1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #include <linux/highmem.h>
10 #include <linux/hrtimer_api.h>
11 #include <linux/ktime_api.h>
12 #include <linux/sched/signal.h>
13 #include <linux/syscalls_api.h>
14 #include <linux/debug_locks.h>
15 #include <linux/prefetch.h>
16 #include <linux/capability.h>
17 #include <linux/pgtable_api.h>
18 #include <linux/wait_bit.h>
19 #include <linux/jiffies.h>
20 #include <linux/spinlock_api.h>
21 #include <linux/cpumask_api.h>
22 #include <linux/lockdep_api.h>
23 #include <linux/hardirq.h>
24 #include <linux/softirq.h>
25 #include <linux/refcount_api.h>
26 #include <linux/topology.h>
27 #include <linux/sched/clock.h>
28 #include <linux/sched/cond_resched.h>
29 #include <linux/sched/cputime.h>
30 #include <linux/sched/debug.h>
31 #include <linux/sched/hotplug.h>
32 #include <linux/sched/init.h>
33 #include <linux/sched/isolation.h>
34 #include <linux/sched/loadavg.h>
35 #include <linux/sched/mm.h>
36 #include <linux/sched/nohz.h>
37 #include <linux/sched/rseq_api.h>
38 #include <linux/sched/rt.h>
40 #include <linux/blkdev.h>
41 #include <linux/context_tracking.h>
42 #include <linux/cpuset.h>
43 #include <linux/delayacct.h>
44 #include <linux/init_task.h>
45 #include <linux/interrupt.h>
46 #include <linux/ioprio.h>
47 #include <linux/kallsyms.h>
48 #include <linux/kcov.h>
49 #include <linux/kprobes.h>
50 #include <linux/llist_api.h>
51 #include <linux/mmu_context.h>
52 #include <linux/mmzone.h>
53 #include <linux/mutex_api.h>
54 #include <linux/nmi.h>
55 #include <linux/nospec.h>
56 #include <linux/perf_event_api.h>
57 #include <linux/profile.h>
58 #include <linux/psi.h>
59 #include <linux/rcuwait_api.h>
60 #include <linux/sched/wake_q.h>
61 #include <linux/scs.h>
62 #include <linux/slab.h>
63 #include <linux/syscalls.h>
64 #include <linux/vtime.h>
65 #include <linux/wait_api.h>
66 #include <linux/workqueue_api.h>
68 #ifdef CONFIG_PREEMPT_DYNAMIC
69 # ifdef CONFIG_GENERIC_ENTRY
70 # include <linux/entry-common.h>
74 #include <uapi/linux/sched/types.h>
76 #include <asm/irq_regs.h>
77 #include <asm/switch_to.h>
80 #define CREATE_TRACE_POINTS
81 #include <linux/sched/rseq_api.h>
82 #include <trace/events/sched.h>
83 #undef CREATE_TRACE_POINTS
87 #include "autogroup.h"
89 #include "autogroup.h"
94 #include "../workqueue_internal.h"
95 #include "../../io_uring/io-wq.h"
96 #include "../smpboot.h"
99 * Export tracepoints that act as a bare tracehook (ie: have no trace event
100 * associated with them) to allow external modules to probe them.
102 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
103 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
104 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
105 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
106 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
107 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
108 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
109 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
110 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
111 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
112 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
116 #ifdef CONFIG_SCHED_DEBUG
118 * Debugging: various feature bits
120 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
121 * sysctl_sched_features, defined in sched.h, to allow constants propagation
122 * at compile time and compiler optimization based on features default.
124 #define SCHED_FEAT(name, enabled) \
125 (1UL << __SCHED_FEAT_##name) * enabled |
126 const_debug unsigned int sysctl_sched_features =
127 #include "features.h"
132 * Print a warning if need_resched is set for the given duration (if
133 * LATENCY_WARN is enabled).
135 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
138 __read_mostly int sysctl_resched_latency_warn_ms = 100;
139 __read_mostly int sysctl_resched_latency_warn_once = 1;
140 #endif /* CONFIG_SCHED_DEBUG */
143 * Number of tasks to iterate in a single balance run.
144 * Limited because this is done with IRQs disabled.
146 const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
148 __read_mostly int scheduler_running;
150 #ifdef CONFIG_SCHED_CORE
152 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
154 /* kernel prio, less is more */
155 static inline int __task_prio(struct task_struct *p)
157 if (p->sched_class == &stop_sched_class) /* trumps deadline */
160 if (rt_prio(p->prio)) /* includes deadline */
161 return p->prio; /* [-1, 99] */
163 if (p->sched_class == &idle_sched_class)
164 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
166 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
176 /* real prio, less is less */
177 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
180 int pa = __task_prio(a), pb = __task_prio(b);
188 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
189 return !dl_time_before(a->dl.deadline, b->dl.deadline);
191 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
192 return cfs_prio_less(a, b, in_fi);
197 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
199 if (a->core_cookie < b->core_cookie)
202 if (a->core_cookie > b->core_cookie)
205 /* flip prio, so high prio is leftmost */
206 if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
212 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
214 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
216 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
219 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
221 const struct task_struct *p = __node_2_sc(node);
222 unsigned long cookie = (unsigned long)key;
224 if (cookie < p->core_cookie)
227 if (cookie > p->core_cookie)
233 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
235 rq->core->core_task_seq++;
240 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
243 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
245 rq->core->core_task_seq++;
247 if (sched_core_enqueued(p)) {
248 rb_erase(&p->core_node, &rq->core_tree);
249 RB_CLEAR_NODE(&p->core_node);
253 * Migrating the last task off the cpu, with the cpu in forced idle
254 * state. Reschedule to create an accounting edge for forced idle,
255 * and re-examine whether the core is still in forced idle state.
257 if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
258 rq->core->core_forceidle_count && rq->curr == rq->idle)
263 * Find left-most (aka, highest priority) task matching @cookie.
265 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
267 struct rb_node *node;
269 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
271 * The idle task always matches any cookie!
274 return idle_sched_class.pick_task(rq);
276 return __node_2_sc(node);
279 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
281 struct rb_node *node = &p->core_node;
283 node = rb_next(node);
287 p = container_of(node, struct task_struct, core_node);
288 if (p->core_cookie != cookie)
295 * Magic required such that:
297 * raw_spin_rq_lock(rq);
299 * raw_spin_rq_unlock(rq);
301 * ends up locking and unlocking the _same_ lock, and all CPUs
302 * always agree on what rq has what lock.
304 * XXX entirely possible to selectively enable cores, don't bother for now.
307 static DEFINE_MUTEX(sched_core_mutex);
308 static atomic_t sched_core_count;
309 static struct cpumask sched_core_mask;
311 static void sched_core_lock(int cpu, unsigned long *flags)
313 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
316 local_irq_save(*flags);
317 for_each_cpu(t, smt_mask)
318 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
321 static void sched_core_unlock(int cpu, unsigned long *flags)
323 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
326 for_each_cpu(t, smt_mask)
327 raw_spin_unlock(&cpu_rq(t)->__lock);
328 local_irq_restore(*flags);
331 static void __sched_core_flip(bool enabled)
339 * Toggle the online cores, one by one.
341 cpumask_copy(&sched_core_mask, cpu_online_mask);
342 for_each_cpu(cpu, &sched_core_mask) {
343 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
345 sched_core_lock(cpu, &flags);
347 for_each_cpu(t, smt_mask)
348 cpu_rq(t)->core_enabled = enabled;
350 cpu_rq(cpu)->core->core_forceidle_start = 0;
352 sched_core_unlock(cpu, &flags);
354 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
358 * Toggle the offline CPUs.
360 for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
361 cpu_rq(cpu)->core_enabled = enabled;
366 static void sched_core_assert_empty(void)
370 for_each_possible_cpu(cpu)
371 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
374 static void __sched_core_enable(void)
376 static_branch_enable(&__sched_core_enabled);
378 * Ensure all previous instances of raw_spin_rq_*lock() have finished
379 * and future ones will observe !sched_core_disabled().
382 __sched_core_flip(true);
383 sched_core_assert_empty();
386 static void __sched_core_disable(void)
388 sched_core_assert_empty();
389 __sched_core_flip(false);
390 static_branch_disable(&__sched_core_enabled);
393 void sched_core_get(void)
395 if (atomic_inc_not_zero(&sched_core_count))
398 mutex_lock(&sched_core_mutex);
399 if (!atomic_read(&sched_core_count))
400 __sched_core_enable();
402 smp_mb__before_atomic();
403 atomic_inc(&sched_core_count);
404 mutex_unlock(&sched_core_mutex);
407 static void __sched_core_put(struct work_struct *work)
409 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
410 __sched_core_disable();
411 mutex_unlock(&sched_core_mutex);
415 void sched_core_put(void)
417 static DECLARE_WORK(_work, __sched_core_put);
420 * "There can be only one"
422 * Either this is the last one, or we don't actually need to do any
423 * 'work'. If it is the last *again*, we rely on
424 * WORK_STRUCT_PENDING_BIT.
426 if (!atomic_add_unless(&sched_core_count, -1, 1))
427 schedule_work(&_work);
430 #else /* !CONFIG_SCHED_CORE */
432 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
434 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
436 #endif /* CONFIG_SCHED_CORE */
439 * Serialization rules:
445 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
448 * rq2->lock where: rq1 < rq2
452 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
453 * local CPU's rq->lock, it optionally removes the task from the runqueue and
454 * always looks at the local rq data structures to find the most eligible task
457 * Task enqueue is also under rq->lock, possibly taken from another CPU.
458 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
459 * the local CPU to avoid bouncing the runqueue state around [ see
460 * ttwu_queue_wakelist() ]
462 * Task wakeup, specifically wakeups that involve migration, are horribly
463 * complicated to avoid having to take two rq->locks.
467 * System-calls and anything external will use task_rq_lock() which acquires
468 * both p->pi_lock and rq->lock. As a consequence the state they change is
469 * stable while holding either lock:
471 * - sched_setaffinity()/
472 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
473 * - set_user_nice(): p->se.load, p->*prio
474 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
475 * p->se.load, p->rt_priority,
476 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
477 * - sched_setnuma(): p->numa_preferred_nid
478 * - sched_move_task(): p->sched_task_group
479 * - uclamp_update_active() p->uclamp*
481 * p->state <- TASK_*:
483 * is changed locklessly using set_current_state(), __set_current_state() or
484 * set_special_state(), see their respective comments, or by
485 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
488 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
490 * is set by activate_task() and cleared by deactivate_task(), under
491 * rq->lock. Non-zero indicates the task is runnable, the special
492 * ON_RQ_MIGRATING state is used for migration without holding both
493 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
495 * p->on_cpu <- { 0, 1 }:
497 * is set by prepare_task() and cleared by finish_task() such that it will be
498 * set before p is scheduled-in and cleared after p is scheduled-out, both
499 * under rq->lock. Non-zero indicates the task is running on its CPU.
501 * [ The astute reader will observe that it is possible for two tasks on one
502 * CPU to have ->on_cpu = 1 at the same time. ]
504 * task_cpu(p): is changed by set_task_cpu(), the rules are:
506 * - Don't call set_task_cpu() on a blocked task:
508 * We don't care what CPU we're not running on, this simplifies hotplug,
509 * the CPU assignment of blocked tasks isn't required to be valid.
511 * - for try_to_wake_up(), called under p->pi_lock:
513 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
515 * - for migration called under rq->lock:
516 * [ see task_on_rq_migrating() in task_rq_lock() ]
518 * o move_queued_task()
521 * - for migration called under double_rq_lock():
523 * o __migrate_swap_task()
524 * o push_rt_task() / pull_rt_task()
525 * o push_dl_task() / pull_dl_task()
526 * o dl_task_offline_migration()
530 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
532 raw_spinlock_t *lock;
534 /* Matches synchronize_rcu() in __sched_core_enable() */
536 if (sched_core_disabled()) {
537 raw_spin_lock_nested(&rq->__lock, subclass);
538 /* preempt_count *MUST* be > 1 */
539 preempt_enable_no_resched();
544 lock = __rq_lockp(rq);
545 raw_spin_lock_nested(lock, subclass);
546 if (likely(lock == __rq_lockp(rq))) {
547 /* preempt_count *MUST* be > 1 */
548 preempt_enable_no_resched();
551 raw_spin_unlock(lock);
555 bool raw_spin_rq_trylock(struct rq *rq)
557 raw_spinlock_t *lock;
560 /* Matches synchronize_rcu() in __sched_core_enable() */
562 if (sched_core_disabled()) {
563 ret = raw_spin_trylock(&rq->__lock);
569 lock = __rq_lockp(rq);
570 ret = raw_spin_trylock(lock);
571 if (!ret || (likely(lock == __rq_lockp(rq)))) {
575 raw_spin_unlock(lock);
579 void raw_spin_rq_unlock(struct rq *rq)
581 raw_spin_unlock(rq_lockp(rq));
586 * double_rq_lock - safely lock two runqueues
588 void double_rq_lock(struct rq *rq1, struct rq *rq2)
590 lockdep_assert_irqs_disabled();
592 if (rq_order_less(rq2, rq1))
595 raw_spin_rq_lock(rq1);
596 if (__rq_lockp(rq1) != __rq_lockp(rq2))
597 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
599 double_rq_clock_clear_update(rq1, rq2);
604 * __task_rq_lock - lock the rq @p resides on.
606 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
611 lockdep_assert_held(&p->pi_lock);
615 raw_spin_rq_lock(rq);
616 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
620 raw_spin_rq_unlock(rq);
622 while (unlikely(task_on_rq_migrating(p)))
628 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
630 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
631 __acquires(p->pi_lock)
637 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
639 raw_spin_rq_lock(rq);
641 * move_queued_task() task_rq_lock()
644 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
645 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
646 * [S] ->cpu = new_cpu [L] task_rq()
650 * If we observe the old CPU in task_rq_lock(), the acquire of
651 * the old rq->lock will fully serialize against the stores.
653 * If we observe the new CPU in task_rq_lock(), the address
654 * dependency headed by '[L] rq = task_rq()' and the acquire
655 * will pair with the WMB to ensure we then also see migrating.
657 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
661 raw_spin_rq_unlock(rq);
662 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
664 while (unlikely(task_on_rq_migrating(p)))
670 * RQ-clock updating methods:
673 static void update_rq_clock_task(struct rq *rq, s64 delta)
676 * In theory, the compile should just see 0 here, and optimize out the call
677 * to sched_rt_avg_update. But I don't trust it...
679 s64 __maybe_unused steal = 0, irq_delta = 0;
681 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
682 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
685 * Since irq_time is only updated on {soft,}irq_exit, we might run into
686 * this case when a previous update_rq_clock() happened inside a
689 * When this happens, we stop ->clock_task and only update the
690 * prev_irq_time stamp to account for the part that fit, so that a next
691 * update will consume the rest. This ensures ->clock_task is
694 * It does however cause some slight miss-attribution of {soft,}irq
695 * time, a more accurate solution would be to update the irq_time using
696 * the current rq->clock timestamp, except that would require using
699 if (irq_delta > delta)
702 rq->prev_irq_time += irq_delta;
705 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
706 if (static_key_false((¶virt_steal_rq_enabled))) {
707 steal = paravirt_steal_clock(cpu_of(rq));
708 steal -= rq->prev_steal_time_rq;
710 if (unlikely(steal > delta))
713 rq->prev_steal_time_rq += steal;
718 rq->clock_task += delta;
720 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
721 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
722 update_irq_load_avg(rq, irq_delta + steal);
724 update_rq_clock_pelt(rq, delta);
727 void update_rq_clock(struct rq *rq)
731 lockdep_assert_rq_held(rq);
733 if (rq->clock_update_flags & RQCF_ACT_SKIP)
736 #ifdef CONFIG_SCHED_DEBUG
737 if (sched_feat(WARN_DOUBLE_CLOCK))
738 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
739 rq->clock_update_flags |= RQCF_UPDATED;
742 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
746 update_rq_clock_task(rq, delta);
749 #ifdef CONFIG_SCHED_HRTICK
751 * Use HR-timers to deliver accurate preemption points.
754 static void hrtick_clear(struct rq *rq)
756 if (hrtimer_active(&rq->hrtick_timer))
757 hrtimer_cancel(&rq->hrtick_timer);
761 * High-resolution timer tick.
762 * Runs from hardirq context with interrupts disabled.
764 static enum hrtimer_restart hrtick(struct hrtimer *timer)
766 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
769 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
773 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
776 return HRTIMER_NORESTART;
781 static void __hrtick_restart(struct rq *rq)
783 struct hrtimer *timer = &rq->hrtick_timer;
784 ktime_t time = rq->hrtick_time;
786 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
790 * called from hardirq (IPI) context
792 static void __hrtick_start(void *arg)
798 __hrtick_restart(rq);
803 * Called to set the hrtick timer state.
805 * called with rq->lock held and irqs disabled
807 void hrtick_start(struct rq *rq, u64 delay)
809 struct hrtimer *timer = &rq->hrtick_timer;
813 * Don't schedule slices shorter than 10000ns, that just
814 * doesn't make sense and can cause timer DoS.
816 delta = max_t(s64, delay, 10000LL);
817 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
820 __hrtick_restart(rq);
822 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
827 * Called to set the hrtick timer state.
829 * called with rq->lock held and irqs disabled
831 void hrtick_start(struct rq *rq, u64 delay)
834 * Don't schedule slices shorter than 10000ns, that just
835 * doesn't make sense. Rely on vruntime for fairness.
837 delay = max_t(u64, delay, 10000LL);
838 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
839 HRTIMER_MODE_REL_PINNED_HARD);
842 #endif /* CONFIG_SMP */
844 static void hrtick_rq_init(struct rq *rq)
847 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
849 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
850 rq->hrtick_timer.function = hrtick;
852 #else /* CONFIG_SCHED_HRTICK */
853 static inline void hrtick_clear(struct rq *rq)
857 static inline void hrtick_rq_init(struct rq *rq)
860 #endif /* CONFIG_SCHED_HRTICK */
863 * cmpxchg based fetch_or, macro so it works for different integer types
865 #define fetch_or(ptr, mask) \
867 typeof(ptr) _ptr = (ptr); \
868 typeof(mask) _mask = (mask); \
869 typeof(*_ptr) _val = *_ptr; \
872 } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
876 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
878 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
879 * this avoids any races wrt polling state changes and thereby avoids
882 static inline bool set_nr_and_not_polling(struct task_struct *p)
884 struct thread_info *ti = task_thread_info(p);
885 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
889 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
891 * If this returns true, then the idle task promises to call
892 * sched_ttwu_pending() and reschedule soon.
894 static bool set_nr_if_polling(struct task_struct *p)
896 struct thread_info *ti = task_thread_info(p);
897 typeof(ti->flags) val = READ_ONCE(ti->flags);
900 if (!(val & _TIF_POLLING_NRFLAG))
902 if (val & _TIF_NEED_RESCHED)
904 if (try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED))
911 static inline bool set_nr_and_not_polling(struct task_struct *p)
913 set_tsk_need_resched(p);
918 static inline bool set_nr_if_polling(struct task_struct *p)
925 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
927 struct wake_q_node *node = &task->wake_q;
930 * Atomically grab the task, if ->wake_q is !nil already it means
931 * it's already queued (either by us or someone else) and will get the
932 * wakeup due to that.
934 * In order to ensure that a pending wakeup will observe our pending
935 * state, even in the failed case, an explicit smp_mb() must be used.
937 smp_mb__before_atomic();
938 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
942 * The head is context local, there can be no concurrency.
945 head->lastp = &node->next;
950 * wake_q_add() - queue a wakeup for 'later' waking.
951 * @head: the wake_q_head to add @task to
952 * @task: the task to queue for 'later' wakeup
954 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
955 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
958 * This function must be used as-if it were wake_up_process(); IOW the task
959 * must be ready to be woken at this location.
961 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
963 if (__wake_q_add(head, task))
964 get_task_struct(task);
968 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
969 * @head: the wake_q_head to add @task to
970 * @task: the task to queue for 'later' wakeup
972 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
973 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
976 * This function must be used as-if it were wake_up_process(); IOW the task
977 * must be ready to be woken at this location.
979 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
980 * that already hold reference to @task can call the 'safe' version and trust
981 * wake_q to do the right thing depending whether or not the @task is already
984 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
986 if (!__wake_q_add(head, task))
987 put_task_struct(task);
990 void wake_up_q(struct wake_q_head *head)
992 struct wake_q_node *node = head->first;
994 while (node != WAKE_Q_TAIL) {
995 struct task_struct *task;
997 task = container_of(node, struct task_struct, wake_q);
998 /* Task can safely be re-inserted now: */
1000 task->wake_q.next = NULL;
1003 * wake_up_process() executes a full barrier, which pairs with
1004 * the queueing in wake_q_add() so as not to miss wakeups.
1006 wake_up_process(task);
1007 put_task_struct(task);
1012 * resched_curr - mark rq's current task 'to be rescheduled now'.
1014 * On UP this means the setting of the need_resched flag, on SMP it
1015 * might also involve a cross-CPU call to trigger the scheduler on
1018 void resched_curr(struct rq *rq)
1020 struct task_struct *curr = rq->curr;
1023 lockdep_assert_rq_held(rq);
1025 if (test_tsk_need_resched(curr))
1030 if (cpu == smp_processor_id()) {
1031 set_tsk_need_resched(curr);
1032 set_preempt_need_resched();
1036 if (set_nr_and_not_polling(curr))
1037 smp_send_reschedule(cpu);
1039 trace_sched_wake_idle_without_ipi(cpu);
1042 void resched_cpu(int cpu)
1044 struct rq *rq = cpu_rq(cpu);
1045 unsigned long flags;
1047 raw_spin_rq_lock_irqsave(rq, flags);
1048 if (cpu_online(cpu) || cpu == smp_processor_id())
1050 raw_spin_rq_unlock_irqrestore(rq, flags);
1054 #ifdef CONFIG_NO_HZ_COMMON
1056 * In the semi idle case, use the nearest busy CPU for migrating timers
1057 * from an idle CPU. This is good for power-savings.
1059 * We don't do similar optimization for completely idle system, as
1060 * selecting an idle CPU will add more delays to the timers than intended
1061 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1063 int get_nohz_timer_target(void)
1065 int i, cpu = smp_processor_id(), default_cpu = -1;
1066 struct sched_domain *sd;
1067 const struct cpumask *hk_mask;
1069 if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1075 hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1078 for_each_domain(cpu, sd) {
1079 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1090 if (default_cpu == -1)
1091 default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1099 * When add_timer_on() enqueues a timer into the timer wheel of an
1100 * idle CPU then this timer might expire before the next timer event
1101 * which is scheduled to wake up that CPU. In case of a completely
1102 * idle system the next event might even be infinite time into the
1103 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1104 * leaves the inner idle loop so the newly added timer is taken into
1105 * account when the CPU goes back to idle and evaluates the timer
1106 * wheel for the next timer event.
1108 static void wake_up_idle_cpu(int cpu)
1110 struct rq *rq = cpu_rq(cpu);
1112 if (cpu == smp_processor_id())
1115 if (set_nr_and_not_polling(rq->idle))
1116 smp_send_reschedule(cpu);
1118 trace_sched_wake_idle_without_ipi(cpu);
1121 static bool wake_up_full_nohz_cpu(int cpu)
1124 * We just need the target to call irq_exit() and re-evaluate
1125 * the next tick. The nohz full kick at least implies that.
1126 * If needed we can still optimize that later with an
1129 if (cpu_is_offline(cpu))
1130 return true; /* Don't try to wake offline CPUs. */
1131 if (tick_nohz_full_cpu(cpu)) {
1132 if (cpu != smp_processor_id() ||
1133 tick_nohz_tick_stopped())
1134 tick_nohz_full_kick_cpu(cpu);
1142 * Wake up the specified CPU. If the CPU is going offline, it is the
1143 * caller's responsibility to deal with the lost wakeup, for example,
1144 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1146 void wake_up_nohz_cpu(int cpu)
1148 if (!wake_up_full_nohz_cpu(cpu))
1149 wake_up_idle_cpu(cpu);
1152 static void nohz_csd_func(void *info)
1154 struct rq *rq = info;
1155 int cpu = cpu_of(rq);
1159 * Release the rq::nohz_csd.
1161 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1162 WARN_ON(!(flags & NOHZ_KICK_MASK));
1164 rq->idle_balance = idle_cpu(cpu);
1165 if (rq->idle_balance && !need_resched()) {
1166 rq->nohz_idle_balance = flags;
1167 raise_softirq_irqoff(SCHED_SOFTIRQ);
1171 #endif /* CONFIG_NO_HZ_COMMON */
1173 #ifdef CONFIG_NO_HZ_FULL
1174 bool sched_can_stop_tick(struct rq *rq)
1176 int fifo_nr_running;
1178 /* Deadline tasks, even if single, need the tick */
1179 if (rq->dl.dl_nr_running)
1183 * If there are more than one RR tasks, we need the tick to affect the
1184 * actual RR behaviour.
1186 if (rq->rt.rr_nr_running) {
1187 if (rq->rt.rr_nr_running == 1)
1194 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1195 * forced preemption between FIFO tasks.
1197 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1198 if (fifo_nr_running)
1202 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1203 * if there's more than one we need the tick for involuntary
1206 if (rq->nr_running > 1)
1211 #endif /* CONFIG_NO_HZ_FULL */
1212 #endif /* CONFIG_SMP */
1214 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1215 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1217 * Iterate task_group tree rooted at *from, calling @down when first entering a
1218 * node and @up when leaving it for the final time.
1220 * Caller must hold rcu_lock or sufficient equivalent.
1222 int walk_tg_tree_from(struct task_group *from,
1223 tg_visitor down, tg_visitor up, void *data)
1225 struct task_group *parent, *child;
1231 ret = (*down)(parent, data);
1234 list_for_each_entry_rcu(child, &parent->children, siblings) {
1241 ret = (*up)(parent, data);
1242 if (ret || parent == from)
1246 parent = parent->parent;
1253 int tg_nop(struct task_group *tg, void *data)
1259 static void set_load_weight(struct task_struct *p, bool update_load)
1261 int prio = p->static_prio - MAX_RT_PRIO;
1262 struct load_weight *load = &p->se.load;
1265 * SCHED_IDLE tasks get minimal weight:
1267 if (task_has_idle_policy(p)) {
1268 load->weight = scale_load(WEIGHT_IDLEPRIO);
1269 load->inv_weight = WMULT_IDLEPRIO;
1274 * SCHED_OTHER tasks have to update their load when changing their
1277 if (update_load && p->sched_class == &fair_sched_class) {
1278 reweight_task(p, prio);
1280 load->weight = scale_load(sched_prio_to_weight[prio]);
1281 load->inv_weight = sched_prio_to_wmult[prio];
1285 #ifdef CONFIG_UCLAMP_TASK
1287 * Serializes updates of utilization clamp values
1289 * The (slow-path) user-space triggers utilization clamp value updates which
1290 * can require updates on (fast-path) scheduler's data structures used to
1291 * support enqueue/dequeue operations.
1292 * While the per-CPU rq lock protects fast-path update operations, user-space
1293 * requests are serialized using a mutex to reduce the risk of conflicting
1294 * updates or API abuses.
1296 static DEFINE_MUTEX(uclamp_mutex);
1298 /* Max allowed minimum utilization */
1299 static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1301 /* Max allowed maximum utilization */
1302 static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1305 * By default RT tasks run at the maximum performance point/capacity of the
1306 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1307 * SCHED_CAPACITY_SCALE.
1309 * This knob allows admins to change the default behavior when uclamp is being
1310 * used. In battery powered devices, particularly, running at the maximum
1311 * capacity and frequency will increase energy consumption and shorten the
1314 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1316 * This knob will not override the system default sched_util_clamp_min defined
1319 static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1321 /* All clamps are required to be less or equal than these values */
1322 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1325 * This static key is used to reduce the uclamp overhead in the fast path. It
1326 * primarily disables the call to uclamp_rq_{inc, dec}() in
1327 * enqueue/dequeue_task().
1329 * This allows users to continue to enable uclamp in their kernel config with
1330 * minimum uclamp overhead in the fast path.
1332 * As soon as userspace modifies any of the uclamp knobs, the static key is
1333 * enabled, since we have an actual users that make use of uclamp
1336 * The knobs that would enable this static key are:
1338 * * A task modifying its uclamp value with sched_setattr().
1339 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1340 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1342 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1344 /* Integer rounded range for each bucket */
1345 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1347 #define for_each_clamp_id(clamp_id) \
1348 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1350 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1352 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1355 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1357 if (clamp_id == UCLAMP_MIN)
1359 return SCHED_CAPACITY_SCALE;
1362 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1363 unsigned int value, bool user_defined)
1365 uc_se->value = value;
1366 uc_se->bucket_id = uclamp_bucket_id(value);
1367 uc_se->user_defined = user_defined;
1370 static inline unsigned int
1371 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1372 unsigned int clamp_value)
1375 * Avoid blocked utilization pushing up the frequency when we go
1376 * idle (which drops the max-clamp) by retaining the last known
1379 if (clamp_id == UCLAMP_MAX) {
1380 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1384 return uclamp_none(UCLAMP_MIN);
1387 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1388 unsigned int clamp_value)
1390 /* Reset max-clamp retention only on idle exit */
1391 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1394 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1398 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1399 unsigned int clamp_value)
1401 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1402 int bucket_id = UCLAMP_BUCKETS - 1;
1405 * Since both min and max clamps are max aggregated, find the
1406 * top most bucket with tasks in.
1408 for ( ; bucket_id >= 0; bucket_id--) {
1409 if (!bucket[bucket_id].tasks)
1411 return bucket[bucket_id].value;
1414 /* No tasks -- default clamp values */
1415 return uclamp_idle_value(rq, clamp_id, clamp_value);
1418 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1420 unsigned int default_util_min;
1421 struct uclamp_se *uc_se;
1423 lockdep_assert_held(&p->pi_lock);
1425 uc_se = &p->uclamp_req[UCLAMP_MIN];
1427 /* Only sync if user didn't override the default */
1428 if (uc_se->user_defined)
1431 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1432 uclamp_se_set(uc_se, default_util_min, false);
1435 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1443 /* Protect updates to p->uclamp_* */
1444 rq = task_rq_lock(p, &rf);
1445 __uclamp_update_util_min_rt_default(p);
1446 task_rq_unlock(rq, p, &rf);
1449 static inline struct uclamp_se
1450 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1452 /* Copy by value as we could modify it */
1453 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1454 #ifdef CONFIG_UCLAMP_TASK_GROUP
1455 unsigned int tg_min, tg_max, value;
1458 * Tasks in autogroups or root task group will be
1459 * restricted by system defaults.
