Loading...
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 "sched.h"
10
11#include <linux/nospec.h>
12
13#include <linux/kcov.h>
14
15#include <asm/switch_to.h>
16#include <asm/tlb.h>
17
18#include "../workqueue_internal.h"
19#include "../smpboot.h"
20
21#include "pelt.h"
22
23#define CREATE_TRACE_POINTS
24#include <trace/events/sched.h>
25
26/*
27 * Export tracepoints that act as a bare tracehook (ie: have no trace event
28 * associated with them) to allow external modules to probe them.
29 */
30EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
31EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
32EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
33EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
34EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
35EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
36
37DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
38
39#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL)
40/*
41 * Debugging: various feature bits
42 *
43 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
44 * sysctl_sched_features, defined in sched.h, to allow constants propagation
45 * at compile time and compiler optimization based on features default.
46 */
47#define SCHED_FEAT(name, enabled) \
48 (1UL << __SCHED_FEAT_##name) * enabled |
49const_debug unsigned int sysctl_sched_features =
50#include "features.h"
51 0;
52#undef SCHED_FEAT
53#endif
54
55/*
56 * Number of tasks to iterate in a single balance run.
57 * Limited because this is done with IRQs disabled.
58 */
59const_debug unsigned int sysctl_sched_nr_migrate = 32;
60
61/*
62 * period over which we measure -rt task CPU usage in us.
63 * default: 1s
64 */
65unsigned int sysctl_sched_rt_period = 1000000;
66
67__read_mostly int scheduler_running;
68
69/*
70 * part of the period that we allow rt tasks to run in us.
71 * default: 0.95s
72 */
73int sysctl_sched_rt_runtime = 950000;
74
75/*
76 * __task_rq_lock - lock the rq @p resides on.
77 */
78struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
79 __acquires(rq->lock)
80{
81 struct rq *rq;
82
83 lockdep_assert_held(&p->pi_lock);
84
85 for (;;) {
86 rq = task_rq(p);
87 raw_spin_lock(&rq->lock);
88 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
89 rq_pin_lock(rq, rf);
90 return rq;
91 }
92 raw_spin_unlock(&rq->lock);
93
94 while (unlikely(task_on_rq_migrating(p)))
95 cpu_relax();
96 }
97}
98
99/*
100 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
101 */
102struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
103 __acquires(p->pi_lock)
104 __acquires(rq->lock)
105{
106 struct rq *rq;
107
108 for (;;) {
109 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
110 rq = task_rq(p);
111 raw_spin_lock(&rq->lock);
112 /*
113 * move_queued_task() task_rq_lock()
114 *
115 * ACQUIRE (rq->lock)
116 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
117 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
118 * [S] ->cpu = new_cpu [L] task_rq()
119 * [L] ->on_rq
120 * RELEASE (rq->lock)
121 *
122 * If we observe the old CPU in task_rq_lock(), the acquire of
123 * the old rq->lock will fully serialize against the stores.
124 *
125 * If we observe the new CPU in task_rq_lock(), the address
126 * dependency headed by '[L] rq = task_rq()' and the acquire
127 * will pair with the WMB to ensure we then also see migrating.
128 */
129 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
130 rq_pin_lock(rq, rf);
131 return rq;
132 }
133 raw_spin_unlock(&rq->lock);
134 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
135
136 while (unlikely(task_on_rq_migrating(p)))
137 cpu_relax();
138 }
139}
140
141/*
142 * RQ-clock updating methods:
143 */
144
145static void update_rq_clock_task(struct rq *rq, s64 delta)
146{
147/*
148 * In theory, the compile should just see 0 here, and optimize out the call
149 * to sched_rt_avg_update. But I don't trust it...
150 */
151 s64 __maybe_unused steal = 0, irq_delta = 0;
152
153#ifdef CONFIG_IRQ_TIME_ACCOUNTING
154 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
155
156 /*
157 * Since irq_time is only updated on {soft,}irq_exit, we might run into
158 * this case when a previous update_rq_clock() happened inside a
159 * {soft,}irq region.
160 *
161 * When this happens, we stop ->clock_task and only update the
162 * prev_irq_time stamp to account for the part that fit, so that a next
163 * update will consume the rest. This ensures ->clock_task is
164 * monotonic.
165 *
166 * It does however cause some slight miss-attribution of {soft,}irq
167 * time, a more accurate solution would be to update the irq_time using
168 * the current rq->clock timestamp, except that would require using
169 * atomic ops.
170 */
171 if (irq_delta > delta)
172 irq_delta = delta;
173
174 rq->prev_irq_time += irq_delta;
175 delta -= irq_delta;
176#endif
177#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
178 if (static_key_false((¶virt_steal_rq_enabled))) {
179 steal = paravirt_steal_clock(cpu_of(rq));
180 steal -= rq->prev_steal_time_rq;
181
182 if (unlikely(steal > delta))
183 steal = delta;
184
185 rq->prev_steal_time_rq += steal;
186 delta -= steal;
187 }
188#endif
189
190 rq->clock_task += delta;
191
192#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
193 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
194 update_irq_load_avg(rq, irq_delta + steal);
195#endif
196 update_rq_clock_pelt(rq, delta);
197}
198
199void update_rq_clock(struct rq *rq)
200{
201 s64 delta;
202
203 lockdep_assert_held(&rq->lock);
204
205 if (rq->clock_update_flags & RQCF_ACT_SKIP)
206 return;
207
208#ifdef CONFIG_SCHED_DEBUG
209 if (sched_feat(WARN_DOUBLE_CLOCK))
210 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
211 rq->clock_update_flags |= RQCF_UPDATED;
212#endif
213
214 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
215 if (delta < 0)
216 return;
217 rq->clock += delta;
218 update_rq_clock_task(rq, delta);
219}
220
221
222#ifdef CONFIG_SCHED_HRTICK
223/*
224 * Use HR-timers to deliver accurate preemption points.
225 */
226
227static void hrtick_clear(struct rq *rq)
228{
229 if (hrtimer_active(&rq->hrtick_timer))
230 hrtimer_cancel(&rq->hrtick_timer);
231}
232
233/*
234 * High-resolution timer tick.
235 * Runs from hardirq context with interrupts disabled.
236 */
237static enum hrtimer_restart hrtick(struct hrtimer *timer)
238{
239 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
240 struct rq_flags rf;
241
242 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
243
244 rq_lock(rq, &rf);
245 update_rq_clock(rq);
246 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
247 rq_unlock(rq, &rf);
248
249 return HRTIMER_NORESTART;
250}
251
252#ifdef CONFIG_SMP
253
254static void __hrtick_restart(struct rq *rq)
255{
256 struct hrtimer *timer = &rq->hrtick_timer;
257
258 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED_HARD);
259}
260
261/*
262 * called from hardirq (IPI) context
263 */
264static void __hrtick_start(void *arg)
265{
266 struct rq *rq = arg;
267 struct rq_flags rf;
268
269 rq_lock(rq, &rf);
270 __hrtick_restart(rq);
271 rq->hrtick_csd_pending = 0;
272 rq_unlock(rq, &rf);
273}
274
275/*
276 * Called to set the hrtick timer state.
277 *
278 * called with rq->lock held and irqs disabled
279 */
280void hrtick_start(struct rq *rq, u64 delay)
281{
282 struct hrtimer *timer = &rq->hrtick_timer;
283 ktime_t time;
284 s64 delta;
285
286 /*
287 * Don't schedule slices shorter than 10000ns, that just
288 * doesn't make sense and can cause timer DoS.
289 */
290 delta = max_t(s64, delay, 10000LL);
291 time = ktime_add_ns(timer->base->get_time(), delta);
292
293 hrtimer_set_expires(timer, time);
294
295 if (rq == this_rq()) {
296 __hrtick_restart(rq);
297 } else if (!rq->hrtick_csd_pending) {
298 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
299 rq->hrtick_csd_pending = 1;
300 }
301}
302
303#else
304/*
305 * Called to set the hrtick timer state.
306 *
307 * called with rq->lock held and irqs disabled
308 */
309void hrtick_start(struct rq *rq, u64 delay)
310{
311 /*
312 * Don't schedule slices shorter than 10000ns, that just
313 * doesn't make sense. Rely on vruntime for fairness.
314 */
315 delay = max_t(u64, delay, 10000LL);
316 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
317 HRTIMER_MODE_REL_PINNED_HARD);
318}
319#endif /* CONFIG_SMP */
320
321static void hrtick_rq_init(struct rq *rq)
322{
323#ifdef CONFIG_SMP
324 rq->hrtick_csd_pending = 0;
325
326 rq->hrtick_csd.flags = 0;
327 rq->hrtick_csd.func = __hrtick_start;
328 rq->hrtick_csd.info = rq;
329#endif
330
331 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
332 rq->hrtick_timer.function = hrtick;
333}
334#else /* CONFIG_SCHED_HRTICK */
335static inline void hrtick_clear(struct rq *rq)
336{
337}
338
339static inline void hrtick_rq_init(struct rq *rq)
340{
341}
342#endif /* CONFIG_SCHED_HRTICK */
343
344/*
345 * cmpxchg based fetch_or, macro so it works for different integer types
346 */
347#define fetch_or(ptr, mask) \
348 ({ \
349 typeof(ptr) _ptr = (ptr); \
350 typeof(mask) _mask = (mask); \
351 typeof(*_ptr) _old, _val = *_ptr; \
352 \
353 for (;;) { \
354 _old = cmpxchg(_ptr, _val, _val | _mask); \
355 if (_old == _val) \
356 break; \
357 _val = _old; \
358 } \
359 _old; \
360})
361
362#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
363/*
364 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
365 * this avoids any races wrt polling state changes and thereby avoids
366 * spurious IPIs.
367 */
368static bool set_nr_and_not_polling(struct task_struct *p)
369{
370 struct thread_info *ti = task_thread_info(p);
371 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
372}
373
374/*
375 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
376 *
377 * If this returns true, then the idle task promises to call
378 * sched_ttwu_pending() and reschedule soon.
379 */
380static bool set_nr_if_polling(struct task_struct *p)
381{
382 struct thread_info *ti = task_thread_info(p);
383 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
384
385 for (;;) {
386 if (!(val & _TIF_POLLING_NRFLAG))
387 return false;
388 if (val & _TIF_NEED_RESCHED)
389 return true;
390 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
391 if (old == val)
392 break;
393 val = old;
394 }
395 return true;
396}
397
398#else
399static bool set_nr_and_not_polling(struct task_struct *p)
400{
401 set_tsk_need_resched(p);
402 return true;
403}
404
405#ifdef CONFIG_SMP
406static bool set_nr_if_polling(struct task_struct *p)
407{
408 return false;
409}
410#endif
411#endif
412
413static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
414{
415 struct wake_q_node *node = &task->wake_q;
416
417 /*
418 * Atomically grab the task, if ->wake_q is !nil already it means
419 * its already queued (either by us or someone else) and will get the
420 * wakeup due to that.
421 *
422 * In order to ensure that a pending wakeup will observe our pending
423 * state, even in the failed case, an explicit smp_mb() must be used.
424 */
425 smp_mb__before_atomic();
426 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
427 return false;
428
429 /*
430 * The head is context local, there can be no concurrency.
431 */
432 *head->lastp = node;
433 head->lastp = &node->next;
434 return true;
435}
436
437/**
438 * wake_q_add() - queue a wakeup for 'later' waking.
439 * @head: the wake_q_head to add @task to
440 * @task: the task to queue for 'later' wakeup
441 *
442 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
443 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
444 * instantly.
445 *
446 * This function must be used as-if it were wake_up_process(); IOW the task
447 * must be ready to be woken at this location.
448 */
449void wake_q_add(struct wake_q_head *head, struct task_struct *task)
450{
451 if (__wake_q_add(head, task))
452 get_task_struct(task);
453}
454
455/**
456 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
457 * @head: the wake_q_head to add @task to
458 * @task: the task to queue for 'later' wakeup
459 *
460 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
461 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
462 * instantly.
463 *
464 * This function must be used as-if it were wake_up_process(); IOW the task
465 * must be ready to be woken at this location.
466 *
467 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
468 * that already hold reference to @task can call the 'safe' version and trust
469 * wake_q to do the right thing depending whether or not the @task is already
470 * queued for wakeup.
471 */
472void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
473{
474 if (!__wake_q_add(head, task))
475 put_task_struct(task);
476}
477
478void wake_up_q(struct wake_q_head *head)
479{
480 struct wake_q_node *node = head->first;
481
482 while (node != WAKE_Q_TAIL) {
483 struct task_struct *task;
484
485 task = container_of(node, struct task_struct, wake_q);
486 BUG_ON(!task);
487 /* Task can safely be re-inserted now: */
488 node = node->next;
489 task->wake_q.next = NULL;
490
491 /*
492 * wake_up_process() executes a full barrier, which pairs with
493 * the queueing in wake_q_add() so as not to miss wakeups.
494 */
495 wake_up_process(task);
496 put_task_struct(task);
497 }
498}
499
500/*
501 * resched_curr - mark rq's current task 'to be rescheduled now'.
502 *
503 * On UP this means the setting of the need_resched flag, on SMP it
504 * might also involve a cross-CPU call to trigger the scheduler on
505 * the target CPU.
506 */
507void resched_curr(struct rq *rq)
508{
509 struct task_struct *curr = rq->curr;
510 int cpu;
511
512 lockdep_assert_held(&rq->lock);
513
514 if (test_tsk_need_resched(curr))
515 return;
516
517 cpu = cpu_of(rq);
518
519 if (cpu == smp_processor_id()) {
520 set_tsk_need_resched(curr);
521 set_preempt_need_resched();
522 return;
523 }
524
525 if (set_nr_and_not_polling(curr))
526 smp_send_reschedule(cpu);
527 else
528 trace_sched_wake_idle_without_ipi(cpu);
529}
530
531void resched_cpu(int cpu)
532{
533 struct rq *rq = cpu_rq(cpu);
534 unsigned long flags;
535
536 raw_spin_lock_irqsave(&rq->lock, flags);
537 if (cpu_online(cpu) || cpu == smp_processor_id())
538 resched_curr(rq);
539 raw_spin_unlock_irqrestore(&rq->lock, flags);
540}
541
542#ifdef CONFIG_SMP
543#ifdef CONFIG_NO_HZ_COMMON
544/*
545 * In the semi idle case, use the nearest busy CPU for migrating timers
546 * from an idle CPU. This is good for power-savings.
547 *
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle CPU will add more delays to the timers than intended
550 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
551 */
552int get_nohz_timer_target(void)
553{
554 int i, cpu = smp_processor_id();
555 struct sched_domain *sd;
556
557 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
558 return cpu;
559
560 rcu_read_lock();
561 for_each_domain(cpu, sd) {
562 for_each_cpu(i, sched_domain_span(sd)) {
563 if (cpu == i)
564 continue;
565
566 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
567 cpu = i;
568 goto unlock;
569 }
570 }
571 }
572
573 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
574 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
575unlock:
576 rcu_read_unlock();
577 return cpu;
578}
579
580/*
581 * When add_timer_on() enqueues a timer into the timer wheel of an
582 * idle CPU then this timer might expire before the next timer event
583 * which is scheduled to wake up that CPU. In case of a completely
584 * idle system the next event might even be infinite time into the
585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
586 * leaves the inner idle loop so the newly added timer is taken into
587 * account when the CPU goes back to idle and evaluates the timer
588 * wheel for the next timer event.
589 */
590static void wake_up_idle_cpu(int cpu)
591{
592 struct rq *rq = cpu_rq(cpu);
593
594 if (cpu == smp_processor_id())
595 return;
596
597 if (set_nr_and_not_polling(rq->idle))
598 smp_send_reschedule(cpu);
599 else
600 trace_sched_wake_idle_without_ipi(cpu);
601}
602
603static bool wake_up_full_nohz_cpu(int cpu)
604{
605 /*
606 * We just need the target to call irq_exit() and re-evaluate
607 * the next tick. The nohz full kick at least implies that.
608 * If needed we can still optimize that later with an
609 * empty IRQ.
610 */
611 if (cpu_is_offline(cpu))
612 return true; /* Don't try to wake offline CPUs. */
613 if (tick_nohz_full_cpu(cpu)) {
614 if (cpu != smp_processor_id() ||
615 tick_nohz_tick_stopped())
616 tick_nohz_full_kick_cpu(cpu);
617 return true;
618 }
619
620 return false;
621}
622
623/*
624 * Wake up the specified CPU. If the CPU is going offline, it is the
625 * caller's responsibility to deal with the lost wakeup, for example,
626 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
627 */
628void wake_up_nohz_cpu(int cpu)
629{
630 if (!wake_up_full_nohz_cpu(cpu))
631 wake_up_idle_cpu(cpu);
632}
633
634static inline bool got_nohz_idle_kick(void)
635{
636 int cpu = smp_processor_id();
637
638 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
639 return false;
640
641 if (idle_cpu(cpu) && !need_resched())
642 return true;
643
644 /*
645 * We can't run Idle Load Balance on this CPU for this time so we
646 * cancel it and clear NOHZ_BALANCE_KICK
647 */
648 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
649 return false;
650}
651
652#else /* CONFIG_NO_HZ_COMMON */
653
654static inline bool got_nohz_idle_kick(void)
655{
656 return false;
657}
658
659#endif /* CONFIG_NO_HZ_COMMON */
660
661#ifdef CONFIG_NO_HZ_FULL
662bool sched_can_stop_tick(struct rq *rq)
663{
664 int fifo_nr_running;
665
666 /* Deadline tasks, even if single, need the tick */
667 if (rq->dl.dl_nr_running)
668 return false;
669
670 /*
671 * If there are more than one RR tasks, we need the tick to effect the
672 * actual RR behaviour.
673 */
674 if (rq->rt.rr_nr_running) {
675 if (rq->rt.rr_nr_running == 1)
676 return true;
677 else
678 return false;
679 }
680
681 /*
682 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
683 * forced preemption between FIFO tasks.
684 */
685 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
686 if (fifo_nr_running)
687 return true;
688
689 /*
690 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
691 * if there's more than one we need the tick for involuntary
692 * preemption.
693 */
694 if (rq->nr_running > 1)
695 return false;
696
697 return true;
698}
699#endif /* CONFIG_NO_HZ_FULL */
700#endif /* CONFIG_SMP */
701
702#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
704/*
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
707 *
708 * Caller must hold rcu_lock or sufficient equivalent.
709 */
710int walk_tg_tree_from(struct task_group *from,
711 tg_visitor down, tg_visitor up, void *data)
712{
713 struct task_group *parent, *child;
714 int ret;
715
716 parent = from;
717
718down:
719 ret = (*down)(parent, data);
720 if (ret)
721 goto out;
722 list_for_each_entry_rcu(child, &parent->children, siblings) {
723 parent = child;
724 goto down;
725
726up:
727 continue;
728 }
729 ret = (*up)(parent, data);
730 if (ret || parent == from)
731 goto out;
732
733 child = parent;
734 parent = parent->parent;
735 if (parent)
736 goto up;
737out:
738 return ret;
739}
740
741int tg_nop(struct task_group *tg, void *data)
742{
743 return 0;
744}
745#endif
746
747static void set_load_weight(struct task_struct *p, bool update_load)
748{
749 int prio = p->static_prio - MAX_RT_PRIO;
750 struct load_weight *load = &p->se.load;
751
752 /*
753 * SCHED_IDLE tasks get minimal weight:
754 */
755 if (task_has_idle_policy(p)) {
756 load->weight = scale_load(WEIGHT_IDLEPRIO);
757 load->inv_weight = WMULT_IDLEPRIO;
758 p->se.runnable_weight = load->weight;
759 return;
760 }
761
762 /*
763 * SCHED_OTHER tasks have to update their load when changing their
764 * weight
765 */
766 if (update_load && p->sched_class == &fair_sched_class) {
767 reweight_task(p, prio);
768 } else {
769 load->weight = scale_load(sched_prio_to_weight[prio]);
770 load->inv_weight = sched_prio_to_wmult[prio];
771 p->se.runnable_weight = load->weight;
772 }
773}
774
775#ifdef CONFIG_UCLAMP_TASK
776/*
777 * Serializes updates of utilization clamp values
778 *
779 * The (slow-path) user-space triggers utilization clamp value updates which
780 * can require updates on (fast-path) scheduler's data structures used to
781 * support enqueue/dequeue operations.
782 * While the per-CPU rq lock protects fast-path update operations, user-space
783 * requests are serialized using a mutex to reduce the risk of conflicting
784 * updates or API abuses.
785 */
786static DEFINE_MUTEX(uclamp_mutex);
787
788/* Max allowed minimum utilization */
789unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
790
791/* Max allowed maximum utilization */
792unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
793
794/* All clamps are required to be less or equal than these values */
795static struct uclamp_se uclamp_default[UCLAMP_CNT];
796
797/* Integer rounded range for each bucket */
798#define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
799
800#define for_each_clamp_id(clamp_id) \
801 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
802
803static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
804{
805 return clamp_value / UCLAMP_BUCKET_DELTA;
806}
807
808static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value)
809{
810 return UCLAMP_BUCKET_DELTA * uclamp_bucket_id(clamp_value);
811}
812
813static inline enum uclamp_id uclamp_none(enum uclamp_id clamp_id)
814{
815 if (clamp_id == UCLAMP_MIN)
816 return 0;
817 return SCHED_CAPACITY_SCALE;
818}
819
820static inline void uclamp_se_set(struct uclamp_se *uc_se,
821 unsigned int value, bool user_defined)
822{
823 uc_se->value = value;
824 uc_se->bucket_id = uclamp_bucket_id(value);
825 uc_se->user_defined = user_defined;
826}
827
828static inline unsigned int
829uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
830 unsigned int clamp_value)
831{
832 /*
833 * Avoid blocked utilization pushing up the frequency when we go
834 * idle (which drops the max-clamp) by retaining the last known
835 * max-clamp.
836 */
837 if (clamp_id == UCLAMP_MAX) {
838 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
839 return clamp_value;
840 }
841
842 return uclamp_none(UCLAMP_MIN);
843}
844
845static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
846 unsigned int clamp_value)
847{
848 /* Reset max-clamp retention only on idle exit */
849 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
850 return;
851
852 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
853}
854
855static inline
856enum uclamp_id uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
857 unsigned int clamp_value)
858{
859 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
860 int bucket_id = UCLAMP_BUCKETS - 1;
861
862 /*
863 * Since both min and max clamps are max aggregated, find the
864 * top most bucket with tasks in.
865 */
866 for ( ; bucket_id >= 0; bucket_id--) {
867 if (!bucket[bucket_id].tasks)
868 continue;
869 return bucket[bucket_id].value;
870 }
871
872 /* No tasks -- default clamp values */
873 return uclamp_idle_value(rq, clamp_id, clamp_value);
874}
875
876static inline struct uclamp_se
877uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
878{
879 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
880#ifdef CONFIG_UCLAMP_TASK_GROUP
881 struct uclamp_se uc_max;
882
883 /*
884 * Tasks in autogroups or root task group will be
885 * restricted by system defaults.
886 */
887 if (task_group_is_autogroup(task_group(p)))
888 return uc_req;
889 if (task_group(p) == &root_task_group)
890 return uc_req;
891
892 uc_max = task_group(p)->uclamp[clamp_id];
893 if (uc_req.value > uc_max.value || !uc_req.user_defined)
894 return uc_max;
895#endif
896
897 return uc_req;
898}
899
900/*
901 * The effective clamp bucket index of a task depends on, by increasing
902 * priority:
903 * - the task specific clamp value, when explicitly requested from userspace
904 * - the task group effective clamp value, for tasks not either in the root
905 * group or in an autogroup
906 * - the system default clamp value, defined by the sysadmin
907 */
908static inline struct uclamp_se
909uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
910{
911 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
912 struct uclamp_se uc_max = uclamp_default[clamp_id];
913
914 /* System default restrictions always apply */
915 if (unlikely(uc_req.value > uc_max.value))
916 return uc_max;
917
918 return uc_req;
919}
920
921enum uclamp_id uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
922{
923 struct uclamp_se uc_eff;
924
925 /* Task currently refcounted: use back-annotated (effective) value */
926 if (p->uclamp[clamp_id].active)
927 return p->uclamp[clamp_id].value;
928
929 uc_eff = uclamp_eff_get(p, clamp_id);
930
931 return uc_eff.value;
932}
933
934/*
935 * When a task is enqueued on a rq, the clamp bucket currently defined by the
936 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
937 * updates the rq's clamp value if required.
938 *
939 * Tasks can have a task-specific value requested from user-space, track
940 * within each bucket the maximum value for tasks refcounted in it.
941 * This "local max aggregation" allows to track the exact "requested" value
942 * for each bucket when all its RUNNABLE tasks require the same clamp.
943 */
944static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
945 enum uclamp_id clamp_id)
946{
947 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
948 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
949 struct uclamp_bucket *bucket;
950
951 lockdep_assert_held(&rq->lock);
952
953 /* Update task effective clamp */
954 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
955
956 bucket = &uc_rq->bucket[uc_se->bucket_id];
957 bucket->tasks++;
958 uc_se->active = true;
959
960 uclamp_idle_reset(rq, clamp_id, uc_se->value);
961
962 /*
963 * Local max aggregation: rq buckets always track the max
964 * "requested" clamp value of its RUNNABLE tasks.
965 */
966 if (bucket->tasks == 1 || uc_se->value > bucket->value)
967 bucket->value = uc_se->value;
968
969 if (uc_se->value > READ_ONCE(uc_rq->value))
970 WRITE_ONCE(uc_rq->value, uc_se->value);
971}
972
973/*
974 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
975 * is released. If this is the last task reference counting the rq's max
976 * active clamp value, then the rq's clamp value is updated.
977 *
978 * Both refcounted tasks and rq's cached clamp values are expected to be
979 * always valid. If it's detected they are not, as defensive programming,
980 * enforce the expected state and warn.
981 */
982static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
983 enum uclamp_id clamp_id)
984{
985 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
986 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
987 struct uclamp_bucket *bucket;
988 unsigned int bkt_clamp;
989 unsigned int rq_clamp;
990
991 lockdep_assert_held(&rq->lock);
992
993 bucket = &uc_rq->bucket[uc_se->bucket_id];
994 SCHED_WARN_ON(!bucket->tasks);
995 if (likely(bucket->tasks))
996 bucket->tasks--;
997 uc_se->active = false;
998
999 /*
1000 * Keep "local max aggregation" simple and accept to (possibly)
1001 * overboost some RUNNABLE tasks in the same bucket.
1002 * The rq clamp bucket value is reset to its base value whenever
1003 * there are no more RUNNABLE tasks refcounting it.
1004 */
1005 if (likely(bucket->tasks))
1006 return;
1007
1008 rq_clamp = READ_ONCE(uc_rq->value);
1009 /*
1010 * Defensive programming: this should never happen. If it happens,
1011 * e.g. due to future modification, warn and fixup the expected value.
1012 */
1013 SCHED_WARN_ON(bucket->value > rq_clamp);
1014 if (bucket->value >= rq_clamp) {
1015 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1016 WRITE_ONCE(uc_rq->value, bkt_clamp);
1017 }
1018}
1019
1020static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1021{
1022 enum uclamp_id clamp_id;
1023
1024 if (unlikely(!p->sched_class->uclamp_enabled))
1025 return;
1026
1027 for_each_clamp_id(clamp_id)
1028 uclamp_rq_inc_id(rq, p, clamp_id);
1029
1030 /* Reset clamp idle holding when there is one RUNNABLE task */
1031 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1032 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1033}
1034
1035static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1036{
1037 enum uclamp_id clamp_id;
1038
1039 if (unlikely(!p->sched_class->uclamp_enabled))
1040 return;
1041
1042 for_each_clamp_id(clamp_id)
1043 uclamp_rq_dec_id(rq, p, clamp_id);
1044}
1045
1046static inline void
1047uclamp_update_active(struct task_struct *p, enum uclamp_id clamp_id)
1048{
1049 struct rq_flags rf;
1050 struct rq *rq;
1051
1052 /*
1053 * Lock the task and the rq where the task is (or was) queued.
1054 *
1055 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1056 * price to pay to safely serialize util_{min,max} updates with
1057 * enqueues, dequeues and migration operations.
1058 * This is the same locking schema used by __set_cpus_allowed_ptr().
1059 */
1060 rq = task_rq_lock(p, &rf);
1061
1062 /*
1063 * Setting the clamp bucket is serialized by task_rq_lock().
1064 * If the task is not yet RUNNABLE and its task_struct is not
1065 * affecting a valid clamp bucket, the next time it's enqueued,
1066 * it will already see the updated clamp bucket value.
1067 */
1068 if (p->uclamp[clamp_id].active) {
1069 uclamp_rq_dec_id(rq, p, clamp_id);
1070 uclamp_rq_inc_id(rq, p, clamp_id);
1071 }
1072
1073 task_rq_unlock(rq, p, &rf);
1074}
1075
1076#ifdef CONFIG_UCLAMP_TASK_GROUP
1077static inline void
1078uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1079 unsigned int clamps)
1080{
1081 enum uclamp_id clamp_id;
1082 struct css_task_iter it;
1083 struct task_struct *p;
1084
1085 css_task_iter_start(css, 0, &it);
1086 while ((p = css_task_iter_next(&it))) {
1087 for_each_clamp_id(clamp_id) {
1088 if ((0x1 << clamp_id) & clamps)
1089 uclamp_update_active(p, clamp_id);
1090 }
1091 }
1092 css_task_iter_end(&it);
1093}
1094
1095static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1096static void uclamp_update_root_tg(void)
1097{
1098 struct task_group *tg = &root_task_group;
1099
1100 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1101 sysctl_sched_uclamp_util_min, false);
1102 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1103 sysctl_sched_uclamp_util_max, false);
1104
1105 rcu_read_lock();
1106 cpu_util_update_eff(&root_task_group.css);
1107 rcu_read_unlock();
1108}
1109#else
1110static void uclamp_update_root_tg(void) { }
1111#endif
1112
1113int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1114 void __user *buffer, size_t *lenp,
1115 loff_t *ppos)
1116{
1117 bool update_root_tg = false;
1118 int old_min, old_max;
1119 int result;
1120
1121 mutex_lock(&uclamp_mutex);
1122 old_min = sysctl_sched_uclamp_util_min;
1123 old_max = sysctl_sched_uclamp_util_max;
1124
1125 result = proc_dointvec(table, write, buffer, lenp, ppos);
1126 if (result)
1127 goto undo;
1128 if (!write)
1129 goto done;
1130
1131 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1132 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE) {
1133 result = -EINVAL;
1134 goto undo;
1135 }
1136
1137 if (old_min != sysctl_sched_uclamp_util_min) {
1138 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1139 sysctl_sched_uclamp_util_min, false);
1140 update_root_tg = true;
1141 }
1142 if (old_max != sysctl_sched_uclamp_util_max) {
1143 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1144 sysctl_sched_uclamp_util_max, false);
1145 update_root_tg = true;
1146 }
1147
1148 if (update_root_tg)
1149 uclamp_update_root_tg();
1150
1151 /*
1152 * We update all RUNNABLE tasks only when task groups are in use.
1153 * Otherwise, keep it simple and do just a lazy update at each next
1154 * task enqueue time.
1155 */
1156
1157 goto done;
1158
1159undo:
1160 sysctl_sched_uclamp_util_min = old_min;
1161 sysctl_sched_uclamp_util_max = old_max;
1162done:
1163 mutex_unlock(&uclamp_mutex);
1164
1165 return result;
1166}
1167
1168static int uclamp_validate(struct task_struct *p,
1169 const struct sched_attr *attr)
1170{
1171 unsigned int lower_bound = p->uclamp_req[UCLAMP_MIN].value;
1172 unsigned int upper_bound = p->uclamp_req[UCLAMP_MAX].value;
1173
1174 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN)
1175 lower_bound = attr->sched_util_min;
1176 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX)
1177 upper_bound = attr->sched_util_max;
1178
1179 if (lower_bound > upper_bound)
1180 return -EINVAL;
1181 if (upper_bound > SCHED_CAPACITY_SCALE)
1182 return -EINVAL;
1183
1184 return 0;
1185}
1186
1187static void __setscheduler_uclamp(struct task_struct *p,
1188 const struct sched_attr *attr)
1189{
1190 enum uclamp_id clamp_id;
1191
1192 /*
1193 * On scheduling class change, reset to default clamps for tasks
1194 * without a task-specific value.
1195 */
1196 for_each_clamp_id(clamp_id) {
1197 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1198 unsigned int clamp_value = uclamp_none(clamp_id);
1199
1200 /* Keep using defined clamps across class changes */
1201 if (uc_se->user_defined)
1202 continue;
1203
1204 /* By default, RT tasks always get 100% boost */
1205 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1206 clamp_value = uclamp_none(UCLAMP_MAX);
1207
1208 uclamp_se_set(uc_se, clamp_value, false);
1209 }
1210
1211 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1212 return;
1213
1214 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1215 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1216 attr->sched_util_min, true);
1217 }
1218
1219 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1220 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1221 attr->sched_util_max, true);
1222 }
1223}
1224
1225static void uclamp_fork(struct task_struct *p)
1226{
1227 enum uclamp_id clamp_id;
1228
1229 for_each_clamp_id(clamp_id)
1230 p->uclamp[clamp_id].active = false;
1231
1232 if (likely(!p->sched_reset_on_fork))
1233 return;
1234
1235 for_each_clamp_id(clamp_id) {
1236 unsigned int clamp_value = uclamp_none(clamp_id);
1237
1238 /* By default, RT tasks always get 100% boost */
1239 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1240 clamp_value = uclamp_none(UCLAMP_MAX);
1241
1242 uclamp_se_set(&p->uclamp_req[clamp_id], clamp_value, false);
1243 }
1244}
1245
1246static void __init init_uclamp(void)
1247{
1248 struct uclamp_se uc_max = {};
1249 enum uclamp_id clamp_id;
1250 int cpu;
1251
1252 mutex_init(&uclamp_mutex);
1253
1254 for_each_possible_cpu(cpu) {
1255 memset(&cpu_rq(cpu)->uclamp, 0, sizeof(struct uclamp_rq));
1256 cpu_rq(cpu)->uclamp_flags = 0;
1257 }
1258
1259 for_each_clamp_id(clamp_id) {
1260 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1261 uclamp_none(clamp_id), false);
1262 }
1263
1264 /* System defaults allow max clamp values for both indexes */
1265 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1266 for_each_clamp_id(clamp_id) {
1267 uclamp_default[clamp_id] = uc_max;
1268#ifdef CONFIG_UCLAMP_TASK_GROUP
1269 root_task_group.uclamp_req[clamp_id] = uc_max;
1270 root_task_group.uclamp[clamp_id] = uc_max;
1271#endif
1272 }
1273}
1274
1275#else /* CONFIG_UCLAMP_TASK */
1276static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1277static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1278static inline int uclamp_validate(struct task_struct *p,
1279 const struct sched_attr *attr)
1280{
1281 return -EOPNOTSUPP;
1282}
1283static void __setscheduler_uclamp(struct task_struct *p,
1284 const struct sched_attr *attr) { }
1285static inline void uclamp_fork(struct task_struct *p) { }
1286static inline void init_uclamp(void) { }
1287#endif /* CONFIG_UCLAMP_TASK */
1288
1289static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1290{
1291 if (!(flags & ENQUEUE_NOCLOCK))
1292 update_rq_clock(rq);
1293
1294 if (!(flags & ENQUEUE_RESTORE)) {
1295 sched_info_queued(rq, p);
1296 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1297 }
1298
1299 uclamp_rq_inc(rq, p);
1300 p->sched_class->enqueue_task(rq, p, flags);
1301}
1302
1303static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1304{
1305 if (!(flags & DEQUEUE_NOCLOCK))
1306 update_rq_clock(rq);
1307
1308 if (!(flags & DEQUEUE_SAVE)) {
1309 sched_info_dequeued(rq, p);
1310 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1311 }
1312
1313 uclamp_rq_dec(rq, p);
1314 p->sched_class->dequeue_task(rq, p, flags);
1315}
1316
1317void activate_task(struct rq *rq, struct task_struct *p, int flags)
1318{
1319 if (task_contributes_to_load(p))
1320 rq->nr_uninterruptible--;
1321
1322 enqueue_task(rq, p, flags);
1323
1324 p->on_rq = TASK_ON_RQ_QUEUED;
1325}
1326
1327void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1328{
1329 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1330
1331 if (task_contributes_to_load(p))
1332 rq->nr_uninterruptible++;
1333
1334 dequeue_task(rq, p, flags);
1335}
1336
1337/*
1338 * __normal_prio - return the priority that is based on the static prio
1339 */
1340static inline int __normal_prio(struct task_struct *p)
1341{
1342 return p->static_prio;
1343}
1344
1345/*
1346 * Calculate the expected normal priority: i.e. priority
1347 * without taking RT-inheritance into account. Might be
1348 * boosted by interactivity modifiers. Changes upon fork,
1349 * setprio syscalls, and whenever the interactivity
1350 * estimator recalculates.
1351 */
1352static inline int normal_prio(struct task_struct *p)
1353{
1354 int prio;
1355
1356 if (task_has_dl_policy(p))
1357 prio = MAX_DL_PRIO-1;
1358 else if (task_has_rt_policy(p))
1359 prio = MAX_RT_PRIO-1 - p->rt_priority;
1360 else
1361 prio = __normal_prio(p);
1362 return prio;
1363}
1364
1365/*
1366 * Calculate the current priority, i.e. the priority
1367 * taken into account by the scheduler. This value might
1368 * be boosted by RT tasks, or might be boosted by
1369 * interactivity modifiers. Will be RT if the task got
1370 * RT-boosted. If not then it returns p->normal_prio.
1371 */
1372static int effective_prio(struct task_struct *p)
1373{
1374 p->normal_prio = normal_prio(p);
1375 /*
1376 * If we are RT tasks or we were boosted to RT priority,
1377 * keep the priority unchanged. Otherwise, update priority
1378 * to the normal priority:
1379 */
1380 if (!rt_prio(p->prio))
1381 return p->normal_prio;
1382 return p->prio;
1383}
1384
1385/**
1386 * task_curr - is this task currently executing on a CPU?
1387 * @p: the task in question.
1388 *
1389 * Return: 1 if the task is currently executing. 0 otherwise.
1390 */
1391inline int task_curr(const struct task_struct *p)
1392{
1393 return cpu_curr(task_cpu(p)) == p;
1394}
1395
1396/*
1397 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1398 * use the balance_callback list if you want balancing.
1399 *
1400 * this means any call to check_class_changed() must be followed by a call to
1401 * balance_callback().
1402 */
1403static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1404 const struct sched_class *prev_class,
1405 int oldprio)
1406{
1407 if (prev_class != p->sched_class) {
1408 if (prev_class->switched_from)
1409 prev_class->switched_from(rq, p);
1410
1411 p->sched_class->switched_to(rq, p);
1412 } else if (oldprio != p->prio || dl_task(p))
1413 p->sched_class->prio_changed(rq, p, oldprio);
1414}
1415
1416void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1417{
1418 const struct sched_class *class;
1419
1420 if (p->sched_class == rq->curr->sched_class) {
1421 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1422 } else {
1423 for_each_class(class) {
1424 if (class == rq->curr->sched_class)
1425 break;
1426 if (class == p->sched_class) {
1427 resched_curr(rq);
1428 break;
1429 }
1430 }
1431 }
1432
1433 /*
1434 * A queue event has occurred, and we're going to schedule. In
1435 * this case, we can save a useless back to back clock update.
1436 */
1437 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1438 rq_clock_skip_update(rq);
1439}
1440
1441#ifdef CONFIG_SMP
1442
1443static inline bool is_per_cpu_kthread(struct task_struct *p)
1444{
1445 if (!(p->flags & PF_KTHREAD))
1446 return false;
1447
1448 if (p->nr_cpus_allowed != 1)
1449 return false;
1450
1451 return true;
1452}
1453
1454/*
1455 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1456 * __set_cpus_allowed_ptr() and select_fallback_rq().
1457 */
1458static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1459{
1460 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1461 return false;
1462
1463 if (is_per_cpu_kthread(p))
1464 return cpu_online(cpu);
1465
1466 return cpu_active(cpu);
1467}
1468
1469/*
1470 * This is how migration works:
1471 *
1472 * 1) we invoke migration_cpu_stop() on the target CPU using
1473 * stop_one_cpu().
1474 * 2) stopper starts to run (implicitly forcing the migrated thread
1475 * off the CPU)
1476 * 3) it checks whether the migrated task is still in the wrong runqueue.
1477 * 4) if it's in the wrong runqueue then the migration thread removes
1478 * it and puts it into the right queue.
1479 * 5) stopper completes and stop_one_cpu() returns and the migration
1480 * is done.
1481 */
1482
1483/*
1484 * move_queued_task - move a queued task to new rq.
1485 *
1486 * Returns (locked) new rq. Old rq's lock is released.
1487 */
1488static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1489 struct task_struct *p, int new_cpu)
1490{
1491 lockdep_assert_held(&rq->lock);
1492
1493 WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
1494 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
1495 set_task_cpu(p, new_cpu);
1496 rq_unlock(rq, rf);
1497
1498 rq = cpu_rq(new_cpu);
1499
1500 rq_lock(rq, rf);
1501 BUG_ON(task_cpu(p) != new_cpu);
1502 enqueue_task(rq, p, 0);
1503 p->on_rq = TASK_ON_RQ_QUEUED;
1504 check_preempt_curr(rq, p, 0);
1505
1506 return rq;
1507}
1508
1509struct migration_arg {
1510 struct task_struct *task;
1511 int dest_cpu;
1512};
1513
1514/*
1515 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1516 * this because either it can't run here any more (set_cpus_allowed()
1517 * away from this CPU, or CPU going down), or because we're
1518 * attempting to rebalance this task on exec (sched_exec).
1519 *
1520 * So we race with normal scheduler movements, but that's OK, as long
1521 * as the task is no longer on this CPU.
1522 */
1523static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1524 struct task_struct *p, int dest_cpu)
1525{
1526 /* Affinity changed (again). */
1527 if (!is_cpu_allowed(p, dest_cpu))
1528 return rq;
1529
1530 update_rq_clock(rq);
1531 rq = move_queued_task(rq, rf, p, dest_cpu);
1532
1533 return rq;
1534}
1535
1536/*
1537 * migration_cpu_stop - this will be executed by a highprio stopper thread
1538 * and performs thread migration by bumping thread off CPU then
1539 * 'pushing' onto another runqueue.
1540 */
1541static int migration_cpu_stop(void *data)
1542{
1543 struct migration_arg *arg = data;
1544 struct task_struct *p = arg->task;
1545 struct rq *rq = this_rq();
1546 struct rq_flags rf;
1547
1548 /*
1549 * The original target CPU might have gone down and we might
1550 * be on another CPU but it doesn't matter.
1551 */
1552 local_irq_disable();
1553 /*
1554 * We need to explicitly wake pending tasks before running
1555 * __migrate_task() such that we will not miss enforcing cpus_ptr
1556 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1557 */
1558 sched_ttwu_pending();
1559
1560 raw_spin_lock(&p->pi_lock);
1561 rq_lock(rq, &rf);
1562 /*
1563 * If task_rq(p) != rq, it cannot be migrated here, because we're
1564 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1565 * we're holding p->pi_lock.
1566 */
1567 if (task_rq(p) == rq) {
1568 if (task_on_rq_queued(p))
1569 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1570 else
1571 p->wake_cpu = arg->dest_cpu;
1572 }
1573 rq_unlock(rq, &rf);
1574 raw_spin_unlock(&p->pi_lock);
1575
1576 local_irq_enable();
1577 return 0;
1578}
1579
1580/*
1581 * sched_class::set_cpus_allowed must do the below, but is not required to
1582 * actually call this function.
1583 */
1584void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1585{
1586 cpumask_copy(&p->cpus_mask, new_mask);
1587 p->nr_cpus_allowed = cpumask_weight(new_mask);
1588}
1589
1590void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1591{
1592 struct rq *rq = task_rq(p);
1593 bool queued, running;
1594
1595 lockdep_assert_held(&p->pi_lock);
1596
1597 queued = task_on_rq_queued(p);
1598 running = task_current(rq, p);
1599
1600 if (queued) {
1601 /*
1602 * Because __kthread_bind() calls this on blocked tasks without
1603 * holding rq->lock.
1604 */
1605 lockdep_assert_held(&rq->lock);
1606 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1607 }
1608 if (running)
1609 put_prev_task(rq, p);
1610
1611 p->sched_class->set_cpus_allowed(p, new_mask);
1612
1613 if (queued)
1614 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1615 if (running)
1616 set_next_task(rq, p);
1617}
1618
1619/*
1620 * Change a given task's CPU affinity. Migrate the thread to a
1621 * proper CPU and schedule it away if the CPU it's executing on
1622 * is removed from the allowed bitmask.
1623 *
1624 * NOTE: the caller must have a valid reference to the task, the
1625 * task must not exit() & deallocate itself prematurely. The
1626 * call is not atomic; no spinlocks may be held.
1627 */
1628static int __set_cpus_allowed_ptr(struct task_struct *p,
1629 const struct cpumask *new_mask, bool check)
1630{
1631 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1632 unsigned int dest_cpu;
1633 struct rq_flags rf;
1634 struct rq *rq;
1635 int ret = 0;
1636
1637 rq = task_rq_lock(p, &rf);
1638 update_rq_clock(rq);
1639
1640 if (p->flags & PF_KTHREAD) {
1641 /*
1642 * Kernel threads are allowed on online && !active CPUs
1643 */
1644 cpu_valid_mask = cpu_online_mask;
1645 }
1646
1647 /*
1648 * Must re-check here, to close a race against __kthread_bind(),
1649 * sched_setaffinity() is not guaranteed to observe the flag.
1650 */
1651 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1652 ret = -EINVAL;
1653 goto out;
1654 }
1655
1656 if (cpumask_equal(p->cpus_ptr, new_mask))
1657 goto out;
1658
1659 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1660 if (dest_cpu >= nr_cpu_ids) {
1661 ret = -EINVAL;
1662 goto out;
1663 }
1664
1665 do_set_cpus_allowed(p, new_mask);
1666
1667 if (p->flags & PF_KTHREAD) {
1668 /*
1669 * For kernel threads that do indeed end up on online &&
1670 * !active we want to ensure they are strict per-CPU threads.
1671 */
1672 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1673 !cpumask_intersects(new_mask, cpu_active_mask) &&
1674 p->nr_cpus_allowed != 1);
1675 }
1676
1677 /* Can the task run on the task's current CPU? If so, we're done */
1678 if (cpumask_test_cpu(task_cpu(p), new_mask))
1679 goto out;
1680
1681 if (task_running(rq, p) || p->state == TASK_WAKING) {
1682 struct migration_arg arg = { p, dest_cpu };
1683 /* Need help from migration thread: drop lock and wait. */
1684 task_rq_unlock(rq, p, &rf);
1685 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1686 return 0;
1687 } else if (task_on_rq_queued(p)) {
1688 /*
1689 * OK, since we're going to drop the lock immediately
1690 * afterwards anyway.
1691 */
1692 rq = move_queued_task(rq, &rf, p, dest_cpu);
1693 }
1694out:
1695 task_rq_unlock(rq, p, &rf);
1696
1697 return ret;
1698}
1699
1700int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1701{
1702 return __set_cpus_allowed_ptr(p, new_mask, false);
1703}
1704EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1705
1706void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1707{
1708#ifdef CONFIG_SCHED_DEBUG
1709 /*
1710 * We should never call set_task_cpu() on a blocked task,
1711 * ttwu() will sort out the placement.
1712 */
1713 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1714 !p->on_rq);
1715
1716 /*
1717 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1718 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1719 * time relying on p->on_rq.
1720 */
1721 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1722 p->sched_class == &fair_sched_class &&
1723 (p->on_rq && !task_on_rq_migrating(p)));
1724
1725#ifdef CONFIG_LOCKDEP
1726 /*
1727 * The caller should hold either p->pi_lock or rq->lock, when changing
1728 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1729 *
1730 * sched_move_task() holds both and thus holding either pins the cgroup,
1731 * see task_group().
1732 *
1733 * Furthermore, all task_rq users should acquire both locks, see
1734 * task_rq_lock().
1735 */
1736 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1737 lockdep_is_held(&task_rq(p)->lock)));
1738#endif
1739 /*
1740 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1741 */
1742 WARN_ON_ONCE(!cpu_online(new_cpu));
1743#endif
1744
1745 trace_sched_migrate_task(p, new_cpu);
1746
1747 if (task_cpu(p) != new_cpu) {
1748 if (p->sched_class->migrate_task_rq)
1749 p->sched_class->migrate_task_rq(p, new_cpu);
1750 p->se.nr_migrations++;
1751 rseq_migrate(p);
1752 perf_event_task_migrate(p);
1753 }
1754
1755 __set_task_cpu(p, new_cpu);
1756}
1757
1758#ifdef CONFIG_NUMA_BALANCING
1759static void __migrate_swap_task(struct task_struct *p, int cpu)
1760{
1761 if (task_on_rq_queued(p)) {
1762 struct rq *src_rq, *dst_rq;
1763 struct rq_flags srf, drf;
1764
1765 src_rq = task_rq(p);
1766 dst_rq = cpu_rq(cpu);
1767
1768 rq_pin_lock(src_rq, &srf);
1769 rq_pin_lock(dst_rq, &drf);
1770
1771 deactivate_task(src_rq, p, 0);
1772 set_task_cpu(p, cpu);
1773 activate_task(dst_rq, p, 0);
1774 check_preempt_curr(dst_rq, p, 0);
1775
1776 rq_unpin_lock(dst_rq, &drf);
1777 rq_unpin_lock(src_rq, &srf);
1778
1779 } else {
1780 /*
1781 * Task isn't running anymore; make it appear like we migrated
1782 * it before it went to sleep. This means on wakeup we make the
1783 * previous CPU our target instead of where it really is.
1784 */
1785 p->wake_cpu = cpu;
1786 }
1787}
1788
1789struct migration_swap_arg {
1790 struct task_struct *src_task, *dst_task;
1791 int src_cpu, dst_cpu;
1792};
1793
1794static int migrate_swap_stop(void *data)
1795{
1796 struct migration_swap_arg *arg = data;
1797 struct rq *src_rq, *dst_rq;
1798 int ret = -EAGAIN;
1799
1800 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1801 return -EAGAIN;
1802
1803 src_rq = cpu_rq(arg->src_cpu);
1804 dst_rq = cpu_rq(arg->dst_cpu);
1805
1806 double_raw_lock(&arg->src_task->pi_lock,
1807 &arg->dst_task->pi_lock);
1808 double_rq_lock(src_rq, dst_rq);
1809
1810 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1811 goto unlock;
1812
1813 if (task_cpu(arg->src_task) != arg->src_cpu)
1814 goto unlock;
1815
1816 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
1817 goto unlock;
1818
1819 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
1820 goto unlock;
1821
1822 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1823 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1824
1825 ret = 0;
1826
1827unlock:
1828 double_rq_unlock(src_rq, dst_rq);
1829 raw_spin_unlock(&arg->dst_task->pi_lock);
1830 raw_spin_unlock(&arg->src_task->pi_lock);
1831
1832 return ret;
1833}
1834
1835/*
1836 * Cross migrate two tasks
1837 */
1838int migrate_swap(struct task_struct *cur, struct task_struct *p,
1839 int target_cpu, int curr_cpu)
1840{
1841 struct migration_swap_arg arg;
1842 int ret = -EINVAL;
1843
1844 arg = (struct migration_swap_arg){
1845 .src_task = cur,
1846 .src_cpu = curr_cpu,
1847 .dst_task = p,
1848 .dst_cpu = target_cpu,
1849 };
1850
1851 if (arg.src_cpu == arg.dst_cpu)
1852 goto out;
1853
1854 /*
1855 * These three tests are all lockless; this is OK since all of them
1856 * will be re-checked with proper locks held further down the line.
1857 */
1858 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1859 goto out;
1860
1861 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
1862 goto out;
1863
1864 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
1865 goto out;
1866
1867 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1868 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1869
1870out:
1871 return ret;
1872}
1873#endif /* CONFIG_NUMA_BALANCING */
1874
1875/*
1876 * wait_task_inactive - wait for a thread to unschedule.
1877 *
1878 * If @match_state is nonzero, it's the @p->state value just checked and
1879 * not expected to change. If it changes, i.e. @p might have woken up,
1880 * then return zero. When we succeed in waiting for @p to be off its CPU,
1881 * we return a positive number (its total switch count). If a second call
1882 * a short while later returns the same number, the caller can be sure that
1883 * @p has remained unscheduled the whole time.
1884 *
1885 * The caller must ensure that the task *will* unschedule sometime soon,
1886 * else this function might spin for a *long* time. This function can't
1887 * be called with interrupts off, or it may introduce deadlock with
1888 * smp_call_function() if an IPI is sent by the same process we are
1889 * waiting to become inactive.
1890 */
1891unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1892{
1893 int running, queued;
1894 struct rq_flags rf;
1895 unsigned long ncsw;
1896 struct rq *rq;
1897
1898 for (;;) {
1899 /*
1900 * We do the initial early heuristics without holding
1901 * any task-queue locks at all. We'll only try to get
1902 * the runqueue lock when things look like they will
1903 * work out!
1904 */
1905 rq = task_rq(p);
1906
1907 /*
1908 * If the task is actively running on another CPU
1909 * still, just relax and busy-wait without holding
1910 * any locks.
1911 *
1912 * NOTE! Since we don't hold any locks, it's not
1913 * even sure that "rq" stays as the right runqueue!
1914 * But we don't care, since "task_running()" will
1915 * return false if the runqueue has changed and p
1916 * is actually now running somewhere else!
1917 */
1918 while (task_running(rq, p)) {
1919 if (match_state && unlikely(p->state != match_state))
1920 return 0;
1921 cpu_relax();
1922 }
1923
1924 /*
1925 * Ok, time to look more closely! We need the rq
1926 * lock now, to be *sure*. If we're wrong, we'll
1927 * just go back and repeat.
1928 */
1929 rq = task_rq_lock(p, &rf);
1930 trace_sched_wait_task(p);
1931 running = task_running(rq, p);
1932 queued = task_on_rq_queued(p);
1933 ncsw = 0;
1934 if (!match_state || p->state == match_state)
1935 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1936 task_rq_unlock(rq, p, &rf);
1937
1938 /*
1939 * If it changed from the expected state, bail out now.
1940 */
1941 if (unlikely(!ncsw))
1942 break;
1943
1944 /*
1945 * Was it really running after all now that we
1946 * checked with the proper locks actually held?
1947 *
1948 * Oops. Go back and try again..
1949 */
1950 if (unlikely(running)) {
1951 cpu_relax();
1952 continue;
1953 }
1954
1955 /*
1956 * It's not enough that it's not actively running,
1957 * it must be off the runqueue _entirely_, and not
1958 * preempted!
1959 *
1960 * So if it was still runnable (but just not actively
1961 * running right now), it's preempted, and we should
1962 * yield - it could be a while.
1963 */
1964 if (unlikely(queued)) {
1965 ktime_t to = NSEC_PER_SEC / HZ;
1966
1967 set_current_state(TASK_UNINTERRUPTIBLE);
1968 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1969 continue;
1970 }
1971
1972 /*
1973 * Ahh, all good. It wasn't running, and it wasn't
1974 * runnable, which means that it will never become
1975 * running in the future either. We're all done!
1976 */
1977 break;
1978 }
1979
1980 return ncsw;
1981}
1982
1983/***
1984 * kick_process - kick a running thread to enter/exit the kernel
1985 * @p: the to-be-kicked thread
1986 *
1987 * Cause a process which is running on another CPU to enter
1988 * kernel-mode, without any delay. (to get signals handled.)
1989 *
1990 * NOTE: this function doesn't have to take the runqueue lock,
1991 * because all it wants to ensure is that the remote task enters
1992 * the kernel. If the IPI races and the task has been migrated
1993 * to another CPU then no harm is done and the purpose has been
1994 * achieved as well.
1995 */
1996void kick_process(struct task_struct *p)
1997{
1998 int cpu;
1999
2000 preempt_disable();
2001 cpu = task_cpu(p);
2002 if ((cpu != smp_processor_id()) && task_curr(p))
2003 smp_send_reschedule(cpu);
2004 preempt_enable();
2005}
2006EXPORT_SYMBOL_GPL(kick_process);
2007
2008/*
2009 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2010 *
2011 * A few notes on cpu_active vs cpu_online:
2012 *
2013 * - cpu_active must be a subset of cpu_online
2014 *
2015 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2016 * see __set_cpus_allowed_ptr(). At this point the newly online
2017 * CPU isn't yet part of the sched domains, and balancing will not
2018 * see it.
2019 *
2020 * - on CPU-down we clear cpu_active() to mask the sched domains and
2021 * avoid the load balancer to place new tasks on the to be removed
2022 * CPU. Existing tasks will remain running there and will be taken
2023 * off.
2024 *
2025 * This means that fallback selection must not select !active CPUs.
2026 * And can assume that any active CPU must be online. Conversely
2027 * select_task_rq() below may allow selection of !active CPUs in order
2028 * to satisfy the above rules.
2029 */
2030static int select_fallback_rq(int cpu, struct task_struct *p)
2031{
2032 int nid = cpu_to_node(cpu);
2033 const struct cpumask *nodemask = NULL;
2034 enum { cpuset, possible, fail } state = cpuset;
2035 int dest_cpu;
2036
2037 /*
2038 * If the node that the CPU is on has been offlined, cpu_to_node()
2039 * will return -1. There is no CPU on the node, and we should
2040 * select the CPU on the other node.
2041 */
2042 if (nid != -1) {
2043 nodemask = cpumask_of_node(nid);
2044
2045 /* Look for allowed, online CPU in same node. */
2046 for_each_cpu(dest_cpu, nodemask) {
2047 if (!cpu_active(dest_cpu))
2048 continue;
2049 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2050 return dest_cpu;
2051 }
2052 }
2053
2054 for (;;) {
2055 /* Any allowed, online CPU? */
2056 for_each_cpu(dest_cpu, p->cpus_ptr) {
2057 if (!is_cpu_allowed(p, dest_cpu))
2058 continue;
2059
2060 goto out;
2061 }
2062
2063 /* No more Mr. Nice Guy. */
2064 switch (state) {
2065 case cpuset:
2066 if (IS_ENABLED(CONFIG_CPUSETS)) {
2067 cpuset_cpus_allowed_fallback(p);
2068 state = possible;
2069 break;
2070 }
2071 /* Fall-through */
2072 case possible:
2073 do_set_cpus_allowed(p, cpu_possible_mask);
2074 state = fail;
2075 break;
2076
2077 case fail:
2078 BUG();
2079 break;
2080 }
2081 }
2082
2083out:
2084 if (state != cpuset) {
2085 /*
2086 * Don't tell them about moving exiting tasks or
2087 * kernel threads (both mm NULL), since they never
2088 * leave kernel.
2089 */
2090 if (p->mm && printk_ratelimit()) {
2091 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2092 task_pid_nr(p), p->comm, cpu);
2093 }
2094 }
2095
2096 return dest_cpu;
2097}
2098
2099/*
2100 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2101 */
2102static inline
2103int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
2104{
2105 lockdep_assert_held(&p->pi_lock);
2106
2107 if (p->nr_cpus_allowed > 1)
2108 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
2109 else
2110 cpu = cpumask_any(p->cpus_ptr);
2111
2112 /*
2113 * In order not to call set_task_cpu() on a blocking task we need
2114 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2115 * CPU.
2116 *
2117 * Since this is common to all placement strategies, this lives here.
2118 *
2119 * [ this allows ->select_task() to simply return task_cpu(p) and
2120 * not worry about this generic constraint ]
2121 */
2122 if (unlikely(!is_cpu_allowed(p, cpu)))
2123 cpu = select_fallback_rq(task_cpu(p), p);
2124
2125 return cpu;
2126}
2127
2128static void update_avg(u64 *avg, u64 sample)
2129{
2130 s64 diff = sample - *avg;
2131 *avg += diff >> 3;
2132}
2133
2134void sched_set_stop_task(int cpu, struct task_struct *stop)
2135{
2136 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2137 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2138
2139 if (stop) {
2140 /*
2141 * Make it appear like a SCHED_FIFO task, its something
2142 * userspace knows about and won't get confused about.
2143 *
2144 * Also, it will make PI more or less work without too
2145 * much confusion -- but then, stop work should not
2146 * rely on PI working anyway.
