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/*
2 * kernel/sched/core.c
3 *
4 * Kernel scheduler and related syscalls
5 *
6 * Copyright (C) 1991-2002 Linus Torvalds
7 *
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
11 * by Andrea Arcangeli
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
27 */
28
29#include <linux/kasan.h>
30#include <linux/mm.h>
31#include <linux/module.h>
32#include <linux/nmi.h>
33#include <linux/init.h>
34#include <linux/uaccess.h>
35#include <linux/highmem.h>
36#include <linux/mmu_context.h>
37#include <linux/interrupt.h>
38#include <linux/capability.h>
39#include <linux/completion.h>
40#include <linux/kernel_stat.h>
41#include <linux/debug_locks.h>
42#include <linux/perf_event.h>
43#include <linux/security.h>
44#include <linux/notifier.h>
45#include <linux/profile.h>
46#include <linux/freezer.h>
47#include <linux/vmalloc.h>
48#include <linux/blkdev.h>
49#include <linux/delay.h>
50#include <linux/pid_namespace.h>
51#include <linux/smp.h>
52#include <linux/threads.h>
53#include <linux/timer.h>
54#include <linux/rcupdate.h>
55#include <linux/cpu.h>
56#include <linux/cpuset.h>
57#include <linux/percpu.h>
58#include <linux/proc_fs.h>
59#include <linux/seq_file.h>
60#include <linux/sysctl.h>
61#include <linux/syscalls.h>
62#include <linux/times.h>
63#include <linux/tsacct_kern.h>
64#include <linux/kprobes.h>
65#include <linux/delayacct.h>
66#include <linux/unistd.h>
67#include <linux/pagemap.h>
68#include <linux/hrtimer.h>
69#include <linux/tick.h>
70#include <linux/ctype.h>
71#include <linux/ftrace.h>
72#include <linux/slab.h>
73#include <linux/init_task.h>
74#include <linux/context_tracking.h>
75#include <linux/compiler.h>
76#include <linux/frame.h>
77#include <linux/prefetch.h>
78#include <linux/mutex.h>
79
80#include <asm/switch_to.h>
81#include <asm/tlb.h>
82#include <asm/irq_regs.h>
83#ifdef CONFIG_PARAVIRT
84#include <asm/paravirt.h>
85#endif
86
87#include "sched.h"
88#include "../workqueue_internal.h"
89#include "../smpboot.h"
90
91#define CREATE_TRACE_POINTS
92#include <trace/events/sched.h>
93
94DEFINE_MUTEX(sched_domains_mutex);
95DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
96
97static void update_rq_clock_task(struct rq *rq, s64 delta);
98
99void update_rq_clock(struct rq *rq)
100{
101 s64 delta;
102
103 lockdep_assert_held(&rq->lock);
104
105 if (rq->clock_skip_update & RQCF_ACT_SKIP)
106 return;
107
108 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
109 if (delta < 0)
110 return;
111 rq->clock += delta;
112 update_rq_clock_task(rq, delta);
113}
114
115/*
116 * Debugging: various feature bits
117 */
118
119#define SCHED_FEAT(name, enabled) \
120 (1UL << __SCHED_FEAT_##name) * enabled |
121
122const_debug unsigned int sysctl_sched_features =
123#include "features.h"
124 0;
125
126#undef SCHED_FEAT
127
128/*
129 * Number of tasks to iterate in a single balance run.
130 * Limited because this is done with IRQs disabled.
131 */
132const_debug unsigned int sysctl_sched_nr_migrate = 32;
133
134/*
135 * period over which we average the RT time consumption, measured
136 * in ms.
137 *
138 * default: 1s
139 */
140const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
141
142/*
143 * period over which we measure -rt task cpu usage in us.
144 * default: 1s
145 */
146unsigned int sysctl_sched_rt_period = 1000000;
147
148__read_mostly int scheduler_running;
149
150/*
151 * part of the period that we allow rt tasks to run in us.
152 * default: 0.95s
153 */
154int sysctl_sched_rt_runtime = 950000;
155
156/* cpus with isolated domains */
157cpumask_var_t cpu_isolated_map;
158
159/*
160 * this_rq_lock - lock this runqueue and disable interrupts.
161 */
162static struct rq *this_rq_lock(void)
163 __acquires(rq->lock)
164{
165 struct rq *rq;
166
167 local_irq_disable();
168 rq = this_rq();
169 raw_spin_lock(&rq->lock);
170
171 return rq;
172}
173
174/*
175 * __task_rq_lock - lock the rq @p resides on.
176 */
177struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
178 __acquires(rq->lock)
179{
180 struct rq *rq;
181
182 lockdep_assert_held(&p->pi_lock);
183
184 for (;;) {
185 rq = task_rq(p);
186 raw_spin_lock(&rq->lock);
187 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
188 rf->cookie = lockdep_pin_lock(&rq->lock);
189 return rq;
190 }
191 raw_spin_unlock(&rq->lock);
192
193 while (unlikely(task_on_rq_migrating(p)))
194 cpu_relax();
195 }
196}
197
198/*
199 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
200 */
201struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
202 __acquires(p->pi_lock)
203 __acquires(rq->lock)
204{
205 struct rq *rq;
206
207 for (;;) {
208 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
209 rq = task_rq(p);
210 raw_spin_lock(&rq->lock);
211 /*
212 * move_queued_task() task_rq_lock()
213 *
214 * ACQUIRE (rq->lock)
215 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
216 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
217 * [S] ->cpu = new_cpu [L] task_rq()
218 * [L] ->on_rq
219 * RELEASE (rq->lock)
220 *
221 * If we observe the old cpu in task_rq_lock, the acquire of
222 * the old rq->lock will fully serialize against the stores.
223 *
224 * If we observe the new cpu in task_rq_lock, the acquire will
225 * pair with the WMB to ensure we must then also see migrating.
226 */
227 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
228 rf->cookie = lockdep_pin_lock(&rq->lock);
229 return rq;
230 }
231 raw_spin_unlock(&rq->lock);
232 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
233
234 while (unlikely(task_on_rq_migrating(p)))
235 cpu_relax();
236 }
237}
238
239#ifdef CONFIG_SCHED_HRTICK
240/*
241 * Use HR-timers to deliver accurate preemption points.
242 */
243
244static void hrtick_clear(struct rq *rq)
245{
246 if (hrtimer_active(&rq->hrtick_timer))
247 hrtimer_cancel(&rq->hrtick_timer);
248}
249
250/*
251 * High-resolution timer tick.
252 * Runs from hardirq context with interrupts disabled.
253 */
254static enum hrtimer_restart hrtick(struct hrtimer *timer)
255{
256 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
257
258 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
259
260 raw_spin_lock(&rq->lock);
261 update_rq_clock(rq);
262 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
263 raw_spin_unlock(&rq->lock);
264
265 return HRTIMER_NORESTART;
266}
267
268#ifdef CONFIG_SMP
269
270static void __hrtick_restart(struct rq *rq)
271{
272 struct hrtimer *timer = &rq->hrtick_timer;
273
274 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
275}
276
277/*
278 * called from hardirq (IPI) context
279 */
280static void __hrtick_start(void *arg)
281{
282 struct rq *rq = arg;
283
284 raw_spin_lock(&rq->lock);
285 __hrtick_restart(rq);
286 rq->hrtick_csd_pending = 0;
287 raw_spin_unlock(&rq->lock);
288}
289
290/*
291 * Called to set the hrtick timer state.
292 *
293 * called with rq->lock held and irqs disabled
294 */
295void hrtick_start(struct rq *rq, u64 delay)
296{
297 struct hrtimer *timer = &rq->hrtick_timer;
298 ktime_t time;
299 s64 delta;
300
301 /*
302 * Don't schedule slices shorter than 10000ns, that just
303 * doesn't make sense and can cause timer DoS.
304 */
305 delta = max_t(s64, delay, 10000LL);
306 time = ktime_add_ns(timer->base->get_time(), delta);
307
308 hrtimer_set_expires(timer, time);
309
310 if (rq == this_rq()) {
311 __hrtick_restart(rq);
312 } else if (!rq->hrtick_csd_pending) {
313 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
314 rq->hrtick_csd_pending = 1;
315 }
316}
317
318#else
319/*
320 * Called to set the hrtick timer state.
321 *
322 * called with rq->lock held and irqs disabled
323 */
324void hrtick_start(struct rq *rq, u64 delay)
325{
326 /*
327 * Don't schedule slices shorter than 10000ns, that just
328 * doesn't make sense. Rely on vruntime for fairness.
329 */
330 delay = max_t(u64, delay, 10000LL);
331 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
332 HRTIMER_MODE_REL_PINNED);
333}
334#endif /* CONFIG_SMP */
335
336static void init_rq_hrtick(struct rq *rq)
337{
338#ifdef CONFIG_SMP
339 rq->hrtick_csd_pending = 0;
340
341 rq->hrtick_csd.flags = 0;
342 rq->hrtick_csd.func = __hrtick_start;
343 rq->hrtick_csd.info = rq;
344#endif
345
346 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
347 rq->hrtick_timer.function = hrtick;
348}
349#else /* CONFIG_SCHED_HRTICK */
350static inline void hrtick_clear(struct rq *rq)
351{
352}
353
354static inline void init_rq_hrtick(struct rq *rq)
355{
356}
357#endif /* CONFIG_SCHED_HRTICK */
358
359/*
360 * cmpxchg based fetch_or, macro so it works for different integer types
361 */
362#define fetch_or(ptr, mask) \
363 ({ \
364 typeof(ptr) _ptr = (ptr); \
365 typeof(mask) _mask = (mask); \
366 typeof(*_ptr) _old, _val = *_ptr; \
367 \
368 for (;;) { \
369 _old = cmpxchg(_ptr, _val, _val | _mask); \
370 if (_old == _val) \
371 break; \
372 _val = _old; \
373 } \
374 _old; \
375})
376
377#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
378/*
379 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
380 * this avoids any races wrt polling state changes and thereby avoids
381 * spurious IPIs.
382 */
383static bool set_nr_and_not_polling(struct task_struct *p)
384{
385 struct thread_info *ti = task_thread_info(p);
386 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
387}
388
389/*
390 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
391 *
392 * If this returns true, then the idle task promises to call
393 * sched_ttwu_pending() and reschedule soon.
394 */
395static bool set_nr_if_polling(struct task_struct *p)
396{
397 struct thread_info *ti = task_thread_info(p);
398 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
399
400 for (;;) {
401 if (!(val & _TIF_POLLING_NRFLAG))
402 return false;
403 if (val & _TIF_NEED_RESCHED)
404 return true;
405 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
406 if (old == val)
407 break;
408 val = old;
409 }
410 return true;
411}
412
413#else
414static bool set_nr_and_not_polling(struct task_struct *p)
415{
416 set_tsk_need_resched(p);
417 return true;
418}
419
420#ifdef CONFIG_SMP
421static bool set_nr_if_polling(struct task_struct *p)
422{
423 return false;
424}
425#endif
426#endif
427
428void wake_q_add(struct wake_q_head *head, struct task_struct *task)
429{
430 struct wake_q_node *node = &task->wake_q;
431
432 /*
433 * Atomically grab the task, if ->wake_q is !nil already it means
434 * its already queued (either by us or someone else) and will get the
435 * wakeup due to that.
436 *
437 * This cmpxchg() implies a full barrier, which pairs with the write
438 * barrier implied by the wakeup in wake_up_q().
439 */
440 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
441 return;
442
443 get_task_struct(task);
444
445 /*
446 * The head is context local, there can be no concurrency.
447 */
448 *head->lastp = node;
449 head->lastp = &node->next;
450}
451
452void wake_up_q(struct wake_q_head *head)
453{
454 struct wake_q_node *node = head->first;
455
456 while (node != WAKE_Q_TAIL) {
457 struct task_struct *task;
458
459 task = container_of(node, struct task_struct, wake_q);
460 BUG_ON(!task);
461 /* task can safely be re-inserted now */
462 node = node->next;
463 task->wake_q.next = NULL;
464
465 /*
466 * wake_up_process() implies a wmb() to pair with the queueing
467 * in wake_q_add() so as not to miss wakeups.
468 */
469 wake_up_process(task);
470 put_task_struct(task);
471 }
472}
473
474/*
475 * resched_curr - mark rq's current task 'to be rescheduled now'.
476 *
477 * On UP this means the setting of the need_resched flag, on SMP it
478 * might also involve a cross-CPU call to trigger the scheduler on
479 * the target CPU.
480 */
481void resched_curr(struct rq *rq)
482{
483 struct task_struct *curr = rq->curr;
484 int cpu;
485
486 lockdep_assert_held(&rq->lock);
487
488 if (test_tsk_need_resched(curr))
489 return;
490
491 cpu = cpu_of(rq);
492
493 if (cpu == smp_processor_id()) {
494 set_tsk_need_resched(curr);
495 set_preempt_need_resched();
496 return;
497 }
498
499 if (set_nr_and_not_polling(curr))
500 smp_send_reschedule(cpu);
501 else
502 trace_sched_wake_idle_without_ipi(cpu);
503}
504
505void resched_cpu(int cpu)
506{
507 struct rq *rq = cpu_rq(cpu);
508 unsigned long flags;
509
510 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
511 return;
512 resched_curr(rq);
513 raw_spin_unlock_irqrestore(&rq->lock, flags);
514}
515
516#ifdef CONFIG_SMP
517#ifdef CONFIG_NO_HZ_COMMON
518/*
519 * In the semi idle case, use the nearest busy cpu for migrating timers
520 * from an idle cpu. This is good for power-savings.
521 *
522 * We don't do similar optimization for completely idle system, as
523 * selecting an idle cpu will add more delays to the timers than intended
524 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
525 */
526int get_nohz_timer_target(void)
527{
528 int i, cpu = smp_processor_id();
529 struct sched_domain *sd;
530
531 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
532 return cpu;
533
534 rcu_read_lock();
535 for_each_domain(cpu, sd) {
536 for_each_cpu(i, sched_domain_span(sd)) {
537 if (cpu == i)
538 continue;
539
540 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
541 cpu = i;
542 goto unlock;
543 }
544 }
545 }
546
547 if (!is_housekeeping_cpu(cpu))
548 cpu = housekeeping_any_cpu();
549unlock:
550 rcu_read_unlock();
551 return cpu;
552}
553/*
554 * When add_timer_on() enqueues a timer into the timer wheel of an
555 * idle CPU then this timer might expire before the next timer event
556 * which is scheduled to wake up that CPU. In case of a completely
557 * idle system the next event might even be infinite time into the
558 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
559 * leaves the inner idle loop so the newly added timer is taken into
560 * account when the CPU goes back to idle and evaluates the timer
561 * wheel for the next timer event.
562 */
563static void wake_up_idle_cpu(int cpu)
564{
565 struct rq *rq = cpu_rq(cpu);
566
567 if (cpu == smp_processor_id())
568 return;
569
570 if (set_nr_and_not_polling(rq->idle))
571 smp_send_reschedule(cpu);
572 else
573 trace_sched_wake_idle_without_ipi(cpu);
574}
575
576static bool wake_up_full_nohz_cpu(int cpu)
577{
578 /*
579 * We just need the target to call irq_exit() and re-evaluate
580 * the next tick. The nohz full kick at least implies that.
581 * If needed we can still optimize that later with an
582 * empty IRQ.
583 */
584 if (cpu_is_offline(cpu))
585 return true; /* Don't try to wake offline CPUs. */
586 if (tick_nohz_full_cpu(cpu)) {
587 if (cpu != smp_processor_id() ||
588 tick_nohz_tick_stopped())
589 tick_nohz_full_kick_cpu(cpu);
590 return true;
591 }
592
593 return false;
594}
595
596/*
597 * Wake up the specified CPU. If the CPU is going offline, it is the
598 * caller's responsibility to deal with the lost wakeup, for example,
599 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
600 */
601void wake_up_nohz_cpu(int cpu)
602{
603 if (!wake_up_full_nohz_cpu(cpu))
604 wake_up_idle_cpu(cpu);
605}
606
607static inline bool got_nohz_idle_kick(void)
608{
609 int cpu = smp_processor_id();
610
611 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
612 return false;
613
614 if (idle_cpu(cpu) && !need_resched())
615 return true;
616
617 /*
618 * We can't run Idle Load Balance on this CPU for this time so we
619 * cancel it and clear NOHZ_BALANCE_KICK
620 */
621 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
622 return false;
623}
624
625#else /* CONFIG_NO_HZ_COMMON */
626
627static inline bool got_nohz_idle_kick(void)
628{
629 return false;
630}
631
632#endif /* CONFIG_NO_HZ_COMMON */
633
634#ifdef CONFIG_NO_HZ_FULL
635bool sched_can_stop_tick(struct rq *rq)
636{
637 int fifo_nr_running;
638
639 /* Deadline tasks, even if single, need the tick */
640 if (rq->dl.dl_nr_running)
641 return false;
642
643 /*
644 * If there are more than one RR tasks, we need the tick to effect the
645 * actual RR behaviour.
646 */
647 if (rq->rt.rr_nr_running) {
648 if (rq->rt.rr_nr_running == 1)
649 return true;
650 else
651 return false;
652 }
653
654 /*
655 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
656 * forced preemption between FIFO tasks.
657 */
658 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
659 if (fifo_nr_running)
660 return true;
661
662 /*
663 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
664 * if there's more than one we need the tick for involuntary
665 * preemption.
666 */
667 if (rq->nr_running > 1)
668 return false;
669
670 return true;
671}
672#endif /* CONFIG_NO_HZ_FULL */
673
674void sched_avg_update(struct rq *rq)
675{
676 s64 period = sched_avg_period();
677
678 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
679 /*
680 * Inline assembly required to prevent the compiler
681 * optimising this loop into a divmod call.
682 * See __iter_div_u64_rem() for another example of this.
683 */
684 asm("" : "+rm" (rq->age_stamp));
685 rq->age_stamp += period;
686 rq->rt_avg /= 2;
687 }
688}
689
690#endif /* CONFIG_SMP */
691
692#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
693 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
694/*
695 * Iterate task_group tree rooted at *from, calling @down when first entering a
696 * node and @up when leaving it for the final time.
697 *
698 * Caller must hold rcu_lock or sufficient equivalent.
699 */
700int walk_tg_tree_from(struct task_group *from,
701 tg_visitor down, tg_visitor up, void *data)
702{
703 struct task_group *parent, *child;
704 int ret;
705
706 parent = from;
707
708down:
709 ret = (*down)(parent, data);
710 if (ret)
711 goto out;
712 list_for_each_entry_rcu(child, &parent->children, siblings) {
713 parent = child;
714 goto down;
715
716up:
717 continue;
718 }
719 ret = (*up)(parent, data);
720 if (ret || parent == from)
721 goto out;
722
723 child = parent;
724 parent = parent->parent;
725 if (parent)
726 goto up;
727out:
728 return ret;
729}
730
731int tg_nop(struct task_group *tg, void *data)
732{
733 return 0;
734}
735#endif
736
737static void set_load_weight(struct task_struct *p)
738{
739 int prio = p->static_prio - MAX_RT_PRIO;
740 struct load_weight *load = &p->se.load;
741
742 /*
743 * SCHED_IDLE tasks get minimal weight:
744 */
745 if (idle_policy(p->policy)) {
746 load->weight = scale_load(WEIGHT_IDLEPRIO);
747 load->inv_weight = WMULT_IDLEPRIO;
748 return;
749 }
750
751 load->weight = scale_load(sched_prio_to_weight[prio]);
752 load->inv_weight = sched_prio_to_wmult[prio];
753}
754
755static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
756{
757 update_rq_clock(rq);
758 if (!(flags & ENQUEUE_RESTORE))
759 sched_info_queued(rq, p);
760 p->sched_class->enqueue_task(rq, p, flags);
761}
762
763static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
764{
765 update_rq_clock(rq);
766 if (!(flags & DEQUEUE_SAVE))
767 sched_info_dequeued(rq, p);
768 p->sched_class->dequeue_task(rq, p, flags);
769}
770
771void activate_task(struct rq *rq, struct task_struct *p, int flags)
772{
773 if (task_contributes_to_load(p))
774 rq->nr_uninterruptible--;
775
776 enqueue_task(rq, p, flags);
777}
778
779void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
780{
781 if (task_contributes_to_load(p))
782 rq->nr_uninterruptible++;
783
784 dequeue_task(rq, p, flags);
785}
786
787static void update_rq_clock_task(struct rq *rq, s64 delta)
788{
789/*
790 * In theory, the compile should just see 0 here, and optimize out the call
791 * to sched_rt_avg_update. But I don't trust it...
792 */
793#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
794 s64 steal = 0, irq_delta = 0;
795#endif
796#ifdef CONFIG_IRQ_TIME_ACCOUNTING
797 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
798
799 /*
800 * Since irq_time is only updated on {soft,}irq_exit, we might run into
801 * this case when a previous update_rq_clock() happened inside a
802 * {soft,}irq region.
803 *
804 * When this happens, we stop ->clock_task and only update the
805 * prev_irq_time stamp to account for the part that fit, so that a next
806 * update will consume the rest. This ensures ->clock_task is
807 * monotonic.
808 *
809 * It does however cause some slight miss-attribution of {soft,}irq
810 * time, a more accurate solution would be to update the irq_time using
811 * the current rq->clock timestamp, except that would require using
812 * atomic ops.
813 */
814 if (irq_delta > delta)
815 irq_delta = delta;
816
817 rq->prev_irq_time += irq_delta;
818 delta -= irq_delta;
819#endif
820#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
821 if (static_key_false((¶virt_steal_rq_enabled))) {
822 steal = paravirt_steal_clock(cpu_of(rq));
823 steal -= rq->prev_steal_time_rq;
824
825 if (unlikely(steal > delta))
826 steal = delta;
827
828 rq->prev_steal_time_rq += steal;
829 delta -= steal;
830 }
831#endif
832
833 rq->clock_task += delta;
834
835#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
836 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
837 sched_rt_avg_update(rq, irq_delta + steal);
838#endif
839}
840
841void sched_set_stop_task(int cpu, struct task_struct *stop)
842{
843 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
844 struct task_struct *old_stop = cpu_rq(cpu)->stop;
845
846 if (stop) {
847 /*
848 * Make it appear like a SCHED_FIFO task, its something
849 * userspace knows about and won't get confused about.
850 *
851 * Also, it will make PI more or less work without too
852 * much confusion -- but then, stop work should not
853 * rely on PI working anyway.
854 */
855 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
856
857 stop->sched_class = &stop_sched_class;
858 }
859
860 cpu_rq(cpu)->stop = stop;
861
862 if (old_stop) {
863 /*
864 * Reset it back to a normal scheduling class so that
865 * it can die in pieces.
866 */
867 old_stop->sched_class = &rt_sched_class;
868 }
869}
870
871/*
872 * __normal_prio - return the priority that is based on the static prio
873 */
874static inline int __normal_prio(struct task_struct *p)
875{
876 return p->static_prio;
877}
878
879/*
880 * Calculate the expected normal priority: i.e. priority
881 * without taking RT-inheritance into account. Might be
882 * boosted by interactivity modifiers. Changes upon fork,
883 * setprio syscalls, and whenever the interactivity
884 * estimator recalculates.
885 */
886static inline int normal_prio(struct task_struct *p)
887{
888 int prio;
889
890 if (task_has_dl_policy(p))
891 prio = MAX_DL_PRIO-1;
892 else if (task_has_rt_policy(p))
893 prio = MAX_RT_PRIO-1 - p->rt_priority;
894 else
895 prio = __normal_prio(p);
896 return prio;
897}
898
899/*
900 * Calculate the current priority, i.e. the priority
901 * taken into account by the scheduler. This value might
902 * be boosted by RT tasks, or might be boosted by
903 * interactivity modifiers. Will be RT if the task got
904 * RT-boosted. If not then it returns p->normal_prio.
905 */
906static int effective_prio(struct task_struct *p)
907{
908 p->normal_prio = normal_prio(p);
909 /*
910 * If we are RT tasks or we were boosted to RT priority,
911 * keep the priority unchanged. Otherwise, update priority
912 * to the normal priority:
913 */
914 if (!rt_prio(p->prio))
915 return p->normal_prio;
916 return p->prio;
917}
918
919/**
920 * task_curr - is this task currently executing on a CPU?
921 * @p: the task in question.
922 *
923 * Return: 1 if the task is currently executing. 0 otherwise.
924 */
925inline int task_curr(const struct task_struct *p)
926{
927 return cpu_curr(task_cpu(p)) == p;
928}
929
930/*
931 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
932 * use the balance_callback list if you want balancing.
933 *
934 * this means any call to check_class_changed() must be followed by a call to
935 * balance_callback().
936 */
937static inline void check_class_changed(struct rq *rq, struct task_struct *p,
938 const struct sched_class *prev_class,
939 int oldprio)
940{
941 if (prev_class != p->sched_class) {
942 if (prev_class->switched_from)
943 prev_class->switched_from(rq, p);
944
945 p->sched_class->switched_to(rq, p);
946 } else if (oldprio != p->prio || dl_task(p))
947 p->sched_class->prio_changed(rq, p, oldprio);
948}
949
950void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
951{
952 const struct sched_class *class;
953
954 if (p->sched_class == rq->curr->sched_class) {
955 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
956 } else {
957 for_each_class(class) {
958 if (class == rq->curr->sched_class)
959 break;
960 if (class == p->sched_class) {
961 resched_curr(rq);
962 break;
963 }
964 }
965 }
966
967 /*
968 * A queue event has occurred, and we're going to schedule. In
969 * this case, we can save a useless back to back clock update.
970 */
971 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
972 rq_clock_skip_update(rq, true);
973}
974
975#ifdef CONFIG_SMP
976/*
977 * This is how migration works:
978 *
979 * 1) we invoke migration_cpu_stop() on the target CPU using
980 * stop_one_cpu().
981 * 2) stopper starts to run (implicitly forcing the migrated thread
982 * off the CPU)
983 * 3) it checks whether the migrated task is still in the wrong runqueue.
984 * 4) if it's in the wrong runqueue then the migration thread removes
985 * it and puts it into the right queue.
986 * 5) stopper completes and stop_one_cpu() returns and the migration
987 * is done.
988 */
989
990/*
991 * move_queued_task - move a queued task to new rq.
992 *
993 * Returns (locked) new rq. Old rq's lock is released.
994 */
995static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
996{
997 lockdep_assert_held(&rq->lock);
998
999 p->on_rq = TASK_ON_RQ_MIGRATING;
1000 dequeue_task(rq, p, 0);
1001 set_task_cpu(p, new_cpu);
1002 raw_spin_unlock(&rq->lock);
1003
1004 rq = cpu_rq(new_cpu);
1005
1006 raw_spin_lock(&rq->lock);
1007 BUG_ON(task_cpu(p) != new_cpu);
1008 enqueue_task(rq, p, 0);
1009 p->on_rq = TASK_ON_RQ_QUEUED;
1010 check_preempt_curr(rq, p, 0);
1011
1012 return rq;
1013}
1014
1015struct migration_arg {
1016 struct task_struct *task;
1017 int dest_cpu;
1018};
1019
1020/*
1021 * Move (not current) task off this cpu, onto dest cpu. We're doing
1022 * this because either it can't run here any more (set_cpus_allowed()
1023 * away from this CPU, or CPU going down), or because we're
1024 * attempting to rebalance this task on exec (sched_exec).
1025 *
1026 * So we race with normal scheduler movements, but that's OK, as long
1027 * as the task is no longer on this CPU.
1028 */
1029static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1030{
1031 if (unlikely(!cpu_active(dest_cpu)))
1032 return rq;
1033
1034 /* Affinity changed (again). */
1035 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1036 return rq;
1037
1038 rq = move_queued_task(rq, p, dest_cpu);
1039
1040 return rq;
1041}
1042
1043/*
1044 * migration_cpu_stop - this will be executed by a highprio stopper thread
1045 * and performs thread migration by bumping thread off CPU then
1046 * 'pushing' onto another runqueue.
1047 */
1048static int migration_cpu_stop(void *data)
1049{
1050 struct migration_arg *arg = data;
1051 struct task_struct *p = arg->task;
1052 struct rq *rq = this_rq();
1053
1054 /*
1055 * The original target cpu might have gone down and we might
1056 * be on another cpu but it doesn't matter.
1057 */
1058 local_irq_disable();
1059 /*
1060 * We need to explicitly wake pending tasks before running
1061 * __migrate_task() such that we will not miss enforcing cpus_allowed
1062 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1063 */
1064 sched_ttwu_pending();
1065
1066 raw_spin_lock(&p->pi_lock);
1067 raw_spin_lock(&rq->lock);
1068 /*
1069 * If task_rq(p) != rq, it cannot be migrated here, because we're
1070 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1071 * we're holding p->pi_lock.
1072 */
1073 if (task_rq(p) == rq) {
1074 if (task_on_rq_queued(p))
1075 rq = __migrate_task(rq, p, arg->dest_cpu);
1076 else
1077 p->wake_cpu = arg->dest_cpu;
1078 }
1079 raw_spin_unlock(&rq->lock);
1080 raw_spin_unlock(&p->pi_lock);
1081
1082 local_irq_enable();
1083 return 0;
1084}
1085
1086/*
1087 * sched_class::set_cpus_allowed must do the below, but is not required to
1088 * actually call this function.
1089 */
1090void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1091{
1092 cpumask_copy(&p->cpus_allowed, new_mask);
1093 p->nr_cpus_allowed = cpumask_weight(new_mask);
1094}
1095
1096void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1097{
1098 struct rq *rq = task_rq(p);
1099 bool queued, running;
1100
1101 lockdep_assert_held(&p->pi_lock);
1102
1103 queued = task_on_rq_queued(p);
1104 running = task_current(rq, p);
1105
1106 if (queued) {
1107 /*
1108 * Because __kthread_bind() calls this on blocked tasks without
1109 * holding rq->lock.
1110 */
1111 lockdep_assert_held(&rq->lock);
1112 dequeue_task(rq, p, DEQUEUE_SAVE);
1113 }
1114 if (running)
1115 put_prev_task(rq, p);
1116
1117 p->sched_class->set_cpus_allowed(p, new_mask);
1118
1119 if (queued)
1120 enqueue_task(rq, p, ENQUEUE_RESTORE);
1121 if (running)
1122 set_curr_task(rq, p);
1123}
1124
1125/*
1126 * Change a given task's CPU affinity. Migrate the thread to a
1127 * proper CPU and schedule it away if the CPU it's executing on
1128 * is removed from the allowed bitmask.
1129 *
1130 * NOTE: the caller must have a valid reference to the task, the
1131 * task must not exit() & deallocate itself prematurely. The
1132 * call is not atomic; no spinlocks may be held.
1133 */
1134static int __set_cpus_allowed_ptr(struct task_struct *p,
1135 const struct cpumask *new_mask, bool check)
1136{
1137 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1138 unsigned int dest_cpu;
1139 struct rq_flags rf;
1140 struct rq *rq;
1141 int ret = 0;
1142
1143 rq = task_rq_lock(p, &rf);
1144
1145 if (p->flags & PF_KTHREAD) {
1146 /*
1147 * Kernel threads are allowed on online && !active CPUs
1148 */
1149 cpu_valid_mask = cpu_online_mask;
1150 }
1151
1152 /*
1153 * Must re-check here, to close a race against __kthread_bind(),
1154 * sched_setaffinity() is not guaranteed to observe the flag.
1155 */
1156 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1157 ret = -EINVAL;
1158 goto out;
1159 }
1160
1161 if (cpumask_equal(&p->cpus_allowed, new_mask))
1162 goto out;
1163
1164 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1165 ret = -EINVAL;
1166 goto out;
1167 }
1168
1169 do_set_cpus_allowed(p, new_mask);
1170
1171 if (p->flags & PF_KTHREAD) {
1172 /*
1173 * For kernel threads that do indeed end up on online &&
1174 * !active we want to ensure they are strict per-cpu threads.
1175 */
1176 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1177 !cpumask_intersects(new_mask, cpu_active_mask) &&
1178 p->nr_cpus_allowed != 1);
1179 }
1180
1181 /* Can the task run on the task's current CPU? If so, we're done */
1182 if (cpumask_test_cpu(task_cpu(p), new_mask))
1183 goto out;
1184
1185 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1186 if (task_running(rq, p) || p->state == TASK_WAKING) {
1187 struct migration_arg arg = { p, dest_cpu };
1188 /* Need help from migration thread: drop lock and wait. */
1189 task_rq_unlock(rq, p, &rf);
1190 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1191 tlb_migrate_finish(p->mm);
1192 return 0;
1193 } else if (task_on_rq_queued(p)) {
1194 /*
1195 * OK, since we're going to drop the lock immediately
1196 * afterwards anyway.