1461 if (task_group_is_autogroup(task_group(p)))
1463 if (task_group(p) == &root_task_group)
1466 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1467 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1468 value = uc_req.value;
1469 value = clamp(value, tg_min, tg_max);
1470 uclamp_se_set(&uc_req, value, false);
1477 * The effective clamp bucket index of a task depends on, by increasing
1479 * - the task specific clamp value, when explicitly requested from userspace
1480 * - the task group effective clamp value, for tasks not either in the root
1481 * group or in an autogroup
1482 * - the system default clamp value, defined by the sysadmin
1484 static inline struct uclamp_se
1485 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1487 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1488 struct uclamp_se uc_max = uclamp_default[clamp_id];
1490 /* System default restrictions always apply */
1491 if (unlikely(uc_req.value > uc_max.value))
1497 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1499 struct uclamp_se uc_eff;
1501 /* Task currently refcounted: use back-annotated (effective) value */
1502 if (p->uclamp[clamp_id].active)
1503 return (unsigned long)p->uclamp[clamp_id].value;
1505 uc_eff = uclamp_eff_get(p, clamp_id);
1507 return (unsigned long)uc_eff.value;
1511 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1512 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1513 * updates the rq's clamp value if required.
1515 * Tasks can have a task-specific value requested from user-space, track
1516 * within each bucket the maximum value for tasks refcounted in it.
1517 * This "local max aggregation" allows to track the exact "requested" value
1518 * for each bucket when all its RUNNABLE tasks require the same clamp.
1520 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1521 enum uclamp_id clamp_id)
1523 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1524 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1525 struct uclamp_bucket *bucket;
1527 lockdep_assert_rq_held(rq);
1529 /* Update task effective clamp */
1530 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1532 bucket = &uc_rq->bucket[uc_se->bucket_id];
1534 uc_se->active = true;
1536 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1539 * Local max aggregation: rq buckets always track the max
1540 * "requested" clamp value of its RUNNABLE tasks.
1542 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1543 bucket->value = uc_se->value;
1545 if (uc_se->value > READ_ONCE(uc_rq->value))
1546 WRITE_ONCE(uc_rq->value, uc_se->value);
1550 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1551 * is released. If this is the last task reference counting the rq's max
1552 * active clamp value, then the rq's clamp value is updated.
1554 * Both refcounted tasks and rq's cached clamp values are expected to be
1555 * always valid. If it's detected they are not, as defensive programming,
1556 * enforce the expected state and warn.
1558 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1559 enum uclamp_id clamp_id)
1561 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1562 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1563 struct uclamp_bucket *bucket;
1564 unsigned int bkt_clamp;
1565 unsigned int rq_clamp;
1567 lockdep_assert_rq_held(rq);
1570 * If sched_uclamp_used was enabled after task @p was enqueued,
1571 * we could end up with unbalanced call to uclamp_rq_dec_id().
1573 * In this case the uc_se->active flag should be false since no uclamp
1574 * accounting was performed at enqueue time and we can just return
1577 * Need to be careful of the following enqueue/dequeue ordering
1581 * // sched_uclamp_used gets enabled
1584 * // Must not decrement bucket->tasks here
1587 * where we could end up with stale data in uc_se and
1588 * bucket[uc_se->bucket_id].
1590 * The following check here eliminates the possibility of such race.
1592 if (unlikely(!uc_se->active))
1595 bucket = &uc_rq->bucket[uc_se->bucket_id];
1597 SCHED_WARN_ON(!bucket->tasks);
1598 if (likely(bucket->tasks))
1601 uc_se->active = false;
1604 * Keep "local max aggregation" simple and accept to (possibly)
1605 * overboost some RUNNABLE tasks in the same bucket.
1606 * The rq clamp bucket value is reset to its base value whenever
1607 * there are no more RUNNABLE tasks refcounting it.
1609 if (likely(bucket->tasks))
1612 rq_clamp = READ_ONCE(uc_rq->value);
1614 * Defensive programming: this should never happen. If it happens,
1615 * e.g. due to future modification, warn and fixup the expected value.
1617 SCHED_WARN_ON(bucket->value > rq_clamp);
1618 if (bucket->value >= rq_clamp) {
1619 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1620 WRITE_ONCE(uc_rq->value, bkt_clamp);
1624 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1626 enum uclamp_id clamp_id;
1629 * Avoid any overhead until uclamp is actually used by the userspace.
1631 * The condition is constructed such that a NOP is generated when
1632 * sched_uclamp_used is disabled.
1634 if (!static_branch_unlikely(&sched_uclamp_used))
1637 if (unlikely(!p->sched_class->uclamp_enabled))
1640 for_each_clamp_id(clamp_id)
1641 uclamp_rq_inc_id(rq, p, clamp_id);
1643 /* Reset clamp idle holding when there is one RUNNABLE task */
1644 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1645 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1648 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1650 enum uclamp_id clamp_id;
1653 * Avoid any overhead until uclamp is actually used by the userspace.
1655 * The condition is constructed such that a NOP is generated when
1656 * sched_uclamp_used is disabled.
1658 if (!static_branch_unlikely(&sched_uclamp_used))
1661 if (unlikely(!p->sched_class->uclamp_enabled))
1664 for_each_clamp_id(clamp_id)
1665 uclamp_rq_dec_id(rq, p, clamp_id);
1668 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1669 enum uclamp_id clamp_id)
1671 if (!p->uclamp[clamp_id].active)
1674 uclamp_rq_dec_id(rq, p, clamp_id);
1675 uclamp_rq_inc_id(rq, p, clamp_id);
1678 * Make sure to clear the idle flag if we've transiently reached 0
1679 * active tasks on rq.
1681 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1682 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1686 uclamp_update_active(struct task_struct *p)
1688 enum uclamp_id clamp_id;
1693 * Lock the task and the rq where the task is (or was) queued.
1695 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1696 * price to pay to safely serialize util_{min,max} updates with
1697 * enqueues, dequeues and migration operations.
1698 * This is the same locking schema used by __set_cpus_allowed_ptr().
1700 rq = task_rq_lock(p, &rf);
1703 * Setting the clamp bucket is serialized by task_rq_lock().
1704 * If the task is not yet RUNNABLE and its task_struct is not
1705 * affecting a valid clamp bucket, the next time it's enqueued,
1706 * it will already see the updated clamp bucket value.
1708 for_each_clamp_id(clamp_id)
1709 uclamp_rq_reinc_id(rq, p, clamp_id);
1711 task_rq_unlock(rq, p, &rf);
1714 #ifdef CONFIG_UCLAMP_TASK_GROUP
1716 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1718 struct css_task_iter it;
1719 struct task_struct *p;
1721 css_task_iter_start(css, 0, &it);
1722 while ((p = css_task_iter_next(&it)))
1723 uclamp_update_active(p);
1724 css_task_iter_end(&it);
1727 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1730 #ifdef CONFIG_SYSCTL
1731 #ifdef CONFIG_UCLAMP_TASK
1732 #ifdef CONFIG_UCLAMP_TASK_GROUP
1733 static void uclamp_update_root_tg(void)
1735 struct task_group *tg = &root_task_group;
1737 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1738 sysctl_sched_uclamp_util_min, false);
1739 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1740 sysctl_sched_uclamp_util_max, false);
1743 cpu_util_update_eff(&root_task_group.css);
1747 static void uclamp_update_root_tg(void) { }
1750 static void uclamp_sync_util_min_rt_default(void)
1752 struct task_struct *g, *p;
1755 * copy_process() sysctl_uclamp
1756 * uclamp_min_rt = X;
1757 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1758 * // link thread smp_mb__after_spinlock()
1759 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1760 * sched_post_fork() for_each_process_thread()
1761 * __uclamp_sync_rt() __uclamp_sync_rt()
1763 * Ensures that either sched_post_fork() will observe the new
1764 * uclamp_min_rt or for_each_process_thread() will observe the new
1767 read_lock(&tasklist_lock);
1768 smp_mb__after_spinlock();
1769 read_unlock(&tasklist_lock);
1772 for_each_process_thread(g, p)
1773 uclamp_update_util_min_rt_default(p);
1777 static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1778 void *buffer, size_t *lenp, loff_t *ppos)
1780 bool update_root_tg = false;
1781 int old_min, old_max, old_min_rt;
1784 mutex_lock(&uclamp_mutex);
1785 old_min = sysctl_sched_uclamp_util_min;
1786 old_max = sysctl_sched_uclamp_util_max;
1787 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1789 result = proc_dointvec(table, write, buffer, lenp, ppos);
1795 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1796 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1797 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1803 if (old_min != sysctl_sched_uclamp_util_min) {
1804 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1805 sysctl_sched_uclamp_util_min, false);
1806 update_root_tg = true;
1808 if (old_max != sysctl_sched_uclamp_util_max) {
1809 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1810 sysctl_sched_uclamp_util_max, false);
1811 update_root_tg = true;
1814 if (update_root_tg) {
1815 static_branch_enable(&sched_uclamp_used);
1816 uclamp_update_root_tg();
1819 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1820 static_branch_enable(&sched_uclamp_used);
1821 uclamp_sync_util_min_rt_default();
1825 * We update all RUNNABLE tasks only when task groups are in use.
1826 * Otherwise, keep it simple and do just a lazy update at each next
1827 * task enqueue time.
1833 sysctl_sched_uclamp_util_min = old_min;
1834 sysctl_sched_uclamp_util_max = old_max;
1835 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1837 mutex_unlock(&uclamp_mutex);
1844 static int uclamp_validate(struct task_struct *p,
1845 const struct sched_attr *attr)
1847 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1848 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1850 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1851 util_min = attr->sched_util_min;
1853 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1857 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1858 util_max = attr->sched_util_max;
1860 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1864 if (util_min != -1 && util_max != -1 && util_min > util_max)
1868 * We have valid uclamp attributes; make sure uclamp is enabled.
1870 * We need to do that here, because enabling static branches is a
1871 * blocking operation which obviously cannot be done while holding
1874 static_branch_enable(&sched_uclamp_used);
1879 static bool uclamp_reset(const struct sched_attr *attr,
1880 enum uclamp_id clamp_id,
1881 struct uclamp_se *uc_se)
1883 /* Reset on sched class change for a non user-defined clamp value. */
1884 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1885 !uc_se->user_defined)
1888 /* Reset on sched_util_{min,max} == -1. */
1889 if (clamp_id == UCLAMP_MIN &&
1890 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1891 attr->sched_util_min == -1) {
1895 if (clamp_id == UCLAMP_MAX &&
1896 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1897 attr->sched_util_max == -1) {
1904 static void __setscheduler_uclamp(struct task_struct *p,
1905 const struct sched_attr *attr)
1907 enum uclamp_id clamp_id;
1909 for_each_clamp_id(clamp_id) {
1910 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1913 if (!uclamp_reset(attr, clamp_id, uc_se))
1917 * RT by default have a 100% boost value that could be modified
1920 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1921 value = sysctl_sched_uclamp_util_min_rt_default;
1923 value = uclamp_none(clamp_id);
1925 uclamp_se_set(uc_se, value, false);
1929 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1932 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1933 attr->sched_util_min != -1) {
1934 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1935 attr->sched_util_min, true);
1938 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1939 attr->sched_util_max != -1) {
1940 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1941 attr->sched_util_max, true);
1945 static void uclamp_fork(struct task_struct *p)
1947 enum uclamp_id clamp_id;
1950 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1951 * as the task is still at its early fork stages.
1953 for_each_clamp_id(clamp_id)
1954 p->uclamp[clamp_id].active = false;
1956 if (likely(!p->sched_reset_on_fork))
1959 for_each_clamp_id(clamp_id) {
1960 uclamp_se_set(&p->uclamp_req[clamp_id],
1961 uclamp_none(clamp_id), false);
1965 static void uclamp_post_fork(struct task_struct *p)
1967 uclamp_update_util_min_rt_default(p);
1970 static void __init init_uclamp_rq(struct rq *rq)
1972 enum uclamp_id clamp_id;
1973 struct uclamp_rq *uc_rq = rq->uclamp;
1975 for_each_clamp_id(clamp_id) {
1976 uc_rq[clamp_id] = (struct uclamp_rq) {
1977 .value = uclamp_none(clamp_id)
1981 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1984 static void __init init_uclamp(void)
1986 struct uclamp_se uc_max = {};
1987 enum uclamp_id clamp_id;
1990 for_each_possible_cpu(cpu)
1991 init_uclamp_rq(cpu_rq(cpu));
1993 for_each_clamp_id(clamp_id) {
1994 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1995 uclamp_none(clamp_id), false);
1998 /* System defaults allow max clamp values for both indexes */
1999 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2000 for_each_clamp_id(clamp_id) {
2001 uclamp_default[clamp_id] = uc_max;
2002 #ifdef CONFIG_UCLAMP_TASK_GROUP
2003 root_task_group.uclamp_req[clamp_id] = uc_max;
2004 root_task_group.uclamp[clamp_id] = uc_max;
2009 #else /* CONFIG_UCLAMP_TASK */
2010 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2011 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
2012 static inline int uclamp_validate(struct task_struct *p,
2013 const struct sched_attr *attr)
2017 static void __setscheduler_uclamp(struct task_struct *p,
2018 const struct sched_attr *attr) { }
2019 static inline void uclamp_fork(struct task_struct *p) { }
2020 static inline void uclamp_post_fork(struct task_struct *p) { }
2021 static inline void init_uclamp(void) { }
2022 #endif /* CONFIG_UCLAMP_TASK */
2024 bool sched_task_on_rq(struct task_struct *p)
2026 return task_on_rq_queued(p);
2029 unsigned long get_wchan(struct task_struct *p)
2031 unsigned long ip = 0;
2034 if (!p || p == current)
2037 /* Only get wchan if task is blocked and we can keep it that way. */
2038 raw_spin_lock_irq(&p->pi_lock);
2039 state = READ_ONCE(p->__state);
2040 smp_rmb(); /* see try_to_wake_up() */
2041 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2042 ip = __get_wchan(p);
2043 raw_spin_unlock_irq(&p->pi_lock);
2048 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2050 if (!(flags & ENQUEUE_NOCLOCK))
2051 update_rq_clock(rq);
2053 if (!(flags & ENQUEUE_RESTORE)) {
2054 sched_info_enqueue(rq, p);
2055 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
2058 uclamp_rq_inc(rq, p);
2059 p->sched_class->enqueue_task(rq, p, flags);
2061 if (sched_core_enabled(rq))
2062 sched_core_enqueue(rq, p);
2065 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2067 if (sched_core_enabled(rq))
2068 sched_core_dequeue(rq, p, flags);
2070 if (!(flags & DEQUEUE_NOCLOCK))
2071 update_rq_clock(rq);
2073 if (!(flags & DEQUEUE_SAVE)) {
2074 sched_info_dequeue(rq, p);
2075 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2078 uclamp_rq_dec(rq, p);
2079 p->sched_class->dequeue_task(rq, p, flags);
2082 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2084 enqueue_task(rq, p, flags);
2086 p->on_rq = TASK_ON_RQ_QUEUED;
2089 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2091 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2093 dequeue_task(rq, p, flags);
2096 static inline int __normal_prio(int policy, int rt_prio, int nice)
2100 if (dl_policy(policy))
2101 prio = MAX_DL_PRIO - 1;
2102 else if (rt_policy(policy))
2103 prio = MAX_RT_PRIO - 1 - rt_prio;
2105 prio = NICE_TO_PRIO(nice);
2111 * Calculate the expected normal priority: i.e. priority
2112 * without taking RT-inheritance into account. Might be
2113 * boosted by interactivity modifiers. Changes upon fork,
2114 * setprio syscalls, and whenever the interactivity
2115 * estimator recalculates.
2117 static inline int normal_prio(struct task_struct *p)
2119 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2123 * Calculate the current priority, i.e. the priority
2124 * taken into account by the scheduler. This value might
2125 * be boosted by RT tasks, or might be boosted by
2126 * interactivity modifiers. Will be RT if the task got
2127 * RT-boosted. If not then it returns p->normal_prio.
2129 static int effective_prio(struct task_struct *p)
2131 p->normal_prio = normal_prio(p);
2133 * If we are RT tasks or we were boosted to RT priority,
2134 * keep the priority unchanged. Otherwise, update priority
2135 * to the normal priority:
2137 if (!rt_prio(p->prio))
2138 return p->normal_prio;
2143 * task_curr - is this task currently executing on a CPU?
2144 * @p: the task in question.
2146 * Return: 1 if the task is currently executing. 0 otherwise.
2148 inline int task_curr(const struct task_struct *p)
2150 return cpu_curr(task_cpu(p)) == p;
2154 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2155 * use the balance_callback list if you want balancing.
2157 * this means any call to check_class_changed() must be followed by a call to
2158 * balance_callback().
2160 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2161 const struct sched_class *prev_class,
2164 if (prev_class != p->sched_class) {
2165 if (prev_class->switched_from)
2166 prev_class->switched_from(rq, p);
2168 p->sched_class->switched_to(rq, p);
2169 } else if (oldprio != p->prio || dl_task(p))
2170 p->sched_class->prio_changed(rq, p, oldprio);
2173 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2175 if (p->sched_class == rq->curr->sched_class)
2176 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2177 else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2181 * A queue event has occurred, and we're going to schedule. In
2182 * this case, we can save a useless back to back clock update.
2184 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2185 rq_clock_skip_update(rq);
2191 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2193 static int __set_cpus_allowed_ptr(struct task_struct *p,
2194 const struct cpumask *new_mask,
2197 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2199 if (likely(!p->migration_disabled))
2202 if (p->cpus_ptr != &p->cpus_mask)
2206 * Violates locking rules! see comment in __do_set_cpus_allowed().
2208 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2211 void migrate_disable(void)
2213 struct task_struct *p = current;
2215 if (p->migration_disabled) {
2216 p->migration_disabled++;
2221 this_rq()->nr_pinned++;
2222 p->migration_disabled = 1;
2225 EXPORT_SYMBOL_GPL(migrate_disable);
2227 void migrate_enable(void)
2229 struct task_struct *p = current;
2231 if (p->migration_disabled > 1) {
2232 p->migration_disabled--;
2236 if (WARN_ON_ONCE(!p->migration_disabled))
2240 * Ensure stop_task runs either before or after this, and that
2241 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2244 if (p->cpus_ptr != &p->cpus_mask)
2245 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2247 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2248 * regular cpus_mask, otherwise things that race (eg.
2249 * select_fallback_rq) get confused.
2252 p->migration_disabled = 0;
2253 this_rq()->nr_pinned--;
2256 EXPORT_SYMBOL_GPL(migrate_enable);
2258 static inline bool rq_has_pinned_tasks(struct rq *rq)
2260 return rq->nr_pinned;
2264 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2265 * __set_cpus_allowed_ptr() and select_fallback_rq().
2267 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2269 /* When not in the task's cpumask, no point in looking further. */
2270 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2273 /* migrate_disabled() must be allowed to finish. */
2274 if (is_migration_disabled(p))
2275 return cpu_online(cpu);
2277 /* Non kernel threads are not allowed during either online or offline. */
2278 if (!(p->flags & PF_KTHREAD))
2279 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2281 /* KTHREAD_IS_PER_CPU is always allowed. */
2282 if (kthread_is_per_cpu(p))
2283 return cpu_online(cpu);
2285 /* Regular kernel threads don't get to stay during offline. */
2289 /* But are allowed during online. */
2290 return cpu_online(cpu);
2294 * This is how migration works:
2296 * 1) we invoke migration_cpu_stop() on the target CPU using
2298 * 2) stopper starts to run (implicitly forcing the migrated thread
2300 * 3) it checks whether the migrated task is still in the wrong runqueue.
2301 * 4) if it's in the wrong runqueue then the migration thread removes
2302 * it and puts it into the right queue.
2303 * 5) stopper completes and stop_one_cpu() returns and the migration
2308 * move_queued_task - move a queued task to new rq.
2310 * Returns (locked) new rq. Old rq's lock is released.
2312 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2313 struct task_struct *p, int new_cpu)
2315 lockdep_assert_rq_held(rq);
2317 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2318 set_task_cpu(p, new_cpu);
2321 rq = cpu_rq(new_cpu);
2324 WARN_ON_ONCE(task_cpu(p) != new_cpu);
2325 activate_task(rq, p, 0);
2326 check_preempt_curr(rq, p, 0);
2331 struct migration_arg {
2332 struct task_struct *task;
2334 struct set_affinity_pending *pending;
2338 * @refs: number of wait_for_completion()
2339 * @stop_pending: is @stop_work in use
2341 struct set_affinity_pending {
2343 unsigned int stop_pending;
2344 struct completion done;
2345 struct cpu_stop_work stop_work;
2346 struct migration_arg arg;
2350 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2351 * this because either it can't run here any more (set_cpus_allowed()
2352 * away from this CPU, or CPU going down), or because we're
2353 * attempting to rebalance this task on exec (sched_exec).
2355 * So we race with normal scheduler movements, but that's OK, as long
2356 * as the task is no longer on this CPU.
2358 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2359 struct task_struct *p, int dest_cpu)
2361 /* Affinity changed (again). */
2362 if (!is_cpu_allowed(p, dest_cpu))
2365 update_rq_clock(rq);
2366 rq = move_queued_task(rq, rf, p, dest_cpu);
2372 * migration_cpu_stop - this will be executed by a highprio stopper thread
2373 * and performs thread migration by bumping thread off CPU then
2374 * 'pushing' onto another runqueue.
2376 static int migration_cpu_stop(void *data)
2378 struct migration_arg *arg = data;
2379 struct set_affinity_pending *pending = arg->pending;
2380 struct task_struct *p = arg->task;
2381 struct rq *rq = this_rq();
2382 bool complete = false;
2386 * The original target CPU might have gone down and we might
2387 * be on another CPU but it doesn't matter.
2389 local_irq_save(rf.flags);
2391 * We need to explicitly wake pending tasks before running
2392 * __migrate_task() such that we will not miss enforcing cpus_ptr
2393 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2395 flush_smp_call_function_queue();
2397 raw_spin_lock(&p->pi_lock);
2401 * If we were passed a pending, then ->stop_pending was set, thus
2402 * p->migration_pending must have remained stable.
2404 WARN_ON_ONCE(pending && pending != p->migration_pending);
2407 * If task_rq(p) != rq, it cannot be migrated here, because we're
2408 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2409 * we're holding p->pi_lock.
2411 if (task_rq(p) == rq) {
2412 if (is_migration_disabled(p))
2416 p->migration_pending = NULL;
2419 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2423 if (task_on_rq_queued(p))
2424 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2426 p->wake_cpu = arg->dest_cpu;
2429 * XXX __migrate_task() can fail, at which point we might end
2430 * up running on a dodgy CPU, AFAICT this can only happen
2431 * during CPU hotplug, at which point we'll get pushed out
2432 * anyway, so it's probably not a big deal.
2435 } else if (pending) {
2437 * This happens when we get migrated between migrate_enable()'s
2438 * preempt_enable() and scheduling the stopper task. At that
2439 * point we're a regular task again and not current anymore.
2441 * A !PREEMPT kernel has a giant hole here, which makes it far
2446 * The task moved before the stopper got to run. We're holding
2447 * ->pi_lock, so the allowed mask is stable - if it got
2448 * somewhere allowed, we're done.
2450 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2451 p->migration_pending = NULL;
2457 * When migrate_enable() hits a rq mis-match we can't reliably
2458 * determine is_migration_disabled() and so have to chase after
2461 WARN_ON_ONCE(!pending->stop_pending);
2462 task_rq_unlock(rq, p, &rf);
2463 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2464 &pending->arg, &pending->stop_work);
2469 pending->stop_pending = false;
2470 task_rq_unlock(rq, p, &rf);
2473 complete_all(&pending->done);
2478 int push_cpu_stop(void *arg)
2480 struct rq *lowest_rq = NULL, *rq = this_rq();
2481 struct task_struct *p = arg;
2483 raw_spin_lock_irq(&p->pi_lock);
2484 raw_spin_rq_lock(rq);
2486 if (task_rq(p) != rq)
2489 if (is_migration_disabled(p)) {
2490 p->migration_flags |= MDF_PUSH;
2494 p->migration_flags &= ~MDF_PUSH;
2496 if (p->sched_class->find_lock_rq)
2497 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2502 // XXX validate p is still the highest prio task
2503 if (task_rq(p) == rq) {
2504 deactivate_task(rq, p, 0);
2505 set_task_cpu(p, lowest_rq->cpu);
2506 activate_task(lowest_rq, p, 0);
2507 resched_curr(lowest_rq);
2510 double_unlock_balance(rq, lowest_rq);
2513 rq->push_busy = false;
2514 raw_spin_rq_unlock(rq);
2515 raw_spin_unlock_irq(&p->pi_lock);
2522 * sched_class::set_cpus_allowed must do the below, but is not required to
2523 * actually call this function.
2525 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2527 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2528 p->cpus_ptr = new_mask;
2532 cpumask_copy(&p->cpus_mask, new_mask);
2533 p->nr_cpus_allowed = cpumask_weight(new_mask);
2537 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2539 struct rq *rq = task_rq(p);
2540 bool queued, running;
2543 * This here violates the locking rules for affinity, since we're only
2544 * supposed to change these variables while holding both rq->lock and
2547 * HOWEVER, it magically works, because ttwu() is the only code that
2548 * accesses these variables under p->pi_lock and only does so after
2549 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2550 * before finish_task().
2552 * XXX do further audits, this smells like something putrid.
2554 if (flags & SCA_MIGRATE_DISABLE)
2555 SCHED_WARN_ON(!p->on_cpu);
2557 lockdep_assert_held(&p->pi_lock);
2559 queued = task_on_rq_queued(p);
2560 running = task_current(rq, p);
2564 * Because __kthread_bind() calls this on blocked tasks without
2567 lockdep_assert_rq_held(rq);
2568 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2571 put_prev_task(rq, p);
2573 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2576 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2578 set_next_task(rq, p);
2581 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2583 __do_set_cpus_allowed(p, new_mask, 0);
2586 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2589 if (!src->user_cpus_ptr)
2592 dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2593 if (!dst->user_cpus_ptr)
2596 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2600 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2602 struct cpumask *user_mask = NULL;
2604 swap(p->user_cpus_ptr, user_mask);
2609 void release_user_cpus_ptr(struct task_struct *p)
2611 kfree(clear_user_cpus_ptr(p));
2615 * This function is wildly self concurrent; here be dragons.
2618 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2619 * designated task is enqueued on an allowed CPU. If that task is currently
2620 * running, we have to kick it out using the CPU stopper.
2622 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2625 * Initial conditions: P0->cpus_mask = [0, 1]
2629 * migrate_disable();
2631 * set_cpus_allowed_ptr(P0, [1]);
2633 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2634 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2635 * This means we need the following scheme:
2639 * migrate_disable();
2641 * set_cpus_allowed_ptr(P0, [1]);
2645 * __set_cpus_allowed_ptr();
2646 * <wakes local stopper>
2647 * `--> <woken on migration completion>
2649 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2650 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2651 * task p are serialized by p->pi_lock, which we can leverage: the one that
2652 * should come into effect at the end of the Migrate-Disable region is the last
2653 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2654 * but we still need to properly signal those waiting tasks at the appropriate
2657 * This is implemented using struct set_affinity_pending. The first
2658 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2659 * setup an instance of that struct and install it on the targeted task_struct.
2660 * Any and all further callers will reuse that instance. Those then wait for
2661 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2662 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2665 * (1) In the cases covered above. There is one more where the completion is
2666 * signaled within affine_move_task() itself: when a subsequent affinity request
2667 * occurs after the stopper bailed out due to the targeted task still being
2668 * Migrate-Disable. Consider:
2670 * Initial conditions: P0->cpus_mask = [0, 1]
2674 * migrate_disable();
2676 * set_cpus_allowed_ptr(P0, [1]);
2679 * migration_cpu_stop()
2680 * is_migration_disabled()
2682 * set_cpus_allowed_ptr(P0, [0, 1]);
2683 * <signal completion>
2686 * Note that the above is safe vs a concurrent migrate_enable(), as any
2687 * pending affinity completion is preceded by an uninstallation of
2688 * p->migration_pending done with p->pi_lock held.
2690 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2691 int dest_cpu, unsigned int flags)
2693 struct set_affinity_pending my_pending = { }, *pending = NULL;
2694 bool stop_pending, complete = false;
2696 /* Can the task run on the task's current CPU? If so, we're done */
2697 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2698 struct task_struct *push_task = NULL;
2700 if ((flags & SCA_MIGRATE_ENABLE) &&
2701 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2702 rq->push_busy = true;
2703 push_task = get_task_struct(p);
2707 * If there are pending waiters, but no pending stop_work,
2708 * then complete now.
2710 pending = p->migration_pending;
2711 if (pending && !pending->stop_pending) {
2712 p->migration_pending = NULL;
2716 task_rq_unlock(rq, p, rf);
2719 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2724 complete_all(&pending->done);
2729 if (!(flags & SCA_MIGRATE_ENABLE)) {
2730 /* serialized by p->pi_lock */
2731 if (!p->migration_pending) {
2732 /* Install the request */
2733 refcount_set(&my_pending.refs, 1);
2734 init_completion(&my_pending.done);
2735 my_pending.arg = (struct migration_arg) {
2737 .dest_cpu = dest_cpu,
2738 .pending = &my_pending,
2741 p->migration_pending = &my_pending;
2743 pending = p->migration_pending;
2744 refcount_inc(&pending->refs);
2746 * Affinity has changed, but we've already installed a
2747 * pending. migration_cpu_stop() *must* see this, else
2748 * we risk a completion of the pending despite having a
2749 * task on a disallowed CPU.
2751 * Serialized by p->pi_lock, so this is safe.
2753 pending->arg.dest_cpu = dest_cpu;
2756 pending = p->migration_pending;
2758 * - !MIGRATE_ENABLE:
2759 * we'll have installed a pending if there wasn't one already.
2762 * we're here because the current CPU isn't matching anymore,
2763 * the only way that can happen is because of a concurrent
2764 * set_cpus_allowed_ptr() call, which should then still be
2765 * pending completion.
2767 * Either way, we really should have a @pending here.
2769 if (WARN_ON_ONCE(!pending)) {
2770 task_rq_unlock(rq, p, rf);
2774 if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2776 * MIGRATE_ENABLE gets here because 'p == current', but for
2777 * anything else we cannot do is_migration_disabled(), punt
2778 * and have the stopper function handle it all race-free.
2780 stop_pending = pending->stop_pending;
2782 pending->stop_pending = true;
2784 if (flags & SCA_MIGRATE_ENABLE)
2785 p->migration_flags &= ~MDF_PUSH;
2787 task_rq_unlock(rq, p, rf);
2789 if (!stop_pending) {
2790 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2791 &pending->arg, &pending->stop_work);
2794 if (flags & SCA_MIGRATE_ENABLE)
2798 if (!is_migration_disabled(p)) {
2799 if (task_on_rq_queued(p))
2800 rq = move_queued_task(rq, rf, p, dest_cpu);
2802 if (!pending->stop_pending) {
2803 p->migration_pending = NULL;
2807 task_rq_unlock(rq, p, rf);
2810 complete_all(&pending->done);
2813 wait_for_completion(&pending->done);
2815 if (refcount_dec_and_test(&pending->refs))
2816 wake_up_var(&pending->refs); /* No UaF, just an address */
2819 * Block the original owner of &pending until all subsequent callers
2820 * have seen the completion and decremented the refcount
2822 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2825 WARN_ON_ONCE(my_pending.stop_pending);
2831 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2833 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2834 const struct cpumask *new_mask,
2837 struct rq_flags *rf)
2838 __releases(rq->lock)
2839 __releases(p->pi_lock)
2841 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2842 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2843 bool kthread = p->flags & PF_KTHREAD;
2844 struct cpumask *user_mask = NULL;
2845 unsigned int dest_cpu;
2848 update_rq_clock(rq);
2850 if (kthread || is_migration_disabled(p)) {
2852 * Kernel threads are allowed on online && !active CPUs,
2853 * however, during cpu-hot-unplug, even these might get pushed
2854 * away if not KTHREAD_IS_PER_CPU.