2147 */
2148 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2149
2150 stop->sched_class = &stop_sched_class;
2151 }
2152
2153 cpu_rq(cpu)->stop = stop;
2154
2155 if (old_stop) {
2156 /*
2157 * Reset it back to a normal scheduling class so that
2158 * it can die in pieces.
2159 */
2160 old_stop->sched_class = &rt_sched_class;
2161 }
2162}
2163
2164#else
2165
2166static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2167 const struct cpumask *new_mask, bool check)
2168{
2169 return set_cpus_allowed_ptr(p, new_mask);
2170}
2171
2172#endif /* CONFIG_SMP */
2173
2174static void
2175ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2176{
2177 struct rq *rq;
2178
2179 if (!schedstat_enabled())
2180 return;
2181
2182 rq = this_rq();
2183
2184#ifdef CONFIG_SMP
2185 if (cpu == rq->cpu) {
2186 __schedstat_inc(rq->ttwu_local);
2187 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2188 } else {
2189 struct sched_domain *sd;
2190
2191 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2192 rcu_read_lock();
2193 for_each_domain(rq->cpu, sd) {
2194 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2195 __schedstat_inc(sd->ttwu_wake_remote);
2196 break;
2197 }
2198 }
2199 rcu_read_unlock();
2200 }
2201
2202 if (wake_flags & WF_MIGRATED)
2203 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2204#endif /* CONFIG_SMP */
2205
2206 __schedstat_inc(rq->ttwu_count);
2207 __schedstat_inc(p->se.statistics.nr_wakeups);
2208
2209 if (wake_flags & WF_SYNC)
2210 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2211}
2212
2213/*
2214 * Mark the task runnable and perform wakeup-preemption.
2215 */
2216static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2217 struct rq_flags *rf)
2218{
2219 check_preempt_curr(rq, p, wake_flags);
2220 p->state = TASK_RUNNING;
2221 trace_sched_wakeup(p);
2222
2223#ifdef CONFIG_SMP
2224 if (p->sched_class->task_woken) {
2225 /*
2226 * Our task @p is fully woken up and running; so its safe to
2227 * drop the rq->lock, hereafter rq is only used for statistics.
2228 */
2229 rq_unpin_lock(rq, rf);
2230 p->sched_class->task_woken(rq, p);
2231 rq_repin_lock(rq, rf);
2232 }
2233
2234 if (rq->idle_stamp) {
2235 u64 delta = rq_clock(rq) - rq->idle_stamp;
2236 u64 max = 2*rq->max_idle_balance_cost;
2237
2238 update_avg(&rq->avg_idle, delta);
2239
2240 if (rq->avg_idle > max)
2241 rq->avg_idle = max;
2242
2243 rq->idle_stamp = 0;
2244 }
2245#endif
2246}
2247
2248static void
2249ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2250 struct rq_flags *rf)
2251{
2252 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2253
2254 lockdep_assert_held(&rq->lock);
2255
2256#ifdef CONFIG_SMP
2257 if (p->sched_contributes_to_load)
2258 rq->nr_uninterruptible--;
2259
2260 if (wake_flags & WF_MIGRATED)
2261 en_flags |= ENQUEUE_MIGRATED;
2262#endif
2263
2264 activate_task(rq, p, en_flags);
2265 ttwu_do_wakeup(rq, p, wake_flags, rf);
2266}
2267
2268/*
2269 * Called in case the task @p isn't fully descheduled from its runqueue,
2270 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2271 * since all we need to do is flip p->state to TASK_RUNNING, since
2272 * the task is still ->on_rq.
2273 */
2274static int ttwu_remote(struct task_struct *p, int wake_flags)
2275{
2276 struct rq_flags rf;
2277 struct rq *rq;
2278 int ret = 0;
2279
2280 rq = __task_rq_lock(p, &rf);
2281 if (task_on_rq_queued(p)) {
2282 /* check_preempt_curr() may use rq clock */
2283 update_rq_clock(rq);
2284 ttwu_do_wakeup(rq, p, wake_flags, &rf);
2285 ret = 1;
2286 }
2287 __task_rq_unlock(rq, &rf);
2288
2289 return ret;
2290}
2291
2292#ifdef CONFIG_SMP
2293void sched_ttwu_pending(void)
2294{
2295 struct rq *rq = this_rq();
2296 struct llist_node *llist = llist_del_all(&rq->wake_list);
2297 struct task_struct *p, *t;
2298 struct rq_flags rf;
2299
2300 if (!llist)
2301 return;
2302
2303 rq_lock_irqsave(rq, &rf);
2304 update_rq_clock(rq);
2305
2306 llist_for_each_entry_safe(p, t, llist, wake_entry)
2307 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
2308
2309 rq_unlock_irqrestore(rq, &rf);
2310}
2311
2312void scheduler_ipi(void)
2313{
2314 /*
2315 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
2316 * TIF_NEED_RESCHED remotely (for the first time) will also send
2317 * this IPI.
2318 */
2319 preempt_fold_need_resched();
2320
2321 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2322 return;
2323
2324 /*
2325 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2326 * traditionally all their work was done from the interrupt return
2327 * path. Now that we actually do some work, we need to make sure
2328 * we do call them.
2329 *
2330 * Some archs already do call them, luckily irq_enter/exit nest
2331 * properly.
2332 *
2333 * Arguably we should visit all archs and update all handlers,
2334 * however a fair share of IPIs are still resched only so this would
2335 * somewhat pessimize the simple resched case.
2336 */
2337 irq_enter();
2338 sched_ttwu_pending();
2339
2340 /*
2341 * Check if someone kicked us for doing the nohz idle load balance.
2342 */
2343 if (unlikely(got_nohz_idle_kick())) {
2344 this_rq()->idle_balance = 1;
2345 raise_softirq_irqoff(SCHED_SOFTIRQ);
2346 }
2347 irq_exit();
2348}
2349
2350static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
2351{
2352 struct rq *rq = cpu_rq(cpu);
2353
2354 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
2355
2356 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
2357 if (!set_nr_if_polling(rq->idle))
2358 smp_send_reschedule(cpu);
2359 else
2360 trace_sched_wake_idle_without_ipi(cpu);
2361 }
2362}
2363
2364void wake_up_if_idle(int cpu)
2365{
2366 struct rq *rq = cpu_rq(cpu);
2367 struct rq_flags rf;
2368
2369 rcu_read_lock();
2370
2371 if (!is_idle_task(rcu_dereference(rq->curr)))
2372 goto out;
2373
2374 if (set_nr_if_polling(rq->idle)) {
2375 trace_sched_wake_idle_without_ipi(cpu);
2376 } else {
2377 rq_lock_irqsave(rq, &rf);
2378 if (is_idle_task(rq->curr))
2379 smp_send_reschedule(cpu);
2380 /* Else CPU is not idle, do nothing here: */
2381 rq_unlock_irqrestore(rq, &rf);
2382 }
2383
2384out:
2385 rcu_read_unlock();
2386}
2387
2388bool cpus_share_cache(int this_cpu, int that_cpu)
2389{
2390 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2391}
2392#endif /* CONFIG_SMP */
2393
2394static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
2395{
2396 struct rq *rq = cpu_rq(cpu);
2397 struct rq_flags rf;
2398
2399#if defined(CONFIG_SMP)
2400 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
2401 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
2402 ttwu_queue_remote(p, cpu, wake_flags);
2403 return;
2404 }
2405#endif
2406
2407 rq_lock(rq, &rf);
2408 update_rq_clock(rq);
2409 ttwu_do_activate(rq, p, wake_flags, &rf);
2410 rq_unlock(rq, &rf);
2411}
2412
2413/*
2414 * Notes on Program-Order guarantees on SMP systems.
2415 *
2416 * MIGRATION
2417 *
2418 * The basic program-order guarantee on SMP systems is that when a task [t]
2419 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2420 * execution on its new CPU [c1].
2421 *
2422 * For migration (of runnable tasks) this is provided by the following means:
2423 *
2424 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2425 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2426 * rq(c1)->lock (if not at the same time, then in that order).
2427 * C) LOCK of the rq(c1)->lock scheduling in task
2428 *
2429 * Release/acquire chaining guarantees that B happens after A and C after B.
2430 * Note: the CPU doing B need not be c0 or c1
2431 *
2432 * Example:
2433 *
2434 * CPU0 CPU1 CPU2
2435 *
2436 * LOCK rq(0)->lock
2437 * sched-out X
2438 * sched-in Y
2439 * UNLOCK rq(0)->lock
2440 *
2441 * LOCK rq(0)->lock // orders against CPU0
2442 * dequeue X
2443 * UNLOCK rq(0)->lock
2444 *
2445 * LOCK rq(1)->lock
2446 * enqueue X
2447 * UNLOCK rq(1)->lock
2448 *
2449 * LOCK rq(1)->lock // orders against CPU2
2450 * sched-out Z
2451 * sched-in X
2452 * UNLOCK rq(1)->lock
2453 *
2454 *
2455 * BLOCKING -- aka. SLEEP + WAKEUP
2456 *
2457 * For blocking we (obviously) need to provide the same guarantee as for
2458 * migration. However the means are completely different as there is no lock
2459 * chain to provide order. Instead we do:
2460 *
2461 * 1) smp_store_release(X->on_cpu, 0)
2462 * 2) smp_cond_load_acquire(!X->on_cpu)
2463 *
2464 * Example:
2465 *
2466 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2467 *
2468 * LOCK rq(0)->lock LOCK X->pi_lock
2469 * dequeue X
2470 * sched-out X
2471 * smp_store_release(X->on_cpu, 0);
2472 *
2473 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2474 * X->state = WAKING
2475 * set_task_cpu(X,2)
2476 *
2477 * LOCK rq(2)->lock
2478 * enqueue X
2479 * X->state = RUNNING
2480 * UNLOCK rq(2)->lock
2481 *
2482 * LOCK rq(2)->lock // orders against CPU1
2483 * sched-out Z
2484 * sched-in X
2485 * UNLOCK rq(2)->lock
2486 *
2487 * UNLOCK X->pi_lock
2488 * UNLOCK rq(0)->lock
2489 *
2490 *
2491 * However, for wakeups there is a second guarantee we must provide, namely we
2492 * must ensure that CONDITION=1 done by the caller can not be reordered with
2493 * accesses to the task state; see try_to_wake_up() and set_current_state().
2494 */
2495
2496/**
2497 * try_to_wake_up - wake up a thread
2498 * @p: the thread to be awakened
2499 * @state: the mask of task states that can be woken
2500 * @wake_flags: wake modifier flags (WF_*)
2501 *
2502 * If (@state & @p->state) @p->state = TASK_RUNNING.
2503 *
2504 * If the task was not queued/runnable, also place it back on a runqueue.
2505 *
2506 * Atomic against schedule() which would dequeue a task, also see
2507 * set_current_state().
2508 *
2509 * This function executes a full memory barrier before accessing the task
2510 * state; see set_current_state().
2511 *
2512 * Return: %true if @p->state changes (an actual wakeup was done),
2513 * %false otherwise.
2514 */
2515static int
2516try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2517{
2518 unsigned long flags;
2519 int cpu, success = 0;
2520
2521 preempt_disable();
2522 if (p == current) {
2523 /*
2524 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2525 * == smp_processor_id()'. Together this means we can special
2526 * case the whole 'p->on_rq && ttwu_remote()' case below
2527 * without taking any locks.
2528 *
2529 * In particular:
2530 * - we rely on Program-Order guarantees for all the ordering,
2531 * - we're serialized against set_special_state() by virtue of
2532 * it disabling IRQs (this allows not taking ->pi_lock).
2533 */
2534 if (!(p->state & state))
2535 goto out;
2536
2537 success = 1;
2538 cpu = task_cpu(p);
2539 trace_sched_waking(p);
2540 p->state = TASK_RUNNING;
2541 trace_sched_wakeup(p);
2542 goto out;
2543 }
2544
2545 /*
2546 * If we are going to wake up a thread waiting for CONDITION we
2547 * need to ensure that CONDITION=1 done by the caller can not be
2548 * reordered with p->state check below. This pairs with mb() in
2549 * set_current_state() the waiting thread does.
2550 */
2551 raw_spin_lock_irqsave(&p->pi_lock, flags);
2552 smp_mb__after_spinlock();
2553 if (!(p->state & state))
2554 goto unlock;
2555
2556 trace_sched_waking(p);
2557
2558 /* We're going to change ->state: */
2559 success = 1;
2560 cpu = task_cpu(p);
2561
2562 /*
2563 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2564 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2565 * in smp_cond_load_acquire() below.
2566 *
2567 * sched_ttwu_pending() try_to_wake_up()
2568 * STORE p->on_rq = 1 LOAD p->state
2569 * UNLOCK rq->lock
2570 *
2571 * __schedule() (switch to task 'p')
2572 * LOCK rq->lock smp_rmb();
2573 * smp_mb__after_spinlock();
2574 * UNLOCK rq->lock
2575 *
2576 * [task p]
2577 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2578 *
2579 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2580 * __schedule(). See the comment for smp_mb__after_spinlock().
2581 */
2582 smp_rmb();
2583 if (p->on_rq && ttwu_remote(p, wake_flags))
2584 goto unlock;
2585
2586#ifdef CONFIG_SMP
2587 /*
2588 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2589 * possible to, falsely, observe p->on_cpu == 0.
2590 *
2591 * One must be running (->on_cpu == 1) in order to remove oneself
2592 * from the runqueue.
2593 *
2594 * __schedule() (switch to task 'p') try_to_wake_up()
2595 * STORE p->on_cpu = 1 LOAD p->on_rq
2596 * UNLOCK rq->lock
2597 *
2598 * __schedule() (put 'p' to sleep)
2599 * LOCK rq->lock smp_rmb();
2600 * smp_mb__after_spinlock();
2601 * STORE p->on_rq = 0 LOAD p->on_cpu
2602 *
2603 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2604 * __schedule(). See the comment for smp_mb__after_spinlock().
2605 */
2606 smp_rmb();
2607
2608 /*
2609 * If the owning (remote) CPU is still in the middle of schedule() with
2610 * this task as prev, wait until its done referencing the task.
2611 *
2612 * Pairs with the smp_store_release() in finish_task().
2613 *
2614 * This ensures that tasks getting woken will be fully ordered against
2615 * their previous state and preserve Program Order.
2616 */
2617 smp_cond_load_acquire(&p->on_cpu, !VAL);
2618
2619 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2620 p->state = TASK_WAKING;
2621
2622 if (p->in_iowait) {
2623 delayacct_blkio_end(p);
2624 atomic_dec(&task_rq(p)->nr_iowait);
2625 }
2626
2627 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2628 if (task_cpu(p) != cpu) {
2629 wake_flags |= WF_MIGRATED;
2630 psi_ttwu_dequeue(p);
2631 set_task_cpu(p, cpu);
2632 }
2633
2634#else /* CONFIG_SMP */
2635
2636 if (p->in_iowait) {
2637 delayacct_blkio_end(p);
2638 atomic_dec(&task_rq(p)->nr_iowait);
2639 }
2640
2641#endif /* CONFIG_SMP */
2642
2643 ttwu_queue(p, cpu, wake_flags);
2644unlock:
2645 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2646out:
2647 if (success)
2648 ttwu_stat(p, cpu, wake_flags);
2649 preempt_enable();
2650
2651 return success;
2652}
2653
2654/**
2655 * wake_up_process - Wake up a specific process
2656 * @p: The process to be woken up.
2657 *
2658 * Attempt to wake up the nominated process and move it to the set of runnable
2659 * processes.
2660 *
2661 * Return: 1 if the process was woken up, 0 if it was already running.
2662 *
2663 * This function executes a full memory barrier before accessing the task state.
2664 */
2665int wake_up_process(struct task_struct *p)
2666{
2667 return try_to_wake_up(p, TASK_NORMAL, 0);
2668}
2669EXPORT_SYMBOL(wake_up_process);
2670
2671int wake_up_state(struct task_struct *p, unsigned int state)
2672{
2673 return try_to_wake_up(p, state, 0);
2674}
2675
2676/*
2677 * Perform scheduler related setup for a newly forked process p.
2678 * p is forked by current.
2679 *
2680 * __sched_fork() is basic setup used by init_idle() too:
2681 */
2682static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2683{
2684 p->on_rq = 0;
2685
2686 p->se.on_rq = 0;
2687 p->se.exec_start = 0;
2688 p->se.sum_exec_runtime = 0;
2689 p->se.prev_sum_exec_runtime = 0;
2690 p->se.nr_migrations = 0;
2691 p->se.vruntime = 0;
2692 INIT_LIST_HEAD(&p->se.group_node);
2693
2694#ifdef CONFIG_FAIR_GROUP_SCHED
2695 p->se.cfs_rq = NULL;
2696#endif
2697
2698#ifdef CONFIG_SCHEDSTATS
2699 /* Even if schedstat is disabled, there should not be garbage */
2700 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2701#endif
2702
2703 RB_CLEAR_NODE(&p->dl.rb_node);
2704 init_dl_task_timer(&p->dl);
2705 init_dl_inactive_task_timer(&p->dl);
2706 __dl_clear_params(p);
2707
2708 INIT_LIST_HEAD(&p->rt.run_list);
2709 p->rt.timeout = 0;
2710 p->rt.time_slice = sched_rr_timeslice;
2711 p->rt.on_rq = 0;
2712 p->rt.on_list = 0;
2713
2714#ifdef CONFIG_PREEMPT_NOTIFIERS
2715 INIT_HLIST_HEAD(&p->preempt_notifiers);
2716#endif
2717
2718#ifdef CONFIG_COMPACTION
2719 p->capture_control = NULL;
2720#endif
2721 init_numa_balancing(clone_flags, p);
2722}
2723
2724DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2725
2726#ifdef CONFIG_NUMA_BALANCING
2727
2728void set_numabalancing_state(bool enabled)
2729{
2730 if (enabled)
2731 static_branch_enable(&sched_numa_balancing);
2732 else
2733 static_branch_disable(&sched_numa_balancing);
2734}
2735
2736#ifdef CONFIG_PROC_SYSCTL
2737int sysctl_numa_balancing(struct ctl_table *table, int write,
2738 void __user *buffer, size_t *lenp, loff_t *ppos)
2739{
2740 struct ctl_table t;
2741 int err;
2742 int state = static_branch_likely(&sched_numa_balancing);
2743
2744 if (write && !capable(CAP_SYS_ADMIN))
2745 return -EPERM;
2746
2747 t = *table;
2748 t.data = &state;
2749 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2750 if (err < 0)
2751 return err;
2752 if (write)
2753 set_numabalancing_state(state);
2754 return err;
2755}
2756#endif
2757#endif
2758
2759#ifdef CONFIG_SCHEDSTATS
2760
2761DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2762static bool __initdata __sched_schedstats = false;
2763
2764static void set_schedstats(bool enabled)
2765{
2766 if (enabled)
2767 static_branch_enable(&sched_schedstats);
2768 else
2769 static_branch_disable(&sched_schedstats);
2770}
2771
2772void force_schedstat_enabled(void)
2773{
2774 if (!schedstat_enabled()) {
2775 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2776 static_branch_enable(&sched_schedstats);
2777 }
2778}
2779
2780static int __init setup_schedstats(char *str)
2781{
2782 int ret = 0;
2783 if (!str)
2784 goto out;
2785
2786 /*
2787 * This code is called before jump labels have been set up, so we can't
2788 * change the static branch directly just yet. Instead set a temporary
2789 * variable so init_schedstats() can do it later.
2790 */
2791 if (!strcmp(str, "enable")) {
2792 __sched_schedstats = true;
2793 ret = 1;
2794 } else if (!strcmp(str, "disable")) {
2795 __sched_schedstats = false;
2796 ret = 1;
2797 }
2798out:
2799 if (!ret)
2800 pr_warn("Unable to parse schedstats=\n");
2801
2802 return ret;
2803}
2804__setup("schedstats=", setup_schedstats);
2805
2806static void __init init_schedstats(void)
2807{
2808 set_schedstats(__sched_schedstats);
2809}
2810
2811#ifdef CONFIG_PROC_SYSCTL
2812int sysctl_schedstats(struct ctl_table *table, int write,
2813 void __user *buffer, size_t *lenp, loff_t *ppos)
2814{
2815 struct ctl_table t;
2816 int err;
2817 int state = static_branch_likely(&sched_schedstats);
2818
2819 if (write && !capable(CAP_SYS_ADMIN))
2820 return -EPERM;
2821
2822 t = *table;
2823 t.data = &state;
2824 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2825 if (err < 0)
2826 return err;
2827 if (write)
2828 set_schedstats(state);
2829 return err;
2830}
2831#endif /* CONFIG_PROC_SYSCTL */
2832#else /* !CONFIG_SCHEDSTATS */
2833static inline void init_schedstats(void) {}
2834#endif /* CONFIG_SCHEDSTATS */
2835
2836/*
2837 * fork()/clone()-time setup:
2838 */
2839int sched_fork(unsigned long clone_flags, struct task_struct *p)
2840{
2841 unsigned long flags;
2842
2843 __sched_fork(clone_flags, p);
2844 /*
2845 * We mark the process as NEW here. This guarantees that
2846 * nobody will actually run it, and a signal or other external
2847 * event cannot wake it up and insert it on the runqueue either.
2848 */
2849 p->state = TASK_NEW;
2850
2851 /*
2852 * Make sure we do not leak PI boosting priority to the child.
2853 */
2854 p->prio = current->normal_prio;
2855
2856 uclamp_fork(p);
2857
2858 /*
2859 * Revert to default priority/policy on fork if requested.
2860 */
2861 if (unlikely(p->sched_reset_on_fork)) {
2862 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2863 p->policy = SCHED_NORMAL;
2864 p->static_prio = NICE_TO_PRIO(0);
2865 p->rt_priority = 0;
2866 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2867 p->static_prio = NICE_TO_PRIO(0);
2868
2869 p->prio = p->normal_prio = __normal_prio(p);
2870 set_load_weight(p, false);
2871
2872 /*
2873 * We don't need the reset flag anymore after the fork. It has
2874 * fulfilled its duty:
2875 */
2876 p->sched_reset_on_fork = 0;
2877 }
2878
2879 if (dl_prio(p->prio))
2880 return -EAGAIN;
2881 else if (rt_prio(p->prio))
2882 p->sched_class = &rt_sched_class;
2883 else
2884 p->sched_class = &fair_sched_class;
2885
2886 init_entity_runnable_average(&p->se);
2887
2888 /*
2889 * The child is not yet in the pid-hash so no cgroup attach races,
2890 * and the cgroup is pinned to this child due to cgroup_fork()
2891 * is ran before sched_fork().
2892 *
2893 * Silence PROVE_RCU.
2894 */
2895 raw_spin_lock_irqsave(&p->pi_lock, flags);
2896 /*
2897 * We're setting the CPU for the first time, we don't migrate,
2898 * so use __set_task_cpu().
2899 */
2900 __set_task_cpu(p, smp_processor_id());
2901 if (p->sched_class->task_fork)
2902 p->sched_class->task_fork(p);
2903 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2904
2905#ifdef CONFIG_SCHED_INFO
2906 if (likely(sched_info_on()))
2907 memset(&p->sched_info, 0, sizeof(p->sched_info));
2908#endif
2909#if defined(CONFIG_SMP)
2910 p->on_cpu = 0;
2911#endif
2912 init_task_preempt_count(p);
2913#ifdef CONFIG_SMP
2914 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2915 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2916#endif
2917 return 0;
2918}
2919
2920unsigned long to_ratio(u64 period, u64 runtime)
2921{
2922 if (runtime == RUNTIME_INF)
2923 return BW_UNIT;
2924
2925 /*
2926 * Doing this here saves a lot of checks in all
2927 * the calling paths, and returning zero seems
2928 * safe for them anyway.
2929 */
2930 if (period == 0)
2931 return 0;
2932
2933 return div64_u64(runtime << BW_SHIFT, period);
2934}
2935
2936/*
2937 * wake_up_new_task - wake up a newly created task for the first time.
2938 *
2939 * This function will do some initial scheduler statistics housekeeping
2940 * that must be done for every newly created context, then puts the task
2941 * on the runqueue and wakes it.
2942 */
2943void wake_up_new_task(struct task_struct *p)
2944{
2945 struct rq_flags rf;
2946 struct rq *rq;
2947
2948 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2949 p->state = TASK_RUNNING;
2950#ifdef CONFIG_SMP
2951 /*
2952 * Fork balancing, do it here and not earlier because:
2953 * - cpus_ptr can change in the fork path
2954 * - any previously selected CPU might disappear through hotplug
2955 *
2956 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2957 * as we're not fully set-up yet.
2958 */
2959 p->recent_used_cpu = task_cpu(p);
2960 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2961#endif
2962 rq = __task_rq_lock(p, &rf);
2963 update_rq_clock(rq);
2964 post_init_entity_util_avg(p);
2965
2966 activate_task(rq, p, ENQUEUE_NOCLOCK);
2967 trace_sched_wakeup_new(p);
2968 check_preempt_curr(rq, p, WF_FORK);
2969#ifdef CONFIG_SMP
2970 if (p->sched_class->task_woken) {
2971 /*
2972 * Nothing relies on rq->lock after this, so its fine to
2973 * drop it.
2974 */
2975 rq_unpin_lock(rq, &rf);
2976 p->sched_class->task_woken(rq, p);
2977 rq_repin_lock(rq, &rf);
2978 }
2979#endif
2980 task_rq_unlock(rq, p, &rf);
2981}
2982
2983#ifdef CONFIG_PREEMPT_NOTIFIERS
2984
2985static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2986
2987void preempt_notifier_inc(void)
2988{
2989 static_branch_inc(&preempt_notifier_key);
2990}
2991EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2992
2993void preempt_notifier_dec(void)
2994{
2995 static_branch_dec(&preempt_notifier_key);
2996}
2997EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2998
2999/**
3000 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3001 * @notifier: notifier struct to register
3002 */
3003void preempt_notifier_register(struct preempt_notifier *notifier)
3004{
3005 if (!static_branch_unlikely(&preempt_notifier_key))
3006 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3007
3008 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3009}
3010EXPORT_SYMBOL_GPL(preempt_notifier_register);
3011
3012/**
3013 * preempt_notifier_unregister - no longer interested in preemption notifications
3014 * @notifier: notifier struct to unregister
3015 *
3016 * This is *not* safe to call from within a preemption notifier.
3017 */
3018void preempt_notifier_unregister(struct preempt_notifier *notifier)
3019{
3020 hlist_del(¬ifier->link);
3021}
3022EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3023
3024static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3025{
3026 struct preempt_notifier *notifier;
3027
3028 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3029 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3030}
3031
3032static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3033{
3034 if (static_branch_unlikely(&preempt_notifier_key))
3035 __fire_sched_in_preempt_notifiers(curr);
3036}
3037
3038static void
3039__fire_sched_out_preempt_notifiers(struct task_struct *curr,
3040 struct task_struct *next)
3041{
3042 struct preempt_notifier *notifier;
3043
3044 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3045 notifier->ops->sched_out(notifier, next);
3046}
3047
3048static __always_inline void
3049fire_sched_out_preempt_notifiers(struct task_struct *curr,
3050 struct task_struct *next)
3051{
3052 if (static_branch_unlikely(&preempt_notifier_key))
3053 __fire_sched_out_preempt_notifiers(curr, next);
3054}
3055
3056#else /* !CONFIG_PREEMPT_NOTIFIERS */
3057
3058static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3059{
3060}
3061
3062static inline void
3063fire_sched_out_preempt_notifiers(struct task_struct *curr,
3064 struct task_struct *next)
3065{
3066}
3067
3068#endif /* CONFIG_PREEMPT_NOTIFIERS */
3069
3070static inline void prepare_task(struct task_struct *next)
3071{
3072#ifdef CONFIG_SMP
3073 /*
3074 * Claim the task as running, we do this before switching to it
3075 * such that any running task will have this set.
3076 */
3077 next->on_cpu = 1;
3078#endif
3079}
3080
3081static inline void finish_task(struct task_struct *prev)
3082{
3083#ifdef CONFIG_SMP
3084 /*
3085 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3086 * We must ensure this doesn't happen until the switch is completely
3087 * finished.
3088 *
3089 * In particular, the load of prev->state in finish_task_switch() must
3090 * happen before this.
3091 *
3092 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3093 */
3094 smp_store_release(&prev->on_cpu, 0);
3095#endif
3096}
3097
3098static inline void
3099prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
3100{
3101 /*
3102 * Since the runqueue lock will be released by the next
3103 * task (which is an invalid locking op but in the case
3104 * of the scheduler it's an obvious special-case), so we
3105 * do an early lockdep release here:
3106 */
3107 rq_unpin_lock(rq, rf);
3108 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3109#ifdef CONFIG_DEBUG_SPINLOCK
3110 /* this is a valid case when another task releases the spinlock */
3111 rq->lock.owner = next;
3112#endif
3113}
3114
3115static inline void finish_lock_switch(struct rq *rq)
3116{
3117 /*
3118 * If we are tracking spinlock dependencies then we have to
3119 * fix up the runqueue lock - which gets 'carried over' from
3120 * prev into current:
3121 */
3122 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
3123 raw_spin_unlock_irq(&rq->lock);
3124}
3125
3126/*
3127 * NOP if the arch has not defined these:
3128 */
3129
3130#ifndef prepare_arch_switch
3131# define prepare_arch_switch(next) do { } while (0)
3132#endif
3133
3134#ifndef finish_arch_post_lock_switch
3135# define finish_arch_post_lock_switch() do { } while (0)
3136#endif
3137
3138/**
3139 * prepare_task_switch - prepare to switch tasks
3140 * @rq: the runqueue preparing to switch
3141 * @prev: the current task that is being switched out
3142 * @next: the task we are going to switch to.
3143 *
3144 * This is called with the rq lock held and interrupts off. It must
3145 * be paired with a subsequent finish_task_switch after the context
3146 * switch.
3147 *
3148 * prepare_task_switch sets up locking and calls architecture specific
3149 * hooks.
3150 */
3151static inline void
3152prepare_task_switch(struct rq *rq, struct task_struct *prev,
3153 struct task_struct *next)
3154{
3155 kcov_prepare_switch(prev);
3156 sched_info_switch(rq, prev, next);
3157 perf_event_task_sched_out(prev, next);
3158 rseq_preempt(prev);
3159 fire_sched_out_preempt_notifiers(prev, next);
3160 prepare_task(next);
3161 prepare_arch_switch(next);
3162}
3163
3164/**
3165 * finish_task_switch - clean up after a task-switch
3166 * @prev: the thread we just switched away from.
3167 *
3168 * finish_task_switch must be called after the context switch, paired
3169 * with a prepare_task_switch call before the context switch.
3170 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3171 * and do any other architecture-specific cleanup actions.
3172 *
3173 * Note that we may have delayed dropping an mm in context_switch(). If
3174 * so, we finish that here outside of the runqueue lock. (Doing it
3175 * with the lock held can cause deadlocks; see schedule() for
3176 * details.)
3177 *
3178 * The context switch have flipped the stack from under us and restored the
3179 * local variables which were saved when this task called schedule() in the
3180 * past. prev == current is still correct but we need to recalculate this_rq
3181 * because prev may have moved to another CPU.
3182 */
3183static struct rq *finish_task_switch(struct task_struct *prev)
3184 __releases(rq->lock)
3185{
3186 struct rq *rq = this_rq();
3187 struct mm_struct *mm = rq->prev_mm;
3188 long prev_state;
3189
3190 /*
3191 * The previous task will have left us with a preempt_count of 2
3192 * because it left us after:
3193 *
3194 * schedule()
3195 * preempt_disable(); // 1
3196 * __schedule()
3197 * raw_spin_lock_irq(&rq->lock) // 2
3198 *
3199 * Also, see FORK_PREEMPT_COUNT.
3200 */
3201 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
3202 "corrupted preempt_count: %s/%d/0x%x\n",
3203 current->comm, current->pid, preempt_count()))
3204 preempt_count_set(FORK_PREEMPT_COUNT);
3205
3206 rq->prev_mm = NULL;
3207
3208 /*
3209 * A task struct has one reference for the use as "current".
3210 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3211 * schedule one last time. The schedule call will never return, and
3212 * the scheduled task must drop that reference.
3213 *
3214 * We must observe prev->state before clearing prev->on_cpu (in
3215 * finish_task), otherwise a concurrent wakeup can get prev
3216 * running on another CPU and we could rave with its RUNNING -> DEAD
3217 * transition, resulting in a double drop.
3218 */
3219 prev_state = prev->state;
3220 vtime_task_switch(prev);
3221 perf_event_task_sched_in(prev, current);
3222 finish_task(prev);
3223 finish_lock_switch(rq);
3224 finish_arch_post_lock_switch();
3225 kcov_finish_switch(current);
3226
3227 fire_sched_in_preempt_notifiers(current);
3228 /*
3229 * When switching through a kernel thread, the loop in
3230 * membarrier_{private,global}_expedited() may have observed that
3231 * kernel thread and not issued an IPI. It is therefore possible to
3232 * schedule between user->kernel->user threads without passing though
3233 * switch_mm(). Membarrier requires a barrier after storing to
3234 * rq->curr, before returning to userspace, so provide them here:
3235 *
3236 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3237 * provided by mmdrop(),
3238 * - a sync_core for SYNC_CORE.
3239 */
3240 if (mm) {
3241 membarrier_mm_sync_core_before_usermode(mm);
3242 mmdrop(mm);
3243 }
3244 if (unlikely(prev_state == TASK_DEAD)) {
3245 if (prev->sched_class->task_dead)
3246 prev->sched_class->task_dead(prev);
3247
3248 /*
3249 * Remove function-return probe instances associated with this
3250 * task and put them back on the free list.
3251 */
3252 kprobe_flush_task(prev);
3253
3254 /* Task is done with its stack. */
3255 put_task_stack(prev);
3256
3257 put_task_struct_rcu_user(prev);
3258 }
3259
3260 tick_nohz_task_switch();
3261 return rq;
3262}
3263
3264#ifdef CONFIG_SMP
3265
3266/* rq->lock is NOT held, but preemption is disabled */
3267static void __balance_callback(struct rq *rq)
3268{
3269 struct callback_head *head, *next;
3270 void (*func)(struct rq *rq);
3271 unsigned long flags;
3272
3273 raw_spin_lock_irqsave(&rq->lock, flags);
3274 head = rq->balance_callback;
3275 rq->balance_callback = NULL;
3276 while (head) {
3277 func = (void (*)(struct rq *))head->func;
3278 next = head->next;
3279 head->next = NULL;
3280 head = next;
3281
3282 func(rq);
3283 }
3284 raw_spin_unlock_irqrestore(&rq->lock, flags);
3285}
3286
3287static inline void balance_callback(struct rq *rq)
3288{
3289 if (unlikely(rq->balance_callback))
3290 __balance_callback(rq);
3291}
3292
3293#else
3294
3295static inline void balance_callback(struct rq *rq)
3296{
3297}
3298
3299#endif
3300
3301/**
3302 * schedule_tail - first thing a freshly forked thread must call.
3303 * @prev: the thread we just switched away from.
3304 */
3305asmlinkage __visible void schedule_tail(struct task_struct *prev)
3306 __releases(rq->lock)
3307{
3308 struct rq *rq;
3309
3310 /*
3311 * New tasks start with FORK_PREEMPT_COUNT, see there and
3312 * finish_task_switch() for details.
3313 *
3314 * finish_task_switch() will drop rq->lock() and lower preempt_count
3315 * and the preempt_enable() will end up enabling preemption (on
3316 * PREEMPT_COUNT kernels).
3317 */
3318
3319 rq = finish_task_switch(prev);
3320 balance_callback(rq);
3321 preempt_enable();
3322
3323 if (current->set_child_tid)
3324 put_user(task_pid_vnr(current), current->set_child_tid);
3325
3326 calculate_sigpending();
3327}
3328
3329/*
3330 * context_switch - switch to the new MM and the new thread's register state.
3331 */
3332static __always_inline struct rq *
3333context_switch(struct rq *rq, struct task_struct *prev,
3334 struct task_struct *next, struct rq_flags *rf)
3335{
3336 prepare_task_switch(rq, prev, next);
3337
3338 /*
3339 * For paravirt, this is coupled with an exit in switch_to to
3340 * combine the page table reload and the switch backend into
3341 * one hypercall.
3342 */
3343 arch_start_context_switch(prev);
3344
3345 /*
3346 * kernel -> kernel lazy + transfer active
3347 * user -> kernel lazy + mmgrab() active
3348 *
3349 * kernel -> user switch + mmdrop() active
3350 * user -> user switch
3351 */
3352 if (!next->mm) { // to kernel
3353 enter_lazy_tlb(prev->active_mm, next);
3354
3355 next->active_mm = prev->active_mm;
3356 if (prev->mm) // from user
3357 mmgrab(prev->active_mm);
3358 else
3359 prev->active_mm = NULL;
3360 } else { // to user
3361 membarrier_switch_mm(rq, prev->active_mm, next->mm);
3362 /*
3363 * sys_membarrier() requires an smp_mb() between setting
3364 * rq->curr / membarrier_switch_mm() and returning to userspace.
3365 *
3366 * The below provides this either through switch_mm(), or in
3367 * case 'prev->active_mm == next->mm' through
3368 * finish_task_switch()'s mmdrop().
3369 */
3370 switch_mm_irqs_off(prev->active_mm, next->mm, next);
3371
3372 if (!prev->mm) { // from kernel
3373 /* will mmdrop() in finish_task_switch(). */
3374 rq->prev_mm = prev->active_mm;
3375 prev->active_mm = NULL;
3376 }
3377 }
3378
3379 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3380
3381 prepare_lock_switch(rq, next, rf);
3382
3383 /* Here we just switch the register state and the stack. */
3384 switch_to(prev, next, prev);
3385 barrier();
3386
3387 return finish_task_switch(prev);
3388}
3389
3390/*
3391 * nr_running and nr_context_switches:
3392 *
3393 * externally visible scheduler statistics: current number of runnable
3394 * threads, total number of context switches performed since bootup.
3395 */
3396unsigned long nr_running(void)
3397{
3398 unsigned long i, sum = 0;
3399
3400 for_each_online_cpu(i)
3401 sum += cpu_rq(i)->nr_running;
3402
3403 return sum;
3404}
3405
3406/*
3407 * Check if only the current task is running on the CPU.
3408 *
3409 * Caution: this function does not check that the caller has disabled
3410 * preemption, thus the result might have a time-of-check-to-time-of-use
3411 * race. The caller is responsible to use it correctly, for example:
3412 *
3413 * - from a non-preemptible section (of course)
3414 *
3415 * - from a thread that is bound to a single CPU
3416 *
3417 * - in a loop with very short iterations (e.g. a polling loop)
3418 */
3419bool single_task_running(void)
3420{
3421 return raw_rq()->nr_running == 1;
3422}
3423EXPORT_SYMBOL(single_task_running);
3424
3425unsigned long long nr_context_switches(void)
3426{
3427 int i;
3428 unsigned long long sum = 0;
3429
3430 for_each_possible_cpu(i)
3431 sum += cpu_rq(i)->nr_switches;
3432
3433 return sum;
3434}
3435
3436/*
3437 * Consumers of these two interfaces, like for example the cpuidle menu
3438 * governor, are using nonsensical data. Preferring shallow idle state selection
3439 * for a CPU that has IO-wait which might not even end up running the task when
3440 * it does become runnable.
3441 */
3442
3443unsigned long nr_iowait_cpu(int cpu)
3444{
3445 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3446}
3447
3448/*
3449 * IO-wait accounting, and how its mostly bollocks (on SMP).
3450 *
3451 * The idea behind IO-wait account is to account the idle time that we could
3452 * have spend running if it were not for IO. That is, if we were to improve the
3453 * storage performance, we'd have a proportional reduction in IO-wait time.
3454 *
3455 * This all works nicely on UP, where, when a task blocks on IO, we account
3456 * idle time as IO-wait, because if the storage were faster, it could've been
3457 * running and we'd not be idle.
3458 *
3459 * This has been extended to SMP, by doing the same for each CPU. This however
3460 * is broken.
3461 *
3462 * Imagine for instance the case where two tasks block on one CPU, only the one
3463 * CPU will have IO-wait accounted, while the other has regular idle. Even
3464 * though, if the storage were faster, both could've ran at the same time,
3465 * utilising both CPUs.
3466 *
3467 * This means, that when looking globally, the current IO-wait accounting on
3468 * SMP is a lower bound, by reason of under accounting.
3469 *
3470 * Worse, since the numbers are provided per CPU, they are sometimes
3471 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3472 * associated with any one particular CPU, it can wake to another CPU than it
3473 * blocked on. This means the per CPU IO-wait number is meaningless.
3474 *
3475 * Task CPU affinities can make all that even more 'interesting'.
3476 */
3477
3478unsigned long nr_iowait(void)
3479{
3480 unsigned long i, sum = 0;
3481
3482 for_each_possible_cpu(i)
3483 sum += nr_iowait_cpu(i);
3484
3485 return sum;
3486}
3487
3488#ifdef CONFIG_SMP
3489
3490/*
3491 * sched_exec - execve() is a valuable balancing opportunity, because at
3492 * this point the task has the smallest effective memory and cache footprint.
3493 */
3494void sched_exec(void)
3495{
3496 struct task_struct *p = current;
3497 unsigned long flags;
3498 int dest_cpu;
3499
3500 raw_spin_lock_irqsave(&p->pi_lock, flags);
3501 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3502 if (dest_cpu == smp_processor_id())
3503 goto unlock;
3504
3505 if (likely(cpu_active(dest_cpu))) {
3506 struct migration_arg arg = { p, dest_cpu };
3507
3508 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3509 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3510 return;
3511 }
3512unlock:
3513 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3514}
3515
3516#endif
3517
3518DEFINE_PER_CPU(struct kernel_stat, kstat);
3519DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3520
3521EXPORT_PER_CPU_SYMBOL(kstat);
3522EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3523
3524/*
3525 * The function fair_sched_class.update_curr accesses the struct curr
3526 * and its field curr->exec_start; when called from task_sched_runtime(),
3527 * we observe a high rate of cache misses in practice.
3528 * Prefetching this data results in improved performance.
3529 */
3530static inline void prefetch_curr_exec_start(struct task_struct *p)
3531{
3532#ifdef CONFIG_FAIR_GROUP_SCHED
3533 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3534#else
3535 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3536#endif
3537 prefetch(curr);
3538 prefetch(&curr->exec_start);
3539}
3540
3541/*
3542 * Return accounted runtime for the task.
3543 * In case the task is currently running, return the runtime plus current's
3544 * pending runtime that have not been accounted yet.
3545 */
3546unsigned long long task_sched_runtime(struct task_struct *p)
3547{
3548 struct rq_flags rf;
3549 struct rq *rq;
3550 u64 ns;
3551
3552#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3553 /*
3554 * 64-bit doesn't need locks to atomically read a 64-bit value.
3555 * So we have a optimization chance when the task's delta_exec is 0.
3556 * Reading ->on_cpu is racy, but this is ok.
3557 *
3558 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3559 * If we race with it entering CPU, unaccounted time is 0. This is
3560 * indistinguishable from the read occurring a few cycles earlier.
3561 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3562 * been accounted, so we're correct here as well.
3563 */
3564 if (!p->on_cpu || !task_on_rq_queued(p))
3565 return p->se.sum_exec_runtime;
3566#endif
3567
3568 rq = task_rq_lock(p, &rf);
3569 /*
3570 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3571 * project cycles that may never be accounted to this
3572 * thread, breaking clock_gettime().
3573 */
3574 if (task_current(rq, p) && task_on_rq_queued(p)) {
3575 prefetch_curr_exec_start(p);
3576 update_rq_clock(rq);
3577 p->sched_class->update_curr(rq);
3578 }
3579 ns = p->se.sum_exec_runtime;
3580 task_rq_unlock(rq, p, &rf);
3581
3582 return ns;
3583}
3584
3585/*
3586 * This function gets called by the timer code, with HZ frequency.
3587 * We call it with interrupts disabled.
3588 */
3589void scheduler_tick(void)
3590{
3591 int cpu = smp_processor_id();
3592 struct rq *rq = cpu_rq(cpu);
3593 struct task_struct *curr = rq->curr;
3594 struct rq_flags rf;
3595
3596 sched_clock_tick();
3597
3598 rq_lock(rq, &rf);
3599
3600 update_rq_clock(rq);
3601 curr->sched_class->task_tick(rq, curr, 0);
3602 calc_global_load_tick(rq);
3603 psi_task_tick(rq);
3604
3605 rq_unlock(rq, &rf);
3606
3607 perf_event_task_tick();
3608
3609#ifdef CONFIG_SMP
3610 rq->idle_balance = idle_cpu(cpu);
3611 trigger_load_balance(rq);
3612#endif
3613}
3614
3615#ifdef CONFIG_NO_HZ_FULL
3616
3617struct tick_work {
3618 int cpu;
3619 atomic_t state;
3620 struct delayed_work work;
3621};
3622/* Values for ->state, see diagram below. */
3623#define TICK_SCHED_REMOTE_OFFLINE 0
3624#define TICK_SCHED_REMOTE_OFFLINING 1
3625#define TICK_SCHED_REMOTE_RUNNING 2
3626
3627/*
3628 * State diagram for ->state:
3629 *
3630 *
3631 * TICK_SCHED_REMOTE_OFFLINE
3632 * | ^
3633 * | |
3634 * | | sched_tick_remote()
3635 * | |
3636 * | |
3637 * +--TICK_SCHED_REMOTE_OFFLINING
3638 * | ^
3639 * | |
3640 * sched_tick_start() | | sched_tick_stop()
3641 * | |
3642 * V |
3643 * TICK_SCHED_REMOTE_RUNNING
3644 *
3645 *
3646 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3647 * and sched_tick_start() are happy to leave the state in RUNNING.
3648 */
3649
3650static struct tick_work __percpu *tick_work_cpu;
3651
3652static void sched_tick_remote(struct work_struct *work)
3653{
3654 struct delayed_work *dwork = to_delayed_work(work);
3655 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3656 int cpu = twork->cpu;
3657 struct rq *rq = cpu_rq(cpu);
3658 struct task_struct *curr;
3659 struct rq_flags rf;
3660 u64 delta;
3661 int os;
3662
3663 /*
3664 * Handle the tick only if it appears the remote CPU is running in full
3665 * dynticks mode. The check is racy by nature, but missing a tick or
3666 * having one too much is no big deal because the scheduler tick updates
3667 * statistics and checks timeslices in a time-independent way, regardless
3668 * of when exactly it is running.
3669 */
3670 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3671 goto out_requeue;
3672
3673 rq_lock_irq(rq, &rf);
3674 curr = rq->curr;
3675 if (is_idle_task(curr) || cpu_is_offline(cpu))
3676 goto out_unlock;
3677
3678 update_rq_clock(rq);
3679 delta = rq_clock_task(rq) - curr->se.exec_start;
3680
3681 /*
3682 * Make sure the next tick runs within a reasonable
3683 * amount of time.
3684 */
3685 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3686 curr->sched_class->task_tick(rq, curr, 0);
3687
3688out_unlock:
3689 rq_unlock_irq(rq, &rf);
3690
3691out_requeue:
3692 /*
3693 * Run the remote tick once per second (1Hz). This arbitrary
3694 * frequency is large enough to avoid overload but short enough
3695 * to keep scheduler internal stats reasonably up to date. But
3696 * first update state to reflect hotplug activity if required.
3697 */
3698 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
3699 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
3700 if (os == TICK_SCHED_REMOTE_RUNNING)
3701 queue_delayed_work(system_unbound_wq, dwork, HZ);
3702}
3703
3704static void sched_tick_start(int cpu)
3705{
3706 int os;
3707 struct tick_work *twork;
3708
3709 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3710 return;
3711
3712 WARN_ON_ONCE(!tick_work_cpu);
3713
3714 twork = per_cpu_ptr(tick_work_cpu, cpu);
3715 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
3716 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
3717 if (os == TICK_SCHED_REMOTE_OFFLINE) {
3718 twork->cpu = cpu;
3719 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3720 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3721 }
3722}
3723
3724#ifdef CONFIG_HOTPLUG_CPU
3725static void sched_tick_stop(int cpu)
3726{
3727 struct tick_work *twork;
3728 int os;
3729
3730 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3731 return;
3732
3733 WARN_ON_ONCE(!tick_work_cpu);
3734
3735 twork = per_cpu_ptr(tick_work_cpu, cpu);
3736 /* There cannot be competing actions, but don't rely on stop-machine. */
3737 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
3738 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
3739 /* Don't cancel, as this would mess up the state machine. */
3740}
3741#endif /* CONFIG_HOTPLUG_CPU */
3742
3743int __init sched_tick_offload_init(void)
3744{
3745 tick_work_cpu = alloc_percpu(struct tick_work);
3746 BUG_ON(!tick_work_cpu);
3747 return 0;
3748}
3749
3750#else /* !CONFIG_NO_HZ_FULL */
3751static inline void sched_tick_start(int cpu) { }
3752static inline void sched_tick_stop(int cpu) { }
3753#endif
3754
3755#if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3756 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3757/*
3758 * If the value passed in is equal to the current preempt count
3759 * then we just disabled preemption. Start timing the latency.
3760 */
3761static inline void preempt_latency_start(int val)
3762{
3763 if (preempt_count() == val) {
3764 unsigned long ip = get_lock_parent_ip();
3765#ifdef CONFIG_DEBUG_PREEMPT
3766 current->preempt_disable_ip = ip;
3767#endif
3768 trace_preempt_off(CALLER_ADDR0, ip);
3769 }
3770}
3771
3772void preempt_count_add(int val)
3773{
3774#ifdef CONFIG_DEBUG_PREEMPT
3775 /*
3776 * Underflow?
3777 */
3778 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3779 return;
3780#endif
3781 __preempt_count_add(val);
3782#ifdef CONFIG_DEBUG_PREEMPT
3783 /*
3784 * Spinlock count overflowing soon?
3785 */
3786 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3787 PREEMPT_MASK - 10);
3788#endif
3789 preempt_latency_start(val);
3790}
3791EXPORT_SYMBOL(preempt_count_add);
3792NOKPROBE_SYMBOL(preempt_count_add);
3793
3794/*
3795 * If the value passed in equals to the current preempt count
3796 * then we just enabled preemption. Stop timing the latency.
3797 */
3798static inline void preempt_latency_stop(int val)
3799{
3800 if (preempt_count() == val)
3801 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3802}
3803
3804void preempt_count_sub(int val)
3805{
3806#ifdef CONFIG_DEBUG_PREEMPT
3807 /*
3808 * Underflow?
3809 */
3810 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3811 return;
3812 /*
3813 * Is the spinlock portion underflowing?
3814 */
3815 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3816 !(preempt_count() & PREEMPT_MASK)))
3817 return;
3818#endif
3819
3820 preempt_latency_stop(val);
3821 __preempt_count_sub(val);
3822}
3823EXPORT_SYMBOL(preempt_count_sub);
3824NOKPROBE_SYMBOL(preempt_count_sub);
3825
3826#else
3827static inline void preempt_latency_start(int val) { }
3828static inline void preempt_latency_stop(int val) { }
3829#endif
3830
3831static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3832{
3833#ifdef CONFIG_DEBUG_PREEMPT
3834 return p->preempt_disable_ip;
3835#else
3836 return 0;
3837#endif
3838}
3839
3840/*
3841 * Print scheduling while atomic bug:
3842 */
3843static noinline void __schedule_bug(struct task_struct *prev)
3844{
3845 /* Save this before calling printk(), since that will clobber it */
3846 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3847
3848 if (oops_in_progress)
3849 return;
3850
3851 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3852 prev->comm, prev->pid, preempt_count());
3853
3854 debug_show_held_locks(prev);
3855 print_modules();
3856 if (irqs_disabled())
3857 print_irqtrace_events(prev);
3858 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3859 && in_atomic_preempt_off()) {
3860 pr_err("Preemption disabled at:");
3861 print_ip_sym(preempt_disable_ip);
3862 pr_cont("\n");
3863 }
3864 if (panic_on_warn)
3865 panic("scheduling while atomic\n");
3866
3867 dump_stack();
3868 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3869}
3870
3871/*
3872 * Various schedule()-time debugging checks and statistics:
3873 */
3874static inline void schedule_debug(struct task_struct *prev, bool preempt)
3875{
3876#ifdef CONFIG_SCHED_STACK_END_CHECK
3877 if (task_stack_end_corrupted(prev))
3878 panic("corrupted stack end detected inside scheduler\n");
3879#endif
3880
3881#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
3882 if (!preempt && prev->state && prev->non_block_count) {
3883 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
3884 prev->comm, prev->pid, prev->non_block_count);
3885 dump_stack();
3886 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3887 }
3888#endif
3889
3890 if (unlikely(in_atomic_preempt_off())) {
3891 __schedule_bug(prev);
3892 preempt_count_set(PREEMPT_DISABLED);
3893 }
3894 rcu_sleep_check();
3895
3896 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3897
3898 schedstat_inc(this_rq()->sched_count);
3899}
3900
3901/*
3902 * Pick up the highest-prio task:
3903 */
3904static inline struct task_struct *
3905pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3906{
3907 const struct sched_class *class;
3908 struct task_struct *p;
3909
3910 /*
3911 * Optimization: we know that if all tasks are in the fair class we can
3912 * call that function directly, but only if the @prev task wasn't of a
3913 * higher scheduling class, because otherwise those loose the
3914 * opportunity to pull in more work from other CPUs.
3915 */
3916 if (likely((prev->sched_class == &idle_sched_class ||
3917 prev->sched_class == &fair_sched_class) &&
3918 rq->nr_running == rq->cfs.h_nr_running)) {
3919
3920 p = fair_sched_class.pick_next_task(rq, prev, rf);
3921 if (unlikely(p == RETRY_TASK))
3922 goto restart;
3923
3924 /* Assumes fair_sched_class->next == idle_sched_class */
3925 if (unlikely(!p))
3926 p = idle_sched_class.pick_next_task(rq, prev, rf);
3927
3928 return p;
3929 }
3930
3931restart:
3932#ifdef CONFIG_SMP
3933 /*
3934 * We must do the balancing pass before put_next_task(), such
3935 * that when we release the rq->lock the task is in the same
3936 * state as before we took rq->lock.
3937 *
3938 * We can terminate the balance pass as soon as we know there is
3939 * a runnable task of @class priority or higher.
3940 */
3941 for_class_range(class, prev->sched_class, &idle_sched_class) {
3942 if (class->balance(rq, prev, rf))
3943 break;
3944 }
3945#endif
3946
3947 put_prev_task(rq, prev);
3948
3949 for_each_class(class) {
3950 p = class->pick_next_task(rq, NULL, NULL);
3951 if (p)
3952 return p;
3953 }
3954
3955 /* The idle class should always have a runnable task: */
3956 BUG();
3957}
3958
3959/*
3960 * __schedule() is the main scheduler function.
3961 *
3962 * The main means of driving the scheduler and thus entering this function are:
3963 *
3964 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3965 *
3966 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3967 * paths. For example, see arch/x86/entry_64.S.
3968 *
3969 * To drive preemption between tasks, the scheduler sets the flag in timer
3970 * interrupt handler scheduler_tick().
3971 *
3972 * 3. Wakeups don't really cause entry into schedule(). They add a
3973 * task to the run-queue and that's it.
3974 *
3975 * Now, if the new task added to the run-queue preempts the current
3976 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3977 * called on the nearest possible occasion:
3978 *
3979 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
3980 *
3981 * - in syscall or exception context, at the next outmost
3982 * preempt_enable(). (this might be as soon as the wake_up()'s
3983 * spin_unlock()!)
3984 *
3985 * - in IRQ context, return from interrupt-handler to
3986 * preemptible context
3987 *
3988 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
3989 * then at the next:
3990 *
3991 * - cond_resched() call
3992 * - explicit schedule() call
3993 * - return from syscall or exception to user-space
3994 * - return from interrupt-handler to user-space
3995 *
3996 * WARNING: must be called with preemption disabled!
3997 */
3998static void __sched notrace __schedule(bool preempt)
3999{
4000 struct task_struct *prev, *next;
4001 unsigned long *switch_count;
4002 struct rq_flags rf;
4003 struct rq *rq;
4004 int cpu;
4005
4006 cpu = smp_processor_id();
4007 rq = cpu_rq(cpu);
4008 prev = rq->curr;
4009
4010 schedule_debug(prev, preempt);
4011
4012 if (sched_feat(HRTICK))
4013 hrtick_clear(rq);
4014
4015 local_irq_disable();
4016 rcu_note_context_switch(preempt);
4017
4018 /*
4019 * Make sure that signal_pending_state()->signal_pending() below
4020 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4021 * done by the caller to avoid the race with signal_wake_up().
4022 *
4023 * The membarrier system call requires a full memory barrier
4024 * after coming from user-space, before storing to rq->curr.
4025 */
4026 rq_lock(rq, &rf);
4027 smp_mb__after_spinlock();
4028
4029 /* Promote REQ to ACT */
4030 rq->clock_update_flags <<= 1;
4031 update_rq_clock(rq);
4032
4033 switch_count = &prev->nivcsw;
4034 if (!preempt && prev->state) {
4035 if (signal_pending_state(prev->state, prev)) {
4036 prev->state = TASK_RUNNING;
4037 } else {
4038 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
4039
4040 if (prev->in_iowait) {
4041 atomic_inc(&rq->nr_iowait);
4042 delayacct_blkio_start();
4043 }
4044 }
4045 switch_count = &prev->nvcsw;
4046 }
4047
4048 next = pick_next_task(rq, prev, &rf);
4049 clear_tsk_need_resched(prev);
4050 clear_preempt_need_resched();
4051
4052 if (likely(prev != next)) {
4053 rq->nr_switches++;
4054 /*
4055 * RCU users of rcu_dereference(rq->curr) may not see
4056 * changes to task_struct made by pick_next_task().
4057 */
4058 RCU_INIT_POINTER(rq->curr, next);
4059 /*
4060 * The membarrier system call requires each architecture
4061 * to have a full memory barrier after updating
4062 * rq->curr, before returning to user-space.
4063 *
4064 * Here are the schemes providing that barrier on the
4065 * various architectures:
4066 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4067 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4068 * - finish_lock_switch() for weakly-ordered
4069 * architectures where spin_unlock is a full barrier,
4070 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4071 * is a RELEASE barrier),
4072 */
4073 ++*switch_count;
4074
4075 trace_sched_switch(preempt, prev, next);
4076
4077 /* Also unlocks the rq: */
4078 rq = context_switch(rq, prev, next, &rf);
4079 } else {
4080 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4081 rq_unlock_irq(rq, &rf);
4082 }
4083
4084 balance_callback(rq);
4085}
4086
4087void __noreturn do_task_dead(void)
4088{
4089 /* Causes final put_task_struct in finish_task_switch(): */
4090 set_special_state(TASK_DEAD);
4091
4092 /* Tell freezer to ignore us: */
4093 current->flags |= PF_NOFREEZE;
4094
4095 __schedule(false);
4096 BUG();
4097
4098 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4099 for (;;)
4100 cpu_relax();
4101}
4102
4103static inline void sched_submit_work(struct task_struct *tsk)
4104{
4105 if (!tsk->state)
4106 return;
4107
4108 /*
4109 * If a worker went to sleep, notify and ask workqueue whether
4110 * it wants to wake up a task to maintain concurrency.
4111 * As this function is called inside the schedule() context,
4112 * we disable preemption to avoid it calling schedule() again
4113 * in the possible wakeup of a kworker.
4114 */
4115 if (tsk->flags & PF_WQ_WORKER) {
4116 preempt_disable();
4117 wq_worker_sleeping(tsk);
4118 preempt_enable_no_resched();
4119 }
4120
4121 if (tsk_is_pi_blocked(tsk))
4122 return;
4123
4124 /*
4125 * If we are going to sleep and we have plugged IO queued,
4126 * make sure to submit it to avoid deadlocks.
4127 */
4128 if (blk_needs_flush_plug(tsk))
4129 blk_schedule_flush_plug(tsk);
4130}
4131
4132static void sched_update_worker(struct task_struct *tsk)
4133{
4134 if (tsk->flags & PF_WQ_WORKER)
4135 wq_worker_running(tsk);
4136}
4137
4138asmlinkage __visible void __sched schedule(void)
4139{
4140 struct task_struct *tsk = current;
4141
4142 sched_submit_work(tsk);
4143 do {
4144 preempt_disable();
4145 __schedule(false);
4146 sched_preempt_enable_no_resched();
4147 } while (need_resched());
4148 sched_update_worker(tsk);
4149}
4150EXPORT_SYMBOL(schedule);
4151
4152/*
4153 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4154 * state (have scheduled out non-voluntarily) by making sure that all
4155 * tasks have either left the run queue or have gone into user space.