1197 */
1198 lockdep_unpin_lock(&rq->lock, rf.cookie);
1199 rq = move_queued_task(rq, p, dest_cpu);
1200 lockdep_repin_lock(&rq->lock, rf.cookie);
1201 }
1202out:
1203 task_rq_unlock(rq, p, &rf);
1204
1205 return ret;
1206}
1207
1208int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1209{
1210 return __set_cpus_allowed_ptr(p, new_mask, false);
1211}
1212EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1213
1214void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1215{
1216#ifdef CONFIG_SCHED_DEBUG
1217 /*
1218 * We should never call set_task_cpu() on a blocked task,
1219 * ttwu() will sort out the placement.
1220 */
1221 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1222 !p->on_rq);
1223
1224 /*
1225 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1226 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1227 * time relying on p->on_rq.
1228 */
1229 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1230 p->sched_class == &fair_sched_class &&
1231 (p->on_rq && !task_on_rq_migrating(p)));
1232
1233#ifdef CONFIG_LOCKDEP
1234 /*
1235 * The caller should hold either p->pi_lock or rq->lock, when changing
1236 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1237 *
1238 * sched_move_task() holds both and thus holding either pins the cgroup,
1239 * see task_group().
1240 *
1241 * Furthermore, all task_rq users should acquire both locks, see
1242 * task_rq_lock().
1243 */
1244 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1245 lockdep_is_held(&task_rq(p)->lock)));
1246#endif
1247#endif
1248
1249 trace_sched_migrate_task(p, new_cpu);
1250
1251 if (task_cpu(p) != new_cpu) {
1252 if (p->sched_class->migrate_task_rq)
1253 p->sched_class->migrate_task_rq(p);
1254 p->se.nr_migrations++;
1255 perf_event_task_migrate(p);
1256 }
1257
1258 __set_task_cpu(p, new_cpu);
1259}
1260
1261static void __migrate_swap_task(struct task_struct *p, int cpu)
1262{
1263 if (task_on_rq_queued(p)) {
1264 struct rq *src_rq, *dst_rq;
1265
1266 src_rq = task_rq(p);
1267 dst_rq = cpu_rq(cpu);
1268
1269 p->on_rq = TASK_ON_RQ_MIGRATING;
1270 deactivate_task(src_rq, p, 0);
1271 set_task_cpu(p, cpu);
1272 activate_task(dst_rq, p, 0);
1273 p->on_rq = TASK_ON_RQ_QUEUED;
1274 check_preempt_curr(dst_rq, p, 0);
1275 } else {
1276 /*
1277 * Task isn't running anymore; make it appear like we migrated
1278 * it before it went to sleep. This means on wakeup we make the
1279 * previous cpu our target instead of where it really is.
1280 */
1281 p->wake_cpu = cpu;
1282 }
1283}
1284
1285struct migration_swap_arg {
1286 struct task_struct *src_task, *dst_task;
1287 int src_cpu, dst_cpu;
1288};
1289
1290static int migrate_swap_stop(void *data)
1291{
1292 struct migration_swap_arg *arg = data;
1293 struct rq *src_rq, *dst_rq;
1294 int ret = -EAGAIN;
1295
1296 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1297 return -EAGAIN;
1298
1299 src_rq = cpu_rq(arg->src_cpu);
1300 dst_rq = cpu_rq(arg->dst_cpu);
1301
1302 double_raw_lock(&arg->src_task->pi_lock,
1303 &arg->dst_task->pi_lock);
1304 double_rq_lock(src_rq, dst_rq);
1305
1306 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1307 goto unlock;
1308
1309 if (task_cpu(arg->src_task) != arg->src_cpu)
1310 goto unlock;
1311
1312 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1313 goto unlock;
1314
1315 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1316 goto unlock;
1317
1318 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1319 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1320
1321 ret = 0;
1322
1323unlock:
1324 double_rq_unlock(src_rq, dst_rq);
1325 raw_spin_unlock(&arg->dst_task->pi_lock);
1326 raw_spin_unlock(&arg->src_task->pi_lock);
1327
1328 return ret;
1329}
1330
1331/*
1332 * Cross migrate two tasks
1333 */
1334int migrate_swap(struct task_struct *cur, struct task_struct *p)
1335{
1336 struct migration_swap_arg arg;
1337 int ret = -EINVAL;
1338
1339 arg = (struct migration_swap_arg){
1340 .src_task = cur,
1341 .src_cpu = task_cpu(cur),
1342 .dst_task = p,
1343 .dst_cpu = task_cpu(p),
1344 };
1345
1346 if (arg.src_cpu == arg.dst_cpu)
1347 goto out;
1348
1349 /*
1350 * These three tests are all lockless; this is OK since all of them
1351 * will be re-checked with proper locks held further down the line.
1352 */
1353 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1354 goto out;
1355
1356 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1357 goto out;
1358
1359 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1360 goto out;
1361
1362 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1363 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1364
1365out:
1366 return ret;
1367}
1368
1369/*
1370 * wait_task_inactive - wait for a thread to unschedule.
1371 *
1372 * If @match_state is nonzero, it's the @p->state value just checked and
1373 * not expected to change. If it changes, i.e. @p might have woken up,
1374 * then return zero. When we succeed in waiting for @p to be off its CPU,
1375 * we return a positive number (its total switch count). If a second call
1376 * a short while later returns the same number, the caller can be sure that
1377 * @p has remained unscheduled the whole time.
1378 *
1379 * The caller must ensure that the task *will* unschedule sometime soon,
1380 * else this function might spin for a *long* time. This function can't
1381 * be called with interrupts off, or it may introduce deadlock with
1382 * smp_call_function() if an IPI is sent by the same process we are
1383 * waiting to become inactive.
1384 */
1385unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1386{
1387 int running, queued;
1388 struct rq_flags rf;
1389 unsigned long ncsw;
1390 struct rq *rq;
1391
1392 for (;;) {
1393 /*
1394 * We do the initial early heuristics without holding
1395 * any task-queue locks at all. We'll only try to get
1396 * the runqueue lock when things look like they will
1397 * work out!
1398 */
1399 rq = task_rq(p);
1400
1401 /*
1402 * If the task is actively running on another CPU
1403 * still, just relax and busy-wait without holding
1404 * any locks.
1405 *
1406 * NOTE! Since we don't hold any locks, it's not
1407 * even sure that "rq" stays as the right runqueue!
1408 * But we don't care, since "task_running()" will
1409 * return false if the runqueue has changed and p
1410 * is actually now running somewhere else!
1411 */
1412 while (task_running(rq, p)) {
1413 if (match_state && unlikely(p->state != match_state))
1414 return 0;
1415 cpu_relax();
1416 }
1417
1418 /*
1419 * Ok, time to look more closely! We need the rq
1420 * lock now, to be *sure*. If we're wrong, we'll
1421 * just go back and repeat.
1422 */
1423 rq = task_rq_lock(p, &rf);
1424 trace_sched_wait_task(p);
1425 running = task_running(rq, p);
1426 queued = task_on_rq_queued(p);
1427 ncsw = 0;
1428 if (!match_state || p->state == match_state)
1429 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1430 task_rq_unlock(rq, p, &rf);
1431
1432 /*
1433 * If it changed from the expected state, bail out now.
1434 */
1435 if (unlikely(!ncsw))
1436 break;
1437
1438 /*
1439 * Was it really running after all now that we
1440 * checked with the proper locks actually held?
1441 *
1442 * Oops. Go back and try again..
1443 */
1444 if (unlikely(running)) {
1445 cpu_relax();
1446 continue;
1447 }
1448
1449 /*
1450 * It's not enough that it's not actively running,
1451 * it must be off the runqueue _entirely_, and not
1452 * preempted!
1453 *
1454 * So if it was still runnable (but just not actively
1455 * running right now), it's preempted, and we should
1456 * yield - it could be a while.
1457 */
1458 if (unlikely(queued)) {
1459 ktime_t to = NSEC_PER_SEC / HZ;
1460
1461 set_current_state(TASK_UNINTERRUPTIBLE);
1462 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1463 continue;
1464 }
1465
1466 /*
1467 * Ahh, all good. It wasn't running, and it wasn't
1468 * runnable, which means that it will never become
1469 * running in the future either. We're all done!
1470 */
1471 break;
1472 }
1473
1474 return ncsw;
1475}
1476
1477/***
1478 * kick_process - kick a running thread to enter/exit the kernel
1479 * @p: the to-be-kicked thread
1480 *
1481 * Cause a process which is running on another CPU to enter
1482 * kernel-mode, without any delay. (to get signals handled.)
1483 *
1484 * NOTE: this function doesn't have to take the runqueue lock,
1485 * because all it wants to ensure is that the remote task enters
1486 * the kernel. If the IPI races and the task has been migrated
1487 * to another CPU then no harm is done and the purpose has been
1488 * achieved as well.
1489 */
1490void kick_process(struct task_struct *p)
1491{
1492 int cpu;
1493
1494 preempt_disable();
1495 cpu = task_cpu(p);
1496 if ((cpu != smp_processor_id()) && task_curr(p))
1497 smp_send_reschedule(cpu);
1498 preempt_enable();
1499}
1500EXPORT_SYMBOL_GPL(kick_process);
1501
1502/*
1503 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1504 *
1505 * A few notes on cpu_active vs cpu_online:
1506 *
1507 * - cpu_active must be a subset of cpu_online
1508 *
1509 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1510 * see __set_cpus_allowed_ptr(). At this point the newly online
1511 * cpu isn't yet part of the sched domains, and balancing will not
1512 * see it.
1513 *
1514 * - on cpu-down we clear cpu_active() to mask the sched domains and
1515 * avoid the load balancer to place new tasks on the to be removed
1516 * cpu. Existing tasks will remain running there and will be taken
1517 * off.
1518 *
1519 * This means that fallback selection must not select !active CPUs.
1520 * And can assume that any active CPU must be online. Conversely
1521 * select_task_rq() below may allow selection of !active CPUs in order
1522 * to satisfy the above rules.
1523 */
1524static int select_fallback_rq(int cpu, struct task_struct *p)
1525{
1526 int nid = cpu_to_node(cpu);
1527 const struct cpumask *nodemask = NULL;
1528 enum { cpuset, possible, fail } state = cpuset;
1529 int dest_cpu;
1530
1531 /*
1532 * If the node that the cpu is on has been offlined, cpu_to_node()
1533 * will return -1. There is no cpu on the node, and we should
1534 * select the cpu on the other node.
1535 */
1536 if (nid != -1) {
1537 nodemask = cpumask_of_node(nid);
1538
1539 /* Look for allowed, online CPU in same node. */
1540 for_each_cpu(dest_cpu, nodemask) {
1541 if (!cpu_active(dest_cpu))
1542 continue;
1543 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1544 return dest_cpu;
1545 }
1546 }
1547
1548 for (;;) {
1549 /* Any allowed, online CPU? */
1550 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1551 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1552 continue;
1553 if (!cpu_online(dest_cpu))
1554 continue;
1555 goto out;
1556 }
1557
1558 /* No more Mr. Nice Guy. */
1559 switch (state) {
1560 case cpuset:
1561 if (IS_ENABLED(CONFIG_CPUSETS)) {
1562 cpuset_cpus_allowed_fallback(p);
1563 state = possible;
1564 break;
1565 }
1566 /* fall-through */
1567 case possible:
1568 do_set_cpus_allowed(p, cpu_possible_mask);
1569 state = fail;
1570 break;
1571
1572 case fail:
1573 BUG();
1574 break;
1575 }
1576 }
1577
1578out:
1579 if (state != cpuset) {
1580 /*
1581 * Don't tell them about moving exiting tasks or
1582 * kernel threads (both mm NULL), since they never
1583 * leave kernel.
1584 */
1585 if (p->mm && printk_ratelimit()) {
1586 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1587 task_pid_nr(p), p->comm, cpu);
1588 }
1589 }
1590
1591 return dest_cpu;
1592}
1593
1594/*
1595 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1596 */
1597static inline
1598int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1599{
1600 lockdep_assert_held(&p->pi_lock);
1601
1602 if (tsk_nr_cpus_allowed(p) > 1)
1603 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1604 else
1605 cpu = cpumask_any(tsk_cpus_allowed(p));
1606
1607 /*
1608 * In order not to call set_task_cpu() on a blocking task we need
1609 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1610 * cpu.
1611 *
1612 * Since this is common to all placement strategies, this lives here.
1613 *
1614 * [ this allows ->select_task() to simply return task_cpu(p) and
1615 * not worry about this generic constraint ]
1616 */
1617 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1618 !cpu_online(cpu)))
1619 cpu = select_fallback_rq(task_cpu(p), p);
1620
1621 return cpu;
1622}
1623
1624static void update_avg(u64 *avg, u64 sample)
1625{
1626 s64 diff = sample - *avg;
1627 *avg += diff >> 3;
1628}
1629
1630#else
1631
1632static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1633 const struct cpumask *new_mask, bool check)
1634{
1635 return set_cpus_allowed_ptr(p, new_mask);
1636}
1637
1638#endif /* CONFIG_SMP */
1639
1640static void
1641ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1642{
1643 struct rq *rq;
1644
1645 if (!schedstat_enabled())
1646 return;
1647
1648 rq = this_rq();
1649
1650#ifdef CONFIG_SMP
1651 if (cpu == rq->cpu) {
1652 schedstat_inc(rq->ttwu_local);
1653 schedstat_inc(p->se.statistics.nr_wakeups_local);
1654 } else {
1655 struct sched_domain *sd;
1656
1657 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1658 rcu_read_lock();
1659 for_each_domain(rq->cpu, sd) {
1660 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1661 schedstat_inc(sd->ttwu_wake_remote);
1662 break;
1663 }
1664 }
1665 rcu_read_unlock();
1666 }
1667
1668 if (wake_flags & WF_MIGRATED)
1669 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1670#endif /* CONFIG_SMP */
1671
1672 schedstat_inc(rq->ttwu_count);
1673 schedstat_inc(p->se.statistics.nr_wakeups);
1674
1675 if (wake_flags & WF_SYNC)
1676 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1677}
1678
1679static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1680{
1681 activate_task(rq, p, en_flags);
1682 p->on_rq = TASK_ON_RQ_QUEUED;
1683
1684 /* if a worker is waking up, notify workqueue */
1685 if (p->flags & PF_WQ_WORKER)
1686 wq_worker_waking_up(p, cpu_of(rq));
1687}
1688
1689/*
1690 * Mark the task runnable and perform wakeup-preemption.
1691 */
1692static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1693 struct pin_cookie cookie)
1694{
1695 check_preempt_curr(rq, p, wake_flags);
1696 p->state = TASK_RUNNING;
1697 trace_sched_wakeup(p);
1698
1699#ifdef CONFIG_SMP
1700 if (p->sched_class->task_woken) {
1701 /*
1702 * Our task @p is fully woken up and running; so its safe to
1703 * drop the rq->lock, hereafter rq is only used for statistics.
1704 */
1705 lockdep_unpin_lock(&rq->lock, cookie);
1706 p->sched_class->task_woken(rq, p);
1707 lockdep_repin_lock(&rq->lock, cookie);
1708 }
1709
1710 if (rq->idle_stamp) {
1711 u64 delta = rq_clock(rq) - rq->idle_stamp;
1712 u64 max = 2*rq->max_idle_balance_cost;
1713
1714 update_avg(&rq->avg_idle, delta);
1715
1716 if (rq->avg_idle > max)
1717 rq->avg_idle = max;
1718
1719 rq->idle_stamp = 0;
1720 }
1721#endif
1722}
1723
1724static void
1725ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1726 struct pin_cookie cookie)
1727{
1728 int en_flags = ENQUEUE_WAKEUP;
1729
1730 lockdep_assert_held(&rq->lock);
1731
1732#ifdef CONFIG_SMP
1733 if (p->sched_contributes_to_load)
1734 rq->nr_uninterruptible--;
1735
1736 if (wake_flags & WF_MIGRATED)
1737 en_flags |= ENQUEUE_MIGRATED;
1738#endif
1739
1740 ttwu_activate(rq, p, en_flags);
1741 ttwu_do_wakeup(rq, p, wake_flags, cookie);
1742}
1743
1744/*
1745 * Called in case the task @p isn't fully descheduled from its runqueue,
1746 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1747 * since all we need to do is flip p->state to TASK_RUNNING, since
1748 * the task is still ->on_rq.
1749 */
1750static int ttwu_remote(struct task_struct *p, int wake_flags)
1751{
1752 struct rq_flags rf;
1753 struct rq *rq;
1754 int ret = 0;
1755
1756 rq = __task_rq_lock(p, &rf);
1757 if (task_on_rq_queued(p)) {
1758 /* check_preempt_curr() may use rq clock */
1759 update_rq_clock(rq);
1760 ttwu_do_wakeup(rq, p, wake_flags, rf.cookie);
1761 ret = 1;
1762 }
1763 __task_rq_unlock(rq, &rf);
1764
1765 return ret;
1766}
1767
1768#ifdef CONFIG_SMP
1769void sched_ttwu_pending(void)
1770{
1771 struct rq *rq = this_rq();
1772 struct llist_node *llist = llist_del_all(&rq->wake_list);
1773 struct pin_cookie cookie;
1774 struct task_struct *p;
1775 unsigned long flags;
1776
1777 if (!llist)
1778 return;
1779
1780 raw_spin_lock_irqsave(&rq->lock, flags);
1781 cookie = lockdep_pin_lock(&rq->lock);
1782
1783 while (llist) {
1784 int wake_flags = 0;
1785
1786 p = llist_entry(llist, struct task_struct, wake_entry);
1787 llist = llist_next(llist);
1788
1789 if (p->sched_remote_wakeup)
1790 wake_flags = WF_MIGRATED;
1791
1792 ttwu_do_activate(rq, p, wake_flags, cookie);
1793 }
1794
1795 lockdep_unpin_lock(&rq->lock, cookie);
1796 raw_spin_unlock_irqrestore(&rq->lock, flags);
1797}
1798
1799void scheduler_ipi(void)
1800{
1801 /*
1802 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1803 * TIF_NEED_RESCHED remotely (for the first time) will also send
1804 * this IPI.
1805 */
1806 preempt_fold_need_resched();
1807
1808 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1809 return;
1810
1811 /*
1812 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1813 * traditionally all their work was done from the interrupt return
1814 * path. Now that we actually do some work, we need to make sure
1815 * we do call them.
1816 *
1817 * Some archs already do call them, luckily irq_enter/exit nest
1818 * properly.
1819 *
1820 * Arguably we should visit all archs and update all handlers,
1821 * however a fair share of IPIs are still resched only so this would
1822 * somewhat pessimize the simple resched case.
1823 */
1824 irq_enter();
1825 sched_ttwu_pending();
1826
1827 /*
1828 * Check if someone kicked us for doing the nohz idle load balance.
1829 */
1830 if (unlikely(got_nohz_idle_kick())) {
1831 this_rq()->idle_balance = 1;
1832 raise_softirq_irqoff(SCHED_SOFTIRQ);
1833 }
1834 irq_exit();
1835}
1836
1837static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1838{
1839 struct rq *rq = cpu_rq(cpu);
1840
1841 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1842
1843 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1844 if (!set_nr_if_polling(rq->idle))
1845 smp_send_reschedule(cpu);
1846 else
1847 trace_sched_wake_idle_without_ipi(cpu);
1848 }
1849}
1850
1851void wake_up_if_idle(int cpu)
1852{
1853 struct rq *rq = cpu_rq(cpu);
1854 unsigned long flags;
1855
1856 rcu_read_lock();
1857
1858 if (!is_idle_task(rcu_dereference(rq->curr)))
1859 goto out;
1860
1861 if (set_nr_if_polling(rq->idle)) {
1862 trace_sched_wake_idle_without_ipi(cpu);
1863 } else {
1864 raw_spin_lock_irqsave(&rq->lock, flags);
1865 if (is_idle_task(rq->curr))
1866 smp_send_reschedule(cpu);
1867 /* Else cpu is not in idle, do nothing here */
1868 raw_spin_unlock_irqrestore(&rq->lock, flags);
1869 }
1870
1871out:
1872 rcu_read_unlock();
1873}
1874
1875bool cpus_share_cache(int this_cpu, int that_cpu)
1876{
1877 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1878}
1879#endif /* CONFIG_SMP */
1880
1881static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1882{
1883 struct rq *rq = cpu_rq(cpu);
1884 struct pin_cookie cookie;
1885
1886#if defined(CONFIG_SMP)
1887 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1888 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1889 ttwu_queue_remote(p, cpu, wake_flags);
1890 return;
1891 }
1892#endif
1893
1894 raw_spin_lock(&rq->lock);
1895 cookie = lockdep_pin_lock(&rq->lock);
1896 ttwu_do_activate(rq, p, wake_flags, cookie);
1897 lockdep_unpin_lock(&rq->lock, cookie);
1898 raw_spin_unlock(&rq->lock);
1899}
1900
1901/*
1902 * Notes on Program-Order guarantees on SMP systems.
1903 *
1904 * MIGRATION
1905 *
1906 * The basic program-order guarantee on SMP systems is that when a task [t]
1907 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1908 * execution on its new cpu [c1].
1909 *
1910 * For migration (of runnable tasks) this is provided by the following means:
1911 *
1912 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1913 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1914 * rq(c1)->lock (if not at the same time, then in that order).
1915 * C) LOCK of the rq(c1)->lock scheduling in task
1916 *
1917 * Transitivity guarantees that B happens after A and C after B.
1918 * Note: we only require RCpc transitivity.
1919 * Note: the cpu doing B need not be c0 or c1
1920 *
1921 * Example:
1922 *
1923 * CPU0 CPU1 CPU2
1924 *
1925 * LOCK rq(0)->lock
1926 * sched-out X
1927 * sched-in Y
1928 * UNLOCK rq(0)->lock
1929 *
1930 * LOCK rq(0)->lock // orders against CPU0
1931 * dequeue X
1932 * UNLOCK rq(0)->lock
1933 *
1934 * LOCK rq(1)->lock
1935 * enqueue X
1936 * UNLOCK rq(1)->lock
1937 *
1938 * LOCK rq(1)->lock // orders against CPU2
1939 * sched-out Z
1940 * sched-in X
1941 * UNLOCK rq(1)->lock
1942 *
1943 *
1944 * BLOCKING -- aka. SLEEP + WAKEUP
1945 *
1946 * For blocking we (obviously) need to provide the same guarantee as for
1947 * migration. However the means are completely different as there is no lock
1948 * chain to provide order. Instead we do:
1949 *
1950 * 1) smp_store_release(X->on_cpu, 0)
1951 * 2) smp_cond_load_acquire(!X->on_cpu)
1952 *
1953 * Example:
1954 *
1955 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1956 *
1957 * LOCK rq(0)->lock LOCK X->pi_lock
1958 * dequeue X
1959 * sched-out X
1960 * smp_store_release(X->on_cpu, 0);
1961 *
1962 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1963 * X->state = WAKING
1964 * set_task_cpu(X,2)
1965 *
1966 * LOCK rq(2)->lock
1967 * enqueue X
1968 * X->state = RUNNING
1969 * UNLOCK rq(2)->lock
1970 *
1971 * LOCK rq(2)->lock // orders against CPU1
1972 * sched-out Z
1973 * sched-in X
1974 * UNLOCK rq(2)->lock
1975 *
1976 * UNLOCK X->pi_lock
1977 * UNLOCK rq(0)->lock
1978 *
1979 *
1980 * However; for wakeups there is a second guarantee we must provide, namely we
1981 * must observe the state that lead to our wakeup. That is, not only must our
1982 * task observe its own prior state, it must also observe the stores prior to
1983 * its wakeup.
1984 *
1985 * This means that any means of doing remote wakeups must order the CPU doing
1986 * the wakeup against the CPU the task is going to end up running on. This,
1987 * however, is already required for the regular Program-Order guarantee above,
1988 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1989 *
1990 */
1991
1992/**
1993 * try_to_wake_up - wake up a thread
1994 * @p: the thread to be awakened
1995 * @state: the mask of task states that can be woken
1996 * @wake_flags: wake modifier flags (WF_*)
1997 *
1998 * If (@state & @p->state) @p->state = TASK_RUNNING.
1999 *
2000 * If the task was not queued/runnable, also place it back on a runqueue.
2001 *
2002 * Atomic against schedule() which would dequeue a task, also see
2003 * set_current_state().
2004 *
2005 * Return: %true if @p->state changes (an actual wakeup was done),
2006 * %false otherwise.
2007 */
2008static int
2009try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2010{
2011 unsigned long flags;
2012 int cpu, success = 0;
2013
2014 /*
2015 * If we are going to wake up a thread waiting for CONDITION we
2016 * need to ensure that CONDITION=1 done by the caller can not be
2017 * reordered with p->state check below. This pairs with mb() in
2018 * set_current_state() the waiting thread does.
2019 */
2020 smp_mb__before_spinlock();
2021 raw_spin_lock_irqsave(&p->pi_lock, flags);
2022 if (!(p->state & state))
2023 goto out;
2024
2025 trace_sched_waking(p);
2026
2027 success = 1; /* we're going to change ->state */
2028 cpu = task_cpu(p);
2029
2030 /*
2031 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2032 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2033 * in smp_cond_load_acquire() below.
2034 *
2035 * sched_ttwu_pending() try_to_wake_up()
2036 * [S] p->on_rq = 1; [L] P->state
2037 * UNLOCK rq->lock -----.
2038 * \
2039 * +--- RMB
2040 * schedule() /
2041 * LOCK rq->lock -----'
2042 * UNLOCK rq->lock
2043 *
2044 * [task p]
2045 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2046 *
2047 * Pairs with the UNLOCK+LOCK on rq->lock from the
2048 * last wakeup of our task and the schedule that got our task
2049 * current.
2050 */
2051 smp_rmb();
2052 if (p->on_rq && ttwu_remote(p, wake_flags))
2053 goto stat;
2054
2055#ifdef CONFIG_SMP
2056 /*
2057 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2058 * possible to, falsely, observe p->on_cpu == 0.
2059 *
2060 * One must be running (->on_cpu == 1) in order to remove oneself
2061 * from the runqueue.
2062 *
2063 * [S] ->on_cpu = 1; [L] ->on_rq
2064 * UNLOCK rq->lock
2065 * RMB
2066 * LOCK rq->lock
2067 * [S] ->on_rq = 0; [L] ->on_cpu
2068 *
2069 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2070 * from the consecutive calls to schedule(); the first switching to our
2071 * task, the second putting it to sleep.
2072 */
2073 smp_rmb();
2074
2075 /*
2076 * If the owning (remote) cpu is still in the middle of schedule() with
2077 * this task as prev, wait until its done referencing the task.
2078 *
2079 * Pairs with the smp_store_release() in finish_lock_switch().
2080 *
2081 * This ensures that tasks getting woken will be fully ordered against
2082 * their previous state and preserve Program Order.
2083 */
2084 smp_cond_load_acquire(&p->on_cpu, !VAL);
2085
2086 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2087 p->state = TASK_WAKING;
2088
2089 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2090 if (task_cpu(p) != cpu) {
2091 wake_flags |= WF_MIGRATED;
2092 set_task_cpu(p, cpu);
2093 }
2094#endif /* CONFIG_SMP */
2095
2096 ttwu_queue(p, cpu, wake_flags);
2097stat:
2098 ttwu_stat(p, cpu, wake_flags);
2099out:
2100 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2101
2102 return success;
2103}
2104
2105/**
2106 * try_to_wake_up_local - try to wake up a local task with rq lock held
2107 * @p: the thread to be awakened
2108 * @cookie: context's cookie for pinning
2109 *
2110 * Put @p on the run-queue if it's not already there. The caller must
2111 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2112 * the current task.
2113 */
2114static void try_to_wake_up_local(struct task_struct *p, struct pin_cookie cookie)
2115{
2116 struct rq *rq = task_rq(p);
2117
2118 if (WARN_ON_ONCE(rq != this_rq()) ||
2119 WARN_ON_ONCE(p == current))
2120 return;
2121
2122 lockdep_assert_held(&rq->lock);
2123
2124 if (!raw_spin_trylock(&p->pi_lock)) {
2125 /*
2126 * This is OK, because current is on_cpu, which avoids it being
2127 * picked for load-balance and preemption/IRQs are still
2128 * disabled avoiding further scheduler activity on it and we've
2129 * not yet picked a replacement task.
2130 */
2131 lockdep_unpin_lock(&rq->lock, cookie);
2132 raw_spin_unlock(&rq->lock);
2133 raw_spin_lock(&p->pi_lock);
2134 raw_spin_lock(&rq->lock);
2135 lockdep_repin_lock(&rq->lock, cookie);
2136 }
2137
2138 if (!(p->state & TASK_NORMAL))
2139 goto out;
2140
2141 trace_sched_waking(p);
2142
2143 if (!task_on_rq_queued(p))
2144 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2145
2146 ttwu_do_wakeup(rq, p, 0, cookie);
2147 ttwu_stat(p, smp_processor_id(), 0);
2148out:
2149 raw_spin_unlock(&p->pi_lock);
2150}
2151
2152/**
2153 * wake_up_process - Wake up a specific process
2154 * @p: The process to be woken up.
2155 *
2156 * Attempt to wake up the nominated process and move it to the set of runnable
2157 * processes.
2158 *
2159 * Return: 1 if the process was woken up, 0 if it was already running.
2160 *
2161 * It may be assumed that this function implies a write memory barrier before
2162 * changing the task state if and only if any tasks are woken up.
2163 */
2164int wake_up_process(struct task_struct *p)
2165{
2166 return try_to_wake_up(p, TASK_NORMAL, 0);
2167}
2168EXPORT_SYMBOL(wake_up_process);
2169
2170int wake_up_state(struct task_struct *p, unsigned int state)
2171{
2172 return try_to_wake_up(p, state, 0);
2173}
2174
2175/*
2176 * This function clears the sched_dl_entity static params.
2177 */
2178void __dl_clear_params(struct task_struct *p)
2179{
2180 struct sched_dl_entity *dl_se = &p->dl;
2181
2182 dl_se->dl_runtime = 0;
2183 dl_se->dl_deadline = 0;
2184 dl_se->dl_period = 0;
2185 dl_se->flags = 0;
2186 dl_se->dl_bw = 0;
2187
2188 dl_se->dl_throttled = 0;
2189 dl_se->dl_yielded = 0;
2190}
2191
2192/*
2193 * Perform scheduler related setup for a newly forked process p.
2194 * p is forked by current.
2195 *
2196 * __sched_fork() is basic setup used by init_idle() too:
2197 */
2198static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2199{
2200 p->on_rq = 0;
2201
2202 p->se.on_rq = 0;
2203 p->se.exec_start = 0;
2204 p->se.sum_exec_runtime = 0;
2205 p->se.prev_sum_exec_runtime = 0;
2206 p->se.nr_migrations = 0;
2207 p->se.vruntime = 0;
2208 INIT_LIST_HEAD(&p->se.group_node);
2209
2210#ifdef CONFIG_FAIR_GROUP_SCHED
2211 p->se.cfs_rq = NULL;
2212#endif
2213
2214#ifdef CONFIG_SCHEDSTATS
2215 /* Even if schedstat is disabled, there should not be garbage */
2216 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2217#endif
2218
2219 RB_CLEAR_NODE(&p->dl.rb_node);
2220 init_dl_task_timer(&p->dl);
2221 __dl_clear_params(p);
2222
2223 INIT_LIST_HEAD(&p->rt.run_list);
2224 p->rt.timeout = 0;
2225 p->rt.time_slice = sched_rr_timeslice;
2226 p->rt.on_rq = 0;
2227 p->rt.on_list = 0;
2228
2229#ifdef CONFIG_PREEMPT_NOTIFIERS
2230 INIT_HLIST_HEAD(&p->preempt_notifiers);
2231#endif
2232
2233#ifdef CONFIG_NUMA_BALANCING
2234 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2235 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2236 p->mm->numa_scan_seq = 0;
2237 }
2238
2239 if (clone_flags & CLONE_VM)
2240 p->numa_preferred_nid = current->numa_preferred_nid;
2241 else
2242 p->numa_preferred_nid = -1;
2243
2244 p->node_stamp = 0ULL;
2245 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2246 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2247 p->numa_work.next = &p->numa_work;
2248 p->numa_faults = NULL;
2249 p->last_task_numa_placement = 0;
2250 p->last_sum_exec_runtime = 0;
2251
2252 p->numa_group = NULL;
2253#endif /* CONFIG_NUMA_BALANCING */
2254}
2255
2256DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2257
2258#ifdef CONFIG_NUMA_BALANCING
2259
2260void set_numabalancing_state(bool enabled)
2261{
2262 if (enabled)
2263 static_branch_enable(&sched_numa_balancing);
2264 else
2265 static_branch_disable(&sched_numa_balancing);
2266}
2267
2268#ifdef CONFIG_PROC_SYSCTL
2269int sysctl_numa_balancing(struct ctl_table *table, int write,
2270 void __user *buffer, size_t *lenp, loff_t *ppos)
2271{
2272 struct ctl_table t;
2273 int err;
2274 int state = static_branch_likely(&sched_numa_balancing);
2275
2276 if (write && !capable(CAP_SYS_ADMIN))
2277 return -EPERM;
2278
2279 t = *table;
2280 t.data = &state;
2281 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2282 if (err < 0)
2283 return err;
2284 if (write)
2285 set_numabalancing_state(state);
2286 return err;
2287}
2288#endif
2289#endif
2290
2291#ifdef CONFIG_SCHEDSTATS
2292
2293DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2294static bool __initdata __sched_schedstats = false;
2295
2296static void set_schedstats(bool enabled)
2297{
2298 if (enabled)
2299 static_branch_enable(&sched_schedstats);
2300 else
2301 static_branch_disable(&sched_schedstats);
2302}
2303
2304void force_schedstat_enabled(void)
2305{
2306 if (!schedstat_enabled()) {
2307 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2308 static_branch_enable(&sched_schedstats);
2309 }
2310}
2311
2312static int __init setup_schedstats(char *str)
2313{
2314 int ret = 0;
2315 if (!str)
2316 goto out;
2317
2318 /*
2319 * This code is called before jump labels have been set up, so we can't
2320 * change the static branch directly just yet. Instead set a temporary
2321 * variable so init_schedstats() can do it later.
