif /* CONFIG_SMP */
 312
 313static void hrtick_rq_init(struct rq *rq)
 314{
 315#ifdef CONFIG_SMP
 316	rq->hrtick_csd_pending = 0;
 317
 318	rq->hrtick_csd.flags = 0;
 319	rq->hrtick_csd.func = __hrtick_start;
 320	rq->hrtick_csd.info = rq;
 321#endif
 322
 323	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 324	rq->hrtick_timer.function = hrtick;
 325}
 326#else	/* CONFIG_SCHED_HRTICK */
 327static inline void hrtick_clear(struct rq *rq)
 328{
 329}
 330
 331static inline void hrtick_rq_init(struct rq *rq)
 332{
 333}
 334#endif	/* CONFIG_SCHED_HRTICK */
 335
 336/*
 337 * cmpxchg based fetch_or, macro so it works for different integer types
 338 */
 339#define fetch_or(ptr, mask)						\
 340	({								\
 341		typeof(ptr) _ptr = (ptr);				\
 342		typeof(mask) _mask = (mask);				\
 343		typeof(*_ptr) _old, _val = *_ptr;			\
 344									\
 345		for (;;) {						\
 346			_old = cmpxchg(_ptr, _val, _val | _mask);	\
 347			if (_old == _val)				\
 348				break;					\
 349			_val = _old;					\
 350		}							\
 351	_old;								\
 352})
 353
 354#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
 355/*
 356 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
 357 * this avoids any races wrt polling state changes and thereby avoids
 358 * spurious IPIs.
 359 */
 360static bool set_nr_and_not_polling(struct task_struct *p)
 361{
 362	struct thread_info *ti = task_thread_info(p);
 363	return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
 364}
 365
 366/*
 367 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
 368 *
 369 * If this returns true, then the idle task promises to call
 370 * sched_ttwu_pending() and reschedule soon.
 371 */
 372static bool set_nr_if_polling(struct task_struct *p)
 373{
 374	struct thread_info *ti = task_thread_info(p);
 375	typeof(ti->flags) old, val = READ_ONCE(ti->flags);
 376
 377	for (;;) {
 378		if (!(val & _TIF_POLLING_NRFLAG))
 379			return false;
 380		if (val & _TIF_NEED_RESCHED)
 381			return true;
 382		old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
 383		if (old == val)
 384			break;
 385		val = old;
 386	}
 387	return true;
 388}
 389
 390#else
 391static bool set_nr_and_not_polling(struct task_struct *p)
 392{
 393	set_tsk_need_resched(p);
 394	return true;
 395}
 396
 397#ifdef CONFIG_SMP
 398static bool set_nr_if_polling(struct task_struct *p)
 399{
 400	return false;
 401}
 402#endif
 403#endif
 404
 405void wake_q_add(struct wake_q_head *head, struct task_struct *task)
 406{
 407	struct wake_q_node *node = &task->wake_q;
 408
 409	/*
 410	 * Atomically grab the task, if ->wake_q is !nil already it means
 411	 * its already queued (either by us or someone else) and will get the
 412	 * wakeup due to that.
 413	 *
 414	 * This cmpxchg() implies a full barrier, which pairs with the write
 415	 * barrier implied by the wakeup in wake_up_q().
 416	 */
 417	if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
 418		return;
 419
 420	get_task_struct(task);
 421
 422	/*
 423	 * The head is context local, there can be no concurrency.
 424	 */
 425	*head->lastp = node;
 426	head->lastp = &node->next;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 427}
 428
 429void wake_up_q(struct wake_q_head *head)
 430{
 431	struct wake_q_node *node = head->first;
 432
 433	while (node != WAKE_Q_TAIL) {
 434		struct task_struct *task;
 435
 436		task = container_of(node, struct task_struct, wake_q);
 437		BUG_ON(!task);
 438		/* Task can safely be re-inserted now: */
 439		node = node->next;
 440		task->wake_q.next = NULL;
 441
 442		/*
 443		 * wake_up_process() implies a wmb() to pair with the queueing
 444		 * in wake_q_add() so as not to miss wakeups.
 445		 */
 446		wake_up_process(task);
 447		put_task_struct(task);
 448	}
 449}
 450
 451/*
 452 * resched_curr - mark rq's current task 'to be rescheduled now'.
 453 *
 454 * On UP this means the setting of the need_resched flag, on SMP it
 455 * might also involve a cross-CPU call to trigger the scheduler on
 456 * the target CPU.
 457 */
 458void resched_curr(struct rq *rq)
 459{
 460	struct task_struct *curr = rq->curr;
 461	int cpu;
 462
 463	lockdep_assert_held(&rq->lock);
 464
 465	if (test_tsk_need_resched(curr))
 466		return;
 467
 468	cpu = cpu_of(rq);
 469
 470	if (cpu == smp_processor_id()) {
 471		set_tsk_need_resched(curr);
 472		set_preempt_need_resched();
 473		return;
 474	}
 475
 476	if (set_nr_and_not_polling(curr))
 477		smp_send_reschedule(cpu);
 478	else
 479		trace_sched_wake_idle_without_ipi(cpu);
 480}
 481
 482void resched_cpu(int cpu)
 483{
 484	struct rq *rq = cpu_rq(cpu);
 485	unsigned long flags;
 486
 487	raw_spin_lock_irqsave(&rq->lock, flags);
 488	if (cpu_online(cpu) || cpu == smp_processor_id())
 489		resched_curr(rq);
 490	raw_spin_unlock_irqrestore(&rq->lock, flags);
 491}
 492
 493#ifdef CONFIG_SMP
 494#ifdef CONFIG_NO_HZ_COMMON
 495/*
 496 * In the semi idle case, use the nearest busy CPU for migrating timers
 497 * from an idle CPU.  This is good for power-savings.
 498 *
 499 * We don't do similar optimization for completely idle system, as
 500 * selecting an idle CPU will add more delays to the timers than intended
 501 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
 502 */
 503int get_nohz_timer_target(void)
 504{
 505	int i, cpu = smp_processor_id();
 506	struct sched_domain *sd;
 
 507
 508	if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
 509		return cpu;
 
 
 
 
 
 
 
 510
 511	rcu_read_lock();
 512	for_each_domain(cpu, sd) {
 513		for_each_cpu(i, sched_domain_span(sd)) {
 514			if (cpu == i)
 515				continue;
 516
 517			if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
 518				cpu = i;
 519				goto unlock;
 520			}
 521		}
 522	}
 523
 524	if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
 525		cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
 526unlock:
 527	rcu_read_unlock();
 528	return cpu;
 529}
 530
 531/*
 532 * When add_timer_on() enqueues a timer into the timer wheel of an
 533 * idle CPU then this timer might expire before the next timer event
 534 * which is scheduled to wake up that CPU. In case of a completely
 535 * idle system the next event might even be infinite time into the
 536 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
 537 * leaves the inner idle loop so the newly added timer is taken into
 538 * account when the CPU goes back to idle and evaluates the timer
 539 * wheel for the next timer event.
 540 */
 541static void wake_up_idle_cpu(int cpu)
 542{
 543	struct rq *rq = cpu_rq(cpu);
 544
 545	if (cpu == smp_processor_id())
 546		return;
 547
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 548	if (set_nr_and_not_polling(rq->idle))
 549		smp_send_reschedule(cpu);
 550	else
 551		trace_sched_wake_idle_without_ipi(cpu);
 552}
 553
 554static bool wake_up_full_nohz_cpu(int cpu)
 555{
 556	/*
 557	 * We just need the target to call irq_exit() and re-evaluate
 558	 * the next tick. The nohz full kick at least implies that.
 559	 * If needed we can still optimize that later with an
 560	 * empty IRQ.
 561	 */
 562	if (cpu_is_offline(cpu))
 563		return true;  /* Don't try to wake offline CPUs. */
 564	if (tick_nohz_full_cpu(cpu)) {
 565		if (cpu != smp_processor_id() ||
 566		    tick_nohz_tick_stopped())
 567			tick_nohz_full_kick_cpu(cpu);
 568		return true;
 569	}
 570
 571	return false;
 572}
 573
 574/*
 575 * Wake up the specified CPU.  If the CPU is going offline, it is the
 576 * caller's responsibility to deal with the lost wakeup, for example,
 577 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
 578 */
 579void wake_up_nohz_cpu(int cpu)
 580{
 581	if (!wake_up_full_nohz_cpu(cpu))
 582		wake_up_idle_cpu(cpu);
 583}
 584
 585static inline bool got_nohz_idle_kick(void)
 586{
 587	int cpu = smp_processor_id();
 588
 589	if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
 590		return false;
 591
 592	if (idle_cpu(cpu) && !need_resched())
 593		return true;
 594
 595	/*
 596	 * We can't run Idle Load Balance on this CPU for this time so we
 597	 * cancel it and clear NOHZ_BALANCE_KICK
 598	 */
 599	atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
 600	return false;
 601}
 602
 603#else /* CONFIG_NO_HZ_COMMON */
 604
 605static inline bool got_nohz_idle_kick(void)
 606{
 607	return false;
 608}
 609
 610#endif /* CONFIG_NO_HZ_COMMON */
 611
 612#ifdef CONFIG_NO_HZ_FULL
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 613bool sched_can_stop_tick(struct rq *rq)
 614{
 615	int fifo_nr_running;
 616
 617	/* Deadline tasks, even if single, need the tick */
 618	if (rq->dl.dl_nr_running)
 619		return false;
 620
 621	/*
 622	 * If there are more than one RR tasks, we need the tick to effect the
 623	 * actual RR behaviour.
 624	 */
 625	if (rq->rt.rr_nr_running) {
 626		if (rq->rt.rr_nr_running == 1)
 627			return true;
 628		else
 629			return false;
 630	}
 631
 632	/*
 633	 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
 634	 * forced preemption between FIFO tasks.
 635	 */
 636	fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
 637	if (fifo_nr_running)
 638		return true;
 639
 640	/*
 641	 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
 642	 * if there's more than one we need the tick for involuntary
 643	 * preemption.
 644	 */
 645	if (rq->nr_running > 1)
 646		return false;
 647
 
 
 
 
 
 
 
 
 
 
 
 
 648	return true;
 649}
 650#endif /* CONFIG_NO_HZ_FULL */
 651
 652void sched_avg_update(struct rq *rq)
 653{
 654	s64 period = sched_avg_period();
 655
 656	while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
 657		/*
 658		 * Inline assembly required to prevent the compiler
 659		 * optimising this loop into a divmod call.
 660		 * See __iter_div_u64_rem() for another example of this.
 661		 */
 662		asm("" : "+rm" (rq->age_stamp));
 663		rq->age_stamp += period;
 664		rq->rt_avg /= 2;
 665	}
 666}
 667
 668#endif /* CONFIG_SMP */
 669
 670#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
 671			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
 672/*
 673 * Iterate task_group tree rooted at *from, calling @down when first entering a
 674 * node and @up when leaving it for the final time.
 675 *
 676 * Caller must hold rcu_lock or sufficient equivalent.
 677 */
 678int walk_tg_tree_from(struct task_group *from,
 679			     tg_visitor down, tg_visitor up, void *data)
 680{
 681	struct task_group *parent, *child;
 682	int ret;
 683
 684	parent = from;
 685
 686down:
 687	ret = (*down)(parent, data);
 688	if (ret)
 689		goto out;
 690	list_for_each_entry_rcu(child, &parent->children, siblings) {
 691		parent = child;
 692		goto down;
 693
 694up:
 695		continue;
 696	}
 697	ret = (*up)(parent, data);
 698	if (ret || parent == from)
 699		goto out;
 700
 701	child = parent;
 702	parent = parent->parent;
 703	if (parent)
 704		goto up;
 705out:
 706	return ret;
 707}
 708
 709int tg_nop(struct task_group *tg, void *data)
 710{
 711	return 0;
 712}
 713#endif
 714
 715static void set_load_weight(struct task_struct *p, bool update_load)
 716{
 717	int prio = p->static_prio - MAX_RT_PRIO;
 718	struct load_weight *load = &p->se.load;
 719
 720	/*
 721	 * SCHED_IDLE tasks get minimal weight:
 722	 */
 723	if (idle_policy(p->policy)) {
 724		load->weight = scale_load(WEIGHT_IDLEPRIO);
 725		load->inv_weight = WMULT_IDLEPRIO;
 726		return;
 727	}
 728
 729	/*
 730	 * SCHED_OTHER tasks have to update their load when changing their
 731	 * weight
 732	 */
 733	if (update_load && p->sched_class == &fair_sched_class) {
 734		reweight_task(p, prio);
 735	} else {
 736		load->weight = scale_load(sched_prio_to_weight[prio]);
 737		load->inv_weight = sched_prio_to_wmult[prio];
 738	}
 739}
 740
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 741static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
 742{
 743	if (!(flags & ENQUEUE_NOCLOCK))
 744		update_rq_clock(rq);
 745
 746	if (!(flags & ENQUEUE_RESTORE))
 747		sched_info_queued(rq, p);
 
 
 748
 
 749	p->sched_class->enqueue_task(rq, p, flags);
 
 
 
 750}
 751
 752static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
 753{
 
 
 
 754	if (!(flags & DEQUEUE_NOCLOCK))
 755		update_rq_clock(rq);
 756
 757	if (!(flags & DEQUEUE_SAVE))
 758		sched_info_dequeued(rq, p);
 
 
 759
 
 760	p->sched_class->dequeue_task(rq, p, flags);
 761}
 762
 763void activate_task(struct rq *rq, struct task_struct *p, int flags)
 764{
 765	if (task_contributes_to_load(p))
 766		rq->nr_uninterruptible--;
 
 
 767
 768	enqueue_task(rq, p, flags);
 
 
 
 769}
 770
 771void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
 772{
 773	if (task_contributes_to_load(p))
 774		rq->nr_uninterruptible++;
 775
 776	dequeue_task(rq, p, flags);
 777}
 778
 779/*
 780 * __normal_prio - return the priority that is based on the static prio
 781 */
 782static inline int __normal_prio(struct task_struct *p)
 783{
 784	return p->static_prio;
 
 
 
 
 
 
 
 
 
 785}
 786
 787/*
 788 * Calculate the expected normal priority: i.e. priority
 789 * without taking RT-inheritance into account. Might be
 790 * boosted by interactivity modifiers. Changes upon fork,
 791 * setprio syscalls, and whenever the interactivity
 792 * estimator recalculates.
 793 */
 794static inline int normal_prio(struct task_struct *p)
 795{
 796	int prio;
 797
 798	if (task_has_dl_policy(p))
 799		prio = MAX_DL_PRIO-1;
 800	else if (task_has_rt_policy(p))
 801		prio = MAX_RT_PRIO-1 - p->rt_priority;
 802	else
 803		prio = __normal_prio(p);
 804	return prio;
 805}
 806
 807/*
 808 * Calculate the current priority, i.e. the priority
 809 * taken into account by the scheduler. This value might
 810 * be boosted by RT tasks, or might be boosted by
 811 * interactivity modifiers. Will be RT if the task got
 812 * RT-boosted. If not then it returns p->normal_prio.
 813 */
 814static int effective_prio(struct task_struct *p)
 815{
 816	p->normal_prio = normal_prio(p);
 817	/*
 818	 * If we are RT tasks or we were boosted to RT priority,
 819	 * keep the priority unchanged. Otherwise, update priority
 820	 * to the normal priority:
 821	 */
 822	if (!rt_prio(p->prio))
 823		return p->normal_prio;
 824	return p->prio;
 825}
 826
 827/**
 828 * task_curr - is this task currently executing on a CPU?
 829 * @p: the task in question.
 830 *
 831 * Return: 1 if the task is currently executing. 0 otherwise.
 832 */
 833inline int task_curr(const struct task_struct *p)
 834{
 835	return cpu_curr(task_cpu(p)) == p;
 836}
 837
 838/*
 839 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
 840 * use the balance_callback list if you want balancing.
 841 *
 842 * this means any call to check_class_changed() must be followed by a call to
 843 * balance_callback().
 844 */
 845static inline void check_class_changed(struct rq *rq, struct task_struct *p,
 846				       const struct sched_class *prev_class,
 847				       int oldprio)
 848{
 849	if (prev_class != p->sched_class) {
 850		if (prev_class->switched_from)
 851			prev_class->switched_from(rq, p);
 852
 853		p->sched_class->switched_to(rq, p);
 854	} else if (oldprio != p->prio || dl_task(p))
 855		p->sched_class->prio_changed(rq, p, oldprio);
 856}
 857
 858void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
 859{
 860	const struct sched_class *class;
 861
 862	if (p->sched_class == rq->curr->sched_class) {
 863		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
 864	} else {
 865		for_each_class(class) {
 866			if (class == rq->curr->sched_class)
 867				break;
 868			if (class == p->sched_class) {
 869				resched_curr(rq);
 870				break;
 871			}
 872		}
 873	}
 874
 875	/*
 876	 * A queue event has occurred, and we're going to schedule.  In
 877	 * this case, we can save a useless back to back clock update.
 878	 */
 879	if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
 880		rq_clock_skip_update(rq);
 881}
 882
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 883#ifdef CONFIG_SMP
 884
 885static inline bool is_per_cpu_kthread(struct task_struct *p)
 
 
 
 
 
 
 886{
 887	if (!(p->flags & PF_KTHREAD))
 888		return false;
 
 
 889
 890	if (p->nr_cpus_allowed != 1)
 891		return false;
 892
 893	return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 894}
 895
 896/*
 897 * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
 898 * __set_cpus_allowed_ptr() and select_fallback_rq().
 899 */
 900static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
 901{
 902	if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
 
 903		return false;
 904
 905	if (is_per_cpu_kthread(p))
 
 
 
 
 
 
 
 
 
 906		return cpu_online(cpu);
 907
 908	return cpu_active(cpu);
 
 
 
 
 
 909}
 910
 911/*
 912 * This is how migration works:
 913 *
 914 * 1) we invoke migration_cpu_stop() on the target CPU using
 915 *    stop_one_cpu().
 916 * 2) stopper starts to run (implicitly forcing the migrated thread
 917 *    off the CPU)
 918 * 3) it checks whether the migrated task is still in the wrong runqueue.
 919 * 4) if it's in the wrong runqueue then the migration thread removes
 920 *    it and puts it into the right queue.
 921 * 5) stopper completes and stop_one_cpu() returns and the migration
 922 *    is done.
 923 */
 924
 925/*
 926 * move_queued_task - move a queued task to new rq.
 927 *
 928 * Returns (locked) new rq. Old rq's lock is released.
 929 */
 930static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
 931				   struct task_struct *p, int new_cpu)
 932{
 933	lockdep_assert_held(&rq->lock);
 934
 935	p->on_rq = TASK_ON_RQ_MIGRATING;
 936	dequeue_task(rq, p, DEQUEUE_NOCLOCK);
 937	set_task_cpu(p, new_cpu);
 938	rq_unlock(rq, rf);
 939
 940	rq = cpu_rq(new_cpu);
 941
 942	rq_lock(rq, rf);
 943	BUG_ON(task_cpu(p) != new_cpu);
 944	enqueue_task(rq, p, 0);
 945	p->on_rq = TASK_ON_RQ_QUEUED;
 946	check_preempt_curr(rq, p, 0);
 947
 948	return rq;
 949}
 950
 951struct migration_arg {
 952	struct task_struct *task;
 953	int dest_cpu;
 
 
 
 
 
 
 
 
 
 
 
 
 
 954};
 955
 956/*
 957 * Move (not current) task off this CPU, onto the destination CPU. We're doing
 958 * this because either it can't run here any more (set_cpus_allowed()
 959 * away from this CPU, or CPU going down), or because we're
 960 * attempting to rebalance this task on exec (sched_exec).
 961 *
 962 * So we race with normal scheduler movements, but that's OK, as long
 963 * as the task is no longer on this CPU.
 964 */
 965static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
 966				 struct task_struct *p, int dest_cpu)
 967{
 968	/* Affinity changed (again). */
 969	if (!is_cpu_allowed(p, dest_cpu))
 970		return rq;
 971
 972	update_rq_clock(rq);
 973	rq = move_queued_task(rq, rf, p, dest_cpu);
 974
 975	return rq;
 976}
 977
 978/*
 979 * migration_cpu_stop - this will be executed by a highprio stopper thread
 980 * and performs thread migration by bumping thread off CPU then
 981 * 'pushing' onto another runqueue.
 982 */
 983static int migration_cpu_stop(void *data)
 984{
 985	struct migration_arg *arg = data;
 
 986	struct task_struct *p = arg->task;
 987	struct rq *rq = this_rq();
 
 988	struct rq_flags rf;
 989
 990	/*
 991	 * The original target CPU might have gone down and we might
 992	 * be on another CPU but it doesn't matter.
 993	 */
 994	local_irq_disable();
 995	/*
 996	 * We need to explicitly wake pending tasks before running
 997	 * __migrate_task() such that we will not miss enforcing cpus_allowed
 998	 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
 999	 */
1000	sched_ttwu_pending();
1001
1002	raw_spin_lock(&p->pi_lock);
1003	rq_lock(rq, &rf);
 
 
 
 
 
 
 
1004	/*
1005	 * If task_rq(p) != rq, it cannot be migrated here, because we're
1006	 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1007	 * we're holding p->pi_lock.
1008	 */
1009	if (task_rq(p) == rq) {
1010		if (task_on_rq_queued(p))
 
 
 
 
 
 
 
 
 
 
 
 
1011			rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1012		else
1013			p->wake_cpu = arg->dest_cpu;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1014	}
1015	rq_unlock(rq, &rf);
1016	raw_spin_unlock(&p->pi_lock);
 
 
 
 
 
1017
1018	local_irq_enable();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1019	return 0;
1020}
1021
1022/*
1023 * sched_class::set_cpus_allowed must do the below, but is not required to
1024 * actually call this function.
1025 */
1026void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1027{
1028	cpumask_copy(&p->cpus_allowed, new_mask);
1029	p->nr_cpus_allowed = cpumask_weight(new_mask);
 
 
 
 
 
 
 
 
 
 
 
1030}
1031
1032void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
 
1033{
1034	struct rq *rq = task_rq(p);
1035	bool queued, running;
1036
1037	lockdep_assert_held(&p->pi_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1038
1039	queued = task_on_rq_queued(p);
1040	running = task_current(rq, p);
1041
1042	if (queued) {
1043		/*
1044		 * Because __kthread_bind() calls this on blocked tasks without
1045		 * holding rq->lock.
1046		 */
1047		lockdep_assert_held(&rq->lock);
1048		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1049	}
1050	if (running)
1051		put_prev_task(rq, p);
1052
1053	p->sched_class->set_cpus_allowed(p, new_mask);
1054
1055	if (queued)
1056		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1057	if (running)
1058		set_curr_task(rq, p);
1059}
1060
1061/*
1062 * Change a given task's CPU affinity. Migrate the thread to a
1063 * proper CPU and schedule it away if the CPU it's executing on
1064 * is removed from the allowed bitmask.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1065 *
1066 * NOTE: the caller must have a valid reference to the task, the
1067 * task must not exit() & deallocate itself prematurely. The
1068 * call is not atomic; no spinlocks may be held.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1069 */
1070static int __set_cpus_allowed_ptr(struct task_struct *p,
1071				  const struct cpumask *new_mask, bool check)
 
 
1072{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1073	const struct cpumask *cpu_valid_mask = cpu_active_mask;
 
1074	unsigned int dest_cpu;
1075	struct rq_flags rf;
1076	struct rq *rq;
1077	int ret = 0;
1078
1079	rq = task_rq_lock(p, &rf);
1080	update_rq_clock(rq);
1081
1082	if (p->flags & PF_KTHREAD) {
1083		/*
1084		 * Kernel threads are allowed on online && !active CPUs
 
 
 
 
 
 
 
1085		 */
1086		cpu_valid_mask = cpu_online_mask;
1087	}
1088
 
 
 
 
 
1089	/*
1090	 * Must re-check here, to close a race against __kthread_bind(),
1091	 * sched_setaffinity() is not guaranteed to observe the flag.
1092	 */
1093	if (check && (p->flags & PF_NO_SETAFFINITY)) {
1094		ret = -EINVAL;
1095		goto out;
1096	}
1097
1098	if (cpumask_equal(&p->cpus_allowed, new_mask))
1099		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
1100
1101	if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
 
 
 
 
 
 
1102		ret = -EINVAL;
1103		goto out;
1104	}
1105
1106	do_set_cpus_allowed(p, new_mask);
1107
1108	if (p->flags & PF_KTHREAD) {
1109		/*
1110		 * For kernel threads that do indeed end up on online &&
1111		 * !active we want to ensure they are strict per-CPU threads.
1112		 */
1113		WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1114			!cpumask_intersects(new_mask, cpu_active_mask) &&
1115			p->nr_cpus_allowed != 1);
1116	}
1117
1118	/* Can the task run on the task's current CPU? If so, we're done */
1119	if (cpumask_test_cpu(task_cpu(p), new_mask))
1120		goto out;
1121
1122	dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1123	if (task_running(rq, p) || p->state == TASK_WAKING) {
1124		struct migration_arg arg = { p, dest_cpu };
1125		/* Need help from migration thread: drop lock and wait. */
1126		task_rq_unlock(rq, p, &rf);
1127		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1128		tlb_migrate_finish(p->mm);
1129		return 0;
1130	} else if (task_on_rq_queued(p)) {
1131		/*
1132		 * OK, since we're going to drop the lock immediately
1133		 * afterwards anyway.
1134		 */
1135		rq = move_queued_task(rq, &rf, p, dest_cpu);
1136	}
1137out:
1138	task_rq_unlock(rq, p, &rf);
1139
1140	return ret;
1141}
1142
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1143int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1144{
1145	return __set_cpus_allowed_ptr(p, new_mask, false);
 
 
 
 
 
1146}
1147EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1148
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1149void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1150{
1151#ifdef CONFIG_SCHED_DEBUG
 
 
1152	/*
1153	 * We should never call set_task_cpu() on a blocked task,
1154	 * ttwu() will sort out the placement.
1155	 */
1156	WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1157			!p->on_rq);
1158
1159	/*
1160	 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1161	 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1162	 * time relying on p->on_rq.
1163	 */
1164	WARN_ON_ONCE(p->state == TASK_RUNNING &&
1165		     p->sched_class == &fair_sched_class &&
1166		     (p->on_rq && !task_on_rq_migrating(p)));
1167
1168#ifdef CONFIG_LOCKDEP
1169	/*
1170	 * The caller should hold either p->pi_lock or rq->lock, when changing
1171	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1172	 *
1173	 * sched_move_task() holds both and thus holding either pins the cgroup,
1174	 * see task_group().
1175	 *
1176	 * Furthermore, all task_rq users should acquire both locks, see
1177	 * task_rq_lock().
1178	 */
1179	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1180				      lockdep_is_held(&task_rq(p)->lock)));
1181#endif
1182	/*
1183	 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1184	 */
1185	WARN_ON_ONCE(!cpu_online(new_cpu));
 
 
1186#endif
1187
1188	trace_sched_migrate_task(p, new_cpu);
1189
1190	if (task_cpu(p) != new_cpu) {
1191		if (p->sched_class->migrate_task_rq)
1192			p->sched_class->migrate_task_rq(p);
1193		p->se.nr_migrations++;
 
 
1194		perf_event_task_migrate(p);
1195	}
1196
1197	__set_task_cpu(p, new_cpu);
1198}
1199
 
1200static void __migrate_swap_task(struct task_struct *p, int cpu)
1201{
1202	if (task_on_rq_queued(p)) {
1203		struct rq *src_rq, *dst_rq;
1204		struct rq_flags srf, drf;
1205
1206		src_rq = task_rq(p);
1207		dst_rq = cpu_rq(cpu);
1208
1209		rq_pin_lock(src_rq, &srf);
1210		rq_pin_lock(dst_rq, &drf);
1211
1212		p->on_rq = TASK_ON_RQ_MIGRATING;
1213		deactivate_task(src_rq, p, 0);
1214		set_task_cpu(p, cpu);
1215		activate_task(dst_rq, p, 0);
1216		p->on_rq = TASK_ON_RQ_QUEUED;
1217		check_preempt_curr(dst_rq, p, 0);
1218
1219		rq_unpin_lock(dst_rq, &drf);
1220		rq_unpin_lock(src_rq, &srf);
1221
1222	} else {
1223		/*
1224		 * Task isn't running anymore; make it appear like we migrated
1225		 * it before it went to sleep. This means on wakeup we make the
1226		 * previous CPU our target instead of where it really is.
1227		 */
1228		p->wake_cpu = cpu;
1229	}
1230}
1231
1232struct migration_swap_arg {
1233	struct task_struct *src_task, *dst_task;
1234	int src_cpu, dst_cpu;
1235};
1236
1237static int migrate_swap_stop(void *data)
1238{
1239	struct migration_swap_arg *arg = data;
1240	struct rq *src_rq, *dst_rq;
1241	int ret = -EAGAIN;
1242
1243	if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1244		return -EAGAIN;
1245
1246	src_rq = cpu_rq(arg->src_cpu);
1247	dst_rq = cpu_rq(arg->dst_cpu);
1248
1249	double_raw_lock(&arg->src_task->pi_lock,
1250			&arg->dst_task->pi_lock);
1251	double_rq_lock(src_rq, dst_rq);
1252
1253	if (task_cpu(arg->dst_task) != arg->dst_cpu)
1254		goto unlock;
1255
1256	if (task_cpu(arg->src_task) != arg->src_cpu)
1257		goto unlock;
1258
1259	if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1260		goto unlock;
1261
1262	if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1263		goto unlock;
1264
1265	__migrate_swap_task(arg->src_task, arg->dst_cpu);
1266	__migrate_swap_task(arg->dst_task, arg->src_cpu);
1267
1268	ret = 0;
1269
1270unlock:
1271	double_rq_unlock(src_rq, dst_rq);
1272	raw_spin_unlock(&arg->dst_task->pi_lock);
1273	raw_spin_unlock(&arg->src_task->pi_lock);
1274
1275	return ret;
1276}
1277
1278/*
1279 * Cross migrate two tasks
1280 */
1281int migrate_swap(struct task_struct *cur, struct task_struct *p)
 
1282{
1283	struct migration_swap_arg arg;
1284	int ret = -EINVAL;
1285
1286	arg = (struct migration_swap_arg){
1287		.src_task = cur,
1288		.src_cpu = task_cpu(cur),
1289		.dst_task = p,
1290		.dst_cpu = task_cpu(p),
1291	};
1292
1293	if (arg.src_cpu == arg.dst_cpu)
1294		goto out;
1295
1296	/*
1297	 * These three tests are all lockless; this is OK since all of them
1298	 * will be re-checked with proper locks held further down the line.
1299	 */
1300	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1301		goto out;
1302
1303	if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1304		goto out;
1305
1306	if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1307		goto out;
1308
1309	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1310	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1311
1312out:
1313	return ret;
1314}
1315
1316/*
1317 * wait_task_inactive - wait for a thread to unschedule.
1318 *
1319 * If @match_state is nonzero, it's the @p->state value just checked and
1320 * not expected to change.  If it changes, i.e. @p might have woken up,
1321 * then return zero.  When we succeed in waiting for @p to be off its CPU,
1322 * we return a positive number (its total switch count).  If a second call
1323 * a short while later returns the same number, the caller can be sure that
1324 * @p has remained unscheduled the whole time.
1325 *
1326 * The caller must ensure that the task *will* unschedule sometime soon,
1327 * else this function might spin for a *long* time. This function can't
1328 * be called with interrupts off, or it may introduce deadlock with
1329 * smp_call_function() if an IPI is sent by the same process we are
1330 * waiting to become inactive.
1331 */
1332unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1333{
1334	int running, queued;
1335	struct rq_flags rf;
1336	unsigned long ncsw;
1337	struct rq *rq;
1338
1339	for (;;) {
1340		/*
1341		 * We do the initial early heuristics without holding
1342		 * any task-queue locks at all. We'll only try to get
1343		 * the runqueue lock when things look like they will
1344		 * work out!
1345		 */
1346		rq = task_rq(p);
1347
1348		/*
1349		 * If the task is actively running on another CPU
1350		 * still, just relax and busy-wait without holding
1351		 * any locks.
1352		 *
1353		 * NOTE! Since we don't hold any locks, it's not
1354		 * even sure that "rq" stays as the right runqueue!
1355		 * But we don't care, since "task_running()" will
1356		 * return false if the runqueue has changed and p
1357		 * is actually now running somewhere else!
1358		 */
1359		while (task_running(rq, p)) {
1360			if (match_state && unlikely(p->state != match_state))
1361				return 0;
1362			cpu_relax();
1363		}
1364
1365		/*
1366		 * Ok, time to look more closely! We need the rq
1367		 * lock now, to be *sure*. If we're wrong, we'll
1368		 * just go back and repeat.
1369		 */
1370		rq = task_rq_lock(p, &rf);
1371		trace_sched_wait_task(p);
1372		running = task_running(rq, p);
1373		queued = task_on_rq_queued(p);
1374		ncsw = 0;
1375		if (!match_state || p->state == match_state)
1376			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1377		task_rq_unlock(rq, p, &rf);
1378
1379		/*
1380		 * If it changed from the expected state, bail out now.
1381		 */
1382		if (unlikely(!ncsw))
1383			break;
1384
1385		/*
1386		 * Was it really running after all now that we
1387		 * checked with the proper locks actually held?
1388		 *
1389		 * Oops. Go back and try again..
1390		 */
1391		if (unlikely(running)) {
1392			cpu_relax();
1393			continue;
1394		}
1395
1396		/*
1397		 * It's not enough that it's not actively running,
1398		 * it must be off the runqueue _entirely_, and not
1399		 * preempted!
1400		 *
1401		 * So if it was still runnable (but just not actively
1402		 * running right now), it's preempted, and we should
1403		 * yield - it could be a while.
1404		 */
1405		if (unlikely(queued)) {
1406			ktime_t to = NSEC_PER_SEC / HZ;
1407
1408			set_current_state(TASK_UNINTERRUPTIBLE);
1409			schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1410			continue;
1411		}
1412
1413		/*
1414		 * Ahh, all good. It wasn't running, and it wasn't
1415		 * runnable, which means that it will never become
1416		 * running in the future either. We're all done!
1417		 */
1418		break;
1419	}
1420
1421	return ncsw;
1422}
1423
1424/***
1425 * kick_process - kick a running thread to enter/exit the kernel
1426 * @p: the to-be-kicked thread
1427 *
1428 * Cause a process which is running on another CPU to enter
1429 * kernel-mode, without any delay. (to get signals handled.)
1430 *
1431 * NOTE: this function doesn't have to take the runqueue lock,
1432 * because all it wants to ensure is that the remote task enters
1433 * the kernel. If the IPI races and the task has been migrated
1434 * to another CPU then no harm is done and the purpose has been
1435 * achieved as well.
1436 */
1437void kick_process(struct task_struct *p)
1438{
1439	int cpu;
 
1440
1441	preempt_disable();
1442	cpu = task_cpu(p);
1443	if ((cpu != smp_processor_id()) && task_curr(p))
1444		smp_send_reschedule(cpu);
1445	preempt_enable();
1446}
1447EXPORT_SYMBOL_GPL(kick_process);
1448
1449/*
1450 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1451 *
1452 * A few notes on cpu_active vs cpu_online:
1453 *
1454 *  - cpu_active must be a subset of cpu_online
1455 *
1456 *  - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1457 *    see __set_cpus_allowed_ptr(). At this point the newly online
1458 *    CPU isn't yet part of the sched domains, and balancing will not
1459 *    see it.
1460 *
1461 *  - on CPU-down we clear cpu_active() to mask the sched domains and
1462 *    avoid the load balancer to place new tasks on the to be removed
1463 *    CPU. Existing tasks will remain running there and will be taken
1464 *    off.
1465 *
1466 * This means that fallback selection must not select !active CPUs.
1467 * And can assume that any active CPU must be online. Conversely
1468 * select_task_rq() below may allow selection of !active CPUs in order
1469 * to satisfy the above rules.
1470 */
1471static int select_fallback_rq(int cpu, struct task_struct *p)
1472{
1473	int nid = cpu_to_node(cpu);
1474	const struct cpumask *nodemask = NULL;
1475	enum { cpuset, possible, fail } state = cpuset;
1476	int dest_cpu;
1477
1478	/*
1479	 * If the node that the CPU is on has been offlined, cpu_to_node()
1480	 * will return -1. There is no CPU on the node, and we should
1481	 * select the CPU on the other node.
1482	 */
1483	if (nid != -1) {
1484		nodemask = cpumask_of_node(nid);
1485
1486		/* Look for allowed, online CPU in same node. */
1487		for_each_cpu(dest_cpu, nodemask) {
1488			if (!cpu_active(dest_cpu))
1489				continue;
1490			if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1491				return dest_cpu;
1492		}
1493	}
1494
1495	for (;;) {
1496		/* Any allowed, online CPU? */
1497		for_each_cpu(dest_cpu, &p->cpus_allowed) {
1498			if (!is_cpu_allowed(p, dest_cpu))
1499				continue;
1500
1501			goto out;
1502		}
1503
1504		/* No more Mr. Nice Guy. */
1505		switch (state) {
1506		case cpuset:
1507			if (IS_ENABLED(CONFIG_CPUSETS)) {
1508				cpuset_cpus_allowed_fallback(p);
1509				state = possible;
1510				break;
1511			}
1512			/* Fall-through */
1513		case possible:
1514			do_set_cpus_allowed(p, cpu_possible_mask);
 
 
 
 
 
 
1515			state = fail;
1516			break;
1517
1518		case fail:
1519			BUG();
1520			break;
1521		}
1522	}
1523
1524out:
1525	if (state != cpuset) {
1526		/*
1527		 * Don't tell them about moving exiting tasks or
1528		 * kernel threads (both mm NULL), since they never
1529		 * leave kernel.
1530		 */
1531		if (p->mm && printk_ratelimit()) {
1532			printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1533					task_pid_nr(p), p->comm, cpu);
1534		}
1535	}
1536
1537	return dest_cpu;
1538}
1539
1540/*
1541 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1542 */
1543static inline
1544int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1545{
1546	lockdep_assert_held(&p->pi_lock);
1547
1548	if (p->nr_cpus_allowed > 1)
1549		cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1550	else
1551		cpu = cpumask_any(&p->cpus_allowed);
1552
1553	/*
1554	 * In order not to call set_task_cpu() on a blocking task we need
1555	 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1556	 * CPU.
1557	 *
1558	 * Since this is common to all placement strategies, this lives here.
1559	 *
1560	 * [ this allows ->select_task() to simply return task_cpu(p) and
1561	 *   not worry about this generic constraint ]
1562	 */
1563	if (unlikely(!is_cpu_allowed(p, cpu)))
1564		cpu = select_fallback_rq(task_cpu(p), p);
1565
1566	return cpu;
1567}
1568
1569static void update_avg(u64 *avg, u64 sample)
1570{
1571	s64 diff = sample - *avg;
1572	*avg += diff >> 3;
1573}
1574
1575void sched_set_stop_task(int cpu, struct task_struct *stop)
1576{
 
1577	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1578	struct task_struct *old_stop = cpu_rq(cpu)->stop;
1579
1580	if (stop) {
1581		/*
1582		 * Make it appear like a SCHED_FIFO task, its something
1583		 * userspace knows about and won't get confused about.
1584		 *
1585		 * Also, it will make PI more or less work without too
1586		 * much confusion -- but then, stop work should not
1587		 * rely on PI working anyway.
1588		 */
1589		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
1590
1591		stop->sched_class = &stop_sched_class;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1592	}
1593
1594	cpu_rq(cpu)->stop = stop;
1595
1596	if (old_stop) {
1597		/*
1598		 * Reset it back to a normal scheduling class so that
1599		 * it can die in pieces.
1600		 */
1601		old_stop->sched_class = &rt_sched_class;
1602	}
1603}
1604
1605#else
1606
1607static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1608					 const struct cpumask *new_mask, bool check)
1609{
1610	return set_cpus_allowed_ptr(p, new_mask);
1611}
1612
1613#endif /* CONFIG_SMP */
 
 
 
 
 
 
 
 
 
 
 
 
1614
1615static void
1616ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1617{
1618	struct rq *rq;
1619
1620	if (!schedstat_enabled())
1621		return;
1622
1623	rq = this_rq();
1624
1625#ifdef CONFIG_SMP
1626	if (cpu == rq->cpu) {
1627		__schedstat_inc(rq->ttwu_local);
1628		__schedstat_inc(p->se.statistics.nr_wakeups_local);
1629	} else {
1630		struct sched_domain *sd;
1631
1632		__schedstat_inc(p->se.statistics.nr_wakeups_remote);
1633		rcu_read_lock();
 
1634		for_each_domain(rq->cpu, sd) {
1635			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1636				__schedstat_inc(sd->ttwu_wake_remote);
1637				break;
1638			}
1639		}
1640		rcu_read_unlock();
1641	}
1642
1643	if (wake_flags & WF_MIGRATED)
1644		__schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1645#endif /* CONFIG_SMP */
1646
1647	__schedstat_inc(rq->ttwu_count);
1648	__schedstat_inc(p->se.statistics.nr_wakeups);
1649
1650	if (wake_flags & WF_SYNC)
1651		__schedstat_inc(p->se.statistics.nr_wakeups_sync);
1652}
1653
1654static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1655{
1656	activate_task(rq, p, en_flags);
1657	p->on_rq = TASK_ON_RQ_QUEUED;
1658
1659	/* If a worker is waking up, notify the workqueue: */
1660	if (p->flags & PF_WQ_WORKER)
1661		wq_worker_waking_up(p, cpu_of(rq));
1662}
1663
1664/*
1665 * Mark the task runnable and perform wakeup-preemption.
1666 */
1667static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1668			   struct rq_flags *rf)
1669{
1670	check_preempt_curr(rq, p, wake_flags);
1671	p->state = TASK_RUNNING;
1672	trace_sched_wakeup(p);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1673
1674#ifdef CONFIG_SMP
1675	if (p->sched_class->task_woken) {
1676		/*
1677		 * Our task @p is fully woken up and running; so its safe to
1678		 * drop the rq->lock, hereafter rq is only used for statistics.
1679		 */
1680		rq_unpin_lock(rq, rf);
1681		p->sched_class->task_woken(rq, p);
1682		rq_repin_lock(rq, rf);
1683	}
1684
1685	if (rq->idle_stamp) {
1686		u64 delta = rq_clock(rq) - rq->idle_stamp;
1687		u64 max = 2*rq->max_idle_balance_cost;
1688
1689		update_avg(&rq->avg_idle, delta);
1690
1691		if (rq->avg_idle > max)
1692			rq->avg_idle = max;
1693
1694		rq->idle_stamp = 0;
1695	}
1696#endif
1697}
1698
1699static void
1700ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1701		 struct rq_flags *rf)
1702{
1703	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1704
1705	lockdep_assert_held(&rq->lock);
1706
1707#ifdef CONFIG_SMP
1708	if (p->sched_contributes_to_load)
1709		rq->nr_uninterruptible--;
1710
1711	if (wake_flags & WF_MIGRATED)
1712		en_flags |= ENQUEUE_MIGRATED;
1713#endif
1714
1715	ttwu_activate(rq, p, en_flags);
1716	ttwu_do_wakeup(rq, p, wake_flags, rf);
1717}
1718
1719/*
1720 * Called in case the task @p isn't fully descheduled from its runqueue,
1721 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1722 * since all we need to do is flip p->state to TASK_RUNNING, since
1723 * the task is still ->on_rq.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1724 */
1725static int ttwu_remote(struct task_struct *p, int wake_flags)
1726{
1727	struct rq_flags rf;
1728	struct rq *rq;
1729	int ret = 0;
1730
1731	rq = __task_rq_lock(p, &rf);
1732	if (task_on_rq_queued(p)) {
1733		/* check_preempt_curr() may use rq clock */
1734		update_rq_clock(rq);
1735		ttwu_do_wakeup(rq, p, wake_flags, &rf);
 
 
 
 
 
 
1736		ret = 1;
1737	}
1738	__task_rq_unlock(rq, &rf);
1739
1740	return ret;
1741}
1742
1743#ifdef CONFIG_SMP
1744void sched_ttwu_pending(void)
1745{
 
1746	struct rq *rq = this_rq();
1747	struct llist_node *llist = llist_del_all(&rq->wake_list);
1748	struct task_struct *p, *t;
1749	struct rq_flags rf;
1750
1751	if (!llist)
1752		return;
1753
1754	rq_lock_irqsave(rq, &rf);
1755	update_rq_clock(rq);
1756
1757	llist_for_each_entry_safe(p, t, llist, wake_entry)
1758		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1759
1760	rq_unlock_irqrestore(rq, &rf);
1761}
1762
1763void scheduler_ipi(void)
1764{
1765	/*
1766	 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1767	 * TIF_NEED_RESCHED remotely (for the first time) will also send
1768	 * this IPI.
1769	 */
1770	preempt_fold_need_resched();
1771
1772	if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1773		return;
1774
1775	/*
1776	 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1777	 * traditionally all their work was done from the interrupt return
1778	 * path. Now that we actually do some work, we need to make sure
1779	 * we do call them.
1780	 *
1781	 * Some archs already do call them, luckily irq_enter/exit nest
1782	 * properly.
1783	 *
1784	 * Arguably we should visit all archs and update all handlers,
1785	 * however a fair share of IPIs are still resched only so this would
1786	 * somewhat pessimize the simple resched case.
1787	 */
1788	irq_enter();
1789	sched_ttwu_pending();
 
1790
1791	/*
1792	 * Check if someone kicked us for doing the nohz idle load balance.
1793	 */
1794	if (unlikely(got_nohz_idle_kick())) {
1795		this_rq()->idle_balance = 1;
1796		raise_softirq_irqoff(SCHED_SOFTIRQ);
 
 
 
 
 
1797	}
1798	irq_exit();
 
1799}
1800
1801static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
 
 
 
 
 
 
1802{
1803	struct rq *rq = cpu_rq(cpu);
1804
1805	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1806
1807	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1808		if (!set_nr_if_polling(rq->idle))
1809			smp_send_reschedule(cpu);
1810		else
1811			trace_sched_wake_idle_without_ipi(cpu);
1812	}
1813}
1814
1815void wake_up_if_idle(int cpu)
1816{
1817	struct rq *rq = cpu_rq(cpu);
1818	struct rq_flags rf;
1819
1820	rcu_read_lock();
1821
1822	if (!is_idle_task(rcu_dereference(rq->curr)))
1823		goto out;
1824
1825	if (set_nr_if_polling(rq->idle)) {
1826		trace_sched_wake_idle_without_ipi(cpu);
1827	} else {
1828		rq_lock_irqsave(rq, &rf);
1829		if (is_idle_task(rq->curr))
1830			smp_send_reschedule(cpu);
1831		/* Else CPU is not idle, do nothing here: */
1832		rq_unlock_irqrestore(rq, &rf);
1833	}
1834
1835out:
1836	rcu_read_unlock();
1837}
1838
1839bool cpus_share_cache(int this_cpu, int that_cpu)
1840{
 
 
 
1841	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1842}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1843#endif /* CONFIG_SMP */
1844
1845static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1846{
1847	struct rq *rq = cpu_rq(cpu);
1848	struct rq_flags rf;
1849
1850#if defined(CONFIG_SMP)
1851	if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1852		sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1853		ttwu_queue_remote(p, cpu, wake_flags);
1854		return;
1855	}
1856#endif
1857
1858	rq_lock(rq, &rf);
1859	update_rq_clock(rq);
1860	ttwu_do_activate(rq, p, wake_flags, &rf);
1861	rq_unlock(rq, &rf);
1862}
1863
1864/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1865 * Notes on Program-Order guarantees on SMP systems.
1866 *
1867 *  MIGRATION
1868 *
1869 * The basic program-order guarantee on SMP systems is that when a task [t]
1870 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1871 * execution on its new CPU [c1].
1872 *
1873 * For migration (of runnable tasks) this is provided by the following means:
1874 *
1875 *  A) UNLOCK of the rq(c0)->lock scheduling out task t
1876 *  B) migration for t is required to synchronize *both* rq(c0)->lock and
1877 *     rq(c1)->lock (if not at the same time, then in that order).
1878 *  C) LOCK of the rq(c1)->lock scheduling in task
1879 *
1880 * Transitivity guarantees that B happens after A and C after B.
1881 * Note: we only require RCpc transitivity.
1882 * Note: the CPU doing B need not be c0 or c1
1883 *
1884 * Example:
1885 *
1886 *   CPU0            CPU1            CPU2
1887 *
1888 *   LOCK rq(0)->lock
1889 *   sched-out X
1890 *   sched-in Y
1891 *   UNLOCK rq(0)->lock
1892 *
1893 *                                   LOCK rq(0)->lock // orders against CPU0
1894 *                                   dequeue X
1895 *                                   UNLOCK rq(0)->lock
1896 *
1897 *                                   LOCK rq(1)->lock
1898 *                                   enqueue X
1899 *                                   UNLOCK rq(1)->lock
1900 *
1901 *                   LOCK rq(1)->lock // orders against CPU2
1902 *                   sched-out Z
1903 *                   sched-in X
1904 *                   UNLOCK rq(1)->lock
1905 *
1906 *
1907 *  BLOCKING -- aka. SLEEP + WAKEUP
1908 *
1909 * For blocking we (obviously) need to provide the same guarantee as for
1910 * migration. However the means are completely different as there is no lock
1911 * chain to provide order. Instead we do:
1912 *
1913 *   1) smp_store_release(X->on_cpu, 0)
1914 *   2) smp_cond_load_acquire(!X->on_cpu)
1915 *
1916 * Example:
1917 *
1918 *   CPU0 (schedule)  CPU1 (try_to_wake_up) CPU2 (schedule)
1919 *
1920 *   LOCK rq(0)->lock LOCK X->pi_lock
1921 *   dequeue X
1922 *   sched-out X
1923 *   smp_store_release(X->on_cpu, 0);
1924 *
1925 *                    smp_cond_load_acquire(&X->on_cpu, !VAL);
1926 *                    X->state = WAKING
1927 *                    set_task_cpu(X,2)
1928 *
1929 *                    LOCK rq(2)->lock
1930 *                    enqueue X
1931 *                    X->state = RUNNING
1932 *                    UNLOCK rq(2)->lock
1933 *
1934 *                                          LOCK rq(2)->lock // orders against CPU1
1935 *                                          sched-out Z
1936 *                                          sched-in X
1937 *                                          UNLOCK rq(2)->lock
1938 *
1939 *                    UNLOCK X->pi_lock
1940 *   UNLOCK rq(0)->lock
1941 *
1942 *
1943 * However; for wakeups there is a second guarantee we must provide, namely we
1944 * must observe the state that lead to our wakeup. That is, not only must our
1945 * task observe its own prior state, it must also observe the stores prior to
1946 * its wakeup.
1947 *
1948 * This means that any means of doing remote wakeups must order the CPU doing
1949 * the wakeup against the CPU the task is going to end up running on. This,
1950 * however, is already required for the regular Program-Order guarantee above,
1951 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1952 *
1953 */
1954
1955/**
1956 * try_to_wake_up - wake up a thread
1957 * @p: the thread to be awakened
1958 * @state: the mask of task states that can be woken
1959 * @wake_flags: wake modifier flags (WF_*)
1960 *
1961 * If (@state & @p->state) @p->state = TASK_RUNNING.
 
 
1962 *
1963 * If the task was not queued/runnable, also place it back on a runqueue.
1964 *
1965 * Atomic against schedule() which would dequeue a task, also see
1966 * set_current_state().
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1967 *
1968 * Return: %true if @p->state changes (an actual wakeup was done),
1969 *	   %false otherwise.
1970 */
1971static int
1972try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1973{
1974	unsigned long flags;
1975	int cpu, success = 0;
1976
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1977	/*
1978	 * If we are going to wake up a thread waiting for CONDITION we
1979	 * need to ensure that CONDITION=1 done by the caller can not be
1980	 * reordered with p->state check below. This pairs with mb() in
1981	 * set_current_state() the waiting thread does.
1982	 */
1983	raw_spin_lock_irqsave(&p->pi_lock, flags);
1984	smp_mb__after_spinlock();
1985	if (!(p->state & state))
1986		goto out;
1987
1988	trace_sched_waking(p);
1989
1990	/* We're going to change ->state: */
1991	success = 1;
1992	cpu = task_cpu(p);
1993
1994	/*
1995	 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1996	 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1997	 * in smp_cond_load_acquire() below.
1998	 *
1999	 * sched_ttwu_pending()                 try_to_wake_up()
2000	 *   [S] p->on_rq = 1;                  [L] P->state
2001	 *       UNLOCK rq->lock  -----.
2002	 *                              \
2003	 *				 +---   RMB
2004	 * schedule()                   /
2005	 *       LOCK rq->lock    -----'
2006	 *       UNLOCK rq->lock
2007	 *
2008	 * [task p]
2009	 *   [S] p->state = UNINTERRUPTIBLE     [L] p->on_rq
2010	 *
2011	 * Pairs with the UNLOCK+LOCK on rq->lock from the
2012	 * last wakeup of our task and the schedule that got our task
2013	 * current.
2014	 */
2015	smp_rmb();
2016	if (p->on_rq && ttwu_remote(p, wake_flags))
2017		goto stat;
2018
2019#ifdef CONFIG_SMP
2020	/*
2021	 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2022	 * possible to, falsely, observe p->on_cpu == 0.
2023	 *
2024	 * One must be running (->on_cpu == 1) in order to remove oneself
2025	 * from the runqueue.
2026	 *
2027	 *  [S] ->on_cpu = 1;	[L] ->on_rq
2028	 *      UNLOCK rq->lock
2029	 *			RMB
2030	 *      LOCK   rq->lock
2031	 *  [S] ->on_rq = 0;    [L] ->on_cpu
2032	 *
2033	 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2034	 * from the consecutive calls to schedule(); the first switching to our
2035	 * task, the second putting it to sleep.
2036	 */
2037	smp_rmb();
2038
2039	/*
2040	 * If the owning (remote) CPU is still in the middle of schedule() with
2041	 * this task as prev, wait until its done referencing the task.
2042	 *
2043	 * Pairs with the smp_store_release() in finish_task().
2044	 *
2045	 * This ensures that tasks getting woken will be fully ordered against
2046	 * their previous state and preserve Program Order.
2047	 */
2048	smp_cond_load_acquire(&p->on_cpu, !VAL);
2049
2050	p->sched_contributes_to_load = !!task_contributes_to_load(p);
2051	p->state = TASK_WAKING;
2052
2053	if (p->in_iowait) {
2054		delayacct_blkio_end(p);
2055		atomic_dec(&task_rq(p)->nr_iowait);
2056	}
 
 
 
2057
2058	cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2059	if (task_cpu(p) != cpu) {
2060		wake_flags |= WF_MIGRATED;
2061		set_task_cpu(p, cpu);
2062	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2063
2064#else /* CONFIG_SMP */
 
 
 
 
 
 
 
 
 
2065
2066	if (p->in_iowait) {
2067		delayacct_blkio_end(p);
2068		atomic_dec(&task_rq(p)->nr_iowait);
2069	}
 
 
2070
 
 
 
 
 
 
2071#endif /* CONFIG_SMP */
2072
2073	ttwu_queue(p, cpu, wake_flags);
2074stat:
2075	ttwu_stat(p, cpu, wake_flags);
2076out:
2077	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
 
2078
2079	return success;
2080}
2081
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2082/**
2083 * try_to_wake_up_local - try to wake up a local task with rq lock held
2084 * @p: the thread to be awakened
2085 * @rf: request-queue flags for pinning
 
 
 
 
 
 
2086 *
2087 * Put @p on the run-queue if it's not already there. The caller must
2088 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2089 * the current task.
2090 */
2091static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2092{
2093	struct rq *rq = task_rq(p);
2094
2095	if (WARN_ON_ONCE(rq != this_rq()) ||
2096	    WARN_ON_ONCE(p == current))
2097		return;
2098
2099	lockdep_assert_held(&rq->lock);
2100
2101	if (!raw_spin_trylock(&p->pi_lock)) {
2102		/*
2103		 * This is OK, because current is on_cpu, which avoids it being
2104		 * picked for load-balance and preemption/IRQs are still
2105		 * disabled avoiding further scheduler activity on it and we've
2106		 * not yet picked a replacement task.
2107		 */
2108		rq_unlock(rq, rf);
2109		raw_spin_lock(&p->pi_lock);
2110		rq_relock(rq, rf);
2111	}
2112
2113	if (!(p->state & TASK_NORMAL))
2114		goto out;
 
 
 
 
 
 
 
 
 
2115
2116	trace_sched_waking(p);
 
2117
2118	if (!task_on_rq_queued(p)) {
2119		if (p->in_iowait) {
2120			delayacct_blkio_end(p);
2121			atomic_dec(&rq->nr_iowait);
2122		}
2123		ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2124	}
2125
2126	ttwu_do_wakeup(rq, p, 0, rf);
2127	ttwu_stat(p, smp_processor_id(), 0);
2128out:
2129	raw_spin_unlock(&p->pi_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2130}
2131
2132/**
2133 * wake_up_process - Wake up a specific process
2134 * @p: The process to be woken up.
2135 *
2136 * Attempt to wake up the nominated process and move it to the set of runnable
2137 * processes.
2138 *
2139 * Return: 1 if the process was woken up, 0 if it was already running.
2140 *
2141 * It may be assumed that this function implies a write memory barrier before
2142 * changing the task state if and only if any tasks are woken up.
2143 */
2144int wake_up_process(struct task_struct *p)
2145{
2146	return try_to_wake_up(p, TASK_NORMAL, 0);
2147}
2148EXPORT_SYMBOL(wake_up_process);
2149
2150int wake_up_state(struct task_struct *p, unsigned int state)
2151{
2152	return try_to_wake_up(p, state, 0);
2153}
2154
2155/*
2156 * Perform scheduler related setup for a newly forked process p.
2157 * p is forked by current.
2158 *
2159 * __sched_fork() is basic setup used by init_idle() too:
2160 */
2161static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2162{
2163	p->on_rq			= 0;
2164
2165	p->se.on_rq			= 0;
2166	p->se.exec_start		= 0;
2167	p->se.sum_exec_runtime		= 0;
2168	p->se.prev_sum_exec_runtime	= 0;
2169	p->se.nr_migrations		= 0;
2170	p->se.vruntime			= 0;
 
 
2171	INIT_LIST_HEAD(&p->se.group_node);
2172
2173#ifdef CONFIG_FAIR_GROUP_SCHED
2174	p->se.cfs_rq			= NULL;
2175#endif
2176
2177#ifdef CONFIG_SCHEDSTATS
2178	/* Even if schedstat is disabled, there should not be garbage */
2179	memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2180#endif
2181
2182	RB_CLEAR_NODE(&p->dl.rb_node);
2183	init_dl_task_timer(&p->dl);
2184	init_dl_inactive_task_timer(&p->dl);
2185	__dl_clear_params(p);
2186
2187	INIT_LIST_HEAD(&p->rt.run_list);
2188	p->rt.timeout		= 0;
2189	p->rt.time_slice	= sched_rr_timeslice;
2190	p->rt.on_rq		= 0;
2191	p->rt.on_list		= 0;
2192
2193#ifdef CONFIG_PREEMPT_NOTIFIERS
2194	INIT_HLIST_HEAD(&p->preempt_notifiers);
2195#endif
2196
2197#ifdef CONFIG_NUMA_BALANCING
2198	if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2199		p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2200		p->mm->numa_scan_seq = 0;
2201	}
2202
2203	if (clone_flags & CLONE_VM)
2204		p->numa_preferred_nid = current->numa_preferred_nid;
2205	else
2206		p->numa_preferred_nid = -1;
2207
2208	p->node_stamp = 0ULL;
2209	p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2210	p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2211	p->numa_work.next = &p->numa_work;
2212	p->numa_faults = NULL;
2213	p->last_task_numa_placement = 0;
2214	p->last_sum_exec_runtime = 0;
2215
2216	p->numa_group = NULL;
2217#endif /* CONFIG_NUMA_BALANCING */
2218}
2219
2220DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2221
2222#ifdef CONFIG_NUMA_BALANCING
2223
2224void set_numabalancing_state(bool enabled)
 
 
2225{
2226	if (enabled)
2227		static_branch_enable(&sched_numa_balancing);
2228	else
2229		static_branch_disable(&sched_numa_balancing);
2230}
2231
 
 
 
 
 
 
 
 
 
2232#ifdef CONFIG_PROC_SYSCTL
2233int sysctl_numa_balancing(struct ctl_table *table, int write,
2234			 void __user *buffer, size_t *lenp, loff_t *ppos)
 
 
 
 
 
 
 
 
 
 
 
2235{
2236	struct ctl_table t;
2237	int err;
2238	int state = static_branch_likely(&sched_numa_balancing);
2239
2240	if (write && !capable(CAP_SYS_ADMIN))
2241		return -EPERM;
2242
2243	t = *table;
2244	t.data = &state;
2245	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2246	if (err < 0)
2247		return err;
2248	if (write)
2249		set_numabalancing_state(state);
 
 
 
 
 
2250	return err;
2251}
2252#endif
2253#endif
2254
2255#ifdef CONFIG_SCHEDSTATS
2256
2257DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2258static bool __initdata __sched_schedstats = false;
2259
2260static void set_schedstats(bool enabled)
2261{
2262	if (enabled)
2263		static_branch_enable(&sched_schedstats);
2264	else
2265		static_branch_disable(&sched_schedstats);
2266}
2267
2268void force_schedstat_enabled(void)
2269{
2270	if (!schedstat_enabled()) {
2271		pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2272		static_branch_enable(&sched_schedstats);
2273	}
2274}
2275
2276static int __init setup_schedstats(char *str)
2277{
2278	int ret = 0;
2279	if (!str)
2280		goto out;
2281
2282	/*
2283	 * This code is called before jump labels have been set up, so we can't
2284	 * change the static branch directly just yet.  Instead set a temporary
2285	 * variable so init_schedstats() can do it later.
2286	 */
2287	if (!strcmp(str, "enable")) {
2288		__sched_schedstats = true;
2289		ret = 1;
2290	} else if (!strcmp(str, "disable")) {
2291		__sched_schedstats = false;
2292		ret = 1;
2293	}
2294out:
2295	if (!ret)
2296		pr_warn("Unable to parse schedstats=\n");
2297
2298	return ret;
2299}
2300__setup("schedstats=", setup_schedstats);
2301
2302static void __init init_schedstats(void)
2303{
2304	set_schedstats(__sched_schedstats);
2305}
2306
2307#ifdef CONFIG_PROC_SYSCTL
2308int sysctl_schedstats(struct ctl_table *table, int write,
2309			 void __user *buffer, size_t *lenp, loff_t *ppos)
2310{
2311	struct ctl_table t;
2312	int err;
2313	int state = static_branch_likely(&sched_schedstats);
2314
2315	if (write && !capable(CAP_SYS_ADMIN))
2316		return -EPERM;
2317
2318	t = *table;
2319	t.data = &state;
2320	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2321	if (err < 0)
2322		return err;
2323	if (write)
2324		set_schedstats(state);
2325	return err;
2326}
2327#endif /* CONFIG_PROC_SYSCTL */
2328#else  /* !CONFIG_SCHEDSTATS */
2329static inline void init_schedstats(void) {}
2330#endif /* CONFIG_SCHEDSTATS */
2331
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2332/*
2333 * fork()/clone()-time setup:
2334 */
2335int sched_fork(unsigned long clone_flags, struct task_struct *p)
2336{
2337	unsigned long flags;
2338	int cpu = get_cpu();
2339
2340	__sched_fork(clone_flags, p);
2341	/*
2342	 * We mark the process as NEW here. This guarantees that
2343	 * nobody will actually run it, and a signal or other external
2344	 * event cannot wake it up and insert it on the runqueue either.
2345	 */
2346	p->state = TASK_NEW;
2347
2348	/*
2349	 * Make sure we do not leak PI boosting priority to the child.
2350	 */
2351	p->prio = current->normal_prio;
2352
 
 
2353	/*
2354	 * Revert to default priority/policy on fork if requested.
2355	 */
2356	if (unlikely(p->sched_reset_on_fork)) {
2357		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2358			p->policy = SCHED_NORMAL;
2359			p->static_prio = NICE_TO_PRIO(0);
2360			p->rt_priority = 0;
2361		} else if (PRIO_TO_NICE(p->static_prio) < 0)
2362			p->static_prio = NICE_TO_PRIO(0);
2363
2364		p->prio = p->normal_prio = __normal_prio(p);
2365		set_load_weight(p, false);
2366
2367		/*
2368		 * We don't need the reset flag anymore after the fork. It has
2369		 * fulfilled its duty:
2370		 */
2371		p->sched_reset_on_fork = 0;
2372	}
2373
2374	if (dl_prio(p->prio)) {
2375		put_cpu();
2376		return -EAGAIN;
2377	} else if (rt_prio(p->prio)) {
2378		p->sched_class = &rt_sched_class;
2379	} else {
2380		p->sched_class = &fair_sched_class;
2381	}
2382
2383	init_entity_runnable_average(&p->se);
2384
2385	/*
2386	 * The child is not yet in the pid-hash so no cgroup attach races,
2387	 * and the cgroup is pinned to this child due to cgroup_fork()
2388	 * is ran before sched_fork().
2389	 *
2390	 * Silence PROVE_RCU.
2391	 */
2392	raw_spin_lock_irqsave(&p->pi_lock, flags);
2393	/*
2394	 * We're setting the CPU for the first time, we don't migrate,
2395	 * so use __set_task_cpu().
2396	 */
2397	__set_task_cpu(p, cpu);
2398	if (p->sched_class->task_fork)
2399		p->sched_class->task_fork(p);
2400	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2401
2402#ifdef CONFIG_SCHED_INFO
2403	if (likely(sched_info_on()))
2404		memset(&p->sched_info, 0, sizeof(p->sched_info));
2405#endif
2406#if defined(CONFIG_SMP)
2407	p->on_cpu = 0;
2408#endif
2409	init_task_preempt_count(p);
2410#ifdef CONFIG_SMP
2411	plist_node_init(&p->pushable_tasks, MAX_PRIO);
2412	RB_CLEAR_NODE(&p->pushable_dl_tasks);
2413#endif
2414
2415	put_cpu();
2416	return 0;
2417}
2418
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2419unsigned long to_ratio(u64 period, u64 runtime)
2420{
2421	if (runtime == RUNTIME_INF)
2422		return BW_UNIT;
2423
2424	/*
2425	 * Doing this here saves a lot of checks in all
2426	 * the calling paths, and returning zero seems
2427	 * safe for them anyway.
2428	 */
2429	if (period == 0)
2430		return 0;
2431
2432	return div64_u64(runtime << BW_SHIFT, period);
2433}
2434
2435/*
2436 * wake_up_new_task - wake up a newly created task for the first time.
2437 *
2438 * This function will do some initial scheduler statistics housekeeping
2439 * that must be done for every newly created context, then puts the task
2440 * on the runqueue and wakes it.
2441 */
2442void wake_up_new_task(struct task_struct *p)
2443{
2444	struct rq_flags rf;
2445	struct rq *rq;
2446
2447	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2448	p->state = TASK_RUNNING;
2449#ifdef CONFIG_SMP
2450	/*
2451	 * Fork balancing, do it here and not earlier because:
2452	 *  - cpus_allowed can change in the fork path
2453	 *  - any previously selected CPU might disappear through hotplug
2454	 *
2455	 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2456	 * as we're not fully set-up yet.
2457	 */
2458	p->recent_used_cpu = task_cpu(p);
2459	__set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
 
2460#endif
2461	rq = __task_rq_lock(p, &rf);
2462	update_rq_clock(rq);
2463	post_init_entity_util_avg(&p->se);
2464
2465	activate_task(rq, p, ENQUEUE_NOCLOCK);
2466	p->on_rq = TASK_ON_RQ_QUEUED;
2467	trace_sched_wakeup_new(p);
2468	check_preempt_curr(rq, p, WF_FORK);
2469#ifdef CONFIG_SMP
2470	if (p->sched_class->task_woken) {
2471		/*
2472		 * Nothing relies on rq->lock after this, so its fine to
2473		 * drop it.
2474		 */
2475		rq_unpin_lock(rq, &rf);
2476		p->sched_class->task_woken(rq, p);
2477		rq_repin_lock(rq, &rf);
2478	}
2479#endif
2480	task_rq_unlock(rq, p, &rf);
2481}
2482
2483#ifdef CONFIG_PREEMPT_NOTIFIERS
2484
2485static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2486
2487void preempt_notifier_inc(void)
2488{
2489	static_branch_inc(&preempt_notifier_key);
2490}
2491EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2492
2493void preempt_notifier_dec(void)
2494{
2495	static_branch_dec(&preempt_notifier_key);
2496}
2497EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2498
2499/**
2500 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2501 * @notifier: notifier struct to register
2502 */
2503void preempt_notifier_register(struct preempt_notifier *notifier)
2504{
2505	if (!static_branch_unlikely(&preempt_notifier_key))
2506		WARN(1, "registering preempt_notifier while notifiers disabled\n");
2507
2508	hlist_add_head(&notifier->link, &current->preempt_notifiers);
2509}
2510EXPORT_SYMBOL_GPL(preempt_notifier_register);
2511
2512/**
2513 * preempt_notifier_unregister - no longer interested in preemption notifications
2514 * @notifier: notifier struct to unregister
2515 *
2516 * This is *not* safe to call from within a preemption notifier.
2517 */
2518void preempt_notifier_unregister(struct preempt_notifier *notifier)
2519{
2520	hlist_del(&notifier->link);
2521}
2522EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2523
2524static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2525{
2526	struct preempt_notifier *notifier;
2527
2528	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2529		notifier->ops->sched_in(notifier, raw_smp_processor_id());
2530}
2531
2532static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2533{
2534	if (static_branch_unlikely(&preempt_notifier_key))
2535		__fire_sched_in_preempt_notifiers(curr);
2536}
2537
2538static void
2539__fire_sched_out_preempt_notifiers(struct task_struct *curr,
2540				   struct task_struct *next)
2541{
2542	struct preempt_notifier *notifier;
2543
2544	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2545		notifier->ops->sched_out(notifier, next);
2546}
2547
2548static __always_inline void
2549fire_sched_out_preempt_notifiers(struct task_struct *curr,
2550				 struct task_struct *next)
2551{
2552	if (static_branch_unlikely(&preempt_notifier_key))
2553		__fire_sched_out_preempt_notifiers(curr, next);
2554}
2555
2556#else /* !CONFIG_PREEMPT_NOTIFIERS */
2557
2558static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2559{
2560}
2561
2562static inline void
2563fire_sched_out_preempt_notifiers(struct task_struct *curr,
2564				 struct task_struct *next)
2565{
2566}
2567
2568#endif /* CONFIG_PREEMPT_NOTIFIERS */
2569
2570static inline void prepare_task(struct task_struct *next)
2571{
2572#ifdef CONFIG_SMP
2573	/*
2574	 * Claim the task as running, we do this before switching to it
2575	 * such that any running task will have this set.
 
 
 
2576	 */
2577	next->on_cpu = 1;
2578#endif
2579}
2580
2581static inline void finish_task(struct task_struct *prev)
2582{
2583#ifdef CONFIG_SMP
2584	/*
2585	 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2586	 * We must ensure this doesn't happen until the switch is completely
 
2587	 * finished.
2588	 *
2589	 * In particular, the load of prev->state in finish_task_switch() must
2590	 * happen before this.
2591	 *
2592	 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2593	 */
2594	smp_store_release(&prev->on_cpu, 0);
2595#endif
2596}
2597
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2598static inline void
2599prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2600{
2601	/*
2602	 * Since the runqueue lock will be released by the next
2603	 * task (which is an invalid locking op but in the case
2604	 * of the scheduler it's an obvious special-case), so we
2605	 * do an early lockdep release here:
2606	 */
2607	rq_unpin_lock(rq, rf);
2608	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2609#ifdef CONFIG_DEBUG_SPINLOCK
2610	/* this is a valid case when another task releases the spinlock */
2611	rq->lock.owner = next;
2612#endif
2613}
2614
2615static inline void finish_lock_switch(struct rq *rq)
2616{
2617	/*
2618	 * If we are tracking spinlock dependencies then we have to
2619	 * fix up the runqueue lock - which gets 'carried over' from
2620	 * prev into current:
2621	 */
2622	spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2623	raw_spin_unlock_irq(&rq->lock);
 
2624}
2625
2626/*
2627 * NOP if the arch has not defined these:
2628 */
2629
2630#ifndef prepare_arch_switch
2631# define prepare_arch_switch(next)	do { } while (0)
2632#endif
2633
2634#ifndef finish_arch_post_lock_switch
2635# define finish_arch_post_lock_switch()	do { } while (0)
2636#endif
2637
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2638/**
2639 * prepare_task_switch - prepare to switch tasks
2640 * @rq: the runqueue preparing to switch
2641 * @prev: the current task that is being switched out
2642 * @next: the task we are going to switch to.
2643 *
2644 * This is called with the rq lock held and interrupts off. It must
2645 * be paired with a subsequent finish_task_switch after the context
2646 * switch.
2647 *
2648 * prepare_task_switch sets up locking and calls architecture specific
2649 * hooks.
2650 */
2651static inline void
2652prepare_task_switch(struct rq *rq, struct task_struct *prev,
2653		    struct task_struct *next)
2654{
 
2655	sched_info_switch(rq, prev, next);
2656	perf_event_task_sched_out(prev, next);
 
2657	fire_sched_out_preempt_notifiers(prev, next);
 
2658	prepare_task(next);
2659	prepare_arch_switch(next);
2660}
2661
2662/**
2663 * finish_task_switch - clean up after a task-switch
2664 * @prev: the thread we just switched away from.
2665 *
2666 * finish_task_switch must be called after the context switch, paired
2667 * with a prepare_task_switch call before the context switch.
2668 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2669 * and do any other architecture-specific cleanup actions.
2670 *
2671 * Note that we may have delayed dropping an mm in context_switch(). If
2672 * so, we finish that here outside of the runqueue lock. (Doing it
2673 * with the lock held can cause deadlocks; see schedule() for
2674 * details.)
2675 *
2676 * The context switch have flipped the stack from under us and restored the
2677 * local variables which were saved when this task called schedule() in the
2678 * past. prev == current is still correct but we need to recalculate this_rq
2679 * because prev may have moved to another CPU.
2680 */
2681static struct rq *finish_task_switch(struct task_struct *prev)
2682	__releases(rq->lock)
2683{
2684	struct rq *rq = this_rq();
2685	struct mm_struct *mm = rq->prev_mm;
2686	long prev_state;
2687
2688	/*
2689	 * The previous task will have left us with a preempt_count of 2
2690	 * because it left us after:
2691	 *
2692	 *	schedule()
2693	 *	  preempt_disable();			// 1
2694	 *	  __schedule()
2695	 *	    raw_spin_lock_irq(&rq->lock)	// 2
2696	 *
2697	 * Also, see FORK_PREEMPT_COUNT.
2698	 */
2699	if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2700		      "corrupted preempt_count: %s/%d/0x%x\n",
2701		      current->comm, current->pid, preempt_count()))
2702		preempt_count_set(FORK_PREEMPT_COUNT);
2703
2704	rq->prev_mm = NULL;
2705
2706	/*
2707	 * A task struct has one reference for the use as "current".
2708	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2709	 * schedule one last time. The schedule call will never return, and
2710	 * the scheduled task must drop that reference.
2711	 *
2712	 * We must observe prev->state before clearing prev->on_cpu (in
2713	 * finish_task), otherwise a concurrent wakeup can get prev
2714	 * running on another CPU and we could rave with its RUNNING -> DEAD
2715	 * transition, resulting in a double drop.
2716	 */
2717	prev_state = prev->state;
2718	vtime_task_switch(prev);
2719	perf_event_task_sched_in(prev, current);
2720	finish_task(prev);
 
2721	finish_lock_switch(rq);
2722	finish_arch_post_lock_switch();
 
 
 
 
 
 
 
 
 
2723
2724	fire_sched_in_preempt_notifiers(current);
2725	/*
2726	 * When switching through a kernel thread, the loop in
2727	 * membarrier_{private,global}_expedited() may have observed that
2728	 * kernel thread and not issued an IPI. It is therefore possible to
2729	 * schedule between user->kernel->user threads without passing though
2730	 * switch_mm(). Membarrier requires a barrier after storing to
2731	 * rq->curr, before returning to userspace, so provide them here:
2732	 *
2733	 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2734	 *   provided by mmdrop(),
2735	 * - a sync_core for SYNC_CORE.
2736	 */
2737	if (mm) {
2738		membarrier_mm_sync_core_before_usermode(mm);
2739		mmdrop(mm);
2740	}
2741	if (unlikely(prev_state & (TASK_DEAD|TASK_PARKED))) {
2742		switch (prev_state) {
2743		case TASK_DEAD:
2744			if (prev->sched_class->task_dead)
2745				prev->sched_class->task_dead(prev);
2746
2747			/*
2748			 * Remove function-return probe instances associated with this
2749			 * task and put them back on the free list.
2750			 */
2751			kprobe_flush_task(prev);
2752
2753			/* Task is done with its stack. */
2754			put_task_stack(prev);
2755
2756			put_task_struct(prev);
2757			break;
2758
2759		case TASK_PARKED:
2760			kthread_park_complete(prev);
2761			break;
2762		}
2763	}
2764
2765	tick_nohz_task_switch();
2766	return rq;
2767}
2768
2769#ifdef CONFIG_SMP
2770
2771/* rq->lock is NOT held, but preemption is disabled */
2772static void __balance_callback(struct rq *rq)
2773{
2774	struct callback_head *head, *next;
2775	void (*func)(struct rq *rq);
2776	unsigned long flags;
2777
2778	raw_spin_lock_irqsave(&rq->lock, flags);
2779	head = rq->balance_callback;
2780	rq->balance_callback = NULL;
2781	while (head) {
2782		func = (void (*)(struct rq *))head->func;
2783		next = head->next;
2784		head->next = NULL;
2785		head = next;
2786
2787		func(rq);
2788	}
2789	raw_spin_unlock_irqrestore(&rq->lock, flags);
2790}
2791
2792static inline void balance_callback(struct rq *rq)
2793{
2794	if (unlikely(rq->balance_callback))
2795		__balance_callback(rq);
2796}
2797
2798#else
2799
2800static inline void balance_callback(struct rq *rq)
2801{
2802}
2803
2804#endif
2805
2806/**
2807 * schedule_tail - first thing a freshly forked thread must call.
2808 * @prev: the thread we just switched away from.
2809 */
2810asmlinkage __visible void schedule_tail(struct task_struct *prev)
2811	__releases(rq->lock)
2812{
2813	struct rq *rq;
2814
2815	/*
2816	 * New tasks start with FORK_PREEMPT_COUNT, see there and
2817	 * finish_task_switch() for details.
2818	 *
2819	 * finish_task_switch() will drop rq->lock() and lower preempt_count
2820	 * and the preempt_enable() will end up enabling preemption (on
2821	 * PREEMPT_COUNT kernels).
2822	 */
2823
2824	rq = finish_task_switch(prev);
2825	balance_callback(rq);
2826	preempt_enable();
2827
2828	if (current->set_child_tid)
2829		put_user(task_pid_vnr(current), current->set_child_tid);
 
 
2830}
2831
2832/*
2833 * context_switch - switch to the new MM and the new thread's register state.
2834 */
2835static __always_inline struct rq *
2836context_switch(struct rq *rq, struct task_struct *prev,
2837	       struct task_struct *next, struct rq_flags *rf)
2838{
2839	struct mm_struct *mm, *oldmm;
2840
2841	prepare_task_switch(rq, prev, next);
2842
2843	mm = next->mm;
2844	oldmm = prev->active_mm;
2845	/*
2846	 * For paravirt, this is coupled with an exit in switch_to to
2847	 * combine the page table reload and the switch backend into
2848	 * one hypercall.
2849	 */
2850	arch_start_context_switch(prev);
2851
2852	/*
2853	 * If mm is non-NULL, we pass through switch_mm(). If mm is
2854	 * NULL, we will pass through mmdrop() in finish_task_switch().
2855	 * Both of these contain the full memory barrier required by
2856	 * membarrier after storing to rq->curr, before returning to
2857	 * user-space.
2858	 */
2859	if (!mm) {
2860		next->active_mm = oldmm;
2861		mmgrab(oldmm);
2862		enter_lazy_tlb(oldmm, next);
2863	} else
2864		switch_mm_irqs_off(oldmm, mm, next);
2865
2866	if (!prev->mm) {
2867		prev->active_mm = NULL;
2868		rq->prev_mm = oldmm;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2869	}
2870
2871	rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
 
2872
2873	prepare_lock_switch(rq, next, rf);
2874
2875	/* Here we just switch the register state and the stack. */
2876	switch_to(prev, next, prev);
2877	barrier();
2878
2879	return finish_task_switch(prev);
2880}
2881
2882/*
2883 * nr_running and nr_context_switches:
2884 *
2885 * externally visible scheduler statistics: current number of runnable
2886 * threads, total number of context switches performed since bootup.
2887 */
2888unsigned long nr_running(void)
2889{
2890	unsigned long i, sum = 0;
2891
2892	for_each_online_cpu(i)
2893		sum += cpu_rq(i)->nr_running;
2894
2895	return sum;
2896}
2897
2898/*
2899 * Check if only the current task is running on the CPU.
2900 *
2901 * Caution: this function does not check that the caller has disabled
2902 * preemption, thus the result might have a time-of-check-to-time-of-use
2903 * race.  The caller is responsible to use it correctly, for example:
2904 *
2905 * - from a non-preemptable section (of course)
2906 *
2907 * - from a thread that is bound to a single CPU
2908 *
2909 * - in a loop with very short iterations (e.g. a polling loop)
2910 */
2911bool single_task_running(void)
2912{
2913	return raw_rq()->nr_running == 1;
2914}
2915EXPORT_SYMBOL(single_task_running);
2916
 
 
 
 
 
2917unsigned long long nr_context_switches(void)
2918{
2919	int i;
2920	unsigned long long sum = 0;
2921
2922	for_each_possible_cpu(i)
2923		sum += cpu_rq(i)->nr_switches;
2924
2925	return sum;
2926}
2927
2928/*
2929 * IO-wait accounting, and how its mostly bollocks (on SMP).
 