4156 * As idle tasks do not do either, they must not ever be preempted
4157 * (schedule out non-voluntarily).
4158 *
4159 * schedule_idle() is similar to schedule_preempt_disable() except that it
4160 * never enables preemption because it does not call sched_submit_work().
4161 */
4162void __sched schedule_idle(void)
4163{
4164 /*
4165 * As this skips calling sched_submit_work(), which the idle task does
4166 * regardless because that function is a nop when the task is in a
4167 * TASK_RUNNING state, make sure this isn't used someplace that the
4168 * current task can be in any other state. Note, idle is always in the
4169 * TASK_RUNNING state.
4170 */
4171 WARN_ON_ONCE(current->state);
4172 do {
4173 __schedule(false);
4174 } while (need_resched());
4175}
4176
4177#ifdef CONFIG_CONTEXT_TRACKING
4178asmlinkage __visible void __sched schedule_user(void)
4179{
4180 /*
4181 * If we come here after a random call to set_need_resched(),
4182 * or we have been woken up remotely but the IPI has not yet arrived,
4183 * we haven't yet exited the RCU idle mode. Do it here manually until
4184 * we find a better solution.
4185 *
4186 * NB: There are buggy callers of this function. Ideally we
4187 * should warn if prev_state != CONTEXT_USER, but that will trigger
4188 * too frequently to make sense yet.
4189 */
4190 enum ctx_state prev_state = exception_enter();
4191 schedule();
4192 exception_exit(prev_state);
4193}
4194#endif
4195
4196/**
4197 * schedule_preempt_disabled - called with preemption disabled
4198 *
4199 * Returns with preemption disabled. Note: preempt_count must be 1
4200 */
4201void __sched schedule_preempt_disabled(void)
4202{
4203 sched_preempt_enable_no_resched();
4204 schedule();
4205 preempt_disable();
4206}
4207
4208static void __sched notrace preempt_schedule_common(void)
4209{
4210 do {
4211 /*
4212 * Because the function tracer can trace preempt_count_sub()
4213 * and it also uses preempt_enable/disable_notrace(), if
4214 * NEED_RESCHED is set, the preempt_enable_notrace() called
4215 * by the function tracer will call this function again and
4216 * cause infinite recursion.
4217 *
4218 * Preemption must be disabled here before the function
4219 * tracer can trace. Break up preempt_disable() into two
4220 * calls. One to disable preemption without fear of being
4221 * traced. The other to still record the preemption latency,
4222 * which can also be traced by the function tracer.
4223 */
4224 preempt_disable_notrace();
4225 preempt_latency_start(1);
4226 __schedule(true);
4227 preempt_latency_stop(1);
4228 preempt_enable_no_resched_notrace();
4229
4230 /*
4231 * Check again in case we missed a preemption opportunity
4232 * between schedule and now.
4233 */
4234 } while (need_resched());
4235}
4236
4237#ifdef CONFIG_PREEMPTION
4238/*
4239 * This is the entry point to schedule() from in-kernel preemption
4240 * off of preempt_enable.
4241 */
4242asmlinkage __visible void __sched notrace preempt_schedule(void)
4243{
4244 /*
4245 * If there is a non-zero preempt_count or interrupts are disabled,
4246 * we do not want to preempt the current task. Just return..
4247 */
4248 if (likely(!preemptible()))
4249 return;
4250
4251 preempt_schedule_common();
4252}
4253NOKPROBE_SYMBOL(preempt_schedule);
4254EXPORT_SYMBOL(preempt_schedule);
4255
4256/**
4257 * preempt_schedule_notrace - preempt_schedule called by tracing
4258 *
4259 * The tracing infrastructure uses preempt_enable_notrace to prevent
4260 * recursion and tracing preempt enabling caused by the tracing
4261 * infrastructure itself. But as tracing can happen in areas coming
4262 * from userspace or just about to enter userspace, a preempt enable
4263 * can occur before user_exit() is called. This will cause the scheduler
4264 * to be called when the system is still in usermode.
4265 *
4266 * To prevent this, the preempt_enable_notrace will use this function
4267 * instead of preempt_schedule() to exit user context if needed before
4268 * calling the scheduler.
4269 */
4270asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4271{
4272 enum ctx_state prev_ctx;
4273
4274 if (likely(!preemptible()))
4275 return;
4276
4277 do {
4278 /*
4279 * Because the function tracer can trace preempt_count_sub()
4280 * and it also uses preempt_enable/disable_notrace(), if
4281 * NEED_RESCHED is set, the preempt_enable_notrace() called
4282 * by the function tracer will call this function again and
4283 * cause infinite recursion.
4284 *
4285 * Preemption must be disabled here before the function
4286 * tracer can trace. Break up preempt_disable() into two
4287 * calls. One to disable preemption without fear of being
4288 * traced. The other to still record the preemption latency,
4289 * which can also be traced by the function tracer.
4290 */
4291 preempt_disable_notrace();
4292 preempt_latency_start(1);
4293 /*
4294 * Needs preempt disabled in case user_exit() is traced
4295 * and the tracer calls preempt_enable_notrace() causing
4296 * an infinite recursion.
4297 */
4298 prev_ctx = exception_enter();
4299 __schedule(true);
4300 exception_exit(prev_ctx);
4301
4302 preempt_latency_stop(1);
4303 preempt_enable_no_resched_notrace();
4304 } while (need_resched());
4305}
4306EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4307
4308#endif /* CONFIG_PREEMPTION */
4309
4310/*
4311 * This is the entry point to schedule() from kernel preemption
4312 * off of irq context.
4313 * Note, that this is called and return with irqs disabled. This will
4314 * protect us against recursive calling from irq.
4315 */
4316asmlinkage __visible void __sched preempt_schedule_irq(void)
4317{
4318 enum ctx_state prev_state;
4319
4320 /* Catch callers which need to be fixed */
4321 BUG_ON(preempt_count() || !irqs_disabled());
4322
4323 prev_state = exception_enter();
4324
4325 do {
4326 preempt_disable();
4327 local_irq_enable();
4328 __schedule(true);
4329 local_irq_disable();
4330 sched_preempt_enable_no_resched();
4331 } while (need_resched());
4332
4333 exception_exit(prev_state);
4334}
4335
4336int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4337 void *key)
4338{
4339 return try_to_wake_up(curr->private, mode, wake_flags);
4340}
4341EXPORT_SYMBOL(default_wake_function);
4342
4343#ifdef CONFIG_RT_MUTEXES
4344
4345static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4346{
4347 if (pi_task)
4348 prio = min(prio, pi_task->prio);
4349
4350 return prio;
4351}
4352
4353static inline int rt_effective_prio(struct task_struct *p, int prio)
4354{
4355 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4356
4357 return __rt_effective_prio(pi_task, prio);
4358}
4359
4360/*
4361 * rt_mutex_setprio - set the current priority of a task
4362 * @p: task to boost
4363 * @pi_task: donor task
4364 *
4365 * This function changes the 'effective' priority of a task. It does
4366 * not touch ->normal_prio like __setscheduler().
4367 *
4368 * Used by the rt_mutex code to implement priority inheritance
4369 * logic. Call site only calls if the priority of the task changed.
4370 */
4371void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4372{
4373 int prio, oldprio, queued, running, queue_flag =
4374 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4375 const struct sched_class *prev_class;
4376 struct rq_flags rf;
4377 struct rq *rq;
4378
4379 /* XXX used to be waiter->prio, not waiter->task->prio */
4380 prio = __rt_effective_prio(pi_task, p->normal_prio);
4381
4382 /*
4383 * If nothing changed; bail early.
4384 */
4385 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4386 return;
4387
4388 rq = __task_rq_lock(p, &rf);
4389 update_rq_clock(rq);
4390 /*
4391 * Set under pi_lock && rq->lock, such that the value can be used under
4392 * either lock.
4393 *
4394 * Note that there is loads of tricky to make this pointer cache work
4395 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4396 * ensure a task is de-boosted (pi_task is set to NULL) before the
4397 * task is allowed to run again (and can exit). This ensures the pointer
4398 * points to a blocked task -- which guaratees the task is present.
4399 */
4400 p->pi_top_task = pi_task;
4401
4402 /*
4403 * For FIFO/RR we only need to set prio, if that matches we're done.
4404 */
4405 if (prio == p->prio && !dl_prio(prio))
4406 goto out_unlock;
4407
4408 /*
4409 * Idle task boosting is a nono in general. There is one
4410 * exception, when PREEMPT_RT and NOHZ is active:
4411 *
4412 * The idle task calls get_next_timer_interrupt() and holds
4413 * the timer wheel base->lock on the CPU and another CPU wants
4414 * to access the timer (probably to cancel it). We can safely
4415 * ignore the boosting request, as the idle CPU runs this code
4416 * with interrupts disabled and will complete the lock
4417 * protected section without being interrupted. So there is no
4418 * real need to boost.
4419 */
4420 if (unlikely(p == rq->idle)) {
4421 WARN_ON(p != rq->curr);
4422 WARN_ON(p->pi_blocked_on);
4423 goto out_unlock;
4424 }
4425
4426 trace_sched_pi_setprio(p, pi_task);
4427 oldprio = p->prio;
4428
4429 if (oldprio == prio)
4430 queue_flag &= ~DEQUEUE_MOVE;
4431
4432 prev_class = p->sched_class;
4433 queued = task_on_rq_queued(p);
4434 running = task_current(rq, p);
4435 if (queued)
4436 dequeue_task(rq, p, queue_flag);
4437 if (running)
4438 put_prev_task(rq, p);
4439
4440 /*
4441 * Boosting condition are:
4442 * 1. -rt task is running and holds mutex A
4443 * --> -dl task blocks on mutex A
4444 *
4445 * 2. -dl task is running and holds mutex A
4446 * --> -dl task blocks on mutex A and could preempt the
4447 * running task
4448 */
4449 if (dl_prio(prio)) {
4450 if (!dl_prio(p->normal_prio) ||
4451 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
4452 p->dl.dl_boosted = 1;
4453 queue_flag |= ENQUEUE_REPLENISH;
4454 } else
4455 p->dl.dl_boosted = 0;
4456 p->sched_class = &dl_sched_class;
4457 } else if (rt_prio(prio)) {
4458 if (dl_prio(oldprio))
4459 p->dl.dl_boosted = 0;
4460 if (oldprio < prio)
4461 queue_flag |= ENQUEUE_HEAD;
4462 p->sched_class = &rt_sched_class;
4463 } else {
4464 if (dl_prio(oldprio))
4465 p->dl.dl_boosted = 0;
4466 if (rt_prio(oldprio))
4467 p->rt.timeout = 0;
4468 p->sched_class = &fair_sched_class;
4469 }
4470
4471 p->prio = prio;
4472
4473 if (queued)
4474 enqueue_task(rq, p, queue_flag);
4475 if (running)
4476 set_next_task(rq, p);
4477
4478 check_class_changed(rq, p, prev_class, oldprio);
4479out_unlock:
4480 /* Avoid rq from going away on us: */
4481 preempt_disable();
4482 __task_rq_unlock(rq, &rf);
4483
4484 balance_callback(rq);
4485 preempt_enable();
4486}
4487#else
4488static inline int rt_effective_prio(struct task_struct *p, int prio)
4489{
4490 return prio;
4491}
4492#endif
4493
4494void set_user_nice(struct task_struct *p, long nice)
4495{
4496 bool queued, running;
4497 int old_prio, delta;
4498 struct rq_flags rf;
4499 struct rq *rq;
4500
4501 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
4502 return;
4503 /*
4504 * We have to be careful, if called from sys_setpriority(),
4505 * the task might be in the middle of scheduling on another CPU.
4506 */
4507 rq = task_rq_lock(p, &rf);
4508 update_rq_clock(rq);
4509
4510 /*
4511 * The RT priorities are set via sched_setscheduler(), but we still
4512 * allow the 'normal' nice value to be set - but as expected
4513 * it wont have any effect on scheduling until the task is
4514 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4515 */
4516 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4517 p->static_prio = NICE_TO_PRIO(nice);
4518 goto out_unlock;
4519 }
4520 queued = task_on_rq_queued(p);
4521 running = task_current(rq, p);
4522 if (queued)
4523 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
4524 if (running)
4525 put_prev_task(rq, p);
4526
4527 p->static_prio = NICE_TO_PRIO(nice);
4528 set_load_weight(p, true);
4529 old_prio = p->prio;
4530 p->prio = effective_prio(p);
4531 delta = p->prio - old_prio;
4532
4533 if (queued) {
4534 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
4535 /*
4536 * If the task increased its priority or is running and
4537 * lowered its priority, then reschedule its CPU:
4538 */
4539 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4540 resched_curr(rq);
4541 }
4542 if (running)
4543 set_next_task(rq, p);
4544out_unlock:
4545 task_rq_unlock(rq, p, &rf);
4546}
4547EXPORT_SYMBOL(set_user_nice);
4548
4549/*
4550 * can_nice - check if a task can reduce its nice value
4551 * @p: task
4552 * @nice: nice value
4553 */
4554int can_nice(const struct task_struct *p, const int nice)
4555{
4556 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4557 int nice_rlim = nice_to_rlimit(nice);
4558
4559 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4560 capable(CAP_SYS_NICE));
4561}
4562
4563#ifdef __ARCH_WANT_SYS_NICE
4564
4565/*
4566 * sys_nice - change the priority of the current process.
4567 * @increment: priority increment
4568 *
4569 * sys_setpriority is a more generic, but much slower function that
4570 * does similar things.
4571 */
4572SYSCALL_DEFINE1(nice, int, increment)
4573{
4574 long nice, retval;
4575
4576 /*
4577 * Setpriority might change our priority at the same moment.
4578 * We don't have to worry. Conceptually one call occurs first
4579 * and we have a single winner.
4580 */
4581 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
4582 nice = task_nice(current) + increment;
4583
4584 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
4585 if (increment < 0 && !can_nice(current, nice))
4586 return -EPERM;
4587
4588 retval = security_task_setnice(current, nice);
4589 if (retval)
4590 return retval;
4591
4592 set_user_nice(current, nice);
4593 return 0;
4594}
4595
4596#endif
4597
4598/**
4599 * task_prio - return the priority value of a given task.
4600 * @p: the task in question.
4601 *
4602 * Return: The priority value as seen by users in /proc.
4603 * RT tasks are offset by -200. Normal tasks are centered
4604 * around 0, value goes from -16 to +15.
4605 */
4606int task_prio(const struct task_struct *p)
4607{
4608 return p->prio - MAX_RT_PRIO;
4609}
4610
4611/**
4612 * idle_cpu - is a given CPU idle currently?
4613 * @cpu: the processor in question.
4614 *
4615 * Return: 1 if the CPU is currently idle. 0 otherwise.
4616 */
4617int idle_cpu(int cpu)
4618{
4619 struct rq *rq = cpu_rq(cpu);
4620
4621 if (rq->curr != rq->idle)
4622 return 0;
4623
4624 if (rq->nr_running)
4625 return 0;
4626
4627#ifdef CONFIG_SMP
4628 if (!llist_empty(&rq->wake_list))
4629 return 0;
4630#endif
4631
4632 return 1;
4633}
4634
4635/**
4636 * available_idle_cpu - is a given CPU idle for enqueuing work.
4637 * @cpu: the CPU in question.
4638 *
4639 * Return: 1 if the CPU is currently idle. 0 otherwise.
4640 */
4641int available_idle_cpu(int cpu)
4642{
4643 if (!idle_cpu(cpu))
4644 return 0;
4645
4646 if (vcpu_is_preempted(cpu))
4647 return 0;
4648
4649 return 1;
4650}
4651
4652/**
4653 * idle_task - return the idle task for a given CPU.
4654 * @cpu: the processor in question.
4655 *
4656 * Return: The idle task for the CPU @cpu.
4657 */
4658struct task_struct *idle_task(int cpu)
4659{
4660 return cpu_rq(cpu)->idle;
4661}
4662
4663/**
4664 * find_process_by_pid - find a process with a matching PID value.
4665 * @pid: the pid in question.
4666 *
4667 * The task of @pid, if found. %NULL otherwise.
4668 */
4669static struct task_struct *find_process_by_pid(pid_t pid)
4670{
4671 return pid ? find_task_by_vpid(pid) : current;
4672}
4673
4674/*
4675 * sched_setparam() passes in -1 for its policy, to let the functions
4676 * it calls know not to change it.
4677 */
4678#define SETPARAM_POLICY -1
4679
4680static void __setscheduler_params(struct task_struct *p,
4681 const struct sched_attr *attr)
4682{
4683 int policy = attr->sched_policy;
4684
4685 if (policy == SETPARAM_POLICY)
4686 policy = p->policy;
4687
4688 p->policy = policy;
4689
4690 if (dl_policy(policy))
4691 __setparam_dl(p, attr);
4692 else if (fair_policy(policy))
4693 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4694
4695 /*
4696 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4697 * !rt_policy. Always setting this ensures that things like
4698 * getparam()/getattr() don't report silly values for !rt tasks.
4699 */
4700 p->rt_priority = attr->sched_priority;
4701 p->normal_prio = normal_prio(p);
4702 set_load_weight(p, true);
4703}
4704
4705/* Actually do priority change: must hold pi & rq lock. */
4706static void __setscheduler(struct rq *rq, struct task_struct *p,
4707 const struct sched_attr *attr, bool keep_boost)
4708{
4709 /*
4710 * If params can't change scheduling class changes aren't allowed
4711 * either.
4712 */
4713 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
4714 return;
4715
4716 __setscheduler_params(p, attr);
4717
4718 /*
4719 * Keep a potential priority boosting if called from
4720 * sched_setscheduler().
4721 */
4722 p->prio = normal_prio(p);
4723 if (keep_boost)
4724 p->prio = rt_effective_prio(p, p->prio);
4725
4726 if (dl_prio(p->prio))
4727 p->sched_class = &dl_sched_class;
4728 else if (rt_prio(p->prio))
4729 p->sched_class = &rt_sched_class;
4730 else
4731 p->sched_class = &fair_sched_class;
4732}
4733
4734/*
4735 * Check the target process has a UID that matches the current process's:
4736 */
4737static bool check_same_owner(struct task_struct *p)
4738{
4739 const struct cred *cred = current_cred(), *pcred;
4740 bool match;
4741
4742 rcu_read_lock();
4743 pcred = __task_cred(p);
4744 match = (uid_eq(cred->euid, pcred->euid) ||
4745 uid_eq(cred->euid, pcred->uid));
4746 rcu_read_unlock();
4747 return match;
4748}
4749
4750static int __sched_setscheduler(struct task_struct *p,
4751 const struct sched_attr *attr,
4752 bool user, bool pi)
4753{
4754 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4755 MAX_RT_PRIO - 1 - attr->sched_priority;
4756 int retval, oldprio, oldpolicy = -1, queued, running;
4757 int new_effective_prio, policy = attr->sched_policy;
4758 const struct sched_class *prev_class;
4759 struct rq_flags rf;
4760 int reset_on_fork;
4761 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4762 struct rq *rq;
4763
4764 /* The pi code expects interrupts enabled */
4765 BUG_ON(pi && in_interrupt());
4766recheck:
4767 /* Double check policy once rq lock held: */
4768 if (policy < 0) {
4769 reset_on_fork = p->sched_reset_on_fork;
4770 policy = oldpolicy = p->policy;
4771 } else {
4772 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4773
4774 if (!valid_policy(policy))
4775 return -EINVAL;
4776 }
4777
4778 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4779 return -EINVAL;
4780
4781 /*
4782 * Valid priorities for SCHED_FIFO and SCHED_RR are
4783 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4784 * SCHED_BATCH and SCHED_IDLE is 0.
4785 */
4786 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4787 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4788 return -EINVAL;
4789 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4790 (rt_policy(policy) != (attr->sched_priority != 0)))
4791 return -EINVAL;
4792
4793 /*
4794 * Allow unprivileged RT tasks to decrease priority:
4795 */
4796 if (user && !capable(CAP_SYS_NICE)) {
4797 if (fair_policy(policy)) {
4798 if (attr->sched_nice < task_nice(p) &&
4799 !can_nice(p, attr->sched_nice))
4800 return -EPERM;
4801 }
4802
4803 if (rt_policy(policy)) {
4804 unsigned long rlim_rtprio =
4805 task_rlimit(p, RLIMIT_RTPRIO);
4806
4807 /* Can't set/change the rt policy: */
4808 if (policy != p->policy && !rlim_rtprio)
4809 return -EPERM;
4810
4811 /* Can't increase priority: */
4812 if (attr->sched_priority > p->rt_priority &&
4813 attr->sched_priority > rlim_rtprio)
4814 return -EPERM;
4815 }
4816
4817 /*
4818 * Can't set/change SCHED_DEADLINE policy at all for now
4819 * (safest behavior); in the future we would like to allow
4820 * unprivileged DL tasks to increase their relative deadline
4821 * or reduce their runtime (both ways reducing utilization)
4822 */
4823 if (dl_policy(policy))
4824 return -EPERM;
4825
4826 /*
4827 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4828 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4829 */
4830 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4831 if (!can_nice(p, task_nice(p)))
4832 return -EPERM;
4833 }
4834
4835 /* Can't change other user's priorities: */
4836 if (!check_same_owner(p))
4837 return -EPERM;
4838
4839 /* Normal users shall not reset the sched_reset_on_fork flag: */
4840 if (p->sched_reset_on_fork && !reset_on_fork)
4841 return -EPERM;
4842 }
4843
4844 if (user) {
4845 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4846 return -EINVAL;
4847
4848 retval = security_task_setscheduler(p);
4849 if (retval)
4850 return retval;
4851 }
4852
4853 /* Update task specific "requested" clamps */
4854 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
4855 retval = uclamp_validate(p, attr);
4856 if (retval)
4857 return retval;
4858 }
4859
4860 if (pi)
4861 cpuset_read_lock();
4862
4863 /*
4864 * Make sure no PI-waiters arrive (or leave) while we are
4865 * changing the priority of the task:
4866 *
4867 * To be able to change p->policy safely, the appropriate
4868 * runqueue lock must be held.
4869 */
4870 rq = task_rq_lock(p, &rf);
4871 update_rq_clock(rq);
4872
4873 /*
4874 * Changing the policy of the stop threads its a very bad idea:
4875 */
4876 if (p == rq->stop) {
4877 retval = -EINVAL;
4878 goto unlock;
4879 }
4880
4881 /*
4882 * If not changing anything there's no need to proceed further,
4883 * but store a possible modification of reset_on_fork.
4884 */
4885 if (unlikely(policy == p->policy)) {
4886 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4887 goto change;
4888 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4889 goto change;
4890 if (dl_policy(policy) && dl_param_changed(p, attr))
4891 goto change;
4892 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
4893 goto change;
4894
4895 p->sched_reset_on_fork = reset_on_fork;
4896 retval = 0;
4897 goto unlock;
4898 }
4899change:
4900
4901 if (user) {
4902#ifdef CONFIG_RT_GROUP_SCHED
4903 /*
4904 * Do not allow realtime tasks into groups that have no runtime
4905 * assigned.
4906 */
4907 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4908 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4909 !task_group_is_autogroup(task_group(p))) {
4910 retval = -EPERM;
4911 goto unlock;
4912 }
4913#endif
4914#ifdef CONFIG_SMP
4915 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4916 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4917 cpumask_t *span = rq->rd->span;
4918
4919 /*
4920 * Don't allow tasks with an affinity mask smaller than
4921 * the entire root_domain to become SCHED_DEADLINE. We
4922 * will also fail if there's no bandwidth available.
4923 */
4924 if (!cpumask_subset(span, p->cpus_ptr) ||
4925 rq->rd->dl_bw.bw == 0) {
4926 retval = -EPERM;
4927 goto unlock;
4928 }
4929 }
4930#endif
4931 }
4932
4933 /* Re-check policy now with rq lock held: */
4934 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4935 policy = oldpolicy = -1;
4936 task_rq_unlock(rq, p, &rf);
4937 if (pi)
4938 cpuset_read_unlock();
4939 goto recheck;
4940 }
4941
4942 /*
4943 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4944 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4945 * is available.
4946 */
4947 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4948 retval = -EBUSY;
4949 goto unlock;
4950 }
4951
4952 p->sched_reset_on_fork = reset_on_fork;
4953 oldprio = p->prio;
4954
4955 if (pi) {
4956 /*
4957 * Take priority boosted tasks into account. If the new
4958 * effective priority is unchanged, we just store the new
4959 * normal parameters and do not touch the scheduler class and
4960 * the runqueue. This will be done when the task deboost
4961 * itself.
4962 */
4963 new_effective_prio = rt_effective_prio(p, newprio);
4964 if (new_effective_prio == oldprio)
4965 queue_flags &= ~DEQUEUE_MOVE;
4966 }
4967
4968 queued = task_on_rq_queued(p);
4969 running = task_current(rq, p);
4970 if (queued)
4971 dequeue_task(rq, p, queue_flags);
4972 if (running)
4973 put_prev_task(rq, p);
4974
4975 prev_class = p->sched_class;
4976
4977 __setscheduler(rq, p, attr, pi);
4978 __setscheduler_uclamp(p, attr);
4979
4980 if (queued) {
4981 /*
4982 * We enqueue to tail when the priority of a task is
4983 * increased (user space view).
4984 */
4985 if (oldprio < p->prio)
4986 queue_flags |= ENQUEUE_HEAD;
4987
4988 enqueue_task(rq, p, queue_flags);
4989 }
4990 if (running)
4991 set_next_task(rq, p);
4992
4993 check_class_changed(rq, p, prev_class, oldprio);
4994
4995 /* Avoid rq from going away on us: */
4996 preempt_disable();
4997 task_rq_unlock(rq, p, &rf);
4998
4999 if (pi) {
5000 cpuset_read_unlock();
5001 rt_mutex_adjust_pi(p);
5002 }
5003
5004 /* Run balance callbacks after we've adjusted the PI chain: */
5005 balance_callback(rq);
5006 preempt_enable();
5007
5008 return 0;
5009
5010unlock:
5011 task_rq_unlock(rq, p, &rf);
5012 if (pi)
5013 cpuset_read_unlock();
5014 return retval;
5015}
5016
5017static int _sched_setscheduler(struct task_struct *p, int policy,
5018 const struct sched_param *param, bool check)
5019{
5020 struct sched_attr attr = {
5021 .sched_policy = policy,
5022 .sched_priority = param->sched_priority,
5023 .sched_nice = PRIO_TO_NICE(p->static_prio),
5024 };
5025
5026 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5027 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
5028 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5029 policy &= ~SCHED_RESET_ON_FORK;
5030 attr.sched_policy = policy;
5031 }
5032
5033 return __sched_setscheduler(p, &attr, check, true);
5034}
5035/**
5036 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5037 * @p: the task in question.
5038 * @policy: new policy.
5039 * @param: structure containing the new RT priority.
5040 *
5041 * Return: 0 on success. An error code otherwise.
5042 *
5043 * NOTE that the task may be already dead.
5044 */
5045int sched_setscheduler(struct task_struct *p, int policy,
5046 const struct sched_param *param)
5047{
5048 return _sched_setscheduler(p, policy, param, true);
5049}
5050EXPORT_SYMBOL_GPL(sched_setscheduler);
5051
5052int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
5053{
5054 return __sched_setscheduler(p, attr, true, true);
5055}
5056EXPORT_SYMBOL_GPL(sched_setattr);
5057
5058int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
5059{
5060 return __sched_setscheduler(p, attr, false, true);
5061}
5062
5063/**
5064 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5065 * @p: the task in question.
5066 * @policy: new policy.
5067 * @param: structure containing the new RT priority.
5068 *
5069 * Just like sched_setscheduler, only don't bother checking if the
5070 * current context has permission. For example, this is needed in
5071 * stop_machine(): we create temporary high priority worker threads,
5072 * but our caller might not have that capability.
5073 *
5074 * Return: 0 on success. An error code otherwise.
5075 */
5076int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5077 const struct sched_param *param)
5078{
5079 return _sched_setscheduler(p, policy, param, false);
5080}
5081EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
5082
5083static int
5084do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5085{
5086 struct sched_param lparam;
5087 struct task_struct *p;
5088 int retval;
5089
5090 if (!param || pid < 0)
5091 return -EINVAL;
5092 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5093 return -EFAULT;
5094
5095 rcu_read_lock();
5096 retval = -ESRCH;
5097 p = find_process_by_pid(pid);
5098 if (likely(p))
5099 get_task_struct(p);
5100 rcu_read_unlock();
5101
5102 if (likely(p)) {
5103 retval = sched_setscheduler(p, policy, &lparam);
5104 put_task_struct(p);
5105 }
5106
5107 return retval;
5108}
5109
5110/*
5111 * Mimics kernel/events/core.c perf_copy_attr().
5112 */
5113static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
5114{
5115 u32 size;
5116 int ret;
5117
5118 /* Zero the full structure, so that a short copy will be nice: */
5119 memset(attr, 0, sizeof(*attr));
5120
5121 ret = get_user(size, &uattr->size);
5122 if (ret)
5123 return ret;
5124
5125 /* ABI compatibility quirk: */
5126 if (!size)
5127 size = SCHED_ATTR_SIZE_VER0;
5128 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
5129 goto err_size;
5130
5131 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
5132 if (ret) {
5133 if (ret == -E2BIG)
5134 goto err_size;
5135 return ret;
5136 }
5137
5138 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
5139 size < SCHED_ATTR_SIZE_VER1)
5140 return -EINVAL;
5141
5142 /*
5143 * XXX: Do we want to be lenient like existing syscalls; or do we want
5144 * to be strict and return an error on out-of-bounds values?
5145 */
5146 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
5147
5148 return 0;
5149
5150err_size:
5151 put_user(sizeof(*attr), &uattr->size);
5152 return -E2BIG;
5153}
5154
5155/**
5156 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5157 * @pid: the pid in question.
5158 * @policy: new policy.
5159 * @param: structure containing the new RT priority.
5160 *
5161 * Return: 0 on success. An error code otherwise.
5162 */
5163SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
5164{
5165 if (policy < 0)
5166 return -EINVAL;
5167
5168 return do_sched_setscheduler(pid, policy, param);
5169}
5170
5171/**
5172 * sys_sched_setparam - set/change the RT priority of a thread
5173 * @pid: the pid in question.
5174 * @param: structure containing the new RT priority.
5175 *
5176 * Return: 0 on success. An error code otherwise.
5177 */
5178SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5179{
5180 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
5181}
5182
5183/**
5184 * sys_sched_setattr - same as above, but with extended sched_attr
5185 * @pid: the pid in question.
5186 * @uattr: structure containing the extended parameters.
5187 * @flags: for future extension.
5188 */
5189SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5190 unsigned int, flags)
5191{
5192 struct sched_attr attr;
5193 struct task_struct *p;
5194 int retval;
5195
5196 if (!uattr || pid < 0 || flags)
5197 return -EINVAL;
5198
5199 retval = sched_copy_attr(uattr, &attr);
5200 if (retval)
5201 return retval;
5202
5203 if ((int)attr.sched_policy < 0)
5204 return -EINVAL;
5205 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5206 attr.sched_policy = SETPARAM_POLICY;
5207
5208 rcu_read_lock();
5209 retval = -ESRCH;
5210 p = find_process_by_pid(pid);
5211 if (likely(p))
5212 get_task_struct(p);
5213 rcu_read_unlock();
5214
5215 if (likely(p)) {
5216 retval = sched_setattr(p, &attr);
5217 put_task_struct(p);
5218 }
5219
5220 return retval;
5221}
5222
5223/**
5224 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5225 * @pid: the pid in question.
5226 *
5227 * Return: On success, the policy of the thread. Otherwise, a negative error
5228 * code.
5229 */
5230SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5231{
5232 struct task_struct *p;
5233 int retval;
5234
5235 if (pid < 0)
5236 return -EINVAL;
5237
5238 retval = -ESRCH;
5239 rcu_read_lock();
5240 p = find_process_by_pid(pid);
5241 if (p) {
5242 retval = security_task_getscheduler(p);
5243 if (!retval)
5244 retval = p->policy
5245 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5246 }
5247 rcu_read_unlock();
5248 return retval;
5249}
5250
5251/**
5252 * sys_sched_getparam - get the RT priority of a thread
5253 * @pid: the pid in question.
5254 * @param: structure containing the RT priority.
5255 *
5256 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5257 * code.
5258 */
5259SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5260{
5261 struct sched_param lp = { .sched_priority = 0 };
5262 struct task_struct *p;
5263 int retval;
5264
5265 if (!param || pid < 0)
5266 return -EINVAL;
5267
5268 rcu_read_lock();
5269 p = find_process_by_pid(pid);
5270 retval = -ESRCH;
5271 if (!p)
5272 goto out_unlock;
5273
5274 retval = security_task_getscheduler(p);
5275 if (retval)
5276 goto out_unlock;
5277
5278 if (task_has_rt_policy(p))
5279 lp.sched_priority = p->rt_priority;
5280 rcu_read_unlock();
5281
5282 /*
5283 * This one might sleep, we cannot do it with a spinlock held ...
5284 */
5285 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5286
5287 return retval;
5288
5289out_unlock:
5290 rcu_read_unlock();
5291 return retval;
5292}
5293
5294/*
5295 * Copy the kernel size attribute structure (which might be larger
5296 * than what user-space knows about) to user-space.
5297 *
5298 * Note that all cases are valid: user-space buffer can be larger or
5299 * smaller than the kernel-space buffer. The usual case is that both
5300 * have the same size.
5301 */
5302static int
5303sched_attr_copy_to_user(struct sched_attr __user *uattr,
5304 struct sched_attr *kattr,
5305 unsigned int usize)
5306{
5307 unsigned int ksize = sizeof(*kattr);
5308
5309 if (!access_ok(uattr, usize))
5310 return -EFAULT;
5311
5312 /*
5313 * sched_getattr() ABI forwards and backwards compatibility:
5314 *
5315 * If usize == ksize then we just copy everything to user-space and all is good.
5316 *
5317 * If usize < ksize then we only copy as much as user-space has space for,
5318 * this keeps ABI compatibility as well. We skip the rest.
5319 *
5320 * If usize > ksize then user-space is using a newer version of the ABI,
5321 * which part the kernel doesn't know about. Just ignore it - tooling can
5322 * detect the kernel's knowledge of attributes from the attr->size value
5323 * which is set to ksize in this case.
5324 */
5325 kattr->size = min(usize, ksize);
5326
5327 if (copy_to_user(uattr, kattr, kattr->size))
5328 return -EFAULT;
5329
5330 return 0;
5331}
5332
5333/**
5334 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5335 * @pid: the pid in question.
5336 * @uattr: structure containing the extended parameters.
5337 * @usize: sizeof(attr) for fwd/bwd comp.
5338 * @flags: for future extension.
5339 */
5340SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5341 unsigned int, usize, unsigned int, flags)
5342{
5343 struct sched_attr kattr = { };
5344 struct task_struct *p;
5345 int retval;
5346
5347 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
5348 usize < SCHED_ATTR_SIZE_VER0 || flags)
5349 return -EINVAL;
5350
5351 rcu_read_lock();
5352 p = find_process_by_pid(pid);
5353 retval = -ESRCH;
5354 if (!p)
5355 goto out_unlock;
5356
5357 retval = security_task_getscheduler(p);
5358 if (retval)
5359 goto out_unlock;
5360
5361 kattr.sched_policy = p->policy;
5362 if (p->sched_reset_on_fork)
5363 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5364 if (task_has_dl_policy(p))
5365 __getparam_dl(p, &kattr);
5366 else if (task_has_rt_policy(p))
5367 kattr.sched_priority = p->rt_priority;
5368 else
5369 kattr.sched_nice = task_nice(p);
5370
5371#ifdef CONFIG_UCLAMP_TASK
5372 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5373 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5374#endif
5375
5376 rcu_read_unlock();
5377
5378 return sched_attr_copy_to_user(uattr, &kattr, usize);
5379
5380out_unlock:
5381 rcu_read_unlock();
5382 return retval;
5383}
5384
5385long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5386{
5387 cpumask_var_t cpus_allowed, new_mask;
5388 struct task_struct *p;
5389 int retval;
5390
5391 rcu_read_lock();
5392
5393 p = find_process_by_pid(pid);
5394 if (!p) {
5395 rcu_read_unlock();
5396 return -ESRCH;
5397 }
5398
5399 /* Prevent p going away */
5400 get_task_struct(p);
5401 rcu_read_unlock();
5402
5403 if (p->flags & PF_NO_SETAFFINITY) {
5404 retval = -EINVAL;
5405 goto out_put_task;
5406 }
5407 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5408 retval = -ENOMEM;
5409 goto out_put_task;
5410 }
5411 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5412 retval = -ENOMEM;
5413 goto out_free_cpus_allowed;
5414 }
5415 retval = -EPERM;
5416 if (!check_same_owner(p)) {
5417 rcu_read_lock();
5418 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5419 rcu_read_unlock();
5420 goto out_free_new_mask;
5421 }
5422 rcu_read_unlock();
5423 }
5424
5425 retval = security_task_setscheduler(p);
5426 if (retval)
5427 goto out_free_new_mask;
5428
5429
5430 cpuset_cpus_allowed(p, cpus_allowed);
5431 cpumask_and(new_mask, in_mask, cpus_allowed);
5432
5433 /*
5434 * Since bandwidth control happens on root_domain basis,
5435 * if admission test is enabled, we only admit -deadline
5436 * tasks allowed to run on all the CPUs in the task's
5437 * root_domain.
5438 */
5439#ifdef CONFIG_SMP
5440 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
5441 rcu_read_lock();
5442 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
5443 retval = -EBUSY;
5444 rcu_read_unlock();
5445 goto out_free_new_mask;
5446 }
5447 rcu_read_unlock();
5448 }
5449#endif
5450again:
5451 retval = __set_cpus_allowed_ptr(p, new_mask, true);
5452
5453 if (!retval) {
5454 cpuset_cpus_allowed(p, cpus_allowed);
5455 if (!cpumask_subset(new_mask, cpus_allowed)) {
5456 /*
5457 * We must have raced with a concurrent cpuset
5458 * update. Just reset the cpus_allowed to the
5459 * cpuset's cpus_allowed
5460 */
5461 cpumask_copy(new_mask, cpus_allowed);
5462 goto again;
5463 }
5464 }
5465out_free_new_mask:
5466 free_cpumask_var(new_mask);
5467out_free_cpus_allowed:
5468 free_cpumask_var(cpus_allowed);
5469out_put_task:
5470 put_task_struct(p);
5471 return retval;
5472}
5473
5474static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5475 struct cpumask *new_mask)
5476{
5477 if (len < cpumask_size())
5478 cpumask_clear(new_mask);
5479 else if (len > cpumask_size())
5480 len = cpumask_size();
5481
5482 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5483}
5484
5485/**
5486 * sys_sched_setaffinity - set the CPU affinity of a process
5487 * @pid: pid of the process
5488 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5489 * @user_mask_ptr: user-space pointer to the new CPU mask
5490 *
5491 * Return: 0 on success. An error code otherwise.
5492 */
5493SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5494 unsigned long __user *, user_mask_ptr)
5495{
5496 cpumask_var_t new_mask;
5497 int retval;
5498
5499 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5500 return -ENOMEM;
5501
5502 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5503 if (retval == 0)
5504 retval = sched_setaffinity(pid, new_mask);
5505 free_cpumask_var(new_mask);
5506 return retval;
5507}
5508
5509long sched_getaffinity(pid_t pid, struct cpumask *mask)
5510{
5511 struct task_struct *p;
5512 unsigned long flags;
5513 int retval;
5514
5515 rcu_read_lock();
5516
5517 retval = -ESRCH;
5518 p = find_process_by_pid(pid);
5519 if (!p)
5520 goto out_unlock;
5521
5522 retval = security_task_getscheduler(p);
5523 if (retval)
5524 goto out_unlock;
5525
5526 raw_spin_lock_irqsave(&p->pi_lock, flags);
5527 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
5528 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5529
5530out_unlock:
5531 rcu_read_unlock();
5532
5533 return retval;
5534}
5535
5536/**
5537 * sys_sched_getaffinity - get the CPU affinity of a process
5538 * @pid: pid of the process
5539 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5540 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5541 *
5542 * Return: size of CPU mask copied to user_mask_ptr on success. An
5543 * error code otherwise.
5544 */
5545SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5546 unsigned long __user *, user_mask_ptr)
5547{
5548 int ret;
5549 cpumask_var_t mask;
5550
5551 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5552 return -EINVAL;
5553 if (len & (sizeof(unsigned long)-1))
5554 return -EINVAL;
5555
5556 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5557 return -ENOMEM;
5558
5559 ret = sched_getaffinity(pid, mask);
5560 if (ret == 0) {
5561 unsigned int retlen = min(len, cpumask_size());
5562
5563 if (copy_to_user(user_mask_ptr, mask, retlen))
5564 ret = -EFAULT;
5565 else
5566 ret = retlen;
5567 }
5568 free_cpumask_var(mask);
5569
5570 return ret;
5571}
5572
5573/**
5574 * sys_sched_yield - yield the current processor to other threads.
5575 *
5576 * This function yields the current CPU to other tasks. If there are no
5577 * other threads running on this CPU then this function will return.
5578 *
5579 * Return: 0.
5580 */
5581static void do_sched_yield(void)
5582{
5583 struct rq_flags rf;
5584 struct rq *rq;
5585
5586 rq = this_rq_lock_irq(&rf);
5587
5588 schedstat_inc(rq->yld_count);
5589 current->sched_class->yield_task(rq);
5590
5591 /*
5592 * Since we are going to call schedule() anyway, there's
5593 * no need to preempt or enable interrupts:
5594 */
5595 preempt_disable();
5596 rq_unlock(rq, &rf);
5597 sched_preempt_enable_no_resched();
5598
5599 schedule();
5600}
5601
5602SYSCALL_DEFINE0(sched_yield)
5603{
5604 do_sched_yield();
5605 return 0;
5606}
5607
5608#ifndef CONFIG_PREEMPTION
5609int __sched _cond_resched(void)
5610{
5611 if (should_resched(0)) {
5612 preempt_schedule_common();
5613 return 1;
5614 }
5615 rcu_all_qs();
5616 return 0;
5617}
5618EXPORT_SYMBOL(_cond_resched);
5619#endif
5620
5621/*
5622 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5623 * call schedule, and on return reacquire the lock.
5624 *
5625 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5626 * operations here to prevent schedule() from being called twice (once via
5627 * spin_unlock(), once by hand).
5628 */
5629int __cond_resched_lock(spinlock_t *lock)
5630{
5631 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5632 int ret = 0;
5633
5634 lockdep_assert_held(lock);
5635
5636 if (spin_needbreak(lock) || resched) {
5637 spin_unlock(lock);
5638 if (resched)
5639 preempt_schedule_common();
5640 else
5641 cpu_relax();
5642 ret = 1;
5643 spin_lock(lock);
5644 }
5645 return ret;
5646}
5647EXPORT_SYMBOL(__cond_resched_lock);
5648
5649/**
5650 * yield - yield the current processor to other threads.
5651 *
5652 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5653 *
5654 * The scheduler is at all times free to pick the calling task as the most
5655 * eligible task to run, if removing the yield() call from your code breaks
5656 * it, its already broken.
5657 *
5658 * Typical broken usage is:
5659 *
5660 * while (!event)
5661 * yield();
5662 *
5663 * where one assumes that yield() will let 'the other' process run that will
5664 * make event true. If the current task is a SCHED_FIFO task that will never
5665 * happen. Never use yield() as a progress guarantee!!
5666 *
5667 * If you want to use yield() to wait for something, use wait_event().
5668 * If you want to use yield() to be 'nice' for others, use cond_resched().
5669 * If you still want to use yield(), do not!
5670 */
5671void __sched yield(void)
5672{
5673 set_current_state(TASK_RUNNING);
5674 do_sched_yield();
5675}
5676EXPORT_SYMBOL(yield);
5677
5678/**
5679 * yield_to - yield the current processor to another thread in
5680 * your thread group, or accelerate that thread toward the
5681 * processor it's on.
5682 * @p: target task
5683 * @preempt: whether task preemption is allowed or not
5684 *
5685 * It's the caller's job to ensure that the target task struct
5686 * can't go away on us before we can do any checks.
5687 *
5688 * Return:
5689 * true (>0) if we indeed boosted the target task.
5690 * false (0) if we failed to boost the target.
5691 * -ESRCH if there's no task to yield to.
5692 */
5693int __sched yield_to(struct task_struct *p, bool preempt)
5694{
5695 struct task_struct *curr = current;
5696 struct rq *rq, *p_rq;
5697 unsigned long flags;
5698 int yielded = 0;
5699
5700 local_irq_save(flags);
5701 rq = this_rq();
5702
5703again:
5704 p_rq = task_rq(p);
5705 /*
5706 * If we're the only runnable task on the rq and target rq also
5707 * has only one task, there's absolutely no point in yielding.
5708 */
5709 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5710 yielded = -ESRCH;
5711 goto out_irq;
5712 }
5713
5714 double_rq_lock(rq, p_rq);
5715 if (task_rq(p) != p_rq) {
5716 double_rq_unlock(rq, p_rq);
5717 goto again;
5718 }
5719
5720 if (!curr->sched_class->yield_to_task)
5721 goto out_unlock;
5722
5723 if (curr->sched_class != p->sched_class)
5724 goto out_unlock;
5725
5726 if (task_running(p_rq, p) || p->state)
5727 goto out_unlock;
5728
5729 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5730 if (yielded) {
5731 schedstat_inc(rq->yld_count);
5732 /*
5733 * Make p's CPU reschedule; pick_next_entity takes care of
5734 * fairness.
5735 */
5736 if (preempt && rq != p_rq)
5737 resched_curr(p_rq);
5738 }
5739
5740out_unlock:
5741 double_rq_unlock(rq, p_rq);
5742out_irq:
5743 local_irq_restore(flags);
5744
5745 if (yielded > 0)
5746 schedule();
5747
5748 return yielded;
5749}
5750EXPORT_SYMBOL_GPL(yield_to);
5751
5752int io_schedule_prepare(void)
5753{
5754 int old_iowait = current->in_iowait;
5755
5756 current->in_iowait = 1;
5757 blk_schedule_flush_plug(current);
5758
5759 return old_iowait;
5760}
5761
5762void io_schedule_finish(int token)
5763{
5764 current->in_iowait = token;
5765}
5766
5767/*
5768 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5769 * that process accounting knows that this is a task in IO wait state.
5770 */
5771long __sched io_schedule_timeout(long timeout)
5772{
5773 int token;
5774 long ret;
5775
5776 token = io_schedule_prepare();
5777 ret = schedule_timeout(timeout);
5778 io_schedule_finish(token);
5779
5780 return ret;
5781}
5782EXPORT_SYMBOL(io_schedule_timeout);
5783
5784void __sched io_schedule(void)
5785{
5786 int token;
5787
5788 token = io_schedule_prepare();
5789 schedule();
5790 io_schedule_finish(token);
5791}
5792EXPORT_SYMBOL(io_schedule);
5793
5794/**
5795 * sys_sched_get_priority_max - return maximum RT priority.
5796 * @policy: scheduling class.
5797 *
5798 * Return: On success, this syscall returns the maximum
5799 * rt_priority that can be used by a given scheduling class.
5800 * On failure, a negative error code is returned.
5801 */
5802SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5803{
5804 int ret = -EINVAL;
5805
5806 switch (policy) {
5807 case SCHED_FIFO:
5808 case SCHED_RR:
5809 ret = MAX_USER_RT_PRIO-1;
5810 break;
5811 case SCHED_DEADLINE:
5812 case SCHED_NORMAL:
5813 case SCHED_BATCH:
5814 case SCHED_IDLE:
5815 ret = 0;
5816 break;
5817 }
5818 return ret;
5819}
5820
5821/**
5822 * sys_sched_get_priority_min - return minimum RT priority.
5823 * @policy: scheduling class.
5824 *
5825 * Return: On success, this syscall returns the minimum
5826 * rt_priority that can be used by a given scheduling class.
5827 * On failure, a negative error code is returned.
5828 */
5829SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5830{
5831 int ret = -EINVAL;
5832
5833 switch (policy) {
5834 case SCHED_FIFO:
5835 case SCHED_RR:
5836 ret = 1;
5837 break;
5838 case SCHED_DEADLINE:
5839 case SCHED_NORMAL:
5840 case SCHED_BATCH:
5841 case SCHED_IDLE:
5842 ret = 0;
5843 }
5844 return ret;
5845}
5846
5847static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5848{
5849 struct task_struct *p;
5850 unsigned int time_slice;
5851 struct rq_flags rf;
5852 struct rq *rq;
5853 int retval;
5854
5855 if (pid < 0)
5856 return -EINVAL;
5857
5858 retval = -ESRCH;
5859 rcu_read_lock();
5860 p = find_process_by_pid(pid);
5861 if (!p)
5862 goto out_unlock;
5863
5864 retval = security_task_getscheduler(p);
5865 if (retval)
5866 goto out_unlock;
5867
5868 rq = task_rq_lock(p, &rf);
5869 time_slice = 0;
5870 if (p->sched_class->get_rr_interval)
5871 time_slice = p->sched_class->get_rr_interval(rq, p);
5872 task_rq_unlock(rq, p, &rf);
5873
5874 rcu_read_unlock();
5875 jiffies_to_timespec64(time_slice, t);
5876 return 0;
5877
5878out_unlock:
5879 rcu_read_unlock();
5880 return retval;
5881}
5882
5883/**
5884 * sys_sched_rr_get_interval - return the default timeslice of a process.
5885 * @pid: pid of the process.
5886 * @interval: userspace pointer to the timeslice value.
5887 *
5888 * this syscall writes the default timeslice value of a given process
5889 * into the user-space timespec buffer. A value of '0' means infinity.
5890 *
5891 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5892 * an error code.
5893 */
5894SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5895 struct __kernel_timespec __user *, interval)
5896{
5897 struct timespec64 t;
5898 int retval = sched_rr_get_interval(pid, &t);
5899
5900 if (retval == 0)
5901 retval = put_timespec64(&t, interval);
5902
5903 return retval;
5904}
5905
5906#ifdef CONFIG_COMPAT_32BIT_TIME
5907SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
5908 struct old_timespec32 __user *, interval)
5909{
5910 struct timespec64 t;
5911 int retval = sched_rr_get_interval(pid, &t);
5912
5913 if (retval == 0)
5914 retval = put_old_timespec32(&t, interval);
5915 return retval;
5916}
5917#endif
5918
5919void sched_show_task(struct task_struct *p)
5920{
5921 unsigned long free = 0;
5922 int ppid;
5923
5924 if (!try_get_task_stack(p))
5925 return;
5926
5927 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5928
5929 if (p->state == TASK_RUNNING)
5930 printk(KERN_CONT " running task ");
5931#ifdef CONFIG_DEBUG_STACK_USAGE
5932 free = stack_not_used(p);
5933#endif
5934 ppid = 0;
5935 rcu_read_lock();
5936 if (pid_alive(p))
5937 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5938 rcu_read_unlock();
5939 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5940 task_pid_nr(p), ppid,
5941 (unsigned long)task_thread_info(p)->flags);
5942
5943 print_worker_info(KERN_INFO, p);
5944 show_stack(p, NULL);
5945 put_task_stack(p);
5946}
5947EXPORT_SYMBOL_GPL(sched_show_task);
5948
5949static inline bool
5950state_filter_match(unsigned long state_filter, struct task_struct *p)
5951{
5952 /* no filter, everything matches */
5953 if (!state_filter)
5954 return true;
5955
5956 /* filter, but doesn't match */
5957 if (!(p->state & state_filter))
5958 return false;
5959
5960 /*
5961 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5962 * TASK_KILLABLE).
5963 */
5964 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5965 return false;
5966
5967 return true;
5968}
5969
5970
5971void show_state_filter(unsigned long state_filter)
5972{
5973 struct task_struct *g, *p;
5974
5975#if BITS_PER_LONG == 32
5976 printk(KERN_INFO
5977 " task PC stack pid father\n");
5978#else
5979 printk(KERN_INFO
5980 " task PC stack pid father\n");
5981#endif
5982 rcu_read_lock();
5983 for_each_process_thread(g, p) {
5984 /*
5985 * reset the NMI-timeout, listing all files on a slow
5986 * console might take a lot of time:
5987 * Also, reset softlockup watchdogs on all CPUs, because
5988 * another CPU might be blocked waiting for us to process
5989 * an IPI.
5990 */
5991 touch_nmi_watchdog();
5992 touch_all_softlockup_watchdogs();
5993 if (state_filter_match(state_filter, p))
5994 sched_show_task(p);
5995 }
5996
5997#ifdef CONFIG_SCHED_DEBUG
5998 if (!state_filter)
5999 sysrq_sched_debug_show();
6000#endif
6001 rcu_read_unlock();
6002 /*
6003 * Only show locks if all tasks are dumped:
6004 */
6005 if (!state_filter)
6006 debug_show_all_locks();
6007}
6008
6009/**
6010 * init_idle - set up an idle thread for a given CPU
6011 * @idle: task in question
6012 * @cpu: CPU the idle task belongs to
6013 *
6014 * NOTE: this function does not set the idle thread's NEED_RESCHED
6015 * flag, to make booting more robust.
6016 */
6017void init_idle(struct task_struct *idle, int cpu)
6018{
6019 struct rq *rq = cpu_rq(cpu);
6020 unsigned long flags;
6021
6022 __sched_fork(0, idle);
6023
6024 raw_spin_lock_irqsave(&idle->pi_lock, flags);
6025 raw_spin_lock(&rq->lock);
6026
6027 idle->state = TASK_RUNNING;
6028 idle->se.exec_start = sched_clock();
6029 idle->flags |= PF_IDLE;
6030
6031 kasan_unpoison_task_stack(idle);
6032
6033#ifdef CONFIG_SMP
6034 /*
6035 * Its possible that init_idle() gets called multiple times on a task,
6036 * in that case do_set_cpus_allowed() will not do the right thing.
6037 *
6038 * And since this is boot we can forgo the serialization.
6039 */
6040 set_cpus_allowed_common(idle, cpumask_of(cpu));
6041#endif
6042 /*
6043 * We're having a chicken and egg problem, even though we are
6044 * holding rq->lock, the CPU isn't yet set to this CPU so the
6045 * lockdep check in task_group() will fail.
6046 *
6047 * Similar case to sched_fork(). / Alternatively we could
6048 * use task_rq_lock() here and obtain the other rq->lock.
6049 *
6050 * Silence PROVE_RCU
6051 */
6052 rcu_read_lock();
6053 __set_task_cpu(idle, cpu);
6054 rcu_read_unlock();
6055
6056 rq->idle = idle;
6057 rcu_assign_pointer(rq->curr, idle);
6058 idle->on_rq = TASK_ON_RQ_QUEUED;
6059#ifdef CONFIG_SMP
6060 idle->on_cpu = 1;
6061#endif
6062 raw_spin_unlock(&rq->lock);
6063 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
6064
6065 /* Set the preempt count _outside_ the spinlocks! */
6066 init_idle_preempt_count(idle, cpu);
6067
6068 /*
6069 * The idle tasks have their own, simple scheduling class:
6070 */
6071 idle->sched_class = &idle_sched_class;
6072 ftrace_graph_init_idle_task(idle, cpu);
6073 vtime_init_idle(idle, cpu);
6074#ifdef CONFIG_SMP
6075 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6076#endif
6077}
6078
6079#ifdef CONFIG_SMP
6080
6081int cpuset_cpumask_can_shrink(const struct cpumask *cur,
6082 const struct cpumask *trial)
6083{
6084 int ret = 1;
6085
6086 if (!cpumask_weight(cur))
6087 return ret;
6088
6089 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
6090
6091 return ret;
6092}
6093
6094int task_can_attach(struct task_struct *p,
6095 const struct cpumask *cs_cpus_allowed)
6096{
6097 int ret = 0;
6098
6099 /*
6100 * Kthreads which disallow setaffinity shouldn't be moved
6101 * to a new cpuset; we don't want to change their CPU
6102 * affinity and isolating such threads by their set of
6103 * allowed nodes is unnecessary. Thus, cpusets are not
6104 * applicable for such threads. This prevents checking for
6105 * success of set_cpus_allowed_ptr() on all attached tasks
6106 * before cpus_mask may be changed.
6107 */
6108 if (p->flags & PF_NO_SETAFFINITY) {
6109 ret = -EINVAL;
6110 goto out;
6111 }
6112
6113 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
6114 cs_cpus_allowed))
6115 ret = dl_task_can_attach(p, cs_cpus_allowed);
6116
6117out:
6118 return ret;
6119}
6120
6121bool sched_smp_initialized __read_mostly;
6122
6123#ifdef CONFIG_NUMA_BALANCING
6124/* Migrate current task p to target_cpu */
6125int migrate_task_to(struct task_struct *p, int target_cpu)
6126{
6127 struct migration_arg arg = { p, target_cpu };
6128 int curr_cpu = task_cpu(p);
6129
6130 if (curr_cpu == target_cpu)
6131 return 0;
6132
6133 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
6134 return -EINVAL;
6135
6136 /* TODO: This is not properly updating schedstats */
6137
6138 trace_sched_move_numa(p, curr_cpu, target_cpu);
6139 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
6140}
6141
6142/*
6143 * Requeue a task on a given node and accurately track the number of NUMA
6144 * tasks on the runqueues
6145 */
6146void sched_setnuma(struct task_struct *p, int nid)
6147{
6148 bool queued, running;
6149 struct rq_flags rf;
6150 struct rq *rq;
6151
6152 rq = task_rq_lock(p, &rf);
6153 queued = task_on_rq_queued(p);
6154 running = task_current(rq, p);
6155
6156 if (queued)
6157 dequeue_task(rq, p, DEQUEUE_SAVE);
6158 if (running)
6159 put_prev_task(rq, p);
6160
6161 p->numa_preferred_nid = nid;
6162
6163 if (queued)
6164 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6165 if (running)
6166 set_next_task(rq, p);
6167 task_rq_unlock(rq, p, &rf);
6168}
6169#endif /* CONFIG_NUMA_BALANCING */
6170
6171#ifdef CONFIG_HOTPLUG_CPU
6172/*
6173 * Ensure that the idle task is using init_mm right before its CPU goes
6174 * offline.
6175 */
6176void idle_task_exit(void)
6177{
6178 struct mm_struct *mm = current->active_mm;
6179
6180 BUG_ON(cpu_online(smp_processor_id()));
6181
6182 if (mm != &init_mm) {
6183 switch_mm(mm, &init_mm, current);
6184 current->active_mm = &init_mm;
6185 finish_arch_post_lock_switch();
6186 }
6187 mmdrop(mm);
6188}
6189
6190/*
6191 * Since this CPU is going 'away' for a while, fold any nr_active delta
6192 * we might have. Assumes we're called after migrate_tasks() so that the
6193 * nr_active count is stable. We need to take the teardown thread which
6194 * is calling this into account, so we hand in adjust = 1 to the load
6195 * calculation.
6196 *
6197 * Also see the comment "Global load-average calculations".
6198 */
6199static void calc_load_migrate(struct rq *rq)
6200{
6201 long delta = calc_load_fold_active(rq, 1);
6202 if (delta)
6203 atomic_long_add(delta, &calc_load_tasks);
6204}
6205
6206static struct task_struct *__pick_migrate_task(struct rq *rq)
6207{
6208 const struct sched_class *class;
6209 struct task_struct *next;
6210
6211 for_each_class(class) {
6212 next = class->pick_next_task(rq, NULL, NULL);
6213 if (next) {
6214 next->sched_class->put_prev_task(rq, next);
6215 return next;
6216 }
6217 }
6218
6219 /* The idle class should always have a runnable task */
6220 BUG();
6221}
6222
6223/*
6224 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6225 * try_to_wake_up()->select_task_rq().
6226 *
6227 * Called with rq->lock held even though we'er in stop_machine() and
6228 * there's no concurrency possible, we hold the required locks anyway
6229 * because of lock validation efforts.