2322 */
2323 if (!strcmp(str, "enable")) {
2324 __sched_schedstats = true;
2325 ret = 1;
2326 } else if (!strcmp(str, "disable")) {
2327 __sched_schedstats = false;
2328 ret = 1;
2329 }
2330out:
2331 if (!ret)
2332 pr_warn("Unable to parse schedstats=\n");
2333
2334 return ret;
2335}
2336__setup("schedstats=", setup_schedstats);
2337
2338static void __init init_schedstats(void)
2339{
2340 set_schedstats(__sched_schedstats);
2341}
2342
2343#ifdef CONFIG_PROC_SYSCTL
2344int sysctl_schedstats(struct ctl_table *table, int write,
2345 void __user *buffer, size_t *lenp, loff_t *ppos)
2346{
2347 struct ctl_table t;
2348 int err;
2349 int state = static_branch_likely(&sched_schedstats);
2350
2351 if (write && !capable(CAP_SYS_ADMIN))
2352 return -EPERM;
2353
2354 t = *table;
2355 t.data = &state;
2356 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2357 if (err < 0)
2358 return err;
2359 if (write)
2360 set_schedstats(state);
2361 return err;
2362}
2363#endif /* CONFIG_PROC_SYSCTL */
2364#else /* !CONFIG_SCHEDSTATS */
2365static inline void init_schedstats(void) {}
2366#endif /* CONFIG_SCHEDSTATS */
2367
2368/*
2369 * fork()/clone()-time setup:
2370 */
2371int sched_fork(unsigned long clone_flags, struct task_struct *p)
2372{
2373 unsigned long flags;
2374 int cpu = get_cpu();
2375
2376 __sched_fork(clone_flags, p);
2377 /*
2378 * We mark the process as NEW here. This guarantees that
2379 * nobody will actually run it, and a signal or other external
2380 * event cannot wake it up and insert it on the runqueue either.
2381 */
2382 p->state = TASK_NEW;
2383
2384 /*
2385 * Make sure we do not leak PI boosting priority to the child.
2386 */
2387 p->prio = current->normal_prio;
2388
2389 /*
2390 * Revert to default priority/policy on fork if requested.
2391 */
2392 if (unlikely(p->sched_reset_on_fork)) {
2393 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2394 p->policy = SCHED_NORMAL;
2395 p->static_prio = NICE_TO_PRIO(0);
2396 p->rt_priority = 0;
2397 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2398 p->static_prio = NICE_TO_PRIO(0);
2399
2400 p->prio = p->normal_prio = __normal_prio(p);
2401 set_load_weight(p);
2402
2403 /*
2404 * We don't need the reset flag anymore after the fork. It has
2405 * fulfilled its duty:
2406 */
2407 p->sched_reset_on_fork = 0;
2408 }
2409
2410 if (dl_prio(p->prio)) {
2411 put_cpu();
2412 return -EAGAIN;
2413 } else if (rt_prio(p->prio)) {
2414 p->sched_class = &rt_sched_class;
2415 } else {
2416 p->sched_class = &fair_sched_class;
2417 }
2418
2419 init_entity_runnable_average(&p->se);
2420
2421 /*
2422 * The child is not yet in the pid-hash so no cgroup attach races,
2423 * and the cgroup is pinned to this child due to cgroup_fork()
2424 * is ran before sched_fork().
2425 *
2426 * Silence PROVE_RCU.
2427 */
2428 raw_spin_lock_irqsave(&p->pi_lock, flags);
2429 /*
2430 * We're setting the cpu for the first time, we don't migrate,
2431 * so use __set_task_cpu().
2432 */
2433 __set_task_cpu(p, cpu);
2434 if (p->sched_class->task_fork)
2435 p->sched_class->task_fork(p);
2436 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2437
2438#ifdef CONFIG_SCHED_INFO
2439 if (likely(sched_info_on()))
2440 memset(&p->sched_info, 0, sizeof(p->sched_info));
2441#endif
2442#if defined(CONFIG_SMP)
2443 p->on_cpu = 0;
2444#endif
2445 init_task_preempt_count(p);
2446#ifdef CONFIG_SMP
2447 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2448 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2449#endif
2450
2451 put_cpu();
2452 return 0;
2453}
2454
2455unsigned long to_ratio(u64 period, u64 runtime)
2456{
2457 if (runtime == RUNTIME_INF)
2458 return 1ULL << 20;
2459
2460 /*
2461 * Doing this here saves a lot of checks in all
2462 * the calling paths, and returning zero seems
2463 * safe for them anyway.
2464 */
2465 if (period == 0)
2466 return 0;
2467
2468 return div64_u64(runtime << 20, period);
2469}
2470
2471#ifdef CONFIG_SMP
2472inline struct dl_bw *dl_bw_of(int i)
2473{
2474 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2475 "sched RCU must be held");
2476 return &cpu_rq(i)->rd->dl_bw;
2477}
2478
2479static inline int dl_bw_cpus(int i)
2480{
2481 struct root_domain *rd = cpu_rq(i)->rd;
2482 int cpus = 0;
2483
2484 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2485 "sched RCU must be held");
2486 for_each_cpu_and(i, rd->span, cpu_active_mask)
2487 cpus++;
2488
2489 return cpus;
2490}
2491#else
2492inline struct dl_bw *dl_bw_of(int i)
2493{
2494 return &cpu_rq(i)->dl.dl_bw;
2495}
2496
2497static inline int dl_bw_cpus(int i)
2498{
2499 return 1;
2500}
2501#endif
2502
2503/*
2504 * We must be sure that accepting a new task (or allowing changing the
2505 * parameters of an existing one) is consistent with the bandwidth
2506 * constraints. If yes, this function also accordingly updates the currently
2507 * allocated bandwidth to reflect the new situation.
2508 *
2509 * This function is called while holding p's rq->lock.
2510 *
2511 * XXX we should delay bw change until the task's 0-lag point, see
2512 * __setparam_dl().
2513 */
2514static int dl_overflow(struct task_struct *p, int policy,
2515 const struct sched_attr *attr)
2516{
2517
2518 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2519 u64 period = attr->sched_period ?: attr->sched_deadline;
2520 u64 runtime = attr->sched_runtime;
2521 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2522 int cpus, err = -1;
2523
2524 /* !deadline task may carry old deadline bandwidth */
2525 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2526 return 0;
2527
2528 /*
2529 * Either if a task, enters, leave, or stays -deadline but changes
2530 * its parameters, we may need to update accordingly the total
2531 * allocated bandwidth of the container.
2532 */
2533 raw_spin_lock(&dl_b->lock);
2534 cpus = dl_bw_cpus(task_cpu(p));
2535 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2536 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2537 __dl_add(dl_b, new_bw);
2538 err = 0;
2539 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2540 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2541 __dl_clear(dl_b, p->dl.dl_bw);
2542 __dl_add(dl_b, new_bw);
2543 err = 0;
2544 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2545 __dl_clear(dl_b, p->dl.dl_bw);
2546 err = 0;
2547 }
2548 raw_spin_unlock(&dl_b->lock);
2549
2550 return err;
2551}
2552
2553extern void init_dl_bw(struct dl_bw *dl_b);
2554
2555/*
2556 * wake_up_new_task - wake up a newly created task for the first time.
2557 *
2558 * This function will do some initial scheduler statistics housekeeping
2559 * that must be done for every newly created context, then puts the task
2560 * on the runqueue and wakes it.
2561 */
2562void wake_up_new_task(struct task_struct *p)
2563{
2564 struct rq_flags rf;
2565 struct rq *rq;
2566
2567 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2568 p->state = TASK_RUNNING;
2569#ifdef CONFIG_SMP
2570 /*
2571 * Fork balancing, do it here and not earlier because:
2572 * - cpus_allowed can change in the fork path
2573 * - any previously selected cpu might disappear through hotplug
2574 *
2575 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2576 * as we're not fully set-up yet.
2577 */
2578 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2579#endif
2580 rq = __task_rq_lock(p, &rf);
2581 post_init_entity_util_avg(&p->se);
2582
2583 activate_task(rq, p, 0);
2584 p->on_rq = TASK_ON_RQ_QUEUED;
2585 trace_sched_wakeup_new(p);
2586 check_preempt_curr(rq, p, WF_FORK);
2587#ifdef CONFIG_SMP
2588 if (p->sched_class->task_woken) {
2589 /*
2590 * Nothing relies on rq->lock after this, so its fine to
2591 * drop it.
2592 */
2593 lockdep_unpin_lock(&rq->lock, rf.cookie);
2594 p->sched_class->task_woken(rq, p);
2595 lockdep_repin_lock(&rq->lock, rf.cookie);
2596 }
2597#endif
2598 task_rq_unlock(rq, p, &rf);
2599}
2600
2601#ifdef CONFIG_PREEMPT_NOTIFIERS
2602
2603static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2604
2605void preempt_notifier_inc(void)
2606{
2607 static_key_slow_inc(&preempt_notifier_key);
2608}
2609EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2610
2611void preempt_notifier_dec(void)
2612{
2613 static_key_slow_dec(&preempt_notifier_key);
2614}
2615EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2616
2617/**
2618 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2619 * @notifier: notifier struct to register
2620 */
2621void preempt_notifier_register(struct preempt_notifier *notifier)
2622{
2623 if (!static_key_false(&preempt_notifier_key))
2624 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2625
2626 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2627}
2628EXPORT_SYMBOL_GPL(preempt_notifier_register);
2629
2630/**
2631 * preempt_notifier_unregister - no longer interested in preemption notifications
2632 * @notifier: notifier struct to unregister
2633 *
2634 * This is *not* safe to call from within a preemption notifier.
2635 */
2636void preempt_notifier_unregister(struct preempt_notifier *notifier)
2637{
2638 hlist_del(¬ifier->link);
2639}
2640EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2641
2642static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2643{
2644 struct preempt_notifier *notifier;
2645
2646 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2647 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2648}
2649
2650static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2651{
2652 if (static_key_false(&preempt_notifier_key))
2653 __fire_sched_in_preempt_notifiers(curr);
2654}
2655
2656static void
2657__fire_sched_out_preempt_notifiers(struct task_struct *curr,
2658 struct task_struct *next)
2659{
2660 struct preempt_notifier *notifier;
2661
2662 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2663 notifier->ops->sched_out(notifier, next);
2664}
2665
2666static __always_inline void
2667fire_sched_out_preempt_notifiers(struct task_struct *curr,
2668 struct task_struct *next)
2669{
2670 if (static_key_false(&preempt_notifier_key))
2671 __fire_sched_out_preempt_notifiers(curr, next);
2672}
2673
2674#else /* !CONFIG_PREEMPT_NOTIFIERS */
2675
2676static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2677{
2678}
2679
2680static inline void
2681fire_sched_out_preempt_notifiers(struct task_struct *curr,
2682 struct task_struct *next)
2683{
2684}
2685
2686#endif /* CONFIG_PREEMPT_NOTIFIERS */
2687
2688/**
2689 * prepare_task_switch - prepare to switch tasks
2690 * @rq: the runqueue preparing to switch
2691 * @prev: the current task that is being switched out
2692 * @next: the task we are going to switch to.
2693 *
2694 * This is called with the rq lock held and interrupts off. It must
2695 * be paired with a subsequent finish_task_switch after the context
2696 * switch.
2697 *
2698 * prepare_task_switch sets up locking and calls architecture specific
2699 * hooks.
2700 */
2701static inline void
2702prepare_task_switch(struct rq *rq, struct task_struct *prev,
2703 struct task_struct *next)
2704{
2705 sched_info_switch(rq, prev, next);
2706 perf_event_task_sched_out(prev, next);
2707 fire_sched_out_preempt_notifiers(prev, next);
2708 prepare_lock_switch(rq, next);
2709 prepare_arch_switch(next);
2710}
2711
2712/**
2713 * finish_task_switch - clean up after a task-switch
2714 * @prev: the thread we just switched away from.
2715 *
2716 * finish_task_switch must be called after the context switch, paired
2717 * with a prepare_task_switch call before the context switch.
2718 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2719 * and do any other architecture-specific cleanup actions.
2720 *
2721 * Note that we may have delayed dropping an mm in context_switch(). If
2722 * so, we finish that here outside of the runqueue lock. (Doing it
2723 * with the lock held can cause deadlocks; see schedule() for
2724 * details.)
2725 *
2726 * The context switch have flipped the stack from under us and restored the
2727 * local variables which were saved when this task called schedule() in the
2728 * past. prev == current is still correct but we need to recalculate this_rq
2729 * because prev may have moved to another CPU.
2730 */
2731static struct rq *finish_task_switch(struct task_struct *prev)
2732 __releases(rq->lock)
2733{
2734 struct rq *rq = this_rq();
2735 struct mm_struct *mm = rq->prev_mm;
2736 long prev_state;
2737
2738 /*
2739 * The previous task will have left us with a preempt_count of 2
2740 * because it left us after:
2741 *
2742 * schedule()
2743 * preempt_disable(); // 1
2744 * __schedule()
2745 * raw_spin_lock_irq(&rq->lock) // 2
2746 *
2747 * Also, see FORK_PREEMPT_COUNT.
2748 */
2749 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2750 "corrupted preempt_count: %s/%d/0x%x\n",
2751 current->comm, current->pid, preempt_count()))
2752 preempt_count_set(FORK_PREEMPT_COUNT);
2753
2754 rq->prev_mm = NULL;
2755
2756 /*
2757 * A task struct has one reference for the use as "current".
2758 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2759 * schedule one last time. The schedule call will never return, and
2760 * the scheduled task must drop that reference.
2761 *
2762 * We must observe prev->state before clearing prev->on_cpu (in
2763 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2764 * running on another CPU and we could rave with its RUNNING -> DEAD
2765 * transition, resulting in a double drop.
2766 */
2767 prev_state = prev->state;
2768 vtime_task_switch(prev);
2769 perf_event_task_sched_in(prev, current);
2770 finish_lock_switch(rq, prev);
2771 finish_arch_post_lock_switch();
2772
2773 fire_sched_in_preempt_notifiers(current);
2774 if (mm)
2775 mmdrop(mm);
2776 if (unlikely(prev_state == TASK_DEAD)) {
2777 if (prev->sched_class->task_dead)
2778 prev->sched_class->task_dead(prev);
2779
2780 /*
2781 * Remove function-return probe instances associated with this
2782 * task and put them back on the free list.
2783 */
2784 kprobe_flush_task(prev);
2785
2786 /* Task is done with its stack. */
2787 put_task_stack(prev);
2788
2789 put_task_struct(prev);
2790 }
2791
2792 tick_nohz_task_switch();
2793 return rq;
2794}
2795
2796#ifdef CONFIG_SMP
2797
2798/* rq->lock is NOT held, but preemption is disabled */
2799static void __balance_callback(struct rq *rq)
2800{
2801 struct callback_head *head, *next;
2802 void (*func)(struct rq *rq);
2803 unsigned long flags;
2804
2805 raw_spin_lock_irqsave(&rq->lock, flags);
2806 head = rq->balance_callback;
2807 rq->balance_callback = NULL;
2808 while (head) {
2809 func = (void (*)(struct rq *))head->func;
2810 next = head->next;
2811 head->next = NULL;
2812 head = next;
2813
2814 func(rq);
2815 }
2816 raw_spin_unlock_irqrestore(&rq->lock, flags);
2817}
2818
2819static inline void balance_callback(struct rq *rq)
2820{
2821 if (unlikely(rq->balance_callback))
2822 __balance_callback(rq);
2823}
2824
2825#else
2826
2827static inline void balance_callback(struct rq *rq)
2828{
2829}
2830
2831#endif
2832
2833/**
2834 * schedule_tail - first thing a freshly forked thread must call.
2835 * @prev: the thread we just switched away from.
2836 */
2837asmlinkage __visible void schedule_tail(struct task_struct *prev)
2838 __releases(rq->lock)
2839{
2840 struct rq *rq;
2841
2842 /*
2843 * New tasks start with FORK_PREEMPT_COUNT, see there and
2844 * finish_task_switch() for details.
2845 *
2846 * finish_task_switch() will drop rq->lock() and lower preempt_count
2847 * and the preempt_enable() will end up enabling preemption (on
2848 * PREEMPT_COUNT kernels).
2849 */
2850
2851 rq = finish_task_switch(prev);
2852 balance_callback(rq);
2853 preempt_enable();
2854
2855 if (current->set_child_tid)
2856 put_user(task_pid_vnr(current), current->set_child_tid);
2857}
2858
2859/*
2860 * context_switch - switch to the new MM and the new thread's register state.
2861 */
2862static __always_inline struct rq *
2863context_switch(struct rq *rq, struct task_struct *prev,
2864 struct task_struct *next, struct pin_cookie cookie)
2865{
2866 struct mm_struct *mm, *oldmm;
2867
2868 prepare_task_switch(rq, prev, next);
2869
2870 mm = next->mm;
2871 oldmm = prev->active_mm;
2872 /*
2873 * For paravirt, this is coupled with an exit in switch_to to
2874 * combine the page table reload and the switch backend into
2875 * one hypercall.
2876 */
2877 arch_start_context_switch(prev);
2878
2879 if (!mm) {
2880 next->active_mm = oldmm;
2881 atomic_inc(&oldmm->mm_count);
2882 enter_lazy_tlb(oldmm, next);
2883 } else
2884 switch_mm_irqs_off(oldmm, mm, next);
2885
2886 if (!prev->mm) {
2887 prev->active_mm = NULL;
2888 rq->prev_mm = oldmm;
2889 }
2890 /*
2891 * Since the runqueue lock will be released by the next
2892 * task (which is an invalid locking op but in the case
2893 * of the scheduler it's an obvious special-case), so we
2894 * do an early lockdep release here:
2895 */
2896 lockdep_unpin_lock(&rq->lock, cookie);
2897 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2898
2899 /* Here we just switch the register state and the stack. */
2900 switch_to(prev, next, prev);
2901 barrier();
2902
2903 return finish_task_switch(prev);
2904}
2905
2906/*
2907 * nr_running and nr_context_switches:
2908 *
2909 * externally visible scheduler statistics: current number of runnable
2910 * threads, total number of context switches performed since bootup.
2911 */
2912unsigned long nr_running(void)
2913{
2914 unsigned long i, sum = 0;
2915
2916 for_each_online_cpu(i)
2917 sum += cpu_rq(i)->nr_running;
2918
2919 return sum;
2920}
2921
2922/*
2923 * Check if only the current task is running on the cpu.
2924 *
2925 * Caution: this function does not check that the caller has disabled
2926 * preemption, thus the result might have a time-of-check-to-time-of-use
2927 * race. The caller is responsible to use it correctly, for example:
2928 *
2929 * - from a non-preemptable section (of course)
2930 *
2931 * - from a thread that is bound to a single CPU
2932 *
2933 * - in a loop with very short iterations (e.g. a polling loop)
2934 */
2935bool single_task_running(void)
2936{
2937 return raw_rq()->nr_running == 1;
2938}
2939EXPORT_SYMBOL(single_task_running);
2940
2941unsigned long long nr_context_switches(void)
2942{
2943 int i;
2944 unsigned long long sum = 0;
2945
2946 for_each_possible_cpu(i)
2947 sum += cpu_rq(i)->nr_switches;
2948
2949 return sum;
2950}
2951
2952unsigned long nr_iowait(void)
2953{
2954 unsigned long i, sum = 0;
2955
2956 for_each_possible_cpu(i)
2957 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2958
2959 return sum;
2960}
2961
2962unsigned long nr_iowait_cpu(int cpu)
2963{
2964 struct rq *this = cpu_rq(cpu);
2965 return atomic_read(&this->nr_iowait);
2966}
2967
2968void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2969{
2970 struct rq *rq = this_rq();
2971 *nr_waiters = atomic_read(&rq->nr_iowait);
2972 *load = rq->load.weight;
2973}
2974
2975#ifdef CONFIG_SMP
2976
2977/*
2978 * sched_exec - execve() is a valuable balancing opportunity, because at
2979 * this point the task has the smallest effective memory and cache footprint.
2980 */
2981void sched_exec(void)
2982{
2983 struct task_struct *p = current;
2984 unsigned long flags;
2985 int dest_cpu;
2986
2987 raw_spin_lock_irqsave(&p->pi_lock, flags);
2988 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2989 if (dest_cpu == smp_processor_id())
2990 goto unlock;
2991
2992 if (likely(cpu_active(dest_cpu))) {
2993 struct migration_arg arg = { p, dest_cpu };
2994
2995 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2996 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2997 return;
2998 }
2999unlock:
3000 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3001}
3002
3003#endif
3004
3005DEFINE_PER_CPU(struct kernel_stat, kstat);
3006DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3007
3008EXPORT_PER_CPU_SYMBOL(kstat);
3009EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3010
3011/*
3012 * The function fair_sched_class.update_curr accesses the struct curr
3013 * and its field curr->exec_start; when called from task_sched_runtime(),
3014 * we observe a high rate of cache misses in practice.
3015 * Prefetching this data results in improved performance.
3016 */
3017static inline void prefetch_curr_exec_start(struct task_struct *p)
3018{
3019#ifdef CONFIG_FAIR_GROUP_SCHED
3020 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3021#else
3022 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3023#endif
3024 prefetch(curr);
3025 prefetch(&curr->exec_start);
3026}
3027
3028/*
3029 * Return accounted runtime for the task.
3030 * In case the task is currently running, return the runtime plus current's
3031 * pending runtime that have not been accounted yet.
3032 */
3033unsigned long long task_sched_runtime(struct task_struct *p)
3034{
3035 struct rq_flags rf;
3036 struct rq *rq;
3037 u64 ns;
3038
3039#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3040 /*
3041 * 64-bit doesn't need locks to atomically read a 64bit value.
3042 * So we have a optimization chance when the task's delta_exec is 0.
3043 * Reading ->on_cpu is racy, but this is ok.
3044 *
3045 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3046 * If we race with it entering cpu, unaccounted time is 0. This is
3047 * indistinguishable from the read occurring a few cycles earlier.
3048 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3049 * been accounted, so we're correct here as well.
3050 */
3051 if (!p->on_cpu || !task_on_rq_queued(p))
3052 return p->se.sum_exec_runtime;
3053#endif
3054
3055 rq = task_rq_lock(p, &rf);
3056 /*
3057 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3058 * project cycles that may never be accounted to this
3059 * thread, breaking clock_gettime().
3060 */
3061 if (task_current(rq, p) && task_on_rq_queued(p)) {
3062 prefetch_curr_exec_start(p);
3063 update_rq_clock(rq);
3064 p->sched_class->update_curr(rq);
3065 }
3066 ns = p->se.sum_exec_runtime;
3067 task_rq_unlock(rq, p, &rf);
3068
3069 return ns;
3070}
3071
3072/*
3073 * This function gets called by the timer code, with HZ frequency.
3074 * We call it with interrupts disabled.
3075 */
3076void scheduler_tick(void)
3077{
3078 int cpu = smp_processor_id();
3079 struct rq *rq = cpu_rq(cpu);
3080 struct task_struct *curr = rq->curr;
3081
3082 sched_clock_tick();
3083
3084 raw_spin_lock(&rq->lock);
3085 update_rq_clock(rq);
3086 curr->sched_class->task_tick(rq, curr, 0);
3087 cpu_load_update_active(rq);
3088 calc_global_load_tick(rq);
3089 raw_spin_unlock(&rq->lock);
3090
3091 perf_event_task_tick();
3092
3093#ifdef CONFIG_SMP
3094 rq->idle_balance = idle_cpu(cpu);
3095 trigger_load_balance(rq);
3096#endif
3097 rq_last_tick_reset(rq);
3098}
3099
3100#ifdef CONFIG_NO_HZ_FULL
3101/**
3102 * scheduler_tick_max_deferment
3103 *
3104 * Keep at least one tick per second when a single
3105 * active task is running because the scheduler doesn't
3106 * yet completely support full dynticks environment.
3107 *
3108 * This makes sure that uptime, CFS vruntime, load
3109 * balancing, etc... continue to move forward, even
3110 * with a very low granularity.
3111 *
3112 * Return: Maximum deferment in nanoseconds.
3113 */
3114u64 scheduler_tick_max_deferment(void)
3115{
3116 struct rq *rq = this_rq();
3117 unsigned long next, now = READ_ONCE(jiffies);
3118
3119 next = rq->last_sched_tick + HZ;
3120
3121 if (time_before_eq(next, now))
3122 return 0;
3123
3124 return jiffies_to_nsecs(next - now);
3125}
3126#endif
3127
3128#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3129 defined(CONFIG_PREEMPT_TRACER))
3130/*
3131 * If the value passed in is equal to the current preempt count
3132 * then we just disabled preemption. Start timing the latency.
3133 */
3134static inline void preempt_latency_start(int val)
3135{
3136 if (preempt_count() == val) {
3137 unsigned long ip = get_lock_parent_ip();
3138#ifdef CONFIG_DEBUG_PREEMPT
3139 current->preempt_disable_ip = ip;
3140#endif
3141 trace_preempt_off(CALLER_ADDR0, ip);
3142 }
3143}
3144
3145void preempt_count_add(int val)
3146{
3147#ifdef CONFIG_DEBUG_PREEMPT
3148 /*
3149 * Underflow?
3150 */
3151 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3152 return;
3153#endif
3154 __preempt_count_add(val);
3155#ifdef CONFIG_DEBUG_PREEMPT
3156 /*
3157 * Spinlock count overflowing soon?
3158 */
3159 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3160 PREEMPT_MASK - 10);
3161#endif
3162 preempt_latency_start(val);
3163}
3164EXPORT_SYMBOL(preempt_count_add);
3165NOKPROBE_SYMBOL(preempt_count_add);
3166
3167/*
3168 * If the value passed in equals to the current preempt count
3169 * then we just enabled preemption. Stop timing the latency.
3170 */
3171static inline void preempt_latency_stop(int val)
3172{
3173 if (preempt_count() == val)
3174 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3175}
3176
3177void preempt_count_sub(int val)
3178{
3179#ifdef CONFIG_DEBUG_PREEMPT
3180 /*
3181 * Underflow?
3182 */
3183 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3184 return;
3185 /*
3186 * Is the spinlock portion underflowing?
3187 */
3188 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3189 !(preempt_count() & PREEMPT_MASK)))
3190 return;
3191#endif
3192
3193 preempt_latency_stop(val);
3194 __preempt_count_sub(val);
3195}
3196EXPORT_SYMBOL(preempt_count_sub);
3197NOKPROBE_SYMBOL(preempt_count_sub);
3198
3199#else
3200static inline void preempt_latency_start(int val) { }
3201static inline void preempt_latency_stop(int val) { }
3202#endif
3203
3204/*
3205 * Print scheduling while atomic bug:
3206 */
3207static noinline void __schedule_bug(struct task_struct *prev)
3208{
3209 /* Save this before calling printk(), since that will clobber it */
3210 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3211
3212 if (oops_in_progress)
3213 return;
3214
3215 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3216 prev->comm, prev->pid, preempt_count());
3217
3218 debug_show_held_locks(prev);
3219 print_modules();
3220 if (irqs_disabled())
3221 print_irqtrace_events(prev);
3222 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3223 && in_atomic_preempt_off()) {
3224 pr_err("Preemption disabled at:");
3225 print_ip_sym(preempt_disable_ip);
3226 pr_cont("\n");
3227 }
3228 if (panic_on_warn)
3229 panic("scheduling while atomic\n");
3230
3231 dump_stack();
3232 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3233}
3234
3235/*
3236 * Various schedule()-time debugging checks and statistics:
3237 */
3238static inline void schedule_debug(struct task_struct *prev)
3239{
3240#ifdef CONFIG_SCHED_STACK_END_CHECK
3241 if (task_stack_end_corrupted(prev))
3242 panic("corrupted stack end detected inside scheduler\n");
3243#endif
3244
3245 if (unlikely(in_atomic_preempt_off())) {
3246 __schedule_bug(prev);
3247 preempt_count_set(PREEMPT_DISABLED);
3248 }
3249 rcu_sleep_check();
3250
3251 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3252
3253 schedstat_inc(this_rq()->sched_count);
3254}
3255
3256/*
3257 * Pick up the highest-prio task:
3258 */
3259static inline struct task_struct *
3260pick_next_task(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
3261{
3262 const struct sched_class *class = &fair_sched_class;
3263 struct task_struct *p;
3264
3265 /*
3266 * Optimization: we know that if all tasks are in
3267 * the fair class we can call that function directly:
3268 */
3269 if (likely(prev->sched_class == class &&
3270 rq->nr_running == rq->cfs.h_nr_running)) {
3271 p = fair_sched_class.pick_next_task(rq, prev, cookie);
3272 if (unlikely(p == RETRY_TASK))
3273 goto again;
3274
3275 /* assumes fair_sched_class->next == idle_sched_class */
3276 if (unlikely(!p))
3277 p = idle_sched_class.pick_next_task(rq, prev, cookie);
3278
3279 return p;
3280 }
3281
3282again:
3283 for_each_class(class) {
3284 p = class->pick_next_task(rq, prev, cookie);
3285 if (p) {
3286 if (unlikely(p == RETRY_TASK))
3287 goto again;
3288 return p;
3289 }
3290 }
3291
3292 BUG(); /* the idle class will always have a runnable task */
3293}
3294
3295/*
3296 * __schedule() is the main scheduler function.
3297 *
3298 * The main means of driving the scheduler and thus entering this function are:
3299 *
3300 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3301 *
3302 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3303 * paths. For example, see arch/x86/entry_64.S.
3304 *
3305 * To drive preemption between tasks, the scheduler sets the flag in timer
3306 * interrupt handler scheduler_tick().
3307 *
3308 * 3. Wakeups don't really cause entry into schedule(). They add a
3309 * task to the run-queue and that's it.
3310 *
3311 * Now, if the new task added to the run-queue preempts the current
3312 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3313 * called on the nearest possible occasion:
3314 *
3315 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3316 *
3317 * - in syscall or exception context, at the next outmost
3318 * preempt_enable(). (this might be as soon as the wake_up()'s
3319 * spin_unlock()!)
3320 *
3321 * - in IRQ context, return from interrupt-handler to
3322 * preemptible context
3323 *
3324 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3325 * then at the next:
3326 *
3327 * - cond_resched() call
3328 * - explicit schedule() call
3329 * - return from syscall or exception to user-space
3330 * - return from interrupt-handler to user-space
3331 *
3332 * WARNING: must be called with preemption disabled!