 
 
 
 
 
 
 
 
 
 
 
2930 *
2931 * The idea behind IO-wait account is to account the idle time that we could
2932 * have spend running if it were not for IO. That is, if we were to improve the
2933 * storage performance, we'd have a proportional reduction in IO-wait time.
2934 *
2935 * This all works nicely on UP, where, when a task blocks on IO, we account
2936 * idle time as IO-wait, because if the storage were faster, it could've been
2937 * running and we'd not be idle.
2938 *
2939 * This has been extended to SMP, by doing the same for each CPU. This however
2940 * is broken.
2941 *
2942 * Imagine for instance the case where two tasks block on one CPU, only the one
2943 * CPU will have IO-wait accounted, while the other has regular idle. Even
2944 * though, if the storage were faster, both could've ran at the same time,
2945 * utilising both CPUs.
2946 *
2947 * This means, that when looking globally, the current IO-wait accounting on
2948 * SMP is a lower bound, by reason of under accounting.
2949 *
2950 * Worse, since the numbers are provided per CPU, they are sometimes
2951 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2952 * associated with any one particular CPU, it can wake to another CPU than it
2953 * blocked on. This means the per CPU IO-wait number is meaningless.
2954 *
2955 * Task CPU affinities can make all that even more 'interesting'.
2956 */
2957
2958unsigned long nr_iowait(void)
2959{
2960	unsigned long i, sum = 0;
2961
2962	for_each_possible_cpu(i)
2963		sum += atomic_read(&cpu_rq(i)->nr_iowait);
2964
2965	return sum;
2966}
2967
2968/*
2969 * Consumers of these two interfaces, like for example the cpufreq menu
2970 * governor are using nonsensical data. Boosting frequency for a CPU that has
2971 * IO-wait which might not even end up running the task when it does become
2972 * runnable.
2973 */
2974
2975unsigned long nr_iowait_cpu(int cpu)
2976{
2977	struct rq *this = cpu_rq(cpu);
2978	return atomic_read(&this->nr_iowait);
2979}
2980
2981void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2982{
2983	struct rq *rq = this_rq();
2984	*nr_waiters = atomic_read(&rq->nr_iowait);
2985	*load = rq->load.weight;
2986}
2987
2988#ifdef CONFIG_SMP
2989
2990/*
2991 * sched_exec - execve() is a valuable balancing opportunity, because at
2992 * this point the task has the smallest effective memory and cache footprint.
2993 */
2994void sched_exec(void)
2995{
2996	struct task_struct *p = current;
2997	unsigned long flags;
2998	int dest_cpu;
2999
3000	raw_spin_lock_irqsave(&p->pi_lock, flags);
3001	dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3002	if (dest_cpu == smp_processor_id())
3003		goto unlock;
3004
3005	if (likely(cpu_active(dest_cpu))) {
3006		struct migration_arg arg = { p, dest_cpu };
3007
3008		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3009		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3010		return;
3011	}
3012unlock:
3013	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3014}
3015
3016#endif
3017
3018DEFINE_PER_CPU(struct kernel_stat, kstat);
3019DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3020
3021EXPORT_PER_CPU_SYMBOL(kstat);
3022EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3023
3024/*
3025 * The function fair_sched_class.update_curr accesses the struct curr
3026 * and its field curr->exec_start; when called from task_sched_runtime(),
3027 * we observe a high rate of cache misses in practice.
3028 * Prefetching this data results in improved performance.
3029 */
3030static inline void prefetch_curr_exec_start(struct task_struct *p)
3031{
3032#ifdef CONFIG_FAIR_GROUP_SCHED
3033	struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3034#else
3035	struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3036#endif
3037	prefetch(curr);
3038	prefetch(&curr->exec_start);
3039}
3040
3041/*
3042 * Return accounted runtime for the task.
3043 * In case the task is currently running, return the runtime plus current's
3044 * pending runtime that have not been accounted yet.
3045 */
3046unsigned long long task_sched_runtime(struct task_struct *p)
3047{
3048	struct rq_flags rf;
3049	struct rq *rq;
3050	u64 ns;
3051
3052#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3053	/*
3054	 * 64-bit doesn't need locks to atomically read a 64-bit value.
3055	 * So we have a optimization chance when the task's delta_exec is 0.
3056	 * Reading ->on_cpu is racy, but this is ok.
3057	 *
3058	 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3059	 * If we race with it entering CPU, unaccounted time is 0. This is
3060	 * indistinguishable from the read occurring a few cycles earlier.
3061	 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3062	 * been accounted, so we're correct here as well.
3063	 */
3064	if (!p->on_cpu || !task_on_rq_queued(p))
3065		return p->se.sum_exec_runtime;
3066#endif
3067
3068	rq = task_rq_lock(p, &rf);
3069	/*
3070	 * Must be ->curr _and_ ->on_rq.  If dequeued, we would
3071	 * project cycles that may never be accounted to this
3072	 * thread, breaking clock_gettime().
3073	 */
3074	if (task_current(rq, p) && task_on_rq_queued(p)) {
3075		prefetch_curr_exec_start(p);
3076		update_rq_clock(rq);
3077		p->sched_class->update_curr(rq);
3078	}
3079	ns = p->se.sum_exec_runtime;
3080	task_rq_unlock(rq, p, &rf);
3081
3082	return ns;
3083}
3084
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3085/*
3086 * This function gets called by the timer code, with HZ frequency.
3087 * We call it with interrupts disabled.
3088 */
3089void scheduler_tick(void)
3090{
3091	int cpu = smp_processor_id();
3092	struct rq *rq = cpu_rq(cpu);
3093	struct task_struct *curr = rq->curr;
3094	struct rq_flags rf;
 
 
 
 
 
3095
3096	sched_clock_tick();
3097
3098	rq_lock(rq, &rf);
3099
3100	update_rq_clock(rq);
 
 
3101	curr->sched_class->task_tick(rq, curr, 0);
3102	cpu_load_update_active(rq);
 
3103	calc_global_load_tick(rq);
 
 
3104
3105	rq_unlock(rq, &rf);
3106
 
 
 
3107	perf_event_task_tick();
3108
 
 
 
3109#ifdef CONFIG_SMP
3110	rq->idle_balance = idle_cpu(cpu);
3111	trigger_load_balance(rq);
3112#endif
3113}
3114
3115#ifdef CONFIG_NO_HZ_FULL
3116
3117struct tick_work {
3118	int			cpu;
 
3119	struct delayed_work	work;
3120};
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3121
3122static struct tick_work __percpu *tick_work_cpu;
3123
3124static void sched_tick_remote(struct work_struct *work)
3125{
3126	struct delayed_work *dwork = to_delayed_work(work);
3127	struct tick_work *twork = container_of(dwork, struct tick_work, work);
3128	int cpu = twork->cpu;
3129	struct rq *rq = cpu_rq(cpu);
3130	struct rq_flags rf;
3131
3132	/*
3133	 * Handle the tick only if it appears the remote CPU is running in full
3134	 * dynticks mode. The check is racy by nature, but missing a tick or
3135	 * having one too much is no big deal because the scheduler tick updates
3136	 * statistics and checks timeslices in a time-independent way, regardless
3137	 * of when exactly it is running.
3138	 */
3139	if (!idle_cpu(cpu) && tick_nohz_tick_stopped_cpu(cpu)) {
3140		struct task_struct *curr;
3141		u64 delta;
3142
3143		rq_lock_irq(rq, &rf);
3144		update_rq_clock(rq);
3145		curr = rq->curr;
3146		delta = rq_clock_task(rq) - curr->se.exec_start;
 
 
 
 
 
 
 
 
3147
3148		/*
3149		 * Make sure the next tick runs within a reasonable
3150		 * amount of time.
3151		 */
3152		WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3153		curr->sched_class->task_tick(rq, curr, 0);
3154		rq_unlock_irq(rq, &rf);
3155	}
3156
3157	/*
3158	 * Run the remote tick once per second (1Hz). This arbitrary
3159	 * frequency is large enough to avoid overload but short enough
3160	 * to keep scheduler internal stats reasonably up to date.
 
3161	 */
3162	queue_delayed_work(system_unbound_wq, dwork, HZ);
 
 
 
3163}
3164
3165static void sched_tick_start(int cpu)
3166{
 
3167	struct tick_work *twork;
3168
3169	if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3170		return;
3171
3172	WARN_ON_ONCE(!tick_work_cpu);
3173
3174	twork = per_cpu_ptr(tick_work_cpu, cpu);
3175	twork->cpu = cpu;
3176	INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3177	queue_delayed_work(system_unbound_wq, &twork->work, HZ);
 
 
 
 
3178}
3179
3180#ifdef CONFIG_HOTPLUG_CPU
3181static void sched_tick_stop(int cpu)
3182{
3183	struct tick_work *twork;
 
3184
3185	if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3186		return;
3187
3188	WARN_ON_ONCE(!tick_work_cpu);
3189
3190	twork = per_cpu_ptr(tick_work_cpu, cpu);
3191	cancel_delayed_work_sync(&twork->work);
 
 
 
3192}
3193#endif /* CONFIG_HOTPLUG_CPU */
3194
3195int __init sched_tick_offload_init(void)
3196{
3197	tick_work_cpu = alloc_percpu(struct tick_work);
3198	BUG_ON(!tick_work_cpu);
3199
3200	return 0;
3201}
3202
3203#else /* !CONFIG_NO_HZ_FULL */
3204static inline void sched_tick_start(int cpu) { }
3205static inline void sched_tick_stop(int cpu) { }
3206#endif
3207
3208#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3209				defined(CONFIG_PREEMPT_TRACER))
3210/*
3211 * If the value passed in is equal to the current preempt count
3212 * then we just disabled preemption. Start timing the latency.
3213 */
3214static inline void preempt_latency_start(int val)
3215{
3216	if (preempt_count() == val) {
3217		unsigned long ip = get_lock_parent_ip();
3218#ifdef CONFIG_DEBUG_PREEMPT
3219		current->preempt_disable_ip = ip;
3220#endif
3221		trace_preempt_off(CALLER_ADDR0, ip);
3222	}
3223}
3224
3225void preempt_count_add(int val)
3226{
3227#ifdef CONFIG_DEBUG_PREEMPT
3228	/*
3229	 * Underflow?
3230	 */
3231	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3232		return;
3233#endif
3234	__preempt_count_add(val);
3235#ifdef CONFIG_DEBUG_PREEMPT
3236	/*
3237	 * Spinlock count overflowing soon?
3238	 */
3239	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3240				PREEMPT_MASK - 10);
3241#endif
3242	preempt_latency_start(val);
3243}
3244EXPORT_SYMBOL(preempt_count_add);
3245NOKPROBE_SYMBOL(preempt_count_add);
3246
3247/*
3248 * If the value passed in equals to the current preempt count
3249 * then we just enabled preemption. Stop timing the latency.
3250 */
3251static inline void preempt_latency_stop(int val)
3252{
3253	if (preempt_count() == val)
3254		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3255}
3256
3257void preempt_count_sub(int val)
3258{
3259#ifdef CONFIG_DEBUG_PREEMPT
3260	/*
3261	 * Underflow?
3262	 */
3263	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3264		return;
3265	/*
3266	 * Is the spinlock portion underflowing?
3267	 */
3268	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3269			!(preempt_count() & PREEMPT_MASK)))
3270		return;
3271#endif
3272
3273	preempt_latency_stop(val);
3274	__preempt_count_sub(val);
3275}
3276EXPORT_SYMBOL(preempt_count_sub);
3277NOKPROBE_SYMBOL(preempt_count_sub);
3278
3279#else
3280static inline void preempt_latency_start(int val) { }
3281static inline void preempt_latency_stop(int val) { }
3282#endif
3283
3284static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3285{
3286#ifdef CONFIG_DEBUG_PREEMPT
3287	return p->preempt_disable_ip;
3288#else
3289	return 0;
3290#endif
3291}
3292
3293/*
3294 * Print scheduling while atomic bug:
3295 */
3296static noinline void __schedule_bug(struct task_struct *prev)
3297{
3298	/* Save this before calling printk(), since that will clobber it */
3299	unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3300
3301	if (oops_in_progress)
3302		return;
3303
3304	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3305		prev->comm, prev->pid, preempt_count());
3306
3307	debug_show_held_locks(prev);
3308	print_modules();
3309	if (irqs_disabled())
3310		print_irqtrace_events(prev);
3311	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3312	    && in_atomic_preempt_off()) {
3313		pr_err("Preemption disabled at:");
3314		print_ip_sym(preempt_disable_ip);
3315		pr_cont("\n");
3316	}
3317	if (panic_on_warn)
3318		panic("scheduling while atomic\n");
3319
3320	dump_stack();
3321	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3322}
3323
3324/*
3325 * Various schedule()-time debugging checks and statistics:
3326 */
3327static inline void schedule_debug(struct task_struct *prev)
3328{
3329#ifdef CONFIG_SCHED_STACK_END_CHECK
3330	if (task_stack_end_corrupted(prev))
3331		panic("corrupted stack end detected inside scheduler\n");
 
 
 
 
 
 
 
 
 
 
 
 
3332#endif
3333
3334	if (unlikely(in_atomic_preempt_off())) {
3335		__schedule_bug(prev);
3336		preempt_count_set(PREEMPT_DISABLED);
3337	}
3338	rcu_sleep_check();
 
3339
3340	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3341
3342	schedstat_inc(this_rq()->sched_count);
3343}
3344
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3345/*
3346 * Pick up the highest-prio task:
3347 */
3348static inline struct task_struct *
3349pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3350{
3351	const struct sched_class *class;
3352	struct task_struct *p;
3353
3354	/*
3355	 * Optimization: we know that if all tasks are in the fair class we can
3356	 * call that function directly, but only if the @prev task wasn't of a
3357	 * higher scheduling class, because otherwise those loose the
3358	 * opportunity to pull in more work from other CPUs.
3359	 */
3360	if (likely((prev->sched_class == &idle_sched_class ||
3361		    prev->sched_class == &fair_sched_class) &&
3362		   rq->nr_running == rq->cfs.h_nr_running)) {
3363
3364		p = fair_sched_class.pick_next_task(rq, prev, rf);
3365		if (unlikely(p == RETRY_TASK))
3366			goto again;
3367
3368		/* Assumes fair_sched_class->next == idle_sched_class */
3369		if (unlikely(!p))
3370			p = idle_sched_class.pick_next_task(rq, prev, rf);
 
 
 
 
 
 
 
 
 
3371
3372		return p;
3373	}
3374
3375again:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3376	for_each_class(class) {
3377		p = class->pick_next_task(rq, prev, rf);
3378		if (p) {
3379			if (unlikely(p == RETRY_TASK))
3380				goto again;
3381			return p;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3382		}
 
 
 
3383	}
3384
3385	/* The idle class should always have a runnable task: */
3386	BUG();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3387}
3388
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3389/*
3390 * __schedule() is the main scheduler function.
3391 *
3392 * The main means of driving the scheduler and thus entering this function are:
3393 *
3394 *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3395 *
3396 *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3397 *      paths. For example, see arch/x86/entry_64.S.
3398 *
3399 *      To drive preemption between tasks, the scheduler sets the flag in timer
3400 *      interrupt handler scheduler_tick().
3401 *
3402 *   3. Wakeups don't really cause entry into schedule(). They add a
3403 *      task to the run-queue and that's it.
3404 *
3405 *      Now, if the new task added to the run-queue preempts the current
3406 *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3407 *      called on the nearest possible occasion:
3408 *
3409 *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
3410 *
3411 *         - in syscall or exception context, at the next outmost
3412 *           preempt_enable(). (this might be as soon as the wake_up()'s
3413 *           spin_unlock()!)
3414 *
3415 *         - in IRQ context, return from interrupt-handler to
3416 *           preemptible context
3417 *
3418 *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3419 *         then at the next:
3420 *
3421 *          - cond_resched() call
3422 *          - explicit schedule() call
3423 *          - return from syscall or exception to user-space
3424 *          - return from interrupt-handler to user-space
3425 *
3426 * WARNING: must be called with preemption disabled!
3427 */
3428static void __sched notrace __schedule(bool preempt)
3429{
3430	struct task_struct *prev, *next;
3431	unsigned long *switch_count;
 
3432	struct rq_flags rf;
3433	struct rq *rq;
3434	int cpu;
3435
3436	cpu = smp_processor_id();
3437	rq = cpu_rq(cpu);
3438	prev = rq->curr;
3439
3440	schedule_debug(prev);
3441
3442	if (sched_feat(HRTICK))
3443		hrtick_clear(rq);
3444
3445	local_irq_disable();
3446	rcu_note_context_switch(preempt);
3447
3448	/*
3449	 * Make sure that signal_pending_state()->signal_pending() below
3450	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3451	 * done by the caller to avoid the race with signal_wake_up().
 
 
 
 
 
 
 
3452	 *
3453	 * The membarrier system call requires a full memory barrier
3454	 * after coming from user-space, before storing to rq->curr.
3455	 */
3456	rq_lock(rq, &rf);
3457	smp_mb__after_spinlock();
3458
3459	/* Promote REQ to ACT */
3460	rq->clock_update_flags <<= 1;
3461	update_rq_clock(rq);
 
3462
3463	switch_count = &prev->nivcsw;
3464	if (!preempt && prev->state) {
3465		if (unlikely(signal_pending_state(prev->state, prev))) {
3466			prev->state = TASK_RUNNING;
 
 
 
 
 
 
3467		} else {
3468			deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3469			prev->on_rq = 0;
 
 
3470
3471			if (prev->in_iowait) {
3472				atomic_inc(&rq->nr_iowait);
3473				delayacct_blkio_start();
3474			}
3475
3476			/*
3477			 * If a worker went to sleep, notify and ask workqueue
3478			 * whether it wants to wake up a task to maintain
3479			 * concurrency.
 
 
 
 
 
 
3480			 */
3481			if (prev->flags & PF_WQ_WORKER) {
3482				struct task_struct *to_wakeup;
3483
3484				to_wakeup = wq_worker_sleeping(prev);
3485				if (to_wakeup)
3486					try_to_wake_up_local(to_wakeup, &rf);
3487			}
3488		}
3489		switch_count = &prev->nvcsw;
3490	}
3491
3492	next = pick_next_task(rq, prev, &rf);
3493	clear_tsk_need_resched(prev);
3494	clear_preempt_need_resched();
 
 
 
3495
3496	if (likely(prev != next)) {
3497		rq->nr_switches++;
3498		rq->curr = next;
 
 
 
 
3499		/*
3500		 * The membarrier system call requires each architecture
3501		 * to have a full memory barrier after updating
3502		 * rq->curr, before returning to user-space.
3503		 *
3504		 * Here are the schemes providing that barrier on the
3505		 * various architectures:
3506		 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3507		 *   switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3508		 * - finish_lock_switch() for weakly-ordered
3509		 *   architectures where spin_unlock is a full barrier,
3510		 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3511		 *   is a RELEASE barrier),
3512		 */
3513		++*switch_count;
3514
3515		trace_sched_switch(preempt, prev, next);
 
 
 
3516
3517		/* Also unlocks the rq: */
3518		rq = context_switch(rq, prev, next, &rf);
3519	} else {
3520		rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3521		rq_unlock_irq(rq, &rf);
 
3522	}
3523
3524	balance_callback(rq);
3525}
3526
3527void __noreturn do_task_dead(void)
3528{
3529	/* Causes final put_task_struct in finish_task_switch(): */
3530	set_special_state(TASK_DEAD);
3531
3532	/* Tell freezer to ignore us: */
3533	current->flags |= PF_NOFREEZE;
3534
3535	__schedule(false);
3536	BUG();
3537
3538	/* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3539	for (;;)
3540		cpu_relax();
3541}
3542
3543static inline void sched_submit_work(struct task_struct *tsk)
3544{
3545	if (!tsk->state || tsk_is_pi_blocked(tsk))
3546		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3547	/*
3548	 * If we are going to sleep and we have plugged IO queued,
3549	 * make sure to submit it to avoid deadlocks.
3550	 */
3551	if (blk_needs_flush_plug(tsk))
3552		blk_schedule_flush_plug(tsk);
 
3553}
3554
3555asmlinkage __visible void __sched schedule(void)
3556{
3557	struct task_struct *tsk = current;
 
 
 
 
 
 
3558
3559	sched_submit_work(tsk);
 
3560	do {
3561		preempt_disable();
3562		__schedule(false);
3563		sched_preempt_enable_no_resched();
3564	} while (need_resched());
3565}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3566EXPORT_SYMBOL(schedule);
3567
3568/*
3569 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3570 * state (have scheduled out non-voluntarily) by making sure that all
3571 * tasks have either left the run queue or have gone into user space.
3572 * As idle tasks do not do either, they must not ever be preempted
3573 * (schedule out non-voluntarily).
3574 *
3575 * schedule_idle() is similar to schedule_preempt_disable() except that it
3576 * never enables preemption because it does not call sched_submit_work().
3577 */
3578void __sched schedule_idle(void)
3579{
3580	/*
3581	 * As this skips calling sched_submit_work(), which the idle task does
3582	 * regardless because that function is a nop when the task is in a
3583	 * TASK_RUNNING state, make sure this isn't used someplace that the
3584	 * current task can be in any other state. Note, idle is always in the
3585	 * TASK_RUNNING state.
3586	 */
3587	WARN_ON_ONCE(current->state);
3588	do {
3589		__schedule(false);
3590	} while (need_resched());
3591}
3592
3593#ifdef CONFIG_CONTEXT_TRACKING
3594asmlinkage __visible void __sched schedule_user(void)
3595{
3596	/*
3597	 * If we come here after a random call to set_need_resched(),
3598	 * or we have been woken up remotely but the IPI has not yet arrived,
3599	 * we haven't yet exited the RCU idle mode. Do it here manually until
3600	 * we find a better solution.
3601	 *
3602	 * NB: There are buggy callers of this function.  Ideally we
3603	 * should warn if prev_state != CONTEXT_USER, but that will trigger
3604	 * too frequently to make sense yet.
3605	 */
3606	enum ctx_state prev_state = exception_enter();
3607	schedule();
3608	exception_exit(prev_state);
3609}
3610#endif
3611
3612/**
3613 * schedule_preempt_disabled - called with preemption disabled
3614 *
3615 * Returns with preemption disabled. Note: preempt_count must be 1
3616 */
3617void __sched schedule_preempt_disabled(void)
3618{
3619	sched_preempt_enable_no_resched();
3620	schedule();
3621	preempt_disable();
3622}
3623
 
 
 
 
 
 
 
 
3624static void __sched notrace preempt_schedule_common(void)
3625{
3626	do {
3627		/*
3628		 * Because the function tracer can trace preempt_count_sub()
3629		 * and it also uses preempt_enable/disable_notrace(), if
3630		 * NEED_RESCHED is set, the preempt_enable_notrace() called
3631		 * by the function tracer will call this function again and
3632		 * cause infinite recursion.
3633		 *
3634		 * Preemption must be disabled here before the function
3635		 * tracer can trace. Break up preempt_disable() into two
3636		 * calls. One to disable preemption without fear of being
3637		 * traced. The other to still record the preemption latency,
3638		 * which can also be traced by the function tracer.
3639		 */
3640		preempt_disable_notrace();
3641		preempt_latency_start(1);
3642		__schedule(true);
3643		preempt_latency_stop(1);
3644		preempt_enable_no_resched_notrace();
3645
3646		/*
3647		 * Check again in case we missed a preemption opportunity
3648		 * between schedule and now.
3649		 */
3650	} while (need_resched());
3651}
3652
3653#ifdef CONFIG_PREEMPT
3654/*
3655 * this is the entry point to schedule() from in-kernel preemption
3656 * off of preempt_enable. Kernel preemptions off return from interrupt
3657 * occur there and call schedule directly.
3658 */
3659asmlinkage __visible void __sched notrace preempt_schedule(void)
3660{
3661	/*
3662	 * If there is a non-zero preempt_count or interrupts are disabled,
3663	 * we do not want to preempt the current task. Just return..
3664	 */
3665	if (likely(!preemptible()))
3666		return;
3667
3668	preempt_schedule_common();
3669}
3670NOKPROBE_SYMBOL(preempt_schedule);
3671EXPORT_SYMBOL(preempt_schedule);
3672
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3673/**
3674 * preempt_schedule_notrace - preempt_schedule called by tracing
3675 *
3676 * The tracing infrastructure uses preempt_enable_notrace to prevent
3677 * recursion and tracing preempt enabling caused by the tracing
3678 * infrastructure itself. But as tracing can happen in areas coming
3679 * from userspace or just about to enter userspace, a preempt enable
3680 * can occur before user_exit() is called. This will cause the scheduler
3681 * to be called when the system is still in usermode.
3682 *
3683 * To prevent this, the preempt_enable_notrace will use this function
3684 * instead of preempt_schedule() to exit user context if needed before
3685 * calling the scheduler.
3686 */
3687asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3688{
3689	enum ctx_state prev_ctx;
3690
3691	if (likely(!preemptible()))
3692		return;
3693
3694	do {
3695		/*
3696		 * Because the function tracer can trace preempt_count_sub()
3697		 * and it also uses preempt_enable/disable_notrace(), if
3698		 * NEED_RESCHED is set, the preempt_enable_notrace() called
3699		 * by the function tracer will call this function again and
3700		 * cause infinite recursion.
3701		 *
3702		 * Preemption must be disabled here before the function
3703		 * tracer can trace. Break up preempt_disable() into two
3704		 * calls. One to disable preemption without fear of being
3705		 * traced. The other to still record the preemption latency,
3706		 * which can also be traced by the function tracer.
3707		 */
3708		preempt_disable_notrace();
3709		preempt_latency_start(1);
3710		/*
3711		 * Needs preempt disabled in case user_exit() is traced
3712		 * and the tracer calls preempt_enable_notrace() causing
3713		 * an infinite recursion.
3714		 */
3715		prev_ctx = exception_enter();
3716		__schedule(true);
3717		exception_exit(prev_ctx);
3718
3719		preempt_latency_stop(1);
3720		preempt_enable_no_resched_notrace();
3721	} while (need_resched());
3722}
3723EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3724
3725#endif /* CONFIG_PREEMPT */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3726
3727/*
3728 * this is the entry point to schedule() from kernel preemption
3729 * off of irq context.
3730 * Note, that this is called and return with irqs disabled. This will
3731 * protect us against recursive calling from irq.
3732 */
3733asmlinkage __visible void __sched preempt_schedule_irq(void)
3734{
3735	enum ctx_state prev_state;
3736
3737	/* Catch callers which need to be fixed */
3738	BUG_ON(preempt_count() || !irqs_disabled());
3739
3740	prev_state = exception_enter();
3741
3742	do {
3743		preempt_disable();
3744		local_irq_enable();
3745		__schedule(true);
3746		local_irq_disable();
3747		sched_preempt_enable_no_resched();
3748	} while (need_resched());
3749
3750	exception_exit(prev_state);
3751}
3752
3753int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3754			  void *key)
3755{
 
3756	return try_to_wake_up(curr->private, mode, wake_flags);
3757}
3758EXPORT_SYMBOL(default_wake_function);
3759
 
 
 
 
 
 
 
 
 
 
 
 
3760#ifdef CONFIG_RT_MUTEXES
3761
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3762static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3763{
3764	if (pi_task)
3765		prio = min(prio, pi_task->prio);
3766
3767	return prio;
3768}
3769
3770static inline int rt_effective_prio(struct task_struct *p, int prio)
3771{
3772	struct task_struct *pi_task = rt_mutex_get_top_task(p);
3773
3774	return __rt_effective_prio(pi_task, prio);
3775}
3776
3777/*
3778 * rt_mutex_setprio - set the current priority of a task
3779 * @p: task to boost
3780 * @pi_task: donor task
3781 *
3782 * This function changes the 'effective' priority of a task. It does
3783 * not touch ->normal_prio like __setscheduler().
3784 *
3785 * Used by the rt_mutex code to implement priority inheritance
3786 * logic. Call site only calls if the priority of the task changed.
3787 */
3788void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3789{
3790	int prio, oldprio, queued, running, queue_flag =
3791		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3792	const struct sched_class *prev_class;
3793	struct rq_flags rf;
3794	struct rq *rq;
3795
3796	/* XXX used to be waiter->prio, not waiter->task->prio */
3797	prio = __rt_effective_prio(pi_task, p->normal_prio);
3798
3799	/*
3800	 * If nothing changed; bail early.
3801	 */
3802	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3803		return;
3804
3805	rq = __task_rq_lock(p, &rf);
3806	update_rq_clock(rq);
3807	/*
3808	 * Set under pi_lock && rq->lock, such that the value can be used under
3809	 * either lock.
3810	 *
3811	 * Note that there is loads of tricky to make this pointer cache work
3812	 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3813	 * ensure a task is de-boosted (pi_task is set to NULL) before the
3814	 * task is allowed to run again (and can exit). This ensures the pointer
3815	 * points to a blocked task -- which guaratees the task is present.
3816	 */
3817	p->pi_top_task = pi_task;
3818
3819	/*
3820	 * For FIFO/RR we only need to set prio, if that matches we're done.
3821	 */
3822	if (prio == p->prio && !dl_prio(prio))
3823		goto out_unlock;
3824
3825	/*
3826	 * Idle task boosting is a nono in general. There is one
3827	 * exception, when PREEMPT_RT and NOHZ is active:
3828	 *
3829	 * The idle task calls get_next_timer_interrupt() and holds
3830	 * the timer wheel base->lock on the CPU and another CPU wants
3831	 * to access the timer (probably to cancel it). We can safely
3832	 * ignore the boosting request, as the idle CPU runs this code
3833	 * with interrupts disabled and will complete the lock
3834	 * protected section without being interrupted. So there is no
3835	 * real need to boost.
3836	 */
3837	if (unlikely(p == rq->idle)) {
3838		WARN_ON(p != rq->curr);
3839		WARN_ON(p->pi_blocked_on);
3840		goto out_unlock;
3841	}
3842
3843	trace_sched_pi_setprio(p, pi_task);
3844	oldprio = p->prio;
3845
3846	if (oldprio == prio)
3847		queue_flag &= ~DEQUEUE_MOVE;
3848
3849	prev_class = p->sched_class;
3850	queued = task_on_rq_queued(p);
3851	running = task_current(rq, p);
3852	if (queued)
3853		dequeue_task(rq, p, queue_flag);
3854	if (running)
3855		put_prev_task(rq, p);
3856
3857	/*
3858	 * Boosting condition are:
3859	 * 1. -rt task is running and holds mutex A
3860	 *      --> -dl task blocks on mutex A
3861	 *
3862	 * 2. -dl task is running and holds mutex A
3863	 *      --> -dl task blocks on mutex A and could preempt the
3864	 *          running task
3865	 */
3866	if (dl_prio(prio)) {
3867		if (!dl_prio(p->normal_prio) ||
3868		    (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3869			p->dl.dl_boosted = 1;
 
3870			queue_flag |= ENQUEUE_REPLENISH;
3871		} else
3872			p->dl.dl_boosted = 0;
3873		p->sched_class = &dl_sched_class;
3874	} else if (rt_prio(prio)) {
3875		if (dl_prio(oldprio))
3876			p->dl.dl_boosted = 0;
3877		if (oldprio < prio)
3878			queue_flag |= ENQUEUE_HEAD;
3879		p->sched_class = &rt_sched_class;
3880	} else {
3881		if (dl_prio(oldprio))
3882			p->dl.dl_boosted = 0;
3883		if (rt_prio(oldprio))
3884			p->rt.timeout = 0;
3885		p->sched_class = &fair_sched_class;
3886	}
3887
3888	p->prio = prio;
3889
3890	if (queued)
3891		enqueue_task(rq, p, queue_flag);
3892	if (running)
3893		set_curr_task(rq, p);
3894
3895	check_class_changed(rq, p, prev_class, oldprio);
3896out_unlock:
3897	/* Avoid rq from going away on us: */
3898	preempt_disable();
3899	__task_rq_unlock(rq, &rf);
3900
3901	balance_callback(rq);
 
 
 
3902	preempt_enable();
3903}
3904#else
3905static inline int rt_effective_prio(struct task_struct *p, int prio)
3906{
3907	return prio;
3908}
3909#endif
3910
3911void set_user_nice(struct task_struct *p, long nice)
3912{
3913	bool queued, running;
3914	int old_prio, delta;
3915	struct rq_flags rf;
3916	struct rq *rq;
 
3917
3918	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3919		return;
3920	/*
3921	 * We have to be careful, if called from sys_setpriority(),
3922	 * the task might be in the middle of scheduling on another CPU.
3923	 */
3924	rq = task_rq_lock(p, &rf);
 
 
3925	update_rq_clock(rq);
3926
3927	/*
3928	 * The RT priorities are set via sched_setscheduler(), but we still
3929	 * allow the 'normal' nice value to be set - but as expected
3930	 * it wont have any effect on scheduling until the task is
3931	 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3932	 */
3933	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3934		p->static_prio = NICE_TO_PRIO(nice);
3935		goto out_unlock;
3936	}
 
3937	queued = task_on_rq_queued(p);
3938	running = task_current(rq, p);
3939	if (queued)
3940		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3941	if (running)
3942		put_prev_task(rq, p);
3943
3944	p->static_prio = NICE_TO_PRIO(nice);
3945	set_load_weight(p, true);
3946	old_prio = p->prio;
3947	p->prio = effective_prio(p);
3948	delta = p->prio - old_prio;
3949
3950	if (queued) {
3951		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3952		/*
3953		 * If the task increased its priority or is running and
3954		 * lowered its priority, then reschedule its CPU:
3955		 */
3956		if (delta < 0 || (delta > 0 && task_running(rq, p)))
3957			resched_curr(rq);
3958	}
3959	if (running)
3960		set_curr_task(rq, p);
3961out_unlock:
3962	task_rq_unlock(rq, p, &rf);
 
 
 
 
3963}
3964EXPORT_SYMBOL(set_user_nice);
3965
3966/*
3967 * can_nice - check if a task can reduce its nice value
 
 
 
3968 * @p: task
3969 * @nice: nice value
3970 */
3971int can_nice(const struct task_struct *p, const int nice)
3972{
3973	/* Convert nice value [19,-20] to rlimit style value [1,40]: */
3974	int nice_rlim = nice_to_rlimit(nice);
3975
3976	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3977		capable(CAP_SYS_NICE));
 
 
 
 
 
 
 
 
 
3978}
3979
3980#ifdef __ARCH_WANT_SYS_NICE
3981
3982/*
3983 * sys_nice - change the priority of the current process.
3984 * @increment: priority increment
3985 *
3986 * sys_setpriority is a more generic, but much slower function that
3987 * does similar things.
3988 */
3989SYSCALL_DEFINE1(nice, int, increment)
3990{
3991	long nice, retval;
3992
3993	/*
3994	 * Setpriority might change our priority at the same moment.
3995	 * We don't have to worry. Conceptually one call occurs first
3996	 * and we have a single winner.
3997	 */
3998	increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3999	nice = task_nice(current) + increment;
4000
4001	nice = clamp_val(nice, MIN_NICE, MAX_NICE);
4002	if (increment < 0 && !can_nice(current, nice))
4003		return -EPERM;
4004
4005	retval = security_task_setnice(current, nice);
4006	if (retval)
4007		return retval;
4008
4009	set_user_nice(current, nice);
4010	return 0;
4011}
4012
4013#endif
4014
4015/**
4016 * task_prio - return the priority value of a given task.
4017 * @p: the task in question.
4018 *
4019 * Return: The priority value as seen by users in /proc.
4020 * RT tasks are offset by -200. Normal tasks are centered
4021 * around 0, value goes from -16 to +15.
 
 
 
 
4022 */
4023int task_prio(const struct task_struct *p)
4024{
4025	return p->prio - MAX_RT_PRIO;
4026}
4027
4028/**
4029 * idle_cpu - is a given CPU idle currently?
4030 * @cpu: the processor in question.
4031 *
4032 * Return: 1 if the CPU is currently idle. 0 otherwise.
4033 */
4034int idle_cpu(int cpu)
4035{
4036	struct rq *rq = cpu_rq(cpu);
4037
4038	if (rq->curr != rq->idle)
4039		return 0;
4040
4041	if (rq->nr_running)
4042		return 0;
4043
4044#ifdef CONFIG_SMP
4045	if (!llist_empty(&rq->wake_list))
4046		return 0;
4047#endif
4048
4049	return 1;
4050}
4051
4052/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4053 * idle_task - return the idle task for a given CPU.
4054 * @cpu: the processor in question.
4055 *
4056 * Return: The idle task for the CPU @cpu.
4057 */
4058struct task_struct *idle_task(int cpu)
4059{
4060	return cpu_rq(cpu)->idle;
4061}
4062
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4063/**
4064 * find_process_by_pid - find a process with a matching PID value.
4065 * @pid: the pid in question.
4066 *
4067 * The task of @pid, if found. %NULL otherwise.
4068 */
4069static struct task_struct *find_process_by_pid(pid_t pid)
4070{
4071	return pid ? find_task_by_vpid(pid) : current;
4072}
4073
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4074/*
4075 * sched_setparam() passes in -1 for its policy, to let the functions
4076 * it calls know not to change it.
4077 */
4078#define SETPARAM_POLICY	-1
4079
4080static void __setscheduler_params(struct task_struct *p,
4081		const struct sched_attr *attr)
4082{
4083	int policy = attr->sched_policy;
4084
4085	if (policy == SETPARAM_POLICY)
4086		policy = p->policy;
4087
4088	p->policy = policy;
4089
4090	if (dl_policy(policy))
4091		__setparam_dl(p, attr);
4092	else if (fair_policy(policy))
4093		p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4094
4095	/*
4096	 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4097	 * !rt_policy. Always setting this ensures that things like
4098	 * getparam()/getattr() don't report silly values for !rt tasks.
4099	 */
4100	p->rt_priority = attr->sched_priority;
4101	p->normal_prio = normal_prio(p);
4102	set_load_weight(p, true);
4103}
4104
4105/* Actually do priority change: must hold pi & rq lock. */
4106static void __setscheduler(struct rq *rq, struct task_struct *p,
4107			   const struct sched_attr *attr, bool keep_boost)
4108{
4109	__setscheduler_params(p, attr);
4110
4111	/*
4112	 * Keep a potential priority boosting if called from
4113	 * sched_setscheduler().
4114	 */
4115	p->prio = normal_prio(p);
4116	if (keep_boost)
4117		p->prio = rt_effective_prio(p, p->prio);
4118
4119	if (dl_prio(p->prio))
4120		p->sched_class = &dl_sched_class;
4121	else if (rt_prio(p->prio))
4122		p->sched_class = &rt_sched_class;
4123	else
4124		p->sched_class = &fair_sched_class;
4125}
4126
4127/*
4128 * Check the target process has a UID that matches the current process's:
4129 */
4130static bool check_same_owner(struct task_struct *p)
4131{
4132	const struct cred *cred = current_cred(), *pcred;
4133	bool match;
4134
4135	rcu_read_lock();
4136	pcred = __task_cred(p);
4137	match = (uid_eq(cred->euid, pcred->euid) ||
4138		 uid_eq(cred->euid, pcred->uid));
4139	rcu_read_unlock();
4140	return match;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4141}
4142
4143static int __sched_setscheduler(struct task_struct *p,
4144				const struct sched_attr *attr,
4145				bool user, bool pi)
4146{
4147	int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4148		      MAX_RT_PRIO - 1 - attr->sched_priority;
4149	int retval, oldprio, oldpolicy = -1, queued, running;
4150	int new_effective_prio, policy = attr->sched_policy;
4151	const struct sched_class *prev_class;
 
4152	struct rq_flags rf;
4153	int reset_on_fork;
4154	int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4155	struct rq *rq;
 
4156
4157	/* The pi code expects interrupts enabled */
4158	BUG_ON(pi && in_interrupt());
4159recheck:
4160	/* Double check policy once rq lock held: */
4161	if (policy < 0) {
4162		reset_on_fork = p->sched_reset_on_fork;
4163		policy = oldpolicy = p->policy;
4164	} else {
4165		reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4166
4167		if (!valid_policy(policy))
4168			return -EINVAL;
4169	}
4170
4171	if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4172		return -EINVAL;
4173
4174	/*
4175	 * Valid priorities for SCHED_FIFO and SCHED_RR are
4176	 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4177	 * SCHED_BATCH and SCHED_IDLE is 0.
4178	 */
4179	if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4180	    (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4181		return -EINVAL;
4182	if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4183	    (rt_policy(policy) != (attr->sched_priority != 0)))
4184		return -EINVAL;
4185
4186	/*
4187	 * Allow unprivileged RT tasks to decrease priority:
4188	 */
4189	if (user && !capable(CAP_SYS_NICE)) {
4190		if (fair_policy(policy)) {
4191			if (attr->sched_nice < task_nice(p) &&
4192			    !can_nice(p, attr->sched_nice))
4193				return -EPERM;
4194		}
4195
4196		if (rt_policy(policy)) {
4197			unsigned long rlim_rtprio =
4198					task_rlimit(p, RLIMIT_RTPRIO);
4199
4200			/* Can't set/change the rt policy: */
4201			if (policy != p->policy && !rlim_rtprio)
4202				return -EPERM;
4203
4204			/* Can't increase priority: */
4205			if (attr->sched_priority > p->rt_priority &&
4206			    attr->sched_priority > rlim_rtprio)
4207				return -EPERM;
4208		}
4209
4210		 /*
4211		  * Can't set/change SCHED_DEADLINE policy at all for now
4212		  * (safest behavior); in the future we would like to allow
4213		  * unprivileged DL tasks to increase their relative deadline
4214		  * or reduce their runtime (both ways reducing utilization)
4215		  */
4216		if (dl_policy(policy))
4217			return -EPERM;
4218
4219		/*
4220		 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4221		 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4222		 */
4223		if (idle_policy(p->policy) && !idle_policy(policy)) {
4224			if (!can_nice(p, task_nice(p)))
4225				return -EPERM;
4226		}
4227
4228		/* Can't change other user's priorities: */
4229		if (!check_same_owner(p))
4230			return -EPERM;
4231
4232		/* Normal users shall not reset the sched_reset_on_fork flag: */
4233		if (p->sched_reset_on_fork && !reset_on_fork)
4234			return -EPERM;
4235	}
4236
4237	if (user) {
 
 
 
 
4238		if (attr->sched_flags & SCHED_FLAG_SUGOV)
4239			return -EINVAL;
4240
4241		retval = security_task_setscheduler(p);
4242		if (retval)
4243			return retval;
4244	}
4245
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4246	/*
4247	 * Make sure no PI-waiters arrive (or leave) while we are
4248	 * changing the priority of the task:
4249	 *
4250	 * To be able to change p->policy safely, the appropriate
4251	 * runqueue lock must be held.
4252	 */
4253	rq = task_rq_lock(p, &rf);
4254	update_rq_clock(rq);
4255
4256	/*
4257	 * Changing the policy of the stop threads its a very bad idea:
4258	 */
4259	if (p == rq->stop) {
4260		task_rq_unlock(rq, p, &rf);
4261		return -EINVAL;
4262	}
4263
4264	/*
4265	 * If not changing anything there's no need to proceed further,
4266	 * but store a possible modification of reset_on_fork.
4267	 */
4268	if (unlikely(policy == p->policy)) {
4269		if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4270			goto change;
4271		if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4272			goto change;
4273		if (dl_policy(policy) && dl_param_changed(p, attr))
4274			goto change;
 
 
4275
4276		p->sched_reset_on_fork = reset_on_fork;
4277		task_rq_unlock(rq, p, &rf);
4278		return 0;
4279	}
4280change:
4281
4282	if (user) {
4283#ifdef CONFIG_RT_GROUP_SCHED
4284		/*
4285		 * Do not allow realtime tasks into groups that have no runtime
4286		 * assigned.
4287		 */
4288		if (rt_bandwidth_enabled() && rt_policy(policy) &&
4289				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4290				!task_group_is_autogroup(task_group(p))) {
4291			task_rq_unlock(rq, p, &rf);
4292			return -EPERM;
4293		}
4294#endif
4295#ifdef CONFIG_SMP
4296		if (dl_bandwidth_enabled() && dl_policy(policy) &&
4297				!(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4298			cpumask_t *span = rq->rd->span;
4299
4300			/*
4301			 * Don't allow tasks with an affinity mask smaller than
4302			 * the entire root_domain to become SCHED_DEADLINE. We
4303			 * will also fail if there's no bandwidth available.
4304			 */
4305			if (!cpumask_subset(span, &p->cpus_allowed) ||
4306			    rq->rd->dl_bw.bw == 0) {
4307				task_rq_unlock(rq, p, &rf);
4308				return -EPERM;
4309			}
4310		}
4311#endif
4312	}
4313
4314	/* Re-check policy now with rq lock held: */
4315	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4316		policy = oldpolicy = -1;
4317		task_rq_unlock(rq, p, &rf);
 
 
4318		goto recheck;
4319	}
4320
4321	/*
4322	 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4323	 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4324	 * is available.
4325	 */
4326	if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4327		task_rq_unlock(rq, p, &rf);
4328		return -EBUSY;
4329	}
4330
4331	p->sched_reset_on_fork = reset_on_fork;
4332	oldprio = p->prio;
4333
 
4334	if (pi) {
4335		/*
4336		 * Take priority boosted tasks into account. If the new
4337		 * effective priority is unchanged, we just store the new
4338		 * normal parameters and do not touch the scheduler class and
4339		 * the runqueue. This will be done when the task deboost
4340		 * itself.
4341		 */
4342		new_effective_prio = rt_effective_prio(p, newprio);
4343		if (new_effective_prio == oldprio)
4344			queue_flags &= ~DEQUEUE_MOVE;
4345	}
4346
4347	queued = task_on_rq_queued(p);
4348	running = task_current(rq, p);
4349	if (queued)
4350		dequeue_task(rq, p, queue_flags);
4351	if (running)
4352		put_prev_task(rq, p);
4353
4354	prev_class = p->sched_class;
4355	__setscheduler(rq, p, attr, pi);
 
 
 
 
 
4356
4357	if (queued) {
4358		/*
4359		 * We enqueue to tail when the priority of a task is
4360		 * increased (user space view).
4361		 */
4362		if (oldprio < p->prio)
4363			queue_flags |= ENQUEUE_HEAD;
4364
4365		enqueue_task(rq, p, queue_flags);
4366	}
4367	if (running)
4368		set_curr_task(rq, p);
4369
4370	check_class_changed(rq, p, prev_class, oldprio);
4371
4372	/* Avoid rq from going away on us: */
4373	preempt_disable();
 
4374	task_rq_unlock(rq, p, &rf);
4375
4376	if (pi)
 
 
4377		rt_mutex_adjust_pi(p);
 
4378
4379	/* Run balance callbacks after we've adjusted the PI chain: */
4380	balance_callback(rq);
4381	preempt_enable();
4382
4383	return 0;
 
 
 
 
 
 
4384}
4385
4386static int _sched_setscheduler(struct task_struct *p, int policy,
4387			       const struct sched_param *param, bool check)
4388{
4389	struct sched_attr attr = {
4390		.sched_policy   = policy,
4391		.sched_priority = param->sched_priority,
4392		.sched_nice	= PRIO_TO_NICE(p->static_prio),
4393	};
4394
4395	/* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4396	if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4397		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4398		policy &= ~SCHED_RESET_ON_FORK;
4399		attr.sched_policy = policy;
4400	}
4401
4402	return __sched_setscheduler(p, &attr, check, true);
4403}
4404/**
4405 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4406 * @p: the task in question.
4407 * @policy: new policy.
4408 * @param: structure containing the new RT priority.
4409 *
 
 
4410 * Return: 0 on success. An error code otherwise.
4411 *
4412 * NOTE that the task may be already dead.
4413 */
4414int sched_setscheduler(struct task_struct *p, int policy,
4415		       const struct sched_param *param)
4416{
4417	return _sched_setscheduler(p, policy, param, true);
4418}
4419EXPORT_SYMBOL_GPL(sched_setscheduler);
4420
4421int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4422{
4423	return __sched_setscheduler(p, attr, true, true);
4424}
4425EXPORT_SYMBOL_GPL(sched_setattr);
4426
4427int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4428{
4429	return __sched_setscheduler(p, attr, false, true);
4430}
 
4431
4432/**
4433 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4434 * @p: the task in question.
4435 * @policy: new policy.
4436 * @param: structure containing the new RT priority.
4437 *
4438 * Just like sched_setscheduler, only don't bother checking if the
4439 * current context has permission.  For example, this is needed in
4440 * stop_machine(): we create temporary high priority worker threads,
4441 * but our caller might not have that capability.
4442 *
4443 * Return: 0 on success. An error code otherwise.
4444 */
4445int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4446			       const struct sched_param *param)
4447{
4448	return _sched_setscheduler(p, policy, param, false);
4449}
4450EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4451
4452static int
4453do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4454{
4455	struct sched_param lparam;
4456	struct task_struct *p;
4457	int retval;
4458
4459	if (!param || pid < 0)
4460		return -EINVAL;
4461	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4462		return -EFAULT;
4463
4464	rcu_read_lock();
4465	retval = -ESRCH;
4466	p = find_process_by_pid(pid);
4467	if (p != NULL)
4468		retval = sched_setscheduler(p, policy, &lparam);
4469	rcu_read_unlock();
4470
4471	return retval;
4472}
4473
4474/*
4475 * Mimics kernel/events/core.c perf_copy_attr().
4476 */
4477static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4478{
4479	u32 size;
4480	int ret;
4481
4482	if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4483		return -EFAULT;
4484
4485	/* Zero the full structure, so that a short copy will be nice: */
4486	memset(attr, 0, sizeof(*attr));
4487
4488	ret = get_user(size, &uattr->size);
4489	if (ret)
4490		return ret;
4491
4492	/* Bail out on silly large: */
4493	if (size > PAGE_SIZE)
4494		goto err_size;
4495
4496	/* ABI compatibility quirk: */
4497	if (!size)
4498		size = SCHED_ATTR_SIZE_VER0;
4499
4500	if (size < SCHED_ATTR_SIZE_VER0)
4501		goto err_size;
4502
4503	/*
4504	 * If we're handed a bigger struct than we know of,
4505	 * ensure all the unknown bits are 0 - i.e. new
4506	 * user-space does not rely on any kernel feature
4507	 * extensions we dont know about yet.
4508	 */
4509	if (size > sizeof(*attr)) {
4510		unsigned char __user *addr;
4511		unsigned char __user *end;
4512		unsigned char val;
4513
4514		addr = (void __user *)uattr + sizeof(*attr);
4515		end  = (void __user *)uattr + size;
4516
4517		for (; addr < end; addr++) {
4518			ret = get_user(val, addr);
4519			if (ret)
4520				return ret;
4521			if (val)
4522				goto err_size;
4523		}
4524		size = sizeof(*attr);
4525	}
4526
4527	ret = copy_from_user(attr, uattr, size);
4528	if (ret)
4529		return -EFAULT;
4530
4531	/*
4532	 * XXX: Do we want to be lenient like existing syscalls; or do we want
4533	 * to be strict and return an error on out-of-bounds values?
4534	 */
4535	attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4536
4537	return 0;
4538
4539err_size:
4540	put_user(sizeof(*attr), &uattr->size);
4541	return -E2BIG;
4542}
4543
 
 
 
 
 
 
 
 
 
 
4544/**
4545 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4546 * @pid: the pid in question.
4547 * @policy: new policy.
4548 * @param: structure containing the new RT priority.
4549 *
4550 * Return: 0 on success. An error code otherwise.
4551 */
4552SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4553{
4554	if (policy < 0)
4555		return -EINVAL;
4556
4557	return do_sched_setscheduler(pid, policy, param);
4558}
4559
4560/**
4561 * sys_sched_setparam - set/change the RT priority of a thread
4562 * @pid: the pid in question.
4563 * @param: structure containing the new RT priority.
4564 *
4565 * Return: 0 on success. An error code otherwise.
4566 */
4567SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4568{
4569	return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4570}
4571
4572/**
4573 * sys_sched_setattr - same as above, but with extended sched_attr
4574 * @pid: the pid in question.
4575 * @uattr: structure containing the extended parameters.
4576 * @flags: for future extension.
4577 */
4578SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4579			       unsigned int, flags)
4580{
4581	struct sched_attr attr;
4582	struct task_struct *p;
4583	int retval;
4584
4585	if (!uattr || pid < 0 || flags)
4586		return -EINVAL;
4587
4588	retval = sched_copy_attr(uattr, &attr);
4589	if (retval)
4590		return retval;
4591
4592	if ((int)attr.sched_policy < 0)
4593		return -EINVAL;
 
 
4594
4595	rcu_read_lock();
4596	retval = -ESRCH;
4597	p = find_process_by_pid(pid);
4598	if (p != NULL)
4599		retval = sched_setattr(p, &attr);
4600	rcu_read_unlock();
4601
4602	return retval;
 
 
 
4603}
4604
4605/**
4606 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4607 * @pid: the pid in question.
4608 *
4609 * Return: On success, the policy of the thread. Otherwise, a negative error
4610 * code.
4611 */
4612SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4613{
4614	struct task_struct *p;
4615	int retval;
4616
4617	if (pid < 0)
4618		return -EINVAL;
4619
4620	retval = -ESRCH;
4621	rcu_read_lock();
4622	p = find_process_by_pid(pid);
4623	if (p) {
4624		retval = security_task_getscheduler(p);
4625		if (!retval)
4626			retval = p->policy
4627				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
 
 
 
4628	}
4629	rcu_read_unlock();
4630	return retval;
4631}
4632
4633/**
4634 * sys_sched_getparam - get the RT priority of a thread
4635 * @pid: the pid in question.
4636 * @param: structure containing the RT priority.
4637 *
4638 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4639 * code.
4640 */
4641SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4642{
4643	struct sched_param lp = { .sched_priority = 0 };
4644	struct task_struct *p;
4645	int retval;
4646
4647	if (!param || pid < 0)
4648		return -EINVAL;
4649
4650	rcu_read_lock();
4651	p = find_process_by_pid(pid);
4652	retval = -ESRCH;
4653	if (!p)
4654		goto out_unlock;
4655
4656	retval = security_task_getscheduler(p);
4657	if (retval)
4658		goto out_unlock;
4659
4660	if (task_has_rt_policy(p))
4661		lp.sched_priority = p->rt_priority;
4662	rcu_read_unlock();
4663
4664	/*
4665	 * This one might sleep, we cannot do it with a spinlock held ...
4666	 */
4667	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4668
4669	return retval;
4670
4671out_unlock:
4672	rcu_read_unlock();
4673	return retval;
4674}
4675
4676static int sched_read_attr(struct sched_attr __user *uattr,
4677			   struct sched_attr *attr,
4678			   unsigned int usize)
 
 
 
 
 
 
 
 
 
4679{
4680	int ret;
4681
4682	if (!access_ok(VERIFY_WRITE, uattr, usize))
4683		return -EFAULT;
4684
4685	/*
4686	 * If we're handed a smaller struct than we know of,
4687	 * ensure all the unknown bits are 0 - i.e. old
4688	 * user-space does not get uncomplete information.
 
 
 
 
 
 
 
 
4689	 */
4690	if (usize < sizeof(*attr)) {
4691		unsigned char *addr;
4692		unsigned char *end;
4693
4694		addr = (void *)attr + usize;
4695		end  = (void *)attr + sizeof(*attr);
4696
4697		for (; addr < end; addr++) {
4698			if (*addr)
4699				return -EFBIG;
4700		}
4701
4702		attr->size = usize;
4703	}
4704
4705	ret = copy_to_user(uattr, attr, attr->size);
4706	if (ret)
4707		return -EFAULT;
4708
4709	return 0;
4710}
4711
4712/**
4713 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4714 * @pid: the pid in question.
4715 * @uattr: structure containing the extended parameters.
4716 * @size: sizeof(attr) for fwd/bwd comp.
4717 * @flags: for future extension.
4718 */
4719SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4720		unsigned int, size, unsigned int, flags)
4721{
4722	struct sched_attr attr = {
4723		.size = sizeof(struct sched_attr),
4724	};
4725	struct task_struct *p;
4726	int retval;
4727
4728	if (!uattr || pid < 0 || size > PAGE_SIZE ||
4729	    size < SCHED_ATTR_SIZE_VER0 || flags)
4730		return -EINVAL;
4731
4732	rcu_read_lock();
4733	p = find_process_by_pid(pid);
4734	retval = -ESRCH;
4735	if (!p)
4736		goto out_unlock;
4737
4738	retval = security_task_getscheduler(p);
4739	if (retval)
4740		goto out_unlock;
4741
4742	attr.sched_policy = p->policy;
4743	if (p->sched_reset_on_fork)
4744		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4745	if (task_has_dl_policy(p))
4746		__getparam_dl(p, &attr);
4747	else if (task_has_rt_policy(p))
4748		attr.sched_priority = p->rt_priority;
4749	else
4750		attr.sched_nice = task_nice(p);
4751
4752	rcu_read_unlock();
 
 
 
 
4753
4754	retval = sched_read_attr(uattr, &attr, size);
4755	return retval;
 
 
 
 
 
 
 
 
4756
4757out_unlock:
4758	rcu_read_unlock();
4759	return retval;
4760}
4761
4762long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
 
4763{
4764	cpumask_var_t cpus_allowed, new_mask;
4765	struct task_struct *p;
4766	int retval;
 
 
 
4767
4768	rcu_read_lock();
 
 
 
 
 
 
 
 
4769
4770	p = find_process_by_pid(pid);
4771	if (!p) {
4772		rcu_read_unlock();
4773		return -ESRCH;
4774	}
4775
4776	/* Prevent p going away */
4777	get_task_struct(p);
4778	rcu_read_unlock();
 
 
 
 
 
4779
4780	if (p->flags & PF_NO_SETAFFINITY) {
4781		retval = -EINVAL;
4782		goto out_put_task;
4783	}
4784	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4785		retval = -ENOMEM;
4786		goto out_put_task;
4787	}
4788	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4789		retval = -ENOMEM;
4790		goto out_free_cpus_allowed;
4791	}
4792	retval = -EPERM;
4793	if (!check_same_owner(p)) {
4794		rcu_read_lock();
4795		if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4796			rcu_read_unlock();
4797			goto out_free_new_mask;
4798		}
4799		rcu_read_unlock();
4800	}
4801
4802	retval = security_task_setscheduler(p);
 
 
 
 
 
 
4803	if (retval)
4804		goto out_free_new_mask;
4805
 
 
 
4806
4807	cpuset_cpus_allowed(p, cpus_allowed);
4808	cpumask_and(new_mask, in_mask, cpus_allowed);
 
 
 
 
 
4809
4810	/*
4811	 * Since bandwidth control happens on root_domain basis,
4812	 * if admission test is enabled, we only admit -deadline
4813	 * tasks allowed to run on all the CPUs in the task's
4814	 * root_domain.
4815	 */
4816#ifdef CONFIG_SMP
4817	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4818		rcu_read_lock();
4819		if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4820			retval = -EBUSY;
4821			rcu_read_unlock();
4822			goto out_free_new_mask;
4823		}
4824		rcu_read_unlock();
4825	}
4826#endif
4827again:
4828	retval = __set_cpus_allowed_ptr(p, new_mask, true);
4829
4830	if (!retval) {
4831		cpuset_cpus_allowed(p, cpus_allowed);
4832		if (!cpumask_subset(new_mask, cpus_allowed)) {
4833			/*
4834			 * We must have raced with a concurrent cpuset
4835			 * update. Just reset the cpus_allowed to the
4836			 * cpuset's cpus_allowed
4837			 */
4838			cpumask_copy(new_mask, cpus_allowed);
4839			goto again;
4840		}
 
 
4841	}
 
4842out_free_new_mask:
4843	free_cpumask_var(new_mask);
4844out_free_cpus_allowed:
4845	free_cpumask_var(cpus_allowed);
4846out_put_task:
4847	put_task_struct(p);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4848	return retval;
4849}
4850
4851static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4852			     struct cpumask *new_mask)
4853{
4854	if (len < cpumask_size())
4855		cpumask_clear(new_mask);
4856	else if (len > cpumask_size())
4857		len = cpumask_size();
4858
4859	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4860}
4861
4862/**
4863 * sys_sched_setaffinity - set the CPU affinity of a process
4864 * @pid: pid of the process
4865 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4866 * @user_mask_ptr: user-space pointer to the new CPU mask
4867 *
4868 * Return: 0 on success. An error code otherwise.
4869 */
4870SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4871		unsigned long __user *, user_mask_ptr)
4872{
4873	cpumask_var_t new_mask;
4874	int retval;
4875
4876	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4877		return -ENOMEM;
4878
4879	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4880	if (retval == 0)
4881		retval = sched_setaffinity(pid, new_mask);
4882	free_cpumask_var(new_mask);
4883	return retval;
4884}
4885
4886long sched_getaffinity(pid_t pid, struct cpumask *mask)
4887{
4888	struct task_struct *p;
4889	unsigned long flags;
4890	int retval;
4891
4892	rcu_read_lock();
4893
4894	retval = -ESRCH;
4895	p = find_process_by_pid(pid);
4896	if (!p)
4897		goto out_unlock;
4898
4899	retval = security_task_getscheduler(p);
4900	if (retval)
4901		goto out_unlock;
4902
4903	raw_spin_lock_irqsave(&p->pi_lock, flags);
4904	cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4905	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4906
4907out_unlock:
4908	rcu_read_unlock();
4909
4910	return retval;
4911}
4912
4913/**
4914 * sys_sched_getaffinity - get the CPU affinity of a process
4915 * @pid: pid of the process
4916 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4917 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4918 *
4919 * Return: size of CPU mask copied to user_mask_ptr on success. An
4920 * error code otherwise.
4921 */
4922SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4923		unsigned long __user *, user_mask_ptr)
4924{
4925	int ret;
4926	cpumask_var_t mask;
4927
4928	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4929		return -EINVAL;
4930	if (len & (sizeof(unsigned long)-1))
4931		return -EINVAL;
4932
4933	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4934		return -ENOMEM;
4935
4936	ret = sched_getaffinity(pid, mask);
4937	if (ret == 0) {
4938		unsigned int retlen = min(len, cpumask_size());
4939
4940		if (copy_to_user(user_mask_ptr, mask, retlen))
4941			ret = -EFAULT;
4942		else
4943			ret = retlen;
4944	}
4945	free_cpumask_var(mask);
4946
4947	return ret;
4948}
4949
4950/**
4951 * sys_sched_yield - yield the current processor to other threads.
4952 *
4953 * This function yields the current CPU to other tasks. If there are no
4954 * other threads running on this CPU then this function will return.
4955 *
4956 * Return: 0.
4957 */
4958static void do_sched_yield(void)
4959{
4960	struct rq_flags rf;
4961	struct rq *rq;
4962
4963	local_irq_disable();
4964	rq = this_rq();
4965	rq_lock(rq, &rf);
4966
4967	schedstat_inc(rq->yld_count);
4968	current->sched_class->yield_task(rq);
4969
4970	/*
4971	 * Since we are going to call schedule() anyway, there's
4972	 * no need to preempt or enable interrupts:
4973	 */
4974	preempt_disable();
4975	rq_unlock(rq, &rf);
4976	sched_preempt_enable_no_resched();
4977
4978	schedule();
4979}
4980
 
 
 
 
 
 
 
 
4981SYSCALL_DEFINE0(sched_yield)
4982{
4983	do_sched_yield();
4984	return 0;
4985}
4986
4987#ifndef CONFIG_PREEMPT
4988int __sched _cond_resched(void)
4989{
4990	if (should_resched(0)) {
4991		preempt_schedule_common();
4992		return 1;
4993	}
 
 
 
 
 
 
 
 
 
 
 
 
4994	rcu_all_qs();
 
4995	return 0;
4996}
4997EXPORT_SYMBOL(_cond_resched);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4998#endif
4999
5000/*
5001 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5002 * call schedule, and on return reacquire the lock.
5003 *
5004 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5005 * operations here to prevent schedule() from being called twice (once via
5006 * spin_unlock(), once by hand).
5007 */
5008int __cond_resched_lock(spinlock_t *lock)
5009{
5010	int resched = should_resched(PREEMPT_LOCK_OFFSET);
5011	int ret = 0;
5012
5013	lockdep_assert_held(lock);
5014
5015	if (spin_needbreak(lock) || resched) {
5016		spin_unlock(lock);
5017		if (resched)
5018			preempt_schedule_common();
5019		else
5020			cpu_relax();
5021		ret = 1;
5022		spin_lock(lock);
5023	}
5024	return ret;
5025}
5026EXPORT_SYMBOL(__cond_resched_lock);
5027
5028int __sched __cond_resched_softirq(void)
5029{
5030	BUG_ON(!in_softirq());
 
5031
5032	if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
5033		local_bh_enable();
5034		preempt_schedule_common();
5035		local_bh_disable();
5036		return 1;
 
 
 
5037	}
5038	return 0;
5039}
5040EXPORT_SYMBOL(__cond_resched_softirq);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5041
5042/**
5043 * yield - yield the current processor to other threads.
5044 *
5045 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5046 *
5047 * The scheduler is at all times free to pick the calling task as the most
5048 * eligible task to run, if removing the yield() call from your code breaks
5049 * it, its already broken.
5050 *
5051 * Typical broken usage is:
5052 *
5053 * while (!event)
5054 *	yield();
5055 *
5056 * where one assumes that yield() will let 'the other' process run that will
5057 * make event true. If the current task is a SCHED_FIFO task that will never
5058 * happen. Never use yield() as a progress guarantee!!
5059 *
5060 * If you want to use yield() to wait for something, use wait_event().
5061 * If you want to use yield() to be 'nice' for others, use cond_resched().
5062 * If you still want to use yield(), do not!
5063 */
5064void __sched yield(void)
5065{
5066	set_current_state(TASK_RUNNING);
5067	do_sched_yield();
5068}
5069EXPORT_SYMBOL(yield);
5070
5071/**
5072 * yield_to - yield the current processor to another thread in
5073 * your thread group, or accelerate that thread toward the
5074 * processor it's on.
5075 * @p: target task
5076 * @preempt: whether task preemption is allowed or not
5077 *
5078 * It's the caller's job to ensure that the target task struct
5079 * can't go away on us before we can do any checks.
5080 *
5081 * Return:
5082 *	true (>0) if we indeed boosted the target task.
5083 *	false (0) if we failed to boost the target.
5084 *	-ESRCH if there's no task to yield to.
5085 */
5086int __sched yield_to(struct task_struct *p, bool preempt)
5087{
5088	struct task_struct *curr = current;
5089	struct rq *rq, *p_rq;
5090	unsigned long flags;
5091	int yielded = 0;
5092
5093	local_irq_save(flags);
5094	rq = this_rq();
5095
5096again:
5097	p_rq = task_rq(p);
5098	/*
5099	 * If we're the only runnable task on the rq and target rq also
5100	 * has only one task, there's absolutely no point in yielding.
5101	 */
5102	if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5103		yielded = -ESRCH;
5104		goto out_irq;
5105	}
5106
5107	double_rq_lock(rq, p_rq);
5108	if (task_rq(p) != p_rq) {
5109		double_rq_unlock(rq, p_rq);
5110		goto again;
5111	}
5112
5113	if (!curr->sched_class->yield_to_task)
5114		goto out_unlock;
5115
5116	if (curr->sched_class != p->sched_class)
5117		goto out_unlock;
5118
5119	if (task_running(p_rq, p) || p->state)
5120		goto out_unlock;
5121
5122	yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5123	if (yielded) {
5124		schedstat_inc(rq->yld_count);
5125		/*
5126		 * Make p's CPU reschedule; pick_next_entity takes care of
5127		 * fairness.
5128		 */
5129		if (preempt && rq != p_rq)
5130			resched_curr(p_rq);
 
5131	}
5132
5133out_unlock:
5134	double_rq_unlock(rq, p_rq);
5135out_irq:
5136	local_irq_restore(flags);
5137
5138	if (yielded > 0)
5139		schedule();
5140
5141	return yielded;
5142}
5143EXPORT_SYMBOL_GPL(yield_to);
5144
5145int io_schedule_prepare(void)
5146{
5147	int old_iowait = current->in_iowait;
5148
5149	current->in_iowait = 1;
5150	blk_schedule_flush_plug(current);
5151
5152	return old_iowait;
5153}
5154
5155void io_schedule_finish(int token)
5156{
5157	current->in_iowait = token;
5158}
5159
5160/*
5161 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5162 * that process accounting knows that this is a task in IO wait state.
5163 */
5164long __sched io_schedule_timeout(long timeout)
5165{
5166	int token;
5167	long ret;
5168
5169	token = io_schedule_prepare();
5170	ret = schedule_timeout(timeout);
5171	io_schedule_finish(token);
5172
5173	return ret;
5174}
5175EXPORT_SYMBOL(io_schedule_timeout);
5176
5177void io_schedule(void)
5178{
5179	int token;
5180
5181	token = io_schedule_prepare();
5182	schedule();
5183	io_schedule_finish(token);
5184}
5185EXPORT_SYMBOL(io_schedule);
5186
5187/**
5188 * sys_sched_get_priority_max - return maximum RT priority.
5189 * @policy: scheduling class.
5190 *
5191 * Return: On success, this syscall returns the maximum
5192 * rt_priority that can be used by a given scheduling class.
5193 * On failure, a negative error code is returned.
5194 */
5195SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5196{
5197	int ret = -EINVAL;
5198
5199	switch (policy) {
5200	case SCHED_FIFO:
5201	case SCHED_RR:
5202		ret = MAX_USER_RT_PRIO-1;
5203		break;
5204	case SCHED_DEADLINE:
5205	case SCHED_NORMAL:
5206	case SCHED_BATCH:
5207	case SCHED_IDLE:
5208		ret = 0;
5209		break;
5210	}
5211	return ret;
5212}
5213
5214/**
5215 * sys_sched_get_priority_min - return minimum RT priority.
5216 * @policy: scheduling class.
5217 *
5218 * Return: On success, this syscall returns the minimum
5219 * rt_priority that can be used by a given scheduling class.
5220 * On failure, a negative error code is returned.
5221 */
5222SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5223{
5224	int ret = -EINVAL;
5225
5226	switch (policy) {
5227	case SCHED_FIFO:
5228	case SCHED_RR:
5229		ret = 1;
5230		break;
5231	case SCHED_DEADLINE:
5232	case SCHED_NORMAL:
5233	case SCHED_BATCH:
5234	case SCHED_IDLE:
5235		ret = 0;
5236	}
5237	return ret;
5238}
5239
5240static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5241{
5242	struct task_struct *p;
5243	unsigned int time_slice;
5244	struct rq_flags rf;
5245	struct rq *rq;
5246	int retval;
5247
5248	if (pid < 0)
5249		return -EINVAL;
5250
5251	retval = -ESRCH;
5252	rcu_read_lock();
5253	p = find_process_by_pid(pid);
5254	if (!p)
5255		goto out_unlock;
5256
5257	retval = security_task_getscheduler(p);
5258	if (retval)
5259		goto out_unlock;
5260
5261	rq = task_rq_lock(p, &rf);
5262	time_slice = 0;
5263	if (p->sched_class->get_rr_interval)
5264		time_slice = p->sched_class->get_rr_interval(rq, p);
5265	task_rq_unlock(rq, p, &rf);
 