6230 */
6231static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
6232{
6233 struct rq *rq = dead_rq;
6234 struct task_struct *next, *stop = rq->stop;
6235 struct rq_flags orf = *rf;
6236 int dest_cpu;
6237
6238 /*
6239 * Fudge the rq selection such that the below task selection loop
6240 * doesn't get stuck on the currently eligible stop task.
6241 *
6242 * We're currently inside stop_machine() and the rq is either stuck
6243 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6244 * either way we should never end up calling schedule() until we're
6245 * done here.
6246 */
6247 rq->stop = NULL;
6248
6249 /*
6250 * put_prev_task() and pick_next_task() sched
6251 * class method both need to have an up-to-date
6252 * value of rq->clock[_task]
6253 */
6254 update_rq_clock(rq);
6255
6256 for (;;) {
6257 /*
6258 * There's this thread running, bail when that's the only
6259 * remaining thread:
6260 */
6261 if (rq->nr_running == 1)
6262 break;
6263
6264 next = __pick_migrate_task(rq);
6265
6266 /*
6267 * Rules for changing task_struct::cpus_mask are holding
6268 * both pi_lock and rq->lock, such that holding either
6269 * stabilizes the mask.
6270 *
6271 * Drop rq->lock is not quite as disastrous as it usually is
6272 * because !cpu_active at this point, which means load-balance
6273 * will not interfere. Also, stop-machine.
6274 */
6275 rq_unlock(rq, rf);
6276 raw_spin_lock(&next->pi_lock);
6277 rq_relock(rq, rf);
6278
6279 /*
6280 * Since we're inside stop-machine, _nothing_ should have
6281 * changed the task, WARN if weird stuff happened, because in
6282 * that case the above rq->lock drop is a fail too.
6283 */
6284 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
6285 raw_spin_unlock(&next->pi_lock);
6286 continue;
6287 }
6288
6289 /* Find suitable destination for @next, with force if needed. */
6290 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
6291 rq = __migrate_task(rq, rf, next, dest_cpu);
6292 if (rq != dead_rq) {
6293 rq_unlock(rq, rf);
6294 rq = dead_rq;
6295 *rf = orf;
6296 rq_relock(rq, rf);
6297 }
6298 raw_spin_unlock(&next->pi_lock);
6299 }
6300
6301 rq->stop = stop;
6302}
6303#endif /* CONFIG_HOTPLUG_CPU */
6304
6305void set_rq_online(struct rq *rq)
6306{
6307 if (!rq->online) {
6308 const struct sched_class *class;
6309
6310 cpumask_set_cpu(rq->cpu, rq->rd->online);
6311 rq->online = 1;
6312
6313 for_each_class(class) {
6314 if (class->rq_online)
6315 class->rq_online(rq);
6316 }
6317 }
6318}
6319
6320void set_rq_offline(struct rq *rq)
6321{
6322 if (rq->online) {
6323 const struct sched_class *class;
6324
6325 for_each_class(class) {
6326 if (class->rq_offline)
6327 class->rq_offline(rq);
6328 }
6329
6330 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6331 rq->online = 0;
6332 }
6333}
6334
6335/*
6336 * used to mark begin/end of suspend/resume:
6337 */
6338static int num_cpus_frozen;
6339
6340/*
6341 * Update cpusets according to cpu_active mask. If cpusets are
6342 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6343 * around partition_sched_domains().
6344 *
6345 * If we come here as part of a suspend/resume, don't touch cpusets because we
6346 * want to restore it back to its original state upon resume anyway.
6347 */
6348static void cpuset_cpu_active(void)
6349{
6350 if (cpuhp_tasks_frozen) {
6351 /*
6352 * num_cpus_frozen tracks how many CPUs are involved in suspend
6353 * resume sequence. As long as this is not the last online
6354 * operation in the resume sequence, just build a single sched
6355 * domain, ignoring cpusets.
6356 */
6357 partition_sched_domains(1, NULL, NULL);
6358 if (--num_cpus_frozen)
6359 return;
6360 /*
6361 * This is the last CPU online operation. So fall through and
6362 * restore the original sched domains by considering the
6363 * cpuset configurations.
6364 */
6365 cpuset_force_rebuild();
6366 }
6367 cpuset_update_active_cpus();
6368}
6369
6370static int cpuset_cpu_inactive(unsigned int cpu)
6371{
6372 if (!cpuhp_tasks_frozen) {
6373 if (dl_cpu_busy(cpu))
6374 return -EBUSY;
6375 cpuset_update_active_cpus();
6376 } else {
6377 num_cpus_frozen++;
6378 partition_sched_domains(1, NULL, NULL);
6379 }
6380 return 0;
6381}
6382
6383int sched_cpu_activate(unsigned int cpu)
6384{
6385 struct rq *rq = cpu_rq(cpu);
6386 struct rq_flags rf;
6387
6388#ifdef CONFIG_SCHED_SMT
6389 /*
6390 * When going up, increment the number of cores with SMT present.
6391 */
6392 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6393 static_branch_inc_cpuslocked(&sched_smt_present);
6394#endif
6395 set_cpu_active(cpu, true);
6396
6397 if (sched_smp_initialized) {
6398 sched_domains_numa_masks_set(cpu);
6399 cpuset_cpu_active();
6400 }
6401
6402 /*
6403 * Put the rq online, if not already. This happens:
6404 *
6405 * 1) In the early boot process, because we build the real domains
6406 * after all CPUs have been brought up.
6407 *
6408 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6409 * domains.
6410 */
6411 rq_lock_irqsave(rq, &rf);
6412 if (rq->rd) {
6413 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6414 set_rq_online(rq);
6415 }
6416 rq_unlock_irqrestore(rq, &rf);
6417
6418 return 0;
6419}
6420
6421int sched_cpu_deactivate(unsigned int cpu)
6422{
6423 int ret;
6424
6425 set_cpu_active(cpu, false);
6426 /*
6427 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6428 * users of this state to go away such that all new such users will
6429 * observe it.
6430 *
6431 * Do sync before park smpboot threads to take care the rcu boost case.
6432 */
6433 synchronize_rcu();
6434
6435#ifdef CONFIG_SCHED_SMT
6436 /*
6437 * When going down, decrement the number of cores with SMT present.
6438 */
6439 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6440 static_branch_dec_cpuslocked(&sched_smt_present);
6441#endif
6442
6443 if (!sched_smp_initialized)
6444 return 0;
6445
6446 ret = cpuset_cpu_inactive(cpu);
6447 if (ret) {
6448 set_cpu_active(cpu, true);
6449 return ret;
6450 }
6451 sched_domains_numa_masks_clear(cpu);
6452 return 0;
6453}
6454
6455static void sched_rq_cpu_starting(unsigned int cpu)
6456{
6457 struct rq *rq = cpu_rq(cpu);
6458
6459 rq->calc_load_update = calc_load_update;
6460 update_max_interval();
6461}
6462
6463int sched_cpu_starting(unsigned int cpu)
6464{
6465 sched_rq_cpu_starting(cpu);
6466 sched_tick_start(cpu);
6467 return 0;
6468}
6469
6470#ifdef CONFIG_HOTPLUG_CPU
6471int sched_cpu_dying(unsigned int cpu)
6472{
6473 struct rq *rq = cpu_rq(cpu);
6474 struct rq_flags rf;
6475
6476 /* Handle pending wakeups and then migrate everything off */
6477 sched_ttwu_pending();
6478 sched_tick_stop(cpu);
6479
6480 rq_lock_irqsave(rq, &rf);
6481 if (rq->rd) {
6482 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6483 set_rq_offline(rq);
6484 }
6485 migrate_tasks(rq, &rf);
6486 BUG_ON(rq->nr_running != 1);
6487 rq_unlock_irqrestore(rq, &rf);
6488
6489 calc_load_migrate(rq);
6490 update_max_interval();
6491 nohz_balance_exit_idle(rq);
6492 hrtick_clear(rq);
6493 return 0;
6494}
6495#endif
6496
6497void __init sched_init_smp(void)
6498{
6499 sched_init_numa();
6500
6501 /*
6502 * There's no userspace yet to cause hotplug operations; hence all the
6503 * CPU masks are stable and all blatant races in the below code cannot
6504 * happen.
6505 */
6506 mutex_lock(&sched_domains_mutex);
6507 sched_init_domains(cpu_active_mask);
6508 mutex_unlock(&sched_domains_mutex);
6509
6510 /* Move init over to a non-isolated CPU */
6511 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
6512 BUG();
6513 sched_init_granularity();
6514
6515 init_sched_rt_class();
6516 init_sched_dl_class();
6517
6518 sched_smp_initialized = true;
6519}
6520
6521static int __init migration_init(void)
6522{
6523 sched_cpu_starting(smp_processor_id());
6524 return 0;
6525}
6526early_initcall(migration_init);
6527
6528#else
6529void __init sched_init_smp(void)
6530{
6531 sched_init_granularity();
6532}
6533#endif /* CONFIG_SMP */
6534
6535int in_sched_functions(unsigned long addr)
6536{
6537 return in_lock_functions(addr) ||
6538 (addr >= (unsigned long)__sched_text_start
6539 && addr < (unsigned long)__sched_text_end);
6540}
6541
6542#ifdef CONFIG_CGROUP_SCHED
6543/*
6544 * Default task group.
6545 * Every task in system belongs to this group at bootup.
6546 */
6547struct task_group root_task_group;
6548LIST_HEAD(task_groups);
6549
6550/* Cacheline aligned slab cache for task_group */
6551static struct kmem_cache *task_group_cache __read_mostly;
6552#endif
6553
6554DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6555DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
6556
6557void __init sched_init(void)
6558{
6559 unsigned long ptr = 0;
6560 int i;
6561
6562 wait_bit_init();
6563
6564#ifdef CONFIG_FAIR_GROUP_SCHED
6565 ptr += 2 * nr_cpu_ids * sizeof(void **);
6566#endif
6567#ifdef CONFIG_RT_GROUP_SCHED
6568 ptr += 2 * nr_cpu_ids * sizeof(void **);
6569#endif
6570 if (ptr) {
6571 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
6572
6573#ifdef CONFIG_FAIR_GROUP_SCHED
6574 root_task_group.se = (struct sched_entity **)ptr;
6575 ptr += nr_cpu_ids * sizeof(void **);
6576
6577 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6578 ptr += nr_cpu_ids * sizeof(void **);
6579
6580#endif /* CONFIG_FAIR_GROUP_SCHED */
6581#ifdef CONFIG_RT_GROUP_SCHED
6582 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6583 ptr += nr_cpu_ids * sizeof(void **);
6584
6585 root_task_group.rt_rq = (struct rt_rq **)ptr;
6586 ptr += nr_cpu_ids * sizeof(void **);
6587
6588#endif /* CONFIG_RT_GROUP_SCHED */
6589 }
6590#ifdef CONFIG_CPUMASK_OFFSTACK
6591 for_each_possible_cpu(i) {
6592 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6593 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6594 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6595 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6596 }
6597#endif /* CONFIG_CPUMASK_OFFSTACK */
6598
6599 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6600 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6601
6602#ifdef CONFIG_SMP
6603 init_defrootdomain();
6604#endif
6605
6606#ifdef CONFIG_RT_GROUP_SCHED
6607 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6608 global_rt_period(), global_rt_runtime());
6609#endif /* CONFIG_RT_GROUP_SCHED */
6610
6611#ifdef CONFIG_CGROUP_SCHED
6612 task_group_cache = KMEM_CACHE(task_group, 0);
6613
6614 list_add(&root_task_group.list, &task_groups);
6615 INIT_LIST_HEAD(&root_task_group.children);
6616 INIT_LIST_HEAD(&root_task_group.siblings);
6617 autogroup_init(&init_task);
6618#endif /* CONFIG_CGROUP_SCHED */
6619
6620 for_each_possible_cpu(i) {
6621 struct rq *rq;
6622
6623 rq = cpu_rq(i);
6624 raw_spin_lock_init(&rq->lock);
6625 rq->nr_running = 0;
6626 rq->calc_load_active = 0;
6627 rq->calc_load_update = jiffies + LOAD_FREQ;
6628 init_cfs_rq(&rq->cfs);
6629 init_rt_rq(&rq->rt);
6630 init_dl_rq(&rq->dl);
6631#ifdef CONFIG_FAIR_GROUP_SCHED
6632 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6633 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6634 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6635 /*
6636 * How much CPU bandwidth does root_task_group get?
6637 *
6638 * In case of task-groups formed thr' the cgroup filesystem, it
6639 * gets 100% of the CPU resources in the system. This overall
6640 * system CPU resource is divided among the tasks of
6641 * root_task_group and its child task-groups in a fair manner,
6642 * based on each entity's (task or task-group's) weight
6643 * (se->load.weight).
6644 *
6645 * In other words, if root_task_group has 10 tasks of weight
6646 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6647 * then A0's share of the CPU resource is:
6648 *
6649 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6650 *
6651 * We achieve this by letting root_task_group's tasks sit
6652 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6653 */
6654 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6655 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6656#endif /* CONFIG_FAIR_GROUP_SCHED */
6657
6658 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6659#ifdef CONFIG_RT_GROUP_SCHED
6660 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6661#endif
6662#ifdef CONFIG_SMP
6663 rq->sd = NULL;
6664 rq->rd = NULL;
6665 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6666 rq->balance_callback = NULL;
6667 rq->active_balance = 0;
6668 rq->next_balance = jiffies;
6669 rq->push_cpu = 0;
6670 rq->cpu = i;
6671 rq->online = 0;
6672 rq->idle_stamp = 0;
6673 rq->avg_idle = 2*sysctl_sched_migration_cost;
6674 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6675
6676 INIT_LIST_HEAD(&rq->cfs_tasks);
6677
6678 rq_attach_root(rq, &def_root_domain);
6679#ifdef CONFIG_NO_HZ_COMMON
6680 rq->last_load_update_tick = jiffies;
6681 rq->last_blocked_load_update_tick = jiffies;
6682 atomic_set(&rq->nohz_flags, 0);
6683#endif
6684#endif /* CONFIG_SMP */
6685 hrtick_rq_init(rq);
6686 atomic_set(&rq->nr_iowait, 0);
6687 }
6688
6689 set_load_weight(&init_task, false);
6690
6691 /*
6692 * The boot idle thread does lazy MMU switching as well:
6693 */
6694 mmgrab(&init_mm);
6695 enter_lazy_tlb(&init_mm, current);
6696
6697 /*
6698 * Make us the idle thread. Technically, schedule() should not be
6699 * called from this thread, however somewhere below it might be,
6700 * but because we are the idle thread, we just pick up running again
6701 * when this runqueue becomes "idle".
6702 */
6703 init_idle(current, smp_processor_id());
6704
6705 calc_load_update = jiffies + LOAD_FREQ;
6706
6707#ifdef CONFIG_SMP
6708 idle_thread_set_boot_cpu();
6709#endif
6710 init_sched_fair_class();
6711
6712 init_schedstats();
6713
6714 psi_init();
6715
6716 init_uclamp();
6717
6718 scheduler_running = 1;
6719}
6720
6721#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6722static inline int preempt_count_equals(int preempt_offset)
6723{
6724 int nested = preempt_count() + rcu_preempt_depth();
6725
6726 return (nested == preempt_offset);
6727}
6728
6729void __might_sleep(const char *file, int line, int preempt_offset)
6730{
6731 /*
6732 * Blocking primitives will set (and therefore destroy) current->state,
6733 * since we will exit with TASK_RUNNING make sure we enter with it,
6734 * otherwise we will destroy state.
6735 */
6736 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6737 "do not call blocking ops when !TASK_RUNNING; "
6738 "state=%lx set at [<%p>] %pS\n",
6739 current->state,
6740 (void *)current->task_state_change,
6741 (void *)current->task_state_change);
6742
6743 ___might_sleep(file, line, preempt_offset);
6744}
6745EXPORT_SYMBOL(__might_sleep);
6746
6747void ___might_sleep(const char *file, int line, int preempt_offset)
6748{
6749 /* Ratelimiting timestamp: */
6750 static unsigned long prev_jiffy;
6751
6752 unsigned long preempt_disable_ip;
6753
6754 /* WARN_ON_ONCE() by default, no rate limit required: */
6755 rcu_sleep_check();
6756
6757 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6758 !is_idle_task(current) && !current->non_block_count) ||
6759 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6760 oops_in_progress)
6761 return;
6762
6763 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6764 return;
6765 prev_jiffy = jiffies;
6766
6767 /* Save this before calling printk(), since that will clobber it: */
6768 preempt_disable_ip = get_preempt_disable_ip(current);
6769
6770 printk(KERN_ERR
6771 "BUG: sleeping function called from invalid context at %s:%d\n",
6772 file, line);
6773 printk(KERN_ERR
6774 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
6775 in_atomic(), irqs_disabled(), current->non_block_count,
6776 current->pid, current->comm);
6777
6778 if (task_stack_end_corrupted(current))
6779 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6780
6781 debug_show_held_locks(current);
6782 if (irqs_disabled())
6783 print_irqtrace_events(current);
6784 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6785 && !preempt_count_equals(preempt_offset)) {
6786 pr_err("Preemption disabled at:");
6787 print_ip_sym(preempt_disable_ip);
6788 pr_cont("\n");
6789 }
6790 dump_stack();
6791 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6792}
6793EXPORT_SYMBOL(___might_sleep);
6794
6795void __cant_sleep(const char *file, int line, int preempt_offset)
6796{
6797 static unsigned long prev_jiffy;
6798
6799 if (irqs_disabled())
6800 return;
6801
6802 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
6803 return;
6804
6805 if (preempt_count() > preempt_offset)
6806 return;
6807
6808 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6809 return;
6810 prev_jiffy = jiffies;
6811
6812 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
6813 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6814 in_atomic(), irqs_disabled(),
6815 current->pid, current->comm);
6816
6817 debug_show_held_locks(current);
6818 dump_stack();
6819 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6820}
6821EXPORT_SYMBOL_GPL(__cant_sleep);
6822#endif
6823
6824#ifdef CONFIG_MAGIC_SYSRQ
6825void normalize_rt_tasks(void)
6826{
6827 struct task_struct *g, *p;
6828 struct sched_attr attr = {
6829 .sched_policy = SCHED_NORMAL,
6830 };
6831
6832 read_lock(&tasklist_lock);
6833 for_each_process_thread(g, p) {
6834 /*
6835 * Only normalize user tasks:
6836 */
6837 if (p->flags & PF_KTHREAD)
6838 continue;
6839
6840 p->se.exec_start = 0;
6841 schedstat_set(p->se.statistics.wait_start, 0);
6842 schedstat_set(p->se.statistics.sleep_start, 0);
6843 schedstat_set(p->se.statistics.block_start, 0);
6844
6845 if (!dl_task(p) && !rt_task(p)) {
6846 /*
6847 * Renice negative nice level userspace
6848 * tasks back to 0:
6849 */
6850 if (task_nice(p) < 0)
6851 set_user_nice(p, 0);
6852 continue;
6853 }
6854
6855 __sched_setscheduler(p, &attr, false, false);
6856 }
6857 read_unlock(&tasklist_lock);
6858}
6859
6860#endif /* CONFIG_MAGIC_SYSRQ */
6861
6862#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6863/*
6864 * These functions are only useful for the IA64 MCA handling, or kdb.
6865 *
6866 * They can only be called when the whole system has been
6867 * stopped - every CPU needs to be quiescent, and no scheduling
6868 * activity can take place. Using them for anything else would
6869 * be a serious bug, and as a result, they aren't even visible
6870 * under any other configuration.
6871 */
6872
6873/**
6874 * curr_task - return the current task for a given CPU.
6875 * @cpu: the processor in question.
6876 *
6877 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6878 *
6879 * Return: The current task for @cpu.
6880 */
6881struct task_struct *curr_task(int cpu)
6882{
6883 return cpu_curr(cpu);
6884}
6885
6886#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6887
6888#ifdef CONFIG_IA64
6889/**
6890 * ia64_set_curr_task - set the current task for a given CPU.
6891 * @cpu: the processor in question.
6892 * @p: the task pointer to set.
6893 *
6894 * Description: This function must only be used when non-maskable interrupts
6895 * are serviced on a separate stack. It allows the architecture to switch the
6896 * notion of the current task on a CPU in a non-blocking manner. This function
6897 * must be called with all CPU's synchronized, and interrupts disabled, the
6898 * and caller must save the original value of the current task (see
6899 * curr_task() above) and restore that value before reenabling interrupts and
6900 * re-starting the system.
6901 *
6902 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6903 */
6904void ia64_set_curr_task(int cpu, struct task_struct *p)
6905{
6906 cpu_curr(cpu) = p;
6907}
6908
6909#endif
6910
6911#ifdef CONFIG_CGROUP_SCHED
6912/* task_group_lock serializes the addition/removal of task groups */
6913static DEFINE_SPINLOCK(task_group_lock);
6914
6915static inline void alloc_uclamp_sched_group(struct task_group *tg,
6916 struct task_group *parent)
6917{
6918#ifdef CONFIG_UCLAMP_TASK_GROUP
6919 enum uclamp_id clamp_id;
6920
6921 for_each_clamp_id(clamp_id) {
6922 uclamp_se_set(&tg->uclamp_req[clamp_id],
6923 uclamp_none(clamp_id), false);
6924 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
6925 }
6926#endif
6927}
6928
6929static void sched_free_group(struct task_group *tg)
6930{
6931 free_fair_sched_group(tg);
6932 free_rt_sched_group(tg);
6933 autogroup_free(tg);
6934 kmem_cache_free(task_group_cache, tg);
6935}
6936
6937/* allocate runqueue etc for a new task group */
6938struct task_group *sched_create_group(struct task_group *parent)
6939{
6940 struct task_group *tg;
6941
6942 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6943 if (!tg)
6944 return ERR_PTR(-ENOMEM);
6945
6946 if (!alloc_fair_sched_group(tg, parent))
6947 goto err;
6948
6949 if (!alloc_rt_sched_group(tg, parent))
6950 goto err;
6951
6952 alloc_uclamp_sched_group(tg, parent);
6953
6954 return tg;
6955
6956err:
6957 sched_free_group(tg);
6958 return ERR_PTR(-ENOMEM);
6959}
6960
6961void sched_online_group(struct task_group *tg, struct task_group *parent)
6962{
6963 unsigned long flags;
6964
6965 spin_lock_irqsave(&task_group_lock, flags);
6966 list_add_rcu(&tg->list, &task_groups);
6967
6968 /* Root should already exist: */
6969 WARN_ON(!parent);
6970
6971 tg->parent = parent;
6972 INIT_LIST_HEAD(&tg->children);
6973 list_add_rcu(&tg->siblings, &parent->children);
6974 spin_unlock_irqrestore(&task_group_lock, flags);
6975
6976 online_fair_sched_group(tg);
6977}
6978
6979/* rcu callback to free various structures associated with a task group */
6980static void sched_free_group_rcu(struct rcu_head *rhp)
6981{
6982 /* Now it should be safe to free those cfs_rqs: */
6983 sched_free_group(container_of(rhp, struct task_group, rcu));
6984}
6985
6986void sched_destroy_group(struct task_group *tg)
6987{
6988 /* Wait for possible concurrent references to cfs_rqs complete: */
6989 call_rcu(&tg->rcu, sched_free_group_rcu);
6990}
6991
6992void sched_offline_group(struct task_group *tg)
6993{
6994 unsigned long flags;
6995
6996 /* End participation in shares distribution: */
6997 unregister_fair_sched_group(tg);
6998
6999 spin_lock_irqsave(&task_group_lock, flags);
7000 list_del_rcu(&tg->list);
7001 list_del_rcu(&tg->siblings);
7002 spin_unlock_irqrestore(&task_group_lock, flags);
7003}
7004
7005static void sched_change_group(struct task_struct *tsk, int type)
7006{
7007 struct task_group *tg;
7008
7009 /*
7010 * All callers are synchronized by task_rq_lock(); we do not use RCU
7011 * which is pointless here. Thus, we pass "true" to task_css_check()
7012 * to prevent lockdep warnings.
7013 */
7014 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7015 struct task_group, css);
7016 tg = autogroup_task_group(tsk, tg);
7017 tsk->sched_task_group = tg;
7018
7019#ifdef CONFIG_FAIR_GROUP_SCHED
7020 if (tsk->sched_class->task_change_group)
7021 tsk->sched_class->task_change_group(tsk, type);
7022 else
7023#endif
7024 set_task_rq(tsk, task_cpu(tsk));
7025}
7026
7027/*
7028 * Change task's runqueue when it moves between groups.
7029 *
7030 * The caller of this function should have put the task in its new group by
7031 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7032 * its new group.
7033 */
7034void sched_move_task(struct task_struct *tsk)
7035{
7036 int queued, running, queue_flags =
7037 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7038 struct rq_flags rf;
7039 struct rq *rq;
7040
7041 rq = task_rq_lock(tsk, &rf);
7042 update_rq_clock(rq);
7043
7044 running = task_current(rq, tsk);
7045 queued = task_on_rq_queued(tsk);
7046
7047 if (queued)
7048 dequeue_task(rq, tsk, queue_flags);
7049 if (running)
7050 put_prev_task(rq, tsk);
7051
7052 sched_change_group(tsk, TASK_MOVE_GROUP);
7053
7054 if (queued)
7055 enqueue_task(rq, tsk, queue_flags);
7056 if (running)
7057 set_next_task(rq, tsk);
7058
7059 task_rq_unlock(rq, tsk, &rf);
7060}
7061
7062static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7063{
7064 return css ? container_of(css, struct task_group, css) : NULL;
7065}
7066
7067static struct cgroup_subsys_state *
7068cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7069{
7070 struct task_group *parent = css_tg(parent_css);
7071 struct task_group *tg;
7072
7073 if (!parent) {
7074 /* This is early initialization for the top cgroup */
7075 return &root_task_group.css;
7076 }
7077
7078 tg = sched_create_group(parent);
7079 if (IS_ERR(tg))
7080 return ERR_PTR(-ENOMEM);
7081
7082 return &tg->css;
7083}
7084
7085/* Expose task group only after completing cgroup initialization */
7086static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7087{
7088 struct task_group *tg = css_tg(css);
7089 struct task_group *parent = css_tg(css->parent);
7090
7091 if (parent)
7092 sched_online_group(tg, parent);
7093 return 0;
7094}
7095
7096static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
7097{
7098 struct task_group *tg = css_tg(css);
7099
7100 sched_offline_group(tg);
7101}
7102
7103static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7104{
7105 struct task_group *tg = css_tg(css);
7106
7107 /*
7108 * Relies on the RCU grace period between css_released() and this.
7109 */
7110 sched_free_group(tg);
7111}
7112
7113/*
7114 * This is called before wake_up_new_task(), therefore we really only
7115 * have to set its group bits, all the other stuff does not apply.
7116 */
7117static void cpu_cgroup_fork(struct task_struct *task)
7118{
7119 struct rq_flags rf;
7120 struct rq *rq;
7121
7122 rq = task_rq_lock(task, &rf);
7123
7124 update_rq_clock(rq);
7125 sched_change_group(task, TASK_SET_GROUP);
7126
7127 task_rq_unlock(rq, task, &rf);
7128}
7129
7130static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
7131{
7132 struct task_struct *task;
7133 struct cgroup_subsys_state *css;
7134 int ret = 0;
7135
7136 cgroup_taskset_for_each(task, css, tset) {
7137#ifdef CONFIG_RT_GROUP_SCHED
7138 if (!sched_rt_can_attach(css_tg(css), task))
7139 return -EINVAL;
7140#endif
7141 /*
7142 * Serialize against wake_up_new_task() such that if its
7143 * running, we're sure to observe its full state.
7144 */
7145 raw_spin_lock_irq(&task->pi_lock);
7146 /*
7147 * Avoid calling sched_move_task() before wake_up_new_task()
7148 * has happened. This would lead to problems with PELT, due to
7149 * move wanting to detach+attach while we're not attached yet.
7150 */
7151 if (task->state == TASK_NEW)
7152 ret = -EINVAL;
7153 raw_spin_unlock_irq(&task->pi_lock);
7154
7155 if (ret)
7156 break;
7157 }
7158 return ret;
7159}
7160
7161static void cpu_cgroup_attach(struct cgroup_taskset *tset)
7162{
7163 struct task_struct *task;
7164 struct cgroup_subsys_state *css;
7165
7166 cgroup_taskset_for_each(task, css, tset)
7167 sched_move_task(task);
7168}
7169
7170#ifdef CONFIG_UCLAMP_TASK_GROUP
7171static void cpu_util_update_eff(struct cgroup_subsys_state *css)
7172{
7173 struct cgroup_subsys_state *top_css = css;
7174 struct uclamp_se *uc_parent = NULL;
7175 struct uclamp_se *uc_se = NULL;
7176 unsigned int eff[UCLAMP_CNT];
7177 enum uclamp_id clamp_id;
7178 unsigned int clamps;
7179
7180 css_for_each_descendant_pre(css, top_css) {
7181 uc_parent = css_tg(css)->parent
7182 ? css_tg(css)->parent->uclamp : NULL;
7183
7184 for_each_clamp_id(clamp_id) {
7185 /* Assume effective clamps matches requested clamps */
7186 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
7187 /* Cap effective clamps with parent's effective clamps */
7188 if (uc_parent &&
7189 eff[clamp_id] > uc_parent[clamp_id].value) {
7190 eff[clamp_id] = uc_parent[clamp_id].value;
7191 }
7192 }
7193 /* Ensure protection is always capped by limit */
7194 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
7195
7196 /* Propagate most restrictive effective clamps */
7197 clamps = 0x0;
7198 uc_se = css_tg(css)->uclamp;
7199 for_each_clamp_id(clamp_id) {
7200 if (eff[clamp_id] == uc_se[clamp_id].value)
7201 continue;
7202 uc_se[clamp_id].value = eff[clamp_id];
7203 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
7204 clamps |= (0x1 << clamp_id);
7205 }
7206 if (!clamps) {
7207 css = css_rightmost_descendant(css);
7208 continue;
7209 }
7210
7211 /* Immediately update descendants RUNNABLE tasks */
7212 uclamp_update_active_tasks(css, clamps);
7213 }
7214}
7215
7216/*
7217 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7218 * C expression. Since there is no way to convert a macro argument (N) into a
7219 * character constant, use two levels of macros.
7220 */
7221#define _POW10(exp) ((unsigned int)1e##exp)
7222#define POW10(exp) _POW10(exp)
7223
7224struct uclamp_request {
7225#define UCLAMP_PERCENT_SHIFT 2
7226#define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7227 s64 percent;
7228 u64 util;
7229 int ret;
7230};
7231
7232static inline struct uclamp_request
7233capacity_from_percent(char *buf)
7234{
7235 struct uclamp_request req = {
7236 .percent = UCLAMP_PERCENT_SCALE,
7237 .util = SCHED_CAPACITY_SCALE,
7238 .ret = 0,
7239 };
7240
7241 buf = strim(buf);
7242 if (strcmp(buf, "max")) {
7243 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
7244 &req.percent);
7245 if (req.ret)
7246 return req;
7247 if (req.percent > UCLAMP_PERCENT_SCALE) {
7248 req.ret = -ERANGE;
7249 return req;
7250 }
7251
7252 req.util = req.percent << SCHED_CAPACITY_SHIFT;
7253 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
7254 }
7255
7256 return req;
7257}
7258
7259static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
7260 size_t nbytes, loff_t off,
7261 enum uclamp_id clamp_id)
7262{
7263 struct uclamp_request req;
7264 struct task_group *tg;
7265
7266 req = capacity_from_percent(buf);
7267 if (req.ret)
7268 return req.ret;
7269
7270 mutex_lock(&uclamp_mutex);
7271 rcu_read_lock();
7272
7273 tg = css_tg(of_css(of));
7274 if (tg->uclamp_req[clamp_id].value != req.util)
7275 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
7276
7277 /*
7278 * Because of not recoverable conversion rounding we keep track of the
7279 * exact requested value
7280 */
7281 tg->uclamp_pct[clamp_id] = req.percent;
7282
7283 /* Update effective clamps to track the most restrictive value */
7284 cpu_util_update_eff(of_css(of));
7285
7286 rcu_read_unlock();
7287 mutex_unlock(&uclamp_mutex);
7288
7289 return nbytes;
7290}
7291
7292static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
7293 char *buf, size_t nbytes,
7294 loff_t off)
7295{
7296 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
7297}
7298
7299static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
7300 char *buf, size_t nbytes,
7301 loff_t off)
7302{
7303 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
7304}
7305
7306static inline void cpu_uclamp_print(struct seq_file *sf,
7307 enum uclamp_id clamp_id)
7308{
7309 struct task_group *tg;
7310 u64 util_clamp;
7311 u64 percent;
7312 u32 rem;
7313
7314 rcu_read_lock();
7315 tg = css_tg(seq_css(sf));
7316 util_clamp = tg->uclamp_req[clamp_id].value;
7317 rcu_read_unlock();
7318
7319 if (util_clamp == SCHED_CAPACITY_SCALE) {
7320 seq_puts(sf, "max\n");
7321 return;
7322 }
7323
7324 percent = tg->uclamp_pct[clamp_id];
7325 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
7326 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
7327}
7328
7329static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
7330{
7331 cpu_uclamp_print(sf, UCLAMP_MIN);
7332 return 0;
7333}
7334
7335static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
7336{
7337 cpu_uclamp_print(sf, UCLAMP_MAX);
7338 return 0;
7339}
7340#endif /* CONFIG_UCLAMP_TASK_GROUP */
7341
7342#ifdef CONFIG_FAIR_GROUP_SCHED
7343static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7344 struct cftype *cftype, u64 shareval)
7345{
7346 if (shareval > scale_load_down(ULONG_MAX))
7347 shareval = MAX_SHARES;
7348 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7349}
7350
7351static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7352 struct cftype *cft)
7353{
7354 struct task_group *tg = css_tg(css);
7355
7356 return (u64) scale_load_down(tg->shares);
7357}
7358
7359#ifdef CONFIG_CFS_BANDWIDTH
7360static DEFINE_MUTEX(cfs_constraints_mutex);
7361
7362const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7363static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7364
7365static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7366
7367static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7368{
7369 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7370 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7371
7372 if (tg == &root_task_group)
7373 return -EINVAL;
7374
7375 /*
7376 * Ensure we have at some amount of bandwidth every period. This is
7377 * to prevent reaching a state of large arrears when throttled via
7378 * entity_tick() resulting in prolonged exit starvation.
7379 */
7380 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7381 return -EINVAL;
7382
7383 /*
7384 * Likewise, bound things on the otherside by preventing insane quota
7385 * periods. This also allows us to normalize in computing quota
7386 * feasibility.
7387 */
7388 if (period > max_cfs_quota_period)
7389 return -EINVAL;
7390
7391 /*
7392 * Prevent race between setting of cfs_rq->runtime_enabled and
7393 * unthrottle_offline_cfs_rqs().
7394 */
7395 get_online_cpus();
7396 mutex_lock(&cfs_constraints_mutex);
7397 ret = __cfs_schedulable(tg, period, quota);
7398 if (ret)
7399 goto out_unlock;
7400
7401 runtime_enabled = quota != RUNTIME_INF;
7402 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7403 /*
7404 * If we need to toggle cfs_bandwidth_used, off->on must occur
7405 * before making related changes, and on->off must occur afterwards
7406 */
7407 if (runtime_enabled && !runtime_was_enabled)
7408 cfs_bandwidth_usage_inc();
7409 raw_spin_lock_irq(&cfs_b->lock);
7410 cfs_b->period = ns_to_ktime(period);
7411 cfs_b->quota = quota;
7412
7413 __refill_cfs_bandwidth_runtime(cfs_b);
7414
7415 /* Restart the period timer (if active) to handle new period expiry: */
7416 if (runtime_enabled)
7417 start_cfs_bandwidth(cfs_b);
7418
7419 raw_spin_unlock_irq(&cfs_b->lock);
7420
7421 for_each_online_cpu(i) {
7422 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7423 struct rq *rq = cfs_rq->rq;
7424 struct rq_flags rf;
7425
7426 rq_lock_irq(rq, &rf);
7427 cfs_rq->runtime_enabled = runtime_enabled;
7428 cfs_rq->runtime_remaining = 0;
7429
7430 if (cfs_rq->throttled)
7431 unthrottle_cfs_rq(cfs_rq);
7432 rq_unlock_irq(rq, &rf);
7433 }
7434 if (runtime_was_enabled && !runtime_enabled)
7435 cfs_bandwidth_usage_dec();
7436out_unlock:
7437 mutex_unlock(&cfs_constraints_mutex);
7438 put_online_cpus();
7439
7440 return ret;
7441}
7442
7443static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7444{
7445 u64 quota, period;
7446
7447 period = ktime_to_ns(tg->cfs_bandwidth.period);
7448 if (cfs_quota_us < 0)
7449 quota = RUNTIME_INF;
7450 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
7451 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7452 else
7453 return -EINVAL;
7454
7455 return tg_set_cfs_bandwidth(tg, period, quota);
7456}
7457
7458static long tg_get_cfs_quota(struct task_group *tg)
7459{
7460 u64 quota_us;
7461
7462 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7463 return -1;
7464
7465 quota_us = tg->cfs_bandwidth.quota;
7466 do_div(quota_us, NSEC_PER_USEC);
7467
7468 return quota_us;
7469}
7470
7471static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7472{
7473 u64 quota, period;
7474
7475 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
7476 return -EINVAL;
7477
7478 period = (u64)cfs_period_us * NSEC_PER_USEC;
7479 quota = tg->cfs_bandwidth.quota;
7480
7481 return tg_set_cfs_bandwidth(tg, period, quota);
7482}
7483
7484static long tg_get_cfs_period(struct task_group *tg)
7485{
7486 u64 cfs_period_us;
7487
7488 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7489 do_div(cfs_period_us, NSEC_PER_USEC);
7490
7491 return cfs_period_us;
7492}
7493
7494static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7495 struct cftype *cft)
7496{
7497 return tg_get_cfs_quota(css_tg(css));
7498}
7499
7500static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7501 struct cftype *cftype, s64 cfs_quota_us)
7502{
7503 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7504}
7505
7506static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7507 struct cftype *cft)
7508{
7509 return tg_get_cfs_period(css_tg(css));
7510}
7511
7512static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7513 struct cftype *cftype, u64 cfs_period_us)
7514{
7515 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7516}
7517
7518struct cfs_schedulable_data {
7519 struct task_group *tg;
7520 u64 period, quota;
7521};
7522
7523/*
7524 * normalize group quota/period to be quota/max_period
7525 * note: units are usecs
7526 */
7527static u64 normalize_cfs_quota(struct task_group *tg,
7528 struct cfs_schedulable_data *d)
7529{
7530 u64 quota, period;
7531
7532 if (tg == d->tg) {
7533 period = d->period;
7534 quota = d->quota;
7535 } else {
7536 period = tg_get_cfs_period(tg);
7537 quota = tg_get_cfs_quota(tg);
7538 }
7539
7540 /* note: these should typically be equivalent */
7541 if (quota == RUNTIME_INF || quota == -1)
7542 return RUNTIME_INF;
7543
7544 return to_ratio(period, quota);
7545}
7546
7547static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7548{
7549 struct cfs_schedulable_data *d = data;
7550 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7551 s64 quota = 0, parent_quota = -1;
7552
7553 if (!tg->parent) {
7554 quota = RUNTIME_INF;
7555 } else {
7556 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7557
7558 quota = normalize_cfs_quota(tg, d);
7559 parent_quota = parent_b->hierarchical_quota;
7560
7561 /*
7562 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7563 * always take the min. On cgroup1, only inherit when no
7564 * limit is set:
7565 */
7566 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
7567 quota = min(quota, parent_quota);
7568 } else {
7569 if (quota == RUNTIME_INF)
7570 quota = parent_quota;
7571 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7572 return -EINVAL;
7573 }
7574 }
7575 cfs_b->hierarchical_quota = quota;
7576
7577 return 0;
7578}
7579
7580static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7581{
7582 int ret;
7583 struct cfs_schedulable_data data = {
7584 .tg = tg,
7585 .period = period,
7586 .quota = quota,
7587 };
7588
7589 if (quota != RUNTIME_INF) {
7590 do_div(data.period, NSEC_PER_USEC);
7591 do_div(data.quota, NSEC_PER_USEC);
7592 }
7593
7594 rcu_read_lock();
7595 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7596 rcu_read_unlock();
7597
7598 return ret;
7599}
7600
7601static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
7602{
7603 struct task_group *tg = css_tg(seq_css(sf));
7604 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7605
7606 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7607 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7608 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7609
7610 if (schedstat_enabled() && tg != &root_task_group) {
7611 u64 ws = 0;
7612 int i;
7613
7614 for_each_possible_cpu(i)
7615 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
7616
7617 seq_printf(sf, "wait_sum %llu\n", ws);
7618 }
7619
7620 return 0;
7621}
7622#endif /* CONFIG_CFS_BANDWIDTH */
7623#endif /* CONFIG_FAIR_GROUP_SCHED */
7624
7625#ifdef CONFIG_RT_GROUP_SCHED
7626static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7627 struct cftype *cft, s64 val)
7628{
7629 return sched_group_set_rt_runtime(css_tg(css), val);
7630}
7631
7632static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7633 struct cftype *cft)
7634{
7635 return sched_group_rt_runtime(css_tg(css));
7636}
7637
7638static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7639 struct cftype *cftype, u64 rt_period_us)
7640{
7641 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7642}
7643
7644static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7645 struct cftype *cft)
7646{
7647 return sched_group_rt_period(css_tg(css));
7648}
7649#endif /* CONFIG_RT_GROUP_SCHED */
7650
7651static struct cftype cpu_legacy_files[] = {
7652#ifdef CONFIG_FAIR_GROUP_SCHED
7653 {
7654 .name = "shares",
7655 .read_u64 = cpu_shares_read_u64,
7656 .write_u64 = cpu_shares_write_u64,
7657 },
7658#endif
7659#ifdef CONFIG_CFS_BANDWIDTH
7660 {
7661 .name = "cfs_quota_us",
7662 .read_s64 = cpu_cfs_quota_read_s64,
7663 .write_s64 = cpu_cfs_quota_write_s64,
7664 },
7665 {
7666 .name = "cfs_period_us",
7667 .read_u64 = cpu_cfs_period_read_u64,
7668 .write_u64 = cpu_cfs_period_write_u64,
7669 },
7670 {
7671 .name = "stat",
7672 .seq_show = cpu_cfs_stat_show,
7673 },
7674#endif
7675#ifdef CONFIG_RT_GROUP_SCHED
7676 {
7677 .name = "rt_runtime_us",
7678 .read_s64 = cpu_rt_runtime_read,
7679 .write_s64 = cpu_rt_runtime_write,
7680 },
7681 {
7682 .name = "rt_period_us",
7683 .read_u64 = cpu_rt_period_read_uint,
7684 .write_u64 = cpu_rt_period_write_uint,
7685 },
7686#endif
7687#ifdef CONFIG_UCLAMP_TASK_GROUP
7688 {
7689 .name = "uclamp.min",
7690 .flags = CFTYPE_NOT_ON_ROOT,
7691 .seq_show = cpu_uclamp_min_show,
7692 .write = cpu_uclamp_min_write,
7693 },
7694 {
7695 .name = "uclamp.max",
7696 .flags = CFTYPE_NOT_ON_ROOT,
7697 .seq_show = cpu_uclamp_max_show,
7698 .write = cpu_uclamp_max_write,
7699 },
7700#endif
7701 { } /* Terminate */
7702};
7703
7704static int cpu_extra_stat_show(struct seq_file *sf,
7705 struct cgroup_subsys_state *css)
7706{
7707#ifdef CONFIG_CFS_BANDWIDTH
7708 {
7709 struct task_group *tg = css_tg(css);
7710 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7711 u64 throttled_usec;
7712
7713 throttled_usec = cfs_b->throttled_time;
7714 do_div(throttled_usec, NSEC_PER_USEC);
7715
7716 seq_printf(sf, "nr_periods %d\n"
7717 "nr_throttled %d\n"
7718 "throttled_usec %llu\n",
7719 cfs_b->nr_periods, cfs_b->nr_throttled,
7720 throttled_usec);
7721 }
7722#endif
7723 return 0;
7724}
7725
7726#ifdef CONFIG_FAIR_GROUP_SCHED
7727static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
7728 struct cftype *cft)
7729{
7730 struct task_group *tg = css_tg(css);
7731 u64 weight = scale_load_down(tg->shares);
7732
7733 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
7734}
7735
7736static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
7737 struct cftype *cft, u64 weight)
7738{
7739 /*
7740 * cgroup weight knobs should use the common MIN, DFL and MAX
7741 * values which are 1, 100 and 10000 respectively. While it loses
7742 * a bit of range on both ends, it maps pretty well onto the shares
7743 * value used by scheduler and the round-trip conversions preserve
7744 * the original value over the entire range.
7745 */
7746 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
7747 return -ERANGE;
7748
7749 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
7750
7751 return sched_group_set_shares(css_tg(css), scale_load(weight));
7752}
7753
7754static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
7755 struct cftype *cft)
7756{
7757 unsigned long weight = scale_load_down(css_tg(css)->shares);
7758 int last_delta = INT_MAX;
7759 int prio, delta;
7760
7761 /* find the closest nice value to the current weight */
7762 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
7763 delta = abs(sched_prio_to_weight[prio] - weight);
7764 if (delta >= last_delta)
7765 break;
7766 last_delta = delta;
7767 }
7768
7769 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
7770}
7771
7772static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
7773 struct cftype *cft, s64 nice)
7774{
7775 unsigned long weight;
7776 int idx;
7777
7778 if (nice < MIN_NICE || nice > MAX_NICE)
7779 return -ERANGE;
7780
7781 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
7782 idx = array_index_nospec(idx, 40);
7783 weight = sched_prio_to_weight[idx];
7784
7785 return sched_group_set_shares(css_tg(css), scale_load(weight));
7786}
7787#endif
7788
7789static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
7790 long period, long quota)
7791{
7792 if (quota < 0)
7793 seq_puts(sf, "max");
7794 else
7795 seq_printf(sf, "%ld", quota);
7796
7797 seq_printf(sf, " %ld\n", period);
7798}
7799
7800/* caller should put the current value in *@periodp before calling */
7801static int __maybe_unused cpu_period_quota_parse(char *buf,
7802 u64 *periodp, u64 *quotap)
7803{
7804 char tok[21]; /* U64_MAX */
7805
7806 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
7807 return -EINVAL;
7808
7809 *periodp *= NSEC_PER_USEC;
7810
7811 if (sscanf(tok, "%llu", quotap))
7812 *quotap *= NSEC_PER_USEC;
7813 else if (!strcmp(tok, "max"))
7814 *quotap = RUNTIME_INF;
7815 else
7816 return -EINVAL;
7817
7818 return 0;
7819}
7820
7821#ifdef CONFIG_CFS_BANDWIDTH
7822static int cpu_max_show(struct seq_file *sf, void *v)
7823{
7824 struct task_group *tg = css_tg(seq_css(sf));
7825
7826 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
7827 return 0;
7828}
7829
7830static ssize_t cpu_max_write(struct kernfs_open_file *of,
7831 char *buf, size_t nbytes, loff_t off)
7832{
7833 struct task_group *tg = css_tg(of_css(of));
7834 u64 period = tg_get_cfs_period(tg);
7835 u64 quota;
7836 int ret;
7837
7838 ret = cpu_period_quota_parse(buf, &period, "a);
7839 if (!ret)
7840 ret = tg_set_cfs_bandwidth(tg, period, quota);
7841 return ret ?: nbytes;
7842}
7843#endif
7844
7845static struct cftype cpu_files[] = {
7846#ifdef CONFIG_FAIR_GROUP_SCHED
7847 {
7848 .name = "weight",
7849 .flags = CFTYPE_NOT_ON_ROOT,
7850 .read_u64 = cpu_weight_read_u64,
7851 .write_u64 = cpu_weight_write_u64,
7852 },
7853 {
7854 .name = "weight.nice",
7855 .flags = CFTYPE_NOT_ON_ROOT,
7856 .read_s64 = cpu_weight_nice_read_s64,
7857 .write_s64 = cpu_weight_nice_write_s64,
7858 },
7859#endif
7860#ifdef CONFIG_CFS_BANDWIDTH
7861 {
7862 .name = "max",
7863 .flags = CFTYPE_NOT_ON_ROOT,
7864 .seq_show = cpu_max_show,
7865 .write = cpu_max_write,
7866 },
7867#endif
7868#ifdef CONFIG_UCLAMP_TASK_GROUP
7869 {
7870 .name = "uclamp.min",
7871 .flags = CFTYPE_NOT_ON_ROOT,
7872 .seq_show = cpu_uclamp_min_show,
7873 .write = cpu_uclamp_min_write,
7874 },
7875 {
7876 .name = "uclamp.max",
7877 .flags = CFTYPE_NOT_ON_ROOT,
7878 .seq_show = cpu_uclamp_max_show,
7879 .write = cpu_uclamp_max_write,
7880 },
7881#endif
7882 { } /* terminate */
7883};
7884
7885struct cgroup_subsys cpu_cgrp_subsys = {
7886 .css_alloc = cpu_cgroup_css_alloc,
7887 .css_online = cpu_cgroup_css_online,
7888 .css_released = cpu_cgroup_css_released,
7889 .css_free = cpu_cgroup_css_free,
7890 .css_extra_stat_show = cpu_extra_stat_show,
7891 .fork = cpu_cgroup_fork,
7892 .can_attach = cpu_cgroup_can_attach,
7893 .attach = cpu_cgroup_attach,
7894 .legacy_cftypes = cpu_legacy_files,
7895 .dfl_cftypes = cpu_files,
7896 .early_init = true,
7897 .threaded = true,
7898};
7899
7900#endif /* CONFIG_CGROUP_SCHED */
7901
7902void dump_cpu_task(int cpu)
7903{
7904 pr_info("Task dump for CPU %d:\n", cpu);
7905 sched_show_task(cpu_curr(cpu));
7906}
7907
7908/*
7909 * Nice levels are multiplicative, with a gentle 10% change for every
7910 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7911 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7912 * that remained on nice 0.
7913 *
7914 * The "10% effect" is relative and cumulative: from _any_ nice level,
7915 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7916 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7917 * If a task goes up by ~10% and another task goes down by ~10% then
7918 * the relative distance between them is ~25%.)
7919 */
7920const int sched_prio_to_weight[40] = {
7921 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7922 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7923 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7924 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7925 /* 0 */ 1024, 820, 655, 526, 423,
7926 /* 5 */ 335, 272, 215, 172, 137,
7927 /* 10 */ 110, 87, 70, 56, 45,
7928 /* 15 */ 36, 29, 23, 18, 15,
7929};
7930
7931/*
7932 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7933 *
7934 * In cases where the weight does not change often, we can use the
7935 * precalculated inverse to speed up arithmetics by turning divisions
7936 * into multiplications:
7937 */
7938const u32 sched_prio_to_wmult[40] = {
7939 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7940 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7941 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7942 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7943 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7944 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7945 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7946 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7947};
7948
7949#undef CREATE_TRACE_POINTS
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * kernel/sched/core.c
4 *
5 * Core kernel scheduler code and related syscalls
6 *
7 * Copyright (C) 1991-2002 Linus Torvalds
8 */
9#define CREATE_TRACE_POINTS
10#include <trace/events/sched.h>
11#undef CREATE_TRACE_POINTS
12
13#include "sched.h"
14
15#include <linux/nospec.h>
16
17#include <linux/kcov.h>
18#include <linux/scs.h>
19
20#include <asm/switch_to.h>
21#include <asm/tlb.h>
22
23#include "../workqueue_internal.h"
24#include "../../fs/io-wq.h"
25#include "../smpboot.h"
26
27#include "pelt.h"
28#include "smp.h"
29
30/*
31 * Export tracepoints that act as a bare tracehook (ie: have no trace event
32 * associated with them) to allow external modules to probe them.
33 */
34EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
35EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
36EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
37EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
38EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
39EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
40EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
41EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
42EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
43
44DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
45
46#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL)
47/*
48 * Debugging: various feature bits
49 *
50 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
51 * sysctl_sched_features, defined in sched.h, to allow constants propagation
52 * at compile time and compiler optimization based on features default.
53 */
54#define SCHED_FEAT(name, enabled) \
55 (1UL << __SCHED_FEAT_##name) * enabled |
56const_debug unsigned int sysctl_sched_features =
57#include "features.h"
58 0;
59#undef SCHED_FEAT
60#endif
61
62/*
63 * Number of tasks to iterate in a single balance run.
64 * Limited because this is done with IRQs disabled.
65 */
66const_debug unsigned int sysctl_sched_nr_migrate = 32;
67
68/*
69 * period over which we measure -rt task CPU usage in us.
70 * default: 1s
71 */
72unsigned int sysctl_sched_rt_period = 1000000;
73
74__read_mostly int scheduler_running;
75
76/*
77 * part of the period that we allow rt tasks to run in us.
78 * default: 0.95s
79 */
80int sysctl_sched_rt_runtime = 950000;
81
82
83/*
84 * Serialization rules:
85 *
86 * Lock order:
87 *
88 * p->pi_lock
89 * rq->lock
90 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
91 *
92 * rq1->lock
93 * rq2->lock where: rq1 < rq2
94 *
95 * Regular state:
96 *
97 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
98 * local CPU's rq->lock, it optionally removes the task from the runqueue and
99 * always looks at the local rq data structures to find the most elegible task
100 * to run next.
101 *
102 * Task enqueue is also under rq->lock, possibly taken from another CPU.
103 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
104 * the local CPU to avoid bouncing the runqueue state around [ see
105 * ttwu_queue_wakelist() ]
106 *
107 * Task wakeup, specifically wakeups that involve migration, are horribly
108 * complicated to avoid having to take two rq->locks.
109 *
110 * Special state:
111 *
112 * System-calls and anything external will use task_rq_lock() which acquires
113 * both p->pi_lock and rq->lock. As a consequence the state they change is
114 * stable while holding either lock:
115 *
116 * - sched_setaffinity()/
117 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
118 * - set_user_nice(): p->se.load, p->*prio
119 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
120 * p->se.load, p->rt_priority,
121 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
122 * - sched_setnuma(): p->numa_preferred_nid
123 * - sched_move_task()/
124 * cpu_cgroup_fork(): p->sched_task_group
125 * - uclamp_update_active() p->uclamp*
126 *
127 * p->state <- TASK_*:
128 *
129 * is changed locklessly using set_current_state(), __set_current_state() or
130 * set_special_state(), see their respective comments, or by
131 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
132 * concurrent self.
133 *
134 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
135 *
136 * is set by activate_task() and cleared by deactivate_task(), under
137 * rq->lock. Non-zero indicates the task is runnable, the special
138 * ON_RQ_MIGRATING state is used for migration without holding both
139 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
140 *
141 * p->on_cpu <- { 0, 1 }:
142 *
143 * is set by prepare_task() and cleared by finish_task() such that it will be
144 * set before p is scheduled-in and cleared after p is scheduled-out, both
145 * under rq->lock. Non-zero indicates the task is running on its CPU.
146 *
147 * [ The astute reader will observe that it is possible for two tasks on one
148 * CPU to have ->on_cpu = 1 at the same time. ]
149 *
150 * task_cpu(p): is changed by set_task_cpu(), the rules are:
151 *
152 * - Don't call set_task_cpu() on a blocked task:
153 *
154 * We don't care what CPU we're not running on, this simplifies hotplug,
155 * the CPU assignment of blocked tasks isn't required to be valid.
156 *
157 * - for try_to_wake_up(), called under p->pi_lock:
158 *
159 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
160 *
161 * - for migration called under rq->lock:
162 * [ see task_on_rq_migrating() in task_rq_lock() ]
163 *
164 * o move_queued_task()
165 * o detach_task()
166 *
167 * - for migration called under double_rq_lock():
168 *
169 * o __migrate_swap_task()
170 * o push_rt_task() / pull_rt_task()
171 * o push_dl_task() / pull_dl_task()
172 * o dl_task_offline_migration()
173 *
174 */
175
176/*
177 * __task_rq_lock - lock the rq @p resides on.
178 */
179struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
180 __acquires(rq->lock)
181{
182 struct rq *rq;
183
184 lockdep_assert_held(&p->pi_lock);
185
186 for (;;) {
187 rq = task_rq(p);
188 raw_spin_lock(&rq->lock);
189 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
190 rq_pin_lock(rq, rf);
191 return rq;
192 }
193 raw_spin_unlock(&rq->lock);
194
195 while (unlikely(task_on_rq_migrating(p)))
196 cpu_relax();
197 }
198}
199
200/*
201 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
202 */
203struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
204 __acquires(p->pi_lock)
205 __acquires(rq->lock)
206{
207 struct rq *rq;
208
209 for (;;) {
210 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
211 rq = task_rq(p);
212 raw_spin_lock(&rq->lock);
213 /*
214 * move_queued_task() task_rq_lock()
215 *
216 * ACQUIRE (rq->lock)
217 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
218 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
219 * [S] ->cpu = new_cpu [L] task_rq()
220 * [L] ->on_rq
221 * RELEASE (rq->lock)
222 *
223 * If we observe the old CPU in task_rq_lock(), the acquire of
224 * the old rq->lock will fully serialize against the stores.
225 *
226 * If we observe the new CPU in task_rq_lock(), the address
227 * dependency headed by '[L] rq = task_rq()' and the acquire
228 * will pair with the WMB to ensure we then also see migrating.
229 */
230 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
231 rq_pin_lock(rq, rf);
232 return rq;
233 }
234 raw_spin_unlock(&rq->lock);
235 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
236
237 while (unlikely(task_on_rq_migrating(p)))
238 cpu_relax();
239 }
240}
241
242/*
243 * RQ-clock updating methods:
244 */
245
246static void update_rq_clock_task(struct rq *rq, s64 delta)
247{
248/*
249 * In theory, the compile should just see 0 here, and optimize out the call
250 * to sched_rt_avg_update. But I don't trust it...
251 */
252 s64 __maybe_unused steal = 0, irq_delta = 0;
253
254#ifdef CONFIG_IRQ_TIME_ACCOUNTING
255 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
256
257 /*
258 * Since irq_time is only updated on {soft,}irq_exit, we might run into
259 * this case when a previous update_rq_clock() happened inside a
260 * {soft,}irq region.
261 *
262 * When this happens, we stop ->clock_task and only update the
263 * prev_irq_time stamp to account for the part that fit, so that a next
264 * update will consume the rest. This ensures ->clock_task is
265 * monotonic.
266 *
267 * It does however cause some slight miss-attribution of {soft,}irq
268 * time, a more accurate solution would be to update the irq_time using
269 * the current rq->clock timestamp, except that would require using
270 * atomic ops.
271 */
272 if (irq_delta > delta)
273 irq_delta = delta;
274
275 rq->prev_irq_time += irq_delta;
276 delta -= irq_delta;
277#endif
278#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
279 if (static_key_false((¶virt_steal_rq_enabled))) {
280 steal = paravirt_steal_clock(cpu_of(rq));
281 steal -= rq->prev_steal_time_rq;
282
283 if (unlikely(steal > delta))
284 steal = delta;
285
286 rq->prev_steal_time_rq += steal;
287 delta -= steal;
288 }
289#endif
290
291 rq->clock_task += delta;
292
293#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
294 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
295 update_irq_load_avg(rq, irq_delta + steal);
296#endif
297 update_rq_clock_pelt(rq, delta);
298}
299
300void update_rq_clock(struct rq *rq)
301{
302 s64 delta;
303
304 lockdep_assert_held(&rq->lock);
305
306 if (rq->clock_update_flags & RQCF_ACT_SKIP)
307 return;
308
309#ifdef CONFIG_SCHED_DEBUG
310 if (sched_feat(WARN_DOUBLE_CLOCK))
311 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
312 rq->clock_update_flags |= RQCF_UPDATED;
313#endif
314
315 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
316 if (delta < 0)
317 return;
318 rq->clock += delta;
319 update_rq_clock_task(rq, delta);
320}
321
322static inline void
323rq_csd_init(struct rq *rq, call_single_data_t *csd, smp_call_func_t func)
324{
325 csd->flags = 0;
326 csd->func = func;
327 csd->info = rq;
328}
329
330#ifdef CONFIG_SCHED_HRTICK
331/*
332 * Use HR-timers to deliver accurate preemption points.
333 */
334
335static void hrtick_clear(struct rq *rq)
336{
337 if (hrtimer_active(&rq->hrtick_timer))
338 hrtimer_cancel(&rq->hrtick_timer);
339}
340
341/*
342 * High-resolution timer tick.
