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/rseq.h>
61 #include <linux/sched/wake_q.h>
62 #include <linux/scs.h>
63 #include <linux/slab.h>
64 #include <linux/syscalls.h>
65 #include <linux/vtime.h>
66 #include <linux/wait_api.h>
67 #include <linux/workqueue_api.h>
69 #ifdef CONFIG_PREEMPT_DYNAMIC
70 # ifdef CONFIG_GENERIC_ENTRY
71 # include <linux/entry-common.h>
75 #include <uapi/linux/sched/types.h>
77 #include <asm/irq_regs.h>
78 #include <asm/switch_to.h>
81 #define CREATE_TRACE_POINTS
82 #include <linux/sched/rseq_api.h>
83 #include <trace/events/sched.h>
84 #include <trace/events/ipi.h>
85 #undef CREATE_TRACE_POINTS
90 #include "autogroup.h"
95 #include "../workqueue_internal.h"
96 #include "../../io_uring/io-wq.h"
97 #include "../smpboot.h"
99 EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpu);
100 EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpumask);
103 * Export tracepoints that act as a bare tracehook (ie: have no trace event
104 * associated with them) to allow external modules to probe them.
106 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
107 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
108 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
109 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
110 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
111 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
112 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
113 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
114 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
115 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
116 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
117 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_compute_energy_tp);
119 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
121 #ifdef CONFIG_SCHED_DEBUG
123 * Debugging: various feature bits
125 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
126 * sysctl_sched_features, defined in sched.h, to allow constants propagation
127 * at compile time and compiler optimization based on features default.
129 #define SCHED_FEAT(name, enabled) \
130 (1UL << __SCHED_FEAT_##name) * enabled |
131 const_debug unsigned int sysctl_sched_features =
132 #include "features.h"
137 * Print a warning if need_resched is set for the given duration (if
138 * LATENCY_WARN is enabled).
140 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
143 __read_mostly int sysctl_resched_latency_warn_ms = 100;
144 __read_mostly int sysctl_resched_latency_warn_once = 1;
145 #endif /* CONFIG_SCHED_DEBUG */
148 * Number of tasks to iterate in a single balance run.
149 * Limited because this is done with IRQs disabled.
151 const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
153 __read_mostly int scheduler_running;
155 #ifdef CONFIG_SCHED_CORE
157 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
159 /* kernel prio, less is more */
160 static inline int __task_prio(const struct task_struct *p)
162 if (p->sched_class == &stop_sched_class) /* trumps deadline */
165 if (rt_prio(p->prio)) /* includes deadline */
166 return p->prio; /* [-1, 99] */
168 if (p->sched_class == &idle_sched_class)
169 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
171 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
181 /* real prio, less is less */
182 static inline bool prio_less(const struct task_struct *a,
183 const struct task_struct *b, bool in_fi)
186 int pa = __task_prio(a), pb = __task_prio(b);
194 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
195 return !dl_time_before(a->dl.deadline, b->dl.deadline);
197 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
198 return cfs_prio_less(a, b, in_fi);
203 static inline bool __sched_core_less(const struct task_struct *a,
204 const struct task_struct *b)
206 if (a->core_cookie < b->core_cookie)
209 if (a->core_cookie > b->core_cookie)
212 /* flip prio, so high prio is leftmost */
213 if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
219 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
221 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
223 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
226 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
228 const struct task_struct *p = __node_2_sc(node);
229 unsigned long cookie = (unsigned long)key;
231 if (cookie < p->core_cookie)
234 if (cookie > p->core_cookie)
240 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
242 rq->core->core_task_seq++;
247 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
250 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
252 rq->core->core_task_seq++;
254 if (sched_core_enqueued(p)) {
255 rb_erase(&p->core_node, &rq->core_tree);
256 RB_CLEAR_NODE(&p->core_node);
260 * Migrating the last task off the cpu, with the cpu in forced idle
261 * state. Reschedule to create an accounting edge for forced idle,
262 * and re-examine whether the core is still in forced idle state.
264 if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
265 rq->core->core_forceidle_count && rq->curr == rq->idle)
269 static int sched_task_is_throttled(struct task_struct *p, int cpu)
271 if (p->sched_class->task_is_throttled)
272 return p->sched_class->task_is_throttled(p, cpu);
277 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
279 struct rb_node *node = &p->core_node;
280 int cpu = task_cpu(p);
283 node = rb_next(node);
287 p = __node_2_sc(node);
288 if (p->core_cookie != cookie)
291 } while (sched_task_is_throttled(p, cpu));
297 * Find left-most (aka, highest priority) and unthrottled task matching @cookie.
298 * If no suitable task is found, NULL will be returned.
300 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
302 struct task_struct *p;
303 struct rb_node *node;
305 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
309 p = __node_2_sc(node);
310 if (!sched_task_is_throttled(p, rq->cpu))
313 return sched_core_next(p, cookie);
317 * Magic required such that:
319 * raw_spin_rq_lock(rq);
321 * raw_spin_rq_unlock(rq);
323 * ends up locking and unlocking the _same_ lock, and all CPUs
324 * always agree on what rq has what lock.
326 * XXX entirely possible to selectively enable cores, don't bother for now.
329 static DEFINE_MUTEX(sched_core_mutex);
330 static atomic_t sched_core_count;
331 static struct cpumask sched_core_mask;
333 static void sched_core_lock(int cpu, unsigned long *flags)
335 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
338 local_irq_save(*flags);
339 for_each_cpu(t, smt_mask)
340 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
343 static void sched_core_unlock(int cpu, unsigned long *flags)
345 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
348 for_each_cpu(t, smt_mask)
349 raw_spin_unlock(&cpu_rq(t)->__lock);
350 local_irq_restore(*flags);
353 static void __sched_core_flip(bool enabled)
361 * Toggle the online cores, one by one.
363 cpumask_copy(&sched_core_mask, cpu_online_mask);
364 for_each_cpu(cpu, &sched_core_mask) {
365 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
367 sched_core_lock(cpu, &flags);
369 for_each_cpu(t, smt_mask)
370 cpu_rq(t)->core_enabled = enabled;
372 cpu_rq(cpu)->core->core_forceidle_start = 0;
374 sched_core_unlock(cpu, &flags);
376 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
380 * Toggle the offline CPUs.
382 for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
383 cpu_rq(cpu)->core_enabled = enabled;
388 static void sched_core_assert_empty(void)
392 for_each_possible_cpu(cpu)
393 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
396 static void __sched_core_enable(void)
398 static_branch_enable(&__sched_core_enabled);
400 * Ensure all previous instances of raw_spin_rq_*lock() have finished
401 * and future ones will observe !sched_core_disabled().
404 __sched_core_flip(true);
405 sched_core_assert_empty();
408 static void __sched_core_disable(void)
410 sched_core_assert_empty();
411 __sched_core_flip(false);
412 static_branch_disable(&__sched_core_enabled);
415 void sched_core_get(void)
417 if (atomic_inc_not_zero(&sched_core_count))
420 mutex_lock(&sched_core_mutex);
421 if (!atomic_read(&sched_core_count))
422 __sched_core_enable();
424 smp_mb__before_atomic();
425 atomic_inc(&sched_core_count);
426 mutex_unlock(&sched_core_mutex);
429 static void __sched_core_put(struct work_struct *work)
431 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
432 __sched_core_disable();
433 mutex_unlock(&sched_core_mutex);
437 void sched_core_put(void)
439 static DECLARE_WORK(_work, __sched_core_put);
442 * "There can be only one"
444 * Either this is the last one, or we don't actually need to do any
445 * 'work'. If it is the last *again*, we rely on
446 * WORK_STRUCT_PENDING_BIT.
448 if (!atomic_add_unless(&sched_core_count, -1, 1))
449 schedule_work(&_work);
452 #else /* !CONFIG_SCHED_CORE */
454 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
456 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
458 #endif /* CONFIG_SCHED_CORE */
461 * Serialization rules:
467 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
470 * rq2->lock where: rq1 < rq2
474 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
475 * local CPU's rq->lock, it optionally removes the task from the runqueue and
476 * always looks at the local rq data structures to find the most eligible task
479 * Task enqueue is also under rq->lock, possibly taken from another CPU.
480 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
481 * the local CPU to avoid bouncing the runqueue state around [ see
482 * ttwu_queue_wakelist() ]
484 * Task wakeup, specifically wakeups that involve migration, are horribly
485 * complicated to avoid having to take two rq->locks.
489 * System-calls and anything external will use task_rq_lock() which acquires
490 * both p->pi_lock and rq->lock. As a consequence the state they change is
491 * stable while holding either lock:
493 * - sched_setaffinity()/
494 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
495 * - set_user_nice(): p->se.load, p->*prio
496 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
497 * p->se.load, p->rt_priority,
498 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
499 * - sched_setnuma(): p->numa_preferred_nid
500 * - sched_move_task(): p->sched_task_group
501 * - uclamp_update_active() p->uclamp*
503 * p->state <- TASK_*:
505 * is changed locklessly using set_current_state(), __set_current_state() or
506 * set_special_state(), see their respective comments, or by
507 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
510 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
512 * is set by activate_task() and cleared by deactivate_task(), under
513 * rq->lock. Non-zero indicates the task is runnable, the special
514 * ON_RQ_MIGRATING state is used for migration without holding both
515 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
517 * p->on_cpu <- { 0, 1 }:
519 * is set by prepare_task() and cleared by finish_task() such that it will be
520 * set before p is scheduled-in and cleared after p is scheduled-out, both
521 * under rq->lock. Non-zero indicates the task is running on its CPU.
523 * [ The astute reader will observe that it is possible for two tasks on one
524 * CPU to have ->on_cpu = 1 at the same time. ]
526 * task_cpu(p): is changed by set_task_cpu(), the rules are:
528 * - Don't call set_task_cpu() on a blocked task:
530 * We don't care what CPU we're not running on, this simplifies hotplug,
531 * the CPU assignment of blocked tasks isn't required to be valid.
533 * - for try_to_wake_up(), called under p->pi_lock:
535 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
537 * - for migration called under rq->lock:
538 * [ see task_on_rq_migrating() in task_rq_lock() ]
540 * o move_queued_task()
543 * - for migration called under double_rq_lock():
545 * o __migrate_swap_task()
546 * o push_rt_task() / pull_rt_task()
547 * o push_dl_task() / pull_dl_task()
548 * o dl_task_offline_migration()
552 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
554 raw_spinlock_t *lock;
556 /* Matches synchronize_rcu() in __sched_core_enable() */
558 if (sched_core_disabled()) {
559 raw_spin_lock_nested(&rq->__lock, subclass);
560 /* preempt_count *MUST* be > 1 */
561 preempt_enable_no_resched();
566 lock = __rq_lockp(rq);
567 raw_spin_lock_nested(lock, subclass);
568 if (likely(lock == __rq_lockp(rq))) {
569 /* preempt_count *MUST* be > 1 */
570 preempt_enable_no_resched();
573 raw_spin_unlock(lock);
577 bool raw_spin_rq_trylock(struct rq *rq)
579 raw_spinlock_t *lock;
582 /* Matches synchronize_rcu() in __sched_core_enable() */
584 if (sched_core_disabled()) {
585 ret = raw_spin_trylock(&rq->__lock);
591 lock = __rq_lockp(rq);
592 ret = raw_spin_trylock(lock);
593 if (!ret || (likely(lock == __rq_lockp(rq)))) {
597 raw_spin_unlock(lock);
601 void raw_spin_rq_unlock(struct rq *rq)
603 raw_spin_unlock(rq_lockp(rq));
608 * double_rq_lock - safely lock two runqueues
610 void double_rq_lock(struct rq *rq1, struct rq *rq2)
612 lockdep_assert_irqs_disabled();
614 if (rq_order_less(rq2, rq1))
617 raw_spin_rq_lock(rq1);
618 if (__rq_lockp(rq1) != __rq_lockp(rq2))
619 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
621 double_rq_clock_clear_update(rq1, rq2);
626 * __task_rq_lock - lock the rq @p resides on.
628 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
633 lockdep_assert_held(&p->pi_lock);
637 raw_spin_rq_lock(rq);
638 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
642 raw_spin_rq_unlock(rq);
644 while (unlikely(task_on_rq_migrating(p)))
650 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
652 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
653 __acquires(p->pi_lock)
659 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
661 raw_spin_rq_lock(rq);
663 * move_queued_task() task_rq_lock()
666 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
667 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
668 * [S] ->cpu = new_cpu [L] task_rq()
672 * If we observe the old CPU in task_rq_lock(), the acquire of
673 * the old rq->lock will fully serialize against the stores.
675 * If we observe the new CPU in task_rq_lock(), the address
676 * dependency headed by '[L] rq = task_rq()' and the acquire
677 * will pair with the WMB to ensure we then also see migrating.
679 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
683 raw_spin_rq_unlock(rq);
684 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
686 while (unlikely(task_on_rq_migrating(p)))
692 * RQ-clock updating methods:
695 static void update_rq_clock_task(struct rq *rq, s64 delta)
698 * In theory, the compile should just see 0 here, and optimize out the call
699 * to sched_rt_avg_update. But I don't trust it...
701 s64 __maybe_unused steal = 0, irq_delta = 0;
703 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
704 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
707 * Since irq_time is only updated on {soft,}irq_exit, we might run into
708 * this case when a previous update_rq_clock() happened inside a
711 * When this happens, we stop ->clock_task and only update the
712 * prev_irq_time stamp to account for the part that fit, so that a next
713 * update will consume the rest. This ensures ->clock_task is
716 * It does however cause some slight miss-attribution of {soft,}irq
717 * time, a more accurate solution would be to update the irq_time using
718 * the current rq->clock timestamp, except that would require using
721 if (irq_delta > delta)
724 rq->prev_irq_time += irq_delta;
726 psi_account_irqtime(rq->curr, irq_delta);
727 delayacct_irq(rq->curr, irq_delta);
729 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
730 if (static_key_false((¶virt_steal_rq_enabled))) {
731 steal = paravirt_steal_clock(cpu_of(rq));
732 steal -= rq->prev_steal_time_rq;
734 if (unlikely(steal > delta))
737 rq->prev_steal_time_rq += steal;
742 rq->clock_task += delta;
744 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
745 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
746 update_irq_load_avg(rq, irq_delta + steal);
748 update_rq_clock_pelt(rq, delta);
751 void update_rq_clock(struct rq *rq)
755 lockdep_assert_rq_held(rq);
757 if (rq->clock_update_flags & RQCF_ACT_SKIP)
760 #ifdef CONFIG_SCHED_DEBUG
761 if (sched_feat(WARN_DOUBLE_CLOCK))
762 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
763 rq->clock_update_flags |= RQCF_UPDATED;
766 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
770 update_rq_clock_task(rq, delta);
773 #ifdef CONFIG_SCHED_HRTICK
775 * Use HR-timers to deliver accurate preemption points.
778 static void hrtick_clear(struct rq *rq)
780 if (hrtimer_active(&rq->hrtick_timer))
781 hrtimer_cancel(&rq->hrtick_timer);
785 * High-resolution timer tick.
786 * Runs from hardirq context with interrupts disabled.
788 static enum hrtimer_restart hrtick(struct hrtimer *timer)
790 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
793 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
797 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
800 return HRTIMER_NORESTART;
805 static void __hrtick_restart(struct rq *rq)
807 struct hrtimer *timer = &rq->hrtick_timer;
808 ktime_t time = rq->hrtick_time;
810 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
814 * called from hardirq (IPI) context
816 static void __hrtick_start(void *arg)
822 __hrtick_restart(rq);
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)
833 struct hrtimer *timer = &rq->hrtick_timer;
837 * Don't schedule slices shorter than 10000ns, that just
838 * doesn't make sense and can cause timer DoS.
840 delta = max_t(s64, delay, 10000LL);
841 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
844 __hrtick_restart(rq);
846 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
851 * Called to set the hrtick timer state.
853 * called with rq->lock held and irqs disabled
855 void hrtick_start(struct rq *rq, u64 delay)
858 * Don't schedule slices shorter than 10000ns, that just
859 * doesn't make sense. Rely on vruntime for fairness.
861 delay = max_t(u64, delay, 10000LL);
862 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
863 HRTIMER_MODE_REL_PINNED_HARD);
866 #endif /* CONFIG_SMP */
868 static void hrtick_rq_init(struct rq *rq)
871 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
873 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
874 rq->hrtick_timer.function = hrtick;
876 #else /* CONFIG_SCHED_HRTICK */
877 static inline void hrtick_clear(struct rq *rq)
881 static inline void hrtick_rq_init(struct rq *rq)
884 #endif /* CONFIG_SCHED_HRTICK */
887 * cmpxchg based fetch_or, macro so it works for different integer types
889 #define fetch_or(ptr, mask) \
891 typeof(ptr) _ptr = (ptr); \
892 typeof(mask) _mask = (mask); \
893 typeof(*_ptr) _val = *_ptr; \
896 } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
900 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
902 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
903 * this avoids any races wrt polling state changes and thereby avoids
906 static inline bool set_nr_and_not_polling(struct task_struct *p)
908 struct thread_info *ti = task_thread_info(p);
909 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
913 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
915 * If this returns true, then the idle task promises to call
916 * sched_ttwu_pending() and reschedule soon.
918 static bool set_nr_if_polling(struct task_struct *p)
920 struct thread_info *ti = task_thread_info(p);
921 typeof(ti->flags) val = READ_ONCE(ti->flags);
924 if (!(val & _TIF_POLLING_NRFLAG))
926 if (val & _TIF_NEED_RESCHED)
928 } while (!try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED));
934 static inline bool set_nr_and_not_polling(struct task_struct *p)
936 set_tsk_need_resched(p);
941 static inline bool set_nr_if_polling(struct task_struct *p)
948 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
950 struct wake_q_node *node = &task->wake_q;
953 * Atomically grab the task, if ->wake_q is !nil already it means
954 * it's already queued (either by us or someone else) and will get the
955 * wakeup due to that.
957 * In order to ensure that a pending wakeup will observe our pending
958 * state, even in the failed case, an explicit smp_mb() must be used.
960 smp_mb__before_atomic();
961 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
965 * The head is context local, there can be no concurrency.
968 head->lastp = &node->next;
973 * wake_q_add() - queue a wakeup for 'later' waking.