3333 */
3334static void __sched notrace __schedule(bool preempt)
3335{
3336 struct task_struct *prev, *next;
3337 unsigned long *switch_count;
3338 struct pin_cookie cookie;
3339 struct rq *rq;
3340 int cpu;
3341
3342 cpu = smp_processor_id();
3343 rq = cpu_rq(cpu);
3344 prev = rq->curr;
3345
3346 schedule_debug(prev);
3347
3348 if (sched_feat(HRTICK))
3349 hrtick_clear(rq);
3350
3351 local_irq_disable();
3352 rcu_note_context_switch();
3353
3354 /*
3355 * Make sure that signal_pending_state()->signal_pending() below
3356 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3357 * done by the caller to avoid the race with signal_wake_up().
3358 */
3359 smp_mb__before_spinlock();
3360 raw_spin_lock(&rq->lock);
3361 cookie = lockdep_pin_lock(&rq->lock);
3362
3363 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3364
3365 switch_count = &prev->nivcsw;
3366 if (!preempt && prev->state) {
3367 if (unlikely(signal_pending_state(prev->state, prev))) {
3368 prev->state = TASK_RUNNING;
3369 } else {
3370 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3371 prev->on_rq = 0;
3372
3373 /*
3374 * If a worker went to sleep, notify and ask workqueue
3375 * whether it wants to wake up a task to maintain
3376 * concurrency.
3377 */
3378 if (prev->flags & PF_WQ_WORKER) {
3379 struct task_struct *to_wakeup;
3380
3381 to_wakeup = wq_worker_sleeping(prev);
3382 if (to_wakeup)
3383 try_to_wake_up_local(to_wakeup, cookie);
3384 }
3385 }
3386 switch_count = &prev->nvcsw;
3387 }
3388
3389 if (task_on_rq_queued(prev))
3390 update_rq_clock(rq);
3391
3392 next = pick_next_task(rq, prev, cookie);
3393 clear_tsk_need_resched(prev);
3394 clear_preempt_need_resched();
3395 rq->clock_skip_update = 0;
3396
3397 if (likely(prev != next)) {
3398 rq->nr_switches++;
3399 rq->curr = next;
3400 ++*switch_count;
3401
3402 trace_sched_switch(preempt, prev, next);
3403 rq = context_switch(rq, prev, next, cookie); /* unlocks the rq */
3404 } else {
3405 lockdep_unpin_lock(&rq->lock, cookie);
3406 raw_spin_unlock_irq(&rq->lock);
3407 }
3408
3409 balance_callback(rq);
3410}
3411
3412void __noreturn do_task_dead(void)
3413{
3414 /*
3415 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3416 * when the following two conditions become true.
3417 * - There is race condition of mmap_sem (It is acquired by
3418 * exit_mm()), and
3419 * - SMI occurs before setting TASK_RUNINNG.
3420 * (or hypervisor of virtual machine switches to other guest)
3421 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3422 *
3423 * To avoid it, we have to wait for releasing tsk->pi_lock which
3424 * is held by try_to_wake_up()
3425 */
3426 smp_mb();
3427 raw_spin_unlock_wait(¤t->pi_lock);
3428
3429 /* causes final put_task_struct in finish_task_switch(). */
3430 __set_current_state(TASK_DEAD);
3431 current->flags |= PF_NOFREEZE; /* tell freezer to ignore us */
3432 __schedule(false);
3433 BUG();
3434 /* Avoid "noreturn function does return". */
3435 for (;;)
3436 cpu_relax(); /* For when BUG is null */
3437}
3438
3439static inline void sched_submit_work(struct task_struct *tsk)
3440{
3441 if (!tsk->state || tsk_is_pi_blocked(tsk))
3442 return;
3443 /*
3444 * If we are going to sleep and we have plugged IO queued,
3445 * make sure to submit it to avoid deadlocks.
3446 */
3447 if (blk_needs_flush_plug(tsk))
3448 blk_schedule_flush_plug(tsk);
3449}
3450
3451asmlinkage __visible void __sched schedule(void)
3452{
3453 struct task_struct *tsk = current;
3454
3455 sched_submit_work(tsk);
3456 do {
3457 preempt_disable();
3458 __schedule(false);
3459 sched_preempt_enable_no_resched();
3460 } while (need_resched());
3461}
3462EXPORT_SYMBOL(schedule);
3463
3464#ifdef CONFIG_CONTEXT_TRACKING
3465asmlinkage __visible void __sched schedule_user(void)
3466{
3467 /*
3468 * If we come here after a random call to set_need_resched(),
3469 * or we have been woken up remotely but the IPI has not yet arrived,
3470 * we haven't yet exited the RCU idle mode. Do it here manually until
3471 * we find a better solution.
3472 *
3473 * NB: There are buggy callers of this function. Ideally we
3474 * should warn if prev_state != CONTEXT_USER, but that will trigger
3475 * too frequently to make sense yet.
3476 */
3477 enum ctx_state prev_state = exception_enter();
3478 schedule();
3479 exception_exit(prev_state);
3480}
3481#endif
3482
3483/**
3484 * schedule_preempt_disabled - called with preemption disabled
3485 *
3486 * Returns with preemption disabled. Note: preempt_count must be 1
3487 */
3488void __sched schedule_preempt_disabled(void)
3489{
3490 sched_preempt_enable_no_resched();
3491 schedule();
3492 preempt_disable();
3493}
3494
3495static void __sched notrace preempt_schedule_common(void)
3496{
3497 do {
3498 /*
3499 * Because the function tracer can trace preempt_count_sub()
3500 * and it also uses preempt_enable/disable_notrace(), if
3501 * NEED_RESCHED is set, the preempt_enable_notrace() called
3502 * by the function tracer will call this function again and
3503 * cause infinite recursion.
3504 *
3505 * Preemption must be disabled here before the function
3506 * tracer can trace. Break up preempt_disable() into two
3507 * calls. One to disable preemption without fear of being
3508 * traced. The other to still record the preemption latency,
3509 * which can also be traced by the function tracer.
3510 */
3511 preempt_disable_notrace();
3512 preempt_latency_start(1);
3513 __schedule(true);
3514 preempt_latency_stop(1);
3515 preempt_enable_no_resched_notrace();
3516
3517 /*
3518 * Check again in case we missed a preemption opportunity
3519 * between schedule and now.
3520 */
3521 } while (need_resched());
3522}
3523
3524#ifdef CONFIG_PREEMPT
3525/*
3526 * this is the entry point to schedule() from in-kernel preemption
3527 * off of preempt_enable. Kernel preemptions off return from interrupt
3528 * occur there and call schedule directly.
3529 */
3530asmlinkage __visible void __sched notrace preempt_schedule(void)
3531{
3532 /*
3533 * If there is a non-zero preempt_count or interrupts are disabled,
3534 * we do not want to preempt the current task. Just return..
3535 */
3536 if (likely(!preemptible()))
3537 return;
3538
3539 preempt_schedule_common();
3540}
3541NOKPROBE_SYMBOL(preempt_schedule);
3542EXPORT_SYMBOL(preempt_schedule);
3543
3544/**
3545 * preempt_schedule_notrace - preempt_schedule called by tracing
3546 *
3547 * The tracing infrastructure uses preempt_enable_notrace to prevent
3548 * recursion and tracing preempt enabling caused by the tracing
3549 * infrastructure itself. But as tracing can happen in areas coming
3550 * from userspace or just about to enter userspace, a preempt enable
3551 * can occur before user_exit() is called. This will cause the scheduler
3552 * to be called when the system is still in usermode.
3553 *
3554 * To prevent this, the preempt_enable_notrace will use this function
3555 * instead of preempt_schedule() to exit user context if needed before
3556 * calling the scheduler.
3557 */
3558asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3559{
3560 enum ctx_state prev_ctx;
3561
3562 if (likely(!preemptible()))
3563 return;
3564
3565 do {
3566 /*
3567 * Because the function tracer can trace preempt_count_sub()
3568 * and it also uses preempt_enable/disable_notrace(), if
3569 * NEED_RESCHED is set, the preempt_enable_notrace() called
3570 * by the function tracer will call this function again and
3571 * cause infinite recursion.
3572 *
3573 * Preemption must be disabled here before the function
3574 * tracer can trace. Break up preempt_disable() into two
3575 * calls. One to disable preemption without fear of being
3576 * traced. The other to still record the preemption latency,
3577 * which can also be traced by the function tracer.
3578 */
3579 preempt_disable_notrace();
3580 preempt_latency_start(1);
3581 /*
3582 * Needs preempt disabled in case user_exit() is traced
3583 * and the tracer calls preempt_enable_notrace() causing
3584 * an infinite recursion.
3585 */
3586 prev_ctx = exception_enter();
3587 __schedule(true);
3588 exception_exit(prev_ctx);
3589
3590 preempt_latency_stop(1);
3591 preempt_enable_no_resched_notrace();
3592 } while (need_resched());
3593}
3594EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3595
3596#endif /* CONFIG_PREEMPT */
3597
3598/*
3599 * this is the entry point to schedule() from kernel preemption
3600 * off of irq context.
3601 * Note, that this is called and return with irqs disabled. This will
3602 * protect us against recursive calling from irq.
3603 */
3604asmlinkage __visible void __sched preempt_schedule_irq(void)
3605{
3606 enum ctx_state prev_state;
3607
3608 /* Catch callers which need to be fixed */
3609 BUG_ON(preempt_count() || !irqs_disabled());
3610
3611 prev_state = exception_enter();
3612
3613 do {
3614 preempt_disable();
3615 local_irq_enable();
3616 __schedule(true);
3617 local_irq_disable();
3618 sched_preempt_enable_no_resched();
3619 } while (need_resched());
3620
3621 exception_exit(prev_state);
3622}
3623
3624int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3625 void *key)
3626{
3627 return try_to_wake_up(curr->private, mode, wake_flags);
3628}
3629EXPORT_SYMBOL(default_wake_function);
3630
3631#ifdef CONFIG_RT_MUTEXES
3632
3633/*
3634 * rt_mutex_setprio - set the current priority of a task
3635 * @p: task
3636 * @prio: prio value (kernel-internal form)
3637 *
3638 * This function changes the 'effective' priority of a task. It does
3639 * not touch ->normal_prio like __setscheduler().
3640 *
3641 * Used by the rt_mutex code to implement priority inheritance
3642 * logic. Call site only calls if the priority of the task changed.
3643 */
3644void rt_mutex_setprio(struct task_struct *p, int prio)
3645{
3646 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3647 const struct sched_class *prev_class;
3648 struct rq_flags rf;
3649 struct rq *rq;
3650
3651 BUG_ON(prio > MAX_PRIO);
3652
3653 rq = __task_rq_lock(p, &rf);
3654
3655 /*
3656 * Idle task boosting is a nono in general. There is one
3657 * exception, when PREEMPT_RT and NOHZ is active:
3658 *
3659 * The idle task calls get_next_timer_interrupt() and holds
3660 * the timer wheel base->lock on the CPU and another CPU wants
3661 * to access the timer (probably to cancel it). We can safely
3662 * ignore the boosting request, as the idle CPU runs this code
3663 * with interrupts disabled and will complete the lock
3664 * protected section without being interrupted. So there is no
3665 * real need to boost.
3666 */
3667 if (unlikely(p == rq->idle)) {
3668 WARN_ON(p != rq->curr);
3669 WARN_ON(p->pi_blocked_on);
3670 goto out_unlock;
3671 }
3672
3673 trace_sched_pi_setprio(p, prio);
3674 oldprio = p->prio;
3675
3676 if (oldprio == prio)
3677 queue_flag &= ~DEQUEUE_MOVE;
3678
3679 prev_class = p->sched_class;
3680 queued = task_on_rq_queued(p);
3681 running = task_current(rq, p);
3682 if (queued)
3683 dequeue_task(rq, p, queue_flag);
3684 if (running)
3685 put_prev_task(rq, p);
3686
3687 /*
3688 * Boosting condition are:
3689 * 1. -rt task is running and holds mutex A
3690 * --> -dl task blocks on mutex A
3691 *
3692 * 2. -dl task is running and holds mutex A
3693 * --> -dl task blocks on mutex A and could preempt the
3694 * running task
3695 */
3696 if (dl_prio(prio)) {
3697 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3698 if (!dl_prio(p->normal_prio) ||
3699 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3700 p->dl.dl_boosted = 1;
3701 queue_flag |= ENQUEUE_REPLENISH;
3702 } else
3703 p->dl.dl_boosted = 0;
3704 p->sched_class = &dl_sched_class;
3705 } else if (rt_prio(prio)) {
3706 if (dl_prio(oldprio))
3707 p->dl.dl_boosted = 0;
3708 if (oldprio < prio)
3709 queue_flag |= ENQUEUE_HEAD;
3710 p->sched_class = &rt_sched_class;
3711 } else {
3712 if (dl_prio(oldprio))
3713 p->dl.dl_boosted = 0;
3714 if (rt_prio(oldprio))
3715 p->rt.timeout = 0;
3716 p->sched_class = &fair_sched_class;
3717 }
3718
3719 p->prio = prio;
3720
3721 if (queued)
3722 enqueue_task(rq, p, queue_flag);
3723 if (running)
3724 set_curr_task(rq, p);
3725
3726 check_class_changed(rq, p, prev_class, oldprio);
3727out_unlock:
3728 preempt_disable(); /* avoid rq from going away on us */
3729 __task_rq_unlock(rq, &rf);
3730
3731 balance_callback(rq);
3732 preempt_enable();
3733}
3734#endif
3735
3736void set_user_nice(struct task_struct *p, long nice)
3737{
3738 bool queued, running;
3739 int old_prio, delta;
3740 struct rq_flags rf;
3741 struct rq *rq;
3742
3743 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3744 return;
3745 /*
3746 * We have to be careful, if called from sys_setpriority(),
3747 * the task might be in the middle of scheduling on another CPU.
3748 */
3749 rq = task_rq_lock(p, &rf);
3750 /*
3751 * The RT priorities are set via sched_setscheduler(), but we still
3752 * allow the 'normal' nice value to be set - but as expected
3753 * it wont have any effect on scheduling until the task is
3754 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3755 */
3756 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3757 p->static_prio = NICE_TO_PRIO(nice);
3758 goto out_unlock;
3759 }
3760 queued = task_on_rq_queued(p);
3761 running = task_current(rq, p);
3762 if (queued)
3763 dequeue_task(rq, p, DEQUEUE_SAVE);
3764 if (running)
3765 put_prev_task(rq, p);
3766
3767 p->static_prio = NICE_TO_PRIO(nice);
3768 set_load_weight(p);
3769 old_prio = p->prio;
3770 p->prio = effective_prio(p);
3771 delta = p->prio - old_prio;
3772
3773 if (queued) {
3774 enqueue_task(rq, p, ENQUEUE_RESTORE);
3775 /*
3776 * If the task increased its priority or is running and
3777 * lowered its priority, then reschedule its CPU:
3778 */
3779 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3780 resched_curr(rq);
3781 }
3782 if (running)
3783 set_curr_task(rq, p);
3784out_unlock:
3785 task_rq_unlock(rq, p, &rf);
3786}
3787EXPORT_SYMBOL(set_user_nice);
3788
3789/*
3790 * can_nice - check if a task can reduce its nice value
3791 * @p: task
3792 * @nice: nice value
3793 */
3794int can_nice(const struct task_struct *p, const int nice)
3795{
3796 /* convert nice value [19,-20] to rlimit style value [1,40] */
3797 int nice_rlim = nice_to_rlimit(nice);
3798
3799 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3800 capable(CAP_SYS_NICE));
3801}
3802
3803#ifdef __ARCH_WANT_SYS_NICE
3804
3805/*
3806 * sys_nice - change the priority of the current process.
3807 * @increment: priority increment
3808 *
3809 * sys_setpriority is a more generic, but much slower function that
3810 * does similar things.
3811 */
3812SYSCALL_DEFINE1(nice, int, increment)
3813{
3814 long nice, retval;
3815
3816 /*
3817 * Setpriority might change our priority at the same moment.
3818 * We don't have to worry. Conceptually one call occurs first
3819 * and we have a single winner.
3820 */
3821 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3822 nice = task_nice(current) + increment;
3823
3824 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3825 if (increment < 0 && !can_nice(current, nice))
3826 return -EPERM;
3827
3828 retval = security_task_setnice(current, nice);
3829 if (retval)
3830 return retval;
3831
3832 set_user_nice(current, nice);
3833 return 0;
3834}
3835
3836#endif
3837
3838/**
3839 * task_prio - return the priority value of a given task.
3840 * @p: the task in question.
3841 *
3842 * Return: The priority value as seen by users in /proc.
3843 * RT tasks are offset by -200. Normal tasks are centered
3844 * around 0, value goes from -16 to +15.
3845 */
3846int task_prio(const struct task_struct *p)
3847{
3848 return p->prio - MAX_RT_PRIO;
3849}
3850
3851/**
3852 * idle_cpu - is a given cpu idle currently?
3853 * @cpu: the processor in question.
3854 *
3855 * Return: 1 if the CPU is currently idle. 0 otherwise.
3856 */
3857int idle_cpu(int cpu)
3858{
3859 struct rq *rq = cpu_rq(cpu);
3860
3861 if (rq->curr != rq->idle)
3862 return 0;
3863
3864 if (rq->nr_running)
3865 return 0;
3866
3867#ifdef CONFIG_SMP
3868 if (!llist_empty(&rq->wake_list))
3869 return 0;
3870#endif
3871
3872 return 1;
3873}
3874
3875/**
3876 * idle_task - return the idle task for a given cpu.
3877 * @cpu: the processor in question.
3878 *
3879 * Return: The idle task for the cpu @cpu.
3880 */
3881struct task_struct *idle_task(int cpu)
3882{
3883 return cpu_rq(cpu)->idle;
3884}
3885
3886/**
3887 * find_process_by_pid - find a process with a matching PID value.
3888 * @pid: the pid in question.
3889 *
3890 * The task of @pid, if found. %NULL otherwise.
3891 */
3892static struct task_struct *find_process_by_pid(pid_t pid)
3893{
3894 return pid ? find_task_by_vpid(pid) : current;
3895}
3896
3897/*
3898 * This function initializes the sched_dl_entity of a newly becoming
3899 * SCHED_DEADLINE task.
3900 *
3901 * Only the static values are considered here, the actual runtime and the
3902 * absolute deadline will be properly calculated when the task is enqueued
3903 * for the first time with its new policy.
3904 */
3905static void
3906__setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3907{
3908 struct sched_dl_entity *dl_se = &p->dl;
3909
3910 dl_se->dl_runtime = attr->sched_runtime;
3911 dl_se->dl_deadline = attr->sched_deadline;
3912 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3913 dl_se->flags = attr->sched_flags;
3914 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3915
3916 /*
3917 * Changing the parameters of a task is 'tricky' and we're not doing
3918 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3919 *
3920 * What we SHOULD do is delay the bandwidth release until the 0-lag
3921 * point. This would include retaining the task_struct until that time
3922 * and change dl_overflow() to not immediately decrement the current
3923 * amount.
3924 *
3925 * Instead we retain the current runtime/deadline and let the new
3926 * parameters take effect after the current reservation period lapses.
3927 * This is safe (albeit pessimistic) because the 0-lag point is always
3928 * before the current scheduling deadline.
3929 *
3930 * We can still have temporary overloads because we do not delay the
3931 * change in bandwidth until that time; so admission control is
3932 * not on the safe side. It does however guarantee tasks will never
3933 * consume more than promised.
3934 */
3935}
3936
3937/*
3938 * sched_setparam() passes in -1 for its policy, to let the functions
3939 * it calls know not to change it.
3940 */
3941#define SETPARAM_POLICY -1
3942
3943static void __setscheduler_params(struct task_struct *p,
3944 const struct sched_attr *attr)
3945{
3946 int policy = attr->sched_policy;
3947
3948 if (policy == SETPARAM_POLICY)
3949 policy = p->policy;
3950
3951 p->policy = policy;
3952
3953 if (dl_policy(policy))
3954 __setparam_dl(p, attr);
3955 else if (fair_policy(policy))
3956 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3957
3958 /*
3959 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3960 * !rt_policy. Always setting this ensures that things like
3961 * getparam()/getattr() don't report silly values for !rt tasks.
3962 */
3963 p->rt_priority = attr->sched_priority;
3964 p->normal_prio = normal_prio(p);
3965 set_load_weight(p);
3966}
3967
3968/* Actually do priority change: must hold pi & rq lock. */
3969static void __setscheduler(struct rq *rq, struct task_struct *p,
3970 const struct sched_attr *attr, bool keep_boost)
3971{
3972 __setscheduler_params(p, attr);
3973
3974 /*
3975 * Keep a potential priority boosting if called from
3976 * sched_setscheduler().
3977 */
3978 if (keep_boost)
3979 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3980 else
3981 p->prio = normal_prio(p);
3982
3983 if (dl_prio(p->prio))
3984 p->sched_class = &dl_sched_class;
3985 else if (rt_prio(p->prio))
3986 p->sched_class = &rt_sched_class;
3987 else
3988 p->sched_class = &fair_sched_class;
3989}
3990
3991static void
3992__getparam_dl(struct task_struct *p, struct sched_attr *attr)
3993{
3994 struct sched_dl_entity *dl_se = &p->dl;
3995
3996 attr->sched_priority = p->rt_priority;
3997 attr->sched_runtime = dl_se->dl_runtime;
3998 attr->sched_deadline = dl_se->dl_deadline;
3999 attr->sched_period = dl_se->dl_period;
4000 attr->sched_flags = dl_se->flags;
4001}
4002
4003/*
4004 * This function validates the new parameters of a -deadline task.
4005 * We ask for the deadline not being zero, and greater or equal
4006 * than the runtime, as well as the period of being zero or
4007 * greater than deadline. Furthermore, we have to be sure that
4008 * user parameters are above the internal resolution of 1us (we
4009 * check sched_runtime only since it is always the smaller one) and
4010 * below 2^63 ns (we have to check both sched_deadline and
4011 * sched_period, as the latter can be zero).
4012 */
4013static bool
4014__checkparam_dl(const struct sched_attr *attr)
4015{
4016 /* deadline != 0 */
4017 if (attr->sched_deadline == 0)
4018 return false;
4019
4020 /*
4021 * Since we truncate DL_SCALE bits, make sure we're at least
4022 * that big.
4023 */
4024 if (attr->sched_runtime < (1ULL << DL_SCALE))
4025 return false;
4026
4027 /*
4028 * Since we use the MSB for wrap-around and sign issues, make
4029 * sure it's not set (mind that period can be equal to zero).
4030 */
4031 if (attr->sched_deadline & (1ULL << 63) ||
4032 attr->sched_period & (1ULL << 63))
4033 return false;
4034
4035 /* runtime <= deadline <= period (if period != 0) */
4036 if ((attr->sched_period != 0 &&
4037 attr->sched_period < attr->sched_deadline) ||
4038 attr->sched_deadline < attr->sched_runtime)
4039 return false;
4040
4041 return true;
4042}
4043
4044/*
4045 * check the target process has a UID that matches the current process's
4046 */
4047static bool check_same_owner(struct task_struct *p)
4048{
4049 const struct cred *cred = current_cred(), *pcred;
4050 bool match;
4051
4052 rcu_read_lock();
4053 pcred = __task_cred(p);
4054 match = (uid_eq(cred->euid, pcred->euid) ||
4055 uid_eq(cred->euid, pcred->uid));
4056 rcu_read_unlock();
4057 return match;
4058}
4059
4060static bool dl_param_changed(struct task_struct *p,
4061 const struct sched_attr *attr)
4062{
4063 struct sched_dl_entity *dl_se = &p->dl;
4064
4065 if (dl_se->dl_runtime != attr->sched_runtime ||
4066 dl_se->dl_deadline != attr->sched_deadline ||
4067 dl_se->dl_period != attr->sched_period ||
4068 dl_se->flags != attr->sched_flags)
4069 return true;
4070
4071 return false;
4072}
4073
4074static int __sched_setscheduler(struct task_struct *p,
4075 const struct sched_attr *attr,
4076 bool user, bool pi)
4077{
4078 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4079 MAX_RT_PRIO - 1 - attr->sched_priority;
4080 int retval, oldprio, oldpolicy = -1, queued, running;
4081 int new_effective_prio, policy = attr->sched_policy;
4082 const struct sched_class *prev_class;
4083 struct rq_flags rf;
4084 int reset_on_fork;
4085 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4086 struct rq *rq;
4087
4088 /* may grab non-irq protected spin_locks */
4089 BUG_ON(in_interrupt());
4090recheck:
4091 /* double check policy once rq lock held */
4092 if (policy < 0) {
4093 reset_on_fork = p->sched_reset_on_fork;
4094 policy = oldpolicy = p->policy;
4095 } else {
4096 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4097
4098 if (!valid_policy(policy))
4099 return -EINVAL;
4100 }
4101
4102 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4103 return -EINVAL;
4104
4105 /*
4106 * Valid priorities for SCHED_FIFO and SCHED_RR are
4107 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4108 * SCHED_BATCH and SCHED_IDLE is 0.
4109 */
4110 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4111 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4112 return -EINVAL;
4113 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4114 (rt_policy(policy) != (attr->sched_priority != 0)))
4115 return -EINVAL;
4116
4117 /*
4118 * Allow unprivileged RT tasks to decrease priority:
4119 */
4120 if (user && !capable(CAP_SYS_NICE)) {
4121 if (fair_policy(policy)) {
4122 if (attr->sched_nice < task_nice(p) &&
4123 !can_nice(p, attr->sched_nice))
4124 return -EPERM;
4125 }
4126
4127 if (rt_policy(policy)) {
4128 unsigned long rlim_rtprio =
4129 task_rlimit(p, RLIMIT_RTPRIO);
4130
4131 /* can't set/change the rt policy */
4132 if (policy != p->policy && !rlim_rtprio)
4133 return -EPERM;
4134
4135 /* can't increase priority */
4136 if (attr->sched_priority > p->rt_priority &&
4137 attr->sched_priority > rlim_rtprio)
4138 return -EPERM;
4139 }
4140
4141 /*
4142 * Can't set/change SCHED_DEADLINE policy at all for now
4143 * (safest behavior); in the future we would like to allow
4144 * unprivileged DL tasks to increase their relative deadline
4145 * or reduce their runtime (both ways reducing utilization)
4146 */
4147 if (dl_policy(policy))
4148 return -EPERM;
4149
4150 /*
4151 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4152 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4153 */
4154 if (idle_policy(p->policy) && !idle_policy(policy)) {
4155 if (!can_nice(p, task_nice(p)))
4156 return -EPERM;
4157 }
4158
4159 /* can't change other user's priorities */
4160 if (!check_same_owner(p))
4161 return -EPERM;
4162
4163 /* Normal users shall not reset the sched_reset_on_fork flag */
4164 if (p->sched_reset_on_fork && !reset_on_fork)
4165 return -EPERM;
4166 }
4167
4168 if (user) {
4169 retval = security_task_setscheduler(p);
4170 if (retval)
4171 return retval;
4172 }
4173
4174 /*
4175 * make sure no PI-waiters arrive (or leave) while we are
4176 * changing the priority of the task:
4177 *
4178 * To be able to change p->policy safely, the appropriate
4179 * runqueue lock must be held.
4180 */
4181 rq = task_rq_lock(p, &rf);
4182
4183 /*
4184 * Changing the policy of the stop threads its a very bad idea
4185 */
4186 if (p == rq->stop) {
4187 task_rq_unlock(rq, p, &rf);
4188 return -EINVAL;
4189 }
4190
4191 /*
4192 * If not changing anything there's no need to proceed further,
4193 * but store a possible modification of reset_on_fork.
4194 */
4195 if (unlikely(policy == p->policy)) {
4196 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4197 goto change;
4198 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4199 goto change;
4200 if (dl_policy(policy) && dl_param_changed(p, attr))
4201 goto change;
4202
4203 p->sched_reset_on_fork = reset_on_fork;
4204 task_rq_unlock(rq, p, &rf);
4205 return 0;
4206 }
4207change:
4208
4209 if (user) {
4210#ifdef CONFIG_RT_GROUP_SCHED
4211 /*
4212 * Do not allow realtime tasks into groups that have no runtime
4213 * assigned.
4214 */
4215 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4216 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4217 !task_group_is_autogroup(task_group(p))) {
4218 task_rq_unlock(rq, p, &rf);
4219 return -EPERM;
4220 }
4221#endif
4222#ifdef CONFIG_SMP
4223 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4224 cpumask_t *span = rq->rd->span;
4225
4226 /*
4227 * Don't allow tasks with an affinity mask smaller than
4228 * the entire root_domain to become SCHED_DEADLINE. We
4229 * will also fail if there's no bandwidth available.
4230 */
4231 if (!cpumask_subset(span, &p->cpus_allowed) ||
4232 rq->rd->dl_bw.bw == 0) {
4233 task_rq_unlock(rq, p, &rf);
4234 return -EPERM;
4235 }
4236 }
4237#endif
4238 }
4239
4240 /* recheck policy now with rq lock held */
4241 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4242 policy = oldpolicy = -1;
4243 task_rq_unlock(rq, p, &rf);
4244 goto recheck;
4245 }
4246
4247 /*
4248 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4249 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4250 * is available.
4251 */
4252 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4253 task_rq_unlock(rq, p, &rf);
4254 return -EBUSY;
4255 }
4256
4257 p->sched_reset_on_fork = reset_on_fork;
4258 oldprio = p->prio;
4259
4260 if (pi) {
4261 /*
4262 * Take priority boosted tasks into account. If the new
4263 * effective priority is unchanged, we just store the new
4264 * normal parameters and do not touch the scheduler class and
4265 * the runqueue. This will be done when the task deboost
4266 * itself.
4267 */
4268 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4269 if (new_effective_prio == oldprio)
4270 queue_flags &= ~DEQUEUE_MOVE;
4271 }
4272
4273 queued = task_on_rq_queued(p);
4274 running = task_current(rq, p);
4275 if (queued)
4276 dequeue_task(rq, p, queue_flags);
4277 if (running)
4278 put_prev_task(rq, p);
4279
4280 prev_class = p->sched_class;
4281 __setscheduler(rq, p, attr, pi);
4282
4283 if (queued) {
4284 /*
4285 * We enqueue to tail when the priority of a task is
4286 * increased (user space view).
4287 */
4288 if (oldprio < p->prio)
4289 queue_flags |= ENQUEUE_HEAD;
4290
4291 enqueue_task(rq, p, queue_flags);
4292 }
4293 if (running)
4294 set_curr_task(rq, p);
4295
4296 check_class_changed(rq, p, prev_class, oldprio);
4297 preempt_disable(); /* avoid rq from going away on us */
4298 task_rq_unlock(rq, p, &rf);
4299
4300 if (pi)
4301 rt_mutex_adjust_pi(p);
4302
4303 /*
4304 * Run balance callbacks after we've adjusted the PI chain.
4305 */
4306 balance_callback(rq);
4307 preempt_enable();
4308
4309 return 0;
4310}
4311
4312static int _sched_setscheduler(struct task_struct *p, int policy,
4313 const struct sched_param *param, bool check)
4314{
4315 struct sched_attr attr = {
4316 .sched_policy = policy,
4317 .sched_priority = param->sched_priority,
4318 .sched_nice = PRIO_TO_NICE(p->static_prio),
4319 };
4320
4321 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4322 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4323 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4324 policy &= ~SCHED_RESET_ON_FORK;
4325 attr.sched_policy = policy;
4326 }
4327
4328 return __sched_setscheduler(p, &attr, check, true);
4329}
4330/**
4331 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4332 * @p: the task in question.
4333 * @policy: new policy.
4334 * @param: structure containing the new RT priority.
4335 *
4336 * Return: 0 on success. An error code otherwise.
4337 *
4338 * NOTE that the task may be already dead.
4339 */
4340int sched_setscheduler(struct task_struct *p, int policy,
4341 const struct sched_param *param)
4342{
4343 return _sched_setscheduler(p, policy, param, true);
4344}
4345EXPORT_SYMBOL_GPL(sched_setscheduler);
4346
4347int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4348{
4349 return __sched_setscheduler(p, attr, true, true);
4350}
4351EXPORT_SYMBOL_GPL(sched_setattr);
4352
4353/**
4354 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4355 * @p: the task in question.
4356 * @policy: new policy.
4357 * @param: structure containing the new RT priority.
4358 *
4359 * Just like sched_setscheduler, only don't bother checking if the
4360 * current context has permission. For example, this is needed in
4361 * stop_machine(): we create temporary high priority worker threads,
4362 * but our caller might not have that capability.
4363 *
4364 * Return: 0 on success. An error code otherwise.