5266
5267	rcu_read_unlock();
5268	jiffies_to_timespec64(time_slice, t);
5269	return 0;
5270
5271out_unlock:
5272	rcu_read_unlock();
5273	return retval;
5274}
5275
5276/**
5277 * sys_sched_rr_get_interval - return the default timeslice of a process.
5278 * @pid: pid of the process.
5279 * @interval: userspace pointer to the timeslice value.
5280 *
5281 * this syscall writes the default timeslice value of a given process
5282 * into the user-space timespec buffer. A value of '0' means infinity.
5283 *
5284 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5285 * an error code.
5286 */
5287SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5288		struct timespec __user *, interval)
5289{
5290	struct timespec64 t;
5291	int retval = sched_rr_get_interval(pid, &t);
5292
5293	if (retval == 0)
5294		retval = put_timespec64(&t, interval);
5295
5296	return retval;
5297}
5298
5299#ifdef CONFIG_COMPAT
5300COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5301		       compat_pid_t, pid,
5302		       struct compat_timespec __user *, interval)
5303{
5304	struct timespec64 t;
5305	int retval = sched_rr_get_interval(pid, &t);
5306
5307	if (retval == 0)
5308		retval = compat_put_timespec64(&t, interval);
5309	return retval;
5310}
5311#endif
5312
5313void sched_show_task(struct task_struct *p)
5314{
5315	unsigned long free = 0;
5316	int ppid;
5317
5318	if (!try_get_task_stack(p))
5319		return;
5320
5321	printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5322
5323	if (p->state == TASK_RUNNING)
5324		printk(KERN_CONT "  running task    ");
5325#ifdef CONFIG_DEBUG_STACK_USAGE
5326	free = stack_not_used(p);
5327#endif
5328	ppid = 0;
5329	rcu_read_lock();
5330	if (pid_alive(p))
5331		ppid = task_pid_nr(rcu_dereference(p->real_parent));
5332	rcu_read_unlock();
5333	printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5334		task_pid_nr(p), ppid,
5335		(unsigned long)task_thread_info(p)->flags);
5336
5337	print_worker_info(KERN_INFO, p);
5338	show_stack(p, NULL);
 
5339	put_task_stack(p);
5340}
5341EXPORT_SYMBOL_GPL(sched_show_task);
5342
5343static inline bool
5344state_filter_match(unsigned long state_filter, struct task_struct *p)
5345{
 
 
5346	/* no filter, everything matches */
5347	if (!state_filter)
5348		return true;
5349
5350	/* filter, but doesn't match */
5351	if (!(p->state & state_filter))
5352		return false;
5353
5354	/*
5355	 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5356	 * TASK_KILLABLE).
5357	 */
5358	if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5359		return false;
5360
5361	return true;
5362}
5363
5364
5365void show_state_filter(unsigned long state_filter)
5366{
5367	struct task_struct *g, *p;
5368
5369#if BITS_PER_LONG == 32
5370	printk(KERN_INFO
5371		"  task                PC stack   pid father\n");
5372#else
5373	printk(KERN_INFO
5374		"  task                        PC stack   pid father\n");
5375#endif
5376	rcu_read_lock();
5377	for_each_process_thread(g, p) {
5378		/*
5379		 * reset the NMI-timeout, listing all files on a slow
5380		 * console might take a lot of time:
5381		 * Also, reset softlockup watchdogs on all CPUs, because
5382		 * another CPU might be blocked waiting for us to process
5383		 * an IPI.
5384		 */
5385		touch_nmi_watchdog();
5386		touch_all_softlockup_watchdogs();
5387		if (state_filter_match(state_filter, p))
5388			sched_show_task(p);
5389	}
5390
5391#ifdef CONFIG_SCHED_DEBUG
5392	if (!state_filter)
5393		sysrq_sched_debug_show();
5394#endif
5395	rcu_read_unlock();
5396	/*
5397	 * Only show locks if all tasks are dumped:
5398	 */
5399	if (!state_filter)
5400		debug_show_all_locks();
5401}
5402
5403/**
5404 * init_idle - set up an idle thread for a given CPU
5405 * @idle: task in question
5406 * @cpu: CPU the idle task belongs to
5407 *
5408 * NOTE: this function does not set the idle thread's NEED_RESCHED
5409 * flag, to make booting more robust.
5410 */
5411void init_idle(struct task_struct *idle, int cpu)
5412{
 
 
 
 
 
 
5413	struct rq *rq = cpu_rq(cpu);
5414	unsigned long flags;
5415
 
 
5416	raw_spin_lock_irqsave(&idle->pi_lock, flags);
5417	raw_spin_lock(&rq->lock);
5418
5419	__sched_fork(0, idle);
5420	idle->state = TASK_RUNNING;
5421	idle->se.exec_start = sched_clock();
5422	idle->flags |= PF_IDLE;
5423
5424	kasan_unpoison_task_stack(idle);
 
 
 
5425
5426#ifdef CONFIG_SMP
5427	/*
5428	 * Its possible that init_idle() gets called multiple times on a task,
5429	 * in that case do_set_cpus_allowed() will not do the right thing.
5430	 *
5431	 * And since this is boot we can forgo the serialization.
5432	 */
5433	set_cpus_allowed_common(idle, cpumask_of(cpu));
5434#endif
5435	/*
5436	 * We're having a chicken and egg problem, even though we are
5437	 * holding rq->lock, the CPU isn't yet set to this CPU so the
5438	 * lockdep check in task_group() will fail.
5439	 *
5440	 * Similar case to sched_fork(). / Alternatively we could
5441	 * use task_rq_lock() here and obtain the other rq->lock.
5442	 *
5443	 * Silence PROVE_RCU
5444	 */
5445	rcu_read_lock();
5446	__set_task_cpu(idle, cpu);
5447	rcu_read_unlock();
5448
5449	rq->curr = rq->idle = idle;
 
5450	idle->on_rq = TASK_ON_RQ_QUEUED;
5451#ifdef CONFIG_SMP
5452	idle->on_cpu = 1;
5453#endif
5454	raw_spin_unlock(&rq->lock);
5455	raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5456
5457	/* Set the preempt count _outside_ the spinlocks! */
5458	init_idle_preempt_count(idle, cpu);
5459
5460	/*
5461	 * The idle tasks have their own, simple scheduling class:
5462	 */
5463	idle->sched_class = &idle_sched_class;
5464	ftrace_graph_init_idle_task(idle, cpu);
5465	vtime_init_idle(idle, cpu);
5466#ifdef CONFIG_SMP
5467	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5468#endif
5469}
5470
5471#ifdef CONFIG_SMP
5472
5473int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5474			      const struct cpumask *trial)
5475{
5476	int ret = 1;
5477
5478	if (!cpumask_weight(cur))
5479		return ret;
5480
5481	ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5482
5483	return ret;
5484}
5485
5486int task_can_attach(struct task_struct *p,
5487		    const struct cpumask *cs_cpus_allowed)
5488{
5489	int ret = 0;
5490
5491	/*
5492	 * Kthreads which disallow setaffinity shouldn't be moved
5493	 * to a new cpuset; we don't want to change their CPU
5494	 * affinity and isolating such threads by their set of
5495	 * allowed nodes is unnecessary.  Thus, cpusets are not
5496	 * applicable for such threads.  This prevents checking for
5497	 * success of set_cpus_allowed_ptr() on all attached tasks
5498	 * before cpus_allowed may be changed.
5499	 */
5500	if (p->flags & PF_NO_SETAFFINITY) {
5501		ret = -EINVAL;
5502		goto out;
5503	}
5504
5505	if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5506					      cs_cpus_allowed))
5507		ret = dl_task_can_attach(p, cs_cpus_allowed);
5508
5509out:
5510	return ret;
5511}
5512
5513bool sched_smp_initialized __read_mostly;
5514
5515#ifdef CONFIG_NUMA_BALANCING
5516/* Migrate current task p to target_cpu */
5517int migrate_task_to(struct task_struct *p, int target_cpu)
5518{
5519	struct migration_arg arg = { p, target_cpu };
5520	int curr_cpu = task_cpu(p);
5521
5522	if (curr_cpu == target_cpu)
5523		return 0;
5524
5525	if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5526		return -EINVAL;
5527
5528	/* TODO: This is not properly updating schedstats */
5529
5530	trace_sched_move_numa(p, curr_cpu, target_cpu);
5531	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5532}
5533
5534/*
5535 * Requeue a task on a given node and accurately track the number of NUMA
5536 * tasks on the runqueues
5537 */
5538void sched_setnuma(struct task_struct *p, int nid)
5539{
5540	bool queued, running;
5541	struct rq_flags rf;
5542	struct rq *rq;
5543
5544	rq = task_rq_lock(p, &rf);
5545	queued = task_on_rq_queued(p);
5546	running = task_current(rq, p);
5547
5548	if (queued)
5549		dequeue_task(rq, p, DEQUEUE_SAVE);
5550	if (running)
5551		put_prev_task(rq, p);
5552
5553	p->numa_preferred_nid = nid;
5554
5555	if (queued)
5556		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5557	if (running)
5558		set_curr_task(rq, p);
5559	task_rq_unlock(rq, p, &rf);
5560}
5561#endif /* CONFIG_NUMA_BALANCING */
5562
5563#ifdef CONFIG_HOTPLUG_CPU
5564/*
5565 * Ensure that the idle task is using init_mm right before its CPU goes
5566 * offline.
5567 */
5568void idle_task_exit(void)
5569{
5570	struct mm_struct *mm = current->active_mm;
5571
5572	BUG_ON(cpu_online(smp_processor_id()));
 
5573
5574	if (mm != &init_mm) {
5575		switch_mm(mm, &init_mm, current);
5576		current->active_mm = &init_mm;
5577		finish_arch_post_lock_switch();
5578	}
5579	mmdrop(mm);
5580}
5581
5582/*
5583 * Since this CPU is going 'away' for a while, fold any nr_active delta
5584 * we might have. Assumes we're called after migrate_tasks() so that the
5585 * nr_active count is stable. We need to take the teardown thread which
5586 * is calling this into account, so we hand in adjust = 1 to the load
5587 * calculation.
5588 *
5589 * Also see the comment "Global load-average calculations".
5590 */
5591static void calc_load_migrate(struct rq *rq)
5592{
5593	long delta = calc_load_fold_active(rq, 1);
5594	if (delta)
5595		atomic_long_add(delta, &calc_load_tasks);
5596}
5597
5598static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5599{
5600}
 
 
 
5601
5602static const struct sched_class fake_sched_class = {
5603	.put_prev_task = put_prev_task_fake,
5604};
5605
5606static struct task_struct fake_task = {
5607	/*
5608	 * Avoid pull_{rt,dl}_task()
5609	 */
5610	.prio = MAX_PRIO + 1,
5611	.sched_class = &fake_sched_class,
5612};
 
 
 
 
 
 
 
 
 
5613
5614/*
5615 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5616 * try_to_wake_up()->select_task_rq().
5617 *
5618 * Called with rq->lock held even though we'er in stop_machine() and
5619 * there's no concurrency possible, we hold the required locks anyway
5620 * because of lock validation efforts.
5621 */
5622static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5623{
5624	struct rq *rq = dead_rq;
5625	struct task_struct *next, *stop = rq->stop;
5626	struct rq_flags orf = *rf;
5627	int dest_cpu;
5628
5629	/*
5630	 * Fudge the rq selection such that the below task selection loop
5631	 * doesn't get stuck on the currently eligible stop task.
5632	 *
5633	 * We're currently inside stop_machine() and the rq is either stuck
5634	 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5635	 * either way we should never end up calling schedule() until we're
5636	 * done here.
5637	 */
5638	rq->stop = NULL;
5639
5640	/*
5641	 * put_prev_task() and pick_next_task() sched
5642	 * class method both need to have an up-to-date
5643	 * value of rq->clock[_task]
5644	 */
5645	update_rq_clock(rq);
5646
5647	for (;;) {
5648		/*
5649		 * There's this thread running, bail when that's the only
5650		 * remaining thread:
5651		 */
5652		if (rq->nr_running == 1)
5653			break;
5654
5655		/*
5656		 * pick_next_task() assumes pinned rq->lock:
5657		 */
5658		next = pick_next_task(rq, &fake_task, rf);
5659		BUG_ON(!next);
5660		put_prev_task(rq, next);
5661
5662		/*
5663		 * Rules for changing task_struct::cpus_allowed are holding
5664		 * both pi_lock and rq->lock, such that holding either
5665		 * stabilizes the mask.
 
 
5666		 *
5667		 * Drop rq->lock is not quite as disastrous as it usually is
5668		 * because !cpu_active at this point, which means load-balance
5669		 * will not interfere. Also, stop-machine.
5670		 */
5671		rq_unlock(rq, rf);
5672		raw_spin_lock(&next->pi_lock);
5673		rq_relock(rq, rf);
5674
5675		/*
5676		 * Since we're inside stop-machine, _nothing_ should have
5677		 * changed the task, WARN if weird stuff happened, because in
5678		 * that case the above rq->lock drop is a fail too.
5679		 */
5680		if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5681			raw_spin_unlock(&next->pi_lock);
5682			continue;
5683		}
 
 
5684
5685		/* Find suitable destination for @next, with force if needed. */
5686		dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5687		rq = __migrate_task(rq, rf, next, dest_cpu);
5688		if (rq != dead_rq) {
5689			rq_unlock(rq, rf);
5690			rq = dead_rq;
5691			*rf = orf;
5692			rq_relock(rq, rf);
5693		}
5694		raw_spin_unlock(&next->pi_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5695	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5696
5697	rq->stop = stop;
 
5698}
 
5699#endif /* CONFIG_HOTPLUG_CPU */
5700
5701void set_rq_online(struct rq *rq)
5702{
5703	if (!rq->online) {
5704		const struct sched_class *class;
5705
5706		cpumask_set_cpu(rq->cpu, rq->rd->online);
5707		rq->online = 1;
5708
5709		for_each_class(class) {
5710			if (class->rq_online)
5711				class->rq_online(rq);
5712		}
5713	}
5714}
5715
5716void set_rq_offline(struct rq *rq)
5717{
5718	if (rq->online) {
5719		const struct sched_class *class;
5720
 
5721		for_each_class(class) {
5722			if (class->rq_offline)
5723				class->rq_offline(rq);
5724		}
5725
5726		cpumask_clear_cpu(rq->cpu, rq->rd->online);
5727		rq->online = 0;
5728	}
5729}
5730
5731static void set_cpu_rq_start_time(unsigned int cpu)
5732{
5733	struct rq *rq = cpu_rq(cpu);
5734
5735	rq->age_stamp = sched_clock_cpu(cpu);
5736}
5737
5738/*
5739 * used to mark begin/end of suspend/resume:
5740 */
5741static int num_cpus_frozen;
5742
5743/*
5744 * Update cpusets according to cpu_active mask.  If cpusets are
5745 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5746 * around partition_sched_domains().
5747 *
5748 * If we come here as part of a suspend/resume, don't touch cpusets because we
5749 * want to restore it back to its original state upon resume anyway.
5750 */
5751static void cpuset_cpu_active(void)
5752{
5753	if (cpuhp_tasks_frozen) {
5754		/*
5755		 * num_cpus_frozen tracks how many CPUs are involved in suspend
5756		 * resume sequence. As long as this is not the last online
5757		 * operation in the resume sequence, just build a single sched
5758		 * domain, ignoring cpusets.
5759		 */
5760		partition_sched_domains(1, NULL, NULL);
5761		if (--num_cpus_frozen)
5762			return;
5763		/*
5764		 * This is the last CPU online operation. So fall through and
5765		 * restore the original sched domains by considering the
5766		 * cpuset configurations.
5767		 */
5768		cpuset_force_rebuild();
5769	}
5770	cpuset_update_active_cpus();
5771}
5772
5773static int cpuset_cpu_inactive(unsigned int cpu)
5774{
5775	if (!cpuhp_tasks_frozen) {
5776		if (dl_cpu_busy(cpu))
5777			return -EBUSY;
 
 
5778		cpuset_update_active_cpus();
5779	} else {
5780		num_cpus_frozen++;
5781		partition_sched_domains(1, NULL, NULL);
5782	}
5783	return 0;
5784}
5785
5786int sched_cpu_activate(unsigned int cpu)
5787{
5788	struct rq *rq = cpu_rq(cpu);
5789	struct rq_flags rf;
5790
 
 
 
 
 
 
 
 
 
 
 
 
 
5791	set_cpu_active(cpu, true);
5792
5793	if (sched_smp_initialized) {
 
5794		sched_domains_numa_masks_set(cpu);
5795		cpuset_cpu_active();
5796	}
5797
5798	/*
5799	 * Put the rq online, if not already. This happens:
5800	 *
5801	 * 1) In the early boot process, because we build the real domains
5802	 *    after all CPUs have been brought up.
5803	 *
5804	 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5805	 *    domains.
5806	 */
5807	rq_lock_irqsave(rq, &rf);
5808	if (rq->rd) {
5809		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5810		set_rq_online(rq);
5811	}
5812	rq_unlock_irqrestore(rq, &rf);
5813
5814	update_max_interval();
5815
5816	return 0;
5817}
5818
5819int sched_cpu_deactivate(unsigned int cpu)
5820{
 
 
5821	int ret;
5822
 
 
 
 
 
 
5823	set_cpu_active(cpu, false);
 
5824	/*
5825	 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5826	 * users of this state to go away such that all new such users will
5827	 * observe it.
 
 
 
 
 
 
 
 
 
 
 
5828	 *
5829	 * Do sync before park smpboot threads to take care the rcu boost case.
5830	 */
5831	synchronize_rcu_mult(call_rcu, call_rcu_sched);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5832
5833	if (!sched_smp_initialized)
5834		return 0;
5835
 
5836	ret = cpuset_cpu_inactive(cpu);
5837	if (ret) {
 
5838		set_cpu_active(cpu, true);
 
5839		return ret;
5840	}
5841	sched_domains_numa_masks_clear(cpu);
5842	return 0;
5843}
5844
5845static void sched_rq_cpu_starting(unsigned int cpu)
5846{
5847	struct rq *rq = cpu_rq(cpu);
5848
5849	rq->calc_load_update = calc_load_update;
5850	update_max_interval();
5851}
5852
5853int sched_cpu_starting(unsigned int cpu)
5854{
5855	set_cpu_rq_start_time(cpu);
5856	sched_rq_cpu_starting(cpu);
5857	sched_tick_start(cpu);
5858	return 0;
5859}
5860
5861#ifdef CONFIG_HOTPLUG_CPU
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5862int sched_cpu_dying(unsigned int cpu)
5863{
5864	struct rq *rq = cpu_rq(cpu);
5865	struct rq_flags rf;
5866
5867	/* Handle pending wakeups and then migrate everything off */
5868	sched_ttwu_pending();
5869	sched_tick_stop(cpu);
5870
5871	rq_lock_irqsave(rq, &rf);
5872	if (rq->rd) {
5873		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5874		set_rq_offline(rq);
5875	}
5876	migrate_tasks(rq, &rf);
5877	BUG_ON(rq->nr_running != 1);
5878	rq_unlock_irqrestore(rq, &rf);
5879
5880	calc_load_migrate(rq);
5881	update_max_interval();
5882	nohz_balance_exit_idle(rq);
5883	hrtick_clear(rq);
 
5884	return 0;
5885}
5886#endif
5887
5888#ifdef CONFIG_SCHED_SMT
5889DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5890
5891static void sched_init_smt(void)
5892{
5893	/*
5894	 * We've enumerated all CPUs and will assume that if any CPU
5895	 * has SMT siblings, CPU0 will too.
5896	 */
5897	if (cpumask_weight(cpu_smt_mask(0)) > 1)
5898		static_branch_enable(&sched_smt_present);
5899}
5900#else
5901static inline void sched_init_smt(void) { }
5902#endif
5903
5904void __init sched_init_smp(void)
5905{
5906	sched_init_numa();
5907
5908	/*
5909	 * There's no userspace yet to cause hotplug operations; hence all the
5910	 * CPU masks are stable and all blatant races in the below code cannot
5911	 * happen.
5912	 */
5913	mutex_lock(&sched_domains_mutex);
5914	sched_init_domains(cpu_active_mask);
5915	mutex_unlock(&sched_domains_mutex);
5916
5917	/* Move init over to a non-isolated CPU */
5918	if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5919		BUG();
 
5920	sched_init_granularity();
5921
5922	init_sched_rt_class();
5923	init_sched_dl_class();
5924
5925	sched_init_smt();
5926
5927	sched_smp_initialized = true;
5928}
5929
5930static int __init migration_init(void)
5931{
5932	sched_rq_cpu_starting(smp_processor_id());
5933	return 0;
5934}
5935early_initcall(migration_init);
5936
5937#else
5938void __init sched_init_smp(void)
5939{
5940	sched_init_granularity();
5941}
5942#endif /* CONFIG_SMP */
5943
5944int in_sched_functions(unsigned long addr)
5945{
5946	return in_lock_functions(addr) ||
5947		(addr >= (unsigned long)__sched_text_start
5948		&& addr < (unsigned long)__sched_text_end);
5949}
5950
5951#ifdef CONFIG_CGROUP_SCHED
5952/*
5953 * Default task group.
5954 * Every task in system belongs to this group at bootup.
5955 */
5956struct task_group root_task_group;
5957LIST_HEAD(task_groups);
5958
5959/* Cacheline aligned slab cache for task_group */
5960static struct kmem_cache *task_group_cache __read_mostly;
5961#endif
5962
5963DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5964DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5965
5966void __init sched_init(void)
5967{
5968	int i, j;
5969	unsigned long alloc_size = 0, ptr;
 
 
 
 
 
 
 
 
5970
5971	sched_clock_init();
5972	wait_bit_init();
5973
5974#ifdef CONFIG_FAIR_GROUP_SCHED
5975	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5976#endif
5977#ifdef CONFIG_RT_GROUP_SCHED
5978	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5979#endif
5980	if (alloc_size) {
5981		ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5982
5983#ifdef CONFIG_FAIR_GROUP_SCHED
5984		root_task_group.se = (struct sched_entity **)ptr;
5985		ptr += nr_cpu_ids * sizeof(void **);
5986
5987		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5988		ptr += nr_cpu_ids * sizeof(void **);
5989
 
 
5990#endif /* CONFIG_FAIR_GROUP_SCHED */
5991#ifdef CONFIG_RT_GROUP_SCHED
5992		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5993		ptr += nr_cpu_ids * sizeof(void **);
5994
5995		root_task_group.rt_rq = (struct rt_rq **)ptr;
5996		ptr += nr_cpu_ids * sizeof(void **);
5997
5998#endif /* CONFIG_RT_GROUP_SCHED */
5999	}
6000#ifdef CONFIG_CPUMASK_OFFSTACK
6001	for_each_possible_cpu(i) {
6002		per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6003			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6004		per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6005			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6006	}
6007#endif /* CONFIG_CPUMASK_OFFSTACK */
6008
6009	init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6010	init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6011
6012#ifdef CONFIG_SMP
6013	init_defrootdomain();
6014#endif
6015
6016#ifdef CONFIG_RT_GROUP_SCHED
6017	init_rt_bandwidth(&root_task_group.rt_bandwidth,
6018			global_rt_period(), global_rt_runtime());
6019#endif /* CONFIG_RT_GROUP_SCHED */
6020
6021#ifdef CONFIG_CGROUP_SCHED
6022	task_group_cache = KMEM_CACHE(task_group, 0);
6023
6024	list_add(&root_task_group.list, &task_groups);
6025	INIT_LIST_HEAD(&root_task_group.children);
6026	INIT_LIST_HEAD(&root_task_group.siblings);
6027	autogroup_init(&init_task);
6028#endif /* CONFIG_CGROUP_SCHED */
6029
6030	for_each_possible_cpu(i) {
6031		struct rq *rq;
6032
6033		rq = cpu_rq(i);
6034		raw_spin_lock_init(&rq->lock);
6035		rq->nr_running = 0;
6036		rq->calc_load_active = 0;
6037		rq->calc_load_update = jiffies + LOAD_FREQ;
6038		init_cfs_rq(&rq->cfs);
6039		init_rt_rq(&rq->rt);
6040		init_dl_rq(&rq->dl);
6041#ifdef CONFIG_FAIR_GROUP_SCHED
6042		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6043		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6044		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6045		/*
6046		 * How much CPU bandwidth does root_task_group get?
6047		 *
6048		 * In case of task-groups formed thr' the cgroup filesystem, it
6049		 * gets 100% of the CPU resources in the system. This overall
6050		 * system CPU resource is divided among the tasks of
6051		 * root_task_group and its child task-groups in a fair manner,
6052		 * based on each entity's (task or task-group's) weight
6053		 * (se->load.weight).
6054		 *
6055		 * In other words, if root_task_group has 10 tasks of weight
6056		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6057		 * then A0's share of the CPU resource is:
6058		 *
6059		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6060		 *
6061		 * We achieve this by letting root_task_group's tasks sit
6062		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6063		 */
6064		init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6065		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6066#endif /* CONFIG_FAIR_GROUP_SCHED */
6067
6068		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6069#ifdef CONFIG_RT_GROUP_SCHED
6070		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6071#endif
6072
6073		for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6074			rq->cpu_load[j] = 0;
6075
6076#ifdef CONFIG_SMP
6077		rq->sd = NULL;
6078		rq->rd = NULL;
6079		rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6080		rq->balance_callback = NULL;
6081		rq->active_balance = 0;
6082		rq->next_balance = jiffies;
6083		rq->push_cpu = 0;
6084		rq->cpu = i;
6085		rq->online = 0;
6086		rq->idle_stamp = 0;
6087		rq->avg_idle = 2*sysctl_sched_migration_cost;
6088		rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6089
6090		INIT_LIST_HEAD(&rq->cfs_tasks);
6091
6092		rq_attach_root(rq, &def_root_domain);
6093#ifdef CONFIG_NO_HZ_COMMON
6094		rq->last_load_update_tick = jiffies;
6095		rq->last_blocked_load_update_tick = jiffies;
6096		atomic_set(&rq->nohz_flags, 0);
 
 
 
 
 
6097#endif
6098#endif /* CONFIG_SMP */
6099		hrtick_rq_init(rq);
6100		atomic_set(&rq->nr_iowait, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
6101	}
6102
6103	set_load_weight(&init_task, false);
6104
6105	/*
6106	 * The boot idle thread does lazy MMU switching as well:
6107	 */
6108	mmgrab(&init_mm);
6109	enter_lazy_tlb(&init_mm, current);
6110
6111	/*
 
 
 
 
 
 
 
 
6112	 * Make us the idle thread. Technically, schedule() should not be
6113	 * called from this thread, however somewhere below it might be,
6114	 * but because we are the idle thread, we just pick up running again
6115	 * when this runqueue becomes "idle".
6116	 */
6117	init_idle(current, smp_processor_id());
6118
6119	calc_load_update = jiffies + LOAD_FREQ;
6120
6121#ifdef CONFIG_SMP
6122	idle_thread_set_boot_cpu();
6123	set_cpu_rq_start_time(smp_processor_id());
6124#endif
6125	init_sched_fair_class();
6126
6127	init_schedstats();
 
 
 
 
6128
6129	scheduler_running = 1;
6130}
6131
6132#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6133static inline int preempt_count_equals(int preempt_offset)
6134{
6135	int nested = preempt_count() + rcu_preempt_depth();
6136
6137	return (nested == preempt_offset);
6138}
6139
6140void __might_sleep(const char *file, int line, int preempt_offset)
6141{
 
6142	/*
6143	 * Blocking primitives will set (and therefore destroy) current->state,
6144	 * since we will exit with TASK_RUNNING make sure we enter with it,
6145	 * otherwise we will destroy state.
6146	 */
6147	WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6148			"do not call blocking ops when !TASK_RUNNING; "
6149			"state=%lx set at [<%p>] %pS\n",
6150			current->state,
6151			(void *)current->task_state_change,
6152			(void *)current->task_state_change);
6153
6154	___might_sleep(file, line, preempt_offset);
6155}
6156EXPORT_SYMBOL(__might_sleep);
6157
6158void ___might_sleep(const char *file, int line, int preempt_offset)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6159{
6160	/* Ratelimiting timestamp: */
6161	static unsigned long prev_jiffy;
6162
6163	unsigned long preempt_disable_ip;
6164
6165	/* WARN_ON_ONCE() by default, no rate limit required: */
6166	rcu_sleep_check();
6167
6168	if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6169	     !is_idle_task(current)) ||
6170	    system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6171	    oops_in_progress)
6172		return;
6173
6174	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6175		return;
6176	prev_jiffy = jiffies;
6177
6178	/* Save this before calling printk(), since that will clobber it: */
6179	preempt_disable_ip = get_preempt_disable_ip(current);
6180
6181	printk(KERN_ERR
6182		"BUG: sleeping function called from invalid context at %s:%d\n",
6183			file, line);
6184	printk(KERN_ERR
6185		"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6186			in_atomic(), irqs_disabled(),
6187			current->pid, current->comm);
 
 
 
 
 
6188
6189	if (task_stack_end_corrupted(current))
6190		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6191
6192	debug_show_held_locks(current);
6193	if (irqs_disabled())
6194		print_irqtrace_events(current);
6195	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6196	    && !preempt_count_equals(preempt_offset)) {
6197		pr_err("Preemption disabled at:");
6198		print_ip_sym(preempt_disable_ip);
6199		pr_cont("\n");
6200	}
6201	dump_stack();
6202	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6203}
6204EXPORT_SYMBOL(___might_sleep);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6205#endif
6206
6207#ifdef CONFIG_MAGIC_SYSRQ
6208void normalize_rt_tasks(void)
6209{
6210	struct task_struct *g, *p;
6211	struct sched_attr attr = {
6212		.sched_policy = SCHED_NORMAL,
6213	};
6214
6215	read_lock(&tasklist_lock);
6216	for_each_process_thread(g, p) {
6217		/*
6218		 * Only normalize user tasks:
6219		 */
6220		if (p->flags & PF_KTHREAD)
6221			continue;
6222
6223		p->se.exec_start = 0;
6224		schedstat_set(p->se.statistics.wait_start,  0);
6225		schedstat_set(p->se.statistics.sleep_start, 0);
6226		schedstat_set(p->se.statistics.block_start, 0);
6227
6228		if (!dl_task(p) && !rt_task(p)) {
6229			/*
6230			 * Renice negative nice level userspace
6231			 * tasks back to 0:
6232			 */
6233			if (task_nice(p) < 0)
6234				set_user_nice(p, 0);
6235			continue;
6236		}
6237
6238		__sched_setscheduler(p, &attr, false, false);
6239	}
6240	read_unlock(&tasklist_lock);
6241}
6242
6243#endif /* CONFIG_MAGIC_SYSRQ */
6244
6245#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6246/*
6247 * These functions are only useful for the IA64 MCA handling, or kdb.
6248 *
6249 * They can only be called when the whole system has been
6250 * stopped - every CPU needs to be quiescent, and no scheduling
6251 * activity can take place. Using them for anything else would
6252 * be a serious bug, and as a result, they aren't even visible
6253 * under any other configuration.
6254 */
6255
6256/**
6257 * curr_task - return the current task for a given CPU.
6258 * @cpu: the processor in question.
6259 *
6260 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6261 *
6262 * Return: The current task for @cpu.
6263 */
6264struct task_struct *curr_task(int cpu)
6265{
6266	return cpu_curr(cpu);
6267}
6268
6269#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6270
6271#ifdef CONFIG_IA64
6272/**
6273 * set_curr_task - set the current task for a given CPU.
6274 * @cpu: the processor in question.
6275 * @p: the task pointer to set.
6276 *
6277 * Description: This function must only be used when non-maskable interrupts
6278 * are serviced on a separate stack. It allows the architecture to switch the
6279 * notion of the current task on a CPU in a non-blocking manner. This function
6280 * must be called with all CPU's synchronized, and interrupts disabled, the
6281 * and caller must save the original value of the current task (see
6282 * curr_task() above) and restore that value before reenabling interrupts and
6283 * re-starting the system.
6284 *
6285 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6286 */
6287void ia64_set_curr_task(int cpu, struct task_struct *p)
6288{
6289	cpu_curr(cpu) = p;
6290}
6291
6292#endif
6293
6294#ifdef CONFIG_CGROUP_SCHED
6295/* task_group_lock serializes the addition/removal of task groups */
6296static DEFINE_SPINLOCK(task_group_lock);
6297
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6298static void sched_free_group(struct task_group *tg)
6299{
6300	free_fair_sched_group(tg);
6301	free_rt_sched_group(tg);
6302	autogroup_free(tg);
6303	kmem_cache_free(task_group_cache, tg);
6304}
6305
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6306/* allocate runqueue etc for a new task group */
6307struct task_group *sched_create_group(struct task_group *parent)
6308{
6309	struct task_group *tg;
6310
6311	tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6312	if (!tg)
6313		return ERR_PTR(-ENOMEM);
6314
6315	if (!alloc_fair_sched_group(tg, parent))
6316		goto err;
6317
6318	if (!alloc_rt_sched_group(tg, parent))
6319		goto err;
6320
 
 
6321	return tg;
6322
6323err:
6324	sched_free_group(tg);
6325	return ERR_PTR(-ENOMEM);
6326}
6327
6328void sched_online_group(struct task_group *tg, struct task_group *parent)
6329{
6330	unsigned long flags;
6331
6332	spin_lock_irqsave(&task_group_lock, flags);
6333	list_add_rcu(&tg->list, &task_groups);
6334
6335	/* Root should already exist: */
6336	WARN_ON(!parent);
6337
6338	tg->parent = parent;
6339	INIT_LIST_HEAD(&tg->children);
6340	list_add_rcu(&tg->siblings, &parent->children);
6341	spin_unlock_irqrestore(&task_group_lock, flags);
6342
6343	online_fair_sched_group(tg);
6344}
6345
6346/* rcu callback to free various structures associated with a task group */
6347static void sched_free_group_rcu(struct rcu_head *rhp)
6348{
6349	/* Now it should be safe to free those cfs_rqs: */
6350	sched_free_group(container_of(rhp, struct task_group, rcu));
6351}
6352
6353void sched_destroy_group(struct task_group *tg)
6354{
6355	/* Wait for possible concurrent references to cfs_rqs complete: */
6356	call_rcu(&tg->rcu, sched_free_group_rcu);
6357}
6358
6359void sched_offline_group(struct task_group *tg)
6360{
6361	unsigned long flags;
6362
6363	/* End participation in shares distribution: */
6364	unregister_fair_sched_group(tg);
6365
 
 
 
 
 
 
 
 
 
 
6366	spin_lock_irqsave(&task_group_lock, flags);
6367	list_del_rcu(&tg->list);
6368	list_del_rcu(&tg->siblings);
6369	spin_unlock_irqrestore(&task_group_lock, flags);
6370}
6371
6372static void sched_change_group(struct task_struct *tsk, int type)
6373{
6374	struct task_group *tg;
6375
6376	/*
6377	 * All callers are synchronized by task_rq_lock(); we do not use RCU
6378	 * which is pointless here. Thus, we pass "true" to task_css_check()
6379	 * to prevent lockdep warnings.
6380	 */
6381	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6382			  struct task_group, css);
6383	tg = autogroup_task_group(tsk, tg);
6384	tsk->sched_task_group = tg;
 
 
 
 
 
 
6385
6386#ifdef CONFIG_FAIR_GROUP_SCHED
6387	if (tsk->sched_class->task_change_group)
6388		tsk->sched_class->task_change_group(tsk, type);
6389	else
6390#endif
6391		set_task_rq(tsk, task_cpu(tsk));
6392}
6393
6394/*
6395 * Change task's runqueue when it moves between groups.
6396 *
6397 * The caller of this function should have put the task in its new group by
6398 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6399 * its new group.
6400 */
6401void sched_move_task(struct task_struct *tsk)
6402{
6403	int queued, running, queue_flags =
6404		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6405	struct rq_flags rf;
6406	struct rq *rq;
6407
6408	rq = task_rq_lock(tsk, &rf);
 
 
 
 
 
 
 
 
 
 
6409	update_rq_clock(rq);
6410
6411	running = task_current(rq, tsk);
6412	queued = task_on_rq_queued(tsk);
6413
6414	if (queued)
6415		dequeue_task(rq, tsk, queue_flags);
6416	if (running)
6417		put_prev_task(rq, tsk);
6418
6419	sched_change_group(tsk, TASK_MOVE_GROUP);
6420
6421	if (queued)
6422		enqueue_task(rq, tsk, queue_flags);
6423	if (running)
6424		set_curr_task(rq, tsk);
6425
6426	task_rq_unlock(rq, tsk, &rf);
 
 
 
 
 
6427}
6428
6429static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6430{
6431	return css ? container_of(css, struct task_group, css) : NULL;
6432}
6433
6434static struct cgroup_subsys_state *
6435cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6436{
6437	struct task_group *parent = css_tg(parent_css);
6438	struct task_group *tg;
6439
6440	if (!parent) {
6441		/* This is early initialization for the top cgroup */
6442		return &root_task_group.css;
6443	}
6444
6445	tg = sched_create_group(parent);
6446	if (IS_ERR(tg))
6447		return ERR_PTR(-ENOMEM);
6448
6449	return &tg->css;
6450}
6451
6452/* Expose task group only after completing cgroup initialization */
6453static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6454{
6455	struct task_group *tg = css_tg(css);
6456	struct task_group *parent = css_tg(css->parent);
6457
6458	if (parent)
6459		sched_online_group(tg, parent);
 
 
 
 
 
 
 
 
6460	return 0;
6461}
6462
6463static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6464{
6465	struct task_group *tg = css_tg(css);
6466
6467	sched_offline_group(tg);
6468}
6469
6470static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6471{
6472	struct task_group *tg = css_tg(css);
6473
6474	/*
6475	 * Relies on the RCU grace period between css_released() and this.
6476	 */
6477	sched_free_group(tg);
6478}
6479
6480/*
6481 * This is called before wake_up_new_task(), therefore we really only
6482 * have to set its group bits, all the other stuff does not apply.
6483 */
6484static void cpu_cgroup_fork(struct task_struct *task)
6485{
6486	struct rq_flags rf;
6487	struct rq *rq;
6488
6489	rq = task_rq_lock(task, &rf);
6490
6491	update_rq_clock(rq);
6492	sched_change_group(task, TASK_SET_GROUP);
6493
6494	task_rq_unlock(rq, task, &rf);
6495}
6496
 
6497static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6498{
6499	struct task_struct *task;
6500	struct cgroup_subsys_state *css;
6501	int ret = 0;
6502
6503	cgroup_taskset_for_each(task, css, tset) {
6504#ifdef CONFIG_RT_GROUP_SCHED
6505		if (!sched_rt_can_attach(css_tg(css), task))
6506			return -EINVAL;
6507#else
6508		/* We don't support RT-tasks being in separate groups */
6509		if (task->sched_class != &fair_sched_class)
6510			return -EINVAL;
6511#endif
6512		/*
6513		 * Serialize against wake_up_new_task() such that if its
6514		 * running, we're sure to observe its full state.
6515		 */
6516		raw_spin_lock_irq(&task->pi_lock);
6517		/*
6518		 * Avoid calling sched_move_task() before wake_up_new_task()
6519		 * has happened. This would lead to problems with PELT, due to
6520		 * move wanting to detach+attach while we're not attached yet.
6521		 */
6522		if (task->state == TASK_NEW)
6523			ret = -EINVAL;
6524		raw_spin_unlock_irq(&task->pi_lock);
6525
6526		if (ret)
6527			break;
6528	}
6529	return ret;
6530}
 
6531
6532static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6533{
6534	struct task_struct *task;
6535	struct cgroup_subsys_state *css;
6536
6537	cgroup_taskset_for_each(task, css, tset)
6538		sched_move_task(task);
6539}
6540
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6541#ifdef CONFIG_FAIR_GROUP_SCHED
6542static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6543				struct cftype *cftype, u64 shareval)
6544{
 
 
6545	return sched_group_set_shares(css_tg(css), scale_load(shareval));
6546}
6547
6548static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6549			       struct cftype *cft)
6550{
6551	struct task_group *tg = css_tg(css);
6552
6553	return (u64) scale_load_down(tg->shares);
6554}
6555
6556#ifdef CONFIG_CFS_BANDWIDTH
6557static DEFINE_MUTEX(cfs_constraints_mutex);
6558
6559const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6560const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
 
 
6561
6562static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6563
6564static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
 
6565{
6566	int i, ret = 0, runtime_enabled, runtime_was_enabled;
6567	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6568
6569	if (tg == &root_task_group)
6570		return -EINVAL;
6571
6572	/*
6573	 * Ensure we have at some amount of bandwidth every period.  This is
6574	 * to prevent reaching a state of large arrears when throttled via
6575	 * entity_tick() resulting in prolonged exit starvation.
6576	 */
6577	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6578		return -EINVAL;
6579
6580	/*
6581	 * Likewise, bound things on the otherside by preventing insane quota
6582	 * periods.  This also allows us to normalize in computing quota
6583	 * feasibility.
6584	 */
6585	if (period > max_cfs_quota_period)
6586		return -EINVAL;
6587
6588	/*
 
 
 
 
 
 
 
 
 
 
6589	 * Prevent race between setting of cfs_rq->runtime_enabled and
6590	 * unthrottle_offline_cfs_rqs().
6591	 */
6592	get_online_cpus();
6593	mutex_lock(&cfs_constraints_mutex);
 
6594	ret = __cfs_schedulable(tg, period, quota);
6595	if (ret)
6596		goto out_unlock;
6597
6598	runtime_enabled = quota != RUNTIME_INF;
6599	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6600	/*
6601	 * If we need to toggle cfs_bandwidth_used, off->on must occur
6602	 * before making related changes, and on->off must occur afterwards
6603	 */
6604	if (runtime_enabled && !runtime_was_enabled)
6605		cfs_bandwidth_usage_inc();
6606	raw_spin_lock_irq(&cfs_b->lock);
6607	cfs_b->period = ns_to_ktime(period);
6608	cfs_b->quota = quota;
6609
6610	__refill_cfs_bandwidth_runtime(cfs_b);
 
 
 
6611
6612	/* Restart the period timer (if active) to handle new period expiry: */
6613	if (runtime_enabled)
6614		start_cfs_bandwidth(cfs_b);
6615
6616	raw_spin_unlock_irq(&cfs_b->lock);
 
 
 
 
 
 
6617
6618	for_each_online_cpu(i) {
6619		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6620		struct rq *rq = cfs_rq->rq;
6621		struct rq_flags rf;
6622
6623		rq_lock_irq(rq, &rf);
6624		cfs_rq->runtime_enabled = runtime_enabled;
6625		cfs_rq->runtime_remaining = 0;
6626
6627		if (cfs_rq->throttled)
6628			unthrottle_cfs_rq(cfs_rq);
6629		rq_unlock_irq(rq, &rf);
6630	}
 
6631	if (runtime_was_enabled && !runtime_enabled)
6632		cfs_bandwidth_usage_dec();
6633out_unlock:
6634	mutex_unlock(&cfs_constraints_mutex);
6635	put_online_cpus();
6636
6637	return ret;
6638}
6639
6640int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6641{
6642	u64 quota, period;
6643
6644	period = ktime_to_ns(tg->cfs_bandwidth.period);
 
6645	if (cfs_quota_us < 0)
6646		quota = RUNTIME_INF;
6647	else
6648		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
 
 
6649
6650	return tg_set_cfs_bandwidth(tg, period, quota);
6651}
6652
6653long tg_get_cfs_quota(struct task_group *tg)
6654{
6655	u64 quota_us;
6656
6657	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6658		return -1;
6659
6660	quota_us = tg->cfs_bandwidth.quota;
6661	do_div(quota_us, NSEC_PER_USEC);
6662
6663	return quota_us;
6664}
6665
6666int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6667{
6668	u64 quota, period;
 
 
 
6669
6670	period = (u64)cfs_period_us * NSEC_PER_USEC;
6671	quota = tg->cfs_bandwidth.quota;
 
6672
6673	return tg_set_cfs_bandwidth(tg, period, quota);
6674}
6675
6676long tg_get_cfs_period(struct task_group *tg)
6677{
6678	u64 cfs_period_us;
6679
6680	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6681	do_div(cfs_period_us, NSEC_PER_USEC);
6682
6683	return cfs_period_us;
6684}
6685
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6686static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6687				  struct cftype *cft)
6688{
6689	return tg_get_cfs_quota(css_tg(css));
6690}
6691
6692static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6693				   struct cftype *cftype, s64 cfs_quota_us)
6694{
6695	return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6696}
6697
6698static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6699				   struct cftype *cft)
6700{
6701	return tg_get_cfs_period(css_tg(css));
6702}
6703
6704static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6705				    struct cftype *cftype, u64 cfs_period_us)
6706{
6707	return tg_set_cfs_period(css_tg(css), cfs_period_us);
6708}
6709
 
 
 
 
 
 
 
 
 
 
 
 
6710struct cfs_schedulable_data {
6711	struct task_group *tg;
6712	u64 period, quota;
6713};
6714
6715/*
6716 * normalize group quota/period to be quota/max_period
6717 * note: units are usecs
6718 */
6719static u64 normalize_cfs_quota(struct task_group *tg,
6720			       struct cfs_schedulable_data *d)
6721{
6722	u64 quota, period;
6723
6724	if (tg == d->tg) {
6725		period = d->period;
6726		quota = d->quota;
6727	} else {
6728		period = tg_get_cfs_period(tg);
6729		quota = tg_get_cfs_quota(tg);
6730	}
6731
6732	/* note: these should typically be equivalent */
6733	if (quota == RUNTIME_INF || quota == -1)
6734		return RUNTIME_INF;
6735
6736	return to_ratio(period, quota);
6737}
6738
6739static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6740{
6741	struct cfs_schedulable_data *d = data;
6742	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6743	s64 quota = 0, parent_quota = -1;
6744
6745	if (!tg->parent) {
6746		quota = RUNTIME_INF;
6747	} else {
6748		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6749
6750		quota = normalize_cfs_quota(tg, d);
6751		parent_quota = parent_b->hierarchical_quota;
6752
6753		/*
6754		 * Ensure max(child_quota) <= parent_quota.  On cgroup2,
6755		 * always take the min.  On cgroup1, only inherit when no
6756		 * limit is set:
 
 
6757		 */
6758		if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6759			quota = min(quota, parent_quota);
 
 
 
6760		} else {
6761			if (quota == RUNTIME_INF)
6762				quota = parent_quota;
6763			else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6764				return -EINVAL;
6765		}
6766	}
6767	cfs_b->hierarchical_quota = quota;
6768
6769	return 0;
6770}
6771
6772static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6773{
6774	int ret;
6775	struct cfs_schedulable_data data = {
6776		.tg = tg,
6777		.period = period,
6778		.quota = quota,
6779	};
6780
6781	if (quota != RUNTIME_INF) {
6782		do_div(data.period, NSEC_PER_USEC);
6783		do_div(data.quota, NSEC_PER_USEC);
6784	}
6785
6786	rcu_read_lock();
6787	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6788	rcu_read_unlock();
6789
6790	return ret;
6791}
6792
6793static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6794{
6795	struct task_group *tg = css_tg(seq_css(sf));
6796	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6797
6798	seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6799	seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6800	seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6801
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6802	return 0;
6803}
6804#endif /* CONFIG_CFS_BANDWIDTH */
6805#endif /* CONFIG_FAIR_GROUP_SCHED */
6806
6807#ifdef CONFIG_RT_GROUP_SCHED
6808static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6809				struct cftype *cft, s64 val)
6810{
6811	return sched_group_set_rt_runtime(css_tg(css), val);
6812}
6813
6814static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6815			       struct cftype *cft)
6816{
6817	return sched_group_rt_runtime(css_tg(css));
6818}
6819
6820static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6821				    struct cftype *cftype, u64 rt_period_us)
6822{
6823	return sched_group_set_rt_period(css_tg(css), rt_period_us);
6824}
6825
6826static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6827				   struct cftype *cft)
6828{
6829	return sched_group_rt_period(css_tg(css));
6830}
6831#endif /* CONFIG_RT_GROUP_SCHED */
6832
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6833static struct cftype cpu_legacy_files[] = {
6834#ifdef CONFIG_FAIR_GROUP_SCHED
6835	{
6836		.name = "shares",
6837		.read_u64 = cpu_shares_read_u64,
6838		.write_u64 = cpu_shares_write_u64,
6839	},
 
 
 
 
 
6840#endif
6841#ifdef CONFIG_CFS_BANDWIDTH
6842	{
6843		.name = "cfs_quota_us",
6844		.read_s64 = cpu_cfs_quota_read_s64,
6845		.write_s64 = cpu_cfs_quota_write_s64,
6846	},
6847	{
6848		.name = "cfs_period_us",
6849		.read_u64 = cpu_cfs_period_read_u64,
6850		.write_u64 = cpu_cfs_period_write_u64,
6851	},
6852	{
 
 
 
 
 
6853		.name = "stat",
6854		.seq_show = cpu_cfs_stat_show,
6855	},
 
 
 
 
6856#endif
6857#ifdef CONFIG_RT_GROUP_SCHED
6858	{
6859		.name = "rt_runtime_us",
6860		.read_s64 = cpu_rt_runtime_read,
6861		.write_s64 = cpu_rt_runtime_write,
6862	},
6863	{
6864		.name = "rt_period_us",
6865		.read_u64 = cpu_rt_period_read_uint,
6866		.write_u64 = cpu_rt_period_write_uint,
6867	},
6868#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6869	{ }	/* Terminate */
6870};
6871
6872static int cpu_extra_stat_show(struct seq_file *sf,
6873			       struct cgroup_subsys_state *css)
6874{
6875#ifdef CONFIG_CFS_BANDWIDTH
6876	{
6877		struct task_group *tg = css_tg(css);
6878		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6879		u64 throttled_usec;
6880
6881		throttled_usec = cfs_b->throttled_time;
6882		do_div(throttled_usec, NSEC_PER_USEC);
 
 
6883
6884		seq_printf(sf, "nr_periods %d\n"
6885			   "nr_throttled %d\n"
6886			   "throttled_usec %llu\n",
 
 
6887			   cfs_b->nr_periods, cfs_b->nr_throttled,
6888			   throttled_usec);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6889	}
6890#endif
6891	return 0;
6892}
6893
6894#ifdef CONFIG_FAIR_GROUP_SCHED
6895static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6896			       struct cftype *cft)
6897{
6898	struct task_group *tg = css_tg(css);
6899	u64 weight = scale_load_down(tg->shares);
6900
6901	return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6902}
6903
6904static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6905				struct cftype *cft, u64 weight)
6906{
6907	/*
6908	 * cgroup weight knobs should use the common MIN, DFL and MAX
6909	 * values which are 1, 100 and 10000 respectively.  While it loses
6910	 * a bit of range on both ends, it maps pretty well onto the shares
6911	 * value used by scheduler and the round-trip conversions preserve
6912	 * the original value over the entire range.
6913	 */
6914	if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6915		return -ERANGE;
6916
6917	weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6918
6919	return sched_group_set_shares(css_tg(css), scale_load(weight));
6920}
6921
6922static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6923				    struct cftype *cft)
6924{
6925	unsigned long weight = scale_load_down(css_tg(css)->shares);
6926	int last_delta = INT_MAX;
6927	int prio, delta;
6928
6929	/* find the closest nice value to the current weight */
6930	for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6931		delta = abs(sched_prio_to_weight[prio] - weight);
6932		if (delta >= last_delta)
6933			break;
6934		last_delta = delta;
6935	}
6936
6937	return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6938}
6939
6940static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6941				     struct cftype *cft, s64 nice)
6942{
6943	unsigned long weight;
6944	int idx;
6945
6946	if (nice < MIN_NICE || nice > MAX_NICE)
6947		return -ERANGE;
6948
6949	idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6950	idx = array_index_nospec(idx, 40);
6951	weight = sched_prio_to_weight[idx];
6952
6953	return sched_group_set_shares(css_tg(css), scale_load(weight));
6954}
6955#endif
6956
6957static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6958						  long period, long quota)
6959{
6960	if (quota < 0)
6961		seq_puts(sf, "max");
6962	else
6963		seq_printf(sf, "%ld", quota);
6964
6965	seq_printf(sf, " %ld\n", period);
6966}
6967
6968/* caller should put the current value in *@periodp before calling */
6969static int __maybe_unused cpu_period_quota_parse(char *buf,
6970						 u64 *periodp, u64 *quotap)
6971{
6972	char tok[21];	/* U64_MAX */
6973
6974	if (!sscanf(buf, "%s %llu", tok, periodp))
6975		return -EINVAL;
6976
6977	*periodp *= NSEC_PER_USEC;
6978
6979	if (sscanf(tok, "%llu", quotap))
6980		*quotap *= NSEC_PER_USEC;
6981	else if (!strcmp(tok, "max"))
6982		*quotap = RUNTIME_INF;
6983	else
6984		return -EINVAL;
6985
6986	return 0;
6987}
6988
6989#ifdef CONFIG_CFS_BANDWIDTH
6990static int cpu_max_show(struct seq_file *sf, void *v)
6991{
6992	struct task_group *tg = css_tg(seq_css(sf));
6993
6994	cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6995	return 0;
6996}
6997
6998static ssize_t cpu_max_write(struct kernfs_open_file *of,
6999			     char *buf, size_t nbytes, loff_t off)
7000{
7001	struct task_group *tg = css_tg(of_css(of));
7002	u64 period = tg_get_cfs_period(tg);
 
7003	u64 quota;
7004	int ret;
7005
7006	ret = cpu_period_quota_parse(buf, &period, &quota);
7007	if (!ret)
7008		ret = tg_set_cfs_bandwidth(tg, period, quota);
7009	return ret ?: nbytes;
7010}
7011#endif
7012
7013static struct cftype cpu_files[] = {
7014#ifdef CONFIG_FAIR_GROUP_SCHED
7015	{
7016		.name = "weight",
7017		.flags = CFTYPE_NOT_ON_ROOT,
7018		.read_u64 = cpu_weight_read_u64,
7019		.write_u64 = cpu_weight_write_u64,
7020	},
7021	{
7022		.name = "weight.nice",
7023		.flags = CFTYPE_NOT_ON_ROOT,
7024		.read_s64 = cpu_weight_nice_read_s64,
7025		.write_s64 = cpu_weight_nice_write_s64,
7026	},
 
 
 
 
 
 
7027#endif
7028#ifdef CONFIG_CFS_BANDWIDTH
7029	{
7030		.name = "max",
7031		.flags = CFTYPE_NOT_ON_ROOT,
7032		.seq_show = cpu_max_show,
7033		.write = cpu_max_write,
7034	},
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7035#endif
7036	{ }	/* terminate */
7037};
7038
7039struct cgroup_subsys cpu_cgrp_subsys = {
7040	.css_alloc	= cpu_cgroup_css_alloc,
7041	.css_online	= cpu_cgroup_css_online,
7042	.css_released	= cpu_cgroup_css_released,
7043	.css_free	= cpu_cgroup_css_free,
7044	.css_extra_stat_show = cpu_extra_stat_show,
7045	.fork		= cpu_cgroup_fork,
 
7046	.can_attach	= cpu_cgroup_can_attach,
 
7047	.attach		= cpu_cgroup_attach,
7048	.legacy_cftypes	= cpu_legacy_files,
7049	.dfl_cftypes	= cpu_files,
7050	.early_init	= true,
7051	.threaded	= true,
7052};
7053
7054#endif	/* CONFIG_CGROUP_SCHED */
7055
7056void dump_cpu_task(int cpu)
7057{
 
 
 
 
 
 
 
 
 
 
 
 
 
7058	pr_info("Task dump for CPU %d:\n", cpu);
7059	sched_show_task(cpu_curr(cpu));
7060}
7061
7062/*
7063 * Nice levels are multiplicative, with a gentle 10% change for every
7064 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7065 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7066 * that remained on nice 0.
7067 *
7068 * The "10% effect" is relative and cumulative: from _any_ nice level,
7069 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7070 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7071 * If a task goes up by ~10% and another task goes down by ~10% then
7072 * the relative distance between them is ~25%.)
7073 */
7074const int sched_prio_to_weight[40] = {
7075 /* -20 */     88761,     71755,     56483,     46273,     36291,
7076 /* -15 */     29154,     23254,     18705,     14949,     11916,
7077 /* -10 */      9548,      7620,      6100,      4904,      3906,
7078 /*  -5 */      3121,      2501,      1991,      1586,      1277,
7079 /*   0 */      1024,       820,       655,       526,       423,
7080 /*   5 */       335,       272,       215,       172,       137,
7081 /*  10 */       110,        87,        70,        56,        45,
7082 /*  15 */        36,        29,        23,        18,        15,
7083};
7084
7085/*
7086 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7087 *
7088 * In cases where the weight does not change often, we can use the
7089 * precalculated inverse to speed up arithmetics by turning divisions
7090 * into multiplications:
7091 */
7092const u32 sched_prio_to_wmult[40] = {
7093 /* -20 */     48388,     59856,     76040,     92818,    118348,
7094 /* -15 */    147320,    184698,    229616,    287308,    360437,
7095 /* -10 */    449829,    563644,    704093,    875809,   1099582,
7096 /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
7097 /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
7098 /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
7099 /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
7100 /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7101};
7102
7103#undef CREATE_TRACE_POINTS
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
v6.8
    1// SPDX-License-Identifier: GPL-2.0-only
    2/*
    3 *  kernel/sched/core.c
    4 *
    5 *  Core kernel scheduler code and related syscalls
    6 *
    7 *  Copyright (C) 1991-2002  Linus Torvalds
    8 */
    9#include <linux/highmem.h>
   10#include <linux/hrtimer_api.h>
   11#include <linux/ktime_api.h>
   12#include <linux/sched/signal.h>
   13#include <linux/syscalls_api.h>
   14#include <linux/debug_locks.h>
   15#include <linux/prefetch.h>
   16#include <linux/capability.h>
   17#include <linux/pgtable_api.h>
   18#include <linux/wait_bit.h>
   19#include <linux/jiffies.h>
   20#include <linux/spinlock_api.h>
   21#include <linux/cpumask_api.h>
   22#include <linux/lockdep_api.h>
   23#include <linux/hardirq.h>
   24#include <linux/softirq.h>
   25#include <linux/refcount_api.h>
   26#include <linux/topology.h>
   27#include <linux/sched/clock.h>
   28#include <linux/sched/cond_resched.h>
   29#include <linux/sched/cputime.h>
   30#include <linux/sched/debug.h>
   31#include <linux/sched/hotplug.h>
   32#include <linux/sched/init.h>
   33#include <linux/sched/isolation.h>
   34#include <linux/sched/loadavg.h>
   35#include <linux/sched/mm.h>
   36#include <linux/sched/nohz.h>
   37#include <linux/sched/rseq_api.h>
   38#include <linux/sched/rt.h>
   39
   40#include <linux/blkdev.h>
   41#include <linux/context_tracking.h>
   42#include <linux/cpuset.h>
   43#include <linux/delayacct.h>
   44#include <linux/init_task.h>
   45#include <linux/interrupt.h>
   46#include <linux/ioprio.h>
   47#include <linux/kallsyms.h>
   48#include <linux/kcov.h>
   49#include <linux/kprobes.h>
   50#include <linux/llist_api.h>
   51#include <linux/mmu_context.h>
   52#include <linux/mmzone.h>
   53#include <linux/mutex_api.h>
   54#include <linux/nmi.h>
   55#include <linux/nospec.h>
   56#include <linux/perf_event_api.h>
   57#include <linux/profile.h>
   58#include <linux/psi.h>
   59#include <linux/rcuwait_api.h>
   60#include <linux/rseq.h>
   61#include <linux/sched/wake_q.h>
   62#include <linux/scs.h>
   63#include <linux/slab.h>
   64#include <linux/syscalls.h>
   65#include <linux/vtime.h>
   66#include <linux/wait_api.h>
   67#include <linux/workqueue_api.h>
   68
   69#ifdef CONFIG_PREEMPT_DYNAMIC
   70# ifdef CONFIG_GENERIC_ENTRY
   71#  include <linux/entry-common.h>
   72# endif
   73#endif
   74
   75#include <uapi/linux/sched/types.h>
   76
   77#include <asm/irq_regs.h>
   78#include <asm/switch_to.h>
   79#include <asm/tlb.h>
   80
   81#define CREATE_TRACE_POINTS
   82#include <linux/sched/rseq_api.h>
   83#include <trace/events/sched.h>
   84#include <trace/events/ipi.h>
   85#undef CREATE_TRACE_POINTS
   86
   87#include "sched.h"
   88#include "stats.h"
   89
   90#include "autogroup.h"
   91#include "pelt.h"
   92#include "smp.h"
   93#include "stats.h"
   94
   95#include "../workqueue_internal.h"
   96#include "../../io_uring/io-wq.h"
   97#include "../smpboot.h"
   98
   99EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpu);
  100EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpumask);
  101
  102/*
  103 * Export tracepoints that act as a bare tracehook (ie: have no trace event
  104 * associated with them) to allow external modules to probe them.
  105 */
  106EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
  107EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
  108EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
  109EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
  110EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
  111EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
  112EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
  113EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
  114EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
  115EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
  116EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
  117EXPORT_TRACEPOINT_SYMBOL_GPL(sched_compute_energy_tp);
  118
  119DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
  120
  121#ifdef CONFIG_SCHED_DEBUG
  122/*
  123 * Debugging: various feature bits
  124 *
  125 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
  126 * sysctl_sched_features, defined in sched.h, to allow constants propagation
  127 * at compile time and compiler optimization based on features default.
  128 */
  129#define SCHED_FEAT(name, enabled)	\
  130	(1UL << __SCHED_FEAT_##name) * enabled |
  131const_debug unsigned int sysctl_sched_features =
  132#include "features.h"
  133	0;
  134#undef SCHED_FEAT
  135
  136/*
  137 * Print a warning if need_resched is set for the given duration (if
  138 * LATENCY_WARN is enabled).
  139 *
  140 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
  141 * per boot.
  142 */
  143__read_mostly int sysctl_resched_latency_warn_ms = 100;
  144__read_mostly int sysctl_resched_latency_warn_once = 1;
  145#endif /* CONFIG_SCHED_DEBUG */
  146
  147/*
  148 * Number of tasks to iterate in a single balance run.
  149 * Limited because this is done with IRQs disabled.
  150 */
  151const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
  152
  153__read_mostly int scheduler_running;
  154
  155#ifdef CONFIG_SCHED_CORE
  156
  157DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
  158
  159/* kernel prio, less is more */
  160static inline int __task_prio(const struct task_struct *p)
  161{
  162	if (p->sched_class == &stop_sched_class) /* trumps deadline */
  163		return -2;
  164
  165	if (rt_prio(p->prio)) /* includes deadline */
  166		return p->prio; /* [-1, 99] */
  167
  168	if (p->sched_class == &idle_sched_class)
  169		return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
  170
  171	return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
  172}
  173
  174/*
  175 * l(a,b)
  176 * le(a,b) := !l(b,a)
  177 * g(a,b)  := l(b,a)
  178 * ge(a,b) := !l(a,b)
  179 */
  180
  181/* real prio, less is less */
  182static inline bool prio_less(const struct task_struct *a,
  183			     const struct task_struct *b, bool in_fi)
  184{
  185
  186	int pa = __task_prio(a), pb = __task_prio(b);
  187
  188	if (-pa < -pb)
  189		return true;
  190
  191	if (-pb < -pa)
  192		return false;
  193
  194	if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
  195		return !dl_time_before(a->dl.deadline, b->dl.deadline);
  196
  197	if (pa == MAX_RT_PRIO + MAX_NICE)	/* fair */
  198		return cfs_prio_less(a, b, in_fi);
  199
  200	return false;
  201}
  202
  203static inline bool __sched_core_less(const struct task_struct *a,
  204				     const struct task_struct *b)
  205{
  206	if (a->core_cookie < b->core_cookie)
  207		return true;
  208
  209	if (a->core_cookie > b->core_cookie)
  210		return false;
  211
  212	/* flip prio, so high prio is leftmost */
  213	if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
  214		return true;
  215
  216	return false;
  217}
  218
  219#define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
  220
  221static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
  222{
  223	return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
  224}
  225
  226static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
  227{
  228	const struct task_struct *p = __node_2_sc(node);
  229	unsigned long cookie = (unsigned long)key;
  230
  231	if (cookie < p->core_cookie)
  232		return -1;
  233
  234	if (cookie > p->core_cookie)
  235		return 1;
  236
  237	return 0;
  238}
  239
  240void sched_core_enqueue(struct rq *rq, struct task_struct *p)
  241{
  242	rq->core->core_task_seq++;
  243
  244	if (!p->core_cookie)
  245		return;
  246
  247	rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
  248}
  249
  250void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
  251{
  252	rq->core->core_task_seq++;
  253
  254	if (sched_core_enqueued(p)) {
  255		rb_erase(&p->core_node, &rq->core_tree);
  256		RB_CLEAR_NODE(&p->core_node);
  257	}
  258
  259	/*
  260	 * Migrating the last task off the cpu, with the cpu in forced idle
  261	 * state. Reschedule to create an accounting edge for forced idle,
  262	 * and re-examine whether the core is still in forced idle state.
  263	 */
  264	if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
  265	    rq->core->core_forceidle_count && rq->curr == rq->idle)
  266		resched_curr(rq);
  267}
  268
  269static int sched_task_is_throttled(struct task_struct *p, int cpu)
  270{
  271	if (p->sched_class->task_is_throttled)
  272		return p->sched_class->task_is_throttled(p, cpu);
  273
  274	return 0;
  275}
  276
  277static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
  278{
  279	struct rb_node *node = &p->core_node;
  280	int cpu = task_cpu(p);
  281
  282	do {
  283		node = rb_next(node);
  284		if (!node)
  285			return NULL;
  286
  287		p = __node_2_sc(node);
  288		if (p->core_cookie != cookie)
  289			return NULL;
  290
  291	} while (sched_task_is_throttled(p, cpu));
  292
  293	return p;
  294}
  295
  296/*
  297 * Find left-most (aka, highest priority) and unthrottled task matching @cookie.
  298 * If no suitable task is found, NULL will be returned.
  299 */
  300static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
  301{
  302	struct task_struct *p;
  303	struct rb_node *node;
  304
  305	node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
  306	if (!node)
  307		return NULL;
  308
  309	p = __node_2_sc(node);
  310	if (!sched_task_is_throttled(p, rq->cpu))
  311		return p;
  312
  313	return sched_core_next(p, cookie);
  314}
  315
  316/*
  317 * Magic required such that:
  318 *
  319 *	raw_spin_rq_lock(rq);
  320 *	...
  321 *	raw_spin_rq_unlock(rq);
  322 *
  323 * ends up locking and unlocking the _same_ lock, and all CPUs
  324 * always agree on what rq has what lock.
  325 *
  326 * XXX entirely possible to selectively enable cores, don't bother for now.
  327 */
  328
  329static DEFINE_MUTEX(sched_core_mutex);
  330static atomic_t sched_core_count;
  331static struct cpumask sched_core_mask;
  332
  333static void sched_core_lock(int cpu, unsigned long *flags)
  334{
  335	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
  336	int t, i = 0;
  337
  338	local_irq_save(*flags);
  339	for_each_cpu(t, smt_mask)
  340		raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
  341}
  342
  343static void sched_core_unlock(int cpu, unsigned long *flags)
  344{
  345	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
  346	int t;
  347
  348	for_each_cpu(t, smt_mask)
  349		raw_spin_unlock(&cpu_rq(t)->__lock);
  350	local_irq_restore(*flags);
  351}
  352
  353static void __sched_core_flip(bool enabled)
  354{
  355	unsigned long flags;
  356	int cpu, t;
  357
  358	cpus_read_lock();
  359
  360	/*
  361	 * Toggle the online cores, one by one.
  362	 */
  363	cpumask_copy(&sched_core_mask, cpu_online_mask);
  364	for_each_cpu(cpu, &sched_core_mask) {
  365		const struct cpumask *smt_mask = cpu_smt_mask(cpu);
  366
  367		sched_core_lock(cpu, &flags);
  368
  369		for_each_cpu(t, smt_mask)
  370			cpu_rq(t)->core_enabled = enabled;
  371
  372		cpu_rq(cpu)->core->core_forceidle_start = 0;
  373
  374		sched_core_unlock(cpu, &flags);
  375
  376		cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
  377	}
  378
  379	/*
  380	 * Toggle the offline CPUs.
  381	 */
  382	for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
  383		cpu_rq(cpu)->core_enabled = enabled;
  384
  385	cpus_read_unlock();
  386}
  387
  388static void sched_core_assert_empty(void)
  389{
  390	int cpu;
  391
  392	for_each_possible_cpu(cpu)
  393		WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
  394}
  395
  396static void __sched_core_enable(void)
  397{
  398	static_branch_enable(&__sched_core_enabled);
  399	/*
  400	 * Ensure all previous instances of raw_spin_rq_*lock() have finished
  401	 * and future ones will observe !sched_core_disabled().
  402	 */
  403	synchronize_rcu();
  404	__sched_core_flip(true);
  405	sched_core_assert_empty();
  406}
  407
  408static void __sched_core_disable(void)
  409{
  410	sched_core_assert_empty();
  411	__sched_core_flip(false);
  412	static_branch_disable(&__sched_core_enabled);
  413}
  414
  415void sched_core_get(void)
  416{
  417	if (atomic_inc_not_zero(&sched_core_count))
  418		return;
  419
  420	mutex_lock(&sched_core_mutex);
  421	if (!atomic_read(&sched_core_count))
  422		__sched_core_enable();
  423
  424	smp_mb__before_atomic();
  425	atomic_inc(&sched_core_count);
  426	mutex_unlock(&sched_core_mutex);
  427}
  428
  429static void __sched_core_put(struct work_struct *work)
  430{
  431	if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
  432		__sched_core_disable();
  433		mutex_unlock(&sched_core_mutex);
  434	}
  435}
  436
  437void sched_core_put(void)
  438{
  439	static DECLARE_WORK(_work, __sched_core_put);
  440
  441	/*
  442	 * "There can be only one"
  443	 *
  444	 * Either this is the last one, or we don't actually need to do any
  445	 * 'work'. If it is the last *again*, we rely on
  446	 * WORK_STRUCT_PENDING_BIT.
  447	 */
  448	if (!atomic_add_unless(&sched_core_count, -1, 1))
  449		schedule_work(&_work);
  450}
  451
  452#else /* !CONFIG_SCHED_CORE */
  453
  454static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
  455static inline void
  456sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
  457
  458#endif /* CONFIG_SCHED_CORE */
  459
  460/*
  461 * Serialization rules:
  462 *
  463 * Lock order:
  464 *
  465 *   p->pi_lock
  466 *     rq->lock
  467 *       hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
  468 *
  469 *  rq1->lock
  470 *    rq2->lock  where: rq1 < rq2
  471 *
  472 * Regular state:
  473 *
  474 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
  475 * local CPU's rq->lock, it optionally removes the task from the runqueue and
  476 * always looks at the local rq data structures to find the most eligible task
  477 * to run next.
  478 *
  479 * Task enqueue is also under rq->lock, possibly taken from another CPU.
  480 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
  481 * the local CPU to avoid bouncing the runqueue state around [ see
  482 * ttwu_queue_wakelist() ]
  483 *
  484 * Task wakeup, specifically wakeups that involve migration, are horribly
  485 * complicated to avoid having to take two rq->locks.
  486 *
  487 * Special state:
  488 *
  489 * System-calls and anything external will use task_rq_lock() which acquires
  490 * both p->pi_lock and rq->lock. As a consequence the state they change is
  491 * stable while holding either lock:
  492 *
  493 *  - sched_setaffinity()/
  494 *    set_cpus_allowed_ptr():	p->cpus_ptr, p->nr_cpus_allowed
  495 *  - set_user_nice():		p->se.load, p->*prio
  496 *  - __sched_setscheduler():	p->sched_class, p->policy, p->*prio,
  497 *				p->se.load, p->rt_priority,
  498 *				p->dl.dl_{runtime, deadline, period, flags, bw, density}
  499 *  - sched_setnuma():		p->numa_preferred_nid
  500 *  - sched_move_task():	p->sched_task_group
  501 *  - uclamp_update_active()	p->uclamp*
  502 *
  503 * p->state <- TASK_*:
  504 *
  505 *   is changed locklessly using set_current_state(), __set_current_state() or
  506 *   set_special_state(), see their respective comments, or by
  507 *   try_to_wake_up(). This latter uses p->pi_lock to serialize against
  508 *   concurrent self.
  509 *
  510 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
  511 *
  512 *   is set by activate_task() and cleared by deactivate_task(), under
  513 *   rq->lock. Non-zero indicates the task is runnable, the special
  514 *   ON_RQ_MIGRATING state is used for migration without holding both
  515 *   rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
  516 *
  517 * p->on_cpu <- { 0, 1 }:
  518 *
  519 *   is set by prepare_task() and cleared by finish_task() such that it will be
  520 *   set before p is scheduled-in and cleared after p is scheduled-out, both
  521 *   under rq->lock. Non-zero indicates the task is running on its CPU.
  522 *
  523 *   [ The astute reader will observe that it is possible for two tasks on one
  524 *     CPU to have ->on_cpu = 1 at the same time. ]
  525 *
  526 * task_cpu(p): is changed by set_task_cpu(), the rules are:
  527 *
  528 *  - Don't call set_task_cpu() on a blocked task:
  529 *
  530 *    We don't care what CPU we're not running on, this simplifies hotplug,
  531 *    the CPU assignment of blocked tasks isn't required to be valid.
  532 *
  533 *  - for try_to_wake_up(), called under p->pi_lock:
  534 *
  535 *    This allows try_to_wake_up() to only take one rq->lock, see its comment.
  536 *
  537 *  - for migration called under rq->lock:
  538 *    [ see task_on_rq_migrating() in task_rq_lock() ]
  539 *
  540 *    o move_queued_task()
  541 *    o detach_task()
  542 *
  543 *  - for migration called under double_rq_lock():
  544 *
  545 *    o __migrate_swap_task()
  546 *    o push_rt_task() / pull_rt_task()
  547 *    o push_dl_task() / pull_dl_task()
  548 *    o dl_task_offline_migration()
  549 *
  550 */
 
  551
  552void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
  553{
  554	raw_spinlock_t *lock;
  555
  556	/* Matches synchronize_rcu() in __sched_core_enable() */
  557	preempt_disable();
  558	if (sched_core_disabled()) {
  559		raw_spin_lock_nested(&rq->__lock, subclass);
  560		/* preempt_count *MUST* be > 1 */
  561		preempt_enable_no_resched();
  562		return;
  563	}
  564
  565	for (;;) {
  566		lock = __rq_lockp(rq);
  567		raw_spin_lock_nested(lock, subclass);
  568		if (likely(lock == __rq_lockp(rq))) {
  569			/* preempt_count *MUST* be > 1 */
  570			preempt_enable_no_resched();
  571			return;
  572		}
  573		raw_spin_unlock(lock);
  574	}
  575}
  576
  577bool raw_spin_rq_trylock(struct rq *rq)
  578{
  579	raw_spinlock_t *lock;
  580	bool ret;
  581
  582	/* Matches synchronize_rcu() in __sched_core_enable() */
  583	preempt_disable();
  584	if (sched_core_disabled()) {
  585		ret = raw_spin_trylock(&rq->__lock);
  586		preempt_enable();
  587		return ret;
  588	}
  589
  590	for (;;) {
  591		lock = __rq_lockp(rq);
  592		ret = raw_spin_trylock(lock);
  593		if (!ret || (likely(lock == __rq_lockp(rq)))) {
  594			preempt_enable();
  595			return ret;
  596		}
  597		raw_spin_unlock(lock);
  598	}
  599}
  600
  601void raw_spin_rq_unlock(struct rq *rq)
  602{
  603	raw_spin_unlock(rq_lockp(rq));
  604}
  605
  606#ifdef CONFIG_SMP
  607/*
  608 * double_rq_lock - safely lock two runqueues
 