974 * @head: the wake_q_head to add @task to
975 * @task: the task to queue for 'later' wakeup
977 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
978 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
981 * This function must be used as-if it were wake_up_process(); IOW the task
982 * must be ready to be woken at this location.
984 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
986 if (__wake_q_add(head, task))
987 get_task_struct(task);
991 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
992 * @head: the wake_q_head to add @task to
993 * @task: the task to queue for 'later' wakeup
995 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
996 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
999 * This function must be used as-if it were wake_up_process(); IOW the task
1000 * must be ready to be woken at this location.
1002 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
1003 * that already hold reference to @task can call the 'safe' version and trust
1004 * wake_q to do the right thing depending whether or not the @task is already
1005 * queued for wakeup.
1007 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
1009 if (!__wake_q_add(head, task))
1010 put_task_struct(task);
1013 void wake_up_q(struct wake_q_head *head)
1015 struct wake_q_node *node = head->first;
1017 while (node != WAKE_Q_TAIL) {
1018 struct task_struct *task;
1020 task = container_of(node, struct task_struct, wake_q);
1021 /* Task can safely be re-inserted now: */
1023 task->wake_q.next = NULL;
1026 * wake_up_process() executes a full barrier, which pairs with
1027 * the queueing in wake_q_add() so as not to miss wakeups.
1029 wake_up_process(task);
1030 put_task_struct(task);
1035 * resched_curr - mark rq's current task 'to be rescheduled now'.
1037 * On UP this means the setting of the need_resched flag, on SMP it
1038 * might also involve a cross-CPU call to trigger the scheduler on
1041 void resched_curr(struct rq *rq)
1043 struct task_struct *curr = rq->curr;
1046 lockdep_assert_rq_held(rq);
1048 if (test_tsk_need_resched(curr))
1053 if (cpu == smp_processor_id()) {
1054 set_tsk_need_resched(curr);
1055 set_preempt_need_resched();
1059 if (set_nr_and_not_polling(curr))
1060 smp_send_reschedule(cpu);
1062 trace_sched_wake_idle_without_ipi(cpu);
1065 void resched_cpu(int cpu)
1067 struct rq *rq = cpu_rq(cpu);
1068 unsigned long flags;
1070 raw_spin_rq_lock_irqsave(rq, flags);
1071 if (cpu_online(cpu) || cpu == smp_processor_id())
1073 raw_spin_rq_unlock_irqrestore(rq, flags);
1077 #ifdef CONFIG_NO_HZ_COMMON
1079 * In the semi idle case, use the nearest busy CPU for migrating timers
1080 * from an idle CPU. This is good for power-savings.
1082 * We don't do similar optimization for completely idle system, as
1083 * selecting an idle CPU will add more delays to the timers than intended
1084 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1086 int get_nohz_timer_target(void)
1088 int i, cpu = smp_processor_id(), default_cpu = -1;
1089 struct sched_domain *sd;
1090 const struct cpumask *hk_mask;
1092 if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1098 hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1102 for_each_domain(cpu, sd) {
1103 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1112 if (default_cpu == -1)
1113 default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1119 * When add_timer_on() enqueues a timer into the timer wheel of an
1120 * idle CPU then this timer might expire before the next timer event
1121 * which is scheduled to wake up that CPU. In case of a completely
1122 * idle system the next event might even be infinite time into the
1123 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1124 * leaves the inner idle loop so the newly added timer is taken into
1125 * account when the CPU goes back to idle and evaluates the timer
1126 * wheel for the next timer event.
1128 static void wake_up_idle_cpu(int cpu)
1130 struct rq *rq = cpu_rq(cpu);
1132 if (cpu == smp_processor_id())
1136 * Set TIF_NEED_RESCHED and send an IPI if in the non-polling
1137 * part of the idle loop. This forces an exit from the idle loop
1138 * and a round trip to schedule(). Now this could be optimized
1139 * because a simple new idle loop iteration is enough to
1140 * re-evaluate the next tick. Provided some re-ordering of tick
1141 * nohz functions that would need to follow TIF_NR_POLLING
1144 * - On most archs, a simple fetch_or on ti::flags with a
1145 * "0" value would be enough to know if an IPI needs to be sent.
1147 * - x86 needs to perform a last need_resched() check between
1148 * monitor and mwait which doesn't take timers into account.
1149 * There a dedicated TIF_TIMER flag would be required to
1150 * fetch_or here and be checked along with TIF_NEED_RESCHED
1153 * However, remote timer enqueue is not such a frequent event
1154 * and testing of the above solutions didn't appear to report
1157 if (set_nr_and_not_polling(rq->idle))
1158 smp_send_reschedule(cpu);
1160 trace_sched_wake_idle_without_ipi(cpu);
1163 static bool wake_up_full_nohz_cpu(int cpu)
1166 * We just need the target to call irq_exit() and re-evaluate
1167 * the next tick. The nohz full kick at least implies that.
1168 * If needed we can still optimize that later with an
1171 if (cpu_is_offline(cpu))
1172 return true; /* Don't try to wake offline CPUs. */
1173 if (tick_nohz_full_cpu(cpu)) {
1174 if (cpu != smp_processor_id() ||
1175 tick_nohz_tick_stopped())
1176 tick_nohz_full_kick_cpu(cpu);
1184 * Wake up the specified CPU. If the CPU is going offline, it is the
1185 * caller's responsibility to deal with the lost wakeup, for example,
1186 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1188 void wake_up_nohz_cpu(int cpu)
1190 if (!wake_up_full_nohz_cpu(cpu))
1191 wake_up_idle_cpu(cpu);
1194 static void nohz_csd_func(void *info)
1196 struct rq *rq = info;
1197 int cpu = cpu_of(rq);
1201 * Release the rq::nohz_csd.
1203 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1204 WARN_ON(!(flags & NOHZ_KICK_MASK));
1206 rq->idle_balance = idle_cpu(cpu);
1207 if (rq->idle_balance && !need_resched()) {
1208 rq->nohz_idle_balance = flags;
1209 raise_softirq_irqoff(SCHED_SOFTIRQ);
1213 #endif /* CONFIG_NO_HZ_COMMON */
1215 #ifdef CONFIG_NO_HZ_FULL
1216 static inline bool __need_bw_check(struct rq *rq, struct task_struct *p)
1218 if (rq->nr_running != 1)
1221 if (p->sched_class != &fair_sched_class)
1224 if (!task_on_rq_queued(p))
1230 bool sched_can_stop_tick(struct rq *rq)
1232 int fifo_nr_running;
1234 /* Deadline tasks, even if single, need the tick */
1235 if (rq->dl.dl_nr_running)
1239 * If there are more than one RR tasks, we need the tick to affect the
1240 * actual RR behaviour.
1242 if (rq->rt.rr_nr_running) {
1243 if (rq->rt.rr_nr_running == 1)
1250 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1251 * forced preemption between FIFO tasks.
1253 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1254 if (fifo_nr_running)
1258 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1259 * if there's more than one we need the tick for involuntary
1262 if (rq->nr_running > 1)
1266 * If there is one task and it has CFS runtime bandwidth constraints
1267 * and it's on the cpu now we don't want to stop the tick.
1268 * This check prevents clearing the bit if a newly enqueued task here is
1269 * dequeued by migrating while the constrained task continues to run.
1270 * E.g. going from 2->1 without going through pick_next_task().
1272 if (sched_feat(HZ_BW) && __need_bw_check(rq, rq->curr)) {
1273 if (cfs_task_bw_constrained(rq->curr))
1279 #endif /* CONFIG_NO_HZ_FULL */
1280 #endif /* CONFIG_SMP */
1282 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1283 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1285 * Iterate task_group tree rooted at *from, calling @down when first entering a
1286 * node and @up when leaving it for the final time.
1288 * Caller must hold rcu_lock or sufficient equivalent.
1290 int walk_tg_tree_from(struct task_group *from,
1291 tg_visitor down, tg_visitor up, void *data)
1293 struct task_group *parent, *child;
1299 ret = (*down)(parent, data);
1302 list_for_each_entry_rcu(child, &parent->children, siblings) {
1309 ret = (*up)(parent, data);
1310 if (ret || parent == from)
1314 parent = parent->parent;
1321 int tg_nop(struct task_group *tg, void *data)
1327 static void set_load_weight(struct task_struct *p, bool update_load)
1329 int prio = p->static_prio - MAX_RT_PRIO;
1330 struct load_weight *load = &p->se.load;
1333 * SCHED_IDLE tasks get minimal weight:
1335 if (task_has_idle_policy(p)) {
1336 load->weight = scale_load(WEIGHT_IDLEPRIO);
1337 load->inv_weight = WMULT_IDLEPRIO;
1342 * SCHED_OTHER tasks have to update their load when changing their
1345 if (update_load && p->sched_class == &fair_sched_class) {
1346 reweight_task(p, prio);
1348 load->weight = scale_load(sched_prio_to_weight[prio]);
1349 load->inv_weight = sched_prio_to_wmult[prio];
1353 #ifdef CONFIG_UCLAMP_TASK
1355 * Serializes updates of utilization clamp values
1357 * The (slow-path) user-space triggers utilization clamp value updates which
1358 * can require updates on (fast-path) scheduler's data structures used to
1359 * support enqueue/dequeue operations.
1360 * While the per-CPU rq lock protects fast-path update operations, user-space
1361 * requests are serialized using a mutex to reduce the risk of conflicting
1362 * updates or API abuses.
1364 static DEFINE_MUTEX(uclamp_mutex);
1366 /* Max allowed minimum utilization */
1367 static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1369 /* Max allowed maximum utilization */
1370 static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1373 * By default RT tasks run at the maximum performance point/capacity of the
1374 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1375 * SCHED_CAPACITY_SCALE.
1377 * This knob allows admins to change the default behavior when uclamp is being
1378 * used. In battery powered devices, particularly, running at the maximum
1379 * capacity and frequency will increase energy consumption and shorten the
1382 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1384 * This knob will not override the system default sched_util_clamp_min defined
1387 static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1389 /* All clamps are required to be less or equal than these values */
1390 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1393 * This static key is used to reduce the uclamp overhead in the fast path. It
1394 * primarily disables the call to uclamp_rq_{inc, dec}() in
1395 * enqueue/dequeue_task().
1397 * This allows users to continue to enable uclamp in their kernel config with
1398 * minimum uclamp overhead in the fast path.
1400 * As soon as userspace modifies any of the uclamp knobs, the static key is
1401 * enabled, since we have an actual users that make use of uclamp
1404 * The knobs that would enable this static key are:
1406 * * A task modifying its uclamp value with sched_setattr().
1407 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1408 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1410 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1412 /* Integer rounded range for each bucket */
1413 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1415 #define for_each_clamp_id(clamp_id) \
1416 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1418 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1420 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1423 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1425 if (clamp_id == UCLAMP_MIN)
1427 return SCHED_CAPACITY_SCALE;
1430 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1431 unsigned int value, bool user_defined)
1433 uc_se->value = value;
1434 uc_se->bucket_id = uclamp_bucket_id(value);
1435 uc_se->user_defined = user_defined;
1438 static inline unsigned int
1439 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1440 unsigned int clamp_value)
1443 * Avoid blocked utilization pushing up the frequency when we go
1444 * idle (which drops the max-clamp) by retaining the last known
1447 if (clamp_id == UCLAMP_MAX) {
1448 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1452 return uclamp_none(UCLAMP_MIN);
1455 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1456 unsigned int clamp_value)
1458 /* Reset max-clamp retention only on idle exit */
1459 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1462 uclamp_rq_set(rq, clamp_id, clamp_value);
1466 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1467 unsigned int clamp_value)
1469 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1470 int bucket_id = UCLAMP_BUCKETS - 1;
1473 * Since both min and max clamps are max aggregated, find the
1474 * top most bucket with tasks in.
1476 for ( ; bucket_id >= 0; bucket_id--) {
1477 if (!bucket[bucket_id].tasks)
1479 return bucket[bucket_id].value;
1482 /* No tasks -- default clamp values */
1483 return uclamp_idle_value(rq, clamp_id, clamp_value);
1486 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1488 unsigned int default_util_min;
1489 struct uclamp_se *uc_se;
1491 lockdep_assert_held(&p->pi_lock);
1493 uc_se = &p->uclamp_req[UCLAMP_MIN];
1495 /* Only sync if user didn't override the default */
1496 if (uc_se->user_defined)
1499 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1500 uclamp_se_set(uc_se, default_util_min, false);
1503 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1508 /* Protect updates to p->uclamp_* */
1509 guard(task_rq_lock)(p);
1510 __uclamp_update_util_min_rt_default(p);
1513 static inline struct uclamp_se
1514 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1516 /* Copy by value as we could modify it */
1517 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1518 #ifdef CONFIG_UCLAMP_TASK_GROUP
1519 unsigned int tg_min, tg_max, value;
1522 * Tasks in autogroups or root task group will be
1523 * restricted by system defaults.
1525 if (task_group_is_autogroup(task_group(p)))
1527 if (task_group(p) == &root_task_group)
1530 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1531 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1532 value = uc_req.value;
1533 value = clamp(value, tg_min, tg_max);
1534 uclamp_se_set(&uc_req, value, false);
1541 * The effective clamp bucket index of a task depends on, by increasing
1543 * - the task specific clamp value, when explicitly requested from userspace
1544 * - the task group effective clamp value, for tasks not either in the root
1545 * group or in an autogroup
1546 * - the system default clamp value, defined by the sysadmin
1548 static inline struct uclamp_se
1549 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1551 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1552 struct uclamp_se uc_max = uclamp_default[clamp_id];
1554 /* System default restrictions always apply */
1555 if (unlikely(uc_req.value > uc_max.value))
1561 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1563 struct uclamp_se uc_eff;
1565 /* Task currently refcounted: use back-annotated (effective) value */
1566 if (p->uclamp[clamp_id].active)
1567 return (unsigned long)p->uclamp[clamp_id].value;
1569 uc_eff = uclamp_eff_get(p, clamp_id);
1571 return (unsigned long)uc_eff.value;
1575 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1576 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1577 * updates the rq's clamp value if required.
1579 * Tasks can have a task-specific value requested from user-space, track
1580 * within each bucket the maximum value for tasks refcounted in it.
1581 * This "local max aggregation" allows to track the exact "requested" value
1582 * for each bucket when all its RUNNABLE tasks require the same clamp.
1584 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1585 enum uclamp_id clamp_id)
1587 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1588 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1589 struct uclamp_bucket *bucket;
1591 lockdep_assert_rq_held(rq);
1593 /* Update task effective clamp */
1594 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1596 bucket = &uc_rq->bucket[uc_se->bucket_id];
1598 uc_se->active = true;
1600 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1603 * Local max aggregation: rq buckets always track the max
1604 * "requested" clamp value of its RUNNABLE tasks.
1606 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1607 bucket->value = uc_se->value;
1609 if (uc_se->value > uclamp_rq_get(rq, clamp_id))
1610 uclamp_rq_set(rq, clamp_id, uc_se->value);
1614 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1615 * is released. If this is the last task reference counting the rq's max
1616 * active clamp value, then the rq's clamp value is updated.
1618 * Both refcounted tasks and rq's cached clamp values are expected to be
1619 * always valid. If it's detected they are not, as defensive programming,
1620 * enforce the expected state and warn.
1622 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1623 enum uclamp_id clamp_id)
1625 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1626 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1627 struct uclamp_bucket *bucket;
1628 unsigned int bkt_clamp;
1629 unsigned int rq_clamp;
1631 lockdep_assert_rq_held(rq);
1634 * If sched_uclamp_used was enabled after task @p was enqueued,
1635 * we could end up with unbalanced call to uclamp_rq_dec_id().
1637 * In this case the uc_se->active flag should be false since no uclamp
1638 * accounting was performed at enqueue time and we can just return
1641 * Need to be careful of the following enqueue/dequeue ordering
1645 * // sched_uclamp_used gets enabled
1648 * // Must not decrement bucket->tasks here
1651 * where we could end up with stale data in uc_se and
1652 * bucket[uc_se->bucket_id].
1654 * The following check here eliminates the possibility of such race.
1656 if (unlikely(!uc_se->active))
1659 bucket = &uc_rq->bucket[uc_se->bucket_id];
1661 SCHED_WARN_ON(!bucket->tasks);
1662 if (likely(bucket->tasks))
1665 uc_se->active = false;
1668 * Keep "local max aggregation" simple and accept to (possibly)
1669 * overboost some RUNNABLE tasks in the same bucket.
1670 * The rq clamp bucket value is reset to its base value whenever
1671 * there are no more RUNNABLE tasks refcounting it.
1673 if (likely(bucket->tasks))
1676 rq_clamp = uclamp_rq_get(rq, clamp_id);
1678 * Defensive programming: this should never happen. If it happens,
1679 * e.g. due to future modification, warn and fixup the expected value.
1681 SCHED_WARN_ON(bucket->value > rq_clamp);
1682 if (bucket->value >= rq_clamp) {
1683 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1684 uclamp_rq_set(rq, clamp_id, bkt_clamp);
1688 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1690 enum uclamp_id clamp_id;
1693 * Avoid any overhead until uclamp is actually used by the userspace.
1695 * The condition is constructed such that a NOP is generated when
1696 * sched_uclamp_used is disabled.
1698 if (!static_branch_unlikely(&sched_uclamp_used))
1701 if (unlikely(!p->sched_class->uclamp_enabled))
1704 for_each_clamp_id(clamp_id)
1705 uclamp_rq_inc_id(rq, p, clamp_id);
1707 /* Reset clamp idle holding when there is one RUNNABLE task */
1708 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1709 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1712 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1714 enum uclamp_id clamp_id;
1717 * Avoid any overhead until uclamp is actually used by the userspace.
1719 * The condition is constructed such that a NOP is generated when
1720 * sched_uclamp_used is disabled.
1722 if (!static_branch_unlikely(&sched_uclamp_used))
1725 if (unlikely(!p->sched_class->uclamp_enabled))
1728 for_each_clamp_id(clamp_id)
1729 uclamp_rq_dec_id(rq, p, clamp_id);
1732 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1733 enum uclamp_id clamp_id)
1735 if (!p->uclamp[clamp_id].active)
1738 uclamp_rq_dec_id(rq, p, clamp_id);
1739 uclamp_rq_inc_id(rq, p, clamp_id);
1742 * Make sure to clear the idle flag if we've transiently reached 0
1743 * active tasks on rq.