4365 */
4366int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4367 const struct sched_param *param)
4368{
4369 return _sched_setscheduler(p, policy, param, false);
4370}
4371EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4372
4373static int
4374do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4375{
4376 struct sched_param lparam;
4377 struct task_struct *p;
4378 int retval;
4379
4380 if (!param || pid < 0)
4381 return -EINVAL;
4382 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4383 return -EFAULT;
4384
4385 rcu_read_lock();
4386 retval = -ESRCH;
4387 p = find_process_by_pid(pid);
4388 if (p != NULL)
4389 retval = sched_setscheduler(p, policy, &lparam);
4390 rcu_read_unlock();
4391
4392 return retval;
4393}
4394
4395/*
4396 * Mimics kernel/events/core.c perf_copy_attr().
4397 */
4398static int sched_copy_attr(struct sched_attr __user *uattr,
4399 struct sched_attr *attr)
4400{
4401 u32 size;
4402 int ret;
4403
4404 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4405 return -EFAULT;
4406
4407 /*
4408 * zero the full structure, so that a short copy will be nice.
4409 */
4410 memset(attr, 0, sizeof(*attr));
4411
4412 ret = get_user(size, &uattr->size);
4413 if (ret)
4414 return ret;
4415
4416 if (size > PAGE_SIZE) /* silly large */
4417 goto err_size;
4418
4419 if (!size) /* abi compat */
4420 size = SCHED_ATTR_SIZE_VER0;
4421
4422 if (size < SCHED_ATTR_SIZE_VER0)
4423 goto err_size;
4424
4425 /*
4426 * If we're handed a bigger struct than we know of,
4427 * ensure all the unknown bits are 0 - i.e. new
4428 * user-space does not rely on any kernel feature
4429 * extensions we dont know about yet.
4430 */
4431 if (size > sizeof(*attr)) {
4432 unsigned char __user *addr;
4433 unsigned char __user *end;
4434 unsigned char val;
4435
4436 addr = (void __user *)uattr + sizeof(*attr);
4437 end = (void __user *)uattr + size;
4438
4439 for (; addr < end; addr++) {
4440 ret = get_user(val, addr);
4441 if (ret)
4442 return ret;
4443 if (val)
4444 goto err_size;
4445 }
4446 size = sizeof(*attr);
4447 }
4448
4449 ret = copy_from_user(attr, uattr, size);
4450 if (ret)
4451 return -EFAULT;
4452
4453 /*
4454 * XXX: do we want to be lenient like existing syscalls; or do we want
4455 * to be strict and return an error on out-of-bounds values?
4456 */
4457 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4458
4459 return 0;
4460
4461err_size:
4462 put_user(sizeof(*attr), &uattr->size);
4463 return -E2BIG;
4464}
4465
4466/**
4467 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4468 * @pid: the pid in question.
4469 * @policy: new policy.
4470 * @param: structure containing the new RT priority.
4471 *
4472 * Return: 0 on success. An error code otherwise.
4473 */
4474SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4475 struct sched_param __user *, param)
4476{
4477 /* negative values for policy are not valid */
4478 if (policy < 0)
4479 return -EINVAL;
4480
4481 return do_sched_setscheduler(pid, policy, param);
4482}
4483
4484/**
4485 * sys_sched_setparam - set/change the RT priority of a thread
4486 * @pid: the pid in question.
4487 * @param: structure containing the new RT priority.
4488 *
4489 * Return: 0 on success. An error code otherwise.
4490 */
4491SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4492{
4493 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4494}
4495
4496/**
4497 * sys_sched_setattr - same as above, but with extended sched_attr
4498 * @pid: the pid in question.
4499 * @uattr: structure containing the extended parameters.
4500 * @flags: for future extension.
4501 */
4502SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4503 unsigned int, flags)
4504{
4505 struct sched_attr attr;
4506 struct task_struct *p;
4507 int retval;
4508
4509 if (!uattr || pid < 0 || flags)
4510 return -EINVAL;
4511
4512 retval = sched_copy_attr(uattr, &attr);
4513 if (retval)
4514 return retval;
4515
4516 if ((int)attr.sched_policy < 0)
4517 return -EINVAL;
4518
4519 rcu_read_lock();
4520 retval = -ESRCH;
4521 p = find_process_by_pid(pid);
4522 if (p != NULL)
4523 retval = sched_setattr(p, &attr);
4524 rcu_read_unlock();
4525
4526 return retval;
4527}
4528
4529/**
4530 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4531 * @pid: the pid in question.
4532 *
4533 * Return: On success, the policy of the thread. Otherwise, a negative error
4534 * code.
4535 */
4536SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4537{
4538 struct task_struct *p;
4539 int retval;
4540
4541 if (pid < 0)
4542 return -EINVAL;
4543
4544 retval = -ESRCH;
4545 rcu_read_lock();
4546 p = find_process_by_pid(pid);
4547 if (p) {
4548 retval = security_task_getscheduler(p);
4549 if (!retval)
4550 retval = p->policy
4551 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4552 }
4553 rcu_read_unlock();
4554 return retval;
4555}
4556
4557/**
4558 * sys_sched_getparam - get the RT priority of a thread
4559 * @pid: the pid in question.
4560 * @param: structure containing the RT priority.
4561 *
4562 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4563 * code.
4564 */
4565SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4566{
4567 struct sched_param lp = { .sched_priority = 0 };
4568 struct task_struct *p;
4569 int retval;
4570
4571 if (!param || pid < 0)
4572 return -EINVAL;
4573
4574 rcu_read_lock();
4575 p = find_process_by_pid(pid);
4576 retval = -ESRCH;
4577 if (!p)
4578 goto out_unlock;
4579
4580 retval = security_task_getscheduler(p);
4581 if (retval)
4582 goto out_unlock;
4583
4584 if (task_has_rt_policy(p))
4585 lp.sched_priority = p->rt_priority;
4586 rcu_read_unlock();
4587
4588 /*
4589 * This one might sleep, we cannot do it with a spinlock held ...
4590 */
4591 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4592
4593 return retval;
4594
4595out_unlock:
4596 rcu_read_unlock();
4597 return retval;
4598}
4599
4600static int sched_read_attr(struct sched_attr __user *uattr,
4601 struct sched_attr *attr,
4602 unsigned int usize)
4603{
4604 int ret;
4605
4606 if (!access_ok(VERIFY_WRITE, uattr, usize))
4607 return -EFAULT;
4608
4609 /*
4610 * If we're handed a smaller struct than we know of,
4611 * ensure all the unknown bits are 0 - i.e. old
4612 * user-space does not get uncomplete information.
4613 */
4614 if (usize < sizeof(*attr)) {
4615 unsigned char *addr;
4616 unsigned char *end;
4617
4618 addr = (void *)attr + usize;
4619 end = (void *)attr + sizeof(*attr);
4620
4621 for (; addr < end; addr++) {
4622 if (*addr)
4623 return -EFBIG;
4624 }
4625
4626 attr->size = usize;
4627 }
4628
4629 ret = copy_to_user(uattr, attr, attr->size);
4630 if (ret)
4631 return -EFAULT;
4632
4633 return 0;
4634}
4635
4636/**
4637 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4638 * @pid: the pid in question.
4639 * @uattr: structure containing the extended parameters.
4640 * @size: sizeof(attr) for fwd/bwd comp.
4641 * @flags: for future extension.
4642 */
4643SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4644 unsigned int, size, unsigned int, flags)
4645{
4646 struct sched_attr attr = {
4647 .size = sizeof(struct sched_attr),
4648 };
4649 struct task_struct *p;
4650 int retval;
4651
4652 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4653 size < SCHED_ATTR_SIZE_VER0 || flags)
4654 return -EINVAL;
4655
4656 rcu_read_lock();
4657 p = find_process_by_pid(pid);
4658 retval = -ESRCH;
4659 if (!p)
4660 goto out_unlock;
4661
4662 retval = security_task_getscheduler(p);
4663 if (retval)
4664 goto out_unlock;
4665
4666 attr.sched_policy = p->policy;
4667 if (p->sched_reset_on_fork)
4668 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4669 if (task_has_dl_policy(p))
4670 __getparam_dl(p, &attr);
4671 else if (task_has_rt_policy(p))
4672 attr.sched_priority = p->rt_priority;
4673 else
4674 attr.sched_nice = task_nice(p);
4675
4676 rcu_read_unlock();
4677
4678 retval = sched_read_attr(uattr, &attr, size);
4679 return retval;
4680
4681out_unlock:
4682 rcu_read_unlock();
4683 return retval;
4684}
4685
4686long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4687{
4688 cpumask_var_t cpus_allowed, new_mask;
4689 struct task_struct *p;
4690 int retval;
4691
4692 rcu_read_lock();
4693
4694 p = find_process_by_pid(pid);
4695 if (!p) {
4696 rcu_read_unlock();
4697 return -ESRCH;
4698 }
4699
4700 /* Prevent p going away */
4701 get_task_struct(p);
4702 rcu_read_unlock();
4703
4704 if (p->flags & PF_NO_SETAFFINITY) {
4705 retval = -EINVAL;
4706 goto out_put_task;
4707 }
4708 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4709 retval = -ENOMEM;
4710 goto out_put_task;
4711 }
4712 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4713 retval = -ENOMEM;
4714 goto out_free_cpus_allowed;
4715 }
4716 retval = -EPERM;
4717 if (!check_same_owner(p)) {
4718 rcu_read_lock();
4719 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4720 rcu_read_unlock();
4721 goto out_free_new_mask;
4722 }
4723 rcu_read_unlock();
4724 }
4725
4726 retval = security_task_setscheduler(p);
4727 if (retval)
4728 goto out_free_new_mask;
4729
4730
4731 cpuset_cpus_allowed(p, cpus_allowed);
4732 cpumask_and(new_mask, in_mask, cpus_allowed);
4733
4734 /*
4735 * Since bandwidth control happens on root_domain basis,
4736 * if admission test is enabled, we only admit -deadline
4737 * tasks allowed to run on all the CPUs in the task's
4738 * root_domain.
4739 */
4740#ifdef CONFIG_SMP
4741 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4742 rcu_read_lock();
4743 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4744 retval = -EBUSY;
4745 rcu_read_unlock();
4746 goto out_free_new_mask;
4747 }
4748 rcu_read_unlock();
4749 }
4750#endif
4751again:
4752 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4753
4754 if (!retval) {
4755 cpuset_cpus_allowed(p, cpus_allowed);
4756 if (!cpumask_subset(new_mask, cpus_allowed)) {
4757 /*
4758 * We must have raced with a concurrent cpuset
4759 * update. Just reset the cpus_allowed to the
4760 * cpuset's cpus_allowed
4761 */
4762 cpumask_copy(new_mask, cpus_allowed);
4763 goto again;
4764 }
4765 }
4766out_free_new_mask:
4767 free_cpumask_var(new_mask);
4768out_free_cpus_allowed:
4769 free_cpumask_var(cpus_allowed);
4770out_put_task:
4771 put_task_struct(p);
4772 return retval;
4773}
4774
4775static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4776 struct cpumask *new_mask)
4777{
4778 if (len < cpumask_size())
4779 cpumask_clear(new_mask);
4780 else if (len > cpumask_size())
4781 len = cpumask_size();
4782
4783 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4784}
4785
4786/**
4787 * sys_sched_setaffinity - set the cpu affinity of a process
4788 * @pid: pid of the process
4789 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4790 * @user_mask_ptr: user-space pointer to the new cpu mask
4791 *
4792 * Return: 0 on success. An error code otherwise.
4793 */
4794SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4795 unsigned long __user *, user_mask_ptr)
4796{
4797 cpumask_var_t new_mask;
4798 int retval;
4799
4800 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4801 return -ENOMEM;
4802
4803 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4804 if (retval == 0)
4805 retval = sched_setaffinity(pid, new_mask);
4806 free_cpumask_var(new_mask);
4807 return retval;
4808}
4809
4810long sched_getaffinity(pid_t pid, struct cpumask *mask)
4811{
4812 struct task_struct *p;
4813 unsigned long flags;
4814 int retval;
4815
4816 rcu_read_lock();
4817
4818 retval = -ESRCH;
4819 p = find_process_by_pid(pid);
4820 if (!p)
4821 goto out_unlock;
4822
4823 retval = security_task_getscheduler(p);
4824 if (retval)
4825 goto out_unlock;
4826
4827 raw_spin_lock_irqsave(&p->pi_lock, flags);
4828 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4829 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4830
4831out_unlock:
4832 rcu_read_unlock();
4833
4834 return retval;
4835}
4836
4837/**
4838 * sys_sched_getaffinity - get the cpu affinity of a process
4839 * @pid: pid of the process
4840 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4841 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4842 *
4843 * Return: size of CPU mask copied to user_mask_ptr on success. An
4844 * error code otherwise.
4845 */
4846SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4847 unsigned long __user *, user_mask_ptr)
4848{
4849 int ret;
4850 cpumask_var_t mask;
4851
4852 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4853 return -EINVAL;
4854 if (len & (sizeof(unsigned long)-1))
4855 return -EINVAL;
4856
4857 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4858 return -ENOMEM;
4859
4860 ret = sched_getaffinity(pid, mask);
4861 if (ret == 0) {
4862 size_t retlen = min_t(size_t, len, cpumask_size());
4863
4864 if (copy_to_user(user_mask_ptr, mask, retlen))
4865 ret = -EFAULT;
4866 else
4867 ret = retlen;
4868 }
4869 free_cpumask_var(mask);
4870
4871 return ret;
4872}
4873
4874/**
4875 * sys_sched_yield - yield the current processor to other threads.
4876 *
4877 * This function yields the current CPU to other tasks. If there are no
4878 * other threads running on this CPU then this function will return.
4879 *
4880 * Return: 0.
4881 */
4882SYSCALL_DEFINE0(sched_yield)
4883{
4884 struct rq *rq = this_rq_lock();
4885
4886 schedstat_inc(rq->yld_count);
4887 current->sched_class->yield_task(rq);
4888
4889 /*
4890 * Since we are going to call schedule() anyway, there's
4891 * no need to preempt or enable interrupts:
4892 */
4893 __release(rq->lock);
4894 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4895 do_raw_spin_unlock(&rq->lock);
4896 sched_preempt_enable_no_resched();
4897
4898 schedule();
4899
4900 return 0;
4901}
4902
4903#ifndef CONFIG_PREEMPT
4904int __sched _cond_resched(void)
4905{
4906 if (should_resched(0)) {
4907 preempt_schedule_common();
4908 return 1;
4909 }
4910 return 0;
4911}
4912EXPORT_SYMBOL(_cond_resched);
4913#endif
4914
4915/*
4916 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4917 * call schedule, and on return reacquire the lock.
4918 *
4919 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4920 * operations here to prevent schedule() from being called twice (once via
4921 * spin_unlock(), once by hand).
4922 */
4923int __cond_resched_lock(spinlock_t *lock)
4924{
4925 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4926 int ret = 0;
4927
4928 lockdep_assert_held(lock);
4929
4930 if (spin_needbreak(lock) || resched) {
4931 spin_unlock(lock);
4932 if (resched)
4933 preempt_schedule_common();
4934 else
4935 cpu_relax();
4936 ret = 1;
4937 spin_lock(lock);
4938 }
4939 return ret;
4940}
4941EXPORT_SYMBOL(__cond_resched_lock);
4942
4943int __sched __cond_resched_softirq(void)
4944{
4945 BUG_ON(!in_softirq());
4946
4947 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4948 local_bh_enable();
4949 preempt_schedule_common();
4950 local_bh_disable();
4951 return 1;
4952 }
4953 return 0;
4954}
4955EXPORT_SYMBOL(__cond_resched_softirq);
4956
4957/**
4958 * yield - yield the current processor to other threads.
4959 *
4960 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4961 *
4962 * The scheduler is at all times free to pick the calling task as the most
4963 * eligible task to run, if removing the yield() call from your code breaks
4964 * it, its already broken.
4965 *
4966 * Typical broken usage is:
4967 *
4968 * while (!event)
4969 * yield();
4970 *
4971 * where one assumes that yield() will let 'the other' process run that will
4972 * make event true. If the current task is a SCHED_FIFO task that will never
4973 * happen. Never use yield() as a progress guarantee!!
4974 *
4975 * If you want to use yield() to wait for something, use wait_event().
4976 * If you want to use yield() to be 'nice' for others, use cond_resched().
4977 * If you still want to use yield(), do not!
4978 */
4979void __sched yield(void)
4980{
4981 set_current_state(TASK_RUNNING);
4982 sys_sched_yield();
4983}
4984EXPORT_SYMBOL(yield);
4985
4986/**
4987 * yield_to - yield the current processor to another thread in
4988 * your thread group, or accelerate that thread toward the
4989 * processor it's on.
4990 * @p: target task
4991 * @preempt: whether task preemption is allowed or not
4992 *
4993 * It's the caller's job to ensure that the target task struct
4994 * can't go away on us before we can do any checks.
4995 *
4996 * Return:
4997 * true (>0) if we indeed boosted the target task.
4998 * false (0) if we failed to boost the target.
4999 * -ESRCH if there's no task to yield to.
5000 */
5001int __sched yield_to(struct task_struct *p, bool preempt)
5002{
5003 struct task_struct *curr = current;
5004 struct rq *rq, *p_rq;
5005 unsigned long flags;
5006 int yielded = 0;
5007
5008 local_irq_save(flags);
5009 rq = this_rq();
5010
5011again:
5012 p_rq = task_rq(p);
5013 /*
5014 * If we're the only runnable task on the rq and target rq also
5015 * has only one task, there's absolutely no point in yielding.
5016 */
5017 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5018 yielded = -ESRCH;
5019 goto out_irq;
5020 }
5021
5022 double_rq_lock(rq, p_rq);
5023 if (task_rq(p) != p_rq) {
5024 double_rq_unlock(rq, p_rq);
5025 goto again;
5026 }
5027
5028 if (!curr->sched_class->yield_to_task)
5029 goto out_unlock;
5030
5031 if (curr->sched_class != p->sched_class)
5032 goto out_unlock;
5033
5034 if (task_running(p_rq, p) || p->state)
5035 goto out_unlock;
5036
5037 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5038 if (yielded) {
5039 schedstat_inc(rq->yld_count);
5040 /*
5041 * Make p's CPU reschedule; pick_next_entity takes care of
5042 * fairness.
5043 */
5044 if (preempt && rq != p_rq)
5045 resched_curr(p_rq);
5046 }
5047
5048out_unlock:
5049 double_rq_unlock(rq, p_rq);
5050out_irq:
5051 local_irq_restore(flags);
5052
5053 if (yielded > 0)
5054 schedule();
5055
5056 return yielded;
5057}
5058EXPORT_SYMBOL_GPL(yield_to);
5059
5060/*
5061 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5062 * that process accounting knows that this is a task in IO wait state.
5063 */
5064long __sched io_schedule_timeout(long timeout)
5065{
5066 int old_iowait = current->in_iowait;
5067 struct rq *rq;
5068 long ret;
5069
5070 current->in_iowait = 1;
5071 blk_schedule_flush_plug(current);
5072
5073 delayacct_blkio_start();
5074 rq = raw_rq();
5075 atomic_inc(&rq->nr_iowait);
5076 ret = schedule_timeout(timeout);
5077 current->in_iowait = old_iowait;
5078 atomic_dec(&rq->nr_iowait);
5079 delayacct_blkio_end();
5080
5081 return ret;
5082}
5083EXPORT_SYMBOL(io_schedule_timeout);
5084
5085/**
5086 * sys_sched_get_priority_max - return maximum RT priority.
5087 * @policy: scheduling class.
5088 *
5089 * Return: On success, this syscall returns the maximum
5090 * rt_priority that can be used by a given scheduling class.
5091 * On failure, a negative error code is returned.
5092 */
5093SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5094{
5095 int ret = -EINVAL;
5096
5097 switch (policy) {
5098 case SCHED_FIFO:
5099 case SCHED_RR:
5100 ret = MAX_USER_RT_PRIO-1;
5101 break;
5102 case SCHED_DEADLINE:
5103 case SCHED_NORMAL:
5104 case SCHED_BATCH:
5105 case SCHED_IDLE:
5106 ret = 0;
5107 break;
5108 }
5109 return ret;
5110}
5111
5112/**
5113 * sys_sched_get_priority_min - return minimum RT priority.
5114 * @policy: scheduling class.
5115 *
5116 * Return: On success, this syscall returns the minimum
5117 * rt_priority that can be used by a given scheduling class.
5118 * On failure, a negative error code is returned.
5119 */
5120SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5121{
5122 int ret = -EINVAL;
5123
5124 switch (policy) {
5125 case SCHED_FIFO:
5126 case SCHED_RR:
5127 ret = 1;
5128 break;
5129 case SCHED_DEADLINE:
5130 case SCHED_NORMAL:
5131 case SCHED_BATCH:
5132 case SCHED_IDLE:
5133 ret = 0;
5134 }
5135 return ret;
5136}
5137
5138/**
5139 * sys_sched_rr_get_interval - return the default timeslice of a process.
5140 * @pid: pid of the process.
5141 * @interval: userspace pointer to the timeslice value.
5142 *
5143 * this syscall writes the default timeslice value of a given process
5144 * into the user-space timespec buffer. A value of '0' means infinity.
5145 *
5146 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5147 * an error code.
5148 */
5149SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5150 struct timespec __user *, interval)
5151{
5152 struct task_struct *p;
5153 unsigned int time_slice;
5154 struct rq_flags rf;
5155 struct timespec t;
5156 struct rq *rq;
5157 int retval;
5158
5159 if (pid < 0)
5160 return -EINVAL;
5161
5162 retval = -ESRCH;
5163 rcu_read_lock();
5164 p = find_process_by_pid(pid);
5165 if (!p)
5166 goto out_unlock;
5167
5168 retval = security_task_getscheduler(p);
5169 if (retval)
5170 goto out_unlock;
5171
5172 rq = task_rq_lock(p, &rf);
5173 time_slice = 0;
5174 if (p->sched_class->get_rr_interval)
5175 time_slice = p->sched_class->get_rr_interval(rq, p);
5176 task_rq_unlock(rq, p, &rf);
5177
5178 rcu_read_unlock();
5179 jiffies_to_timespec(time_slice, &t);
5180 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5181 return retval;
5182
5183out_unlock:
5184 rcu_read_unlock();
5185 return retval;
5186}
5187
5188static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5189
5190void sched_show_task(struct task_struct *p)
5191{
5192 unsigned long free = 0;
5193 int ppid;
5194 unsigned long state = p->state;
5195
5196 if (!try_get_task_stack(p))
5197 return;
5198 if (state)
5199 state = __ffs(state) + 1;
5200 printk(KERN_INFO "%-15.15s %c", p->comm,
5201 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5202 if (state == TASK_RUNNING)
5203 printk(KERN_CONT " running task ");
5204#ifdef CONFIG_DEBUG_STACK_USAGE
5205 free = stack_not_used(p);
5206#endif
5207 ppid = 0;
5208 rcu_read_lock();
5209 if (pid_alive(p))
5210 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5211 rcu_read_unlock();
5212 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5213 task_pid_nr(p), ppid,
5214 (unsigned long)task_thread_info(p)->flags);
5215
5216 print_worker_info(KERN_INFO, p);
5217 show_stack(p, NULL);
5218 put_task_stack(p);
5219}
5220
5221void show_state_filter(unsigned long state_filter)
5222{
5223 struct task_struct *g, *p;
5224
5225#if BITS_PER_LONG == 32
5226 printk(KERN_INFO
5227 " task PC stack pid father\n");
5228#else
5229 printk(KERN_INFO
5230 " task PC stack pid father\n");
5231#endif
5232 rcu_read_lock();
5233 for_each_process_thread(g, p) {
5234 /*
5235 * reset the NMI-timeout, listing all files on a slow
5236 * console might take a lot of time:
5237 * Also, reset softlockup watchdogs on all CPUs, because
5238 * another CPU might be blocked waiting for us to process
5239 * an IPI.
5240 */
5241 touch_nmi_watchdog();
5242 touch_all_softlockup_watchdogs();
5243 if (!state_filter || (p->state & state_filter))
5244 sched_show_task(p);
5245 }
5246
5247#ifdef CONFIG_SCHED_DEBUG
5248 if (!state_filter)
5249 sysrq_sched_debug_show();
5250#endif
5251 rcu_read_unlock();
5252 /*
5253 * Only show locks if all tasks are dumped:
5254 */
5255 if (!state_filter)
5256 debug_show_all_locks();
5257}
5258
5259void init_idle_bootup_task(struct task_struct *idle)
5260{
5261 idle->sched_class = &idle_sched_class;
5262}
5263
5264/**
5265 * init_idle - set up an idle thread for a given CPU
5266 * @idle: task in question
5267 * @cpu: cpu the idle task belongs to
5268 *
5269 * NOTE: this function does not set the idle thread's NEED_RESCHED
5270 * flag, to make booting more robust.
5271 */
5272void init_idle(struct task_struct *idle, int cpu)
5273{
5274 struct rq *rq = cpu_rq(cpu);
5275 unsigned long flags;
5276
5277 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5278 raw_spin_lock(&rq->lock);
5279
5280 __sched_fork(0, idle);
5281 idle->state = TASK_RUNNING;
5282 idle->se.exec_start = sched_clock();
5283 idle->flags |= PF_IDLE;
5284
5285 kasan_unpoison_task_stack(idle);
5286
5287#ifdef CONFIG_SMP
5288 /*
5289 * Its possible that init_idle() gets called multiple times on a task,
5290 * in that case do_set_cpus_allowed() will not do the right thing.
5291 *
5292 * And since this is boot we can forgo the serialization.
5293 */
5294 set_cpus_allowed_common(idle, cpumask_of(cpu));
5295#endif
5296 /*
5297 * We're having a chicken and egg problem, even though we are
5298 * holding rq->lock, the cpu isn't yet set to this cpu so the
5299 * lockdep check in task_group() will fail.
5300 *
5301 * Similar case to sched_fork(). / Alternatively we could
5302 * use task_rq_lock() here and obtain the other rq->lock.
5303 *
5304 * Silence PROVE_RCU
5305 */
5306 rcu_read_lock();
5307 __set_task_cpu(idle, cpu);
5308 rcu_read_unlock();
5309
5310 rq->curr = rq->idle = idle;
5311 idle->on_rq = TASK_ON_RQ_QUEUED;
5312#ifdef CONFIG_SMP
5313 idle->on_cpu = 1;
5314#endif
5315 raw_spin_unlock(&rq->lock);
5316 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5317
5318 /* Set the preempt count _outside_ the spinlocks! */
5319 init_idle_preempt_count(idle, cpu);
5320
5321 /*
5322 * The idle tasks have their own, simple scheduling class:
5323 */
5324 idle->sched_class = &idle_sched_class;
5325 ftrace_graph_init_idle_task(idle, cpu);
5326 vtime_init_idle(idle, cpu);
5327#ifdef CONFIG_SMP
5328 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5329#endif
5330}
5331
5332int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5333 const struct cpumask *trial)
5334{
5335 int ret = 1, trial_cpus;
5336 struct dl_bw *cur_dl_b;
5337 unsigned long flags;
5338
5339 if (!cpumask_weight(cur))
5340 return ret;
5341
5342 rcu_read_lock_sched();
5343 cur_dl_b = dl_bw_of(cpumask_any(cur));
5344 trial_cpus = cpumask_weight(trial);
5345
5346 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5347 if (cur_dl_b->bw != -1 &&
5348 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5349 ret = 0;
5350 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5351 rcu_read_unlock_sched();
5352
5353 return ret;
5354}
5355
5356int task_can_attach(struct task_struct *p,
5357 const struct cpumask *cs_cpus_allowed)
5358{
5359 int ret = 0;
5360
5361 /*
5362 * Kthreads which disallow setaffinity shouldn't be moved
5363 * to a new cpuset; we don't want to change their cpu
5364 * affinity and isolating such threads by their set of
5365 * allowed nodes is unnecessary. Thus, cpusets are not
5366 * applicable for such threads. This prevents checking for
5367 * success of set_cpus_allowed_ptr() on all attached tasks
5368 * before cpus_allowed may be changed.
5369 */
5370 if (p->flags & PF_NO_SETAFFINITY) {
5371 ret = -EINVAL;
5372 goto out;
5373 }
5374
5375#ifdef CONFIG_SMP
5376 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5377 cs_cpus_allowed)) {
5378 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5379 cs_cpus_allowed);
5380 struct dl_bw *dl_b;
5381 bool overflow;
5382 int cpus;
5383 unsigned long flags;
5384
5385 rcu_read_lock_sched();
5386 dl_b = dl_bw_of(dest_cpu);
5387 raw_spin_lock_irqsave(&dl_b->lock, flags);
5388 cpus = dl_bw_cpus(dest_cpu);
5389 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5390 if (overflow)
5391 ret = -EBUSY;
5392 else {
5393 /*
5394 * We reserve space for this task in the destination
5395 * root_domain, as we can't fail after this point.
5396 * We will free resources in the source root_domain
5397 * later on (see set_cpus_allowed_dl()).
5398 */
5399 __dl_add(dl_b, p->dl.dl_bw);
5400 }
5401 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5402 rcu_read_unlock_sched();
5403
5404 }
5405#endif
5406out:
5407 return ret;
5408}
5409
5410#ifdef CONFIG_SMP
5411
5412static bool sched_smp_initialized __read_mostly;
5413
5414#ifdef CONFIG_NUMA_BALANCING
5415/* Migrate current task p to target_cpu */
5416int migrate_task_to(struct task_struct *p, int target_cpu)
5417{
5418 struct migration_arg arg = { p, target_cpu };
5419 int curr_cpu = task_cpu(p);
5420
5421 if (curr_cpu == target_cpu)
5422 return 0;
5423
5424 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5425 return -EINVAL;
5426
5427 /* TODO: This is not properly updating schedstats */
5428
5429 trace_sched_move_numa(p, curr_cpu, target_cpu);
5430 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5431}
5432
5433/*
5434 * Requeue a task on a given node and accurately track the number of NUMA
5435 * tasks on the runqueues
5436 */
5437void sched_setnuma(struct task_struct *p, int nid)
5438{
5439 bool queued, running;
5440 struct rq_flags rf;
5441 struct rq *rq;
5442
5443 rq = task_rq_lock(p, &rf);
5444 queued = task_on_rq_queued(p);
5445 running = task_current(rq, p);
5446
5447 if (queued)
5448 dequeue_task(rq, p, DEQUEUE_SAVE);
5449 if (running)
5450 put_prev_task(rq, p);
5451
5452 p->numa_preferred_nid = nid;
5453
5454 if (queued)
5455 enqueue_task(rq, p, ENQUEUE_RESTORE);
5456 if (running)
5457 set_curr_task(rq, p);
5458 task_rq_unlock(rq, p, &rf);
5459}
5460#endif /* CONFIG_NUMA_BALANCING */
5461
5462#ifdef CONFIG_HOTPLUG_CPU
5463/*
5464 * Ensures that the idle task is using init_mm right before its cpu goes
5465 * offline.
5466 */
5467void idle_task_exit(void)
5468{
5469 struct mm_struct *mm = current->active_mm;
5470
5471 BUG_ON(cpu_online(smp_processor_id()));
5472
5473 if (mm != &init_mm) {
5474 switch_mm_irqs_off(mm, &init_mm, current);
5475 finish_arch_post_lock_switch();
5476 }
5477 mmdrop(mm);
5478}
5479
5480/*
5481 * Since this CPU is going 'away' for a while, fold any nr_active delta
5482 * we might have. Assumes we're called after migrate_tasks() so that the
5483 * nr_active count is stable. We need to take the teardown thread which
5484 * is calling this into account, so we hand in adjust = 1 to the load
5485 * calculation.
5486 *
5487 * Also see the comment "Global load-average calculations".
5488 */
5489static void calc_load_migrate(struct rq *rq)
5490{
5491 long delta = calc_load_fold_active(rq, 1);
5492 if (delta)
5493 atomic_long_add(delta, &calc_load_tasks);
5494}
5495
5496static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5497{
5498}
5499
5500static const struct sched_class fake_sched_class = {
5501 .put_prev_task = put_prev_task_fake,
5502};
5503
5504static struct task_struct fake_task = {
5505 /*
5506 * Avoid pull_{rt,dl}_task()
5507 */
5508 .prio = MAX_PRIO + 1,
5509 .sched_class = &fake_sched_class,
5510};
5511
5512/*
5513 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5514 * try_to_wake_up()->select_task_rq().
5515 *
5516 * Called with rq->lock held even though we'er in stop_machine() and
5517 * there's no concurrency possible, we hold the required locks anyway
5518 * because of lock validation efforts.