2856 * Specifically, migration_disabled() tasks must not fail the
2857 * cpumask_any_and_distribute() pick below, esp. so on
2858 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2859 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2861 cpu_valid_mask = cpu_online_mask;
2864 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2870 * Must re-check here, to close a race against __kthread_bind(),
2871 * sched_setaffinity() is not guaranteed to observe the flag.
2873 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2878 if (!(flags & SCA_MIGRATE_ENABLE)) {
2879 if (cpumask_equal(&p->cpus_mask, new_mask))
2882 if (WARN_ON_ONCE(p == current &&
2883 is_migration_disabled(p) &&
2884 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2891 * Picking a ~random cpu helps in cases where we are changing affinity
2892 * for groups of tasks (ie. cpuset), so that load balancing is not
2893 * immediately required to distribute the tasks within their new mask.
2895 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2896 if (dest_cpu >= nr_cpu_ids) {
2901 __do_set_cpus_allowed(p, new_mask, flags);
2903 if (flags & SCA_USER)
2904 user_mask = clear_user_cpus_ptr(p);
2906 ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2913 task_rq_unlock(rq, p, rf);
2919 * Change a given task's CPU affinity. Migrate the thread to a
2920 * proper CPU and schedule it away if the CPU it's executing on
2921 * is removed from the allowed bitmask.
2923 * NOTE: the caller must have a valid reference to the task, the
2924 * task must not exit() & deallocate itself prematurely. The
2925 * call is not atomic; no spinlocks may be held.
2927 static int __set_cpus_allowed_ptr(struct task_struct *p,
2928 const struct cpumask *new_mask, u32 flags)
2933 rq = task_rq_lock(p, &rf);
2934 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2937 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2939 return __set_cpus_allowed_ptr(p, new_mask, 0);
2941 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2944 * Change a given task's CPU affinity to the intersection of its current
2945 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2946 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2947 * If the resulting mask is empty, leave the affinity unchanged and return
2950 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2951 struct cpumask *new_mask,
2952 const struct cpumask *subset_mask)
2954 struct cpumask *user_mask = NULL;
2959 if (!p->user_cpus_ptr) {
2960 user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2965 rq = task_rq_lock(p, &rf);
2968 * Forcefully restricting the affinity of a deadline task is
2969 * likely to cause problems, so fail and noisily override the
2972 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2977 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2983 * We're about to butcher the task affinity, so keep track of what
2984 * the user asked for in case we're able to restore it later on.
2987 cpumask_copy(user_mask, p->cpus_ptr);
2988 p->user_cpus_ptr = user_mask;
2991 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
2994 task_rq_unlock(rq, p, &rf);
3000 * Restrict the CPU affinity of task @p so that it is a subset of
3001 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
3002 * old affinity mask. If the resulting mask is empty, we warn and walk
3003 * up the cpuset hierarchy until we find a suitable mask.
3005 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3007 cpumask_var_t new_mask;
3008 const struct cpumask *override_mask = task_cpu_possible_mask(p);
3010 alloc_cpumask_var(&new_mask, GFP_KERNEL);
3013 * __migrate_task() can fail silently in the face of concurrent
3014 * offlining of the chosen destination CPU, so take the hotplug
3015 * lock to ensure that the migration succeeds.
3018 if (!cpumask_available(new_mask))
3021 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3025 * We failed to find a valid subset of the affinity mask for the
3026 * task, so override it based on its cpuset hierarchy.
3028 cpuset_cpus_allowed(p, new_mask);
3029 override_mask = new_mask;
3032 if (printk_ratelimit()) {
3033 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3034 task_pid_nr(p), p->comm,
3035 cpumask_pr_args(override_mask));
3038 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3041 free_cpumask_var(new_mask);
3045 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
3048 * Restore the affinity of a task @p which was previously restricted by a
3049 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
3050 * @p->user_cpus_ptr.
3052 * It is the caller's responsibility to serialise this with any calls to
3053 * force_compatible_cpus_allowed_ptr(@p).
3055 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3057 struct cpumask *user_mask = p->user_cpus_ptr;
3058 unsigned long flags;
3061 * Try to restore the old affinity mask. If this fails, then
3062 * we free the mask explicitly to avoid it being inherited across
3063 * a subsequent fork().
3065 if (!user_mask || !__sched_setaffinity(p, user_mask))
3068 raw_spin_lock_irqsave(&p->pi_lock, flags);
3069 user_mask = clear_user_cpus_ptr(p);
3070 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3075 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3077 #ifdef CONFIG_SCHED_DEBUG
3078 unsigned int state = READ_ONCE(p->__state);
3081 * We should never call set_task_cpu() on a blocked task,
3082 * ttwu() will sort out the placement.
3084 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3087 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3088 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3089 * time relying on p->on_rq.
3091 WARN_ON_ONCE(state == TASK_RUNNING &&
3092 p->sched_class == &fair_sched_class &&
3093 (p->on_rq && !task_on_rq_migrating(p)));
3095 #ifdef CONFIG_LOCKDEP
3097 * The caller should hold either p->pi_lock or rq->lock, when changing
3098 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3100 * sched_move_task() holds both and thus holding either pins the cgroup,
3103 * Furthermore, all task_rq users should acquire both locks, see
3106 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3107 lockdep_is_held(__rq_lockp(task_rq(p)))));
3110 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3112 WARN_ON_ONCE(!cpu_online(new_cpu));
3114 WARN_ON_ONCE(is_migration_disabled(p));
3117 trace_sched_migrate_task(p, new_cpu);
3119 if (task_cpu(p) != new_cpu) {
3120 if (p->sched_class->migrate_task_rq)
3121 p->sched_class->migrate_task_rq(p, new_cpu);
3122 p->se.nr_migrations++;
3124 perf_event_task_migrate(p);
3127 __set_task_cpu(p, new_cpu);
3130 #ifdef CONFIG_NUMA_BALANCING
3131 static void __migrate_swap_task(struct task_struct *p, int cpu)
3133 if (task_on_rq_queued(p)) {
3134 struct rq *src_rq, *dst_rq;
3135 struct rq_flags srf, drf;
3137 src_rq = task_rq(p);
3138 dst_rq = cpu_rq(cpu);
3140 rq_pin_lock(src_rq, &srf);
3141 rq_pin_lock(dst_rq, &drf);
3143 deactivate_task(src_rq, p, 0);
3144 set_task_cpu(p, cpu);
3145 activate_task(dst_rq, p, 0);
3146 check_preempt_curr(dst_rq, p, 0);
3148 rq_unpin_lock(dst_rq, &drf);
3149 rq_unpin_lock(src_rq, &srf);
3153 * Task isn't running anymore; make it appear like we migrated
3154 * it before it went to sleep. This means on wakeup we make the
3155 * previous CPU our target instead of where it really is.
3161 struct migration_swap_arg {
3162 struct task_struct *src_task, *dst_task;
3163 int src_cpu, dst_cpu;
3166 static int migrate_swap_stop(void *data)
3168 struct migration_swap_arg *arg = data;
3169 struct rq *src_rq, *dst_rq;
3172 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3175 src_rq = cpu_rq(arg->src_cpu);
3176 dst_rq = cpu_rq(arg->dst_cpu);
3178 double_raw_lock(&arg->src_task->pi_lock,
3179 &arg->dst_task->pi_lock);
3180 double_rq_lock(src_rq, dst_rq);
3182 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3185 if (task_cpu(arg->src_task) != arg->src_cpu)
3188 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3191 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3194 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3195 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3200 double_rq_unlock(src_rq, dst_rq);
3201 raw_spin_unlock(&arg->dst_task->pi_lock);
3202 raw_spin_unlock(&arg->src_task->pi_lock);
3208 * Cross migrate two tasks
3210 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3211 int target_cpu, int curr_cpu)
3213 struct migration_swap_arg arg;
3216 arg = (struct migration_swap_arg){
3218 .src_cpu = curr_cpu,
3220 .dst_cpu = target_cpu,
3223 if (arg.src_cpu == arg.dst_cpu)
3227 * These three tests are all lockless; this is OK since all of them
3228 * will be re-checked with proper locks held further down the line.
3230 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3233 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3236 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3239 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3240 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3245 #endif /* CONFIG_NUMA_BALANCING */
3248 * wait_task_inactive - wait for a thread to unschedule.
3250 * Wait for the thread to block in any of the states set in @match_state.
3251 * If it changes, i.e. @p might have woken up, then return zero. When we
3252 * succeed in waiting for @p to be off its CPU, we return a positive number
3253 * (its total switch count). If a second call a short while later returns the
3254 * same number, the caller can be sure that @p has remained unscheduled the
3257 * The caller must ensure that the task *will* unschedule sometime soon,
3258 * else this function might spin for a *long* time. This function can't
3259 * be called with interrupts off, or it may introduce deadlock with
3260 * smp_call_function() if an IPI is sent by the same process we are
3261 * waiting to become inactive.
3263 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3265 int running, queued;
3272 * We do the initial early heuristics without holding
3273 * any task-queue locks at all. We'll only try to get
3274 * the runqueue lock when things look like they will
3280 * If the task is actively running on another CPU
3281 * still, just relax and busy-wait without holding
3284 * NOTE! Since we don't hold any locks, it's not
3285 * even sure that "rq" stays as the right runqueue!
3286 * But we don't care, since "task_on_cpu()" will
3287 * return false if the runqueue has changed and p
3288 * is actually now running somewhere else!
3290 while (task_on_cpu(rq, p)) {
3291 if (!(READ_ONCE(p->__state) & match_state))
3297 * Ok, time to look more closely! We need the rq
3298 * lock now, to be *sure*. If we're wrong, we'll
3299 * just go back and repeat.
3301 rq = task_rq_lock(p, &rf);
3302 trace_sched_wait_task(p);
3303 running = task_on_cpu(rq, p);
3304 queued = task_on_rq_queued(p);
3306 if (READ_ONCE(p->__state) & match_state)
3307 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3308 task_rq_unlock(rq, p, &rf);
3311 * If it changed from the expected state, bail out now.
3313 if (unlikely(!ncsw))
3317 * Was it really running after all now that we
3318 * checked with the proper locks actually held?
3320 * Oops. Go back and try again..
3322 if (unlikely(running)) {
3328 * It's not enough that it's not actively running,
3329 * it must be off the runqueue _entirely_, and not
3332 * So if it was still runnable (but just not actively
3333 * running right now), it's preempted, and we should
3334 * yield - it could be a while.
3336 if (unlikely(queued)) {
3337 ktime_t to = NSEC_PER_SEC / HZ;
3339 set_current_state(TASK_UNINTERRUPTIBLE);
3340 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3345 * Ahh, all good. It wasn't running, and it wasn't
3346 * runnable, which means that it will never become
3347 * running in the future either. We're all done!
3356 * kick_process - kick a running thread to enter/exit the kernel
3357 * @p: the to-be-kicked thread
3359 * Cause a process which is running on another CPU to enter
3360 * kernel-mode, without any delay. (to get signals handled.)
3362 * NOTE: this function doesn't have to take the runqueue lock,
3363 * because all it wants to ensure is that the remote task enters
3364 * the kernel. If the IPI races and the task has been migrated
3365 * to another CPU then no harm is done and the purpose has been
3368 void kick_process(struct task_struct *p)
3374 if ((cpu != smp_processor_id()) && task_curr(p))
3375 smp_send_reschedule(cpu);
3378 EXPORT_SYMBOL_GPL(kick_process);
3381 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3383 * A few notes on cpu_active vs cpu_online:
3385 * - cpu_active must be a subset of cpu_online
3387 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3388 * see __set_cpus_allowed_ptr(). At this point the newly online
3389 * CPU isn't yet part of the sched domains, and balancing will not
3392 * - on CPU-down we clear cpu_active() to mask the sched domains and
3393 * avoid the load balancer to place new tasks on the to be removed
3394 * CPU. Existing tasks will remain running there and will be taken
3397 * This means that fallback selection must not select !active CPUs.
3398 * And can assume that any active CPU must be online. Conversely
3399 * select_task_rq() below may allow selection of !active CPUs in order
3400 * to satisfy the above rules.
3402 static int select_fallback_rq(int cpu, struct task_struct *p)
3404 int nid = cpu_to_node(cpu);
3405 const struct cpumask *nodemask = NULL;
3406 enum { cpuset, possible, fail } state = cpuset;
3410 * If the node that the CPU is on has been offlined, cpu_to_node()
3411 * will return -1. There is no CPU on the node, and we should
3412 * select the CPU on the other node.
3415 nodemask = cpumask_of_node(nid);
3417 /* Look for allowed, online CPU in same node. */
3418 for_each_cpu(dest_cpu, nodemask) {
3419 if (is_cpu_allowed(p, dest_cpu))
3425 /* Any allowed, online CPU? */
3426 for_each_cpu(dest_cpu, p->cpus_ptr) {
3427 if (!is_cpu_allowed(p, dest_cpu))
3433 /* No more Mr. Nice Guy. */
3436 if (cpuset_cpus_allowed_fallback(p)) {
3443 * XXX When called from select_task_rq() we only
3444 * hold p->pi_lock and again violate locking order.
3446 * More yuck to audit.
3448 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3458 if (state != cpuset) {
3460 * Don't tell them about moving exiting tasks or
3461 * kernel threads (both mm NULL), since they never
3464 if (p->mm && printk_ratelimit()) {
3465 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3466 task_pid_nr(p), p->comm, cpu);
3474 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3477 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3479 lockdep_assert_held(&p->pi_lock);
3481 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3482 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3484 cpu = cpumask_any(p->cpus_ptr);
3487 * In order not to call set_task_cpu() on a blocking task we need
3488 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3491 * Since this is common to all placement strategies, this lives here.
3493 * [ this allows ->select_task() to simply return task_cpu(p) and
3494 * not worry about this generic constraint ]
3496 if (unlikely(!is_cpu_allowed(p, cpu)))
3497 cpu = select_fallback_rq(task_cpu(p), p);
3502 void sched_set_stop_task(int cpu, struct task_struct *stop)
3504 static struct lock_class_key stop_pi_lock;
3505 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3506 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3510 * Make it appear like a SCHED_FIFO task, its something
3511 * userspace knows about and won't get confused about.
3513 * Also, it will make PI more or less work without too
3514 * much confusion -- but then, stop work should not
3515 * rely on PI working anyway.
3517 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3519 stop->sched_class = &stop_sched_class;
3522 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3523 * adjust the effective priority of a task. As a result,
3524 * rt_mutex_setprio() can trigger (RT) balancing operations,
3525 * which can then trigger wakeups of the stop thread to push
3526 * around the current task.
3528 * The stop task itself will never be part of the PI-chain, it
3529 * never blocks, therefore that ->pi_lock recursion is safe.
3530 * Tell lockdep about this by placing the stop->pi_lock in its
3533 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3536 cpu_rq(cpu)->stop = stop;
3540 * Reset it back to a normal scheduling class so that
3541 * it can die in pieces.
3543 old_stop->sched_class = &rt_sched_class;
3547 #else /* CONFIG_SMP */
3549 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3550 const struct cpumask *new_mask,
3553 return set_cpus_allowed_ptr(p, new_mask);
3556 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3558 static inline bool rq_has_pinned_tasks(struct rq *rq)
3563 #endif /* !CONFIG_SMP */
3566 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3570 if (!schedstat_enabled())
3576 if (cpu == rq->cpu) {
3577 __schedstat_inc(rq->ttwu_local);
3578 __schedstat_inc(p->stats.nr_wakeups_local);
3580 struct sched_domain *sd;
3582 __schedstat_inc(p->stats.nr_wakeups_remote);
3584 for_each_domain(rq->cpu, sd) {
3585 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3586 __schedstat_inc(sd->ttwu_wake_remote);
3593 if (wake_flags & WF_MIGRATED)
3594 __schedstat_inc(p->stats.nr_wakeups_migrate);
3595 #endif /* CONFIG_SMP */
3597 __schedstat_inc(rq->ttwu_count);
3598 __schedstat_inc(p->stats.nr_wakeups);
3600 if (wake_flags & WF_SYNC)
3601 __schedstat_inc(p->stats.nr_wakeups_sync);
3605 * Mark the task runnable and perform wakeup-preemption.
3607 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3608 struct rq_flags *rf)
3610 check_preempt_curr(rq, p, wake_flags);
3611 WRITE_ONCE(p->__state, TASK_RUNNING);
3612 trace_sched_wakeup(p);
3615 if (p->sched_class->task_woken) {
3617 * Our task @p is fully woken up and running; so it's safe to
3618 * drop the rq->lock, hereafter rq is only used for statistics.
3620 rq_unpin_lock(rq, rf);
3621 p->sched_class->task_woken(rq, p);
3622 rq_repin_lock(rq, rf);
3625 if (rq->idle_stamp) {
3626 u64 delta = rq_clock(rq) - rq->idle_stamp;
3627 u64 max = 2*rq->max_idle_balance_cost;
3629 update_avg(&rq->avg_idle, delta);
3631 if (rq->avg_idle > max)
3634 rq->wake_stamp = jiffies;
3635 rq->wake_avg_idle = rq->avg_idle / 2;
3643 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3644 struct rq_flags *rf)
3646 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3648 lockdep_assert_rq_held(rq);
3650 if (p->sched_contributes_to_load)
3651 rq->nr_uninterruptible--;
3654 if (wake_flags & WF_MIGRATED)
3655 en_flags |= ENQUEUE_MIGRATED;
3659 delayacct_blkio_end(p);
3660 atomic_dec(&task_rq(p)->nr_iowait);
3663 activate_task(rq, p, en_flags);
3664 ttwu_do_wakeup(rq, p, wake_flags, rf);
3668 * Consider @p being inside a wait loop:
3671 * set_current_state(TASK_UNINTERRUPTIBLE);
3678 * __set_current_state(TASK_RUNNING);
3680 * between set_current_state() and schedule(). In this case @p is still
3681 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3684 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3685 * then schedule() must still happen and p->state can be changed to
3686 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3687 * need to do a full wakeup with enqueue.
3689 * Returns: %true when the wakeup is done,
3692 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3698 rq = __task_rq_lock(p, &rf);
3699 if (task_on_rq_queued(p)) {
3700 /* check_preempt_curr() may use rq clock */
3701 update_rq_clock(rq);
3702 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3705 __task_rq_unlock(rq, &rf);
3711 void sched_ttwu_pending(void *arg)
3713 struct llist_node *llist = arg;
3714 struct rq *rq = this_rq();
3715 struct task_struct *p, *t;
3722 * rq::ttwu_pending racy indication of out-standing wakeups.
3723 * Races such that false-negatives are possible, since they
3724 * are shorter lived that false-positives would be.
3726 WRITE_ONCE(rq->ttwu_pending, 0);
3728 rq_lock_irqsave(rq, &rf);
3729 update_rq_clock(rq);
3731 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3732 if (WARN_ON_ONCE(p->on_cpu))
3733 smp_cond_load_acquire(&p->on_cpu, !VAL);
3735 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3736 set_task_cpu(p, cpu_of(rq));
3738 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3741 rq_unlock_irqrestore(rq, &rf);
3744 void send_call_function_single_ipi(int cpu)
3746 struct rq *rq = cpu_rq(cpu);
3748 if (!set_nr_if_polling(rq->idle))
3749 arch_send_call_function_single_ipi(cpu);
3751 trace_sched_wake_idle_without_ipi(cpu);
3755 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3756 * necessary. The wakee CPU on receipt of the IPI will queue the task
3757 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3758 * of the wakeup instead of the waker.
3760 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3762 struct rq *rq = cpu_rq(cpu);
3764 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3766 WRITE_ONCE(rq->ttwu_pending, 1);
3767 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3770 void wake_up_if_idle(int cpu)
3772 struct rq *rq = cpu_rq(cpu);
3777 if (!is_idle_task(rcu_dereference(rq->curr)))
3780 rq_lock_irqsave(rq, &rf);
3781 if (is_idle_task(rq->curr))
3783 /* Else CPU is not idle, do nothing here: */
3784 rq_unlock_irqrestore(rq, &rf);
3790 bool cpus_share_cache(int this_cpu, int that_cpu)
3792 if (this_cpu == that_cpu)
3795 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3798 static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3801 * Do not complicate things with the async wake_list while the CPU is
3804 if (!cpu_active(cpu))
3807 /* Ensure the task will still be allowed to run on the CPU. */
3808 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3812 * If the CPU does not share cache, then queue the task on the
3813 * remote rqs wakelist to avoid accessing remote data.
3815 if (!cpus_share_cache(smp_processor_id(), cpu))
3818 if (cpu == smp_processor_id())
3822 * If the wakee cpu is idle, or the task is descheduling and the
3823 * only running task on the CPU, then use the wakelist to offload
3824 * the task activation to the idle (or soon-to-be-idle) CPU as
3825 * the current CPU is likely busy. nr_running is checked to
3826 * avoid unnecessary task stacking.
3828 * Note that we can only get here with (wakee) p->on_rq=0,
3829 * p->on_cpu can be whatever, we've done the dequeue, so
3830 * the wakee has been accounted out of ->nr_running.
3832 if (!cpu_rq(cpu)->nr_running)
3838 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3840 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
3841 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3842 __ttwu_queue_wakelist(p, cpu, wake_flags);
3849 #else /* !CONFIG_SMP */
3851 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3856 #endif /* CONFIG_SMP */
3858 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3860 struct rq *rq = cpu_rq(cpu);
3863 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3867 update_rq_clock(rq);
3868 ttwu_do_activate(rq, p, wake_flags, &rf);
3873 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3875 * The caller holds p::pi_lock if p != current or has preemption
3876 * disabled when p == current.
3878 * The rules of PREEMPT_RT saved_state:
3880 * The related locking code always holds p::pi_lock when updating
3881 * p::saved_state, which means the code is fully serialized in both cases.
3883 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3884 * bits set. This allows to distinguish all wakeup scenarios.
3886 static __always_inline
3887 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3889 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3890 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3891 state != TASK_RTLOCK_WAIT);
3894 if (READ_ONCE(p->__state) & state) {
3899 #ifdef CONFIG_PREEMPT_RT
3901 * Saved state preserves the task state across blocking on
3902 * an RT lock. If the state matches, set p::saved_state to
3903 * TASK_RUNNING, but do not wake the task because it waits
3904 * for a lock wakeup. Also indicate success because from
3905 * the regular waker's point of view this has succeeded.
3907 * After acquiring the lock the task will restore p::__state
3908 * from p::saved_state which ensures that the regular
3909 * wakeup is not lost. The restore will also set
3910 * p::saved_state to TASK_RUNNING so any further tests will
3911 * not result in false positives vs. @success
3913 if (p->saved_state & state) {
3914 p->saved_state = TASK_RUNNING;
3922 * Notes on Program-Order guarantees on SMP systems.
3926 * The basic program-order guarantee on SMP systems is that when a task [t]
3927 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3928 * execution on its new CPU [c1].
3930 * For migration (of runnable tasks) this is provided by the following means:
3932 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3933 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3934 * rq(c1)->lock (if not at the same time, then in that order).
3935 * C) LOCK of the rq(c1)->lock scheduling in task
3937 * Release/acquire chaining guarantees that B happens after A and C after B.
3938 * Note: the CPU doing B need not be c0 or c1
3947 * UNLOCK rq(0)->lock
3949 * LOCK rq(0)->lock // orders against CPU0
3951 * UNLOCK rq(0)->lock
3955 * UNLOCK rq(1)->lock
3957 * LOCK rq(1)->lock // orders against CPU2
3960 * UNLOCK rq(1)->lock
3963 * BLOCKING -- aka. SLEEP + WAKEUP
3965 * For blocking we (obviously) need to provide the same guarantee as for
3966 * migration. However the means are completely different as there is no lock
3967 * chain to provide order. Instead we do:
3969 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3970 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3974 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3976 * LOCK rq(0)->lock LOCK X->pi_lock
3979 * smp_store_release(X->on_cpu, 0);
3981 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3987 * X->state = RUNNING
3988 * UNLOCK rq(2)->lock
3990 * LOCK rq(2)->lock // orders against CPU1
3993 * UNLOCK rq(2)->lock
3996 * UNLOCK rq(0)->lock
3999 * However, for wakeups there is a second guarantee we must provide, namely we
4000 * must ensure that CONDITION=1 done by the caller can not be reordered with
4001 * accesses to the task state; see try_to_wake_up() and set_current_state().
4005 * try_to_wake_up - wake up a thread
4006 * @p: the thread to be awakened
4007 * @state: the mask of task states that can be woken
4008 * @wake_flags: wake modifier flags (WF_*)
4010 * Conceptually does:
4012 * If (@state & @p->state) @p->state = TASK_RUNNING.
4014 * If the task was not queued/runnable, also place it back on a runqueue.
4016 * This function is atomic against schedule() which would dequeue the task.
4018 * It issues a full memory barrier before accessing @p->state, see the comment
4019 * with set_current_state().
4021 * Uses p->pi_lock to serialize against concurrent wake-ups.
4023 * Relies on p->pi_lock stabilizing:
4026 * - p->sched_task_group
4027 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4029 * Tries really hard to only take one task_rq(p)->lock for performance.
4030 * Takes rq->lock in:
4031 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4032 * - ttwu_queue() -- new rq, for enqueue of the task;
4033 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4035 * As a consequence we race really badly with just about everything. See the
4036 * many memory barriers and their comments for details.
4038 * Return: %true if @p->state changes (an actual wakeup was done),
4042 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4044 unsigned long flags;
4045 int cpu, success = 0;
4050 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4051 * == smp_processor_id()'. Together this means we can special
4052 * case the whole 'p->on_rq && ttwu_runnable()' case below
4053 * without taking any locks.
4056 * - we rely on Program-Order guarantees for all the ordering,
4057 * - we're serialized against set_special_state() by virtue of
4058 * it disabling IRQs (this allows not taking ->pi_lock).
4060 if (!ttwu_state_match(p, state, &success))
4063 trace_sched_waking(p);
4064 WRITE_ONCE(p->__state, TASK_RUNNING);
4065 trace_sched_wakeup(p);
4070 * If we are going to wake up a thread waiting for CONDITION we
4071 * need to ensure that CONDITION=1 done by the caller can not be
4072 * reordered with p->state check below. This pairs with smp_store_mb()
4073 * in set_current_state() that the waiting thread does.
4075 raw_spin_lock_irqsave(&p->pi_lock, flags);
4076 smp_mb__after_spinlock();
4077 if (!ttwu_state_match(p, state, &success))
4080 trace_sched_waking(p);
4083 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4084 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4085 * in smp_cond_load_acquire() below.
4087 * sched_ttwu_pending() try_to_wake_up()
4088 * STORE p->on_rq = 1 LOAD p->state
4091 * __schedule() (switch to task 'p')
4092 * LOCK rq->lock smp_rmb();
4093 * smp_mb__after_spinlock();
4097 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4099 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4100 * __schedule(). See the comment for smp_mb__after_spinlock().
4102 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4105 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4110 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4111 * possible to, falsely, observe p->on_cpu == 0.
4113 * One must be running (->on_cpu == 1) in order to remove oneself
4114 * from the runqueue.
4116 * __schedule() (switch to task 'p') try_to_wake_up()
4117 * STORE p->on_cpu = 1 LOAD p->on_rq
4120 * __schedule() (put 'p' to sleep)
4121 * LOCK rq->lock smp_rmb();
4122 * smp_mb__after_spinlock();
4123 * STORE p->on_rq = 0 LOAD p->on_cpu
4125 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4126 * __schedule(). See the comment for smp_mb__after_spinlock().
4128 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4129 * schedule()'s deactivate_task() has 'happened' and p will no longer
4130 * care about it's own p->state. See the comment in __schedule().
4132 smp_acquire__after_ctrl_dep();
4135 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4136 * == 0), which means we need to do an enqueue, change p->state to
4137 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4138 * enqueue, such as ttwu_queue_wakelist().
4140 WRITE_ONCE(p->__state, TASK_WAKING);
4143 * If the owning (remote) CPU is still in the middle of schedule() with
4144 * this task as prev, considering queueing p on the remote CPUs wake_list
4145 * which potentially sends an IPI instead of spinning on p->on_cpu to
4146 * let the waker make forward progress. This is safe because IRQs are
4147 * disabled and the IPI will deliver after on_cpu is cleared.
4149 * Ensure we load task_cpu(p) after p->on_cpu:
4151 * set_task_cpu(p, cpu);
4152 * STORE p->cpu = @cpu
4153 * __schedule() (switch to task 'p')
4155 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4156 * STORE p->on_cpu = 1 LOAD p->cpu
4158 * to ensure we observe the correct CPU on which the task is currently
4161 if (smp_load_acquire(&p->on_cpu) &&
4162 ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4166 * If the owning (remote) CPU is still in the middle of schedule() with
4167 * this task as prev, wait until it's done referencing the task.
4169 * Pairs with the smp_store_release() in finish_task().
4171 * This ensures that tasks getting woken will be fully ordered against
4172 * their previous state and preserve Program Order.
4174 smp_cond_load_acquire(&p->on_cpu, !VAL);
4176 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4177 if (task_cpu(p) != cpu) {
4179 delayacct_blkio_end(p);
4180 atomic_dec(&task_rq(p)->nr_iowait);
4183 wake_flags |= WF_MIGRATED;
4184 psi_ttwu_dequeue(p);
4185 set_task_cpu(p, cpu);
4189 #endif /* CONFIG_SMP */
4191 ttwu_queue(p, cpu, wake_flags);
4193 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4196 ttwu_stat(p, task_cpu(p), wake_flags);
4203 * task_call_func - Invoke a function on task in fixed state
4204 * @p: Process for which the function is to be invoked, can be @current.
4205 * @func: Function to invoke.
4206 * @arg: Argument to function.
4208 * Fix the task in it's current state by avoiding wakeups and or rq operations
4209 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4210 * to work out what the state is, if required. Given that @func can be invoked
4211 * with a runqueue lock held, it had better be quite lightweight.
4214 * Whatever @func returns
4216 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4218 struct rq *rq = NULL;
4223 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4225 state = READ_ONCE(p->__state);
4228 * Ensure we load p->on_rq after p->__state, otherwise it would be
4229 * possible to, falsely, observe p->on_rq == 0.
4231 * See try_to_wake_up() for a longer comment.
4236 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4237 * the task is blocked. Make sure to check @state since ttwu() can drop
4238 * locks at the end, see ttwu_queue_wakelist().