1745 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1746 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1750 uclamp_update_active(struct task_struct *p)
1752 enum uclamp_id clamp_id;
1757 * Lock the task and the rq where the task is (or was) queued.
1759 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1760 * price to pay to safely serialize util_{min,max} updates with
1761 * enqueues, dequeues and migration operations.
1762 * This is the same locking schema used by __set_cpus_allowed_ptr().
1764 rq = task_rq_lock(p, &rf);
1767 * Setting the clamp bucket is serialized by task_rq_lock().
1768 * If the task is not yet RUNNABLE and its task_struct is not
1769 * affecting a valid clamp bucket, the next time it's enqueued,
1770 * it will already see the updated clamp bucket value.
1772 for_each_clamp_id(clamp_id)
1773 uclamp_rq_reinc_id(rq, p, clamp_id);
1775 task_rq_unlock(rq, p, &rf);
1778 #ifdef CONFIG_UCLAMP_TASK_GROUP
1780 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1782 struct css_task_iter it;
1783 struct task_struct *p;
1785 css_task_iter_start(css, 0, &it);
1786 while ((p = css_task_iter_next(&it)))
1787 uclamp_update_active(p);
1788 css_task_iter_end(&it);
1791 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1794 #ifdef CONFIG_SYSCTL
1795 #ifdef CONFIG_UCLAMP_TASK_GROUP
1796 static void uclamp_update_root_tg(void)
1798 struct task_group *tg = &root_task_group;
1800 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1801 sysctl_sched_uclamp_util_min, false);
1802 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1803 sysctl_sched_uclamp_util_max, false);
1806 cpu_util_update_eff(&root_task_group.css);
1809 static void uclamp_update_root_tg(void) { }
1812 static void uclamp_sync_util_min_rt_default(void)
1814 struct task_struct *g, *p;
1817 * copy_process() sysctl_uclamp
1818 * uclamp_min_rt = X;
1819 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1820 * // link thread smp_mb__after_spinlock()
1821 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1822 * sched_post_fork() for_each_process_thread()
1823 * __uclamp_sync_rt() __uclamp_sync_rt()
1825 * Ensures that either sched_post_fork() will observe the new
1826 * uclamp_min_rt or for_each_process_thread() will observe the new
1829 read_lock(&tasklist_lock);
1830 smp_mb__after_spinlock();
1831 read_unlock(&tasklist_lock);
1834 for_each_process_thread(g, p)
1835 uclamp_update_util_min_rt_default(p);
1838 static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1839 void *buffer, size_t *lenp, loff_t *ppos)
1841 bool update_root_tg = false;
1842 int old_min, old_max, old_min_rt;
1845 guard(mutex)(&uclamp_mutex);
1847 old_min = sysctl_sched_uclamp_util_min;
1848 old_max = sysctl_sched_uclamp_util_max;
1849 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1851 result = proc_dointvec(table, write, buffer, lenp, ppos);
1857 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1858 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1859 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1865 if (old_min != sysctl_sched_uclamp_util_min) {
1866 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1867 sysctl_sched_uclamp_util_min, false);
1868 update_root_tg = true;
1870 if (old_max != sysctl_sched_uclamp_util_max) {
1871 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1872 sysctl_sched_uclamp_util_max, false);
1873 update_root_tg = true;
1876 if (update_root_tg) {
1877 static_branch_enable(&sched_uclamp_used);
1878 uclamp_update_root_tg();
1881 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1882 static_branch_enable(&sched_uclamp_used);
1883 uclamp_sync_util_min_rt_default();
1887 * We update all RUNNABLE tasks only when task groups are in use.
1888 * Otherwise, keep it simple and do just a lazy update at each next
1889 * task enqueue time.
1894 sysctl_sched_uclamp_util_min = old_min;
1895 sysctl_sched_uclamp_util_max = old_max;
1896 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1901 static int uclamp_validate(struct task_struct *p,
1902 const struct sched_attr *attr)
1904 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1905 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1907 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1908 util_min = attr->sched_util_min;
1910 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1914 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1915 util_max = attr->sched_util_max;
1917 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1921 if (util_min != -1 && util_max != -1 && util_min > util_max)
1925 * We have valid uclamp attributes; make sure uclamp is enabled.
1927 * We need to do that here, because enabling static branches is a
1928 * blocking operation which obviously cannot be done while holding
1931 static_branch_enable(&sched_uclamp_used);
1936 static bool uclamp_reset(const struct sched_attr *attr,
1937 enum uclamp_id clamp_id,
1938 struct uclamp_se *uc_se)
1940 /* Reset on sched class change for a non user-defined clamp value. */
1941 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1942 !uc_se->user_defined)
1945 /* Reset on sched_util_{min,max} == -1. */
1946 if (clamp_id == UCLAMP_MIN &&
1947 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1948 attr->sched_util_min == -1) {
1952 if (clamp_id == UCLAMP_MAX &&
1953 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1954 attr->sched_util_max == -1) {
1961 static void __setscheduler_uclamp(struct task_struct *p,
1962 const struct sched_attr *attr)
1964 enum uclamp_id clamp_id;
1966 for_each_clamp_id(clamp_id) {
1967 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1970 if (!uclamp_reset(attr, clamp_id, uc_se))
1974 * RT by default have a 100% boost value that could be modified
1977 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1978 value = sysctl_sched_uclamp_util_min_rt_default;
1980 value = uclamp_none(clamp_id);
1982 uclamp_se_set(uc_se, value, false);
1986 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1989 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1990 attr->sched_util_min != -1) {
1991 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1992 attr->sched_util_min, true);
1995 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1996 attr->sched_util_max != -1) {
1997 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1998 attr->sched_util_max, true);
2002 static void uclamp_fork(struct task_struct *p)
2004 enum uclamp_id clamp_id;
2007 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
2008 * as the task is still at its early fork stages.
2010 for_each_clamp_id(clamp_id)
2011 p->uclamp[clamp_id].active = false;
2013 if (likely(!p->sched_reset_on_fork))
2016 for_each_clamp_id(clamp_id) {
2017 uclamp_se_set(&p->uclamp_req[clamp_id],
2018 uclamp_none(clamp_id), false);
2022 static void uclamp_post_fork(struct task_struct *p)
2024 uclamp_update_util_min_rt_default(p);
2027 static void __init init_uclamp_rq(struct rq *rq)
2029 enum uclamp_id clamp_id;
2030 struct uclamp_rq *uc_rq = rq->uclamp;
2032 for_each_clamp_id(clamp_id) {
2033 uc_rq[clamp_id] = (struct uclamp_rq) {
2034 .value = uclamp_none(clamp_id)
2038 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
2041 static void __init init_uclamp(void)
2043 struct uclamp_se uc_max = {};
2044 enum uclamp_id clamp_id;
2047 for_each_possible_cpu(cpu)
2048 init_uclamp_rq(cpu_rq(cpu));
2050 for_each_clamp_id(clamp_id) {
2051 uclamp_se_set(&init_task.uclamp_req[clamp_id],
2052 uclamp_none(clamp_id), false);
2055 /* System defaults allow max clamp values for both indexes */
2056 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2057 for_each_clamp_id(clamp_id) {
2058 uclamp_default[clamp_id] = uc_max;
2059 #ifdef CONFIG_UCLAMP_TASK_GROUP
2060 root_task_group.uclamp_req[clamp_id] = uc_max;
2061 root_task_group.uclamp[clamp_id] = uc_max;
2066 #else /* !CONFIG_UCLAMP_TASK */
2067 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2068 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
2069 static inline int uclamp_validate(struct task_struct *p,
2070 const struct sched_attr *attr)
2074 static void __setscheduler_uclamp(struct task_struct *p,
2075 const struct sched_attr *attr) { }
2076 static inline void uclamp_fork(struct task_struct *p) { }
2077 static inline void uclamp_post_fork(struct task_struct *p) { }
2078 static inline void init_uclamp(void) { }
2079 #endif /* CONFIG_UCLAMP_TASK */
2081 bool sched_task_on_rq(struct task_struct *p)
2083 return task_on_rq_queued(p);
2086 unsigned long get_wchan(struct task_struct *p)
2088 unsigned long ip = 0;
2091 if (!p || p == current)
2094 /* Only get wchan if task is blocked and we can keep it that way. */
2095 raw_spin_lock_irq(&p->pi_lock);
2096 state = READ_ONCE(p->__state);
2097 smp_rmb(); /* see try_to_wake_up() */
2098 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2099 ip = __get_wchan(p);
2100 raw_spin_unlock_irq(&p->pi_lock);
2105 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2107 if (!(flags & ENQUEUE_NOCLOCK))
2108 update_rq_clock(rq);
2110 if (!(flags & ENQUEUE_RESTORE)) {
2111 sched_info_enqueue(rq, p);
2112 psi_enqueue(p, (flags & ENQUEUE_WAKEUP) && !(flags & ENQUEUE_MIGRATED));
2115 uclamp_rq_inc(rq, p);
2116 p->sched_class->enqueue_task(rq, p, flags);
2118 if (sched_core_enabled(rq))
2119 sched_core_enqueue(rq, p);
2122 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2124 if (sched_core_enabled(rq))
2125 sched_core_dequeue(rq, p, flags);
2127 if (!(flags & DEQUEUE_NOCLOCK))
2128 update_rq_clock(rq);
2130 if (!(flags & DEQUEUE_SAVE)) {
2131 sched_info_dequeue(rq, p);
2132 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2135 uclamp_rq_dec(rq, p);
2136 p->sched_class->dequeue_task(rq, p, flags);
2139 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2141 if (task_on_rq_migrating(p))
2142 flags |= ENQUEUE_MIGRATED;
2143 if (flags & ENQUEUE_MIGRATED)
2144 sched_mm_cid_migrate_to(rq, p);
2146 enqueue_task(rq, p, flags);
2148 WRITE_ONCE(p->on_rq, TASK_ON_RQ_QUEUED);
2149 ASSERT_EXCLUSIVE_WRITER(p->on_rq);
2152 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2154 WRITE_ONCE(p->on_rq, (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING);
2155 ASSERT_EXCLUSIVE_WRITER(p->on_rq);
2157 dequeue_task(rq, p, flags);
2160 static inline int __normal_prio(int policy, int rt_prio, int nice)
2164 if (dl_policy(policy))
2165 prio = MAX_DL_PRIO - 1;
2166 else if (rt_policy(policy))
2167 prio = MAX_RT_PRIO - 1 - rt_prio;
2169 prio = NICE_TO_PRIO(nice);
2175 * Calculate the expected normal priority: i.e. priority
2176 * without taking RT-inheritance into account. Might be
2177 * boosted by interactivity modifiers. Changes upon fork,
2178 * setprio syscalls, and whenever the interactivity
2179 * estimator recalculates.
2181 static inline int normal_prio(struct task_struct *p)
2183 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2187 * Calculate the current priority, i.e. the priority
2188 * taken into account by the scheduler. This value might
2189 * be boosted by RT tasks, or might be boosted by
2190 * interactivity modifiers. Will be RT if the task got
2191 * RT-boosted. If not then it returns p->normal_prio.
2193 static int effective_prio(struct task_struct *p)
2195 p->normal_prio = normal_prio(p);
2197 * If we are RT tasks or we were boosted to RT priority,
2198 * keep the priority unchanged. Otherwise, update priority
2199 * to the normal priority:
2201 if (!rt_prio(p->prio))
2202 return p->normal_prio;
2207 * task_curr - is this task currently executing on a CPU?
2208 * @p: the task in question.
2210 * Return: 1 if the task is currently executing. 0 otherwise.
2212 inline int task_curr(const struct task_struct *p)
2214 return cpu_curr(task_cpu(p)) == p;
2218 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2219 * use the balance_callback list if you want balancing.
2221 * this means any call to check_class_changed() must be followed by a call to
2222 * balance_callback().
2224 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2225 const struct sched_class *prev_class,
2228 if (prev_class != p->sched_class) {
2229 if (prev_class->switched_from)
2230 prev_class->switched_from(rq, p);
2232 p->sched_class->switched_to(rq, p);
2233 } else if (oldprio != p->prio || dl_task(p))
2234 p->sched_class->prio_changed(rq, p, oldprio);
2237 void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags)
2239 if (p->sched_class == rq->curr->sched_class)
2240 rq->curr->sched_class->wakeup_preempt(rq, p, flags);
2241 else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2245 * A queue event has occurred, and we're going to schedule. In
2246 * this case, we can save a useless back to back clock update.
2248 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2249 rq_clock_skip_update(rq);
2252 static __always_inline
2253 int __task_state_match(struct task_struct *p, unsigned int state)
2255 if (READ_ONCE(p->__state) & state)
2258 if (READ_ONCE(p->saved_state) & state)
2264 static __always_inline
2265 int task_state_match(struct task_struct *p, unsigned int state)
2268 * Serialize against current_save_and_set_rtlock_wait_state(),
2269 * current_restore_rtlock_saved_state(), and __refrigerator().
2271 guard(raw_spinlock_irq)(&p->pi_lock);
2272 return __task_state_match(p, state);
2276 * wait_task_inactive - wait for a thread to unschedule.
2278 * Wait for the thread to block in any of the states set in @match_state.
2279 * If it changes, i.e. @p might have woken up, then return zero. When we
2280 * succeed in waiting for @p to be off its CPU, we return a positive number
2281 * (its total switch count). If a second call a short while later returns the
2282 * same number, the caller can be sure that @p has remained unscheduled the
2285 * The caller must ensure that the task *will* unschedule sometime soon,
2286 * else this function might spin for a *long* time. This function can't
2287 * be called with interrupts off, or it may introduce deadlock with
2288 * smp_call_function() if an IPI is sent by the same process we are
2289 * waiting to become inactive.
2291 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
2293 int running, queued, match;
2300 * We do the initial early heuristics without holding
2301 * any task-queue locks at all. We'll only try to get
2302 * the runqueue lock when things look like they will
2308 * If the task is actively running on another CPU
2309 * still, just relax and busy-wait without holding
2312 * NOTE! Since we don't hold any locks, it's not
2313 * even sure that "rq" stays as the right runqueue!
2314 * But we don't care, since "task_on_cpu()" will
2315 * return false if the runqueue has changed and p
2316 * is actually now running somewhere else!
2318 while (task_on_cpu(rq, p)) {
2319 if (!task_state_match(p, match_state))
2325 * Ok, time to look more closely! We need the rq
2326 * lock now, to be *sure*. If we're wrong, we'll
2327 * just go back and repeat.
2329 rq = task_rq_lock(p, &rf);
2330 trace_sched_wait_task(p);
2331 running = task_on_cpu(rq, p);
2332 queued = task_on_rq_queued(p);
2334 if ((match = __task_state_match(p, match_state))) {
2336 * When matching on p->saved_state, consider this task
2337 * still queued so it will wait.
2341 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2343 task_rq_unlock(rq, p, &rf);
2346 * If it changed from the expected state, bail out now.
2348 if (unlikely(!ncsw))
2352 * Was it really running after all now that we
2353 * checked with the proper locks actually held?
2355 * Oops. Go back and try again..
2357 if (unlikely(running)) {
2363 * It's not enough that it's not actively running,
2364 * it must be off the runqueue _entirely_, and not
2367 * So if it was still runnable (but just not actively
2368 * running right now), it's preempted, and we should
2369 * yield - it could be a while.
2371 if (unlikely(queued)) {
2372 ktime_t to = NSEC_PER_SEC / HZ;
2374 set_current_state(TASK_UNINTERRUPTIBLE);
2375 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
2380 * Ahh, all good. It wasn't running, and it wasn't
2381 * runnable, which means that it will never become
2382 * running in the future either. We're all done!
2393 __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
2395 static int __set_cpus_allowed_ptr(struct task_struct *p,
2396 struct affinity_context *ctx);
2398 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2400 struct affinity_context ac = {
2401 .new_mask = cpumask_of(rq->cpu),
2402 .flags = SCA_MIGRATE_DISABLE,
2405 if (likely(!p->migration_disabled))
2408 if (p->cpus_ptr != &p->cpus_mask)
2412 * Violates locking rules! see comment in __do_set_cpus_allowed().
2414 __do_set_cpus_allowed(p, &ac);
2417 void migrate_disable(void)
2419 struct task_struct *p = current;
2421 if (p->migration_disabled) {
2422 p->migration_disabled++;
2427 this_rq()->nr_pinned++;
2428 p->migration_disabled = 1;
2430 EXPORT_SYMBOL_GPL(migrate_disable);
2432 void migrate_enable(void)
2434 struct task_struct *p = current;
2435 struct affinity_context ac = {
2436 .new_mask = &p->cpus_mask,
2437 .flags = SCA_MIGRATE_ENABLE,
2440 if (p->migration_disabled > 1) {
2441 p->migration_disabled--;
2445 if (WARN_ON_ONCE(!p->migration_disabled))
2449 * Ensure stop_task runs either before or after this, and that
2450 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2453 if (p->cpus_ptr != &p->cpus_mask)
2454 __set_cpus_allowed_ptr(p, &ac);
2456 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2457 * regular cpus_mask, otherwise things that race (eg.
2458 * select_fallback_rq) get confused.
2461 p->migration_disabled = 0;
2462 this_rq()->nr_pinned--;
2464 EXPORT_SYMBOL_GPL(migrate_enable);
2466 static inline bool rq_has_pinned_tasks(struct rq *rq)
2468 return rq->nr_pinned;
2472 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2473 * __set_cpus_allowed_ptr() and select_fallback_rq().
2475 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2477 /* When not in the task's cpumask, no point in looking further. */
2478 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2481 /* migrate_disabled() must be allowed to finish. */
2482 if (is_migration_disabled(p))
2483 return cpu_online(cpu);
2485 /* Non kernel threads are not allowed during either online or offline. */
2486 if (!(p->flags & PF_KTHREAD))
2487 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2489 /* KTHREAD_IS_PER_CPU is always allowed. */
2490 if (kthread_is_per_cpu(p))
2491 return cpu_online(cpu);
2493 /* Regular kernel threads don't get to stay during offline. */
2497 /* But are allowed during online. */
2498 return cpu_online(cpu);
2502 * This is how migration works:
2504 * 1) we invoke migration_cpu_stop() on the target CPU using
2506 * 2) stopper starts to run (implicitly forcing the migrated thread
2508 * 3) it checks whether the migrated task is still in the wrong runqueue.