5519 */
5520static void migrate_tasks(struct rq *dead_rq)
5521{
5522 struct rq *rq = dead_rq;
5523 struct task_struct *next, *stop = rq->stop;
5524 struct pin_cookie cookie;
5525 int dest_cpu;
5526
5527 /*
5528 * Fudge the rq selection such that the below task selection loop
5529 * doesn't get stuck on the currently eligible stop task.
5530 *
5531 * We're currently inside stop_machine() and the rq is either stuck
5532 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5533 * either way we should never end up calling schedule() until we're
5534 * done here.
5535 */
5536 rq->stop = NULL;
5537
5538 /*
5539 * put_prev_task() and pick_next_task() sched
5540 * class method both need to have an up-to-date
5541 * value of rq->clock[_task]
5542 */
5543 update_rq_clock(rq);
5544
5545 for (;;) {
5546 /*
5547 * There's this thread running, bail when that's the only
5548 * remaining thread.
5549 */
5550 if (rq->nr_running == 1)
5551 break;
5552
5553 /*
5554 * pick_next_task assumes pinned rq->lock.
5555 */
5556 cookie = lockdep_pin_lock(&rq->lock);
5557 next = pick_next_task(rq, &fake_task, cookie);
5558 BUG_ON(!next);
5559 next->sched_class->put_prev_task(rq, next);
5560
5561 /*
5562 * Rules for changing task_struct::cpus_allowed are holding
5563 * both pi_lock and rq->lock, such that holding either
5564 * stabilizes the mask.
5565 *
5566 * Drop rq->lock is not quite as disastrous as it usually is
5567 * because !cpu_active at this point, which means load-balance
5568 * will not interfere. Also, stop-machine.
5569 */
5570 lockdep_unpin_lock(&rq->lock, cookie);
5571 raw_spin_unlock(&rq->lock);
5572 raw_spin_lock(&next->pi_lock);
5573 raw_spin_lock(&rq->lock);
5574
5575 /*
5576 * Since we're inside stop-machine, _nothing_ should have
5577 * changed the task, WARN if weird stuff happened, because in
5578 * that case the above rq->lock drop is a fail too.
5579 */
5580 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5581 raw_spin_unlock(&next->pi_lock);
5582 continue;
5583 }
5584
5585 /* Find suitable destination for @next, with force if needed. */
5586 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5587
5588 rq = __migrate_task(rq, next, dest_cpu);
5589 if (rq != dead_rq) {
5590 raw_spin_unlock(&rq->lock);
5591 rq = dead_rq;
5592 raw_spin_lock(&rq->lock);
5593 }
5594 raw_spin_unlock(&next->pi_lock);
5595 }
5596
5597 rq->stop = stop;
5598}
5599#endif /* CONFIG_HOTPLUG_CPU */
5600
5601static void set_rq_online(struct rq *rq)
5602{
5603 if (!rq->online) {
5604 const struct sched_class *class;
5605
5606 cpumask_set_cpu(rq->cpu, rq->rd->online);
5607 rq->online = 1;
5608
5609 for_each_class(class) {
5610 if (class->rq_online)
5611 class->rq_online(rq);
5612 }
5613 }
5614}
5615
5616static void set_rq_offline(struct rq *rq)
5617{
5618 if (rq->online) {
5619 const struct sched_class *class;
5620
5621 for_each_class(class) {
5622 if (class->rq_offline)
5623 class->rq_offline(rq);
5624 }
5625
5626 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5627 rq->online = 0;
5628 }
5629}
5630
5631static void set_cpu_rq_start_time(unsigned int cpu)
5632{
5633 struct rq *rq = cpu_rq(cpu);
5634
5635 rq->age_stamp = sched_clock_cpu(cpu);
5636}
5637
5638static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5639
5640#ifdef CONFIG_SCHED_DEBUG
5641
5642static __read_mostly int sched_debug_enabled;
5643
5644static int __init sched_debug_setup(char *str)
5645{
5646 sched_debug_enabled = 1;
5647
5648 return 0;
5649}
5650early_param("sched_debug", sched_debug_setup);
5651
5652static inline bool sched_debug(void)
5653{
5654 return sched_debug_enabled;
5655}
5656
5657static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5658 struct cpumask *groupmask)
5659{
5660 struct sched_group *group = sd->groups;
5661
5662 cpumask_clear(groupmask);
5663
5664 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5665
5666 if (!(sd->flags & SD_LOAD_BALANCE)) {
5667 printk("does not load-balance\n");
5668 if (sd->parent)
5669 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5670 " has parent");
5671 return -1;
5672 }
5673
5674 printk(KERN_CONT "span %*pbl level %s\n",
5675 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5676
5677 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5678 printk(KERN_ERR "ERROR: domain->span does not contain "
5679 "CPU%d\n", cpu);
5680 }
5681 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5682 printk(KERN_ERR "ERROR: domain->groups does not contain"
5683 " CPU%d\n", cpu);
5684 }
5685
5686 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5687 do {
5688 if (!group) {
5689 printk("\n");
5690 printk(KERN_ERR "ERROR: group is NULL\n");
5691 break;
5692 }
5693
5694 if (!cpumask_weight(sched_group_cpus(group))) {
5695 printk(KERN_CONT "\n");
5696 printk(KERN_ERR "ERROR: empty group\n");
5697 break;
5698 }
5699
5700 if (!(sd->flags & SD_OVERLAP) &&
5701 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5702 printk(KERN_CONT "\n");
5703 printk(KERN_ERR "ERROR: repeated CPUs\n");
5704 break;
5705 }
5706
5707 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5708
5709 printk(KERN_CONT " %*pbl",
5710 cpumask_pr_args(sched_group_cpus(group)));
5711 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5712 printk(KERN_CONT " (cpu_capacity = %lu)",
5713 group->sgc->capacity);
5714 }
5715
5716 group = group->next;
5717 } while (group != sd->groups);
5718 printk(KERN_CONT "\n");
5719
5720 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5721 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5722
5723 if (sd->parent &&
5724 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5725 printk(KERN_ERR "ERROR: parent span is not a superset "
5726 "of domain->span\n");
5727 return 0;
5728}
5729
5730static void sched_domain_debug(struct sched_domain *sd, int cpu)
5731{
5732 int level = 0;
5733
5734 if (!sched_debug_enabled)
5735 return;
5736
5737 if (!sd) {
5738 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5739 return;
5740 }
5741
5742 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5743
5744 for (;;) {
5745 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5746 break;
5747 level++;
5748 sd = sd->parent;
5749 if (!sd)
5750 break;
5751 }
5752}
5753#else /* !CONFIG_SCHED_DEBUG */
5754
5755# define sched_debug_enabled 0
5756# define sched_domain_debug(sd, cpu) do { } while (0)
5757static inline bool sched_debug(void)
5758{
5759 return false;
5760}
5761#endif /* CONFIG_SCHED_DEBUG */
5762
5763static int sd_degenerate(struct sched_domain *sd)
5764{
5765 if (cpumask_weight(sched_domain_span(sd)) == 1)
5766 return 1;
5767
5768 /* Following flags need at least 2 groups */
5769 if (sd->flags & (SD_LOAD_BALANCE |
5770 SD_BALANCE_NEWIDLE |
5771 SD_BALANCE_FORK |
5772 SD_BALANCE_EXEC |
5773 SD_SHARE_CPUCAPACITY |
5774 SD_ASYM_CPUCAPACITY |
5775 SD_SHARE_PKG_RESOURCES |
5776 SD_SHARE_POWERDOMAIN)) {
5777 if (sd->groups != sd->groups->next)
5778 return 0;
5779 }
5780
5781 /* Following flags don't use groups */
5782 if (sd->flags & (SD_WAKE_AFFINE))
5783 return 0;
5784
5785 return 1;
5786}
5787
5788static int
5789sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5790{
5791 unsigned long cflags = sd->flags, pflags = parent->flags;
5792
5793 if (sd_degenerate(parent))
5794 return 1;
5795
5796 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5797 return 0;
5798
5799 /* Flags needing groups don't count if only 1 group in parent */
5800 if (parent->groups == parent->groups->next) {
5801 pflags &= ~(SD_LOAD_BALANCE |
5802 SD_BALANCE_NEWIDLE |
5803 SD_BALANCE_FORK |
5804 SD_BALANCE_EXEC |
5805 SD_ASYM_CPUCAPACITY |
5806 SD_SHARE_CPUCAPACITY |
5807 SD_SHARE_PKG_RESOURCES |
5808 SD_PREFER_SIBLING |
5809 SD_SHARE_POWERDOMAIN);
5810 if (nr_node_ids == 1)
5811 pflags &= ~SD_SERIALIZE;
5812 }
5813 if (~cflags & pflags)
5814 return 0;
5815
5816 return 1;
5817}
5818
5819static void free_rootdomain(struct rcu_head *rcu)
5820{
5821 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5822
5823 cpupri_cleanup(&rd->cpupri);
5824 cpudl_cleanup(&rd->cpudl);
5825 free_cpumask_var(rd->dlo_mask);
5826 free_cpumask_var(rd->rto_mask);
5827 free_cpumask_var(rd->online);
5828 free_cpumask_var(rd->span);
5829 kfree(rd);
5830}
5831
5832static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5833{
5834 struct root_domain *old_rd = NULL;
5835 unsigned long flags;
5836
5837 raw_spin_lock_irqsave(&rq->lock, flags);
5838
5839 if (rq->rd) {
5840 old_rd = rq->rd;
5841
5842 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5843 set_rq_offline(rq);
5844
5845 cpumask_clear_cpu(rq->cpu, old_rd->span);
5846
5847 /*
5848 * If we dont want to free the old_rd yet then
5849 * set old_rd to NULL to skip the freeing later
5850 * in this function:
5851 */
5852 if (!atomic_dec_and_test(&old_rd->refcount))
5853 old_rd = NULL;
5854 }
5855
5856 atomic_inc(&rd->refcount);
5857 rq->rd = rd;
5858
5859 cpumask_set_cpu(rq->cpu, rd->span);
5860 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5861 set_rq_online(rq);
5862
5863 raw_spin_unlock_irqrestore(&rq->lock, flags);
5864
5865 if (old_rd)
5866 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5867}
5868
5869static int init_rootdomain(struct root_domain *rd)
5870{
5871 memset(rd, 0, sizeof(*rd));
5872
5873 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5874 goto out;
5875 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5876 goto free_span;
5877 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5878 goto free_online;
5879 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5880 goto free_dlo_mask;
5881
5882 init_dl_bw(&rd->dl_bw);
5883 if (cpudl_init(&rd->cpudl) != 0)
5884 goto free_dlo_mask;
5885
5886 if (cpupri_init(&rd->cpupri) != 0)
5887 goto free_rto_mask;
5888 return 0;
5889
5890free_rto_mask:
5891 free_cpumask_var(rd->rto_mask);
5892free_dlo_mask:
5893 free_cpumask_var(rd->dlo_mask);
5894free_online:
5895 free_cpumask_var(rd->online);
5896free_span:
5897 free_cpumask_var(rd->span);
5898out:
5899 return -ENOMEM;
5900}
5901
5902/*
5903 * By default the system creates a single root-domain with all cpus as
5904 * members (mimicking the global state we have today).
5905 */
5906struct root_domain def_root_domain;
5907
5908static void init_defrootdomain(void)
5909{
5910 init_rootdomain(&def_root_domain);
5911
5912 atomic_set(&def_root_domain.refcount, 1);
5913}
5914
5915static struct root_domain *alloc_rootdomain(void)
5916{
5917 struct root_domain *rd;
5918
5919 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5920 if (!rd)
5921 return NULL;
5922
5923 if (init_rootdomain(rd) != 0) {
5924 kfree(rd);
5925 return NULL;
5926 }
5927
5928 return rd;
5929}
5930
5931static void free_sched_groups(struct sched_group *sg, int free_sgc)
5932{
5933 struct sched_group *tmp, *first;
5934
5935 if (!sg)
5936 return;
5937
5938 first = sg;
5939 do {
5940 tmp = sg->next;
5941
5942 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5943 kfree(sg->sgc);
5944
5945 kfree(sg);
5946 sg = tmp;
5947 } while (sg != first);
5948}
5949
5950static void destroy_sched_domain(struct sched_domain *sd)
5951{
5952 /*
5953 * If its an overlapping domain it has private groups, iterate and
5954 * nuke them all.
5955 */
5956 if (sd->flags & SD_OVERLAP) {
5957 free_sched_groups(sd->groups, 1);
5958 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5959 kfree(sd->groups->sgc);
5960 kfree(sd->groups);
5961 }
5962 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
5963 kfree(sd->shared);
5964 kfree(sd);
5965}
5966
5967static void destroy_sched_domains_rcu(struct rcu_head *rcu)
5968{
5969 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5970
5971 while (sd) {
5972 struct sched_domain *parent = sd->parent;
5973 destroy_sched_domain(sd);
5974 sd = parent;
5975 }
5976}
5977
5978static void destroy_sched_domains(struct sched_domain *sd)
5979{
5980 if (sd)
5981 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
5982}
5983
5984/*
5985 * Keep a special pointer to the highest sched_domain that has
5986 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5987 * allows us to avoid some pointer chasing select_idle_sibling().
5988 *
5989 * Also keep a unique ID per domain (we use the first cpu number in
5990 * the cpumask of the domain), this allows us to quickly tell if
5991 * two cpus are in the same cache domain, see cpus_share_cache().
5992 */
5993DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5994DEFINE_PER_CPU(int, sd_llc_size);
5995DEFINE_PER_CPU(int, sd_llc_id);
5996DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
5997DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5998DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5999
6000static void update_top_cache_domain(int cpu)
6001{
6002 struct sched_domain_shared *sds = NULL;
6003 struct sched_domain *sd;
6004 int id = cpu;
6005 int size = 1;
6006
6007 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6008 if (sd) {
6009 id = cpumask_first(sched_domain_span(sd));
6010 size = cpumask_weight(sched_domain_span(sd));
6011 sds = sd->shared;
6012 }
6013
6014 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6015 per_cpu(sd_llc_size, cpu) = size;
6016 per_cpu(sd_llc_id, cpu) = id;
6017 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
6018
6019 sd = lowest_flag_domain(cpu, SD_NUMA);
6020 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6021
6022 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6023 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6024}
6025
6026/*
6027 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6028 * hold the hotplug lock.
6029 */
6030static void
6031cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6032{
6033 struct rq *rq = cpu_rq(cpu);
6034 struct sched_domain *tmp;
6035
6036 /* Remove the sched domains which do not contribute to scheduling. */
6037 for (tmp = sd; tmp; ) {
6038 struct sched_domain *parent = tmp->parent;
6039 if (!parent)
6040 break;
6041
6042 if (sd_parent_degenerate(tmp, parent)) {
6043 tmp->parent = parent->parent;
6044 if (parent->parent)
6045 parent->parent->child = tmp;
6046 /*
6047 * Transfer SD_PREFER_SIBLING down in case of a
6048 * degenerate parent; the spans match for this
6049 * so the property transfers.
6050 */
6051 if (parent->flags & SD_PREFER_SIBLING)
6052 tmp->flags |= SD_PREFER_SIBLING;
6053 destroy_sched_domain(parent);
6054 } else
6055 tmp = tmp->parent;
6056 }
6057
6058 if (sd && sd_degenerate(sd)) {
6059 tmp = sd;
6060 sd = sd->parent;
6061 destroy_sched_domain(tmp);
6062 if (sd)
6063 sd->child = NULL;
6064 }
6065
6066 sched_domain_debug(sd, cpu);
6067
6068 rq_attach_root(rq, rd);
6069 tmp = rq->sd;
6070 rcu_assign_pointer(rq->sd, sd);
6071 destroy_sched_domains(tmp);
6072
6073 update_top_cache_domain(cpu);
6074}
6075
6076/* Setup the mask of cpus configured for isolated domains */
6077static int __init isolated_cpu_setup(char *str)
6078{
6079 int ret;
6080
6081 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6082 ret = cpulist_parse(str, cpu_isolated_map);
6083 if (ret) {
6084 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
6085 return 0;
6086 }
6087 return 1;
6088}
6089__setup("isolcpus=", isolated_cpu_setup);
6090
6091struct s_data {
6092 struct sched_domain ** __percpu sd;
6093 struct root_domain *rd;
6094};
6095
6096enum s_alloc {
6097 sa_rootdomain,
6098 sa_sd,
6099 sa_sd_storage,
6100 sa_none,
6101};
6102
6103/*
6104 * Build an iteration mask that can exclude certain CPUs from the upwards
6105 * domain traversal.
6106 *
6107 * Asymmetric node setups can result in situations where the domain tree is of
6108 * unequal depth, make sure to skip domains that already cover the entire
6109 * range.
6110 *
6111 * In that case build_sched_domains() will have terminated the iteration early
6112 * and our sibling sd spans will be empty. Domains should always include the
6113 * cpu they're built on, so check that.
6114 *
6115 */
6116static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6117{
6118 const struct cpumask *span = sched_domain_span(sd);
6119 struct sd_data *sdd = sd->private;
6120 struct sched_domain *sibling;
6121 int i;
6122
6123 for_each_cpu(i, span) {
6124 sibling = *per_cpu_ptr(sdd->sd, i);
6125 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6126 continue;
6127
6128 cpumask_set_cpu(i, sched_group_mask(sg));
6129 }
6130}
6131
6132/*
6133 * Return the canonical balance cpu for this group, this is the first cpu
6134 * of this group that's also in the iteration mask.
6135 */
6136int group_balance_cpu(struct sched_group *sg)
6137{
6138 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6139}
6140
6141static int
6142build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6143{
6144 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6145 const struct cpumask *span = sched_domain_span(sd);
6146 struct cpumask *covered = sched_domains_tmpmask;
6147 struct sd_data *sdd = sd->private;
6148 struct sched_domain *sibling;
6149 int i;
6150
6151 cpumask_clear(covered);
6152
6153 for_each_cpu(i, span) {
6154 struct cpumask *sg_span;
6155
6156 if (cpumask_test_cpu(i, covered))
6157 continue;
6158
6159 sibling = *per_cpu_ptr(sdd->sd, i);
6160
6161 /* See the comment near build_group_mask(). */
6162 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6163 continue;
6164
6165 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6166 GFP_KERNEL, cpu_to_node(cpu));
6167
6168 if (!sg)
6169 goto fail;
6170
6171 sg_span = sched_group_cpus(sg);
6172 if (sibling->child)
6173 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6174 else
6175 cpumask_set_cpu(i, sg_span);
6176
6177 cpumask_or(covered, covered, sg_span);
6178
6179 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6180 if (atomic_inc_return(&sg->sgc->ref) == 1)
6181 build_group_mask(sd, sg);
6182
6183 /*
6184 * Initialize sgc->capacity such that even if we mess up the
6185 * domains and no possible iteration will get us here, we won't
6186 * die on a /0 trap.
6187 */
6188 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6189 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
6190
6191 /*
6192 * Make sure the first group of this domain contains the
6193 * canonical balance cpu. Otherwise the sched_domain iteration
6194 * breaks. See update_sg_lb_stats().
6195 */
6196 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6197 group_balance_cpu(sg) == cpu)
6198 groups = sg;
6199
6200 if (!first)
6201 first = sg;
6202 if (last)
6203 last->next = sg;
6204 last = sg;
6205 last->next = first;
6206 }
6207 sd->groups = groups;
6208
6209 return 0;
6210
6211fail:
6212 free_sched_groups(first, 0);
6213
6214 return -ENOMEM;
6215}
6216
6217static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6218{
6219 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6220 struct sched_domain *child = sd->child;
6221
6222 if (child)
6223 cpu = cpumask_first(sched_domain_span(child));
6224
6225 if (sg) {
6226 *sg = *per_cpu_ptr(sdd->sg, cpu);
6227 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6228 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6229 }
6230
6231 return cpu;
6232}
6233
6234/*
6235 * build_sched_groups will build a circular linked list of the groups
6236 * covered by the given span, and will set each group's ->cpumask correctly,
6237 * and ->cpu_capacity to 0.
6238 *
6239 * Assumes the sched_domain tree is fully constructed
6240 */
6241static int
6242build_sched_groups(struct sched_domain *sd, int cpu)
6243{
6244 struct sched_group *first = NULL, *last = NULL;
6245 struct sd_data *sdd = sd->private;
6246 const struct cpumask *span = sched_domain_span(sd);
6247 struct cpumask *covered;
6248 int i;
6249
6250 get_group(cpu, sdd, &sd->groups);
6251 atomic_inc(&sd->groups->ref);
6252
6253 if (cpu != cpumask_first(span))
6254 return 0;
6255
6256 lockdep_assert_held(&sched_domains_mutex);
6257 covered = sched_domains_tmpmask;
6258
6259 cpumask_clear(covered);
6260
6261 for_each_cpu(i, span) {
6262 struct sched_group *sg;
6263 int group, j;
6264
6265 if (cpumask_test_cpu(i, covered))
6266 continue;
6267
6268 group = get_group(i, sdd, &sg);
6269 cpumask_setall(sched_group_mask(sg));
6270
6271 for_each_cpu(j, span) {
6272 if (get_group(j, sdd, NULL) != group)
6273 continue;
6274
6275 cpumask_set_cpu(j, covered);
6276 cpumask_set_cpu(j, sched_group_cpus(sg));
6277 }
6278
6279 if (!first)
6280 first = sg;
6281 if (last)
6282 last->next = sg;
6283 last = sg;
6284 }
6285 last->next = first;
6286
6287 return 0;
6288}
6289
6290/*
6291 * Initialize sched groups cpu_capacity.
6292 *
6293 * cpu_capacity indicates the capacity of sched group, which is used while
6294 * distributing the load between different sched groups in a sched domain.
6295 * Typically cpu_capacity for all the groups in a sched domain will be same
6296 * unless there are asymmetries in the topology. If there are asymmetries,
6297 * group having more cpu_capacity will pickup more load compared to the
6298 * group having less cpu_capacity.
6299 */
6300static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6301{
6302 struct sched_group *sg = sd->groups;
6303
6304 WARN_ON(!sg);
6305
6306 do {
6307 int cpu, max_cpu = -1;
6308
6309 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6310
6311 if (!(sd->flags & SD_ASYM_PACKING))
6312 goto next;
6313
6314 for_each_cpu(cpu, sched_group_cpus(sg)) {
6315 if (max_cpu < 0)
6316 max_cpu = cpu;
6317 else if (sched_asym_prefer(cpu, max_cpu))
6318 max_cpu = cpu;
6319 }
6320 sg->asym_prefer_cpu = max_cpu;
6321
6322next:
6323 sg = sg->next;
6324 } while (sg != sd->groups);
6325
6326 if (cpu != group_balance_cpu(sg))
6327 return;
6328
6329 update_group_capacity(sd, cpu);
6330}
6331
6332/*
6333 * Initializers for schedule domains
6334 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6335 */
6336
6337static int default_relax_domain_level = -1;
6338int sched_domain_level_max;
6339
6340static int __init setup_relax_domain_level(char *str)
6341{
6342 if (kstrtoint(str, 0, &default_relax_domain_level))
6343 pr_warn("Unable to set relax_domain_level\n");
6344
6345 return 1;
6346}
6347__setup("relax_domain_level=", setup_relax_domain_level);
6348
6349static void set_domain_attribute(struct sched_domain *sd,
6350 struct sched_domain_attr *attr)
6351{
6352 int request;
6353
6354 if (!attr || attr->relax_domain_level < 0) {
6355 if (default_relax_domain_level < 0)
6356 return;
6357 else
6358 request = default_relax_domain_level;
6359 } else
6360 request = attr->relax_domain_level;
6361 if (request < sd->level) {
6362 /* turn off idle balance on this domain */
6363 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6364 } else {
6365 /* turn on idle balance on this domain */
6366 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6367 }
6368}
6369
6370static void __sdt_free(const struct cpumask *cpu_map);
6371static int __sdt_alloc(const struct cpumask *cpu_map);
6372
6373static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6374 const struct cpumask *cpu_map)
6375{
6376 switch (what) {
6377 case sa_rootdomain:
6378 if (!atomic_read(&d->rd->refcount))
6379 free_rootdomain(&d->rd->rcu); /* fall through */
6380 case sa_sd:
6381 free_percpu(d->sd); /* fall through */
6382 case sa_sd_storage:
6383 __sdt_free(cpu_map); /* fall through */
6384 case sa_none:
6385 break;
6386 }
6387}
6388
6389static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6390 const struct cpumask *cpu_map)
6391{
6392 memset(d, 0, sizeof(*d));
6393
6394 if (__sdt_alloc(cpu_map))
6395 return sa_sd_storage;
6396 d->sd = alloc_percpu(struct sched_domain *);
6397 if (!d->sd)
6398 return sa_sd_storage;
6399 d->rd = alloc_rootdomain();
6400 if (!d->rd)
6401 return sa_sd;
6402 return sa_rootdomain;
6403}
6404
6405/*
6406 * NULL the sd_data elements we've used to build the sched_domain and
6407 * sched_group structure so that the subsequent __free_domain_allocs()
6408 * will not free the data we're using.
6409 */
6410static void claim_allocations(int cpu, struct sched_domain *sd)
6411{
6412 struct sd_data *sdd = sd->private;
6413
6414 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6415 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6416
6417 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
6418 *per_cpu_ptr(sdd->sds, cpu) = NULL;
6419
6420 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6421 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6422
6423 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6424 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6425}
6426
6427#ifdef CONFIG_NUMA
6428static int sched_domains_numa_levels;
6429enum numa_topology_type sched_numa_topology_type;
6430static int *sched_domains_numa_distance;
6431int sched_max_numa_distance;
6432static struct cpumask ***sched_domains_numa_masks;
6433static int sched_domains_curr_level;
6434#endif
6435
6436/*
6437 * SD_flags allowed in topology descriptions.
6438 *
6439 * These flags are purely descriptive of the topology and do not prescribe
6440 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6441 * function:
6442 *
6443 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6444 * SD_SHARE_PKG_RESOURCES - describes shared caches
6445 * SD_NUMA - describes NUMA topologies
6446 * SD_SHARE_POWERDOMAIN - describes shared power domain
6447 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6448 *
6449 * Odd one out, which beside describing the topology has a quirk also
6450 * prescribes the desired behaviour that goes along with it:
6451 *
6452 * SD_ASYM_PACKING - describes SMT quirks
6453 */
6454#define TOPOLOGY_SD_FLAGS \
6455 (SD_SHARE_CPUCAPACITY | \
6456 SD_SHARE_PKG_RESOURCES | \
6457 SD_NUMA | \
6458 SD_ASYM_PACKING | \
6459 SD_ASYM_CPUCAPACITY | \
6460 SD_SHARE_POWERDOMAIN)
6461
6462static struct sched_domain *
6463sd_init(struct sched_domain_topology_level *tl,
6464 const struct cpumask *cpu_map,
6465 struct sched_domain *child, int cpu)
6466{
6467 struct sd_data *sdd = &tl->data;
6468 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6469 int sd_id, sd_weight, sd_flags = 0;
6470
6471#ifdef CONFIG_NUMA
6472 /*
6473 * Ugly hack to pass state to sd_numa_mask()...
6474 */
6475 sched_domains_curr_level = tl->numa_level;
6476#endif
6477
6478 sd_weight = cpumask_weight(tl->mask(cpu));
6479
6480 if (tl->sd_flags)
6481 sd_flags = (*tl->sd_flags)();
6482 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6483 "wrong sd_flags in topology description\n"))
6484 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6485
6486 *sd = (struct sched_domain){
6487 .min_interval = sd_weight,
6488 .max_interval = 2*sd_weight,
6489 .busy_factor = 32,
6490 .imbalance_pct = 125,
6491
6492 .cache_nice_tries = 0,
6493 .busy_idx = 0,
6494 .idle_idx = 0,
6495 .newidle_idx = 0,
6496 .wake_idx = 0,
6497 .forkexec_idx = 0,
6498
6499 .flags = 1*SD_LOAD_BALANCE
6500 | 1*SD_BALANCE_NEWIDLE
6501 | 1*SD_BALANCE_EXEC
6502 | 1*SD_BALANCE_FORK
6503 | 0*SD_BALANCE_WAKE
6504 | 1*SD_WAKE_AFFINE
6505 | 0*SD_SHARE_CPUCAPACITY
6506 | 0*SD_SHARE_PKG_RESOURCES
6507 | 0*SD_SERIALIZE
6508 | 0*SD_PREFER_SIBLING
6509 | 0*SD_NUMA
6510 | sd_flags
6511 ,
6512
6513 .last_balance = jiffies,
6514 .balance_interval = sd_weight,
6515 .smt_gain = 0,
6516 .max_newidle_lb_cost = 0,
6517 .next_decay_max_lb_cost = jiffies,
6518 .child = child,
6519#ifdef CONFIG_SCHED_DEBUG
6520 .name = tl->name,
6521#endif
6522 };
6523
6524 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6525 sd_id = cpumask_first(sched_domain_span(sd));
6526
6527 /*
6528 * Convert topological properties into behaviour.
6529 */
6530
6531 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6532 struct sched_domain *t = sd;
6533
6534 for_each_lower_domain(t)
6535 t->flags |= SD_BALANCE_WAKE;
6536 }
6537
6538 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6539 sd->flags |= SD_PREFER_SIBLING;
6540 sd->imbalance_pct = 110;
6541 sd->smt_gain = 1178; /* ~15% */
6542
6543 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6544 sd->imbalance_pct = 117;
6545 sd->cache_nice_tries = 1;
6546 sd->busy_idx = 2;
6547
6548#ifdef CONFIG_NUMA
6549 } else if (sd->flags & SD_NUMA) {
6550 sd->cache_nice_tries = 2;
6551 sd->busy_idx = 3;
6552 sd->idle_idx = 2;
6553
6554 sd->flags |= SD_SERIALIZE;
6555 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6556 sd->flags &= ~(SD_BALANCE_EXEC |
6557 SD_BALANCE_FORK |
6558 SD_WAKE_AFFINE);
6559 }
6560
6561#endif
6562 } else {
6563 sd->flags |= SD_PREFER_SIBLING;
6564 sd->cache_nice_tries = 1;
6565 sd->busy_idx = 2;
6566 sd->idle_idx = 1;
6567 }
6568
6569 /*
6570 * For all levels sharing cache; connect a sched_domain_shared
6571 * instance.
6572 */
6573 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6574 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
6575 atomic_inc(&sd->shared->ref);
6576 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
6577 }
6578
6579 sd->private = sdd;
6580
6581 return sd;
6582}
6583
6584/*
6585 * Topology list, bottom-up.
6586 */
6587static struct sched_domain_topology_level default_topology[] = {
6588#ifdef CONFIG_SCHED_SMT
6589 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6590#endif
6591#ifdef CONFIG_SCHED_MC
6592 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6593#endif
6594 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6595 { NULL, },
6596};
6597
6598static struct sched_domain_topology_level *sched_domain_topology =
6599 default_topology;
6600
6601#define for_each_sd_topology(tl) \
6602 for (tl = sched_domain_topology; tl->mask; tl++)
6603
6604void set_sched_topology(struct sched_domain_topology_level *tl)
6605{
6606 if (WARN_ON_ONCE(sched_smp_initialized))
6607 return;
6608
6609 sched_domain_topology = tl;
6610}
6611
6612#ifdef CONFIG_NUMA
6613
6614static const struct cpumask *sd_numa_mask(int cpu)
6615{
6616 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6617}
6618
6619static void sched_numa_warn(const char *str)
6620{
6621 static int done = false;
6622 int i,j;
6623
6624 if (done)
6625 return;
6626
6627 done = true;
6628
6629 printk(KERN_WARNING "ERROR: %s\n\n", str);
6630
6631 for (i = 0; i < nr_node_ids; i++) {
6632 printk(KERN_WARNING " ");
6633 for (j = 0; j < nr_node_ids; j++)
6634 printk(KERN_CONT "%02d ", node_distance(i,j));
6635 printk(KERN_CONT "\n");
6636 }
6637 printk(KERN_WARNING "\n");
6638}
6639
6640bool find_numa_distance(int distance)
6641{
6642 int i;
6643
6644 if (distance == node_distance(0, 0))
6645 return true;
6646
6647 for (i = 0; i < sched_domains_numa_levels; i++) {
6648 if (sched_domains_numa_distance[i] == distance)
6649 return true;
6650 }
6651
6652 return false;
6653}
6654
6655/*
6656 * A system can have three types of NUMA topology:
6657 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6658 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6659 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6660 *
6661 * The difference between a glueless mesh topology and a backplane
6662 * topology lies in whether communication between not directly
6663 * connected nodes goes through intermediary nodes (where programs
6664 * could run), or through backplane controllers. This affects
6665 * placement of programs.