343 * Runs from hardirq context with interrupts disabled.
344 */
345static enum hrtimer_restart hrtick(struct hrtimer *timer)
346{
347 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
348 struct rq_flags rf;
349
350 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
351
352 rq_lock(rq, &rf);
353 update_rq_clock(rq);
354 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
355 rq_unlock(rq, &rf);
356
357 return HRTIMER_NORESTART;
358}
359
360#ifdef CONFIG_SMP
361
362static void __hrtick_restart(struct rq *rq)
363{
364 struct hrtimer *timer = &rq->hrtick_timer;
365
366 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED_HARD);
367}
368
369/*
370 * called from hardirq (IPI) context
371 */
372static void __hrtick_start(void *arg)
373{
374 struct rq *rq = arg;
375 struct rq_flags rf;
376
377 rq_lock(rq, &rf);
378 __hrtick_restart(rq);
379 rq_unlock(rq, &rf);
380}
381
382/*
383 * Called to set the hrtick timer state.
384 *
385 * called with rq->lock held and irqs disabled
386 */
387void hrtick_start(struct rq *rq, u64 delay)
388{
389 struct hrtimer *timer = &rq->hrtick_timer;
390 ktime_t time;
391 s64 delta;
392
393 /*
394 * Don't schedule slices shorter than 10000ns, that just
395 * doesn't make sense and can cause timer DoS.
396 */
397 delta = max_t(s64, delay, 10000LL);
398 time = ktime_add_ns(timer->base->get_time(), delta);
399
400 hrtimer_set_expires(timer, time);
401
402 if (rq == this_rq())
403 __hrtick_restart(rq);
404 else
405 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
406}
407
408#else
409/*
410 * Called to set the hrtick timer state.
411 *
412 * called with rq->lock held and irqs disabled
413 */
414void hrtick_start(struct rq *rq, u64 delay)
415{
416 /*
417 * Don't schedule slices shorter than 10000ns, that just
418 * doesn't make sense. Rely on vruntime for fairness.
419 */
420 delay = max_t(u64, delay, 10000LL);
421 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
422 HRTIMER_MODE_REL_PINNED_HARD);
423}
424
425#endif /* CONFIG_SMP */
426
427static void hrtick_rq_init(struct rq *rq)
428{
429#ifdef CONFIG_SMP
430 rq_csd_init(rq, &rq->hrtick_csd, __hrtick_start);
431#endif
432 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
433 rq->hrtick_timer.function = hrtick;
434}
435#else /* CONFIG_SCHED_HRTICK */
436static inline void hrtick_clear(struct rq *rq)
437{
438}
439
440static inline void hrtick_rq_init(struct rq *rq)
441{
442}
443#endif /* CONFIG_SCHED_HRTICK */
444
445/*
446 * cmpxchg based fetch_or, macro so it works for different integer types
447 */
448#define fetch_or(ptr, mask) \
449 ({ \
450 typeof(ptr) _ptr = (ptr); \
451 typeof(mask) _mask = (mask); \
452 typeof(*_ptr) _old, _val = *_ptr; \
453 \
454 for (;;) { \
455 _old = cmpxchg(_ptr, _val, _val | _mask); \
456 if (_old == _val) \
457 break; \
458 _val = _old; \
459 } \
460 _old; \
461})
462
463#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
464/*
465 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
466 * this avoids any races wrt polling state changes and thereby avoids
467 * spurious IPIs.
468 */
469static bool set_nr_and_not_polling(struct task_struct *p)
470{
471 struct thread_info *ti = task_thread_info(p);
472 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
473}
474
475/*
476 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
477 *
478 * If this returns true, then the idle task promises to call
479 * sched_ttwu_pending() and reschedule soon.
480 */
481static bool set_nr_if_polling(struct task_struct *p)
482{
483 struct thread_info *ti = task_thread_info(p);
484 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
485
486 for (;;) {
487 if (!(val & _TIF_POLLING_NRFLAG))
488 return false;
489 if (val & _TIF_NEED_RESCHED)
490 return true;
491 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
492 if (old == val)
493 break;
494 val = old;
495 }
496 return true;
497}
498
499#else
500static bool set_nr_and_not_polling(struct task_struct *p)
501{
502 set_tsk_need_resched(p);
503 return true;
504}
505
506#ifdef CONFIG_SMP
507static bool set_nr_if_polling(struct task_struct *p)
508{
509 return false;
510}
511#endif
512#endif
513
514static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
515{
516 struct wake_q_node *node = &task->wake_q;
517
518 /*
519 * Atomically grab the task, if ->wake_q is !nil already it means
520 * its already queued (either by us or someone else) and will get the
521 * wakeup due to that.
522 *
523 * In order to ensure that a pending wakeup will observe our pending
524 * state, even in the failed case, an explicit smp_mb() must be used.
525 */
526 smp_mb__before_atomic();
527 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
528 return false;
529
530 /*
531 * The head is context local, there can be no concurrency.
532 */
533 *head->lastp = node;
534 head->lastp = &node->next;
535 return true;
536}
537
538/**
539 * wake_q_add() - queue a wakeup for 'later' waking.
540 * @head: the wake_q_head to add @task to
541 * @task: the task to queue for 'later' wakeup
542 *
543 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
544 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
545 * instantly.
546 *
547 * This function must be used as-if it were wake_up_process(); IOW the task
548 * must be ready to be woken at this location.
549 */
550void wake_q_add(struct wake_q_head *head, struct task_struct *task)
551{
552 if (__wake_q_add(head, task))
553 get_task_struct(task);
554}
555
556/**
557 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
558 * @head: the wake_q_head to add @task to
559 * @task: the task to queue for 'later' wakeup
560 *
561 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
562 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
563 * instantly.
564 *
565 * This function must be used as-if it were wake_up_process(); IOW the task
566 * must be ready to be woken at this location.
567 *
568 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
569 * that already hold reference to @task can call the 'safe' version and trust
570 * wake_q to do the right thing depending whether or not the @task is already
571 * queued for wakeup.
572 */
573void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
574{
575 if (!__wake_q_add(head, task))
576 put_task_struct(task);
577}
578
579void wake_up_q(struct wake_q_head *head)
580{
581 struct wake_q_node *node = head->first;
582
583 while (node != WAKE_Q_TAIL) {
584 struct task_struct *task;
585
586 task = container_of(node, struct task_struct, wake_q);
587 BUG_ON(!task);
588 /* Task can safely be re-inserted now: */
589 node = node->next;
590 task->wake_q.next = NULL;
591
592 /*
593 * wake_up_process() executes a full barrier, which pairs with
594 * the queueing in wake_q_add() so as not to miss wakeups.
595 */
596 wake_up_process(task);
597 put_task_struct(task);
598 }
599}
600
601/*
602 * resched_curr - mark rq's current task 'to be rescheduled now'.
603 *
604 * On UP this means the setting of the need_resched flag, on SMP it
605 * might also involve a cross-CPU call to trigger the scheduler on
606 * the target CPU.
607 */
608void resched_curr(struct rq *rq)
609{
610 struct task_struct *curr = rq->curr;
611 int cpu;
612
613 lockdep_assert_held(&rq->lock);
614
615 if (test_tsk_need_resched(curr))
616 return;
617
618 cpu = cpu_of(rq);
619
620 if (cpu == smp_processor_id()) {
621 set_tsk_need_resched(curr);
622 set_preempt_need_resched();
623 return;
624 }
625
626 if (set_nr_and_not_polling(curr))
627 smp_send_reschedule(cpu);
628 else
629 trace_sched_wake_idle_without_ipi(cpu);
630}
631
632void resched_cpu(int cpu)
633{
634 struct rq *rq = cpu_rq(cpu);
635 unsigned long flags;
636
637 raw_spin_lock_irqsave(&rq->lock, flags);
638 if (cpu_online(cpu) || cpu == smp_processor_id())
639 resched_curr(rq);
640 raw_spin_unlock_irqrestore(&rq->lock, flags);
641}
642
643#ifdef CONFIG_SMP
644#ifdef CONFIG_NO_HZ_COMMON
645/*
646 * In the semi idle case, use the nearest busy CPU for migrating timers
647 * from an idle CPU. This is good for power-savings.
648 *
649 * We don't do similar optimization for completely idle system, as
650 * selecting an idle CPU will add more delays to the timers than intended
651 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
652 */
653int get_nohz_timer_target(void)
654{
655 int i, cpu = smp_processor_id(), default_cpu = -1;
656 struct sched_domain *sd;
657
658 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
659 if (!idle_cpu(cpu))
660 return cpu;
661 default_cpu = cpu;
662 }
663
664 rcu_read_lock();
665 for_each_domain(cpu, sd) {
666 for_each_cpu_and(i, sched_domain_span(sd),
667 housekeeping_cpumask(HK_FLAG_TIMER)) {
668 if (cpu == i)
669 continue;
670
671 if (!idle_cpu(i)) {
672 cpu = i;
673 goto unlock;
674 }
675 }
676 }
677
678 if (default_cpu == -1)
679 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
680 cpu = default_cpu;
681unlock:
682 rcu_read_unlock();
683 return cpu;
684}
685
686/*
687 * When add_timer_on() enqueues a timer into the timer wheel of an
688 * idle CPU then this timer might expire before the next timer event
689 * which is scheduled to wake up that CPU. In case of a completely
690 * idle system the next event might even be infinite time into the
691 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
692 * leaves the inner idle loop so the newly added timer is taken into
693 * account when the CPU goes back to idle and evaluates the timer
694 * wheel for the next timer event.
695 */
696static void wake_up_idle_cpu(int cpu)
697{
698 struct rq *rq = cpu_rq(cpu);
699
700 if (cpu == smp_processor_id())
701 return;
702
703 if (set_nr_and_not_polling(rq->idle))
704 smp_send_reschedule(cpu);
705 else
706 trace_sched_wake_idle_without_ipi(cpu);
707}
708
709static bool wake_up_full_nohz_cpu(int cpu)
710{
711 /*
712 * We just need the target to call irq_exit() and re-evaluate
713 * the next tick. The nohz full kick at least implies that.
714 * If needed we can still optimize that later with an
715 * empty IRQ.
716 */
717 if (cpu_is_offline(cpu))
718 return true; /* Don't try to wake offline CPUs. */
719 if (tick_nohz_full_cpu(cpu)) {
720 if (cpu != smp_processor_id() ||
721 tick_nohz_tick_stopped())
722 tick_nohz_full_kick_cpu(cpu);
723 return true;
724 }
725
726 return false;
727}
728
729/*
730 * Wake up the specified CPU. If the CPU is going offline, it is the
731 * caller's responsibility to deal with the lost wakeup, for example,
732 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
733 */
734void wake_up_nohz_cpu(int cpu)
735{
736 if (!wake_up_full_nohz_cpu(cpu))
737 wake_up_idle_cpu(cpu);
738}
739
740static void nohz_csd_func(void *info)
741{
742 struct rq *rq = info;
743 int cpu = cpu_of(rq);
744 unsigned int flags;
745
746 /*
747 * Release the rq::nohz_csd.
748 */
749 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
750 WARN_ON(!(flags & NOHZ_KICK_MASK));
751
752 rq->idle_balance = idle_cpu(cpu);
753 if (rq->idle_balance && !need_resched()) {
754 rq->nohz_idle_balance = flags;
755 raise_softirq_irqoff(SCHED_SOFTIRQ);
756 }
757}
758
759#endif /* CONFIG_NO_HZ_COMMON */
760
761#ifdef CONFIG_NO_HZ_FULL
762bool sched_can_stop_tick(struct rq *rq)
763{
764 int fifo_nr_running;
765
766 /* Deadline tasks, even if single, need the tick */
767 if (rq->dl.dl_nr_running)
768 return false;
769
770 /*
771 * If there are more than one RR tasks, we need the tick to effect the
772 * actual RR behaviour.
773 */
774 if (rq->rt.rr_nr_running) {
775 if (rq->rt.rr_nr_running == 1)
776 return true;
777 else
778 return false;
779 }
780
781 /*
782 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
783 * forced preemption between FIFO tasks.
784 */
785 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
786 if (fifo_nr_running)
787 return true;
788
789 /*
790 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
791 * if there's more than one we need the tick for involuntary
792 * preemption.
793 */
794 if (rq->nr_running > 1)
795 return false;
796
797 return true;
798}
799#endif /* CONFIG_NO_HZ_FULL */
800#endif /* CONFIG_SMP */
801
802#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
803 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
804/*
805 * Iterate task_group tree rooted at *from, calling @down when first entering a
806 * node and @up when leaving it for the final time.
807 *
808 * Caller must hold rcu_lock or sufficient equivalent.
809 */
810int walk_tg_tree_from(struct task_group *from,
811 tg_visitor down, tg_visitor up, void *data)
812{
813 struct task_group *parent, *child;
814 int ret;
815
816 parent = from;
817
818down:
819 ret = (*down)(parent, data);
820 if (ret)
821 goto out;
822 list_for_each_entry_rcu(child, &parent->children, siblings) {
823 parent = child;
824 goto down;
825
826up:
827 continue;
828 }
829 ret = (*up)(parent, data);
830 if (ret || parent == from)
831 goto out;
832
833 child = parent;
834 parent = parent->parent;
835 if (parent)
836 goto up;
837out:
838 return ret;
839}
840
841int tg_nop(struct task_group *tg, void *data)
842{
843 return 0;
844}
845#endif
846
847static void set_load_weight(struct task_struct *p, bool update_load)
848{
849 int prio = p->static_prio - MAX_RT_PRIO;
850 struct load_weight *load = &p->se.load;
851
852 /*
853 * SCHED_IDLE tasks get minimal weight:
854 */
855 if (task_has_idle_policy(p)) {
856 load->weight = scale_load(WEIGHT_IDLEPRIO);
857 load->inv_weight = WMULT_IDLEPRIO;
858 return;
859 }
860
861 /*
862 * SCHED_OTHER tasks have to update their load when changing their
863 * weight
864 */
865 if (update_load && p->sched_class == &fair_sched_class) {
866 reweight_task(p, prio);
867 } else {
868 load->weight = scale_load(sched_prio_to_weight[prio]);
869 load->inv_weight = sched_prio_to_wmult[prio];
870 }
871}
872
873#ifdef CONFIG_UCLAMP_TASK
874/*
875 * Serializes updates of utilization clamp values
876 *
877 * The (slow-path) user-space triggers utilization clamp value updates which
878 * can require updates on (fast-path) scheduler's data structures used to
879 * support enqueue/dequeue operations.
880 * While the per-CPU rq lock protects fast-path update operations, user-space
881 * requests are serialized using a mutex to reduce the risk of conflicting
882 * updates or API abuses.
883 */
884static DEFINE_MUTEX(uclamp_mutex);
885
886/* Max allowed minimum utilization */
887unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
888
889/* Max allowed maximum utilization */
890unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
891
892/*
893 * By default RT tasks run at the maximum performance point/capacity of the
894 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
895 * SCHED_CAPACITY_SCALE.
896 *
897 * This knob allows admins to change the default behavior when uclamp is being
898 * used. In battery powered devices, particularly, running at the maximum
899 * capacity and frequency will increase energy consumption and shorten the
900 * battery life.
901 *
902 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
903 *
904 * This knob will not override the system default sched_util_clamp_min defined
905 * above.
906 */
907unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
908
909/* All clamps are required to be less or equal than these values */
910static struct uclamp_se uclamp_default[UCLAMP_CNT];
911
912/*
913 * This static key is used to reduce the uclamp overhead in the fast path. It
914 * primarily disables the call to uclamp_rq_{inc, dec}() in
915 * enqueue/dequeue_task().
916 *
917 * This allows users to continue to enable uclamp in their kernel config with
918 * minimum uclamp overhead in the fast path.
919 *
920 * As soon as userspace modifies any of the uclamp knobs, the static key is
921 * enabled, since we have an actual users that make use of uclamp
922 * functionality.
923 *
924 * The knobs that would enable this static key are:
925 *
926 * * A task modifying its uclamp value with sched_setattr().
927 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
928 * * An admin modifying the cgroup cpu.uclamp.{min, max}
929 */
930DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
931
932/* Integer rounded range for each bucket */
933#define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
934
935#define for_each_clamp_id(clamp_id) \
936 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
937
938static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
939{
940 return clamp_value / UCLAMP_BUCKET_DELTA;
941}
942
943static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value)
944{
945 return UCLAMP_BUCKET_DELTA * uclamp_bucket_id(clamp_value);
946}
947
948static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
949{
950 if (clamp_id == UCLAMP_MIN)
951 return 0;
952 return SCHED_CAPACITY_SCALE;
953}
954
955static inline void uclamp_se_set(struct uclamp_se *uc_se,
956 unsigned int value, bool user_defined)
957{
958 uc_se->value = value;
959 uc_se->bucket_id = uclamp_bucket_id(value);
960 uc_se->user_defined = user_defined;
961}
962
963static inline unsigned int
964uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
965 unsigned int clamp_value)
966{
967 /*
968 * Avoid blocked utilization pushing up the frequency when we go
969 * idle (which drops the max-clamp) by retaining the last known
970 * max-clamp.
971 */
972 if (clamp_id == UCLAMP_MAX) {
973 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
974 return clamp_value;
975 }
976
977 return uclamp_none(UCLAMP_MIN);
978}
979
980static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
981 unsigned int clamp_value)
982{
983 /* Reset max-clamp retention only on idle exit */
984 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
985 return;
986
987 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
988}
989
990static inline
991unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
992 unsigned int clamp_value)
993{
994 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
995 int bucket_id = UCLAMP_BUCKETS - 1;
996
997 /*
998 * Since both min and max clamps are max aggregated, find the
999 * top most bucket with tasks in.
1000 */
1001 for ( ; bucket_id >= 0; bucket_id--) {
1002 if (!bucket[bucket_id].tasks)
1003 continue;
1004 return bucket[bucket_id].value;
1005 }
1006
1007 /* No tasks -- default clamp values */
1008 return uclamp_idle_value(rq, clamp_id, clamp_value);
1009}
1010
1011static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1012{
1013 unsigned int default_util_min;
1014 struct uclamp_se *uc_se;
1015
1016 lockdep_assert_held(&p->pi_lock);
1017
1018 uc_se = &p->uclamp_req[UCLAMP_MIN];
1019
1020 /* Only sync if user didn't override the default */
1021 if (uc_se->user_defined)
1022 return;
1023
1024 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1025 uclamp_se_set(uc_se, default_util_min, false);
1026}
1027
1028static void uclamp_update_util_min_rt_default(struct task_struct *p)
1029{
1030 struct rq_flags rf;
1031 struct rq *rq;
1032
1033 if (!rt_task(p))
1034 return;
1035
1036 /* Protect updates to p->uclamp_* */
1037 rq = task_rq_lock(p, &rf);
1038 __uclamp_update_util_min_rt_default(p);
1039 task_rq_unlock(rq, p, &rf);
1040}
1041
1042static void uclamp_sync_util_min_rt_default(void)
1043{
1044 struct task_struct *g, *p;
1045
1046 /*
1047 * copy_process() sysctl_uclamp
1048 * uclamp_min_rt = X;
1049 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1050 * // link thread smp_mb__after_spinlock()
1051 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1052 * sched_post_fork() for_each_process_thread()
1053 * __uclamp_sync_rt() __uclamp_sync_rt()
1054 *
1055 * Ensures that either sched_post_fork() will observe the new
1056 * uclamp_min_rt or for_each_process_thread() will observe the new
1057 * task.
1058 */
1059 read_lock(&tasklist_lock);
1060 smp_mb__after_spinlock();
1061 read_unlock(&tasklist_lock);
1062
1063 rcu_read_lock();
1064 for_each_process_thread(g, p)
1065 uclamp_update_util_min_rt_default(p);
1066 rcu_read_unlock();
1067}
1068
1069static inline struct uclamp_se
1070uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1071{
1072 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1073#ifdef CONFIG_UCLAMP_TASK_GROUP
1074 struct uclamp_se uc_max;
1075
1076 /*
1077 * Tasks in autogroups or root task group will be
1078 * restricted by system defaults.
1079 */
1080 if (task_group_is_autogroup(task_group(p)))
1081 return uc_req;
1082 if (task_group(p) == &root_task_group)
1083 return uc_req;
1084
1085 uc_max = task_group(p)->uclamp[clamp_id];
1086 if (uc_req.value > uc_max.value || !uc_req.user_defined)
1087 return uc_max;
1088#endif
1089
1090 return uc_req;
1091}
1092
1093/*
1094 * The effective clamp bucket index of a task depends on, by increasing
1095 * priority:
1096 * - the task specific clamp value, when explicitly requested from userspace
1097 * - the task group effective clamp value, for tasks not either in the root
1098 * group or in an autogroup
1099 * - the system default clamp value, defined by the sysadmin
1100 */
1101static inline struct uclamp_se
1102uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1103{
1104 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1105 struct uclamp_se uc_max = uclamp_default[clamp_id];
1106
1107 /* System default restrictions always apply */
1108 if (unlikely(uc_req.value > uc_max.value))
1109 return uc_max;
1110
1111 return uc_req;
1112}
1113
1114unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1115{
1116 struct uclamp_se uc_eff;
1117
1118 /* Task currently refcounted: use back-annotated (effective) value */
1119 if (p->uclamp[clamp_id].active)
1120 return (unsigned long)p->uclamp[clamp_id].value;
1121
1122 uc_eff = uclamp_eff_get(p, clamp_id);
1123
1124 return (unsigned long)uc_eff.value;
1125}
1126
1127/*
1128 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1129 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1130 * updates the rq's clamp value if required.
1131 *
1132 * Tasks can have a task-specific value requested from user-space, track
1133 * within each bucket the maximum value for tasks refcounted in it.
1134 * This "local max aggregation" allows to track the exact "requested" value
1135 * for each bucket when all its RUNNABLE tasks require the same clamp.
1136 */
1137static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1138 enum uclamp_id clamp_id)
1139{
1140 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1141 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1142 struct uclamp_bucket *bucket;
1143
1144 lockdep_assert_held(&rq->lock);
1145
1146 /* Update task effective clamp */
1147 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1148
1149 bucket = &uc_rq->bucket[uc_se->bucket_id];
1150 bucket->tasks++;
1151 uc_se->active = true;
1152
1153 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1154
1155 /*
1156 * Local max aggregation: rq buckets always track the max
1157 * "requested" clamp value of its RUNNABLE tasks.
1158 */
1159 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1160 bucket->value = uc_se->value;
1161
1162 if (uc_se->value > READ_ONCE(uc_rq->value))
1163 WRITE_ONCE(uc_rq->value, uc_se->value);
1164}
1165
1166/*
1167 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1168 * is released. If this is the last task reference counting the rq's max
1169 * active clamp value, then the rq's clamp value is updated.
1170 *
1171 * Both refcounted tasks and rq's cached clamp values are expected to be
1172 * always valid. If it's detected they are not, as defensive programming,
1173 * enforce the expected state and warn.
1174 */
1175static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1176 enum uclamp_id clamp_id)
1177{
1178 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1179 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1180 struct uclamp_bucket *bucket;
1181 unsigned int bkt_clamp;
1182 unsigned int rq_clamp;
1183
1184 lockdep_assert_held(&rq->lock);
1185
1186 /*
1187 * If sched_uclamp_used was enabled after task @p was enqueued,
1188 * we could end up with unbalanced call to uclamp_rq_dec_id().
1189 *
1190 * In this case the uc_se->active flag should be false since no uclamp
1191 * accounting was performed at enqueue time and we can just return
1192 * here.
1193 *
1194 * Need to be careful of the following enqeueue/dequeue ordering
1195 * problem too
1196 *
1197 * enqueue(taskA)
1198 * // sched_uclamp_used gets enabled
1199 * enqueue(taskB)
1200 * dequeue(taskA)
1201 * // Must not decrement bukcet->tasks here
1202 * dequeue(taskB)
1203 *
1204 * where we could end up with stale data in uc_se and
1205 * bucket[uc_se->bucket_id].
1206 *
1207 * The following check here eliminates the possibility of such race.
1208 */
1209 if (unlikely(!uc_se->active))
1210 return;
1211
1212 bucket = &uc_rq->bucket[uc_se->bucket_id];
1213
1214 SCHED_WARN_ON(!bucket->tasks);
1215 if (likely(bucket->tasks))
1216 bucket->tasks--;
1217
1218 uc_se->active = false;
1219
1220 /*
1221 * Keep "local max aggregation" simple and accept to (possibly)
1222 * overboost some RUNNABLE tasks in the same bucket.
1223 * The rq clamp bucket value is reset to its base value whenever
1224 * there are no more RUNNABLE tasks refcounting it.
1225 */
1226 if (likely(bucket->tasks))
1227 return;
1228
1229 rq_clamp = READ_ONCE(uc_rq->value);
1230 /*
1231 * Defensive programming: this should never happen. If it happens,
1232 * e.g. due to future modification, warn and fixup the expected value.
1233 */
1234 SCHED_WARN_ON(bucket->value > rq_clamp);
1235 if (bucket->value >= rq_clamp) {
1236 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1237 WRITE_ONCE(uc_rq->value, bkt_clamp);
1238 }
1239}
1240
1241static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1242{
1243 enum uclamp_id clamp_id;
1244
1245 /*
1246 * Avoid any overhead until uclamp is actually used by the userspace.
1247 *
1248 * The condition is constructed such that a NOP is generated when
1249 * sched_uclamp_used is disabled.
1250 */
1251 if (!static_branch_unlikely(&sched_uclamp_used))
1252 return;
1253
1254 if (unlikely(!p->sched_class->uclamp_enabled))
1255 return;
1256
1257 for_each_clamp_id(clamp_id)
1258 uclamp_rq_inc_id(rq, p, clamp_id);
1259
1260 /* Reset clamp idle holding when there is one RUNNABLE task */
1261 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1262 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1263}
1264
1265static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1266{
1267 enum uclamp_id clamp_id;
1268
1269 /*
1270 * Avoid any overhead until uclamp is actually used by the userspace.
1271 *
1272 * The condition is constructed such that a NOP is generated when
1273 * sched_uclamp_used is disabled.
1274 */
1275 if (!static_branch_unlikely(&sched_uclamp_used))
1276 return;
1277
1278 if (unlikely(!p->sched_class->uclamp_enabled))
1279 return;
1280
1281 for_each_clamp_id(clamp_id)
1282 uclamp_rq_dec_id(rq, p, clamp_id);
1283}
1284
1285static inline void
1286uclamp_update_active(struct task_struct *p, enum uclamp_id clamp_id)
1287{
1288 struct rq_flags rf;
1289 struct rq *rq;
1290
1291 /*
1292 * Lock the task and the rq where the task is (or was) queued.
1293 *
1294 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1295 * price to pay to safely serialize util_{min,max} updates with
1296 * enqueues, dequeues and migration operations.
1297 * This is the same locking schema used by __set_cpus_allowed_ptr().
1298 */
1299 rq = task_rq_lock(p, &rf);
1300
1301 /*
1302 * Setting the clamp bucket is serialized by task_rq_lock().
1303 * If the task is not yet RUNNABLE and its task_struct is not
1304 * affecting a valid clamp bucket, the next time it's enqueued,
1305 * it will already see the updated clamp bucket value.
1306 */
1307 if (p->uclamp[clamp_id].active) {
1308 uclamp_rq_dec_id(rq, p, clamp_id);
1309 uclamp_rq_inc_id(rq, p, clamp_id);
1310 }
1311
1312 task_rq_unlock(rq, p, &rf);
1313}
1314
1315#ifdef CONFIG_UCLAMP_TASK_GROUP
1316static inline void
1317uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1318 unsigned int clamps)
1319{
1320 enum uclamp_id clamp_id;
1321 struct css_task_iter it;
1322 struct task_struct *p;
1323
1324 css_task_iter_start(css, 0, &it);
1325 while ((p = css_task_iter_next(&it))) {
1326 for_each_clamp_id(clamp_id) {
1327 if ((0x1 << clamp_id) & clamps)
1328 uclamp_update_active(p, clamp_id);
1329 }
1330 }
1331 css_task_iter_end(&it);
1332}
1333
1334static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1335static void uclamp_update_root_tg(void)
1336{
1337 struct task_group *tg = &root_task_group;
1338
1339 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1340 sysctl_sched_uclamp_util_min, false);
1341 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1342 sysctl_sched_uclamp_util_max, false);
1343
1344 rcu_read_lock();
1345 cpu_util_update_eff(&root_task_group.css);
1346 rcu_read_unlock();
1347}
1348#else
1349static void uclamp_update_root_tg(void) { }
1350#endif
1351
1352int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1353 void *buffer, size_t *lenp, loff_t *ppos)
1354{
1355 bool update_root_tg = false;
1356 int old_min, old_max, old_min_rt;
1357 int result;
1358
1359 mutex_lock(&uclamp_mutex);
1360 old_min = sysctl_sched_uclamp_util_min;
1361 old_max = sysctl_sched_uclamp_util_max;
1362 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1363
1364 result = proc_dointvec(table, write, buffer, lenp, ppos);
1365 if (result)
1366 goto undo;
1367 if (!write)
1368 goto done;
1369
1370 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1371 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1372 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1373
1374 result = -EINVAL;
1375 goto undo;
1376 }
1377
1378 if (old_min != sysctl_sched_uclamp_util_min) {
1379 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1380 sysctl_sched_uclamp_util_min, false);
1381 update_root_tg = true;
1382 }
1383 if (old_max != sysctl_sched_uclamp_util_max) {
1384 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1385 sysctl_sched_uclamp_util_max, false);
1386 update_root_tg = true;
1387 }
1388
1389 if (update_root_tg) {
1390 static_branch_enable(&sched_uclamp_used);
1391 uclamp_update_root_tg();
1392 }
1393
1394 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1395 static_branch_enable(&sched_uclamp_used);
1396 uclamp_sync_util_min_rt_default();
1397 }
1398
1399 /*
1400 * We update all RUNNABLE tasks only when task groups are in use.
1401 * Otherwise, keep it simple and do just a lazy update at each next
1402 * task enqueue time.
1403 */
1404
1405 goto done;
1406
1407undo:
1408 sysctl_sched_uclamp_util_min = old_min;
1409 sysctl_sched_uclamp_util_max = old_max;
1410 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1411done:
1412 mutex_unlock(&uclamp_mutex);
1413
1414 return result;
1415}
1416
1417static int uclamp_validate(struct task_struct *p,
1418 const struct sched_attr *attr)
1419{
1420 unsigned int lower_bound = p->uclamp_req[UCLAMP_MIN].value;
1421 unsigned int upper_bound = p->uclamp_req[UCLAMP_MAX].value;
1422
1423 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN)
1424 lower_bound = attr->sched_util_min;
1425 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX)
1426 upper_bound = attr->sched_util_max;
1427
1428 if (lower_bound > upper_bound)
1429 return -EINVAL;
1430 if (upper_bound > SCHED_CAPACITY_SCALE)
1431 return -EINVAL;
1432
1433 /*
1434 * We have valid uclamp attributes; make sure uclamp is enabled.
1435 *
1436 * We need to do that here, because enabling static branches is a
1437 * blocking operation which obviously cannot be done while holding
1438 * scheduler locks.
1439 */
1440 static_branch_enable(&sched_uclamp_used);
1441
1442 return 0;
1443}
1444
1445static void __setscheduler_uclamp(struct task_struct *p,
1446 const struct sched_attr *attr)
1447{
1448 enum uclamp_id clamp_id;
1449
1450 /*
1451 * On scheduling class change, reset to default clamps for tasks
1452 * without a task-specific value.
1453 */
1454 for_each_clamp_id(clamp_id) {
1455 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1456
1457 /* Keep using defined clamps across class changes */
1458 if (uc_se->user_defined)
1459 continue;
1460
1461 /*
1462 * RT by default have a 100% boost value that could be modified
1463 * at runtime.
1464 */
1465 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1466 __uclamp_update_util_min_rt_default(p);
1467 else
1468 uclamp_se_set(uc_se, uclamp_none(clamp_id), false);
1469
1470 }
1471
1472 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1473 return;
1474
1475 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1476 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1477 attr->sched_util_min, true);
1478 }
1479
1480 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1481 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1482 attr->sched_util_max, true);
1483 }
1484}
1485
1486static void uclamp_fork(struct task_struct *p)
1487{
1488 enum uclamp_id clamp_id;
1489
1490 /*
1491 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1492 * as the task is still at its early fork stages.
1493 */
1494 for_each_clamp_id(clamp_id)
1495 p->uclamp[clamp_id].active = false;
1496
1497 if (likely(!p->sched_reset_on_fork))
1498 return;
1499
1500 for_each_clamp_id(clamp_id) {
1501 uclamp_se_set(&p->uclamp_req[clamp_id],
1502 uclamp_none(clamp_id), false);
1503 }
1504}
1505
1506static void uclamp_post_fork(struct task_struct *p)
1507{
1508 uclamp_update_util_min_rt_default(p);
1509}
1510
1511static void __init init_uclamp_rq(struct rq *rq)
1512{
1513 enum uclamp_id clamp_id;
1514 struct uclamp_rq *uc_rq = rq->uclamp;
1515
1516 for_each_clamp_id(clamp_id) {
1517 uc_rq[clamp_id] = (struct uclamp_rq) {
1518 .value = uclamp_none(clamp_id)
1519 };
1520 }
1521
1522 rq->uclamp_flags = 0;
1523}
1524
1525static void __init init_uclamp(void)
1526{
1527 struct uclamp_se uc_max = {};
1528 enum uclamp_id clamp_id;
1529 int cpu;
1530
1531 for_each_possible_cpu(cpu)
1532 init_uclamp_rq(cpu_rq(cpu));
1533
1534 for_each_clamp_id(clamp_id) {
1535 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1536 uclamp_none(clamp_id), false);
1537 }
1538
1539 /* System defaults allow max clamp values for both indexes */
1540 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1541 for_each_clamp_id(clamp_id) {
1542 uclamp_default[clamp_id] = uc_max;
1543#ifdef CONFIG_UCLAMP_TASK_GROUP
1544 root_task_group.uclamp_req[clamp_id] = uc_max;
1545 root_task_group.uclamp[clamp_id] = uc_max;
1546#endif
1547 }
1548}
1549
1550#else /* CONFIG_UCLAMP_TASK */
1551static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1552static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1553static inline int uclamp_validate(struct task_struct *p,
1554 const struct sched_attr *attr)
1555{
1556 return -EOPNOTSUPP;
1557}
1558static void __setscheduler_uclamp(struct task_struct *p,
1559 const struct sched_attr *attr) { }
1560static inline void uclamp_fork(struct task_struct *p) { }
1561static inline void uclamp_post_fork(struct task_struct *p) { }
1562static inline void init_uclamp(void) { }
1563#endif /* CONFIG_UCLAMP_TASK */
1564
1565static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1566{
1567 if (!(flags & ENQUEUE_NOCLOCK))
1568 update_rq_clock(rq);
1569
1570 if (!(flags & ENQUEUE_RESTORE)) {
1571 sched_info_queued(rq, p);
1572 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1573 }
1574
1575 uclamp_rq_inc(rq, p);
1576 p->sched_class->enqueue_task(rq, p, flags);
1577}
1578
1579static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1580{
1581 if (!(flags & DEQUEUE_NOCLOCK))
1582 update_rq_clock(rq);
1583
1584 if (!(flags & DEQUEUE_SAVE)) {
1585 sched_info_dequeued(rq, p);
1586 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1587 }
1588
1589 uclamp_rq_dec(rq, p);
1590 p->sched_class->dequeue_task(rq, p, flags);
1591}
1592
1593void activate_task(struct rq *rq, struct task_struct *p, int flags)
1594{
1595 enqueue_task(rq, p, flags);
1596
1597 p->on_rq = TASK_ON_RQ_QUEUED;
1598}
1599
1600void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1601{
1602 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1603
1604 dequeue_task(rq, p, flags);
1605}
1606
1607/*
1608 * __normal_prio - return the priority that is based on the static prio
1609 */
1610static inline int __normal_prio(struct task_struct *p)
1611{
1612 return p->static_prio;
1613}
1614
1615/*
1616 * Calculate the expected normal priority: i.e. priority
1617 * without taking RT-inheritance into account. Might be
1618 * boosted by interactivity modifiers. Changes upon fork,
1619 * setprio syscalls, and whenever the interactivity
1620 * estimator recalculates.
1621 */
1622static inline int normal_prio(struct task_struct *p)
1623{
1624 int prio;
1625
1626 if (task_has_dl_policy(p))
1627 prio = MAX_DL_PRIO-1;
1628 else if (task_has_rt_policy(p))
1629 prio = MAX_RT_PRIO-1 - p->rt_priority;
1630 else
1631 prio = __normal_prio(p);
1632 return prio;
1633}
1634
1635/*
1636 * Calculate the current priority, i.e. the priority
1637 * taken into account by the scheduler. This value might
1638 * be boosted by RT tasks, or might be boosted by
1639 * interactivity modifiers. Will be RT if the task got
1640 * RT-boosted. If not then it returns p->normal_prio.
1641 */
1642static int effective_prio(struct task_struct *p)
1643{
1644 p->normal_prio = normal_prio(p);
1645 /*
1646 * If we are RT tasks or we were boosted to RT priority,
1647 * keep the priority unchanged. Otherwise, update priority
1648 * to the normal priority:
1649 */
1650 if (!rt_prio(p->prio))
1651 return p->normal_prio;
1652 return p->prio;
1653}
1654
1655/**
1656 * task_curr - is this task currently executing on a CPU?
1657 * @p: the task in question.
1658 *
1659 * Return: 1 if the task is currently executing. 0 otherwise.
1660 */
1661inline int task_curr(const struct task_struct *p)
1662{
1663 return cpu_curr(task_cpu(p)) == p;
1664}
1665
1666/*
1667 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1668 * use the balance_callback list if you want balancing.
1669 *
1670 * this means any call to check_class_changed() must be followed by a call to
1671 * balance_callback().
1672 */
1673static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1674 const struct sched_class *prev_class,
1675 int oldprio)
1676{
1677 if (prev_class != p->sched_class) {
1678 if (prev_class->switched_from)
1679 prev_class->switched_from(rq, p);
1680
1681 p->sched_class->switched_to(rq, p);
1682 } else if (oldprio != p->prio || dl_task(p))
1683 p->sched_class->prio_changed(rq, p, oldprio);
1684}
1685
1686void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1687{
1688 if (p->sched_class == rq->curr->sched_class)
1689 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1690 else if (p->sched_class > rq->curr->sched_class)
1691 resched_curr(rq);
1692
1693 /*
1694 * A queue event has occurred, and we're going to schedule. In
1695 * this case, we can save a useless back to back clock update.
1696 */
1697 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1698 rq_clock_skip_update(rq);
1699}
1700
1701#ifdef CONFIG_SMP
1702
1703/*
1704 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1705 * __set_cpus_allowed_ptr() and select_fallback_rq().
1706 */
1707static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1708{
1709 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1710 return false;
1711
1712 if (is_per_cpu_kthread(p))
1713 return cpu_online(cpu);
1714
1715 return cpu_active(cpu);
1716}
1717
1718/*
1719 * This is how migration works:
1720 *
1721 * 1) we invoke migration_cpu_stop() on the target CPU using
1722 * stop_one_cpu().
1723 * 2) stopper starts to run (implicitly forcing the migrated thread
1724 * off the CPU)
1725 * 3) it checks whether the migrated task is still in the wrong runqueue.
1726 * 4) if it's in the wrong runqueue then the migration thread removes
1727 * it and puts it into the right queue.
1728 * 5) stopper completes and stop_one_cpu() returns and the migration
1729 * is done.
1730 */
1731
1732/*
1733 * move_queued_task - move a queued task to new rq.
1734 *
1735 * Returns (locked) new rq. Old rq's lock is released.
1736 */
1737static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1738 struct task_struct *p, int new_cpu)
1739{
1740 lockdep_assert_held(&rq->lock);
1741
1742 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
1743 set_task_cpu(p, new_cpu);
1744 rq_unlock(rq, rf);
1745
1746 rq = cpu_rq(new_cpu);
1747
1748 rq_lock(rq, rf);
1749 BUG_ON(task_cpu(p) != new_cpu);
1750 activate_task(rq, p, 0);
1751 check_preempt_curr(rq, p, 0);
1752
1753 return rq;
1754}
1755
1756struct migration_arg {
1757 struct task_struct *task;
1758 int dest_cpu;
1759};
1760
1761/*
1762 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1763 * this because either it can't run here any more (set_cpus_allowed()
1764 * away from this CPU, or CPU going down), or because we're
1765 * attempting to rebalance this task on exec (sched_exec).
1766 *
1767 * So we race with normal scheduler movements, but that's OK, as long
1768 * as the task is no longer on this CPU.
1769 */
1770static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1771 struct task_struct *p, int dest_cpu)
1772{
1773 /* Affinity changed (again). */
1774 if (!is_cpu_allowed(p, dest_cpu))
1775 return rq;
1776
1777 update_rq_clock(rq);
1778 rq = move_queued_task(rq, rf, p, dest_cpu);
1779
1780 return rq;
1781}
1782
1783/*
1784 * migration_cpu_stop - this will be executed by a highprio stopper thread
1785 * and performs thread migration by bumping thread off CPU then
1786 * 'pushing' onto another runqueue.
1787 */
1788static int migration_cpu_stop(void *data)
1789{
1790 struct migration_arg *arg = data;
1791 struct task_struct *p = arg->task;
1792 struct rq *rq = this_rq();
1793 struct rq_flags rf;
1794
1795 /*
1796 * The original target CPU might have gone down and we might
1797 * be on another CPU but it doesn't matter.
1798 */
1799 local_irq_disable();
1800 /*
1801 * We need to explicitly wake pending tasks before running
1802 * __migrate_task() such that we will not miss enforcing cpus_ptr
1803 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1804 */
1805 flush_smp_call_function_from_idle();
1806
1807 raw_spin_lock(&p->pi_lock);
1808 rq_lock(rq, &rf);
1809 /*
1810 * If task_rq(p) != rq, it cannot be migrated here, because we're
1811 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1812 * we're holding p->pi_lock.
1813 */
1814 if (task_rq(p) == rq) {
1815 if (task_on_rq_queued(p))
1816 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1817 else
1818 p->wake_cpu = arg->dest_cpu;
1819 }
1820 rq_unlock(rq, &rf);
1821 raw_spin_unlock(&p->pi_lock);
1822
1823 local_irq_enable();
1824 return 0;
1825}
1826
1827/*
1828 * sched_class::set_cpus_allowed must do the below, but is not required to
1829 * actually call this function.
1830 */
1831void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1832{
1833 cpumask_copy(&p->cpus_mask, new_mask);
1834 p->nr_cpus_allowed = cpumask_weight(new_mask);
1835}
1836
1837void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1838{
1839 struct rq *rq = task_rq(p);
1840 bool queued, running;
1841
1842 lockdep_assert_held(&p->pi_lock);
1843
1844 queued = task_on_rq_queued(p);
1845 running = task_current(rq, p);
1846
1847 if (queued) {
1848 /*
1849 * Because __kthread_bind() calls this on blocked tasks without
1850 * holding rq->lock.
1851 */
1852 lockdep_assert_held(&rq->lock);
1853 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1854 }
1855 if (running)
1856 put_prev_task(rq, p);
1857
1858 p->sched_class->set_cpus_allowed(p, new_mask);
1859
1860 if (queued)
1861 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1862 if (running)
1863 set_next_task(rq, p);
1864}
1865
1866/*
1867 * Change a given task's CPU affinity. Migrate the thread to a
1868 * proper CPU and schedule it away if the CPU it's executing on
1869 * is removed from the allowed bitmask.
1870 *
1871 * NOTE: the caller must have a valid reference to the task, the
1872 * task must not exit() & deallocate itself prematurely. The
1873 * call is not atomic; no spinlocks may be held.
1874 */
1875static int __set_cpus_allowed_ptr(struct task_struct *p,
1876 const struct cpumask *new_mask, bool check)
1877{
1878 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1879 unsigned int dest_cpu;
1880 struct rq_flags rf;
1881 struct rq *rq;
1882 int ret = 0;
1883
1884 rq = task_rq_lock(p, &rf);
1885 update_rq_clock(rq);
1886
1887 if (p->flags & PF_KTHREAD) {
1888 /*
1889 * Kernel threads are allowed on online && !active CPUs
1890 */
1891 cpu_valid_mask = cpu_online_mask;
1892 }
1893
1894 /*
1895 * Must re-check here, to close a race against __kthread_bind(),
1896 * sched_setaffinity() is not guaranteed to observe the flag.
1897 */
1898 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1899 ret = -EINVAL;
1900 goto out;
1901 }
1902
1903 if (cpumask_equal(&p->cpus_mask, new_mask))
1904 goto out;
1905
1906 /*
1907 * Picking a ~random cpu helps in cases where we are changing affinity
1908 * for groups of tasks (ie. cpuset), so that load balancing is not
1909 * immediately required to distribute the tasks within their new mask.
1910 */
1911 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
1912 if (dest_cpu >= nr_cpu_ids) {
1913 ret = -EINVAL;
1914 goto out;
1915 }
1916
1917 do_set_cpus_allowed(p, new_mask);
1918
1919 if (p->flags & PF_KTHREAD) {
1920 /*
1921 * For kernel threads that do indeed end up on online &&
1922 * !active we want to ensure they are strict per-CPU threads.
1923 */
1924 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1925 !cpumask_intersects(new_mask, cpu_active_mask) &&
1926 p->nr_cpus_allowed != 1);
1927 }
1928
1929 /* Can the task run on the task's current CPU? If so, we're done */
1930 if (cpumask_test_cpu(task_cpu(p), new_mask))
1931 goto out;
1932
1933 if (task_running(rq, p) || p->state == TASK_WAKING) {
1934 struct migration_arg arg = { p, dest_cpu };
1935 /* Need help from migration thread: drop lock and wait. */
1936 task_rq_unlock(rq, p, &rf);
1937 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1938 return 0;
1939 } else if (task_on_rq_queued(p)) {
1940 /*
1941 * OK, since we're going to drop the lock immediately
1942 * afterwards anyway.
1943 */
1944 rq = move_queued_task(rq, &rf, p, dest_cpu);
1945 }
1946out:
1947 task_rq_unlock(rq, p, &rf);
1948
1949 return ret;
1950}
1951
1952int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1953{
1954 return __set_cpus_allowed_ptr(p, new_mask, false);
1955}
1956EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1957
1958void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1959{
1960#ifdef CONFIG_SCHED_DEBUG
1961 /*
1962 * We should never call set_task_cpu() on a blocked task,
1963 * ttwu() will sort out the placement.
1964 */
1965 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1966 !p->on_rq);
1967
1968 /*
1969 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1970 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1971 * time relying on p->on_rq.
1972 */
1973 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1974 p->sched_class == &fair_sched_class &&
1975 (p->on_rq && !task_on_rq_migrating(p)));
1976
1977#ifdef CONFIG_LOCKDEP
1978 /*
1979 * The caller should hold either p->pi_lock or rq->lock, when changing
1980 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1981 *
1982 * sched_move_task() holds both and thus holding either pins the cgroup,
1983 * see task_group().
1984 *
1985 * Furthermore, all task_rq users should acquire both locks, see
1986 * task_rq_lock().
1987 */
1988 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1989 lockdep_is_held(&task_rq(p)->lock)));
1990#endif
1991 /*
1992 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1993 */
1994 WARN_ON_ONCE(!cpu_online(new_cpu));
1995#endif
1996
1997 trace_sched_migrate_task(p, new_cpu);
1998
1999 if (task_cpu(p) != new_cpu) {
2000 if (p->sched_class->migrate_task_rq)
2001 p->sched_class->migrate_task_rq(p, new_cpu);
2002 p->se.nr_migrations++;
2003 rseq_migrate(p);
2004 perf_event_task_migrate(p);
2005 }
2006
2007 __set_task_cpu(p, new_cpu);
2008}
2009
2010#ifdef CONFIG_NUMA_BALANCING
2011static void __migrate_swap_task(struct task_struct *p, int cpu)
2012{
2013 if (task_on_rq_queued(p)) {
2014 struct rq *src_rq, *dst_rq;
2015 struct rq_flags srf, drf;
2016
2017 src_rq = task_rq(p);
2018 dst_rq = cpu_rq(cpu);
2019
2020 rq_pin_lock(src_rq, &srf);
2021 rq_pin_lock(dst_rq, &drf);
2022
2023 deactivate_task(src_rq, p, 0);
2024 set_task_cpu(p, cpu);
2025 activate_task(dst_rq, p, 0);
2026 check_preempt_curr(dst_rq, p, 0);
2027
2028 rq_unpin_lock(dst_rq, &drf);
2029 rq_unpin_lock(src_rq, &srf);
2030
2031 } else {
2032 /*
2033 * Task isn't running anymore; make it appear like we migrated
2034 * it before it went to sleep. This means on wakeup we make the
2035 * previous CPU our target instead of where it really is.
2036 */
2037 p->wake_cpu = cpu;
2038 }
2039}
2040
2041struct migration_swap_arg {
2042 struct task_struct *src_task, *dst_task;
2043 int src_cpu, dst_cpu;
2044};
2045
2046static int migrate_swap_stop(void *data)
2047{
2048 struct migration_swap_arg *arg = data;
2049 struct rq *src_rq, *dst_rq;
2050 int ret = -EAGAIN;
2051
2052 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
2053 return -EAGAIN;
2054
2055 src_rq = cpu_rq(arg->src_cpu);
2056 dst_rq = cpu_rq(arg->dst_cpu);
2057
2058 double_raw_lock(&arg->src_task->pi_lock,
2059 &arg->dst_task->pi_lock);
2060 double_rq_lock(src_rq, dst_rq);
2061
2062 if (task_cpu(arg->dst_task) != arg->dst_cpu)
2063 goto unlock;
2064
2065 if (task_cpu(arg->src_task) != arg->src_cpu)
2066 goto unlock;
2067
2068 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
2069 goto unlock;
2070
2071 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
2072 goto unlock;
2073
2074 __migrate_swap_task(arg->src_task, arg->dst_cpu);
2075 __migrate_swap_task(arg->dst_task, arg->src_cpu);
2076
2077 ret = 0;
2078
2079unlock:
2080 double_rq_unlock(src_rq, dst_rq);
2081 raw_spin_unlock(&arg->dst_task->pi_lock);
2082 raw_spin_unlock(&arg->src_task->pi_lock);
2083
2084 return ret;
2085}
2086
2087/*
2088 * Cross migrate two tasks
2089 */
2090int migrate_swap(struct task_struct *cur, struct task_struct *p,
2091 int target_cpu, int curr_cpu)
2092{
2093 struct migration_swap_arg arg;
2094 int ret = -EINVAL;
2095
2096 arg = (struct migration_swap_arg){
2097 .src_task = cur,
2098 .src_cpu = curr_cpu,
2099 .dst_task = p,
2100 .dst_cpu = target_cpu,
2101 };
2102
2103 if (arg.src_cpu == arg.dst_cpu)
2104 goto out;
2105
2106 /*
2107 * These three tests are all lockless; this is OK since all of them
2108 * will be re-checked with proper locks held further down the line.
2109 */
2110 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
2111 goto out;
2112
2113 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
2114 goto out;
2115
2116 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
2117 goto out;
2118
2119 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
2120 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
2121
2122out:
2123 return ret;
2124}
2125#endif /* CONFIG_NUMA_BALANCING */
2126
2127/*
2128 * wait_task_inactive - wait for a thread to unschedule.
2129 *
2130 * If @match_state is nonzero, it's the @p->state value just checked and
2131 * not expected to change. If it changes, i.e. @p might have woken up,
2132 * then return zero. When we succeed in waiting for @p to be off its CPU,
2133 * we return a positive number (its total switch count). If a second call
2134 * a short while later returns the same number, the caller can be sure that
2135 * @p has remained unscheduled the whole time.
2136 *
2137 * The caller must ensure that the task *will* unschedule sometime soon,
2138 * else this function might spin for a *long* time. This function can't
2139 * be called with interrupts off, or it may introduce deadlock with
2140 * smp_call_function() if an IPI is sent by the same process we are
2141 * waiting to become inactive.
2142 */
2143unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2144{
2145 int running, queued;
2146 struct rq_flags rf;
2147 unsigned long ncsw;
2148 struct rq *rq;
2149
2150 for (;;) {
2151 /*
2152 * We do the initial early heuristics without holding
2153 * any task-queue locks at all. We'll only try to get
2154 * the runqueue lock when things look like they will
2155 * work out!
2156 */
2157 rq = task_rq(p);
2158
2159 /*
2160 * If the task is actively running on another CPU
2161 * still, just relax and busy-wait without holding
2162 * any locks.
2163 *
2164 * NOTE! Since we don't hold any locks, it's not
2165 * even sure that "rq" stays as the right runqueue!
2166 * But we don't care, since "task_running()" will
2167 * return false if the runqueue has changed and p
2168 * is actually now running somewhere else!
2169 */
2170 while (task_running(rq, p)) {
2171 if (match_state && unlikely(p->state != match_state))
2172 return 0;
2173 cpu_relax();
2174 }
2175
2176 /*
2177 * Ok, time to look more closely! We need the rq
2178 * lock now, to be *sure*. If we're wrong, we'll
2179 * just go back and repeat.
2180 */
2181 rq = task_rq_lock(p, &rf);
2182 trace_sched_wait_task(p);
2183 running = task_running(rq, p);
2184 queued = task_on_rq_queued(p);
2185 ncsw = 0;
2186 if (!match_state || p->state == match_state)
2187 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2188 task_rq_unlock(rq, p, &rf);
2189
2190 /*
2191 * If it changed from the expected state, bail out now.
2192 */
2193 if (unlikely(!ncsw))
2194 break;
2195
2196 /*
2197 * Was it really running after all now that we
2198 * checked with the proper locks actually held?
2199 *
2200 * Oops. Go back and try again..
2201 */
2202 if (unlikely(running)) {
2203 cpu_relax();
2204 continue;
2205 }
2206
2207 /*
2208 * It's not enough that it's not actively running,
2209 * it must be off the runqueue _entirely_, and not
2210 * preempted!
2211 *
2212 * So if it was still runnable (but just not actively
2213 * running right now), it's preempted, and we should
2214 * yield - it could be a while.
2215 */
2216 if (unlikely(queued)) {
2217 ktime_t to = NSEC_PER_SEC / HZ;
2218
2219 set_current_state(TASK_UNINTERRUPTIBLE);
2220 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2221 continue;
2222 }
2223
2224 /*
2225 * Ahh, all good. It wasn't running, and it wasn't
2226 * runnable, which means that it will never become
2227 * running in the future either. We're all done!
2228 */
2229 break;
2230 }
2231
2232 return ncsw;
2233}
2234
2235/***
2236 * kick_process - kick a running thread to enter/exit the kernel
2237 * @p: the to-be-kicked thread
2238 *
2239 * Cause a process which is running on another CPU to enter
2240 * kernel-mode, without any delay. (to get signals handled.)
2241 *
2242 * NOTE: this function doesn't have to take the runqueue lock,
2243 * because all it wants to ensure is that the remote task enters
2244 * the kernel. If the IPI races and the task has been migrated
2245 * to another CPU then no harm is done and the purpose has been
2246 * achieved as well.
2247 */
2248void kick_process(struct task_struct *p)
2249{
2250 int cpu;
2251
2252 preempt_disable();
2253 cpu = task_cpu(p);
2254 if ((cpu != smp_processor_id()) && task_curr(p))
2255 smp_send_reschedule(cpu);
2256 preempt_enable();
2257}
2258EXPORT_SYMBOL_GPL(kick_process);
2259
2260/*
2261 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2262 *
2263 * A few notes on cpu_active vs cpu_online:
2264 *
2265 * - cpu_active must be a subset of cpu_online
2266 *
2267 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2268 * see __set_cpus_allowed_ptr(). At this point the newly online
2269 * CPU isn't yet part of the sched domains, and balancing will not
2270 * see it.
2271 *
2272 * - on CPU-down we clear cpu_active() to mask the sched domains and
2273 * avoid the load balancer to place new tasks on the to be removed
2274 * CPU. Existing tasks will remain running there and will be taken
2275 * off.
2276 *
2277 * This means that fallback selection must not select !active CPUs.
2278 * And can assume that any active CPU must be online. Conversely
2279 * select_task_rq() below may allow selection of !active CPUs in order
2280 * to satisfy the above rules.
2281 */
2282static int select_fallback_rq(int cpu, struct task_struct *p)
2283{
2284 int nid = cpu_to_node(cpu);
2285 const struct cpumask *nodemask = NULL;
2286 enum { cpuset, possible, fail } state = cpuset;
2287 int dest_cpu;
2288
2289 /*
2290 * If the node that the CPU is on has been offlined, cpu_to_node()
2291 * will return -1. There is no CPU on the node, and we should
2292 * select the CPU on the other node.
2293 */
2294 if (nid != -1) {
2295 nodemask = cpumask_of_node(nid);
2296
2297 /* Look for allowed, online CPU in same node. */
2298 for_each_cpu(dest_cpu, nodemask) {
2299 if (!cpu_active(dest_cpu))
2300 continue;
2301 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2302 return dest_cpu;
2303 }
2304 }
2305
2306 for (;;) {
2307 /* Any allowed, online CPU? */
2308 for_each_cpu(dest_cpu, p->cpus_ptr) {
2309 if (!is_cpu_allowed(p, dest_cpu))
2310 continue;
2311
2312 goto out;
2313 }
2314
2315 /* No more Mr. Nice Guy. */
2316 switch (state) {
2317 case cpuset:
2318 if (IS_ENABLED(CONFIG_CPUSETS)) {
2319 cpuset_cpus_allowed_fallback(p);
2320 state = possible;
2321 break;
2322 }
2323 fallthrough;
2324 case possible:
2325 do_set_cpus_allowed(p, cpu_possible_mask);
2326 state = fail;
2327 break;
2328
2329 case fail:
2330 BUG();
2331 break;
2332 }
2333 }
2334
2335out:
2336 if (state != cpuset) {
2337 /*
2338 * Don't tell them about moving exiting tasks or
2339 * kernel threads (both mm NULL), since they never
2340 * leave kernel.
2341 */
2342 if (p->mm && printk_ratelimit()) {
2343 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2344 task_pid_nr(p), p->comm, cpu);
2345 }
2346 }
2347
2348 return dest_cpu;
2349}
2350
2351/*
2352 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2353 */
2354static inline
2355int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
2356{
2357 lockdep_assert_held(&p->pi_lock);
2358
2359 if (p->nr_cpus_allowed > 1)
2360 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
2361 else
2362 cpu = cpumask_any(p->cpus_ptr);
2363
2364 /*
2365 * In order not to call set_task_cpu() on a blocking task we need
2366 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2367 * CPU.
2368 *
2369 * Since this is common to all placement strategies, this lives here.
2370 *
2371 * [ this allows ->select_task() to simply return task_cpu(p) and
2372 * not worry about this generic constraint ]
2373 */
2374 if (unlikely(!is_cpu_allowed(p, cpu)))
2375 cpu = select_fallback_rq(task_cpu(p), p);
2376
2377 return cpu;
2378}
2379
2380void sched_set_stop_task(int cpu, struct task_struct *stop)
2381{
2382 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2383 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2384
2385 if (stop) {
2386 /*
2387 * Make it appear like a SCHED_FIFO task, its something
2388 * userspace knows about and won't get confused about.
2389 *
2390 * Also, it will make PI more or less work without too
2391 * much confusion -- but then, stop work should not
2392 * rely on PI working anyway.
2393 */
2394 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2395
2396 stop->sched_class = &stop_sched_class;
2397 }
2398
2399 cpu_rq(cpu)->stop = stop;
2400
2401 if (old_stop) {
2402 /*
2403 * Reset it back to a normal scheduling class so that
2404 * it can die in pieces.