2509 * 4) if it's in the wrong runqueue then the migration thread removes
2510 * it and puts it into the right queue.
2511 * 5) stopper completes and stop_one_cpu() returns and the migration
2516 * move_queued_task - move a queued task to new rq.
2518 * Returns (locked) new rq. Old rq's lock is released.
2520 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2521 struct task_struct *p, int new_cpu)
2523 lockdep_assert_rq_held(rq);
2525 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2526 set_task_cpu(p, new_cpu);
2529 rq = cpu_rq(new_cpu);
2532 WARN_ON_ONCE(task_cpu(p) != new_cpu);
2533 activate_task(rq, p, 0);
2534 wakeup_preempt(rq, p, 0);
2539 struct migration_arg {
2540 struct task_struct *task;
2542 struct set_affinity_pending *pending;
2546 * @refs: number of wait_for_completion()
2547 * @stop_pending: is @stop_work in use
2549 struct set_affinity_pending {
2551 unsigned int stop_pending;
2552 struct completion done;
2553 struct cpu_stop_work stop_work;
2554 struct migration_arg arg;
2558 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2559 * this because either it can't run here any more (set_cpus_allowed()
2560 * away from this CPU, or CPU going down), or because we're
2561 * attempting to rebalance this task on exec (sched_exec).
2563 * So we race with normal scheduler movements, but that's OK, as long
2564 * as the task is no longer on this CPU.
2566 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2567 struct task_struct *p, int dest_cpu)
2569 /* Affinity changed (again). */
2570 if (!is_cpu_allowed(p, dest_cpu))
2573 rq = move_queued_task(rq, rf, p, dest_cpu);
2579 * migration_cpu_stop - this will be executed by a highprio stopper thread
2580 * and performs thread migration by bumping thread off CPU then
2581 * 'pushing' onto another runqueue.
2583 static int migration_cpu_stop(void *data)
2585 struct migration_arg *arg = data;
2586 struct set_affinity_pending *pending = arg->pending;
2587 struct task_struct *p = arg->task;
2588 struct rq *rq = this_rq();
2589 bool complete = false;
2593 * The original target CPU might have gone down and we might
2594 * be on another CPU but it doesn't matter.
2596 local_irq_save(rf.flags);
2598 * We need to explicitly wake pending tasks before running
2599 * __migrate_task() such that we will not miss enforcing cpus_ptr
2600 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2602 flush_smp_call_function_queue();
2604 raw_spin_lock(&p->pi_lock);
2608 * If we were passed a pending, then ->stop_pending was set, thus
2609 * p->migration_pending must have remained stable.
2611 WARN_ON_ONCE(pending && pending != p->migration_pending);
2614 * If task_rq(p) != rq, it cannot be migrated here, because we're
2615 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2616 * we're holding p->pi_lock.
2618 if (task_rq(p) == rq) {
2619 if (is_migration_disabled(p))
2623 p->migration_pending = NULL;
2626 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2630 if (task_on_rq_queued(p)) {
2631 update_rq_clock(rq);
2632 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2634 p->wake_cpu = arg->dest_cpu;
2638 * XXX __migrate_task() can fail, at which point we might end
2639 * up running on a dodgy CPU, AFAICT this can only happen
2640 * during CPU hotplug, at which point we'll get pushed out
2641 * anyway, so it's probably not a big deal.
2644 } else if (pending) {
2646 * This happens when we get migrated between migrate_enable()'s
2647 * preempt_enable() and scheduling the stopper task. At that
2648 * point we're a regular task again and not current anymore.
2650 * A !PREEMPT kernel has a giant hole here, which makes it far
2655 * The task moved before the stopper got to run. We're holding
2656 * ->pi_lock, so the allowed mask is stable - if it got
2657 * somewhere allowed, we're done.
2659 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2660 p->migration_pending = NULL;
2666 * When migrate_enable() hits a rq mis-match we can't reliably
2667 * determine is_migration_disabled() and so have to chase after
2670 WARN_ON_ONCE(!pending->stop_pending);
2672 task_rq_unlock(rq, p, &rf);
2673 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2674 &pending->arg, &pending->stop_work);
2680 pending->stop_pending = false;
2681 task_rq_unlock(rq, p, &rf);
2684 complete_all(&pending->done);
2689 int push_cpu_stop(void *arg)
2691 struct rq *lowest_rq = NULL, *rq = this_rq();
2692 struct task_struct *p = arg;
2694 raw_spin_lock_irq(&p->pi_lock);
2695 raw_spin_rq_lock(rq);
2697 if (task_rq(p) != rq)
2700 if (is_migration_disabled(p)) {
2701 p->migration_flags |= MDF_PUSH;
2705 p->migration_flags &= ~MDF_PUSH;
2707 if (p->sched_class->find_lock_rq)
2708 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2713 // XXX validate p is still the highest prio task
2714 if (task_rq(p) == rq) {
2715 deactivate_task(rq, p, 0);
2716 set_task_cpu(p, lowest_rq->cpu);
2717 activate_task(lowest_rq, p, 0);
2718 resched_curr(lowest_rq);
2721 double_unlock_balance(rq, lowest_rq);
2724 rq->push_busy = false;
2725 raw_spin_rq_unlock(rq);
2726 raw_spin_unlock_irq(&p->pi_lock);
2733 * sched_class::set_cpus_allowed must do the below, but is not required to
2734 * actually call this function.
2736 void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
2738 if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2739 p->cpus_ptr = ctx->new_mask;
2743 cpumask_copy(&p->cpus_mask, ctx->new_mask);
2744 p->nr_cpus_allowed = cpumask_weight(ctx->new_mask);
2747 * Swap in a new user_cpus_ptr if SCA_USER flag set
2749 if (ctx->flags & SCA_USER)
2750 swap(p->user_cpus_ptr, ctx->user_mask);
2754 __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
2756 struct rq *rq = task_rq(p);
2757 bool queued, running;
2760 * This here violates the locking rules for affinity, since we're only
2761 * supposed to change these variables while holding both rq->lock and
2764 * HOWEVER, it magically works, because ttwu() is the only code that
2765 * accesses these variables under p->pi_lock and only does so after
2766 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2767 * before finish_task().
2769 * XXX do further audits, this smells like something putrid.
2771 if (ctx->flags & SCA_MIGRATE_DISABLE)
2772 SCHED_WARN_ON(!p->on_cpu);
2774 lockdep_assert_held(&p->pi_lock);
2776 queued = task_on_rq_queued(p);
2777 running = task_current(rq, p);
2781 * Because __kthread_bind() calls this on blocked tasks without
2784 lockdep_assert_rq_held(rq);
2785 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2788 put_prev_task(rq, p);
2790 p->sched_class->set_cpus_allowed(p, ctx);
2793 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2795 set_next_task(rq, p);
2799 * Used for kthread_bind() and select_fallback_rq(), in both cases the user
2800 * affinity (if any) should be destroyed too.
2802 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2804 struct affinity_context ac = {
2805 .new_mask = new_mask,
2807 .flags = SCA_USER, /* clear the user requested mask */
2809 union cpumask_rcuhead {
2811 struct rcu_head rcu;
2814 __do_set_cpus_allowed(p, &ac);
2817 * Because this is called with p->pi_lock held, it is not possible
2818 * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
2821 kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
2824 static cpumask_t *alloc_user_cpus_ptr(int node)
2827 * See do_set_cpus_allowed() above for the rcu_head usage.
2829 int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
2831 return kmalloc_node(size, GFP_KERNEL, node);
2834 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2837 cpumask_t *user_mask;
2838 unsigned long flags;
2841 * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2842 * may differ by now due to racing.
2844 dst->user_cpus_ptr = NULL;
2847 * This check is racy and losing the race is a valid situation.
2848 * It is not worth the extra overhead of taking the pi_lock on
2851 if (data_race(!src->user_cpus_ptr))
2854 user_mask = alloc_user_cpus_ptr(node);
2859 * Use pi_lock to protect content of user_cpus_ptr
2861 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2862 * do_set_cpus_allowed().
2864 raw_spin_lock_irqsave(&src->pi_lock, flags);
2865 if (src->user_cpus_ptr) {
2866 swap(dst->user_cpus_ptr, user_mask);
2867 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2869 raw_spin_unlock_irqrestore(&src->pi_lock, flags);
2871 if (unlikely(user_mask))
2877 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2879 struct cpumask *user_mask = NULL;
2881 swap(p->user_cpus_ptr, user_mask);
2886 void release_user_cpus_ptr(struct task_struct *p)
2888 kfree(clear_user_cpus_ptr(p));
2892 * This function is wildly self concurrent; here be dragons.
2895 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2896 * designated task is enqueued on an allowed CPU. If that task is currently
2897 * running, we have to kick it out using the CPU stopper.
2899 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2902 * Initial conditions: P0->cpus_mask = [0, 1]
2906 * migrate_disable();
2908 * set_cpus_allowed_ptr(P0, [1]);
2910 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2911 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2912 * This means we need the following scheme:
2916 * migrate_disable();
2918 * set_cpus_allowed_ptr(P0, [1]);
2922 * __set_cpus_allowed_ptr();
2923 * <wakes local stopper>
2924 * `--> <woken on migration completion>
2926 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2927 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2928 * task p are serialized by p->pi_lock, which we can leverage: the one that
2929 * should come into effect at the end of the Migrate-Disable region is the last
2930 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2931 * but we still need to properly signal those waiting tasks at the appropriate
2934 * This is implemented using struct set_affinity_pending. The first
2935 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2936 * setup an instance of that struct and install it on the targeted task_struct.
2937 * Any and all further callers will reuse that instance. Those then wait for
2938 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2939 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2942 * (1) In the cases covered above. There is one more where the completion is
2943 * signaled within affine_move_task() itself: when a subsequent affinity request
2944 * occurs after the stopper bailed out due to the targeted task still being
2945 * Migrate-Disable. Consider:
2947 * Initial conditions: P0->cpus_mask = [0, 1]
2951 * migrate_disable();
2953 * set_cpus_allowed_ptr(P0, [1]);
2956 * migration_cpu_stop()
2957 * is_migration_disabled()
2959 * set_cpus_allowed_ptr(P0, [0, 1]);
2960 * <signal completion>
2963 * Note that the above is safe vs a concurrent migrate_enable(), as any
2964 * pending affinity completion is preceded by an uninstallation of
2965 * p->migration_pending done with p->pi_lock held.
2967 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2968 int dest_cpu, unsigned int flags)
2969 __releases(rq->lock)
2970 __releases(p->pi_lock)
2972 struct set_affinity_pending my_pending = { }, *pending = NULL;
2973 bool stop_pending, complete = false;
2975 /* Can the task run on the task's current CPU? If so, we're done */
2976 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2977 struct task_struct *push_task = NULL;
2979 if ((flags & SCA_MIGRATE_ENABLE) &&
2980 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2981 rq->push_busy = true;
2982 push_task = get_task_struct(p);
2986 * If there are pending waiters, but no pending stop_work,
2987 * then complete now.
2989 pending = p->migration_pending;
2990 if (pending && !pending->stop_pending) {
2991 p->migration_pending = NULL;
2996 task_rq_unlock(rq, p, rf);
2998 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
3004 complete_all(&pending->done);
3009 if (!(flags & SCA_MIGRATE_ENABLE)) {
3010 /* serialized by p->pi_lock */
3011 if (!p->migration_pending) {
3012 /* Install the request */
3013 refcount_set(&my_pending.refs, 1);
3014 init_completion(&my_pending.done);
3015 my_pending.arg = (struct migration_arg) {
3017 .dest_cpu = dest_cpu,
3018 .pending = &my_pending,
3021 p->migration_pending = &my_pending;
3023 pending = p->migration_pending;
3024 refcount_inc(&pending->refs);
3026 * Affinity has changed, but we've already installed a
3027 * pending. migration_cpu_stop() *must* see this, else
3028 * we risk a completion of the pending despite having a
3029 * task on a disallowed CPU.
3031 * Serialized by p->pi_lock, so this is safe.
3033 pending->arg.dest_cpu = dest_cpu;
3036 pending = p->migration_pending;
3038 * - !MIGRATE_ENABLE:
3039 * we'll have installed a pending if there wasn't one already.
3042 * we're here because the current CPU isn't matching anymore,
3043 * the only way that can happen is because of a concurrent
3044 * set_cpus_allowed_ptr() call, which should then still be
3045 * pending completion.
3047 * Either way, we really should have a @pending here.
3049 if (WARN_ON_ONCE(!pending)) {
3050 task_rq_unlock(rq, p, rf);
3054 if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
3056 * MIGRATE_ENABLE gets here because 'p == current', but for
3057 * anything else we cannot do is_migration_disabled(), punt
3058 * and have the stopper function handle it all race-free.
3060 stop_pending = pending->stop_pending;
3062 pending->stop_pending = true;
3064 if (flags & SCA_MIGRATE_ENABLE)
3065 p->migration_flags &= ~MDF_PUSH;
3068 task_rq_unlock(rq, p, rf);
3069 if (!stop_pending) {
3070 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
3071 &pending->arg, &pending->stop_work);
3075 if (flags & SCA_MIGRATE_ENABLE)
3079 if (!is_migration_disabled(p)) {
3080 if (task_on_rq_queued(p))
3081 rq = move_queued_task(rq, rf, p, dest_cpu);
3083 if (!pending->stop_pending) {
3084 p->migration_pending = NULL;
3088 task_rq_unlock(rq, p, rf);
3091 complete_all(&pending->done);
3094 wait_for_completion(&pending->done);
3096 if (refcount_dec_and_test(&pending->refs))
3097 wake_up_var(&pending->refs); /* No UaF, just an address */
3100 * Block the original owner of &pending until all subsequent callers
3101 * have seen the completion and decremented the refcount
3103 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
3106 WARN_ON_ONCE(my_pending.stop_pending);
3112 * Called with both p->pi_lock and rq->lock held; drops both before returning.
3114 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
3115 struct affinity_context *ctx,
3117 struct rq_flags *rf)
3118 __releases(rq->lock)
3119 __releases(p->pi_lock)
3121 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
3122 const struct cpumask *cpu_valid_mask = cpu_active_mask;
3123 bool kthread = p->flags & PF_KTHREAD;
3124 unsigned int dest_cpu;
3127 update_rq_clock(rq);
3129 if (kthread || is_migration_disabled(p)) {
3131 * Kernel threads are allowed on online && !active CPUs,
3132 * however, during cpu-hot-unplug, even these might get pushed
3133 * away if not KTHREAD_IS_PER_CPU.
3135 * Specifically, migration_disabled() tasks must not fail the
3136 * cpumask_any_and_distribute() pick below, esp. so on
3137 * SCA_MIGRATE_ENABLE, otherwise we'll not call
3138 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
3140 cpu_valid_mask = cpu_online_mask;
3143 if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) {
3149 * Must re-check here, to close a race against __kthread_bind(),
3150 * sched_setaffinity() is not guaranteed to observe the flag.
3152 if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
3157 if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
3158 if (cpumask_equal(&p->cpus_mask, ctx->new_mask)) {
3159 if (ctx->flags & SCA_USER)
3160 swap(p->user_cpus_ptr, ctx->user_mask);
3164 if (WARN_ON_ONCE(p == current &&
3165 is_migration_disabled(p) &&
3166 !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
3173 * Picking a ~random cpu helps in cases where we are changing affinity
3174 * for groups of tasks (ie. cpuset), so that load balancing is not
3175 * immediately required to distribute the tasks within their new mask.
3177 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
3178 if (dest_cpu >= nr_cpu_ids) {
3183 __do_set_cpus_allowed(p, ctx);
3185 return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
3188 task_rq_unlock(rq, p, rf);
3194 * Change a given task's CPU affinity. Migrate the thread to a
3195 * proper CPU and schedule it away if the CPU it's executing on
3196 * is removed from the allowed bitmask.
3198 * NOTE: the caller must have a valid reference to the task, the
3199 * task must not exit() & deallocate itself prematurely. The
3200 * call is not atomic; no spinlocks may be held.
3202 static int __set_cpus_allowed_ptr(struct task_struct *p,
3203 struct affinity_context *ctx)
3208 rq = task_rq_lock(p, &rf);
3210 * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
3213 if (p->user_cpus_ptr &&
3214 !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
3215 cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr))
3216 ctx->new_mask = rq->scratch_mask;
3218 return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf);
3221 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3223 struct affinity_context ac = {
3224 .new_mask = new_mask,
3228 return __set_cpus_allowed_ptr(p, &ac);
3230 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
3233 * Change a given task's CPU affinity to the intersection of its current
3234 * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
3235 * If user_cpus_ptr is defined, use it as the basis for restricting CPU
3236 * affinity or use cpu_online_mask instead.
3238 * If the resulting mask is empty, leave the affinity unchanged and return
3241 static int restrict_cpus_allowed_ptr(struct task_struct *p,
3242 struct cpumask *new_mask,
3243 const struct cpumask *subset_mask)
3245 struct affinity_context ac = {
3246 .new_mask = new_mask,
3253 rq = task_rq_lock(p, &rf);
3256 * Forcefully restricting the affinity of a deadline task is
3257 * likely to cause problems, so fail and noisily override the
3260 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
3265 if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) {
3270 return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf);
3273 task_rq_unlock(rq, p, &rf);
3278 * Restrict the CPU affinity of task @p so that it is a subset of
3279 * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
3280 * old affinity mask. If the resulting mask is empty, we warn and walk
3281 * up the cpuset hierarchy until we find a suitable mask.
3283 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3285 cpumask_var_t new_mask;
3286 const struct cpumask *override_mask = task_cpu_possible_mask(p);
3288 alloc_cpumask_var(&new_mask, GFP_KERNEL);
3291 * __migrate_task() can fail silently in the face of concurrent
3292 * offlining of the chosen destination CPU, so take the hotplug
3293 * lock to ensure that the migration succeeds.
3296 if (!cpumask_available(new_mask))
3299 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3303 * We failed to find a valid subset of the affinity mask for the
3304 * task, so override it based on its cpuset hierarchy.