6666 *
6667 * The type of topology can be discerned with the following tests:
6668 * - If the maximum distance between any nodes is 1 hop, the system
6669 * is directly connected.
6670 * - If for two nodes A and B, located N > 1 hops away from each other,
6671 * there is an intermediary node C, which is < N hops away from both
6672 * nodes A and B, the system is a glueless mesh.
6673 */
6674static void init_numa_topology_type(void)
6675{
6676 int a, b, c, n;
6677
6678 n = sched_max_numa_distance;
6679
6680 if (sched_domains_numa_levels <= 1) {
6681 sched_numa_topology_type = NUMA_DIRECT;
6682 return;
6683 }
6684
6685 for_each_online_node(a) {
6686 for_each_online_node(b) {
6687 /* Find two nodes furthest removed from each other. */
6688 if (node_distance(a, b) < n)
6689 continue;
6690
6691 /* Is there an intermediary node between a and b? */
6692 for_each_online_node(c) {
6693 if (node_distance(a, c) < n &&
6694 node_distance(b, c) < n) {
6695 sched_numa_topology_type =
6696 NUMA_GLUELESS_MESH;
6697 return;
6698 }
6699 }
6700
6701 sched_numa_topology_type = NUMA_BACKPLANE;
6702 return;
6703 }
6704 }
6705}
6706
6707static void sched_init_numa(void)
6708{
6709 int next_distance, curr_distance = node_distance(0, 0);
6710 struct sched_domain_topology_level *tl;
6711 int level = 0;
6712 int i, j, k;
6713
6714 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6715 if (!sched_domains_numa_distance)
6716 return;
6717
6718 /*
6719 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6720 * unique distances in the node_distance() table.
6721 *
6722 * Assumes node_distance(0,j) includes all distances in
6723 * node_distance(i,j) in order to avoid cubic time.
6724 */
6725 next_distance = curr_distance;
6726 for (i = 0; i < nr_node_ids; i++) {
6727 for (j = 0; j < nr_node_ids; j++) {
6728 for (k = 0; k < nr_node_ids; k++) {
6729 int distance = node_distance(i, k);
6730
6731 if (distance > curr_distance &&
6732 (distance < next_distance ||
6733 next_distance == curr_distance))
6734 next_distance = distance;
6735
6736 /*
6737 * While not a strong assumption it would be nice to know
6738 * about cases where if node A is connected to B, B is not
6739 * equally connected to A.
6740 */
6741 if (sched_debug() && node_distance(k, i) != distance)
6742 sched_numa_warn("Node-distance not symmetric");
6743
6744 if (sched_debug() && i && !find_numa_distance(distance))
6745 sched_numa_warn("Node-0 not representative");
6746 }
6747 if (next_distance != curr_distance) {
6748 sched_domains_numa_distance[level++] = next_distance;
6749 sched_domains_numa_levels = level;
6750 curr_distance = next_distance;
6751 } else break;
6752 }
6753
6754 /*
6755 * In case of sched_debug() we verify the above assumption.
6756 */
6757 if (!sched_debug())
6758 break;
6759 }
6760
6761 if (!level)
6762 return;
6763
6764 /*
6765 * 'level' contains the number of unique distances, excluding the
6766 * identity distance node_distance(i,i).
6767 *
6768 * The sched_domains_numa_distance[] array includes the actual distance
6769 * numbers.
6770 */
6771
6772 /*
6773 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6774 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6775 * the array will contain less then 'level' members. This could be
6776 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6777 * in other functions.
6778 *
6779 * We reset it to 'level' at the end of this function.
6780 */
6781 sched_domains_numa_levels = 0;
6782
6783 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6784 if (!sched_domains_numa_masks)
6785 return;
6786
6787 /*
6788 * Now for each level, construct a mask per node which contains all
6789 * cpus of nodes that are that many hops away from us.
6790 */
6791 for (i = 0; i < level; i++) {
6792 sched_domains_numa_masks[i] =
6793 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6794 if (!sched_domains_numa_masks[i])
6795 return;
6796
6797 for (j = 0; j < nr_node_ids; j++) {
6798 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6799 if (!mask)
6800 return;
6801
6802 sched_domains_numa_masks[i][j] = mask;
6803
6804 for_each_node(k) {
6805 if (node_distance(j, k) > sched_domains_numa_distance[i])
6806 continue;
6807
6808 cpumask_or(mask, mask, cpumask_of_node(k));
6809 }
6810 }
6811 }
6812
6813 /* Compute default topology size */
6814 for (i = 0; sched_domain_topology[i].mask; i++);
6815
6816 tl = kzalloc((i + level + 1) *
6817 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6818 if (!tl)
6819 return;
6820
6821 /*
6822 * Copy the default topology bits..
6823 */
6824 for (i = 0; sched_domain_topology[i].mask; i++)
6825 tl[i] = sched_domain_topology[i];
6826
6827 /*
6828 * .. and append 'j' levels of NUMA goodness.
6829 */
6830 for (j = 0; j < level; i++, j++) {
6831 tl[i] = (struct sched_domain_topology_level){
6832 .mask = sd_numa_mask,
6833 .sd_flags = cpu_numa_flags,
6834 .flags = SDTL_OVERLAP,
6835 .numa_level = j,
6836 SD_INIT_NAME(NUMA)
6837 };
6838 }
6839
6840 sched_domain_topology = tl;
6841
6842 sched_domains_numa_levels = level;
6843 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6844
6845 init_numa_topology_type();
6846}
6847
6848static void sched_domains_numa_masks_set(unsigned int cpu)
6849{
6850 int node = cpu_to_node(cpu);
6851 int i, j;
6852
6853 for (i = 0; i < sched_domains_numa_levels; i++) {
6854 for (j = 0; j < nr_node_ids; j++) {
6855 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6856 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6857 }
6858 }
6859}
6860
6861static void sched_domains_numa_masks_clear(unsigned int cpu)
6862{
6863 int i, j;
6864
6865 for (i = 0; i < sched_domains_numa_levels; i++) {
6866 for (j = 0; j < nr_node_ids; j++)
6867 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6868 }
6869}
6870
6871#else
6872static inline void sched_init_numa(void) { }
6873static void sched_domains_numa_masks_set(unsigned int cpu) { }
6874static void sched_domains_numa_masks_clear(unsigned int cpu) { }
6875#endif /* CONFIG_NUMA */
6876
6877static int __sdt_alloc(const struct cpumask *cpu_map)
6878{
6879 struct sched_domain_topology_level *tl;
6880 int j;
6881
6882 for_each_sd_topology(tl) {
6883 struct sd_data *sdd = &tl->data;
6884
6885 sdd->sd = alloc_percpu(struct sched_domain *);
6886 if (!sdd->sd)
6887 return -ENOMEM;
6888
6889 sdd->sds = alloc_percpu(struct sched_domain_shared *);
6890 if (!sdd->sds)
6891 return -ENOMEM;
6892
6893 sdd->sg = alloc_percpu(struct sched_group *);
6894 if (!sdd->sg)
6895 return -ENOMEM;
6896
6897 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6898 if (!sdd->sgc)
6899 return -ENOMEM;
6900
6901 for_each_cpu(j, cpu_map) {
6902 struct sched_domain *sd;
6903 struct sched_domain_shared *sds;
6904 struct sched_group *sg;
6905 struct sched_group_capacity *sgc;
6906
6907 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6908 GFP_KERNEL, cpu_to_node(j));
6909 if (!sd)
6910 return -ENOMEM;
6911
6912 *per_cpu_ptr(sdd->sd, j) = sd;
6913
6914 sds = kzalloc_node(sizeof(struct sched_domain_shared),
6915 GFP_KERNEL, cpu_to_node(j));
6916 if (!sds)
6917 return -ENOMEM;
6918
6919 *per_cpu_ptr(sdd->sds, j) = sds;
6920
6921 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6922 GFP_KERNEL, cpu_to_node(j));
6923 if (!sg)
6924 return -ENOMEM;
6925
6926 sg->next = sg;
6927
6928 *per_cpu_ptr(sdd->sg, j) = sg;
6929
6930 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6931 GFP_KERNEL, cpu_to_node(j));
6932 if (!sgc)
6933 return -ENOMEM;
6934
6935 *per_cpu_ptr(sdd->sgc, j) = sgc;
6936 }
6937 }
6938
6939 return 0;
6940}
6941
6942static void __sdt_free(const struct cpumask *cpu_map)
6943{
6944 struct sched_domain_topology_level *tl;
6945 int j;
6946
6947 for_each_sd_topology(tl) {
6948 struct sd_data *sdd = &tl->data;
6949
6950 for_each_cpu(j, cpu_map) {
6951 struct sched_domain *sd;
6952
6953 if (sdd->sd) {
6954 sd = *per_cpu_ptr(sdd->sd, j);
6955 if (sd && (sd->flags & SD_OVERLAP))
6956 free_sched_groups(sd->groups, 0);
6957 kfree(*per_cpu_ptr(sdd->sd, j));
6958 }
6959
6960 if (sdd->sds)
6961 kfree(*per_cpu_ptr(sdd->sds, j));
6962 if (sdd->sg)
6963 kfree(*per_cpu_ptr(sdd->sg, j));
6964 if (sdd->sgc)
6965 kfree(*per_cpu_ptr(sdd->sgc, j));
6966 }
6967 free_percpu(sdd->sd);
6968 sdd->sd = NULL;
6969 free_percpu(sdd->sds);
6970 sdd->sds = NULL;
6971 free_percpu(sdd->sg);
6972 sdd->sg = NULL;
6973 free_percpu(sdd->sgc);
6974 sdd->sgc = NULL;
6975 }
6976}
6977
6978struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6979 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6980 struct sched_domain *child, int cpu)
6981{
6982 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
6983
6984 if (child) {
6985 sd->level = child->level + 1;
6986 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6987 child->parent = sd;
6988
6989 if (!cpumask_subset(sched_domain_span(child),
6990 sched_domain_span(sd))) {
6991 pr_err("BUG: arch topology borken\n");
6992#ifdef CONFIG_SCHED_DEBUG
6993 pr_err(" the %s domain not a subset of the %s domain\n",
6994 child->name, sd->name);
6995#endif
6996 /* Fixup, ensure @sd has at least @child cpus. */
6997 cpumask_or(sched_domain_span(sd),
6998 sched_domain_span(sd),
6999 sched_domain_span(child));
7000 }
7001
7002 }
7003 set_domain_attribute(sd, attr);
7004
7005 return sd;
7006}
7007
7008/*
7009 * Build sched domains for a given set of cpus and attach the sched domains
7010 * to the individual cpus
7011 */
7012static int build_sched_domains(const struct cpumask *cpu_map,
7013 struct sched_domain_attr *attr)
7014{
7015 enum s_alloc alloc_state;
7016 struct sched_domain *sd;
7017 struct s_data d;
7018 struct rq *rq = NULL;
7019 int i, ret = -ENOMEM;
7020
7021 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7022 if (alloc_state != sa_rootdomain)
7023 goto error;
7024
7025 /* Set up domains for cpus specified by the cpu_map. */
7026 for_each_cpu(i, cpu_map) {
7027 struct sched_domain_topology_level *tl;
7028
7029 sd = NULL;
7030 for_each_sd_topology(tl) {
7031 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7032 if (tl == sched_domain_topology)
7033 *per_cpu_ptr(d.sd, i) = sd;
7034 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7035 sd->flags |= SD_OVERLAP;
7036 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7037 break;
7038 }
7039 }
7040
7041 /* Build the groups for the domains */
7042 for_each_cpu(i, cpu_map) {
7043 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7044 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7045 if (sd->flags & SD_OVERLAP) {
7046 if (build_overlap_sched_groups(sd, i))
7047 goto error;
7048 } else {
7049 if (build_sched_groups(sd, i))
7050 goto error;
7051 }
7052 }
7053 }
7054
7055 /* Calculate CPU capacity for physical packages and nodes */
7056 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7057 if (!cpumask_test_cpu(i, cpu_map))
7058 continue;
7059
7060 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7061 claim_allocations(i, sd);
7062 init_sched_groups_capacity(i, sd);
7063 }
7064 }
7065
7066 /* Attach the domains */
7067 rcu_read_lock();
7068 for_each_cpu(i, cpu_map) {
7069 rq = cpu_rq(i);
7070 sd = *per_cpu_ptr(d.sd, i);
7071
7072 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
7073 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
7074 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
7075
7076 cpu_attach_domain(sd, d.rd, i);
7077 }
7078 rcu_read_unlock();
7079
7080 if (rq && sched_debug_enabled) {
7081 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
7082 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
7083 }
7084
7085 ret = 0;
7086error:
7087 __free_domain_allocs(&d, alloc_state, cpu_map);
7088 return ret;
7089}
7090
7091static cpumask_var_t *doms_cur; /* current sched domains */
7092static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7093static struct sched_domain_attr *dattr_cur;
7094 /* attribues of custom domains in 'doms_cur' */
7095
7096/*
7097 * Special case: If a kmalloc of a doms_cur partition (array of
7098 * cpumask) fails, then fallback to a single sched domain,
7099 * as determined by the single cpumask fallback_doms.
7100 */
7101static cpumask_var_t fallback_doms;
7102
7103/*
7104 * arch_update_cpu_topology lets virtualized architectures update the
7105 * cpu core maps. It is supposed to return 1 if the topology changed
7106 * or 0 if it stayed the same.
7107 */
7108int __weak arch_update_cpu_topology(void)
7109{
7110 return 0;
7111}
7112
7113cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7114{
7115 int i;
7116 cpumask_var_t *doms;
7117
7118 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7119 if (!doms)
7120 return NULL;
7121 for (i = 0; i < ndoms; i++) {
7122 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7123 free_sched_domains(doms, i);
7124 return NULL;
7125 }
7126 }
7127 return doms;
7128}
7129
7130void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7131{
7132 unsigned int i;
7133 for (i = 0; i < ndoms; i++)
7134 free_cpumask_var(doms[i]);
7135 kfree(doms);
7136}
7137
7138/*
7139 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7140 * For now this just excludes isolated cpus, but could be used to
7141 * exclude other special cases in the future.
7142 */
7143static int init_sched_domains(const struct cpumask *cpu_map)
7144{
7145 int err;
7146
7147 arch_update_cpu_topology();
7148 ndoms_cur = 1;
7149 doms_cur = alloc_sched_domains(ndoms_cur);
7150 if (!doms_cur)
7151 doms_cur = &fallback_doms;
7152 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7153 err = build_sched_domains(doms_cur[0], NULL);
7154 register_sched_domain_sysctl();
7155
7156 return err;
7157}
7158
7159/*
7160 * Detach sched domains from a group of cpus specified in cpu_map
7161 * These cpus will now be attached to the NULL domain
7162 */
7163static void detach_destroy_domains(const struct cpumask *cpu_map)
7164{
7165 int i;
7166
7167 rcu_read_lock();
7168 for_each_cpu(i, cpu_map)
7169 cpu_attach_domain(NULL, &def_root_domain, i);
7170 rcu_read_unlock();
7171}
7172
7173/* handle null as "default" */
7174static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7175 struct sched_domain_attr *new, int idx_new)
7176{
7177 struct sched_domain_attr tmp;
7178
7179 /* fast path */
7180 if (!new && !cur)
7181 return 1;
7182
7183 tmp = SD_ATTR_INIT;
7184 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7185 new ? (new + idx_new) : &tmp,
7186 sizeof(struct sched_domain_attr));
7187}
7188
7189/*
7190 * Partition sched domains as specified by the 'ndoms_new'
7191 * cpumasks in the array doms_new[] of cpumasks. This compares
7192 * doms_new[] to the current sched domain partitioning, doms_cur[].
7193 * It destroys each deleted domain and builds each new domain.
7194 *
7195 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7196 * The masks don't intersect (don't overlap.) We should setup one
7197 * sched domain for each mask. CPUs not in any of the cpumasks will
7198 * not be load balanced. If the same cpumask appears both in the
7199 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7200 * it as it is.
7201 *
7202 * The passed in 'doms_new' should be allocated using
7203 * alloc_sched_domains. This routine takes ownership of it and will
7204 * free_sched_domains it when done with it. If the caller failed the
7205 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7206 * and partition_sched_domains() will fallback to the single partition
7207 * 'fallback_doms', it also forces the domains to be rebuilt.
7208 *
7209 * If doms_new == NULL it will be replaced with cpu_online_mask.
7210 * ndoms_new == 0 is a special case for destroying existing domains,
7211 * and it will not create the default domain.
7212 *
7213 * Call with hotplug lock held
7214 */
7215void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7216 struct sched_domain_attr *dattr_new)
7217{
7218 int i, j, n;
7219 int new_topology;
7220
7221 mutex_lock(&sched_domains_mutex);
7222
7223 /* always unregister in case we don't destroy any domains */
7224 unregister_sched_domain_sysctl();
7225
7226 /* Let architecture update cpu core mappings. */
7227 new_topology = arch_update_cpu_topology();
7228
7229 n = doms_new ? ndoms_new : 0;
7230
7231 /* Destroy deleted domains */
7232 for (i = 0; i < ndoms_cur; i++) {
7233 for (j = 0; j < n && !new_topology; j++) {
7234 if (cpumask_equal(doms_cur[i], doms_new[j])
7235 && dattrs_equal(dattr_cur, i, dattr_new, j))
7236 goto match1;
7237 }
7238 /* no match - a current sched domain not in new doms_new[] */
7239 detach_destroy_domains(doms_cur[i]);
7240match1:
7241 ;
7242 }
7243
7244 n = ndoms_cur;
7245 if (doms_new == NULL) {
7246 n = 0;
7247 doms_new = &fallback_doms;
7248 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7249 WARN_ON_ONCE(dattr_new);
7250 }
7251
7252 /* Build new domains */
7253 for (i = 0; i < ndoms_new; i++) {
7254 for (j = 0; j < n && !new_topology; j++) {
7255 if (cpumask_equal(doms_new[i], doms_cur[j])
7256 && dattrs_equal(dattr_new, i, dattr_cur, j))
7257 goto match2;
7258 }
7259 /* no match - add a new doms_new */
7260 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7261match2:
7262 ;
7263 }
7264
7265 /* Remember the new sched domains */
7266 if (doms_cur != &fallback_doms)
7267 free_sched_domains(doms_cur, ndoms_cur);
7268 kfree(dattr_cur); /* kfree(NULL) is safe */
7269 doms_cur = doms_new;
7270 dattr_cur = dattr_new;
7271 ndoms_cur = ndoms_new;
7272
7273 register_sched_domain_sysctl();
7274
7275 mutex_unlock(&sched_domains_mutex);
7276}
7277
7278static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7279
7280/*
7281 * Update cpusets according to cpu_active mask. If cpusets are
7282 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7283 * around partition_sched_domains().
7284 *
7285 * If we come here as part of a suspend/resume, don't touch cpusets because we
7286 * want to restore it back to its original state upon resume anyway.
7287 */
7288static void cpuset_cpu_active(void)
7289{
7290 if (cpuhp_tasks_frozen) {
7291 /*
7292 * num_cpus_frozen tracks how many CPUs are involved in suspend
7293 * resume sequence. As long as this is not the last online
7294 * operation in the resume sequence, just build a single sched
7295 * domain, ignoring cpusets.
7296 */
7297 num_cpus_frozen--;
7298 if (likely(num_cpus_frozen)) {
7299 partition_sched_domains(1, NULL, NULL);
7300 return;
7301 }
7302 /*
7303 * This is the last CPU online operation. So fall through and
7304 * restore the original sched domains by considering the
7305 * cpuset configurations.
7306 */
7307 }
7308 cpuset_update_active_cpus(true);
7309}
7310
7311static int cpuset_cpu_inactive(unsigned int cpu)
7312{
7313 unsigned long flags;
7314 struct dl_bw *dl_b;
7315 bool overflow;
7316 int cpus;
7317
7318 if (!cpuhp_tasks_frozen) {
7319 rcu_read_lock_sched();
7320 dl_b = dl_bw_of(cpu);
7321
7322 raw_spin_lock_irqsave(&dl_b->lock, flags);
7323 cpus = dl_bw_cpus(cpu);
7324 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7325 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7326
7327 rcu_read_unlock_sched();
7328
7329 if (overflow)
7330 return -EBUSY;
7331 cpuset_update_active_cpus(false);
7332 } else {
7333 num_cpus_frozen++;
7334 partition_sched_domains(1, NULL, NULL);
7335 }
7336 return 0;
7337}
7338
7339int sched_cpu_activate(unsigned int cpu)
7340{
7341 struct rq *rq = cpu_rq(cpu);
7342 unsigned long flags;
7343
7344 set_cpu_active(cpu, true);
7345
7346 if (sched_smp_initialized) {
7347 sched_domains_numa_masks_set(cpu);
7348 cpuset_cpu_active();
7349 }
7350
7351 /*
7352 * Put the rq online, if not already. This happens:
7353 *
7354 * 1) In the early boot process, because we build the real domains
7355 * after all cpus have been brought up.
7356 *
7357 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7358 * domains.
7359 */
7360 raw_spin_lock_irqsave(&rq->lock, flags);
7361 if (rq->rd) {
7362 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7363 set_rq_online(rq);
7364 }
7365 raw_spin_unlock_irqrestore(&rq->lock, flags);
7366
7367 update_max_interval();
7368
7369 return 0;
7370}
7371
7372int sched_cpu_deactivate(unsigned int cpu)
7373{
7374 int ret;
7375
7376 set_cpu_active(cpu, false);
7377 /*
7378 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7379 * users of this state to go away such that all new such users will
7380 * observe it.
7381 *
7382 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7383 * not imply sync_sched(), so wait for both.
7384 *
7385 * Do sync before park smpboot threads to take care the rcu boost case.
7386 */
7387 if (IS_ENABLED(CONFIG_PREEMPT))
7388 synchronize_rcu_mult(call_rcu, call_rcu_sched);
7389 else
7390 synchronize_rcu();
7391
7392 if (!sched_smp_initialized)
7393 return 0;
7394
7395 ret = cpuset_cpu_inactive(cpu);
7396 if (ret) {
7397 set_cpu_active(cpu, true);
7398 return ret;
7399 }
7400 sched_domains_numa_masks_clear(cpu);
7401 return 0;
7402}
7403
7404static void sched_rq_cpu_starting(unsigned int cpu)
7405{
7406 struct rq *rq = cpu_rq(cpu);
7407
7408 rq->calc_load_update = calc_load_update;
7409 update_max_interval();
7410}
7411
7412int sched_cpu_starting(unsigned int cpu)
7413{
7414 set_cpu_rq_start_time(cpu);
7415 sched_rq_cpu_starting(cpu);
7416 return 0;
7417}
7418
7419#ifdef CONFIG_HOTPLUG_CPU
7420int sched_cpu_dying(unsigned int cpu)
7421{
7422 struct rq *rq = cpu_rq(cpu);
7423 unsigned long flags;
7424
7425 /* Handle pending wakeups and then migrate everything off */
7426 sched_ttwu_pending();
7427 raw_spin_lock_irqsave(&rq->lock, flags);
7428 if (rq->rd) {
7429 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7430 set_rq_offline(rq);
7431 }
7432 migrate_tasks(rq);
7433 BUG_ON(rq->nr_running != 1);
7434 raw_spin_unlock_irqrestore(&rq->lock, flags);
7435 calc_load_migrate(rq);
7436 update_max_interval();
7437 nohz_balance_exit_idle(cpu);
7438 hrtick_clear(rq);
7439 return 0;
7440}
7441#endif
7442
7443#ifdef CONFIG_SCHED_SMT
7444DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7445
7446static void sched_init_smt(void)
7447{
7448 /*
7449 * We've enumerated all CPUs and will assume that if any CPU
7450 * has SMT siblings, CPU0 will too.
7451 */
7452 if (cpumask_weight(cpu_smt_mask(0)) > 1)
7453 static_branch_enable(&sched_smt_present);
7454}
7455#else
7456static inline void sched_init_smt(void) { }
7457#endif
7458
7459void __init sched_init_smp(void)
7460{
7461 cpumask_var_t non_isolated_cpus;
7462
7463 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7464 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7465
7466 sched_init_numa();
7467
7468 /*
7469 * There's no userspace yet to cause hotplug operations; hence all the
7470 * cpu masks are stable and all blatant races in the below code cannot
7471 * happen.
7472 */
7473 mutex_lock(&sched_domains_mutex);
7474 init_sched_domains(cpu_active_mask);
7475 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7476 if (cpumask_empty(non_isolated_cpus))
7477 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7478 mutex_unlock(&sched_domains_mutex);
7479
7480 /* Move init over to a non-isolated CPU */
7481 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7482 BUG();
7483 sched_init_granularity();
7484 free_cpumask_var(non_isolated_cpus);
7485
7486 init_sched_rt_class();
7487 init_sched_dl_class();
7488
7489 sched_init_smt();
7490
7491 sched_smp_initialized = true;
7492}
7493
7494static int __init migration_init(void)
7495{
7496 sched_rq_cpu_starting(smp_processor_id());
7497 return 0;
7498}
7499early_initcall(migration_init);
7500
7501#else
7502void __init sched_init_smp(void)
7503{
7504 sched_init_granularity();
7505}
7506#endif /* CONFIG_SMP */
7507
7508int in_sched_functions(unsigned long addr)
7509{
7510 return in_lock_functions(addr) ||
7511 (addr >= (unsigned long)__sched_text_start
7512 && addr < (unsigned long)__sched_text_end);
7513}
7514
7515#ifdef CONFIG_CGROUP_SCHED
7516/*
7517 * Default task group.
7518 * Every task in system belongs to this group at bootup.
7519 */
7520struct task_group root_task_group;
7521LIST_HEAD(task_groups);
7522
7523/* Cacheline aligned slab cache for task_group */
7524static struct kmem_cache *task_group_cache __read_mostly;
7525#endif
7526
7527DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7528DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
7529
7530#define WAIT_TABLE_BITS 8
7531#define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
7532static wait_queue_head_t bit_wait_table[WAIT_TABLE_SIZE] __cacheline_aligned;
7533
7534wait_queue_head_t *bit_waitqueue(void *word, int bit)
7535{
7536 const int shift = BITS_PER_LONG == 32 ? 5 : 6;
7537 unsigned long val = (unsigned long)word << shift | bit;
7538
7539 return bit_wait_table + hash_long(val, WAIT_TABLE_BITS);
7540}
7541EXPORT_SYMBOL(bit_waitqueue);
7542
7543void __init sched_init(void)
7544{
7545 int i, j;
7546 unsigned long alloc_size = 0, ptr;
7547
7548 for (i = 0; i < WAIT_TABLE_SIZE; i++)
7549 init_waitqueue_head(bit_wait_table + i);
7550
7551#ifdef CONFIG_FAIR_GROUP_SCHED
7552 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7553#endif
7554#ifdef CONFIG_RT_GROUP_SCHED
7555 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7556#endif
7557 if (alloc_size) {
7558 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7559
7560#ifdef CONFIG_FAIR_GROUP_SCHED
7561 root_task_group.se = (struct sched_entity **)ptr;
7562 ptr += nr_cpu_ids * sizeof(void **);
7563
7564 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7565 ptr += nr_cpu_ids * sizeof(void **);
7566
7567#endif /* CONFIG_FAIR_GROUP_SCHED */
7568#ifdef CONFIG_RT_GROUP_SCHED
7569 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7570 ptr += nr_cpu_ids * sizeof(void **);
7571
7572 root_task_group.rt_rq = (struct rt_rq **)ptr;
7573 ptr += nr_cpu_ids * sizeof(void **);
7574
7575#endif /* CONFIG_RT_GROUP_SCHED */
7576 }
7577#ifdef CONFIG_CPUMASK_OFFSTACK
7578 for_each_possible_cpu(i) {
7579 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7580 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7581 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
7582 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7583 }
7584#endif /* CONFIG_CPUMASK_OFFSTACK */
7585
7586 init_rt_bandwidth(&def_rt_bandwidth,
7587 global_rt_period(), global_rt_runtime());
7588 init_dl_bandwidth(&def_dl_bandwidth,
7589 global_rt_period(), global_rt_runtime());
7590
7591#ifdef CONFIG_SMP
7592 init_defrootdomain();
7593#endif
7594
7595#ifdef CONFIG_RT_GROUP_SCHED
7596 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7597 global_rt_period(), global_rt_runtime());
7598#endif /* CONFIG_RT_GROUP_SCHED */
7599
7600#ifdef CONFIG_CGROUP_SCHED
7601 task_group_cache = KMEM_CACHE(task_group, 0);
7602
7603 list_add(&root_task_group.list, &task_groups);
7604 INIT_LIST_HEAD(&root_task_group.children);
7605 INIT_LIST_HEAD(&root_task_group.siblings);
7606 autogroup_init(&init_task);
7607#endif /* CONFIG_CGROUP_SCHED */
7608
7609 for_each_possible_cpu(i) {
7610 struct rq *rq;
7611
7612 rq = cpu_rq(i);
7613 raw_spin_lock_init(&rq->lock);
7614 rq->nr_running = 0;
7615 rq->calc_load_active = 0;
7616 rq->calc_load_update = jiffies + LOAD_FREQ;
7617 init_cfs_rq(&rq->cfs);
7618 init_rt_rq(&rq->rt);
7619 init_dl_rq(&rq->dl);
7620#ifdef CONFIG_FAIR_GROUP_SCHED
7621 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7622 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7623 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7624 /*
7625 * How much cpu bandwidth does root_task_group get?
7626 *
7627 * In case of task-groups formed thr' the cgroup filesystem, it
7628 * gets 100% of the cpu resources in the system. This overall
7629 * system cpu resource is divided among the tasks of
7630 * root_task_group and its child task-groups in a fair manner,
7631 * based on each entity's (task or task-group's) weight
7632 * (se->load.weight).
7633 *
7634 * In other words, if root_task_group has 10 tasks of weight
7635 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7636 * then A0's share of the cpu resource is:
7637 *
7638 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7639 *
7640 * We achieve this by letting root_task_group's tasks sit
7641 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7642 */
7643 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7644 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7645#endif /* CONFIG_FAIR_GROUP_SCHED */
7646
7647 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7648#ifdef CONFIG_RT_GROUP_SCHED
7649 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7650#endif
7651
7652 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7653 rq->cpu_load[j] = 0;
7654
7655#ifdef CONFIG_SMP
7656 rq->sd = NULL;
7657 rq->rd = NULL;
7658 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7659 rq->balance_callback = NULL;
7660 rq->active_balance = 0;
7661 rq->next_balance = jiffies;
7662 rq->push_cpu = 0;
7663 rq->cpu = i;
7664 rq->online = 0;
7665 rq->idle_stamp = 0;
7666 rq->avg_idle = 2*sysctl_sched_migration_cost;
7667 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7668
7669 INIT_LIST_HEAD(&rq->cfs_tasks);
7670
7671 rq_attach_root(rq, &def_root_domain);
7672#ifdef CONFIG_NO_HZ_COMMON
7673 rq->last_load_update_tick = jiffies;
7674 rq->nohz_flags = 0;
7675#endif
7676#ifdef CONFIG_NO_HZ_FULL
7677 rq->last_sched_tick = 0;
7678#endif
7679#endif /* CONFIG_SMP */
7680 init_rq_hrtick(rq);
7681 atomic_set(&rq->nr_iowait, 0);
7682 }
7683
7684 set_load_weight(&init_task);
7685
7686 /*
7687 * The boot idle thread does lazy MMU switching as well:
7688 */
7689 atomic_inc(&init_mm.mm_count);
7690 enter_lazy_tlb(&init_mm, current);
7691
7692 /*
7693 * Make us the idle thread. Technically, schedule() should not be
7694 * called from this thread, however somewhere below it might be,
7695 * but because we are the idle thread, we just pick up running again
7696 * when this runqueue becomes "idle".
7697 */
7698 init_idle(current, smp_processor_id());
7699
7700 calc_load_update = jiffies + LOAD_FREQ;
7701
7702#ifdef CONFIG_SMP
7703 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7704 /* May be allocated at isolcpus cmdline parse time */
7705 if (cpu_isolated_map == NULL)
7706 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7707 idle_thread_set_boot_cpu();
7708 set_cpu_rq_start_time(smp_processor_id());
7709#endif
7710 init_sched_fair_class();
7711
7712 init_schedstats();
7713
7714 scheduler_running = 1;
7715}
7716
7717#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7718static inline int preempt_count_equals(int preempt_offset)
7719{
7720 int nested = preempt_count() + rcu_preempt_depth();
7721
7722 return (nested == preempt_offset);
7723}
7724
7725void __might_sleep(const char *file, int line, int preempt_offset)
7726{
7727 /*
7728 * Blocking primitives will set (and therefore destroy) current->state,
7729 * since we will exit with TASK_RUNNING make sure we enter with it,
7730 * otherwise we will destroy state.