4240 if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4241 rq = __task_rq_lock(p, &rf);
4244 * At this point the task is pinned; either:
4245 * - blocked and we're holding off wakeups (pi->lock)
4246 * - woken, and we're holding off enqueue (rq->lock)
4247 * - queued, and we're holding off schedule (rq->lock)
4248 * - running, and we're holding off de-schedule (rq->lock)
4250 * The called function (@func) can use: task_curr(), p->on_rq and
4251 * p->__state to differentiate between these states.
4258 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4263 * cpu_curr_snapshot - Return a snapshot of the currently running task
4264 * @cpu: The CPU on which to snapshot the task.
4266 * Returns the task_struct pointer of the task "currently" running on
4267 * the specified CPU. If the same task is running on that CPU throughout,
4268 * the return value will be a pointer to that task's task_struct structure.
4269 * If the CPU did any context switches even vaguely concurrently with the
4270 * execution of this function, the return value will be a pointer to the
4271 * task_struct structure of a randomly chosen task that was running on
4272 * that CPU somewhere around the time that this function was executing.
4274 * If the specified CPU was offline, the return value is whatever it
4275 * is, perhaps a pointer to the task_struct structure of that CPU's idle
4276 * task, but there is no guarantee. Callers wishing a useful return
4277 * value must take some action to ensure that the specified CPU remains
4278 * online throughout.
4280 * This function executes full memory barriers before and after fetching
4281 * the pointer, which permits the caller to confine this function's fetch
4282 * with respect to the caller's accesses to other shared variables.
4284 struct task_struct *cpu_curr_snapshot(int cpu)
4286 struct task_struct *t;
4288 smp_mb(); /* Pairing determined by caller's synchronization design. */
4289 t = rcu_dereference(cpu_curr(cpu));
4290 smp_mb(); /* Pairing determined by caller's synchronization design. */
4295 * wake_up_process - Wake up a specific process
4296 * @p: The process to be woken up.
4298 * Attempt to wake up the nominated process and move it to the set of runnable
4301 * Return: 1 if the process was woken up, 0 if it was already running.
4303 * This function executes a full memory barrier before accessing the task state.
4305 int wake_up_process(struct task_struct *p)
4307 return try_to_wake_up(p, TASK_NORMAL, 0);
4309 EXPORT_SYMBOL(wake_up_process);
4311 int wake_up_state(struct task_struct *p, unsigned int state)
4313 return try_to_wake_up(p, state, 0);
4317 * Perform scheduler related setup for a newly forked process p.
4318 * p is forked by current.
4320 * __sched_fork() is basic setup used by init_idle() too:
4322 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4327 p->se.exec_start = 0;
4328 p->se.sum_exec_runtime = 0;
4329 p->se.prev_sum_exec_runtime = 0;
4330 p->se.nr_migrations = 0;
4332 INIT_LIST_HEAD(&p->se.group_node);
4334 #ifdef CONFIG_FAIR_GROUP_SCHED
4335 p->se.cfs_rq = NULL;
4338 #ifdef CONFIG_SCHEDSTATS
4339 /* Even if schedstat is disabled, there should not be garbage */
4340 memset(&p->stats, 0, sizeof(p->stats));
4343 RB_CLEAR_NODE(&p->dl.rb_node);
4344 init_dl_task_timer(&p->dl);
4345 init_dl_inactive_task_timer(&p->dl);
4346 __dl_clear_params(p);
4348 INIT_LIST_HEAD(&p->rt.run_list);
4350 p->rt.time_slice = sched_rr_timeslice;
4354 #ifdef CONFIG_PREEMPT_NOTIFIERS
4355 INIT_HLIST_HEAD(&p->preempt_notifiers);
4358 #ifdef CONFIG_COMPACTION
4359 p->capture_control = NULL;
4361 init_numa_balancing(clone_flags, p);
4363 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4364 p->migration_pending = NULL;
4368 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4370 #ifdef CONFIG_NUMA_BALANCING
4372 int sysctl_numa_balancing_mode;
4374 static void __set_numabalancing_state(bool enabled)
4377 static_branch_enable(&sched_numa_balancing);
4379 static_branch_disable(&sched_numa_balancing);
4382 void set_numabalancing_state(bool enabled)
4385 sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4387 sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4388 __set_numabalancing_state(enabled);
4391 #ifdef CONFIG_PROC_SYSCTL
4392 int sysctl_numa_balancing(struct ctl_table *table, int write,
4393 void *buffer, size_t *lenp, loff_t *ppos)
4397 int state = sysctl_numa_balancing_mode;
4399 if (write && !capable(CAP_SYS_ADMIN))
4404 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4408 sysctl_numa_balancing_mode = state;
4409 __set_numabalancing_state(state);
4416 #ifdef CONFIG_SCHEDSTATS
4418 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4420 static void set_schedstats(bool enabled)
4423 static_branch_enable(&sched_schedstats);
4425 static_branch_disable(&sched_schedstats);
4428 void force_schedstat_enabled(void)
4430 if (!schedstat_enabled()) {
4431 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4432 static_branch_enable(&sched_schedstats);
4436 static int __init setup_schedstats(char *str)
4442 if (!strcmp(str, "enable")) {
4443 set_schedstats(true);
4445 } else if (!strcmp(str, "disable")) {
4446 set_schedstats(false);
4451 pr_warn("Unable to parse schedstats=\n");
4455 __setup("schedstats=", setup_schedstats);
4457 #ifdef CONFIG_PROC_SYSCTL
4458 static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4459 size_t *lenp, loff_t *ppos)
4463 int state = static_branch_likely(&sched_schedstats);
4465 if (write && !capable(CAP_SYS_ADMIN))
4470 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4474 set_schedstats(state);
4477 #endif /* CONFIG_PROC_SYSCTL */
4478 #endif /* CONFIG_SCHEDSTATS */
4480 #ifdef CONFIG_SYSCTL
4481 static struct ctl_table sched_core_sysctls[] = {
4482 #ifdef CONFIG_SCHEDSTATS
4484 .procname = "sched_schedstats",
4486 .maxlen = sizeof(unsigned int),
4488 .proc_handler = sysctl_schedstats,
4489 .extra1 = SYSCTL_ZERO,
4490 .extra2 = SYSCTL_ONE,
4492 #endif /* CONFIG_SCHEDSTATS */
4493 #ifdef CONFIG_UCLAMP_TASK
4495 .procname = "sched_util_clamp_min",
4496 .data = &sysctl_sched_uclamp_util_min,
4497 .maxlen = sizeof(unsigned int),
4499 .proc_handler = sysctl_sched_uclamp_handler,
4502 .procname = "sched_util_clamp_max",
4503 .data = &sysctl_sched_uclamp_util_max,
4504 .maxlen = sizeof(unsigned int),
4506 .proc_handler = sysctl_sched_uclamp_handler,
4509 .procname = "sched_util_clamp_min_rt_default",
4510 .data = &sysctl_sched_uclamp_util_min_rt_default,
4511 .maxlen = sizeof(unsigned int),
4513 .proc_handler = sysctl_sched_uclamp_handler,
4515 #endif /* CONFIG_UCLAMP_TASK */
4518 static int __init sched_core_sysctl_init(void)
4520 register_sysctl_init("kernel", sched_core_sysctls);
4523 late_initcall(sched_core_sysctl_init);
4524 #endif /* CONFIG_SYSCTL */
4527 * fork()/clone()-time setup:
4529 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4531 __sched_fork(clone_flags, p);
4533 * We mark the process as NEW here. This guarantees that
4534 * nobody will actually run it, and a signal or other external
4535 * event cannot wake it up and insert it on the runqueue either.
4537 p->__state = TASK_NEW;
4540 * Make sure we do not leak PI boosting priority to the child.
4542 p->prio = current->normal_prio;
4547 * Revert to default priority/policy on fork if requested.
4549 if (unlikely(p->sched_reset_on_fork)) {
4550 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4551 p->policy = SCHED_NORMAL;
4552 p->static_prio = NICE_TO_PRIO(0);
4554 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4555 p->static_prio = NICE_TO_PRIO(0);
4557 p->prio = p->normal_prio = p->static_prio;
4558 set_load_weight(p, false);
4561 * We don't need the reset flag anymore after the fork. It has
4562 * fulfilled its duty:
4564 p->sched_reset_on_fork = 0;
4567 if (dl_prio(p->prio))
4569 else if (rt_prio(p->prio))
4570 p->sched_class = &rt_sched_class;
4572 p->sched_class = &fair_sched_class;
4574 init_entity_runnable_average(&p->se);
4577 #ifdef CONFIG_SCHED_INFO
4578 if (likely(sched_info_on()))
4579 memset(&p->sched_info, 0, sizeof(p->sched_info));
4581 #if defined(CONFIG_SMP)
4584 init_task_preempt_count(p);
4586 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4587 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4592 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4594 unsigned long flags;
4597 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4598 * required yet, but lockdep gets upset if rules are violated.
4600 raw_spin_lock_irqsave(&p->pi_lock, flags);
4601 #ifdef CONFIG_CGROUP_SCHED
4603 struct task_group *tg;
4604 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4605 struct task_group, css);
4606 tg = autogroup_task_group(p, tg);
4607 p->sched_task_group = tg;
4612 * We're setting the CPU for the first time, we don't migrate,
4613 * so use __set_task_cpu().
4615 __set_task_cpu(p, smp_processor_id());
4616 if (p->sched_class->task_fork)
4617 p->sched_class->task_fork(p);
4618 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4621 void sched_post_fork(struct task_struct *p)
4623 uclamp_post_fork(p);
4626 unsigned long to_ratio(u64 period, u64 runtime)
4628 if (runtime == RUNTIME_INF)
4632 * Doing this here saves a lot of checks in all
4633 * the calling paths, and returning zero seems
4634 * safe for them anyway.
4639 return div64_u64(runtime << BW_SHIFT, period);
4643 * wake_up_new_task - wake up a newly created task for the first time.
4645 * This function will do some initial scheduler statistics housekeeping
4646 * that must be done for every newly created context, then puts the task
4647 * on the runqueue and wakes it.
4649 void wake_up_new_task(struct task_struct *p)
4654 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4655 WRITE_ONCE(p->__state, TASK_RUNNING);
4658 * Fork balancing, do it here and not earlier because:
4659 * - cpus_ptr can change in the fork path
4660 * - any previously selected CPU might disappear through hotplug
4662 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4663 * as we're not fully set-up yet.
4665 p->recent_used_cpu = task_cpu(p);
4667 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4669 rq = __task_rq_lock(p, &rf);
4670 update_rq_clock(rq);
4671 post_init_entity_util_avg(p);
4673 activate_task(rq, p, ENQUEUE_NOCLOCK);
4674 trace_sched_wakeup_new(p);
4675 check_preempt_curr(rq, p, WF_FORK);
4677 if (p->sched_class->task_woken) {
4679 * Nothing relies on rq->lock after this, so it's fine to
4682 rq_unpin_lock(rq, &rf);
4683 p->sched_class->task_woken(rq, p);
4684 rq_repin_lock(rq, &rf);
4687 task_rq_unlock(rq, p, &rf);
4690 #ifdef CONFIG_PREEMPT_NOTIFIERS
4692 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4694 void preempt_notifier_inc(void)
4696 static_branch_inc(&preempt_notifier_key);
4698 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4700 void preempt_notifier_dec(void)
4702 static_branch_dec(&preempt_notifier_key);
4704 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4707 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4708 * @notifier: notifier struct to register
4710 void preempt_notifier_register(struct preempt_notifier *notifier)
4712 if (!static_branch_unlikely(&preempt_notifier_key))
4713 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4715 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4717 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4720 * preempt_notifier_unregister - no longer interested in preemption notifications
4721 * @notifier: notifier struct to unregister
4723 * This is *not* safe to call from within a preemption notifier.
4725 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4727 hlist_del(¬ifier->link);
4729 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4731 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4733 struct preempt_notifier *notifier;
4735 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4736 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4739 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4741 if (static_branch_unlikely(&preempt_notifier_key))
4742 __fire_sched_in_preempt_notifiers(curr);
4746 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4747 struct task_struct *next)
4749 struct preempt_notifier *notifier;
4751 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4752 notifier->ops->sched_out(notifier, next);
4755 static __always_inline void
4756 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4757 struct task_struct *next)
4759 if (static_branch_unlikely(&preempt_notifier_key))
4760 __fire_sched_out_preempt_notifiers(curr, next);
4763 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4765 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4770 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4771 struct task_struct *next)
4775 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4777 static inline void prepare_task(struct task_struct *next)
4781 * Claim the task as running, we do this before switching to it
4782 * such that any running task will have this set.
4784 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4785 * its ordering comment.
4787 WRITE_ONCE(next->on_cpu, 1);
4791 static inline void finish_task(struct task_struct *prev)
4795 * This must be the very last reference to @prev from this CPU. After
4796 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4797 * must ensure this doesn't happen until the switch is completely
4800 * In particular, the load of prev->state in finish_task_switch() must
4801 * happen before this.
4803 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4805 smp_store_release(&prev->on_cpu, 0);
4811 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4813 void (*func)(struct rq *rq);
4814 struct callback_head *next;
4816 lockdep_assert_rq_held(rq);
4819 func = (void (*)(struct rq *))head->func;
4828 static void balance_push(struct rq *rq);
4831 * balance_push_callback is a right abuse of the callback interface and plays
4832 * by significantly different rules.
4834 * Where the normal balance_callback's purpose is to be ran in the same context
4835 * that queued it (only later, when it's safe to drop rq->lock again),
4836 * balance_push_callback is specifically targeted at __schedule().
4838 * This abuse is tolerated because it places all the unlikely/odd cases behind
4839 * a single test, namely: rq->balance_callback == NULL.
4841 struct callback_head balance_push_callback = {
4843 .func = (void (*)(struct callback_head *))balance_push,
4846 static inline struct callback_head *
4847 __splice_balance_callbacks(struct rq *rq, bool split)
4849 struct callback_head *head = rq->balance_callback;
4854 lockdep_assert_rq_held(rq);
4856 * Must not take balance_push_callback off the list when
4857 * splice_balance_callbacks() and balance_callbacks() are not
4858 * in the same rq->lock section.
4860 * In that case it would be possible for __schedule() to interleave
4861 * and observe the list empty.
4863 if (split && head == &balance_push_callback)
4866 rq->balance_callback = NULL;
4871 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4873 return __splice_balance_callbacks(rq, true);
4876 static void __balance_callbacks(struct rq *rq)
4878 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
4881 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4883 unsigned long flags;
4885 if (unlikely(head)) {
4886 raw_spin_rq_lock_irqsave(rq, flags);
4887 do_balance_callbacks(rq, head);
4888 raw_spin_rq_unlock_irqrestore(rq, flags);
4894 static inline void __balance_callbacks(struct rq *rq)
4898 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4903 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4910 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4913 * Since the runqueue lock will be released by the next
4914 * task (which is an invalid locking op but in the case
4915 * of the scheduler it's an obvious special-case), so we
4916 * do an early lockdep release here:
4918 rq_unpin_lock(rq, rf);
4919 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4920 #ifdef CONFIG_DEBUG_SPINLOCK
4921 /* this is a valid case when another task releases the spinlock */
4922 rq_lockp(rq)->owner = next;
4926 static inline void finish_lock_switch(struct rq *rq)
4929 * If we are tracking spinlock dependencies then we have to
4930 * fix up the runqueue lock - which gets 'carried over' from
4931 * prev into current:
4933 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4934 __balance_callbacks(rq);
4935 raw_spin_rq_unlock_irq(rq);
4939 * NOP if the arch has not defined these:
4942 #ifndef prepare_arch_switch
4943 # define prepare_arch_switch(next) do { } while (0)
4946 #ifndef finish_arch_post_lock_switch
4947 # define finish_arch_post_lock_switch() do { } while (0)
4950 static inline void kmap_local_sched_out(void)
4952 #ifdef CONFIG_KMAP_LOCAL
4953 if (unlikely(current->kmap_ctrl.idx))
4954 __kmap_local_sched_out();
4958 static inline void kmap_local_sched_in(void)
4960 #ifdef CONFIG_KMAP_LOCAL
4961 if (unlikely(current->kmap_ctrl.idx))
4962 __kmap_local_sched_in();
4967 * prepare_task_switch - prepare to switch tasks
4968 * @rq: the runqueue preparing to switch
4969 * @prev: the current task that is being switched out
4970 * @next: the task we are going to switch to.
4972 * This is called with the rq lock held and interrupts off. It must
4973 * be paired with a subsequent finish_task_switch after the context
4976 * prepare_task_switch sets up locking and calls architecture specific
4980 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4981 struct task_struct *next)
4983 kcov_prepare_switch(prev);
4984 sched_info_switch(rq, prev, next);
4985 perf_event_task_sched_out(prev, next);
4987 fire_sched_out_preempt_notifiers(prev, next);
4988 kmap_local_sched_out();
4990 prepare_arch_switch(next);
4994 * finish_task_switch - clean up after a task-switch
4995 * @prev: the thread we just switched away from.
4997 * finish_task_switch must be called after the context switch, paired
4998 * with a prepare_task_switch call before the context switch.
4999 * finish_task_switch will reconcile locking set up by prepare_task_switch,
5000 * and do any other architecture-specific cleanup actions.
5002 * Note that we may have delayed dropping an mm in context_switch(). If
5003 * so, we finish that here outside of the runqueue lock. (Doing it
5004 * with the lock held can cause deadlocks; see schedule() for
5007 * The context switch have flipped the stack from under us and restored the
5008 * local variables which were saved when this task called schedule() in the
5009 * past. prev == current is still correct but we need to recalculate this_rq
5010 * because prev may have moved to another CPU.
5012 static struct rq *finish_task_switch(struct task_struct *prev)
5013 __releases(rq->lock)
5015 struct rq *rq = this_rq();
5016 struct mm_struct *mm = rq->prev_mm;
5017 unsigned int prev_state;
5020 * The previous task will have left us with a preempt_count of 2
5021 * because it left us after:
5024 * preempt_disable(); // 1
5026 * raw_spin_lock_irq(&rq->lock) // 2
5028 * Also, see FORK_PREEMPT_COUNT.
5030 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5031 "corrupted preempt_count: %s/%d/0x%x\n",
5032 current->comm, current->pid, preempt_count()))
5033 preempt_count_set(FORK_PREEMPT_COUNT);
5038 * A task struct has one reference for the use as "current".
5039 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5040 * schedule one last time. The schedule call will never return, and
5041 * the scheduled task must drop that reference.
5043 * We must observe prev->state before clearing prev->on_cpu (in
5044 * finish_task), otherwise a concurrent wakeup can get prev
5045 * running on another CPU and we could rave with its RUNNING -> DEAD
5046 * transition, resulting in a double drop.
5048 prev_state = READ_ONCE(prev->__state);
5049 vtime_task_switch(prev);
5050 perf_event_task_sched_in(prev, current);
5052 tick_nohz_task_switch();
5053 finish_lock_switch(rq);
5054 finish_arch_post_lock_switch();
5055 kcov_finish_switch(current);
5057 * kmap_local_sched_out() is invoked with rq::lock held and
5058 * interrupts disabled. There is no requirement for that, but the
5059 * sched out code does not have an interrupt enabled section.
5060 * Restoring the maps on sched in does not require interrupts being
5063 kmap_local_sched_in();
5065 fire_sched_in_preempt_notifiers(current);
5067 * When switching through a kernel thread, the loop in
5068 * membarrier_{private,global}_expedited() may have observed that
5069 * kernel thread and not issued an IPI. It is therefore possible to
5070 * schedule between user->kernel->user threads without passing though
5071 * switch_mm(). Membarrier requires a barrier after storing to
5072 * rq->curr, before returning to userspace, so provide them here:
5074 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5075 * provided by mmdrop(),
5076 * - a sync_core for SYNC_CORE.
5079 membarrier_mm_sync_core_before_usermode(mm);
5082 if (unlikely(prev_state == TASK_DEAD)) {
5083 if (prev->sched_class->task_dead)
5084 prev->sched_class->task_dead(prev);
5086 /* Task is done with its stack. */
5087 put_task_stack(prev);
5089 put_task_struct_rcu_user(prev);
5096 * schedule_tail - first thing a freshly forked thread must call.
5097 * @prev: the thread we just switched away from.
5099 asmlinkage __visible void schedule_tail(struct task_struct *prev)
5100 __releases(rq->lock)
5103 * New tasks start with FORK_PREEMPT_COUNT, see there and
5104 * finish_task_switch() for details.
5106 * finish_task_switch() will drop rq->lock() and lower preempt_count
5107 * and the preempt_enable() will end up enabling preemption (on
5108 * PREEMPT_COUNT kernels).
5111 finish_task_switch(prev);
5114 if (current->set_child_tid)
5115 put_user(task_pid_vnr(current), current->set_child_tid);
5117 calculate_sigpending();
5121 * context_switch - switch to the new MM and the new thread's register state.
5123 static __always_inline struct rq *
5124 context_switch(struct rq *rq, struct task_struct *prev,
5125 struct task_struct *next, struct rq_flags *rf)
5127 prepare_task_switch(rq, prev, next);
5130 * For paravirt, this is coupled with an exit in switch_to to
5131 * combine the page table reload and the switch backend into
5134 arch_start_context_switch(prev);
5137 * kernel -> kernel lazy + transfer active
5138 * user -> kernel lazy + mmgrab() active
5140 * kernel -> user switch + mmdrop() active
5141 * user -> user switch
5143 if (!next->mm) { // to kernel
5144 enter_lazy_tlb(prev->active_mm, next);
5146 next->active_mm = prev->active_mm;
5147 if (prev->mm) // from user
5148 mmgrab(prev->active_mm);
5150 prev->active_mm = NULL;
5152 membarrier_switch_mm(rq, prev->active_mm, next->mm);
5154 * sys_membarrier() requires an smp_mb() between setting
5155 * rq->curr / membarrier_switch_mm() and returning to userspace.
5157 * The below provides this either through switch_mm(), or in
5158 * case 'prev->active_mm == next->mm' through
5159 * finish_task_switch()'s mmdrop().
5161 switch_mm_irqs_off(prev->active_mm, next->mm, next);
5163 if (!prev->mm) { // from kernel
5164 /* will mmdrop() in finish_task_switch(). */
5165 rq->prev_mm = prev->active_mm;
5166 prev->active_mm = NULL;
5170 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5172 prepare_lock_switch(rq, next, rf);
5174 /* Here we just switch the register state and the stack. */
5175 switch_to(prev, next, prev);
5178 return finish_task_switch(prev);
5182 * nr_running and nr_context_switches:
5184 * externally visible scheduler statistics: current number of runnable
5185 * threads, total number of context switches performed since bootup.
5187 unsigned int nr_running(void)
5189 unsigned int i, sum = 0;
5191 for_each_online_cpu(i)
5192 sum += cpu_rq(i)->nr_running;
5198 * Check if only the current task is running on the CPU.
5200 * Caution: this function does not check that the caller has disabled
5201 * preemption, thus the result might have a time-of-check-to-time-of-use
5202 * race. The caller is responsible to use it correctly, for example:
5204 * - from a non-preemptible section (of course)
5206 * - from a thread that is bound to a single CPU
5208 * - in a loop with very short iterations (e.g. a polling loop)
5210 bool single_task_running(void)
5212 return raw_rq()->nr_running == 1;
5214 EXPORT_SYMBOL(single_task_running);
5216 unsigned long long nr_context_switches(void)
5219 unsigned long long sum = 0;
5221 for_each_possible_cpu(i)
5222 sum += cpu_rq(i)->nr_switches;
5228 * Consumers of these two interfaces, like for example the cpuidle menu
5229 * governor, are using nonsensical data. Preferring shallow idle state selection
5230 * for a CPU that has IO-wait which might not even end up running the task when
5231 * it does become runnable.
5234 unsigned int nr_iowait_cpu(int cpu)
5236 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5240 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5242 * The idea behind IO-wait account is to account the idle time that we could
5243 * have spend running if it were not for IO. That is, if we were to improve the
5244 * storage performance, we'd have a proportional reduction in IO-wait time.
5246 * This all works nicely on UP, where, when a task blocks on IO, we account
5247 * idle time as IO-wait, because if the storage were faster, it could've been
5248 * running and we'd not be idle.
5250 * This has been extended to SMP, by doing the same for each CPU. This however
5253 * Imagine for instance the case where two tasks block on one CPU, only the one
5254 * CPU will have IO-wait accounted, while the other has regular idle. Even
5255 * though, if the storage were faster, both could've ran at the same time,
5256 * utilising both CPUs.
5258 * This means, that when looking globally, the current IO-wait accounting on
5259 * SMP is a lower bound, by reason of under accounting.
5261 * Worse, since the numbers are provided per CPU, they are sometimes
5262 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5263 * associated with any one particular CPU, it can wake to another CPU than it
5264 * blocked on. This means the per CPU IO-wait number is meaningless.
5266 * Task CPU affinities can make all that even more 'interesting'.
5269 unsigned int nr_iowait(void)
5271 unsigned int i, sum = 0;
5273 for_each_possible_cpu(i)
5274 sum += nr_iowait_cpu(i);
5282 * sched_exec - execve() is a valuable balancing opportunity, because at
5283 * this point the task has the smallest effective memory and cache footprint.
5285 void sched_exec(void)
5287 struct task_struct *p = current;
5288 unsigned long flags;
5291 raw_spin_lock_irqsave(&p->pi_lock, flags);
5292 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5293 if (dest_cpu == smp_processor_id())
5296 if (likely(cpu_active(dest_cpu))) {
5297 struct migration_arg arg = { p, dest_cpu };
5299 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5300 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5304 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5309 DEFINE_PER_CPU(struct kernel_stat, kstat);
5310 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5312 EXPORT_PER_CPU_SYMBOL(kstat);
5313 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5316 * The function fair_sched_class.update_curr accesses the struct curr
5317 * and its field curr->exec_start; when called from task_sched_runtime(),
5318 * we observe a high rate of cache misses in practice.
5319 * Prefetching this data results in improved performance.
5321 static inline void prefetch_curr_exec_start(struct task_struct *p)
5323 #ifdef CONFIG_FAIR_GROUP_SCHED
5324 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5326 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5329 prefetch(&curr->exec_start);
5333 * Return accounted runtime for the task.
5334 * In case the task is currently running, return the runtime plus current's
5335 * pending runtime that have not been accounted yet.
5337 unsigned long long task_sched_runtime(struct task_struct *p)
5343 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5345 * 64-bit doesn't need locks to atomically read a 64-bit value.
5346 * So we have a optimization chance when the task's delta_exec is 0.
5347 * Reading ->on_cpu is racy, but this is ok.
5349 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5350 * If we race with it entering CPU, unaccounted time is 0. This is
5351 * indistinguishable from the read occurring a few cycles earlier.
5352 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5353 * been accounted, so we're correct here as well.
5355 if (!p->on_cpu || !task_on_rq_queued(p))
5356 return p->se.sum_exec_runtime;
5359 rq = task_rq_lock(p, &rf);
5361 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5362 * project cycles that may never be accounted to this
5363 * thread, breaking clock_gettime().
5365 if (task_current(rq, p) && task_on_rq_queued(p)) {
5366 prefetch_curr_exec_start(p);
5367 update_rq_clock(rq);
5368 p->sched_class->update_curr(rq);
5370 ns = p->se.sum_exec_runtime;
5371 task_rq_unlock(rq, p, &rf);
5376 #ifdef CONFIG_SCHED_DEBUG
5377 static u64 cpu_resched_latency(struct rq *rq)
5379 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5380 u64 resched_latency, now = rq_clock(rq);
5381 static bool warned_once;
5383 if (sysctl_resched_latency_warn_once && warned_once)
5386 if (!need_resched() || !latency_warn_ms)
5389 if (system_state == SYSTEM_BOOTING)
5392 if (!rq->last_seen_need_resched_ns) {
5393 rq->last_seen_need_resched_ns = now;
5394 rq->ticks_without_resched = 0;
5398 rq->ticks_without_resched++;
5399 resched_latency = now - rq->last_seen_need_resched_ns;
5400 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5405 return resched_latency;
5408 static int __init setup_resched_latency_warn_ms(char *str)
5412 if ((kstrtol(str, 0, &val))) {
5413 pr_warn("Unable to set resched_latency_warn_ms\n");
5417 sysctl_resched_latency_warn_ms = val;
5420 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5422 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5423 #endif /* CONFIG_SCHED_DEBUG */
5426 * This function gets called by the timer code, with HZ frequency.
5427 * We call it with interrupts disabled.
5429 void scheduler_tick(void)
5431 int cpu = smp_processor_id();
5432 struct rq *rq = cpu_rq(cpu);
5433 struct task_struct *curr = rq->curr;
5435 unsigned long thermal_pressure;
5436 u64 resched_latency;
5438 arch_scale_freq_tick();
5443 update_rq_clock(rq);
5444 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5445 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5446 curr->sched_class->task_tick(rq, curr, 0);
5447 if (sched_feat(LATENCY_WARN))
5448 resched_latency = cpu_resched_latency(rq);
5449 calc_global_load_tick(rq);
5450 sched_core_tick(rq);
5454 if (sched_feat(LATENCY_WARN) && resched_latency)
5455 resched_latency_warn(cpu, resched_latency);
5457 perf_event_task_tick();
5460 rq->idle_balance = idle_cpu(cpu);
5461 trigger_load_balance(rq);
5465 #ifdef CONFIG_NO_HZ_FULL
5470 struct delayed_work work;
5472 /* Values for ->state, see diagram below. */
5473 #define TICK_SCHED_REMOTE_OFFLINE 0
5474 #define TICK_SCHED_REMOTE_OFFLINING 1
5475 #define TICK_SCHED_REMOTE_RUNNING 2
5478 * State diagram for ->state:
5481 * TICK_SCHED_REMOTE_OFFLINE
5484 * | | sched_tick_remote()
5487 * +--TICK_SCHED_REMOTE_OFFLINING
5490 * sched_tick_start() | | sched_tick_stop()
5493 * TICK_SCHED_REMOTE_RUNNING
5496 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5497 * and sched_tick_start() are happy to leave the state in RUNNING.
5500 static struct tick_work __percpu *tick_work_cpu;
5502 static void sched_tick_remote(struct work_struct *work)
5504 struct delayed_work *dwork = to_delayed_work(work);
5505 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5506 int cpu = twork->cpu;
5507 struct rq *rq = cpu_rq(cpu);
5508 struct task_struct *curr;
5514 * Handle the tick only if it appears the remote CPU is running in full
5515 * dynticks mode. The check is racy by nature, but missing a tick or
5516 * having one too much is no big deal because the scheduler tick updates
5517 * statistics and checks timeslices in a time-independent way, regardless
5518 * of when exactly it is running.
5520 if (!tick_nohz_tick_stopped_cpu(cpu))
5523 rq_lock_irq(rq, &rf);
5525 if (cpu_is_offline(cpu))
5528 update_rq_clock(rq);
5530 if (!is_idle_task(curr)) {
5532 * Make sure the next tick runs within a reasonable
5535 delta = rq_clock_task(rq) - curr->se.exec_start;
5536 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5538 curr->sched_class->task_tick(rq, curr, 0);
5540 calc_load_nohz_remote(rq);
5542 rq_unlock_irq(rq, &rf);
5546 * Run the remote tick once per second (1Hz). This arbitrary
5547 * frequency is large enough to avoid overload but short enough
5548 * to keep scheduler internal stats reasonably up to date. But
5549 * first update state to reflect hotplug activity if required.