  609 */
  610void double_rq_lock(struct rq *rq1, struct rq *rq2)
  611{
  612	lockdep_assert_irqs_disabled();
  613
  614	if (rq_order_less(rq2, rq1))
  615		swap(rq1, rq2);
  616
  617	raw_spin_rq_lock(rq1);
  618	if (__rq_lockp(rq1) != __rq_lockp(rq2))
  619		raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
  620
  621	double_rq_clock_clear_update(rq1, rq2);
  622}
  623#endif
  624
  625/*
  626 * __task_rq_lock - lock the rq @p resides on.
  627 */
  628struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
  629	__acquires(rq->lock)
  630{
  631	struct rq *rq;
  632
  633	lockdep_assert_held(&p->pi_lock);
  634
  635	for (;;) {
  636		rq = task_rq(p);
  637		raw_spin_rq_lock(rq);
  638		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
  639			rq_pin_lock(rq, rf);
  640			return rq;
  641		}
  642		raw_spin_rq_unlock(rq);
  643
  644		while (unlikely(task_on_rq_migrating(p)))
  645			cpu_relax();
  646	}
  647}
  648
  649/*
  650 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
  651 */
  652struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
  653	__acquires(p->pi_lock)
  654	__acquires(rq->lock)
  655{
  656	struct rq *rq;
  657
  658	for (;;) {
  659		raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
  660		rq = task_rq(p);
  661		raw_spin_rq_lock(rq);
  662		/*
  663		 *	move_queued_task()		task_rq_lock()
  664		 *
  665		 *	ACQUIRE (rq->lock)
  666		 *	[S] ->on_rq = MIGRATING		[L] rq = task_rq()
  667		 *	WMB (__set_task_cpu())		ACQUIRE (rq->lock);
  668		 *	[S] ->cpu = new_cpu		[L] task_rq()
  669		 *					[L] ->on_rq
  670		 *	RELEASE (rq->lock)
  671		 *
  672		 * If we observe the old CPU in task_rq_lock(), the acquire of
  673		 * the old rq->lock will fully serialize against the stores.
  674		 *
  675		 * If we observe the new CPU in task_rq_lock(), the address
  676		 * dependency headed by '[L] rq = task_rq()' and the acquire
  677		 * will pair with the WMB to ensure we then also see migrating.
  678		 */
  679		if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
  680			rq_pin_lock(rq, rf);
  681			return rq;
  682		}
  683		raw_spin_rq_unlock(rq);
  684		raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
  685
  686		while (unlikely(task_on_rq_migrating(p)))
  687			cpu_relax();
  688	}
  689}
  690
  691/*
  692 * RQ-clock updating methods:
  693 */
  694
  695static void update_rq_clock_task(struct rq *rq, s64 delta)
  696{
  697/*
  698 * In theory, the compile should just see 0 here, and optimize out the call
  699 * to sched_rt_avg_update. But I don't trust it...
  700 */
  701	s64 __maybe_unused steal = 0, irq_delta = 0;
  702
 
  703#ifdef CONFIG_IRQ_TIME_ACCOUNTING
  704	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
  705
  706	/*
  707	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
  708	 * this case when a previous update_rq_clock() happened inside a
  709	 * {soft,}irq region.
  710	 *
  711	 * When this happens, we stop ->clock_task and only update the
  712	 * prev_irq_time stamp to account for the part that fit, so that a next
  713	 * update will consume the rest. This ensures ->clock_task is
  714	 * monotonic.
  715	 *
  716	 * It does however cause some slight miss-attribution of {soft,}irq
  717	 * time, a more accurate solution would be to update the irq_time using
  718	 * the current rq->clock timestamp, except that would require using
  719	 * atomic ops.
  720	 */
  721	if (irq_delta > delta)
  722		irq_delta = delta;
  723
  724	rq->prev_irq_time += irq_delta;
  725	delta -= irq_delta;
  726	psi_account_irqtime(rq->curr, irq_delta);
  727	delayacct_irq(rq->curr, irq_delta);
  728#endif
  729#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
  730	if (static_key_false((&paravirt_steal_rq_enabled))) {
  731		steal = paravirt_steal_clock(cpu_of(rq));
  732		steal -= rq->prev_steal_time_rq;
  733
  734		if (unlikely(steal > delta))
  735			steal = delta;
  736
  737		rq->prev_steal_time_rq += steal;
  738		delta -= steal;
  739	}
  740#endif
  741
  742	rq->clock_task += delta;
  743
  744#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
  745	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
  746		update_irq_load_avg(rq, irq_delta + steal);
  747#endif
  748	update_rq_clock_pelt(rq, delta);
  749}
  750
  751void update_rq_clock(struct rq *rq)
  752{
  753	s64 delta;
  754
  755	lockdep_assert_rq_held(rq);
  756
  757	if (rq->clock_update_flags & RQCF_ACT_SKIP)
  758		return;
  759
  760#ifdef CONFIG_SCHED_DEBUG
  761	if (sched_feat(WARN_DOUBLE_CLOCK))
  762		SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
  763	rq->clock_update_flags |= RQCF_UPDATED;
  764#endif
  765
  766	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
  767	if (delta < 0)
  768		return;
  769	rq->clock += delta;
  770	update_rq_clock_task(rq, delta);
  771}
  772
 
  773#ifdef CONFIG_SCHED_HRTICK
  774/*
  775 * Use HR-timers to deliver accurate preemption points.
  776 */
  777
  778static void hrtick_clear(struct rq *rq)
  779{
  780	if (hrtimer_active(&rq->hrtick_timer))
  781		hrtimer_cancel(&rq->hrtick_timer);
  782}
  783
  784/*
  785 * High-resolution timer tick.
  786 * Runs from hardirq context with interrupts disabled.
  787 */
  788static enum hrtimer_restart hrtick(struct hrtimer *timer)
  789{
  790	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
  791	struct rq_flags rf;
  792
  793	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
  794
  795	rq_lock(rq, &rf);
  796	update_rq_clock(rq);
  797	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
  798	rq_unlock(rq, &rf);
  799
  800	return HRTIMER_NORESTART;
  801}
  802
  803#ifdef CONFIG_SMP
  804
  805static void __hrtick_restart(struct rq *rq)
  806{
  807	struct hrtimer *timer = &rq->hrtick_timer;
  808	ktime_t time = rq->hrtick_time;
  809
  810	hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
  811}
  812
  813/*
  814 * called from hardirq (IPI) context
  815 */
  816static void __hrtick_start(void *arg)
  817{
  818	struct rq *rq = arg;
  819	struct rq_flags rf;
  820
  821	rq_lock(rq, &rf);
  822	__hrtick_restart(rq);
 
  823	rq_unlock(rq, &rf);
  824}
  825
  826/*
  827 * Called to set the hrtick timer state.
  828 *
  829 * called with rq->lock held and irqs disabled
  830 */
  831void hrtick_start(struct rq *rq, u64 delay)
  832{
  833	struct hrtimer *timer = &rq->hrtick_timer;
 
  834	s64 delta;
  835
  836	/*
  837	 * Don't schedule slices shorter than 10000ns, that just
  838	 * doesn't make sense and can cause timer DoS.
  839	 */
  840	delta = max_t(s64, delay, 10000LL);
  841	rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
  842
  843	if (rq == this_rq())
 
 
  844		__hrtick_restart(rq);
  845	else
  846		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
 
 
  847}
  848
  849#else
  850/*
  851 * Called to set the hrtick timer state.
  852 *
  853 * called with rq->lock held and irqs disabled
  854 */
  855void hrtick_start(struct rq *rq, u64 delay)
  856{
  857	/*
  858	 * Don't schedule slices shorter than 10000ns, that just
  859	 * doesn't make sense. Rely on vruntime for fairness.
  860	 */
  861	delay = max_t(u64, delay, 10000LL);
  862	hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
  863		      HRTIMER_MODE_REL_PINNED_HARD);
  864}
  865
  866#endif /* CONFIG_SMP */
  867
  868static void hrtick_rq_init(struct rq *rq)
  869{
  870#ifdef CONFIG_SMP
  871	INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
 
 
 
 
  872#endif
  873	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
 
  874	rq->hrtick_timer.function = hrtick;
  875}
  876#else	/* CONFIG_SCHED_HRTICK */
  877static inline void hrtick_clear(struct rq *rq)
  878{
  879}
  880
  881static inline void hrtick_rq_init(struct rq *rq)
  882{
  883}
  884#endif	/* CONFIG_SCHED_HRTICK */
  885
  886/*
  887 * cmpxchg based fetch_or, macro so it works for different integer types
  888 */
  889#define fetch_or(ptr, mask)						\
  890	({								\
  891		typeof(ptr) _ptr = (ptr);				\
  892		typeof(mask) _mask = (mask);				\
  893		typeof(*_ptr) _val = *_ptr;				\
  894									\
  895		do {							\
  896		} while (!try_cmpxchg(_ptr, &_val, _val | _mask));	\
  897	_val;								\
 
 
 
 
  898})
  899
  900#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
  901/*
  902 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
  903 * this avoids any races wrt polling state changes and thereby avoids
  904 * spurious IPIs.
  905 */
  906static inline bool set_nr_and_not_polling(struct task_struct *p)
  907{
  908	struct thread_info *ti = task_thread_info(p);
  909	return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
  910}
  911
  912/*
  913 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
  914 *
  915 * If this returns true, then the idle task promises to call
  916 * sched_ttwu_pending() and reschedule soon.
  917 */
  918static bool set_nr_if_polling(struct task_struct *p)
  919{
  920	struct thread_info *ti = task_thread_info(p);
  921	typeof(ti->flags) val = READ_ONCE(ti->flags);
  922
  923	do {
  924		if (!(val & _TIF_POLLING_NRFLAG))
  925			return false;
  926		if (val & _TIF_NEED_RESCHED)
  927			return true;
  928	} while (!try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED));
  929
 
 
 
  930	return true;
  931}
  932
  933#else
  934static inline bool set_nr_and_not_polling(struct task_struct *p)
  935{
  936	set_tsk_need_resched(p);
  937	return true;
  938}
  939
  940#ifdef CONFIG_SMP
  941static inline bool set_nr_if_polling(struct task_struct *p)
  942{
  943	return false;
  944}
  945#endif
  946#endif
  947
  948static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
  949{
  950	struct wake_q_node *node = &task->wake_q;
  951
  952	/*
  953	 * Atomically grab the task, if ->wake_q is !nil already it means
  954	 * it's already queued (either by us or someone else) and will get the
  955	 * wakeup due to that.
  956	 *
  957	 * In order to ensure that a pending wakeup will observe our pending
  958	 * state, even in the failed case, an explicit smp_mb() must be used.
  959	 */
  960	smp_mb__before_atomic();
  961	if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
  962		return false;
 
  963
  964	/*
  965	 * The head is context local, there can be no concurrency.
  966	 */
  967	*head->lastp = node;
  968	head->lastp = &node->next;
  969	return true;
  970}
  971
  972/**
  973 * wake_q_add() - queue a wakeup for 'later' waking.
  974 * @head: the wake_q_head to add @task to
  975 * @task: the task to queue for 'later' wakeup
  976 *
  977 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
  978 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
  979 * instantly.
  980 *
  981 * This function must be used as-if it were wake_up_process(); IOW the task
  982 * must be ready to be woken at this location.
  983 */
  984void wake_q_add(struct wake_q_head *head, struct task_struct *task)
  985{
  986	if (__wake_q_add(head, task))
  987		get_task_struct(task);
  988}
  989
  990/**
  991 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
  992 * @head: the wake_q_head to add @task to
  993 * @task: the task to queue for 'later' wakeup
  994 *
  995 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
  996 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
  997 * instantly.
  998 *
  999 * This function must be used as-if it were wake_up_process(); IOW the task
 1000 * must be ready to be woken at this location.
 1001 *
 1002 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
 1003 * that already hold reference to @task can call the 'safe' version and trust
 1004 * wake_q to do the right thing depending whether or not the @task is already
 1005 * queued for wakeup.
 1006 */
 1007void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
 1008{
 1009	if (!__wake_q_add(head, task))
 1010		put_task_struct(task);
 1011}
 1012
 1013void wake_up_q(struct wake_q_head *head)
 1014{
 1015	struct wake_q_node *node = head->first;
 1016
 1017	while (node != WAKE_Q_TAIL) {
 1018		struct task_struct *task;
 1019
 1020		task = container_of(node, struct task_struct, wake_q);
 
 1021		/* Task can safely be re-inserted now: */
 1022		node = node->next;
 1023		task->wake_q.next = NULL;
 1024
 1025		/*
 1026		 * wake_up_process() executes a full barrier, which pairs with
 1027		 * the queueing in wake_q_add() so as not to miss wakeups.
 1028		 */
 1029		wake_up_process(task);
 1030		put_task_struct(task);
 1031	}
 1032}
 1033
 1034/*
 1035 * resched_curr - mark rq's current task 'to be rescheduled now'.
 1036 *
 1037 * On UP this means the setting of the need_resched flag, on SMP it
 1038 * might also involve a cross-CPU call to trigger the scheduler on
 1039 * the target CPU.
 1040 */
 1041void resched_curr(struct rq *rq)
 1042{
 1043	struct task_struct *curr = rq->curr;
 1044	int cpu;
 1045
 1046	lockdep_assert_rq_held(rq);
 1047
 1048	if (test_tsk_need_resched(curr))
 1049		return;
 1050
 1051	cpu = cpu_of(rq);
 1052
 1053	if (cpu == smp_processor_id()) {
 1054		set_tsk_need_resched(curr);
 1055		set_preempt_need_resched();
 1056		return;
 1057	}
 1058
 1059	if (set_nr_and_not_polling(curr))
 1060		smp_send_reschedule(cpu);
 1061	else
 1062		trace_sched_wake_idle_without_ipi(cpu);
 1063}
 1064
 1065void resched_cpu(int cpu)
 1066{
 1067	struct rq *rq = cpu_rq(cpu);
 1068	unsigned long flags;
 1069
 1070	raw_spin_rq_lock_irqsave(rq, flags);
 1071	if (cpu_online(cpu) || cpu == smp_processor_id())
 1072		resched_curr(rq);
 1073	raw_spin_rq_unlock_irqrestore(rq, flags);
 1074}
 1075
 1076#ifdef CONFIG_SMP
 1077#ifdef CONFIG_NO_HZ_COMMON
 1078/*
 1079 * In the semi idle case, use the nearest busy CPU for migrating timers
 1080 * from an idle CPU.  This is good for power-savings.
 1081 *
 1082 * We don't do similar optimization for completely idle system, as
 1083 * selecting an idle CPU will add more delays to the timers than intended
 1084 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
 1085 */
 1086int get_nohz_timer_target(void)
 1087{
 1088	int i, cpu = smp_processor_id(), default_cpu = -1;
 1089	struct sched_domain *sd;
 1090	const struct cpumask *hk_mask;
 1091
 1092	if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
 1093		if (!idle_cpu(cpu))
 1094			return cpu;
 1095		default_cpu = cpu;
 1096	}
 1097
 1098	hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
 1099
 1100	guard(rcu)();
 1101
 
 1102	for_each_domain(cpu, sd) {
 1103		for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
 1104			if (cpu == i)
 1105				continue;
 1106
 1107			if (!idle_cpu(i))
 1108				return i;
 
 
 1109		}
 1110	}
 1111
 1112	if (default_cpu == -1)
 1113		default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
 1114
 1115	return default_cpu;
 
 1116}
 1117
 1118/*
 1119 * When add_timer_on() enqueues a timer into the timer wheel of an
 1120 * idle CPU then this timer might expire before the next timer event
 1121 * which is scheduled to wake up that CPU. In case of a completely
 1122 * idle system the next event might even be infinite time into the
 1123 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
 1124 * leaves the inner idle loop so the newly added timer is taken into
 1125 * account when the CPU goes back to idle and evaluates the timer
 1126 * wheel for the next timer event.
 1127 */
 1128static void wake_up_idle_cpu(int cpu)
 1129{
 1130	struct rq *rq = cpu_rq(cpu);
 1131
 1132	if (cpu == smp_processor_id())
 1133		return;
 1134
 1135	/*
 1136	 * Set TIF_NEED_RESCHED and send an IPI if in the non-polling
 1137	 * part of the idle loop. This forces an exit from the idle loop
 1138	 * and a round trip to schedule(). Now this could be optimized
 1139	 * because a simple new idle loop iteration is enough to
 1140	 * re-evaluate the next tick. Provided some re-ordering of tick
 1141	 * nohz functions that would need to follow TIF_NR_POLLING
 1142	 * clearing:
 1143	 *
 1144	 * - On most archs, a simple fetch_or on ti::flags with a
 1145	 *   "0" value would be enough to know if an IPI needs to be sent.
 1146	 *
 1147	 * - x86 needs to perform a last need_resched() check between
 1148	 *   monitor and mwait which doesn't take timers into account.
 1149	 *   There a dedicated TIF_TIMER flag would be required to
 1150	 *   fetch_or here and be checked along with TIF_NEED_RESCHED
 1151	 *   before mwait().
 1152	 *
 1153	 * However, remote timer enqueue is not such a frequent event
 1154	 * and testing of the above solutions didn't appear to report
 1155	 * much benefits.
 1156	 */
 1157	if (set_nr_and_not_polling(rq->idle))
 1158		smp_send_reschedule(cpu);
 1159	else
 1160		trace_sched_wake_idle_without_ipi(cpu);
 1161}
 1162
 1163static bool wake_up_full_nohz_cpu(int cpu)
 1164{
 1165	/*
 1166	 * We just need the target to call irq_exit() and re-evaluate
 1167	 * the next tick. The nohz full kick at least implies that.
 1168	 * If needed we can still optimize that later with an
 1169	 * empty IRQ.
 1170	 */
 1171	if (cpu_is_offline(cpu))
 1172		return true;  /* Don't try to wake offline CPUs. */
 1173	if (tick_nohz_full_cpu(cpu)) {
 1174		if (cpu != smp_processor_id() ||
 1175		    tick_nohz_tick_stopped())
 1176			tick_nohz_full_kick_cpu(cpu);
 1177		return true;
 1178	}
 1179
 1180	return false;
 1181}
 1182
 1183/*
 1184 * Wake up the specified CPU.  If the CPU is going offline, it is the
 1185 * caller's responsibility to deal with the lost wakeup, for example,
 1186 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
 1187 */
 1188void wake_up_nohz_cpu(int cpu)
 1189{
 1190	if (!wake_up_full_nohz_cpu(cpu))
 1191		wake_up_idle_cpu(cpu);
 1192}
 1193
 1194static void nohz_csd_func(void *info)
 1195{
 1196	struct rq *rq = info;
 1197	int cpu = cpu_of(rq);
 1198	unsigned int flags;
 
 
 
 
 1199
 1200	/*
 1201	 * Release the rq::nohz_csd.
 
 1202	 */
 1203	flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
 1204	WARN_ON(!(flags & NOHZ_KICK_MASK));
 
 1205
 1206	rq->idle_balance = idle_cpu(cpu);
 1207	if (rq->idle_balance && !need_resched()) {
 1208		rq->nohz_idle_balance = flags;
 1209		raise_softirq_irqoff(SCHED_SOFTIRQ);
 1210	}
 1211}
 1212
 1213#endif /* CONFIG_NO_HZ_COMMON */
 1214
 1215#ifdef CONFIG_NO_HZ_FULL
 1216static inline bool __need_bw_check(struct rq *rq, struct task_struct *p)
 1217{
 1218	if (rq->nr_running != 1)
 1219		return false;
 1220
 1221	if (p->sched_class != &fair_sched_class)
 1222		return false;
 1223
 1224	if (!task_on_rq_queued(p))
 1225		return false;
 1226
 1227	return true;
 1228}
 1229
 1230bool sched_can_stop_tick(struct rq *rq)
 1231{
 1232	int fifo_nr_running;
 1233
 1234	/* Deadline tasks, even if single, need the tick */
 1235	if (rq->dl.dl_nr_running)
 1236		return false;
 1237
 1238	/*
 1239	 * If there are more than one RR tasks, we need the tick to affect the
 1240	 * actual RR behaviour.
 1241	 */
 1242	if (rq->rt.rr_nr_running) {
 1243		if (rq->rt.rr_nr_running == 1)
 1244			return true;
 1245		else
 1246			return false;
 1247	}
 1248
 1249	/*
 1250	 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
 1251	 * forced preemption between FIFO tasks.
 1252	 */
 1253	fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
 1254	if (fifo_nr_running)
 1255		return true;
 1256
 1257	/*
 1258	 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
 1259	 * if there's more than one we need the tick for involuntary
 1260	 * preemption.
 1261	 */
 1262	if (rq->nr_running > 1)
 1263		return false;
 1264
 1265	/*
 1266	 * If there is one task and it has CFS runtime bandwidth constraints
 1267	 * and it's on the cpu now we don't want to stop the tick.
 1268	 * This check prevents clearing the bit if a newly enqueued task here is
 1269	 * dequeued by migrating while the constrained task continues to run.
 1270	 * E.g. going from 2->1 without going through pick_next_task().
 1271	 */
 1272	if (sched_feat(HZ_BW) && __need_bw_check(rq, rq->curr)) {
 1273		if (cfs_task_bw_constrained(rq->curr))
 1274			return false;
 1275	}
 1276
 1277	return true;
 1278}
 1279#endif /* CONFIG_NO_HZ_FULL */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 1280#endif /* CONFIG_SMP */
 1281
 1282#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
 1283			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
 1284/*
 1285 * Iterate task_group tree rooted at *from, calling @down when first entering a
 1286 * node and @up when leaving it for the final time.
 1287 *
 1288 * Caller must hold rcu_lock or sufficient equivalent.
 1289 */
 1290int walk_tg_tree_from(struct task_group *from,
 1291			     tg_visitor down, tg_visitor up, void *data)
 1292{
 1293	struct task_group *parent, *child;
 1294	int ret;
 1295
 1296	parent = from;
 1297
 1298down:
 1299	ret = (*down)(parent, data);
 1300	if (ret)
 1301		goto out;
 1302	list_for_each_entry_rcu(child, &parent->children, siblings) {
 1303		parent = child;
 1304		goto down;
 1305
 1306up:
 1307		continue;
 1308	}
 1309	ret = (*up)(parent, data);
 1310	if (ret || parent == from)
 1311		goto out;
 1312
 1313	child = parent;
 1314	parent = parent->parent;
 1315	if (parent)
 1316		goto up;
 1317out:
 1318	return ret;
 1319}
 1320
 1321int tg_nop(struct task_group *tg, void *data)
 1322{
 1323	return 0;
 1324}
 1325#endif
 1326
 1327static void set_load_weight(struct task_struct *p, bool update_load)
 1328{
 1329	int prio = p->static_prio - MAX_RT_PRIO;
 1330	struct load_weight *load = &p->se.load;
 1331
 1332	/*
 1333	 * SCHED_IDLE tasks get minimal weight:
 1334	 */
 1335	if (task_has_idle_policy(p)) {
 1336		load->weight = scale_load(WEIGHT_IDLEPRIO);
 1337		load->inv_weight = WMULT_IDLEPRIO;
 1338		return;
 1339	}
 1340
 1341	/*
 1342	 * SCHED_OTHER tasks have to update their load when changing their
 1343	 * weight
 1344	 */
 1345	if (update_load && p->sched_class == &fair_sched_class) {
 1346		reweight_task(p, prio);
 1347	} else {
 1348		load->weight = scale_load(sched_prio_to_weight[prio]);
 1349		load->inv_weight = sched_prio_to_wmult[prio];
 1350	}
 1351}
 1352
 1353#ifdef CONFIG_UCLAMP_TASK
 1354/*
 1355 * Serializes updates of utilization clamp values
 1356 *
 1357 * The (slow-path) user-space triggers utilization clamp value updates which
 1358 * can require updates on (fast-path) scheduler's data structures used to
 1359 * support enqueue/dequeue operations.
 1360 * While the per-CPU rq lock protects fast-path update operations, user-space
 1361 * requests are serialized using a mutex to reduce the risk of conflicting
 1362 * updates or API abuses.
 1363 */
 1364static DEFINE_MUTEX(uclamp_mutex);
 1365
 1366/* Max allowed minimum utilization */
 1367static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
 1368
 1369/* Max allowed maximum utilization */
 1370static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
 1371
 1372/*
 1373 * By default RT tasks run at the maximum performance point/capacity of the
 1374 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
 1375 * SCHED_CAPACITY_SCALE.
 1376 *
 1377 * This knob allows admins to change the default behavior when uclamp is being
 1378 * used. In battery powered devices, particularly, running at the maximum
 1379 * capacity and frequency will increase energy consumption and shorten the
 1380 * battery life.
 1381 *
 1382 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
 1383 *
 1384 * This knob will not override the system default sched_util_clamp_min defined
 1385 * above.
 1386 */
 1387static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
 1388
 1389/* All clamps are required to be less or equal than these values */
 1390static struct uclamp_se uclamp_default[UCLAMP_CNT];
 1391
 1392/*
 1393 * This static key is used to reduce the uclamp overhead in the fast path. It
 1394 * primarily disables the call to uclamp_rq_{inc, dec}() in
 1395 * enqueue/dequeue_task().
 1396 *
 1397 * This allows users to continue to enable uclamp in their kernel config with
 1398 * minimum uclamp overhead in the fast path.
 1399 *
 1400 * As soon as userspace modifies any of the uclamp knobs, the static key is
 1401 * enabled, since we have an actual users that make use of uclamp
 1402 * functionality.
 1403 *
 1404 * The knobs that would enable this static key are:
 1405 *
 1406 *   * A task modifying its uclamp value with sched_setattr().
 1407 *   * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
 1408 *   * An admin modifying the cgroup cpu.uclamp.{min, max}
 1409 */
 1410DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
 1411
 1412/* Integer rounded range for each bucket */
 1413#define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
 1414
 1415#define for_each_clamp_id(clamp_id) \
 1416	for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
 1417
 1418static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
 1419{
 1420	return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
 1421}
 1422
 1423static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
 1424{
 1425	if (clamp_id == UCLAMP_MIN)
 1426		return 0;
 1427	return SCHED_CAPACITY_SCALE;
 1428}
 1429
 1430static inline void uclamp_se_set(struct uclamp_se *uc_se,
 1431				 unsigned int value, bool user_defined)
 1432{
 1433	uc_se->value = value;
 1434	uc_se->bucket_id = uclamp_bucket_id(value);
 1435	uc_se->user_defined = user_defined;
 1436}
 1437
 1438static inline unsigned int
 1439uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
 1440		  unsigned int clamp_value)
 1441{
 1442	/*
 1443	 * Avoid blocked utilization pushing up the frequency when we go
 1444	 * idle (which drops the max-clamp) by retaining the last known
 1445	 * max-clamp.
 1446	 */
 1447	if (clamp_id == UCLAMP_MAX) {
 1448		rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
 1449		return clamp_value;
 1450	}
 1451
 1452	return uclamp_none(UCLAMP_MIN);
 1453}
 1454
 1455static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
 1456				     unsigned int clamp_value)
 1457{
 1458	/* Reset max-clamp retention only on idle exit */
 1459	if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
 1460		return;
 1461
 1462	uclamp_rq_set(rq, clamp_id, clamp_value);
 1463}
 1464
 1465static inline
 1466unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
 1467				   unsigned int clamp_value)
 1468{
 1469	struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
 1470	int bucket_id = UCLAMP_BUCKETS - 1;
 1471
 1472	/*
 1473	 * Since both min and max clamps are max aggregated, find the
 1474	 * top most bucket with tasks in.
 1475	 */
 1476	for ( ; bucket_id >= 0; bucket_id--) {
 1477		if (!bucket[bucket_id].tasks)
 1478			continue;
 1479		return bucket[bucket_id].value;
 1480	}
 1481
 1482	/* No tasks -- default clamp values */
 1483	return uclamp_idle_value(rq, clamp_id, clamp_value);
 1484}
 1485
 1486static void __uclamp_update_util_min_rt_default(struct task_struct *p)
 1487{
 1488	unsigned int default_util_min;
 1489	struct uclamp_se *uc_se;
 1490
 1491	lockdep_assert_held(&p->pi_lock);
 1492
 1493	uc_se = &p->uclamp_req[UCLAMP_MIN];
 1494
 1495	/* Only sync if user didn't override the default */
 1496	if (uc_se->user_defined)
 1497		return;
 1498
 1499	default_util_min = sysctl_sched_uclamp_util_min_rt_default;
 1500	uclamp_se_set(uc_se, default_util_min, false);
 1501}
 1502
 1503static void uclamp_update_util_min_rt_default(struct task_struct *p)
 1504{
 1505	if (!rt_task(p))
 1506		return;
 1507
 1508	/* Protect updates to p->uclamp_* */
 1509	guard(task_rq_lock)(p);
 1510	__uclamp_update_util_min_rt_default(p);
 1511}
 1512
 1513static inline struct uclamp_se
 1514uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
 1515{
 1516	/* Copy by value as we could modify it */
 1517	struct uclamp_se uc_req = p->uclamp_req[clamp_id];
 1518#ifdef CONFIG_UCLAMP_TASK_GROUP
 1519	unsigned int tg_min, tg_max, value;
 1520
 1521	/*
 1522	 * Tasks in autogroups or root task group will be
 1523	 * restricted by system defaults.
 1524	 */
 1525	if (task_group_is_autogroup(task_group(p)))
 1526		return uc_req;
 1527	if (task_group(p) == &root_task_group)
 1528		return uc_req;
 1529
 1530	tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
 1531	tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
 1532	value = uc_req.value;
 1533	value = clamp(value, tg_min, tg_max);
 1534	uclamp_se_set(&uc_req, value, false);
 1535#endif
 1536
 1537	return uc_req;
 1538}
 1539
 1540/*
 1541 * The effective clamp bucket index of a task depends on, by increasing
 1542 * priority:
 1543 * - the task specific clamp value, when explicitly requested from userspace
 1544 * - the task group effective clamp value, for tasks not either in the root
 1545 *   group or in an autogroup
 1546 * - the system default clamp value, defined by the sysadmin
 1547 */
 1548static inline struct uclamp_se
 1549uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
 1550{
 1551	struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
 1552	struct uclamp_se uc_max = uclamp_default[clamp_id];
 1553
 1554	/* System default restrictions always apply */
 1555	if (unlikely(uc_req.value > uc_max.value))
 1556		return uc_max;
 1557
 1558	return uc_req;
 1559}
 1560
 1561unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
 1562{
 1563	struct uclamp_se uc_eff;
 1564
 1565	/* Task currently refcounted: use back-annotated (effective) value */
 1566	if (p->uclamp[clamp_id].active)
 1567		return (unsigned long)p->uclamp[clamp_id].value;
 1568
 1569	uc_eff = uclamp_eff_get(p, clamp_id);
 1570
 1571	return (unsigned long)uc_eff.value;
 1572}
 1573
 1574/*
 1575 * When a task is enqueued on a rq, the clamp bucket currently defined by the
 1576 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
 1577 * updates the rq's clamp value if required.
 1578 *
 1579 * Tasks can have a task-specific value requested from user-space, track
 1580 * within each bucket the maximum value for tasks refcounted in it.
 1581 * This "local max aggregation" allows to track the exact "requested" value
 1582 * for each bucket when all its RUNNABLE tasks require the same clamp.
 1583 */
 1584static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
 1585				    enum uclamp_id clamp_id)
 1586{
 1587	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
 1588	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
 1589	struct uclamp_bucket *bucket;
 1590
 1591	lockdep_assert_rq_held(rq);
 1592
 1593	/* Update task effective clamp */
 1594	p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
 1595
 1596	bucket = &uc_rq->bucket[uc_se->bucket_id];
 1597	bucket->tasks++;
 1598	uc_se->active = true;
 1599
 1600	uclamp_idle_reset(rq, clamp_id, uc_se->value);
 1601
 1602	/*
 1603	 * Local max aggregation: rq buckets always track the max
 1604	 * "requested" clamp value of its RUNNABLE tasks.
 1605	 */
 1606	if (bucket->tasks == 1 || uc_se->value > bucket->value)
 1607		bucket->value = uc_se->value;
 1608
 1609	if (uc_se->value > uclamp_rq_get(rq, clamp_id))
 1610		uclamp_rq_set(rq, clamp_id, uc_se->value);
 1611}
 1612
 1613/*
 1614 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
 1615 * is released. If this is the last task reference counting the rq's max
 1616 * active clamp value, then the rq's clamp value is updated.
 1617 *
 1618 * Both refcounted tasks and rq's cached clamp values are expected to be
 1619 * always valid. If it's detected they are not, as defensive programming,
 1620 * enforce the expected state and warn.
 1621 */
 1622static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
 1623				    enum uclamp_id clamp_id)
 1624{
 1625	struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
 1626	struct uclamp_se *uc_se = &p->uclamp[clamp_id];
 1627	struct uclamp_bucket *bucket;
 1628	unsigned int bkt_clamp;
 1629	unsigned int rq_clamp;
 1630
 1631	lockdep_assert_rq_held(rq);
 1632
 1633	/*
 1634	 * If sched_uclamp_used was enabled after task @p was enqueued,
 1635	 * we could end up with unbalanced call to uclamp_rq_dec_id().
 1636	 *
 1637	 * In this case the uc_se->active flag should be false since no uclamp
 1638	 * accounting was performed at enqueue time and we can just return
 1639	 * here.
 1640	 *
 1641	 * Need to be careful of the following enqueue/dequeue ordering
 1642	 * problem too
 1643	 *
 1644	 *	enqueue(taskA)
 1645	 *	// sched_uclamp_used gets enabled
 1646	 *	enqueue(taskB)
 1647	 *	dequeue(taskA)
 1648	 *	// Must not decrement bucket->tasks here
 1649	 *	dequeue(taskB)
 1650	 *
 1651	 * where we could end up with stale data in uc_se and
 1652	 * bucket[uc_se->bucket_id].
 1653	 *
 1654	 * The following check here eliminates the possibility of such race.
 1655	 */
 1656	if (unlikely(!uc_se->active))
 1657		return;
 1658
 1659	bucket = &uc_rq->bucket[uc_se->bucket_id];
 1660
 1661	SCHED_WARN_ON(!bucket->tasks);
 1662	if (likely(bucket->tasks))
 1663		bucket->tasks--;
 1664
 1665	uc_se->active = false;
 1666
 1667	/*
 1668	 * Keep "local max aggregation" simple and accept to (possibly)
 1669	 * overboost some RUNNABLE tasks in the same bucket.
 1670	 * The rq clamp bucket value is reset to its base value whenever
 1671	 * there are no more RUNNABLE tasks refcounting it.
 1672	 */
 1673	if (likely(bucket->tasks))
 1674		return;
 1675
 1676	rq_clamp = uclamp_rq_get(rq, clamp_id);
 1677	/*
 1678	 * Defensive programming: this should never happen. If it happens,
 1679	 * e.g. due to future modification, warn and fixup the expected value.
 1680	 */
 1681	SCHED_WARN_ON(bucket->value > rq_clamp);
 1682	if (bucket->value >= rq_clamp) {
 1683		bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
 1684		uclamp_rq_set(rq, clamp_id, bkt_clamp);
 1685	}
 1686}
 1687
 1688static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
 1689{
 1690	enum uclamp_id clamp_id;
 1691
 1692	/*
 1693	 * Avoid any overhead until uclamp is actually used by the userspace.
 1694	 *
 1695	 * The condition is constructed such that a NOP is generated when
 1696	 * sched_uclamp_used is disabled.
 1697	 */
 1698	if (!static_branch_unlikely(&sched_uclamp_used))
 1699		return;
 1700
 1701	if (unlikely(!p->sched_class->uclamp_enabled))
 1702		return;
 1703
 1704	for_each_clamp_id(clamp_id)
 1705		uclamp_rq_inc_id(rq, p, clamp_id);
 1706
 1707	/* Reset clamp idle holding when there is one RUNNABLE task */
 1708	if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
 1709		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
 1710}
 1711
 1712static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
 1713{
 1714	enum uclamp_id clamp_id;
 1715
 1716	/*
 1717	 * Avoid any overhead until uclamp is actually used by the userspace.
 1718	 *
 1719	 * The condition is constructed such that a NOP is generated when
 1720	 * sched_uclamp_used is disabled.
 1721	 */
 1722	if (!static_branch_unlikely(&sched_uclamp_used))
 1723		return;
 1724
 1725	if (unlikely(!p->sched_class->uclamp_enabled))
 1726		return;
 1727
 1728	for_each_clamp_id(clamp_id)
 1729		uclamp_rq_dec_id(rq, p, clamp_id);
 1730}
 1731
 1732static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
 1733				      enum uclamp_id clamp_id)
 1734{
 1735	if (!p->uclamp[clamp_id].active)
 1736		return;
 1737
 1738	uclamp_rq_dec_id(rq, p, clamp_id);
 1739	uclamp_rq_inc_id(rq, p, clamp_id);
 1740
 1741	/*
 1742	 * Make sure to clear the idle flag if we've transiently reached 0
 1743	 * active tasks on rq.
 1744	 */
 1745	if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
 1746		rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
 1747}
 1748
 1749static inline void
 1750uclamp_update_active(struct task_struct *p)
 1751{
 1752	enum uclamp_id clamp_id;
 1753	struct rq_flags rf;
 1754	struct rq *rq;
 1755
 1756	/*
 1757	 * Lock the task and the rq where the task is (or was) queued.
 1758	 *
 1759	 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
 1760	 * price to pay to safely serialize util_{min,max} updates with
 1761	 * enqueues, dequeues and migration operations.
 1762	 * This is the same locking schema used by __set_cpus_allowed_ptr().
 1763	 */
 1764	rq = task_rq_lock(p, &rf);
 1765
 1766	/*
 1767	 * Setting the clamp bucket is serialized by task_rq_lock().
 1768	 * If the task is not yet RUNNABLE and its task_struct is not
 1769	 * affecting a valid clamp bucket, the next time it's enqueued,
 1770	 * it will already see the updated clamp bucket value.
 1771	 */
 1772	for_each_clamp_id(clamp_id)
 1773		uclamp_rq_reinc_id(rq, p, clamp_id);
 1774
 1775	task_rq_unlock(rq, p, &rf);
 1776}
 1777
 1778#ifdef CONFIG_UCLAMP_TASK_GROUP
 1779static inline void
 1780uclamp_update_active_tasks(struct cgroup_subsys_state *css)
 1781{
 1782	struct css_task_iter it;
 1783	struct task_struct *p;
 1784
 1785	css_task_iter_start(css, 0, &it);
 1786	while ((p = css_task_iter_next(&it)))
 1787		uclamp_update_active(p);
 1788	css_task_iter_end(&it);
 1789}
 1790
 1791static void cpu_util_update_eff(struct cgroup_subsys_state *css);
 1792#endif
 1793
 1794#ifdef CONFIG_SYSCTL
 1795#ifdef CONFIG_UCLAMP_TASK
 1796#ifdef CONFIG_UCLAMP_TASK_GROUP
 1797static void uclamp_update_root_tg(void)
 1798{
 1799	struct task_group *tg = &root_task_group;
 1800
 1801	uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
 1802		      sysctl_sched_uclamp_util_min, false);
 1803	uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
 1804		      sysctl_sched_uclamp_util_max, false);
 1805
 1806	guard(rcu)();
 1807	cpu_util_update_eff(&root_task_group.css);
 1808}
 1809#else
 1810static void uclamp_update_root_tg(void) { }
 1811#endif
 1812
 1813static void uclamp_sync_util_min_rt_default(void)
 1814{
 1815	struct task_struct *g, *p;
 1816
 1817	/*
 1818	 * copy_process()			sysctl_uclamp
 1819	 *					  uclamp_min_rt = X;
 1820	 *   write_lock(&tasklist_lock)		  read_lock(&tasklist_lock)
 1821	 *   // link thread			  smp_mb__after_spinlock()
 1822	 *   write_unlock(&tasklist_lock)	  read_unlock(&tasklist_lock);
 1823	 *   sched_post_fork()			  for_each_process_thread()
 1824	 *     __uclamp_sync_rt()		    __uclamp_sync_rt()
 1825	 *
 1826	 * Ensures that either sched_post_fork() will observe the new
 1827	 * uclamp_min_rt or for_each_process_thread() will observe the new
 1828	 * task.
 1829	 */
 1830	read_lock(&tasklist_lock);
 1831	smp_mb__after_spinlock();
 1832	read_unlock(&tasklist_lock);
 1833
 1834	guard(rcu)();
 1835	for_each_process_thread(g, p)
 1836		uclamp_update_util_min_rt_default(p);
 1837}
 1838
 1839static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
 1840				void *buffer, size_t *lenp, loff_t *ppos)
 1841{
 1842	bool update_root_tg = false;
 1843	int old_min, old_max, old_min_rt;
 1844	int result;
 1845
 1846	guard(mutex)(&uclamp_mutex);
 1847
 1848	old_min = sysctl_sched_uclamp_util_min;
 1849	old_max = sysctl_sched_uclamp_util_max;
 1850	old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
 1851
 1852	result = proc_dointvec(table, write, buffer, lenp, ppos);
 1853	if (result)
 1854		goto undo;
 1855	if (!write)
 1856		return 0;
 1857
 1858	if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
 1859	    sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE	||
 1860	    sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
 1861
 1862		result = -EINVAL;
 1863		goto undo;
 1864	}
 1865
 1866	if (old_min != sysctl_sched_uclamp_util_min) {
 1867		uclamp_se_set(&uclamp_default[UCLAMP_MIN],
 1868			      sysctl_sched_uclamp_util_min, false);
 1869		update_root_tg = true;
 1870	}
 1871	if (old_max != sysctl_sched_uclamp_util_max) {
 1872		uclamp_se_set(&uclamp_default[UCLAMP_MAX],
 1873			      sysctl_sched_uclamp_util_max, false);
 1874		update_root_tg = true;
 1875	}
 1876
 1877	if (update_root_tg) {
 1878		static_branch_enable(&sched_uclamp_used);
 1879		uclamp_update_root_tg();
 1880	}
 1881
 1882	if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
 1883		static_branch_enable(&sched_uclamp_used);
 1884		uclamp_sync_util_min_rt_default();
 1885	}
 1886
 1887	/*
 1888	 * We update all RUNNABLE tasks only when task groups are in use.
 1889	 * Otherwise, keep it simple and do just a lazy update at each next
 1890	 * task enqueue time.
 1891	 */
 1892	return 0;
 1893
 1894undo:
 1895	sysctl_sched_uclamp_util_min = old_min;
 1896	sysctl_sched_uclamp_util_max = old_max;
 1897	sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
 1898	return result;
 1899}
 1900#endif
 1901#endif
 1902
 1903static int uclamp_validate(struct task_struct *p,
 1904			   const struct sched_attr *attr)
 1905{
 1906	int util_min = p->uclamp_req[UCLAMP_MIN].value;
 1907	int util_max = p->uclamp_req[UCLAMP_MAX].value;
 1908
 1909	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
 1910		util_min = attr->sched_util_min;
 1911
 1912		if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
 1913			return -EINVAL;
 1914	}
 1915
 1916	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
 1917		util_max = attr->sched_util_max;
 1918
 1919		if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
 1920			return -EINVAL;
 1921	}
 1922
 1923	if (util_min != -1 && util_max != -1 && util_min > util_max)
 1924		return -EINVAL;
 1925
 1926	/*
 1927	 * We have valid uclamp attributes; make sure uclamp is enabled.
 1928	 *
 1929	 * We need to do that here, because enabling static branches is a
 1930	 * blocking operation which obviously cannot be done while holding
 1931	 * scheduler locks.
 1932	 */
 1933	static_branch_enable(&sched_uclamp_used);
 1934
 1935	return 0;
 1936}
 1937
 1938static bool uclamp_reset(const struct sched_attr *attr,
 1939			 enum uclamp_id clamp_id,
 1940			 struct uclamp_se *uc_se)
 1941{
 1942	/* Reset on sched class change for a non user-defined clamp value. */
 1943	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
 1944	    !uc_se->user_defined)
 1945		return true;
 1946
 1947	/* Reset on sched_util_{min,max} == -1. */
 1948	if (clamp_id == UCLAMP_MIN &&
 1949	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
 1950	    attr->sched_util_min == -1) {
 1951		return true;
 1952	}
 1953
 1954	if (clamp_id == UCLAMP_MAX &&
 1955	    attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
 1956	    attr->sched_util_max == -1) {
 1957		return true;
 1958	}
 1959
 1960	return false;
 1961}
 1962
 1963static void __setscheduler_uclamp(struct task_struct *p,
 1964				  const struct sched_attr *attr)
 1965{
 1966	enum uclamp_id clamp_id;
 1967
 1968	for_each_clamp_id(clamp_id) {
 1969		struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
 1970		unsigned int value;
 1971
 1972		if (!uclamp_reset(attr, clamp_id, uc_se))
 1973			continue;
 1974
 1975		/*
 1976		 * RT by default have a 100% boost value that could be modified
 1977		 * at runtime.
 1978		 */
 1979		if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
 1980			value = sysctl_sched_uclamp_util_min_rt_default;
 1981		else
 1982			value = uclamp_none(clamp_id);
 1983
 1984		uclamp_se_set(uc_se, value, false);
 1985
 1986	}
 1987
 1988	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
 1989		return;
 1990
 1991	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
 1992	    attr->sched_util_min != -1) {
 1993		uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
 1994			      attr->sched_util_min, true);
 1995	}
 1996
 1997	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
 1998	    attr->sched_util_max != -1) {
 1999		uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
 2000			      attr->sched_util_max, true);
 2001	}
 2002}
 2003
 2004static void uclamp_fork(struct task_struct *p)
 2005{
 2006	enum uclamp_id clamp_id;
 2007
 2008	/*
 2009	 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
 2010	 * as the task is still at its early fork stages.
 2011	 */
 2012	for_each_clamp_id(clamp_id)
 2013		p->uclamp[clamp_id].active = false;
 2014
 2015	if (likely(!p->sched_reset_on_fork))
 2016		return;
 2017
 2018	for_each_clamp_id(clamp_id) {
 2019		uclamp_se_set(&p->uclamp_req[clamp_id],
 2020			      uclamp_none(clamp_id), false);
 2021	}
 2022}
 2023
 2024static void uclamp_post_fork(struct task_struct *p)
 2025{
 2026	uclamp_update_util_min_rt_default(p);
 2027}
 2028
 2029static void __init init_uclamp_rq(struct rq *rq)
 2030{
 2031	enum uclamp_id clamp_id;
 2032	struct uclamp_rq *uc_rq = rq->uclamp;
 2033
 2034	for_each_clamp_id(clamp_id) {
 2035		uc_rq[clamp_id] = (struct uclamp_rq) {
 2036			.value = uclamp_none(clamp_id)
 2037		};
 2038	}
 2039
 2040	rq->uclamp_flags = UCLAMP_FLAG_IDLE;
 2041}
 2042
 2043static void __init init_uclamp(void)
 2044{
 2045	struct uclamp_se uc_max = {};
 2046	enum uclamp_id clamp_id;
 2047	int cpu;
 2048
 2049	for_each_possible_cpu(cpu)
 2050		init_uclamp_rq(cpu_rq(cpu));
 2051
 2052	for_each_clamp_id(clamp_id) {
 2053		uclamp_se_set(&init_task.uclamp_req[clamp_id],
 2054			      uclamp_none(clamp_id), false);
 2055	}
 2056
 2057	/* System defaults allow max clamp values for both indexes */
 2058	uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
 2059	for_each_clamp_id(clamp_id) {
 2060		uclamp_default[clamp_id] = uc_max;
 2061#ifdef CONFIG_UCLAMP_TASK_GROUP
 2062		root_task_group.uclamp_req[clamp_id] = uc_max;
 2063		root_task_group.uclamp[clamp_id] = uc_max;
 2064#endif
 2065	}
 2066}
 2067
 2068#else /* CONFIG_UCLAMP_TASK */
 2069static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
 2070static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
 2071static inline int uclamp_validate(struct task_struct *p,
 2072				  const struct sched_attr *attr)
 2073{
 2074	return -EOPNOTSUPP;
 2075}
 2076static void __setscheduler_uclamp(struct task_struct *p,
 2077				  const struct sched_attr *attr) { }
 2078static inline void uclamp_fork(struct task_struct *p) { }
 2079static inline void uclamp_post_fork(struct task_struct *p) { }
 2080static inline void init_uclamp(void) { }
 2081#endif /* CONFIG_UCLAMP_TASK */
 2082
 2083bool sched_task_on_rq(struct task_struct *p)
 2084{
 2085	return task_on_rq_queued(p);
 2086}
 2087
 2088unsigned long get_wchan(struct task_struct *p)
 2089{
 2090	unsigned long ip = 0;
 2091	unsigned int state;
 2092
 2093	if (!p || p == current)
 2094		return 0;
 2095
 2096	/* Only get wchan if task is blocked and we can keep it that way. */
 2097	raw_spin_lock_irq(&p->pi_lock);
 2098	state = READ_ONCE(p->__state);
 2099	smp_rmb(); /* see try_to_wake_up() */
 2100	if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
 2101		ip = __get_wchan(p);
 2102	raw_spin_unlock_irq(&p->pi_lock);
 2103
 2104	return ip;
 2105}
 2106
 2107static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
 2108{
 2109	if (!(flags & ENQUEUE_NOCLOCK))
 2110		update_rq_clock(rq);
 2111
 2112	if (!(flags & ENQUEUE_RESTORE)) {
 2113		sched_info_enqueue(rq, p);
 2114		psi_enqueue(p, (flags & ENQUEUE_WAKEUP) && !(flags & ENQUEUE_MIGRATED));
 2115	}
 2116
 2117	uclamp_rq_inc(rq, p);
 2118	p->sched_class->enqueue_task(rq, p, flags);
 2119
 2120	if (sched_core_enabled(rq))
 2121		sched_core_enqueue(rq, p);
 2122}
 2123
 2124static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
 2125{
 2126	if (sched_core_enabled(rq))
 2127		sched_core_dequeue(rq, p, flags);
 2128
 2129	if (!(flags & DEQUEUE_NOCLOCK))
 2130		update_rq_clock(rq);
 2131
 2132	if (!(flags & DEQUEUE_SAVE)) {
 2133		sched_info_dequeue(rq, p);
 2134		psi_dequeue(p, flags & DEQUEUE_SLEEP);
 2135	}
 2136
 2137	uclamp_rq_dec(rq, p);
 2138	p->sched_class->dequeue_task(rq, p, flags);
 2139}
 2140
 2141void activate_task(struct rq *rq, struct task_struct *p, int flags)
 2142{
 2143	if (task_on_rq_migrating(p))
 2144		flags |= ENQUEUE_MIGRATED;
 2145	if (flags & ENQUEUE_MIGRATED)
 2146		sched_mm_cid_migrate_to(rq, p);
 2147
 2148	enqueue_task(rq, p, flags);
 2149
 2150	WRITE_ONCE(p->on_rq, TASK_ON_RQ_QUEUED);
 2151	ASSERT_EXCLUSIVE_WRITER(p->on_rq);
 2152}
 2153
 2154void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
 2155{
 2156	WRITE_ONCE(p->on_rq, (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING);
 2157	ASSERT_EXCLUSIVE_WRITER(p->on_rq);
 2158
 2159	dequeue_task(rq, p, flags);
 2160}
 2161
 2162static inline int __normal_prio(int policy, int rt_prio, int nice)
 
 
 
 2163{
 2164	int prio;
 2165
 2166	if (dl_policy(policy))
 2167		prio = MAX_DL_PRIO - 1;
 2168	else if (rt_policy(policy))
 2169		prio = MAX_RT_PRIO - 1 - rt_prio;
 2170	else
 2171		prio = NICE_TO_PRIO(nice);
 2172
 2173	return prio;
 2174}
 2175
 2176/*
 2177 * Calculate the expected normal priority: i.e. priority
 2178 * without taking RT-inheritance into account. Might be
 2179 * boosted by interactivity modifiers. Changes upon fork,
 2180 * setprio syscalls, and whenever the interactivity
 2181 * estimator recalculates.
 2182 */
 2183static inline int normal_prio(struct task_struct *p)
 2184{
 2185	return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
 
 
 
 
 
 
 
 
 2186}
 2187
 2188/*
 2189 * Calculate the current priority, i.e. the priority
 2190 * taken into account by the scheduler. This value might
 2191 * be boosted by RT tasks, or might be boosted by
 2192 * interactivity modifiers. Will be RT if the task got
 2193 * RT-boosted. If not then it returns p->normal_prio.
 2194 */
 2195static int effective_prio(struct task_struct *p)
 2196{
 2197	p->normal_prio = normal_prio(p);
 2198	/*
 2199	 * If we are RT tasks or we were boosted to RT priority,
 2200	 * keep the priority unchanged. Otherwise, update priority
 2201	 * to the normal priority:
 2202	 */
 2203	if (!rt_prio(p->prio))
 2204		return p->normal_prio;
 2205	return p->prio;
 2206}
 2207
 2208/**
 2209 * task_curr - is this task currently executing on a CPU?
 2210 * @p: the task in question.
 2211 *
 2212 * Return: 1 if the task is currently executing. 0 otherwise.
 2213 */
 2214inline int task_curr(const struct task_struct *p)
 2215{
 2216	return cpu_curr(task_cpu(p)) == p;
 2217}
 2218
 2219/*
 2220 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
 2221 * use the balance_callback list if you want balancing.
 2222 *
 2223 * this means any call to check_class_changed() must be followed by a call to
 2224 * balance_callback().
 2225 */
 2226static inline void check_class_changed(struct rq *rq, struct task_struct *p,
 2227				       const struct sched_class *prev_class,
 2228				       int oldprio)
 2229{
 2230	if (prev_class != p->sched_class) {
 2231		if (prev_class->switched_from)
 2232			prev_class->switched_from(rq, p);
 2233
 2234		p->sched_class->switched_to(rq, p);
 2235	} else if (oldprio != p->prio || dl_task(p))
 2236		p->sched_class->prio_changed(rq, p, oldprio);
 2237}
 2238
 2239void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags)
 2240{
 2241	if (p->sched_class == rq->curr->sched_class)
 2242		rq->curr->sched_class->wakeup_preempt(rq, p, flags);
 2243	else if (sched_class_above(p->sched_class, rq->curr->sched_class))
 2244		resched_curr(rq);
 
 
 
 
 
 
 
 
 
 
 2245
 2246	/*
 2247	 * A queue event has occurred, and we're going to schedule.  In
 2248	 * this case, we can save a useless back to back clock update.
 2249	 */
 2250	if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
 2251		rq_clock_skip_update(rq);
 2252}
 2253
 2254static __always_inline
 2255int __task_state_match(struct task_struct *p, unsigned int state)
 2256{
 2257	if (READ_ONCE(p->__state) & state)
 2258		return 1;
 2259
 2260	if (READ_ONCE(p->saved_state) & state)
 2261		return -1;
 2262
 2263	return 0;
 2264}
 2265
 2266static __always_inline
 2267int task_state_match(struct task_struct *p, unsigned int state)
 2268{
 2269	/*
 2270	 * Serialize against current_save_and_set_rtlock_wait_state(),
 2271	 * current_restore_rtlock_saved_state(), and __refrigerator().
 2272	 */
 2273	guard(raw_spinlock_irq)(&p->pi_lock);
 2274	return __task_state_match(p, state);
 2275}
 2276
 2277/*
 2278 * wait_task_inactive - wait for a thread to unschedule.
 2279 *
 2280 * Wait for the thread to block in any of the states set in @match_state.
 2281 * If it changes, i.e. @p might have woken up, then return zero.  When we
 2282 * succeed in waiting for @p to be off its CPU, we return a positive number
 2283 * (its total switch count).  If a second call a short while later returns the
 2284 * same number, the caller can be sure that @p has remained unscheduled the
 2285 * whole time.
 2286 *
 2287 * The caller must ensure that the task *will* unschedule sometime soon,
 2288 * else this function might spin for a *long* time. This function can't
 2289 * be called with interrupts off, or it may introduce deadlock with
 2290 * smp_call_function() if an IPI is sent by the same process we are
 2291 * waiting to become inactive.
 2292 */
 2293unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
 2294{
 2295	int running, queued, match;
 2296	struct rq_flags rf;
 2297	unsigned long ncsw;
 2298	struct rq *rq;
 2299
 2300	for (;;) {
 2301		/*
 2302		 * We do the initial early heuristics without holding
 2303		 * any task-queue locks at all. We'll only try to get
 2304		 * the runqueue lock when things look like they will
 2305		 * work out!
 2306		 */
 2307		rq = task_rq(p);
 2308
 2309		/*
 2310		 * If the task is actively running on another CPU
 2311		 * still, just relax and busy-wait without holding
 2312		 * any locks.
 2313		 *
 2314		 * NOTE! Since we don't hold any locks, it's not
 2315		 * even sure that "rq" stays as the right runqueue!
 2316		 * But we don't care, since "task_on_cpu()" will
 2317		 * return false if the runqueue has changed and p
 2318		 * is actually now running somewhere else!
 2319		 */
 2320		while (task_on_cpu(rq, p)) {
 2321			if (!task_state_match(p, match_state))
 2322				return 0;
 2323			cpu_relax();
 2324		}
 2325
 2326		/*
 2327		 * Ok, time to look more closely! We need the rq
 2328		 * lock now, to be *sure*. If we're wrong, we'll
 2329		 * just go back and repeat.
 2330		 */
 2331		rq = task_rq_lock(p, &rf);
 2332		trace_sched_wait_task(p);
 2333		running = task_on_cpu(rq, p);
 2334		queued = task_on_rq_queued(p);
 2335		ncsw = 0;
 2336		if ((match = __task_state_match(p, match_state))) {
 2337			/*
 2338			 * When matching on p->saved_state, consider this task
 2339			 * still queued so it will wait.
 2340			 */
 2341			if (match < 0)
 2342				queued = 1;
 2343			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
 2344		}
 2345		task_rq_unlock(rq, p, &rf);
 2346
 2347		/*
 2348		 * If it changed from the expected state, bail out now.
 2349		 */
 2350		if (unlikely(!ncsw))
 2351			break;
 2352
 2353		/*
 2354		 * Was it really running after all now that we
 2355		 * checked with the proper locks actually held?
 2356		 *
 2357		 * Oops. Go back and try again..
 2358		 */
 2359		if (unlikely(running)) {
 2360			cpu_relax();
 2361			continue;
 2362		}
 2363
 2364		/*
 2365		 * It's not enough that it's not actively running,
 2366		 * it must be off the runqueue _entirely_, and not
 2367		 * preempted!
 2368		 *
 2369		 * So if it was still runnable (but just not actively
 2370		 * running right now), it's preempted, and we should
 2371		 * yield - it could be a while.
 2372		 */
 2373		if (unlikely(queued)) {
 2374			ktime_t to = NSEC_PER_SEC / HZ;
 2375
 2376			set_current_state(TASK_UNINTERRUPTIBLE);
 2377			schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
 2378			continue;
 2379		}
 2380
 2381		/*
 2382		 * Ahh, all good. It wasn't running, and it wasn't
 2383		 * runnable, which means that it will never become
 2384		 * running in the future either. We're all done!
 2385		 */
 2386		break;
 2387	}
 2388
 2389	return ncsw;
 2390}
 2391
 2392#ifdef CONFIG_SMP
 2393
 2394static void
 2395__do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
 2396
 2397static int __set_cpus_allowed_ptr(struct task_struct *p,
 2398				  struct affinity_context *ctx);
 2399
 2400static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
 2401{
 2402	struct affinity_context ac = {
 2403		.new_mask  = cpumask_of(rq->cpu),
 2404		.flags     = SCA_MIGRATE_DISABLE,
 2405	};
 2406
 2407	if (likely(!p->migration_disabled))
 2408		return;
 2409
 2410	if (p->cpus_ptr != &p->cpus_mask)
 2411		return;
 2412
 2413	/*
 2414	 * Violates locking rules! see comment in __do_set_cpus_allowed().
 2415	 */
 2416	__do_set_cpus_allowed(p, &ac);
 2417}
 2418
 2419void migrate_disable(void)
 2420{
 2421	struct task_struct *p = current;
 2422
 2423	if (p->migration_disabled) {
 2424		p->migration_disabled++;
 2425		return;
 2426	}
 2427
 2428	guard(preempt)();
 2429	this_rq()->nr_pinned++;
 2430	p->migration_disabled = 1;
 2431}
 2432EXPORT_SYMBOL_GPL(migrate_disable);
 2433
 2434void migrate_enable(void)
 2435{
 2436	struct task_struct *p = current;
 2437	struct affinity_context ac = {
 2438		.new_mask  = &p->cpus_mask,
 2439		.flags     = SCA_MIGRATE_ENABLE,
 2440	};
 2441
 2442	if (p->migration_disabled > 1) {
 2443		p->migration_disabled--;
 2444		return;
 2445	}
 2446
 2447	if (WARN_ON_ONCE(!p->migration_disabled))
 2448		return;
 2449
 2450	/*
 2451	 * Ensure stop_task runs either before or after this, and that
 2452	 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
 2453	 */
 2454	guard(preempt)();
 2455	if (p->cpus_ptr != &p->cpus_mask)
 2456		__set_cpus_allowed_ptr(p, &ac);
 2457	/*
 2458	 * Mustn't clear migration_disabled() until cpus_ptr points back at the
 2459	 * regular cpus_mask, otherwise things that race (eg.
 2460	 * select_fallback_rq) get confused.
 2461	 */
 2462	barrier();
 2463	p->migration_disabled = 0;
 2464	this_rq()->nr_pinned--;
 2465}
 2466EXPORT_SYMBOL_GPL(migrate_enable);
 2467
 2468static inline bool rq_has_pinned_tasks(struct rq *rq)
 2469{
 2470	return rq->nr_pinned;
 2471}
 2472
 2473/*
 2474 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
 2475 * __set_cpus_allowed_ptr() and select_fallback_rq().
 2476 */
 2477static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
 2478{
 2479	/* When not in the task's cpumask, no point in looking further. */
 2480	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
 2481		return false;
 2482
 2483	/* migrate_disabled() must be allowed to finish. */
 2484	if (is_migration_disabled(p))
 2485		return cpu_online(cpu);
 2486
 2487	/* Non kernel threads are not allowed during either online or offline. */
 2488	if (!(p->flags & PF_KTHREAD))
 2489		return cpu_active(cpu) && task_cpu_possible(cpu, p);
 2490
 2491	/* KTHREAD_IS_PER_CPU is always allowed. */
 2492	if (kthread_is_per_cpu(p))
 2493		return cpu_online(cpu);
 2494
 2495	/* Regular kernel threads don't get to stay during offline. */
 2496	if (cpu_dying(cpu))
 2497		return false;
 2498
 2499	/* But are allowed during online. */
 2500	return cpu_online(cpu);
 2501}
 2502
 2503/*
 2504 * This is how migration works:
 2505 *
 2506 * 1) we invoke migration_cpu_stop() on the target CPU using
 2507 *    stop_one_cpu().
 2508 * 2) stopper starts to run (implicitly forcing the migrated thread
 2509 *    off the CPU)
 2510 * 3) it checks whether the migrated task is still in the wrong runqueue.
 2511 * 4) if it's in the wrong runqueue then the migration thread removes
 2512 *    it and puts it into the right queue.
 2513 * 5) stopper completes and stop_one_cpu() returns and the migration
 2514 *    is done.
 2515 */
 2516
 2517/*
 2518 * move_queued_task - move a queued task to new rq.
 2519 *
 2520 * Returns (locked) new rq. Old rq's lock is released.
 2521 */
 2522static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
 2523				   struct task_struct *p, int new_cpu)
 2524{
 2525	lockdep_assert_rq_held(rq);
 2526
 2527	deactivate_task(rq, p, DEQUEUE_NOCLOCK);
 
 2528	set_task_cpu(p, new_cpu);
 2529	rq_unlock(rq, rf);
 2530
 2531	rq = cpu_rq(new_cpu);
 2532
 2533	rq_lock(rq, rf);
 2534	WARN_ON_ONCE(task_cpu(p) != new_cpu);
 2535	activate_task(rq, p, 0);
 2536	wakeup_preempt(rq, p, 0);
 
 2537
 2538	return rq;
 2539}
 2540
 2541struct migration_arg {
 2542	struct task_struct		*task;
 2543	int				dest_cpu;
 2544	struct set_affinity_pending	*pending;
 2545};
 2546
 2547/*
 2548 * @refs: number of wait_for_completion()
 2549 * @stop_pending: is @stop_work in use
 2550 */
 2551struct set_affinity_pending {
 2552	refcount_t		refs;
 2553	unsigned int		stop_pending;
 2554	struct completion	done;
 2555	struct cpu_stop_work	stop_work;
 2556	struct migration_arg	arg;
 2557};
 2558
 2559/*
 2560 * Move (not current) task off this CPU, onto the destination CPU. We're doing
 2561 * this because either it can't run here any more (set_cpus_allowed()
 2562 * away from this CPU, or CPU going down), or because we're
 2563 * attempting to rebalance this task on exec (sched_exec).
 2564 *
 2565 * So we race with normal scheduler movements, but that's OK, as long
 2566 * as the task is no longer on this CPU.
 2567 */
 2568static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
 2569				 struct task_struct *p, int dest_cpu)
 2570{
 2571	/* Affinity changed (again). */
 2572	if (!is_cpu_allowed(p, dest_cpu))
 2573		return rq;
 2574
 
 2575	rq = move_queued_task(rq, rf, p, dest_cpu);
 2576
 2577	return rq;
 2578}
 2579
 2580/*
 2581 * migration_cpu_stop - this will be executed by a highprio stopper thread
 2582 * and performs thread migration by bumping thread off CPU then
 2583 * 'pushing' onto another runqueue.
 2584 */
 2585static int migration_cpu_stop(void *data)
 2586{
 2587	struct migration_arg *arg = data;
 2588	struct set_affinity_pending *pending = arg->pending;
 2589	struct task_struct *p = arg->task;
 2590	struct rq *rq = this_rq();
 2591	bool complete = false;
 2592	struct rq_flags rf;
 2593
 2594	/*
 2595	 * The original target CPU might have gone down and we might
 2596	 * be on another CPU but it doesn't matter.
 2597	 */
 2598	local_irq_save(rf.flags);
 2599	/*
 2600	 * We need to explicitly wake pending tasks before running
 2601	 * __migrate_task() such that we will not miss enforcing cpus_ptr
 2602	 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
 2603	 */
 2604	flush_smp_call_function_queue();
 2605
 2606	raw_spin_lock(&p->pi_lock);
 2607	rq_lock(rq, &rf);
 2608
 2609	/*
 2610	 * If we were passed a pending, then ->stop_pending was set, thus
 2611	 * p->migration_pending must have remained stable.
 2612	 */
 2613	WARN_ON_ONCE(pending && pending != p->migration_pending);
 2614
 2615	/*
 2616	 * If task_rq(p) != rq, it cannot be migrated here, because we're
 2617	 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
 2618	 * we're holding p->pi_lock.
 2619	 */
 2620	if (task_rq(p) == rq) {
 2621		if (is_migration_disabled(p))
 2622			goto out;
 2623
 2624		if (pending) {
 2625			p->migration_pending = NULL;
 2626			complete = true;
 2627
 2628			if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
 2629				goto out;
 2630		}
 2631
 2632		if (task_on_rq_queued(p)) {
 2633			update_rq_clock(rq);
 2634			rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
 2635		} else {
 2636			p->wake_cpu = arg->dest_cpu;
 2637		}
 2638
 2639		/*
 2640		 * XXX __migrate_task() can fail, at which point we might end
 2641		 * up running on a dodgy CPU, AFAICT this can only happen
 2642		 * during CPU hotplug, at which point we'll get pushed out
 2643		 * anyway, so it's probably not a big deal.
 2644		 */
 2645
 2646	} else if (pending) {
 2647		/*
 2648		 * This happens when we get migrated between migrate_enable()'s
 2649		 * preempt_enable() and scheduling the stopper task. At that
 2650		 * point we're a regular task again and not current anymore.
 2651		 *
 2652		 * A !PREEMPT kernel has a giant hole here, which makes it far
 2653		 * more likely.
 2654		 */
 2655
 2656		/*
 2657		 * The task moved before the stopper got to run. We're holding
 2658		 * ->pi_lock, so the allowed mask is stable - if it got
 2659		 * somewhere allowed, we're done.
 2660		 */
 2661		if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
 2662			p->migration_pending = NULL;
 2663			complete = true;
 2664			goto out;
 2665		}
 2666
 2667		/*
 2668		 * When migrate_enable() hits a rq mis-match we can't reliably
 2669		 * determine is_migration_disabled() and so have to chase after
 2670		 * it.
 2671		 */
 2672		WARN_ON_ONCE(!pending->stop_pending);
 2673		preempt_disable();
 2674		task_rq_unlock(rq, p, &rf);
 2675		stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
 2676				    &pending->arg, &pending->stop_work);
 2677		preempt_enable();
 2678		return 0;
 2679	}
 2680out:
 2681	if (pending)
 2682		pending->stop_pending = false;
 2683	task_rq_unlock(rq, p, &rf);
 2684
 2685	if (complete)
 2686		complete_all(&pending->done);
 2687
 2688	return 0;
 2689}
 2690
 2691int push_cpu_stop(void *arg)
 2692{
 2693	struct rq *lowest_rq = NULL, *rq = this_rq();
 2694	struct task_struct *p = arg;
 2695
 2696	raw_spin_lock_irq(&p->pi_lock);
 2697	raw_spin_rq_lock(rq);
 2698
 2699	if (task_rq(p) != rq)
 2700		goto out_unlock;
 2701
 2702	if (is_migration_disabled(p)) {
 2703		p->migration_flags |= MDF_PUSH;
 2704		goto out_unlock;
 2705	}
 2706
 2707	p->migration_flags &= ~MDF_PUSH;
 2708
 2709	if (p->sched_class->find_lock_rq)
 2710		lowest_rq = p->sched_class->find_lock_rq(p, rq);
 2711
 2712	if (!lowest_rq)
 2713		goto out_unlock;
 2714
 2715	// XXX validate p is still the highest prio task
 2716	if (task_rq(p) == rq) {
 2717		deactivate_task(rq, p, 0);
 2718		set_task_cpu(p, lowest_rq->cpu);
 2719		activate_task(lowest_rq, p, 0);
 2720		resched_curr(lowest_rq);
 2721	}
 2722
 2723	double_unlock_balance(rq, lowest_rq);
 2724
 2725out_unlock:
 2726	rq->push_busy = false;
 2727	raw_spin_rq_unlock(rq);
 2728	raw_spin_unlock_irq(&p->pi_lock);
 2729
 2730	put_task_struct(p);
 2731	return 0;
 2732}
 2733
 2734/*
 2735 * sched_class::set_cpus_allowed must do the below, but is not required to
 2736 * actually call this function.
 2737 */
 2738void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
 2739{
 2740	if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
 2741		p->cpus_ptr = ctx->new_mask;
 2742		return;
 2743	}
 2744
 2745	cpumask_copy(&p->cpus_mask, ctx->new_mask);
 2746	p->nr_cpus_allowed = cpumask_weight(ctx->new_mask);
 2747
 2748	/*
 2749	 * Swap in a new user_cpus_ptr if SCA_USER flag set
 2750	 */
 2751	if (ctx->flags & SCA_USER)
 2752		swap(p->user_cpus_ptr, ctx->user_mask);
 2753}
 2754
 2755static void
 2756__do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
 2757{
 2758	struct rq *rq = task_rq(p);
 2759	bool queued, running;
 2760
 2761	/*
 2762	 * This here violates the locking rules for affinity, since we're only
 2763	 * supposed to change these variables while holding both rq->lock and
 2764	 * p->pi_lock.
 2765	 *
 2766	 * HOWEVER, it magically works, because ttwu() is the only code that
 2767	 * accesses these variables under p->pi_lock and only does so after
 2768	 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
 2769	 * before finish_task().
 2770	 *
 2771	 * XXX do further audits, this smells like something putrid.
 2772	 */
 2773	if (ctx->flags & SCA_MIGRATE_DISABLE)
 2774		SCHED_WARN_ON(!p->on_cpu);
 2775	else
 2776		lockdep_assert_held(&p->pi_lock);
 2777
 2778	queued = task_on_rq_queued(p);
 2779	running = task_current(rq, p);
 2780
 2781	if (queued) {
 2782		/*
 2783		 * Because __kthread_bind() calls this on blocked tasks without
 2784		 * holding rq->lock.
 2785		 */
 2786		lockdep_assert_rq_held(rq);
 2787		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
 2788	}
 2789	if (running)
 2790		put_prev_task(rq, p);
 2791
 2792	p->sched_class->set_cpus_allowed(p, ctx);
 2793
 2794	if (queued)
 2795		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
 2796	if (running)
 2797		set_next_task(rq, p);
 2798}
 2799
 2800/*
 2801 * Used for kthread_bind() and select_fallback_rq(), in both cases the user
 2802 * affinity (if any) should be destroyed too.
 2803 */
 2804void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
 2805{
 2806	struct affinity_context ac = {
 2807		.new_mask  = new_mask,
 2808		.user_mask = NULL,
 2809		.flags     = SCA_USER,	/* clear the user requested mask */
 2810	};
 2811	union cpumask_rcuhead {
 2812		cpumask_t cpumask;
 2813		struct rcu_head rcu;
 2814	};
 2815
 2816	__do_set_cpus_allowed(p, &ac);
 2817
 2818	/*
 2819	 * Because this is called with p->pi_lock held, it is not possible
 2820	 * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
 2821	 * kfree_rcu().
 2822	 */
 2823	kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
 2824}
 2825
 2826static cpumask_t *alloc_user_cpus_ptr(int node)
 2827{
 2828	/*
 2829	 * See do_set_cpus_allowed() above for the rcu_head usage.
 2830	 */
 2831	int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
 2832
 2833	return kmalloc_node(size, GFP_KERNEL, node);
 2834}
 2835
 2836int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
 2837		      int node)
 2838{
 2839	cpumask_t *user_mask;
 2840	unsigned long flags;
 2841
 2842	/*
 2843	 * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
 2844	 * may differ by now due to racing.
 2845	 */
 2846	dst->user_cpus_ptr = NULL;
 2847
 2848	/*
 2849	 * This check is racy and losing the race is a valid situation.
 2850	 * It is not worth the extra overhead of taking the pi_lock on
 2851	 * every fork/clone.
 2852	 */
 2853	if (data_race(!src->user_cpus_ptr))
 2854		return 0;
 2855
 2856	user_mask = alloc_user_cpus_ptr(node);
 2857	if (!user_mask)
 2858		return -ENOMEM;
 2859
 2860	/*
 2861	 * Use pi_lock to protect content of user_cpus_ptr
 2862	 *
 2863	 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
 2864	 * do_set_cpus_allowed().
 2865	 */
 2866	raw_spin_lock_irqsave(&src->pi_lock, flags);
 2867	if (src->user_cpus_ptr) {
 2868		swap(dst->user_cpus_ptr, user_mask);
 2869		cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
 2870	}
 2871	raw_spin_unlock_irqrestore(&src->pi_lock, flags);
 2872
 2873	if (unlikely(user_mask))
 2874		kfree(user_mask);
 2875
 2876	return 0;
 2877}
 2878
 2879static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
 2880{
 2881	struct cpumask *user_mask = NULL;
 2882
 2883	swap(p->user_cpus_ptr, user_mask);
 2884
 2885	return user_mask;
 2886}
 2887
 2888void release_user_cpus_ptr(struct task_struct *p)
 2889{
 2890	kfree(clear_user_cpus_ptr(p));
 2891}
 2892
 2893/*
 2894 * This function is wildly self concurrent; here be dragons.
 2895 *
 2896 *
 2897 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
 2898 * designated task is enqueued on an allowed CPU. If that task is currently
 2899 * running, we have to kick it out using the CPU stopper.
 2900 *
 2901 * Migrate-Disable comes along and tramples all over our nice sandcastle.
 2902 * Consider:
 2903 *
 2904 *     Initial conditions: P0->cpus_mask = [0, 1]
 2905 *
 2906 *     P0@CPU0                  P1
 2907 *
 2908 *     migrate_disable();
 2909 *     <preempted>
 2910 *                              set_cpus_allowed_ptr(P0, [1]);
 2911 *
 2912 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
 2913 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
 2914 * This means we need the following scheme:
 2915 *
 2916 *     P0@CPU0                  P1
 2917 *
 2918 *     migrate_disable();
 2919 *     <preempted>
 2920 *                              set_cpus_allowed_ptr(P0, [1]);
 2921 *                                <blocks>
 2922 *     <resumes>
 2923 *     migrate_enable();
 2924 *       __set_cpus_allowed_ptr();
 2925 *       <wakes local stopper>
 2926 *                         `--> <woken on migration completion>
 2927 *
 2928 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
 2929 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
 2930 * task p are serialized by p->pi_lock, which we can leverage: the one that
 2931 * should come into effect at the end of the Migrate-Disable region is the last
 2932 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
 2933 * but we still need to properly signal those waiting tasks at the appropriate
 2934 * moment.
 2935 *
 2936 * This is implemented using struct set_affinity_pending. The first
 2937 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
 2938 * setup an instance of that struct and install it on the targeted task_struct.
 2939 * Any and all further callers will reuse that instance. Those then wait for
 2940 * a completion signaled at the tail of the CPU stopper callback (1), triggered
 2941 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
 2942 *
 2943 *
 2944 * (1) In the cases covered above. There is one more where the completion is
 2945 * signaled within affine_move_task() itself: when a subsequent affinity request
 2946 * occurs after the stopper bailed out due to the targeted task still being
 2947 * Migrate-Disable. Consider:
 2948 *
 2949 *     Initial conditions: P0->cpus_mask = [0, 1]
 2950 *
 2951 *     CPU0		  P1				P2
 2952 *     <P0>
 2953 *       migrate_disable();
 2954 *       <preempted>
 2955 *                        set_cpus_allowed_ptr(P0, [1]);
 2956 *                          <blocks>
 2957 *     <migration/0>
 2958 *       migration_cpu_stop()
 2959 *         is_migration_disabled()
 2960 *           <bails>
 2961 *                                                       set_cpus_allowed_ptr(P0, [0, 1]);
 2962 *                                                         <signal completion>
 2963 *                          <awakes>
 2964 *
 2965 * Note that the above is safe vs a concurrent migrate_enable(), as any
 2966 * pending affinity completion is preceded by an uninstallation of
 2967 * p->migration_pending done with p->pi_lock held.
 2968 */
 2969static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
 2970			    int dest_cpu, unsigned int flags)
 2971	__releases(rq->lock)
 2972	__releases(p->pi_lock)
 2973{
 2974	struct set_affinity_pending my_pending = { }, *pending = NULL;
 2975	bool stop_pending, complete = false;
 2976
 2977	/* Can the task run on the task's current CPU? If so, we're done */
 2978	if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
 2979		struct task_struct *push_task = NULL;
 2980
 2981		if ((flags & SCA_MIGRATE_ENABLE) &&
 2982		    (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
 2983			rq->push_busy = true;
 2984			push_task = get_task_struct(p);
 2985		}
 2986
 2987		/*
 2988		 * If there are pending waiters, but no pending stop_work,
 2989		 * then complete now.
 2990		 */
 2991		pending = p->migration_pending;
 2992		if (pending && !pending->stop_pending) {
 2993			p->migration_pending = NULL;
 2994			complete = true;
 2995		}
 2996
 2997		preempt_disable();
 2998		task_rq_unlock(rq, p, rf);
 2999		if (push_task) {
 3000			stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
 3001					    p, &rq->push_work);
 3002		}
 3003		preempt_enable();
 3004
 3005		if (complete)
 3006			complete_all(&pending->done);
 3007
 3008		return 0;
 3009	}
 3010
 3011	if (!(flags & SCA_MIGRATE_ENABLE)) {
 3012		/* serialized by p->pi_lock */
 3013		if (!p->migration_pending) {
 3014			/* Install the request */
 3015			refcount_set(&my_pending.refs, 1);
 3016			init_completion(&my_pending.done);
 3017			my_pending.arg = (struct migration_arg) {
 3018				.task = p,
 3019				.dest_cpu = dest_cpu,
 3020				.pending = &my_pending,
 3021			};
 3022
 3023			p->migration_pending = &my_pending;
 3024		} else {
 3025			pending = p->migration_pending;
 3026			refcount_inc(&pending->refs);
 3027			/*
 3028			 * Affinity has changed, but we've already installed a
 3029			 * pending. migration_cpu_stop() *must* see this, else
 3030			 * we risk a completion of the pending despite having a
 3031			 * task on a disallowed CPU.
 3032			 *
 3033			 * Serialized by p->pi_lock, so this is safe.
 3034			 */
 3035			pending->arg.dest_cpu = dest_cpu;
 3036		}
 3037	}
 3038	pending = p->migration_pending;
 3039	/*
 3040	 * - !MIGRATE_ENABLE:
 3041	 *   we'll have installed a pending if there wasn't one already.
 3042	 *
 3043	 * - MIGRATE_ENABLE:
 3044	 *   we're here because the current CPU isn't matching anymore,
 3045	 *   the only way that can happen is because of a concurrent
 3046	 *   set_cpus_allowed_ptr() call, which should then still be
 3047	 *   pending completion.
 3048	 *
 3049	 * Either way, we really should have a @pending here.
 3050	 */
 3051	if (WARN_ON_ONCE(!pending)) {
 3052		task_rq_unlock(rq, p, rf);
 3053		return -EINVAL;
 3054	}
 3055
 3056	if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
 3057		/*
 3058		 * MIGRATE_ENABLE gets here because 'p == current', but for
 3059		 * anything else we cannot do is_migration_disabled(), punt
 3060		 * and have the stopper function handle it all race-free.
 3061		 */
 3062		stop_pending = pending->stop_pending;
 3063		if (!stop_pending)
 3064			pending->stop_pending = true;
 3065
 3066		if (flags & SCA_MIGRATE_ENABLE)
 3067			p->migration_flags &= ~MDF_PUSH;
 3068
 3069		preempt_disable();
 3070		task_rq_unlock(rq, p, rf);
 3071		if (!stop_pending) {
 3072			stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
 3073					    &pending->arg, &pending->stop_work);
 3074		}
 3075		preempt_enable();
 3076
 3077		if (flags & SCA_MIGRATE_ENABLE)
 3078			return 0;
 3079	} else {
 3080
 3081		if (!is_migration_disabled(p)) {
 3082			if (task_on_rq_queued(p))
 3083				rq = move_queued_task(rq, rf, p, dest_cpu);
 3084
 3085			if (!pending->stop_pending) {
 3086				p->migration_pending = NULL;
 3087				complete = true;
 3088			}
 3089		}
 3090		task_rq_unlock(rq, p, rf);
 3091
 3092		if (complete)
 3093			complete_all(&pending->done);
 3094	}
 3095
 3096	wait_for_completion(&pending->done);
 3097
 3098	if (refcount_dec_and_test(&pending->refs))
 3099		wake_up_var(&pending->refs); /* No UaF, just an address */
 3100
 3101	/*
 3102	 * Block the original owner of &pending until all subsequent callers
 3103	 * have seen the completion and decremented the refcount
 3104	 */
 3105	wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
 3106
 3107	/* ARGH */
 3108	WARN_ON_ONCE(my_pending.stop_pending);
 3109
 3110	return 0;
 3111}
 3112
 3113/*
 3114 * Called with both p->pi_lock and rq->lock held; drops both before returning.
 3115 */
 3116static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
 3117					 struct affinity_context *ctx,
 3118					 struct rq *rq,
 3119					 struct rq_flags *rf)
 3120	__releases(rq->lock)
 3121	__releases(p->pi_lock)
 3122{
 3123	const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
 3124	const struct cpumask *cpu_valid_mask = cpu_active_mask;
 3125	bool kthread = p->flags & PF_KTHREAD;
 3126	unsigned int dest_cpu;
 