7731 */
7732 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7733 "do not call blocking ops when !TASK_RUNNING; "
7734 "state=%lx set at [<%p>] %pS\n",
7735 current->state,
7736 (void *)current->task_state_change,
7737 (void *)current->task_state_change);
7738
7739 ___might_sleep(file, line, preempt_offset);
7740}
7741EXPORT_SYMBOL(__might_sleep);
7742
7743void ___might_sleep(const char *file, int line, int preempt_offset)
7744{
7745 static unsigned long prev_jiffy; /* ratelimiting */
7746 unsigned long preempt_disable_ip;
7747
7748 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7749 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7750 !is_idle_task(current)) ||
7751 system_state != SYSTEM_RUNNING || oops_in_progress)
7752 return;
7753 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7754 return;
7755 prev_jiffy = jiffies;
7756
7757 /* Save this before calling printk(), since that will clobber it */
7758 preempt_disable_ip = get_preempt_disable_ip(current);
7759
7760 printk(KERN_ERR
7761 "BUG: sleeping function called from invalid context at %s:%d\n",
7762 file, line);
7763 printk(KERN_ERR
7764 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7765 in_atomic(), irqs_disabled(),
7766 current->pid, current->comm);
7767
7768 if (task_stack_end_corrupted(current))
7769 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7770
7771 debug_show_held_locks(current);
7772 if (irqs_disabled())
7773 print_irqtrace_events(current);
7774 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
7775 && !preempt_count_equals(preempt_offset)) {
7776 pr_err("Preemption disabled at:");
7777 print_ip_sym(preempt_disable_ip);
7778 pr_cont("\n");
7779 }
7780 dump_stack();
7781 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7782}
7783EXPORT_SYMBOL(___might_sleep);
7784#endif
7785
7786#ifdef CONFIG_MAGIC_SYSRQ
7787void normalize_rt_tasks(void)
7788{
7789 struct task_struct *g, *p;
7790 struct sched_attr attr = {
7791 .sched_policy = SCHED_NORMAL,
7792 };
7793
7794 read_lock(&tasklist_lock);
7795 for_each_process_thread(g, p) {
7796 /*
7797 * Only normalize user tasks:
7798 */
7799 if (p->flags & PF_KTHREAD)
7800 continue;
7801
7802 p->se.exec_start = 0;
7803 schedstat_set(p->se.statistics.wait_start, 0);
7804 schedstat_set(p->se.statistics.sleep_start, 0);
7805 schedstat_set(p->se.statistics.block_start, 0);
7806
7807 if (!dl_task(p) && !rt_task(p)) {
7808 /*
7809 * Renice negative nice level userspace
7810 * tasks back to 0:
7811 */
7812 if (task_nice(p) < 0)
7813 set_user_nice(p, 0);
7814 continue;
7815 }
7816
7817 __sched_setscheduler(p, &attr, false, false);
7818 }
7819 read_unlock(&tasklist_lock);
7820}
7821
7822#endif /* CONFIG_MAGIC_SYSRQ */
7823
7824#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7825/*
7826 * These functions are only useful for the IA64 MCA handling, or kdb.
7827 *
7828 * They can only be called when the whole system has been
7829 * stopped - every CPU needs to be quiescent, and no scheduling
7830 * activity can take place. Using them for anything else would
7831 * be a serious bug, and as a result, they aren't even visible
7832 * under any other configuration.
7833 */
7834
7835/**
7836 * curr_task - return the current task for a given cpu.
7837 * @cpu: the processor in question.
7838 *
7839 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7840 *
7841 * Return: The current task for @cpu.
7842 */
7843struct task_struct *curr_task(int cpu)
7844{
7845 return cpu_curr(cpu);
7846}
7847
7848#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7849
7850#ifdef CONFIG_IA64
7851/**
7852 * set_curr_task - set the current task for a given cpu.
7853 * @cpu: the processor in question.
7854 * @p: the task pointer to set.
7855 *
7856 * Description: This function must only be used when non-maskable interrupts
7857 * are serviced on a separate stack. It allows the architecture to switch the
7858 * notion of the current task on a cpu in a non-blocking manner. This function
7859 * must be called with all CPU's synchronized, and interrupts disabled, the
7860 * and caller must save the original value of the current task (see
7861 * curr_task() above) and restore that value before reenabling interrupts and
7862 * re-starting the system.
7863 *
7864 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7865 */
7866void ia64_set_curr_task(int cpu, struct task_struct *p)
7867{
7868 cpu_curr(cpu) = p;
7869}
7870
7871#endif
7872
7873#ifdef CONFIG_CGROUP_SCHED
7874/* task_group_lock serializes the addition/removal of task groups */
7875static DEFINE_SPINLOCK(task_group_lock);
7876
7877static void sched_free_group(struct task_group *tg)
7878{
7879 free_fair_sched_group(tg);
7880 free_rt_sched_group(tg);
7881 autogroup_free(tg);
7882 kmem_cache_free(task_group_cache, tg);
7883}
7884
7885/* allocate runqueue etc for a new task group */
7886struct task_group *sched_create_group(struct task_group *parent)
7887{
7888 struct task_group *tg;
7889
7890 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7891 if (!tg)
7892 return ERR_PTR(-ENOMEM);
7893
7894 if (!alloc_fair_sched_group(tg, parent))
7895 goto err;
7896
7897 if (!alloc_rt_sched_group(tg, parent))
7898 goto err;
7899
7900 return tg;
7901
7902err:
7903 sched_free_group(tg);
7904 return ERR_PTR(-ENOMEM);
7905}
7906
7907void sched_online_group(struct task_group *tg, struct task_group *parent)
7908{
7909 unsigned long flags;
7910
7911 spin_lock_irqsave(&task_group_lock, flags);
7912 list_add_rcu(&tg->list, &task_groups);
7913
7914 WARN_ON(!parent); /* root should already exist */
7915
7916 tg->parent = parent;
7917 INIT_LIST_HEAD(&tg->children);
7918 list_add_rcu(&tg->siblings, &parent->children);
7919 spin_unlock_irqrestore(&task_group_lock, flags);
7920
7921 online_fair_sched_group(tg);
7922}
7923
7924/* rcu callback to free various structures associated with a task group */
7925static void sched_free_group_rcu(struct rcu_head *rhp)
7926{
7927 /* now it should be safe to free those cfs_rqs */
7928 sched_free_group(container_of(rhp, struct task_group, rcu));
7929}
7930
7931void sched_destroy_group(struct task_group *tg)
7932{
7933 /* wait for possible concurrent references to cfs_rqs complete */
7934 call_rcu(&tg->rcu, sched_free_group_rcu);
7935}
7936
7937void sched_offline_group(struct task_group *tg)
7938{
7939 unsigned long flags;
7940
7941 /* end participation in shares distribution */
7942 unregister_fair_sched_group(tg);
7943
7944 spin_lock_irqsave(&task_group_lock, flags);
7945 list_del_rcu(&tg->list);
7946 list_del_rcu(&tg->siblings);
7947 spin_unlock_irqrestore(&task_group_lock, flags);
7948}
7949
7950static void sched_change_group(struct task_struct *tsk, int type)
7951{
7952 struct task_group *tg;
7953
7954 /*
7955 * All callers are synchronized by task_rq_lock(); we do not use RCU
7956 * which is pointless here. Thus, we pass "true" to task_css_check()
7957 * to prevent lockdep warnings.
7958 */
7959 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7960 struct task_group, css);
7961 tg = autogroup_task_group(tsk, tg);
7962 tsk->sched_task_group = tg;
7963
7964#ifdef CONFIG_FAIR_GROUP_SCHED
7965 if (tsk->sched_class->task_change_group)
7966 tsk->sched_class->task_change_group(tsk, type);
7967 else
7968#endif
7969 set_task_rq(tsk, task_cpu(tsk));
7970}
7971
7972/*
7973 * Change task's runqueue when it moves between groups.
7974 *
7975 * The caller of this function should have put the task in its new group by
7976 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7977 * its new group.
7978 */
7979void sched_move_task(struct task_struct *tsk)
7980{
7981 int queued, running;
7982 struct rq_flags rf;
7983 struct rq *rq;
7984
7985 rq = task_rq_lock(tsk, &rf);
7986
7987 running = task_current(rq, tsk);
7988 queued = task_on_rq_queued(tsk);
7989
7990 if (queued)
7991 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7992 if (unlikely(running))
7993 put_prev_task(rq, tsk);
7994
7995 sched_change_group(tsk, TASK_MOVE_GROUP);
7996
7997 if (queued)
7998 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7999 if (unlikely(running))
8000 set_curr_task(rq, tsk);
8001
8002 task_rq_unlock(rq, tsk, &rf);
8003}
8004#endif /* CONFIG_CGROUP_SCHED */
8005
8006#ifdef CONFIG_RT_GROUP_SCHED
8007/*
8008 * Ensure that the real time constraints are schedulable.
8009 */
8010static DEFINE_MUTEX(rt_constraints_mutex);
8011
8012/* Must be called with tasklist_lock held */
8013static inline int tg_has_rt_tasks(struct task_group *tg)
8014{
8015 struct task_struct *g, *p;
8016
8017 /*
8018 * Autogroups do not have RT tasks; see autogroup_create().
8019 */
8020 if (task_group_is_autogroup(tg))
8021 return 0;
8022
8023 for_each_process_thread(g, p) {
8024 if (rt_task(p) && task_group(p) == tg)
8025 return 1;
8026 }
8027
8028 return 0;
8029}
8030
8031struct rt_schedulable_data {
8032 struct task_group *tg;
8033 u64 rt_period;
8034 u64 rt_runtime;
8035};
8036
8037static int tg_rt_schedulable(struct task_group *tg, void *data)
8038{
8039 struct rt_schedulable_data *d = data;
8040 struct task_group *child;
8041 unsigned long total, sum = 0;
8042 u64 period, runtime;
8043
8044 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8045 runtime = tg->rt_bandwidth.rt_runtime;
8046
8047 if (tg == d->tg) {
8048 period = d->rt_period;
8049 runtime = d->rt_runtime;
8050 }
8051
8052 /*
8053 * Cannot have more runtime than the period.
8054 */
8055 if (runtime > period && runtime != RUNTIME_INF)
8056 return -EINVAL;
8057
8058 /*
8059 * Ensure we don't starve existing RT tasks.
8060 */
8061 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8062 return -EBUSY;
8063
8064 total = to_ratio(period, runtime);
8065
8066 /*
8067 * Nobody can have more than the global setting allows.
8068 */
8069 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8070 return -EINVAL;
8071
8072 /*
8073 * The sum of our children's runtime should not exceed our own.
8074 */
8075 list_for_each_entry_rcu(child, &tg->children, siblings) {
8076 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8077 runtime = child->rt_bandwidth.rt_runtime;
8078
8079 if (child == d->tg) {
8080 period = d->rt_period;
8081 runtime = d->rt_runtime;
8082 }
8083
8084 sum += to_ratio(period, runtime);
8085 }
8086
8087 if (sum > total)
8088 return -EINVAL;
8089
8090 return 0;
8091}
8092
8093static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8094{
8095 int ret;
8096
8097 struct rt_schedulable_data data = {
8098 .tg = tg,
8099 .rt_period = period,
8100 .rt_runtime = runtime,
8101 };
8102
8103 rcu_read_lock();
8104 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8105 rcu_read_unlock();
8106
8107 return ret;
8108}
8109
8110static int tg_set_rt_bandwidth(struct task_group *tg,
8111 u64 rt_period, u64 rt_runtime)
8112{
8113 int i, err = 0;
8114
8115 /*
8116 * Disallowing the root group RT runtime is BAD, it would disallow the
8117 * kernel creating (and or operating) RT threads.
8118 */
8119 if (tg == &root_task_group && rt_runtime == 0)
8120 return -EINVAL;
8121
8122 /* No period doesn't make any sense. */
8123 if (rt_period == 0)
8124 return -EINVAL;
8125
8126 mutex_lock(&rt_constraints_mutex);
8127 read_lock(&tasklist_lock);
8128 err = __rt_schedulable(tg, rt_period, rt_runtime);
8129 if (err)
8130 goto unlock;
8131
8132 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8133 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8134 tg->rt_bandwidth.rt_runtime = rt_runtime;
8135
8136 for_each_possible_cpu(i) {
8137 struct rt_rq *rt_rq = tg->rt_rq[i];
8138
8139 raw_spin_lock(&rt_rq->rt_runtime_lock);
8140 rt_rq->rt_runtime = rt_runtime;
8141 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8142 }
8143 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8144unlock:
8145 read_unlock(&tasklist_lock);
8146 mutex_unlock(&rt_constraints_mutex);
8147
8148 return err;
8149}
8150
8151static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8152{
8153 u64 rt_runtime, rt_period;
8154
8155 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8156 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8157 if (rt_runtime_us < 0)
8158 rt_runtime = RUNTIME_INF;
8159
8160 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8161}
8162
8163static long sched_group_rt_runtime(struct task_group *tg)
8164{
8165 u64 rt_runtime_us;
8166
8167 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8168 return -1;
8169
8170 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8171 do_div(rt_runtime_us, NSEC_PER_USEC);
8172 return rt_runtime_us;
8173}
8174
8175static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8176{
8177 u64 rt_runtime, rt_period;
8178
8179 rt_period = rt_period_us * NSEC_PER_USEC;
8180 rt_runtime = tg->rt_bandwidth.rt_runtime;
8181
8182 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8183}
8184
8185static long sched_group_rt_period(struct task_group *tg)
8186{
8187 u64 rt_period_us;
8188
8189 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8190 do_div(rt_period_us, NSEC_PER_USEC);
8191 return rt_period_us;
8192}
8193#endif /* CONFIG_RT_GROUP_SCHED */
8194
8195#ifdef CONFIG_RT_GROUP_SCHED
8196static int sched_rt_global_constraints(void)
8197{
8198 int ret = 0;
8199
8200 mutex_lock(&rt_constraints_mutex);
8201 read_lock(&tasklist_lock);
8202 ret = __rt_schedulable(NULL, 0, 0);
8203 read_unlock(&tasklist_lock);
8204 mutex_unlock(&rt_constraints_mutex);
8205
8206 return ret;
8207}
8208
8209static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8210{
8211 /* Don't accept realtime tasks when there is no way for them to run */
8212 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8213 return 0;
8214
8215 return 1;
8216}
8217
8218#else /* !CONFIG_RT_GROUP_SCHED */
8219static int sched_rt_global_constraints(void)
8220{
8221 unsigned long flags;
8222 int i;
8223
8224 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8225 for_each_possible_cpu(i) {
8226 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8227
8228 raw_spin_lock(&rt_rq->rt_runtime_lock);
8229 rt_rq->rt_runtime = global_rt_runtime();
8230 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8231 }
8232 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8233
8234 return 0;
8235}
8236#endif /* CONFIG_RT_GROUP_SCHED */
8237
8238static int sched_dl_global_validate(void)
8239{
8240 u64 runtime = global_rt_runtime();
8241 u64 period = global_rt_period();
8242 u64 new_bw = to_ratio(period, runtime);
8243 struct dl_bw *dl_b;
8244 int cpu, ret = 0;
8245 unsigned long flags;
8246
8247 /*
8248 * Here we want to check the bandwidth not being set to some
8249 * value smaller than the currently allocated bandwidth in
8250 * any of the root_domains.
8251 *
8252 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8253 * cycling on root_domains... Discussion on different/better
8254 * solutions is welcome!
8255 */
8256 for_each_possible_cpu(cpu) {
8257 rcu_read_lock_sched();
8258 dl_b = dl_bw_of(cpu);
8259
8260 raw_spin_lock_irqsave(&dl_b->lock, flags);
8261 if (new_bw < dl_b->total_bw)
8262 ret = -EBUSY;
8263 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8264
8265 rcu_read_unlock_sched();
8266
8267 if (ret)
8268 break;
8269 }
8270
8271 return ret;
8272}
8273
8274static void sched_dl_do_global(void)
8275{
8276 u64 new_bw = -1;
8277 struct dl_bw *dl_b;
8278 int cpu;
8279 unsigned long flags;
8280
8281 def_dl_bandwidth.dl_period = global_rt_period();
8282 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8283
8284 if (global_rt_runtime() != RUNTIME_INF)
8285 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8286
8287 /*
8288 * FIXME: As above...
8289 */
8290 for_each_possible_cpu(cpu) {
8291 rcu_read_lock_sched();
8292 dl_b = dl_bw_of(cpu);
8293
8294 raw_spin_lock_irqsave(&dl_b->lock, flags);
8295 dl_b->bw = new_bw;
8296 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8297
8298 rcu_read_unlock_sched();
8299 }
8300}
8301
8302static int sched_rt_global_validate(void)
8303{
8304 if (sysctl_sched_rt_period <= 0)
8305 return -EINVAL;
8306
8307 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8308 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8309 return -EINVAL;
8310
8311 return 0;
8312}
8313
8314static void sched_rt_do_global(void)
8315{
8316 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8317 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8318}
8319
8320int sched_rt_handler(struct ctl_table *table, int write,
8321 void __user *buffer, size_t *lenp,
8322 loff_t *ppos)
8323{
8324 int old_period, old_runtime;
8325 static DEFINE_MUTEX(mutex);
8326 int ret;
8327
8328 mutex_lock(&mutex);
8329 old_period = sysctl_sched_rt_period;
8330 old_runtime = sysctl_sched_rt_runtime;
8331
8332 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8333
8334 if (!ret && write) {
8335 ret = sched_rt_global_validate();
8336 if (ret)
8337 goto undo;
8338
8339 ret = sched_dl_global_validate();
8340 if (ret)
8341 goto undo;
8342
8343 ret = sched_rt_global_constraints();
8344 if (ret)
8345 goto undo;
8346
8347 sched_rt_do_global();
8348 sched_dl_do_global();
8349 }
8350 if (0) {
8351undo:
8352 sysctl_sched_rt_period = old_period;
8353 sysctl_sched_rt_runtime = old_runtime;
8354 }
8355 mutex_unlock(&mutex);
8356
8357 return ret;
8358}
8359
8360int sched_rr_handler(struct ctl_table *table, int write,
8361 void __user *buffer, size_t *lenp,
8362 loff_t *ppos)
8363{
8364 int ret;
8365 static DEFINE_MUTEX(mutex);
8366
8367 mutex_lock(&mutex);
8368 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8369 /* make sure that internally we keep jiffies */
8370 /* also, writing zero resets timeslice to default */
8371 if (!ret && write) {
8372 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8373 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8374 }
8375 mutex_unlock(&mutex);
8376 return ret;
8377}
8378
8379#ifdef CONFIG_CGROUP_SCHED
8380
8381static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8382{
8383 return css ? container_of(css, struct task_group, css) : NULL;
8384}
8385
8386static struct cgroup_subsys_state *
8387cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8388{
8389 struct task_group *parent = css_tg(parent_css);
8390 struct task_group *tg;
8391
8392 if (!parent) {
8393 /* This is early initialization for the top cgroup */
8394 return &root_task_group.css;
8395 }
8396
8397 tg = sched_create_group(parent);
8398 if (IS_ERR(tg))
8399 return ERR_PTR(-ENOMEM);
8400
8401 sched_online_group(tg, parent);
8402
8403 return &tg->css;
8404}
8405
8406static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8407{
8408 struct task_group *tg = css_tg(css);
8409
8410 sched_offline_group(tg);
8411}
8412
8413static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8414{
8415 struct task_group *tg = css_tg(css);
8416
8417 /*
8418 * Relies on the RCU grace period between css_released() and this.
8419 */
8420 sched_free_group(tg);
8421}
8422
8423/*
8424 * This is called before wake_up_new_task(), therefore we really only
8425 * have to set its group bits, all the other stuff does not apply.
8426 */
8427static void cpu_cgroup_fork(struct task_struct *task)
8428{
8429 struct rq_flags rf;
8430 struct rq *rq;
8431
8432 rq = task_rq_lock(task, &rf);
8433
8434 sched_change_group(task, TASK_SET_GROUP);
8435
8436 task_rq_unlock(rq, task, &rf);
8437}
8438
8439static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8440{
8441 struct task_struct *task;
8442 struct cgroup_subsys_state *css;
8443 int ret = 0;
8444
8445 cgroup_taskset_for_each(task, css, tset) {
8446#ifdef CONFIG_RT_GROUP_SCHED
8447 if (!sched_rt_can_attach(css_tg(css), task))
8448 return -EINVAL;
8449#else
8450 /* We don't support RT-tasks being in separate groups */
8451 if (task->sched_class != &fair_sched_class)
8452 return -EINVAL;
8453#endif
8454 /*
8455 * Serialize against wake_up_new_task() such that if its
8456 * running, we're sure to observe its full state.
8457 */
8458 raw_spin_lock_irq(&task->pi_lock);
8459 /*
8460 * Avoid calling sched_move_task() before wake_up_new_task()
8461 * has happened. This would lead to problems with PELT, due to
8462 * move wanting to detach+attach while we're not attached yet.
8463 */
8464 if (task->state == TASK_NEW)
8465 ret = -EINVAL;
8466 raw_spin_unlock_irq(&task->pi_lock);
8467
8468 if (ret)
8469 break;
8470 }
8471 return ret;
8472}
8473
8474static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8475{
8476 struct task_struct *task;
8477 struct cgroup_subsys_state *css;
8478
8479 cgroup_taskset_for_each(task, css, tset)
8480 sched_move_task(task);
8481}
8482
8483#ifdef CONFIG_FAIR_GROUP_SCHED
8484static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8485 struct cftype *cftype, u64 shareval)
8486{
8487 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8488}
8489
8490static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8491 struct cftype *cft)
8492{
8493 struct task_group *tg = css_tg(css);
8494
8495 return (u64) scale_load_down(tg->shares);
8496}
8497
8498#ifdef CONFIG_CFS_BANDWIDTH
8499static DEFINE_MUTEX(cfs_constraints_mutex);
8500
8501const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8502const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8503
8504static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8505
8506static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8507{
8508 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8509 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8510
8511 if (tg == &root_task_group)
8512 return -EINVAL;
8513
8514 /*
8515 * Ensure we have at some amount of bandwidth every period. This is
8516 * to prevent reaching a state of large arrears when throttled via
8517 * entity_tick() resulting in prolonged exit starvation.
8518 */
8519 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8520 return -EINVAL;
8521
8522 /*
8523 * Likewise, bound things on the otherside by preventing insane quota
8524 * periods. This also allows us to normalize in computing quota
8525 * feasibility.
8526 */
8527 if (period > max_cfs_quota_period)
8528 return -EINVAL;
8529
8530 /*
8531 * Prevent race between setting of cfs_rq->runtime_enabled and
8532 * unthrottle_offline_cfs_rqs().
8533 */
8534 get_online_cpus();
8535 mutex_lock(&cfs_constraints_mutex);
8536 ret = __cfs_schedulable(tg, period, quota);
8537 if (ret)
8538 goto out_unlock;
8539
8540 runtime_enabled = quota != RUNTIME_INF;
8541 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8542 /*
8543 * If we need to toggle cfs_bandwidth_used, off->on must occur
8544 * before making related changes, and on->off must occur afterwards
8545 */
8546 if (runtime_enabled && !runtime_was_enabled)
8547 cfs_bandwidth_usage_inc();
8548 raw_spin_lock_irq(&cfs_b->lock);
8549 cfs_b->period = ns_to_ktime(period);
8550 cfs_b->quota = quota;
8551
8552 __refill_cfs_bandwidth_runtime(cfs_b);
8553 /* restart the period timer (if active) to handle new period expiry */
8554 if (runtime_enabled)
8555 start_cfs_bandwidth(cfs_b);
8556 raw_spin_unlock_irq(&cfs_b->lock);
8557
8558 for_each_online_cpu(i) {
8559 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8560 struct rq *rq = cfs_rq->rq;
8561
8562 raw_spin_lock_irq(&rq->lock);
8563 cfs_rq->runtime_enabled = runtime_enabled;
8564 cfs_rq->runtime_remaining = 0;
8565
8566 if (cfs_rq->throttled)
8567 unthrottle_cfs_rq(cfs_rq);
8568 raw_spin_unlock_irq(&rq->lock);
8569 }
8570 if (runtime_was_enabled && !runtime_enabled)
8571 cfs_bandwidth_usage_dec();
8572out_unlock:
8573 mutex_unlock(&cfs_constraints_mutex);
8574 put_online_cpus();
8575
8576 return ret;
8577}
8578
8579int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8580{
8581 u64 quota, period;
8582
8583 period = ktime_to_ns(tg->cfs_bandwidth.period);
8584 if (cfs_quota_us < 0)
8585 quota = RUNTIME_INF;
8586 else
8587 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8588
8589 return tg_set_cfs_bandwidth(tg, period, quota);
8590}
8591
8592long tg_get_cfs_quota(struct task_group *tg)
8593{
8594 u64 quota_us;
8595
8596 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8597 return -1;
8598
8599 quota_us = tg->cfs_bandwidth.quota;
8600 do_div(quota_us, NSEC_PER_USEC);
8601
8602 return quota_us;
8603}
8604
8605int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8606{
8607 u64 quota, period;
8608
8609 period = (u64)cfs_period_us * NSEC_PER_USEC;
8610 quota = tg->cfs_bandwidth.quota;
8611
8612 return tg_set_cfs_bandwidth(tg, period, quota);
8613}
8614
8615long tg_get_cfs_period(struct task_group *tg)
8616{
8617 u64 cfs_period_us;
8618
8619 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8620 do_div(cfs_period_us, NSEC_PER_USEC);
8621
8622 return cfs_period_us;
8623}
8624
8625static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8626 struct cftype *cft)
8627{
8628 return tg_get_cfs_quota(css_tg(css));
8629}
8630
8631static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8632 struct cftype *cftype, s64 cfs_quota_us)
8633{
8634 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8635}
8636
8637static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8638 struct cftype *cft)
8639{
8640 return tg_get_cfs_period(css_tg(css));
8641}
8642
8643static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8644 struct cftype *cftype, u64 cfs_period_us)
8645{
8646 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8647}
8648
8649struct cfs_schedulable_data {
8650 struct task_group *tg;
8651 u64 period, quota;
8652};
8653
8654/*
8655 * normalize group quota/period to be quota/max_period
8656 * note: units are usecs
8657 */
8658static u64 normalize_cfs_quota(struct task_group *tg,
8659 struct cfs_schedulable_data *d)
8660{
8661 u64 quota, period;
8662
8663 if (tg == d->tg) {
8664 period = d->period;
8665 quota = d->quota;
8666 } else {
8667 period = tg_get_cfs_period(tg);
8668 quota = tg_get_cfs_quota(tg);
8669 }
8670
8671 /* note: these should typically be equivalent */
8672 if (quota == RUNTIME_INF || quota == -1)
8673 return RUNTIME_INF;
8674
8675 return to_ratio(period, quota);
8676}
8677
8678static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8679{
8680 struct cfs_schedulable_data *d = data;
8681 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8682 s64 quota = 0, parent_quota = -1;
8683
8684 if (!tg->parent) {
8685 quota = RUNTIME_INF;
8686 } else {
8687 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8688
8689 quota = normalize_cfs_quota(tg, d);
8690 parent_quota = parent_b->hierarchical_quota;
8691
8692 /*
8693 * ensure max(child_quota) <= parent_quota, inherit when no
8694 * limit is set
8695 */
8696 if (quota == RUNTIME_INF)
8697 quota = parent_quota;
8698 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8699 return -EINVAL;
8700 }
8701 cfs_b->hierarchical_quota = quota;
8702
8703 return 0;
8704}
8705
8706static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8707{
8708 int ret;
8709 struct cfs_schedulable_data data = {
8710 .tg = tg,
8711 .period = period,
8712 .quota = quota,
8713 };
8714
8715 if (quota != RUNTIME_INF) {
8716 do_div(data.period, NSEC_PER_USEC);
8717 do_div(data.quota, NSEC_PER_USEC);
8718 }
8719
8720 rcu_read_lock();
8721 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8722 rcu_read_unlock();
8723
8724 return ret;
8725}
8726
8727static int cpu_stats_show(struct seq_file *sf, void *v)
8728{
8729 struct task_group *tg = css_tg(seq_css(sf));
8730 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8731
8732 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8733 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8734 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8735
8736 return 0;
8737}
8738#endif /* CONFIG_CFS_BANDWIDTH */
8739#endif /* CONFIG_FAIR_GROUP_SCHED */
8740
8741#ifdef CONFIG_RT_GROUP_SCHED
8742static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8743 struct cftype *cft, s64 val)
8744{
8745 return sched_group_set_rt_runtime(css_tg(css), val);
8746}
8747
8748static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8749 struct cftype *cft)
8750{
8751 return sched_group_rt_runtime(css_tg(css));
8752}
8753
8754static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8755 struct cftype *cftype, u64 rt_period_us)
8756{
8757 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8758}
8759
8760static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8761 struct cftype *cft)
8762{
8763 return sched_group_rt_period(css_tg(css));
8764}
8765#endif /* CONFIG_RT_GROUP_SCHED */
8766
8767static struct cftype cpu_files[] = {
8768#ifdef CONFIG_FAIR_GROUP_SCHED
8769 {
8770 .name = "shares",
8771 .read_u64 = cpu_shares_read_u64,
8772 .write_u64 = cpu_shares_write_u64,
8773 },
8774#endif
8775#ifdef CONFIG_CFS_BANDWIDTH
8776 {
8777 .name = "cfs_quota_us",
8778 .read_s64 = cpu_cfs_quota_read_s64,
8779 .write_s64 = cpu_cfs_quota_write_s64,
8780 },
8781 {
8782 .name = "cfs_period_us",
8783 .read_u64 = cpu_cfs_period_read_u64,
8784 .write_u64 = cpu_cfs_period_write_u64,
8785 },
8786 {
8787 .name = "stat",
8788 .seq_show = cpu_stats_show,
8789 },
8790#endif
8791#ifdef CONFIG_RT_GROUP_SCHED
8792 {
8793 .name = "rt_runtime_us",
8794 .read_s64 = cpu_rt_runtime_read,
8795 .write_s64 = cpu_rt_runtime_write,
8796 },
8797 {
8798 .name = "rt_period_us",
8799 .read_u64 = cpu_rt_period_read_uint,
8800 .write_u64 = cpu_rt_period_write_uint,
8801 },
8802#endif
8803 { } /* terminate */
8804};
8805
8806struct cgroup_subsys cpu_cgrp_subsys = {
8807 .css_alloc = cpu_cgroup_css_alloc,
8808 .css_released = cpu_cgroup_css_released,
8809 .css_free = cpu_cgroup_css_free,
8810 .fork = cpu_cgroup_fork,
8811 .can_attach = cpu_cgroup_can_attach,
8812 .attach = cpu_cgroup_attach,
8813 .legacy_cftypes = cpu_files,
8814 .early_init = true,
8815};
8816
8817#endif /* CONFIG_CGROUP_SCHED */
8818
8819void dump_cpu_task(int cpu)
8820{
8821 pr_info("Task dump for CPU %d:\n", cpu);
8822 sched_show_task(cpu_curr(cpu));
8823}
8824
8825/*
8826 * Nice levels are multiplicative, with a gentle 10% change for every
8827 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8828 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8829 * that remained on nice 0.
8830 *
8831 * The "10% effect" is relative and cumulative: from _any_ nice level,
8832 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8833 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8834 * If a task goes up by ~10% and another task goes down by ~10% then
8835 * the relative distance between them is ~25%.)
8836 */
8837const int sched_prio_to_weight[40] = {
8838 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8839 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8840 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8841 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8842 /* 0 */ 1024, 820, 655, 526, 423,
8843 /* 5 */ 335, 272, 215, 172, 137,
8844 /* 10 */ 110, 87, 70, 56, 45,
8845 /* 15 */ 36, 29, 23, 18, 15,
8846};
8847
8848/*
8849 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8850 *
8851 * In cases where the weight does not change often, we can use the
8852 * precalculated inverse to speed up arithmetics by turning divisions
8853 * into multiplications:
8854 */
8855const u32 sched_prio_to_wmult[40] = {
8856 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8857 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8858 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8859 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8860 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8861 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8862 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8863 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8864};