2405 */
2406 old_stop->sched_class = &rt_sched_class;
2407 }
2408}
2409
2410#else
2411
2412static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2413 const struct cpumask *new_mask, bool check)
2414{
2415 return set_cpus_allowed_ptr(p, new_mask);
2416}
2417
2418#endif /* CONFIG_SMP */
2419
2420static void
2421ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2422{
2423 struct rq *rq;
2424
2425 if (!schedstat_enabled())
2426 return;
2427
2428 rq = this_rq();
2429
2430#ifdef CONFIG_SMP
2431 if (cpu == rq->cpu) {
2432 __schedstat_inc(rq->ttwu_local);
2433 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2434 } else {
2435 struct sched_domain *sd;
2436
2437 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2438 rcu_read_lock();
2439 for_each_domain(rq->cpu, sd) {
2440 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2441 __schedstat_inc(sd->ttwu_wake_remote);
2442 break;
2443 }
2444 }
2445 rcu_read_unlock();
2446 }
2447
2448 if (wake_flags & WF_MIGRATED)
2449 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2450#endif /* CONFIG_SMP */
2451
2452 __schedstat_inc(rq->ttwu_count);
2453 __schedstat_inc(p->se.statistics.nr_wakeups);
2454
2455 if (wake_flags & WF_SYNC)
2456 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2457}
2458
2459/*
2460 * Mark the task runnable and perform wakeup-preemption.
2461 */
2462static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2463 struct rq_flags *rf)
2464{
2465 check_preempt_curr(rq, p, wake_flags);
2466 p->state = TASK_RUNNING;
2467 trace_sched_wakeup(p);
2468
2469#ifdef CONFIG_SMP
2470 if (p->sched_class->task_woken) {
2471 /*
2472 * Our task @p is fully woken up and running; so its safe to
2473 * drop the rq->lock, hereafter rq is only used for statistics.
2474 */
2475 rq_unpin_lock(rq, rf);
2476 p->sched_class->task_woken(rq, p);
2477 rq_repin_lock(rq, rf);
2478 }
2479
2480 if (rq->idle_stamp) {
2481 u64 delta = rq_clock(rq) - rq->idle_stamp;
2482 u64 max = 2*rq->max_idle_balance_cost;
2483
2484 update_avg(&rq->avg_idle, delta);
2485
2486 if (rq->avg_idle > max)
2487 rq->avg_idle = max;
2488
2489 rq->idle_stamp = 0;
2490 }
2491#endif
2492}
2493
2494static void
2495ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2496 struct rq_flags *rf)
2497{
2498 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2499
2500 lockdep_assert_held(&rq->lock);
2501
2502 if (p->sched_contributes_to_load)
2503 rq->nr_uninterruptible--;
2504
2505#ifdef CONFIG_SMP
2506 if (wake_flags & WF_MIGRATED)
2507 en_flags |= ENQUEUE_MIGRATED;
2508#endif
2509
2510 activate_task(rq, p, en_flags);
2511 ttwu_do_wakeup(rq, p, wake_flags, rf);
2512}
2513
2514/*
2515 * Consider @p being inside a wait loop:
2516 *
2517 * for (;;) {
2518 * set_current_state(TASK_UNINTERRUPTIBLE);
2519 *
2520 * if (CONDITION)
2521 * break;
2522 *
2523 * schedule();
2524 * }
2525 * __set_current_state(TASK_RUNNING);
2526 *
2527 * between set_current_state() and schedule(). In this case @p is still
2528 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
2529 * an atomic manner.
2530 *
2531 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
2532 * then schedule() must still happen and p->state can be changed to
2533 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
2534 * need to do a full wakeup with enqueue.
2535 *
2536 * Returns: %true when the wakeup is done,
2537 * %false otherwise.
2538 */
2539static int ttwu_runnable(struct task_struct *p, int wake_flags)
2540{
2541 struct rq_flags rf;
2542 struct rq *rq;
2543 int ret = 0;
2544
2545 rq = __task_rq_lock(p, &rf);
2546 if (task_on_rq_queued(p)) {
2547 /* check_preempt_curr() may use rq clock */
2548 update_rq_clock(rq);
2549 ttwu_do_wakeup(rq, p, wake_flags, &rf);
2550 ret = 1;
2551 }
2552 __task_rq_unlock(rq, &rf);
2553
2554 return ret;
2555}
2556
2557#ifdef CONFIG_SMP
2558void sched_ttwu_pending(void *arg)
2559{
2560 struct llist_node *llist = arg;
2561 struct rq *rq = this_rq();
2562 struct task_struct *p, *t;
2563 struct rq_flags rf;
2564
2565 if (!llist)
2566 return;
2567
2568 /*
2569 * rq::ttwu_pending racy indication of out-standing wakeups.
2570 * Races such that false-negatives are possible, since they
2571 * are shorter lived that false-positives would be.
2572 */
2573 WRITE_ONCE(rq->ttwu_pending, 0);
2574
2575 rq_lock_irqsave(rq, &rf);
2576 update_rq_clock(rq);
2577
2578 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
2579 if (WARN_ON_ONCE(p->on_cpu))
2580 smp_cond_load_acquire(&p->on_cpu, !VAL);
2581
2582 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
2583 set_task_cpu(p, cpu_of(rq));
2584
2585 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
2586 }
2587
2588 rq_unlock_irqrestore(rq, &rf);
2589}
2590
2591void send_call_function_single_ipi(int cpu)
2592{
2593 struct rq *rq = cpu_rq(cpu);
2594
2595 if (!set_nr_if_polling(rq->idle))
2596 arch_send_call_function_single_ipi(cpu);
2597 else
2598 trace_sched_wake_idle_without_ipi(cpu);
2599}
2600
2601/*
2602 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
2603 * necessary. The wakee CPU on receipt of the IPI will queue the task
2604 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
2605 * of the wakeup instead of the waker.
2606 */
2607static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
2608{
2609 struct rq *rq = cpu_rq(cpu);
2610
2611 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
2612
2613 WRITE_ONCE(rq->ttwu_pending, 1);
2614 __smp_call_single_queue(cpu, &p->wake_entry.llist);
2615}
2616
2617void wake_up_if_idle(int cpu)
2618{
2619 struct rq *rq = cpu_rq(cpu);
2620 struct rq_flags rf;
2621
2622 rcu_read_lock();
2623
2624 if (!is_idle_task(rcu_dereference(rq->curr)))
2625 goto out;
2626
2627 if (set_nr_if_polling(rq->idle)) {
2628 trace_sched_wake_idle_without_ipi(cpu);
2629 } else {
2630 rq_lock_irqsave(rq, &rf);
2631 if (is_idle_task(rq->curr))
2632 smp_send_reschedule(cpu);
2633 /* Else CPU is not idle, do nothing here: */
2634 rq_unlock_irqrestore(rq, &rf);
2635 }
2636
2637out:
2638 rcu_read_unlock();
2639}
2640
2641bool cpus_share_cache(int this_cpu, int that_cpu)
2642{
2643 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2644}
2645
2646static inline bool ttwu_queue_cond(int cpu, int wake_flags)
2647{
2648 /*
2649 * If the CPU does not share cache, then queue the task on the
2650 * remote rqs wakelist to avoid accessing remote data.
2651 */
2652 if (!cpus_share_cache(smp_processor_id(), cpu))
2653 return true;
2654
2655 /*
2656 * If the task is descheduling and the only running task on the
2657 * CPU then use the wakelist to offload the task activation to
2658 * the soon-to-be-idle CPU as the current CPU is likely busy.
2659 * nr_running is checked to avoid unnecessary task stacking.
2660 */
2661 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
2662 return true;
2663
2664 return false;
2665}
2666
2667static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
2668{
2669 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
2670 if (WARN_ON_ONCE(cpu == smp_processor_id()))
2671 return false;
2672
2673 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
2674 __ttwu_queue_wakelist(p, cpu, wake_flags);
2675 return true;
2676 }
2677
2678 return false;
2679}
2680
2681#else /* !CONFIG_SMP */
2682
2683static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
2684{
2685 return false;
2686}
2687
2688#endif /* CONFIG_SMP */
2689
2690static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
2691{
2692 struct rq *rq = cpu_rq(cpu);
2693 struct rq_flags rf;
2694
2695 if (ttwu_queue_wakelist(p, cpu, wake_flags))
2696 return;
2697
2698 rq_lock(rq, &rf);
2699 update_rq_clock(rq);
2700 ttwu_do_activate(rq, p, wake_flags, &rf);
2701 rq_unlock(rq, &rf);
2702}
2703
2704/*
2705 * Notes on Program-Order guarantees on SMP systems.
2706 *
2707 * MIGRATION
2708 *
2709 * The basic program-order guarantee on SMP systems is that when a task [t]
2710 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2711 * execution on its new CPU [c1].
2712 *
2713 * For migration (of runnable tasks) this is provided by the following means:
2714 *
2715 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2716 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2717 * rq(c1)->lock (if not at the same time, then in that order).
2718 * C) LOCK of the rq(c1)->lock scheduling in task
2719 *
2720 * Release/acquire chaining guarantees that B happens after A and C after B.
2721 * Note: the CPU doing B need not be c0 or c1
2722 *
2723 * Example:
2724 *
2725 * CPU0 CPU1 CPU2
2726 *
2727 * LOCK rq(0)->lock
2728 * sched-out X
2729 * sched-in Y
2730 * UNLOCK rq(0)->lock
2731 *
2732 * LOCK rq(0)->lock // orders against CPU0
2733 * dequeue X
2734 * UNLOCK rq(0)->lock
2735 *
2736 * LOCK rq(1)->lock
2737 * enqueue X
2738 * UNLOCK rq(1)->lock
2739 *
2740 * LOCK rq(1)->lock // orders against CPU2
2741 * sched-out Z
2742 * sched-in X
2743 * UNLOCK rq(1)->lock
2744 *
2745 *
2746 * BLOCKING -- aka. SLEEP + WAKEUP
2747 *
2748 * For blocking we (obviously) need to provide the same guarantee as for
2749 * migration. However the means are completely different as there is no lock
2750 * chain to provide order. Instead we do:
2751 *
2752 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
2753 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
2754 *
2755 * Example:
2756 *
2757 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2758 *
2759 * LOCK rq(0)->lock LOCK X->pi_lock
2760 * dequeue X
2761 * sched-out X
2762 * smp_store_release(X->on_cpu, 0);
2763 *
2764 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2765 * X->state = WAKING
2766 * set_task_cpu(X,2)
2767 *
2768 * LOCK rq(2)->lock
2769 * enqueue X
2770 * X->state = RUNNING
2771 * UNLOCK rq(2)->lock
2772 *
2773 * LOCK rq(2)->lock // orders against CPU1
2774 * sched-out Z
2775 * sched-in X
2776 * UNLOCK rq(2)->lock
2777 *
2778 * UNLOCK X->pi_lock
2779 * UNLOCK rq(0)->lock
2780 *
2781 *
2782 * However, for wakeups there is a second guarantee we must provide, namely we
2783 * must ensure that CONDITION=1 done by the caller can not be reordered with
2784 * accesses to the task state; see try_to_wake_up() and set_current_state().
2785 */
2786
2787/**
2788 * try_to_wake_up - wake up a thread
2789 * @p: the thread to be awakened
2790 * @state: the mask of task states that can be woken
2791 * @wake_flags: wake modifier flags (WF_*)
2792 *
2793 * Conceptually does:
2794 *
2795 * If (@state & @p->state) @p->state = TASK_RUNNING.
2796 *
2797 * If the task was not queued/runnable, also place it back on a runqueue.
2798 *
2799 * This function is atomic against schedule() which would dequeue the task.
2800 *
2801 * It issues a full memory barrier before accessing @p->state, see the comment
2802 * with set_current_state().
2803 *
2804 * Uses p->pi_lock to serialize against concurrent wake-ups.
2805 *
2806 * Relies on p->pi_lock stabilizing:
2807 * - p->sched_class
2808 * - p->cpus_ptr
2809 * - p->sched_task_group
2810 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
2811 *
2812 * Tries really hard to only take one task_rq(p)->lock for performance.
2813 * Takes rq->lock in:
2814 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
2815 * - ttwu_queue() -- new rq, for enqueue of the task;
2816 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
2817 *
2818 * As a consequence we race really badly with just about everything. See the
2819 * many memory barriers and their comments for details.
2820 *
2821 * Return: %true if @p->state changes (an actual wakeup was done),
2822 * %false otherwise.
2823 */
2824static int
2825try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2826{
2827 unsigned long flags;
2828 int cpu, success = 0;
2829
2830 preempt_disable();
2831 if (p == current) {
2832 /*
2833 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2834 * == smp_processor_id()'. Together this means we can special
2835 * case the whole 'p->on_rq && ttwu_runnable()' case below
2836 * without taking any locks.
2837 *
2838 * In particular:
2839 * - we rely on Program-Order guarantees for all the ordering,
2840 * - we're serialized against set_special_state() by virtue of
2841 * it disabling IRQs (this allows not taking ->pi_lock).
2842 */
2843 if (!(p->state & state))
2844 goto out;
2845
2846 success = 1;
2847 trace_sched_waking(p);
2848 p->state = TASK_RUNNING;
2849 trace_sched_wakeup(p);
2850 goto out;
2851 }
2852
2853 /*
2854 * If we are going to wake up a thread waiting for CONDITION we
2855 * need to ensure that CONDITION=1 done by the caller can not be
2856 * reordered with p->state check below. This pairs with smp_store_mb()
2857 * in set_current_state() that the waiting thread does.
2858 */
2859 raw_spin_lock_irqsave(&p->pi_lock, flags);
2860 smp_mb__after_spinlock();
2861 if (!(p->state & state))
2862 goto unlock;
2863
2864 trace_sched_waking(p);
2865
2866 /* We're going to change ->state: */
2867 success = 1;
2868
2869 /*
2870 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2871 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2872 * in smp_cond_load_acquire() below.
2873 *
2874 * sched_ttwu_pending() try_to_wake_up()
2875 * STORE p->on_rq = 1 LOAD p->state
2876 * UNLOCK rq->lock
2877 *
2878 * __schedule() (switch to task 'p')
2879 * LOCK rq->lock smp_rmb();
2880 * smp_mb__after_spinlock();
2881 * UNLOCK rq->lock
2882 *
2883 * [task p]
2884 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2885 *
2886 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2887 * __schedule(). See the comment for smp_mb__after_spinlock().
2888 *
2889 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
2890 */
2891 smp_rmb();
2892 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
2893 goto unlock;
2894
2895 if (p->in_iowait) {
2896 delayacct_blkio_end(p);
2897 atomic_dec(&task_rq(p)->nr_iowait);
2898 }
2899
2900#ifdef CONFIG_SMP
2901 /*
2902 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2903 * possible to, falsely, observe p->on_cpu == 0.
2904 *
2905 * One must be running (->on_cpu == 1) in order to remove oneself
2906 * from the runqueue.
2907 *
2908 * __schedule() (switch to task 'p') try_to_wake_up()
2909 * STORE p->on_cpu = 1 LOAD p->on_rq
2910 * UNLOCK rq->lock
2911 *
2912 * __schedule() (put 'p' to sleep)
2913 * LOCK rq->lock smp_rmb();
2914 * smp_mb__after_spinlock();
2915 * STORE p->on_rq = 0 LOAD p->on_cpu
2916 *
2917 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2918 * __schedule(). See the comment for smp_mb__after_spinlock().
2919 *
2920 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
2921 * schedule()'s deactivate_task() has 'happened' and p will no longer
2922 * care about it's own p->state. See the comment in __schedule().
2923 */
2924 smp_acquire__after_ctrl_dep();
2925
2926 /*
2927 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
2928 * == 0), which means we need to do an enqueue, change p->state to
2929 * TASK_WAKING such that we can unlock p->pi_lock before doing the
2930 * enqueue, such as ttwu_queue_wakelist().
2931 */
2932 p->state = TASK_WAKING;
2933
2934 /*
2935 * If the owning (remote) CPU is still in the middle of schedule() with
2936 * this task as prev, considering queueing p on the remote CPUs wake_list
2937 * which potentially sends an IPI instead of spinning on p->on_cpu to
2938 * let the waker make forward progress. This is safe because IRQs are
2939 * disabled and the IPI will deliver after on_cpu is cleared.
2940 *
2941 * Ensure we load task_cpu(p) after p->on_cpu:
2942 *
2943 * set_task_cpu(p, cpu);
2944 * STORE p->cpu = @cpu
2945 * __schedule() (switch to task 'p')
2946 * LOCK rq->lock
2947 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
2948 * STORE p->on_cpu = 1 LOAD p->cpu
2949 *
2950 * to ensure we observe the correct CPU on which the task is currently
2951 * scheduling.
2952 */
2953 if (smp_load_acquire(&p->on_cpu) &&
2954 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
2955 goto unlock;
2956
2957 /*
2958 * If the owning (remote) CPU is still in the middle of schedule() with
2959 * this task as prev, wait until its done referencing the task.
2960 *
2961 * Pairs with the smp_store_release() in finish_task().
2962 *
2963 * This ensures that tasks getting woken will be fully ordered against
2964 * their previous state and preserve Program Order.
2965 */
2966 smp_cond_load_acquire(&p->on_cpu, !VAL);
2967
2968 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2969 if (task_cpu(p) != cpu) {
2970 wake_flags |= WF_MIGRATED;
2971 psi_ttwu_dequeue(p);
2972 set_task_cpu(p, cpu);
2973 }
2974#else
2975 cpu = task_cpu(p);
2976#endif /* CONFIG_SMP */
2977
2978 ttwu_queue(p, cpu, wake_flags);
2979unlock:
2980 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2981out:
2982 if (success)
2983 ttwu_stat(p, task_cpu(p), wake_flags);
2984 preempt_enable();
2985
2986 return success;
2987}
2988
2989/**
2990 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
2991 * @p: Process for which the function is to be invoked.
2992 * @func: Function to invoke.
2993 * @arg: Argument to function.
2994 *
2995 * If the specified task can be quickly locked into a definite state
2996 * (either sleeping or on a given runqueue), arrange to keep it in that
2997 * state while invoking @func(@arg). This function can use ->on_rq and
2998 * task_curr() to work out what the state is, if required. Given that
2999 * @func can be invoked with a runqueue lock held, it had better be quite
3000 * lightweight.
3001 *
3002 * Returns:
3003 * @false if the task slipped out from under the locks.
3004 * @true if the task was locked onto a runqueue or is sleeping.
3005 * However, @func can override this by returning @false.
3006 */
3007bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
3008{
3009 bool ret = false;
3010 struct rq_flags rf;
3011 struct rq *rq;
3012
3013 lockdep_assert_irqs_enabled();
3014 raw_spin_lock_irq(&p->pi_lock);
3015 if (p->on_rq) {
3016 rq = __task_rq_lock(p, &rf);
3017 if (task_rq(p) == rq)
3018 ret = func(p, arg);
3019 rq_unlock(rq, &rf);
3020 } else {
3021 switch (p->state) {
3022 case TASK_RUNNING:
3023 case TASK_WAKING:
3024 break;
3025 default:
3026 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3027 if (!p->on_rq)
3028 ret = func(p, arg);
3029 }
3030 }
3031 raw_spin_unlock_irq(&p->pi_lock);
3032 return ret;
3033}
3034
3035/**
3036 * wake_up_process - Wake up a specific process
3037 * @p: The process to be woken up.
3038 *
3039 * Attempt to wake up the nominated process and move it to the set of runnable
3040 * processes.
3041 *
3042 * Return: 1 if the process was woken up, 0 if it was already running.
3043 *
3044 * This function executes a full memory barrier before accessing the task state.
3045 */
3046int wake_up_process(struct task_struct *p)
3047{
3048 return try_to_wake_up(p, TASK_NORMAL, 0);
3049}
3050EXPORT_SYMBOL(wake_up_process);
3051
3052int wake_up_state(struct task_struct *p, unsigned int state)
3053{
3054 return try_to_wake_up(p, state, 0);
3055}
3056
3057/*
3058 * Perform scheduler related setup for a newly forked process p.
3059 * p is forked by current.
3060 *
3061 * __sched_fork() is basic setup used by init_idle() too:
3062 */
3063static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
3064{
3065 p->on_rq = 0;
3066
3067 p->se.on_rq = 0;
3068 p->se.exec_start = 0;
3069 p->se.sum_exec_runtime = 0;
3070 p->se.prev_sum_exec_runtime = 0;
3071 p->se.nr_migrations = 0;
3072 p->se.vruntime = 0;
3073 INIT_LIST_HEAD(&p->se.group_node);
3074
3075#ifdef CONFIG_FAIR_GROUP_SCHED
3076 p->se.cfs_rq = NULL;
3077#endif
3078
3079#ifdef CONFIG_SCHEDSTATS
3080 /* Even if schedstat is disabled, there should not be garbage */
3081 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
3082#endif
3083
3084 RB_CLEAR_NODE(&p->dl.rb_node);
3085 init_dl_task_timer(&p->dl);
3086 init_dl_inactive_task_timer(&p->dl);
3087 __dl_clear_params(p);
3088
3089 INIT_LIST_HEAD(&p->rt.run_list);
3090 p->rt.timeout = 0;
3091 p->rt.time_slice = sched_rr_timeslice;
3092 p->rt.on_rq = 0;
3093 p->rt.on_list = 0;
3094
3095#ifdef CONFIG_PREEMPT_NOTIFIERS
3096 INIT_HLIST_HEAD(&p->preempt_notifiers);
3097#endif
3098
3099#ifdef CONFIG_COMPACTION
3100 p->capture_control = NULL;
3101#endif
3102 init_numa_balancing(clone_flags, p);
3103#ifdef CONFIG_SMP
3104 p->wake_entry.u_flags = CSD_TYPE_TTWU;
3105#endif
3106}
3107
3108DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
3109
3110#ifdef CONFIG_NUMA_BALANCING
3111
3112void set_numabalancing_state(bool enabled)
3113{
3114 if (enabled)
3115 static_branch_enable(&sched_numa_balancing);
3116 else
3117 static_branch_disable(&sched_numa_balancing);
3118}
3119
3120#ifdef CONFIG_PROC_SYSCTL
3121int sysctl_numa_balancing(struct ctl_table *table, int write,
3122 void *buffer, size_t *lenp, loff_t *ppos)
3123{
3124 struct ctl_table t;
3125 int err;
3126 int state = static_branch_likely(&sched_numa_balancing);
3127
3128 if (write && !capable(CAP_SYS_ADMIN))
3129 return -EPERM;
3130
3131 t = *table;
3132 t.data = &state;
3133 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3134 if (err < 0)
3135 return err;
3136 if (write)
3137 set_numabalancing_state(state);
3138 return err;
3139}
3140#endif
3141#endif
3142
3143#ifdef CONFIG_SCHEDSTATS
3144
3145DEFINE_STATIC_KEY_FALSE(sched_schedstats);
3146static bool __initdata __sched_schedstats = false;
3147
3148static void set_schedstats(bool enabled)
3149{
3150 if (enabled)
3151 static_branch_enable(&sched_schedstats);
3152 else
3153 static_branch_disable(&sched_schedstats);
3154}
3155
3156void force_schedstat_enabled(void)
3157{
3158 if (!schedstat_enabled()) {
3159 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
3160 static_branch_enable(&sched_schedstats);
3161 }
3162}
3163
3164static int __init setup_schedstats(char *str)
3165{
3166 int ret = 0;
3167 if (!str)
3168 goto out;
3169
3170 /*
3171 * This code is called before jump labels have been set up, so we can't
3172 * change the static branch directly just yet. Instead set a temporary
3173 * variable so init_schedstats() can do it later.
3174 */
3175 if (!strcmp(str, "enable")) {
3176 __sched_schedstats = true;
3177 ret = 1;
3178 } else if (!strcmp(str, "disable")) {
3179 __sched_schedstats = false;
3180 ret = 1;
3181 }
3182out:
3183 if (!ret)
3184 pr_warn("Unable to parse schedstats=\n");
3185
3186 return ret;
3187}
3188__setup("schedstats=", setup_schedstats);
3189
3190static void __init init_schedstats(void)
3191{
3192 set_schedstats(__sched_schedstats);
3193}
3194
3195#ifdef CONFIG_PROC_SYSCTL
3196int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
3197 size_t *lenp, loff_t *ppos)
3198{
3199 struct ctl_table t;
3200 int err;
3201 int state = static_branch_likely(&sched_schedstats);
3202
3203 if (write && !capable(CAP_SYS_ADMIN))
3204 return -EPERM;
3205
3206 t = *table;
3207 t.data = &state;
3208 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3209 if (err < 0)
3210 return err;
3211 if (write)
3212 set_schedstats(state);
3213 return err;
3214}
3215#endif /* CONFIG_PROC_SYSCTL */
3216#else /* !CONFIG_SCHEDSTATS */
3217static inline void init_schedstats(void) {}
3218#endif /* CONFIG_SCHEDSTATS */
3219
3220/*
3221 * fork()/clone()-time setup:
3222 */
3223int sched_fork(unsigned long clone_flags, struct task_struct *p)
3224{
3225 unsigned long flags;
3226
3227 __sched_fork(clone_flags, p);
3228 /*
3229 * We mark the process as NEW here. This guarantees that
3230 * nobody will actually run it, and a signal or other external
3231 * event cannot wake it up and insert it on the runqueue either.
3232 */
3233 p->state = TASK_NEW;
3234
3235 /*
3236 * Make sure we do not leak PI boosting priority to the child.
3237 */
3238 p->prio = current->normal_prio;
3239
3240 uclamp_fork(p);
3241
3242 /*
3243 * Revert to default priority/policy on fork if requested.
3244 */
3245 if (unlikely(p->sched_reset_on_fork)) {
3246 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3247 p->policy = SCHED_NORMAL;
3248 p->static_prio = NICE_TO_PRIO(0);
3249 p->rt_priority = 0;
3250 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3251 p->static_prio = NICE_TO_PRIO(0);
3252
3253 p->prio = p->normal_prio = __normal_prio(p);
3254 set_load_weight(p, false);
3255
3256 /*
3257 * We don't need the reset flag anymore after the fork. It has
3258 * fulfilled its duty:
3259 */
3260 p->sched_reset_on_fork = 0;
3261 }
3262
3263 if (dl_prio(p->prio))
3264 return -EAGAIN;
3265 else if (rt_prio(p->prio))
3266 p->sched_class = &rt_sched_class;
3267 else
3268 p->sched_class = &fair_sched_class;
3269
3270 init_entity_runnable_average(&p->se);
3271
3272 /*
3273 * The child is not yet in the pid-hash so no cgroup attach races,
3274 * and the cgroup is pinned to this child due to cgroup_fork()
3275 * is ran before sched_fork().
3276 *
3277 * Silence PROVE_RCU.
3278 */
3279 raw_spin_lock_irqsave(&p->pi_lock, flags);
3280 rseq_migrate(p);
3281 /*
3282 * We're setting the CPU for the first time, we don't migrate,
3283 * so use __set_task_cpu().
3284 */
3285 __set_task_cpu(p, smp_processor_id());
3286 if (p->sched_class->task_fork)
3287 p->sched_class->task_fork(p);
3288 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3289
3290#ifdef CONFIG_SCHED_INFO
3291 if (likely(sched_info_on()))
3292 memset(&p->sched_info, 0, sizeof(p->sched_info));
3293#endif
3294#if defined(CONFIG_SMP)
3295 p->on_cpu = 0;
3296#endif
3297 init_task_preempt_count(p);
3298#ifdef CONFIG_SMP
3299 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3300 RB_CLEAR_NODE(&p->pushable_dl_tasks);
3301#endif
3302 return 0;
3303}
3304
3305void sched_post_fork(struct task_struct *p)
3306{
3307 uclamp_post_fork(p);
3308}
3309
3310unsigned long to_ratio(u64 period, u64 runtime)
3311{
3312 if (runtime == RUNTIME_INF)
3313 return BW_UNIT;
3314
3315 /*
3316 * Doing this here saves a lot of checks in all
3317 * the calling paths, and returning zero seems
3318 * safe for them anyway.
3319 */
3320 if (period == 0)
3321 return 0;
3322
3323 return div64_u64(runtime << BW_SHIFT, period);
3324}
3325
3326/*
3327 * wake_up_new_task - wake up a newly created task for the first time.
3328 *
3329 * This function will do some initial scheduler statistics housekeeping
3330 * that must be done for every newly created context, then puts the task
3331 * on the runqueue and wakes it.
3332 */
3333void wake_up_new_task(struct task_struct *p)
3334{
3335 struct rq_flags rf;
3336 struct rq *rq;
3337
3338 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3339 p->state = TASK_RUNNING;
3340#ifdef CONFIG_SMP
3341 /*
3342 * Fork balancing, do it here and not earlier because:
3343 * - cpus_ptr can change in the fork path
3344 * - any previously selected CPU might disappear through hotplug
3345 *
3346 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3347 * as we're not fully set-up yet.
3348 */
3349 p->recent_used_cpu = task_cpu(p);
3350 rseq_migrate(p);
3351 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
3352#endif
3353 rq = __task_rq_lock(p, &rf);
3354 update_rq_clock(rq);
3355 post_init_entity_util_avg(p);
3356
3357 activate_task(rq, p, ENQUEUE_NOCLOCK);
3358 trace_sched_wakeup_new(p);
3359 check_preempt_curr(rq, p, WF_FORK);
3360#ifdef CONFIG_SMP
3361 if (p->sched_class->task_woken) {
3362 /*
3363 * Nothing relies on rq->lock after this, so its fine to
3364 * drop it.
3365 */
3366 rq_unpin_lock(rq, &rf);
3367 p->sched_class->task_woken(rq, p);
3368 rq_repin_lock(rq, &rf);
3369 }
3370#endif
3371 task_rq_unlock(rq, p, &rf);
3372}
3373
3374#ifdef CONFIG_PREEMPT_NOTIFIERS
3375
3376static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
3377
3378void preempt_notifier_inc(void)
3379{
3380 static_branch_inc(&preempt_notifier_key);
3381}
3382EXPORT_SYMBOL_GPL(preempt_notifier_inc);
3383
3384void preempt_notifier_dec(void)
3385{
3386 static_branch_dec(&preempt_notifier_key);
3387}
3388EXPORT_SYMBOL_GPL(preempt_notifier_dec);
3389
3390/**
3391 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3392 * @notifier: notifier struct to register
3393 */
3394void preempt_notifier_register(struct preempt_notifier *notifier)
3395{
3396 if (!static_branch_unlikely(&preempt_notifier_key))
3397 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3398
3399 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3400}
3401EXPORT_SYMBOL_GPL(preempt_notifier_register);
3402
3403/**
3404 * preempt_notifier_unregister - no longer interested in preemption notifications
3405 * @notifier: notifier struct to unregister
3406 *
3407 * This is *not* safe to call from within a preemption notifier.
3408 */
3409void preempt_notifier_unregister(struct preempt_notifier *notifier)
3410{
3411 hlist_del(¬ifier->link);
3412}
3413EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3414
3415static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3416{
3417 struct preempt_notifier *notifier;
3418
3419 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3420 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3421}
3422
3423static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3424{
3425 if (static_branch_unlikely(&preempt_notifier_key))
3426 __fire_sched_in_preempt_notifiers(curr);
3427}
3428
3429static void
3430__fire_sched_out_preempt_notifiers(struct task_struct *curr,
3431 struct task_struct *next)
3432{
3433 struct preempt_notifier *notifier;
3434
3435 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3436 notifier->ops->sched_out(notifier, next);
3437}
3438
3439static __always_inline void
3440fire_sched_out_preempt_notifiers(struct task_struct *curr,
3441 struct task_struct *next)
3442{
3443 if (static_branch_unlikely(&preempt_notifier_key))
3444 __fire_sched_out_preempt_notifiers(curr, next);
3445}
3446
3447#else /* !CONFIG_PREEMPT_NOTIFIERS */
3448
3449static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3450{
3451}
3452
3453static inline void
3454fire_sched_out_preempt_notifiers(struct task_struct *curr,
3455 struct task_struct *next)
3456{
3457}
3458
3459#endif /* CONFIG_PREEMPT_NOTIFIERS */
3460
3461static inline void prepare_task(struct task_struct *next)
3462{
3463#ifdef CONFIG_SMP
3464 /*
3465 * Claim the task as running, we do this before switching to it
3466 * such that any running task will have this set.
3467 *
3468 * See the ttwu() WF_ON_CPU case and its ordering comment.
3469 */
3470 WRITE_ONCE(next->on_cpu, 1);
3471#endif
3472}
3473
3474static inline void finish_task(struct task_struct *prev)
3475{
3476#ifdef CONFIG_SMP
3477 /*
3478 * This must be the very last reference to @prev from this CPU. After
3479 * p->on_cpu is cleared, the task can be moved to a different CPU. We
3480 * must ensure this doesn't happen until the switch is completely
3481 * finished.
3482 *
3483 * In particular, the load of prev->state in finish_task_switch() must
3484 * happen before this.
3485 *
3486 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3487 */
3488 smp_store_release(&prev->on_cpu, 0);
3489#endif
3490}
3491
3492static inline void
3493prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
3494{
3495 /*
3496 * Since the runqueue lock will be released by the next
3497 * task (which is an invalid locking op but in the case
3498 * of the scheduler it's an obvious special-case), so we
3499 * do an early lockdep release here:
3500 */
3501 rq_unpin_lock(rq, rf);
3502 spin_release(&rq->lock.dep_map, _THIS_IP_);
3503#ifdef CONFIG_DEBUG_SPINLOCK
3504 /* this is a valid case when another task releases the spinlock */
3505 rq->lock.owner = next;
3506#endif
3507}
3508
3509static inline void finish_lock_switch(struct rq *rq)
3510{
3511 /*
3512 * If we are tracking spinlock dependencies then we have to
3513 * fix up the runqueue lock - which gets 'carried over' from
3514 * prev into current:
3515 */
3516 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
3517 raw_spin_unlock_irq(&rq->lock);
3518}
3519
3520/*
3521 * NOP if the arch has not defined these:
3522 */
3523
3524#ifndef prepare_arch_switch
3525# define prepare_arch_switch(next) do { } while (0)
3526#endif
3527
3528#ifndef finish_arch_post_lock_switch
3529# define finish_arch_post_lock_switch() do { } while (0)
3530#endif
3531
3532/**
3533 * prepare_task_switch - prepare to switch tasks
3534 * @rq: the runqueue preparing to switch
3535 * @prev: the current task that is being switched out
3536 * @next: the task we are going to switch to.
3537 *
3538 * This is called with the rq lock held and interrupts off. It must
3539 * be paired with a subsequent finish_task_switch after the context
3540 * switch.
3541 *
3542 * prepare_task_switch sets up locking and calls architecture specific
3543 * hooks.
3544 */
3545static inline void
3546prepare_task_switch(struct rq *rq, struct task_struct *prev,
3547 struct task_struct *next)
3548{
3549 kcov_prepare_switch(prev);
3550 sched_info_switch(rq, prev, next);
3551 perf_event_task_sched_out(prev, next);
3552 rseq_preempt(prev);
3553 fire_sched_out_preempt_notifiers(prev, next);
3554 prepare_task(next);
3555 prepare_arch_switch(next);
3556}
3557
3558/**
3559 * finish_task_switch - clean up after a task-switch
3560 * @prev: the thread we just switched away from.
3561 *
3562 * finish_task_switch must be called after the context switch, paired
3563 * with a prepare_task_switch call before the context switch.
3564 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3565 * and do any other architecture-specific cleanup actions.
3566 *
3567 * Note that we may have delayed dropping an mm in context_switch(). If
3568 * so, we finish that here outside of the runqueue lock. (Doing it
3569 * with the lock held can cause deadlocks; see schedule() for
3570 * details.)
3571 *
3572 * The context switch have flipped the stack from under us and restored the
3573 * local variables which were saved when this task called schedule() in the
3574 * past. prev == current is still correct but we need to recalculate this_rq
3575 * because prev may have moved to another CPU.
3576 */
3577static struct rq *finish_task_switch(struct task_struct *prev)
3578 __releases(rq->lock)
3579{
3580 struct rq *rq = this_rq();
3581 struct mm_struct *mm = rq->prev_mm;
3582 long prev_state;
3583
3584 /*
3585 * The previous task will have left us with a preempt_count of 2
3586 * because it left us after:
3587 *
3588 * schedule()
3589 * preempt_disable(); // 1
3590 * __schedule()
3591 * raw_spin_lock_irq(&rq->lock) // 2
3592 *
3593 * Also, see FORK_PREEMPT_COUNT.
3594 */
3595 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
3596 "corrupted preempt_count: %s/%d/0x%x\n",
3597 current->comm, current->pid, preempt_count()))
3598 preempt_count_set(FORK_PREEMPT_COUNT);
3599
3600 rq->prev_mm = NULL;
3601
3602 /*
3603 * A task struct has one reference for the use as "current".
3604 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3605 * schedule one last time. The schedule call will never return, and
3606 * the scheduled task must drop that reference.
3607 *
3608 * We must observe prev->state before clearing prev->on_cpu (in
3609 * finish_task), otherwise a concurrent wakeup can get prev
3610 * running on another CPU and we could rave with its RUNNING -> DEAD
3611 * transition, resulting in a double drop.
3612 */
3613 prev_state = prev->state;
3614 vtime_task_switch(prev);
3615 perf_event_task_sched_in(prev, current);
3616 finish_task(prev);
3617 finish_lock_switch(rq);
3618 finish_arch_post_lock_switch();
3619 kcov_finish_switch(current);
3620
3621 fire_sched_in_preempt_notifiers(current);
3622 /*
3623 * When switching through a kernel thread, the loop in
3624 * membarrier_{private,global}_expedited() may have observed that
3625 * kernel thread and not issued an IPI. It is therefore possible to
3626 * schedule between user->kernel->user threads without passing though
3627 * switch_mm(). Membarrier requires a barrier after storing to
3628 * rq->curr, before returning to userspace, so provide them here:
3629 *
3630 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3631 * provided by mmdrop(),
3632 * - a sync_core for SYNC_CORE.
3633 */
3634 if (mm) {
3635 membarrier_mm_sync_core_before_usermode(mm);
3636 mmdrop(mm);
3637 }
3638 if (unlikely(prev_state == TASK_DEAD)) {
3639 if (prev->sched_class->task_dead)
3640 prev->sched_class->task_dead(prev);
3641
3642 /*
3643 * Remove function-return probe instances associated with this
3644 * task and put them back on the free list.
3645 */
3646 kprobe_flush_task(prev);
3647
3648 /* Task is done with its stack. */
3649 put_task_stack(prev);
3650
3651 put_task_struct_rcu_user(prev);
3652 }
3653
3654 tick_nohz_task_switch();
3655 return rq;
3656}
3657
3658#ifdef CONFIG_SMP
3659
3660/* rq->lock is NOT held, but preemption is disabled */
3661static void __balance_callback(struct rq *rq)
3662{
3663 struct callback_head *head, *next;
3664 void (*func)(struct rq *rq);
3665 unsigned long flags;
3666
3667 raw_spin_lock_irqsave(&rq->lock, flags);
3668 head = rq->balance_callback;
3669 rq->balance_callback = NULL;
3670 while (head) {
3671 func = (void (*)(struct rq *))head->func;
3672 next = head->next;
3673 head->next = NULL;
3674 head = next;
3675
3676 func(rq);
3677 }
3678 raw_spin_unlock_irqrestore(&rq->lock, flags);
3679}
3680
3681static inline void balance_callback(struct rq *rq)
3682{
3683 if (unlikely(rq->balance_callback))
3684 __balance_callback(rq);
3685}
3686
3687#else
3688
3689static inline void balance_callback(struct rq *rq)
3690{
3691}
3692
3693#endif
3694
3695/**
3696 * schedule_tail - first thing a freshly forked thread must call.
3697 * @prev: the thread we just switched away from.
3698 */
3699asmlinkage __visible void schedule_tail(struct task_struct *prev)
3700 __releases(rq->lock)
3701{
3702 struct rq *rq;
3703
3704 /*
3705 * New tasks start with FORK_PREEMPT_COUNT, see there and
3706 * finish_task_switch() for details.
3707 *
3708 * finish_task_switch() will drop rq->lock() and lower preempt_count
3709 * and the preempt_enable() will end up enabling preemption (on
3710 * PREEMPT_COUNT kernels).
3711 */
3712
3713 rq = finish_task_switch(prev);
3714 balance_callback(rq);
3715 preempt_enable();
3716
3717 if (current->set_child_tid)
3718 put_user(task_pid_vnr(current), current->set_child_tid);
3719
3720 calculate_sigpending();
3721}
3722
3723/*
3724 * context_switch - switch to the new MM and the new thread's register state.
3725 */
3726static __always_inline struct rq *
3727context_switch(struct rq *rq, struct task_struct *prev,
3728 struct task_struct *next, struct rq_flags *rf)
3729{
3730 prepare_task_switch(rq, prev, next);
3731
3732 /*
3733 * For paravirt, this is coupled with an exit in switch_to to
3734 * combine the page table reload and the switch backend into
3735 * one hypercall.
3736 */
3737 arch_start_context_switch(prev);
3738
3739 /*
3740 * kernel -> kernel lazy + transfer active
3741 * user -> kernel lazy + mmgrab() active
3742 *
3743 * kernel -> user switch + mmdrop() active
3744 * user -> user switch
3745 */
3746 if (!next->mm) { // to kernel
3747 enter_lazy_tlb(prev->active_mm, next);
3748
3749 next->active_mm = prev->active_mm;
3750 if (prev->mm) // from user
3751 mmgrab(prev->active_mm);
3752 else
3753 prev->active_mm = NULL;
3754 } else { // to user
3755 membarrier_switch_mm(rq, prev->active_mm, next->mm);
3756 /*
3757 * sys_membarrier() requires an smp_mb() between setting
3758 * rq->curr / membarrier_switch_mm() and returning to userspace.
3759 *
3760 * The below provides this either through switch_mm(), or in
3761 * case 'prev->active_mm == next->mm' through
3762 * finish_task_switch()'s mmdrop().
3763 */
3764 switch_mm_irqs_off(prev->active_mm, next->mm, next);
3765
3766 if (!prev->mm) { // from kernel
3767 /* will mmdrop() in finish_task_switch(). */
3768 rq->prev_mm = prev->active_mm;
3769 prev->active_mm = NULL;
3770 }
3771 }
3772
3773 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3774
3775 prepare_lock_switch(rq, next, rf);
3776
3777 /* Here we just switch the register state and the stack. */
3778 switch_to(prev, next, prev);
3779 barrier();
3780
3781 return finish_task_switch(prev);
3782}
3783
3784/*
3785 * nr_running and nr_context_switches:
3786 *
3787 * externally visible scheduler statistics: current number of runnable
3788 * threads, total number of context switches performed since bootup.
3789 */
3790unsigned long nr_running(void)
3791{
3792 unsigned long i, sum = 0;
3793
3794 for_each_online_cpu(i)
3795 sum += cpu_rq(i)->nr_running;
3796
3797 return sum;
3798}
3799
3800/*
3801 * Check if only the current task is running on the CPU.
3802 *
3803 * Caution: this function does not check that the caller has disabled
3804 * preemption, thus the result might have a time-of-check-to-time-of-use
3805 * race. The caller is responsible to use it correctly, for example:
3806 *
3807 * - from a non-preemptible section (of course)
3808 *
3809 * - from a thread that is bound to a single CPU
3810 *
3811 * - in a loop with very short iterations (e.g. a polling loop)
3812 */
3813bool single_task_running(void)
3814{
3815 return raw_rq()->nr_running == 1;
3816}
3817EXPORT_SYMBOL(single_task_running);
3818
3819unsigned long long nr_context_switches(void)
3820{
3821 int i;
3822 unsigned long long sum = 0;
3823
3824 for_each_possible_cpu(i)
3825 sum += cpu_rq(i)->nr_switches;
3826
3827 return sum;
3828}
3829
3830/*
3831 * Consumers of these two interfaces, like for example the cpuidle menu
3832 * governor, are using nonsensical data. Preferring shallow idle state selection
3833 * for a CPU that has IO-wait which might not even end up running the task when
3834 * it does become runnable.
3835 */
3836
3837unsigned long nr_iowait_cpu(int cpu)
3838{
3839 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3840}
3841
3842/*
3843 * IO-wait accounting, and how its mostly bollocks (on SMP).
3844 *
3845 * The idea behind IO-wait account is to account the idle time that we could
3846 * have spend running if it were not for IO. That is, if we were to improve the
3847 * storage performance, we'd have a proportional reduction in IO-wait time.
3848 *
3849 * This all works nicely on UP, where, when a task blocks on IO, we account
3850 * idle time as IO-wait, because if the storage were faster, it could've been
3851 * running and we'd not be idle.
3852 *
3853 * This has been extended to SMP, by doing the same for each CPU. This however
3854 * is broken.
3855 *
3856 * Imagine for instance the case where two tasks block on one CPU, only the one
3857 * CPU will have IO-wait accounted, while the other has regular idle. Even
3858 * though, if the storage were faster, both could've ran at the same time,
3859 * utilising both CPUs.
3860 *
3861 * This means, that when looking globally, the current IO-wait accounting on
3862 * SMP is a lower bound, by reason of under accounting.
3863 *
3864 * Worse, since the numbers are provided per CPU, they are sometimes
3865 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3866 * associated with any one particular CPU, it can wake to another CPU than it
3867 * blocked on. This means the per CPU IO-wait number is meaningless.
3868 *
3869 * Task CPU affinities can make all that even more 'interesting'.
3870 */
3871
3872unsigned long nr_iowait(void)
3873{
3874 unsigned long i, sum = 0;
3875
3876 for_each_possible_cpu(i)
3877 sum += nr_iowait_cpu(i);
3878
3879 return sum;
3880}
3881
3882#ifdef CONFIG_SMP
3883
3884/*
3885 * sched_exec - execve() is a valuable balancing opportunity, because at
3886 * this point the task has the smallest effective memory and cache footprint.
3887 */
3888void sched_exec(void)
3889{
3890 struct task_struct *p = current;
3891 unsigned long flags;
3892 int dest_cpu;
3893
3894 raw_spin_lock_irqsave(&p->pi_lock, flags);
3895 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3896 if (dest_cpu == smp_processor_id())
3897 goto unlock;
3898
3899 if (likely(cpu_active(dest_cpu))) {
3900 struct migration_arg arg = { p, dest_cpu };
3901
3902 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3903 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3904 return;
3905 }
3906unlock:
3907 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3908}
3909
3910#endif
3911
3912DEFINE_PER_CPU(struct kernel_stat, kstat);
3913DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3914
3915EXPORT_PER_CPU_SYMBOL(kstat);
3916EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3917
3918/*
3919 * The function fair_sched_class.update_curr accesses the struct curr
3920 * and its field curr->exec_start; when called from task_sched_runtime(),
3921 * we observe a high rate of cache misses in practice.
3922 * Prefetching this data results in improved performance.
3923 */
3924static inline void prefetch_curr_exec_start(struct task_struct *p)
3925{
3926#ifdef CONFIG_FAIR_GROUP_SCHED
3927 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3928#else
3929 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3930#endif
3931 prefetch(curr);
3932 prefetch(&curr->exec_start);
3933}
3934
3935/*
3936 * Return accounted runtime for the task.
3937 * In case the task is currently running, return the runtime plus current's
3938 * pending runtime that have not been accounted yet.
3939 */
3940unsigned long long task_sched_runtime(struct task_struct *p)
3941{
3942 struct rq_flags rf;
3943 struct rq *rq;
3944 u64 ns;
3945
3946#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3947 /*
3948 * 64-bit doesn't need locks to atomically read a 64-bit value.
3949 * So we have a optimization chance when the task's delta_exec is 0.
3950 * Reading ->on_cpu is racy, but this is ok.
3951 *
3952 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3953 * If we race with it entering CPU, unaccounted time is 0. This is
3954 * indistinguishable from the read occurring a few cycles earlier.
3955 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3956 * been accounted, so we're correct here as well.
3957 */
3958 if (!p->on_cpu || !task_on_rq_queued(p))
3959 return p->se.sum_exec_runtime;
3960#endif
3961
3962 rq = task_rq_lock(p, &rf);
3963 /*
3964 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3965 * project cycles that may never be accounted to this
3966 * thread, breaking clock_gettime().
3967 */
3968 if (task_current(rq, p) && task_on_rq_queued(p)) {
3969 prefetch_curr_exec_start(p);
3970 update_rq_clock(rq);
3971 p->sched_class->update_curr(rq);
3972 }
3973 ns = p->se.sum_exec_runtime;
3974 task_rq_unlock(rq, p, &rf);
3975
3976 return ns;
3977}
3978
3979/*
3980 * This function gets called by the timer code, with HZ frequency.
3981 * We call it with interrupts disabled.
3982 */
3983void scheduler_tick(void)
3984{
3985 int cpu = smp_processor_id();
3986 struct rq *rq = cpu_rq(cpu);
3987 struct task_struct *curr = rq->curr;
3988 struct rq_flags rf;
3989 unsigned long thermal_pressure;
3990
3991 arch_scale_freq_tick();
3992 sched_clock_tick();
3993
3994 rq_lock(rq, &rf);
3995
3996 update_rq_clock(rq);
3997 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
3998 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
3999 curr->sched_class->task_tick(rq, curr, 0);
4000 calc_global_load_tick(rq);
4001 psi_task_tick(rq);
4002
4003 rq_unlock(rq, &rf);
4004
4005 perf_event_task_tick();
4006
4007#ifdef CONFIG_SMP
4008 rq->idle_balance = idle_cpu(cpu);
4009 trigger_load_balance(rq);
4010#endif
4011}
4012
4013#ifdef CONFIG_NO_HZ_FULL
4014
4015struct tick_work {
4016 int cpu;
4017 atomic_t state;
4018 struct delayed_work work;
4019};
4020/* Values for ->state, see diagram below. */
4021#define TICK_SCHED_REMOTE_OFFLINE 0
4022#define TICK_SCHED_REMOTE_OFFLINING 1
4023#define TICK_SCHED_REMOTE_RUNNING 2
4024
4025/*
4026 * State diagram for ->state:
4027 *
4028 *
4029 * TICK_SCHED_REMOTE_OFFLINE
4030 * | ^
4031 * | |
4032 * | | sched_tick_remote()
4033 * | |
4034 * | |
4035 * +--TICK_SCHED_REMOTE_OFFLINING
4036 * | ^
4037 * | |
4038 * sched_tick_start() | | sched_tick_stop()
4039 * | |
4040 * V |
4041 * TICK_SCHED_REMOTE_RUNNING
4042 *
4043 *
4044 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
4045 * and sched_tick_start() are happy to leave the state in RUNNING.
4046 */
4047
4048static struct tick_work __percpu *tick_work_cpu;
4049
4050static void sched_tick_remote(struct work_struct *work)
4051{
4052 struct delayed_work *dwork = to_delayed_work(work);
4053 struct tick_work *twork = container_of(dwork, struct tick_work, work);
4054 int cpu = twork->cpu;
4055 struct rq *rq = cpu_rq(cpu);
4056 struct task_struct *curr;
4057 struct rq_flags rf;
4058 u64 delta;
4059 int os;
4060
4061 /*
4062 * Handle the tick only if it appears the remote CPU is running in full
4063 * dynticks mode. The check is racy by nature, but missing a tick or
4064 * having one too much is no big deal because the scheduler tick updates
4065 * statistics and checks timeslices in a time-independent way, regardless
4066 * of when exactly it is running.
4067 */
4068 if (!tick_nohz_tick_stopped_cpu(cpu))
4069 goto out_requeue;
4070
4071 rq_lock_irq(rq, &rf);
4072 curr = rq->curr;
4073 if (cpu_is_offline(cpu))
4074 goto out_unlock;
4075
4076 update_rq_clock(rq);
4077
4078 if (!is_idle_task(curr)) {
4079 /*
4080 * Make sure the next tick runs within a reasonable
4081 * amount of time.
4082 */
4083 delta = rq_clock_task(rq) - curr->se.exec_start;
4084 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
4085 }
4086 curr->sched_class->task_tick(rq, curr, 0);
4087
4088 calc_load_nohz_remote(rq);
4089out_unlock:
4090 rq_unlock_irq(rq, &rf);
4091out_requeue:
4092
4093 /*
4094 * Run the remote tick once per second (1Hz). This arbitrary
4095 * frequency is large enough to avoid overload but short enough
4096 * to keep scheduler internal stats reasonably up to date. But
4097 * first update state to reflect hotplug activity if required.
4098 */
4099 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
4100 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
4101 if (os == TICK_SCHED_REMOTE_RUNNING)
4102 queue_delayed_work(system_unbound_wq, dwork, HZ);
4103}
4104
4105static void sched_tick_start(int cpu)
4106{
4107 int os;
4108 struct tick_work *twork;
4109
4110 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4111 return;
4112
4113 WARN_ON_ONCE(!tick_work_cpu);
4114
4115 twork = per_cpu_ptr(tick_work_cpu, cpu);
4116 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
4117 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
4118 if (os == TICK_SCHED_REMOTE_OFFLINE) {
4119 twork->cpu = cpu;
4120 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
4121 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
4122 }
4123}
4124
4125#ifdef CONFIG_HOTPLUG_CPU
4126static void sched_tick_stop(int cpu)
4127{
4128 struct tick_work *twork;
4129 int os;
4130
4131 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4132 return;
4133
4134 WARN_ON_ONCE(!tick_work_cpu);
4135
4136 twork = per_cpu_ptr(tick_work_cpu, cpu);
4137 /* There cannot be competing actions, but don't rely on stop-machine. */
4138 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
4139 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
4140 /* Don't cancel, as this would mess up the state machine. */
4141}
4142#endif /* CONFIG_HOTPLUG_CPU */
4143
4144int __init sched_tick_offload_init(void)
4145{
4146 tick_work_cpu = alloc_percpu(struct tick_work);
4147 BUG_ON(!tick_work_cpu);
4148 return 0;
4149}
4150
4151#else /* !CONFIG_NO_HZ_FULL */
4152static inline void sched_tick_start(int cpu) { }
4153static inline void sched_tick_stop(int cpu) { }
4154#endif
4155
4156#if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
4157 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
4158/*
4159 * If the value passed in is equal to the current preempt count
4160 * then we just disabled preemption. Start timing the latency.
4161 */
4162static inline void preempt_latency_start(int val)
4163{
4164 if (preempt_count() == val) {
4165 unsigned long ip = get_lock_parent_ip();
4166#ifdef CONFIG_DEBUG_PREEMPT
4167 current->preempt_disable_ip = ip;
4168#endif
4169 trace_preempt_off(CALLER_ADDR0, ip);
4170 }
4171}
4172
4173void preempt_count_add(int val)
4174{
4175#ifdef CONFIG_DEBUG_PREEMPT
4176 /*
4177 * Underflow?
4178 */
4179 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4180 return;
4181#endif
4182 __preempt_count_add(val);
4183#ifdef CONFIG_DEBUG_PREEMPT
4184 /*
4185 * Spinlock count overflowing soon?
4186 */
4187 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4188 PREEMPT_MASK - 10);
4189#endif
4190 preempt_latency_start(val);
4191}
4192EXPORT_SYMBOL(preempt_count_add);
4193NOKPROBE_SYMBOL(preempt_count_add);
4194
4195/*
4196 * If the value passed in equals to the current preempt count
4197 * then we just enabled preemption. Stop timing the latency.
4198 */
4199static inline void preempt_latency_stop(int val)
4200{
4201 if (preempt_count() == val)
4202 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
4203}
4204
4205void preempt_count_sub(int val)
4206{
4207#ifdef CONFIG_DEBUG_PREEMPT
4208 /*
4209 * Underflow?
4210 */
4211 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4212 return;
4213 /*
4214 * Is the spinlock portion underflowing?
4215 */
4216 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4217 !(preempt_count() & PREEMPT_MASK)))
4218 return;
4219#endif
4220
4221 preempt_latency_stop(val);
4222 __preempt_count_sub(val);
4223}
4224EXPORT_SYMBOL(preempt_count_sub);
4225NOKPROBE_SYMBOL(preempt_count_sub);
4226
4227#else
4228static inline void preempt_latency_start(int val) { }
4229static inline void preempt_latency_stop(int val) { }
4230#endif
4231
4232static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
4233{
4234#ifdef CONFIG_DEBUG_PREEMPT
4235 return p->preempt_disable_ip;
4236#else
4237 return 0;
4238#endif
4239}
4240
4241/*
4242 * Print scheduling while atomic bug:
4243 */
4244static noinline void __schedule_bug(struct task_struct *prev)
4245{
4246 /* Save this before calling printk(), since that will clobber it */
4247 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
4248
4249 if (oops_in_progress)
4250 return;
4251
4252 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4253 prev->comm, prev->pid, preempt_count());
4254
4255 debug_show_held_locks(prev);
4256 print_modules();
4257 if (irqs_disabled())
4258 print_irqtrace_events(prev);
4259 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
4260 && in_atomic_preempt_off()) {
4261 pr_err("Preemption disabled at:");
4262 print_ip_sym(KERN_ERR, preempt_disable_ip);
4263 }
4264 if (panic_on_warn)
4265 panic("scheduling while atomic\n");
4266
4267 dump_stack();
4268 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4269}
4270
4271/*
4272 * Various schedule()-time debugging checks and statistics:
4273 */
4274static inline void schedule_debug(struct task_struct *prev, bool preempt)
4275{
4276#ifdef CONFIG_SCHED_STACK_END_CHECK
4277 if (task_stack_end_corrupted(prev))
4278 panic("corrupted stack end detected inside scheduler\n");
4279
4280 if (task_scs_end_corrupted(prev))
4281 panic("corrupted shadow stack detected inside scheduler\n");
4282#endif
4283
4284#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
4285 if (!preempt && prev->state && prev->non_block_count) {
4286 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
4287 prev->comm, prev->pid, prev->non_block_count);
4288 dump_stack();
4289 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4290 }
4291#endif
4292
4293 if (unlikely(in_atomic_preempt_off())) {
4294 __schedule_bug(prev);
4295 preempt_count_set(PREEMPT_DISABLED);
4296 }
4297 rcu_sleep_check();
4298
4299 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4300
4301 schedstat_inc(this_rq()->sched_count);
4302}
4303
4304static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
4305 struct rq_flags *rf)
4306{
4307#ifdef CONFIG_SMP
4308 const struct sched_class *class;
4309 /*
4310 * We must do the balancing pass before put_prev_task(), such
4311 * that when we release the rq->lock the task is in the same
4312 * state as before we took rq->lock.
4313 *
4314 * We can terminate the balance pass as soon as we know there is
4315 * a runnable task of @class priority or higher.
4316 */
4317 for_class_range(class, prev->sched_class, &idle_sched_class) {
4318 if (class->balance(rq, prev, rf))
4319 break;
4320 }
4321#endif
4322
4323 put_prev_task(rq, prev);
4324}
4325
4326/*
4327 * Pick up the highest-prio task:
4328 */
4329static inline struct task_struct *
4330pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
4331{
4332 const struct sched_class *class;
4333 struct task_struct *p;
4334
4335 /*
4336 * Optimization: we know that if all tasks are in the fair class we can
4337 * call that function directly, but only if the @prev task wasn't of a
4338 * higher scheduling class, because otherwise those loose the
4339 * opportunity to pull in more work from other CPUs.
4340 */
4341 if (likely(prev->sched_class <= &fair_sched_class &&
4342 rq->nr_running == rq->cfs.h_nr_running)) {
4343
4344 p = pick_next_task_fair(rq, prev, rf);
4345 if (unlikely(p == RETRY_TASK))
4346 goto restart;
4347
4348 /* Assumes fair_sched_class->next == idle_sched_class */
4349 if (!p) {
4350 put_prev_task(rq, prev);
4351 p = pick_next_task_idle(rq);
4352 }
4353
4354 return p;
4355 }
4356
4357restart:
4358 put_prev_task_balance(rq, prev, rf);
4359
4360 for_each_class(class) {
4361 p = class->pick_next_task(rq);
4362 if (p)
4363 return p;
4364 }
4365
4366 /* The idle class should always have a runnable task: */
4367 BUG();
4368}
4369
4370/*
4371 * __schedule() is the main scheduler function.
4372 *
4373 * The main means of driving the scheduler and thus entering this function are:
4374 *
4375 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4376 *
4377 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4378 * paths. For example, see arch/x86/entry_64.S.
4379 *
4380 * To drive preemption between tasks, the scheduler sets the flag in timer
4381 * interrupt handler scheduler_tick().
4382 *
4383 * 3. Wakeups don't really cause entry into schedule(). They add a
4384 * task to the run-queue and that's it.
4385 *
4386 * Now, if the new task added to the run-queue preempts the current
4387 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4388 * called on the nearest possible occasion:
4389 *
4390 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4391 *
4392 * - in syscall or exception context, at the next outmost
4393 * preempt_enable(). (this might be as soon as the wake_up()'s
4394 * spin_unlock()!)