 
 3127	int ret = 0;
 3128
 
 3129	update_rq_clock(rq);
 3130
 3131	if (kthread || is_migration_disabled(p)) {
 3132		/*
 3133		 * Kernel threads are allowed on online && !active CPUs,
 3134		 * however, during cpu-hot-unplug, even these might get pushed
 3135		 * away if not KTHREAD_IS_PER_CPU.
 3136		 *
 3137		 * Specifically, migration_disabled() tasks must not fail the
 3138		 * cpumask_any_and_distribute() pick below, esp. so on
 3139		 * SCA_MIGRATE_ENABLE, otherwise we'll not call
 3140		 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
 3141		 */
 3142		cpu_valid_mask = cpu_online_mask;
 3143	}
 3144
 3145	if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) {
 3146		ret = -EINVAL;
 3147		goto out;
 3148	}
 3149
 3150	/*
 3151	 * Must re-check here, to close a race against __kthread_bind(),
 3152	 * sched_setaffinity() is not guaranteed to observe the flag.
 3153	 */
 3154	if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
 3155		ret = -EINVAL;
 3156		goto out;
 3157	}
 3158
 3159	if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
 3160		if (cpumask_equal(&p->cpus_mask, ctx->new_mask)) {
 3161			if (ctx->flags & SCA_USER)
 3162				swap(p->user_cpus_ptr, ctx->user_mask);
 3163			goto out;
 3164		}
 3165
 3166		if (WARN_ON_ONCE(p == current &&
 3167				 is_migration_disabled(p) &&
 3168				 !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
 3169			ret = -EBUSY;
 3170			goto out;
 3171		}
 3172	}
 3173
 3174	/*
 3175	 * Picking a ~random cpu helps in cases where we are changing affinity
 3176	 * for groups of tasks (ie. cpuset), so that load balancing is not
 3177	 * immediately required to distribute the tasks within their new mask.
 3178	 */
 3179	dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
 3180	if (dest_cpu >= nr_cpu_ids) {
 3181		ret = -EINVAL;
 3182		goto out;
 3183	}
 3184
 3185	__do_set_cpus_allowed(p, ctx);
 3186
 3187	return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
 
 
 
 
 
 
 
 
 
 
 
 
 3188
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 3189out:
 3190	task_rq_unlock(rq, p, rf);
 3191
 3192	return ret;
 3193}
 3194
 3195/*
 3196 * Change a given task's CPU affinity. Migrate the thread to a
 3197 * proper CPU and schedule it away if the CPU it's executing on
 3198 * is removed from the allowed bitmask.
 3199 *
 3200 * NOTE: the caller must have a valid reference to the task, the
 3201 * task must not exit() & deallocate itself prematurely. The
 3202 * call is not atomic; no spinlocks may be held.
 3203 */
 3204static int __set_cpus_allowed_ptr(struct task_struct *p,
 3205				  struct affinity_context *ctx)
 3206{
 3207	struct rq_flags rf;
 3208	struct rq *rq;
 3209
 3210	rq = task_rq_lock(p, &rf);
 3211	/*
 3212	 * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
 3213	 * flags are set.
 3214	 */
 3215	if (p->user_cpus_ptr &&
 3216	    !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
 3217	    cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr))
 3218		ctx->new_mask = rq->scratch_mask;
 3219
 3220	return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf);
 3221}
 3222
 3223int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
 3224{
 3225	struct affinity_context ac = {
 3226		.new_mask  = new_mask,
 3227		.flags     = 0,
 3228	};
 3229
 3230	return __set_cpus_allowed_ptr(p, &ac);
 3231}
 3232EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
 3233
 3234/*
 3235 * Change a given task's CPU affinity to the intersection of its current
 3236 * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
 3237 * If user_cpus_ptr is defined, use it as the basis for restricting CPU
 3238 * affinity or use cpu_online_mask instead.
 3239 *
 3240 * If the resulting mask is empty, leave the affinity unchanged and return
 3241 * -EINVAL.
 3242 */
 3243static int restrict_cpus_allowed_ptr(struct task_struct *p,
 3244				     struct cpumask *new_mask,
 3245				     const struct cpumask *subset_mask)
 3246{
 3247	struct affinity_context ac = {
 3248		.new_mask  = new_mask,
 3249		.flags     = 0,
 3250	};
 3251	struct rq_flags rf;
 3252	struct rq *rq;
 3253	int err;
 3254
 3255	rq = task_rq_lock(p, &rf);
 3256
 3257	/*
 3258	 * Forcefully restricting the affinity of a deadline task is
 3259	 * likely to cause problems, so fail and noisily override the
 3260	 * mask entirely.
 3261	 */
 3262	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
 3263		err = -EPERM;
 3264		goto err_unlock;
 3265	}
 3266
 3267	if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) {
 3268		err = -EINVAL;
 3269		goto err_unlock;
 3270	}
 3271
 3272	return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf);
 3273
 3274err_unlock:
 3275	task_rq_unlock(rq, p, &rf);
 3276	return err;
 3277}
 3278
 3279/*
 3280 * Restrict the CPU affinity of task @p so that it is a subset of
 3281 * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
 3282 * old affinity mask. If the resulting mask is empty, we warn and walk
 3283 * up the cpuset hierarchy until we find a suitable mask.
 3284 */
 3285void force_compatible_cpus_allowed_ptr(struct task_struct *p)
 3286{
 3287	cpumask_var_t new_mask;
 3288	const struct cpumask *override_mask = task_cpu_possible_mask(p);
 3289
 3290	alloc_cpumask_var(&new_mask, GFP_KERNEL);
 3291
 3292	/*
 3293	 * __migrate_task() can fail silently in the face of concurrent
 3294	 * offlining of the chosen destination CPU, so take the hotplug
 3295	 * lock to ensure that the migration succeeds.
 3296	 */
 3297	cpus_read_lock();
 3298	if (!cpumask_available(new_mask))
 3299		goto out_set_mask;
 3300
 3301	if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
 3302		goto out_free_mask;
 3303
 3304	/*
 3305	 * We failed to find a valid subset of the affinity mask for the
 3306	 * task, so override it based on its cpuset hierarchy.
 3307	 */
 3308	cpuset_cpus_allowed(p, new_mask);
 3309	override_mask = new_mask;
 3310
 3311out_set_mask:
 3312	if (printk_ratelimit()) {
 3313		printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
 3314				task_pid_nr(p), p->comm,
 3315				cpumask_pr_args(override_mask));
 3316	}
 3317
 3318	WARN_ON(set_cpus_allowed_ptr(p, override_mask));
 3319out_free_mask:
 3320	cpus_read_unlock();
 3321	free_cpumask_var(new_mask);
 3322}
 3323
 3324static int
 3325__sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
 3326
 3327/*
 3328 * Restore the affinity of a task @p which was previously restricted by a
 3329 * call to force_compatible_cpus_allowed_ptr().
 3330 *
 3331 * It is the caller's responsibility to serialise this with any calls to
 3332 * force_compatible_cpus_allowed_ptr(@p).
 3333 */
 3334void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
 3335{
 3336	struct affinity_context ac = {
 3337		.new_mask  = task_user_cpus(p),
 3338		.flags     = 0,
 3339	};
 3340	int ret;
 3341
 3342	/*
 3343	 * Try to restore the old affinity mask with __sched_setaffinity().
 3344	 * Cpuset masking will be done there too.
 3345	 */
 3346	ret = __sched_setaffinity(p, &ac);
 3347	WARN_ON_ONCE(ret);
 3348}
 3349
 3350void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
 3351{
 3352#ifdef CONFIG_SCHED_DEBUG
 3353	unsigned int state = READ_ONCE(p->__state);
 3354
 3355	/*
 3356	 * We should never call set_task_cpu() on a blocked task,
 3357	 * ttwu() will sort out the placement.
 3358	 */
 3359	WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
 
 3360
 3361	/*
 3362	 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
 3363	 * because schedstat_wait_{start,end} rebase migrating task's wait_start
 3364	 * time relying on p->on_rq.
 3365	 */
 3366	WARN_ON_ONCE(state == TASK_RUNNING &&
 3367		     p->sched_class == &fair_sched_class &&
 3368		     (p->on_rq && !task_on_rq_migrating(p)));
 3369
 3370#ifdef CONFIG_LOCKDEP
 3371	/*
 3372	 * The caller should hold either p->pi_lock or rq->lock, when changing
 3373	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
 3374	 *
 3375	 * sched_move_task() holds both and thus holding either pins the cgroup,
 3376	 * see task_group().
 3377	 *
 3378	 * Furthermore, all task_rq users should acquire both locks, see
 3379	 * task_rq_lock().
 3380	 */
 3381	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
 3382				      lockdep_is_held(__rq_lockp(task_rq(p)))));
 3383#endif
 3384	/*
 3385	 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
 3386	 */
 3387	WARN_ON_ONCE(!cpu_online(new_cpu));
 3388
 3389	WARN_ON_ONCE(is_migration_disabled(p));
 3390#endif
 3391
 3392	trace_sched_migrate_task(p, new_cpu);
 3393
 3394	if (task_cpu(p) != new_cpu) {
 3395		if (p->sched_class->migrate_task_rq)
 3396			p->sched_class->migrate_task_rq(p, new_cpu);
 3397		p->se.nr_migrations++;
 3398		rseq_migrate(p);
 3399		sched_mm_cid_migrate_from(p);
 3400		perf_event_task_migrate(p);
 3401	}
 3402
 3403	__set_task_cpu(p, new_cpu);
 3404}
 3405
 3406#ifdef CONFIG_NUMA_BALANCING
 3407static void __migrate_swap_task(struct task_struct *p, int cpu)
 3408{
 3409	if (task_on_rq_queued(p)) {
 3410		struct rq *src_rq, *dst_rq;
 3411		struct rq_flags srf, drf;
 3412
 3413		src_rq = task_rq(p);
 3414		dst_rq = cpu_rq(cpu);
 3415
 3416		rq_pin_lock(src_rq, &srf);
 3417		rq_pin_lock(dst_rq, &drf);
 3418
 
 3419		deactivate_task(src_rq, p, 0);
 3420		set_task_cpu(p, cpu);
 3421		activate_task(dst_rq, p, 0);
 3422		wakeup_preempt(dst_rq, p, 0);
 
 3423
 3424		rq_unpin_lock(dst_rq, &drf);
 3425		rq_unpin_lock(src_rq, &srf);
 3426
 3427	} else {
 3428		/*
 3429		 * Task isn't running anymore; make it appear like we migrated
 3430		 * it before it went to sleep. This means on wakeup we make the
 3431		 * previous CPU our target instead of where it really is.
 3432		 */
 3433		p->wake_cpu = cpu;
 3434	}
 3435}
 3436
 3437struct migration_swap_arg {
 3438	struct task_struct *src_task, *dst_task;
 3439	int src_cpu, dst_cpu;
 3440};
 3441
 3442static int migrate_swap_stop(void *data)
 3443{
 3444	struct migration_swap_arg *arg = data;
 3445	struct rq *src_rq, *dst_rq;
 
 3446
 3447	if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
 3448		return -EAGAIN;
 3449
 3450	src_rq = cpu_rq(arg->src_cpu);
 3451	dst_rq = cpu_rq(arg->dst_cpu);
 3452
 3453	guard(double_raw_spinlock)(&arg->src_task->pi_lock, &arg->dst_task->pi_lock);
 3454	guard(double_rq_lock)(src_rq, dst_rq);
 
 3455
 3456	if (task_cpu(arg->dst_task) != arg->dst_cpu)
 3457		return -EAGAIN;
 3458
 3459	if (task_cpu(arg->src_task) != arg->src_cpu)
 3460		return -EAGAIN;
 3461
 3462	if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
 3463		return -EAGAIN;
 3464
 3465	if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
 3466		return -EAGAIN;
 3467
 3468	__migrate_swap_task(arg->src_task, arg->dst_cpu);
 3469	__migrate_swap_task(arg->dst_task, arg->src_cpu);
 3470
 3471	return 0;
 
 
 
 
 
 
 
 3472}
 3473
 3474/*
 3475 * Cross migrate two tasks
 3476 */
 3477int migrate_swap(struct task_struct *cur, struct task_struct *p,
 3478		int target_cpu, int curr_cpu)
 3479{
 3480	struct migration_swap_arg arg;
 3481	int ret = -EINVAL;
 3482
 3483	arg = (struct migration_swap_arg){
 3484		.src_task = cur,
 3485		.src_cpu = curr_cpu,
 3486		.dst_task = p,
 3487		.dst_cpu = target_cpu,
 3488	};
 3489
 3490	if (arg.src_cpu == arg.dst_cpu)
 3491		goto out;
 3492
 3493	/*
 3494	 * These three tests are all lockless; this is OK since all of them
 3495	 * will be re-checked with proper locks held further down the line.
 3496	 */
 3497	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
 3498		goto out;
 3499
 3500	if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
 3501		goto out;
 3502
 3503	if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
 3504		goto out;
 3505
 3506	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
 3507	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
 3508
 3509out:
 3510	return ret;
 3511}
 3512#endif /* CONFIG_NUMA_BALANCING */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 3513
 3514/***
 3515 * kick_process - kick a running thread to enter/exit the kernel
 3516 * @p: the to-be-kicked thread
 3517 *
 3518 * Cause a process which is running on another CPU to enter
 3519 * kernel-mode, without any delay. (to get signals handled.)
 3520 *
 3521 * NOTE: this function doesn't have to take the runqueue lock,
 3522 * because all it wants to ensure is that the remote task enters
 3523 * the kernel. If the IPI races and the task has been migrated
 3524 * to another CPU then no harm is done and the purpose has been
 3525 * achieved as well.
 3526 */
 3527void kick_process(struct task_struct *p)
 3528{
 3529	guard(preempt)();
 3530	int cpu = task_cpu(p);
 3531
 
 
 3532	if ((cpu != smp_processor_id()) && task_curr(p))
 3533		smp_send_reschedule(cpu);
 
 3534}
 3535EXPORT_SYMBOL_GPL(kick_process);
 3536
 3537/*
 3538 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
 3539 *
 3540 * A few notes on cpu_active vs cpu_online:
 3541 *
 3542 *  - cpu_active must be a subset of cpu_online
 3543 *
 3544 *  - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
 3545 *    see __set_cpus_allowed_ptr(). At this point the newly online
 3546 *    CPU isn't yet part of the sched domains, and balancing will not
 3547 *    see it.
 3548 *
 3549 *  - on CPU-down we clear cpu_active() to mask the sched domains and
 3550 *    avoid the load balancer to place new tasks on the to be removed
 3551 *    CPU. Existing tasks will remain running there and will be taken
 3552 *    off.
 3553 *
 3554 * This means that fallback selection must not select !active CPUs.
 3555 * And can assume that any active CPU must be online. Conversely
 3556 * select_task_rq() below may allow selection of !active CPUs in order
 3557 * to satisfy the above rules.
 3558 */
 3559static int select_fallback_rq(int cpu, struct task_struct *p)
 3560{
 3561	int nid = cpu_to_node(cpu);
 3562	const struct cpumask *nodemask = NULL;
 3563	enum { cpuset, possible, fail } state = cpuset;
 3564	int dest_cpu;
 3565
 3566	/*
 3567	 * If the node that the CPU is on has been offlined, cpu_to_node()
 3568	 * will return -1. There is no CPU on the node, and we should
 3569	 * select the CPU on the other node.
 3570	 */
 3571	if (nid != -1) {
 3572		nodemask = cpumask_of_node(nid);
 3573
 3574		/* Look for allowed, online CPU in same node. */
 3575		for_each_cpu(dest_cpu, nodemask) {
 3576			if (is_cpu_allowed(p, dest_cpu))
 
 
 3577				return dest_cpu;
 3578		}
 3579	}
 3580
 3581	for (;;) {
 3582		/* Any allowed, online CPU? */
 3583		for_each_cpu(dest_cpu, p->cpus_ptr) {
 3584			if (!is_cpu_allowed(p, dest_cpu))
 3585				continue;
 3586
 3587			goto out;
 3588		}
 3589
 3590		/* No more Mr. Nice Guy. */
 3591		switch (state) {
 3592		case cpuset:
 3593			if (cpuset_cpus_allowed_fallback(p)) {
 
 3594				state = possible;
 3595				break;
 3596			}
 3597			fallthrough;
 3598		case possible:
 3599			/*
 3600			 * XXX When called from select_task_rq() we only
 3601			 * hold p->pi_lock and again violate locking order.
 3602			 *
 3603			 * More yuck to audit.
 3604			 */
 3605			do_set_cpus_allowed(p, task_cpu_possible_mask(p));
 3606			state = fail;
 3607			break;
 
 3608		case fail:
 3609			BUG();
 3610			break;
 3611		}
 3612	}
 3613
 3614out:
 3615	if (state != cpuset) {
 3616		/*
 3617		 * Don't tell them about moving exiting tasks or
 3618		 * kernel threads (both mm NULL), since they never
 3619		 * leave kernel.
 3620		 */
 3621		if (p->mm && printk_ratelimit()) {
 3622			printk_deferred("process %d (%s) no longer affine to cpu%d\n",
 3623					task_pid_nr(p), p->comm, cpu);
 3624		}
 3625	}
 3626
 3627	return dest_cpu;
 3628}
 3629
 3630/*
 3631 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
 3632 */
 3633static inline
 3634int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
 3635{
 3636	lockdep_assert_held(&p->pi_lock);
 3637
 3638	if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
 3639		cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
 3640	else
 3641		cpu = cpumask_any(p->cpus_ptr);
 3642
 3643	/*
 3644	 * In order not to call set_task_cpu() on a blocking task we need
 3645	 * to rely on ttwu() to place the task on a valid ->cpus_ptr
 3646	 * CPU.
 3647	 *
 3648	 * Since this is common to all placement strategies, this lives here.
 3649	 *
 3650	 * [ this allows ->select_task() to simply return task_cpu(p) and
 3651	 *   not worry about this generic constraint ]
 3652	 */
 3653	if (unlikely(!is_cpu_allowed(p, cpu)))
 3654		cpu = select_fallback_rq(task_cpu(p), p);
 3655
 3656	return cpu;
 3657}
 3658
 
 
 
 
 
 
 3659void sched_set_stop_task(int cpu, struct task_struct *stop)
 3660{
 3661	static struct lock_class_key stop_pi_lock;
 3662	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
 3663	struct task_struct *old_stop = cpu_rq(cpu)->stop;
 3664
 3665	if (stop) {
 3666		/*
 3667		 * Make it appear like a SCHED_FIFO task, its something
 3668		 * userspace knows about and won't get confused about.
 3669		 *
 3670		 * Also, it will make PI more or less work without too
 3671		 * much confusion -- but then, stop work should not
 3672		 * rely on PI working anyway.
 3673		 */
 3674		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
 3675
 3676		stop->sched_class = &stop_sched_class;
 3677
 3678		/*
 3679		 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
 3680		 * adjust the effective priority of a task. As a result,
 3681		 * rt_mutex_setprio() can trigger (RT) balancing operations,
 3682		 * which can then trigger wakeups of the stop thread to push
 3683		 * around the current task.
 3684		 *
 3685		 * The stop task itself will never be part of the PI-chain, it
 3686		 * never blocks, therefore that ->pi_lock recursion is safe.
 3687		 * Tell lockdep about this by placing the stop->pi_lock in its
 3688		 * own class.
 3689		 */
 3690		lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
 3691	}
 3692
 3693	cpu_rq(cpu)->stop = stop;
 3694
 3695	if (old_stop) {
 3696		/*
 3697		 * Reset it back to a normal scheduling class so that
 3698		 * it can die in pieces.
 3699		 */
 3700		old_stop->sched_class = &rt_sched_class;
 3701	}
 3702}
 3703
 3704#else /* CONFIG_SMP */
 3705
 3706static inline int __set_cpus_allowed_ptr(struct task_struct *p,
 3707					 struct affinity_context *ctx)
 3708{
 3709	return set_cpus_allowed_ptr(p, ctx->new_mask);
 3710}
 3711
 3712static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
 3713
 3714static inline bool rq_has_pinned_tasks(struct rq *rq)
 3715{
 3716	return false;
 3717}
 3718
 3719static inline cpumask_t *alloc_user_cpus_ptr(int node)
 3720{
 3721	return NULL;
 3722}
 3723
 3724#endif /* !CONFIG_SMP */
 3725
 3726static void
 3727ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
 3728{
 3729	struct rq *rq;
 3730
 3731	if (!schedstat_enabled())
 3732		return;
 3733
 3734	rq = this_rq();
 3735
 3736#ifdef CONFIG_SMP
 3737	if (cpu == rq->cpu) {
 3738		__schedstat_inc(rq->ttwu_local);
 3739		__schedstat_inc(p->stats.nr_wakeups_local);
 3740	} else {
 3741		struct sched_domain *sd;
 3742
 3743		__schedstat_inc(p->stats.nr_wakeups_remote);
 3744
 3745		guard(rcu)();
 3746		for_each_domain(rq->cpu, sd) {
 3747			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
 3748				__schedstat_inc(sd->ttwu_wake_remote);
 3749				break;
 3750			}
 3751		}
 
 3752	}
 3753
 3754	if (wake_flags & WF_MIGRATED)
 3755		__schedstat_inc(p->stats.nr_wakeups_migrate);
 3756#endif /* CONFIG_SMP */
 3757
 3758	__schedstat_inc(rq->ttwu_count);
 3759	__schedstat_inc(p->stats.nr_wakeups);
 3760
 3761	if (wake_flags & WF_SYNC)
 3762		__schedstat_inc(p->stats.nr_wakeups_sync);
 
 
 
 
 
 
 
 
 
 
 3763}
 3764
 3765/*
 3766 * Mark the task runnable.
 3767 */
 3768static inline void ttwu_do_wakeup(struct task_struct *p)
 
 3769{
 3770	WRITE_ONCE(p->__state, TASK_RUNNING);
 
 3771	trace_sched_wakeup(p);
 3772}
 3773
 3774static void
 3775ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
 3776		 struct rq_flags *rf)
 3777{
 3778	int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
 3779
 3780	lockdep_assert_rq_held(rq);
 3781
 3782	if (p->sched_contributes_to_load)
 3783		rq->nr_uninterruptible--;
 3784
 3785#ifdef CONFIG_SMP
 3786	if (wake_flags & WF_MIGRATED)
 3787		en_flags |= ENQUEUE_MIGRATED;
 3788	else
 3789#endif
 3790	if (p->in_iowait) {
 3791		delayacct_blkio_end(p);
 3792		atomic_dec(&task_rq(p)->nr_iowait);
 3793	}
 3794
 3795	activate_task(rq, p, en_flags);
 3796	wakeup_preempt(rq, p, wake_flags);
 3797
 3798	ttwu_do_wakeup(p);
 3799
 3800#ifdef CONFIG_SMP
 3801	if (p->sched_class->task_woken) {
 3802		/*
 3803		 * Our task @p is fully woken up and running; so it's safe to
 3804		 * drop the rq->lock, hereafter rq is only used for statistics.
 3805		 */
 3806		rq_unpin_lock(rq, rf);
 3807		p->sched_class->task_woken(rq, p);
 3808		rq_repin_lock(rq, rf);
 3809	}
 3810
 3811	if (rq->idle_stamp) {
 3812		u64 delta = rq_clock(rq) - rq->idle_stamp;
 3813		u64 max = 2*rq->max_idle_balance_cost;
 3814
 3815		update_avg(&rq->avg_idle, delta);
 3816
 3817		if (rq->avg_idle > max)
 3818			rq->avg_idle = max;
 3819
 3820		rq->idle_stamp = 0;
 3821	}
 3822#endif
 
 
 
 
 
 
 
 3823
 3824	p->dl_server = NULL;
 
 
 
 
 
 
 
 
 
 
 
 3825}
 3826
 3827/*
 3828 * Consider @p being inside a wait loop:
 3829 *
 3830 *   for (;;) {
 3831 *      set_current_state(TASK_UNINTERRUPTIBLE);
 3832 *
 3833 *      if (CONDITION)
 3834 *         break;
 3835 *
 3836 *      schedule();
 3837 *   }
 3838 *   __set_current_state(TASK_RUNNING);
 3839 *
 3840 * between set_current_state() and schedule(). In this case @p is still
 3841 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
 3842 * an atomic manner.
 3843 *
 3844 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
 3845 * then schedule() must still happen and p->state can be changed to
 3846 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
 3847 * need to do a full wakeup with enqueue.
 3848 *
 3849 * Returns: %true when the wakeup is done,
 3850 *          %false otherwise.
 3851 */
 3852static int ttwu_runnable(struct task_struct *p, int wake_flags)
 3853{
 3854	struct rq_flags rf;
 3855	struct rq *rq;
 3856	int ret = 0;
 3857
 3858	rq = __task_rq_lock(p, &rf);
 3859	if (task_on_rq_queued(p)) {
 3860		if (!task_on_cpu(rq, p)) {
 3861			/*
 3862			 * When on_rq && !on_cpu the task is preempted, see if
 3863			 * it should preempt the task that is current now.
 3864			 */
 3865			update_rq_clock(rq);
 3866			wakeup_preempt(rq, p, wake_flags);
 3867		}
 3868		ttwu_do_wakeup(p);
 3869		ret = 1;
 3870	}
 3871	__task_rq_unlock(rq, &rf);
 3872
 3873	return ret;
 3874}
 3875
 3876#ifdef CONFIG_SMP
 3877void sched_ttwu_pending(void *arg)
 3878{
 3879	struct llist_node *llist = arg;
 3880	struct rq *rq = this_rq();
 
 3881	struct task_struct *p, *t;
 3882	struct rq_flags rf;
 3883
 3884	if (!llist)
 3885		return;
 3886
 3887	rq_lock_irqsave(rq, &rf);
 3888	update_rq_clock(rq);
 3889
 3890	llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
 3891		if (WARN_ON_ONCE(p->on_cpu))
 3892			smp_cond_load_acquire(&p->on_cpu, !VAL);
 
 
 3893
 3894		if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
 3895			set_task_cpu(p, cpu_of(rq));
 
 
 
 
 
 
 3896
 3897		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
 3898	}
 3899
 3900	/*
 3901	 * Must be after enqueueing at least once task such that
 3902	 * idle_cpu() does not observe a false-negative -- if it does,
 3903	 * it is possible for select_idle_siblings() to stack a number
 3904	 * of tasks on this CPU during that window.
 3905	 *
 3906	 * It is ok to clear ttwu_pending when another task pending.
 3907	 * We will receive IPI after local irq enabled and then enqueue it.
 3908	 * Since now nr_running > 0, idle_cpu() will always get correct result.
 
 
 
 3909	 */
 3910	WRITE_ONCE(rq->ttwu_pending, 0);
 3911	rq_unlock_irqrestore(rq, &rf);
 3912}
 3913
 3914/*
 3915 * Prepare the scene for sending an IPI for a remote smp_call
 3916 *
 3917 * Returns true if the caller can proceed with sending the IPI.
 3918 * Returns false otherwise.
 3919 */
 3920bool call_function_single_prep_ipi(int cpu)
 3921{
 3922	if (set_nr_if_polling(cpu_rq(cpu)->idle)) {
 3923		trace_sched_wake_idle_without_ipi(cpu);
 3924		return false;
 3925	}
 3926
 3927	return true;
 3928}
 3929
 3930/*
 3931 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
 3932 * necessary. The wakee CPU on receipt of the IPI will queue the task
 3933 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
 3934 * of the wakeup instead of the waker.
 3935 */
 3936static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
 3937{
 3938	struct rq *rq = cpu_rq(cpu);
 3939
 3940	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
 3941
 3942	WRITE_ONCE(rq->ttwu_pending, 1);
 3943	__smp_call_single_queue(cpu, &p->wake_entry.llist);
 
 
 
 
 3944}
 3945
 3946void wake_up_if_idle(int cpu)
 3947{
 3948	struct rq *rq = cpu_rq(cpu);
 
 3949
 3950	guard(rcu)();
 3951	if (is_idle_task(rcu_dereference(rq->curr))) {
 3952		guard(rq_lock_irqsave)(rq);
 
 
 
 
 
 
 3953		if (is_idle_task(rq->curr))
 3954			resched_curr(rq);
 
 
 3955	}
 
 
 
 3956}
 3957
 3958bool cpus_share_cache(int this_cpu, int that_cpu)
 3959{
 3960	if (this_cpu == that_cpu)
 3961		return true;
 3962
 3963	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
 3964}
 3965
 3966/*
 3967 * Whether CPUs are share cache resources, which means LLC on non-cluster
 3968 * machines and LLC tag or L2 on machines with clusters.
 3969 */
 3970bool cpus_share_resources(int this_cpu, int that_cpu)
 3971{
 3972	if (this_cpu == that_cpu)
 3973		return true;
 3974
 3975	return per_cpu(sd_share_id, this_cpu) == per_cpu(sd_share_id, that_cpu);
 3976}
 3977
 3978static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
 3979{
 3980	/*
 3981	 * Do not complicate things with the async wake_list while the CPU is
 3982	 * in hotplug state.
 3983	 */
 3984	if (!cpu_active(cpu))
 3985		return false;
 3986
 3987	/* Ensure the task will still be allowed to run on the CPU. */
 3988	if (!cpumask_test_cpu(cpu, p->cpus_ptr))
 3989		return false;
 3990
 3991	/*
 3992	 * If the CPU does not share cache, then queue the task on the
 3993	 * remote rqs wakelist to avoid accessing remote data.
 3994	 */
 3995	if (!cpus_share_cache(smp_processor_id(), cpu))
 3996		return true;
 3997
 3998	if (cpu == smp_processor_id())
 3999		return false;
 4000
 4001	/*
 4002	 * If the wakee cpu is idle, or the task is descheduling and the
 4003	 * only running task on the CPU, then use the wakelist to offload
 4004	 * the task activation to the idle (or soon-to-be-idle) CPU as
 4005	 * the current CPU is likely busy. nr_running is checked to
 4006	 * avoid unnecessary task stacking.
 4007	 *
 4008	 * Note that we can only get here with (wakee) p->on_rq=0,
 4009	 * p->on_cpu can be whatever, we've done the dequeue, so
 4010	 * the wakee has been accounted out of ->nr_running.
 4011	 */
 4012	if (!cpu_rq(cpu)->nr_running)
 4013		return true;
 4014
 4015	return false;
 4016}
 4017
 4018static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
 4019{
 4020	if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
 4021		sched_clock_cpu(cpu); /* Sync clocks across CPUs */
 4022		__ttwu_queue_wakelist(p, cpu, wake_flags);
 4023		return true;
 4024	}
 4025
 4026	return false;
 4027}
 4028
 4029#else /* !CONFIG_SMP */
 4030
 4031static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
 4032{
 4033	return false;
 4034}
 4035
 4036#endif /* CONFIG_SMP */
 4037
 4038static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
 4039{
 4040	struct rq *rq = cpu_rq(cpu);
 4041	struct rq_flags rf;
 4042
 4043	if (ttwu_queue_wakelist(p, cpu, wake_flags))
 
 
 
 4044		return;
 
 
 4045
 4046	rq_lock(rq, &rf);
 4047	update_rq_clock(rq);
 4048	ttwu_do_activate(rq, p, wake_flags, &rf);
 4049	rq_unlock(rq, &rf);
 4050}
 4051
 4052/*
 4053 * Invoked from try_to_wake_up() to check whether the task can be woken up.
 4054 *
 4055 * The caller holds p::pi_lock if p != current or has preemption
 4056 * disabled when p == current.
 4057 *
 4058 * The rules of saved_state:
 4059 *
 4060 *   The related locking code always holds p::pi_lock when updating
 4061 *   p::saved_state, which means the code is fully serialized in both cases.
 4062 *
 4063 *   For PREEMPT_RT, the lock wait and lock wakeups happen via TASK_RTLOCK_WAIT.
 4064 *   No other bits set. This allows to distinguish all wakeup scenarios.
 4065 *
 4066 *   For FREEZER, the wakeup happens via TASK_FROZEN. No other bits set. This
 4067 *   allows us to prevent early wakeup of tasks before they can be run on
 4068 *   asymmetric ISA architectures (eg ARMv9).
 4069 */
 4070static __always_inline
 4071bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
 4072{
 4073	int match;
 4074
 4075	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
 4076		WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
 4077			     state != TASK_RTLOCK_WAIT);
 4078	}
 4079
 4080	*success = !!(match = __task_state_match(p, state));
 4081
 4082	/*
 4083	 * Saved state preserves the task state across blocking on
 4084	 * an RT lock or TASK_FREEZABLE tasks.  If the state matches,
 4085	 * set p::saved_state to TASK_RUNNING, but do not wake the task
 4086	 * because it waits for a lock wakeup or __thaw_task(). Also
 4087	 * indicate success because from the regular waker's point of
 4088	 * view this has succeeded.
 4089	 *
 4090	 * After acquiring the lock the task will restore p::__state
 4091	 * from p::saved_state which ensures that the regular
 4092	 * wakeup is not lost. The restore will also set
 4093	 * p::saved_state to TASK_RUNNING so any further tests will
 4094	 * not result in false positives vs. @success
 4095	 */
 4096	if (match < 0)
 4097		p->saved_state = TASK_RUNNING;
 4098
 4099	return match > 0;
 4100}
 4101
 4102/*
 4103 * Notes on Program-Order guarantees on SMP systems.
 4104 *
 4105 *  MIGRATION
 4106 *
 4107 * The basic program-order guarantee on SMP systems is that when a task [t]
 4108 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
 4109 * execution on its new CPU [c1].
 4110 *
 4111 * For migration (of runnable tasks) this is provided by the following means:
 4112 *
 4113 *  A) UNLOCK of the rq(c0)->lock scheduling out task t
 4114 *  B) migration for t is required to synchronize *both* rq(c0)->lock and
 4115 *     rq(c1)->lock (if not at the same time, then in that order).
 4116 *  C) LOCK of the rq(c1)->lock scheduling in task
 4117 *
 4118 * Release/acquire chaining guarantees that B happens after A and C after B.
 
 4119 * Note: the CPU doing B need not be c0 or c1
 4120 *
 4121 * Example:
 4122 *
 4123 *   CPU0            CPU1            CPU2
 4124 *
 4125 *   LOCK rq(0)->lock
 4126 *   sched-out X
 4127 *   sched-in Y
 4128 *   UNLOCK rq(0)->lock
 4129 *
 4130 *                                   LOCK rq(0)->lock // orders against CPU0
 4131 *                                   dequeue X
 4132 *                                   UNLOCK rq(0)->lock
 4133 *
 4134 *                                   LOCK rq(1)->lock
 4135 *                                   enqueue X
 4136 *                                   UNLOCK rq(1)->lock
 4137 *
 4138 *                   LOCK rq(1)->lock // orders against CPU2
 4139 *                   sched-out Z
 4140 *                   sched-in X
 4141 *                   UNLOCK rq(1)->lock
 4142 *
 4143 *
 4144 *  BLOCKING -- aka. SLEEP + WAKEUP
 4145 *
 4146 * For blocking we (obviously) need to provide the same guarantee as for
 4147 * migration. However the means are completely different as there is no lock
 4148 * chain to provide order. Instead we do:
 4149 *
 4150 *   1) smp_store_release(X->on_cpu, 0)   -- finish_task()
 4151 *   2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
 4152 *
 4153 * Example:
 4154 *
 4155 *   CPU0 (schedule)  CPU1 (try_to_wake_up) CPU2 (schedule)
 4156 *
 4157 *   LOCK rq(0)->lock LOCK X->pi_lock
 4158 *   dequeue X
 4159 *   sched-out X
 4160 *   smp_store_release(X->on_cpu, 0);
 4161 *
 4162 *                    smp_cond_load_acquire(&X->on_cpu, !VAL);
 4163 *                    X->state = WAKING
 4164 *                    set_task_cpu(X,2)
 4165 *
 4166 *                    LOCK rq(2)->lock
 4167 *                    enqueue X
 4168 *                    X->state = RUNNING
 4169 *                    UNLOCK rq(2)->lock
 4170 *
 4171 *                                          LOCK rq(2)->lock // orders against CPU1
 4172 *                                          sched-out Z
 4173 *                                          sched-in X
 4174 *                                          UNLOCK rq(2)->lock
 4175 *
 4176 *                    UNLOCK X->pi_lock
 4177 *   UNLOCK rq(0)->lock
 4178 *
 4179 *
 4180 * However, for wakeups there is a second guarantee we must provide, namely we
 4181 * must ensure that CONDITION=1 done by the caller can not be reordered with
 4182 * accesses to the task state; see try_to_wake_up() and set_current_state().
 
 
 
 
 
 
 
 4183 */
 4184
 4185/**
 4186 * try_to_wake_up - wake up a thread
 4187 * @p: the thread to be awakened
 4188 * @state: the mask of task states that can be woken
 4189 * @wake_flags: wake modifier flags (WF_*)
 4190 *
 4191 * Conceptually does:
 4192 *
 4193 *   If (@state & @p->state) @p->state = TASK_RUNNING.
 4194 *
 4195 * If the task was not queued/runnable, also place it back on a runqueue.
 4196 *
 4197 * This function is atomic against schedule() which would dequeue the task.
 4198 *
 4199 * It issues a full memory barrier before accessing @p->state, see the comment
 4200 * with set_current_state().
 4201 *
 4202 * Uses p->pi_lock to serialize against concurrent wake-ups.
 4203 *
 4204 * Relies on p->pi_lock stabilizing:
 4205 *  - p->sched_class
 4206 *  - p->cpus_ptr
 4207 *  - p->sched_task_group
 4208 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
 4209 *
 4210 * Tries really hard to only take one task_rq(p)->lock for performance.
 4211 * Takes rq->lock in:
 4212 *  - ttwu_runnable()    -- old rq, unavoidable, see comment there;
 4213 *  - ttwu_queue()       -- new rq, for enqueue of the task;
 4214 *  - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
 4215 *
 4216 * As a consequence we race really badly with just about everything. See the
 4217 * many memory barriers and their comments for details.
 4218 *
 4219 * Return: %true if @p->state changes (an actual wakeup was done),
 4220 *	   %false otherwise.
 4221 */
 4222int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
 
 4223{
 4224	guard(preempt)();
 4225	int cpu, success = 0;
 4226
 4227	if (p == current) {
 4228		/*
 4229		 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
 4230		 * == smp_processor_id()'. Together this means we can special
 4231		 * case the whole 'p->on_rq && ttwu_runnable()' case below
 4232		 * without taking any locks.
 4233		 *
 4234		 * In particular:
 4235		 *  - we rely on Program-Order guarantees for all the ordering,
 4236		 *  - we're serialized against set_special_state() by virtue of
 4237		 *    it disabling IRQs (this allows not taking ->pi_lock).
 4238		 */
 4239		if (!ttwu_state_match(p, state, &success))
 4240			goto out;
 4241
 4242		trace_sched_waking(p);
 4243		ttwu_do_wakeup(p);
 4244		goto out;
 4245	}
 4246
 4247	/*
 4248	 * If we are going to wake up a thread waiting for CONDITION we
 4249	 * need to ensure that CONDITION=1 done by the caller can not be
 4250	 * reordered with p->state check below. This pairs with smp_store_mb()
 4251	 * in set_current_state() that the waiting thread does.
 4252	 */
 4253	scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
 4254		smp_mb__after_spinlock();
 4255		if (!ttwu_state_match(p, state, &success))
 4256			break;
 4257
 4258		trace_sched_waking(p);
 4259
 4260		/*
 4261		 * Ensure we load p->on_rq _after_ p->state, otherwise it would
 4262		 * be possible to, falsely, observe p->on_rq == 0 and get stuck
 4263		 * in smp_cond_load_acquire() below.
 4264		 *
 4265		 * sched_ttwu_pending()			try_to_wake_up()
 4266		 *   STORE p->on_rq = 1			  LOAD p->state
 4267		 *   UNLOCK rq->lock
 4268		 *
 4269		 * __schedule() (switch to task 'p')
 4270		 *   LOCK rq->lock			  smp_rmb();
 4271		 *   smp_mb__after_spinlock();
 4272		 *   UNLOCK rq->lock
 4273		 *
 4274		 * [task p]
 4275		 *   STORE p->state = UNINTERRUPTIBLE	  LOAD p->on_rq
 4276		 *
 4277		 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
 4278		 * __schedule().  See the comment for smp_mb__after_spinlock().
 4279		 *
 4280		 * A similar smp_rmb() lives in __task_needs_rq_lock().
 4281		 */
 4282		smp_rmb();
 4283		if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
 4284			break;
 
 
 
 4285
 4286#ifdef CONFIG_SMP
 4287		/*
 4288		 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
 4289		 * possible to, falsely, observe p->on_cpu == 0.
 4290		 *
 4291		 * One must be running (->on_cpu == 1) in order to remove oneself
 4292		 * from the runqueue.
 4293		 *
 4294		 * __schedule() (switch to task 'p')	try_to_wake_up()
 4295		 *   STORE p->on_cpu = 1		  LOAD p->on_rq
 4296		 *   UNLOCK rq->lock
 4297		 *
 4298		 * __schedule() (put 'p' to sleep)
 4299		 *   LOCK rq->lock			  smp_rmb();
 4300		 *   smp_mb__after_spinlock();
 4301		 *   STORE p->on_rq = 0			  LOAD p->on_cpu
 4302		 *
 4303		 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
 4304		 * __schedule().  See the comment for smp_mb__after_spinlock().
 4305		 *
 4306		 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
 4307		 * schedule()'s deactivate_task() has 'happened' and p will no longer
 4308		 * care about it's own p->state. See the comment in __schedule().
 4309		 */
 4310		smp_acquire__after_ctrl_dep();
 
 
 
 
 
 
 
 
 4311
 4312		/*
 4313		 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
 4314		 * == 0), which means we need to do an enqueue, change p->state to
 4315		 * TASK_WAKING such that we can unlock p->pi_lock before doing the
 4316		 * enqueue, such as ttwu_queue_wakelist().
 4317		 */
 4318		WRITE_ONCE(p->__state, TASK_WAKING);
 4319
 4320		/*
 4321		 * If the owning (remote) CPU is still in the middle of schedule() with
 4322		 * this task as prev, considering queueing p on the remote CPUs wake_list
 4323		 * which potentially sends an IPI instead of spinning on p->on_cpu to
 4324		 * let the waker make forward progress. This is safe because IRQs are
 4325		 * disabled and the IPI will deliver after on_cpu is cleared.
 4326		 *
 4327		 * Ensure we load task_cpu(p) after p->on_cpu:
 4328		 *
 4329		 * set_task_cpu(p, cpu);
 4330		 *   STORE p->cpu = @cpu
 4331		 * __schedule() (switch to task 'p')
 4332		 *   LOCK rq->lock
 4333		 *   smp_mb__after_spin_lock()		smp_cond_load_acquire(&p->on_cpu)
 4334		 *   STORE p->on_cpu = 1		LOAD p->cpu
 4335		 *
 4336		 * to ensure we observe the correct CPU on which the task is currently
 4337		 * scheduling.
 4338		 */
 4339		if (smp_load_acquire(&p->on_cpu) &&
 4340		    ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
 4341			break;
 4342
 4343		/*
 4344		 * If the owning (remote) CPU is still in the middle of schedule() with
 4345		 * this task as prev, wait until it's done referencing the task.
 4346		 *
 4347		 * Pairs with the smp_store_release() in finish_task().
 4348		 *
 4349		 * This ensures that tasks getting woken will be fully ordered against
 4350		 * their previous state and preserve Program Order.
 4351		 */
 4352		smp_cond_load_acquire(&p->on_cpu, !VAL);
 4353
 4354		cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
 4355		if (task_cpu(p) != cpu) {
 4356			if (p->in_iowait) {
 4357				delayacct_blkio_end(p);
 4358				atomic_dec(&task_rq(p)->nr_iowait);
 4359			}
 4360
 4361			wake_flags |= WF_MIGRATED;
 4362			psi_ttwu_dequeue(p);
 4363			set_task_cpu(p, cpu);
 4364		}
 4365#else
 4366		cpu = task_cpu(p);
 4367#endif /* CONFIG_SMP */
 4368
 4369		ttwu_queue(p, cpu, wake_flags);
 4370	}
 
 4371out:
 4372	if (success)
 4373		ttwu_stat(p, task_cpu(p), wake_flags);
 4374
 4375	return success;
 4376}
 4377
 4378static bool __task_needs_rq_lock(struct task_struct *p)
 4379{
 4380	unsigned int state = READ_ONCE(p->__state);
 4381
 4382	/*
 4383	 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
 4384	 * the task is blocked. Make sure to check @state since ttwu() can drop
 4385	 * locks at the end, see ttwu_queue_wakelist().
 4386	 */
 4387	if (state == TASK_RUNNING || state == TASK_WAKING)
 4388		return true;
 4389
 4390	/*
 4391	 * Ensure we load p->on_rq after p->__state, otherwise it would be
 4392	 * possible to, falsely, observe p->on_rq == 0.
 4393	 *
 4394	 * See try_to_wake_up() for a longer comment.
 4395	 */
 4396	smp_rmb();
 4397	if (p->on_rq)
 4398		return true;
 4399
 4400#ifdef CONFIG_SMP
 4401	/*
 4402	 * Ensure the task has finished __schedule() and will not be referenced
 4403	 * anymore. Again, see try_to_wake_up() for a longer comment.
 4404	 */
 4405	smp_rmb();
 4406	smp_cond_load_acquire(&p->on_cpu, !VAL);
 4407#endif
 4408
 4409	return false;
 4410}
 4411
 4412/**
 4413 * task_call_func - Invoke a function on task in fixed state
 4414 * @p: Process for which the function is to be invoked, can be @current.
 4415 * @func: Function to invoke.
 4416 * @arg: Argument to function.
 4417 *
 4418 * Fix the task in it's current state by avoiding wakeups and or rq operations
 4419 * and call @func(@arg) on it.  This function can use ->on_rq and task_curr()
 4420 * to work out what the state is, if required.  Given that @func can be invoked
 4421 * with a runqueue lock held, it had better be quite lightweight.
 4422 *
 4423 * Returns:
 4424 *   Whatever @func returns
 
 4425 */
 4426int task_call_func(struct task_struct *p, task_call_f func, void *arg)
 4427{
 4428	struct rq *rq = NULL;
 4429	struct rq_flags rf;
 4430	int ret;
 
 
 4431
 4432	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
 4433
 4434	if (__task_needs_rq_lock(p))
 4435		rq = __task_rq_lock(p, &rf);
 
 
 
 
 
 
 
 
 
 4436
 4437	/*
 4438	 * At this point the task is pinned; either:
 4439	 *  - blocked and we're holding off wakeups	 (pi->lock)
 4440	 *  - woken, and we're holding off enqueue	 (rq->lock)
 4441	 *  - queued, and we're holding off schedule	 (rq->lock)
 4442	 *  - running, and we're holding off de-schedule (rq->lock)
 4443	 *
 4444	 * The called function (@func) can use: task_curr(), p->on_rq and
 4445	 * p->__state to differentiate between these states.
 4446	 */
 4447	ret = func(p, arg);
 4448
 4449	if (rq)
 4450		rq_unlock(rq, &rf);
 4451
 4452	raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
 4453	return ret;
 4454}
 
 
 
 
 4455
 4456/**
 4457 * cpu_curr_snapshot - Return a snapshot of the currently running task
 4458 * @cpu: The CPU on which to snapshot the task.
 4459 *
 4460 * Returns the task_struct pointer of the task "currently" running on
 4461 * the specified CPU.  If the same task is running on that CPU throughout,
 4462 * the return value will be a pointer to that task's task_struct structure.
 4463 * If the CPU did any context switches even vaguely concurrently with the
 4464 * execution of this function, the return value will be a pointer to the
 4465 * task_struct structure of a randomly chosen task that was running on
 4466 * that CPU somewhere around the time that this function was executing.
 4467 *
 4468 * If the specified CPU was offline, the return value is whatever it
 4469 * is, perhaps a pointer to the task_struct structure of that CPU's idle
 4470 * task, but there is no guarantee.  Callers wishing a useful return
 4471 * value must take some action to ensure that the specified CPU remains
 4472 * online throughout.
 4473 *
 4474 * This function executes full memory barriers before and after fetching
 4475 * the pointer, which permits the caller to confine this function's fetch
 4476 * with respect to the caller's accesses to other shared variables.
 4477 */
 4478struct task_struct *cpu_curr_snapshot(int cpu)
 4479{
 4480	struct task_struct *t;
 4481
 4482	smp_mb(); /* Pairing determined by caller's synchronization design. */
 4483	t = rcu_dereference(cpu_curr(cpu));
 4484	smp_mb(); /* Pairing determined by caller's synchronization design. */
 4485	return t;
 4486}
 4487
 4488/**
 4489 * wake_up_process - Wake up a specific process
 4490 * @p: The process to be woken up.
 4491 *
 4492 * Attempt to wake up the nominated process and move it to the set of runnable
 4493 * processes.
 4494 *
 4495 * Return: 1 if the process was woken up, 0 if it was already running.
 4496 *
 4497 * This function executes a full memory barrier before accessing the task state.
 
 4498 */
 4499int wake_up_process(struct task_struct *p)
 4500{
 4501	return try_to_wake_up(p, TASK_NORMAL, 0);
 4502}
 4503EXPORT_SYMBOL(wake_up_process);
 4504
 4505int wake_up_state(struct task_struct *p, unsigned int state)
 4506{
 4507	return try_to_wake_up(p, state, 0);
 4508}
 4509
 4510/*
 4511 * Perform scheduler related setup for a newly forked process p.
 4512 * p is forked by current.
 4513 *
 4514 * __sched_fork() is basic setup used by init_idle() too:
 4515 */
 4516static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
 4517{
 4518	p->on_rq			= 0;
 4519
 4520	p->se.on_rq			= 0;
 4521	p->se.exec_start		= 0;
 4522	p->se.sum_exec_runtime		= 0;
 4523	p->se.prev_sum_exec_runtime	= 0;
 4524	p->se.nr_migrations		= 0;
 4525	p->se.vruntime			= 0;
 4526	p->se.vlag			= 0;
 4527	p->se.slice			= sysctl_sched_base_slice;
 4528	INIT_LIST_HEAD(&p->se.group_node);
 4529
 4530#ifdef CONFIG_FAIR_GROUP_SCHED
 4531	p->se.cfs_rq			= NULL;
 4532#endif
 4533
 4534#ifdef CONFIG_SCHEDSTATS
 4535	/* Even if schedstat is disabled, there should not be garbage */
 4536	memset(&p->stats, 0, sizeof(p->stats));
 4537#endif
 4538
 4539	init_dl_entity(&p->dl);
 
 
 
 4540
 4541	INIT_LIST_HEAD(&p->rt.run_list);
 4542	p->rt.timeout		= 0;
 4543	p->rt.time_slice	= sched_rr_timeslice;
 4544	p->rt.on_rq		= 0;
 4545	p->rt.on_list		= 0;
 4546
 4547#ifdef CONFIG_PREEMPT_NOTIFIERS
 4548	INIT_HLIST_HEAD(&p->preempt_notifiers);
 4549#endif
 4550
 4551#ifdef CONFIG_COMPACTION
 4552	p->capture_control = NULL;
 4553#endif
 4554	init_numa_balancing(clone_flags, p);
 4555#ifdef CONFIG_SMP
 4556	p->wake_entry.u_flags = CSD_TYPE_TTWU;
 4557	p->migration_pending = NULL;
 4558#endif
 4559	init_sched_mm_cid(p);
 
 
 
 
 
 
 
 
 
 
 
 
 4560}
 4561
 4562DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
 4563
 4564#ifdef CONFIG_NUMA_BALANCING
 4565
 4566int sysctl_numa_balancing_mode;
 4567
 4568static void __set_numabalancing_state(bool enabled)
 4569{
 4570	if (enabled)
 4571		static_branch_enable(&sched_numa_balancing);
 4572	else
 4573		static_branch_disable(&sched_numa_balancing);
 4574}
 4575
 4576void set_numabalancing_state(bool enabled)
 4577{
 4578	if (enabled)
 4579		sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
 4580	else
 4581		sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
 4582	__set_numabalancing_state(enabled);
 4583}
 4584
 4585#ifdef CONFIG_PROC_SYSCTL
 4586static void reset_memory_tiering(void)
 4587{
 4588	struct pglist_data *pgdat;
 4589
 4590	for_each_online_pgdat(pgdat) {
 4591		pgdat->nbp_threshold = 0;
 4592		pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
 4593		pgdat->nbp_th_start = jiffies_to_msecs(jiffies);
 4594	}
 4595}
 4596
 4597static int sysctl_numa_balancing(struct ctl_table *table, int write,
 4598			  void *buffer, size_t *lenp, loff_t *ppos)
 4599{
 4600	struct ctl_table t;
 4601	int err;
 4602	int state = sysctl_numa_balancing_mode;
 4603
 4604	if (write && !capable(CAP_SYS_ADMIN))
 4605		return -EPERM;
 4606
 4607	t = *table;
 4608	t.data = &state;
 4609	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
 4610	if (err < 0)
 4611		return err;
 4612	if (write) {
 4613		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
 4614		    (state & NUMA_BALANCING_MEMORY_TIERING))
 4615			reset_memory_tiering();
 4616		sysctl_numa_balancing_mode = state;
 4617		__set_numabalancing_state(state);
 4618	}
 4619	return err;
 4620}
 4621#endif
 4622#endif
 4623
 4624#ifdef CONFIG_SCHEDSTATS
 4625
 4626DEFINE_STATIC_KEY_FALSE(sched_schedstats);
 
 4627
 4628static void set_schedstats(bool enabled)
 4629{
 4630	if (enabled)
 4631		static_branch_enable(&sched_schedstats);
 4632	else
 4633		static_branch_disable(&sched_schedstats);
 4634}
 4635
 4636void force_schedstat_enabled(void)
 4637{
 4638	if (!schedstat_enabled()) {
 4639		pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
 4640		static_branch_enable(&sched_schedstats);
 4641	}
 4642}
 4643
 4644static int __init setup_schedstats(char *str)
 4645{
 4646	int ret = 0;
 4647	if (!str)
 4648		goto out;
 4649
 
 
 
 
 
 4650	if (!strcmp(str, "enable")) {
 4651		set_schedstats(true);
 4652		ret = 1;
 4653	} else if (!strcmp(str, "disable")) {
 4654		set_schedstats(false);
 4655		ret = 1;
 4656	}
 4657out:
 4658	if (!ret)
 4659		pr_warn("Unable to parse schedstats=\n");
 4660
 4661	return ret;
 4662}
 4663__setup("schedstats=", setup_schedstats);
 4664
 
 
 
 
 
 4665#ifdef CONFIG_PROC_SYSCTL
 4666static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
 4667		size_t *lenp, loff_t *ppos)
 4668{
 4669	struct ctl_table t;
 4670	int err;
 4671	int state = static_branch_likely(&sched_schedstats);
 4672
 4673	if (write && !capable(CAP_SYS_ADMIN))
 4674		return -EPERM;
 4675
 4676	t = *table;
 4677	t.data = &state;
 4678	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
 4679	if (err < 0)
 4680		return err;
 4681	if (write)
 4682		set_schedstats(state);
 4683	return err;
 4684}
 4685#endif /* CONFIG_PROC_SYSCTL */
 
 
 4686#endif /* CONFIG_SCHEDSTATS */
 4687
 4688#ifdef CONFIG_SYSCTL
 4689static struct ctl_table sched_core_sysctls[] = {
 4690#ifdef CONFIG_SCHEDSTATS
 4691	{
 4692		.procname       = "sched_schedstats",
 4693		.data           = NULL,
 4694		.maxlen         = sizeof(unsigned int),
 4695		.mode           = 0644,
 4696		.proc_handler   = sysctl_schedstats,
 4697		.extra1         = SYSCTL_ZERO,
 4698		.extra2         = SYSCTL_ONE,
 4699	},
 4700#endif /* CONFIG_SCHEDSTATS */
 4701#ifdef CONFIG_UCLAMP_TASK
 4702	{
 4703		.procname       = "sched_util_clamp_min",
 4704		.data           = &sysctl_sched_uclamp_util_min,
 4705		.maxlen         = sizeof(unsigned int),
 4706		.mode           = 0644,
 4707		.proc_handler   = sysctl_sched_uclamp_handler,
 4708	},
 4709	{
 4710		.procname       = "sched_util_clamp_max",
 4711		.data           = &sysctl_sched_uclamp_util_max,
 4712		.maxlen         = sizeof(unsigned int),
 4713		.mode           = 0644,
 4714		.proc_handler   = sysctl_sched_uclamp_handler,
 4715	},
 4716	{
 4717		.procname       = "sched_util_clamp_min_rt_default",
 4718		.data           = &sysctl_sched_uclamp_util_min_rt_default,
 4719		.maxlen         = sizeof(unsigned int),
 4720		.mode           = 0644,
 4721		.proc_handler   = sysctl_sched_uclamp_handler,
 4722	},
 4723#endif /* CONFIG_UCLAMP_TASK */
 4724#ifdef CONFIG_NUMA_BALANCING
 4725	{
 4726		.procname	= "numa_balancing",
 4727		.data		= NULL, /* filled in by handler */
 4728		.maxlen		= sizeof(unsigned int),
 4729		.mode		= 0644,
 4730		.proc_handler	= sysctl_numa_balancing,
 4731		.extra1		= SYSCTL_ZERO,
 4732		.extra2		= SYSCTL_FOUR,
 4733	},
 4734#endif /* CONFIG_NUMA_BALANCING */
 4735	{}
 4736};
 4737static int __init sched_core_sysctl_init(void)
 4738{
 4739	register_sysctl_init("kernel", sched_core_sysctls);
 4740	return 0;
 4741}
 4742late_initcall(sched_core_sysctl_init);
 4743#endif /* CONFIG_SYSCTL */
 4744
 4745/*
 4746 * fork()/clone()-time setup:
 4747 */
 4748int sched_fork(unsigned long clone_flags, struct task_struct *p)
 4749{
 
 
 
 4750	__sched_fork(clone_flags, p);
 4751	/*
 4752	 * We mark the process as NEW here. This guarantees that
 4753	 * nobody will actually run it, and a signal or other external
 4754	 * event cannot wake it up and insert it on the runqueue either.
 4755	 */
 4756	p->__state = TASK_NEW;
 4757
 4758	/*
 4759	 * Make sure we do not leak PI boosting priority to the child.
 4760	 */
 4761	p->prio = current->normal_prio;
 4762
 4763	uclamp_fork(p);
 4764
 4765	/*
 4766	 * Revert to default priority/policy on fork if requested.
 4767	 */
 4768	if (unlikely(p->sched_reset_on_fork)) {
 4769		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
 4770			p->policy = SCHED_NORMAL;
 4771			p->static_prio = NICE_TO_PRIO(0);
 4772			p->rt_priority = 0;
 4773		} else if (PRIO_TO_NICE(p->static_prio) < 0)
 4774			p->static_prio = NICE_TO_PRIO(0);
 4775
 4776		p->prio = p->normal_prio = p->static_prio;
 4777		set_load_weight(p, false);
 4778
 4779		/*
 4780		 * We don't need the reset flag anymore after the fork. It has
 4781		 * fulfilled its duty:
 4782		 */
 4783		p->sched_reset_on_fork = 0;
 4784	}
 4785
 4786	if (dl_prio(p->prio))
 
 4787		return -EAGAIN;
 4788	else if (rt_prio(p->prio))
 4789		p->sched_class = &rt_sched_class;
 4790	else
 4791		p->sched_class = &fair_sched_class;
 
 4792
 4793	init_entity_runnable_average(&p->se);
 4794
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 4795
 4796#ifdef CONFIG_SCHED_INFO
 4797	if (likely(sched_info_on()))
 4798		memset(&p->sched_info, 0, sizeof(p->sched_info));
 4799#endif
 4800#if defined(CONFIG_SMP)
 4801	p->on_cpu = 0;
 4802#endif
 4803	init_task_preempt_count(p);
 4804#ifdef CONFIG_SMP
 4805	plist_node_init(&p->pushable_tasks, MAX_PRIO);
 4806	RB_CLEAR_NODE(&p->pushable_dl_tasks);
 4807#endif
 
 
 4808	return 0;
 4809}
 4810
 4811void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
 4812{
 4813	unsigned long flags;
 4814
 4815	/*
 4816	 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
 4817	 * required yet, but lockdep gets upset if rules are violated.
 4818	 */
 4819	raw_spin_lock_irqsave(&p->pi_lock, flags);
 4820#ifdef CONFIG_CGROUP_SCHED
 4821	if (1) {
 4822		struct task_group *tg;
 4823		tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
 4824				  struct task_group, css);
 4825		tg = autogroup_task_group(p, tg);
 4826		p->sched_task_group = tg;
 4827	}
 4828#endif
 4829	rseq_migrate(p);
 4830	/*
 4831	 * We're setting the CPU for the first time, we don't migrate,
 4832	 * so use __set_task_cpu().
 4833	 */
 4834	__set_task_cpu(p, smp_processor_id());
 4835	if (p->sched_class->task_fork)
 4836		p->sched_class->task_fork(p);
 4837	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
 4838}
 4839
 4840void sched_post_fork(struct task_struct *p)
 4841{
 4842	uclamp_post_fork(p);
 4843}
 4844
 4845unsigned long to_ratio(u64 period, u64 runtime)
 4846{
 4847	if (runtime == RUNTIME_INF)
 4848		return BW_UNIT;
 4849
 4850	/*
 4851	 * Doing this here saves a lot of checks in all
 4852	 * the calling paths, and returning zero seems
 4853	 * safe for them anyway.
 4854	 */
 4855	if (period == 0)
 4856		return 0;
 4857
 4858	return div64_u64(runtime << BW_SHIFT, period);
 4859}
 4860
 4861/*
 4862 * wake_up_new_task - wake up a newly created task for the first time.
 4863 *
 4864 * This function will do some initial scheduler statistics housekeeping
 4865 * that must be done for every newly created context, then puts the task
 4866 * on the runqueue and wakes it.
 4867 */
 4868void wake_up_new_task(struct task_struct *p)
 4869{
 4870	struct rq_flags rf;
 4871	struct rq *rq;
 4872
 4873	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
 4874	WRITE_ONCE(p->__state, TASK_RUNNING);
 4875#ifdef CONFIG_SMP
 4876	/*
 4877	 * Fork balancing, do it here and not earlier because:
 4878	 *  - cpus_ptr can change in the fork path
 4879	 *  - any previously selected CPU might disappear through hotplug
 4880	 *
 4881	 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
 4882	 * as we're not fully set-up yet.
 4883	 */
 4884	p->recent_used_cpu = task_cpu(p);
 4885	rseq_migrate(p);
 4886	__set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
 4887#endif
 4888	rq = __task_rq_lock(p, &rf);
 4889	update_rq_clock(rq);
 4890	post_init_entity_util_avg(p);
 4891
 4892	activate_task(rq, p, ENQUEUE_NOCLOCK);
 
 4893	trace_sched_wakeup_new(p);
 4894	wakeup_preempt(rq, p, WF_FORK);
 4895#ifdef CONFIG_SMP
 4896	if (p->sched_class->task_woken) {
 4897		/*
 4898		 * Nothing relies on rq->lock after this, so it's fine to
 4899		 * drop it.
 4900		 */
 4901		rq_unpin_lock(rq, &rf);
 4902		p->sched_class->task_woken(rq, p);
 4903		rq_repin_lock(rq, &rf);
 4904	}
 4905#endif
 4906	task_rq_unlock(rq, p, &rf);
 4907}
 4908
 4909#ifdef CONFIG_PREEMPT_NOTIFIERS
 4910
 4911static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
 4912
 4913void preempt_notifier_inc(void)
 4914{
 4915	static_branch_inc(&preempt_notifier_key);
 4916}
 4917EXPORT_SYMBOL_GPL(preempt_notifier_inc);
 4918
 4919void preempt_notifier_dec(void)
 4920{
 4921	static_branch_dec(&preempt_notifier_key);
 4922}
 4923EXPORT_SYMBOL_GPL(preempt_notifier_dec);
 4924
 4925/**
 4926 * preempt_notifier_register - tell me when current is being preempted & rescheduled
 4927 * @notifier: notifier struct to register
 4928 */
 4929void preempt_notifier_register(struct preempt_notifier *notifier)
 4930{
 4931	if (!static_branch_unlikely(&preempt_notifier_key))
 4932		WARN(1, "registering preempt_notifier while notifiers disabled\n");
 4933
 4934	hlist_add_head(&notifier->link, &current->preempt_notifiers);
 4935}
 4936EXPORT_SYMBOL_GPL(preempt_notifier_register);
 4937
 4938/**
 4939 * preempt_notifier_unregister - no longer interested in preemption notifications
 4940 * @notifier: notifier struct to unregister
 4941 *
 4942 * This is *not* safe to call from within a preemption notifier.
 4943 */
 4944void preempt_notifier_unregister(struct preempt_notifier *notifier)
 4945{
 4946	hlist_del(&notifier->link);
 4947}
 4948EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
 4949
 4950static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
 4951{
 4952	struct preempt_notifier *notifier;
 4953
 4954	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
 4955		notifier->ops->sched_in(notifier, raw_smp_processor_id());
 4956}
 4957
 4958static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
 4959{
 4960	if (static_branch_unlikely(&preempt_notifier_key))
 4961		__fire_sched_in_preempt_notifiers(curr);
 4962}
 4963
 4964static void
 4965__fire_sched_out_preempt_notifiers(struct task_struct *curr,
 4966				   struct task_struct *next)
 4967{
 4968	struct preempt_notifier *notifier;
 4969
 4970	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
 4971		notifier->ops->sched_out(notifier, next);
 4972}
 4973
 4974static __always_inline void
 4975fire_sched_out_preempt_notifiers(struct task_struct *curr,
 4976				 struct task_struct *next)
 4977{
 4978	if (static_branch_unlikely(&preempt_notifier_key))
 4979		__fire_sched_out_preempt_notifiers(curr, next);
 4980}
 4981
 4982#else /* !CONFIG_PREEMPT_NOTIFIERS */
 4983
 4984static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
 4985{
 4986}
 4987
 4988static inline void
 4989fire_sched_out_preempt_notifiers(struct task_struct *curr,
 4990				 struct task_struct *next)
 4991{
 4992}
 4993
 4994#endif /* CONFIG_PREEMPT_NOTIFIERS */
 4995
 4996static inline void prepare_task(struct task_struct *next)
 4997{
 4998#ifdef CONFIG_SMP
 4999	/*
 5000	 * Claim the task as running, we do this before switching to it
 5001	 * such that any running task will have this set.
 5002	 *
 5003	 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
 5004	 * its ordering comment.
 5005	 */
 5006	WRITE_ONCE(next->on_cpu, 1);
 5007#endif
 5008}
 5009
 5010static inline void finish_task(struct task_struct *prev)
 5011{
 5012#ifdef CONFIG_SMP
 5013	/*
 5014	 * This must be the very last reference to @prev from this CPU. After
 5015	 * p->on_cpu is cleared, the task can be moved to a different CPU. We
 5016	 * must ensure this doesn't happen until the switch is completely
 5017	 * finished.
 5018	 *
 5019	 * In particular, the load of prev->state in finish_task_switch() must
 5020	 * happen before this.
 5021	 *
 5022	 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
 5023	 */
 5024	smp_store_release(&prev->on_cpu, 0);
 5025#endif
 5026}
 5027
 5028#ifdef CONFIG_SMP
 5029
 5030static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
 5031{
 5032	void (*func)(struct rq *rq);
 5033	struct balance_callback *next;
 5034
 5035	lockdep_assert_rq_held(rq);
 5036
 5037	while (head) {
 5038		func = (void (*)(struct rq *))head->func;
 5039		next = head->next;
 5040		head->next = NULL;
 5041		head = next;
 5042
 5043		func(rq);
 5044	}
 5045}
 5046
 5047static void balance_push(struct rq *rq);
 5048
 5049/*
 5050 * balance_push_callback is a right abuse of the callback interface and plays
 5051 * by significantly different rules.
 5052 *
 5053 * Where the normal balance_callback's purpose is to be ran in the same context
 5054 * that queued it (only later, when it's safe to drop rq->lock again),
 5055 * balance_push_callback is specifically targeted at __schedule().
 5056 *
 5057 * This abuse is tolerated because it places all the unlikely/odd cases behind
 5058 * a single test, namely: rq->balance_callback == NULL.
 5059 */
 5060struct balance_callback balance_push_callback = {
 5061	.next = NULL,
 5062	.func = balance_push,
 5063};
 5064
 5065static inline struct balance_callback *
 5066__splice_balance_callbacks(struct rq *rq, bool split)
 5067{
 5068	struct balance_callback *head = rq->balance_callback;
 5069
 5070	if (likely(!head))
 5071		return NULL;
 5072
 5073	lockdep_assert_rq_held(rq);
 5074	/*
 5075	 * Must not take balance_push_callback off the list when
 5076	 * splice_balance_callbacks() and balance_callbacks() are not
 5077	 * in the same rq->lock section.
 5078	 *
 5079	 * In that case it would be possible for __schedule() to interleave
 5080	 * and observe the list empty.
 5081	 */
 5082	if (split && head == &balance_push_callback)
 5083		head = NULL;
 5084	else
 5085		rq->balance_callback = NULL;
 5086
 5087	return head;
 5088}
 5089
 5090static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
 5091{
 5092	return __splice_balance_callbacks(rq, true);
 5093}
 5094
 5095static void __balance_callbacks(struct rq *rq)
 5096{
 5097	do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
 5098}
 5099
 5100static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
 5101{
 5102	unsigned long flags;
 5103
 5104	if (unlikely(head)) {
 5105		raw_spin_rq_lock_irqsave(rq, flags);
 5106		do_balance_callbacks(rq, head);
 5107		raw_spin_rq_unlock_irqrestore(rq, flags);
 5108	}
 5109}
 5110
 5111#else
 5112
 5113static inline void __balance_callbacks(struct rq *rq)
 5114{
 5115}
 5116
 5117static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
 5118{
 5119	return NULL;
 5120}
 5121
 5122static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
 5123{
 5124}
 5125
 5126#endif
 5127
 5128static inline void
 5129prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
 5130{
 5131	/*
 5132	 * Since the runqueue lock will be released by the next
 5133	 * task (which is an invalid locking op but in the case
 5134	 * of the scheduler it's an obvious special-case), so we
 5135	 * do an early lockdep release here:
 5136	 */
 5137	rq_unpin_lock(rq, rf);
 5138	spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
 5139#ifdef CONFIG_DEBUG_SPINLOCK
 5140	/* this is a valid case when another task releases the spinlock */
 5141	rq_lockp(rq)->owner = next;
 5142#endif
 5143}
 5144
 5145static inline void finish_lock_switch(struct rq *rq)
 5146{
 5147	/*
 5148	 * If we are tracking spinlock dependencies then we have to
 5149	 * fix up the runqueue lock - which gets 'carried over' from
 5150	 * prev into current:
 5151	 */
 5152	spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
 5153	__balance_callbacks(rq);
 5154	raw_spin_rq_unlock_irq(rq);
 5155}
 5156
 5157/*
 5158 * NOP if the arch has not defined these:
 5159 */
 5160
 5161#ifndef prepare_arch_switch
 5162# define prepare_arch_switch(next)	do { } while (0)
 5163#endif
 5164
 5165#ifndef finish_arch_post_lock_switch
 5166# define finish_arch_post_lock_switch()	do { } while (0)
 5167#endif
 5168
 5169static inline void kmap_local_sched_out(void)
 5170{
 5171#ifdef CONFIG_KMAP_LOCAL
 5172	if (unlikely(current->kmap_ctrl.idx))
 5173		__kmap_local_sched_out();
 5174#endif
 5175}
 5176
 5177static inline void kmap_local_sched_in(void)
 5178{
 5179#ifdef CONFIG_KMAP_LOCAL
 5180	if (unlikely(current->kmap_ctrl.idx))
 5181		__kmap_local_sched_in();
 5182#endif
 5183}
 5184
 5185/**
 5186 * prepare_task_switch - prepare to switch tasks
 5187 * @rq: the runqueue preparing to switch
 5188 * @prev: the current task that is being switched out
 5189 * @next: the task we are going to switch to.
 5190 *
 5191 * This is called with the rq lock held and interrupts off. It must
 5192 * be paired with a subsequent finish_task_switch after the context
 5193 * switch.
 5194 *
 5195 * prepare_task_switch sets up locking and calls architecture specific
 5196 * hooks.
 5197 */
 5198static inline void
 5199prepare_task_switch(struct rq *rq, struct task_struct *prev,
 5200		    struct task_struct *next)
 5201{
 5202	kcov_prepare_switch(prev);
 5203	sched_info_switch(rq, prev, next);
 5204	perf_event_task_sched_out(prev, next);
 5205	rseq_preempt(prev);
 5206	fire_sched_out_preempt_notifiers(prev, next);
 5207	kmap_local_sched_out();
 5208	prepare_task(next);
 5209	prepare_arch_switch(next);
 5210}
 5211
 5212/**
 5213 * finish_task_switch - clean up after a task-switch
 5214 * @prev: the thread we just switched away from.
 5215 *
 5216 * finish_task_switch must be called after the context switch, paired
 5217 * with a prepare_task_switch call before the context switch.
 5218 * finish_task_switch will reconcile locking set up by prepare_task_switch,
 5219 * and do any other architecture-specific cleanup actions.
 5220 *
 5221 * Note that we may have delayed dropping an mm in context_switch(). If
 5222 * so, we finish that here outside of the runqueue lock. (Doing it
 5223 * with the lock held can cause deadlocks; see schedule() for
 5224 * details.)
 5225 *
 5226 * The context switch have flipped the stack from under us and restored the
 5227 * local variables which were saved when this task called schedule() in the
 5228 * past. prev == current is still correct but we need to recalculate this_rq
 5229 * because prev may have moved to another CPU.
 5230 */
 5231static struct rq *finish_task_switch(struct task_struct *prev)
 5232	__releases(rq->lock)
 5233{
 5234	struct rq *rq = this_rq();
 5235	struct mm_struct *mm = rq->prev_mm;
 5236	unsigned int prev_state;
 5237
 5238	/*
 5239	 * The previous task will have left us with a preempt_count of 2
 5240	 * because it left us after:
 5241	 *
 5242	 *	schedule()
 5243	 *	  preempt_disable();			// 1
 5244	 *	  __schedule()
 5245	 *	    raw_spin_lock_irq(&rq->lock)	// 2
 5246	 *
 5247	 * Also, see FORK_PREEMPT_COUNT.
 5248	 */
 5249	if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
 5250		      "corrupted preempt_count: %s/%d/0x%x\n",
 5251		      current->comm, current->pid, preempt_count()))
 5252		preempt_count_set(FORK_PREEMPT_COUNT);
 5253
 5254	rq->prev_mm = NULL;
 5255
 5256	/*
 5257	 * A task struct has one reference for the use as "current".
 5258	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
 5259	 * schedule one last time. The schedule call will never return, and
 5260	 * the scheduled task must drop that reference.
 5261	 *
 5262	 * We must observe prev->state before clearing prev->on_cpu (in
 5263	 * finish_task), otherwise a concurrent wakeup can get prev
 5264	 * running on another CPU and we could rave with its RUNNING -> DEAD
 5265	 * transition, resulting in a double drop.
 5266	 */
 5267	prev_state = READ_ONCE(prev->__state);
 5268	vtime_task_switch(prev);
 5269	perf_event_task_sched_in(prev, current);
 5270	finish_task(prev);
 5271	tick_nohz_task_switch();
 5272	finish_lock_switch(rq);
 5273	finish_arch_post_lock_switch();
 5274	kcov_finish_switch(current);
 5275	/*
 5276	 * kmap_local_sched_out() is invoked with rq::lock held and
 5277	 * interrupts disabled. There is no requirement for that, but the
 5278	 * sched out code does not have an interrupt enabled section.
 5279	 * Restoring the maps on sched in does not require interrupts being
 5280	 * disabled either.
 5281	 */
 5282	kmap_local_sched_in();
 5283
 5284	fire_sched_in_preempt_notifiers(current);
 5285	/*
 5286	 * When switching through a kernel thread, the loop in
 5287	 * membarrier_{private,global}_expedited() may have observed that
 5288	 * kernel thread and not issued an IPI. It is therefore possible to
 5289	 * schedule between user->kernel->user threads without passing though
 5290	 * switch_mm(). Membarrier requires a barrier after storing to
 5291	 * rq->curr, before returning to userspace, so provide them here:
 5292	 *
 5293	 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
 5294	 *   provided by mmdrop_lazy_tlb(),
 5295	 * - a sync_core for SYNC_CORE.
 5296	 */
 5297	if (mm) {
 5298		membarrier_mm_sync_core_before_usermode(mm);
 5299		mmdrop_lazy_tlb_sched(mm);
 5300	}
 