3306 cpuset_cpus_allowed(p, new_mask);
3307 override_mask = new_mask;
3310 if (printk_ratelimit()) {
3311 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3312 task_pid_nr(p), p->comm,
3313 cpumask_pr_args(override_mask));
3316 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3319 free_cpumask_var(new_mask);
3323 __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
3326 * Restore the affinity of a task @p which was previously restricted by a
3327 * call to force_compatible_cpus_allowed_ptr().
3329 * It is the caller's responsibility to serialise this with any calls to
3330 * force_compatible_cpus_allowed_ptr(@p).
3332 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3334 struct affinity_context ac = {
3335 .new_mask = task_user_cpus(p),
3341 * Try to restore the old affinity mask with __sched_setaffinity().
3342 * Cpuset masking will be done there too.
3344 ret = __sched_setaffinity(p, &ac);
3348 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3350 #ifdef CONFIG_SCHED_DEBUG
3351 unsigned int state = READ_ONCE(p->__state);
3354 * We should never call set_task_cpu() on a blocked task,
3355 * ttwu() will sort out the placement.
3357 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3360 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3361 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3362 * time relying on p->on_rq.
3364 WARN_ON_ONCE(state == TASK_RUNNING &&
3365 p->sched_class == &fair_sched_class &&
3366 (p->on_rq && !task_on_rq_migrating(p)));
3368 #ifdef CONFIG_LOCKDEP
3370 * The caller should hold either p->pi_lock or rq->lock, when changing
3371 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3373 * sched_move_task() holds both and thus holding either pins the cgroup,
3376 * Furthermore, all task_rq users should acquire both locks, see
3379 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3380 lockdep_is_held(__rq_lockp(task_rq(p)))));
3383 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3385 WARN_ON_ONCE(!cpu_online(new_cpu));
3387 WARN_ON_ONCE(is_migration_disabled(p));
3390 trace_sched_migrate_task(p, new_cpu);
3392 if (task_cpu(p) != new_cpu) {
3393 if (p->sched_class->migrate_task_rq)
3394 p->sched_class->migrate_task_rq(p, new_cpu);
3395 p->se.nr_migrations++;
3397 sched_mm_cid_migrate_from(p);
3398 perf_event_task_migrate(p);
3401 __set_task_cpu(p, new_cpu);
3404 #ifdef CONFIG_NUMA_BALANCING
3405 static void __migrate_swap_task(struct task_struct *p, int cpu)
3407 if (task_on_rq_queued(p)) {
3408 struct rq *src_rq, *dst_rq;
3409 struct rq_flags srf, drf;
3411 src_rq = task_rq(p);
3412 dst_rq = cpu_rq(cpu);
3414 rq_pin_lock(src_rq, &srf);
3415 rq_pin_lock(dst_rq, &drf);
3417 deactivate_task(src_rq, p, 0);
3418 set_task_cpu(p, cpu);
3419 activate_task(dst_rq, p, 0);
3420 wakeup_preempt(dst_rq, p, 0);
3422 rq_unpin_lock(dst_rq, &drf);
3423 rq_unpin_lock(src_rq, &srf);
3427 * Task isn't running anymore; make it appear like we migrated
3428 * it before it went to sleep. This means on wakeup we make the
3429 * previous CPU our target instead of where it really is.
3435 struct migration_swap_arg {
3436 struct task_struct *src_task, *dst_task;
3437 int src_cpu, dst_cpu;
3440 static int migrate_swap_stop(void *data)
3442 struct migration_swap_arg *arg = data;
3443 struct rq *src_rq, *dst_rq;
3445 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3448 src_rq = cpu_rq(arg->src_cpu);
3449 dst_rq = cpu_rq(arg->dst_cpu);
3451 guard(double_raw_spinlock)(&arg->src_task->pi_lock, &arg->dst_task->pi_lock);
3452 guard(double_rq_lock)(src_rq, dst_rq);
3454 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3457 if (task_cpu(arg->src_task) != arg->src_cpu)
3460 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3463 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3466 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3467 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3473 * Cross migrate two tasks
3475 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3476 int target_cpu, int curr_cpu)
3478 struct migration_swap_arg arg;
3481 arg = (struct migration_swap_arg){
3483 .src_cpu = curr_cpu,
3485 .dst_cpu = target_cpu,
3488 if (arg.src_cpu == arg.dst_cpu)
3492 * These three tests are all lockless; this is OK since all of them
3493 * will be re-checked with proper locks held further down the line.
3495 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3498 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3501 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3504 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3505 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3510 #endif /* CONFIG_NUMA_BALANCING */
3513 * kick_process - kick a running thread to enter/exit the kernel
3514 * @p: the to-be-kicked thread
3516 * Cause a process which is running on another CPU to enter
3517 * kernel-mode, without any delay. (to get signals handled.)
3519 * NOTE: this function doesn't have to take the runqueue lock,
3520 * because all it wants to ensure is that the remote task enters
3521 * the kernel. If the IPI races and the task has been migrated
3522 * to another CPU then no harm is done and the purpose has been
3525 void kick_process(struct task_struct *p)
3528 int cpu = task_cpu(p);
3530 if ((cpu != smp_processor_id()) && task_curr(p))
3531 smp_send_reschedule(cpu);
3533 EXPORT_SYMBOL_GPL(kick_process);
3536 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3538 * A few notes on cpu_active vs cpu_online:
3540 * - cpu_active must be a subset of cpu_online
3542 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3543 * see __set_cpus_allowed_ptr(). At this point the newly online
3544 * CPU isn't yet part of the sched domains, and balancing will not
3547 * - on CPU-down we clear cpu_active() to mask the sched domains and
3548 * avoid the load balancer to place new tasks on the to be removed
3549 * CPU. Existing tasks will remain running there and will be taken
3552 * This means that fallback selection must not select !active CPUs.
3553 * And can assume that any active CPU must be online. Conversely
3554 * select_task_rq() below may allow selection of !active CPUs in order
3555 * to satisfy the above rules.
3557 static int select_fallback_rq(int cpu, struct task_struct *p)
3559 int nid = cpu_to_node(cpu);
3560 const struct cpumask *nodemask = NULL;
3561 enum { cpuset, possible, fail } state = cpuset;
3565 * If the node that the CPU is on has been offlined, cpu_to_node()
3566 * will return -1. There is no CPU on the node, and we should
3567 * select the CPU on the other node.
3570 nodemask = cpumask_of_node(nid);
3572 /* Look for allowed, online CPU in same node. */
3573 for_each_cpu(dest_cpu, nodemask) {
3574 if (is_cpu_allowed(p, dest_cpu))
3580 /* Any allowed, online CPU? */
3581 for_each_cpu(dest_cpu, p->cpus_ptr) {
3582 if (!is_cpu_allowed(p, dest_cpu))
3588 /* No more Mr. Nice Guy. */
3591 if (cpuset_cpus_allowed_fallback(p)) {
3598 * XXX When called from select_task_rq() we only
3599 * hold p->pi_lock and again violate locking order.
3601 * More yuck to audit.
3603 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3613 if (state != cpuset) {
3615 * Don't tell them about moving exiting tasks or
3616 * kernel threads (both mm NULL), since they never
3619 if (p->mm && printk_ratelimit()) {
3620 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3621 task_pid_nr(p), p->comm, cpu);
3629 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3632 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3634 lockdep_assert_held(&p->pi_lock);
3636 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3637 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3639 cpu = cpumask_any(p->cpus_ptr);
3642 * In order not to call set_task_cpu() on a blocking task we need
3643 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3646 * Since this is common to all placement strategies, this lives here.
3648 * [ this allows ->select_task() to simply return task_cpu(p) and
3649 * not worry about this generic constraint ]
3651 if (unlikely(!is_cpu_allowed(p, cpu)))
3652 cpu = select_fallback_rq(task_cpu(p), p);
3657 void sched_set_stop_task(int cpu, struct task_struct *stop)
3659 static struct lock_class_key stop_pi_lock;
3660 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3661 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3665 * Make it appear like a SCHED_FIFO task, its something
3666 * userspace knows about and won't get confused about.
3668 * Also, it will make PI more or less work without too
3669 * much confusion -- but then, stop work should not
3670 * rely on PI working anyway.
3672 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3674 stop->sched_class = &stop_sched_class;
3677 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3678 * adjust the effective priority of a task. As a result,
3679 * rt_mutex_setprio() can trigger (RT) balancing operations,
3680 * which can then trigger wakeups of the stop thread to push
3681 * around the current task.
3683 * The stop task itself will never be part of the PI-chain, it
3684 * never blocks, therefore that ->pi_lock recursion is safe.
3685 * Tell lockdep about this by placing the stop->pi_lock in its
3688 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3691 cpu_rq(cpu)->stop = stop;
3695 * Reset it back to a normal scheduling class so that
3696 * it can die in pieces.
3698 old_stop->sched_class = &rt_sched_class;
3702 #else /* CONFIG_SMP */
3704 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3705 struct affinity_context *ctx)
3707 return set_cpus_allowed_ptr(p, ctx->new_mask);
3710 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3712 static inline bool rq_has_pinned_tasks(struct rq *rq)
3717 static inline cpumask_t *alloc_user_cpus_ptr(int node)
3722 #endif /* !CONFIG_SMP */
3725 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3729 if (!schedstat_enabled())
3735 if (cpu == rq->cpu) {
3736 __schedstat_inc(rq->ttwu_local);
3737 __schedstat_inc(p->stats.nr_wakeups_local);
3739 struct sched_domain *sd;
3741 __schedstat_inc(p->stats.nr_wakeups_remote);
3744 for_each_domain(rq->cpu, sd) {
3745 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3746 __schedstat_inc(sd->ttwu_wake_remote);
3752 if (wake_flags & WF_MIGRATED)
3753 __schedstat_inc(p->stats.nr_wakeups_migrate);
3754 #endif /* CONFIG_SMP */
3756 __schedstat_inc(rq->ttwu_count);
3757 __schedstat_inc(p->stats.nr_wakeups);
3759 if (wake_flags & WF_SYNC)
3760 __schedstat_inc(p->stats.nr_wakeups_sync);
3764 * Mark the task runnable.
3766 static inline void ttwu_do_wakeup(struct task_struct *p)
3768 WRITE_ONCE(p->__state, TASK_RUNNING);
3769 trace_sched_wakeup(p);
3773 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3774 struct rq_flags *rf)
3776 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3778 lockdep_assert_rq_held(rq);
3780 if (p->sched_contributes_to_load)
3781 rq->nr_uninterruptible--;
3784 if (wake_flags & WF_MIGRATED)
3785 en_flags |= ENQUEUE_MIGRATED;
3789 delayacct_blkio_end(p);
3790 atomic_dec(&task_rq(p)->nr_iowait);
3793 activate_task(rq, p, en_flags);
3794 wakeup_preempt(rq, p, wake_flags);
3799 if (p->sched_class->task_woken) {
3801 * Our task @p is fully woken up and running; so it's safe to
3802 * drop the rq->lock, hereafter rq is only used for statistics.
3804 rq_unpin_lock(rq, rf);
3805 p->sched_class->task_woken(rq, p);
3806 rq_repin_lock(rq, rf);
3809 if (rq->idle_stamp) {
3810 u64 delta = rq_clock(rq) - rq->idle_stamp;
3811 u64 max = 2*rq->max_idle_balance_cost;
3813 update_avg(&rq->avg_idle, delta);
3815 if (rq->avg_idle > max)
3822 p->dl_server = NULL;
3826 * Consider @p being inside a wait loop:
3829 * set_current_state(TASK_UNINTERRUPTIBLE);
3836 * __set_current_state(TASK_RUNNING);
3838 * between set_current_state() and schedule(). In this case @p is still
3839 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3842 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3843 * then schedule() must still happen and p->state can be changed to
3844 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3845 * need to do a full wakeup with enqueue.
3847 * Returns: %true when the wakeup is done,
3850 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3856 rq = __task_rq_lock(p, &rf);
3857 if (task_on_rq_queued(p)) {
3858 if (!task_on_cpu(rq, p)) {
3860 * When on_rq && !on_cpu the task is preempted, see if
3861 * it should preempt the task that is current now.
3863 update_rq_clock(rq);
3864 wakeup_preempt(rq, p, wake_flags);
3869 __task_rq_unlock(rq, &rf);
3875 void sched_ttwu_pending(void *arg)
3877 struct llist_node *llist = arg;
3878 struct rq *rq = this_rq();
3879 struct task_struct *p, *t;
3885 rq_lock_irqsave(rq, &rf);
3886 update_rq_clock(rq);
3888 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3889 if (WARN_ON_ONCE(p->on_cpu))
3890 smp_cond_load_acquire(&p->on_cpu, !VAL);
3892 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3893 set_task_cpu(p, cpu_of(rq));
3895 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3899 * Must be after enqueueing at least once task such that
3900 * idle_cpu() does not observe a false-negative -- if it does,
3901 * it is possible for select_idle_siblings() to stack a number
3902 * of tasks on this CPU during that window.
3904 * It is ok to clear ttwu_pending when another task pending.
3905 * We will receive IPI after local irq enabled and then enqueue it.
3906 * Since now nr_running > 0, idle_cpu() will always get correct result.
3908 WRITE_ONCE(rq->ttwu_pending, 0);
3909 rq_unlock_irqrestore(rq, &rf);
3913 * Prepare the scene for sending an IPI for a remote smp_call
3915 * Returns true if the caller can proceed with sending the IPI.
3916 * Returns false otherwise.
3918 bool call_function_single_prep_ipi(int cpu)
3920 if (set_nr_if_polling(cpu_rq(cpu)->idle)) {
3921 trace_sched_wake_idle_without_ipi(cpu);
3929 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3930 * necessary. The wakee CPU on receipt of the IPI will queue the task
3931 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3932 * of the wakeup instead of the waker.
3934 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3936 struct rq *rq = cpu_rq(cpu);
3938 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3940 WRITE_ONCE(rq->ttwu_pending, 1);
3941 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3944 void wake_up_if_idle(int cpu)
3946 struct rq *rq = cpu_rq(cpu);
3949 if (is_idle_task(rcu_dereference(rq->curr))) {
3950 guard(rq_lock_irqsave)(rq);
3951 if (is_idle_task(rq->curr))
3956 bool cpus_equal_capacity(int this_cpu, int that_cpu)
3958 if (!sched_asym_cpucap_active())
3961 if (this_cpu == that_cpu)
3964 return arch_scale_cpu_capacity(this_cpu) == arch_scale_cpu_capacity(that_cpu);
3967 bool cpus_share_cache(int this_cpu, int that_cpu)
3969 if (this_cpu == that_cpu)
3972 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3976 * Whether CPUs are share cache resources, which means LLC on non-cluster
3977 * machines and LLC tag or L2 on machines with clusters.
3979 bool cpus_share_resources(int this_cpu, int that_cpu)
3981 if (this_cpu == that_cpu)
3984 return per_cpu(sd_share_id, this_cpu) == per_cpu(sd_share_id, that_cpu);
3987 static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3990 * Do not complicate things with the async wake_list while the CPU is
3993 if (!cpu_active(cpu))
3996 /* Ensure the task will still be allowed to run on the CPU. */
3997 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
4001 * If the CPU does not share cache, then queue the task on the
4002 * remote rqs wakelist to avoid accessing remote data.
4004 if (!cpus_share_cache(smp_processor_id(), cpu))
4007 if (cpu == smp_processor_id())
4011 * If the wakee cpu is idle, or the task is descheduling and the
4012 * only running task on the CPU, then use the wakelist to offload
4013 * the task activation to the idle (or soon-to-be-idle) CPU as
4014 * the current CPU is likely busy. nr_running is checked to
4015 * avoid unnecessary task stacking.
4017 * Note that we can only get here with (wakee) p->on_rq=0,
4018 * p->on_cpu can be whatever, we've done the dequeue, so
4019 * the wakee has been accounted out of ->nr_running.
4021 if (!cpu_rq(cpu)->nr_running)
4027 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4029 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
4030 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
4031 __ttwu_queue_wakelist(p, cpu, wake_flags);
4038 #else /* !CONFIG_SMP */
4040 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4045 #endif /* CONFIG_SMP */
4047 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
4049 struct rq *rq = cpu_rq(cpu);
4052 if (ttwu_queue_wakelist(p, cpu, wake_flags))
4056 update_rq_clock(rq);
4057 ttwu_do_activate(rq, p, wake_flags, &rf);
4062 * Invoked from try_to_wake_up() to check whether the task can be woken up.
4064 * The caller holds p::pi_lock if p != current or has preemption
4065 * disabled when p == current.
4067 * The rules of saved_state:
4069 * The related locking code always holds p::pi_lock when updating
4070 * p::saved_state, which means the code is fully serialized in both cases.
4072 * For PREEMPT_RT, the lock wait and lock wakeups happen via TASK_RTLOCK_WAIT.
4073 * No other bits set. This allows to distinguish all wakeup scenarios.
4075 * For FREEZER, the wakeup happens via TASK_FROZEN. No other bits set. This
4076 * allows us to prevent early wakeup of tasks before they can be run on
4077 * asymmetric ISA architectures (eg ARMv9).
4079 static __always_inline
4080 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
4084 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
4085 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
4086 state != TASK_RTLOCK_WAIT);
4089 *success = !!(match = __task_state_match(p, state));
4092 * Saved state preserves the task state across blocking on
4093 * an RT lock or TASK_FREEZABLE tasks. If the state matches,
4094 * set p::saved_state to TASK_RUNNING, but do not wake the task
4095 * because it waits for a lock wakeup or __thaw_task(). Also
4096 * indicate success because from the regular waker's point of
4097 * view this has succeeded.
4099 * After acquiring the lock the task will restore p::__state
4100 * from p::saved_state which ensures that the regular
4101 * wakeup is not lost. The restore will also set
4102 * p::saved_state to TASK_RUNNING so any further tests will
4103 * not result in false positives vs. @success
4106 p->saved_state = TASK_RUNNING;
4112 * Notes on Program-Order guarantees on SMP systems.
4116 * The basic program-order guarantee on SMP systems is that when a task [t]
4117 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
4118 * execution on its new CPU [c1].
4120 * For migration (of runnable tasks) this is provided by the following means:
4122 * A) UNLOCK of the rq(c0)->lock scheduling out task t
4123 * B) migration for t is required to synchronize *both* rq(c0)->lock and
4124 * rq(c1)->lock (if not at the same time, then in that order).