4395 *
4396 * - in IRQ context, return from interrupt-handler to
4397 * preemptible context
4398 *
4399 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4400 * then at the next:
4401 *
4402 * - cond_resched() call
4403 * - explicit schedule() call
4404 * - return from syscall or exception to user-space
4405 * - return from interrupt-handler to user-space
4406 *
4407 * WARNING: must be called with preemption disabled!
4408 */
4409static void __sched notrace __schedule(bool preempt)
4410{
4411 struct task_struct *prev, *next;
4412 unsigned long *switch_count;
4413 unsigned long prev_state;
4414 struct rq_flags rf;
4415 struct rq *rq;
4416 int cpu;
4417
4418 cpu = smp_processor_id();
4419 rq = cpu_rq(cpu);
4420 prev = rq->curr;
4421
4422 schedule_debug(prev, preempt);
4423
4424 if (sched_feat(HRTICK))
4425 hrtick_clear(rq);
4426
4427 local_irq_disable();
4428 rcu_note_context_switch(preempt);
4429
4430 /*
4431 * Make sure that signal_pending_state()->signal_pending() below
4432 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4433 * done by the caller to avoid the race with signal_wake_up():
4434 *
4435 * __set_current_state(@state) signal_wake_up()
4436 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
4437 * wake_up_state(p, state)
4438 * LOCK rq->lock LOCK p->pi_state
4439 * smp_mb__after_spinlock() smp_mb__after_spinlock()
4440 * if (signal_pending_state()) if (p->state & @state)
4441 *
4442 * Also, the membarrier system call requires a full memory barrier
4443 * after coming from user-space, before storing to rq->curr.
4444 */
4445 rq_lock(rq, &rf);
4446 smp_mb__after_spinlock();
4447
4448 /* Promote REQ to ACT */
4449 rq->clock_update_flags <<= 1;
4450 update_rq_clock(rq);
4451
4452 switch_count = &prev->nivcsw;
4453
4454 /*
4455 * We must load prev->state once (task_struct::state is volatile), such
4456 * that:
4457 *
4458 * - we form a control dependency vs deactivate_task() below.
4459 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
4460 */
4461 prev_state = prev->state;
4462 if (!preempt && prev_state) {
4463 if (signal_pending_state(prev_state, prev)) {
4464 prev->state = TASK_RUNNING;
4465 } else {
4466 prev->sched_contributes_to_load =
4467 (prev_state & TASK_UNINTERRUPTIBLE) &&
4468 !(prev_state & TASK_NOLOAD) &&
4469 !(prev->flags & PF_FROZEN);
4470
4471 if (prev->sched_contributes_to_load)
4472 rq->nr_uninterruptible++;
4473
4474 /*
4475 * __schedule() ttwu()
4476 * prev_state = prev->state; if (p->on_rq && ...)
4477 * if (prev_state) goto out;
4478 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
4479 * p->state = TASK_WAKING
4480 *
4481 * Where __schedule() and ttwu() have matching control dependencies.
4482 *
4483 * After this, schedule() must not care about p->state any more.
4484 */
4485 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
4486
4487 if (prev->in_iowait) {
4488 atomic_inc(&rq->nr_iowait);
4489 delayacct_blkio_start();
4490 }
4491 }
4492 switch_count = &prev->nvcsw;
4493 }
4494
4495 next = pick_next_task(rq, prev, &rf);
4496 clear_tsk_need_resched(prev);
4497 clear_preempt_need_resched();
4498
4499 if (likely(prev != next)) {
4500 rq->nr_switches++;
4501 /*
4502 * RCU users of rcu_dereference(rq->curr) may not see
4503 * changes to task_struct made by pick_next_task().
4504 */
4505 RCU_INIT_POINTER(rq->curr, next);
4506 /*
4507 * The membarrier system call requires each architecture
4508 * to have a full memory barrier after updating
4509 * rq->curr, before returning to user-space.
4510 *
4511 * Here are the schemes providing that barrier on the
4512 * various architectures:
4513 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4514 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4515 * - finish_lock_switch() for weakly-ordered
4516 * architectures where spin_unlock is a full barrier,
4517 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4518 * is a RELEASE barrier),
4519 */
4520 ++*switch_count;
4521
4522 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
4523
4524 trace_sched_switch(preempt, prev, next);
4525
4526 /* Also unlocks the rq: */
4527 rq = context_switch(rq, prev, next, &rf);
4528 } else {
4529 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4530 rq_unlock_irq(rq, &rf);
4531 }
4532
4533 balance_callback(rq);
4534}
4535
4536void __noreturn do_task_dead(void)
4537{
4538 /* Causes final put_task_struct in finish_task_switch(): */
4539 set_special_state(TASK_DEAD);
4540
4541 /* Tell freezer to ignore us: */
4542 current->flags |= PF_NOFREEZE;
4543
4544 __schedule(false);
4545 BUG();
4546
4547 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4548 for (;;)
4549 cpu_relax();
4550}
4551
4552static inline void sched_submit_work(struct task_struct *tsk)
4553{
4554 if (!tsk->state)
4555 return;
4556
4557 /*
4558 * If a worker went to sleep, notify and ask workqueue whether
4559 * it wants to wake up a task to maintain concurrency.
4560 * As this function is called inside the schedule() context,
4561 * we disable preemption to avoid it calling schedule() again
4562 * in the possible wakeup of a kworker and because wq_worker_sleeping()
4563 * requires it.
4564 */
4565 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4566 preempt_disable();
4567 if (tsk->flags & PF_WQ_WORKER)
4568 wq_worker_sleeping(tsk);
4569 else
4570 io_wq_worker_sleeping(tsk);
4571 preempt_enable_no_resched();
4572 }
4573
4574 if (tsk_is_pi_blocked(tsk))
4575 return;
4576
4577 /*
4578 * If we are going to sleep and we have plugged IO queued,
4579 * make sure to submit it to avoid deadlocks.
4580 */
4581 if (blk_needs_flush_plug(tsk))
4582 blk_schedule_flush_plug(tsk);
4583}
4584
4585static void sched_update_worker(struct task_struct *tsk)
4586{
4587 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4588 if (tsk->flags & PF_WQ_WORKER)
4589 wq_worker_running(tsk);
4590 else
4591 io_wq_worker_running(tsk);
4592 }
4593}
4594
4595asmlinkage __visible void __sched schedule(void)
4596{
4597 struct task_struct *tsk = current;
4598
4599 sched_submit_work(tsk);
4600 do {
4601 preempt_disable();
4602 __schedule(false);
4603 sched_preempt_enable_no_resched();
4604 } while (need_resched());
4605 sched_update_worker(tsk);
4606}
4607EXPORT_SYMBOL(schedule);
4608
4609/*
4610 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4611 * state (have scheduled out non-voluntarily) by making sure that all
4612 * tasks have either left the run queue or have gone into user space.
4613 * As idle tasks do not do either, they must not ever be preempted
4614 * (schedule out non-voluntarily).
4615 *
4616 * schedule_idle() is similar to schedule_preempt_disable() except that it
4617 * never enables preemption because it does not call sched_submit_work().
4618 */
4619void __sched schedule_idle(void)
4620{
4621 /*
4622 * As this skips calling sched_submit_work(), which the idle task does
4623 * regardless because that function is a nop when the task is in a
4624 * TASK_RUNNING state, make sure this isn't used someplace that the
4625 * current task can be in any other state. Note, idle is always in the
4626 * TASK_RUNNING state.
4627 */
4628 WARN_ON_ONCE(current->state);
4629 do {
4630 __schedule(false);
4631 } while (need_resched());
4632}
4633
4634#ifdef CONFIG_CONTEXT_TRACKING
4635asmlinkage __visible void __sched schedule_user(void)
4636{
4637 /*
4638 * If we come here after a random call to set_need_resched(),
4639 * or we have been woken up remotely but the IPI has not yet arrived,
4640 * we haven't yet exited the RCU idle mode. Do it here manually until
4641 * we find a better solution.
4642 *
4643 * NB: There are buggy callers of this function. Ideally we
4644 * should warn if prev_state != CONTEXT_USER, but that will trigger
4645 * too frequently to make sense yet.
4646 */
4647 enum ctx_state prev_state = exception_enter();
4648 schedule();
4649 exception_exit(prev_state);
4650}
4651#endif
4652
4653/**
4654 * schedule_preempt_disabled - called with preemption disabled
4655 *
4656 * Returns with preemption disabled. Note: preempt_count must be 1
4657 */
4658void __sched schedule_preempt_disabled(void)
4659{
4660 sched_preempt_enable_no_resched();
4661 schedule();
4662 preempt_disable();
4663}
4664
4665static void __sched notrace preempt_schedule_common(void)
4666{
4667 do {
4668 /*
4669 * Because the function tracer can trace preempt_count_sub()
4670 * and it also uses preempt_enable/disable_notrace(), if
4671 * NEED_RESCHED is set, the preempt_enable_notrace() called
4672 * by the function tracer will call this function again and
4673 * cause infinite recursion.
4674 *
4675 * Preemption must be disabled here before the function
4676 * tracer can trace. Break up preempt_disable() into two
4677 * calls. One to disable preemption without fear of being
4678 * traced. The other to still record the preemption latency,
4679 * which can also be traced by the function tracer.
4680 */
4681 preempt_disable_notrace();
4682 preempt_latency_start(1);
4683 __schedule(true);
4684 preempt_latency_stop(1);
4685 preempt_enable_no_resched_notrace();
4686
4687 /*
4688 * Check again in case we missed a preemption opportunity
4689 * between schedule and now.
4690 */
4691 } while (need_resched());
4692}
4693
4694#ifdef CONFIG_PREEMPTION
4695/*
4696 * This is the entry point to schedule() from in-kernel preemption
4697 * off of preempt_enable.
4698 */
4699asmlinkage __visible void __sched notrace preempt_schedule(void)
4700{
4701 /*
4702 * If there is a non-zero preempt_count or interrupts are disabled,
4703 * we do not want to preempt the current task. Just return..
4704 */
4705 if (likely(!preemptible()))
4706 return;
4707
4708 preempt_schedule_common();
4709}
4710NOKPROBE_SYMBOL(preempt_schedule);
4711EXPORT_SYMBOL(preempt_schedule);
4712
4713/**
4714 * preempt_schedule_notrace - preempt_schedule called by tracing
4715 *
4716 * The tracing infrastructure uses preempt_enable_notrace to prevent
4717 * recursion and tracing preempt enabling caused by the tracing
4718 * infrastructure itself. But as tracing can happen in areas coming
4719 * from userspace or just about to enter userspace, a preempt enable
4720 * can occur before user_exit() is called. This will cause the scheduler
4721 * to be called when the system is still in usermode.
4722 *
4723 * To prevent this, the preempt_enable_notrace will use this function
4724 * instead of preempt_schedule() to exit user context if needed before
4725 * calling the scheduler.
4726 */
4727asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4728{
4729 enum ctx_state prev_ctx;
4730
4731 if (likely(!preemptible()))
4732 return;
4733
4734 do {
4735 /*
4736 * Because the function tracer can trace preempt_count_sub()
4737 * and it also uses preempt_enable/disable_notrace(), if
4738 * NEED_RESCHED is set, the preempt_enable_notrace() called
4739 * by the function tracer will call this function again and
4740 * cause infinite recursion.
4741 *
4742 * Preemption must be disabled here before the function
4743 * tracer can trace. Break up preempt_disable() into two
4744 * calls. One to disable preemption without fear of being
4745 * traced. The other to still record the preemption latency,
4746 * which can also be traced by the function tracer.
4747 */
4748 preempt_disable_notrace();
4749 preempt_latency_start(1);
4750 /*
4751 * Needs preempt disabled in case user_exit() is traced
4752 * and the tracer calls preempt_enable_notrace() causing
4753 * an infinite recursion.
4754 */
4755 prev_ctx = exception_enter();
4756 __schedule(true);
4757 exception_exit(prev_ctx);
4758
4759 preempt_latency_stop(1);
4760 preempt_enable_no_resched_notrace();
4761 } while (need_resched());
4762}
4763EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4764
4765#endif /* CONFIG_PREEMPTION */
4766
4767/*
4768 * This is the entry point to schedule() from kernel preemption
4769 * off of irq context.
4770 * Note, that this is called and return with irqs disabled. This will
4771 * protect us against recursive calling from irq.
4772 */
4773asmlinkage __visible void __sched preempt_schedule_irq(void)
4774{
4775 enum ctx_state prev_state;
4776
4777 /* Catch callers which need to be fixed */
4778 BUG_ON(preempt_count() || !irqs_disabled());
4779
4780 prev_state = exception_enter();
4781
4782 do {
4783 preempt_disable();
4784 local_irq_enable();
4785 __schedule(true);
4786 local_irq_disable();
4787 sched_preempt_enable_no_resched();
4788 } while (need_resched());
4789
4790 exception_exit(prev_state);
4791}
4792
4793int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4794 void *key)
4795{
4796 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
4797 return try_to_wake_up(curr->private, mode, wake_flags);
4798}
4799EXPORT_SYMBOL(default_wake_function);
4800
4801#ifdef CONFIG_RT_MUTEXES
4802
4803static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4804{
4805 if (pi_task)
4806 prio = min(prio, pi_task->prio);
4807
4808 return prio;
4809}
4810
4811static inline int rt_effective_prio(struct task_struct *p, int prio)
4812{
4813 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4814
4815 return __rt_effective_prio(pi_task, prio);
4816}
4817
4818/*
4819 * rt_mutex_setprio - set the current priority of a task
4820 * @p: task to boost
4821 * @pi_task: donor task
4822 *
4823 * This function changes the 'effective' priority of a task. It does
4824 * not touch ->normal_prio like __setscheduler().
4825 *
4826 * Used by the rt_mutex code to implement priority inheritance
4827 * logic. Call site only calls if the priority of the task changed.
4828 */
4829void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4830{
4831 int prio, oldprio, queued, running, queue_flag =
4832 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4833 const struct sched_class *prev_class;
4834 struct rq_flags rf;
4835 struct rq *rq;
4836
4837 /* XXX used to be waiter->prio, not waiter->task->prio */
4838 prio = __rt_effective_prio(pi_task, p->normal_prio);
4839
4840 /*
4841 * If nothing changed; bail early.
4842 */
4843 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4844 return;
4845
4846 rq = __task_rq_lock(p, &rf);
4847 update_rq_clock(rq);
4848 /*
4849 * Set under pi_lock && rq->lock, such that the value can be used under
4850 * either lock.
4851 *
4852 * Note that there is loads of tricky to make this pointer cache work
4853 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4854 * ensure a task is de-boosted (pi_task is set to NULL) before the
4855 * task is allowed to run again (and can exit). This ensures the pointer
4856 * points to a blocked task -- which guaratees the task is present.
4857 */
4858 p->pi_top_task = pi_task;
4859
4860 /*
4861 * For FIFO/RR we only need to set prio, if that matches we're done.
4862 */
4863 if (prio == p->prio && !dl_prio(prio))
4864 goto out_unlock;
4865
4866 /*
4867 * Idle task boosting is a nono in general. There is one
4868 * exception, when PREEMPT_RT and NOHZ is active:
4869 *
4870 * The idle task calls get_next_timer_interrupt() and holds
4871 * the timer wheel base->lock on the CPU and another CPU wants
4872 * to access the timer (probably to cancel it). We can safely
4873 * ignore the boosting request, as the idle CPU runs this code
4874 * with interrupts disabled and will complete the lock
4875 * protected section without being interrupted. So there is no
4876 * real need to boost.
4877 */
4878 if (unlikely(p == rq->idle)) {
4879 WARN_ON(p != rq->curr);
4880 WARN_ON(p->pi_blocked_on);
4881 goto out_unlock;
4882 }
4883
4884 trace_sched_pi_setprio(p, pi_task);
4885 oldprio = p->prio;
4886
4887 if (oldprio == prio)
4888 queue_flag &= ~DEQUEUE_MOVE;
4889
4890 prev_class = p->sched_class;
4891 queued = task_on_rq_queued(p);
4892 running = task_current(rq, p);
4893 if (queued)
4894 dequeue_task(rq, p, queue_flag);
4895 if (running)
4896 put_prev_task(rq, p);
4897
4898 /*
4899 * Boosting condition are:
4900 * 1. -rt task is running and holds mutex A
4901 * --> -dl task blocks on mutex A
4902 *
4903 * 2. -dl task is running and holds mutex A
4904 * --> -dl task blocks on mutex A and could preempt the
4905 * running task
4906 */
4907 if (dl_prio(prio)) {
4908 if (!dl_prio(p->normal_prio) ||
4909 (pi_task && dl_prio(pi_task->prio) &&
4910 dl_entity_preempt(&pi_task->dl, &p->dl))) {
4911 p->dl.dl_boosted = 1;
4912 queue_flag |= ENQUEUE_REPLENISH;
4913 } else
4914 p->dl.dl_boosted = 0;
4915 p->sched_class = &dl_sched_class;
4916 } else if (rt_prio(prio)) {
4917 if (dl_prio(oldprio))
4918 p->dl.dl_boosted = 0;
4919 if (oldprio < prio)
4920 queue_flag |= ENQUEUE_HEAD;
4921 p->sched_class = &rt_sched_class;
4922 } else {
4923 if (dl_prio(oldprio))
4924 p->dl.dl_boosted = 0;
4925 if (rt_prio(oldprio))
4926 p->rt.timeout = 0;
4927 p->sched_class = &fair_sched_class;
4928 }
4929
4930 p->prio = prio;
4931
4932 if (queued)
4933 enqueue_task(rq, p, queue_flag);
4934 if (running)
4935 set_next_task(rq, p);
4936
4937 check_class_changed(rq, p, prev_class, oldprio);
4938out_unlock:
4939 /* Avoid rq from going away on us: */
4940 preempt_disable();
4941 __task_rq_unlock(rq, &rf);
4942
4943 balance_callback(rq);
4944 preempt_enable();
4945}
4946#else
4947static inline int rt_effective_prio(struct task_struct *p, int prio)
4948{
4949 return prio;
4950}
4951#endif
4952
4953void set_user_nice(struct task_struct *p, long nice)
4954{
4955 bool queued, running;
4956 int old_prio;
4957 struct rq_flags rf;
4958 struct rq *rq;
4959
4960 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
4961 return;
4962 /*
4963 * We have to be careful, if called from sys_setpriority(),
4964 * the task might be in the middle of scheduling on another CPU.
4965 */
4966 rq = task_rq_lock(p, &rf);
4967 update_rq_clock(rq);
4968
4969 /*
4970 * The RT priorities are set via sched_setscheduler(), but we still
4971 * allow the 'normal' nice value to be set - but as expected
4972 * it wont have any effect on scheduling until the task is
4973 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4974 */
4975 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4976 p->static_prio = NICE_TO_PRIO(nice);
4977 goto out_unlock;
4978 }
4979 queued = task_on_rq_queued(p);
4980 running = task_current(rq, p);
4981 if (queued)
4982 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
4983 if (running)
4984 put_prev_task(rq, p);
4985
4986 p->static_prio = NICE_TO_PRIO(nice);
4987 set_load_weight(p, true);
4988 old_prio = p->prio;
4989 p->prio = effective_prio(p);
4990
4991 if (queued)
4992 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
4993 if (running)
4994 set_next_task(rq, p);
4995
4996 /*
4997 * If the task increased its priority or is running and
4998 * lowered its priority, then reschedule its CPU:
4999 */
5000 p->sched_class->prio_changed(rq, p, old_prio);
5001
5002out_unlock:
5003 task_rq_unlock(rq, p, &rf);
5004}
5005EXPORT_SYMBOL(set_user_nice);
5006
5007/*
5008 * can_nice - check if a task can reduce its nice value
5009 * @p: task
5010 * @nice: nice value
5011 */
5012int can_nice(const struct task_struct *p, const int nice)
5013{
5014 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
5015 int nice_rlim = nice_to_rlimit(nice);
5016
5017 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5018 capable(CAP_SYS_NICE));
5019}
5020
5021#ifdef __ARCH_WANT_SYS_NICE
5022
5023/*
5024 * sys_nice - change the priority of the current process.
5025 * @increment: priority increment
5026 *
5027 * sys_setpriority is a more generic, but much slower function that
5028 * does similar things.
5029 */
5030SYSCALL_DEFINE1(nice, int, increment)
5031{
5032 long nice, retval;
5033
5034 /*
5035 * Setpriority might change our priority at the same moment.
5036 * We don't have to worry. Conceptually one call occurs first
5037 * and we have a single winner.
5038 */
5039 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
5040 nice = task_nice(current) + increment;
5041
5042 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
5043 if (increment < 0 && !can_nice(current, nice))
5044 return -EPERM;
5045
5046 retval = security_task_setnice(current, nice);
5047 if (retval)
5048 return retval;
5049
5050 set_user_nice(current, nice);
5051 return 0;
5052}
5053
5054#endif
5055
5056/**
5057 * task_prio - return the priority value of a given task.
5058 * @p: the task in question.
5059 *
5060 * Return: The priority value as seen by users in /proc.
5061 * RT tasks are offset by -200. Normal tasks are centered
5062 * around 0, value goes from -16 to +15.
5063 */
5064int task_prio(const struct task_struct *p)
5065{
5066 return p->prio - MAX_RT_PRIO;
5067}
5068
5069/**
5070 * idle_cpu - is a given CPU idle currently?
5071 * @cpu: the processor in question.
5072 *
5073 * Return: 1 if the CPU is currently idle. 0 otherwise.
5074 */
5075int idle_cpu(int cpu)
5076{
5077 struct rq *rq = cpu_rq(cpu);
5078
5079 if (rq->curr != rq->idle)
5080 return 0;
5081
5082 if (rq->nr_running)
5083 return 0;
5084
5085#ifdef CONFIG_SMP
5086 if (rq->ttwu_pending)
5087 return 0;
5088#endif
5089
5090 return 1;
5091}
5092
5093/**
5094 * available_idle_cpu - is a given CPU idle for enqueuing work.
5095 * @cpu: the CPU in question.
5096 *
5097 * Return: 1 if the CPU is currently idle. 0 otherwise.
5098 */
5099int available_idle_cpu(int cpu)
5100{
5101 if (!idle_cpu(cpu))
5102 return 0;
5103
5104 if (vcpu_is_preempted(cpu))
5105 return 0;
5106
5107 return 1;
5108}
5109
5110/**
5111 * idle_task - return the idle task for a given CPU.
5112 * @cpu: the processor in question.
5113 *
5114 * Return: The idle task for the CPU @cpu.
5115 */
5116struct task_struct *idle_task(int cpu)
5117{
5118 return cpu_rq(cpu)->idle;
5119}
5120
5121/**
5122 * find_process_by_pid - find a process with a matching PID value.
5123 * @pid: the pid in question.
5124 *
5125 * The task of @pid, if found. %NULL otherwise.
5126 */
5127static struct task_struct *find_process_by_pid(pid_t pid)
5128{
5129 return pid ? find_task_by_vpid(pid) : current;
5130}
5131
5132/*
5133 * sched_setparam() passes in -1 for its policy, to let the functions
5134 * it calls know not to change it.
5135 */
5136#define SETPARAM_POLICY -1
5137
5138static void __setscheduler_params(struct task_struct *p,
5139 const struct sched_attr *attr)
5140{
5141 int policy = attr->sched_policy;
5142
5143 if (policy == SETPARAM_POLICY)
5144 policy = p->policy;
5145
5146 p->policy = policy;
5147
5148 if (dl_policy(policy))
5149 __setparam_dl(p, attr);
5150 else if (fair_policy(policy))
5151 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
5152
5153 /*
5154 * __sched_setscheduler() ensures attr->sched_priority == 0 when
5155 * !rt_policy. Always setting this ensures that things like
5156 * getparam()/getattr() don't report silly values for !rt tasks.
5157 */
5158 p->rt_priority = attr->sched_priority;
5159 p->normal_prio = normal_prio(p);
5160 set_load_weight(p, true);
5161}
5162
5163/* Actually do priority change: must hold pi & rq lock. */
5164static void __setscheduler(struct rq *rq, struct task_struct *p,
5165 const struct sched_attr *attr, bool keep_boost)
5166{
5167 /*
5168 * If params can't change scheduling class changes aren't allowed
5169 * either.
5170 */
5171 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
5172 return;
5173
5174 __setscheduler_params(p, attr);
5175
5176 /*
5177 * Keep a potential priority boosting if called from
5178 * sched_setscheduler().
5179 */
5180 p->prio = normal_prio(p);
5181 if (keep_boost)
5182 p->prio = rt_effective_prio(p, p->prio);
5183
5184 if (dl_prio(p->prio))
5185 p->sched_class = &dl_sched_class;
5186 else if (rt_prio(p->prio))
5187 p->sched_class = &rt_sched_class;
5188 else
5189 p->sched_class = &fair_sched_class;
5190}
5191
5192/*
5193 * Check the target process has a UID that matches the current process's:
5194 */
5195static bool check_same_owner(struct task_struct *p)
5196{
5197 const struct cred *cred = current_cred(), *pcred;
5198 bool match;
5199
5200 rcu_read_lock();
5201 pcred = __task_cred(p);
5202 match = (uid_eq(cred->euid, pcred->euid) ||
5203 uid_eq(cred->euid, pcred->uid));
5204 rcu_read_unlock();
5205 return match;
5206}
5207
5208static int __sched_setscheduler(struct task_struct *p,
5209 const struct sched_attr *attr,
5210 bool user, bool pi)
5211{
5212 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
5213 MAX_RT_PRIO - 1 - attr->sched_priority;
5214 int retval, oldprio, oldpolicy = -1, queued, running;
5215 int new_effective_prio, policy = attr->sched_policy;
5216 const struct sched_class *prev_class;
5217 struct rq_flags rf;
5218 int reset_on_fork;
5219 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
5220 struct rq *rq;
5221
5222 /* The pi code expects interrupts enabled */
5223 BUG_ON(pi && in_interrupt());
5224recheck:
5225 /* Double check policy once rq lock held: */
5226 if (policy < 0) {
5227 reset_on_fork = p->sched_reset_on_fork;
5228 policy = oldpolicy = p->policy;
5229 } else {
5230 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
5231
5232 if (!valid_policy(policy))
5233 return -EINVAL;
5234 }
5235
5236 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
5237 return -EINVAL;
5238
5239 /*
5240 * Valid priorities for SCHED_FIFO and SCHED_RR are
5241 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5242 * SCHED_BATCH and SCHED_IDLE is 0.
5243 */
5244 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
5245 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
5246 return -EINVAL;
5247 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
5248 (rt_policy(policy) != (attr->sched_priority != 0)))
5249 return -EINVAL;
5250
5251 /*
5252 * Allow unprivileged RT tasks to decrease priority:
5253 */
5254 if (user && !capable(CAP_SYS_NICE)) {
5255 if (fair_policy(policy)) {
5256 if (attr->sched_nice < task_nice(p) &&
5257 !can_nice(p, attr->sched_nice))
5258 return -EPERM;
5259 }
5260
5261 if (rt_policy(policy)) {
5262 unsigned long rlim_rtprio =
5263 task_rlimit(p, RLIMIT_RTPRIO);
5264
5265 /* Can't set/change the rt policy: */
5266 if (policy != p->policy && !rlim_rtprio)
5267 return -EPERM;
5268
5269 /* Can't increase priority: */
5270 if (attr->sched_priority > p->rt_priority &&
5271 attr->sched_priority > rlim_rtprio)
5272 return -EPERM;
5273 }
5274
5275 /*
5276 * Can't set/change SCHED_DEADLINE policy at all for now
5277 * (safest behavior); in the future we would like to allow
5278 * unprivileged DL tasks to increase their relative deadline
5279 * or reduce their runtime (both ways reducing utilization)
5280 */
5281 if (dl_policy(policy))
5282 return -EPERM;
5283
5284 /*
5285 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5286 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5287 */
5288 if (task_has_idle_policy(p) && !idle_policy(policy)) {
5289 if (!can_nice(p, task_nice(p)))
5290 return -EPERM;
5291 }
5292
5293 /* Can't change other user's priorities: */
5294 if (!check_same_owner(p))
5295 return -EPERM;
5296
5297 /* Normal users shall not reset the sched_reset_on_fork flag: */
5298 if (p->sched_reset_on_fork && !reset_on_fork)
5299 return -EPERM;
5300 }
5301
5302 if (user) {
5303 if (attr->sched_flags & SCHED_FLAG_SUGOV)
5304 return -EINVAL;
5305
5306 retval = security_task_setscheduler(p);
5307 if (retval)
5308 return retval;
5309 }
5310
5311 /* Update task specific "requested" clamps */
5312 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
5313 retval = uclamp_validate(p, attr);
5314 if (retval)
5315 return retval;
5316 }
5317
5318 if (pi)
5319 cpuset_read_lock();
5320
5321 /*
5322 * Make sure no PI-waiters arrive (or leave) while we are
5323 * changing the priority of the task:
5324 *
5325 * To be able to change p->policy safely, the appropriate
5326 * runqueue lock must be held.
5327 */
5328 rq = task_rq_lock(p, &rf);
5329 update_rq_clock(rq);
5330
5331 /*
5332 * Changing the policy of the stop threads its a very bad idea:
5333 */
5334 if (p == rq->stop) {
5335 retval = -EINVAL;
5336 goto unlock;
5337 }
5338
5339 /*
5340 * If not changing anything there's no need to proceed further,
5341 * but store a possible modification of reset_on_fork.
5342 */
5343 if (unlikely(policy == p->policy)) {
5344 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
5345 goto change;
5346 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
5347 goto change;
5348 if (dl_policy(policy) && dl_param_changed(p, attr))
5349 goto change;
5350 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
5351 goto change;
5352
5353 p->sched_reset_on_fork = reset_on_fork;
5354 retval = 0;
5355 goto unlock;
5356 }
5357change:
5358
5359 if (user) {
5360#ifdef CONFIG_RT_GROUP_SCHED
5361 /*
5362 * Do not allow realtime tasks into groups that have no runtime
5363 * assigned.
5364 */
5365 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5366 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5367 !task_group_is_autogroup(task_group(p))) {
5368 retval = -EPERM;
5369 goto unlock;
5370 }
5371#endif
5372#ifdef CONFIG_SMP
5373 if (dl_bandwidth_enabled() && dl_policy(policy) &&
5374 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
5375 cpumask_t *span = rq->rd->span;
5376
5377 /*
5378 * Don't allow tasks with an affinity mask smaller than
5379 * the entire root_domain to become SCHED_DEADLINE. We
5380 * will also fail if there's no bandwidth available.
5381 */
5382 if (!cpumask_subset(span, p->cpus_ptr) ||
5383 rq->rd->dl_bw.bw == 0) {
5384 retval = -EPERM;
5385 goto unlock;
5386 }
5387 }
5388#endif
5389 }
5390
5391 /* Re-check policy now with rq lock held: */
5392 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5393 policy = oldpolicy = -1;
5394 task_rq_unlock(rq, p, &rf);
5395 if (pi)
5396 cpuset_read_unlock();
5397 goto recheck;
5398 }
5399
5400 /*
5401 * If setscheduling to SCHED_DEADLINE (or changing the parameters
5402 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
5403 * is available.
5404 */
5405 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
5406 retval = -EBUSY;
5407 goto unlock;
5408 }
5409
5410 p->sched_reset_on_fork = reset_on_fork;
5411 oldprio = p->prio;
5412
5413 if (pi) {
5414 /*
5415 * Take priority boosted tasks into account. If the new
5416 * effective priority is unchanged, we just store the new
5417 * normal parameters and do not touch the scheduler class and
5418 * the runqueue. This will be done when the task deboost
5419 * itself.
5420 */
5421 new_effective_prio = rt_effective_prio(p, newprio);
5422 if (new_effective_prio == oldprio)
5423 queue_flags &= ~DEQUEUE_MOVE;
5424 }
5425
5426 queued = task_on_rq_queued(p);
5427 running = task_current(rq, p);
5428 if (queued)
5429 dequeue_task(rq, p, queue_flags);
5430 if (running)
5431 put_prev_task(rq, p);
5432
5433 prev_class = p->sched_class;
5434
5435 __setscheduler(rq, p, attr, pi);
5436 __setscheduler_uclamp(p, attr);
5437
5438 if (queued) {
5439 /*
5440 * We enqueue to tail when the priority of a task is
5441 * increased (user space view).
5442 */
5443 if (oldprio < p->prio)
5444 queue_flags |= ENQUEUE_HEAD;
5445
5446 enqueue_task(rq, p, queue_flags);
5447 }
5448 if (running)
5449 set_next_task(rq, p);
5450
5451 check_class_changed(rq, p, prev_class, oldprio);
5452
5453 /* Avoid rq from going away on us: */
5454 preempt_disable();
5455 task_rq_unlock(rq, p, &rf);
5456
5457 if (pi) {
5458 cpuset_read_unlock();
5459 rt_mutex_adjust_pi(p);
5460 }
5461
5462 /* Run balance callbacks after we've adjusted the PI chain: */
5463 balance_callback(rq);
5464 preempt_enable();
5465
5466 return 0;
5467
5468unlock:
5469 task_rq_unlock(rq, p, &rf);
5470 if (pi)
5471 cpuset_read_unlock();
5472 return retval;
5473}
5474
5475static int _sched_setscheduler(struct task_struct *p, int policy,
5476 const struct sched_param *param, bool check)
5477{
5478 struct sched_attr attr = {
5479 .sched_policy = policy,
5480 .sched_priority = param->sched_priority,
5481 .sched_nice = PRIO_TO_NICE(p->static_prio),
5482 };
5483
5484 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5485 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
5486 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5487 policy &= ~SCHED_RESET_ON_FORK;
5488 attr.sched_policy = policy;
5489 }
5490
5491 return __sched_setscheduler(p, &attr, check, true);
5492}
5493/**
5494 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5495 * @p: the task in question.
5496 * @policy: new policy.
5497 * @param: structure containing the new RT priority.
5498 *
5499 * Use sched_set_fifo(), read its comment.
5500 *
5501 * Return: 0 on success. An error code otherwise.
5502 *
5503 * NOTE that the task may be already dead.
5504 */
5505int sched_setscheduler(struct task_struct *p, int policy,
5506 const struct sched_param *param)
5507{
5508 return _sched_setscheduler(p, policy, param, true);
5509}
5510
5511int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
5512{
5513 return __sched_setscheduler(p, attr, true, true);
5514}
5515
5516int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
5517{
5518 return __sched_setscheduler(p, attr, false, true);
5519}
5520
5521/**
5522 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5523 * @p: the task in question.
5524 * @policy: new policy.
5525 * @param: structure containing the new RT priority.
5526 *
5527 * Just like sched_setscheduler, only don't bother checking if the
5528 * current context has permission. For example, this is needed in
5529 * stop_machine(): we create temporary high priority worker threads,
5530 * but our caller might not have that capability.
5531 *
5532 * Return: 0 on success. An error code otherwise.
5533 */
5534int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5535 const struct sched_param *param)
5536{
5537 return _sched_setscheduler(p, policy, param, false);
5538}
5539
5540/*
5541 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
5542 * incapable of resource management, which is the one thing an OS really should
5543 * be doing.
5544 *
5545 * This is of course the reason it is limited to privileged users only.
5546 *
5547 * Worse still; it is fundamentally impossible to compose static priority
5548 * workloads. You cannot take two correctly working static prio workloads
5549 * and smash them together and still expect them to work.
5550 *
5551 * For this reason 'all' FIFO tasks the kernel creates are basically at:
5552 *
5553 * MAX_RT_PRIO / 2
5554 *
5555 * The administrator _MUST_ configure the system, the kernel simply doesn't
5556 * know enough information to make a sensible choice.
5557 */
5558void sched_set_fifo(struct task_struct *p)
5559{
5560 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
5561 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
5562}
5563EXPORT_SYMBOL_GPL(sched_set_fifo);
5564
5565/*
5566 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
5567 */
5568void sched_set_fifo_low(struct task_struct *p)
5569{
5570 struct sched_param sp = { .sched_priority = 1 };
5571 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
5572}
5573EXPORT_SYMBOL_GPL(sched_set_fifo_low);
5574
5575void sched_set_normal(struct task_struct *p, int nice)
5576{
5577 struct sched_attr attr = {
5578 .sched_policy = SCHED_NORMAL,
5579 .sched_nice = nice,
5580 };
5581 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
5582}
5583EXPORT_SYMBOL_GPL(sched_set_normal);
5584
5585static int
5586do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5587{
5588 struct sched_param lparam;
5589 struct task_struct *p;
5590 int retval;
5591
5592 if (!param || pid < 0)
5593 return -EINVAL;
5594 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5595 return -EFAULT;
5596
5597 rcu_read_lock();
5598 retval = -ESRCH;
5599 p = find_process_by_pid(pid);
5600 if (likely(p))
5601 get_task_struct(p);
5602 rcu_read_unlock();
5603
5604 if (likely(p)) {
5605 retval = sched_setscheduler(p, policy, &lparam);
5606 put_task_struct(p);
5607 }
5608
5609 return retval;
5610}
5611
5612/*
5613 * Mimics kernel/events/core.c perf_copy_attr().
5614 */
5615static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
5616{
5617 u32 size;
5618 int ret;
5619
5620 /* Zero the full structure, so that a short copy will be nice: */
5621 memset(attr, 0, sizeof(*attr));
5622
5623 ret = get_user(size, &uattr->size);
5624 if (ret)
5625 return ret;
5626
5627 /* ABI compatibility quirk: */
5628 if (!size)
5629 size = SCHED_ATTR_SIZE_VER0;
5630 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
5631 goto err_size;
5632
5633 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
5634 if (ret) {
5635 if (ret == -E2BIG)
5636 goto err_size;
5637 return ret;
5638 }
5639
5640 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
5641 size < SCHED_ATTR_SIZE_VER1)
5642 return -EINVAL;
5643
5644 /*
5645 * XXX: Do we want to be lenient like existing syscalls; or do we want
5646 * to be strict and return an error on out-of-bounds values?
5647 */
5648 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
5649
5650 return 0;
5651
5652err_size:
5653 put_user(sizeof(*attr), &uattr->size);
5654 return -E2BIG;
5655}
5656
5657/**
5658 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5659 * @pid: the pid in question.
5660 * @policy: new policy.
5661 * @param: structure containing the new RT priority.
5662 *
5663 * Return: 0 on success. An error code otherwise.
5664 */
5665SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
5666{
5667 if (policy < 0)
5668 return -EINVAL;
5669
5670 return do_sched_setscheduler(pid, policy, param);
5671}
5672
5673/**
5674 * sys_sched_setparam - set/change the RT priority of a thread
5675 * @pid: the pid in question.
5676 * @param: structure containing the new RT priority.
5677 *
5678 * Return: 0 on success. An error code otherwise.
5679 */
5680SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5681{
5682 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
5683}
5684
5685/**
5686 * sys_sched_setattr - same as above, but with extended sched_attr
5687 * @pid: the pid in question.
5688 * @uattr: structure containing the extended parameters.
5689 * @flags: for future extension.
5690 */
5691SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5692 unsigned int, flags)
5693{
5694 struct sched_attr attr;
5695 struct task_struct *p;
5696 int retval;
5697
5698 if (!uattr || pid < 0 || flags)
5699 return -EINVAL;
5700
5701 retval = sched_copy_attr(uattr, &attr);
5702 if (retval)
5703 return retval;
5704
5705 if ((int)attr.sched_policy < 0)
5706 return -EINVAL;
5707 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5708 attr.sched_policy = SETPARAM_POLICY;
5709
5710 rcu_read_lock();
5711 retval = -ESRCH;
5712 p = find_process_by_pid(pid);
5713 if (likely(p))
5714 get_task_struct(p);
5715 rcu_read_unlock();
5716
5717 if (likely(p)) {
5718 retval = sched_setattr(p, &attr);
5719 put_task_struct(p);
5720 }
5721
5722 return retval;
5723}
5724
5725/**
5726 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5727 * @pid: the pid in question.
5728 *
5729 * Return: On success, the policy of the thread. Otherwise, a negative error
5730 * code.
5731 */
5732SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5733{
5734 struct task_struct *p;
5735 int retval;
5736
5737 if (pid < 0)
5738 return -EINVAL;
5739
5740 retval = -ESRCH;
5741 rcu_read_lock();
5742 p = find_process_by_pid(pid);
5743 if (p) {
5744 retval = security_task_getscheduler(p);
5745 if (!retval)
5746 retval = p->policy
5747 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5748 }
5749 rcu_read_unlock();
5750 return retval;
5751}
5752
5753/**
5754 * sys_sched_getparam - get the RT priority of a thread
5755 * @pid: the pid in question.
5756 * @param: structure containing the RT priority.
5757 *
5758 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5759 * code.
5760 */
5761SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5762{
5763 struct sched_param lp = { .sched_priority = 0 };
5764 struct task_struct *p;
5765 int retval;
5766
5767 if (!param || pid < 0)
5768 return -EINVAL;
5769
5770 rcu_read_lock();
5771 p = find_process_by_pid(pid);
5772 retval = -ESRCH;
5773 if (!p)
5774 goto out_unlock;
5775
5776 retval = security_task_getscheduler(p);
5777 if (retval)
5778 goto out_unlock;
5779
5780 if (task_has_rt_policy(p))
5781 lp.sched_priority = p->rt_priority;
5782 rcu_read_unlock();
5783
5784 /*
5785 * This one might sleep, we cannot do it with a spinlock held ...
5786 */
5787 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5788
5789 return retval;
5790
5791out_unlock:
5792 rcu_read_unlock();
5793 return retval;
5794}
5795
5796/*
5797 * Copy the kernel size attribute structure (which might be larger
5798 * than what user-space knows about) to user-space.
5799 *
5800 * Note that all cases are valid: user-space buffer can be larger or
5801 * smaller than the kernel-space buffer. The usual case is that both
5802 * have the same size.
5803 */
5804static int
5805sched_attr_copy_to_user(struct sched_attr __user *uattr,
5806 struct sched_attr *kattr,
5807 unsigned int usize)
5808{
5809 unsigned int ksize = sizeof(*kattr);
5810
5811 if (!access_ok(uattr, usize))
5812 return -EFAULT;
5813
5814 /*
5815 * sched_getattr() ABI forwards and backwards compatibility:
5816 *
5817 * If usize == ksize then we just copy everything to user-space and all is good.
5818 *
5819 * If usize < ksize then we only copy as much as user-space has space for,
5820 * this keeps ABI compatibility as well. We skip the rest.
5821 *
5822 * If usize > ksize then user-space is using a newer version of the ABI,
5823 * which part the kernel doesn't know about. Just ignore it - tooling can
5824 * detect the kernel's knowledge of attributes from the attr->size value
5825 * which is set to ksize in this case.
5826 */
5827 kattr->size = min(usize, ksize);
5828
5829 if (copy_to_user(uattr, kattr, kattr->size))
5830 return -EFAULT;
5831
5832 return 0;
5833}
5834
5835/**
5836 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5837 * @pid: the pid in question.
5838 * @uattr: structure containing the extended parameters.
5839 * @usize: sizeof(attr) for fwd/bwd comp.
5840 * @flags: for future extension.
5841 */
5842SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5843 unsigned int, usize, unsigned int, flags)
5844{
5845 struct sched_attr kattr = { };
5846 struct task_struct *p;
5847 int retval;
5848
5849 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
5850 usize < SCHED_ATTR_SIZE_VER0 || flags)
5851 return -EINVAL;
5852
5853 rcu_read_lock();
5854 p = find_process_by_pid(pid);
5855 retval = -ESRCH;
5856 if (!p)
5857 goto out_unlock;
5858
5859 retval = security_task_getscheduler(p);
5860 if (retval)
5861 goto out_unlock;
5862
5863 kattr.sched_policy = p->policy;
5864 if (p->sched_reset_on_fork)
5865 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5866 if (task_has_dl_policy(p))
5867 __getparam_dl(p, &kattr);
5868 else if (task_has_rt_policy(p))
5869 kattr.sched_priority = p->rt_priority;
5870 else
5871 kattr.sched_nice = task_nice(p);
5872
5873#ifdef CONFIG_UCLAMP_TASK
5874 /*
5875 * This could race with another potential updater, but this is fine
5876 * because it'll correctly read the old or the new value. We don't need
5877 * to guarantee who wins the race as long as it doesn't return garbage.
5878 */
5879 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5880 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5881#endif
5882
5883 rcu_read_unlock();
5884
5885 return sched_attr_copy_to_user(uattr, &kattr, usize);
5886
5887out_unlock:
5888 rcu_read_unlock();
5889 return retval;
5890}
5891
5892long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5893{
5894 cpumask_var_t cpus_allowed, new_mask;
5895 struct task_struct *p;
5896 int retval;
5897
5898 rcu_read_lock();
5899
5900 p = find_process_by_pid(pid);
5901 if (!p) {
5902 rcu_read_unlock();
5903 return -ESRCH;
5904 }
5905
5906 /* Prevent p going away */
5907 get_task_struct(p);
5908 rcu_read_unlock();
5909
5910 if (p->flags & PF_NO_SETAFFINITY) {
5911 retval = -EINVAL;
5912 goto out_put_task;
5913 }
5914 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5915 retval = -ENOMEM;
5916 goto out_put_task;
5917 }
5918 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5919 retval = -ENOMEM;
5920 goto out_free_cpus_allowed;
5921 }
5922 retval = -EPERM;
5923 if (!check_same_owner(p)) {
5924 rcu_read_lock();
5925 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5926 rcu_read_unlock();
5927 goto out_free_new_mask;
5928 }
5929 rcu_read_unlock();
5930 }
5931
5932 retval = security_task_setscheduler(p);
5933 if (retval)
5934 goto out_free_new_mask;
5935
5936
5937 cpuset_cpus_allowed(p, cpus_allowed);
5938 cpumask_and(new_mask, in_mask, cpus_allowed);
5939
5940 /*
5941 * Since bandwidth control happens on root_domain basis,
5942 * if admission test is enabled, we only admit -deadline
5943 * tasks allowed to run on all the CPUs in the task's
5944 * root_domain.
5945 */
5946#ifdef CONFIG_SMP
5947 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
5948 rcu_read_lock();
5949 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
5950 retval = -EBUSY;
5951 rcu_read_unlock();
5952 goto out_free_new_mask;
5953 }
5954 rcu_read_unlock();
5955 }
5956#endif
5957again:
5958 retval = __set_cpus_allowed_ptr(p, new_mask, true);
5959
5960 if (!retval) {
5961 cpuset_cpus_allowed(p, cpus_allowed);
5962 if (!cpumask_subset(new_mask, cpus_allowed)) {
5963 /*
5964 * We must have raced with a concurrent cpuset
5965 * update. Just reset the cpus_allowed to the
5966 * cpuset's cpus_allowed
5967 */
5968 cpumask_copy(new_mask, cpus_allowed);
5969 goto again;
5970 }
5971 }
5972out_free_new_mask:
5973 free_cpumask_var(new_mask);
5974out_free_cpus_allowed:
5975 free_cpumask_var(cpus_allowed);
5976out_put_task:
5977 put_task_struct(p);
5978 return retval;
5979}
5980
5981static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5982 struct cpumask *new_mask)
5983{
5984 if (len < cpumask_size())
5985 cpumask_clear(new_mask);
5986 else if (len > cpumask_size())
5987 len = cpumask_size();
5988
5989 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5990}
5991
5992/**
5993 * sys_sched_setaffinity - set the CPU affinity of a process
5994 * @pid: pid of the process
5995 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5996 * @user_mask_ptr: user-space pointer to the new CPU mask
5997 *
5998 * Return: 0 on success. An error code otherwise.
5999 */
6000SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6001 unsigned long __user *, user_mask_ptr)
6002{
6003 cpumask_var_t new_mask;
6004 int retval;
6005
6006 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6007 return -ENOMEM;
6008
6009 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6010 if (retval == 0)
6011 retval = sched_setaffinity(pid, new_mask);
6012 free_cpumask_var(new_mask);
6013 return retval;
6014}
6015
6016long sched_getaffinity(pid_t pid, struct cpumask *mask)
6017{
6018 struct task_struct *p;
6019 unsigned long flags;
6020 int retval;
6021
6022 rcu_read_lock();
6023
6024 retval = -ESRCH;
6025 p = find_process_by_pid(pid);
6026 if (!p)
6027 goto out_unlock;
6028
6029 retval = security_task_getscheduler(p);
6030 if (retval)
6031 goto out_unlock;
6032
6033 raw_spin_lock_irqsave(&p->pi_lock, flags);
6034 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
6035 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6036
6037out_unlock:
6038 rcu_read_unlock();
6039
6040 return retval;
6041}
6042
6043/**
6044 * sys_sched_getaffinity - get the CPU affinity of a process
6045 * @pid: pid of the process
6046 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6047 * @user_mask_ptr: user-space pointer to hold the current CPU mask
6048 *
6049 * Return: size of CPU mask copied to user_mask_ptr on success. An
6050 * error code otherwise.
6051 */
6052SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6053 unsigned long __user *, user_mask_ptr)
6054{
6055 int ret;
6056 cpumask_var_t mask;
6057
6058 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
6059 return -EINVAL;
6060 if (len & (sizeof(unsigned long)-1))
6061 return -EINVAL;
6062
6063 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6064 return -ENOMEM;
6065
6066 ret = sched_getaffinity(pid, mask);
6067 if (ret == 0) {
6068 unsigned int retlen = min(len, cpumask_size());
6069
6070 if (copy_to_user(user_mask_ptr, mask, retlen))
6071 ret = -EFAULT;
6072 else
6073 ret = retlen;
6074 }
6075 free_cpumask_var(mask);
6076
6077 return ret;
6078}
6079
6080/**
6081 * sys_sched_yield - yield the current processor to other threads.
6082 *
6083 * This function yields the current CPU to other tasks. If there are no
6084 * other threads running on this CPU then this function will return.
6085 *
6086 * Return: 0.
6087 */
6088static void do_sched_yield(void)
6089{
6090 struct rq_flags rf;
6091 struct rq *rq;
6092
6093 rq = this_rq_lock_irq(&rf);
6094
6095 schedstat_inc(rq->yld_count);
6096 current->sched_class->yield_task(rq);
6097
6098 /*
6099 * Since we are going to call schedule() anyway, there's
6100 * no need to preempt or enable interrupts:
6101 */
6102 preempt_disable();
6103 rq_unlock(rq, &rf);
6104 sched_preempt_enable_no_resched();
6105
6106 schedule();
6107}
6108
6109SYSCALL_DEFINE0(sched_yield)
6110{
6111 do_sched_yield();
6112 return 0;
6113}
6114
6115#ifndef CONFIG_PREEMPTION
6116int __sched _cond_resched(void)
6117{
6118 if (should_resched(0)) {
6119 preempt_schedule_common();
6120 return 1;
6121 }
6122 rcu_all_qs();
6123 return 0;
6124}
6125EXPORT_SYMBOL(_cond_resched);
6126#endif
6127
6128/*
6129 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6130 * call schedule, and on return reacquire the lock.
6131 *
6132 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
6133 * operations here to prevent schedule() from being called twice (once via
6134 * spin_unlock(), once by hand).
6135 */
6136int __cond_resched_lock(spinlock_t *lock)
6137{
6138 int resched = should_resched(PREEMPT_LOCK_OFFSET);
6139 int ret = 0;
6140
6141 lockdep_assert_held(lock);
6142
6143 if (spin_needbreak(lock) || resched) {
6144 spin_unlock(lock);
6145 if (resched)
6146 preempt_schedule_common();
6147 else
6148 cpu_relax();
6149 ret = 1;
6150 spin_lock(lock);
6151 }
6152 return ret;
6153}
6154EXPORT_SYMBOL(__cond_resched_lock);
6155
6156/**
6157 * yield - yield the current processor to other threads.
6158 *
6159 * Do not ever use this function, there's a 99% chance you're doing it wrong.
6160 *
6161 * The scheduler is at all times free to pick the calling task as the most
6162 * eligible task to run, if removing the yield() call from your code breaks
6163 * it, its already broken.
6164 *
6165 * Typical broken usage is:
6166 *
6167 * while (!event)
6168 * yield();
6169 *
6170 * where one assumes that yield() will let 'the other' process run that will
6171 * make event true. If the current task is a SCHED_FIFO task that will never
6172 * happen. Never use yield() as a progress guarantee!!
6173 *
6174 * If you want to use yield() to wait for something, use wait_event().
6175 * If you want to use yield() to be 'nice' for others, use cond_resched().
6176 * If you still want to use yield(), do not!
6177 */
6178void __sched yield(void)
6179{
6180 set_current_state(TASK_RUNNING);
6181 do_sched_yield();
6182}
6183EXPORT_SYMBOL(yield);
6184
6185/**
6186 * yield_to - yield the current processor to another thread in
6187 * your thread group, or accelerate that thread toward the
6188 * processor it's on.
6189 * @p: target task
6190 * @preempt: whether task preemption is allowed or not
6191 *
6192 * It's the caller's job to ensure that the target task struct
6193 * can't go away on us before we can do any checks.
6194 *
6195 * Return:
6196 * true (>0) if we indeed boosted the target task.
6197 * false (0) if we failed to boost the target.
6198 * -ESRCH if there's no task to yield to.
6199 */
6200int __sched yield_to(struct task_struct *p, bool preempt)
6201{
6202 struct task_struct *curr = current;
6203 struct rq *rq, *p_rq;
6204 unsigned long flags;
6205 int yielded = 0;
6206
6207 local_irq_save(flags);
6208 rq = this_rq();
6209
6210again:
6211 p_rq = task_rq(p);
6212 /*
6213 * If we're the only runnable task on the rq and target rq also
6214 * has only one task, there's absolutely no point in yielding.
6215 */
6216 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
6217 yielded = -ESRCH;
6218 goto out_irq;
6219 }
6220
6221 double_rq_lock(rq, p_rq);
6222 if (task_rq(p) != p_rq) {
6223 double_rq_unlock(rq, p_rq);
6224 goto again;
6225 }
6226
6227 if (!curr->sched_class->yield_to_task)
6228 goto out_unlock;
6229
6230 if (curr->sched_class != p->sched_class)
6231 goto out_unlock;
6232
6233 if (task_running(p_rq, p) || p->state)
6234 goto out_unlock;
6235
6236 yielded = curr->sched_class->yield_to_task(rq, p);
6237 if (yielded) {
6238 schedstat_inc(rq->yld_count);
6239 /*
6240 * Make p's CPU reschedule; pick_next_entity takes care of
6241 * fairness.
6242 */
6243 if (preempt && rq != p_rq)
6244 resched_curr(p_rq);
6245 }
6246
6247out_unlock:
6248 double_rq_unlock(rq, p_rq);
6249out_irq:
6250 local_irq_restore(flags);
6251
6252 if (yielded > 0)
6253 schedule();
6254
6255 return yielded;
6256}
6257EXPORT_SYMBOL_GPL(yield_to);
6258
6259int io_schedule_prepare(void)
6260{
6261 int old_iowait = current->in_iowait;
6262
6263 current->in_iowait = 1;
6264 blk_schedule_flush_plug(current);
6265
6266 return old_iowait;
6267}
6268
6269void io_schedule_finish(int token)
6270{
6271 current->in_iowait = token;
6272}
6273
6274/*
6275 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6276 * that process accounting knows that this is a task in IO wait state.
6277 */
6278long __sched io_schedule_timeout(long timeout)
6279{
6280 int token;
6281 long ret;
6282
6283 token = io_schedule_prepare();
6284 ret = schedule_timeout(timeout);
6285 io_schedule_finish(token);
6286
6287 return ret;
6288}
6289EXPORT_SYMBOL(io_schedule_timeout);
6290
6291void __sched io_schedule(void)
6292{
6293 int token;
6294
6295 token = io_schedule_prepare();
6296 schedule();
6297 io_schedule_finish(token);
6298}
6299EXPORT_SYMBOL(io_schedule);
6300
6301/**
6302 * sys_sched_get_priority_max - return maximum RT priority.
6303 * @policy: scheduling class.
6304 *
6305 * Return: On success, this syscall returns the maximum
6306 * rt_priority that can be used by a given scheduling class.
6307 * On failure, a negative error code is returned.
6308 */
6309SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6310{
6311 int ret = -EINVAL;
6312
6313 switch (policy) {
6314 case SCHED_FIFO:
6315 case SCHED_RR:
6316 ret = MAX_USER_RT_PRIO-1;
6317 break;
6318 case SCHED_DEADLINE:
6319 case SCHED_NORMAL:
6320 case SCHED_BATCH:
6321 case SCHED_IDLE:
6322 ret = 0;
6323 break;
6324 }
6325 return ret;
6326}
6327
6328/**
6329 * sys_sched_get_priority_min - return minimum RT priority.
6330 * @policy: scheduling class.
6331 *
6332 * Return: On success, this syscall returns the minimum
6333 * rt_priority that can be used by a given scheduling class.
6334 * On failure, a negative error code is returned.
6335 */
6336SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6337{
6338 int ret = -EINVAL;
6339
6340 switch (policy) {
6341 case SCHED_FIFO:
6342 case SCHED_RR:
6343 ret = 1;
6344 break;
6345 case SCHED_DEADLINE:
6346 case SCHED_NORMAL:
6347 case SCHED_BATCH:
6348 case SCHED_IDLE:
6349 ret = 0;
6350 }
6351 return ret;
6352}
6353
6354static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
6355{
6356 struct task_struct *p;
6357 unsigned int time_slice;
6358 struct rq_flags rf;
6359 struct rq *rq;
6360 int retval;
6361
6362 if (pid < 0)
6363 return -EINVAL;
6364
6365 retval = -ESRCH;
6366 rcu_read_lock();
6367 p = find_process_by_pid(pid);
6368 if (!p)
6369 goto out_unlock;
6370
6371 retval = security_task_getscheduler(p);
6372 if (retval)
6373 goto out_unlock;
6374
6375 rq = task_rq_lock(p, &rf);
6376 time_slice = 0;
6377 if (p->sched_class->get_rr_interval)
6378 time_slice = p->sched_class->get_rr_interval(rq, p);
6379 task_rq_unlock(rq, p, &rf);
6380
6381 rcu_read_unlock();
6382 jiffies_to_timespec64(time_slice, t);
6383 return 0;
6384
6385out_unlock:
6386 rcu_read_unlock();
6387 return retval;
6388}
6389
6390/**
6391 * sys_sched_rr_get_interval - return the default timeslice of a process.
6392 * @pid: pid of the process.
6393 * @interval: userspace pointer to the timeslice value.
6394 *
6395 * this syscall writes the default timeslice value of a given process
6396 * into the user-space timespec buffer. A value of '0' means infinity.
6397 *
6398 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
6399 * an error code.
6400 */
6401SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6402 struct __kernel_timespec __user *, interval)
6403{
6404 struct timespec64 t;
6405 int retval = sched_rr_get_interval(pid, &t);
6406
6407 if (retval == 0)
6408 retval = put_timespec64(&t, interval);
6409
6410 return retval;
6411}
6412
6413#ifdef CONFIG_COMPAT_32BIT_TIME
6414SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
6415 struct old_timespec32 __user *, interval)
6416{
6417 struct timespec64 t;
6418 int retval = sched_rr_get_interval(pid, &t);
6419
6420 if (retval == 0)
6421 retval = put_old_timespec32(&t, interval);
6422 return retval;
6423}
6424#endif
6425
6426void sched_show_task(struct task_struct *p)
6427{
6428 unsigned long free = 0;
6429 int ppid;
6430
6431 if (!try_get_task_stack(p))
6432 return;
6433
6434 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
6435
6436 if (p->state == TASK_RUNNING)
6437 pr_cont(" running task ");
6438#ifdef CONFIG_DEBUG_STACK_USAGE
6439 free = stack_not_used(p);
6440#endif
6441 ppid = 0;
6442 rcu_read_lock();
6443 if (pid_alive(p))
6444 ppid = task_pid_nr(rcu_dereference(p->real_parent));
6445 rcu_read_unlock();
6446 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
6447 free, task_pid_nr(p), ppid,
6448 (unsigned long)task_thread_info(p)->flags);
6449
6450 print_worker_info(KERN_INFO, p);
6451 show_stack(p, NULL, KERN_INFO);
6452 put_task_stack(p);
6453}
6454EXPORT_SYMBOL_GPL(sched_show_task);
6455
6456static inline bool
6457state_filter_match(unsigned long state_filter, struct task_struct *p)
6458{
6459 /* no filter, everything matches */
6460 if (!state_filter)
6461 return true;
6462
6463 /* filter, but doesn't match */
6464 if (!(p->state & state_filter))
6465 return false;
6466
6467 /*
6468 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
6469 * TASK_KILLABLE).