5551 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5552 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5553 if (os == TICK_SCHED_REMOTE_RUNNING)
5554 queue_delayed_work(system_unbound_wq, dwork, HZ);
5557 static void sched_tick_start(int cpu)
5560 struct tick_work *twork;
5562 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5565 WARN_ON_ONCE(!tick_work_cpu);
5567 twork = per_cpu_ptr(tick_work_cpu, cpu);
5568 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5569 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5570 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5572 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5573 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5577 #ifdef CONFIG_HOTPLUG_CPU
5578 static void sched_tick_stop(int cpu)
5580 struct tick_work *twork;
5583 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5586 WARN_ON_ONCE(!tick_work_cpu);
5588 twork = per_cpu_ptr(tick_work_cpu, cpu);
5589 /* There cannot be competing actions, but don't rely on stop-machine. */
5590 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5591 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5592 /* Don't cancel, as this would mess up the state machine. */
5594 #endif /* CONFIG_HOTPLUG_CPU */
5596 int __init sched_tick_offload_init(void)
5598 tick_work_cpu = alloc_percpu(struct tick_work);
5599 BUG_ON(!tick_work_cpu);
5603 #else /* !CONFIG_NO_HZ_FULL */
5604 static inline void sched_tick_start(int cpu) { }
5605 static inline void sched_tick_stop(int cpu) { }
5608 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5609 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5611 * If the value passed in is equal to the current preempt count
5612 * then we just disabled preemption. Start timing the latency.
5614 static inline void preempt_latency_start(int val)
5616 if (preempt_count() == val) {
5617 unsigned long ip = get_lock_parent_ip();
5618 #ifdef CONFIG_DEBUG_PREEMPT
5619 current->preempt_disable_ip = ip;
5621 trace_preempt_off(CALLER_ADDR0, ip);
5625 void preempt_count_add(int val)
5627 #ifdef CONFIG_DEBUG_PREEMPT
5631 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5634 __preempt_count_add(val);
5635 #ifdef CONFIG_DEBUG_PREEMPT
5637 * Spinlock count overflowing soon?
5639 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5642 preempt_latency_start(val);
5644 EXPORT_SYMBOL(preempt_count_add);
5645 NOKPROBE_SYMBOL(preempt_count_add);
5648 * If the value passed in equals to the current preempt count
5649 * then we just enabled preemption. Stop timing the latency.
5651 static inline void preempt_latency_stop(int val)
5653 if (preempt_count() == val)
5654 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5657 void preempt_count_sub(int val)
5659 #ifdef CONFIG_DEBUG_PREEMPT
5663 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5666 * Is the spinlock portion underflowing?
5668 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5669 !(preempt_count() & PREEMPT_MASK)))
5673 preempt_latency_stop(val);
5674 __preempt_count_sub(val);
5676 EXPORT_SYMBOL(preempt_count_sub);
5677 NOKPROBE_SYMBOL(preempt_count_sub);
5680 static inline void preempt_latency_start(int val) { }
5681 static inline void preempt_latency_stop(int val) { }
5684 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5686 #ifdef CONFIG_DEBUG_PREEMPT
5687 return p->preempt_disable_ip;
5694 * Print scheduling while atomic bug:
5696 static noinline void __schedule_bug(struct task_struct *prev)
5698 /* Save this before calling printk(), since that will clobber it */
5699 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5701 if (oops_in_progress)
5704 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5705 prev->comm, prev->pid, preempt_count());
5707 debug_show_held_locks(prev);
5709 if (irqs_disabled())
5710 print_irqtrace_events(prev);
5711 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5712 && in_atomic_preempt_off()) {
5713 pr_err("Preemption disabled at:");
5714 print_ip_sym(KERN_ERR, preempt_disable_ip);
5717 panic("scheduling while atomic\n");
5720 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5724 * Various schedule()-time debugging checks and statistics:
5726 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5728 #ifdef CONFIG_SCHED_STACK_END_CHECK
5729 if (task_stack_end_corrupted(prev))
5730 panic("corrupted stack end detected inside scheduler\n");
5732 if (task_scs_end_corrupted(prev))
5733 panic("corrupted shadow stack detected inside scheduler\n");
5736 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5737 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5738 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5739 prev->comm, prev->pid, prev->non_block_count);
5741 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5745 if (unlikely(in_atomic_preempt_off())) {
5746 __schedule_bug(prev);
5747 preempt_count_set(PREEMPT_DISABLED);
5750 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5752 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5754 schedstat_inc(this_rq()->sched_count);
5757 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5758 struct rq_flags *rf)
5761 const struct sched_class *class;
5763 * We must do the balancing pass before put_prev_task(), such
5764 * that when we release the rq->lock the task is in the same
5765 * state as before we took rq->lock.
5767 * We can terminate the balance pass as soon as we know there is
5768 * a runnable task of @class priority or higher.
5770 for_class_range(class, prev->sched_class, &idle_sched_class) {
5771 if (class->balance(rq, prev, rf))
5776 put_prev_task(rq, prev);
5780 * Pick up the highest-prio task:
5782 static inline struct task_struct *
5783 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5785 const struct sched_class *class;
5786 struct task_struct *p;
5789 * Optimization: we know that if all tasks are in the fair class we can
5790 * call that function directly, but only if the @prev task wasn't of a
5791 * higher scheduling class, because otherwise those lose the
5792 * opportunity to pull in more work from other CPUs.
5794 if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
5795 rq->nr_running == rq->cfs.h_nr_running)) {
5797 p = pick_next_task_fair(rq, prev, rf);
5798 if (unlikely(p == RETRY_TASK))
5801 /* Assume the next prioritized class is idle_sched_class */
5803 put_prev_task(rq, prev);
5804 p = pick_next_task_idle(rq);
5811 put_prev_task_balance(rq, prev, rf);
5813 for_each_class(class) {
5814 p = class->pick_next_task(rq);
5819 BUG(); /* The idle class should always have a runnable task. */
5822 #ifdef CONFIG_SCHED_CORE
5823 static inline bool is_task_rq_idle(struct task_struct *t)
5825 return (task_rq(t)->idle == t);
5828 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5830 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5833 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5835 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5838 return a->core_cookie == b->core_cookie;
5841 static inline struct task_struct *pick_task(struct rq *rq)
5843 const struct sched_class *class;
5844 struct task_struct *p;
5846 for_each_class(class) {
5847 p = class->pick_task(rq);
5852 BUG(); /* The idle class should always have a runnable task. */
5855 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5857 static void queue_core_balance(struct rq *rq);
5859 static struct task_struct *
5860 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5862 struct task_struct *next, *p, *max = NULL;
5863 const struct cpumask *smt_mask;
5864 bool fi_before = false;
5865 bool core_clock_updated = (rq == rq->core);
5866 unsigned long cookie;
5867 int i, cpu, occ = 0;
5871 if (!sched_core_enabled(rq))
5872 return __pick_next_task(rq, prev, rf);
5876 /* Stopper task is switching into idle, no need core-wide selection. */
5877 if (cpu_is_offline(cpu)) {
5879 * Reset core_pick so that we don't enter the fastpath when
5880 * coming online. core_pick would already be migrated to
5881 * another cpu during offline.
5883 rq->core_pick = NULL;
5884 return __pick_next_task(rq, prev, rf);
5888 * If there were no {en,de}queues since we picked (IOW, the task
5889 * pointers are all still valid), and we haven't scheduled the last
5890 * pick yet, do so now.
5892 * rq->core_pick can be NULL if no selection was made for a CPU because
5893 * it was either offline or went offline during a sibling's core-wide
5894 * selection. In this case, do a core-wide selection.
5896 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5897 rq->core->core_pick_seq != rq->core_sched_seq &&
5899 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5901 next = rq->core_pick;
5903 put_prev_task(rq, prev);
5904 set_next_task(rq, next);
5907 rq->core_pick = NULL;
5911 put_prev_task_balance(rq, prev, rf);
5913 smt_mask = cpu_smt_mask(cpu);
5914 need_sync = !!rq->core->core_cookie;
5917 rq->core->core_cookie = 0UL;
5918 if (rq->core->core_forceidle_count) {
5919 if (!core_clock_updated) {
5920 update_rq_clock(rq->core);
5921 core_clock_updated = true;
5923 sched_core_account_forceidle(rq);
5924 /* reset after accounting force idle */
5925 rq->core->core_forceidle_start = 0;
5926 rq->core->core_forceidle_count = 0;
5927 rq->core->core_forceidle_occupation = 0;
5933 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5935 * @task_seq guards the task state ({en,de}queues)
5936 * @pick_seq is the @task_seq we did a selection on
5937 * @sched_seq is the @pick_seq we scheduled
5939 * However, preemptions can cause multiple picks on the same task set.
5940 * 'Fix' this by also increasing @task_seq for every pick.
5942 rq->core->core_task_seq++;
5945 * Optimize for common case where this CPU has no cookies
5946 * and there are no cookied tasks running on siblings.
5949 next = pick_task(rq);
5950 if (!next->core_cookie) {
5951 rq->core_pick = NULL;
5953 * For robustness, update the min_vruntime_fi for
5954 * unconstrained picks as well.
5956 WARN_ON_ONCE(fi_before);
5957 task_vruntime_update(rq, next, false);
5963 * For each thread: do the regular task pick and find the max prio task
5966 * Tie-break prio towards the current CPU
5968 for_each_cpu_wrap(i, smt_mask, cpu) {
5972 * Current cpu always has its clock updated on entrance to
5973 * pick_next_task(). If the current cpu is not the core,
5974 * the core may also have been updated above.
5976 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
5977 update_rq_clock(rq_i);
5979 p = rq_i->core_pick = pick_task(rq_i);
5980 if (!max || prio_less(max, p, fi_before))
5984 cookie = rq->core->core_cookie = max->core_cookie;
5987 * For each thread: try and find a runnable task that matches @max or
5990 for_each_cpu(i, smt_mask) {
5992 p = rq_i->core_pick;
5994 if (!cookie_equals(p, cookie)) {
5997 p = sched_core_find(rq_i, cookie);
5999 p = idle_sched_class.pick_task(rq_i);
6002 rq_i->core_pick = p;
6004 if (p == rq_i->idle) {
6005 if (rq_i->nr_running) {
6006 rq->core->core_forceidle_count++;
6008 rq->core->core_forceidle_seq++;
6015 if (schedstat_enabled() && rq->core->core_forceidle_count) {
6016 rq->core->core_forceidle_start = rq_clock(rq->core);
6017 rq->core->core_forceidle_occupation = occ;
6020 rq->core->core_pick_seq = rq->core->core_task_seq;
6021 next = rq->core_pick;
6022 rq->core_sched_seq = rq->core->core_pick_seq;
6024 /* Something should have been selected for current CPU */
6025 WARN_ON_ONCE(!next);
6028 * Reschedule siblings
6030 * NOTE: L1TF -- at this point we're no longer running the old task and
6031 * sending an IPI (below) ensures the sibling will no longer be running
6032 * their task. This ensures there is no inter-sibling overlap between
6033 * non-matching user state.
6035 for_each_cpu(i, smt_mask) {
6039 * An online sibling might have gone offline before a task
6040 * could be picked for it, or it might be offline but later
6041 * happen to come online, but its too late and nothing was
6042 * picked for it. That's Ok - it will pick tasks for itself,
6045 if (!rq_i->core_pick)
6049 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6050 * fi_before fi update?
6056 if (!(fi_before && rq->core->core_forceidle_count))
6057 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6059 rq_i->core_pick->core_occupation = occ;
6062 rq_i->core_pick = NULL;
6066 /* Did we break L1TF mitigation requirements? */
6067 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6069 if (rq_i->curr == rq_i->core_pick) {
6070 rq_i->core_pick = NULL;
6078 set_next_task(rq, next);
6080 if (rq->core->core_forceidle_count && next == rq->idle)
6081 queue_core_balance(rq);
6086 static bool try_steal_cookie(int this, int that)
6088 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6089 struct task_struct *p;
6090 unsigned long cookie;
6091 bool success = false;
6093 local_irq_disable();
6094 double_rq_lock(dst, src);
6096 cookie = dst->core->core_cookie;
6100 if (dst->curr != dst->idle)
6103 p = sched_core_find(src, cookie);
6108 if (p == src->core_pick || p == src->curr)
6111 if (!is_cpu_allowed(p, this))
6114 if (p->core_occupation > dst->idle->core_occupation)
6117 deactivate_task(src, p, 0);
6118 set_task_cpu(p, this);
6119 activate_task(dst, p, 0);
6127 p = sched_core_next(p, cookie);
6131 double_rq_unlock(dst, src);
6137 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6141 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
6148 if (try_steal_cookie(cpu, i))
6155 static void sched_core_balance(struct rq *rq)
6157 struct sched_domain *sd;
6158 int cpu = cpu_of(rq);
6162 raw_spin_rq_unlock_irq(rq);
6163 for_each_domain(cpu, sd) {
6167 if (steal_cookie_task(cpu, sd))
6170 raw_spin_rq_lock_irq(rq);
6175 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
6177 static void queue_core_balance(struct rq *rq)
6179 if (!sched_core_enabled(rq))
6182 if (!rq->core->core_cookie)
6185 if (!rq->nr_running) /* not forced idle */
6188 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6191 static void sched_core_cpu_starting(unsigned int cpu)
6193 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6194 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6195 unsigned long flags;
6198 sched_core_lock(cpu, &flags);
6200 WARN_ON_ONCE(rq->core != rq);
6202 /* if we're the first, we'll be our own leader */
6203 if (cpumask_weight(smt_mask) == 1)
6206 /* find the leader */
6207 for_each_cpu(t, smt_mask) {
6211 if (rq->core == rq) {
6217 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6220 /* install and validate core_rq */
6221 for_each_cpu(t, smt_mask) {
6227 WARN_ON_ONCE(rq->core != core_rq);
6231 sched_core_unlock(cpu, &flags);
6234 static void sched_core_cpu_deactivate(unsigned int cpu)
6236 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6237 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6238 unsigned long flags;
6241 sched_core_lock(cpu, &flags);
6243 /* if we're the last man standing, nothing to do */
6244 if (cpumask_weight(smt_mask) == 1) {
6245 WARN_ON_ONCE(rq->core != rq);
6249 /* if we're not the leader, nothing to do */
6253 /* find a new leader */
6254 for_each_cpu(t, smt_mask) {
6257 core_rq = cpu_rq(t);
6261 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6264 /* copy the shared state to the new leader */
6265 core_rq->core_task_seq = rq->core_task_seq;
6266 core_rq->core_pick_seq = rq->core_pick_seq;
6267 core_rq->core_cookie = rq->core_cookie;
6268 core_rq->core_forceidle_count = rq->core_forceidle_count;
6269 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6270 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6273 * Accounting edge for forced idle is handled in pick_next_task().
6274 * Don't need another one here, since the hotplug thread shouldn't
6277 core_rq->core_forceidle_start = 0;
6279 /* install new leader */
6280 for_each_cpu(t, smt_mask) {
6286 sched_core_unlock(cpu, &flags);
6289 static inline void sched_core_cpu_dying(unsigned int cpu)
6291 struct rq *rq = cpu_rq(cpu);
6297 #else /* !CONFIG_SCHED_CORE */
6299 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6300 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6301 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6303 static struct task_struct *
6304 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6306 return __pick_next_task(rq, prev, rf);
6309 #endif /* CONFIG_SCHED_CORE */
6312 * Constants for the sched_mode argument of __schedule().
6314 * The mode argument allows RT enabled kernels to differentiate a
6315 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6316 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6317 * optimize the AND operation out and just check for zero.
6320 #define SM_PREEMPT 0x1
6321 #define SM_RTLOCK_WAIT 0x2
6323 #ifndef CONFIG_PREEMPT_RT
6324 # define SM_MASK_PREEMPT (~0U)
6326 # define SM_MASK_PREEMPT SM_PREEMPT
6330 * __schedule() is the main scheduler function.
6332 * The main means of driving the scheduler and thus entering this function are:
6334 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6336 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6337 * paths. For example, see arch/x86/entry_64.S.
6339 * To drive preemption between tasks, the scheduler sets the flag in timer
6340 * interrupt handler scheduler_tick().
6342 * 3. Wakeups don't really cause entry into schedule(). They add a
6343 * task to the run-queue and that's it.
6345 * Now, if the new task added to the run-queue preempts the current
6346 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6347 * called on the nearest possible occasion:
6349 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6351 * - in syscall or exception context, at the next outmost
6352 * preempt_enable(). (this might be as soon as the wake_up()'s
6355 * - in IRQ context, return from interrupt-handler to
6356 * preemptible context
6358 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6361 * - cond_resched() call
6362 * - explicit schedule() call
6363 * - return from syscall or exception to user-space
6364 * - return from interrupt-handler to user-space
6366 * WARNING: must be called with preemption disabled!
6368 static void __sched notrace __schedule(unsigned int sched_mode)
6370 struct task_struct *prev, *next;
6371 unsigned long *switch_count;
6372 unsigned long prev_state;
6377 cpu = smp_processor_id();
6381 schedule_debug(prev, !!sched_mode);
6383 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6386 local_irq_disable();
6387 rcu_note_context_switch(!!sched_mode);
6390 * Make sure that signal_pending_state()->signal_pending() below
6391 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6392 * done by the caller to avoid the race with signal_wake_up():
6394 * __set_current_state(@state) signal_wake_up()
6395 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6396 * wake_up_state(p, state)
6397 * LOCK rq->lock LOCK p->pi_state
6398 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6399 * if (signal_pending_state()) if (p->state & @state)
6401 * Also, the membarrier system call requires a full memory barrier
6402 * after coming from user-space, before storing to rq->curr.
6405 smp_mb__after_spinlock();
6407 /* Promote REQ to ACT */
6408 rq->clock_update_flags <<= 1;
6409 update_rq_clock(rq);
6411 switch_count = &prev->nivcsw;
6414 * We must load prev->state once (task_struct::state is volatile), such
6415 * that we form a control dependency vs deactivate_task() below.
6417 prev_state = READ_ONCE(prev->__state);
6418 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6419 if (signal_pending_state(prev_state, prev)) {
6420 WRITE_ONCE(prev->__state, TASK_RUNNING);
6422 prev->sched_contributes_to_load =
6423 (prev_state & TASK_UNINTERRUPTIBLE) &&
6424 !(prev_state & TASK_NOLOAD) &&
6425 !(prev_state & TASK_FROZEN);
6427 if (prev->sched_contributes_to_load)
6428 rq->nr_uninterruptible++;
6431 * __schedule() ttwu()
6432 * prev_state = prev->state; if (p->on_rq && ...)
6433 * if (prev_state) goto out;
6434 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6435 * p->state = TASK_WAKING
6437 * Where __schedule() and ttwu() have matching control dependencies.
6439 * After this, schedule() must not care about p->state any more.
6441 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6443 if (prev->in_iowait) {
6444 atomic_inc(&rq->nr_iowait);
6445 delayacct_blkio_start();
6448 switch_count = &prev->nvcsw;
6451 next = pick_next_task(rq, prev, &rf);
6452 clear_tsk_need_resched(prev);
6453 clear_preempt_need_resched();
6454 #ifdef CONFIG_SCHED_DEBUG
6455 rq->last_seen_need_resched_ns = 0;
6458 if (likely(prev != next)) {
6461 * RCU users of rcu_dereference(rq->curr) may not see
6462 * changes to task_struct made by pick_next_task().
6464 RCU_INIT_POINTER(rq->curr, next);
6466 * The membarrier system call requires each architecture
6467 * to have a full memory barrier after updating
6468 * rq->curr, before returning to user-space.
6470 * Here are the schemes providing that barrier on the
6471 * various architectures:
6472 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6473 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6474 * - finish_lock_switch() for weakly-ordered
6475 * architectures where spin_unlock is a full barrier,
6476 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6477 * is a RELEASE barrier),
6481 migrate_disable_switch(rq, prev);
6482 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6484 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6486 /* Also unlocks the rq: */
6487 rq = context_switch(rq, prev, next, &rf);
6489 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6491 rq_unpin_lock(rq, &rf);
6492 __balance_callbacks(rq);
6493 raw_spin_rq_unlock_irq(rq);
6497 void __noreturn do_task_dead(void)
6499 /* Causes final put_task_struct in finish_task_switch(): */
6500 set_special_state(TASK_DEAD);
6502 /* Tell freezer to ignore us: */
6503 current->flags |= PF_NOFREEZE;
6505 __schedule(SM_NONE);
6508 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6513 static inline void sched_submit_work(struct task_struct *tsk)
6515 unsigned int task_flags;
6517 if (task_is_running(tsk))
6520 task_flags = tsk->flags;
6522 * If a worker goes to sleep, notify and ask workqueue whether it
6523 * wants to wake up a task to maintain concurrency.
6525 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6526 if (task_flags & PF_WQ_WORKER)
6527 wq_worker_sleeping(tsk);
6529 io_wq_worker_sleeping(tsk);
6533 * spinlock and rwlock must not flush block requests. This will
6534 * deadlock if the callback attempts to acquire a lock which is
6537 SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6540 * If we are going to sleep and we have plugged IO queued,
6541 * make sure to submit it to avoid deadlocks.
6543 blk_flush_plug(tsk->plug, true);
6546 static void sched_update_worker(struct task_struct *tsk)
6548 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6549 if (tsk->flags & PF_WQ_WORKER)
6550 wq_worker_running(tsk);
6552 io_wq_worker_running(tsk);
6556 asmlinkage __visible void __sched schedule(void)
6558 struct task_struct *tsk = current;
6560 sched_submit_work(tsk);
6563 __schedule(SM_NONE);
6564 sched_preempt_enable_no_resched();
6565 } while (need_resched());
6566 sched_update_worker(tsk);
6568 EXPORT_SYMBOL(schedule);
6571 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6572 * state (have scheduled out non-voluntarily) by making sure that all
6573 * tasks have either left the run queue or have gone into user space.
6574 * As idle tasks do not do either, they must not ever be preempted
6575 * (schedule out non-voluntarily).
6577 * schedule_idle() is similar to schedule_preempt_disable() except that it
6578 * never enables preemption because it does not call sched_submit_work().
6580 void __sched schedule_idle(void)
6583 * As this skips calling sched_submit_work(), which the idle task does
6584 * regardless because that function is a nop when the task is in a
6585 * TASK_RUNNING state, make sure this isn't used someplace that the
6586 * current task can be in any other state. Note, idle is always in the
6587 * TASK_RUNNING state.
6589 WARN_ON_ONCE(current->__state);
6591 __schedule(SM_NONE);
6592 } while (need_resched());
6595 #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6596 asmlinkage __visible void __sched schedule_user(void)
6599 * If we come here after a random call to set_need_resched(),
6600 * or we have been woken up remotely but the IPI has not yet arrived,
6601 * we haven't yet exited the RCU idle mode. Do it here manually until
6602 * we find a better solution.
6604 * NB: There are buggy callers of this function. Ideally we
6605 * should warn if prev_state != CONTEXT_USER, but that will trigger
6606 * too frequently to make sense yet.
6608 enum ctx_state prev_state = exception_enter();
6610 exception_exit(prev_state);
6615 * schedule_preempt_disabled - called with preemption disabled
6617 * Returns with preemption disabled. Note: preempt_count must be 1
6619 void __sched schedule_preempt_disabled(void)
6621 sched_preempt_enable_no_resched();
6626 #ifdef CONFIG_PREEMPT_RT
6627 void __sched notrace schedule_rtlock(void)
6631 __schedule(SM_RTLOCK_WAIT);
6632 sched_preempt_enable_no_resched();
6633 } while (need_resched());
6635 NOKPROBE_SYMBOL(schedule_rtlock);
6638 static void __sched notrace preempt_schedule_common(void)
6642 * Because the function tracer can trace preempt_count_sub()
6643 * and it also uses preempt_enable/disable_notrace(), if
6644 * NEED_RESCHED is set, the preempt_enable_notrace() called
6645 * by the function tracer will call this function again and
6646 * cause infinite recursion.
6648 * Preemption must be disabled here before the function
6649 * tracer can trace. Break up preempt_disable() into two
6650 * calls. One to disable preemption without fear of being
6651 * traced. The other to still record the preemption latency,
6652 * which can also be traced by the function tracer.
6654 preempt_disable_notrace();
6655 preempt_latency_start(1);
6656 __schedule(SM_PREEMPT);
6657 preempt_latency_stop(1);
6658 preempt_enable_no_resched_notrace();
6661 * Check again in case we missed a preemption opportunity
6662 * between schedule and now.
6664 } while (need_resched());
6667 #ifdef CONFIG_PREEMPTION
6669 * This is the entry point to schedule() from in-kernel preemption
6670 * off of preempt_enable.
6672 asmlinkage __visible void __sched notrace preempt_schedule(void)
6675 * If there is a non-zero preempt_count or interrupts are disabled,
6676 * we do not want to preempt the current task. Just return..
6678 if (likely(!preemptible()))
6680 preempt_schedule_common();
6682 NOKPROBE_SYMBOL(preempt_schedule);
6683 EXPORT_SYMBOL(preempt_schedule);
6685 #ifdef CONFIG_PREEMPT_DYNAMIC
6686 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6687 #ifndef preempt_schedule_dynamic_enabled
6688 #define preempt_schedule_dynamic_enabled preempt_schedule
6689 #define preempt_schedule_dynamic_disabled NULL
6691 DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6692 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6693 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6694 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6695 void __sched notrace dynamic_preempt_schedule(void)
6697 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6701 NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6702 EXPORT_SYMBOL(dynamic_preempt_schedule);
6707 * preempt_schedule_notrace - preempt_schedule called by tracing
6709 * The tracing infrastructure uses preempt_enable_notrace to prevent
6710 * recursion and tracing preempt enabling caused by the tracing
6711 * infrastructure itself. But as tracing can happen in areas coming
6712 * from userspace or just about to enter userspace, a preempt enable
6713 * can occur before user_exit() is called. This will cause the scheduler
6714 * to be called when the system is still in usermode.
6716 * To prevent this, the preempt_enable_notrace will use this function
6717 * instead of preempt_schedule() to exit user context if needed before
6718 * calling the scheduler.
6720 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6722 enum ctx_state prev_ctx;
6724 if (likely(!preemptible()))
6729 * Because the function tracer can trace preempt_count_sub()
6730 * and it also uses preempt_enable/disable_notrace(), if
6731 * NEED_RESCHED is set, the preempt_enable_notrace() called
6732 * by the function tracer will call this function again and
6733 * cause infinite recursion.
6735 * Preemption must be disabled here before the function
6736 * tracer can trace. Break up preempt_disable() into two
6737 * calls. One to disable preemption without fear of being
6738 * traced. The other to still record the preemption latency,
6739 * which can also be traced by the function tracer.
6741 preempt_disable_notrace();
6742 preempt_latency_start(1);
6744 * Needs preempt disabled in case user_exit() is traced
6745 * and the tracer calls preempt_enable_notrace() causing
6746 * an infinite recursion.
6748 prev_ctx = exception_enter();
6749 __schedule(SM_PREEMPT);
6750 exception_exit(prev_ctx);
6752 preempt_latency_stop(1);
6753 preempt_enable_no_resched_notrace();
6754 } while (need_resched());
6756 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6758 #ifdef CONFIG_PREEMPT_DYNAMIC
6759 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6760 #ifndef preempt_schedule_notrace_dynamic_enabled
6761 #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6762 #define preempt_schedule_notrace_dynamic_disabled NULL
6764 DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6765 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6766 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6767 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6768 void __sched notrace dynamic_preempt_schedule_notrace(void)
6770 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6772 preempt_schedule_notrace();
6774 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6775 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6779 #endif /* CONFIG_PREEMPTION */
6782 * This is the entry point to schedule() from kernel preemption
6783 * off of irq context.
6784 * Note, that this is called and return with irqs disabled. This will
6785 * protect us against recursive calling from irq.
6787 asmlinkage __visible void __sched preempt_schedule_irq(void)
6789 enum ctx_state prev_state;
6791 /* Catch callers which need to be fixed */
6792 BUG_ON(preempt_count() || !irqs_disabled());
6794 prev_state = exception_enter();
6799 __schedule(SM_PREEMPT);
6800 local_irq_disable();
6801 sched_preempt_enable_no_resched();
6802 } while (need_resched());
6804 exception_exit(prev_state);
6807 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6810 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6811 return try_to_wake_up(curr->private, mode, wake_flags);
6813 EXPORT_SYMBOL(default_wake_function);
6815 static void __setscheduler_prio(struct task_struct *p, int prio)
6818 p->sched_class = &dl_sched_class;
6819 else if (rt_prio(prio))
6820 p->sched_class = &rt_sched_class;
6822 p->sched_class = &fair_sched_class;
6827 #ifdef CONFIG_RT_MUTEXES
6829 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6832 prio = min(prio, pi_task->prio);
6837 static inline int rt_effective_prio(struct task_struct *p, int prio)
6839 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6841 return __rt_effective_prio(pi_task, prio);
6845 * rt_mutex_setprio - set the current priority of a task
6847 * @pi_task: donor task
6849 * This function changes the 'effective' priority of a task. It does
6850 * not touch ->normal_prio like __setscheduler().
6852 * Used by the rt_mutex code to implement priority inheritance
6853 * logic. Call site only calls if the priority of the task changed.
6855 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6857 int prio, oldprio, queued, running, queue_flag =
6858 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6859 const struct sched_class *prev_class;
6863 /* XXX used to be waiter->prio, not waiter->task->prio */
6864 prio = __rt_effective_prio(pi_task, p->normal_prio);
6867 * If nothing changed; bail early.
6869 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6872 rq = __task_rq_lock(p, &rf);
6873 update_rq_clock(rq);
6875 * Set under pi_lock && rq->lock, such that the value can be used under
6878 * Note that there is loads of tricky to make this pointer cache work
6879 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6880 * ensure a task is de-boosted (pi_task is set to NULL) before the
6881 * task is allowed to run again (and can exit). This ensures the pointer
6882 * points to a blocked task -- which guarantees the task is present.
6884 p->pi_top_task = pi_task;
6887 * For FIFO/RR we only need to set prio, if that matches we're done.
6889 if (prio == p->prio && !dl_prio(prio))
6893 * Idle task boosting is a nono in general. There is one
6894 * exception, when PREEMPT_RT and NOHZ is active:
6896 * The idle task calls get_next_timer_interrupt() and holds
6897 * the timer wheel base->lock on the CPU and another CPU wants
6898 * to access the timer (probably to cancel it). We can safely
6899 * ignore the boosting request, as the idle CPU runs this code
6900 * with interrupts disabled and will complete the lock
6901 * protected section without being interrupted. So there is no
6902 * real need to boost.