4125 * C) LOCK of the rq(c1)->lock scheduling in task
4127 * Release/acquire chaining guarantees that B happens after A and C after B.
4128 * Note: the CPU doing B need not be c0 or c1
4137 * UNLOCK rq(0)->lock
4139 * LOCK rq(0)->lock // orders against CPU0
4141 * UNLOCK rq(0)->lock
4145 * UNLOCK rq(1)->lock
4147 * LOCK rq(1)->lock // orders against CPU2
4150 * UNLOCK rq(1)->lock
4153 * BLOCKING -- aka. SLEEP + WAKEUP
4155 * For blocking we (obviously) need to provide the same guarantee as for
4156 * migration. However the means are completely different as there is no lock
4157 * chain to provide order. Instead we do:
4159 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
4160 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
4164 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
4166 * LOCK rq(0)->lock LOCK X->pi_lock
4169 * smp_store_release(X->on_cpu, 0);
4171 * smp_cond_load_acquire(&X->on_cpu, !VAL);
4177 * X->state = RUNNING
4178 * UNLOCK rq(2)->lock
4180 * LOCK rq(2)->lock // orders against CPU1
4183 * UNLOCK rq(2)->lock
4186 * UNLOCK rq(0)->lock
4189 * However, for wakeups there is a second guarantee we must provide, namely we
4190 * must ensure that CONDITION=1 done by the caller can not be reordered with
4191 * accesses to the task state; see try_to_wake_up() and set_current_state().
4195 * try_to_wake_up - wake up a thread
4196 * @p: the thread to be awakened
4197 * @state: the mask of task states that can be woken
4198 * @wake_flags: wake modifier flags (WF_*)
4200 * Conceptually does:
4202 * If (@state & @p->state) @p->state = TASK_RUNNING.
4204 * If the task was not queued/runnable, also place it back on a runqueue.
4206 * This function is atomic against schedule() which would dequeue the task.
4208 * It issues a full memory barrier before accessing @p->state, see the comment
4209 * with set_current_state().
4211 * Uses p->pi_lock to serialize against concurrent wake-ups.
4213 * Relies on p->pi_lock stabilizing:
4216 * - p->sched_task_group
4217 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4219 * Tries really hard to only take one task_rq(p)->lock for performance.
4220 * Takes rq->lock in:
4221 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4222 * - ttwu_queue() -- new rq, for enqueue of the task;
4223 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4225 * As a consequence we race really badly with just about everything. See the
4226 * many memory barriers and their comments for details.
4228 * Return: %true if @p->state changes (an actual wakeup was done),
4231 int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4234 int cpu, success = 0;
4238 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4239 * == smp_processor_id()'. Together this means we can special
4240 * case the whole 'p->on_rq && ttwu_runnable()' case below
4241 * without taking any locks.
4244 * - we rely on Program-Order guarantees for all the ordering,
4245 * - we're serialized against set_special_state() by virtue of
4246 * it disabling IRQs (this allows not taking ->pi_lock).
4248 if (!ttwu_state_match(p, state, &success))
4251 trace_sched_waking(p);
4257 * If we are going to wake up a thread waiting for CONDITION we
4258 * need to ensure that CONDITION=1 done by the caller can not be
4259 * reordered with p->state check below. This pairs with smp_store_mb()
4260 * in set_current_state() that the waiting thread does.
4262 scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
4263 smp_mb__after_spinlock();
4264 if (!ttwu_state_match(p, state, &success))
4267 trace_sched_waking(p);
4270 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4271 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4272 * in smp_cond_load_acquire() below.
4274 * sched_ttwu_pending() try_to_wake_up()
4275 * STORE p->on_rq = 1 LOAD p->state
4278 * __schedule() (switch to task 'p')
4279 * LOCK rq->lock smp_rmb();
4280 * smp_mb__after_spinlock();
4284 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4286 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4287 * __schedule(). See the comment for smp_mb__after_spinlock().
4289 * A similar smp_rmb() lives in __task_needs_rq_lock().
4292 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4297 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4298 * possible to, falsely, observe p->on_cpu == 0.
4300 * One must be running (->on_cpu == 1) in order to remove oneself
4301 * from the runqueue.
4303 * __schedule() (switch to task 'p') try_to_wake_up()
4304 * STORE p->on_cpu = 1 LOAD p->on_rq
4307 * __schedule() (put 'p' to sleep)
4308 * LOCK rq->lock smp_rmb();
4309 * smp_mb__after_spinlock();
4310 * STORE p->on_rq = 0 LOAD p->on_cpu
4312 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4313 * __schedule(). See the comment for smp_mb__after_spinlock().
4315 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4316 * schedule()'s deactivate_task() has 'happened' and p will no longer
4317 * care about it's own p->state. See the comment in __schedule().
4319 smp_acquire__after_ctrl_dep();
4322 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4323 * == 0), which means we need to do an enqueue, change p->state to
4324 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4325 * enqueue, such as ttwu_queue_wakelist().
4327 WRITE_ONCE(p->__state, TASK_WAKING);
4330 * If the owning (remote) CPU is still in the middle of schedule() with
4331 * this task as prev, considering queueing p on the remote CPUs wake_list
4332 * which potentially sends an IPI instead of spinning on p->on_cpu to
4333 * let the waker make forward progress. This is safe because IRQs are
4334 * disabled and the IPI will deliver after on_cpu is cleared.
4336 * Ensure we load task_cpu(p) after p->on_cpu:
4338 * set_task_cpu(p, cpu);
4339 * STORE p->cpu = @cpu
4340 * __schedule() (switch to task 'p')
4342 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4343 * STORE p->on_cpu = 1 LOAD p->cpu
4345 * to ensure we observe the correct CPU on which the task is currently
4348 if (smp_load_acquire(&p->on_cpu) &&
4349 ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4353 * If the owning (remote) CPU is still in the middle of schedule() with
4354 * this task as prev, wait until it's done referencing the task.
4356 * Pairs with the smp_store_release() in finish_task().
4358 * This ensures that tasks getting woken will be fully ordered against
4359 * their previous state and preserve Program Order.
4361 smp_cond_load_acquire(&p->on_cpu, !VAL);
4363 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4364 if (task_cpu(p) != cpu) {
4366 delayacct_blkio_end(p);
4367 atomic_dec(&task_rq(p)->nr_iowait);
4370 wake_flags |= WF_MIGRATED;
4371 psi_ttwu_dequeue(p);
4372 set_task_cpu(p, cpu);
4376 #endif /* CONFIG_SMP */
4378 ttwu_queue(p, cpu, wake_flags);
4382 ttwu_stat(p, task_cpu(p), wake_flags);
4387 static bool __task_needs_rq_lock(struct task_struct *p)
4389 unsigned int state = READ_ONCE(p->__state);
4392 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4393 * the task is blocked. Make sure to check @state since ttwu() can drop
4394 * locks at the end, see ttwu_queue_wakelist().
4396 if (state == TASK_RUNNING || state == TASK_WAKING)
4400 * Ensure we load p->on_rq after p->__state, otherwise it would be
4401 * possible to, falsely, observe p->on_rq == 0.
4403 * See try_to_wake_up() for a longer comment.
4411 * Ensure the task has finished __schedule() and will not be referenced
4412 * anymore. Again, see try_to_wake_up() for a longer comment.
4415 smp_cond_load_acquire(&p->on_cpu, !VAL);
4422 * task_call_func - Invoke a function on task in fixed state
4423 * @p: Process for which the function is to be invoked, can be @current.
4424 * @func: Function to invoke.
4425 * @arg: Argument to function.
4427 * Fix the task in it's current state by avoiding wakeups and or rq operations
4428 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4429 * to work out what the state is, if required. Given that @func can be invoked
4430 * with a runqueue lock held, it had better be quite lightweight.
4433 * Whatever @func returns
4435 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4437 struct rq *rq = NULL;
4441 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4443 if (__task_needs_rq_lock(p))
4444 rq = __task_rq_lock(p, &rf);
4447 * At this point the task is pinned; either:
4448 * - blocked and we're holding off wakeups (pi->lock)
4449 * - woken, and we're holding off enqueue (rq->lock)
4450 * - queued, and we're holding off schedule (rq->lock)
4451 * - running, and we're holding off de-schedule (rq->lock)
4453 * The called function (@func) can use: task_curr(), p->on_rq and
4454 * p->__state to differentiate between these states.
4461 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4466 * cpu_curr_snapshot - Return a snapshot of the currently running task
4467 * @cpu: The CPU on which to snapshot the task.
4469 * Returns the task_struct pointer of the task "currently" running on
4470 * the specified CPU. If the same task is running on that CPU throughout,
4471 * the return value will be a pointer to that task's task_struct structure.
4472 * If the CPU did any context switches even vaguely concurrently with the
4473 * execution of this function, the return value will be a pointer to the
4474 * task_struct structure of a randomly chosen task that was running on
4475 * that CPU somewhere around the time that this function was executing.
4477 * If the specified CPU was offline, the return value is whatever it
4478 * is, perhaps a pointer to the task_struct structure of that CPU's idle
4479 * task, but there is no guarantee. Callers wishing a useful return
4480 * value must take some action to ensure that the specified CPU remains
4481 * online throughout.
4483 * This function executes full memory barriers before and after fetching
4484 * the pointer, which permits the caller to confine this function's fetch
4485 * with respect to the caller's accesses to other shared variables.
4487 struct task_struct *cpu_curr_snapshot(int cpu)
4489 struct task_struct *t;
4491 smp_mb(); /* Pairing determined by caller's synchronization design. */
4492 t = rcu_dereference(cpu_curr(cpu));
4493 smp_mb(); /* Pairing determined by caller's synchronization design. */
4498 * wake_up_process - Wake up a specific process
4499 * @p: The process to be woken up.
4501 * Attempt to wake up the nominated process and move it to the set of runnable
4504 * Return: 1 if the process was woken up, 0 if it was already running.
4506 * This function executes a full memory barrier before accessing the task state.
4508 int wake_up_process(struct task_struct *p)
4510 return try_to_wake_up(p, TASK_NORMAL, 0);
4512 EXPORT_SYMBOL(wake_up_process);
4514 int wake_up_state(struct task_struct *p, unsigned int state)
4516 return try_to_wake_up(p, state, 0);
4520 * Perform scheduler related setup for a newly forked process p.
4521 * p is forked by current.
4523 * __sched_fork() is basic setup used by init_idle() too:
4525 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4530 p->se.exec_start = 0;
4531 p->se.sum_exec_runtime = 0;
4532 p->se.prev_sum_exec_runtime = 0;
4533 p->se.nr_migrations = 0;
4536 p->se.slice = sysctl_sched_base_slice;
4537 INIT_LIST_HEAD(&p->se.group_node);
4539 #ifdef CONFIG_FAIR_GROUP_SCHED
4540 p->se.cfs_rq = NULL;
4543 #ifdef CONFIG_SCHEDSTATS
4544 /* Even if schedstat is disabled, there should not be garbage */
4545 memset(&p->stats, 0, sizeof(p->stats));
4548 init_dl_entity(&p->dl);
4550 INIT_LIST_HEAD(&p->rt.run_list);
4552 p->rt.time_slice = sched_rr_timeslice;
4556 #ifdef CONFIG_PREEMPT_NOTIFIERS
4557 INIT_HLIST_HEAD(&p->preempt_notifiers);
4560 #ifdef CONFIG_COMPACTION
4561 p->capture_control = NULL;
4563 init_numa_balancing(clone_flags, p);
4565 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4566 p->migration_pending = NULL;
4568 init_sched_mm_cid(p);
4571 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4573 #ifdef CONFIG_NUMA_BALANCING
4575 int sysctl_numa_balancing_mode;
4577 static void __set_numabalancing_state(bool enabled)
4580 static_branch_enable(&sched_numa_balancing);
4582 static_branch_disable(&sched_numa_balancing);
4585 void set_numabalancing_state(bool enabled)
4588 sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4590 sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4591 __set_numabalancing_state(enabled);
4594 #ifdef CONFIG_PROC_SYSCTL
4595 static void reset_memory_tiering(void)
4597 struct pglist_data *pgdat;
4599 for_each_online_pgdat(pgdat) {
4600 pgdat->nbp_threshold = 0;
4601 pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
4602 pgdat->nbp_th_start = jiffies_to_msecs(jiffies);
4606 static int sysctl_numa_balancing(struct ctl_table *table, int write,
4607 void *buffer, size_t *lenp, loff_t *ppos)
4611 int state = sysctl_numa_balancing_mode;
4613 if (write && !capable(CAP_SYS_ADMIN))
4618 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4622 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4623 (state & NUMA_BALANCING_MEMORY_TIERING))
4624 reset_memory_tiering();
4625 sysctl_numa_balancing_mode = state;
4626 __set_numabalancing_state(state);
4633 #ifdef CONFIG_SCHEDSTATS
4635 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4637 static void set_schedstats(bool enabled)
4640 static_branch_enable(&sched_schedstats);
4642 static_branch_disable(&sched_schedstats);
4645 void force_schedstat_enabled(void)
4647 if (!schedstat_enabled()) {
4648 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4649 static_branch_enable(&sched_schedstats);
4653 static int __init setup_schedstats(char *str)
4659 if (!strcmp(str, "enable")) {
4660 set_schedstats(true);
4662 } else if (!strcmp(str, "disable")) {
4663 set_schedstats(false);
4668 pr_warn("Unable to parse schedstats=\n");
4672 __setup("schedstats=", setup_schedstats);
4674 #ifdef CONFIG_PROC_SYSCTL
4675 static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4676 size_t *lenp, loff_t *ppos)
4680 int state = static_branch_likely(&sched_schedstats);
4682 if (write && !capable(CAP_SYS_ADMIN))
4687 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4691 set_schedstats(state);
4694 #endif /* CONFIG_PROC_SYSCTL */
4695 #endif /* CONFIG_SCHEDSTATS */
4697 #ifdef CONFIG_SYSCTL
4698 static struct ctl_table sched_core_sysctls[] = {
4699 #ifdef CONFIG_SCHEDSTATS
4701 .procname = "sched_schedstats",
4703 .maxlen = sizeof(unsigned int),
4705 .proc_handler = sysctl_schedstats,
4706 .extra1 = SYSCTL_ZERO,
4707 .extra2 = SYSCTL_ONE,
4709 #endif /* CONFIG_SCHEDSTATS */
4710 #ifdef CONFIG_UCLAMP_TASK
4712 .procname = "sched_util_clamp_min",
4713 .data = &sysctl_sched_uclamp_util_min,
4714 .maxlen = sizeof(unsigned int),
4716 .proc_handler = sysctl_sched_uclamp_handler,
4719 .procname = "sched_util_clamp_max",
4720 .data = &sysctl_sched_uclamp_util_max,
4721 .maxlen = sizeof(unsigned int),
4723 .proc_handler = sysctl_sched_uclamp_handler,
4726 .procname = "sched_util_clamp_min_rt_default",
4727 .data = &sysctl_sched_uclamp_util_min_rt_default,
4728 .maxlen = sizeof(unsigned int),
4730 .proc_handler = sysctl_sched_uclamp_handler,
4732 #endif /* CONFIG_UCLAMP_TASK */
4733 #ifdef CONFIG_NUMA_BALANCING
4735 .procname = "numa_balancing",
4736 .data = NULL, /* filled in by handler */
4737 .maxlen = sizeof(unsigned int),
4739 .proc_handler = sysctl_numa_balancing,
4740 .extra1 = SYSCTL_ZERO,
4741 .extra2 = SYSCTL_FOUR,
4743 #endif /* CONFIG_NUMA_BALANCING */
4746 static int __init sched_core_sysctl_init(void)
4748 register_sysctl_init("kernel", sched_core_sysctls);
4751 late_initcall(sched_core_sysctl_init);
4752 #endif /* CONFIG_SYSCTL */
4755 * fork()/clone()-time setup:
4757 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4759 __sched_fork(clone_flags, p);
4761 * We mark the process as NEW here. This guarantees that
4762 * nobody will actually run it, and a signal or other external
4763 * event cannot wake it up and insert it on the runqueue either.
4765 p->__state = TASK_NEW;
4768 * Make sure we do not leak PI boosting priority to the child.
4770 p->prio = current->normal_prio;
4775 * Revert to default priority/policy on fork if requested.
4777 if (unlikely(p->sched_reset_on_fork)) {
4778 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4779 p->policy = SCHED_NORMAL;
4780 p->static_prio = NICE_TO_PRIO(0);
4782 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4783 p->static_prio = NICE_TO_PRIO(0);
4785 p->prio = p->normal_prio = p->static_prio;
4786 set_load_weight(p, false);
4789 * We don't need the reset flag anymore after the fork. It has
4790 * fulfilled its duty:
4792 p->sched_reset_on_fork = 0;
4795 if (dl_prio(p->prio))
4797 else if (rt_prio(p->prio))
4798 p->sched_class = &rt_sched_class;
4800 p->sched_class = &fair_sched_class;
4802 init_entity_runnable_average(&p->se);
4805 #ifdef CONFIG_SCHED_INFO
4806 if (likely(sched_info_on()))
4807 memset(&p->sched_info, 0, sizeof(p->sched_info));
4809 #if defined(CONFIG_SMP)
4812 init_task_preempt_count(p);
4814 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4815 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4820 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4822 unsigned long flags;
4825 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4826 * required yet, but lockdep gets upset if rules are violated.
4828 raw_spin_lock_irqsave(&p->pi_lock, flags);
4829 #ifdef CONFIG_CGROUP_SCHED
4831 struct task_group *tg;
4832 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4833 struct task_group, css);
4834 tg = autogroup_task_group(p, tg);
4835 p->sched_task_group = tg;
4840 * We're setting the CPU for the first time, we don't migrate,
4841 * so use __set_task_cpu().
4843 __set_task_cpu(p, smp_processor_id());
4844 if (p->sched_class->task_fork)
4845 p->sched_class->task_fork(p);
4846 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4849 void sched_post_fork(struct task_struct *p)
4851 uclamp_post_fork(p);
4854 unsigned long to_ratio(u64 period, u64 runtime)
4856 if (runtime == RUNTIME_INF)
4860 * Doing this here saves a lot of checks in all
4861 * the calling paths, and returning zero seems
4862 * safe for them anyway.