6470 */
6471 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
6472 return false;
6473
6474 return true;
6475}
6476
6477
6478void show_state_filter(unsigned long state_filter)
6479{
6480 struct task_struct *g, *p;
6481
6482 rcu_read_lock();
6483 for_each_process_thread(g, p) {
6484 /*
6485 * reset the NMI-timeout, listing all files on a slow
6486 * console might take a lot of time:
6487 * Also, reset softlockup watchdogs on all CPUs, because
6488 * another CPU might be blocked waiting for us to process
6489 * an IPI.
6490 */
6491 touch_nmi_watchdog();
6492 touch_all_softlockup_watchdogs();
6493 if (state_filter_match(state_filter, p))
6494 sched_show_task(p);
6495 }
6496
6497#ifdef CONFIG_SCHED_DEBUG
6498 if (!state_filter)
6499 sysrq_sched_debug_show();
6500#endif
6501 rcu_read_unlock();
6502 /*
6503 * Only show locks if all tasks are dumped:
6504 */
6505 if (!state_filter)
6506 debug_show_all_locks();
6507}
6508
6509/**
6510 * init_idle - set up an idle thread for a given CPU
6511 * @idle: task in question
6512 * @cpu: CPU the idle task belongs to
6513 *
6514 * NOTE: this function does not set the idle thread's NEED_RESCHED
6515 * flag, to make booting more robust.
6516 */
6517void init_idle(struct task_struct *idle, int cpu)
6518{
6519 struct rq *rq = cpu_rq(cpu);
6520 unsigned long flags;
6521
6522 __sched_fork(0, idle);
6523
6524 raw_spin_lock_irqsave(&idle->pi_lock, flags);
6525 raw_spin_lock(&rq->lock);
6526
6527 idle->state = TASK_RUNNING;
6528 idle->se.exec_start = sched_clock();
6529 idle->flags |= PF_IDLE;
6530
6531 scs_task_reset(idle);
6532 kasan_unpoison_task_stack(idle);
6533
6534#ifdef CONFIG_SMP
6535 /*
6536 * Its possible that init_idle() gets called multiple times on a task,
6537 * in that case do_set_cpus_allowed() will not do the right thing.
6538 *
6539 * And since this is boot we can forgo the serialization.
6540 */
6541 set_cpus_allowed_common(idle, cpumask_of(cpu));
6542#endif
6543 /*
6544 * We're having a chicken and egg problem, even though we are
6545 * holding rq->lock, the CPU isn't yet set to this CPU so the
6546 * lockdep check in task_group() will fail.
6547 *
6548 * Similar case to sched_fork(). / Alternatively we could
6549 * use task_rq_lock() here and obtain the other rq->lock.
6550 *
6551 * Silence PROVE_RCU
6552 */
6553 rcu_read_lock();
6554 __set_task_cpu(idle, cpu);
6555 rcu_read_unlock();
6556
6557 rq->idle = idle;
6558 rcu_assign_pointer(rq->curr, idle);
6559 idle->on_rq = TASK_ON_RQ_QUEUED;
6560#ifdef CONFIG_SMP
6561 idle->on_cpu = 1;
6562#endif
6563 raw_spin_unlock(&rq->lock);
6564 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
6565
6566 /* Set the preempt count _outside_ the spinlocks! */
6567 init_idle_preempt_count(idle, cpu);
6568
6569 /*
6570 * The idle tasks have their own, simple scheduling class:
6571 */
6572 idle->sched_class = &idle_sched_class;
6573 ftrace_graph_init_idle_task(idle, cpu);
6574 vtime_init_idle(idle, cpu);
6575#ifdef CONFIG_SMP
6576 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6577#endif
6578}
6579
6580#ifdef CONFIG_SMP
6581
6582int cpuset_cpumask_can_shrink(const struct cpumask *cur,
6583 const struct cpumask *trial)
6584{
6585 int ret = 1;
6586
6587 if (!cpumask_weight(cur))
6588 return ret;
6589
6590 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
6591
6592 return ret;
6593}
6594
6595int task_can_attach(struct task_struct *p,
6596 const struct cpumask *cs_cpus_allowed)
6597{
6598 int ret = 0;
6599
6600 /*
6601 * Kthreads which disallow setaffinity shouldn't be moved
6602 * to a new cpuset; we don't want to change their CPU
6603 * affinity and isolating such threads by their set of
6604 * allowed nodes is unnecessary. Thus, cpusets are not
6605 * applicable for such threads. This prevents checking for
6606 * success of set_cpus_allowed_ptr() on all attached tasks
6607 * before cpus_mask may be changed.
6608 */
6609 if (p->flags & PF_NO_SETAFFINITY) {
6610 ret = -EINVAL;
6611 goto out;
6612 }
6613
6614 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
6615 cs_cpus_allowed))
6616 ret = dl_task_can_attach(p, cs_cpus_allowed);
6617
6618out:
6619 return ret;
6620}
6621
6622bool sched_smp_initialized __read_mostly;
6623
6624#ifdef CONFIG_NUMA_BALANCING
6625/* Migrate current task p to target_cpu */
6626int migrate_task_to(struct task_struct *p, int target_cpu)
6627{
6628 struct migration_arg arg = { p, target_cpu };
6629 int curr_cpu = task_cpu(p);
6630
6631 if (curr_cpu == target_cpu)
6632 return 0;
6633
6634 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
6635 return -EINVAL;
6636
6637 /* TODO: This is not properly updating schedstats */
6638
6639 trace_sched_move_numa(p, curr_cpu, target_cpu);
6640 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
6641}
6642
6643/*
6644 * Requeue a task on a given node and accurately track the number of NUMA
6645 * tasks on the runqueues
6646 */
6647void sched_setnuma(struct task_struct *p, int nid)
6648{
6649 bool queued, running;
6650 struct rq_flags rf;
6651 struct rq *rq;
6652
6653 rq = task_rq_lock(p, &rf);
6654 queued = task_on_rq_queued(p);
6655 running = task_current(rq, p);
6656
6657 if (queued)
6658 dequeue_task(rq, p, DEQUEUE_SAVE);
6659 if (running)
6660 put_prev_task(rq, p);
6661
6662 p->numa_preferred_nid = nid;
6663
6664 if (queued)
6665 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6666 if (running)
6667 set_next_task(rq, p);
6668 task_rq_unlock(rq, p, &rf);
6669}
6670#endif /* CONFIG_NUMA_BALANCING */
6671
6672#ifdef CONFIG_HOTPLUG_CPU
6673/*
6674 * Ensure that the idle task is using init_mm right before its CPU goes
6675 * offline.
6676 */
6677void idle_task_exit(void)
6678{
6679 struct mm_struct *mm = current->active_mm;
6680
6681 BUG_ON(cpu_online(smp_processor_id()));
6682 BUG_ON(current != this_rq()->idle);
6683
6684 if (mm != &init_mm) {
6685 switch_mm(mm, &init_mm, current);
6686 finish_arch_post_lock_switch();
6687 }
6688
6689 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
6690}
6691
6692/*
6693 * Since this CPU is going 'away' for a while, fold any nr_active delta
6694 * we might have. Assumes we're called after migrate_tasks() so that the
6695 * nr_active count is stable. We need to take the teardown thread which
6696 * is calling this into account, so we hand in adjust = 1 to the load
6697 * calculation.
6698 *
6699 * Also see the comment "Global load-average calculations".
6700 */
6701static void calc_load_migrate(struct rq *rq)
6702{
6703 long delta = calc_load_fold_active(rq, 1);
6704 if (delta)
6705 atomic_long_add(delta, &calc_load_tasks);
6706}
6707
6708static struct task_struct *__pick_migrate_task(struct rq *rq)
6709{
6710 const struct sched_class *class;
6711 struct task_struct *next;
6712
6713 for_each_class(class) {
6714 next = class->pick_next_task(rq);
6715 if (next) {
6716 next->sched_class->put_prev_task(rq, next);
6717 return next;
6718 }
6719 }
6720
6721 /* The idle class should always have a runnable task */
6722 BUG();
6723}
6724
6725/*
6726 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6727 * try_to_wake_up()->select_task_rq().
6728 *
6729 * Called with rq->lock held even though we'er in stop_machine() and
6730 * there's no concurrency possible, we hold the required locks anyway
6731 * because of lock validation efforts.
6732 */
6733static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
6734{
6735 struct rq *rq = dead_rq;
6736 struct task_struct *next, *stop = rq->stop;
6737 struct rq_flags orf = *rf;
6738 int dest_cpu;
6739
6740 /*
6741 * Fudge the rq selection such that the below task selection loop
6742 * doesn't get stuck on the currently eligible stop task.
6743 *
6744 * We're currently inside stop_machine() and the rq is either stuck
6745 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6746 * either way we should never end up calling schedule() until we're
6747 * done here.
6748 */
6749 rq->stop = NULL;
6750
6751 /*
6752 * put_prev_task() and pick_next_task() sched
6753 * class method both need to have an up-to-date
6754 * value of rq->clock[_task]
6755 */
6756 update_rq_clock(rq);
6757
6758 for (;;) {
6759 /*
6760 * There's this thread running, bail when that's the only
6761 * remaining thread:
6762 */
6763 if (rq->nr_running == 1)
6764 break;
6765
6766 next = __pick_migrate_task(rq);
6767
6768 /*
6769 * Rules for changing task_struct::cpus_mask are holding
6770 * both pi_lock and rq->lock, such that holding either
6771 * stabilizes the mask.
6772 *
6773 * Drop rq->lock is not quite as disastrous as it usually is
6774 * because !cpu_active at this point, which means load-balance
6775 * will not interfere. Also, stop-machine.
6776 */
6777 rq_unlock(rq, rf);
6778 raw_spin_lock(&next->pi_lock);
6779 rq_relock(rq, rf);
6780
6781 /*
6782 * Since we're inside stop-machine, _nothing_ should have
6783 * changed the task, WARN if weird stuff happened, because in
6784 * that case the above rq->lock drop is a fail too.
6785 */
6786 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
6787 raw_spin_unlock(&next->pi_lock);
6788 continue;
6789 }
6790
6791 /* Find suitable destination for @next, with force if needed. */
6792 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
6793 rq = __migrate_task(rq, rf, next, dest_cpu);
6794 if (rq != dead_rq) {
6795 rq_unlock(rq, rf);
6796 rq = dead_rq;
6797 *rf = orf;
6798 rq_relock(rq, rf);
6799 }
6800 raw_spin_unlock(&next->pi_lock);
6801 }
6802
6803 rq->stop = stop;
6804}
6805#endif /* CONFIG_HOTPLUG_CPU */
6806
6807void set_rq_online(struct rq *rq)
6808{
6809 if (!rq->online) {
6810 const struct sched_class *class;
6811
6812 cpumask_set_cpu(rq->cpu, rq->rd->online);
6813 rq->online = 1;
6814
6815 for_each_class(class) {
6816 if (class->rq_online)
6817 class->rq_online(rq);
6818 }
6819 }
6820}
6821
6822void set_rq_offline(struct rq *rq)
6823{
6824 if (rq->online) {
6825 const struct sched_class *class;
6826
6827 for_each_class(class) {
6828 if (class->rq_offline)
6829 class->rq_offline(rq);
6830 }
6831
6832 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6833 rq->online = 0;
6834 }
6835}
6836
6837/*
6838 * used to mark begin/end of suspend/resume:
6839 */
6840static int num_cpus_frozen;
6841
6842/*
6843 * Update cpusets according to cpu_active mask. If cpusets are
6844 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6845 * around partition_sched_domains().
6846 *
6847 * If we come here as part of a suspend/resume, don't touch cpusets because we
6848 * want to restore it back to its original state upon resume anyway.
6849 */
6850static void cpuset_cpu_active(void)
6851{
6852 if (cpuhp_tasks_frozen) {
6853 /*
6854 * num_cpus_frozen tracks how many CPUs are involved in suspend
6855 * resume sequence. As long as this is not the last online
6856 * operation in the resume sequence, just build a single sched
6857 * domain, ignoring cpusets.
6858 */
6859 partition_sched_domains(1, NULL, NULL);
6860 if (--num_cpus_frozen)
6861 return;
6862 /*
6863 * This is the last CPU online operation. So fall through and
6864 * restore the original sched domains by considering the
6865 * cpuset configurations.
6866 */
6867 cpuset_force_rebuild();
6868 }
6869 cpuset_update_active_cpus();
6870}
6871
6872static int cpuset_cpu_inactive(unsigned int cpu)
6873{
6874 if (!cpuhp_tasks_frozen) {
6875 if (dl_cpu_busy(cpu))
6876 return -EBUSY;
6877 cpuset_update_active_cpus();
6878 } else {
6879 num_cpus_frozen++;
6880 partition_sched_domains(1, NULL, NULL);
6881 }
6882 return 0;
6883}
6884
6885int sched_cpu_activate(unsigned int cpu)
6886{
6887 struct rq *rq = cpu_rq(cpu);
6888 struct rq_flags rf;
6889
6890#ifdef CONFIG_SCHED_SMT
6891 /*
6892 * When going up, increment the number of cores with SMT present.
6893 */
6894 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6895 static_branch_inc_cpuslocked(&sched_smt_present);
6896#endif
6897 set_cpu_active(cpu, true);
6898
6899 if (sched_smp_initialized) {
6900 sched_domains_numa_masks_set(cpu);
6901 cpuset_cpu_active();
6902 }
6903
6904 /*
6905 * Put the rq online, if not already. This happens:
6906 *
6907 * 1) In the early boot process, because we build the real domains
6908 * after all CPUs have been brought up.
6909 *
6910 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6911 * domains.
6912 */
6913 rq_lock_irqsave(rq, &rf);
6914 if (rq->rd) {
6915 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6916 set_rq_online(rq);
6917 }
6918 rq_unlock_irqrestore(rq, &rf);
6919
6920 return 0;
6921}
6922
6923int sched_cpu_deactivate(unsigned int cpu)
6924{
6925 int ret;
6926
6927 set_cpu_active(cpu, false);
6928 /*
6929 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6930 * users of this state to go away such that all new such users will
6931 * observe it.
6932 *
6933 * Do sync before park smpboot threads to take care the rcu boost case.
6934 */
6935 synchronize_rcu();
6936
6937#ifdef CONFIG_SCHED_SMT
6938 /*
6939 * When going down, decrement the number of cores with SMT present.
6940 */
6941 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6942 static_branch_dec_cpuslocked(&sched_smt_present);
6943#endif
6944
6945 if (!sched_smp_initialized)
6946 return 0;
6947
6948 ret = cpuset_cpu_inactive(cpu);
6949 if (ret) {
6950 set_cpu_active(cpu, true);
6951 return ret;
6952 }
6953 sched_domains_numa_masks_clear(cpu);
6954 return 0;
6955}
6956
6957static void sched_rq_cpu_starting(unsigned int cpu)
6958{
6959 struct rq *rq = cpu_rq(cpu);
6960
6961 rq->calc_load_update = calc_load_update;
6962 update_max_interval();
6963}
6964
6965int sched_cpu_starting(unsigned int cpu)
6966{
6967 sched_rq_cpu_starting(cpu);
6968 sched_tick_start(cpu);
6969 return 0;
6970}
6971
6972#ifdef CONFIG_HOTPLUG_CPU
6973int sched_cpu_dying(unsigned int cpu)
6974{
6975 struct rq *rq = cpu_rq(cpu);
6976 struct rq_flags rf;
6977
6978 /* Handle pending wakeups and then migrate everything off */
6979 sched_tick_stop(cpu);
6980
6981 rq_lock_irqsave(rq, &rf);
6982 if (rq->rd) {
6983 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6984 set_rq_offline(rq);
6985 }
6986 migrate_tasks(rq, &rf);
6987 BUG_ON(rq->nr_running != 1);
6988 rq_unlock_irqrestore(rq, &rf);
6989
6990 calc_load_migrate(rq);
6991 update_max_interval();
6992 nohz_balance_exit_idle(rq);
6993 hrtick_clear(rq);
6994 return 0;
6995}
6996#endif
6997
6998void __init sched_init_smp(void)
6999{
7000 sched_init_numa();
7001
7002 /*
7003 * There's no userspace yet to cause hotplug operations; hence all the
7004 * CPU masks are stable and all blatant races in the below code cannot
7005 * happen.
7006 */
7007 mutex_lock(&sched_domains_mutex);
7008 sched_init_domains(cpu_active_mask);
7009 mutex_unlock(&sched_domains_mutex);
7010
7011 /* Move init over to a non-isolated CPU */
7012 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
7013 BUG();
7014 sched_init_granularity();
7015
7016 init_sched_rt_class();
7017 init_sched_dl_class();
7018
7019 sched_smp_initialized = true;
7020}
7021
7022static int __init migration_init(void)
7023{
7024 sched_cpu_starting(smp_processor_id());
7025 return 0;
7026}
7027early_initcall(migration_init);
7028
7029#else
7030void __init sched_init_smp(void)
7031{
7032 sched_init_granularity();
7033}
7034#endif /* CONFIG_SMP */
7035
7036int in_sched_functions(unsigned long addr)
7037{
7038 return in_lock_functions(addr) ||
7039 (addr >= (unsigned long)__sched_text_start
7040 && addr < (unsigned long)__sched_text_end);
7041}
7042
7043#ifdef CONFIG_CGROUP_SCHED
7044/*
7045 * Default task group.
7046 * Every task in system belongs to this group at bootup.
7047 */
7048struct task_group root_task_group;
7049LIST_HEAD(task_groups);
7050
7051/* Cacheline aligned slab cache for task_group */
7052static struct kmem_cache *task_group_cache __read_mostly;
7053#endif
7054
7055DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7056DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
7057
7058void __init sched_init(void)
7059{
7060 unsigned long ptr = 0;
7061 int i;
7062
7063 /* Make sure the linker didn't screw up */
7064 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
7065 &fair_sched_class + 1 != &rt_sched_class ||
7066 &rt_sched_class + 1 != &dl_sched_class);
7067#ifdef CONFIG_SMP
7068 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
7069#endif
7070
7071 wait_bit_init();
7072
7073#ifdef CONFIG_FAIR_GROUP_SCHED
7074 ptr += 2 * nr_cpu_ids * sizeof(void **);
7075#endif
7076#ifdef CONFIG_RT_GROUP_SCHED
7077 ptr += 2 * nr_cpu_ids * sizeof(void **);
7078#endif
7079 if (ptr) {
7080 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
7081
7082#ifdef CONFIG_FAIR_GROUP_SCHED
7083 root_task_group.se = (struct sched_entity **)ptr;
7084 ptr += nr_cpu_ids * sizeof(void **);
7085
7086 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7087 ptr += nr_cpu_ids * sizeof(void **);
7088
7089 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7090 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7091#endif /* CONFIG_FAIR_GROUP_SCHED */
7092#ifdef CONFIG_RT_GROUP_SCHED
7093 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7094 ptr += nr_cpu_ids * sizeof(void **);
7095
7096 root_task_group.rt_rq = (struct rt_rq **)ptr;
7097 ptr += nr_cpu_ids * sizeof(void **);
7098
7099#endif /* CONFIG_RT_GROUP_SCHED */
7100 }
7101#ifdef CONFIG_CPUMASK_OFFSTACK
7102 for_each_possible_cpu(i) {
7103 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7104 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7105 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
7106 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7107 }
7108#endif /* CONFIG_CPUMASK_OFFSTACK */
7109
7110 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
7111 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
7112
7113#ifdef CONFIG_SMP
7114 init_defrootdomain();
7115#endif
7116
7117#ifdef CONFIG_RT_GROUP_SCHED
7118 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7119 global_rt_period(), global_rt_runtime());
7120#endif /* CONFIG_RT_GROUP_SCHED */
7121
7122#ifdef CONFIG_CGROUP_SCHED
7123 task_group_cache = KMEM_CACHE(task_group, 0);
7124
7125 list_add(&root_task_group.list, &task_groups);
7126 INIT_LIST_HEAD(&root_task_group.children);
7127 INIT_LIST_HEAD(&root_task_group.siblings);
7128 autogroup_init(&init_task);
7129#endif /* CONFIG_CGROUP_SCHED */
7130
7131 for_each_possible_cpu(i) {
7132 struct rq *rq;
7133
7134 rq = cpu_rq(i);
7135 raw_spin_lock_init(&rq->lock);
7136 rq->nr_running = 0;
7137 rq->calc_load_active = 0;
7138 rq->calc_load_update = jiffies + LOAD_FREQ;
7139 init_cfs_rq(&rq->cfs);
7140 init_rt_rq(&rq->rt);
7141 init_dl_rq(&rq->dl);
7142#ifdef CONFIG_FAIR_GROUP_SCHED
7143 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7144 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7145 /*
7146 * How much CPU bandwidth does root_task_group get?
7147 *
7148 * In case of task-groups formed thr' the cgroup filesystem, it
7149 * gets 100% of the CPU resources in the system. This overall
7150 * system CPU resource is divided among the tasks of
7151 * root_task_group and its child task-groups in a fair manner,
7152 * based on each entity's (task or task-group's) weight
7153 * (se->load.weight).
7154 *
7155 * In other words, if root_task_group has 10 tasks of weight
7156 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7157 * then A0's share of the CPU resource is:
7158 *
7159 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7160 *
7161 * We achieve this by letting root_task_group's tasks sit
7162 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7163 */
7164 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7165#endif /* CONFIG_FAIR_GROUP_SCHED */
7166
7167 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7168#ifdef CONFIG_RT_GROUP_SCHED
7169 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7170#endif
7171#ifdef CONFIG_SMP
7172 rq->sd = NULL;
7173 rq->rd = NULL;
7174 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7175 rq->balance_callback = NULL;
7176 rq->active_balance = 0;
7177 rq->next_balance = jiffies;
7178 rq->push_cpu = 0;
7179 rq->cpu = i;
7180 rq->online = 0;
7181 rq->idle_stamp = 0;
7182 rq->avg_idle = 2*sysctl_sched_migration_cost;
7183 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7184
7185 INIT_LIST_HEAD(&rq->cfs_tasks);
7186
7187 rq_attach_root(rq, &def_root_domain);
7188#ifdef CONFIG_NO_HZ_COMMON
7189 rq->last_blocked_load_update_tick = jiffies;
7190 atomic_set(&rq->nohz_flags, 0);
7191
7192 rq_csd_init(rq, &rq->nohz_csd, nohz_csd_func);
7193#endif
7194#endif /* CONFIG_SMP */
7195 hrtick_rq_init(rq);
7196 atomic_set(&rq->nr_iowait, 0);
7197 }
7198
7199 set_load_weight(&init_task, false);
7200
7201 /*
7202 * The boot idle thread does lazy MMU switching as well:
7203 */
7204 mmgrab(&init_mm);
7205 enter_lazy_tlb(&init_mm, current);
7206
7207 /*
7208 * Make us the idle thread. Technically, schedule() should not be
7209 * called from this thread, however somewhere below it might be,
7210 * but because we are the idle thread, we just pick up running again
7211 * when this runqueue becomes "idle".
7212 */
7213 init_idle(current, smp_processor_id());
7214
7215 calc_load_update = jiffies + LOAD_FREQ;
7216
7217#ifdef CONFIG_SMP
7218 idle_thread_set_boot_cpu();
7219#endif
7220 init_sched_fair_class();
7221
7222 init_schedstats();
7223
7224 psi_init();
7225
7226 init_uclamp();
7227
7228 scheduler_running = 1;
7229}
7230
7231#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7232static inline int preempt_count_equals(int preempt_offset)
7233{
7234 int nested = preempt_count() + rcu_preempt_depth();
7235
7236 return (nested == preempt_offset);
7237}
7238
7239void __might_sleep(const char *file, int line, int preempt_offset)
7240{
7241 /*
7242 * Blocking primitives will set (and therefore destroy) current->state,
7243 * since we will exit with TASK_RUNNING make sure we enter with it,
7244 * otherwise we will destroy state.
7245 */
7246 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7247 "do not call blocking ops when !TASK_RUNNING; "
7248 "state=%lx set at [<%p>] %pS\n",
7249 current->state,
7250 (void *)current->task_state_change,
7251 (void *)current->task_state_change);
7252
7253 ___might_sleep(file, line, preempt_offset);
7254}
7255EXPORT_SYMBOL(__might_sleep);
7256
7257void ___might_sleep(const char *file, int line, int preempt_offset)
7258{
7259 /* Ratelimiting timestamp: */
7260 static unsigned long prev_jiffy;
7261
7262 unsigned long preempt_disable_ip;
7263
7264 /* WARN_ON_ONCE() by default, no rate limit required: */
7265 rcu_sleep_check();
7266
7267 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7268 !is_idle_task(current) && !current->non_block_count) ||
7269 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
7270 oops_in_progress)
7271 return;
7272
7273 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7274 return;
7275 prev_jiffy = jiffies;
7276
7277 /* Save this before calling printk(), since that will clobber it: */
7278 preempt_disable_ip = get_preempt_disable_ip(current);
7279
7280 printk(KERN_ERR
7281 "BUG: sleeping function called from invalid context at %s:%d\n",
7282 file, line);
7283 printk(KERN_ERR
7284 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
7285 in_atomic(), irqs_disabled(), current->non_block_count,
7286 current->pid, current->comm);
7287
7288 if (task_stack_end_corrupted(current))
7289 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7290
7291 debug_show_held_locks(current);
7292 if (irqs_disabled())
7293 print_irqtrace_events(current);
7294 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
7295 && !preempt_count_equals(preempt_offset)) {
7296 pr_err("Preemption disabled at:");
7297 print_ip_sym(KERN_ERR, preempt_disable_ip);
7298 }
7299 dump_stack();
7300 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7301}
7302EXPORT_SYMBOL(___might_sleep);
7303
7304void __cant_sleep(const char *file, int line, int preempt_offset)
7305{
7306 static unsigned long prev_jiffy;
7307
7308 if (irqs_disabled())
7309 return;
7310
7311 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
7312 return;
7313
7314 if (preempt_count() > preempt_offset)
7315 return;
7316
7317 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7318 return;
7319 prev_jiffy = jiffies;
7320
7321 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
7322 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7323 in_atomic(), irqs_disabled(),
7324 current->pid, current->comm);
7325
7326 debug_show_held_locks(current);
7327 dump_stack();
7328 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7329}
7330EXPORT_SYMBOL_GPL(__cant_sleep);
7331#endif
7332
7333#ifdef CONFIG_MAGIC_SYSRQ
7334void normalize_rt_tasks(void)
7335{
7336 struct task_struct *g, *p;
7337 struct sched_attr attr = {
7338 .sched_policy = SCHED_NORMAL,
7339 };
7340
7341 read_lock(&tasklist_lock);
7342 for_each_process_thread(g, p) {
7343 /*
7344 * Only normalize user tasks:
7345 */
7346 if (p->flags & PF_KTHREAD)
7347 continue;
7348
7349 p->se.exec_start = 0;
7350 schedstat_set(p->se.statistics.wait_start, 0);
7351 schedstat_set(p->se.statistics.sleep_start, 0);
7352 schedstat_set(p->se.statistics.block_start, 0);
7353
7354 if (!dl_task(p) && !rt_task(p)) {
7355 /*
7356 * Renice negative nice level userspace
7357 * tasks back to 0:
7358 */
7359 if (task_nice(p) < 0)
7360 set_user_nice(p, 0);
7361 continue;
7362 }
7363
7364 __sched_setscheduler(p, &attr, false, false);
7365 }
7366 read_unlock(&tasklist_lock);
7367}
7368
7369#endif /* CONFIG_MAGIC_SYSRQ */
7370
7371#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7372/*
7373 * These functions are only useful for the IA64 MCA handling, or kdb.
7374 *
7375 * They can only be called when the whole system has been
7376 * stopped - every CPU needs to be quiescent, and no scheduling
7377 * activity can take place. Using them for anything else would
7378 * be a serious bug, and as a result, they aren't even visible
7379 * under any other configuration.
7380 */
7381
7382/**
7383 * curr_task - return the current task for a given CPU.
7384 * @cpu: the processor in question.
7385 *
7386 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7387 *
7388 * Return: The current task for @cpu.
7389 */
7390struct task_struct *curr_task(int cpu)
7391{
7392 return cpu_curr(cpu);
7393}
7394
7395#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7396
7397#ifdef CONFIG_IA64
7398/**
7399 * ia64_set_curr_task - set the current task for a given CPU.
7400 * @cpu: the processor in question.
7401 * @p: the task pointer to set.
7402 *
7403 * Description: This function must only be used when non-maskable interrupts
7404 * are serviced on a separate stack. It allows the architecture to switch the
7405 * notion of the current task on a CPU in a non-blocking manner. This function
7406 * must be called with all CPU's synchronized, and interrupts disabled, the
7407 * and caller must save the original value of the current task (see
7408 * curr_task() above) and restore that value before reenabling interrupts and
7409 * re-starting the system.
7410 *
7411 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7412 */
7413void ia64_set_curr_task(int cpu, struct task_struct *p)
7414{
7415 cpu_curr(cpu) = p;
7416}
7417
7418#endif
7419
7420#ifdef CONFIG_CGROUP_SCHED
7421/* task_group_lock serializes the addition/removal of task groups */
7422static DEFINE_SPINLOCK(task_group_lock);
7423
7424static inline void alloc_uclamp_sched_group(struct task_group *tg,
7425 struct task_group *parent)
7426{
7427#ifdef CONFIG_UCLAMP_TASK_GROUP
7428 enum uclamp_id clamp_id;
7429
7430 for_each_clamp_id(clamp_id) {
7431 uclamp_se_set(&tg->uclamp_req[clamp_id],
7432 uclamp_none(clamp_id), false);
7433 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
7434 }
7435#endif
7436}
7437
7438static void sched_free_group(struct task_group *tg)
7439{
7440 free_fair_sched_group(tg);
7441 free_rt_sched_group(tg);
7442 autogroup_free(tg);
7443 kmem_cache_free(task_group_cache, tg);
7444}
7445
7446/* allocate runqueue etc for a new task group */
7447struct task_group *sched_create_group(struct task_group *parent)
7448{
7449 struct task_group *tg;
7450
7451 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7452 if (!tg)
7453 return ERR_PTR(-ENOMEM);
7454
7455 if (!alloc_fair_sched_group(tg, parent))
7456 goto err;
7457
7458 if (!alloc_rt_sched_group(tg, parent))
7459 goto err;
7460
7461 alloc_uclamp_sched_group(tg, parent);
7462
7463 return tg;
7464
7465err:
7466 sched_free_group(tg);
7467 return ERR_PTR(-ENOMEM);
7468}
7469
7470void sched_online_group(struct task_group *tg, struct task_group *parent)
7471{
7472 unsigned long flags;
7473
7474 spin_lock_irqsave(&task_group_lock, flags);
7475 list_add_rcu(&tg->list, &task_groups);
7476
7477 /* Root should already exist: */
7478 WARN_ON(!parent);
7479
7480 tg->parent = parent;
7481 INIT_LIST_HEAD(&tg->children);
7482 list_add_rcu(&tg->siblings, &parent->children);
7483 spin_unlock_irqrestore(&task_group_lock, flags);
7484
7485 online_fair_sched_group(tg);
7486}
7487
7488/* rcu callback to free various structures associated with a task group */
7489static void sched_free_group_rcu(struct rcu_head *rhp)
7490{
7491 /* Now it should be safe to free those cfs_rqs: */
7492 sched_free_group(container_of(rhp, struct task_group, rcu));
7493}
7494
7495void sched_destroy_group(struct task_group *tg)
7496{
7497 /* Wait for possible concurrent references to cfs_rqs complete: */
7498 call_rcu(&tg->rcu, sched_free_group_rcu);
7499}
7500
7501void sched_offline_group(struct task_group *tg)
7502{
7503 unsigned long flags;
7504
7505 /* End participation in shares distribution: */
7506 unregister_fair_sched_group(tg);
7507
7508 spin_lock_irqsave(&task_group_lock, flags);
7509 list_del_rcu(&tg->list);
7510 list_del_rcu(&tg->siblings);
7511 spin_unlock_irqrestore(&task_group_lock, flags);
7512}
7513
7514static void sched_change_group(struct task_struct *tsk, int type)
7515{
7516 struct task_group *tg;
7517
7518 /*
7519 * All callers are synchronized by task_rq_lock(); we do not use RCU
7520 * which is pointless here. Thus, we pass "true" to task_css_check()
7521 * to prevent lockdep warnings.
7522 */
7523 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7524 struct task_group, css);
7525 tg = autogroup_task_group(tsk, tg);
7526 tsk->sched_task_group = tg;
7527
7528#ifdef CONFIG_FAIR_GROUP_SCHED
7529 if (tsk->sched_class->task_change_group)
7530 tsk->sched_class->task_change_group(tsk, type);
7531 else
7532#endif
7533 set_task_rq(tsk, task_cpu(tsk));
7534}
7535
7536/*
7537 * Change task's runqueue when it moves between groups.
7538 *
7539 * The caller of this function should have put the task in its new group by
7540 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7541 * its new group.
7542 */
7543void sched_move_task(struct task_struct *tsk)
7544{
7545 int queued, running, queue_flags =
7546 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7547 struct rq_flags rf;
7548 struct rq *rq;
7549
7550 rq = task_rq_lock(tsk, &rf);
7551 update_rq_clock(rq);
7552
7553 running = task_current(rq, tsk);
7554 queued = task_on_rq_queued(tsk);
7555
7556 if (queued)
7557 dequeue_task(rq, tsk, queue_flags);
7558 if (running)
7559 put_prev_task(rq, tsk);
7560
7561 sched_change_group(tsk, TASK_MOVE_GROUP);
7562
7563 if (queued)
7564 enqueue_task(rq, tsk, queue_flags);
7565 if (running) {
7566 set_next_task(rq, tsk);
7567 /*
7568 * After changing group, the running task may have joined a
7569 * throttled one but it's still the running task. Trigger a
7570 * resched to make sure that task can still run.
7571 */
7572 resched_curr(rq);
7573 }
7574
7575 task_rq_unlock(rq, tsk, &rf);
7576}
7577
7578static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7579{
7580 return css ? container_of(css, struct task_group, css) : NULL;
7581}
7582
7583static struct cgroup_subsys_state *
7584cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7585{
7586 struct task_group *parent = css_tg(parent_css);
7587 struct task_group *tg;
7588
7589 if (!parent) {
7590 /* This is early initialization for the top cgroup */
7591 return &root_task_group.css;
7592 }
7593
7594 tg = sched_create_group(parent);
7595 if (IS_ERR(tg))
7596 return ERR_PTR(-ENOMEM);
7597
7598 return &tg->css;
7599}
7600
7601/* Expose task group only after completing cgroup initialization */
7602static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7603{
7604 struct task_group *tg = css_tg(css);
7605 struct task_group *parent = css_tg(css->parent);
7606
7607 if (parent)
7608 sched_online_group(tg, parent);
7609
7610#ifdef CONFIG_UCLAMP_TASK_GROUP
7611 /* Propagate the effective uclamp value for the new group */
7612 cpu_util_update_eff(css);
7613#endif
7614
7615 return 0;
7616}
7617
7618static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
7619{
7620 struct task_group *tg = css_tg(css);
7621
7622 sched_offline_group(tg);
7623}
7624
7625static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7626{
7627 struct task_group *tg = css_tg(css);
7628
7629 /*
7630 * Relies on the RCU grace period between css_released() and this.
7631 */
7632 sched_free_group(tg);
7633}
7634
7635/*
7636 * This is called before wake_up_new_task(), therefore we really only
7637 * have to set its group bits, all the other stuff does not apply.
7638 */
7639static void cpu_cgroup_fork(struct task_struct *task)
7640{
7641 struct rq_flags rf;
7642 struct rq *rq;
7643
7644 rq = task_rq_lock(task, &rf);
7645
7646 update_rq_clock(rq);
7647 sched_change_group(task, TASK_SET_GROUP);
7648
7649 task_rq_unlock(rq, task, &rf);
7650}
7651
7652static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
7653{
7654 struct task_struct *task;
7655 struct cgroup_subsys_state *css;
7656 int ret = 0;
7657
7658 cgroup_taskset_for_each(task, css, tset) {
7659#ifdef CONFIG_RT_GROUP_SCHED
7660 if (!sched_rt_can_attach(css_tg(css), task))
7661 return -EINVAL;
7662#endif
7663 /*
7664 * Serialize against wake_up_new_task() such that if its
7665 * running, we're sure to observe its full state.
7666 */
7667 raw_spin_lock_irq(&task->pi_lock);
7668 /*
7669 * Avoid calling sched_move_task() before wake_up_new_task()
7670 * has happened. This would lead to problems with PELT, due to
7671 * move wanting to detach+attach while we're not attached yet.
7672 */
7673 if (task->state == TASK_NEW)
7674 ret = -EINVAL;
7675 raw_spin_unlock_irq(&task->pi_lock);
7676
7677 if (ret)
7678 break;
7679 }
7680 return ret;
7681}
7682
7683static void cpu_cgroup_attach(struct cgroup_taskset *tset)
7684{
7685 struct task_struct *task;
7686 struct cgroup_subsys_state *css;
7687
7688 cgroup_taskset_for_each(task, css, tset)
7689 sched_move_task(task);
7690}
7691
7692#ifdef CONFIG_UCLAMP_TASK_GROUP
7693static void cpu_util_update_eff(struct cgroup_subsys_state *css)
7694{
7695 struct cgroup_subsys_state *top_css = css;
7696 struct uclamp_se *uc_parent = NULL;
7697 struct uclamp_se *uc_se = NULL;
7698 unsigned int eff[UCLAMP_CNT];
7699 enum uclamp_id clamp_id;
7700 unsigned int clamps;
7701
7702 css_for_each_descendant_pre(css, top_css) {
7703 uc_parent = css_tg(css)->parent
7704 ? css_tg(css)->parent->uclamp : NULL;
7705
7706 for_each_clamp_id(clamp_id) {
7707 /* Assume effective clamps matches requested clamps */
7708 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
7709 /* Cap effective clamps with parent's effective clamps */
7710 if (uc_parent &&
7711 eff[clamp_id] > uc_parent[clamp_id].value) {
7712 eff[clamp_id] = uc_parent[clamp_id].value;
7713 }
7714 }
7715 /* Ensure protection is always capped by limit */
7716 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
7717
7718 /* Propagate most restrictive effective clamps */
7719 clamps = 0x0;
7720 uc_se = css_tg(css)->uclamp;
7721 for_each_clamp_id(clamp_id) {
7722 if (eff[clamp_id] == uc_se[clamp_id].value)
7723 continue;
7724 uc_se[clamp_id].value = eff[clamp_id];
7725 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
7726 clamps |= (0x1 << clamp_id);
7727 }
7728 if (!clamps) {
7729 css = css_rightmost_descendant(css);
7730 continue;
7731 }
7732
7733 /* Immediately update descendants RUNNABLE tasks */
7734 uclamp_update_active_tasks(css, clamps);
7735 }
7736}
7737
7738/*
7739 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7740 * C expression. Since there is no way to convert a macro argument (N) into a
7741 * character constant, use two levels of macros.
7742 */
7743#define _POW10(exp) ((unsigned int)1e##exp)
7744#define POW10(exp) _POW10(exp)
7745
7746struct uclamp_request {
7747#define UCLAMP_PERCENT_SHIFT 2
7748#define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7749 s64 percent;
7750 u64 util;
7751 int ret;
7752};
7753
7754static inline struct uclamp_request
7755capacity_from_percent(char *buf)
7756{
7757 struct uclamp_request req = {
7758 .percent = UCLAMP_PERCENT_SCALE,
7759 .util = SCHED_CAPACITY_SCALE,
7760 .ret = 0,
7761 };
7762
7763 buf = strim(buf);
7764 if (strcmp(buf, "max")) {
7765 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
7766 &req.percent);
7767 if (req.ret)
7768 return req;
7769 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
7770 req.ret = -ERANGE;
7771 return req;
7772 }
7773
7774 req.util = req.percent << SCHED_CAPACITY_SHIFT;
7775 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
7776 }
7777
7778 return req;
7779}
7780
7781static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
7782 size_t nbytes, loff_t off,
7783 enum uclamp_id clamp_id)
7784{
7785 struct uclamp_request req;
7786 struct task_group *tg;
7787
7788 req = capacity_from_percent(buf);
7789 if (req.ret)
7790 return req.ret;
7791
7792 static_branch_enable(&sched_uclamp_used);
7793
7794 mutex_lock(&uclamp_mutex);
7795 rcu_read_lock();
7796
7797 tg = css_tg(of_css(of));
7798 if (tg->uclamp_req[clamp_id].value != req.util)
7799 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
7800
7801 /*
7802 * Because of not recoverable conversion rounding we keep track of the
7803 * exact requested value
7804 */
7805 tg->uclamp_pct[clamp_id] = req.percent;
7806
7807 /* Update effective clamps to track the most restrictive value */
7808 cpu_util_update_eff(of_css(of));
7809
7810 rcu_read_unlock();
7811 mutex_unlock(&uclamp_mutex);
7812
7813 return nbytes;
7814}
7815
7816static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
7817 char *buf, size_t nbytes,
7818 loff_t off)
7819{
7820 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
7821}
7822
7823static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
7824 char *buf, size_t nbytes,
7825 loff_t off)
7826{
7827 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
7828}
7829
7830static inline void cpu_uclamp_print(struct seq_file *sf,
7831 enum uclamp_id clamp_id)
7832{
7833 struct task_group *tg;
7834 u64 util_clamp;
7835 u64 percent;
7836 u32 rem;
7837
7838 rcu_read_lock();
7839 tg = css_tg(seq_css(sf));
7840 util_clamp = tg->uclamp_req[clamp_id].value;
7841 rcu_read_unlock();
7842
7843 if (util_clamp == SCHED_CAPACITY_SCALE) {
7844 seq_puts(sf, "max\n");
7845 return;
7846 }
7847
7848 percent = tg->uclamp_pct[clamp_id];
7849 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
7850 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
7851}
7852
7853static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
7854{
7855 cpu_uclamp_print(sf, UCLAMP_MIN);
7856 return 0;
7857}
7858
7859static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
7860{
7861 cpu_uclamp_print(sf, UCLAMP_MAX);
7862 return 0;
7863}
7864#endif /* CONFIG_UCLAMP_TASK_GROUP */
7865
7866#ifdef CONFIG_FAIR_GROUP_SCHED
7867static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7868 struct cftype *cftype, u64 shareval)
7869{
7870 if (shareval > scale_load_down(ULONG_MAX))
7871 shareval = MAX_SHARES;
7872 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7873}
7874
7875static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7876 struct cftype *cft)
7877{
7878 struct task_group *tg = css_tg(css);
7879
7880 return (u64) scale_load_down(tg->shares);
7881}
7882
7883#ifdef CONFIG_CFS_BANDWIDTH
7884static DEFINE_MUTEX(cfs_constraints_mutex);
7885
7886const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7887static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7888/* More than 203 days if BW_SHIFT equals 20. */
7889static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
7890
7891static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7892
7893static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7894{
7895 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7896 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7897
7898 if (tg == &root_task_group)
7899 return -EINVAL;
7900
7901 /*
7902 * Ensure we have at some amount of bandwidth every period. This is
7903 * to prevent reaching a state of large arrears when throttled via
7904 * entity_tick() resulting in prolonged exit starvation.
7905 */
7906 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7907 return -EINVAL;
7908
7909 /*
7910 * Likewise, bound things on the otherside by preventing insane quota
7911 * periods. This also allows us to normalize in computing quota
7912 * feasibility.
7913 */
7914 if (period > max_cfs_quota_period)
7915 return -EINVAL;
7916
7917 /*
7918 * Bound quota to defend quota against overflow during bandwidth shift.
7919 */
7920 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
7921 return -EINVAL;
7922
7923 /*
7924 * Prevent race between setting of cfs_rq->runtime_enabled and
7925 * unthrottle_offline_cfs_rqs().
7926 */
7927 get_online_cpus();
7928 mutex_lock(&cfs_constraints_mutex);
7929 ret = __cfs_schedulable(tg, period, quota);
7930 if (ret)
7931 goto out_unlock;
7932
7933 runtime_enabled = quota != RUNTIME_INF;
7934 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7935 /*
7936 * If we need to toggle cfs_bandwidth_used, off->on must occur
7937 * before making related changes, and on->off must occur afterwards
7938 */
7939 if (runtime_enabled && !runtime_was_enabled)
7940 cfs_bandwidth_usage_inc();
7941 raw_spin_lock_irq(&cfs_b->lock);
7942 cfs_b->period = ns_to_ktime(period);
7943 cfs_b->quota = quota;
7944
7945 __refill_cfs_bandwidth_runtime(cfs_b);
7946
7947 /* Restart the period timer (if active) to handle new period expiry: */
7948 if (runtime_enabled)
7949 start_cfs_bandwidth(cfs_b);
7950
7951 raw_spin_unlock_irq(&cfs_b->lock);
7952
7953 for_each_online_cpu(i) {
7954 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7955 struct rq *rq = cfs_rq->rq;
7956 struct rq_flags rf;
7957
7958 rq_lock_irq(rq, &rf);
7959 cfs_rq->runtime_enabled = runtime_enabled;
7960 cfs_rq->runtime_remaining = 0;
7961
7962 if (cfs_rq->throttled)
7963 unthrottle_cfs_rq(cfs_rq);
7964 rq_unlock_irq(rq, &rf);
7965 }
7966 if (runtime_was_enabled && !runtime_enabled)
7967 cfs_bandwidth_usage_dec();
7968out_unlock:
7969 mutex_unlock(&cfs_constraints_mutex);
7970 put_online_cpus();
7971
7972 return ret;
7973}
7974
7975static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7976{
7977 u64 quota, period;
7978
7979 period = ktime_to_ns(tg->cfs_bandwidth.period);
7980 if (cfs_quota_us < 0)
7981 quota = RUNTIME_INF;
7982 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
7983 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7984 else
7985 return -EINVAL;
7986
7987 return tg_set_cfs_bandwidth(tg, period, quota);
7988}
7989
7990static long tg_get_cfs_quota(struct task_group *tg)
7991{
7992 u64 quota_us;
7993
7994 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7995 return -1;
7996
7997 quota_us = tg->cfs_bandwidth.quota;
7998 do_div(quota_us, NSEC_PER_USEC);
7999
8000 return quota_us;
8001}
8002
8003static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8004{
8005 u64 quota, period;
8006
8007 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
8008 return -EINVAL;
8009
8010 period = (u64)cfs_period_us * NSEC_PER_USEC;
8011 quota = tg->cfs_bandwidth.quota;
8012
8013 return tg_set_cfs_bandwidth(tg, period, quota);
8014}
8015
8016static long tg_get_cfs_period(struct task_group *tg)
8017{
8018 u64 cfs_period_us;
8019
8020 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8021 do_div(cfs_period_us, NSEC_PER_USEC);
8022
8023 return cfs_period_us;
8024}
8025
8026static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8027 struct cftype *cft)
8028{
8029 return tg_get_cfs_quota(css_tg(css));
8030}
8031
8032static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8033 struct cftype *cftype, s64 cfs_quota_us)
8034{
8035 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8036}
8037
8038static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8039 struct cftype *cft)
8040{
8041 return tg_get_cfs_period(css_tg(css));
8042}
8043
8044static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8045 struct cftype *cftype, u64 cfs_period_us)
8046{
8047 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8048}
8049
8050struct cfs_schedulable_data {
8051 struct task_group *tg;
8052 u64 period, quota;
8053};
8054
8055/*
8056 * normalize group quota/period to be quota/max_period
8057 * note: units are usecs
8058 */
8059static u64 normalize_cfs_quota(struct task_group *tg,
8060 struct cfs_schedulable_data *d)
8061{
8062 u64 quota, period;
8063
8064 if (tg == d->tg) {
8065 period = d->period;
8066 quota = d->quota;
8067 } else {
8068 period = tg_get_cfs_period(tg);
8069 quota = tg_get_cfs_quota(tg);
8070 }
8071
8072 /* note: these should typically be equivalent */
8073 if (quota == RUNTIME_INF || quota == -1)
8074 return RUNTIME_INF;
8075
8076 return to_ratio(period, quota);
8077}
8078
8079static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8080{
8081 struct cfs_schedulable_data *d = data;
8082 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8083 s64 quota = 0, parent_quota = -1;
8084
8085 if (!tg->parent) {
8086 quota = RUNTIME_INF;
8087 } else {
8088 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8089
8090 quota = normalize_cfs_quota(tg, d);
8091 parent_quota = parent_b->hierarchical_quota;
8092
8093 /*
8094 * Ensure max(child_quota) <= parent_quota. On cgroup2,
8095 * always take the min. On cgroup1, only inherit when no
8096 * limit is set:
8097 */
8098 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
8099 quota = min(quota, parent_quota);
8100 } else {
8101 if (quota == RUNTIME_INF)
8102 quota = parent_quota;
8103 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8104 return -EINVAL;
8105 }
8106 }
8107 cfs_b->hierarchical_quota = quota;
8108
8109 return 0;
8110}
8111
8112static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8113{
8114 int ret;
8115 struct cfs_schedulable_data data = {
8116 .tg = tg,
8117 .period = period,
8118 .quota = quota,
8119 };
8120
8121 if (quota != RUNTIME_INF) {
8122 do_div(data.period, NSEC_PER_USEC);
8123 do_div(data.quota, NSEC_PER_USEC);
8124 }
8125
8126 rcu_read_lock();
8127 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8128 rcu_read_unlock();
8129
8130 return ret;
8131}
8132
8133static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
8134{
8135 struct task_group *tg = css_tg(seq_css(sf));
8136 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8137
8138 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8139 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8140 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8141
8142 if (schedstat_enabled() && tg != &root_task_group) {
8143 u64 ws = 0;
8144 int i;
8145
8146 for_each_possible_cpu(i)
8147 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
8148
8149 seq_printf(sf, "wait_sum %llu\n", ws);
8150 }
8151
8152 return 0;
8153}
8154#endif /* CONFIG_CFS_BANDWIDTH */
8155#endif /* CONFIG_FAIR_GROUP_SCHED */
8156
8157#ifdef CONFIG_RT_GROUP_SCHED
8158static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8159 struct cftype *cft, s64 val)
8160{
8161 return sched_group_set_rt_runtime(css_tg(css), val);
8162}
8163
8164static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8165 struct cftype *cft)
8166{
8167 return sched_group_rt_runtime(css_tg(css));
8168}
8169
8170static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8171 struct cftype *cftype, u64 rt_period_us)
8172{
8173 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8174}
8175
8176static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8177 struct cftype *cft)
8178{
8179 return sched_group_rt_period(css_tg(css));
8180}
8181#endif /* CONFIG_RT_GROUP_SCHED */
8182
8183static struct cftype cpu_legacy_files[] = {
8184#ifdef CONFIG_FAIR_GROUP_SCHED
8185 {
8186 .name = "shares",
8187 .read_u64 = cpu_shares_read_u64,
8188 .write_u64 = cpu_shares_write_u64,
8189 },
8190#endif
8191#ifdef CONFIG_CFS_BANDWIDTH
8192 {
8193 .name = "cfs_quota_us",
8194 .read_s64 = cpu_cfs_quota_read_s64,
8195 .write_s64 = cpu_cfs_quota_write_s64,
8196 },
8197 {
8198 .name = "cfs_period_us",
8199 .read_u64 = cpu_cfs_period_read_u64,
8200 .write_u64 = cpu_cfs_period_write_u64,
8201 },
8202 {
8203 .name = "stat",
8204 .seq_show = cpu_cfs_stat_show,
8205 },
8206#endif
8207#ifdef CONFIG_RT_GROUP_SCHED
8208 {
8209 .name = "rt_runtime_us",
8210 .read_s64 = cpu_rt_runtime_read,
8211 .write_s64 = cpu_rt_runtime_write,
8212 },
8213 {
8214 .name = "rt_period_us",
8215 .read_u64 = cpu_rt_period_read_uint,
8216 .write_u64 = cpu_rt_period_write_uint,
8217 },
8218#endif
8219#ifdef CONFIG_UCLAMP_TASK_GROUP
8220 {
8221 .name = "uclamp.min",
8222 .flags = CFTYPE_NOT_ON_ROOT,
8223 .seq_show = cpu_uclamp_min_show,
8224 .write = cpu_uclamp_min_write,
8225 },
8226 {
8227 .name = "uclamp.max",
8228 .flags = CFTYPE_NOT_ON_ROOT,
8229 .seq_show = cpu_uclamp_max_show,
8230 .write = cpu_uclamp_max_write,
8231 },
8232#endif
8233 { } /* Terminate */
8234};
8235
8236static int cpu_extra_stat_show(struct seq_file *sf,
8237 struct cgroup_subsys_state *css)
8238{
8239#ifdef CONFIG_CFS_BANDWIDTH
8240 {
8241 struct task_group *tg = css_tg(css);
8242 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8243 u64 throttled_usec;
8244
8245 throttled_usec = cfs_b->throttled_time;
8246 do_div(throttled_usec, NSEC_PER_USEC);
8247
8248 seq_printf(sf, "nr_periods %d\n"
8249 "nr_throttled %d\n"
8250 "throttled_usec %llu\n",
8251 cfs_b->nr_periods, cfs_b->nr_throttled,
8252 throttled_usec);
8253 }
8254#endif
8255 return 0;
8256}
8257
8258#ifdef CONFIG_FAIR_GROUP_SCHED
8259static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
8260 struct cftype *cft)
8261{
8262 struct task_group *tg = css_tg(css);
8263 u64 weight = scale_load_down(tg->shares);
8264
8265 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
8266}
8267
8268static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
8269 struct cftype *cft, u64 weight)
8270{
8271 /*
8272 * cgroup weight knobs should use the common MIN, DFL and MAX
8273 * values which are 1, 100 and 10000 respectively. While it loses
8274 * a bit of range on both ends, it maps pretty well onto the shares
8275 * value used by scheduler and the round-trip conversions preserve
8276 * the original value over the entire range.
8277 */
8278 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
8279 return -ERANGE;
8280
8281 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
8282
8283 return sched_group_set_shares(css_tg(css), scale_load(weight));
8284}
8285
8286static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
8287 struct cftype *cft)
8288{
8289 unsigned long weight = scale_load_down(css_tg(css)->shares);
8290 int last_delta = INT_MAX;
8291 int prio, delta;
8292
8293 /* find the closest nice value to the current weight */
8294 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
8295 delta = abs(sched_prio_to_weight[prio] - weight);
8296 if (delta >= last_delta)
8297 break;
8298 last_delta = delta;
8299 }
8300
8301 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
8302}
8303
8304static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
8305 struct cftype *cft, s64 nice)
8306{
8307 unsigned long weight;
8308 int idx;
8309
8310 if (nice < MIN_NICE || nice > MAX_NICE)
8311 return -ERANGE;
8312
8313 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
8314 idx = array_index_nospec(idx, 40);
8315 weight = sched_prio_to_weight[idx];
8316
8317 return sched_group_set_shares(css_tg(css), scale_load(weight));
8318}
8319#endif
8320
8321static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
8322 long period, long quota)
8323{
8324 if (quota < 0)
8325 seq_puts(sf, "max");
8326 else
8327 seq_printf(sf, "%ld", quota);
8328
8329 seq_printf(sf, " %ld\n", period);
8330}
8331
8332/* caller should put the current value in *@periodp before calling */
8333static int __maybe_unused cpu_period_quota_parse(char *buf,
8334 u64 *periodp, u64 *quotap)
8335{
8336 char tok[21]; /* U64_MAX */
8337
8338 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
8339 return -EINVAL;
8340
8341 *periodp *= NSEC_PER_USEC;
8342
8343 if (sscanf(tok, "%llu", quotap))
8344 *quotap *= NSEC_PER_USEC;
8345 else if (!strcmp(tok, "max"))
8346 *quotap = RUNTIME_INF;
8347 else
8348 return -EINVAL;
8349
8350 return 0;
8351}
8352
8353#ifdef CONFIG_CFS_BANDWIDTH
8354static int cpu_max_show(struct seq_file *sf, void *v)
8355{
8356 struct task_group *tg = css_tg(seq_css(sf));
8357
8358 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
8359 return 0;
8360}
8361
8362static ssize_t cpu_max_write(struct kernfs_open_file *of,
8363 char *buf, size_t nbytes, loff_t off)
8364{
8365 struct task_group *tg = css_tg(of_css(of));
8366 u64 period = tg_get_cfs_period(tg);
8367 u64 quota;
8368 int ret;
8369
8370 ret = cpu_period_quota_parse(buf, &period, "a);
8371 if (!ret)
8372 ret = tg_set_cfs_bandwidth(tg, period, quota);
8373 return ret ?: nbytes;
8374}
8375#endif
8376
8377static struct cftype cpu_files[] = {
8378#ifdef CONFIG_FAIR_GROUP_SCHED
8379 {
8380 .name = "weight",
8381 .flags = CFTYPE_NOT_ON_ROOT,
8382 .read_u64 = cpu_weight_read_u64,
8383 .write_u64 = cpu_weight_write_u64,
8384 },
8385 {
8386 .name = "weight.nice",
8387 .flags = CFTYPE_NOT_ON_ROOT,
8388 .read_s64 = cpu_weight_nice_read_s64,
8389 .write_s64 = cpu_weight_nice_write_s64,
8390 },
8391#endif
8392#ifdef CONFIG_CFS_BANDWIDTH
8393 {
8394 .name = "max",
8395 .flags = CFTYPE_NOT_ON_ROOT,
8396 .seq_show = cpu_max_show,
8397 .write = cpu_max_write,
8398 },
8399#endif
8400#ifdef CONFIG_UCLAMP_TASK_GROUP
8401 {
8402 .name = "uclamp.min",
8403 .flags = CFTYPE_NOT_ON_ROOT,
8404 .seq_show = cpu_uclamp_min_show,
8405 .write = cpu_uclamp_min_write,
8406 },
8407 {
8408 .name = "uclamp.max",
8409 .flags = CFTYPE_NOT_ON_ROOT,
8410 .seq_show = cpu_uclamp_max_show,
8411 .write = cpu_uclamp_max_write,
8412 },
8413#endif
8414 { } /* terminate */
8415};
8416
8417struct cgroup_subsys cpu_cgrp_subsys = {
8418 .css_alloc = cpu_cgroup_css_alloc,
8419 .css_online = cpu_cgroup_css_online,
8420 .css_released = cpu_cgroup_css_released,
8421 .css_free = cpu_cgroup_css_free,
8422 .css_extra_stat_show = cpu_extra_stat_show,
8423 .fork = cpu_cgroup_fork,
8424 .can_attach = cpu_cgroup_can_attach,
8425 .attach = cpu_cgroup_attach,
8426 .legacy_cftypes = cpu_legacy_files,
8427 .dfl_cftypes = cpu_files,
8428 .early_init = true,
8429 .threaded = true,
8430};
8431
8432#endif /* CONFIG_CGROUP_SCHED */
8433
8434void dump_cpu_task(int cpu)
8435{
8436 pr_info("Task dump for CPU %d:\n", cpu);
8437 sched_show_task(cpu_curr(cpu));
8438}
8439
8440/*
8441 * Nice levels are multiplicative, with a gentle 10% change for every
8442 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8443 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8444 * that remained on nice 0.
8445 *
8446 * The "10% effect" is relative and cumulative: from _any_ nice level,
8447 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8448 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8449 * If a task goes up by ~10% and another task goes down by ~10% then
8450 * the relative distance between them is ~25%.)
8451 */
8452const int sched_prio_to_weight[40] = {
8453 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8454 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8455 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8456 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8457 /* 0 */ 1024, 820, 655, 526, 423,
8458 /* 5 */ 335, 272, 215, 172, 137,
8459 /* 10 */ 110, 87, 70, 56, 45,
8460 /* 15 */ 36, 29, 23, 18, 15,
8461};
8462
8463/*
8464 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8465 *
8466 * In cases where the weight does not change often, we can use the
8467 * precalculated inverse to speed up arithmetics by turning divisions
8468 * into multiplications:
8469 */
8470const u32 sched_prio_to_wmult[40] = {
8471 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8472 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8473 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8474 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8475 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8476 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8477 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8478 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8479};
8480
8481void call_trace_sched_update_nr_running(struct rq *rq, int count)
8482{
8483 trace_sched_update_nr_running_tp(rq, count);
8484}