 
 
 
 
 5301
 5302	if (unlikely(prev_state == TASK_DEAD)) {
 5303		if (prev->sched_class->task_dead)
 5304			prev->sched_class->task_dead(prev);
 
 
 5305
 5306		/* Task is done with its stack. */
 5307		put_task_stack(prev);
 5308
 5309		put_task_struct_rcu_user(prev);
 
 
 
 
 
 
 5310	}
 5311
 
 5312	return rq;
 5313}
 5314
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 5315/**
 5316 * schedule_tail - first thing a freshly forked thread must call.
 5317 * @prev: the thread we just switched away from.
 5318 */
 5319asmlinkage __visible void schedule_tail(struct task_struct *prev)
 5320	__releases(rq->lock)
 5321{
 
 
 5322	/*
 5323	 * New tasks start with FORK_PREEMPT_COUNT, see there and
 5324	 * finish_task_switch() for details.
 5325	 *
 5326	 * finish_task_switch() will drop rq->lock() and lower preempt_count
 5327	 * and the preempt_enable() will end up enabling preemption (on
 5328	 * PREEMPT_COUNT kernels).
 5329	 */
 5330
 5331	finish_task_switch(prev);
 
 5332	preempt_enable();
 5333
 5334	if (current->set_child_tid)
 5335		put_user(task_pid_vnr(current), current->set_child_tid);
 5336
 5337	calculate_sigpending();
 5338}
 5339
 5340/*
 5341 * context_switch - switch to the new MM and the new thread's register state.
 5342 */
 5343static __always_inline struct rq *
 5344context_switch(struct rq *rq, struct task_struct *prev,
 5345	       struct task_struct *next, struct rq_flags *rf)
 5346{
 
 
 5347	prepare_task_switch(rq, prev, next);
 5348
 
 
 5349	/*
 5350	 * For paravirt, this is coupled with an exit in switch_to to
 5351	 * combine the page table reload and the switch backend into
 5352	 * one hypercall.
 5353	 */
 5354	arch_start_context_switch(prev);
 5355
 5356	/*
 5357	 * kernel -> kernel   lazy + transfer active
 5358	 *   user -> kernel   lazy + mmgrab_lazy_tlb() active
 5359	 *
 5360	 * kernel ->   user   switch + mmdrop_lazy_tlb() active
 5361	 *   user ->   user   switch
 5362	 *
 5363	 * switch_mm_cid() needs to be updated if the barriers provided
 5364	 * by context_switch() are modified.
 5365	 */
 5366	if (!next->mm) {                                // to kernel
 5367		enter_lazy_tlb(prev->active_mm, next);
 5368
 5369		next->active_mm = prev->active_mm;
 5370		if (prev->mm)                           // from user
 5371			mmgrab_lazy_tlb(prev->active_mm);
 5372		else
 5373			prev->active_mm = NULL;
 5374	} else {                                        // to user
 5375		membarrier_switch_mm(rq, prev->active_mm, next->mm);
 5376		/*
 5377		 * sys_membarrier() requires an smp_mb() between setting
 5378		 * rq->curr / membarrier_switch_mm() and returning to userspace.
 5379		 *
 5380		 * The below provides this either through switch_mm(), or in
 5381		 * case 'prev->active_mm == next->mm' through
 5382		 * finish_task_switch()'s mmdrop().
 5383		 */
 5384		switch_mm_irqs_off(prev->active_mm, next->mm, next);
 5385		lru_gen_use_mm(next->mm);
 5386
 5387		if (!prev->mm) {                        // from kernel
 5388			/* will mmdrop_lazy_tlb() in finish_task_switch(). */
 5389			rq->prev_mm = prev->active_mm;
 5390			prev->active_mm = NULL;
 5391		}
 5392	}
 5393
 5394	/* switch_mm_cid() requires the memory barriers above. */
 5395	switch_mm_cid(rq, prev, next);
 5396
 5397	prepare_lock_switch(rq, next, rf);
 5398
 5399	/* Here we just switch the register state and the stack. */
 5400	switch_to(prev, next, prev);
 5401	barrier();
 5402
 5403	return finish_task_switch(prev);
 5404}
 5405
 5406/*
 5407 * nr_running and nr_context_switches:
 5408 *
 5409 * externally visible scheduler statistics: current number of runnable
 5410 * threads, total number of context switches performed since bootup.
 5411 */
 5412unsigned int nr_running(void)
 5413{
 5414	unsigned int i, sum = 0;
 5415
 5416	for_each_online_cpu(i)
 5417		sum += cpu_rq(i)->nr_running;
 5418
 5419	return sum;
 5420}
 5421
 5422/*
 5423 * Check if only the current task is running on the CPU.
 5424 *
 5425 * Caution: this function does not check that the caller has disabled
 5426 * preemption, thus the result might have a time-of-check-to-time-of-use
 5427 * race.  The caller is responsible to use it correctly, for example:
 5428 *
 5429 * - from a non-preemptible section (of course)
 5430 *
 5431 * - from a thread that is bound to a single CPU
 5432 *
 5433 * - in a loop with very short iterations (e.g. a polling loop)
 5434 */
 5435bool single_task_running(void)
 5436{
 5437	return raw_rq()->nr_running == 1;
 5438}
 5439EXPORT_SYMBOL(single_task_running);
 5440
 5441unsigned long long nr_context_switches_cpu(int cpu)
 5442{
 5443	return cpu_rq(cpu)->nr_switches;
 5444}
 5445
 5446unsigned long long nr_context_switches(void)
 5447{
 5448	int i;
 5449	unsigned long long sum = 0;
 5450
 5451	for_each_possible_cpu(i)
 5452		sum += cpu_rq(i)->nr_switches;
 5453
 5454	return sum;
 5455}
 5456
 5457/*
 5458 * Consumers of these two interfaces, like for example the cpuidle menu
 5459 * governor, are using nonsensical data. Preferring shallow idle state selection
 5460 * for a CPU that has IO-wait which might not even end up running the task when
 5461 * it does become runnable.
 5462 */
 5463
 5464unsigned int nr_iowait_cpu(int cpu)
 5465{
 5466	return atomic_read(&cpu_rq(cpu)->nr_iowait);
 5467}
 5468
 5469/*
 5470 * IO-wait accounting, and how it's mostly bollocks (on SMP).
 5471 *
 5472 * The idea behind IO-wait account is to account the idle time that we could
 5473 * have spend running if it were not for IO. That is, if we were to improve the
 5474 * storage performance, we'd have a proportional reduction in IO-wait time.
 5475 *
 5476 * This all works nicely on UP, where, when a task blocks on IO, we account
 5477 * idle time as IO-wait, because if the storage were faster, it could've been
 5478 * running and we'd not be idle.
 5479 *
 5480 * This has been extended to SMP, by doing the same for each CPU. This however
 5481 * is broken.
 5482 *
 5483 * Imagine for instance the case where two tasks block on one CPU, only the one
 5484 * CPU will have IO-wait accounted, while the other has regular idle. Even
 5485 * though, if the storage were faster, both could've ran at the same time,
 5486 * utilising both CPUs.
 5487 *
 5488 * This means, that when looking globally, the current IO-wait accounting on
 5489 * SMP is a lower bound, by reason of under accounting.
 5490 *
 5491 * Worse, since the numbers are provided per CPU, they are sometimes
 5492 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
 5493 * associated with any one particular CPU, it can wake to another CPU than it
 5494 * blocked on. This means the per CPU IO-wait number is meaningless.
 5495 *
 5496 * Task CPU affinities can make all that even more 'interesting'.
 5497 */
 5498
 5499unsigned int nr_iowait(void)
 5500{
 5501	unsigned int i, sum = 0;
 5502
 5503	for_each_possible_cpu(i)
 5504		sum += nr_iowait_cpu(i);
 5505
 5506	return sum;
 5507}
 5508
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 5509#ifdef CONFIG_SMP
 5510
 5511/*
 5512 * sched_exec - execve() is a valuable balancing opportunity, because at
 5513 * this point the task has the smallest effective memory and cache footprint.
 5514 */
 5515void sched_exec(void)
 5516{
 5517	struct task_struct *p = current;
 5518	struct migration_arg arg;
 5519	int dest_cpu;
 5520
 5521	scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
 5522		dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
 5523		if (dest_cpu == smp_processor_id())
 5524			return;
 5525
 5526		if (unlikely(!cpu_active(dest_cpu)))
 5527			return;
 5528
 5529		arg = (struct migration_arg){ p, dest_cpu };
 
 
 5530	}
 5531	stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
 
 5532}
 5533
 5534#endif
 5535
 5536DEFINE_PER_CPU(struct kernel_stat, kstat);
 5537DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
 5538
 5539EXPORT_PER_CPU_SYMBOL(kstat);
 5540EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
 5541
 5542/*
 5543 * The function fair_sched_class.update_curr accesses the struct curr
 5544 * and its field curr->exec_start; when called from task_sched_runtime(),
 5545 * we observe a high rate of cache misses in practice.
 5546 * Prefetching this data results in improved performance.
 5547 */
 5548static inline void prefetch_curr_exec_start(struct task_struct *p)
 5549{
 5550#ifdef CONFIG_FAIR_GROUP_SCHED
 5551	struct sched_entity *curr = (&p->se)->cfs_rq->curr;
 5552#else
 5553	struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
 5554#endif
 5555	prefetch(curr);
 5556	prefetch(&curr->exec_start);
 5557}
 5558
 5559/*
 5560 * Return accounted runtime for the task.
 5561 * In case the task is currently running, return the runtime plus current's
 5562 * pending runtime that have not been accounted yet.
 5563 */
 5564unsigned long long task_sched_runtime(struct task_struct *p)
 5565{
 5566	struct rq_flags rf;
 5567	struct rq *rq;
 5568	u64 ns;
 5569
 5570#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
 5571	/*
 5572	 * 64-bit doesn't need locks to atomically read a 64-bit value.
 5573	 * So we have a optimization chance when the task's delta_exec is 0.
 5574	 * Reading ->on_cpu is racy, but this is ok.
 5575	 *
 5576	 * If we race with it leaving CPU, we'll take a lock. So we're correct.
 5577	 * If we race with it entering CPU, unaccounted time is 0. This is
 5578	 * indistinguishable from the read occurring a few cycles earlier.
 5579	 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
 5580	 * been accounted, so we're correct here as well.
 5581	 */
 5582	if (!p->on_cpu || !task_on_rq_queued(p))
 5583		return p->se.sum_exec_runtime;
 5584#endif
 5585
 5586	rq = task_rq_lock(p, &rf);
 5587	/*
 5588	 * Must be ->curr _and_ ->on_rq.  If dequeued, we would
 5589	 * project cycles that may never be accounted to this
 5590	 * thread, breaking clock_gettime().
 5591	 */
 5592	if (task_current(rq, p) && task_on_rq_queued(p)) {
 5593		prefetch_curr_exec_start(p);
 5594		update_rq_clock(rq);
 5595		p->sched_class->update_curr(rq);
 5596	}
 5597	ns = p->se.sum_exec_runtime;
 5598	task_rq_unlock(rq, p, &rf);
 5599
 5600	return ns;
 5601}
 5602
 5603#ifdef CONFIG_SCHED_DEBUG
 5604static u64 cpu_resched_latency(struct rq *rq)
 5605{
 5606	int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
 5607	u64 resched_latency, now = rq_clock(rq);
 5608	static bool warned_once;
 5609
 5610	if (sysctl_resched_latency_warn_once && warned_once)
 5611		return 0;
 5612
 5613	if (!need_resched() || !latency_warn_ms)
 5614		return 0;
 5615
 5616	if (system_state == SYSTEM_BOOTING)
 5617		return 0;
 5618
 5619	if (!rq->last_seen_need_resched_ns) {
 5620		rq->last_seen_need_resched_ns = now;
 5621		rq->ticks_without_resched = 0;
 5622		return 0;
 5623	}
 5624
 5625	rq->ticks_without_resched++;
 5626	resched_latency = now - rq->last_seen_need_resched_ns;
 5627	if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
 5628		return 0;
 5629
 5630	warned_once = true;
 5631
 5632	return resched_latency;
 5633}
 5634
 5635static int __init setup_resched_latency_warn_ms(char *str)
 5636{
 5637	long val;
 5638
 5639	if ((kstrtol(str, 0, &val))) {
 5640		pr_warn("Unable to set resched_latency_warn_ms\n");
 5641		return 1;
 5642	}
 5643
 5644	sysctl_resched_latency_warn_ms = val;
 5645	return 1;
 5646}
 5647__setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
 5648#else
 5649static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
 5650#endif /* CONFIG_SCHED_DEBUG */
 5651
 5652/*
 5653 * This function gets called by the timer code, with HZ frequency.
 5654 * We call it with interrupts disabled.
 5655 */
 5656void scheduler_tick(void)
 5657{
 5658	int cpu = smp_processor_id();
 5659	struct rq *rq = cpu_rq(cpu);
 5660	struct task_struct *curr = rq->curr;
 5661	struct rq_flags rf;
 5662	unsigned long thermal_pressure;
 5663	u64 resched_latency;
 5664
 5665	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
 5666		arch_scale_freq_tick();
 5667
 5668	sched_clock_tick();
 5669
 5670	rq_lock(rq, &rf);
 5671
 5672	update_rq_clock(rq);
 5673	thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
 5674	update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
 5675	curr->sched_class->task_tick(rq, curr, 0);
 5676	if (sched_feat(LATENCY_WARN))
 5677		resched_latency = cpu_resched_latency(rq);
 5678	calc_global_load_tick(rq);
 5679	sched_core_tick(rq);
 5680	task_tick_mm_cid(rq, curr);
 5681
 5682	rq_unlock(rq, &rf);
 5683
 5684	if (sched_feat(LATENCY_WARN) && resched_latency)
 5685		resched_latency_warn(cpu, resched_latency);
 5686
 5687	perf_event_task_tick();
 5688
 5689	if (curr->flags & PF_WQ_WORKER)
 5690		wq_worker_tick(curr);
 5691
 5692#ifdef CONFIG_SMP
 5693	rq->idle_balance = idle_cpu(cpu);
 5694	trigger_load_balance(rq);
 5695#endif
 5696}
 5697
 5698#ifdef CONFIG_NO_HZ_FULL
 5699
 5700struct tick_work {
 5701	int			cpu;
 5702	atomic_t		state;
 5703	struct delayed_work	work;
 5704};
 5705/* Values for ->state, see diagram below. */
 5706#define TICK_SCHED_REMOTE_OFFLINE	0
 5707#define TICK_SCHED_REMOTE_OFFLINING	1
 5708#define TICK_SCHED_REMOTE_RUNNING	2
 5709
 5710/*
 5711 * State diagram for ->state:
 5712 *
 5713 *
 5714 *          TICK_SCHED_REMOTE_OFFLINE
 5715 *                    |   ^
 5716 *                    |   |
 5717 *                    |   | sched_tick_remote()
 5718 *                    |   |
 5719 *                    |   |
 5720 *                    +--TICK_SCHED_REMOTE_OFFLINING
 5721 *                    |   ^
 5722 *                    |   |
 5723 * sched_tick_start() |   | sched_tick_stop()
 5724 *                    |   |
 5725 *                    V   |
 5726 *          TICK_SCHED_REMOTE_RUNNING
 5727 *
 5728 *
 5729 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
 5730 * and sched_tick_start() are happy to leave the state in RUNNING.
 5731 */
 5732
 5733static struct tick_work __percpu *tick_work_cpu;
 5734
 5735static void sched_tick_remote(struct work_struct *work)
 5736{
 5737	struct delayed_work *dwork = to_delayed_work(work);
 5738	struct tick_work *twork = container_of(dwork, struct tick_work, work);
 5739	int cpu = twork->cpu;
 5740	struct rq *rq = cpu_rq(cpu);
 5741	int os;
 5742
 5743	/*
 5744	 * Handle the tick only if it appears the remote CPU is running in full
 5745	 * dynticks mode. The check is racy by nature, but missing a tick or
 5746	 * having one too much is no big deal because the scheduler tick updates
 5747	 * statistics and checks timeslices in a time-independent way, regardless
 5748	 * of when exactly it is running.
 5749	 */
 5750	if (tick_nohz_tick_stopped_cpu(cpu)) {
 5751		guard(rq_lock_irq)(rq);
 5752		struct task_struct *curr = rq->curr;
 5753
 5754		if (cpu_online(cpu)) {
 5755			update_rq_clock(rq);
 5756
 5757			if (!is_idle_task(curr)) {
 5758				/*
 5759				 * Make sure the next tick runs within a
 5760				 * reasonable amount of time.
 5761				 */
 5762				u64 delta = rq_clock_task(rq) - curr->se.exec_start;
 5763				WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
 5764			}
 5765			curr->sched_class->task_tick(rq, curr, 0);
 5766
 5767			calc_load_nohz_remote(rq);
 5768		}
 
 
 
 
 
 5769	}
 5770
 5771	/*
 5772	 * Run the remote tick once per second (1Hz). This arbitrary
 5773	 * frequency is large enough to avoid overload but short enough
 5774	 * to keep scheduler internal stats reasonably up to date.  But
 5775	 * first update state to reflect hotplug activity if required.
 5776	 */
 5777	os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
 5778	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
 5779	if (os == TICK_SCHED_REMOTE_RUNNING)
 5780		queue_delayed_work(system_unbound_wq, dwork, HZ);
 5781}
 5782
 5783static void sched_tick_start(int cpu)
 5784{
 5785	int os;
 5786	struct tick_work *twork;
 5787
 5788	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
 5789		return;
 5790
 5791	WARN_ON_ONCE(!tick_work_cpu);
 5792
 5793	twork = per_cpu_ptr(tick_work_cpu, cpu);
 5794	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
 5795	WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
 5796	if (os == TICK_SCHED_REMOTE_OFFLINE) {
 5797		twork->cpu = cpu;
 5798		INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
 5799		queue_delayed_work(system_unbound_wq, &twork->work, HZ);
 5800	}
 5801}
 5802
 5803#ifdef CONFIG_HOTPLUG_CPU
 5804static void sched_tick_stop(int cpu)
 5805{
 5806	struct tick_work *twork;
 5807	int os;
 5808
 5809	if (housekeeping_cpu(cpu, HK_TYPE_TICK))
 5810		return;
 5811
 5812	WARN_ON_ONCE(!tick_work_cpu);
 5813
 5814	twork = per_cpu_ptr(tick_work_cpu, cpu);
 5815	/* There cannot be competing actions, but don't rely on stop-machine. */
 5816	os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
 5817	WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
 5818	/* Don't cancel, as this would mess up the state machine. */
 5819}
 5820#endif /* CONFIG_HOTPLUG_CPU */
 5821
 5822int __init sched_tick_offload_init(void)
 5823{
 5824	tick_work_cpu = alloc_percpu(struct tick_work);
 5825	BUG_ON(!tick_work_cpu);
 
 5826	return 0;
 5827}
 5828
 5829#else /* !CONFIG_NO_HZ_FULL */
 5830static inline void sched_tick_start(int cpu) { }
 5831static inline void sched_tick_stop(int cpu) { }
 5832#endif
 5833
 5834#if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
 5835				defined(CONFIG_TRACE_PREEMPT_TOGGLE))
 5836/*
 5837 * If the value passed in is equal to the current preempt count
 5838 * then we just disabled preemption. Start timing the latency.
 5839 */
 5840static inline void preempt_latency_start(int val)
 5841{
 5842	if (preempt_count() == val) {
 5843		unsigned long ip = get_lock_parent_ip();
 5844#ifdef CONFIG_DEBUG_PREEMPT
 5845		current->preempt_disable_ip = ip;
 5846#endif
 5847		trace_preempt_off(CALLER_ADDR0, ip);
 5848	}
 5849}
 5850
 5851void preempt_count_add(int val)
 5852{
 5853#ifdef CONFIG_DEBUG_PREEMPT
 5854	/*
 5855	 * Underflow?
 5856	 */
 5857	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
 5858		return;
 5859#endif
 5860	__preempt_count_add(val);
 5861#ifdef CONFIG_DEBUG_PREEMPT
 5862	/*
 5863	 * Spinlock count overflowing soon?
 5864	 */
 5865	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
 5866				PREEMPT_MASK - 10);
 5867#endif
 5868	preempt_latency_start(val);
 5869}
 5870EXPORT_SYMBOL(preempt_count_add);
 5871NOKPROBE_SYMBOL(preempt_count_add);
 5872
 5873/*
 5874 * If the value passed in equals to the current preempt count
 5875 * then we just enabled preemption. Stop timing the latency.
 5876 */
 5877static inline void preempt_latency_stop(int val)
 5878{
 5879	if (preempt_count() == val)
 5880		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
 5881}
 5882
 5883void preempt_count_sub(int val)
 5884{
 5885#ifdef CONFIG_DEBUG_PREEMPT
 5886	/*
 5887	 * Underflow?
 5888	 */
 5889	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
 5890		return;
 5891	/*
 5892	 * Is the spinlock portion underflowing?
 5893	 */
 5894	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
 5895			!(preempt_count() & PREEMPT_MASK)))
 5896		return;
 5897#endif
 5898
 5899	preempt_latency_stop(val);
 5900	__preempt_count_sub(val);
 5901}
 5902EXPORT_SYMBOL(preempt_count_sub);
 5903NOKPROBE_SYMBOL(preempt_count_sub);
 5904
 5905#else
 5906static inline void preempt_latency_start(int val) { }
 5907static inline void preempt_latency_stop(int val) { }
 5908#endif
 5909
 5910static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
 5911{
 5912#ifdef CONFIG_DEBUG_PREEMPT
 5913	return p->preempt_disable_ip;
 5914#else
 5915	return 0;
 5916#endif
 5917}
 5918
 5919/*
 5920 * Print scheduling while atomic bug:
 5921 */
 5922static noinline void __schedule_bug(struct task_struct *prev)
 5923{
 5924	/* Save this before calling printk(), since that will clobber it */
 5925	unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
 5926
 5927	if (oops_in_progress)
 5928		return;
 5929
 5930	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
 5931		prev->comm, prev->pid, preempt_count());
 5932
 5933	debug_show_held_locks(prev);
 5934	print_modules();
 5935	if (irqs_disabled())
 5936		print_irqtrace_events(prev);
 5937	if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
 
 5938		pr_err("Preemption disabled at:");
 5939		print_ip_sym(KERN_ERR, preempt_disable_ip);
 
 5940	}
 5941	check_panic_on_warn("scheduling while atomic");
 
 5942
 5943	dump_stack();
 5944	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 5945}
 5946
 5947/*
 5948 * Various schedule()-time debugging checks and statistics:
 5949 */
 5950static inline void schedule_debug(struct task_struct *prev, bool preempt)
 5951{
 5952#ifdef CONFIG_SCHED_STACK_END_CHECK
 5953	if (task_stack_end_corrupted(prev))
 5954		panic("corrupted stack end detected inside scheduler\n");
 5955
 5956	if (task_scs_end_corrupted(prev))
 5957		panic("corrupted shadow stack detected inside scheduler\n");
 5958#endif
 5959
 5960#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
 5961	if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
 5962		printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
 5963			prev->comm, prev->pid, prev->non_block_count);
 5964		dump_stack();
 5965		add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
 5966	}
 5967#endif
 5968
 5969	if (unlikely(in_atomic_preempt_off())) {
 5970		__schedule_bug(prev);
 5971		preempt_count_set(PREEMPT_DISABLED);
 5972	}
 5973	rcu_sleep_check();
 5974	SCHED_WARN_ON(ct_state() == CONTEXT_USER);
 5975
 5976	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
 5977
 5978	schedstat_inc(this_rq()->sched_count);
 5979}
 5980
 5981static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
 5982				  struct rq_flags *rf)
 5983{
 5984#ifdef CONFIG_SMP
 5985	const struct sched_class *class;
 5986	/*
 5987	 * We must do the balancing pass before put_prev_task(), such
 5988	 * that when we release the rq->lock the task is in the same
 5989	 * state as before we took rq->lock.
 5990	 *
 5991	 * We can terminate the balance pass as soon as we know there is
 5992	 * a runnable task of @class priority or higher.
 5993	 */
 5994	for_class_range(class, prev->sched_class, &idle_sched_class) {
 5995		if (class->balance(rq, prev, rf))
 5996			break;
 5997	}
 5998#endif
 5999
 6000	put_prev_task(rq, prev);
 6001}
 6002
 6003/*
 6004 * Pick up the highest-prio task:
 6005 */
 6006static inline struct task_struct *
 6007__pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
 6008{
 6009	const struct sched_class *class;
 6010	struct task_struct *p;
 6011
 6012	/*
 6013	 * Optimization: we know that if all tasks are in the fair class we can
 6014	 * call that function directly, but only if the @prev task wasn't of a
 6015	 * higher scheduling class, because otherwise those lose the
 6016	 * opportunity to pull in more work from other CPUs.
 6017	 */
 6018	if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
 
 6019		   rq->nr_running == rq->cfs.h_nr_running)) {
 6020
 6021		p = pick_next_task_fair(rq, prev, rf);
 6022		if (unlikely(p == RETRY_TASK))
 6023			goto restart;
 6024
 6025		/* Assume the next prioritized class is idle_sched_class */
 6026		if (!p) {
 6027			put_prev_task(rq, prev);
 6028			p = pick_next_task_idle(rq);
 6029		}
 6030
 6031		/*
 6032		 * This is the fast path; it cannot be a DL server pick;
 6033		 * therefore even if @p == @prev, ->dl_server must be NULL.
 6034		 */
 6035		if (p->dl_server)
 6036			p->dl_server = NULL;
 6037
 6038		return p;
 6039	}
 6040
 6041restart:
 6042	put_prev_task_balance(rq, prev, rf);
 6043
 6044	/*
 6045	 * We've updated @prev and no longer need the server link, clear it.
 6046	 * Must be done before ->pick_next_task() because that can (re)set
 6047	 * ->dl_server.
 6048	 */
 6049	if (prev->dl_server)
 6050		prev->dl_server = NULL;
 6051
 6052	for_each_class(class) {
 6053		p = class->pick_next_task(rq);
 6054		if (p)
 6055			return p;
 6056	}
 6057
 6058	BUG(); /* The idle class should always have a runnable task. */
 6059}
 6060
 6061#ifdef CONFIG_SCHED_CORE
 6062static inline bool is_task_rq_idle(struct task_struct *t)
 6063{
 6064	return (task_rq(t)->idle == t);
 6065}
 6066
 6067static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
 6068{
 6069	return is_task_rq_idle(a) || (a->core_cookie == cookie);
 6070}
 6071
 6072static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
 6073{
 6074	if (is_task_rq_idle(a) || is_task_rq_idle(b))
 6075		return true;
 6076
 6077	return a->core_cookie == b->core_cookie;
 6078}
 6079
 6080static inline struct task_struct *pick_task(struct rq *rq)
 6081{
 6082	const struct sched_class *class;
 6083	struct task_struct *p;
 6084
 6085	for_each_class(class) {
 6086		p = class->pick_task(rq);
 6087		if (p)
 
 
 6088			return p;
 6089	}
 6090
 6091	BUG(); /* The idle class should always have a runnable task. */
 6092}
 6093
 6094extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
 6095
 6096static void queue_core_balance(struct rq *rq);
 6097
 6098static struct task_struct *
 6099pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
 6100{
 6101	struct task_struct *next, *p, *max = NULL;
 6102	const struct cpumask *smt_mask;
 6103	bool fi_before = false;
 6104	bool core_clock_updated = (rq == rq->core);
 6105	unsigned long cookie;
 6106	int i, cpu, occ = 0;
 6107	struct rq *rq_i;
 6108	bool need_sync;
 6109
 6110	if (!sched_core_enabled(rq))
 6111		return __pick_next_task(rq, prev, rf);
 6112
 6113	cpu = cpu_of(rq);
 6114
 6115	/* Stopper task is switching into idle, no need core-wide selection. */
 6116	if (cpu_is_offline(cpu)) {
 6117		/*
 6118		 * Reset core_pick so that we don't enter the fastpath when
 6119		 * coming online. core_pick would already be migrated to
 6120		 * another cpu during offline.
 6121		 */
 6122		rq->core_pick = 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		if (next != prev) {
 6142			put_prev_task(rq, prev);
 6143			set_next_task(rq, next);
 6144		}
 6145
 6146		rq->core_pick = NULL;
 6147		goto out;
 6148	}
 6149
 6150	put_prev_task_balance(rq, prev, rf);
 6151
 6152	smt_mask = cpu_smt_mask(cpu);
 6153	need_sync = !!rq->core->core_cookie;
 6154
 6155	/* reset state */
 6156	rq->core->core_cookie = 0UL;
 6157	if (rq->core->core_forceidle_count) {
 6158		if (!core_clock_updated) {
 6159			update_rq_clock(rq->core);
 6160			core_clock_updated = true;
 6161		}
 6162		sched_core_account_forceidle(rq);
 6163		/* reset after accounting force idle */
 6164		rq->core->core_forceidle_start = 0;
 6165		rq->core->core_forceidle_count = 0;
 6166		rq->core->core_forceidle_occupation = 0;
 6167		need_sync = true;
 6168		fi_before = true;
 6169	}
 6170
 6171	/*
 6172	 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
 6173	 *
 6174	 * @task_seq guards the task state ({en,de}queues)
 6175	 * @pick_seq is the @task_seq we did a selection on
 6176	 * @sched_seq is the @pick_seq we scheduled
 6177	 *
 6178	 * However, preemptions can cause multiple picks on the same task set.
 6179	 * 'Fix' this by also increasing @task_seq for every pick.
 6180	 */
 6181	rq->core->core_task_seq++;
 6182
 6183	/*
 6184	 * Optimize for common case where this CPU has no cookies
 6185	 * and there are no cookied tasks running on siblings.
 6186	 */
 6187	if (!need_sync) {
 6188		next = pick_task(rq);
 6189		if (!next->core_cookie) {
 6190			rq->core_pick = NULL;
 6191			/*
 6192			 * For robustness, update the min_vruntime_fi for
 6193			 * unconstrained picks as well.
 6194			 */
 6195			WARN_ON_ONCE(fi_before);
 6196			task_vruntime_update(rq, next, false);
 6197			goto out_set_next;
 6198		}
 6199	}
 6200
 6201	/*
 6202	 * For each thread: do the regular task pick and find the max prio task
 6203	 * amongst them.
 6204	 *
 6205	 * Tie-break prio towards the current CPU
 6206	 */
 6207	for_each_cpu_wrap(i, smt_mask, cpu) {
 6208		rq_i = cpu_rq(i);
 6209
 6210		/*
 6211		 * Current cpu always has its clock updated on entrance to
 6212		 * pick_next_task(). If the current cpu is not the core,
 6213		 * the core may also have been updated above.
 6214		 */
 6215		if (i != cpu && (rq_i != rq->core || !core_clock_updated))
 6216			update_rq_clock(rq_i);
 6217
 6218		p = rq_i->core_pick = pick_task(rq_i);
 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
 6243		if (p == rq_i->idle) {
 6244			if (rq_i->nr_running) {
 6245				rq->core->core_forceidle_count++;
 6246				if (!fi_before)
 6247					rq->core->core_forceidle_seq++;
 6248			}
 6249		} else {
 6250			occ++;
 6251		}
 6252	}
 6253
 6254	if (schedstat_enabled() && rq->core->core_forceidle_count) {
 6255		rq->core->core_forceidle_start = rq_clock(rq->core);
 6256		rq->core->core_forceidle_occupation = occ;
 6257	}
 6258
 6259	rq->core->core_pick_seq = rq->core->core_task_seq;
 6260	next = rq->core_pick;
 6261	rq->core_sched_seq = rq->core->core_pick_seq;
 6262
 6263	/* Something should have been selected for current CPU */
 6264	WARN_ON_ONCE(!next);
 6265
 6266	/*
 6267	 * Reschedule siblings
 6268	 *
 6269	 * NOTE: L1TF -- at this point we're no longer running the old task and
 6270	 * sending an IPI (below) ensures the sibling will no longer be running
 6271	 * their task. This ensures there is no inter-sibling overlap between
 6272	 * non-matching user state.
 6273	 */
 6274	for_each_cpu(i, smt_mask) {
 6275		rq_i = cpu_rq(i);
 6276
 6277		/*
 6278		 * An online sibling might have gone offline before a task
 6279		 * could be picked for it, or it might be offline but later
 6280		 * happen to come online, but its too late and nothing was
 6281		 * picked for it.  That's Ok - it will pick tasks for itself,
 6282		 * so ignore it.
 6283		 */
 6284		if (!rq_i->core_pick)
 6285			continue;
 6286
 6287		/*
 6288		 * Update for new !FI->FI transitions, or if continuing to be in !FI:
 6289		 * fi_before     fi      update?
 6290		 *  0            0       1
 6291		 *  0            1       1
 6292		 *  1            0       1
 6293		 *  1            1       0
 6294		 */
 6295		if (!(fi_before && rq->core->core_forceidle_count))
 6296			task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
 6297
 6298		rq_i->core_pick->core_occupation = occ;
 6299
 6300		if (i == cpu) {
 6301			rq_i->core_pick = NULL;
 6302			continue;
 6303		}
 6304
 6305		/* Did we break L1TF mitigation requirements? */
 6306		WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
 6307
 6308		if (rq_i->curr == rq_i->core_pick) {
 6309			rq_i->core_pick = NULL;
 6310			continue;
 6311		}
 6312
 6313		resched_curr(rq_i);
 6314	}
 6315
 6316out_set_next:
 6317	set_next_task(rq, next);
 6318out:
 6319	if (rq->core->core_forceidle_count && next == rq->idle)
 6320		queue_core_balance(rq);
 6321
 6322	return next;
 6323}
 6324
 6325static bool try_steal_cookie(int this, int that)
 6326{
 6327	struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
 6328	struct task_struct *p;
 6329	unsigned long cookie;
 6330	bool success = false;
 6331
 6332	guard(irq)();
 6333	guard(double_rq_lock)(dst, src);
 6334
 6335	cookie = dst->core->core_cookie;
 6336	if (!cookie)
 6337		return false;
 6338
 6339	if (dst->curr != dst->idle)
 6340		return false;
 6341
 6342	p = sched_core_find(src, cookie);
 6343	if (!p)
 6344		return false;
 6345
 6346	do {
 6347		if (p == src->core_pick || p == src->curr)
 6348			goto next;
 6349
 6350		if (!is_cpu_allowed(p, this))
 6351			goto next;
 6352
 6353		if (p->core_occupation > dst->idle->core_occupation)
 6354			goto next;
 6355		/*
 6356		 * sched_core_find() and sched_core_next() will ensure
 6357		 * that task @p is not throttled now, we also need to
 6358		 * check whether the runqueue of the destination CPU is
 6359		 * being throttled.
 6360		 */
 6361		if (sched_task_is_throttled(p, this))
 6362			goto next;
 6363
 6364		deactivate_task(src, p, 0);
 6365		set_task_cpu(p, this);
 6366		activate_task(dst, p, 0);
 6367
 6368		resched_curr(dst);
 6369
 6370		success = true;
 6371		break;
 6372
 6373next:
 6374		p = sched_core_next(p, cookie);
 6375	} while (p);
 6376
 6377	return success;
 6378}
 6379
 6380static bool steal_cookie_task(int cpu, struct sched_domain *sd)
 6381{
 6382	int i;
 6383
 6384	for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) {
 6385		if (i == cpu)
 6386			continue;
 6387
 6388		if (need_resched())
 6389			break;
 6390
 6391		if (try_steal_cookie(cpu, i))
 6392			return true;
 6393	}
 6394
 6395	return false;
 6396}
 6397
 6398static void sched_core_balance(struct rq *rq)
 6399{
 6400	struct sched_domain *sd;
 6401	int cpu = cpu_of(rq);
 6402
 6403	guard(preempt)();
 6404	guard(rcu)();
 6405
 6406	raw_spin_rq_unlock_irq(rq);
 6407	for_each_domain(cpu, sd) {
 6408		if (need_resched())
 6409			break;
 6410
 6411		if (steal_cookie_task(cpu, sd))
 6412			break;
 6413	}
 6414	raw_spin_rq_lock_irq(rq);
 6415}
 6416
 6417static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
 6418
 6419static void queue_core_balance(struct rq *rq)
 6420{
 6421	if (!sched_core_enabled(rq))
 6422		return;
 6423
 6424	if (!rq->core->core_cookie)
 6425		return;
 6426
 6427	if (!rq->nr_running) /* not forced idle */
 6428		return;
 6429
 6430	queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
 6431}
 6432
 6433DEFINE_LOCK_GUARD_1(core_lock, int,
 6434		    sched_core_lock(*_T->lock, &_T->flags),
 6435		    sched_core_unlock(*_T->lock, &_T->flags),
 6436		    unsigned long flags)
 6437
 6438static void sched_core_cpu_starting(unsigned int cpu)
 6439{
 6440	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
 6441	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
 6442	int t;
 6443
 6444	guard(core_lock)(&cpu);
 6445
 6446	WARN_ON_ONCE(rq->core != rq);
 6447
 6448	/* if we're the first, we'll be our own leader */
 6449	if (cpumask_weight(smt_mask) == 1)
 6450		return;
 6451
 6452	/* find the leader */
 6453	for_each_cpu(t, smt_mask) {
 6454		if (t == cpu)
 6455			continue;
 6456		rq = cpu_rq(t);
 6457		if (rq->core == rq) {
 6458			core_rq = rq;
 6459			break;
 6460		}
 6461	}
 6462
 6463	if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
 6464		return;
 6465
 6466	/* install and validate core_rq */
 6467	for_each_cpu(t, smt_mask) {
 6468		rq = cpu_rq(t);
 6469
 6470		if (t == cpu)
 6471			rq->core = core_rq;
 6472
 6473		WARN_ON_ONCE(rq->core != core_rq);
 6474	}
 6475}
 6476
 6477static void sched_core_cpu_deactivate(unsigned int cpu)
 6478{
 6479	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
 6480	struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
 6481	int t;
 6482
 6483	guard(core_lock)(&cpu);
 6484
 6485	/* if we're the last man standing, nothing to do */
 6486	if (cpumask_weight(smt_mask) == 1) {
 6487		WARN_ON_ONCE(rq->core != rq);
 6488		return;
 6489	}
 6490
 6491	/* if we're not the leader, nothing to do */
 6492	if (rq->core != rq)
 6493		return;
 6494
 6495	/* find a new leader */
 6496	for_each_cpu(t, smt_mask) {
 6497		if (t == cpu)
 6498			continue;
 6499		core_rq = cpu_rq(t);
 6500		break;
 6501	}
 6502
 6503	if (WARN_ON_ONCE(!core_rq)) /* impossible */
 6504		return;
 6505
 6506	/* copy the shared state to the new leader */
 6507	core_rq->core_task_seq             = rq->core_task_seq;
 6508	core_rq->core_pick_seq             = rq->core_pick_seq;
 6509	core_rq->core_cookie               = rq->core_cookie;
 6510	core_rq->core_forceidle_count      = rq->core_forceidle_count;
 6511	core_rq->core_forceidle_seq        = rq->core_forceidle_seq;
 6512	core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
 6513
 6514	/*
 6515	 * Accounting edge for forced idle is handled in pick_next_task().
 6516	 * Don't need another one here, since the hotplug thread shouldn't
 6517	 * have a cookie.
 6518	 */
 6519	core_rq->core_forceidle_start = 0;
 6520
 6521	/* install new leader */
 6522	for_each_cpu(t, smt_mask) {
 6523		rq = cpu_rq(t);
 6524		rq->core = core_rq;
 6525	}
 6526}
 6527
 6528static inline void sched_core_cpu_dying(unsigned int cpu)
 6529{
 6530	struct rq *rq = cpu_rq(cpu);
 6531
 6532	if (rq->core != rq)
 6533		rq->core = rq;
 6534}
 6535
 6536#else /* !CONFIG_SCHED_CORE */
 6537
 6538static inline void sched_core_cpu_starting(unsigned int cpu) {}
 6539static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
 6540static inline void sched_core_cpu_dying(unsigned int cpu) {}
 6541
 6542static struct task_struct *
 6543pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
 6544{
 6545	return __pick_next_task(rq, prev, rf);
 6546}
 6547
 6548#endif /* CONFIG_SCHED_CORE */
 6549
 6550/*
 6551 * Constants for the sched_mode argument of __schedule().
 6552 *
 6553 * The mode argument allows RT enabled kernels to differentiate a
 6554 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
 6555 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
 6556 * optimize the AND operation out and just check for zero.
 6557 */
 6558#define SM_NONE			0x0
 6559#define SM_PREEMPT		0x1
 6560#define SM_RTLOCK_WAIT		0x2
 6561
 6562#ifndef CONFIG_PREEMPT_RT
 6563# define SM_MASK_PREEMPT	(~0U)
 6564#else
 6565# define SM_MASK_PREEMPT	SM_PREEMPT
 6566#endif
 6567
 6568/*
 6569 * __schedule() is the main scheduler function.
 6570 *
 6571 * The main means of driving the scheduler and thus entering this function are:
 6572 *
 6573 *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
 6574 *
 6575 *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
 6576 *      paths. For example, see arch/x86/entry_64.S.
 6577 *
 6578 *      To drive preemption between tasks, the scheduler sets the flag in timer
 6579 *      interrupt handler scheduler_tick().
 6580 *
 6581 *   3. Wakeups don't really cause entry into schedule(). They add a
 6582 *      task to the run-queue and that's it.
 6583 *
 6584 *      Now, if the new task added to the run-queue preempts the current
 6585 *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
 6586 *      called on the nearest possible occasion:
 6587 *
 6588 *       - If the kernel is preemptible (CONFIG_PREEMPTION=y):
 6589 *
 6590 *         - in syscall or exception context, at the next outmost
 6591 *           preempt_enable(). (this might be as soon as the wake_up()'s
 6592 *           spin_unlock()!)
 6593 *
 6594 *         - in IRQ context, return from interrupt-handler to
 6595 *           preemptible context
 6596 *
 6597 *       - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
 6598 *         then at the next:
 6599 *
 6600 *          - cond_resched() call
 6601 *          - explicit schedule() call
 6602 *          - return from syscall or exception to user-space
 6603 *          - return from interrupt-handler to user-space
 6604 *
 6605 * WARNING: must be called with preemption disabled!
 6606 */
 6607static void __sched notrace __schedule(unsigned int sched_mode)
 6608{
 6609	struct task_struct *prev, *next;
 6610	unsigned long *switch_count;
 6611	unsigned long prev_state;
 6612	struct rq_flags rf;
 6613	struct rq *rq;
 6614	int cpu;
 6615
 6616	cpu = smp_processor_id();
 6617	rq = cpu_rq(cpu);
 6618	prev = rq->curr;
 6619
 6620	schedule_debug(prev, !!sched_mode);
 6621
 6622	if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
 6623		hrtick_clear(rq);
 6624
 6625	local_irq_disable();
 6626	rcu_note_context_switch(!!sched_mode);
 6627
 6628	/*
 6629	 * Make sure that signal_pending_state()->signal_pending() below
 6630	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
 6631	 * done by the caller to avoid the race with signal_wake_up():
 6632	 *
 6633	 * __set_current_state(@state)		signal_wake_up()
 6634	 * schedule()				  set_tsk_thread_flag(p, TIF_SIGPENDING)
 6635	 *					  wake_up_state(p, state)
 6636	 *   LOCK rq->lock			    LOCK p->pi_state
 6637	 *   smp_mb__after_spinlock()		    smp_mb__after_spinlock()
 6638	 *     if (signal_pending_state())	    if (p->state & @state)
 6639	 *
 6640	 * Also, the membarrier system call requires a full memory barrier
 6641	 * after coming from user-space, before storing to rq->curr.
 6642	 */
 6643	rq_lock(rq, &rf);
 6644	smp_mb__after_spinlock();
 6645
 6646	/* Promote REQ to ACT */
 6647	rq->clock_update_flags <<= 1;
 6648	update_rq_clock(rq);
 6649	rq->clock_update_flags = RQCF_UPDATED;
 6650
 6651	switch_count = &prev->nivcsw;
 6652
 6653	/*
 6654	 * We must load prev->state once (task_struct::state is volatile), such
 6655	 * that we form a control dependency vs deactivate_task() below.
 6656	 */
 6657	prev_state = READ_ONCE(prev->__state);
 6658	if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
 6659		if (signal_pending_state(prev_state, prev)) {
 6660			WRITE_ONCE(prev->__state, TASK_RUNNING);
 6661		} else {
 6662			prev->sched_contributes_to_load =
 6663				(prev_state & TASK_UNINTERRUPTIBLE) &&
 6664				!(prev_state & TASK_NOLOAD) &&
 6665				!(prev_state & TASK_FROZEN);
 6666
 6667			if (prev->sched_contributes_to_load)
 6668				rq->nr_uninterruptible++;
 
 
 6669
 6670			/*
 6671			 * __schedule()			ttwu()
 6672			 *   prev_state = prev->state;    if (p->on_rq && ...)
 6673			 *   if (prev_state)		    goto out;
 6674			 *     p->on_rq = 0;		  smp_acquire__after_ctrl_dep();
 6675			 *				  p->state = TASK_WAKING
 6676			 *
 6677			 * Where __schedule() and ttwu() have matching control dependencies.
 6678			 *
 6679			 * After this, schedule() must not care about p->state any more.
 6680			 */
 6681			deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
 
 6682
 6683			if (prev->in_iowait) {
 6684				atomic_inc(&rq->nr_iowait);
 6685				delayacct_blkio_start();
 6686			}
 6687		}
 6688		switch_count = &prev->nvcsw;
 6689	}
 6690
 6691	next = pick_next_task(rq, prev, &rf);
 6692	clear_tsk_need_resched(prev);
 6693	clear_preempt_need_resched();
 6694#ifdef CONFIG_SCHED_DEBUG
 6695	rq->last_seen_need_resched_ns = 0;
 6696#endif
 6697
 6698	if (likely(prev != next)) {
 6699		rq->nr_switches++;
 6700		/*
 6701		 * RCU users of rcu_dereference(rq->curr) may not see
 6702		 * changes to task_struct made by pick_next_task().
 6703		 */
 6704		RCU_INIT_POINTER(rq->curr, next);
 6705		/*
 6706		 * The membarrier system call requires each architecture
 6707		 * to have a full memory barrier after updating
 6708		 * rq->curr, before returning to user-space.
 6709		 *
 6710		 * Here are the schemes providing that barrier on the
 6711		 * various architectures:
 6712		 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
 6713		 *   switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
 6714		 * - finish_lock_switch() for weakly-ordered
 6715		 *   architectures where spin_unlock is a full barrier,
 6716		 * - switch_to() for arm64 (weakly-ordered, spin_unlock
 6717		 *   is a RELEASE barrier),
 6718		 */
 6719		++*switch_count;
 6720
 6721		migrate_disable_switch(rq, prev);
 6722		psi_sched_switch(prev, next, !task_on_rq_queued(prev));
 6723
 6724		trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
 6725
 6726		/* Also unlocks the rq: */
 6727		rq = context_switch(rq, prev, next, &rf);
 6728	} else {
 6729		rq_unpin_lock(rq, &rf);
 6730		__balance_callbacks(rq);
 6731		raw_spin_rq_unlock_irq(rq);
 6732	}
 
 
 6733}
 6734
 6735void __noreturn do_task_dead(void)
 6736{
 6737	/* Causes final put_task_struct in finish_task_switch(): */
 6738	set_special_state(TASK_DEAD);
 6739
 6740	/* Tell freezer to ignore us: */
 6741	current->flags |= PF_NOFREEZE;
 6742
 6743	__schedule(SM_NONE);
 6744	BUG();
 6745
 6746	/* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
 6747	for (;;)
 6748		cpu_relax();
 6749}
 6750
 6751static inline void sched_submit_work(struct task_struct *tsk)
 6752{
 6753	static DEFINE_WAIT_OVERRIDE_MAP(sched_map, LD_WAIT_CONFIG);
 6754	unsigned int task_flags;
 6755
 6756	/*
 6757	 * Establish LD_WAIT_CONFIG context to ensure none of the code called
 6758	 * will use a blocking primitive -- which would lead to recursion.
 6759	 */
 6760	lock_map_acquire_try(&sched_map);
 6761
 6762	task_flags = tsk->flags;
 6763	/*
 6764	 * If a worker goes to sleep, notify and ask workqueue whether it
 6765	 * wants to wake up a task to maintain concurrency.
 6766	 */
 6767	if (task_flags & PF_WQ_WORKER)
 6768		wq_worker_sleeping(tsk);
 6769	else if (task_flags & PF_IO_WORKER)
 6770		io_wq_worker_sleeping(tsk);
 6771
 6772	/*
 6773	 * spinlock and rwlock must not flush block requests.  This will
 6774	 * deadlock if the callback attempts to acquire a lock which is
 6775	 * already acquired.
 6776	 */
 6777	SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
 6778
 6779	/*
 6780	 * If we are going to sleep and we have plugged IO queued,
 6781	 * make sure to submit it to avoid deadlocks.
 6782	 */
 6783	blk_flush_plug(tsk->plug, true);
 6784
 6785	lock_map_release(&sched_map);
 6786}
 6787
 6788static void sched_update_worker(struct task_struct *tsk)
 6789{
 6790	if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
 6791		if (tsk->flags & PF_WQ_WORKER)
 6792			wq_worker_running(tsk);
 6793		else
 6794			io_wq_worker_running(tsk);
 6795	}
 6796}
 6797
 6798static __always_inline void __schedule_loop(unsigned int sched_mode)
 6799{
 6800	do {
 6801		preempt_disable();
 6802		__schedule(sched_mode);
 6803		sched_preempt_enable_no_resched();
 6804	} while (need_resched());
 6805}
 6806
 6807asmlinkage __visible void __sched schedule(void)
 6808{
 6809	struct task_struct *tsk = current;
 6810
 6811#ifdef CONFIG_RT_MUTEXES
 6812	lockdep_assert(!tsk->sched_rt_mutex);
 6813#endif
 6814
 6815	if (!task_is_running(tsk))
 6816		sched_submit_work(tsk);
 6817	__schedule_loop(SM_NONE);
 6818	sched_update_worker(tsk);
 6819}
 6820EXPORT_SYMBOL(schedule);
 6821
 6822/*
 6823 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
 6824 * state (have scheduled out non-voluntarily) by making sure that all
 6825 * tasks have either left the run queue or have gone into user space.
 6826 * As idle tasks do not do either, they must not ever be preempted
 6827 * (schedule out non-voluntarily).
 6828 *
 6829 * schedule_idle() is similar to schedule_preempt_disable() except that it
 6830 * never enables preemption because it does not call sched_submit_work().
 6831 */
 6832void __sched schedule_idle(void)
 6833{
 6834	/*
 6835	 * As this skips calling sched_submit_work(), which the idle task does
 6836	 * regardless because that function is a nop when the task is in a
 6837	 * TASK_RUNNING state, make sure this isn't used someplace that the
 6838	 * current task can be in any other state. Note, idle is always in the
 6839	 * TASK_RUNNING state.
 6840	 */
 6841	WARN_ON_ONCE(current->__state);
 6842	do {
 6843		__schedule(SM_NONE);
 6844	} while (need_resched());
 6845}
 6846
 6847#if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
 6848asmlinkage __visible void __sched schedule_user(void)
 6849{
 6850	/*
 6851	 * If we come here after a random call to set_need_resched(),
 6852	 * or we have been woken up remotely but the IPI has not yet arrived,
 6853	 * we haven't yet exited the RCU idle mode. Do it here manually until
 6854	 * we find a better solution.
 6855	 *
 6856	 * NB: There are buggy callers of this function.  Ideally we
 6857	 * should warn if prev_state != CONTEXT_USER, but that will trigger
 6858	 * too frequently to make sense yet.
 6859	 */
 6860	enum ctx_state prev_state = exception_enter();
 6861	schedule();
 6862	exception_exit(prev_state);
 6863}
 6864#endif
 6865
 6866/**
 6867 * schedule_preempt_disabled - called with preemption disabled
 6868 *
 6869 * Returns with preemption disabled. Note: preempt_count must be 1
 6870 */
 6871void __sched schedule_preempt_disabled(void)
 6872{
 6873	sched_preempt_enable_no_resched();
 6874	schedule();
 6875	preempt_disable();
 6876}
 6877
 6878#ifdef CONFIG_PREEMPT_RT
 6879void __sched notrace schedule_rtlock(void)
 6880{
 6881	__schedule_loop(SM_RTLOCK_WAIT);
 6882}
 6883NOKPROBE_SYMBOL(schedule_rtlock);
 6884#endif
 6885
 6886static void __sched notrace preempt_schedule_common(void)
 6887{
 6888	do {
 6889		/*
 6890		 * Because the function tracer can trace preempt_count_sub()
 6891		 * and it also uses preempt_enable/disable_notrace(), if
 6892		 * NEED_RESCHED is set, the preempt_enable_notrace() called
 6893		 * by the function tracer will call this function again and
 6894		 * cause infinite recursion.
 6895		 *
 6896		 * Preemption must be disabled here before the function
 6897		 * tracer can trace. Break up preempt_disable() into two
 6898		 * calls. One to disable preemption without fear of being
 6899		 * traced. The other to still record the preemption latency,
 6900		 * which can also be traced by the function tracer.
 6901		 */
 6902		preempt_disable_notrace();
 6903		preempt_latency_start(1);
 6904		__schedule(SM_PREEMPT);
 6905		preempt_latency_stop(1);
 6906		preempt_enable_no_resched_notrace();
 6907
 6908		/*
 6909		 * Check again in case we missed a preemption opportunity
 6910		 * between schedule and now.
 6911		 */
 6912	} while (need_resched());
 6913}
 6914
 6915#ifdef CONFIG_PREEMPTION
 6916/*
 6917 * This is the entry point to schedule() from in-kernel preemption
 6918 * off of preempt_enable.
 
 6919 */
 6920asmlinkage __visible void __sched notrace preempt_schedule(void)
 6921{
 6922	/*
 6923	 * If there is a non-zero preempt_count or interrupts are disabled,
 6924	 * we do not want to preempt the current task. Just return..
 6925	 */
 6926	if (likely(!preemptible()))
 6927		return;
 
 6928	preempt_schedule_common();
 6929}
 6930NOKPROBE_SYMBOL(preempt_schedule);
 6931EXPORT_SYMBOL(preempt_schedule);
 6932
 6933#ifdef CONFIG_PREEMPT_DYNAMIC
 6934#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 6935#ifndef preempt_schedule_dynamic_enabled
 6936#define preempt_schedule_dynamic_enabled	preempt_schedule
 6937#define preempt_schedule_dynamic_disabled	NULL
 6938#endif
 6939DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
 6940EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
 6941#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 6942static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
 6943void __sched notrace dynamic_preempt_schedule(void)
 6944{
 6945	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
 6946		return;
 6947	preempt_schedule();
 6948}
 6949NOKPROBE_SYMBOL(dynamic_preempt_schedule);
 6950EXPORT_SYMBOL(dynamic_preempt_schedule);
 6951#endif
 6952#endif
 6953
 6954/**
 6955 * preempt_schedule_notrace - preempt_schedule called by tracing
 6956 *
 6957 * The tracing infrastructure uses preempt_enable_notrace to prevent
 6958 * recursion and tracing preempt enabling caused by the tracing
 6959 * infrastructure itself. But as tracing can happen in areas coming
 6960 * from userspace or just about to enter userspace, a preempt enable
 6961 * can occur before user_exit() is called. This will cause the scheduler
 6962 * to be called when the system is still in usermode.
 6963 *
 6964 * To prevent this, the preempt_enable_notrace will use this function
 6965 * instead of preempt_schedule() to exit user context if needed before
 6966 * calling the scheduler.
 6967 */
 6968asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
 6969{
 6970	enum ctx_state prev_ctx;
 6971
 6972	if (likely(!preemptible()))
 6973		return;
 6974
 6975	do {
 6976		/*
 6977		 * Because the function tracer can trace preempt_count_sub()
 6978		 * and it also uses preempt_enable/disable_notrace(), if
 6979		 * NEED_RESCHED is set, the preempt_enable_notrace() called
 6980		 * by the function tracer will call this function again and
 6981		 * cause infinite recursion.
 6982		 *
 6983		 * Preemption must be disabled here before the function
 6984		 * tracer can trace. Break up preempt_disable() into two
 6985		 * calls. One to disable preemption without fear of being
 6986		 * traced. The other to still record the preemption latency,
 6987		 * which can also be traced by the function tracer.
 6988		 */
 6989		preempt_disable_notrace();
 6990		preempt_latency_start(1);
 6991		/*
 6992		 * Needs preempt disabled in case user_exit() is traced
 6993		 * and the tracer calls preempt_enable_notrace() causing
 6994		 * an infinite recursion.
 6995		 */
 6996		prev_ctx = exception_enter();
 6997		__schedule(SM_PREEMPT);
 6998		exception_exit(prev_ctx);
 6999
 7000		preempt_latency_stop(1);
 7001		preempt_enable_no_resched_notrace();
 7002	} while (need_resched());
 7003}
 7004EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
 7005
 7006#ifdef CONFIG_PREEMPT_DYNAMIC
 7007#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 7008#ifndef preempt_schedule_notrace_dynamic_enabled
 7009#define preempt_schedule_notrace_dynamic_enabled	preempt_schedule_notrace
 7010#define preempt_schedule_notrace_dynamic_disabled	NULL
 7011#endif
 7012DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
 7013EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
 7014#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 7015static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
 7016void __sched notrace dynamic_preempt_schedule_notrace(void)
 7017{
 7018	if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
 7019		return;
 7020	preempt_schedule_notrace();
 7021}
 7022NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
 7023EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
 7024#endif
 7025#endif
 7026
 7027#endif /* CONFIG_PREEMPTION */
 7028
 7029/*
 7030 * This is the entry point to schedule() from kernel preemption
 7031 * off of irq context.
 7032 * Note, that this is called and return with irqs disabled. This will
 7033 * protect us against recursive calling from irq.
 7034 */
 7035asmlinkage __visible void __sched preempt_schedule_irq(void)
 7036{
 7037	enum ctx_state prev_state;
 7038
 7039	/* Catch callers which need to be fixed */
 7040	BUG_ON(preempt_count() || !irqs_disabled());
 7041
 7042	prev_state = exception_enter();
 7043
 7044	do {
 7045		preempt_disable();
 7046		local_irq_enable();
 7047		__schedule(SM_PREEMPT);
 7048		local_irq_disable();
 7049		sched_preempt_enable_no_resched();
 7050	} while (need_resched());
 7051
 7052	exception_exit(prev_state);
 7053}
 7054
 7055int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
 7056			  void *key)
 7057{
 7058	WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~(WF_SYNC|WF_CURRENT_CPU));
 7059	return try_to_wake_up(curr->private, mode, wake_flags);
 7060}
 7061EXPORT_SYMBOL(default_wake_function);
 7062
 7063static void __setscheduler_prio(struct task_struct *p, int prio)
 7064{
 7065	if (dl_prio(prio))
 7066		p->sched_class = &dl_sched_class;
 7067	else if (rt_prio(prio))
 7068		p->sched_class = &rt_sched_class;
 7069	else
 7070		p->sched_class = &fair_sched_class;
 7071
 7072	p->prio = prio;
 7073}
 7074
 7075#ifdef CONFIG_RT_MUTEXES
 7076
 7077/*
 7078 * Would be more useful with typeof()/auto_type but they don't mix with
 7079 * bit-fields. Since it's a local thing, use int. Keep the generic sounding
 7080 * name such that if someone were to implement this function we get to compare
 7081 * notes.
 7082 */
 7083#define fetch_and_set(x, v) ({ int _x = (x); (x) = (v); _x; })
 7084
 7085void rt_mutex_pre_schedule(void)
 7086{
 7087	lockdep_assert(!fetch_and_set(current->sched_rt_mutex, 1));
 7088	sched_submit_work(current);
 7089}
 7090
 7091void rt_mutex_schedule(void)
 7092{
 7093	lockdep_assert(current->sched_rt_mutex);
 7094	__schedule_loop(SM_NONE);
 7095}
 7096
 7097void rt_mutex_post_schedule(void)
 7098{
 7099	sched_update_worker(current);
 7100	lockdep_assert(fetch_and_set(current->sched_rt_mutex, 0));
 7101}
 7102
 7103static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
 7104{
 7105	if (pi_task)
 7106		prio = min(prio, pi_task->prio);
 7107
 7108	return prio;
 7109}
 7110
 7111static inline int rt_effective_prio(struct task_struct *p, int prio)
 7112{
 7113	struct task_struct *pi_task = rt_mutex_get_top_task(p);
 7114
 7115	return __rt_effective_prio(pi_task, prio);
 7116}
 7117
 7118/*
 7119 * rt_mutex_setprio - set the current priority of a task
 7120 * @p: task to boost
 7121 * @pi_task: donor task
 7122 *
 7123 * This function changes the 'effective' priority of a task. It does
 7124 * not touch ->normal_prio like __setscheduler().
 7125 *
 7126 * Used by the rt_mutex code to implement priority inheritance
 7127 * logic. Call site only calls if the priority of the task changed.
 7128 */
 7129void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
 7130{
 7131	int prio, oldprio, queued, running, queue_flag =
 7132		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
 7133	const struct sched_class *prev_class;
 7134	struct rq_flags rf;
 7135	struct rq *rq;
 7136
 7137	/* XXX used to be waiter->prio, not waiter->task->prio */
 7138	prio = __rt_effective_prio(pi_task, p->normal_prio);
 7139
 7140	/*
 7141	 * If nothing changed; bail early.
 7142	 */
 7143	if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
 7144		return;
 7145
 7146	rq = __task_rq_lock(p, &rf);
 7147	update_rq_clock(rq);
 7148	/*
 7149	 * Set under pi_lock && rq->lock, such that the value can be used under
 7150	 * either lock.
 7151	 *
 7152	 * Note that there is loads of tricky to make this pointer cache work
 7153	 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
 7154	 * ensure a task is de-boosted (pi_task is set to NULL) before the
 7155	 * task is allowed to run again (and can exit). This ensures the pointer
 7156	 * points to a blocked task -- which guarantees the task is present.
 7157	 */
 7158	p->pi_top_task = pi_task;
 7159
 7160	/*
 7161	 * For FIFO/RR we only need to set prio, if that matches we're done.
 7162	 */
 7163	if (prio == p->prio && !dl_prio(prio))
 7164		goto out_unlock;
 7165
 7166	/*
 7167	 * Idle task boosting is a nono in general. There is one
 7168	 * exception, when PREEMPT_RT and NOHZ is active:
 7169	 *
 7170	 * The idle task calls get_next_timer_interrupt() and holds
 7171	 * the timer wheel base->lock on the CPU and another CPU wants
 7172	 * to access the timer (probably to cancel it). We can safely
 7173	 * ignore the boosting request, as the idle CPU runs this code
 7174	 * with interrupts disabled and will complete the lock
 7175	 * protected section without being interrupted. So there is no
 7176	 * real need to boost.
 7177	 */
 7178	if (unlikely(p == rq->idle)) {
 7179		WARN_ON(p != rq->curr);
 7180		WARN_ON(p->pi_blocked_on);
 7181		goto out_unlock;
 7182	}
 7183
 7184	trace_sched_pi_setprio(p, pi_task);
 7185	oldprio = p->prio;
 7186
 7187	if (oldprio == prio)
 7188		queue_flag &= ~DEQUEUE_MOVE;
 7189
 7190	prev_class = p->sched_class;
 7191	queued = task_on_rq_queued(p);
 7192	running = task_current(rq, p);
 7193	if (queued)
 7194		dequeue_task(rq, p, queue_flag);
 7195	if (running)
 7196		put_prev_task(rq, p);
 7197
 7198	/*
 7199	 * Boosting condition are:
 7200	 * 1. -rt task is running and holds mutex A
 7201	 *      --> -dl task blocks on mutex A
 7202	 *
 7203	 * 2. -dl task is running and holds mutex A
 7204	 *      --> -dl task blocks on mutex A and could preempt the
 7205	 *          running task
 7206	 */
 7207	if (dl_prio(prio)) {
 7208		if (!dl_prio(p->normal_prio) ||
 7209		    (pi_task && dl_prio(pi_task->prio) &&
 7210		     dl_entity_preempt(&pi_task->dl, &p->dl))) {
 7211			p->dl.pi_se = pi_task->dl.pi_se;
 7212			queue_flag |= ENQUEUE_REPLENISH;
 7213		} else {
 7214			p->dl.pi_se = &p->dl;
 7215		}
 7216	} else if (rt_prio(prio)) {
 7217		if (dl_prio(oldprio))
 7218			p->dl.pi_se = &p->dl;
 7219		if (oldprio < prio)
 7220			queue_flag |= ENQUEUE_HEAD;
 
 7221	} else {
 7222		if (dl_prio(oldprio))
 7223			p->dl.pi_se = &p->dl;
 7224		if (rt_prio(oldprio))
 7225			p->rt.timeout = 0;
 
 7226	}
 7227
 7228	__setscheduler_prio(p, prio);
 7229
 7230	if (queued)
 7231		enqueue_task(rq, p, queue_flag);
 7232	if (running)
 7233		set_next_task(rq, p);
 7234
 7235	check_class_changed(rq, p, prev_class, oldprio);
 7236out_unlock:
 7237	/* Avoid rq from going away on us: */
 7238	preempt_disable();
 
 7239
 7240	rq_unpin_lock(rq, &rf);
 7241	__balance_callbacks(rq);
 7242	raw_spin_rq_unlock(rq);
 7243
 7244	preempt_enable();
 7245}
 7246#else
 7247static inline int rt_effective_prio(struct task_struct *p, int prio)
 7248{
 7249	return prio;
 7250}
 7251#endif
 7252
 7253void set_user_nice(struct task_struct *p, long nice)
 7254{
 7255	bool queued, running;
 
 
 7256	struct rq *rq;
 7257	int old_prio;
 7258
 7259	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
 7260		return;
 7261	/*
 7262	 * We have to be careful, if called from sys_setpriority(),
 7263	 * the task might be in the middle of scheduling on another CPU.
 7264	 */
 7265	CLASS(task_rq_lock, rq_guard)(p);
 7266	rq = rq_guard.rq;
 7267
 7268	update_rq_clock(rq);
 7269
 7270	/*
 7271	 * The RT priorities are set via sched_setscheduler(), but we still
 7272	 * allow the 'normal' nice value to be set - but as expected
 7273	 * it won't have any effect on scheduling until the task is
 7274	 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
 7275	 */
 7276	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
 7277		p->static_prio = NICE_TO_PRIO(nice);
 7278		return;
 7279	}
 7280
 7281	queued = task_on_rq_queued(p);
 7282	running = task_current(rq, p);
 7283	if (queued)
 7284		dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
 7285	if (running)
 7286		put_prev_task(rq, p);
 7287
 7288	p->static_prio = NICE_TO_PRIO(nice);
 7289	set_load_weight(p, true);
 7290	old_prio = p->prio;
 7291	p->prio = effective_prio(p);
 
 7292
 7293	if (queued)
 7294		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
 
 
 
 
 
 
 