4867 return div64_u64(runtime << BW_SHIFT, period);
4871 * wake_up_new_task - wake up a newly created task for the first time.
4873 * This function will do some initial scheduler statistics housekeeping
4874 * that must be done for every newly created context, then puts the task
4875 * on the runqueue and wakes it.
4877 void wake_up_new_task(struct task_struct *p)
4882 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4883 WRITE_ONCE(p->__state, TASK_RUNNING);
4886 * Fork balancing, do it here and not earlier because:
4887 * - cpus_ptr can change in the fork path
4888 * - any previously selected CPU might disappear through hotplug
4890 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4891 * as we're not fully set-up yet.
4893 p->recent_used_cpu = task_cpu(p);
4895 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4897 rq = __task_rq_lock(p, &rf);
4898 update_rq_clock(rq);
4899 post_init_entity_util_avg(p);
4901 activate_task(rq, p, ENQUEUE_NOCLOCK);
4902 trace_sched_wakeup_new(p);
4903 wakeup_preempt(rq, p, WF_FORK);
4905 if (p->sched_class->task_woken) {
4907 * Nothing relies on rq->lock after this, so it's fine to
4910 rq_unpin_lock(rq, &rf);
4911 p->sched_class->task_woken(rq, p);
4912 rq_repin_lock(rq, &rf);
4915 task_rq_unlock(rq, p, &rf);
4918 #ifdef CONFIG_PREEMPT_NOTIFIERS
4920 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4922 void preempt_notifier_inc(void)
4924 static_branch_inc(&preempt_notifier_key);
4926 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4928 void preempt_notifier_dec(void)
4930 static_branch_dec(&preempt_notifier_key);
4932 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4935 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4936 * @notifier: notifier struct to register
4938 void preempt_notifier_register(struct preempt_notifier *notifier)
4940 if (!static_branch_unlikely(&preempt_notifier_key))
4941 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4943 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4945 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4948 * preempt_notifier_unregister - no longer interested in preemption notifications
4949 * @notifier: notifier struct to unregister
4951 * This is *not* safe to call from within a preemption notifier.
4953 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4955 hlist_del(¬ifier->link);
4957 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4959 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4961 struct preempt_notifier *notifier;
4963 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4964 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4967 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4969 if (static_branch_unlikely(&preempt_notifier_key))
4970 __fire_sched_in_preempt_notifiers(curr);
4974 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4975 struct task_struct *next)
4977 struct preempt_notifier *notifier;
4979 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4980 notifier->ops->sched_out(notifier, next);
4983 static __always_inline void
4984 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4985 struct task_struct *next)
4987 if (static_branch_unlikely(&preempt_notifier_key))
4988 __fire_sched_out_preempt_notifiers(curr, next);
4991 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4993 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4998 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4999 struct task_struct *next)
5003 #endif /* CONFIG_PREEMPT_NOTIFIERS */
5005 static inline void prepare_task(struct task_struct *next)
5009 * Claim the task as running, we do this before switching to it
5010 * such that any running task will have this set.
5012 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
5013 * its ordering comment.
5015 WRITE_ONCE(next->on_cpu, 1);
5019 static inline void finish_task(struct task_struct *prev)
5023 * This must be the very last reference to @prev from this CPU. After
5024 * p->on_cpu is cleared, the task can be moved to a different CPU. We
5025 * must ensure this doesn't happen until the switch is completely
5028 * In particular, the load of prev->state in finish_task_switch() must
5029 * happen before this.
5031 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
5033 smp_store_release(&prev->on_cpu, 0);
5039 static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
5041 void (*func)(struct rq *rq);
5042 struct balance_callback *next;
5044 lockdep_assert_rq_held(rq);
5047 func = (void (*)(struct rq *))head->func;
5056 static void balance_push(struct rq *rq);
5059 * balance_push_callback is a right abuse of the callback interface and plays
5060 * by significantly different rules.
5062 * Where the normal balance_callback's purpose is to be ran in the same context
5063 * that queued it (only later, when it's safe to drop rq->lock again),
5064 * balance_push_callback is specifically targeted at __schedule().
5066 * This abuse is tolerated because it places all the unlikely/odd cases behind
5067 * a single test, namely: rq->balance_callback == NULL.
5069 struct balance_callback balance_push_callback = {
5071 .func = balance_push,
5074 static inline struct balance_callback *
5075 __splice_balance_callbacks(struct rq *rq, bool split)
5077 struct balance_callback *head = rq->balance_callback;
5082 lockdep_assert_rq_held(rq);
5084 * Must not take balance_push_callback off the list when
5085 * splice_balance_callbacks() and balance_callbacks() are not
5086 * in the same rq->lock section.
5088 * In that case it would be possible for __schedule() to interleave
5089 * and observe the list empty.
5091 if (split && head == &balance_push_callback)
5094 rq->balance_callback = NULL;
5099 static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5101 return __splice_balance_callbacks(rq, true);
5104 static void __balance_callbacks(struct rq *rq)
5106 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
5109 static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5111 unsigned long flags;
5113 if (unlikely(head)) {
5114 raw_spin_rq_lock_irqsave(rq, flags);
5115 do_balance_callbacks(rq, head);
5116 raw_spin_rq_unlock_irqrestore(rq, flags);
5122 static inline void __balance_callbacks(struct rq *rq)
5126 static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5131 static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5138 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
5141 * Since the runqueue lock will be released by the next
5142 * task (which is an invalid locking op but in the case
5143 * of the scheduler it's an obvious special-case), so we
5144 * do an early lockdep release here:
5146 rq_unpin_lock(rq, rf);
5147 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
5148 #ifdef CONFIG_DEBUG_SPINLOCK
5149 /* this is a valid case when another task releases the spinlock */
5150 rq_lockp(rq)->owner = next;
5154 static inline void finish_lock_switch(struct rq *rq)
5157 * If we are tracking spinlock dependencies then we have to
5158 * fix up the runqueue lock - which gets 'carried over' from
5159 * prev into current:
5161 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
5162 __balance_callbacks(rq);
5163 raw_spin_rq_unlock_irq(rq);
5167 * NOP if the arch has not defined these:
5170 #ifndef prepare_arch_switch
5171 # define prepare_arch_switch(next) do { } while (0)
5174 #ifndef finish_arch_post_lock_switch
5175 # define finish_arch_post_lock_switch() do { } while (0)
5178 static inline void kmap_local_sched_out(void)
5180 #ifdef CONFIG_KMAP_LOCAL
5181 if (unlikely(current->kmap_ctrl.idx))
5182 __kmap_local_sched_out();
5186 static inline void kmap_local_sched_in(void)
5188 #ifdef CONFIG_KMAP_LOCAL
5189 if (unlikely(current->kmap_ctrl.idx))
5190 __kmap_local_sched_in();
5195 * prepare_task_switch - prepare to switch tasks
5196 * @rq: the runqueue preparing to switch
5197 * @prev: the current task that is being switched out
5198 * @next: the task we are going to switch to.
5200 * This is called with the rq lock held and interrupts off. It must
5201 * be paired with a subsequent finish_task_switch after the context
5204 * prepare_task_switch sets up locking and calls architecture specific
5208 prepare_task_switch(struct rq *rq, struct task_struct *prev,
5209 struct task_struct *next)
5211 kcov_prepare_switch(prev);
5212 sched_info_switch(rq, prev, next);
5213 perf_event_task_sched_out(prev, next);
5215 fire_sched_out_preempt_notifiers(prev, next);
5216 kmap_local_sched_out();
5218 prepare_arch_switch(next);
5222 * finish_task_switch - clean up after a task-switch
5223 * @prev: the thread we just switched away from.
5225 * finish_task_switch must be called after the context switch, paired
5226 * with a prepare_task_switch call before the context switch.
5227 * finish_task_switch will reconcile locking set up by prepare_task_switch,
5228 * and do any other architecture-specific cleanup actions.
5230 * Note that we may have delayed dropping an mm in context_switch(). If
5231 * so, we finish that here outside of the runqueue lock. (Doing it
5232 * with the lock held can cause deadlocks; see schedule() for
5235 * The context switch have flipped the stack from under us and restored the
5236 * local variables which were saved when this task called schedule() in the
5237 * past. prev == current is still correct but we need to recalculate this_rq
5238 * because prev may have moved to another CPU.
5240 static struct rq *finish_task_switch(struct task_struct *prev)
5241 __releases(rq->lock)
5243 struct rq *rq = this_rq();
5244 struct mm_struct *mm = rq->prev_mm;
5245 unsigned int prev_state;
5248 * The previous task will have left us with a preempt_count of 2
5249 * because it left us after:
5252 * preempt_disable(); // 1
5254 * raw_spin_lock_irq(&rq->lock) // 2
5256 * Also, see FORK_PREEMPT_COUNT.
5258 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5259 "corrupted preempt_count: %s/%d/0x%x\n",
5260 current->comm, current->pid, preempt_count()))
5261 preempt_count_set(FORK_PREEMPT_COUNT);
5266 * A task struct has one reference for the use as "current".
5267 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5268 * schedule one last time. The schedule call will never return, and
5269 * the scheduled task must drop that reference.
5271 * We must observe prev->state before clearing prev->on_cpu (in
5272 * finish_task), otherwise a concurrent wakeup can get prev
5273 * running on another CPU and we could rave with its RUNNING -> DEAD
5274 * transition, resulting in a double drop.
5276 prev_state = READ_ONCE(prev->__state);
5277 vtime_task_switch(prev);
5278 perf_event_task_sched_in(prev, current);
5280 tick_nohz_task_switch();
5281 finish_lock_switch(rq);
5282 finish_arch_post_lock_switch();
5283 kcov_finish_switch(current);
5285 * kmap_local_sched_out() is invoked with rq::lock held and
5286 * interrupts disabled. There is no requirement for that, but the
5287 * sched out code does not have an interrupt enabled section.
5288 * Restoring the maps on sched in does not require interrupts being
5291 kmap_local_sched_in();
5293 fire_sched_in_preempt_notifiers(current);
5295 * When switching through a kernel thread, the loop in
5296 * membarrier_{private,global}_expedited() may have observed that
5297 * kernel thread and not issued an IPI. It is therefore possible to
5298 * schedule between user->kernel->user threads without passing though
5299 * switch_mm(). Membarrier requires a barrier after storing to
5300 * rq->curr, before returning to userspace, so provide them here:
5302 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5303 * provided by mmdrop_lazy_tlb(),
5304 * - a sync_core for SYNC_CORE.
5307 membarrier_mm_sync_core_before_usermode(mm);
5308 mmdrop_lazy_tlb_sched(mm);
5311 if (unlikely(prev_state == TASK_DEAD)) {
5312 if (prev->sched_class->task_dead)
5313 prev->sched_class->task_dead(prev);
5315 /* Task is done with its stack. */
5316 put_task_stack(prev);
5318 put_task_struct_rcu_user(prev);
5325 * schedule_tail - first thing a freshly forked thread must call.
5326 * @prev: the thread we just switched away from.
5328 asmlinkage __visible void schedule_tail(struct task_struct *prev)
5329 __releases(rq->lock)
5332 * New tasks start with FORK_PREEMPT_COUNT, see there and
5333 * finish_task_switch() for details.
5335 * finish_task_switch() will drop rq->lock() and lower preempt_count
5336 * and the preempt_enable() will end up enabling preemption (on
5337 * PREEMPT_COUNT kernels).
5340 finish_task_switch(prev);
5343 if (current->set_child_tid)
5344 put_user(task_pid_vnr(current), current->set_child_tid);
5346 calculate_sigpending();
5350 * context_switch - switch to the new MM and the new thread's register state.
5352 static __always_inline struct rq *
5353 context_switch(struct rq *rq, struct task_struct *prev,
5354 struct task_struct *next, struct rq_flags *rf)
5356 prepare_task_switch(rq, prev, next);
5359 * For paravirt, this is coupled with an exit in switch_to to
5360 * combine the page table reload and the switch backend into
5363 arch_start_context_switch(prev);
5366 * kernel -> kernel lazy + transfer active
5367 * user -> kernel lazy + mmgrab_lazy_tlb() active
5369 * kernel -> user switch + mmdrop_lazy_tlb() active
5370 * user -> user switch
5372 * switch_mm_cid() needs to be updated if the barriers provided
5373 * by context_switch() are modified.
5375 if (!next->mm) { // to kernel
5376 enter_lazy_tlb(prev->active_mm, next);
5378 next->active_mm = prev->active_mm;
5379 if (prev->mm) // from user
5380 mmgrab_lazy_tlb(prev->active_mm);
5382 prev->active_mm = NULL;
5384 membarrier_switch_mm(rq, prev->active_mm, next->mm);
5386 * sys_membarrier() requires an smp_mb() between setting
5387 * rq->curr / membarrier_switch_mm() and returning to userspace.
5389 * The below provides this either through switch_mm(), or in
5390 * case 'prev->active_mm == next->mm' through
5391 * finish_task_switch()'s mmdrop().
5393 switch_mm_irqs_off(prev->active_mm, next->mm, next);
5394 lru_gen_use_mm(next->mm);
5396 if (!prev->mm) { // from kernel
5397 /* will mmdrop_lazy_tlb() in finish_task_switch(). */
5398 rq->prev_mm = prev->active_mm;
5399 prev->active_mm = NULL;
5403 /* switch_mm_cid() requires the memory barriers above. */
5404 switch_mm_cid(rq, prev, next);
5406 prepare_lock_switch(rq, next, rf);
5408 /* Here we just switch the register state and the stack. */
5409 switch_to(prev, next, prev);
5412 return finish_task_switch(prev);
5416 * nr_running and nr_context_switches:
5418 * externally visible scheduler statistics: current number of runnable
5419 * threads, total number of context switches performed since bootup.
5421 unsigned int nr_running(void)
5423 unsigned int i, sum = 0;
5425 for_each_online_cpu(i)
5426 sum += cpu_rq(i)->nr_running;
5432 * Check if only the current task is running on the CPU.
5434 * Caution: this function does not check that the caller has disabled
5435 * preemption, thus the result might have a time-of-check-to-time-of-use
5436 * race. The caller is responsible to use it correctly, for example:
5438 * - from a non-preemptible section (of course)
5440 * - from a thread that is bound to a single CPU
5442 * - in a loop with very short iterations (e.g. a polling loop)
5444 bool single_task_running(void)
5446 return raw_rq()->nr_running == 1;
5448 EXPORT_SYMBOL(single_task_running);
5450 unsigned long long nr_context_switches_cpu(int cpu)
5452 return cpu_rq(cpu)->nr_switches;
5455 unsigned long long nr_context_switches(void)
5458 unsigned long long sum = 0;
5460 for_each_possible_cpu(i)
5461 sum += cpu_rq(i)->nr_switches;
5467 * Consumers of these two interfaces, like for example the cpuidle menu
5468 * governor, are using nonsensical data. Preferring shallow idle state selection
5469 * for a CPU that has IO-wait which might not even end up running the task when
5470 * it does become runnable.
5473 unsigned int nr_iowait_cpu(int cpu)
5475 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5479 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5481 * The idea behind IO-wait account is to account the idle time that we could
5482 * have spend running if it were not for IO. That is, if we were to improve the
5483 * storage performance, we'd have a proportional reduction in IO-wait time.
5485 * This all works nicely on UP, where, when a task blocks on IO, we account
5486 * idle time as IO-wait, because if the storage were faster, it could've been
5487 * running and we'd not be idle.
5489 * This has been extended to SMP, by doing the same for each CPU. This however
5492 * Imagine for instance the case where two tasks block on one CPU, only the one
5493 * CPU will have IO-wait accounted, while the other has regular idle. Even
5494 * though, if the storage were faster, both could've ran at the same time,
5495 * utilising both CPUs.
5497 * This means, that when looking globally, the current IO-wait accounting on
5498 * SMP is a lower bound, by reason of under accounting.
5500 * Worse, since the numbers are provided per CPU, they are sometimes
5501 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5502 * associated with any one particular CPU, it can wake to another CPU than it
5503 * blocked on. This means the per CPU IO-wait number is meaningless.
5505 * Task CPU affinities can make all that even more 'interesting'.
5508 unsigned int nr_iowait(void)
5510 unsigned int i, sum = 0;
5512 for_each_possible_cpu(i)
5513 sum += nr_iowait_cpu(i);
5521 * sched_exec - execve() is a valuable balancing opportunity, because at
5522 * this point the task has the smallest effective memory and cache footprint.
5524 void sched_exec(void)
5526 struct task_struct *p = current;
5527 struct migration_arg arg;
5530 scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
5531 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5532 if (dest_cpu == smp_processor_id())
5535 if (unlikely(!cpu_active(dest_cpu)))
5538 arg = (struct migration_arg){ p, dest_cpu };
5540 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5545 DEFINE_PER_CPU(struct kernel_stat, kstat);
5546 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5548 EXPORT_PER_CPU_SYMBOL(kstat);
5549 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5552 * The function fair_sched_class.update_curr accesses the struct curr
5553 * and its field curr->exec_start; when called from task_sched_runtime(),
5554 * we observe a high rate of cache misses in practice.
5555 * Prefetching this data results in improved performance.
5557 static inline void prefetch_curr_exec_start(struct task_struct *p)
5559 #ifdef CONFIG_FAIR_GROUP_SCHED
5560 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5562 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5565 prefetch(&curr->exec_start);
5569 * Return accounted runtime for the task.
5570 * In case the task is currently running, return the runtime plus current's
5571 * pending runtime that have not been accounted yet.
5573 unsigned long long task_sched_runtime(struct task_struct *p)
5579 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5581 * 64-bit doesn't need locks to atomically read a 64-bit value.
5582 * So we have a optimization chance when the task's delta_exec is 0.
5583 * Reading ->on_cpu is racy, but this is ok.
5585 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5586 * If we race with it entering CPU, unaccounted time is 0. This is
5587 * indistinguishable from the read occurring a few cycles earlier.
5588 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5589 * been accounted, so we're correct here as well.
5591 if (!p->on_cpu || !task_on_rq_queued(p))
5592 return p->se.sum_exec_runtime;
5595 rq = task_rq_lock(p, &rf);
5597 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5598 * project cycles that may never be accounted to this
5599 * thread, breaking clock_gettime().