6904 if (unlikely(p == rq->idle)) {
6905 WARN_ON(p != rq->curr);
6906 WARN_ON(p->pi_blocked_on);
6910 trace_sched_pi_setprio(p, pi_task);
6913 if (oldprio == prio)
6914 queue_flag &= ~DEQUEUE_MOVE;
6916 prev_class = p->sched_class;
6917 queued = task_on_rq_queued(p);
6918 running = task_current(rq, p);
6920 dequeue_task(rq, p, queue_flag);
6922 put_prev_task(rq, p);
6925 * Boosting condition are:
6926 * 1. -rt task is running and holds mutex A
6927 * --> -dl task blocks on mutex A
6929 * 2. -dl task is running and holds mutex A
6930 * --> -dl task blocks on mutex A and could preempt the
6933 if (dl_prio(prio)) {
6934 if (!dl_prio(p->normal_prio) ||
6935 (pi_task && dl_prio(pi_task->prio) &&
6936 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6937 p->dl.pi_se = pi_task->dl.pi_se;
6938 queue_flag |= ENQUEUE_REPLENISH;
6940 p->dl.pi_se = &p->dl;
6942 } else if (rt_prio(prio)) {
6943 if (dl_prio(oldprio))
6944 p->dl.pi_se = &p->dl;
6946 queue_flag |= ENQUEUE_HEAD;
6948 if (dl_prio(oldprio))
6949 p->dl.pi_se = &p->dl;
6950 if (rt_prio(oldprio))
6954 __setscheduler_prio(p, prio);
6957 enqueue_task(rq, p, queue_flag);
6959 set_next_task(rq, p);
6961 check_class_changed(rq, p, prev_class, oldprio);
6963 /* Avoid rq from going away on us: */
6966 rq_unpin_lock(rq, &rf);
6967 __balance_callbacks(rq);
6968 raw_spin_rq_unlock(rq);
6973 static inline int rt_effective_prio(struct task_struct *p, int prio)
6979 void set_user_nice(struct task_struct *p, long nice)
6981 bool queued, running;
6986 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6989 * We have to be careful, if called from sys_setpriority(),
6990 * the task might be in the middle of scheduling on another CPU.
6992 rq = task_rq_lock(p, &rf);
6993 update_rq_clock(rq);
6996 * The RT priorities are set via sched_setscheduler(), but we still
6997 * allow the 'normal' nice value to be set - but as expected
6998 * it won't have any effect on scheduling until the task is
6999 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7001 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7002 p->static_prio = NICE_TO_PRIO(nice);
7005 queued = task_on_rq_queued(p);
7006 running = task_current(rq, p);
7008 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7010 put_prev_task(rq, p);
7012 p->static_prio = NICE_TO_PRIO(nice);
7013 set_load_weight(p, true);
7015 p->prio = effective_prio(p);
7018 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7020 set_next_task(rq, p);
7023 * If the task increased its priority or is running and
7024 * lowered its priority, then reschedule its CPU:
7026 p->sched_class->prio_changed(rq, p, old_prio);
7029 task_rq_unlock(rq, p, &rf);
7031 EXPORT_SYMBOL(set_user_nice);
7034 * is_nice_reduction - check if nice value is an actual reduction
7036 * Similar to can_nice() but does not perform a capability check.
7041 static bool is_nice_reduction(const struct task_struct *p, const int nice)
7043 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7044 int nice_rlim = nice_to_rlimit(nice);
7046 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7050 * can_nice - check if a task can reduce its nice value
7054 int can_nice(const struct task_struct *p, const int nice)
7056 return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7059 #ifdef __ARCH_WANT_SYS_NICE
7062 * sys_nice - change the priority of the current process.
7063 * @increment: priority increment
7065 * sys_setpriority is a more generic, but much slower function that
7066 * does similar things.
7068 SYSCALL_DEFINE1(nice, int, increment)
7073 * Setpriority might change our priority at the same moment.
7074 * We don't have to worry. Conceptually one call occurs first
7075 * and we have a single winner.
7077 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7078 nice = task_nice(current) + increment;
7080 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7081 if (increment < 0 && !can_nice(current, nice))
7084 retval = security_task_setnice(current, nice);
7088 set_user_nice(current, nice);
7095 * task_prio - return the priority value of a given task.
7096 * @p: the task in question.
7098 * Return: The priority value as seen by users in /proc.
7100 * sched policy return value kernel prio user prio/nice
7102 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7103 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7104 * deadline -101 -1 0
7106 int task_prio(const struct task_struct *p)
7108 return p->prio - MAX_RT_PRIO;
7112 * idle_cpu - is a given CPU idle currently?
7113 * @cpu: the processor in question.
7115 * Return: 1 if the CPU is currently idle. 0 otherwise.
7117 int idle_cpu(int cpu)
7119 struct rq *rq = cpu_rq(cpu);
7121 if (rq->curr != rq->idle)
7128 if (rq->ttwu_pending)
7136 * available_idle_cpu - is a given CPU idle for enqueuing work.
7137 * @cpu: the CPU in question.
7139 * Return: 1 if the CPU is currently idle. 0 otherwise.
7141 int available_idle_cpu(int cpu)
7146 if (vcpu_is_preempted(cpu))
7153 * idle_task - return the idle task for a given CPU.
7154 * @cpu: the processor in question.
7156 * Return: The idle task for the CPU @cpu.
7158 struct task_struct *idle_task(int cpu)
7160 return cpu_rq(cpu)->idle;
7165 * This function computes an effective utilization for the given CPU, to be
7166 * used for frequency selection given the linear relation: f = u * f_max.
7168 * The scheduler tracks the following metrics:
7170 * cpu_util_{cfs,rt,dl,irq}()
7173 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7174 * synchronized windows and are thus directly comparable.
7176 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7177 * which excludes things like IRQ and steal-time. These latter are then accrued
7178 * in the irq utilization.
7180 * The DL bandwidth number otoh is not a measured metric but a value computed
7181 * based on the task model parameters and gives the minimal utilization
7182 * required to meet deadlines.
7184 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7185 enum cpu_util_type type,
7186 struct task_struct *p)
7188 unsigned long dl_util, util, irq, max;
7189 struct rq *rq = cpu_rq(cpu);
7191 max = arch_scale_cpu_capacity(cpu);
7193 if (!uclamp_is_used() &&
7194 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7199 * Early check to see if IRQ/steal time saturates the CPU, can be
7200 * because of inaccuracies in how we track these -- see
7201 * update_irq_load_avg().
7203 irq = cpu_util_irq(rq);
7204 if (unlikely(irq >= max))
7208 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7209 * CFS tasks and we use the same metric to track the effective
7210 * utilization (PELT windows are synchronized) we can directly add them
7211 * to obtain the CPU's actual utilization.
7213 * CFS and RT utilization can be boosted or capped, depending on
7214 * utilization clamp constraints requested by currently RUNNABLE
7216 * When there are no CFS RUNNABLE tasks, clamps are released and
7217 * frequency will be gracefully reduced with the utilization decay.
7219 util = util_cfs + cpu_util_rt(rq);
7220 if (type == FREQUENCY_UTIL)
7221 util = uclamp_rq_util_with(rq, util, p);
7223 dl_util = cpu_util_dl(rq);
7226 * For frequency selection we do not make cpu_util_dl() a permanent part
7227 * of this sum because we want to use cpu_bw_dl() later on, but we need
7228 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7229 * that we select f_max when there is no idle time.
7231 * NOTE: numerical errors or stop class might cause us to not quite hit
7232 * saturation when we should -- something for later.
7234 if (util + dl_util >= max)
7238 * OTOH, for energy computation we need the estimated running time, so
7239 * include util_dl and ignore dl_bw.
7241 if (type == ENERGY_UTIL)
7245 * There is still idle time; further improve the number by using the
7246 * irq metric. Because IRQ/steal time is hidden from the task clock we
7247 * need to scale the task numbers:
7250 * U' = irq + --------- * U
7253 util = scale_irq_capacity(util, irq, max);
7257 * Bandwidth required by DEADLINE must always be granted while, for
7258 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7259 * to gracefully reduce the frequency when no tasks show up for longer
7262 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7263 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7264 * an interface. So, we only do the latter for now.
7266 if (type == FREQUENCY_UTIL)
7267 util += cpu_bw_dl(rq);
7269 return min(max, util);
7272 unsigned long sched_cpu_util(int cpu)
7274 return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
7276 #endif /* CONFIG_SMP */
7279 * find_process_by_pid - find a process with a matching PID value.
7280 * @pid: the pid in question.
7282 * The task of @pid, if found. %NULL otherwise.
7284 static struct task_struct *find_process_by_pid(pid_t pid)
7286 return pid ? find_task_by_vpid(pid) : current;
7290 * sched_setparam() passes in -1 for its policy, to let the functions
7291 * it calls know not to change it.
7293 #define SETPARAM_POLICY -1
7295 static void __setscheduler_params(struct task_struct *p,
7296 const struct sched_attr *attr)
7298 int policy = attr->sched_policy;
7300 if (policy == SETPARAM_POLICY)
7305 if (dl_policy(policy))
7306 __setparam_dl(p, attr);
7307 else if (fair_policy(policy))
7308 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7311 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7312 * !rt_policy. Always setting this ensures that things like
7313 * getparam()/getattr() don't report silly values for !rt tasks.
7315 p->rt_priority = attr->sched_priority;
7316 p->normal_prio = normal_prio(p);
7317 set_load_weight(p, true);
7321 * Check the target process has a UID that matches the current process's:
7323 static bool check_same_owner(struct task_struct *p)
7325 const struct cred *cred = current_cred(), *pcred;
7329 pcred = __task_cred(p);
7330 match = (uid_eq(cred->euid, pcred->euid) ||
7331 uid_eq(cred->euid, pcred->uid));
7337 * Allow unprivileged RT tasks to decrease priority.
7338 * Only issue a capable test if needed and only once to avoid an audit
7339 * event on permitted non-privileged operations:
7341 static int user_check_sched_setscheduler(struct task_struct *p,
7342 const struct sched_attr *attr,
7343 int policy, int reset_on_fork)
7345 if (fair_policy(policy)) {
7346 if (attr->sched_nice < task_nice(p) &&
7347 !is_nice_reduction(p, attr->sched_nice))
7351 if (rt_policy(policy)) {
7352 unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7354 /* Can't set/change the rt policy: */
7355 if (policy != p->policy && !rlim_rtprio)
7358 /* Can't increase priority: */
7359 if (attr->sched_priority > p->rt_priority &&
7360 attr->sched_priority > rlim_rtprio)
7365 * Can't set/change SCHED_DEADLINE policy at all for now
7366 * (safest behavior); in the future we would like to allow
7367 * unprivileged DL tasks to increase their relative deadline
7368 * or reduce their runtime (both ways reducing utilization)
7370 if (dl_policy(policy))
7374 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7375 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7377 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7378 if (!is_nice_reduction(p, task_nice(p)))
7382 /* Can't change other user's priorities: */
7383 if (!check_same_owner(p))
7386 /* Normal users shall not reset the sched_reset_on_fork flag: */
7387 if (p->sched_reset_on_fork && !reset_on_fork)
7393 if (!capable(CAP_SYS_NICE))
7399 static int __sched_setscheduler(struct task_struct *p,
7400 const struct sched_attr *attr,
7403 int oldpolicy = -1, policy = attr->sched_policy;
7404 int retval, oldprio, newprio, queued, running;
7405 const struct sched_class *prev_class;
7406 struct callback_head *head;
7409 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7412 /* The pi code expects interrupts enabled */
7413 BUG_ON(pi && in_interrupt());
7415 /* Double check policy once rq lock held: */
7417 reset_on_fork = p->sched_reset_on_fork;
7418 policy = oldpolicy = p->policy;
7420 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7422 if (!valid_policy(policy))
7426 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7430 * Valid priorities for SCHED_FIFO and SCHED_RR are
7431 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7432 * SCHED_BATCH and SCHED_IDLE is 0.
7434 if (attr->sched_priority > MAX_RT_PRIO-1)
7436 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7437 (rt_policy(policy) != (attr->sched_priority != 0)))
7441 retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7445 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7448 retval = security_task_setscheduler(p);
7453 /* Update task specific "requested" clamps */
7454 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7455 retval = uclamp_validate(p, attr);
7464 * Make sure no PI-waiters arrive (or leave) while we are
7465 * changing the priority of the task:
7467 * To be able to change p->policy safely, the appropriate
7468 * runqueue lock must be held.
7470 rq = task_rq_lock(p, &rf);
7471 update_rq_clock(rq);
7474 * Changing the policy of the stop threads its a very bad idea:
7476 if (p == rq->stop) {
7482 * If not changing anything there's no need to proceed further,
7483 * but store a possible modification of reset_on_fork.
7485 if (unlikely(policy == p->policy)) {
7486 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7488 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7490 if (dl_policy(policy) && dl_param_changed(p, attr))
7492 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7495 p->sched_reset_on_fork = reset_on_fork;
7502 #ifdef CONFIG_RT_GROUP_SCHED
7504 * Do not allow realtime tasks into groups that have no runtime
7507 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7508 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7509 !task_group_is_autogroup(task_group(p))) {
7515 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7516 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7517 cpumask_t *span = rq->rd->span;
7520 * Don't allow tasks with an affinity mask smaller than
7521 * the entire root_domain to become SCHED_DEADLINE. We
7522 * will also fail if there's no bandwidth available.
7524 if (!cpumask_subset(span, p->cpus_ptr) ||
7525 rq->rd->dl_bw.bw == 0) {
7533 /* Re-check policy now with rq lock held: */
7534 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7535 policy = oldpolicy = -1;
7536 task_rq_unlock(rq, p, &rf);
7538 cpuset_read_unlock();
7543 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7544 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7547 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7552 p->sched_reset_on_fork = reset_on_fork;
7555 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7558 * Take priority boosted tasks into account. If the new
7559 * effective priority is unchanged, we just store the new
7560 * normal parameters and do not touch the scheduler class and
7561 * the runqueue. This will be done when the task deboost
7564 newprio = rt_effective_prio(p, newprio);
7565 if (newprio == oldprio)
7566 queue_flags &= ~DEQUEUE_MOVE;
7569 queued = task_on_rq_queued(p);
7570 running = task_current(rq, p);
7572 dequeue_task(rq, p, queue_flags);
7574 put_prev_task(rq, p);
7576 prev_class = p->sched_class;
7578 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7579 __setscheduler_params(p, attr);
7580 __setscheduler_prio(p, newprio);
7582 __setscheduler_uclamp(p, attr);
7586 * We enqueue to tail when the priority of a task is
7587 * increased (user space view).
7589 if (oldprio < p->prio)
7590 queue_flags |= ENQUEUE_HEAD;
7592 enqueue_task(rq, p, queue_flags);
7595 set_next_task(rq, p);
7597 check_class_changed(rq, p, prev_class, oldprio);
7599 /* Avoid rq from going away on us: */
7601 head = splice_balance_callbacks(rq);
7602 task_rq_unlock(rq, p, &rf);
7605 cpuset_read_unlock();
7606 rt_mutex_adjust_pi(p);
7609 /* Run balance callbacks after we've adjusted the PI chain: */
7610 balance_callbacks(rq, head);
7616 task_rq_unlock(rq, p, &rf);
7618 cpuset_read_unlock();
7622 static int _sched_setscheduler(struct task_struct *p, int policy,
7623 const struct sched_param *param, bool check)
7625 struct sched_attr attr = {
7626 .sched_policy = policy,
7627 .sched_priority = param->sched_priority,
7628 .sched_nice = PRIO_TO_NICE(p->static_prio),
7631 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7632 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7633 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7634 policy &= ~SCHED_RESET_ON_FORK;
7635 attr.sched_policy = policy;
7638 return __sched_setscheduler(p, &attr, check, true);
7641 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7642 * @p: the task in question.
7643 * @policy: new policy.
7644 * @param: structure containing the new RT priority.
7646 * Use sched_set_fifo(), read its comment.
7648 * Return: 0 on success. An error code otherwise.
7650 * NOTE that the task may be already dead.
7652 int sched_setscheduler(struct task_struct *p, int policy,
7653 const struct sched_param *param)
7655 return _sched_setscheduler(p, policy, param, true);
7658 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7660 return __sched_setscheduler(p, attr, true, true);
7663 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7665 return __sched_setscheduler(p, attr, false, true);
7667 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7670 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7671 * @p: the task in question.
7672 * @policy: new policy.
7673 * @param: structure containing the new RT priority.
7675 * Just like sched_setscheduler, only don't bother checking if the
7676 * current context has permission. For example, this is needed in
7677 * stop_machine(): we create temporary high priority worker threads,
7678 * but our caller might not have that capability.
7680 * Return: 0 on success. An error code otherwise.
7682 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7683 const struct sched_param *param)
7685 return _sched_setscheduler(p, policy, param, false);
7689 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7690 * incapable of resource management, which is the one thing an OS really should
7693 * This is of course the reason it is limited to privileged users only.
7695 * Worse still; it is fundamentally impossible to compose static priority
7696 * workloads. You cannot take two correctly working static prio workloads
7697 * and smash them together and still expect them to work.
7699 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7703 * The administrator _MUST_ configure the system, the kernel simply doesn't
7704 * know enough information to make a sensible choice.
7706 void sched_set_fifo(struct task_struct *p)
7708 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7709 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7711 EXPORT_SYMBOL_GPL(sched_set_fifo);
7714 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7716 void sched_set_fifo_low(struct task_struct *p)
7718 struct sched_param sp = { .sched_priority = 1 };
7719 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7721 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7723 void sched_set_normal(struct task_struct *p, int nice)
7725 struct sched_attr attr = {
7726 .sched_policy = SCHED_NORMAL,
7729 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7731 EXPORT_SYMBOL_GPL(sched_set_normal);
7734 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7736 struct sched_param lparam;
7737 struct task_struct *p;
7740 if (!param || pid < 0)
7742 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7747 p = find_process_by_pid(pid);
7753 retval = sched_setscheduler(p, policy, &lparam);
7761 * Mimics kernel/events/core.c perf_copy_attr().
7763 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7768 /* Zero the full structure, so that a short copy will be nice: */
7769 memset(attr, 0, sizeof(*attr));
7771 ret = get_user(size, &uattr->size);
7775 /* ABI compatibility quirk: */
7777 size = SCHED_ATTR_SIZE_VER0;
7778 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7781 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7788 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7789 size < SCHED_ATTR_SIZE_VER1)
7793 * XXX: Do we want to be lenient like existing syscalls; or do we want
7794 * to be strict and return an error on out-of-bounds values?
7796 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7801 put_user(sizeof(*attr), &uattr->size);
7805 static void get_params(struct task_struct *p, struct sched_attr *attr)
7807 if (task_has_dl_policy(p))
7808 __getparam_dl(p, attr);
7809 else if (task_has_rt_policy(p))
7810 attr->sched_priority = p->rt_priority;
7812 attr->sched_nice = task_nice(p);
7816 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7817 * @pid: the pid in question.
7818 * @policy: new policy.
7819 * @param: structure containing the new RT priority.
7821 * Return: 0 on success. An error code otherwise.
7823 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7828 return do_sched_setscheduler(pid, policy, param);
7832 * sys_sched_setparam - set/change the RT priority of a thread
7833 * @pid: the pid in question.
7834 * @param: structure containing the new RT priority.
7836 * Return: 0 on success. An error code otherwise.
7838 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7840 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7844 * sys_sched_setattr - same as above, but with extended sched_attr
7845 * @pid: the pid in question.
7846 * @uattr: structure containing the extended parameters.
7847 * @flags: for future extension.
7849 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7850 unsigned int, flags)
7852 struct sched_attr attr;
7853 struct task_struct *p;
7856 if (!uattr || pid < 0 || flags)
7859 retval = sched_copy_attr(uattr, &attr);
7863 if ((int)attr.sched_policy < 0)
7865 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7866 attr.sched_policy = SETPARAM_POLICY;
7870 p = find_process_by_pid(pid);
7876 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7877 get_params(p, &attr);
7878 retval = sched_setattr(p, &attr);
7886 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7887 * @pid: the pid in question.
7889 * Return: On success, the policy of the thread. Otherwise, a negative error
7892 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7894 struct task_struct *p;
7902 p = find_process_by_pid(pid);
7904 retval = security_task_getscheduler(p);
7907 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7914 * sys_sched_getparam - get the RT priority of a thread
7915 * @pid: the pid in question.
7916 * @param: structure containing the RT priority.
7918 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7921 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7923 struct sched_param lp = { .sched_priority = 0 };
7924 struct task_struct *p;
7927 if (!param || pid < 0)
7931 p = find_process_by_pid(pid);
7936 retval = security_task_getscheduler(p);
7940 if (task_has_rt_policy(p))
7941 lp.sched_priority = p->rt_priority;
7945 * This one might sleep, we cannot do it with a spinlock held ...
7947 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7957 * Copy the kernel size attribute structure (which might be larger
7958 * than what user-space knows about) to user-space.
7960 * Note that all cases are valid: user-space buffer can be larger or
7961 * smaller than the kernel-space buffer. The usual case is that both
7962 * have the same size.
7965 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7966 struct sched_attr *kattr,
7969 unsigned int ksize = sizeof(*kattr);
7971 if (!access_ok(uattr, usize))
7975 * sched_getattr() ABI forwards and backwards compatibility:
7977 * If usize == ksize then we just copy everything to user-space and all is good.
7979 * If usize < ksize then we only copy as much as user-space has space for,
7980 * this keeps ABI compatibility as well. We skip the rest.
7982 * If usize > ksize then user-space is using a newer version of the ABI,
7983 * which part the kernel doesn't know about. Just ignore it - tooling can
7984 * detect the kernel's knowledge of attributes from the attr->size value
7985 * which is set to ksize in this case.
7987 kattr->size = min(usize, ksize);
7989 if (copy_to_user(uattr, kattr, kattr->size))
7996 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7997 * @pid: the pid in question.
7998 * @uattr: structure containing the extended parameters.
7999 * @usize: sizeof(attr) for fwd/bwd comp.
8000 * @flags: for future extension.
8002 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8003 unsigned int, usize, unsigned int, flags)
8005 struct sched_attr kattr = { };
8006 struct task_struct *p;
8009 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8010 usize < SCHED_ATTR_SIZE_VER0 || flags)
8014 p = find_process_by_pid(pid);
8019 retval = security_task_getscheduler(p);
8023 kattr.sched_policy = p->policy;
8024 if (p->sched_reset_on_fork)
8025 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8026 get_params(p, &kattr);
8027 kattr.sched_flags &= SCHED_FLAG_ALL;
8029 #ifdef CONFIG_UCLAMP_TASK
8031 * This could race with another potential updater, but this is fine
8032 * because it'll correctly read the old or the new value. We don't need
8033 * to guarantee who wins the race as long as it doesn't return garbage.
8035 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8036 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8041 return sched_attr_copy_to_user(uattr, &kattr, usize);
8049 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8054 * If the task isn't a deadline task or admission control is
8055 * disabled then we don't care about affinity changes.
8057 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8061 * Since bandwidth control happens on root_domain basis,
8062 * if admission test is enabled, we only admit -deadline
8063 * tasks allowed to run on all the CPUs in the task's
8067 if (!cpumask_subset(task_rq(p)->rd->span, mask))
8075 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
8078 cpumask_var_t cpus_allowed, new_mask;
8080 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8083 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8085 goto out_free_cpus_allowed;
8088 cpuset_cpus_allowed(p, cpus_allowed);
8089 cpumask_and(new_mask, mask, cpus_allowed);
8091 retval = dl_task_check_affinity(p, new_mask);
8093 goto out_free_new_mask;
8095 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
8097 goto out_free_new_mask;
8099 cpuset_cpus_allowed(p, cpus_allowed);
8100 if (!cpumask_subset(new_mask, cpus_allowed)) {
8102 * We must have raced with a concurrent cpuset update.
8103 * Just reset the cpumask to the cpuset's cpus_allowed.
8105 cpumask_copy(new_mask, cpus_allowed);
8110 free_cpumask_var(new_mask);
8111 out_free_cpus_allowed:
8112 free_cpumask_var(cpus_allowed);
8116 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8118 struct task_struct *p;
8123 p = find_process_by_pid(pid);
8129 /* Prevent p going away */
8133 if (p->flags & PF_NO_SETAFFINITY) {
8138 if (!check_same_owner(p)) {
8140 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8148 retval = security_task_setscheduler(p);
8152 retval = __sched_setaffinity(p, in_mask);
8158 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8159 struct cpumask *new_mask)
8161 if (len < cpumask_size())
8162 cpumask_clear(new_mask);
8163 else if (len > cpumask_size())
8164 len = cpumask_size();
8166 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8170 * sys_sched_setaffinity - set the CPU affinity of a process
8171 * @pid: pid of the process
8172 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8173 * @user_mask_ptr: user-space pointer to the new CPU mask
8175 * Return: 0 on success. An error code otherwise.
8177 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8178 unsigned long __user *, user_mask_ptr)
8180 cpumask_var_t new_mask;
8183 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8186 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8188 retval = sched_setaffinity(pid, new_mask);
8189 free_cpumask_var(new_mask);
8193 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8195 struct task_struct *p;
8196 unsigned long flags;
8202 p = find_process_by_pid(pid);
8206 retval = security_task_getscheduler(p);
8210 raw_spin_lock_irqsave(&p->pi_lock, flags);
8211 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8212 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8221 * sys_sched_getaffinity - get the CPU affinity of a process
8222 * @pid: pid of the process
8223 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8224 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8226 * Return: size of CPU mask copied to user_mask_ptr on success. An
8227 * error code otherwise.
8229 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8230 unsigned long __user *, user_mask_ptr)
8235 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8237 if (len & (sizeof(unsigned long)-1))
8240 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8243 ret = sched_getaffinity(pid, mask);
8245 unsigned int retlen = min(len, cpumask_size());
8247 if (copy_to_user(user_mask_ptr, mask, retlen))
8252 free_cpumask_var(mask);
8257 static void do_sched_yield(void)
8262 rq = this_rq_lock_irq(&rf);
8264 schedstat_inc(rq->yld_count);
8265 current->sched_class->yield_task(rq);
8268 rq_unlock_irq(rq, &rf);
8269 sched_preempt_enable_no_resched();
8275 * sys_sched_yield - yield the current processor to other threads.
8277 * This function yields the current CPU to other tasks. If there are no
8278 * other threads running on this CPU then this function will return.
8282 SYSCALL_DEFINE0(sched_yield)
8288 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8289 int __sched __cond_resched(void)
8291 if (should_resched(0)) {
8292 preempt_schedule_common();
8296 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8297 * whether the current CPU is in an RCU read-side critical section,
8298 * so the tick can report quiescent states even for CPUs looping
8299 * in kernel context. In contrast, in non-preemptible kernels,
8300 * RCU readers leave no in-memory hints, which means that CPU-bound
8301 * processes executing in kernel context might never report an
8302 * RCU quiescent state. Therefore, the following code causes
8303 * cond_resched() to report a quiescent state, but only when RCU
8304 * is in urgent need of one.
8306 #ifndef CONFIG_PREEMPT_RCU
8311 EXPORT_SYMBOL(__cond_resched);
8314 #ifdef CONFIG_PREEMPT_DYNAMIC
8315 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8316 #define cond_resched_dynamic_enabled __cond_resched
8317 #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8318 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8319 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8321 #define might_resched_dynamic_enabled __cond_resched
8322 #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8323 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8324 EXPORT_STATIC_CALL_TRAMP(might_resched);
8325 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8326 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8327 int __sched dynamic_cond_resched(void)
8329 if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8331 return __cond_resched();
8333 EXPORT_SYMBOL(dynamic_cond_resched);
8335 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8336 int __sched dynamic_might_resched(void)
8338 if (!static_branch_unlikely(&sk_dynamic_might_resched))
8340 return __cond_resched();
8342 EXPORT_SYMBOL(dynamic_might_resched);
8347 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8348 * call schedule, and on return reacquire the lock.
8350 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8351 * operations here to prevent schedule() from being called twice (once via
8352 * spin_unlock(), once by hand).
8354 int __cond_resched_lock(spinlock_t *lock)
8356 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8359 lockdep_assert_held(lock);
8361 if (spin_needbreak(lock) || resched) {
8363 if (!_cond_resched())
8370 EXPORT_SYMBOL(__cond_resched_lock);
8372 int __cond_resched_rwlock_read(rwlock_t *lock)
8374 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8377 lockdep_assert_held_read(lock);
8379 if (rwlock_needbreak(lock) || resched) {
8381 if (!_cond_resched())
8388 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8390 int __cond_resched_rwlock_write(rwlock_t *lock)
8392 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8395 lockdep_assert_held_write(lock);
8397 if (rwlock_needbreak(lock) || resched) {
8399 if (!_cond_resched())
8406 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8408 #ifdef CONFIG_PREEMPT_DYNAMIC
8410 #ifdef CONFIG_GENERIC_ENTRY
8411 #include <linux/entry-common.h>
8417 * SC:preempt_schedule
8418 * SC:preempt_schedule_notrace
8419 * SC:irqentry_exit_cond_resched
8423 * cond_resched <- __cond_resched
8424 * might_resched <- RET0
8425 * preempt_schedule <- NOP
8426 * preempt_schedule_notrace <- NOP
8427 * irqentry_exit_cond_resched <- NOP
8430 * cond_resched <- __cond_resched
8431 * might_resched <- __cond_resched
8432 * preempt_schedule <- NOP
8433 * preempt_schedule_notrace <- NOP
8434 * irqentry_exit_cond_resched <- NOP
8437 * cond_resched <- RET0
8438 * might_resched <- RET0
8439 * preempt_schedule <- preempt_schedule
8440 * preempt_schedule_notrace <- preempt_schedule_notrace
8441 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8445 preempt_dynamic_undefined = -1,
8446 preempt_dynamic_none,
8447 preempt_dynamic_voluntary,
8448 preempt_dynamic_full,
8451 int preempt_dynamic_mode = preempt_dynamic_undefined;
8453 int sched_dynamic_mode(const char *str)
8455 if (!strcmp(str, "none"))
8456 return preempt_dynamic_none;
8458 if (!strcmp(str, "voluntary"))
8459 return preempt_dynamic_voluntary;
8461 if (!strcmp(str, "full"))
8462 return preempt_dynamic_full;
8467 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8468 #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8469 #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8470 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8471 #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8472 #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8474 #error "Unsupported PREEMPT_DYNAMIC mechanism"
8477 void sched_dynamic_update(int mode)
8480 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8481 * the ZERO state, which is invalid.