 7295	if (running)
 7296		set_next_task(rq, p);
 7297
 7298	/*
 7299	 * If the task increased its priority or is running and
 7300	 * lowered its priority, then reschedule its CPU:
 7301	 */
 7302	p->sched_class->prio_changed(rq, p, old_prio);
 7303}
 7304EXPORT_SYMBOL(set_user_nice);
 7305
 7306/*
 7307 * is_nice_reduction - check if nice value is an actual reduction
 7308 *
 7309 * Similar to can_nice() but does not perform a capability check.
 7310 *
 7311 * @p: task
 7312 * @nice: nice value
 7313 */
 7314static bool is_nice_reduction(const struct task_struct *p, const int nice)
 7315{
 7316	/* Convert nice value [19,-20] to rlimit style value [1,40]: */
 7317	int nice_rlim = nice_to_rlimit(nice);
 7318
 7319	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
 7320}
 7321
 7322/*
 7323 * can_nice - check if a task can reduce its nice value
 7324 * @p: task
 7325 * @nice: nice value
 7326 */
 7327int can_nice(const struct task_struct *p, const int nice)
 7328{
 7329	return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
 7330}
 7331
 7332#ifdef __ARCH_WANT_SYS_NICE
 7333
 7334/*
 7335 * sys_nice - change the priority of the current process.
 7336 * @increment: priority increment
 7337 *
 7338 * sys_setpriority is a more generic, but much slower function that
 7339 * does similar things.
 7340 */
 7341SYSCALL_DEFINE1(nice, int, increment)
 7342{
 7343	long nice, retval;
 7344
 7345	/*
 7346	 * Setpriority might change our priority at the same moment.
 7347	 * We don't have to worry. Conceptually one call occurs first
 7348	 * and we have a single winner.
 7349	 */
 7350	increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
 7351	nice = task_nice(current) + increment;
 7352
 7353	nice = clamp_val(nice, MIN_NICE, MAX_NICE);
 7354	if (increment < 0 && !can_nice(current, nice))
 7355		return -EPERM;
 7356
 7357	retval = security_task_setnice(current, nice);
 7358	if (retval)
 7359		return retval;
 7360
 7361	set_user_nice(current, nice);
 7362	return 0;
 7363}
 7364
 7365#endif
 7366
 7367/**
 7368 * task_prio - return the priority value of a given task.
 7369 * @p: the task in question.
 7370 *
 7371 * Return: The priority value as seen by users in /proc.
 7372 *
 7373 * sched policy         return value   kernel prio    user prio/nice
 7374 *
 7375 * normal, batch, idle     [0 ... 39]  [100 ... 139]          0/[-20 ... 19]
 7376 * fifo, rr             [-2 ... -100]     [98 ... 0]  [1 ... 99]
 7377 * deadline                     -101             -1           0
 7378 */
 7379int task_prio(const struct task_struct *p)
 7380{
 7381	return p->prio - MAX_RT_PRIO;
 7382}
 7383
 7384/**
 7385 * idle_cpu - is a given CPU idle currently?
 7386 * @cpu: the processor in question.
 7387 *
 7388 * Return: 1 if the CPU is currently idle. 0 otherwise.
 7389 */
 7390int idle_cpu(int cpu)
 7391{
 7392	struct rq *rq = cpu_rq(cpu);
 7393
 7394	if (rq->curr != rq->idle)
 7395		return 0;
 7396
 7397	if (rq->nr_running)
 7398		return 0;
 7399
 7400#ifdef CONFIG_SMP
 7401	if (rq->ttwu_pending)
 7402		return 0;
 7403#endif
 7404
 7405	return 1;
 7406}
 7407
 7408/**
 7409 * available_idle_cpu - is a given CPU idle for enqueuing work.
 7410 * @cpu: the CPU in question.
 7411 *
 7412 * Return: 1 if the CPU is currently idle. 0 otherwise.
 7413 */
 7414int available_idle_cpu(int cpu)
 7415{
 7416	if (!idle_cpu(cpu))
 7417		return 0;
 7418
 7419	if (vcpu_is_preempted(cpu))
 7420		return 0;
 7421
 7422	return 1;
 7423}
 7424
 7425/**
 7426 * idle_task - return the idle task for a given CPU.
 7427 * @cpu: the processor in question.
 7428 *
 7429 * Return: The idle task for the CPU @cpu.
 7430 */
 7431struct task_struct *idle_task(int cpu)
 7432{
 7433	return cpu_rq(cpu)->idle;
 7434}
 7435
 7436#ifdef CONFIG_SCHED_CORE
 7437int sched_core_idle_cpu(int cpu)
 7438{
 7439	struct rq *rq = cpu_rq(cpu);
 7440
 7441	if (sched_core_enabled(rq) && rq->curr == rq->idle)
 7442		return 1;
 7443
 7444	return idle_cpu(cpu);
 7445}
 7446
 7447#endif
 7448
 7449#ifdef CONFIG_SMP
 7450/*
 7451 * This function computes an effective utilization for the given CPU, to be
 7452 * used for frequency selection given the linear relation: f = u * f_max.
 7453 *
 7454 * The scheduler tracks the following metrics:
 7455 *
 7456 *   cpu_util_{cfs,rt,dl,irq}()
 7457 *   cpu_bw_dl()
 7458 *
 7459 * Where the cfs,rt and dl util numbers are tracked with the same metric and
 7460 * synchronized windows and are thus directly comparable.
 7461 *
 7462 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
 7463 * which excludes things like IRQ and steal-time. These latter are then accrued
 7464 * in the irq utilization.
 7465 *
 7466 * The DL bandwidth number otoh is not a measured metric but a value computed
 7467 * based on the task model parameters and gives the minimal utilization
 7468 * required to meet deadlines.
 7469 */
 7470unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
 7471				 unsigned long *min,
 7472				 unsigned long *max)
 7473{
 7474	unsigned long util, irq, scale;
 7475	struct rq *rq = cpu_rq(cpu);
 7476
 7477	scale = arch_scale_cpu_capacity(cpu);
 7478
 7479	/*
 7480	 * Early check to see if IRQ/steal time saturates the CPU, can be
 7481	 * because of inaccuracies in how we track these -- see
 7482	 * update_irq_load_avg().
 7483	 */
 7484	irq = cpu_util_irq(rq);
 7485	if (unlikely(irq >= scale)) {
 7486		if (min)
 7487			*min = scale;
 7488		if (max)
 7489			*max = scale;
 7490		return scale;
 7491	}
 7492
 7493	if (min) {
 7494		/*
 7495		 * The minimum utilization returns the highest level between:
 7496		 * - the computed DL bandwidth needed with the IRQ pressure which
 7497		 *   steals time to the deadline task.
 7498		 * - The minimum performance requirement for CFS and/or RT.
 7499		 */
 7500		*min = max(irq + cpu_bw_dl(rq), uclamp_rq_get(rq, UCLAMP_MIN));
 7501
 7502		/*
 7503		 * When an RT task is runnable and uclamp is not used, we must
 7504		 * ensure that the task will run at maximum compute capacity.
 7505		 */
 7506		if (!uclamp_is_used() && rt_rq_is_runnable(&rq->rt))
 7507			*min = max(*min, scale);
 7508	}
 7509
 7510	/*
 7511	 * Because the time spend on RT/DL tasks is visible as 'lost' time to
 7512	 * CFS tasks and we use the same metric to track the effective
 7513	 * utilization (PELT windows are synchronized) we can directly add them
 7514	 * to obtain the CPU's actual utilization.
 7515	 */
 7516	util = util_cfs + cpu_util_rt(rq);
 7517	util += cpu_util_dl(rq);
 7518
 7519	/*
 7520	 * The maximum hint is a soft bandwidth requirement, which can be lower
 7521	 * than the actual utilization because of uclamp_max requirements.
 7522	 */
 7523	if (max)
 7524		*max = min(scale, uclamp_rq_get(rq, UCLAMP_MAX));
 7525
 7526	if (util >= scale)
 7527		return scale;
 7528
 7529	/*
 7530	 * There is still idle time; further improve the number by using the
 7531	 * irq metric. Because IRQ/steal time is hidden from the task clock we
 7532	 * need to scale the task numbers:
 7533	 *
 7534	 *              max - irq
 7535	 *   U' = irq + --------- * U
 7536	 *                 max
 7537	 */
 7538	util = scale_irq_capacity(util, irq, scale);
 7539	util += irq;
 7540
 7541	return min(scale, util);
 7542}
 7543
 7544unsigned long sched_cpu_util(int cpu)
 7545{
 7546	return effective_cpu_util(cpu, cpu_util_cfs(cpu), NULL, NULL);
 7547}
 7548#endif /* CONFIG_SMP */
 7549
 7550/**
 7551 * find_process_by_pid - find a process with a matching PID value.
 7552 * @pid: the pid in question.
 7553 *
 7554 * The task of @pid, if found. %NULL otherwise.
 7555 */
 7556static struct task_struct *find_process_by_pid(pid_t pid)
 7557{
 7558	return pid ? find_task_by_vpid(pid) : current;
 7559}
 7560
 7561static struct task_struct *find_get_task(pid_t pid)
 7562{
 7563	struct task_struct *p;
 7564	guard(rcu)();
 7565
 7566	p = find_process_by_pid(pid);
 7567	if (likely(p))
 7568		get_task_struct(p);
 7569
 7570	return p;
 7571}
 7572
 7573DEFINE_CLASS(find_get_task, struct task_struct *, if (_T) put_task_struct(_T),
 7574	     find_get_task(pid), pid_t pid)
 7575
 7576/*
 7577 * sched_setparam() passes in -1 for its policy, to let the functions
 7578 * it calls know not to change it.
 7579 */
 7580#define SETPARAM_POLICY	-1
 7581
 7582static void __setscheduler_params(struct task_struct *p,
 7583		const struct sched_attr *attr)
 7584{
 7585	int policy = attr->sched_policy;
 7586
 7587	if (policy == SETPARAM_POLICY)
 7588		policy = p->policy;
 7589
 7590	p->policy = policy;
 7591
 7592	if (dl_policy(policy))
 7593		__setparam_dl(p, attr);
 7594	else if (fair_policy(policy))
 7595		p->static_prio = NICE_TO_PRIO(attr->sched_nice);
 7596
 7597	/*
 7598	 * __sched_setscheduler() ensures attr->sched_priority == 0 when
 7599	 * !rt_policy. Always setting this ensures that things like
 7600	 * getparam()/getattr() don't report silly values for !rt tasks.
 7601	 */
 7602	p->rt_priority = attr->sched_priority;
 7603	p->normal_prio = normal_prio(p);
 7604	set_load_weight(p, true);
 7605}
 7606
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 7607/*
 7608 * Check the target process has a UID that matches the current process's:
 7609 */
 7610static bool check_same_owner(struct task_struct *p)
 7611{
 7612	const struct cred *cred = current_cred(), *pcred;
 7613	guard(rcu)();
 7614
 
 7615	pcred = __task_cred(p);
 7616	return (uid_eq(cred->euid, pcred->euid) ||
 7617		uid_eq(cred->euid, pcred->uid));
 7618}
 7619
 7620/*
 7621 * Allow unprivileged RT tasks to decrease priority.
 7622 * Only issue a capable test if needed and only once to avoid an audit
 7623 * event on permitted non-privileged operations:
 7624 */
 7625static int user_check_sched_setscheduler(struct task_struct *p,
 7626					 const struct sched_attr *attr,
 7627					 int policy, int reset_on_fork)
 7628{
 7629	if (fair_policy(policy)) {
 7630		if (attr->sched_nice < task_nice(p) &&
 7631		    !is_nice_reduction(p, attr->sched_nice))
 7632			goto req_priv;
 7633	}
 7634
 7635	if (rt_policy(policy)) {
 7636		unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
 7637
 7638		/* Can't set/change the rt policy: */
 7639		if (policy != p->policy && !rlim_rtprio)
 7640			goto req_priv;
 7641
 7642		/* Can't increase priority: */
 7643		if (attr->sched_priority > p->rt_priority &&
 7644		    attr->sched_priority > rlim_rtprio)
 7645			goto req_priv;
 7646	}
 7647
 7648	/*
 7649	 * Can't set/change SCHED_DEADLINE policy at all for now
 7650	 * (safest behavior); in the future we would like to allow
 7651	 * unprivileged DL tasks to increase their relative deadline
 7652	 * or reduce their runtime (both ways reducing utilization)
 7653	 */
 7654	if (dl_policy(policy))
 7655		goto req_priv;
 7656
 7657	/*
 7658	 * Treat SCHED_IDLE as nice 20. Only allow a switch to
 7659	 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
 7660	 */
 7661	if (task_has_idle_policy(p) && !idle_policy(policy)) {
 7662		if (!is_nice_reduction(p, task_nice(p)))
 7663			goto req_priv;
 7664	}
 7665
 7666	/* Can't change other user's priorities: */
 7667	if (!check_same_owner(p))
 7668		goto req_priv;
 7669
 7670	/* Normal users shall not reset the sched_reset_on_fork flag: */
 7671	if (p->sched_reset_on_fork && !reset_on_fork)
 7672		goto req_priv;
 7673
 7674	return 0;
 7675
 7676req_priv:
 7677	if (!capable(CAP_SYS_NICE))
 7678		return -EPERM;
 7679
 7680	return 0;
 7681}
 7682
 7683static int __sched_setscheduler(struct task_struct *p,
 7684				const struct sched_attr *attr,
 7685				bool user, bool pi)
 7686{
 7687	int oldpolicy = -1, policy = attr->sched_policy;
 7688	int retval, oldprio, newprio, queued, running;
 
 
 7689	const struct sched_class *prev_class;
 7690	struct balance_callback *head;
 7691	struct rq_flags rf;
 7692	int reset_on_fork;
 7693	int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
 7694	struct rq *rq;
 7695	bool cpuset_locked = false;
 7696
 7697	/* The pi code expects interrupts enabled */
 7698	BUG_ON(pi && in_interrupt());
 7699recheck:
 7700	/* Double check policy once rq lock held: */
 7701	if (policy < 0) {
 7702		reset_on_fork = p->sched_reset_on_fork;
 7703		policy = oldpolicy = p->policy;
 7704	} else {
 7705		reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
 7706
 7707		if (!valid_policy(policy))
 7708			return -EINVAL;
 7709	}
 7710
 7711	if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
 7712		return -EINVAL;
 7713
 7714	/*
 7715	 * Valid priorities for SCHED_FIFO and SCHED_RR are
 7716	 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
 7717	 * SCHED_BATCH and SCHED_IDLE is 0.
 7718	 */
 7719	if (attr->sched_priority > MAX_RT_PRIO-1)
 
 7720		return -EINVAL;
 7721	if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
 7722	    (rt_policy(policy) != (attr->sched_priority != 0)))
 7723		return -EINVAL;
 7724
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 7725	if (user) {
 7726		retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
 7727		if (retval)
 7728			return retval;
 7729
 7730		if (attr->sched_flags & SCHED_FLAG_SUGOV)
 7731			return -EINVAL;
 7732
 7733		retval = security_task_setscheduler(p);
 7734		if (retval)
 7735			return retval;
 7736	}
 7737
 7738	/* Update task specific "requested" clamps */
 7739	if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
 7740		retval = uclamp_validate(p, attr);
 7741		if (retval)
 7742			return retval;
 7743	}
 7744
 7745	/*
 7746	 * SCHED_DEADLINE bandwidth accounting relies on stable cpusets
 7747	 * information.
 7748	 */
 7749	if (dl_policy(policy) || dl_policy(p->policy)) {
 7750		cpuset_locked = true;
 7751		cpuset_lock();
 7752	}
 7753
 7754	/*
 7755	 * Make sure no PI-waiters arrive (or leave) while we are
 7756	 * changing the priority of the task:
 7757	 *
 7758	 * To be able to change p->policy safely, the appropriate
 7759	 * runqueue lock must be held.
 7760	 */
 7761	rq = task_rq_lock(p, &rf);
 7762	update_rq_clock(rq);
 7763
 7764	/*
 7765	 * Changing the policy of the stop threads its a very bad idea:
 7766	 */
 7767	if (p == rq->stop) {
 7768		retval = -EINVAL;
 7769		goto unlock;
 7770	}
 7771
 7772	/*
 7773	 * If not changing anything there's no need to proceed further,
 7774	 * but store a possible modification of reset_on_fork.
 7775	 */
 7776	if (unlikely(policy == p->policy)) {
 7777		if (fair_policy(policy) && attr->sched_nice != task_nice(p))
 7778			goto change;
 7779		if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
 7780			goto change;
 7781		if (dl_policy(policy) && dl_param_changed(p, attr))
 7782			goto change;
 7783		if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
 7784			goto change;
 7785
 7786		p->sched_reset_on_fork = reset_on_fork;
 7787		retval = 0;
 7788		goto unlock;
 7789	}
 7790change:
 7791
 7792	if (user) {
 7793#ifdef CONFIG_RT_GROUP_SCHED
 7794		/*
 7795		 * Do not allow realtime tasks into groups that have no runtime
 7796		 * assigned.
 7797		 */
 7798		if (rt_bandwidth_enabled() && rt_policy(policy) &&
 7799				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
 7800				!task_group_is_autogroup(task_group(p))) {
 7801			retval = -EPERM;
 7802			goto unlock;
 7803		}
 7804#endif
 7805#ifdef CONFIG_SMP
 7806		if (dl_bandwidth_enabled() && dl_policy(policy) &&
 7807				!(attr->sched_flags & SCHED_FLAG_SUGOV)) {
 7808			cpumask_t *span = rq->rd->span;
 7809
 7810			/*
 7811			 * Don't allow tasks with an affinity mask smaller than
 7812			 * the entire root_domain to become SCHED_DEADLINE. We
 7813			 * will also fail if there's no bandwidth available.
 7814			 */
 7815			if (!cpumask_subset(span, p->cpus_ptr) ||
 7816			    rq->rd->dl_bw.bw == 0) {
 7817				retval = -EPERM;
 7818				goto unlock;
 7819			}
 7820		}
 7821#endif
 7822	}
 7823
 7824	/* Re-check policy now with rq lock held: */
 7825	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
 7826		policy = oldpolicy = -1;
 7827		task_rq_unlock(rq, p, &rf);
 7828		if (cpuset_locked)
 7829			cpuset_unlock();
 7830		goto recheck;
 7831	}
 7832
 7833	/*
 7834	 * If setscheduling to SCHED_DEADLINE (or changing the parameters
 7835	 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
 7836	 * is available.
 7837	 */
 7838	if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
 7839		retval = -EBUSY;
 7840		goto unlock;
 7841	}
 7842
 7843	p->sched_reset_on_fork = reset_on_fork;
 7844	oldprio = p->prio;
 7845
 7846	newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
 7847	if (pi) {
 7848		/*
 7849		 * Take priority boosted tasks into account. If the new
 7850		 * effective priority is unchanged, we just store the new
 7851		 * normal parameters and do not touch the scheduler class and
 7852		 * the runqueue. This will be done when the task deboost
 7853		 * itself.
 7854		 */
 7855		newprio = rt_effective_prio(p, newprio);
 7856		if (newprio == oldprio)
 7857			queue_flags &= ~DEQUEUE_MOVE;
 7858	}
 7859
 7860	queued = task_on_rq_queued(p);
 7861	running = task_current(rq, p);
 7862	if (queued)
 7863		dequeue_task(rq, p, queue_flags);
 7864	if (running)
 7865		put_prev_task(rq, p);
 7866
 7867	prev_class = p->sched_class;
 7868
 7869	if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
 7870		__setscheduler_params(p, attr);
 7871		__setscheduler_prio(p, newprio);
 7872	}
 7873	__setscheduler_uclamp(p, attr);
 7874
 7875	if (queued) {
 7876		/*
 7877		 * We enqueue to tail when the priority of a task is
 7878		 * increased (user space view).
 7879		 */
 7880		if (oldprio < p->prio)
 7881			queue_flags |= ENQUEUE_HEAD;
 7882
 7883		enqueue_task(rq, p, queue_flags);
 7884	}
 7885	if (running)
 7886		set_next_task(rq, p);
 7887
 7888	check_class_changed(rq, p, prev_class, oldprio);
 7889
 7890	/* Avoid rq from going away on us: */
 7891	preempt_disable();
 7892	head = splice_balance_callbacks(rq);
 7893	task_rq_unlock(rq, p, &rf);
 7894
 7895	if (pi) {
 7896		if (cpuset_locked)
 7897			cpuset_unlock();
 7898		rt_mutex_adjust_pi(p);
 7899	}
 7900
 7901	/* Run balance callbacks after we've adjusted the PI chain: */
 7902	balance_callbacks(rq, head);
 7903	preempt_enable();
 7904
 7905	return 0;
 7906
 7907unlock:
 7908	task_rq_unlock(rq, p, &rf);
 7909	if (cpuset_locked)
 7910		cpuset_unlock();
 7911	return retval;
 7912}
 7913
 7914static int _sched_setscheduler(struct task_struct *p, int policy,
 7915			       const struct sched_param *param, bool check)
 7916{
 7917	struct sched_attr attr = {
 7918		.sched_policy   = policy,
 7919		.sched_priority = param->sched_priority,
 7920		.sched_nice	= PRIO_TO_NICE(p->static_prio),
 7921	};
 7922
 7923	/* Fixup the legacy SCHED_RESET_ON_FORK hack. */
 7924	if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
 7925		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
 7926		policy &= ~SCHED_RESET_ON_FORK;
 7927		attr.sched_policy = policy;
 7928	}
 7929
 7930	return __sched_setscheduler(p, &attr, check, true);
 7931}
 7932/**
 7933 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
 7934 * @p: the task in question.
 7935 * @policy: new policy.
 7936 * @param: structure containing the new RT priority.
 7937 *
 7938 * Use sched_set_fifo(), read its comment.
 7939 *
 7940 * Return: 0 on success. An error code otherwise.
 7941 *
 7942 * NOTE that the task may be already dead.
 7943 */
 7944int sched_setscheduler(struct task_struct *p, int policy,
 7945		       const struct sched_param *param)
 7946{
 7947	return _sched_setscheduler(p, policy, param, true);
 7948}
 
 7949
 7950int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
 7951{
 7952	return __sched_setscheduler(p, attr, true, true);
 7953}
 
 7954
 7955int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
 7956{
 7957	return __sched_setscheduler(p, attr, false, true);
 7958}
 7959EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
 7960
 7961/**
 7962 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
 7963 * @p: the task in question.
 7964 * @policy: new policy.
 7965 * @param: structure containing the new RT priority.
 7966 *
 7967 * Just like sched_setscheduler, only don't bother checking if the
 7968 * current context has permission.  For example, this is needed in
 7969 * stop_machine(): we create temporary high priority worker threads,
 7970 * but our caller might not have that capability.
 7971 *
 7972 * Return: 0 on success. An error code otherwise.
 7973 */
 7974int sched_setscheduler_nocheck(struct task_struct *p, int policy,
 7975			       const struct sched_param *param)
 7976{
 7977	return _sched_setscheduler(p, policy, param, false);
 7978}
 7979
 7980/*
 7981 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
 7982 * incapable of resource management, which is the one thing an OS really should
 7983 * be doing.
 7984 *
 7985 * This is of course the reason it is limited to privileged users only.
 7986 *
 7987 * Worse still; it is fundamentally impossible to compose static priority
 7988 * workloads. You cannot take two correctly working static prio workloads
 7989 * and smash them together and still expect them to work.
 7990 *
 7991 * For this reason 'all' FIFO tasks the kernel creates are basically at:
 7992 *
 7993 *   MAX_RT_PRIO / 2
 7994 *
 7995 * The administrator _MUST_ configure the system, the kernel simply doesn't
 7996 * know enough information to make a sensible choice.
 7997 */
 7998void sched_set_fifo(struct task_struct *p)
 7999{
 8000	struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
 8001	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
 8002}
 8003EXPORT_SYMBOL_GPL(sched_set_fifo);
 8004
 8005/*
 8006 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
 8007 */
 8008void sched_set_fifo_low(struct task_struct *p)
 8009{
 8010	struct sched_param sp = { .sched_priority = 1 };
 8011	WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
 8012}
 8013EXPORT_SYMBOL_GPL(sched_set_fifo_low);
 8014
 8015void sched_set_normal(struct task_struct *p, int nice)
 8016{
 8017	struct sched_attr attr = {
 8018		.sched_policy = SCHED_NORMAL,
 8019		.sched_nice = nice,
 8020	};
 8021	WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
 8022}
 8023EXPORT_SYMBOL_GPL(sched_set_normal);
 8024
 8025static int
 8026do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
 8027{
 8028	struct sched_param lparam;
 
 
 8029
 8030	if (!param || pid < 0)
 8031		return -EINVAL;
 8032	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
 8033		return -EFAULT;
 8034
 8035	CLASS(find_get_task, p)(pid);
 8036	if (!p)
 8037		return -ESRCH;
 
 
 
 8038
 8039	return sched_setscheduler(p, policy, &lparam);
 8040}
 8041
 8042/*
 8043 * Mimics kernel/events/core.c perf_copy_attr().
 8044 */
 8045static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
 8046{
 8047	u32 size;
 8048	int ret;
 8049
 
 
 
 8050	/* Zero the full structure, so that a short copy will be nice: */
 8051	memset(attr, 0, sizeof(*attr));
 8052
 8053	ret = get_user(size, &uattr->size);
 8054	if (ret)
 8055		return ret;
 8056
 
 
 
 
 8057	/* ABI compatibility quirk: */
 8058	if (!size)
 8059		size = SCHED_ATTR_SIZE_VER0;
 8060	if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
 
 8061		goto err_size;
 8062
 8063	ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
 8064	if (ret) {
 8065		if (ret == -E2BIG)
 8066			goto err_size;
 8067		return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 8068	}
 8069
 8070	if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
 8071	    size < SCHED_ATTR_SIZE_VER1)
 8072		return -EINVAL;
 8073
 8074	/*
 8075	 * XXX: Do we want to be lenient like existing syscalls; or do we want
 8076	 * to be strict and return an error on out-of-bounds values?
 8077	 */
 8078	attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
 8079
 8080	return 0;
 8081
 8082err_size:
 8083	put_user(sizeof(*attr), &uattr->size);
 8084	return -E2BIG;
 8085}
 8086
 8087static void get_params(struct task_struct *p, struct sched_attr *attr)
 8088{
 8089	if (task_has_dl_policy(p))
 8090		__getparam_dl(p, attr);
 8091	else if (task_has_rt_policy(p))
 8092		attr->sched_priority = p->rt_priority;
 8093	else
 8094		attr->sched_nice = task_nice(p);
 8095}
 8096
 8097/**
 8098 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
 8099 * @pid: the pid in question.
 8100 * @policy: new policy.
 8101 * @param: structure containing the new RT priority.
 8102 *
 8103 * Return: 0 on success. An error code otherwise.
 8104 */
 8105SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
 8106{
 8107	if (policy < 0)
 8108		return -EINVAL;
 8109
 8110	return do_sched_setscheduler(pid, policy, param);
 8111}
 8112
 8113/**
 8114 * sys_sched_setparam - set/change the RT priority of a thread
 8115 * @pid: the pid in question.
 8116 * @param: structure containing the new RT priority.
 8117 *
 8118 * Return: 0 on success. An error code otherwise.
 8119 */
 8120SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
 8121{
 8122	return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
 8123}
 8124
 8125/**
 8126 * sys_sched_setattr - same as above, but with extended sched_attr
 8127 * @pid: the pid in question.
 8128 * @uattr: structure containing the extended parameters.
 8129 * @flags: for future extension.
 8130 */
 8131SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
 8132			       unsigned int, flags)
 8133{
 8134	struct sched_attr attr;
 
 8135	int retval;
 8136
 8137	if (!uattr || pid < 0 || flags)
 8138		return -EINVAL;
 8139
 8140	retval = sched_copy_attr(uattr, &attr);
 8141	if (retval)
 8142		return retval;
 8143
 8144	if ((int)attr.sched_policy < 0)
 8145		return -EINVAL;
 8146	if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
 8147		attr.sched_policy = SETPARAM_POLICY;
 8148
 8149	CLASS(find_get_task, p)(pid);
 8150	if (!p)
 8151		return -ESRCH;
 
 
 
 8152
 8153	if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
 8154		get_params(p, &attr);
 8155
 8156	return sched_setattr(p, &attr);
 8157}
 8158
 8159/**
 8160 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
 8161 * @pid: the pid in question.
 8162 *
 8163 * Return: On success, the policy of the thread. Otherwise, a negative error
 8164 * code.
 8165 */
 8166SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
 8167{
 8168	struct task_struct *p;
 8169	int retval;
 8170
 8171	if (pid < 0)
 8172		return -EINVAL;
 8173
 8174	guard(rcu)();
 
 8175	p = find_process_by_pid(pid);
 8176	if (!p)
 8177		return -ESRCH;
 8178
 8179	retval = security_task_getscheduler(p);
 8180	if (!retval) {
 8181		retval = p->policy;
 8182		if (p->sched_reset_on_fork)
 8183			retval |= SCHED_RESET_ON_FORK;
 8184	}
 
 8185	return retval;
 8186}
 8187
 8188/**
 8189 * sys_sched_getparam - get the RT priority of a thread
 8190 * @pid: the pid in question.
 8191 * @param: structure containing the RT priority.
 8192 *
 8193 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
 8194 * code.
 8195 */
 8196SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
 8197{
 8198	struct sched_param lp = { .sched_priority = 0 };
 8199	struct task_struct *p;
 8200	int retval;
 8201
 8202	if (!param || pid < 0)
 8203		return -EINVAL;
 8204
 8205	scoped_guard (rcu) {
 8206		p = find_process_by_pid(pid);
 8207		if (!p)
 8208			return -ESRCH;
 
 8209
 8210		retval = security_task_getscheduler(p);
 8211		if (retval)
 8212			return retval;
 8213
 8214		if (task_has_rt_policy(p))
 8215			lp.sched_priority = p->rt_priority;
 8216	}
 8217
 8218	/*
 8219	 * This one might sleep, we cannot do it with a spinlock held ...
 8220	 */
 8221	return copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
 
 
 
 
 
 
 8222}
 8223
 8224/*
 8225 * Copy the kernel size attribute structure (which might be larger
 8226 * than what user-space knows about) to user-space.
 8227 *
 8228 * Note that all cases are valid: user-space buffer can be larger or
 8229 * smaller than the kernel-space buffer. The usual case is that both
 8230 * have the same size.
 8231 */
 8232static int
 8233sched_attr_copy_to_user(struct sched_attr __user *uattr,
 8234			struct sched_attr *kattr,
 8235			unsigned int usize)
 8236{
 8237	unsigned int ksize = sizeof(*kattr);
 8238
 8239	if (!access_ok(uattr, usize))
 8240		return -EFAULT;
 8241
 8242	/*
 8243	 * sched_getattr() ABI forwards and backwards compatibility:
 8244	 *
 8245	 * If usize == ksize then we just copy everything to user-space and all is good.
 8246	 *
 8247	 * If usize < ksize then we only copy as much as user-space has space for,
 8248	 * this keeps ABI compatibility as well. We skip the rest.
 8249	 *
 8250	 * If usize > ksize then user-space is using a newer version of the ABI,
 8251	 * which part the kernel doesn't know about. Just ignore it - tooling can
 8252	 * detect the kernel's knowledge of attributes from the attr->size value
 8253	 * which is set to ksize in this case.
 8254	 */
 8255	kattr->size = min(usize, ksize);
 
 
 
 
 
 
 
 
 
 
 
 
 
 8256
 8257	if (copy_to_user(uattr, kattr, kattr->size))
 
 8258		return -EFAULT;
 8259
 8260	return 0;
 8261}
 8262
 8263/**
 8264 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
 8265 * @pid: the pid in question.
 8266 * @uattr: structure containing the extended parameters.
 8267 * @usize: sizeof(attr) for fwd/bwd comp.
 8268 * @flags: for future extension.
 8269 */
 8270SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
 8271		unsigned int, usize, unsigned int, flags)
 8272{
 8273	struct sched_attr kattr = { };
 
 
 8274	struct task_struct *p;
 8275	int retval;
 8276
 8277	if (!uattr || pid < 0 || usize > PAGE_SIZE ||
 8278	    usize < SCHED_ATTR_SIZE_VER0 || flags)
 8279		return -EINVAL;
 8280
 8281	scoped_guard (rcu) {
 8282		p = find_process_by_pid(pid);
 8283		if (!p)
 8284			return -ESRCH;
 
 8285
 8286		retval = security_task_getscheduler(p);
 8287		if (retval)
 8288			return retval;
 
 
 
 
 
 
 
 
 
 
 8289
 8290		kattr.sched_policy = p->policy;
 8291		if (p->sched_reset_on_fork)
 8292			kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
 8293		get_params(p, &kattr);
 8294		kattr.sched_flags &= SCHED_FLAG_ALL;
 8295
 8296#ifdef CONFIG_UCLAMP_TASK
 8297		/*
 8298		 * This could race with another potential updater, but this is fine
 8299		 * because it'll correctly read the old or the new value. We don't need
 8300		 * to guarantee who wins the race as long as it doesn't return garbage.
 8301		 */
 8302		kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
 8303		kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
 8304#endif
 8305	}
 8306
 8307	return sched_attr_copy_to_user(uattr, &kattr, usize);
 
 
 8308}
 8309
 8310#ifdef CONFIG_SMP
 8311int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
 8312{
 8313	/*
 8314	 * If the task isn't a deadline task or admission control is
 8315	 * disabled then we don't care about affinity changes.
 8316	 */
 8317	if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
 8318		return 0;
 8319
 8320	/*
 8321	 * Since bandwidth control happens on root_domain basis,
 8322	 * if admission test is enabled, we only admit -deadline
 8323	 * tasks allowed to run on all the CPUs in the task's
 8324	 * root_domain.
 8325	 */
 8326	guard(rcu)();
 8327	if (!cpumask_subset(task_rq(p)->rd->span, mask))
 8328		return -EBUSY;
 8329
 8330	return 0;
 8331}
 8332#endif
 
 
 8333
 8334static int
 8335__sched_setaffinity(struct task_struct *p, struct affinity_context *ctx)
 8336{
 8337	int retval;
 8338	cpumask_var_t cpus_allowed, new_mask;
 8339
 8340	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
 8341		return -ENOMEM;
 8342
 
 
 
 
 
 
 
 
 8343	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
 8344		retval = -ENOMEM;
 8345		goto out_free_cpus_allowed;
 8346	}
 
 
 
 
 
 
 
 
 
 8347
 8348	cpuset_cpus_allowed(p, cpus_allowed);
 8349	cpumask_and(new_mask, ctx->new_mask, cpus_allowed);
 8350
 8351	ctx->new_mask = new_mask;
 8352	ctx->flags |= SCA_CHECK;
 8353
 8354	retval = dl_task_check_affinity(p, new_mask);
 8355	if (retval)
 8356		goto out_free_new_mask;
 8357
 8358	retval = __set_cpus_allowed_ptr(p, ctx);
 8359	if (retval)
 8360		goto out_free_new_mask;
 8361
 8362	cpuset_cpus_allowed(p, cpus_allowed);
 8363	if (!cpumask_subset(new_mask, cpus_allowed)) {
 8364		/*
 8365		 * We must have raced with a concurrent cpuset update.
 8366		 * Just reset the cpumask to the cpuset's cpus_allowed.
 8367		 */
 8368		cpumask_copy(new_mask, cpus_allowed);
 8369
 8370		/*
 8371		 * If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
 8372		 * will restore the previous user_cpus_ptr value.
 8373		 *
 8374		 * In the unlikely event a previous user_cpus_ptr exists,
 8375		 * we need to further restrict the mask to what is allowed
 8376		 * by that old user_cpus_ptr.
 8377		 */
 8378		if (unlikely((ctx->flags & SCA_USER) && ctx->user_mask)) {
 8379			bool empty = !cpumask_and(new_mask, new_mask,
 8380						  ctx->user_mask);
 
 
 
 
 
 
 
 
 8381
 8382			if (WARN_ON_ONCE(empty))
 8383				cpumask_copy(new_mask, cpus_allowed);
 
 
 
 
 
 
 
 
 8384		}
 8385		__set_cpus_allowed_ptr(p, ctx);
 8386		retval = -EINVAL;
 8387	}
 8388
 8389out_free_new_mask:
 8390	free_cpumask_var(new_mask);
 8391out_free_cpus_allowed:
 8392	free_cpumask_var(cpus_allowed);
 8393	return retval;
 8394}
 8395
 8396long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
 8397{
 8398	struct affinity_context ac;
 8399	struct cpumask *user_mask;
 8400	int retval;
 8401
 8402	CLASS(find_get_task, p)(pid);
 8403	if (!p)
 8404		return -ESRCH;
 8405
 8406	if (p->flags & PF_NO_SETAFFINITY)
 8407		return -EINVAL;
 8408
 8409	if (!check_same_owner(p)) {
 8410		guard(rcu)();
 8411		if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE))
 8412			return -EPERM;
 8413	}
 8414
 8415	retval = security_task_setscheduler(p);
 8416	if (retval)
 8417		return retval;
 8418
 8419	/*
 8420	 * With non-SMP configs, user_cpus_ptr/user_mask isn't used and
 8421	 * alloc_user_cpus_ptr() returns NULL.
 8422	 */
 8423	user_mask = alloc_user_cpus_ptr(NUMA_NO_NODE);
 8424	if (user_mask) {
 8425		cpumask_copy(user_mask, in_mask);
 8426	} else if (IS_ENABLED(CONFIG_SMP)) {
 8427		return -ENOMEM;
 8428	}
 8429
 8430	ac = (struct affinity_context){
 8431		.new_mask  = in_mask,
 8432		.user_mask = user_mask,
 8433		.flags     = SCA_USER,
 8434	};
 8435
 8436	retval = __sched_setaffinity(p, &ac);
 8437	kfree(ac.user_mask);
 8438
 8439	return retval;
 8440}
 8441
 8442static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
 8443			     struct cpumask *new_mask)
 8444{
 8445	if (len < cpumask_size())
 8446		cpumask_clear(new_mask);
 8447	else if (len > cpumask_size())
 8448		len = cpumask_size();
 8449
 8450	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
 8451}
 8452
 8453/**
 8454 * sys_sched_setaffinity - set the CPU affinity of a process
 8455 * @pid: pid of the process
 8456 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
 8457 * @user_mask_ptr: user-space pointer to the new CPU mask
 8458 *
 8459 * Return: 0 on success. An error code otherwise.
 8460 */
 8461SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
 8462		unsigned long __user *, user_mask_ptr)
 8463{
 8464	cpumask_var_t new_mask;
 8465	int retval;
 8466
 8467	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
 8468		return -ENOMEM;
 8469
 8470	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
 8471	if (retval == 0)
 8472		retval = sched_setaffinity(pid, new_mask);
 8473	free_cpumask_var(new_mask);
 8474	return retval;
 8475}
 8476
 8477long sched_getaffinity(pid_t pid, struct cpumask *mask)
 8478{
 8479	struct task_struct *p;
 
 8480	int retval;
 8481
 8482	guard(rcu)();
 
 
 8483	p = find_process_by_pid(pid);
 8484	if (!p)
 8485		return -ESRCH;
 8486
 8487	retval = security_task_getscheduler(p);
 8488	if (retval)
 8489		return retval;
 8490
 8491	guard(raw_spinlock_irqsave)(&p->pi_lock);
 8492	cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
 
 8493
 8494	return 0;
 
 
 
 8495}
 8496
 8497/**
 8498 * sys_sched_getaffinity - get the CPU affinity of a process
 8499 * @pid: pid of the process
 8500 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
 8501 * @user_mask_ptr: user-space pointer to hold the current CPU mask
 8502 *
 8503 * Return: size of CPU mask copied to user_mask_ptr on success. An
 8504 * error code otherwise.
 8505 */
 8506SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
 8507		unsigned long __user *, user_mask_ptr)
 8508{
 8509	int ret;
 8510	cpumask_var_t mask;
 8511
 8512	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
 8513		return -EINVAL;
 8514	if (len & (sizeof(unsigned long)-1))
 8515		return -EINVAL;
 8516
 8517	if (!zalloc_cpumask_var(&mask, GFP_KERNEL))
 8518		return -ENOMEM;
 8519
 8520	ret = sched_getaffinity(pid, mask);
 8521	if (ret == 0) {
 8522		unsigned int retlen = min(len, cpumask_size());
 8523
 8524		if (copy_to_user(user_mask_ptr, cpumask_bits(mask), retlen))
 8525			ret = -EFAULT;
 8526		else
 8527			ret = retlen;
 8528	}
 8529	free_cpumask_var(mask);
 8530
 8531	return ret;
 8532}
 8533
 
 
 
 
 
 
 
 
 8534static void do_sched_yield(void)
 8535{
 8536	struct rq_flags rf;
 8537	struct rq *rq;
 8538
 8539	rq = this_rq_lock_irq(&rf);
 
 
 8540
 8541	schedstat_inc(rq->yld_count);
 8542	current->sched_class->yield_task(rq);
 8543
 
 
 
 
 8544	preempt_disable();
 8545	rq_unlock_irq(rq, &rf);
 8546	sched_preempt_enable_no_resched();
 8547
 8548	schedule();
 8549}
 8550
 8551/**
 8552 * sys_sched_yield - yield the current processor to other threads.
 8553 *
 8554 * This function yields the current CPU to other tasks. If there are no
 8555 * other threads running on this CPU then this function will return.
 8556 *
 8557 * Return: 0.
 8558 */
 8559SYSCALL_DEFINE0(sched_yield)
 8560{
 8561	do_sched_yield();
 8562	return 0;
 8563}
 8564
 8565#if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
 8566int __sched __cond_resched(void)
 8567{
 8568	if (should_resched(0)) {
 8569		preempt_schedule_common();
 8570		return 1;
 8571	}
 8572	/*
 8573	 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
 8574	 * whether the current CPU is in an RCU read-side critical section,
 8575	 * so the tick can report quiescent states even for CPUs looping
 8576	 * in kernel context.  In contrast, in non-preemptible kernels,
 8577	 * RCU readers leave no in-memory hints, which means that CPU-bound
 8578	 * processes executing in kernel context might never report an
 8579	 * RCU quiescent state.  Therefore, the following code causes
 8580	 * cond_resched() to report a quiescent state, but only when RCU
 8581	 * is in urgent need of one.
 8582	 */
 8583#ifndef CONFIG_PREEMPT_RCU
 8584	rcu_all_qs();
 8585#endif
 8586	return 0;
 8587}
 8588EXPORT_SYMBOL(__cond_resched);
 8589#endif
 8590
 8591#ifdef CONFIG_PREEMPT_DYNAMIC
 8592#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 8593#define cond_resched_dynamic_enabled	__cond_resched
 8594#define cond_resched_dynamic_disabled	((void *)&__static_call_return0)
 8595DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
 8596EXPORT_STATIC_CALL_TRAMP(cond_resched);
 8597
 8598#define might_resched_dynamic_enabled	__cond_resched
 8599#define might_resched_dynamic_disabled	((void *)&__static_call_return0)
 8600DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
 8601EXPORT_STATIC_CALL_TRAMP(might_resched);
 8602#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 8603static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
 8604int __sched dynamic_cond_resched(void)
 8605{
 8606	klp_sched_try_switch();
 8607	if (!static_branch_unlikely(&sk_dynamic_cond_resched))
 8608		return 0;
 8609	return __cond_resched();
 8610}
 8611EXPORT_SYMBOL(dynamic_cond_resched);
 8612
 8613static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
 8614int __sched dynamic_might_resched(void)
 8615{
 8616	if (!static_branch_unlikely(&sk_dynamic_might_resched))
 8617		return 0;
 8618	return __cond_resched();
 8619}
 8620EXPORT_SYMBOL(dynamic_might_resched);
 8621#endif
 8622#endif
 8623
 8624/*
 8625 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
 8626 * call schedule, and on return reacquire the lock.
 8627 *
 8628 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
 8629 * operations here to prevent schedule() from being called twice (once via
 8630 * spin_unlock(), once by hand).
 8631 */
 8632int __cond_resched_lock(spinlock_t *lock)
 8633{
 8634	int resched = should_resched(PREEMPT_LOCK_OFFSET);
 8635	int ret = 0;
 8636
 8637	lockdep_assert_held(lock);
 8638
 8639	if (spin_needbreak(lock) || resched) {
 8640		spin_unlock(lock);
 8641		if (!_cond_resched())
 
 
 8642			cpu_relax();
 8643		ret = 1;
 8644		spin_lock(lock);
 8645	}
 8646	return ret;
 8647}
 8648EXPORT_SYMBOL(__cond_resched_lock);
 8649
 8650int __cond_resched_rwlock_read(rwlock_t *lock)
 8651{
 8652	int resched = should_resched(PREEMPT_LOCK_OFFSET);
 8653	int ret = 0;
 8654
 8655	lockdep_assert_held_read(lock);
 8656
 8657	if (rwlock_needbreak(lock) || resched) {
 8658		read_unlock(lock);
 8659		if (!_cond_resched())
 8660			cpu_relax();
 8661		ret = 1;
 8662		read_lock(lock);
 8663	}
 8664	return ret;
 8665}
 8666EXPORT_SYMBOL(__cond_resched_rwlock_read);
 8667
 8668int __cond_resched_rwlock_write(rwlock_t *lock)
 8669{
 8670	int resched = should_resched(PREEMPT_LOCK_OFFSET);
 8671	int ret = 0;
 8672
 8673	lockdep_assert_held_write(lock);
 8674
 8675	if (rwlock_needbreak(lock) || resched) {
 8676		write_unlock(lock);
 8677		if (!_cond_resched())
 8678			cpu_relax();
 8679		ret = 1;
 8680		write_lock(lock);
 8681	}
 8682	return ret;
 8683}
 8684EXPORT_SYMBOL(__cond_resched_rwlock_write);
 8685
 8686#ifdef CONFIG_PREEMPT_DYNAMIC
 8687
 8688#ifdef CONFIG_GENERIC_ENTRY
 8689#include <linux/entry-common.h>
 8690#endif
 8691
 8692/*
 8693 * SC:cond_resched
 8694 * SC:might_resched
 8695 * SC:preempt_schedule
 8696 * SC:preempt_schedule_notrace
 8697 * SC:irqentry_exit_cond_resched
 8698 *
 8699 *
 8700 * NONE:
 8701 *   cond_resched               <- __cond_resched
 8702 *   might_resched              <- RET0
 8703 *   preempt_schedule           <- NOP
 8704 *   preempt_schedule_notrace   <- NOP
 8705 *   irqentry_exit_cond_resched <- NOP
 8706 *
 8707 * VOLUNTARY:
 8708 *   cond_resched               <- __cond_resched
 8709 *   might_resched              <- __cond_resched
 8710 *   preempt_schedule           <- NOP
 8711 *   preempt_schedule_notrace   <- NOP
 8712 *   irqentry_exit_cond_resched <- NOP
 8713 *
 8714 * FULL:
 8715 *   cond_resched               <- RET0
 8716 *   might_resched              <- RET0
 8717 *   preempt_schedule           <- preempt_schedule
 8718 *   preempt_schedule_notrace   <- preempt_schedule_notrace
 8719 *   irqentry_exit_cond_resched <- irqentry_exit_cond_resched
 8720 */
 8721
 8722enum {
 8723	preempt_dynamic_undefined = -1,
 8724	preempt_dynamic_none,
 8725	preempt_dynamic_voluntary,
 8726	preempt_dynamic_full,
 8727};
 8728
 8729int preempt_dynamic_mode = preempt_dynamic_undefined;
 8730
 8731int sched_dynamic_mode(const char *str)
 8732{
 8733	if (!strcmp(str, "none"))
 8734		return preempt_dynamic_none;
 8735
 8736	if (!strcmp(str, "voluntary"))
 8737		return preempt_dynamic_voluntary;
 8738
 8739	if (!strcmp(str, "full"))
 8740		return preempt_dynamic_full;
 8741
 8742	return -EINVAL;
 8743}
 8744
 8745#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
 8746#define preempt_dynamic_enable(f)	static_call_update(f, f##_dynamic_enabled)
 8747#define preempt_dynamic_disable(f)	static_call_update(f, f##_dynamic_disabled)
 8748#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
 8749#define preempt_dynamic_enable(f)	static_key_enable(&sk_dynamic_##f.key)
 8750#define preempt_dynamic_disable(f)	static_key_disable(&sk_dynamic_##f.key)
 8751#else
 8752#error "Unsupported PREEMPT_DYNAMIC mechanism"
 8753#endif
 8754
 8755static DEFINE_MUTEX(sched_dynamic_mutex);
 8756static bool klp_override;
 8757
 8758static void __sched_dynamic_update(int mode)
 8759{
 8760	/*
 8761	 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
 8762	 * the ZERO state, which is invalid.
 8763	 */
 8764	if (!klp_override)
 8765		preempt_dynamic_enable(cond_resched);
 8766	preempt_dynamic_enable(might_resched);
 8767	preempt_dynamic_enable(preempt_schedule);
 8768	preempt_dynamic_enable(preempt_schedule_notrace);
 8769	preempt_dynamic_enable(irqentry_exit_cond_resched);
 8770
 8771	switch (mode) {
 8772	case preempt_dynamic_none:
 8773		if (!klp_override)
 8774			preempt_dynamic_enable(cond_resched);
 8775		preempt_dynamic_disable(might_resched);
 8776		preempt_dynamic_disable(preempt_schedule);
 8777		preempt_dynamic_disable(preempt_schedule_notrace);
 8778		preempt_dynamic_disable(irqentry_exit_cond_resched);
 8779		if (mode != preempt_dynamic_mode)
 8780			pr_info("Dynamic Preempt: none\n");
 8781		break;
 8782
 8783	case preempt_dynamic_voluntary:
 8784		if (!klp_override)
 8785			preempt_dynamic_enable(cond_resched);
 8786		preempt_dynamic_enable(might_resched);
 8787		preempt_dynamic_disable(preempt_schedule);
 8788		preempt_dynamic_disable(preempt_schedule_notrace);
 8789		preempt_dynamic_disable(irqentry_exit_cond_resched);
 8790		if (mode != preempt_dynamic_mode)
 8791			pr_info("Dynamic Preempt: voluntary\n");
 8792		break;
 8793
 8794	case preempt_dynamic_full:
 8795		if (!klp_override)
 8796			preempt_dynamic_disable(cond_resched);
 8797		preempt_dynamic_disable(might_resched);
 8798		preempt_dynamic_enable(preempt_schedule);
 8799		preempt_dynamic_enable(preempt_schedule_notrace);
 8800		preempt_dynamic_enable(irqentry_exit_cond_resched);
 8801		if (mode != preempt_dynamic_mode)
 8802			pr_info("Dynamic Preempt: full\n");
 8803		break;
 8804	}
 8805
 8806	preempt_dynamic_mode = mode;
 8807}
 8808
 8809void sched_dynamic_update(int mode)
 8810{
 8811	mutex_lock(&sched_dynamic_mutex);
 8812	__sched_dynamic_update(mode);
 8813	mutex_unlock(&sched_dynamic_mutex);
 8814}
 8815
 8816#ifdef CONFIG_HAVE_PREEMPT_DYNAMIC_CALL
 8817
 8818static int klp_cond_resched(void)
 8819{
 8820	__klp_sched_try_switch();
 8821	return __cond_resched();
 8822}
 8823
 8824void sched_dynamic_klp_enable(void)
 8825{
 8826	mutex_lock(&sched_dynamic_mutex);
 8827
 8828	klp_override = true;
 8829	static_call_update(cond_resched, klp_cond_resched);
 8830
 8831	mutex_unlock(&sched_dynamic_mutex);
 8832}
 8833
 8834void sched_dynamic_klp_disable(void)
 8835{
 8836	mutex_lock(&sched_dynamic_mutex);
 8837
 8838	klp_override = false;
 8839	__sched_dynamic_update(preempt_dynamic_mode);
 8840
 8841	mutex_unlock(&sched_dynamic_mutex);
 8842}
 8843
 8844#endif /* CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
 8845
 8846static int __init setup_preempt_mode(char *str)
 8847{
 8848	int mode = sched_dynamic_mode(str);
 8849	if (mode < 0) {
 8850		pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
 8851		return 0;
 8852	}
 8853
 8854	sched_dynamic_update(mode);
 8855	return 1;
 8856}
 8857__setup("preempt=", setup_preempt_mode);
 8858
 8859static void __init preempt_dynamic_init(void)
 8860{
 8861	if (preempt_dynamic_mode == preempt_dynamic_undefined) {
 8862		if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
 8863			sched_dynamic_update(preempt_dynamic_none);
 8864		} else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
 8865			sched_dynamic_update(preempt_dynamic_voluntary);
 8866		} else {
 8867			/* Default static call setting, nothing to do */
 8868			WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
 8869			preempt_dynamic_mode = preempt_dynamic_full;
 8870			pr_info("Dynamic Preempt: full\n");
 8871		}
 8872	}
 8873}
 8874
 8875#define PREEMPT_MODEL_ACCESSOR(mode) \
 8876	bool preempt_model_##mode(void)						 \
 8877	{									 \
 8878		WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
 8879		return preempt_dynamic_mode == preempt_dynamic_##mode;		 \
 8880	}									 \
 8881	EXPORT_SYMBOL_GPL(preempt_model_##mode)
 8882
 8883PREEMPT_MODEL_ACCESSOR(none);
 8884PREEMPT_MODEL_ACCESSOR(voluntary);
 8885PREEMPT_MODEL_ACCESSOR(full);
 8886
 8887#else /* !CONFIG_PREEMPT_DYNAMIC */
 8888
 8889static inline void preempt_dynamic_init(void) { }
 8890
 8891#endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
 8892
 8893/**
 8894 * yield - yield the current processor to other threads.
 8895 *
 8896 * Do not ever use this function, there's a 99% chance you're doing it wrong.
 8897 *
 8898 * The scheduler is at all times free to pick the calling task as the most
 8899 * eligible task to run, if removing the yield() call from your code breaks
 8900 * it, it's already broken.
 8901 *
 8902 * Typical broken usage is:
 8903 *
 8904 * while (!event)
 8905 *	yield();
 8906 *
 8907 * where one assumes that yield() will let 'the other' process run that will
 8908 * make event true. If the current task is a SCHED_FIFO task that will never
 8909 * happen. Never use yield() as a progress guarantee!!
 8910 *
 8911 * If you want to use yield() to wait for something, use wait_event().
 8912 * If you want to use yield() to be 'nice' for others, use cond_resched().
 8913 * If you still want to use yield(), do not!
 8914 */
 8915void __sched yield(void)
 8916{
 8917	set_current_state(TASK_RUNNING);
 8918	do_sched_yield();
 8919}
 8920EXPORT_SYMBOL(yield);
 8921
 8922/**
 8923 * yield_to - yield the current processor to another thread in
 8924 * your thread group, or accelerate that thread toward the
 8925 * processor it's on.
 8926 * @p: target task
 8927 * @preempt: whether task preemption is allowed or not
 8928 *
 8929 * It's the caller's job to ensure that the target task struct
 8930 * can't go away on us before we can do any checks.
 8931 *
 8932 * Return:
 8933 *	true (>0) if we indeed boosted the target task.
 8934 *	false (0) if we failed to boost the target.
 8935 *	-ESRCH if there's no task to yield to.
 8936 */
 8937int __sched yield_to(struct task_struct *p, bool preempt)
 8938{
 8939	struct task_struct *curr = current;
 8940	struct rq *rq, *p_rq;
 
 8941	int yielded = 0;
 8942
 8943	scoped_guard (irqsave) {
 8944		rq = this_rq();
 8945
 8946again:
 8947		p_rq = task_rq(p);
 8948		/*
 8949		 * If we're the only runnable task on the rq and target rq also
 8950		 * has only one task, there's absolutely no point in yielding.
 8951		 */
 8952		if (rq->nr_running == 1 && p_rq->nr_running == 1)
 8953			return -ESRCH;
 
 
 8954
 8955		guard(double_rq_lock)(rq, p_rq);
 8956		if (task_rq(p) != p_rq)
 8957			goto again;
 
 
 8958
 8959		if (!curr->sched_class->yield_to_task)
 8960			return 0;
 8961
 8962		if (curr->sched_class != p->sched_class)
 8963			return 0;
 8964
 8965		if (task_on_cpu(p_rq, p) || !task_is_running(p))
 8966			return 0;
 8967
 8968		yielded = curr->sched_class->yield_to_task(rq, p);
 8969		if (yielded) {
 8970			schedstat_inc(rq->yld_count);
 8971			/*
 8972			 * Make p's CPU reschedule; pick_next_entity
 8973			 * takes care of fairness.
 8974			 */
 8975			if (preempt && rq != p_rq)
 8976				resched_curr(p_rq);
 8977		}
 8978	}
 8979
 8980	if (yielded)
 
 
 
 
 
 8981		schedule();
 8982
 8983	return yielded;
 8984}
 8985EXPORT_SYMBOL_GPL(yield_to);
 8986
 8987int io_schedule_prepare(void)
 8988{
 8989	int old_iowait = current->in_iowait;
 8990
 8991	current->in_iowait = 1;
 8992	blk_flush_plug(current->plug, true);
 
 8993	return old_iowait;
 8994}
 8995
 8996void io_schedule_finish(int token)
 8997{
 8998	current->in_iowait = token;
 8999}
 9000
 9001/*
 9002 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
 9003 * that process accounting knows that this is a task in IO wait state.
 9004 */
 9005long __sched io_schedule_timeout(long timeout)
 9006{
 9007	int token;
 9008	long ret;
 9009
 9010	token = io_schedule_prepare();
 9011	ret = schedule_timeout(timeout);
 9012	io_schedule_finish(token);
 9013
 9014	return ret;
 9015}
 9016EXPORT_SYMBOL(io_schedule_timeout);
 9017
 9018void __sched io_schedule(void)
 9019{
 9020	int token;
 9021
 9022	token = io_schedule_prepare();
 9023	schedule();
 9024	io_schedule_finish(token);
 9025}
 9026EXPORT_SYMBOL(io_schedule);
 9027
 9028/**
 9029 * sys_sched_get_priority_max - return maximum RT priority.
 9030 * @policy: scheduling class.
 9031 *
 9032 * Return: On success, this syscall returns the maximum
 9033 * rt_priority that can be used by a given scheduling class.
 9034 * On failure, a negative error code is returned.
 9035 */
 9036SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
 9037{
 9038	int ret = -EINVAL;
 9039
 9040	switch (policy) {
 9041	case SCHED_FIFO:
 9042	case SCHED_RR:
 9043		ret = MAX_RT_PRIO-1;
 9044		break;
 9045	case SCHED_DEADLINE:
 9046	case SCHED_NORMAL:
 9047	case SCHED_BATCH:
 9048	case SCHED_IDLE:
 9049		ret = 0;
 9050		break;
 9051	}
 9052	return ret;
 9053}
 9054
 9055/**
 9056 * sys_sched_get_priority_min - return minimum RT priority.
 9057 * @policy: scheduling class.
 9058 *
 9059 * Return: On success, this syscall returns the minimum
 9060 * rt_priority that can be used by a given scheduling class.
 9061 * On failure, a negative error code is returned.
 9062 */
 9063SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
 9064{
 9065	int ret = -EINVAL;
 9066
 9067	switch (policy) {
 9068	case SCHED_FIFO:
 9069	case SCHED_RR:
 9070		ret = 1;
 9071		break;
 9072	case SCHED_DEADLINE:
 9073	case SCHED_NORMAL:
 9074	case SCHED_BATCH:
 9075	case SCHED_IDLE:
 9076		ret = 0;
 9077	}
 9078	return ret;
 9079}
 9080
 9081static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
 9082{
 9083	unsigned int time_slice = 0;
 
 
 
 9084	int retval;
 9085
 9086	if (pid < 0)
 9087		return -EINVAL;
 9088
 9089	scoped_guard (rcu) {
 9090		struct task_struct *p = find_process_by_pid(pid);
 9091		if (!p)
 9092			return -ESRCH;
 
 9093
 9094		retval = security_task_getscheduler(p);
 9095		if (retval)
 9096			return retval;
 9097
 9098		scoped_guard (task_rq_lock, p) {
 9099			struct rq *rq = scope.rq;
 9100			if (p->sched_class->get_rr_interval)
 9101				time_slice = p->sched_class->get_rr_interval(rq, p);
 9102		}
 9103	}
 9104
 
 9105	jiffies_to_timespec64(time_slice, t);
 9106	return 0;
 
 
 
 
 9107}
 9108
 9109/**
 9110 * sys_sched_rr_get_interval - return the default timeslice of a process.
 9111 * @pid: pid of the process.
 9112 * @interval: userspace pointer to the timeslice value.
 9113 *
 9114 * this syscall writes the default timeslice value of a given process
 9115 * into the user-space timespec buffer. A value of '0' means infinity.
 9116 *
 9117 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
 9118 * an error code.
 9119 */
 9120SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
 9121		struct __kernel_timespec __user *, interval)
 9122{
 9123	struct timespec64 t;
 9124	int retval = sched_rr_get_interval(pid, &t);
 9125
 9126	if (retval == 0)
 9127		retval = put_timespec64(&t, interval);
 9128
 9129	return retval;
 9130}
 9131
 9132#ifdef CONFIG_COMPAT_32BIT_TIME
 9133SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
 9134		struct old_timespec32 __user *, interval)
 
 9135{
 9136	struct timespec64 t;
 9137	int retval = sched_rr_get_interval(pid, &t);
 9138
 9139	if (retval == 0)
 9140		retval = put_old_timespec32(&t, interval);
 9141	return retval;
 9142}
 9143#endif
 9144
 9145void sched_show_task(struct task_struct *p)
 9146{
 9147	unsigned long free = 0;
 9148	int ppid;
 9149
 9150	if (!try_get_task_stack(p))
 9151		return;
 9152
 9153	pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
 9154
 9155	if (task_is_running(p))
 9156		pr_cont("  running task    ");
 9157#ifdef CONFIG_DEBUG_STACK_USAGE
 9158	free = stack_not_used(p);
 9159#endif
 9160	ppid = 0;
 9161	rcu_read_lock();
 9162	if (pid_alive(p))
 9163		ppid = task_pid_nr(rcu_dereference(p->real_parent));
 9164	rcu_read_unlock();
 9165	pr_cont(" stack:%-5lu pid:%-5d tgid:%-5d ppid:%-6d flags:0x%08lx\n",
 9166		free, task_pid_nr(p), task_tgid_nr(p),
 9167		ppid, read_task_thread_flags(p));
 9168
 9169	print_worker_info(KERN_INFO, p);
 9170	print_stop_info(KERN_INFO, p);
 9171	show_stack(p, NULL, KERN_INFO);
 9172	put_task_stack(p);
 9173}
 9174EXPORT_SYMBOL_GPL(sched_show_task);
 9175
 9176static inline bool
 9177state_filter_match(unsigned long state_filter, struct task_struct *p)
 9178{
 9179	unsigned int state = READ_ONCE(p->__state);
 9180
 9181	/* no filter, everything matches */
 9182	if (!state_filter)
 9183		return true;
 9184
 9185	/* filter, but doesn't match */
 9186	if (!(state & state_filter))
 9187		return false;
 9188
 9189	/*
 9190	 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
 9191	 * TASK_KILLABLE).
 9192	 */
 9193	if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
 9194		return false;
 9195
 9196	return true;
 9197}
 9198
 9199
 9200void show_state_filter(unsigned int state_filter)
 9201{
 9202	struct task_struct *g, *p;
 9203
 
 
 
 
 
 
 
 9204	rcu_read_lock();
 9205	for_each_process_thread(g, p) {
 9206		/*
 9207		 * reset the NMI-timeout, listing all files on a slow
 9208		 * console might take a lot of time:
 9209		 * Also, reset softlockup watchdogs on all CPUs, because
 9210		 * another CPU might be blocked waiting for us to process
 9211		 * an IPI.
 9212		 */
 9213		touch_nmi_watchdog();
 9214		touch_all_softlockup_watchdogs();
 9215		if (state_filter_match(state_filter, p))
 9216			sched_show_task(p);
 9217	}
 9218
 9219#ifdef CONFIG_SCHED_DEBUG
 9220	if (!state_filter)
 9221		sysrq_sched_debug_show();
 9222#endif
 9223	rcu_read_unlock();
 9224	/*
 9225	 * Only show locks if all tasks are dumped:
 9226	 */
 9227	if (!state_filter)
 9228		debug_show_all_locks();
 9229}
 9230
 9231/**
 9232 * init_idle - set up an idle thread for a given CPU
 9233 * @idle: task in question
 9234 * @cpu: CPU the idle task belongs to
 9235 *
 9236 * NOTE: this function does not set the idle thread's NEED_RESCHED
 9237 * flag, to make booting more robust.
 9238 */
 9239void __init init_idle(struct task_struct *idle, int cpu)
 9240{
 9241#ifdef CONFIG_SMP
 9242	struct affinity_context ac = (struct affinity_context) {
 9243		.new_mask  = cpumask_of(cpu),
 9244		.flags     = 0,
 9245	};
 9246#endif
 9247	struct rq *rq = cpu_rq(cpu);
 9248	unsigned long flags;
 9249
 9250	__sched_fork(0, idle);
 9251
 9252	raw_spin_lock_irqsave(&idle->pi_lock, flags);
 9253	raw_spin_rq_lock(rq);
 9254
 9255	idle->__state = TASK_RUNNING;
 
 9256	idle->se.exec_start = sched_clock();
 9257	/*
 9258	 * PF_KTHREAD should already be set at this point; regardless, make it
 9259	 * look like a proper per-CPU kthread.
 9260	 */
 9261	idle->flags |= PF_KTHREAD | PF_NO_SETAFFINITY;
 9262	kthread_set_per_cpu(idle, cpu);
 9263
 9264#ifdef CONFIG_SMP
 9265	/*
 9266	 * It's possible that init_idle() gets called multiple times on a task,
 9267	 * in that case do_set_cpus_allowed() will not do the right thing.
 9268	 *
 9269	 * And since this is boot we can forgo the serialization.
 9270	 */
 9271	set_cpus_allowed_common(idle, &ac);
 9272#endif
 9273	/*
 9274	 * We're having a chicken and egg problem, even though we are
 9275	 * holding rq->lock, the CPU isn't yet set to this CPU so the
 9276	 * lockdep check in task_group() will fail.
 9277	 *
 9278	 * Similar case to sched_fork(). / Alternatively we could
 9279	 * use task_rq_lock() here and obtain the other rq->lock.
 9280	 *
 9281	 * Silence PROVE_RCU
 9282	 */
 9283	rcu_read_lock();
 9284	__set_task_cpu(idle, cpu);
 9285	rcu_read_unlock();
 9286
 9287	rq->idle = idle;
 9288	rcu_assign_pointer(rq->curr, idle);
 9289	idle->on_rq = TASK_ON_RQ_QUEUED;
 9290#ifdef CONFIG_SMP
 9291	idle->on_cpu = 1;
 9292#endif
 9293	raw_spin_rq_unlock(rq);
 9294	raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
 9295
 9296	/* Set the preempt count _outside_ the spinlocks! */
 9297	init_idle_preempt_count(idle, cpu);
 9298
 9299	/*
 9300	 * The idle tasks have their own, simple scheduling class:
 9301	 */
 9302	idle->sched_class = &idle_sched_class;
 9303	ftrace_graph_init_idle_task(idle, cpu);
 9304	vtime_init_idle(idle, cpu);
 9305#ifdef CONFIG_SMP
 9306	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
 9307#endif
 9308}
 9309
 9310#ifdef CONFIG_SMP
 9311
 9312int cpuset_cpumask_can_shrink(const struct cpumask *cur,
 9313			      const struct cpumask *trial)
 9314{
 9315	int ret = 1;
 9316
 9317	if (cpumask_empty(cur))
 9318		return ret;
 9319
 9320	ret = dl_cpuset_cpumask_can_shrink(cur, trial);
 9321
 9322	return ret;
 9323}
 9324
 9325int task_can_attach(struct task_struct *p)
 
 9326{
 9327	int ret = 0;
 9328
 9329	/*
 9330	 * Kthreads which disallow setaffinity shouldn't be moved
 9331	 * to a new cpuset; we don't want to change their CPU
 9332	 * affinity and isolating such threads by their set of
 9333	 * allowed nodes is unnecessary.  Thus, cpusets are not
 9334	 * applicable for such threads.  This prevents checking for
 9335	 * success of set_cpus_allowed_ptr() on all attached tasks
 9336	 * before cpus_mask may be changed.
 9337	 */
 9338	if (p->flags & PF_NO_SETAFFINITY)
 9339		ret = -EINVAL;
 
 
 9340
 
 
 
 
 
 9341	return ret;
 9342}
 9343
 9344bool sched_smp_initialized __read_mostly;
 9345
 9346#ifdef CONFIG_NUMA_BALANCING
 9347/* Migrate current task p to target_cpu */
 9348int migrate_task_to(struct task_struct *p, int target_cpu)
 9349{
 9350	struct migration_arg arg = { p, target_cpu };
 9351	int curr_cpu = task_cpu(p);
 9352
 9353	if (curr_cpu == target_cpu)
 9354		return 0;
 9355
 9356	if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
 9357		return -EINVAL;
 9358
 9359	/* TODO: This is not properly updating schedstats */
 9360
 9361	trace_sched_move_numa(p, curr_cpu, target_cpu);
 9362	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
 9363}
 9364
 9365/*
 9366 * Requeue a task on a given node and accurately track the number of NUMA
 9367 * tasks on the runqueues
 9368 */
 9369void sched_setnuma(struct task_struct *p, int nid)
 9370{
 9371	bool queued, running;
 9372	struct rq_flags rf;
 9373	struct rq *rq;
 9374
 9375	rq = task_rq_lock(p, &rf);
 9376	queued = task_on_rq_queued(p);
 9377	running = task_current(rq, p);
 9378
 9379	if (queued)
 9380		dequeue_task(rq, p, DEQUEUE_SAVE);
 9381	if (running)
 9382		put_prev_task(rq, p);
 9383
 9384	p->numa_preferred_nid = nid;
 9385
 9386	if (queued)
 9387		enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
 9388	if (running)
 9389		set_next_task(rq, p);
 9390	task_rq_unlock(rq, p, &rf);
 9391}
 9392#endif /* CONFIG_NUMA_BALANCING */
 9393
 9394#ifdef CONFIG_HOTPLUG_CPU
 9395/*
 9396 * Ensure that the idle task is using init_mm right before its CPU goes
 9397 * offline.
 9398 */
 9399void idle_task_exit(void)
 9400{
 9401	struct mm_struct *mm = current->active_mm;
 9402
 9403	BUG_ON(cpu_online(smp_processor_id()));
 9404	BUG_ON(current != this_rq()->idle);
 9405
 9406	if (mm != &init_mm) {
 9407		switch_mm(mm, &init_mm, current);
 
 9408		finish_arch_post_lock_switch();
 9409	}
 
 
 9410
 9411	/* finish_cpu(), as ran on the BP, will clean up the active_mm state */
 
 
 
 
 
 
 
 
 
 
 
 
 
 9412}
 9413
 9414static int __balance_push_cpu_stop(void *arg)
 9415{
 9416	struct task_struct *p = arg;
 9417	struct rq *rq = this_rq();
 9418	struct rq_flags rf;
 9419	int cpu;
 9420
 9421	raw_spin_lock_irq(&p->pi_lock);
 9422	rq_lock(rq, &rf);
 
 9423
 9424	update_rq_clock(rq);
 9425
 9426	if (task_rq(p) == rq && task_on_rq_queued(p)) {
 9427		cpu = select_fallback_rq(rq->cpu, p);
 9428		rq = __migrate_task(rq, &rf, p, cpu);
 9429	}
 9430
 9431	rq_unlock(rq, &rf);
 9432	raw_spin_unlock_irq(&p->pi_lock);
 9433
 9434	put_task_struct(p);
 9435
 9436	return 0;
 9437}
 9438
 9439static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
 9440
 9441/*
 9442 * Ensure we only run per-cpu kthreads once the CPU goes !active.
 
 9443 *
 9444 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
 9445 * effective when the hotplug motion is down.
 9446 */
 9447static void balance_push(struct rq *rq)
 9448{
 9449	struct task_struct *push_task = rq->curr;
 9450
 9451	lockdep_assert_rq_held(rq);
 
 
 9452
 9453	/*
 9454	 * Ensure the thing is persistent until balance_push_set(.on = false);
 
 
 
 
 
 
 9455	 */
 9456	rq->balance_callback = &balance_push_callback;
 9457
 9458	/*
 9459	 * Only active while going offline and when invoked on the outgoing
 9460	 * CPU.
 
