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