8483 preempt_dynamic_enable(cond_resched);
8484 preempt_dynamic_enable(might_resched);
8485 preempt_dynamic_enable(preempt_schedule);
8486 preempt_dynamic_enable(preempt_schedule_notrace);
8487 preempt_dynamic_enable(irqentry_exit_cond_resched);
8490 case preempt_dynamic_none:
8491 preempt_dynamic_enable(cond_resched);
8492 preempt_dynamic_disable(might_resched);
8493 preempt_dynamic_disable(preempt_schedule);
8494 preempt_dynamic_disable(preempt_schedule_notrace);
8495 preempt_dynamic_disable(irqentry_exit_cond_resched);
8496 pr_info("Dynamic Preempt: none\n");
8499 case preempt_dynamic_voluntary:
8500 preempt_dynamic_enable(cond_resched);
8501 preempt_dynamic_enable(might_resched);
8502 preempt_dynamic_disable(preempt_schedule);
8503 preempt_dynamic_disable(preempt_schedule_notrace);
8504 preempt_dynamic_disable(irqentry_exit_cond_resched);
8505 pr_info("Dynamic Preempt: voluntary\n");
8508 case preempt_dynamic_full:
8509 preempt_dynamic_disable(cond_resched);
8510 preempt_dynamic_disable(might_resched);
8511 preempt_dynamic_enable(preempt_schedule);
8512 preempt_dynamic_enable(preempt_schedule_notrace);
8513 preempt_dynamic_enable(irqentry_exit_cond_resched);
8514 pr_info("Dynamic Preempt: full\n");
8518 preempt_dynamic_mode = mode;
8521 static int __init setup_preempt_mode(char *str)
8523 int mode = sched_dynamic_mode(str);
8525 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8529 sched_dynamic_update(mode);
8532 __setup("preempt=", setup_preempt_mode);
8534 static void __init preempt_dynamic_init(void)
8536 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8537 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8538 sched_dynamic_update(preempt_dynamic_none);
8539 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8540 sched_dynamic_update(preempt_dynamic_voluntary);
8542 /* Default static call setting, nothing to do */
8543 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8544 preempt_dynamic_mode = preempt_dynamic_full;
8545 pr_info("Dynamic Preempt: full\n");
8550 #define PREEMPT_MODEL_ACCESSOR(mode) \
8551 bool preempt_model_##mode(void) \
8553 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8554 return preempt_dynamic_mode == preempt_dynamic_##mode; \
8556 EXPORT_SYMBOL_GPL(preempt_model_##mode)
8558 PREEMPT_MODEL_ACCESSOR(none);
8559 PREEMPT_MODEL_ACCESSOR(voluntary);
8560 PREEMPT_MODEL_ACCESSOR(full);
8562 #else /* !CONFIG_PREEMPT_DYNAMIC */
8564 static inline void preempt_dynamic_init(void) { }
8566 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8569 * yield - yield the current processor to other threads.
8571 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8573 * The scheduler is at all times free to pick the calling task as the most
8574 * eligible task to run, if removing the yield() call from your code breaks
8575 * it, it's already broken.
8577 * Typical broken usage is:
8582 * where one assumes that yield() will let 'the other' process run that will
8583 * make event true. If the current task is a SCHED_FIFO task that will never
8584 * happen. Never use yield() as a progress guarantee!!
8586 * If you want to use yield() to wait for something, use wait_event().
8587 * If you want to use yield() to be 'nice' for others, use cond_resched().
8588 * If you still want to use yield(), do not!
8590 void __sched yield(void)
8592 set_current_state(TASK_RUNNING);
8595 EXPORT_SYMBOL(yield);
8598 * yield_to - yield the current processor to another thread in
8599 * your thread group, or accelerate that thread toward the
8600 * processor it's on.
8602 * @preempt: whether task preemption is allowed or not
8604 * It's the caller's job to ensure that the target task struct
8605 * can't go away on us before we can do any checks.
8608 * true (>0) if we indeed boosted the target task.
8609 * false (0) if we failed to boost the target.
8610 * -ESRCH if there's no task to yield to.
8612 int __sched yield_to(struct task_struct *p, bool preempt)
8614 struct task_struct *curr = current;
8615 struct rq *rq, *p_rq;
8616 unsigned long flags;
8619 local_irq_save(flags);
8625 * If we're the only runnable task on the rq and target rq also
8626 * has only one task, there's absolutely no point in yielding.
8628 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8633 double_rq_lock(rq, p_rq);
8634 if (task_rq(p) != p_rq) {
8635 double_rq_unlock(rq, p_rq);
8639 if (!curr->sched_class->yield_to_task)
8642 if (curr->sched_class != p->sched_class)
8645 if (task_on_cpu(p_rq, p) || !task_is_running(p))
8648 yielded = curr->sched_class->yield_to_task(rq, p);
8650 schedstat_inc(rq->yld_count);
8652 * Make p's CPU reschedule; pick_next_entity takes care of
8655 if (preempt && rq != p_rq)
8660 double_rq_unlock(rq, p_rq);
8662 local_irq_restore(flags);
8669 EXPORT_SYMBOL_GPL(yield_to);
8671 int io_schedule_prepare(void)
8673 int old_iowait = current->in_iowait;
8675 current->in_iowait = 1;
8676 blk_flush_plug(current->plug, true);
8680 void io_schedule_finish(int token)
8682 current->in_iowait = token;
8686 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8687 * that process accounting knows that this is a task in IO wait state.
8689 long __sched io_schedule_timeout(long timeout)
8694 token = io_schedule_prepare();
8695 ret = schedule_timeout(timeout);
8696 io_schedule_finish(token);
8700 EXPORT_SYMBOL(io_schedule_timeout);
8702 void __sched io_schedule(void)
8706 token = io_schedule_prepare();
8708 io_schedule_finish(token);
8710 EXPORT_SYMBOL(io_schedule);
8713 * sys_sched_get_priority_max - return maximum RT priority.
8714 * @policy: scheduling class.
8716 * Return: On success, this syscall returns the maximum
8717 * rt_priority that can be used by a given scheduling class.
8718 * On failure, a negative error code is returned.
8720 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8727 ret = MAX_RT_PRIO-1;
8729 case SCHED_DEADLINE:
8740 * sys_sched_get_priority_min - return minimum RT priority.
8741 * @policy: scheduling class.
8743 * Return: On success, this syscall returns the minimum
8744 * rt_priority that can be used by a given scheduling class.
8745 * On failure, a negative error code is returned.
8747 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8756 case SCHED_DEADLINE:
8765 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8767 struct task_struct *p;
8768 unsigned int time_slice;
8778 p = find_process_by_pid(pid);
8782 retval = security_task_getscheduler(p);
8786 rq = task_rq_lock(p, &rf);
8788 if (p->sched_class->get_rr_interval)
8789 time_slice = p->sched_class->get_rr_interval(rq, p);
8790 task_rq_unlock(rq, p, &rf);
8793 jiffies_to_timespec64(time_slice, t);
8802 * sys_sched_rr_get_interval - return the default timeslice of a process.
8803 * @pid: pid of the process.
8804 * @interval: userspace pointer to the timeslice value.
8806 * this syscall writes the default timeslice value of a given process
8807 * into the user-space timespec buffer. A value of '0' means infinity.
8809 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8812 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8813 struct __kernel_timespec __user *, interval)
8815 struct timespec64 t;
8816 int retval = sched_rr_get_interval(pid, &t);
8819 retval = put_timespec64(&t, interval);
8824 #ifdef CONFIG_COMPAT_32BIT_TIME
8825 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8826 struct old_timespec32 __user *, interval)
8828 struct timespec64 t;
8829 int retval = sched_rr_get_interval(pid, &t);
8832 retval = put_old_timespec32(&t, interval);
8837 void sched_show_task(struct task_struct *p)
8839 unsigned long free = 0;
8842 if (!try_get_task_stack(p))
8845 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8847 if (task_is_running(p))
8848 pr_cont(" running task ");
8849 #ifdef CONFIG_DEBUG_STACK_USAGE
8850 free = stack_not_used(p);
8855 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8857 pr_cont(" stack:%-5lu pid:%-5d ppid:%-6d flags:0x%08lx\n",
8858 free, task_pid_nr(p), ppid,
8859 read_task_thread_flags(p));
8861 print_worker_info(KERN_INFO, p);
8862 print_stop_info(KERN_INFO, p);
8863 show_stack(p, NULL, KERN_INFO);
8866 EXPORT_SYMBOL_GPL(sched_show_task);
8869 state_filter_match(unsigned long state_filter, struct task_struct *p)
8871 unsigned int state = READ_ONCE(p->__state);
8873 /* no filter, everything matches */
8877 /* filter, but doesn't match */
8878 if (!(state & state_filter))
8882 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8885 if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
8892 void show_state_filter(unsigned int state_filter)
8894 struct task_struct *g, *p;
8897 for_each_process_thread(g, p) {
8899 * reset the NMI-timeout, listing all files on a slow
8900 * console might take a lot of time:
8901 * Also, reset softlockup watchdogs on all CPUs, because
8902 * another CPU might be blocked waiting for us to process
8905 touch_nmi_watchdog();
8906 touch_all_softlockup_watchdogs();
8907 if (state_filter_match(state_filter, p))
8911 #ifdef CONFIG_SCHED_DEBUG
8913 sysrq_sched_debug_show();
8917 * Only show locks if all tasks are dumped:
8920 debug_show_all_locks();
8924 * init_idle - set up an idle thread for a given CPU
8925 * @idle: task in question
8926 * @cpu: CPU the idle task belongs to
8928 * NOTE: this function does not set the idle thread's NEED_RESCHED
8929 * flag, to make booting more robust.
8931 void __init init_idle(struct task_struct *idle, int cpu)
8933 struct rq *rq = cpu_rq(cpu);
8934 unsigned long flags;
8936 __sched_fork(0, idle);
8938 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8939 raw_spin_rq_lock(rq);
8941 idle->__state = TASK_RUNNING;
8942 idle->se.exec_start = sched_clock();
8944 * PF_KTHREAD should already be set at this point; regardless, make it
8945 * look like a proper per-CPU kthread.
8947 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8948 kthread_set_per_cpu(idle, cpu);
8952 * It's possible that init_idle() gets called multiple times on a task,
8953 * in that case do_set_cpus_allowed() will not do the right thing.
8955 * And since this is boot we can forgo the serialization.
8957 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8960 * We're having a chicken and egg problem, even though we are
8961 * holding rq->lock, the CPU isn't yet set to this CPU so the
8962 * lockdep check in task_group() will fail.
8964 * Similar case to sched_fork(). / Alternatively we could
8965 * use task_rq_lock() here and obtain the other rq->lock.
8970 __set_task_cpu(idle, cpu);
8974 rcu_assign_pointer(rq->curr, idle);
8975 idle->on_rq = TASK_ON_RQ_QUEUED;
8979 raw_spin_rq_unlock(rq);
8980 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8982 /* Set the preempt count _outside_ the spinlocks! */
8983 init_idle_preempt_count(idle, cpu);
8986 * The idle tasks have their own, simple scheduling class:
8988 idle->sched_class = &idle_sched_class;
8989 ftrace_graph_init_idle_task(idle, cpu);
8990 vtime_init_idle(idle, cpu);
8992 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8998 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8999 const struct cpumask *trial)
9003 if (cpumask_empty(cur))
9006 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9011 int task_can_attach(struct task_struct *p,
9012 const struct cpumask *cs_effective_cpus)
9017 * Kthreads which disallow setaffinity shouldn't be moved
9018 * to a new cpuset; we don't want to change their CPU
9019 * affinity and isolating such threads by their set of
9020 * allowed nodes is unnecessary. Thus, cpusets are not
9021 * applicable for such threads. This prevents checking for
9022 * success of set_cpus_allowed_ptr() on all attached tasks
9023 * before cpus_mask may be changed.
9025 if (p->flags & PF_NO_SETAFFINITY) {
9030 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
9031 cs_effective_cpus)) {
9032 int cpu = cpumask_any_and(cpu_active_mask, cs_effective_cpus);
9034 if (unlikely(cpu >= nr_cpu_ids))
9036 ret = dl_cpu_busy(cpu, p);
9043 bool sched_smp_initialized __read_mostly;
9045 #ifdef CONFIG_NUMA_BALANCING
9046 /* Migrate current task p to target_cpu */
9047 int migrate_task_to(struct task_struct *p, int target_cpu)
9049 struct migration_arg arg = { p, target_cpu };
9050 int curr_cpu = task_cpu(p);
9052 if (curr_cpu == target_cpu)
9055 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9058 /* TODO: This is not properly updating schedstats */
9060 trace_sched_move_numa(p, curr_cpu, target_cpu);
9061 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9065 * Requeue a task on a given node and accurately track the number of NUMA
9066 * tasks on the runqueues
9068 void sched_setnuma(struct task_struct *p, int nid)
9070 bool queued, running;
9074 rq = task_rq_lock(p, &rf);
9075 queued = task_on_rq_queued(p);
9076 running = task_current(rq, p);
9079 dequeue_task(rq, p, DEQUEUE_SAVE);
9081 put_prev_task(rq, p);
9083 p->numa_preferred_nid = nid;
9086 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9088 set_next_task(rq, p);
9089 task_rq_unlock(rq, p, &rf);
9091 #endif /* CONFIG_NUMA_BALANCING */
9093 #ifdef CONFIG_HOTPLUG_CPU
9095 * Ensure that the idle task is using init_mm right before its CPU goes
9098 void idle_task_exit(void)
9100 struct mm_struct *mm = current->active_mm;
9102 BUG_ON(cpu_online(smp_processor_id()));
9103 BUG_ON(current != this_rq()->idle);
9105 if (mm != &init_mm) {
9106 switch_mm(mm, &init_mm, current);
9107 finish_arch_post_lock_switch();
9110 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9113 static int __balance_push_cpu_stop(void *arg)
9115 struct task_struct *p = arg;
9116 struct rq *rq = this_rq();
9120 raw_spin_lock_irq(&p->pi_lock);
9123 update_rq_clock(rq);
9125 if (task_rq(p) == rq && task_on_rq_queued(p)) {
9126 cpu = select_fallback_rq(rq->cpu, p);
9127 rq = __migrate_task(rq, &rf, p, cpu);
9131 raw_spin_unlock_irq(&p->pi_lock);
9138 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9141 * Ensure we only run per-cpu kthreads once the CPU goes !active.
9143 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9144 * effective when the hotplug motion is down.
9146 static void balance_push(struct rq *rq)
9148 struct task_struct *push_task = rq->curr;
9150 lockdep_assert_rq_held(rq);
9153 * Ensure the thing is persistent until balance_push_set(.on = false);
9155 rq->balance_callback = &balance_push_callback;
9158 * Only active while going offline and when invoked on the outgoing
9161 if (!cpu_dying(rq->cpu) || rq != this_rq())
9165 * Both the cpu-hotplug and stop task are in this case and are
9166 * required to complete the hotplug process.
9168 if (kthread_is_per_cpu(push_task) ||
9169 is_migration_disabled(push_task)) {
9172 * If this is the idle task on the outgoing CPU try to wake
9173 * up the hotplug control thread which might wait for the
9174 * last task to vanish. The rcuwait_active() check is
9175 * accurate here because the waiter is pinned on this CPU
9176 * and can't obviously be running in parallel.
9178 * On RT kernels this also has to check whether there are
9179 * pinned and scheduled out tasks on the runqueue. They
9180 * need to leave the migrate disabled section first.
9182 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9183 rcuwait_active(&rq->hotplug_wait)) {
9184 raw_spin_rq_unlock(rq);
9185 rcuwait_wake_up(&rq->hotplug_wait);
9186 raw_spin_rq_lock(rq);
9191 get_task_struct(push_task);
9193 * Temporarily drop rq->lock such that we can wake-up the stop task.
9194 * Both preemption and IRQs are still disabled.
9196 raw_spin_rq_unlock(rq);
9197 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9198 this_cpu_ptr(&push_work));
9200 * At this point need_resched() is true and we'll take the loop in
9201 * schedule(). The next pick is obviously going to be the stop task
9202 * which kthread_is_per_cpu() and will push this task away.
9204 raw_spin_rq_lock(rq);
9207 static void balance_push_set(int cpu, bool on)
9209 struct rq *rq = cpu_rq(cpu);
9212 rq_lock_irqsave(rq, &rf);
9214 WARN_ON_ONCE(rq->balance_callback);
9215 rq->balance_callback = &balance_push_callback;
9216 } else if (rq->balance_callback == &balance_push_callback) {
9217 rq->balance_callback = NULL;
9219 rq_unlock_irqrestore(rq, &rf);
9223 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9224 * inactive. All tasks which are not per CPU kernel threads are either
9225 * pushed off this CPU now via balance_push() or placed on a different CPU
9226 * during wakeup. Wait until the CPU is quiescent.
9228 static void balance_hotplug_wait(void)
9230 struct rq *rq = this_rq();
9232 rcuwait_wait_event(&rq->hotplug_wait,
9233 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9234 TASK_UNINTERRUPTIBLE);
9239 static inline void balance_push(struct rq *rq)
9243 static inline void balance_push_set(int cpu, bool on)
9247 static inline void balance_hotplug_wait(void)
9251 #endif /* CONFIG_HOTPLUG_CPU */
9253 void set_rq_online(struct rq *rq)
9256 const struct sched_class *class;
9258 cpumask_set_cpu(rq->cpu, rq->rd->online);
9261 for_each_class(class) {
9262 if (class->rq_online)
9263 class->rq_online(rq);
9268 void set_rq_offline(struct rq *rq)
9271 const struct sched_class *class;
9273 for_each_class(class) {
9274 if (class->rq_offline)
9275 class->rq_offline(rq);
9278 cpumask_clear_cpu(rq->cpu, rq->rd->online);
9284 * used to mark begin/end of suspend/resume:
9286 static int num_cpus_frozen;
9289 * Update cpusets according to cpu_active mask. If cpusets are
9290 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9291 * around partition_sched_domains().
9293 * If we come here as part of a suspend/resume, don't touch cpusets because we
9294 * want to restore it back to its original state upon resume anyway.
9296 static void cpuset_cpu_active(void)
9298 if (cpuhp_tasks_frozen) {
9300 * num_cpus_frozen tracks how many CPUs are involved in suspend
9301 * resume sequence. As long as this is not the last online
9302 * operation in the resume sequence, just build a single sched
9303 * domain, ignoring cpusets.
9305 partition_sched_domains(1, NULL, NULL);
9306 if (--num_cpus_frozen)
9309 * This is the last CPU online operation. So fall through and
9310 * restore the original sched domains by considering the
9311 * cpuset configurations.
9313 cpuset_force_rebuild();
9315 cpuset_update_active_cpus();
9318 static int cpuset_cpu_inactive(unsigned int cpu)
9320 if (!cpuhp_tasks_frozen) {
9321 int ret = dl_cpu_busy(cpu, NULL);
9325 cpuset_update_active_cpus();
9328 partition_sched_domains(1, NULL, NULL);
9333 int sched_cpu_activate(unsigned int cpu)
9335 struct rq *rq = cpu_rq(cpu);
9339 * Clear the balance_push callback and prepare to schedule
9342 balance_push_set(cpu, false);
9344 #ifdef CONFIG_SCHED_SMT
9346 * When going up, increment the number of cores with SMT present.
9348 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9349 static_branch_inc_cpuslocked(&sched_smt_present);
9351 set_cpu_active(cpu, true);
9353 if (sched_smp_initialized) {
9354 sched_update_numa(cpu, true);
9355 sched_domains_numa_masks_set(cpu);
9356 cpuset_cpu_active();
9360 * Put the rq online, if not already. This happens:
9362 * 1) In the early boot process, because we build the real domains
9363 * after all CPUs have been brought up.
9365 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9368 rq_lock_irqsave(rq, &rf);
9370 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9373 rq_unlock_irqrestore(rq, &rf);
9378 int sched_cpu_deactivate(unsigned int cpu)
9380 struct rq *rq = cpu_rq(cpu);
9385 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9386 * load balancing when not active
9388 nohz_balance_exit_idle(rq);
9390 set_cpu_active(cpu, false);
9393 * From this point forward, this CPU will refuse to run any task that
9394 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9395 * push those tasks away until this gets cleared, see
9396 * sched_cpu_dying().
9398 balance_push_set(cpu, true);
9401 * We've cleared cpu_active_mask / set balance_push, wait for all
9402 * preempt-disabled and RCU users of this state to go away such that
9403 * all new such users will observe it.
9405 * Specifically, we rely on ttwu to no longer target this CPU, see
9406 * ttwu_queue_cond() and is_cpu_allowed().
9408 * Do sync before park smpboot threads to take care the rcu boost case.
9412 rq_lock_irqsave(rq, &rf);
9414 update_rq_clock(rq);
9415 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9418 rq_unlock_irqrestore(rq, &rf);
9420 #ifdef CONFIG_SCHED_SMT
9422 * When going down, decrement the number of cores with SMT present.
9424 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9425 static_branch_dec_cpuslocked(&sched_smt_present);
9427 sched_core_cpu_deactivate(cpu);
9430 if (!sched_smp_initialized)
9433 sched_update_numa(cpu, false);
9434 ret = cpuset_cpu_inactive(cpu);
9436 balance_push_set(cpu, false);
9437 set_cpu_active(cpu, true);
9438 sched_update_numa(cpu, true);
9441 sched_domains_numa_masks_clear(cpu);
9445 static void sched_rq_cpu_starting(unsigned int cpu)
9447 struct rq *rq = cpu_rq(cpu);
9449 rq->calc_load_update = calc_load_update;
9450 update_max_interval();
9453 int sched_cpu_starting(unsigned int cpu)
9455 sched_core_cpu_starting(cpu);
9456 sched_rq_cpu_starting(cpu);
9457 sched_tick_start(cpu);
9461 #ifdef CONFIG_HOTPLUG_CPU
9464 * Invoked immediately before the stopper thread is invoked to bring the
9465 * CPU down completely. At this point all per CPU kthreads except the
9466 * hotplug thread (current) and the stopper thread (inactive) have been
9467 * either parked or have been unbound from the outgoing CPU. Ensure that
9468 * any of those which might be on the way out are gone.
9470 * If after this point a bound task is being woken on this CPU then the
9471 * responsible hotplug callback has failed to do it's job.
9472 * sched_cpu_dying() will catch it with the appropriate fireworks.
9474 int sched_cpu_wait_empty(unsigned int cpu)
9476 balance_hotplug_wait();
9481 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9482 * might have. Called from the CPU stopper task after ensuring that the
9483 * stopper is the last running task on the CPU, so nr_active count is
9484 * stable. We need to take the teardown thread which is calling this into
9485 * account, so we hand in adjust = 1 to the load calculation.
9487 * Also see the comment "Global load-average calculations".
9489 static void calc_load_migrate(struct rq *rq)
9491 long delta = calc_load_fold_active(rq, 1);
9494 atomic_long_add(delta, &calc_load_tasks);
9497 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9499 struct task_struct *g, *p;
9500 int cpu = cpu_of(rq);
9502 lockdep_assert_rq_held(rq);
9504 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9505 for_each_process_thread(g, p) {
9506 if (task_cpu(p) != cpu)
9509 if (!task_on_rq_queued(p))
9512 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9516 int sched_cpu_dying(unsigned int cpu)
9518 struct rq *rq = cpu_rq(cpu);
9521 /* Handle pending wakeups and then migrate everything off */
9522 sched_tick_stop(cpu);
9524 rq_lock_irqsave(rq, &rf);
9525 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9526 WARN(true, "Dying CPU not properly vacated!");
9527 dump_rq_tasks(rq, KERN_WARNING);
9529 rq_unlock_irqrestore(rq, &rf);
9531 calc_load_migrate(rq);
9532 update_max_interval();
9534 sched_core_cpu_dying(cpu);
9539 void __init sched_init_smp(void)
9541 sched_init_numa(NUMA_NO_NODE);
9544 * There's no userspace yet to cause hotplug operations; hence all the
9545 * CPU masks are stable and all blatant races in the below code cannot
9548 mutex_lock(&sched_domains_mutex);
9549 sched_init_domains(cpu_active_mask);
9550 mutex_unlock(&sched_domains_mutex);
9552 /* Move init over to a non-isolated CPU */
9553 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9555 current->flags &= ~PF_NO_SETAFFINITY;
9556 sched_init_granularity();
9558 init_sched_rt_class();
9559 init_sched_dl_class();
9561 sched_smp_initialized = true;
9564 static int __init migration_init(void)
9566 sched_cpu_starting(smp_processor_id());
9569 early_initcall(migration_init);
9572 void __init sched_init_smp(void)
9574 sched_init_granularity();
9576 #endif /* CONFIG_SMP */
9578 int in_sched_functions(unsigned long addr)
9580 return in_lock_functions(addr) ||
9581 (addr >= (unsigned long)__sched_text_start
9582 && addr < (unsigned long)__sched_text_end);
9585 #ifdef CONFIG_CGROUP_SCHED
9587 * Default task group.
9588 * Every task in system belongs to this group at bootup.
9590 struct task_group root_task_group;
9591 LIST_HEAD(task_groups);
9593 /* Cacheline aligned slab cache for task_group */
9594 static struct kmem_cache *task_group_cache __read_mostly;
9597 void __init sched_init(void)
9599 unsigned long ptr = 0;
9602 /* Make sure the linker didn't screw up */
9603 BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9604 &fair_sched_class != &rt_sched_class + 1 ||
9605 &rt_sched_class != &dl_sched_class + 1);
9607 BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9612 #ifdef CONFIG_FAIR_GROUP_SCHED
9613 ptr += 2 * nr_cpu_ids * sizeof(void **);
9615 #ifdef CONFIG_RT_GROUP_SCHED
9616 ptr += 2 * nr_cpu_ids * sizeof(void **);
9619 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9621 #ifdef CONFIG_FAIR_GROUP_SCHED
9622 root_task_group.se = (struct sched_entity **)ptr;
9623 ptr += nr_cpu_ids * sizeof(void **);
9625 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9626 ptr += nr_cpu_ids * sizeof(void **);
9628 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9629 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9630 #endif /* CONFIG_FAIR_GROUP_SCHED */
9631 #ifdef CONFIG_RT_GROUP_SCHED
9632 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9633 ptr += nr_cpu_ids * sizeof(void **);
9635 root_task_group.rt_rq = (struct rt_rq **)ptr;
9636 ptr += nr_cpu_ids * sizeof(void **);
9638 #endif /* CONFIG_RT_GROUP_SCHED */
9641 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9644 init_defrootdomain();
9647 #ifdef CONFIG_RT_GROUP_SCHED
9648 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9649 global_rt_period(), global_rt_runtime());
9650 #endif /* CONFIG_RT_GROUP_SCHED */
9652 #ifdef CONFIG_CGROUP_SCHED
9653 task_group_cache = KMEM_CACHE(task_group, 0);
9655 list_add(&root_task_group.list, &task_groups);
9656 INIT_LIST_HEAD(&root_task_group.children);
9657 INIT_LIST_HEAD(&root_task_group.siblings);
9658 autogroup_init(&init_task);
9659 #endif /* CONFIG_CGROUP_SCHED */
9661 for_each_possible_cpu(i) {
9665 raw_spin_lock_init(&rq->__lock);
9667 rq->calc_load_active = 0;
9668 rq->calc_load_update = jiffies + LOAD_FREQ;
9669 init_cfs_rq(&rq->cfs);
9670 init_rt_rq(&rq->rt);
9671 init_dl_rq(&rq->dl);
9672 #ifdef CONFIG_FAIR_GROUP_SCHED
9673 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9674 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9676 * How much CPU bandwidth does root_task_group get?
9678 * In case of task-groups formed thr' the cgroup filesystem, it
9679 * gets 100% of the CPU resources in the system. This overall
9680 * system CPU resource is divided among the tasks of
9681 * root_task_group and its child task-groups in a fair manner,
9682 * based on each entity's (task or task-group's) weight
9683 * (se->load.weight).
9685 * In other words, if root_task_group has 10 tasks of weight
9686 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9687 * then A0's share of the CPU resource is:
9689 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9691 * We achieve this by letting root_task_group's tasks sit
9692 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9694 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9695 #endif /* CONFIG_FAIR_GROUP_SCHED */
9697 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9698 #ifdef CONFIG_RT_GROUP_SCHED
9699 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9704 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9705 rq->balance_callback = &balance_push_callback;
9706 rq->active_balance = 0;
9707 rq->next_balance = jiffies;
9712 rq->avg_idle = 2*sysctl_sched_migration_cost;
9713 rq->wake_stamp = jiffies;
9714 rq->wake_avg_idle = rq->avg_idle;
9715 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9717 INIT_LIST_HEAD(&rq->cfs_tasks);
9719 rq_attach_root(rq, &def_root_domain);
9720 #ifdef CONFIG_NO_HZ_COMMON
9721 rq->last_blocked_load_update_tick = jiffies;
9722 atomic_set(&rq->nohz_flags, 0);
9724 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9726 #ifdef CONFIG_HOTPLUG_CPU
9727 rcuwait_init(&rq->hotplug_wait);
9729 #endif /* CONFIG_SMP */
9731 atomic_set(&rq->nr_iowait, 0);
9733 #ifdef CONFIG_SCHED_CORE
9735 rq->core_pick = NULL;
9736 rq->core_enabled = 0;
9737 rq->core_tree = RB_ROOT;
9738 rq->core_forceidle_count = 0;
9739 rq->core_forceidle_occupation = 0;
9740 rq->core_forceidle_start = 0;
9742 rq->core_cookie = 0UL;
9746 set_load_weight(&init_task, false);
9749 * The boot idle thread does lazy MMU switching as well:
9752 enter_lazy_tlb(&init_mm, current);
9755 * The idle task doesn't need the kthread struct to function, but it
9756 * is dressed up as a per-CPU kthread and thus needs to play the part
9757 * if we want to avoid special-casing it in code that deals with per-CPU
9760 WARN_ON(!set_kthread_struct(current));
9763 * Make us the idle thread. Technically, schedule() should not be
9764 * called from this thread, however somewhere below it might be,
9765 * but because we are the idle thread, we just pick up running again
9766 * when this runqueue becomes "idle".
9768 init_idle(current, smp_processor_id());
9770 calc_load_update = jiffies + LOAD_FREQ;
9773 idle_thread_set_boot_cpu();
9774 balance_push_set(smp_processor_id(), false);
9776 init_sched_fair_class();
9782 preempt_dynamic_init();
9784 scheduler_running = 1;
9787 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9789 void __might_sleep(const char *file, int line)
9791 unsigned int state = get_current_state();
9793 * Blocking primitives will set (and therefore destroy) current->state,
9794 * since we will exit with TASK_RUNNING make sure we enter with it,
9795 * otherwise we will destroy state.