 9461	 */
 9462	if (!cpu_dying(rq->cpu) || rq != this_rq())
 9463		return;
 
 
 
 
 
 
 
 9464
 9465	/*
 9466	 * Both the cpu-hotplug and stop task are in this case and are
 9467	 * required to complete the hotplug process.
 9468	 */
 9469	if (kthread_is_per_cpu(push_task) ||
 9470	    is_migration_disabled(push_task)) {
 9471
 9472		/*
 9473		 * If this is the idle task on the outgoing CPU try to wake
 9474		 * up the hotplug control thread which might wait for the
 9475		 * last task to vanish. The rcuwait_active() check is
 9476		 * accurate here because the waiter is pinned on this CPU
 9477		 * and can't obviously be running in parallel.
 9478		 *
 9479		 * On RT kernels this also has to check whether there are
 9480		 * pinned and scheduled out tasks on the runqueue. They
 9481		 * need to leave the migrate disabled section first.
 9482		 */
 9483		if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
 9484		    rcuwait_active(&rq->hotplug_wait)) {
 9485			raw_spin_rq_unlock(rq);
 9486			rcuwait_wake_up(&rq->hotplug_wait);
 9487			raw_spin_rq_lock(rq);
 
 
 
 
 
 
 
 9488		}
 9489		return;
 9490	}
 9491
 9492	get_task_struct(push_task);
 9493	/*
 9494	 * Temporarily drop rq->lock such that we can wake-up the stop task.
 9495	 * Both preemption and IRQs are still disabled.
 9496	 */
 9497	preempt_disable();
 9498	raw_spin_rq_unlock(rq);
 9499	stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
 9500			    this_cpu_ptr(&push_work));
 9501	preempt_enable();
 9502	/*
 9503	 * At this point need_resched() is true and we'll take the loop in
 9504	 * schedule(). The next pick is obviously going to be the stop task
 9505	 * which kthread_is_per_cpu() and will push this task away.
 9506	 */
 9507	raw_spin_rq_lock(rq);
 9508}
 9509
 9510static void balance_push_set(int cpu, bool on)
 9511{
 9512	struct rq *rq = cpu_rq(cpu);
 9513	struct rq_flags rf;
 9514
 9515	rq_lock_irqsave(rq, &rf);
 9516	if (on) {
 9517		WARN_ON_ONCE(rq->balance_callback);
 9518		rq->balance_callback = &balance_push_callback;
 9519	} else if (rq->balance_callback == &balance_push_callback) {
 9520		rq->balance_callback = NULL;
 9521	}
 9522	rq_unlock_irqrestore(rq, &rf);
 9523}
 9524
 9525/*
 9526 * Invoked from a CPUs hotplug control thread after the CPU has been marked
 9527 * inactive. All tasks which are not per CPU kernel threads are either
 9528 * pushed off this CPU now via balance_push() or placed on a different CPU
 9529 * during wakeup. Wait until the CPU is quiescent.
 9530 */
 9531static void balance_hotplug_wait(void)
 9532{
 9533	struct rq *rq = this_rq();
 9534
 9535	rcuwait_wait_event(&rq->hotplug_wait,
 9536			   rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
 9537			   TASK_UNINTERRUPTIBLE);
 9538}
 9539
 9540#else
 9541
 9542static inline void balance_push(struct rq *rq)
 9543{
 9544}
 9545
 9546static inline void balance_push_set(int cpu, bool on)
 9547{
 9548}
 9549
 9550static inline void balance_hotplug_wait(void)
 9551{
 9552}
 9553
 9554#endif /* CONFIG_HOTPLUG_CPU */
 9555
 9556void set_rq_online(struct rq *rq)
 9557{
 9558	if (!rq->online) {
 9559		const struct sched_class *class;
 9560
 9561		cpumask_set_cpu(rq->cpu, rq->rd->online);
 9562		rq->online = 1;
 9563
 9564		for_each_class(class) {
 9565			if (class->rq_online)
 9566				class->rq_online(rq);
 9567		}
 9568	}
 9569}
 9570
 9571void set_rq_offline(struct rq *rq)
 9572{
 9573	if (rq->online) {
 9574		const struct sched_class *class;
 9575
 9576		update_rq_clock(rq);
 9577		for_each_class(class) {
 9578			if (class->rq_offline)
 9579				class->rq_offline(rq);
 9580		}
 9581
 9582		cpumask_clear_cpu(rq->cpu, rq->rd->online);
 9583		rq->online = 0;
 9584	}
 9585}
 9586
 
 
 
 
 
 
 
 9587/*
 9588 * used to mark begin/end of suspend/resume:
 9589 */
 9590static int num_cpus_frozen;
 9591
 9592/*
 9593 * Update cpusets according to cpu_active mask.  If cpusets are
 9594 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
 9595 * around partition_sched_domains().
 9596 *
 9597 * If we come here as part of a suspend/resume, don't touch cpusets because we
 9598 * want to restore it back to its original state upon resume anyway.
 9599 */
 9600static void cpuset_cpu_active(void)
 9601{
 9602	if (cpuhp_tasks_frozen) {
 9603		/*
 9604		 * num_cpus_frozen tracks how many CPUs are involved in suspend
 9605		 * resume sequence. As long as this is not the last online
 9606		 * operation in the resume sequence, just build a single sched
 9607		 * domain, ignoring cpusets.
 9608		 */
 9609		partition_sched_domains(1, NULL, NULL);
 9610		if (--num_cpus_frozen)
 9611			return;
 9612		/*
 9613		 * This is the last CPU online operation. So fall through and
 9614		 * restore the original sched domains by considering the
 9615		 * cpuset configurations.
 9616		 */
 9617		cpuset_force_rebuild();
 9618	}
 9619	cpuset_update_active_cpus();
 9620}
 9621
 9622static int cpuset_cpu_inactive(unsigned int cpu)
 9623{
 9624	if (!cpuhp_tasks_frozen) {
 9625		int ret = dl_bw_check_overflow(cpu);
 9626
 9627		if (ret)
 9628			return ret;
 9629		cpuset_update_active_cpus();
 9630	} else {
 9631		num_cpus_frozen++;
 9632		partition_sched_domains(1, NULL, NULL);
 9633	}
 9634	return 0;
 9635}
 9636
 9637int sched_cpu_activate(unsigned int cpu)
 9638{
 9639	struct rq *rq = cpu_rq(cpu);
 9640	struct rq_flags rf;
 9641
 9642	/*
 9643	 * Clear the balance_push callback and prepare to schedule
 9644	 * regular tasks.
 9645	 */
 9646	balance_push_set(cpu, false);
 9647
 9648#ifdef CONFIG_SCHED_SMT
 9649	/*
 9650	 * When going up, increment the number of cores with SMT present.
 9651	 */
 9652	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
 9653		static_branch_inc_cpuslocked(&sched_smt_present);
 9654#endif
 9655	set_cpu_active(cpu, true);
 9656
 9657	if (sched_smp_initialized) {
 9658		sched_update_numa(cpu, true);
 9659		sched_domains_numa_masks_set(cpu);
 9660		cpuset_cpu_active();
 9661	}
 9662
 9663	/*
 9664	 * Put the rq online, if not already. This happens:
 9665	 *
 9666	 * 1) In the early boot process, because we build the real domains
 9667	 *    after all CPUs have been brought up.
 9668	 *
 9669	 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
 9670	 *    domains.
 9671	 */
 9672	rq_lock_irqsave(rq, &rf);
 9673	if (rq->rd) {
 9674		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
 9675		set_rq_online(rq);
 9676	}
 9677	rq_unlock_irqrestore(rq, &rf);
 9678
 
 
 9679	return 0;
 9680}
 9681
 9682int sched_cpu_deactivate(unsigned int cpu)
 9683{
 9684	struct rq *rq = cpu_rq(cpu);
 9685	struct rq_flags rf;
 9686	int ret;
 9687
 9688	/*
 9689	 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
 9690	 * load balancing when not active
 9691	 */
 9692	nohz_balance_exit_idle(rq);
 9693
 9694	set_cpu_active(cpu, false);
 9695
 9696	/*
 9697	 * From this point forward, this CPU will refuse to run any task that
 9698	 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
 9699	 * push those tasks away until this gets cleared, see
 9700	 * sched_cpu_dying().
 9701	 */
 9702	balance_push_set(cpu, true);
 9703
 9704	/*
 9705	 * We've cleared cpu_active_mask / set balance_push, wait for all
 9706	 * preempt-disabled and RCU users of this state to go away such that
 9707	 * all new such users will observe it.
 9708	 *
 9709	 * Specifically, we rely on ttwu to no longer target this CPU, see
 9710	 * ttwu_queue_cond() and is_cpu_allowed().
 9711	 *
 9712	 * Do sync before park smpboot threads to take care the rcu boost case.
 9713	 */
 9714	synchronize_rcu();
 9715
 9716	rq_lock_irqsave(rq, &rf);
 9717	if (rq->rd) {
 9718		BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
 9719		set_rq_offline(rq);
 9720	}
 9721	rq_unlock_irqrestore(rq, &rf);
 9722
 9723#ifdef CONFIG_SCHED_SMT
 9724	/*
 9725	 * When going down, decrement the number of cores with SMT present.
 9726	 */
 9727	if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
 9728		static_branch_dec_cpuslocked(&sched_smt_present);
 9729
 9730	sched_core_cpu_deactivate(cpu);
 9731#endif
 9732
 9733	if (!sched_smp_initialized)
 9734		return 0;
 9735
 9736	sched_update_numa(cpu, false);
 9737	ret = cpuset_cpu_inactive(cpu);
 9738	if (ret) {
 9739		balance_push_set(cpu, false);
 9740		set_cpu_active(cpu, true);
 9741		sched_update_numa(cpu, true);
 9742		return ret;
 9743	}
 9744	sched_domains_numa_masks_clear(cpu);
 9745	return 0;
 9746}
 9747
 9748static void sched_rq_cpu_starting(unsigned int cpu)
 9749{
 9750	struct rq *rq = cpu_rq(cpu);
 9751
 9752	rq->calc_load_update = calc_load_update;
 9753	update_max_interval();
 9754}
 9755
 9756int sched_cpu_starting(unsigned int cpu)
 9757{
 9758	sched_core_cpu_starting(cpu);
 9759	sched_rq_cpu_starting(cpu);
 9760	sched_tick_start(cpu);
 9761	return 0;
 9762}
 9763
 9764#ifdef CONFIG_HOTPLUG_CPU
 9765
 9766/*
 9767 * Invoked immediately before the stopper thread is invoked to bring the
 9768 * CPU down completely. At this point all per CPU kthreads except the
 9769 * hotplug thread (current) and the stopper thread (inactive) have been
 9770 * either parked or have been unbound from the outgoing CPU. Ensure that
 9771 * any of those which might be on the way out are gone.
 9772 *
 9773 * If after this point a bound task is being woken on this CPU then the
 9774 * responsible hotplug callback has failed to do it's job.
 9775 * sched_cpu_dying() will catch it with the appropriate fireworks.
 9776 */
 9777int sched_cpu_wait_empty(unsigned int cpu)
 9778{
 9779	balance_hotplug_wait();
 9780	return 0;
 9781}
 9782
 9783/*
 9784 * Since this CPU is going 'away' for a while, fold any nr_active delta we
 9785 * might have. Called from the CPU stopper task after ensuring that the
 9786 * stopper is the last running task on the CPU, so nr_active count is
 9787 * stable. We need to take the teardown thread which is calling this into
 9788 * account, so we hand in adjust = 1 to the load calculation.
 9789 *
 9790 * Also see the comment "Global load-average calculations".
 9791 */
 9792static void calc_load_migrate(struct rq *rq)
 9793{
 9794	long delta = calc_load_fold_active(rq, 1);
 9795
 9796	if (delta)
 9797		atomic_long_add(delta, &calc_load_tasks);
 9798}
 9799
 9800static void dump_rq_tasks(struct rq *rq, const char *loglvl)
 9801{
 9802	struct task_struct *g, *p;
 9803	int cpu = cpu_of(rq);
 9804
 9805	lockdep_assert_rq_held(rq);
 9806
 9807	printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
 9808	for_each_process_thread(g, p) {
 9809		if (task_cpu(p) != cpu)
 9810			continue;
 9811
 9812		if (!task_on_rq_queued(p))
 9813			continue;
 9814
 9815		printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
 9816	}
 9817}
 9818
 9819int sched_cpu_dying(unsigned int cpu)
 9820{
 9821	struct rq *rq = cpu_rq(cpu);
 9822	struct rq_flags rf;
 9823
 9824	/* Handle pending wakeups and then migrate everything off */
 
 9825	sched_tick_stop(cpu);
 9826
 9827	rq_lock_irqsave(rq, &rf);
 9828	if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
 9829		WARN(true, "Dying CPU not properly vacated!");
 9830		dump_rq_tasks(rq, KERN_WARNING);
 9831	}
 
 
 9832	rq_unlock_irqrestore(rq, &rf);
 9833
 9834	calc_load_migrate(rq);
 9835	update_max_interval();
 
 9836	hrtick_clear(rq);
 9837	sched_core_cpu_dying(cpu);
 9838	return 0;
 9839}
 9840#endif
 9841
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 9842void __init sched_init_smp(void)
 9843{
 9844	sched_init_numa(NUMA_NO_NODE);
 9845
 9846	/*
 9847	 * There's no userspace yet to cause hotplug operations; hence all the
 9848	 * CPU masks are stable and all blatant races in the below code cannot
 9849	 * happen.
 9850	 */
 9851	mutex_lock(&sched_domains_mutex);
 9852	sched_init_domains(cpu_active_mask);
 9853	mutex_unlock(&sched_domains_mutex);
 9854
 9855	/* Move init over to a non-isolated CPU */
 9856	if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
 9857		BUG();
 9858	current->flags &= ~PF_NO_SETAFFINITY;
 9859	sched_init_granularity();
 9860
 9861	init_sched_rt_class();
 9862	init_sched_dl_class();
 9863
 
 
 9864	sched_smp_initialized = true;
 9865}
 9866
 9867static int __init migration_init(void)
 9868{
 9869	sched_cpu_starting(smp_processor_id());
 9870	return 0;
 9871}
 9872early_initcall(migration_init);
 9873
 9874#else
 9875void __init sched_init_smp(void)
 9876{
 9877	sched_init_granularity();
 9878}
 9879#endif /* CONFIG_SMP */
 9880
 9881int in_sched_functions(unsigned long addr)
 9882{
 9883	return in_lock_functions(addr) ||
 9884		(addr >= (unsigned long)__sched_text_start
 9885		&& addr < (unsigned long)__sched_text_end);
 9886}
 9887
 9888#ifdef CONFIG_CGROUP_SCHED
 9889/*
 9890 * Default task group.
 9891 * Every task in system belongs to this group at bootup.
 9892 */
 9893struct task_group root_task_group;
 9894LIST_HEAD(task_groups);
 9895
 9896/* Cacheline aligned slab cache for task_group */
 9897static struct kmem_cache *task_group_cache __ro_after_init;
 9898#endif
 9899
 
 
 
 9900void __init sched_init(void)
 9901{
 9902	unsigned long ptr = 0;
 9903	int i;
 9904
 9905	/* Make sure the linker didn't screw up */
 9906	BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
 9907	       &fair_sched_class != &rt_sched_class + 1 ||
 9908	       &rt_sched_class   != &dl_sched_class + 1);
 9909#ifdef CONFIG_SMP
 9910	BUG_ON(&dl_sched_class != &stop_sched_class + 1);
 9911#endif
 9912
 
 9913	wait_bit_init();
 9914
 9915#ifdef CONFIG_FAIR_GROUP_SCHED
 9916	ptr += 2 * nr_cpu_ids * sizeof(void **);
 9917#endif
 9918#ifdef CONFIG_RT_GROUP_SCHED
 9919	ptr += 2 * nr_cpu_ids * sizeof(void **);
 9920#endif
 9921	if (ptr) {
 9922		ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
 9923
 9924#ifdef CONFIG_FAIR_GROUP_SCHED
 9925		root_task_group.se = (struct sched_entity **)ptr;
 9926		ptr += nr_cpu_ids * sizeof(void **);
 9927
 9928		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
 9929		ptr += nr_cpu_ids * sizeof(void **);
 9930
 9931		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
 9932		init_cfs_bandwidth(&root_task_group.cfs_bandwidth, NULL);
 9933#endif /* CONFIG_FAIR_GROUP_SCHED */
 9934#ifdef CONFIG_RT_GROUP_SCHED
 9935		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
 9936		ptr += nr_cpu_ids * sizeof(void **);
 9937
 9938		root_task_group.rt_rq = (struct rt_rq **)ptr;
 9939		ptr += nr_cpu_ids * sizeof(void **);
 9940
 9941#endif /* CONFIG_RT_GROUP_SCHED */
 9942	}
 
 
 
 
 
 
 
 
 9943
 9944	init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
 
 9945
 9946#ifdef CONFIG_SMP
 9947	init_defrootdomain();
 9948#endif
 9949
 9950#ifdef CONFIG_RT_GROUP_SCHED
 9951	init_rt_bandwidth(&root_task_group.rt_bandwidth,
 9952			global_rt_period(), global_rt_runtime());
 9953#endif /* CONFIG_RT_GROUP_SCHED */
 9954
 9955#ifdef CONFIG_CGROUP_SCHED
 9956	task_group_cache = KMEM_CACHE(task_group, 0);
 9957
 9958	list_add(&root_task_group.list, &task_groups);
 9959	INIT_LIST_HEAD(&root_task_group.children);
 9960	INIT_LIST_HEAD(&root_task_group.siblings);
 9961	autogroup_init(&init_task);
 9962#endif /* CONFIG_CGROUP_SCHED */
 9963
 9964	for_each_possible_cpu(i) {
 9965		struct rq *rq;
 9966
 9967		rq = cpu_rq(i);
 9968		raw_spin_lock_init(&rq->__lock);
 9969		rq->nr_running = 0;
 9970		rq->calc_load_active = 0;
 9971		rq->calc_load_update = jiffies + LOAD_FREQ;
 9972		init_cfs_rq(&rq->cfs);
 9973		init_rt_rq(&rq->rt);
 9974		init_dl_rq(&rq->dl);
 9975#ifdef CONFIG_FAIR_GROUP_SCHED
 
 9976		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
 9977		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
 9978		/*
 9979		 * How much CPU bandwidth does root_task_group get?
 9980		 *
 9981		 * In case of task-groups formed thr' the cgroup filesystem, it
 9982		 * gets 100% of the CPU resources in the system. This overall
 9983		 * system CPU resource is divided among the tasks of
 9984		 * root_task_group and its child task-groups in a fair manner,
 9985		 * based on each entity's (task or task-group's) weight
 9986		 * (se->load.weight).
 9987		 *
 9988		 * In other words, if root_task_group has 10 tasks of weight
 9989		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
 9990		 * then A0's share of the CPU resource is:
 9991		 *
 9992		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
 9993		 *
 9994		 * We achieve this by letting root_task_group's tasks sit
 9995		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
 9996		 */
 
 9997		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
 9998#endif /* CONFIG_FAIR_GROUP_SCHED */
 9999
10000		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
10001#ifdef CONFIG_RT_GROUP_SCHED
10002		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
10003#endif
 
 
 
 
10004#ifdef CONFIG_SMP
10005		rq->sd = NULL;
10006		rq->rd = NULL;
10007		rq->cpu_capacity = SCHED_CAPACITY_SCALE;
10008		rq->balance_callback = &balance_push_callback;
10009		rq->active_balance = 0;
10010		rq->next_balance = jiffies;
10011		rq->push_cpu = 0;
10012		rq->cpu = i;
10013		rq->online = 0;
10014		rq->idle_stamp = 0;
10015		rq->avg_idle = 2*sysctl_sched_migration_cost;
10016		rq->max_idle_balance_cost = sysctl_sched_migration_cost;
10017
10018		INIT_LIST_HEAD(&rq->cfs_tasks);
10019
10020		rq_attach_root(rq, &def_root_domain);
10021#ifdef CONFIG_NO_HZ_COMMON
 
10022		rq->last_blocked_load_update_tick = jiffies;
10023		atomic_set(&rq->nohz_flags, 0);
10024
10025		INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
10026#endif
10027#ifdef CONFIG_HOTPLUG_CPU
10028		rcuwait_init(&rq->hotplug_wait);
10029#endif
10030#endif /* CONFIG_SMP */
10031		hrtick_rq_init(rq);
10032		atomic_set(&rq->nr_iowait, 0);
10033
10034#ifdef CONFIG_SCHED_CORE
10035		rq->core = rq;
10036		rq->core_pick = NULL;
10037		rq->core_enabled = 0;
10038		rq->core_tree = RB_ROOT;
10039		rq->core_forceidle_count = 0;
10040		rq->core_forceidle_occupation = 0;
10041		rq->core_forceidle_start = 0;
10042
10043		rq->core_cookie = 0UL;
10044#endif
10045		zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i));
10046	}
10047
10048	set_load_weight(&init_task, false);
10049
10050	/*
10051	 * The boot idle thread does lazy MMU switching as well:
10052	 */
10053	mmgrab_lazy_tlb(&init_mm);
10054	enter_lazy_tlb(&init_mm, current);
10055
10056	/*
10057	 * The idle task doesn't need the kthread struct to function, but it
10058	 * is dressed up as a per-CPU kthread and thus needs to play the part
10059	 * if we want to avoid special-casing it in code that deals with per-CPU
10060	 * kthreads.
10061	 */
10062	WARN_ON(!set_kthread_struct(current));
10063
10064	/*
10065	 * Make us the idle thread. Technically, schedule() should not be
10066	 * called from this thread, however somewhere below it might be,
10067	 * but because we are the idle thread, we just pick up running again
10068	 * when this runqueue becomes "idle".
10069	 */
10070	init_idle(current, smp_processor_id());
10071
10072	calc_load_update = jiffies + LOAD_FREQ;
10073
10074#ifdef CONFIG_SMP
10075	idle_thread_set_boot_cpu();
10076	balance_push_set(smp_processor_id(), false);
10077#endif
10078	init_sched_fair_class();
10079
10080	psi_init();
10081
10082	init_uclamp();
10083
10084	preempt_dynamic_init();
10085
10086	scheduler_running = 1;
10087}
10088
10089#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
 
 
 
 
 
 
10090
10091void __might_sleep(const char *file, int line)
10092{
10093	unsigned int state = get_current_state();
10094	/*
10095	 * Blocking primitives will set (and therefore destroy) current->state,
10096	 * since we will exit with TASK_RUNNING make sure we enter with it,
10097	 * otherwise we will destroy state.
10098	 */
10099	WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
10100			"do not call blocking ops when !TASK_RUNNING; "
10101			"state=%x set at [<%p>] %pS\n", state,
 
10102			(void *)current->task_state_change,
10103			(void *)current->task_state_change);
10104
10105	__might_resched(file, line, 0);
10106}
10107EXPORT_SYMBOL(__might_sleep);
10108
10109static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
10110{
10111	if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
10112		return;
10113
10114	if (preempt_count() == preempt_offset)
10115		return;
10116
10117	pr_err("Preemption disabled at:");
10118	print_ip_sym(KERN_ERR, ip);
10119}
10120
10121static inline bool resched_offsets_ok(unsigned int offsets)
10122{
10123	unsigned int nested = preempt_count();
10124
10125	nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
10126
10127	return nested == offsets;
10128}
10129
10130void __might_resched(const char *file, int line, unsigned int offsets)
10131{
10132	/* Ratelimiting timestamp: */
10133	static unsigned long prev_jiffy;
10134
10135	unsigned long preempt_disable_ip;
10136
10137	/* WARN_ON_ONCE() by default, no rate limit required: */
10138	rcu_sleep_check();
10139
10140	if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
10141	     !is_idle_task(current) && !current->non_block_count) ||
10142	    system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
10143	    oops_in_progress)
10144		return;
10145
10146	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10147		return;
10148	prev_jiffy = jiffies;
10149
10150	/* Save this before calling printk(), since that will clobber it: */
10151	preempt_disable_ip = get_preempt_disable_ip(current);
10152
10153	pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10154	       file, line);
10155	pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10156	       in_atomic(), irqs_disabled(), current->non_block_count,
10157	       current->pid, current->comm);
10158	pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
10159	       offsets & MIGHT_RESCHED_PREEMPT_MASK);
10160
10161	if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
10162		pr_err("RCU nest depth: %d, expected: %u\n",
10163		       rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
10164	}
10165
10166	if (task_stack_end_corrupted(current))
10167		pr_emerg("Thread overran stack, or stack corrupted\n");
10168
10169	debug_show_held_locks(current);
10170	if (irqs_disabled())
10171		print_irqtrace_events(current);
10172
10173	print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
10174				 preempt_disable_ip);
10175
 
 
10176	dump_stack();
10177	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10178}
10179EXPORT_SYMBOL(__might_resched);
10180
10181void __cant_sleep(const char *file, int line, int preempt_offset)
10182{
10183	static unsigned long prev_jiffy;
10184
10185	if (irqs_disabled())
10186		return;
10187
10188	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10189		return;
10190
10191	if (preempt_count() > preempt_offset)
10192		return;
10193
10194	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10195		return;
10196	prev_jiffy = jiffies;
10197
10198	printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
10199	printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10200			in_atomic(), irqs_disabled(),
10201			current->pid, current->comm);
10202
10203	debug_show_held_locks(current);
10204	dump_stack();
10205	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10206}
10207EXPORT_SYMBOL_GPL(__cant_sleep);
10208
10209#ifdef CONFIG_SMP
10210void __cant_migrate(const char *file, int line)
10211{
10212	static unsigned long prev_jiffy;
10213
10214	if (irqs_disabled())
10215		return;
10216
10217	if (is_migration_disabled(current))
10218		return;
10219
10220	if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10221		return;
10222
10223	if (preempt_count() > 0)
10224		return;
10225
10226	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10227		return;
10228	prev_jiffy = jiffies;
10229
10230	pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
10231	pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10232	       in_atomic(), irqs_disabled(), is_migration_disabled(current),
10233	       current->pid, current->comm);
10234
10235	debug_show_held_locks(current);
10236	dump_stack();
10237	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10238}
10239EXPORT_SYMBOL_GPL(__cant_migrate);
10240#endif
10241#endif
10242
10243#ifdef CONFIG_MAGIC_SYSRQ
10244void normalize_rt_tasks(void)
10245{
10246	struct task_struct *g, *p;
10247	struct sched_attr attr = {
10248		.sched_policy = SCHED_NORMAL,
10249	};
10250
10251	read_lock(&tasklist_lock);
10252	for_each_process_thread(g, p) {
10253		/*
10254		 * Only normalize user tasks:
10255		 */
10256		if (p->flags & PF_KTHREAD)
10257			continue;
10258
10259		p->se.exec_start = 0;
10260		schedstat_set(p->stats.wait_start,  0);
10261		schedstat_set(p->stats.sleep_start, 0);
10262		schedstat_set(p->stats.block_start, 0);
10263
10264		if (!dl_task(p) && !rt_task(p)) {
10265			/*
10266			 * Renice negative nice level userspace
10267			 * tasks back to 0:
10268			 */
10269			if (task_nice(p) < 0)
10270				set_user_nice(p, 0);
10271			continue;
10272		}
10273
10274		__sched_setscheduler(p, &attr, false, false);
10275	}
10276	read_unlock(&tasklist_lock);
10277}
10278
10279#endif /* CONFIG_MAGIC_SYSRQ */
10280
10281#if defined(CONFIG_KGDB_KDB)
10282/*
10283 * These functions are only useful for kdb.
10284 *
10285 * They can only be called when the whole system has been
10286 * stopped - every CPU needs to be quiescent, and no scheduling
10287 * activity can take place. Using them for anything else would
10288 * be a serious bug, and as a result, they aren't even visible
10289 * under any other configuration.
10290 */
10291
10292/**
10293 * curr_task - return the current task for a given CPU.
10294 * @cpu: the processor in question.
10295 *
10296 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10297 *
10298 * Return: The current task for @cpu.
10299 */
10300struct task_struct *curr_task(int cpu)
10301{
10302	return cpu_curr(cpu);
10303}
10304
10305#endif /* defined(CONFIG_KGDB_KDB) */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
10306
10307#ifdef CONFIG_CGROUP_SCHED
10308/* task_group_lock serializes the addition/removal of task groups */
10309static DEFINE_SPINLOCK(task_group_lock);
10310
10311static inline void alloc_uclamp_sched_group(struct task_group *tg,
10312					    struct task_group *parent)
10313{
10314#ifdef CONFIG_UCLAMP_TASK_GROUP
10315	enum uclamp_id clamp_id;
10316
10317	for_each_clamp_id(clamp_id) {
10318		uclamp_se_set(&tg->uclamp_req[clamp_id],
10319			      uclamp_none(clamp_id), false);
10320		tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10321	}
10322#endif
10323}
10324
10325static void sched_free_group(struct task_group *tg)
10326{
10327	free_fair_sched_group(tg);
10328	free_rt_sched_group(tg);
10329	autogroup_free(tg);
10330	kmem_cache_free(task_group_cache, tg);
10331}
10332
10333static void sched_free_group_rcu(struct rcu_head *rcu)
10334{
10335	sched_free_group(container_of(rcu, struct task_group, rcu));
10336}
10337
10338static void sched_unregister_group(struct task_group *tg)
10339{
10340	unregister_fair_sched_group(tg);
10341	unregister_rt_sched_group(tg);
10342	/*
10343	 * We have to wait for yet another RCU grace period to expire, as
10344	 * print_cfs_stats() might run concurrently.
10345	 */
10346	call_rcu(&tg->rcu, sched_free_group_rcu);
10347}
10348
10349/* allocate runqueue etc for a new task group */
10350struct task_group *sched_create_group(struct task_group *parent)
10351{
10352	struct task_group *tg;
10353
10354	tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10355	if (!tg)
10356		return ERR_PTR(-ENOMEM);
10357
10358	if (!alloc_fair_sched_group(tg, parent))
10359		goto err;
10360
10361	if (!alloc_rt_sched_group(tg, parent))
10362		goto err;
10363
10364	alloc_uclamp_sched_group(tg, parent);
10365
10366	return tg;
10367
10368err:
10369	sched_free_group(tg);
10370	return ERR_PTR(-ENOMEM);
10371}
10372
10373void sched_online_group(struct task_group *tg, struct task_group *parent)
10374{
10375	unsigned long flags;
10376
10377	spin_lock_irqsave(&task_group_lock, flags);
10378	list_add_rcu(&tg->list, &task_groups);
10379
10380	/* Root should already exist: */
10381	WARN_ON(!parent);
10382
10383	tg->parent = parent;
10384	INIT_LIST_HEAD(&tg->children);
10385	list_add_rcu(&tg->siblings, &parent->children);
10386	spin_unlock_irqrestore(&task_group_lock, flags);
10387
10388	online_fair_sched_group(tg);
10389}
10390
10391/* rcu callback to free various structures associated with a task group */
10392static void sched_unregister_group_rcu(struct rcu_head *rhp)
10393{
10394	/* Now it should be safe to free those cfs_rqs: */
10395	sched_unregister_group(container_of(rhp, struct task_group, rcu));
10396}
10397
10398void sched_destroy_group(struct task_group *tg)
10399{
10400	/* Wait for possible concurrent references to cfs_rqs complete: */
10401	call_rcu(&tg->rcu, sched_unregister_group_rcu);
10402}
10403
10404void sched_release_group(struct task_group *tg)
10405{
10406	unsigned long flags;
10407
10408	/*
10409	 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10410	 * sched_cfs_period_timer()).
10411	 *
10412	 * For this to be effective, we have to wait for all pending users of
10413	 * this task group to leave their RCU critical section to ensure no new
10414	 * user will see our dying task group any more. Specifically ensure
10415	 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10416	 *
10417	 * We therefore defer calling unregister_fair_sched_group() to
10418	 * sched_unregister_group() which is guarantied to get called only after the
10419	 * current RCU grace period has expired.
10420	 */
10421	spin_lock_irqsave(&task_group_lock, flags);
10422	list_del_rcu(&tg->list);
10423	list_del_rcu(&tg->siblings);
10424	spin_unlock_irqrestore(&task_group_lock, flags);
10425}
10426
10427static struct task_group *sched_get_task_group(struct task_struct *tsk)
10428{
10429	struct task_group *tg;
10430
10431	/*
10432	 * All callers are synchronized by task_rq_lock(); we do not use RCU
10433	 * which is pointless here. Thus, we pass "true" to task_css_check()
10434	 * to prevent lockdep warnings.
10435	 */
10436	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10437			  struct task_group, css);
10438	tg = autogroup_task_group(tsk, tg);
10439
10440	return tg;
10441}
10442
10443static void sched_change_group(struct task_struct *tsk, struct task_group *group)
10444{
10445	tsk->sched_task_group = group;
10446
10447#ifdef CONFIG_FAIR_GROUP_SCHED
10448	if (tsk->sched_class->task_change_group)
10449		tsk->sched_class->task_change_group(tsk);
10450	else
10451#endif
10452		set_task_rq(tsk, task_cpu(tsk));
10453}
10454
10455/*
10456 * Change task's runqueue when it moves between groups.
10457 *
10458 * The caller of this function should have put the task in its new group by
10459 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10460 * its new group.
10461 */
10462void sched_move_task(struct task_struct *tsk)
10463{
10464	int queued, running, queue_flags =
10465		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10466	struct task_group *group;
10467	struct rq *rq;
10468
10469	CLASS(task_rq_lock, rq_guard)(tsk);
10470	rq = rq_guard.rq;
10471
10472	/*
10473	 * Esp. with SCHED_AUTOGROUP enabled it is possible to get superfluous
10474	 * group changes.
10475	 */
10476	group = sched_get_task_group(tsk);
10477	if (group == tsk->sched_task_group)
10478		return;
10479
10480	update_rq_clock(rq);
10481
10482	running = task_current(rq, tsk);
10483	queued = task_on_rq_queued(tsk);
10484
10485	if (queued)
10486		dequeue_task(rq, tsk, queue_flags);
10487	if (running)
10488		put_prev_task(rq, tsk);
10489
10490	sched_change_group(tsk, group);
10491
10492	if (queued)
10493		enqueue_task(rq, tsk, queue_flags);
10494	if (running) {
10495		set_next_task(rq, tsk);
10496		/*
10497		 * After changing group, the running task may have joined a
10498		 * throttled one but it's still the running task. Trigger a
10499		 * resched to make sure that task can still run.
10500		 */
10501		resched_curr(rq);
10502	}
10503}
10504
10505static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10506{
10507	return css ? container_of(css, struct task_group, css) : NULL;
10508}
10509
10510static struct cgroup_subsys_state *
10511cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10512{
10513	struct task_group *parent = css_tg(parent_css);
10514	struct task_group *tg;
10515
10516	if (!parent) {
10517		/* This is early initialization for the top cgroup */
10518		return &root_task_group.css;
10519	}
10520
10521	tg = sched_create_group(parent);
10522	if (IS_ERR(tg))
10523		return ERR_PTR(-ENOMEM);
10524
10525	return &tg->css;
10526}
10527
10528/* Expose task group only after completing cgroup initialization */
10529static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10530{
10531	struct task_group *tg = css_tg(css);
10532	struct task_group *parent = css_tg(css->parent);
10533
10534	if (parent)
10535		sched_online_group(tg, parent);
10536
10537#ifdef CONFIG_UCLAMP_TASK_GROUP
10538	/* Propagate the effective uclamp value for the new group */
10539	guard(mutex)(&uclamp_mutex);
10540	guard(rcu)();
10541	cpu_util_update_eff(css);
10542#endif
10543
10544	return 0;
10545}
10546
10547static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10548{
10549	struct task_group *tg = css_tg(css);
10550
10551	sched_release_group(tg);
10552}
10553
10554static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10555{
10556	struct task_group *tg = css_tg(css);
10557
10558	/*
10559	 * Relies on the RCU grace period between css_released() and this.
10560	 */
10561	sched_unregister_group(tg);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
10562}
10563
10564#ifdef CONFIG_RT_GROUP_SCHED
10565static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10566{
10567	struct task_struct *task;
10568	struct cgroup_subsys_state *css;
 
10569
10570	cgroup_taskset_for_each(task, css, tset) {
 
10571		if (!sched_rt_can_attach(css_tg(css), task))
10572			return -EINVAL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
10573	}
10574	return 0;
10575}
10576#endif
10577
10578static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10579{
10580	struct task_struct *task;
10581	struct cgroup_subsys_state *css;
10582
10583	cgroup_taskset_for_each(task, css, tset)
10584		sched_move_task(task);
10585}
10586
10587#ifdef CONFIG_UCLAMP_TASK_GROUP
10588static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10589{
10590	struct cgroup_subsys_state *top_css = css;
10591	struct uclamp_se *uc_parent = NULL;
10592	struct uclamp_se *uc_se = NULL;
10593	unsigned int eff[UCLAMP_CNT];
10594	enum uclamp_id clamp_id;
10595	unsigned int clamps;
10596
10597	lockdep_assert_held(&uclamp_mutex);
10598	SCHED_WARN_ON(!rcu_read_lock_held());
10599
10600	css_for_each_descendant_pre(css, top_css) {
10601		uc_parent = css_tg(css)->parent
10602			? css_tg(css)->parent->uclamp : NULL;
10603
10604		for_each_clamp_id(clamp_id) {
10605			/* Assume effective clamps matches requested clamps */
10606			eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10607			/* Cap effective clamps with parent's effective clamps */
10608			if (uc_parent &&
10609			    eff[clamp_id] > uc_parent[clamp_id].value) {
10610				eff[clamp_id] = uc_parent[clamp_id].value;
10611			}
10612		}
10613		/* Ensure protection is always capped by limit */
10614		eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10615
10616		/* Propagate most restrictive effective clamps */
10617		clamps = 0x0;
10618		uc_se = css_tg(css)->uclamp;
10619		for_each_clamp_id(clamp_id) {
10620			if (eff[clamp_id] == uc_se[clamp_id].value)
10621				continue;
10622			uc_se[clamp_id].value = eff[clamp_id];
10623			uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10624			clamps |= (0x1 << clamp_id);
10625		}
10626		if (!clamps) {
10627			css = css_rightmost_descendant(css);
10628			continue;
10629		}
10630
10631		/* Immediately update descendants RUNNABLE tasks */
10632		uclamp_update_active_tasks(css);
10633	}
10634}
10635
10636/*
10637 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10638 * C expression. Since there is no way to convert a macro argument (N) into a
10639 * character constant, use two levels of macros.
10640 */
10641#define _POW10(exp) ((unsigned int)1e##exp)
10642#define POW10(exp) _POW10(exp)
10643
10644struct uclamp_request {
10645#define UCLAMP_PERCENT_SHIFT	2
10646#define UCLAMP_PERCENT_SCALE	(100 * POW10(UCLAMP_PERCENT_SHIFT))
10647	s64 percent;
10648	u64 util;
10649	int ret;
10650};
10651
10652static inline struct uclamp_request
10653capacity_from_percent(char *buf)
10654{
10655	struct uclamp_request req = {
10656		.percent = UCLAMP_PERCENT_SCALE,
10657		.util = SCHED_CAPACITY_SCALE,
10658		.ret = 0,
10659	};
10660
10661	buf = strim(buf);
10662	if (strcmp(buf, "max")) {
10663		req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10664					     &req.percent);
10665		if (req.ret)
10666			return req;
10667		if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10668			req.ret = -ERANGE;
10669			return req;
10670		}
10671
10672		req.util = req.percent << SCHED_CAPACITY_SHIFT;
10673		req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10674	}
10675
10676	return req;
10677}
10678
10679static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10680				size_t nbytes, loff_t off,
10681				enum uclamp_id clamp_id)
10682{
10683	struct uclamp_request req;
10684	struct task_group *tg;
10685
10686	req = capacity_from_percent(buf);
10687	if (req.ret)
10688		return req.ret;
10689
10690	static_branch_enable(&sched_uclamp_used);
10691
10692	guard(mutex)(&uclamp_mutex);
10693	guard(rcu)();
10694
10695	tg = css_tg(of_css(of));
10696	if (tg->uclamp_req[clamp_id].value != req.util)
10697		uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10698
10699	/*
10700	 * Because of not recoverable conversion rounding we keep track of the
10701	 * exact requested value
10702	 */
10703	tg->uclamp_pct[clamp_id] = req.percent;
10704
10705	/* Update effective clamps to track the most restrictive value */
10706	cpu_util_update_eff(of_css(of));
10707
10708	return nbytes;
10709}
10710
10711static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10712				    char *buf, size_t nbytes,
10713				    loff_t off)
10714{
10715	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10716}
10717
10718static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10719				    char *buf, size_t nbytes,
10720				    loff_t off)
10721{
10722	return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10723}
10724
10725static inline void cpu_uclamp_print(struct seq_file *sf,
10726				    enum uclamp_id clamp_id)
10727{
10728	struct task_group *tg;
10729	u64 util_clamp;
10730	u64 percent;
10731	u32 rem;
10732
10733	scoped_guard (rcu) {
10734		tg = css_tg(seq_css(sf));
10735		util_clamp = tg->uclamp_req[clamp_id].value;
10736	}
10737
10738	if (util_clamp == SCHED_CAPACITY_SCALE) {
10739		seq_puts(sf, "max\n");
10740		return;
10741	}
10742
10743	percent = tg->uclamp_pct[clamp_id];
10744	percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10745	seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10746}
10747
10748static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10749{
10750	cpu_uclamp_print(sf, UCLAMP_MIN);
10751	return 0;
10752}
10753
10754static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10755{
10756	cpu_uclamp_print(sf, UCLAMP_MAX);
10757	return 0;
10758}
10759#endif /* CONFIG_UCLAMP_TASK_GROUP */
10760
10761#ifdef CONFIG_FAIR_GROUP_SCHED
10762static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10763				struct cftype *cftype, u64 shareval)
10764{
10765	if (shareval > scale_load_down(ULONG_MAX))
10766		shareval = MAX_SHARES;
10767	return sched_group_set_shares(css_tg(css), scale_load(shareval));
10768}
10769
10770static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10771			       struct cftype *cft)
10772{
10773	struct task_group *tg = css_tg(css);
10774
10775	return (u64) scale_load_down(tg->shares);
10776}
10777
10778#ifdef CONFIG_CFS_BANDWIDTH
10779static DEFINE_MUTEX(cfs_constraints_mutex);
10780
10781const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10782static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10783/* More than 203 days if BW_SHIFT equals 20. */
10784static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10785
10786static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10787
10788static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10789				u64 burst)
10790{
10791	int i, ret = 0, runtime_enabled, runtime_was_enabled;
10792	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10793
10794	if (tg == &root_task_group)
10795		return -EINVAL;
10796
10797	/*
10798	 * Ensure we have at some amount of bandwidth every period.  This is
10799	 * to prevent reaching a state of large arrears when throttled via
10800	 * entity_tick() resulting in prolonged exit starvation.
10801	 */
10802	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10803		return -EINVAL;
10804
10805	/*
10806	 * Likewise, bound things on the other side by preventing insane quota
10807	 * periods.  This also allows us to normalize in computing quota
10808	 * feasibility.
10809	 */
10810	if (period > max_cfs_quota_period)
10811		return -EINVAL;
10812
10813	/*
10814	 * Bound quota to defend quota against overflow during bandwidth shift.
10815	 */
10816	if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10817		return -EINVAL;
10818
10819	if (quota != RUNTIME_INF && (burst > quota ||
10820				     burst + quota > max_cfs_runtime))
10821		return -EINVAL;
10822
10823	/*
10824	 * Prevent race between setting of cfs_rq->runtime_enabled and
10825	 * unthrottle_offline_cfs_rqs().
10826	 */
10827	guard(cpus_read_lock)();
10828	guard(mutex)(&cfs_constraints_mutex);
10829
10830	ret = __cfs_schedulable(tg, period, quota);
10831	if (ret)
10832		return ret;
10833
10834	runtime_enabled = quota != RUNTIME_INF;
10835	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10836	/*
10837	 * If we need to toggle cfs_bandwidth_used, off->on must occur
10838	 * before making related changes, and on->off must occur afterwards
10839	 */
10840	if (runtime_enabled && !runtime_was_enabled)
10841		cfs_bandwidth_usage_inc();
 
 
 
10842
10843	scoped_guard (raw_spinlock_irq, &cfs_b->lock) {
10844		cfs_b->period = ns_to_ktime(period);
10845		cfs_b->quota = quota;
10846		cfs_b->burst = burst;
10847
10848		__refill_cfs_bandwidth_runtime(cfs_b);
 
 
10849
10850		/*
10851		 * Restart the period timer (if active) to handle new
10852		 * period expiry:
10853		 */
10854		if (runtime_enabled)
10855			start_cfs_bandwidth(cfs_b);
10856	}
10857
10858	for_each_online_cpu(i) {
10859		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10860		struct rq *rq = cfs_rq->rq;
 
10861
10862		guard(rq_lock_irq)(rq);
10863		cfs_rq->runtime_enabled = runtime_enabled;
10864		cfs_rq->runtime_remaining = 0;
10865
10866		if (cfs_rq->throttled)
10867			unthrottle_cfs_rq(cfs_rq);
 
10868	}
10869
10870	if (runtime_was_enabled && !runtime_enabled)
10871		cfs_bandwidth_usage_dec();
 
 
 
10872
10873	return 0;
10874}
10875
10876static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10877{
10878	u64 quota, period, burst;
10879
10880	period = ktime_to_ns(tg->cfs_bandwidth.period);
10881	burst = tg->cfs_bandwidth.burst;
10882	if (cfs_quota_us < 0)
10883		quota = RUNTIME_INF;
10884	else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10885		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10886	else
10887		return -EINVAL;
10888
10889	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10890}
10891
10892static long tg_get_cfs_quota(struct task_group *tg)
10893{
10894	u64 quota_us;
10895
10896	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10897		return -1;
10898
10899	quota_us = tg->cfs_bandwidth.quota;
10900	do_div(quota_us, NSEC_PER_USEC);
10901
10902	return quota_us;
10903}
10904
10905static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10906{
10907	u64 quota, period, burst;
10908
10909	if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10910		return -EINVAL;
10911
10912	period = (u64)cfs_period_us * NSEC_PER_USEC;
10913	quota = tg->cfs_bandwidth.quota;
10914	burst = tg->cfs_bandwidth.burst;
10915
10916	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10917}
10918
10919static long tg_get_cfs_period(struct task_group *tg)
10920{
10921	u64 cfs_period_us;
10922
10923	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10924	do_div(cfs_period_us, NSEC_PER_USEC);
10925
10926	return cfs_period_us;
10927}
10928
10929static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10930{
10931	u64 quota, period, burst;
10932
10933	if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10934		return -EINVAL;
10935
10936	burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10937	period = ktime_to_ns(tg->cfs_bandwidth.period);
10938	quota = tg->cfs_bandwidth.quota;
10939
10940	return tg_set_cfs_bandwidth(tg, period, quota, burst);
10941}
10942
10943static long tg_get_cfs_burst(struct task_group *tg)
10944{
10945	u64 burst_us;
10946
10947	burst_us = tg->cfs_bandwidth.burst;
10948	do_div(burst_us, NSEC_PER_USEC);
10949
10950	return burst_us;
10951}
10952
10953static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10954				  struct cftype *cft)
10955{
10956	return tg_get_cfs_quota(css_tg(css));
10957}
10958
10959static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10960				   struct cftype *cftype, s64 cfs_quota_us)
10961{
10962	return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10963}
10964
10965static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10966				   struct cftype *cft)
10967{
10968	return tg_get_cfs_period(css_tg(css));
10969}
10970
10971static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10972				    struct cftype *cftype, u64 cfs_period_us)
10973{
10974	return tg_set_cfs_period(css_tg(css), cfs_period_us);
10975}
10976
10977static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10978				  struct cftype *cft)
10979{
10980	return tg_get_cfs_burst(css_tg(css));
10981}
10982
10983static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10984				   struct cftype *cftype, u64 cfs_burst_us)
10985{
10986	return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10987}
10988
10989struct cfs_schedulable_data {
10990	struct task_group *tg;
10991	u64 period, quota;
10992};
10993
10994/*
10995 * normalize group quota/period to be quota/max_period
10996 * note: units are usecs
10997 */
10998static u64 normalize_cfs_quota(struct task_group *tg,
10999			       struct cfs_schedulable_data *d)
11000{
11001	u64 quota, period;
11002
11003	if (tg == d->tg) {
11004		period = d->period;
11005		quota = d->quota;
11006	} else {
11007		period = tg_get_cfs_period(tg);
11008		quota = tg_get_cfs_quota(tg);
11009	}
11010
11011	/* note: these should typically be equivalent */
11012	if (quota == RUNTIME_INF || quota == -1)
11013		return RUNTIME_INF;
11014
11015	return to_ratio(period, quota);
11016}
11017
11018static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
11019{
11020	struct cfs_schedulable_data *d = data;
11021	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11022	s64 quota = 0, parent_quota = -1;
11023
11024	if (!tg->parent) {
11025		quota = RUNTIME_INF;
11026	} else {
11027		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
11028
11029		quota = normalize_cfs_quota(tg, d);
11030		parent_quota = parent_b->hierarchical_quota;
11031
11032		/*
11033		 * Ensure max(child_quota) <= parent_quota.  On cgroup2,
11034		 * always take the non-RUNTIME_INF min.  On cgroup1, only
11035		 * inherit when no limit is set. In both cases this is used
11036		 * by the scheduler to determine if a given CFS task has a
11037		 * bandwidth constraint at some higher level.
11038		 */
11039		if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
11040			if (quota == RUNTIME_INF)
11041				quota = parent_quota;
11042			else if (parent_quota != RUNTIME_INF)
11043				quota = min(quota, parent_quota);
11044		} else {
11045			if (quota == RUNTIME_INF)
11046				quota = parent_quota;
11047			else if (parent_quota != RUNTIME_INF && quota > parent_quota)
11048				return -EINVAL;
11049		}
11050	}
11051	cfs_b->hierarchical_quota = quota;
11052
11053	return 0;
11054}
11055
11056static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
11057{
 
11058	struct cfs_schedulable_data data = {
11059		.tg = tg,
11060		.period = period,
11061		.quota = quota,
11062	};
11063
11064	if (quota != RUNTIME_INF) {
11065		do_div(data.period, NSEC_PER_USEC);
11066		do_div(data.quota, NSEC_PER_USEC);
11067	}
11068
11069	guard(rcu)();
11070	return walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
 
 
 
11071}
11072
11073static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
11074{
11075	struct task_group *tg = css_tg(seq_css(sf));
11076	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11077
11078	seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
11079	seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
11080	seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
11081
11082	if (schedstat_enabled() && tg != &root_task_group) {
11083		struct sched_statistics *stats;
11084		u64 ws = 0;
11085		int i;
11086
11087		for_each_possible_cpu(i) {
11088			stats = __schedstats_from_se(tg->se[i]);
11089			ws += schedstat_val(stats->wait_sum);
11090		}
11091
11092		seq_printf(sf, "wait_sum %llu\n", ws);
11093	}
11094
11095	seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
11096	seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
11097
11098	return 0;
11099}
11100
11101static u64 throttled_time_self(struct task_group *tg)
11102{
11103	int i;
11104	u64 total = 0;
11105
11106	for_each_possible_cpu(i) {
11107		total += READ_ONCE(tg->cfs_rq[i]->throttled_clock_self_time);
11108	}
11109
11110	return total;
11111}
11112
11113static int cpu_cfs_local_stat_show(struct seq_file *sf, void *v)
11114{
11115	struct task_group *tg = css_tg(seq_css(sf));
11116
11117	seq_printf(sf, "throttled_time %llu\n", throttled_time_self(tg));
11118
11119	return 0;
11120}
11121#endif /* CONFIG_CFS_BANDWIDTH */
11122#endif /* CONFIG_FAIR_GROUP_SCHED */
11123
11124#ifdef CONFIG_RT_GROUP_SCHED
11125static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
11126				struct cftype *cft, s64 val)
11127{
11128	return sched_group_set_rt_runtime(css_tg(css), val);
11129}
11130
11131static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
11132			       struct cftype *cft)
11133{
11134	return sched_group_rt_runtime(css_tg(css));
11135}
11136
11137static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
11138				    struct cftype *cftype, u64 rt_period_us)
11139{
11140	return sched_group_set_rt_period(css_tg(css), rt_period_us);
11141}
11142
11143static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
11144				   struct cftype *cft)
11145{
11146	return sched_group_rt_period(css_tg(css));
11147}
11148#endif /* CONFIG_RT_GROUP_SCHED */
11149
11150#ifdef CONFIG_FAIR_GROUP_SCHED
11151static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
11152			       struct cftype *cft)
11153{
11154	return css_tg(css)->idle;
11155}
11156
11157static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
11158				struct cftype *cft, s64 idle)
11159{
11160	return sched_group_set_idle(css_tg(css), idle);
11161}
11162#endif
11163
11164static struct cftype cpu_legacy_files[] = {
11165#ifdef CONFIG_FAIR_GROUP_SCHED
11166	{
11167		.name = "shares",
11168		.read_u64 = cpu_shares_read_u64,
11169		.write_u64 = cpu_shares_write_u64,
11170	},
11171	{
11172		.name = "idle",
11173		.read_s64 = cpu_idle_read_s64,
11174		.write_s64 = cpu_idle_write_s64,
11175	},
11176#endif
11177#ifdef CONFIG_CFS_BANDWIDTH
11178	{
11179		.name = "cfs_quota_us",
11180		.read_s64 = cpu_cfs_quota_read_s64,
11181		.write_s64 = cpu_cfs_quota_write_s64,
11182	},
11183	{
11184		.name = "cfs_period_us",
11185		.read_u64 = cpu_cfs_period_read_u64,
11186		.write_u64 = cpu_cfs_period_write_u64,
11187	},
11188	{
11189		.name = "cfs_burst_us",
11190		.read_u64 = cpu_cfs_burst_read_u64,
11191		.write_u64 = cpu_cfs_burst_write_u64,
11192	},
11193	{
11194		.name = "stat",
11195		.seq_show = cpu_cfs_stat_show,
11196	},
11197	{
11198		.name = "stat.local",
11199		.seq_show = cpu_cfs_local_stat_show,
11200	},
11201#endif
11202#ifdef CONFIG_RT_GROUP_SCHED
11203	{
11204		.name = "rt_runtime_us",
11205		.read_s64 = cpu_rt_runtime_read,
11206		.write_s64 = cpu_rt_runtime_write,
11207	},
11208	{
11209		.name = "rt_period_us",
11210		.read_u64 = cpu_rt_period_read_uint,
11211		.write_u64 = cpu_rt_period_write_uint,
11212	},
11213#endif
11214#ifdef CONFIG_UCLAMP_TASK_GROUP
11215	{
11216		.name = "uclamp.min",
11217		.flags = CFTYPE_NOT_ON_ROOT,
11218		.seq_show = cpu_uclamp_min_show,
11219		.write = cpu_uclamp_min_write,
11220	},
11221	{
11222		.name = "uclamp.max",
11223		.flags = CFTYPE_NOT_ON_ROOT,
11224		.seq_show = cpu_uclamp_max_show,
11225		.write = cpu_uclamp_max_write,
11226	},
11227#endif
11228	{ }	/* Terminate */
11229};
11230
11231static int cpu_extra_stat_show(struct seq_file *sf,
11232			       struct cgroup_subsys_state *css)
11233{
11234#ifdef CONFIG_CFS_BANDWIDTH
11235	{
11236		struct task_group *tg = css_tg(css);
11237		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11238		u64 throttled_usec, burst_usec;
11239
11240		throttled_usec = cfs_b->throttled_time;
11241		do_div(throttled_usec, NSEC_PER_USEC);
11242		burst_usec = cfs_b->burst_time;
11243		do_div(burst_usec, NSEC_PER_USEC);
11244
11245		seq_printf(sf, "nr_periods %d\n"
11246			   "nr_throttled %d\n"
11247			   "throttled_usec %llu\n"
11248			   "nr_bursts %d\n"
11249			   "burst_usec %llu\n",
11250			   cfs_b->nr_periods, cfs_b->nr_throttled,
11251			   throttled_usec, cfs_b->nr_burst, burst_usec);
11252	}
11253#endif
11254	return 0;
11255}
11256
11257static int cpu_local_stat_show(struct seq_file *sf,
11258			       struct cgroup_subsys_state *css)
11259{
11260#ifdef CONFIG_CFS_BANDWIDTH
11261	{
11262		struct task_group *tg = css_tg(css);
11263		u64 throttled_self_usec;
11264
11265		throttled_self_usec = throttled_time_self(tg);
11266		do_div(throttled_self_usec, NSEC_PER_USEC);
11267
11268		seq_printf(sf, "throttled_usec %llu\n",
11269			   throttled_self_usec);
11270	}
11271#endif
11272	return 0;
11273}
11274
11275#ifdef CONFIG_FAIR_GROUP_SCHED
11276static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
11277			       struct cftype *cft)
11278{
11279	struct task_group *tg = css_tg(css);
11280	u64 weight = scale_load_down(tg->shares);
11281
11282	return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11283}
11284
11285static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11286				struct cftype *cft, u64 weight)
11287{
11288	/*
11289	 * cgroup weight knobs should use the common MIN, DFL and MAX
11290	 * values which are 1, 100 and 10000 respectively.  While it loses
11291	 * a bit of range on both ends, it maps pretty well onto the shares
11292	 * value used by scheduler and the round-trip conversions preserve
11293	 * the original value over the entire range.
11294	 */
11295	if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11296		return -ERANGE;
11297
11298	weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11299
11300	return sched_group_set_shares(css_tg(css), scale_load(weight));
11301}
11302
11303static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11304				    struct cftype *cft)
11305{
11306	unsigned long weight = scale_load_down(css_tg(css)->shares);
11307	int last_delta = INT_MAX;
11308	int prio, delta;
11309
11310	/* find the closest nice value to the current weight */
11311	for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11312		delta = abs(sched_prio_to_weight[prio] - weight);
11313		if (delta >= last_delta)
11314			break;
11315		last_delta = delta;
11316	}
11317
11318	return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11319}
11320
11321static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11322				     struct cftype *cft, s64 nice)
11323{
11324	unsigned long weight;
11325	int idx;
11326
11327	if (nice < MIN_NICE || nice > MAX_NICE)
11328		return -ERANGE;
11329
11330	idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11331	idx = array_index_nospec(idx, 40);
11332	weight = sched_prio_to_weight[idx];
11333
11334	return sched_group_set_shares(css_tg(css), scale_load(weight));
11335}
11336#endif
11337
11338static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11339						  long period, long quota)
11340{
11341	if (quota < 0)
11342		seq_puts(sf, "max");
11343	else
11344		seq_printf(sf, "%ld", quota);
11345
11346	seq_printf(sf, " %ld\n", period);
11347}
11348
11349/* caller should put the current value in *@periodp before calling */
11350static int __maybe_unused cpu_period_quota_parse(char *buf,
11351						 u64 *periodp, u64 *quotap)
11352{
11353	char tok[21];	/* U64_MAX */
11354
11355	if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11356		return -EINVAL;
11357
11358	*periodp *= NSEC_PER_USEC;
11359
11360	if (sscanf(tok, "%llu", quotap))
11361		*quotap *= NSEC_PER_USEC;
11362	else if (!strcmp(tok, "max"))
11363		*quotap = RUNTIME_INF;
11364	else
11365		return -EINVAL;
11366
11367	return 0;
11368}
11369
11370#ifdef CONFIG_CFS_BANDWIDTH
11371static int cpu_max_show(struct seq_file *sf, void *v)
11372{
11373	struct task_group *tg = css_tg(seq_css(sf));
11374
11375	cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11376	return 0;
11377}
11378
11379static ssize_t cpu_max_write(struct kernfs_open_file *of,
11380			     char *buf, size_t nbytes, loff_t off)
11381{
11382	struct task_group *tg = css_tg(of_css(of));
11383	u64 period = tg_get_cfs_period(tg);
11384	u64 burst = tg_get_cfs_burst(tg);
11385	u64 quota;
11386	int ret;
11387
11388	ret = cpu_period_quota_parse(buf, &period, &quota);
11389	if (!ret)
11390		ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11391	return ret ?: nbytes;
11392}
11393#endif
11394
11395static struct cftype cpu_files[] = {
11396#ifdef CONFIG_FAIR_GROUP_SCHED
11397	{
11398		.name = "weight",
11399		.flags = CFTYPE_NOT_ON_ROOT,
11400		.read_u64 = cpu_weight_read_u64,
11401		.write_u64 = cpu_weight_write_u64,
11402	},
11403	{
11404		.name = "weight.nice",
11405		.flags = CFTYPE_NOT_ON_ROOT,
11406		.read_s64 = cpu_weight_nice_read_s64,
11407		.write_s64 = cpu_weight_nice_write_s64,
11408	},
11409	{
11410		.name = "idle",
11411		.flags = CFTYPE_NOT_ON_ROOT,
11412		.read_s64 = cpu_idle_read_s64,
11413		.write_s64 = cpu_idle_write_s64,
11414	},
11415#endif
11416#ifdef CONFIG_CFS_BANDWIDTH
11417	{
11418		.name = "max",
11419		.flags = CFTYPE_NOT_ON_ROOT,
11420		.seq_show = cpu_max_show,
11421		.write = cpu_max_write,
11422	},
11423	{
11424		.name = "max.burst",
11425		.flags = CFTYPE_NOT_ON_ROOT,
11426		.read_u64 = cpu_cfs_burst_read_u64,
11427		.write_u64 = cpu_cfs_burst_write_u64,
11428	},
11429#endif
11430#ifdef CONFIG_UCLAMP_TASK_GROUP
11431	{
11432		.name = "uclamp.min",
11433		.flags = CFTYPE_NOT_ON_ROOT,
11434		.seq_show = cpu_uclamp_min_show,
11435		.write = cpu_uclamp_min_write,
11436	},
11437	{
11438		.name = "uclamp.max",
11439		.flags = CFTYPE_NOT_ON_ROOT,
11440		.seq_show = cpu_uclamp_max_show,
11441		.write = cpu_uclamp_max_write,
11442	},
11443#endif
11444	{ }	/* terminate */
11445};
11446
11447struct cgroup_subsys cpu_cgrp_subsys = {
11448	.css_alloc	= cpu_cgroup_css_alloc,
11449	.css_online	= cpu_cgroup_css_online,
11450	.css_released	= cpu_cgroup_css_released,
11451	.css_free	= cpu_cgroup_css_free,
11452	.css_extra_stat_show = cpu_extra_stat_show,
11453	.css_local_stat_show = cpu_local_stat_show,
11454#ifdef CONFIG_RT_GROUP_SCHED
11455	.can_attach	= cpu_cgroup_can_attach,
11456#endif
11457	.attach		= cpu_cgroup_attach,
11458	.legacy_cftypes	= cpu_legacy_files,
11459	.dfl_cftypes	= cpu_files,
11460	.early_init	= true,
11461	.threaded	= true,
11462};
11463
11464#endif	/* CONFIG_CGROUP_SCHED */
11465
11466void dump_cpu_task(int cpu)
11467{
11468	if (cpu == smp_processor_id() && in_hardirq()) {
11469		struct pt_regs *regs;
11470
11471		regs = get_irq_regs();
11472		if (regs) {
11473			show_regs(regs);
11474			return;
11475		}
11476	}
11477
11478	if (trigger_single_cpu_backtrace(cpu))
11479		return;
11480
11481	pr_info("Task dump for CPU %d:\n", cpu);
11482	sched_show_task(cpu_curr(cpu));
11483}
11484
11485/*
11486 * Nice levels are multiplicative, with a gentle 10% change for every
11487 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11488 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11489 * that remained on nice 0.
11490 *
11491 * The "10% effect" is relative and cumulative: from _any_ nice level,
11492 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11493 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11494 * If a task goes up by ~10% and another task goes down by ~10% then
11495 * the relative distance between them is ~25%.)
11496 */
11497const int sched_prio_to_weight[40] = {
11498 /* -20 */     88761,     71755,     56483,     46273,     36291,
11499 /* -15 */     29154,     23254,     18705,     14949,     11916,
11500 /* -10 */      9548,      7620,      6100,      4904,      3906,
11501 /*  -5 */      3121,      2501,      1991,      1586,      1277,
11502 /*   0 */      1024,       820,       655,       526,       423,
11503 /*   5 */       335,       272,       215,       172,       137,
11504 /*  10 */       110,        87,        70,        56,        45,
11505 /*  15 */        36,        29,        23,        18,        15,
11506};
11507
11508/*
11509 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11510 *
11511 * In cases where the weight does not change often, we can use the
11512 * precalculated inverse to speed up arithmetics by turning divisions
11513 * into multiplications:
11514 */
11515const u32 sched_prio_to_wmult[40] = {
11516 /* -20 */     48388,     59856,     76040,     92818,    118348,
11517 /* -15 */    147320,    184698,    229616,    287308,    360437,
11518 /* -10 */    449829,    563644,    704093,    875809,   1099582,
11519 /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
11520 /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
11521 /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
11522 /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
11523 /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11524};
11525
11526void call_trace_sched_update_nr_running(struct rq *rq, int count)
11527{
11528        trace_sched_update_nr_running_tp(rq, count);
11529}
11530
11531#ifdef CONFIG_SCHED_MM_CID
11532
11533/*
11534 * @cid_lock: Guarantee forward-progress of cid allocation.
11535 *
11536 * Concurrency ID allocation within a bitmap is mostly lock-free. The cid_lock
11537 * is only used when contention is detected by the lock-free allocation so
11538 * forward progress can be guaranteed.
11539 */
11540DEFINE_RAW_SPINLOCK(cid_lock);
11541
11542/*
11543 * @use_cid_lock: Select cid allocation behavior: lock-free vs spinlock.
11544 *
11545 * When @use_cid_lock is 0, the cid allocation is lock-free. When contention is
11546 * detected, it is set to 1 to ensure that all newly coming allocations are
11547 * serialized by @cid_lock until the allocation which detected contention
11548 * completes and sets @use_cid_lock back to 0. This guarantees forward progress
11549 * of a cid allocation.
11550 */
11551int use_cid_lock;
11552
11553/*
11554 * mm_cid remote-clear implements a lock-free algorithm to clear per-mm/cpu cid
11555 * concurrently with respect to the execution of the source runqueue context
11556 * switch.
11557 *
11558 * There is one basic properties we want to guarantee here:
11559 *
11560 * (1) Remote-clear should _never_ mark a per-cpu cid UNSET when it is actively
11561 * used by a task. That would lead to concurrent allocation of the cid and
11562 * userspace corruption.
11563 *
11564 * Provide this guarantee by introducing a Dekker memory ordering to guarantee
11565 * that a pair of loads observe at least one of a pair of stores, which can be
11566 * shown as:
11567 *
11568 *      X = Y = 0
11569 *
11570 *      w[X]=1          w[Y]=1
11571 *      MB              MB
11572 *      r[Y]=y          r[X]=x
11573 *
11574 * Which guarantees that x==0 && y==0 is impossible. But rather than using
11575 * values 0 and 1, this algorithm cares about specific state transitions of the
11576 * runqueue current task (as updated by the scheduler context switch), and the
11577 * per-mm/cpu cid value.
11578 *
11579 * Let's introduce task (Y) which has task->mm == mm and task (N) which has
11580 * task->mm != mm for the rest of the discussion. There are two scheduler state
11581 * transitions on context switch we care about:
11582 *
11583 * (TSA) Store to rq->curr with transition from (N) to (Y)
11584 *
11585 * (TSB) Store to rq->curr with transition from (Y) to (N)
11586 *
11587 * On the remote-clear side, there is one transition we care about:
11588 *
11589 * (TMA) cmpxchg to *pcpu_cid to set the LAZY flag
11590 *
11591 * There is also a transition to UNSET state which can be performed from all
11592 * sides (scheduler, remote-clear). It is always performed with a cmpxchg which
11593 * guarantees that only a single thread will succeed:
11594 *
11595 * (TMB) cmpxchg to *pcpu_cid to mark UNSET
11596 *
11597 * Just to be clear, what we do _not_ want to happen is a transition to UNSET
11598 * when a thread is actively using the cid (property (1)).
11599 *
11600 * Let's looks at the relevant combinations of TSA/TSB, and TMA transitions.
11601 *
11602 * Scenario A) (TSA)+(TMA) (from next task perspective)
11603 *
11604 * CPU0                                      CPU1
11605 *
11606 * Context switch CS-1                       Remote-clear
11607 *   - store to rq->curr: (N)->(Y) (TSA)     - cmpxchg to *pcpu_id to LAZY (TMA)
11608 *                                             (implied barrier after cmpxchg)
11609 *   - switch_mm_cid()
11610 *     - memory barrier (see switch_mm_cid()
11611 *       comment explaining how this barrier
11612 *       is combined with other scheduler
11613 *       barriers)
11614 *     - mm_cid_get (next)
11615 *       - READ_ONCE(*pcpu_cid)              - rcu_dereference(src_rq->curr)
11616 *
11617 * This Dekker ensures that either task (Y) is observed by the
11618 * rcu_dereference() or the LAZY flag is observed by READ_ONCE(), or both are
11619 * observed.
11620 *
11621 * If task (Y) store is observed by rcu_dereference(), it means that there is
11622 * still an active task on the cpu. Remote-clear will therefore not transition
11623 * to UNSET, which fulfills property (1).
11624 *
11625 * If task (Y) is not observed, but the lazy flag is observed by READ_ONCE(),
11626 * it will move its state to UNSET, which clears the percpu cid perhaps
11627 * uselessly (which is not an issue for correctness). Because task (Y) is not
11628 * observed, CPU1 can move ahead to set the state to UNSET. Because moving
11629 * state to UNSET is done with a cmpxchg expecting that the old state has the
11630 * LAZY flag set, only one thread will successfully UNSET.
11631 *
11632 * If both states (LAZY flag and task (Y)) are observed, the thread on CPU0
11633 * will observe the LAZY flag and transition to UNSET (perhaps uselessly), and
11634 * CPU1 will observe task (Y) and do nothing more, which is fine.
11635 *
11636 * What we are effectively preventing with this Dekker is a scenario where
11637 * neither LAZY flag nor store (Y) are observed, which would fail property (1)
11638 * because this would UNSET a cid which is actively used.
11639 */
11640
11641void sched_mm_cid_migrate_from(struct task_struct *t)
11642{
11643	t->migrate_from_cpu = task_cpu(t);
11644}
11645
11646static
11647int __sched_mm_cid_migrate_from_fetch_cid(struct rq *src_rq,
11648					  struct task_struct *t,
11649					  struct mm_cid *src_pcpu_cid)
11650{
11651	struct mm_struct *mm = t->mm;
11652	struct task_struct *src_task;
11653	int src_cid, last_mm_cid;
11654
11655	if (!mm)
11656		return -1;
11657
11658	last_mm_cid = t->last_mm_cid;
11659	/*
11660	 * If the migrated task has no last cid, or if the current
11661	 * task on src rq uses the cid, it means the source cid does not need
11662	 * to be moved to the destination cpu.
11663	 */
11664	if (last_mm_cid == -1)
11665		return -1;
11666	src_cid = READ_ONCE(src_pcpu_cid->cid);
11667	if (!mm_cid_is_valid(src_cid) || last_mm_cid != src_cid)
11668		return -1;
11669
11670	/*
11671	 * If we observe an active task using the mm on this rq, it means we
11672	 * are not the last task to be migrated from this cpu for this mm, so
11673	 * there is no need to move src_cid to the destination cpu.
11674	 */
11675	guard(rcu)();
11676	src_task = rcu_dereference(src_rq->curr);
11677	if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11678		t->last_mm_cid = -1;
11679		return -1;
11680	}
11681
11682	return src_cid;
11683}
11684
11685static
11686int __sched_mm_cid_migrate_from_try_steal_cid(struct rq *src_rq,
11687					      struct task_struct *t,
11688					      struct mm_cid *src_pcpu_cid,
11689					      int src_cid)
11690{
11691	struct task_struct *src_task;
11692	struct mm_struct *mm = t->mm;
11693	int lazy_cid;
11694
11695	if (src_cid == -1)
11696		return -1;
11697
11698	/*
11699	 * Attempt to clear the source cpu cid to move it to the destination
11700	 * cpu.
11701	 */
11702	lazy_cid = mm_cid_set_lazy_put(src_cid);
11703	if (!try_cmpxchg(&src_pcpu_cid->cid, &src_cid, lazy_cid))
11704		return -1;
11705
11706	/*
11707	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11708	 * rq->curr->mm matches the scheduler barrier in context_switch()
11709	 * between store to rq->curr and load of prev and next task's
11710	 * per-mm/cpu cid.
11711	 *
11712	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11713	 * rq->curr->mm_cid_active matches the barrier in
11714	 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11715	 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11716	 * load of per-mm/cpu cid.
11717	 */
11718
11719	/*
11720	 * If we observe an active task using the mm on this rq after setting
11721	 * the lazy-put flag, this task will be responsible for transitioning
11722	 * from lazy-put flag set to MM_CID_UNSET.
11723	 */
11724	scoped_guard (rcu) {
11725		src_task = rcu_dereference(src_rq->curr);
11726		if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11727			/*
11728			 * We observed an active task for this mm, there is therefore
11729			 * no point in moving this cid to the destination cpu.
11730			 */
11731			t->last_mm_cid = -1;
11732			return -1;
11733		}
11734	}
11735
11736	/*
11737	 * The src_cid is unused, so it can be unset.
11738	 */
11739	if (!try_cmpxchg(&src_pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11740		return -1;
11741	return src_cid;
11742}
11743
11744/*
11745 * Migration to dst cpu. Called with dst_rq lock held.
11746 * Interrupts are disabled, which keeps the window of cid ownership without the
11747 * source rq lock held small.
11748 */
11749void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t)
11750{
11751	struct mm_cid *src_pcpu_cid, *dst_pcpu_cid;
11752	struct mm_struct *mm = t->mm;
11753	int src_cid, dst_cid, src_cpu;
11754	struct rq *src_rq;
11755
11756	lockdep_assert_rq_held(dst_rq);
11757
11758	if (!mm)
11759		return;
11760	src_cpu = t->migrate_from_cpu;
11761	if (src_cpu == -1) {
11762		t->last_mm_cid = -1;
11763		return;
11764	}
11765	/*
11766	 * Move the src cid if the dst cid is unset. This keeps id
11767	 * allocation closest to 0 in cases where few threads migrate around
11768	 * many cpus.
11769	 *
11770	 * If destination cid is already set, we may have to just clear
11771	 * the src cid to ensure compactness in frequent migrations
11772	 * scenarios.
11773	 *
11774	 * It is not useful to clear the src cid when the number of threads is
11775	 * greater or equal to the number of allowed cpus, because user-space
11776	 * can expect that the number of allowed cids can reach the number of
11777	 * allowed cpus.
11778	 */
11779	dst_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(dst_rq));
11780	dst_cid = READ_ONCE(dst_pcpu_cid->cid);
11781	if (!mm_cid_is_unset(dst_cid) &&
11782	    atomic_read(&mm->mm_users) >= t->nr_cpus_allowed)
11783		return;
11784	src_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, src_cpu);
11785	src_rq = cpu_rq(src_cpu);
11786	src_cid = __sched_mm_cid_migrate_from_fetch_cid(src_rq, t, src_pcpu_cid);
11787	if (src_cid == -1)
11788		return;
11789	src_cid = __sched_mm_cid_migrate_from_try_steal_cid(src_rq, t, src_pcpu_cid,
11790							    src_cid);
11791	if (src_cid == -1)
11792		return;
11793	if (!mm_cid_is_unset(dst_cid)) {
11794		__mm_cid_put(mm, src_cid);
11795		return;
11796	}
11797	/* Move src_cid to dst cpu. */
11798	mm_cid_snapshot_time(dst_rq, mm);
11799	WRITE_ONCE(dst_pcpu_cid->cid, src_cid);
11800}
11801
11802static void sched_mm_cid_remote_clear(struct mm_struct *mm, struct mm_cid *pcpu_cid,
11803				      int cpu)
11804{
11805	struct rq *rq = cpu_rq(cpu);
11806	struct task_struct *t;
11807	int cid, lazy_cid;
11808
11809	cid = READ_ONCE(pcpu_cid->cid);
11810	if (!mm_cid_is_valid(cid))
11811		return;
11812
11813	/*
11814	 * Clear the cpu cid if it is set to keep cid allocation compact.  If
11815	 * there happens to be other tasks left on the source cpu using this
11816	 * mm, the next task using this mm will reallocate its cid on context
11817	 * switch.
11818	 */
11819	lazy_cid = mm_cid_set_lazy_put(cid);
11820	if (!try_cmpxchg(&pcpu_cid->cid, &cid, lazy_cid))
11821		return;
11822
11823	/*
11824	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11825	 * rq->curr->mm matches the scheduler barrier in context_switch()
11826	 * between store to rq->curr and load of prev and next task's
11827	 * per-mm/cpu cid.
11828	 *
11829	 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11830	 * rq->curr->mm_cid_active matches the barrier in
11831	 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11832	 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11833	 * load of per-mm/cpu cid.
11834	 */
11835
11836	/*
11837	 * If we observe an active task using the mm on this rq after setting
11838	 * the lazy-put flag, that task will be responsible for transitioning
11839	 * from lazy-put flag set to MM_CID_UNSET.
11840	 */
11841	scoped_guard (rcu) {
11842		t = rcu_dereference(rq->curr);
11843		if (READ_ONCE(t->mm_cid_active) && t->mm == mm)
11844			return;
11845	}
11846
11847	/*
11848	 * The cid is unused, so it can be unset.
11849	 * Disable interrupts to keep the window of cid ownership without rq
11850	 * lock small.
11851	 */
11852	scoped_guard (irqsave) {
11853		if (try_cmpxchg(&pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11854			__mm_cid_put(mm, cid);
11855	}
11856}
11857
11858static void sched_mm_cid_remote_clear_old(struct mm_struct *mm, int cpu)
11859{
11860	struct rq *rq = cpu_rq(cpu);
11861	struct mm_cid *pcpu_cid;
11862	struct task_struct *curr;
11863	u64 rq_clock;
11864
11865	/*
11866	 * rq->clock load is racy on 32-bit but one spurious clear once in a
11867	 * while is irrelevant.
11868	 */
11869	rq_clock = READ_ONCE(rq->clock);
11870	pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11871
11872	/*
11873	 * In order to take care of infrequently scheduled tasks, bump the time
11874	 * snapshot associated with this cid if an active task using the mm is
11875	 * observed on this rq.
11876	 */
11877	scoped_guard (rcu) {
11878		curr = rcu_dereference(rq->curr);
11879		if (READ_ONCE(curr->mm_cid_active) && curr->mm == mm) {
11880			WRITE_ONCE(pcpu_cid->time, rq_clock);
11881			return;
11882		}
11883	}
11884
11885	if (rq_clock < pcpu_cid->time + SCHED_MM_CID_PERIOD_NS)
11886		return;
11887	sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11888}
11889
11890static void sched_mm_cid_remote_clear_weight(struct mm_struct *mm, int cpu,
11891					     int weight)
11892{
11893	struct mm_cid *pcpu_cid;
11894	int cid;
11895
11896	pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11897	cid = READ_ONCE(pcpu_cid->cid);
11898	if (!mm_cid_is_valid(cid) || cid < weight)
11899		return;
11900	sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11901}
11902
11903static void task_mm_cid_work(struct callback_head *work)
11904{
11905	unsigned long now = jiffies, old_scan, next_scan;
11906	struct task_struct *t = current;
11907	struct cpumask *cidmask;
11908	struct mm_struct *mm;
11909	int weight, cpu;
11910
11911	SCHED_WARN_ON(t != container_of(work, struct task_struct, cid_work));
11912
11913	work->next = work;	/* Prevent double-add */
11914	if (t->flags & PF_EXITING)
11915		return;
11916	mm = t->mm;
11917	if (!mm)
11918		return;
11919	old_scan = READ_ONCE(mm->mm_cid_next_scan);
11920	next_scan = now + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11921	if (!old_scan) {
11922		unsigned long res;
11923
11924		res = cmpxchg(&mm->mm_cid_next_scan, old_scan, next_scan);
11925		if (res != old_scan)
11926			old_scan = res;
11927		else
11928			old_scan = next_scan;
11929	}
11930	if (time_before(now, old_scan))
11931		return;
11932	if (!try_cmpxchg(&mm->mm_cid_next_scan, &old_scan, next_scan))
11933		return;
11934	cidmask = mm_cidmask(mm);
11935	/* Clear cids that were not recently used. */
11936	for_each_possible_cpu(cpu)
11937		sched_mm_cid_remote_clear_old(mm, cpu);
11938	weight = cpumask_weight(cidmask);
11939	/*
11940	 * Clear cids that are greater or equal to the cidmask weight to
11941	 * recompact it.
11942	 */
11943	for_each_possible_cpu(cpu)
11944		sched_mm_cid_remote_clear_weight(mm, cpu, weight);
11945}
11946
11947void init_sched_mm_cid(struct task_struct *t)
11948{
11949	struct mm_struct *mm = t->mm;
11950	int mm_users = 0;
11951
11952	if (mm) {
11953		mm_users = atomic_read(&mm->mm_users);
11954		if (mm_users == 1)
11955			mm->mm_cid_next_scan = jiffies + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11956	}
11957	t->cid_work.next = &t->cid_work;	/* Protect against double add */
11958	init_task_work(&t->cid_work, task_mm_cid_work);
11959}
11960
11961void task_tick_mm_cid(struct rq *rq, struct task_struct *curr)
11962{
11963	struct callback_head *work = &curr->cid_work;
11964	unsigned long now = jiffies;
11965
11966	if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) ||
11967	    work->next != work)
11968		return;
11969	if (time_before(now, READ_ONCE(curr->mm->mm_cid_next_scan)))
11970		return;
11971	task_work_add(curr, work, TWA_RESUME);
11972}
11973
11974void sched_mm_cid_exit_signals(struct task_struct *t)
11975{
11976	struct mm_struct *mm = t->mm;
11977	struct rq *rq;
11978
11979	if (!mm)
11980		return;
11981
11982	preempt_disable();
11983	rq = this_rq();
11984	guard(rq_lock_irqsave)(rq);
11985	preempt_enable_no_resched();	/* holding spinlock */
11986	WRITE_ONCE(t->mm_cid_active, 0);
11987	/*
11988	 * Store t->mm_cid_active before loading per-mm/cpu cid.
11989	 * Matches barrier in sched_mm_cid_remote_clear_old().
11990	 */
11991	smp_mb();
11992	mm_cid_put(mm);
11993	t->last_mm_cid = t->mm_cid = -1;
11994}
11995
11996void sched_mm_cid_before_execve(struct task_struct *t)
11997{
11998	struct mm_struct *mm = t->mm;
11999	struct rq *rq;
12000
12001	if (!mm)
12002		return;
12003
12004	preempt_disable();
12005	rq = this_rq();
12006	guard(rq_lock_irqsave)(rq);
12007	preempt_enable_no_resched();	/* holding spinlock */
12008	WRITE_ONCE(t->mm_cid_active, 0);
12009	/*
12010	 * Store t->mm_cid_active before loading per-mm/cpu cid.
12011	 * Matches barrier in sched_mm_cid_remote_clear_old().
12012	 */
12013	smp_mb();
12014	mm_cid_put(mm);
12015	t->last_mm_cid = t->mm_cid = -1;
12016}
12017
12018void sched_mm_cid_after_execve(struct task_struct *t)
12019{
12020	struct mm_struct *mm = t->mm;
12021	struct rq *rq;
12022
12023	if (!mm)
12024		return;
12025
12026	preempt_disable();
12027	rq = this_rq();
12028	scoped_guard (rq_lock_irqsave, rq) {
12029		preempt_enable_no_resched();	/* holding spinlock */
12030		WRITE_ONCE(t->mm_cid_active, 1);
12031		/*
12032		 * Store t->mm_cid_active before loading per-mm/cpu cid.
12033		 * Matches barrier in sched_mm_cid_remote_clear_old().
12034		 */
12035		smp_mb();
12036		t->last_mm_cid = t->mm_cid = mm_cid_get(rq, mm);
12037	}
12038	rseq_set_notify_resume(t);
12039}
12040
12041void sched_mm_cid_fork(struct task_struct *t)
12042{
12043	WARN_ON_ONCE(!t->mm || t->mm_cid != -1);
12044	t->mm_cid_active = 1;
12045}
12046#endif