9797 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9798 "do not call blocking ops when !TASK_RUNNING; "
9799 "state=%x set at [<%p>] %pS\n", state,
9800 (void *)current->task_state_change,
9801 (void *)current->task_state_change);
9803 __might_resched(file, line, 0);
9805 EXPORT_SYMBOL(__might_sleep);
9807 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9809 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9812 if (preempt_count() == preempt_offset)
9815 pr_err("Preemption disabled at:");
9816 print_ip_sym(KERN_ERR, ip);
9819 static inline bool resched_offsets_ok(unsigned int offsets)
9821 unsigned int nested = preempt_count();
9823 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9825 return nested == offsets;
9828 void __might_resched(const char *file, int line, unsigned int offsets)
9830 /* Ratelimiting timestamp: */
9831 static unsigned long prev_jiffy;
9833 unsigned long preempt_disable_ip;
9835 /* WARN_ON_ONCE() by default, no rate limit required: */
9838 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9839 !is_idle_task(current) && !current->non_block_count) ||
9840 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9844 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9846 prev_jiffy = jiffies;
9848 /* Save this before calling printk(), since that will clobber it: */
9849 preempt_disable_ip = get_preempt_disable_ip(current);
9851 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9853 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9854 in_atomic(), irqs_disabled(), current->non_block_count,
9855 current->pid, current->comm);
9856 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9857 offsets & MIGHT_RESCHED_PREEMPT_MASK);
9859 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9860 pr_err("RCU nest depth: %d, expected: %u\n",
9861 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9864 if (task_stack_end_corrupted(current))
9865 pr_emerg("Thread overran stack, or stack corrupted\n");
9867 debug_show_held_locks(current);
9868 if (irqs_disabled())
9869 print_irqtrace_events(current);
9871 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9872 preempt_disable_ip);
9875 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9877 EXPORT_SYMBOL(__might_resched);
9879 void __cant_sleep(const char *file, int line, int preempt_offset)
9881 static unsigned long prev_jiffy;
9883 if (irqs_disabled())
9886 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9889 if (preempt_count() > preempt_offset)
9892 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9894 prev_jiffy = jiffies;
9896 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9897 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9898 in_atomic(), irqs_disabled(),
9899 current->pid, current->comm);
9901 debug_show_held_locks(current);
9903 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9905 EXPORT_SYMBOL_GPL(__cant_sleep);
9908 void __cant_migrate(const char *file, int line)
9910 static unsigned long prev_jiffy;
9912 if (irqs_disabled())
9915 if (is_migration_disabled(current))
9918 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9921 if (preempt_count() > 0)
9924 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9926 prev_jiffy = jiffies;
9928 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9929 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9930 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9931 current->pid, current->comm);
9933 debug_show_held_locks(current);
9935 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9937 EXPORT_SYMBOL_GPL(__cant_migrate);
9941 #ifdef CONFIG_MAGIC_SYSRQ
9942 void normalize_rt_tasks(void)
9944 struct task_struct *g, *p;
9945 struct sched_attr attr = {
9946 .sched_policy = SCHED_NORMAL,
9949 read_lock(&tasklist_lock);
9950 for_each_process_thread(g, p) {
9952 * Only normalize user tasks:
9954 if (p->flags & PF_KTHREAD)
9957 p->se.exec_start = 0;
9958 schedstat_set(p->stats.wait_start, 0);
9959 schedstat_set(p->stats.sleep_start, 0);
9960 schedstat_set(p->stats.block_start, 0);
9962 if (!dl_task(p) && !rt_task(p)) {
9964 * Renice negative nice level userspace
9967 if (task_nice(p) < 0)
9968 set_user_nice(p, 0);
9972 __sched_setscheduler(p, &attr, false, false);
9974 read_unlock(&tasklist_lock);
9977 #endif /* CONFIG_MAGIC_SYSRQ */
9979 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9981 * These functions are only useful for the IA64 MCA handling, or kdb.
9983 * They can only be called when the whole system has been
9984 * stopped - every CPU needs to be quiescent, and no scheduling
9985 * activity can take place. Using them for anything else would
9986 * be a serious bug, and as a result, they aren't even visible
9987 * under any other configuration.
9991 * curr_task - return the current task for a given CPU.
9992 * @cpu: the processor in question.
9994 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9996 * Return: The current task for @cpu.
9998 struct task_struct *curr_task(int cpu)
10000 return cpu_curr(cpu);
10003 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
10007 * ia64_set_curr_task - set the current task for a given CPU.
10008 * @cpu: the processor in question.
10009 * @p: the task pointer to set.
10011 * Description: This function must only be used when non-maskable interrupts
10012 * are serviced on a separate stack. It allows the architecture to switch the
10013 * notion of the current task on a CPU in a non-blocking manner. This function
10014 * must be called with all CPU's synchronized, and interrupts disabled, the
10015 * and caller must save the original value of the current task (see
10016 * curr_task() above) and restore that value before reenabling interrupts and
10017 * re-starting the system.
10019 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10021 void ia64_set_curr_task(int cpu, struct task_struct *p)
10028 #ifdef CONFIG_CGROUP_SCHED
10029 /* task_group_lock serializes the addition/removal of task groups */
10030 static DEFINE_SPINLOCK(task_group_lock);
10032 static inline void alloc_uclamp_sched_group(struct task_group *tg,
10033 struct task_group *parent)
10035 #ifdef CONFIG_UCLAMP_TASK_GROUP
10036 enum uclamp_id clamp_id;
10038 for_each_clamp_id(clamp_id) {
10039 uclamp_se_set(&tg->uclamp_req[clamp_id],
10040 uclamp_none(clamp_id), false);
10041 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10046 static void sched_free_group(struct task_group *tg)
10048 free_fair_sched_group(tg);
10049 free_rt_sched_group(tg);
10050 autogroup_free(tg);
10051 kmem_cache_free(task_group_cache, tg);
10054 static void sched_free_group_rcu(struct rcu_head *rcu)
10056 sched_free_group(container_of(rcu, struct task_group, rcu));
10059 static void sched_unregister_group(struct task_group *tg)
10061 unregister_fair_sched_group(tg);
10062 unregister_rt_sched_group(tg);
10064 * We have to wait for yet another RCU grace period to expire, as
10065 * print_cfs_stats() might run concurrently.
10067 call_rcu(&tg->rcu, sched_free_group_rcu);
10070 /* allocate runqueue etc for a new task group */
10071 struct task_group *sched_create_group(struct task_group *parent)
10073 struct task_group *tg;
10075 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10077 return ERR_PTR(-ENOMEM);
10079 if (!alloc_fair_sched_group(tg, parent))
10082 if (!alloc_rt_sched_group(tg, parent))
10085 alloc_uclamp_sched_group(tg, parent);
10090 sched_free_group(tg);
10091 return ERR_PTR(-ENOMEM);
10094 void sched_online_group(struct task_group *tg, struct task_group *parent)
10096 unsigned long flags;
10098 spin_lock_irqsave(&task_group_lock, flags);
10099 list_add_rcu(&tg->list, &task_groups);
10101 /* Root should already exist: */
10104 tg->parent = parent;
10105 INIT_LIST_HEAD(&tg->children);
10106 list_add_rcu(&tg->siblings, &parent->children);
10107 spin_unlock_irqrestore(&task_group_lock, flags);
10109 online_fair_sched_group(tg);
10112 /* rcu callback to free various structures associated with a task group */
10113 static void sched_unregister_group_rcu(struct rcu_head *rhp)
10115 /* Now it should be safe to free those cfs_rqs: */
10116 sched_unregister_group(container_of(rhp, struct task_group, rcu));
10119 void sched_destroy_group(struct task_group *tg)
10121 /* Wait for possible concurrent references to cfs_rqs complete: */
10122 call_rcu(&tg->rcu, sched_unregister_group_rcu);
10125 void sched_release_group(struct task_group *tg)
10127 unsigned long flags;
10130 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10131 * sched_cfs_period_timer()).
10133 * For this to be effective, we have to wait for all pending users of
10134 * this task group to leave their RCU critical section to ensure no new
10135 * user will see our dying task group any more. Specifically ensure
10136 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10138 * We therefore defer calling unregister_fair_sched_group() to
10139 * sched_unregister_group() which is guarantied to get called only after the
10140 * current RCU grace period has expired.
10142 spin_lock_irqsave(&task_group_lock, flags);
10143 list_del_rcu(&tg->list);
10144 list_del_rcu(&tg->siblings);
10145 spin_unlock_irqrestore(&task_group_lock, flags);
10148 static void sched_change_group(struct task_struct *tsk)
10150 struct task_group *tg;
10153 * All callers are synchronized by task_rq_lock(); we do not use RCU
10154 * which is pointless here. Thus, we pass "true" to task_css_check()
10155 * to prevent lockdep warnings.
10157 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10158 struct task_group, css);
10159 tg = autogroup_task_group(tsk, tg);
10160 tsk->sched_task_group = tg;
10162 #ifdef CONFIG_FAIR_GROUP_SCHED
10163 if (tsk->sched_class->task_change_group)
10164 tsk->sched_class->task_change_group(tsk);
10167 set_task_rq(tsk, task_cpu(tsk));
10171 * Change task's runqueue when it moves between groups.
10173 * The caller of this function should have put the task in its new group by
10174 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10177 void sched_move_task(struct task_struct *tsk)
10179 int queued, running, queue_flags =
10180 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10181 struct rq_flags rf;
10184 rq = task_rq_lock(tsk, &rf);
10185 update_rq_clock(rq);
10187 running = task_current(rq, tsk);
10188 queued = task_on_rq_queued(tsk);
10191 dequeue_task(rq, tsk, queue_flags);
10193 put_prev_task(rq, tsk);
10195 sched_change_group(tsk);
10198 enqueue_task(rq, tsk, queue_flags);
10200 set_next_task(rq, tsk);
10202 * After changing group, the running task may have joined a
10203 * throttled one but it's still the running task. Trigger a
10204 * resched to make sure that task can still run.
10209 task_rq_unlock(rq, tsk, &rf);
10212 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10214 return css ? container_of(css, struct task_group, css) : NULL;
10217 static struct cgroup_subsys_state *
10218 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10220 struct task_group *parent = css_tg(parent_css);
10221 struct task_group *tg;
10224 /* This is early initialization for the top cgroup */
10225 return &root_task_group.css;
10228 tg = sched_create_group(parent);
10230 return ERR_PTR(-ENOMEM);
10235 /* Expose task group only after completing cgroup initialization */
10236 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10238 struct task_group *tg = css_tg(css);
10239 struct task_group *parent = css_tg(css->parent);
10242 sched_online_group(tg, parent);
10244 #ifdef CONFIG_UCLAMP_TASK_GROUP
10245 /* Propagate the effective uclamp value for the new group */
10246 mutex_lock(&uclamp_mutex);
10248 cpu_util_update_eff(css);
10250 mutex_unlock(&uclamp_mutex);
10256 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10258 struct task_group *tg = css_tg(css);
10260 sched_release_group(tg);
10263 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10265 struct task_group *tg = css_tg(css);
10268 * Relies on the RCU grace period between css_released() and this.
10270 sched_unregister_group(tg);
10273 #ifdef CONFIG_RT_GROUP_SCHED
10274 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10276 struct task_struct *task;
10277 struct cgroup_subsys_state *css;
10279 cgroup_taskset_for_each(task, css, tset) {
10280 if (!sched_rt_can_attach(css_tg(css), task))
10287 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10289 struct task_struct *task;
10290 struct cgroup_subsys_state *css;
10292 cgroup_taskset_for_each(task, css, tset)
10293 sched_move_task(task);
10296 #ifdef CONFIG_UCLAMP_TASK_GROUP
10297 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10299 struct cgroup_subsys_state *top_css = css;
10300 struct uclamp_se *uc_parent = NULL;
10301 struct uclamp_se *uc_se = NULL;
10302 unsigned int eff[UCLAMP_CNT];
10303 enum uclamp_id clamp_id;
10304 unsigned int clamps;
10306 lockdep_assert_held(&uclamp_mutex);
10307 SCHED_WARN_ON(!rcu_read_lock_held());
10309 css_for_each_descendant_pre(css, top_css) {
10310 uc_parent = css_tg(css)->parent
10311 ? css_tg(css)->parent->uclamp : NULL;
10313 for_each_clamp_id(clamp_id) {
10314 /* Assume effective clamps matches requested clamps */
10315 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10316 /* Cap effective clamps with parent's effective clamps */
10318 eff[clamp_id] > uc_parent[clamp_id].value) {
10319 eff[clamp_id] = uc_parent[clamp_id].value;
10322 /* Ensure protection is always capped by limit */
10323 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10325 /* Propagate most restrictive effective clamps */
10327 uc_se = css_tg(css)->uclamp;
10328 for_each_clamp_id(clamp_id) {
10329 if (eff[clamp_id] == uc_se[clamp_id].value)
10331 uc_se[clamp_id].value = eff[clamp_id];
10332 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10333 clamps |= (0x1 << clamp_id);
10336 css = css_rightmost_descendant(css);
10340 /* Immediately update descendants RUNNABLE tasks */
10341 uclamp_update_active_tasks(css);
10346 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10347 * C expression. Since there is no way to convert a macro argument (N) into a
10348 * character constant, use two levels of macros.
10350 #define _POW10(exp) ((unsigned int)1e##exp)
10351 #define POW10(exp) _POW10(exp)
10353 struct uclamp_request {
10354 #define UCLAMP_PERCENT_SHIFT 2
10355 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10361 static inline struct uclamp_request
10362 capacity_from_percent(char *buf)
10364 struct uclamp_request req = {
10365 .percent = UCLAMP_PERCENT_SCALE,
10366 .util = SCHED_CAPACITY_SCALE,
10371 if (strcmp(buf, "max")) {
10372 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10376 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10381 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10382 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10388 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10389 size_t nbytes, loff_t off,
10390 enum uclamp_id clamp_id)
10392 struct uclamp_request req;
10393 struct task_group *tg;
10395 req = capacity_from_percent(buf);
10399 static_branch_enable(&sched_uclamp_used);
10401 mutex_lock(&uclamp_mutex);
10404 tg = css_tg(of_css(of));
10405 if (tg->uclamp_req[clamp_id].value != req.util)
10406 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10409 * Because of not recoverable conversion rounding we keep track of the
10410 * exact requested value
10412 tg->uclamp_pct[clamp_id] = req.percent;
10414 /* Update effective clamps to track the most restrictive value */
10415 cpu_util_update_eff(of_css(of));
10418 mutex_unlock(&uclamp_mutex);
10423 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10424 char *buf, size_t nbytes,
10427 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10430 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10431 char *buf, size_t nbytes,
10434 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10437 static inline void cpu_uclamp_print(struct seq_file *sf,
10438 enum uclamp_id clamp_id)
10440 struct task_group *tg;
10446 tg = css_tg(seq_css(sf));
10447 util_clamp = tg->uclamp_req[clamp_id].value;
10450 if (util_clamp == SCHED_CAPACITY_SCALE) {
10451 seq_puts(sf, "max\n");
10455 percent = tg->uclamp_pct[clamp_id];
10456 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10457 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10460 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10462 cpu_uclamp_print(sf, UCLAMP_MIN);
10466 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10468 cpu_uclamp_print(sf, UCLAMP_MAX);
10471 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10473 #ifdef CONFIG_FAIR_GROUP_SCHED
10474 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10475 struct cftype *cftype, u64 shareval)
10477 if (shareval > scale_load_down(ULONG_MAX))
10478 shareval = MAX_SHARES;
10479 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10482 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10483 struct cftype *cft)
10485 struct task_group *tg = css_tg(css);
10487 return (u64) scale_load_down(tg->shares);
10490 #ifdef CONFIG_CFS_BANDWIDTH
10491 static DEFINE_MUTEX(cfs_constraints_mutex);
10493 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10494 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10495 /* More than 203 days if BW_SHIFT equals 20. */
10496 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10498 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10500 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10503 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10504 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10506 if (tg == &root_task_group)
10510 * Ensure we have at some amount of bandwidth every period. This is
10511 * to prevent reaching a state of large arrears when throttled via
10512 * entity_tick() resulting in prolonged exit starvation.
10514 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10518 * Likewise, bound things on the other side by preventing insane quota
10519 * periods. This also allows us to normalize in computing quota
10522 if (period > max_cfs_quota_period)
10526 * Bound quota to defend quota against overflow during bandwidth shift.
10528 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10531 if (quota != RUNTIME_INF && (burst > quota ||
10532 burst + quota > max_cfs_runtime))
10536 * Prevent race between setting of cfs_rq->runtime_enabled and
10537 * unthrottle_offline_cfs_rqs().
10540 mutex_lock(&cfs_constraints_mutex);
10541 ret = __cfs_schedulable(tg, period, quota);
10545 runtime_enabled = quota != RUNTIME_INF;
10546 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10548 * If we need to toggle cfs_bandwidth_used, off->on must occur
10549 * before making related changes, and on->off must occur afterwards
10551 if (runtime_enabled && !runtime_was_enabled)
10552 cfs_bandwidth_usage_inc();
10553 raw_spin_lock_irq(&cfs_b->lock);
10554 cfs_b->period = ns_to_ktime(period);
10555 cfs_b->quota = quota;
10556 cfs_b->burst = burst;
10558 __refill_cfs_bandwidth_runtime(cfs_b);
10560 /* Restart the period timer (if active) to handle new period expiry: */
10561 if (runtime_enabled)
10562 start_cfs_bandwidth(cfs_b);
10564 raw_spin_unlock_irq(&cfs_b->lock);
10566 for_each_online_cpu(i) {
10567 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10568 struct rq *rq = cfs_rq->rq;
10569 struct rq_flags rf;
10571 rq_lock_irq(rq, &rf);
10572 cfs_rq->runtime_enabled = runtime_enabled;
10573 cfs_rq->runtime_remaining = 0;
10575 if (cfs_rq->throttled)
10576 unthrottle_cfs_rq(cfs_rq);
10577 rq_unlock_irq(rq, &rf);
10579 if (runtime_was_enabled && !runtime_enabled)
10580 cfs_bandwidth_usage_dec();
10582 mutex_unlock(&cfs_constraints_mutex);
10583 cpus_read_unlock();
10588 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10590 u64 quota, period, burst;
10592 period = ktime_to_ns(tg->cfs_bandwidth.period);
10593 burst = tg->cfs_bandwidth.burst;
10594 if (cfs_quota_us < 0)
10595 quota = RUNTIME_INF;
10596 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10597 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10601 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10604 static long tg_get_cfs_quota(struct task_group *tg)
10608 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10611 quota_us = tg->cfs_bandwidth.quota;
10612 do_div(quota_us, NSEC_PER_USEC);
10617 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10619 u64 quota, period, burst;
10621 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10624 period = (u64)cfs_period_us * NSEC_PER_USEC;
10625 quota = tg->cfs_bandwidth.quota;
10626 burst = tg->cfs_bandwidth.burst;
10628 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10631 static long tg_get_cfs_period(struct task_group *tg)
10635 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10636 do_div(cfs_period_us, NSEC_PER_USEC);
10638 return cfs_period_us;
10641 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10643 u64 quota, period, burst;
10645 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10648 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10649 period = ktime_to_ns(tg->cfs_bandwidth.period);
10650 quota = tg->cfs_bandwidth.quota;
10652 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10655 static long tg_get_cfs_burst(struct task_group *tg)
10659 burst_us = tg->cfs_bandwidth.burst;
10660 do_div(burst_us, NSEC_PER_USEC);
10665 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10666 struct cftype *cft)
10668 return tg_get_cfs_quota(css_tg(css));
10671 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10672 struct cftype *cftype, s64 cfs_quota_us)
10674 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10677 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10678 struct cftype *cft)
10680 return tg_get_cfs_period(css_tg(css));
10683 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10684 struct cftype *cftype, u64 cfs_period_us)
10686 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10689 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10690 struct cftype *cft)
10692 return tg_get_cfs_burst(css_tg(css));
10695 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10696 struct cftype *cftype, u64 cfs_burst_us)
10698 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10701 struct cfs_schedulable_data {
10702 struct task_group *tg;
10707 * normalize group quota/period to be quota/max_period
10708 * note: units are usecs
10710 static u64 normalize_cfs_quota(struct task_group *tg,
10711 struct cfs_schedulable_data *d)
10716 period = d->period;
10719 period = tg_get_cfs_period(tg);
10720 quota = tg_get_cfs_quota(tg);
10723 /* note: these should typically be equivalent */
10724 if (quota == RUNTIME_INF || quota == -1)
10725 return RUNTIME_INF;
10727 return to_ratio(period, quota);
10730 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10732 struct cfs_schedulable_data *d = data;
10733 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10734 s64 quota = 0, parent_quota = -1;
10737 quota = RUNTIME_INF;
10739 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10741 quota = normalize_cfs_quota(tg, d);
10742 parent_quota = parent_b->hierarchical_quota;
10745 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10746 * always take the min. On cgroup1, only inherit when no
10749 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10750 quota = min(quota, parent_quota);
10752 if (quota == RUNTIME_INF)
10753 quota = parent_quota;
10754 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10758 cfs_b->hierarchical_quota = quota;
10763 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10766 struct cfs_schedulable_data data = {
10772 if (quota != RUNTIME_INF) {
10773 do_div(data.period, NSEC_PER_USEC);
10774 do_div(data.quota, NSEC_PER_USEC);
10778 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10784 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10786 struct task_group *tg = css_tg(seq_css(sf));
10787 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10789 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10790 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10791 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10793 if (schedstat_enabled() && tg != &root_task_group) {
10794 struct sched_statistics *stats;
10798 for_each_possible_cpu(i) {
10799 stats = __schedstats_from_se(tg->se[i]);
10800 ws += schedstat_val(stats->wait_sum);
10803 seq_printf(sf, "wait_sum %llu\n", ws);
10806 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10807 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10811 #endif /* CONFIG_CFS_BANDWIDTH */
10812 #endif /* CONFIG_FAIR_GROUP_SCHED */
10814 #ifdef CONFIG_RT_GROUP_SCHED
10815 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10816 struct cftype *cft, s64 val)
10818 return sched_group_set_rt_runtime(css_tg(css), val);
10821 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10822 struct cftype *cft)
10824 return sched_group_rt_runtime(css_tg(css));
10827 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10828 struct cftype *cftype, u64 rt_period_us)
10830 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10833 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10834 struct cftype *cft)
10836 return sched_group_rt_period(css_tg(css));
10838 #endif /* CONFIG_RT_GROUP_SCHED */
10840 #ifdef CONFIG_FAIR_GROUP_SCHED
10841 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10842 struct cftype *cft)
10844 return css_tg(css)->idle;
10847 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10848 struct cftype *cft, s64 idle)
10850 return sched_group_set_idle(css_tg(css), idle);
10854 static struct cftype cpu_legacy_files[] = {
10855 #ifdef CONFIG_FAIR_GROUP_SCHED
10858 .read_u64 = cpu_shares_read_u64,
10859 .write_u64 = cpu_shares_write_u64,
10863 .read_s64 = cpu_idle_read_s64,
10864 .write_s64 = cpu_idle_write_s64,
10867 #ifdef CONFIG_CFS_BANDWIDTH
10869 .name = "cfs_quota_us",
10870 .read_s64 = cpu_cfs_quota_read_s64,
10871 .write_s64 = cpu_cfs_quota_write_s64,
10874 .name = "cfs_period_us",
10875 .read_u64 = cpu_cfs_period_read_u64,
10876 .write_u64 = cpu_cfs_period_write_u64,
10879 .name = "cfs_burst_us",
10880 .read_u64 = cpu_cfs_burst_read_u64,
10881 .write_u64 = cpu_cfs_burst_write_u64,
10885 .seq_show = cpu_cfs_stat_show,
10888 #ifdef CONFIG_RT_GROUP_SCHED
10890 .name = "rt_runtime_us",
10891 .read_s64 = cpu_rt_runtime_read,
10892 .write_s64 = cpu_rt_runtime_write,
10895 .name = "rt_period_us",
10896 .read_u64 = cpu_rt_period_read_uint,
10897 .write_u64 = cpu_rt_period_write_uint,
10900 #ifdef CONFIG_UCLAMP_TASK_GROUP
10902 .name = "uclamp.min",
10903 .flags = CFTYPE_NOT_ON_ROOT,
10904 .seq_show = cpu_uclamp_min_show,
10905 .write = cpu_uclamp_min_write,
10908 .name = "uclamp.max",
10909 .flags = CFTYPE_NOT_ON_ROOT,
10910 .seq_show = cpu_uclamp_max_show,
10911 .write = cpu_uclamp_max_write,
10914 { } /* Terminate */
10917 static int cpu_extra_stat_show(struct seq_file *sf,
10918 struct cgroup_subsys_state *css)
10920 #ifdef CONFIG_CFS_BANDWIDTH
10922 struct task_group *tg = css_tg(css);
10923 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10924 u64 throttled_usec, burst_usec;
10926 throttled_usec = cfs_b->throttled_time;
10927 do_div(throttled_usec, NSEC_PER_USEC);
10928 burst_usec = cfs_b->burst_time;
10929 do_div(burst_usec, NSEC_PER_USEC);
10931 seq_printf(sf, "nr_periods %d\n"
10932 "nr_throttled %d\n"
10933 "throttled_usec %llu\n"
10935 "burst_usec %llu\n",
10936 cfs_b->nr_periods, cfs_b->nr_throttled,
10937 throttled_usec, cfs_b->nr_burst, burst_usec);
10943 #ifdef CONFIG_FAIR_GROUP_SCHED
10944 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10945 struct cftype *cft)
10947 struct task_group *tg = css_tg(css);
10948 u64 weight = scale_load_down(tg->shares);
10950 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10953 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10954 struct cftype *cft, u64 weight)
10957 * cgroup weight knobs should use the common MIN, DFL and MAX
10958 * values which are 1, 100 and 10000 respectively. While it loses
10959 * a bit of range on both ends, it maps pretty well onto the shares
10960 * value used by scheduler and the round-trip conversions preserve
10961 * the original value over the entire range.
10963 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10966 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10968 return sched_group_set_shares(css_tg(css), scale_load(weight));
10971 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10972 struct cftype *cft)
10974 unsigned long weight = scale_load_down(css_tg(css)->shares);
10975 int last_delta = INT_MAX;
10978 /* find the closest nice value to the current weight */
10979 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10980 delta = abs(sched_prio_to_weight[prio] - weight);
10981 if (delta >= last_delta)
10983 last_delta = delta;
10986 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10989 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10990 struct cftype *cft, s64 nice)
10992 unsigned long weight;
10995 if (nice < MIN_NICE || nice > MAX_NICE)
10998 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10999 idx = array_index_nospec(idx, 40);
11000 weight = sched_prio_to_weight[idx];
11002 return sched_group_set_shares(css_tg(css), scale_load(weight));
11006 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11007 long period, long quota)
11010 seq_puts(sf, "max");
11012 seq_printf(sf, "%ld", quota);
11014 seq_printf(sf, " %ld\n", period);
11017 /* caller should put the current value in *@periodp before calling */
11018 static int __maybe_unused cpu_period_quota_parse(char *buf,
11019 u64 *periodp, u64 *quotap)
11021 char tok[21]; /* U64_MAX */
11023 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11026 *periodp *= NSEC_PER_USEC;
11028 if (sscanf(tok, "%llu", quotap))
11029 *quotap *= NSEC_PER_USEC;
11030 else if (!strcmp(tok, "max"))
11031 *quotap = RUNTIME_INF;
11038 #ifdef CONFIG_CFS_BANDWIDTH
11039 static int cpu_max_show(struct seq_file *sf, void *v)
11041 struct task_group *tg = css_tg(seq_css(sf));
11043 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11047 static ssize_t cpu_max_write(struct kernfs_open_file *of,
11048 char *buf, size_t nbytes, loff_t off)
11050 struct task_group *tg = css_tg(of_css(of));
11051 u64 period = tg_get_cfs_period(tg);
11052 u64 burst = tg_get_cfs_burst(tg);
11056 ret = cpu_period_quota_parse(buf, &period, "a);
11058 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11059 return ret ?: nbytes;
11063 static struct cftype cpu_files[] = {
11064 #ifdef CONFIG_FAIR_GROUP_SCHED
11067 .flags = CFTYPE_NOT_ON_ROOT,
11068 .read_u64 = cpu_weight_read_u64,
11069 .write_u64 = cpu_weight_write_u64,
11072 .name = "weight.nice",
11073 .flags = CFTYPE_NOT_ON_ROOT,
11074 .read_s64 = cpu_weight_nice_read_s64,
11075 .write_s64 = cpu_weight_nice_write_s64,
11079 .flags = CFTYPE_NOT_ON_ROOT,
11080 .read_s64 = cpu_idle_read_s64,
11081 .write_s64 = cpu_idle_write_s64,
11084 #ifdef CONFIG_CFS_BANDWIDTH
11087 .flags = CFTYPE_NOT_ON_ROOT,
11088 .seq_show = cpu_max_show,
11089 .write = cpu_max_write,
11092 .name = "max.burst",
11093 .flags = CFTYPE_NOT_ON_ROOT,
11094 .read_u64 = cpu_cfs_burst_read_u64,
11095 .write_u64 = cpu_cfs_burst_write_u64,
11098 #ifdef CONFIG_UCLAMP_TASK_GROUP
11100 .name = "uclamp.min",
11101 .flags = CFTYPE_NOT_ON_ROOT,
11102 .seq_show = cpu_uclamp_min_show,
11103 .write = cpu_uclamp_min_write,
11106 .name = "uclamp.max",
11107 .flags = CFTYPE_NOT_ON_ROOT,
11108 .seq_show = cpu_uclamp_max_show,
11109 .write = cpu_uclamp_max_write,
11112 { } /* terminate */
11115 struct cgroup_subsys cpu_cgrp_subsys = {
11116 .css_alloc = cpu_cgroup_css_alloc,
11117 .css_online = cpu_cgroup_css_online,
11118 .css_released = cpu_cgroup_css_released,
11119 .css_free = cpu_cgroup_css_free,
11120 .css_extra_stat_show = cpu_extra_stat_show,
11121 #ifdef CONFIG_RT_GROUP_SCHED
11122 .can_attach = cpu_cgroup_can_attach,
11124 .attach = cpu_cgroup_attach,
11125 .legacy_cftypes = cpu_legacy_files,
11126 .dfl_cftypes = cpu_files,
11127 .early_init = true,
11131 #endif /* CONFIG_CGROUP_SCHED */
11133 void dump_cpu_task(int cpu)
11135 if (cpu == smp_processor_id() && in_hardirq()) {
11136 struct pt_regs *regs;
11138 regs = get_irq_regs();
11145 if (trigger_single_cpu_backtrace(cpu))
11148 pr_info("Task dump for CPU %d:\n", cpu);
11149 sched_show_task(cpu_curr(cpu));
11153 * Nice levels are multiplicative, with a gentle 10% change for every
11154 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11155 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11156 * that remained on nice 0.
11158 * The "10% effect" is relative and cumulative: from _any_ nice level,
11159 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11160 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11161 * If a task goes up by ~10% and another task goes down by ~10% then
11162 * the relative distance between them is ~25%.)
11164 const int sched_prio_to_weight[40] = {
11165 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11166 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11167 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11168 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11169 /* 0 */ 1024, 820, 655, 526, 423,
11170 /* 5 */ 335, 272, 215, 172, 137,
11171 /* 10 */ 110, 87, 70, 56, 45,
11172 /* 15 */ 36, 29, 23, 18, 15,
11176 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11178 * In cases where the weight does not change often, we can use the
11179 * precalculated inverse to speed up arithmetics by turning divisions
11180 * into multiplications:
11182 const u32 sched_prio_to_wmult[40] = {
11183 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11184 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11185 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11186 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11187 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11188 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11189 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11190 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11193 void call_trace_sched_update_nr_running(struct rq *rq, int count)
11195 trace_sched_update_nr_running_tp(rq, count);