5601 if (task_current(rq, p) && task_on_rq_queued(p)) {
5602 prefetch_curr_exec_start(p);
5603 update_rq_clock(rq);
5604 p->sched_class->update_curr(rq);
5606 ns = p->se.sum_exec_runtime;
5607 task_rq_unlock(rq, p, &rf);
5612 #ifdef CONFIG_SCHED_DEBUG
5613 static u64 cpu_resched_latency(struct rq *rq)
5615 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5616 u64 resched_latency, now = rq_clock(rq);
5617 static bool warned_once;
5619 if (sysctl_resched_latency_warn_once && warned_once)
5622 if (!need_resched() || !latency_warn_ms)
5625 if (system_state == SYSTEM_BOOTING)
5628 if (!rq->last_seen_need_resched_ns) {
5629 rq->last_seen_need_resched_ns = now;
5630 rq->ticks_without_resched = 0;
5634 rq->ticks_without_resched++;
5635 resched_latency = now - rq->last_seen_need_resched_ns;
5636 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5641 return resched_latency;
5644 static int __init setup_resched_latency_warn_ms(char *str)
5648 if ((kstrtol(str, 0, &val))) {
5649 pr_warn("Unable to set resched_latency_warn_ms\n");
5653 sysctl_resched_latency_warn_ms = val;
5656 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5658 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5659 #endif /* CONFIG_SCHED_DEBUG */
5662 * This function gets called by the timer code, with HZ frequency.
5663 * We call it with interrupts disabled.
5665 void scheduler_tick(void)
5667 int cpu = smp_processor_id();
5668 struct rq *rq = cpu_rq(cpu);
5669 struct task_struct *curr = rq->curr;
5671 unsigned long thermal_pressure;
5672 u64 resched_latency;
5674 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5675 arch_scale_freq_tick();
5681 update_rq_clock(rq);
5682 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5683 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5684 curr->sched_class->task_tick(rq, curr, 0);
5685 if (sched_feat(LATENCY_WARN))
5686 resched_latency = cpu_resched_latency(rq);
5687 calc_global_load_tick(rq);
5688 sched_core_tick(rq);
5689 task_tick_mm_cid(rq, curr);
5693 if (sched_feat(LATENCY_WARN) && resched_latency)
5694 resched_latency_warn(cpu, resched_latency);
5696 perf_event_task_tick();
5698 if (curr->flags & PF_WQ_WORKER)
5699 wq_worker_tick(curr);
5702 rq->idle_balance = idle_cpu(cpu);
5703 trigger_load_balance(rq);
5707 #ifdef CONFIG_NO_HZ_FULL
5712 struct delayed_work work;
5714 /* Values for ->state, see diagram below. */
5715 #define TICK_SCHED_REMOTE_OFFLINE 0
5716 #define TICK_SCHED_REMOTE_OFFLINING 1
5717 #define TICK_SCHED_REMOTE_RUNNING 2
5720 * State diagram for ->state:
5723 * TICK_SCHED_REMOTE_OFFLINE
5726 * | | sched_tick_remote()
5729 * +--TICK_SCHED_REMOTE_OFFLINING
5732 * sched_tick_start() | | sched_tick_stop()
5735 * TICK_SCHED_REMOTE_RUNNING
5738 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5739 * and sched_tick_start() are happy to leave the state in RUNNING.
5742 static struct tick_work __percpu *tick_work_cpu;
5744 static void sched_tick_remote(struct work_struct *work)
5746 struct delayed_work *dwork = to_delayed_work(work);
5747 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5748 int cpu = twork->cpu;
5749 struct rq *rq = cpu_rq(cpu);
5753 * Handle the tick only if it appears the remote CPU is running in full
5754 * dynticks mode. The check is racy by nature, but missing a tick or
5755 * having one too much is no big deal because the scheduler tick updates
5756 * statistics and checks timeslices in a time-independent way, regardless
5757 * of when exactly it is running.
5759 if (tick_nohz_tick_stopped_cpu(cpu)) {
5760 guard(rq_lock_irq)(rq);
5761 struct task_struct *curr = rq->curr;
5763 if (cpu_online(cpu)) {
5764 update_rq_clock(rq);
5766 if (!is_idle_task(curr)) {
5768 * Make sure the next tick runs within a
5769 * reasonable amount of time.
5771 u64 delta = rq_clock_task(rq) - curr->se.exec_start;
5772 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5774 curr->sched_class->task_tick(rq, curr, 0);
5776 calc_load_nohz_remote(rq);
5781 * Run the remote tick once per second (1Hz). This arbitrary
5782 * frequency is large enough to avoid overload but short enough
5783 * to keep scheduler internal stats reasonably up to date. But
5784 * first update state to reflect hotplug activity if required.
5786 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5787 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5788 if (os == TICK_SCHED_REMOTE_RUNNING)
5789 queue_delayed_work(system_unbound_wq, dwork, HZ);
5792 static void sched_tick_start(int cpu)
5795 struct tick_work *twork;
5797 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5800 WARN_ON_ONCE(!tick_work_cpu);
5802 twork = per_cpu_ptr(tick_work_cpu, cpu);
5803 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5804 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5805 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5807 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5808 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5812 #ifdef CONFIG_HOTPLUG_CPU
5813 static void sched_tick_stop(int cpu)
5815 struct tick_work *twork;
5818 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5821 WARN_ON_ONCE(!tick_work_cpu);
5823 twork = per_cpu_ptr(tick_work_cpu, cpu);
5824 /* There cannot be competing actions, but don't rely on stop-machine. */
5825 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5826 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5827 /* Don't cancel, as this would mess up the state machine. */
5829 #endif /* CONFIG_HOTPLUG_CPU */
5831 int __init sched_tick_offload_init(void)
5833 tick_work_cpu = alloc_percpu(struct tick_work);
5834 BUG_ON(!tick_work_cpu);
5838 #else /* !CONFIG_NO_HZ_FULL */
5839 static inline void sched_tick_start(int cpu) { }
5840 static inline void sched_tick_stop(int cpu) { }
5843 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5844 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5846 * If the value passed in is equal to the current preempt count
5847 * then we just disabled preemption. Start timing the latency.
5849 static inline void preempt_latency_start(int val)
5851 if (preempt_count() == val) {
5852 unsigned long ip = get_lock_parent_ip();
5853 #ifdef CONFIG_DEBUG_PREEMPT
5854 current->preempt_disable_ip = ip;
5856 trace_preempt_off(CALLER_ADDR0, ip);
5860 void preempt_count_add(int val)
5862 #ifdef CONFIG_DEBUG_PREEMPT
5866 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5869 __preempt_count_add(val);
5870 #ifdef CONFIG_DEBUG_PREEMPT
5872 * Spinlock count overflowing soon?
5874 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5877 preempt_latency_start(val);
5879 EXPORT_SYMBOL(preempt_count_add);
5880 NOKPROBE_SYMBOL(preempt_count_add);
5883 * If the value passed in equals to the current preempt count
5884 * then we just enabled preemption. Stop timing the latency.
5886 static inline void preempt_latency_stop(int val)
5888 if (preempt_count() == val)
5889 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5892 void preempt_count_sub(int val)
5894 #ifdef CONFIG_DEBUG_PREEMPT
5898 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5901 * Is the spinlock portion underflowing?
5903 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5904 !(preempt_count() & PREEMPT_MASK)))
5908 preempt_latency_stop(val);
5909 __preempt_count_sub(val);
5911 EXPORT_SYMBOL(preempt_count_sub);
5912 NOKPROBE_SYMBOL(preempt_count_sub);
5915 static inline void preempt_latency_start(int val) { }
5916 static inline void preempt_latency_stop(int val) { }
5919 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5921 #ifdef CONFIG_DEBUG_PREEMPT
5922 return p->preempt_disable_ip;
5929 * Print scheduling while atomic bug:
5931 static noinline void __schedule_bug(struct task_struct *prev)
5933 /* Save this before calling printk(), since that will clobber it */
5934 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5936 if (oops_in_progress)
5939 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5940 prev->comm, prev->pid, preempt_count());
5942 debug_show_held_locks(prev);
5944 if (irqs_disabled())
5945 print_irqtrace_events(prev);
5946 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
5947 pr_err("Preemption disabled at:");
5948 print_ip_sym(KERN_ERR, preempt_disable_ip);
5950 check_panic_on_warn("scheduling while atomic");
5953 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5957 * Various schedule()-time debugging checks and statistics:
5959 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5961 #ifdef CONFIG_SCHED_STACK_END_CHECK
5962 if (task_stack_end_corrupted(prev))
5963 panic("corrupted stack end detected inside scheduler\n");
5965 if (task_scs_end_corrupted(prev))
5966 panic("corrupted shadow stack detected inside scheduler\n");
5969 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5970 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5971 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5972 prev->comm, prev->pid, prev->non_block_count);
5974 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5978 if (unlikely(in_atomic_preempt_off())) {
5979 __schedule_bug(prev);
5980 preempt_count_set(PREEMPT_DISABLED);
5983 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5985 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5987 schedstat_inc(this_rq()->sched_count);
5990 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5991 struct rq_flags *rf)
5994 const struct sched_class *class;
5996 * We must do the balancing pass before put_prev_task(), such
5997 * that when we release the rq->lock the task is in the same
5998 * state as before we took rq->lock.
6000 * We can terminate the balance pass as soon as we know there is
6001 * a runnable task of @class priority or higher.
6003 for_class_range(class, prev->sched_class, &idle_sched_class) {
6004 if (class->balance(rq, prev, rf))
6009 put_prev_task(rq, prev);
6013 * Pick up the highest-prio task:
6015 static inline struct task_struct *
6016 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6018 const struct sched_class *class;
6019 struct task_struct *p;
6022 * Optimization: we know that if all tasks are in the fair class we can
6023 * call that function directly, but only if the @prev task wasn't of a
6024 * higher scheduling class, because otherwise those lose the
6025 * opportunity to pull in more work from other CPUs.
6027 if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
6028 rq->nr_running == rq->cfs.h_nr_running)) {
6030 p = pick_next_task_fair(rq, prev, rf);
6031 if (unlikely(p == RETRY_TASK))
6034 /* Assume the next prioritized class is idle_sched_class */
6036 put_prev_task(rq, prev);
6037 p = pick_next_task_idle(rq);
6041 * This is the fast path; it cannot be a DL server pick;
6042 * therefore even if @p == @prev, ->dl_server must be NULL.
6045 p->dl_server = NULL;
6051 put_prev_task_balance(rq, prev, rf);
6054 * We've updated @prev and no longer need the server link, clear it.
6055 * Must be done before ->pick_next_task() because that can (re)set
6058 if (prev->dl_server)
6059 prev->dl_server = NULL;
6061 for_each_class(class) {
6062 p = class->pick_next_task(rq);
6067 BUG(); /* The idle class should always have a runnable task. */
6070 #ifdef CONFIG_SCHED_CORE
6071 static inline bool is_task_rq_idle(struct task_struct *t)
6073 return (task_rq(t)->idle == t);
6076 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
6078 return is_task_rq_idle(a) || (a->core_cookie == cookie);
6081 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
6083 if (is_task_rq_idle(a) || is_task_rq_idle(b))
6086 return a->core_cookie == b->core_cookie;
6089 static inline struct task_struct *pick_task(struct rq *rq)
6091 const struct sched_class *class;
6092 struct task_struct *p;
6094 for_each_class(class) {
6095 p = class->pick_task(rq);
6100 BUG(); /* The idle class should always have a runnable task. */
6103 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
6105 static void queue_core_balance(struct rq *rq);
6107 static struct task_struct *
6108 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6110 struct task_struct *next, *p, *max = NULL;
6111 const struct cpumask *smt_mask;
6112 bool fi_before = false;
6113 bool core_clock_updated = (rq == rq->core);
6114 unsigned long cookie;
6115 int i, cpu, occ = 0;
6119 if (!sched_core_enabled(rq))
6120 return __pick_next_task(rq, prev, rf);
6124 /* Stopper task is switching into idle, no need core-wide selection. */
6125 if (cpu_is_offline(cpu)) {
6127 * Reset core_pick so that we don't enter the fastpath when
6128 * coming online. core_pick would already be migrated to
6129 * another cpu during offline.
6131 rq->core_pick = NULL;
6132 return __pick_next_task(rq, prev, rf);
6136 * If there were no {en,de}queues since we picked (IOW, the task
6137 * pointers are all still valid), and we haven't scheduled the last
6138 * pick yet, do so now.
6140 * rq->core_pick can be NULL if no selection was made for a CPU because
6141 * it was either offline or went offline during a sibling's core-wide
6142 * selection. In this case, do a core-wide selection.
6144 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
6145 rq->core->core_pick_seq != rq->core_sched_seq &&
6147 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
6149 next = rq->core_pick;
6151 put_prev_task(rq, prev);
6152 set_next_task(rq, next);
6155 rq->core_pick = NULL;
6159 put_prev_task_balance(rq, prev, rf);
6161 smt_mask = cpu_smt_mask(cpu);
6162 need_sync = !!rq->core->core_cookie;
6165 rq->core->core_cookie = 0UL;
6166 if (rq->core->core_forceidle_count) {
6167 if (!core_clock_updated) {
6168 update_rq_clock(rq->core);
6169 core_clock_updated = true;
6171 sched_core_account_forceidle(rq);
6172 /* reset after accounting force idle */
6173 rq->core->core_forceidle_start = 0;
6174 rq->core->core_forceidle_count = 0;
6175 rq->core->core_forceidle_occupation = 0;
6181 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
6183 * @task_seq guards the task state ({en,de}queues)
6184 * @pick_seq is the @task_seq we did a selection on
6185 * @sched_seq is the @pick_seq we scheduled
6187 * However, preemptions can cause multiple picks on the same task set.
6188 * 'Fix' this by also increasing @task_seq for every pick.
6190 rq->core->core_task_seq++;
6193 * Optimize for common case where this CPU has no cookies
6194 * and there are no cookied tasks running on siblings.
6197 next = pick_task(rq);
6198 if (!next->core_cookie) {
6199 rq->core_pick = NULL;
6201 * For robustness, update the min_vruntime_fi for
6202 * unconstrained picks as well.
6204 WARN_ON_ONCE(fi_before);
6205 task_vruntime_update(rq, next, false);
6211 * For each thread: do the regular task pick and find the max prio task
6214 * Tie-break prio towards the current CPU
6216 for_each_cpu_wrap(i, smt_mask, cpu) {
6220 * Current cpu always has its clock updated on entrance to
6221 * pick_next_task(). If the current cpu is not the core,
6222 * the core may also have been updated above.
6224 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
6225 update_rq_clock(rq_i);
6227 p = rq_i->core_pick = pick_task(rq_i);
6228 if (!max || prio_less(max, p, fi_before))
6232 cookie = rq->core->core_cookie = max->core_cookie;
6235 * For each thread: try and find a runnable task that matches @max or
6238 for_each_cpu(i, smt_mask) {
6240 p = rq_i->core_pick;
6242 if (!cookie_equals(p, cookie)) {
6245 p = sched_core_find(rq_i, cookie);
6247 p = idle_sched_class.pick_task(rq_i);
6250 rq_i->core_pick = p;
6252 if (p == rq_i->idle) {
6253 if (rq_i->nr_running) {
6254 rq->core->core_forceidle_count++;
6256 rq->core->core_forceidle_seq++;
6263 if (schedstat_enabled() && rq->core->core_forceidle_count) {
6264 rq->core->core_forceidle_start = rq_clock(rq->core);
6265 rq->core->core_forceidle_occupation = occ;
6268 rq->core->core_pick_seq = rq->core->core_task_seq;
6269 next = rq->core_pick;
6270 rq->core_sched_seq = rq->core->core_pick_seq;
6272 /* Something should have been selected for current CPU */
6273 WARN_ON_ONCE(!next);
6276 * Reschedule siblings
6278 * NOTE: L1TF -- at this point we're no longer running the old task and
6279 * sending an IPI (below) ensures the sibling will no longer be running
6280 * their task. This ensures there is no inter-sibling overlap between
6281 * non-matching user state.
6283 for_each_cpu(i, smt_mask) {
6287 * An online sibling might have gone offline before a task
6288 * could be picked for it, or it might be offline but later
6289 * happen to come online, but its too late and nothing was
6290 * picked for it. That's Ok - it will pick tasks for itself,
6293 if (!rq_i->core_pick)
6297 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6298 * fi_before fi update?
6304 if (!(fi_before && rq->core->core_forceidle_count))
6305 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6307 rq_i->core_pick->core_occupation = occ;
6310 rq_i->core_pick = NULL;
6314 /* Did we break L1TF mitigation requirements? */
6315 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6317 if (rq_i->curr == rq_i->core_pick) {
6318 rq_i->core_pick = NULL;
6326 set_next_task(rq, next);
6328 if (rq->core->core_forceidle_count && next == rq->idle)
6329 queue_core_balance(rq);
6334 static bool try_steal_cookie(int this, int that)
6336 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6337 struct task_struct *p;
6338 unsigned long cookie;
6339 bool success = false;
6342 guard(double_rq_lock)(dst, src);
6344 cookie = dst->core->core_cookie;
6348 if (dst->curr != dst->idle)
6351 p = sched_core_find(src, cookie);
6356 if (p == src->core_pick || p == src->curr)
6359 if (!is_cpu_allowed(p, this))
6362 if (p->core_occupation > dst->idle->core_occupation)
6365 * sched_core_find() and sched_core_next() will ensure
6366 * that task @p is not throttled now, we also need to
6367 * check whether the runqueue of the destination CPU is
6370 if (sched_task_is_throttled(p, this))
6373 deactivate_task(src, p, 0);
6374 set_task_cpu(p, this);
6375 activate_task(dst, p, 0);
6383 p = sched_core_next(p, cookie);
6389 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6393 for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) {
6400 if (try_steal_cookie(cpu, i))
6407 static void sched_core_balance(struct rq *rq)
6409 struct sched_domain *sd;
6410 int cpu = cpu_of(rq);
6415 raw_spin_rq_unlock_irq(rq);
6416 for_each_domain(cpu, sd) {
6420 if (steal_cookie_task(cpu, sd))
6423 raw_spin_rq_lock_irq(rq);
6426 static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
6428 static void queue_core_balance(struct rq *rq)
6430 if (!sched_core_enabled(rq))
6433 if (!rq->core->core_cookie)
6436 if (!rq->nr_running) /* not forced idle */
6439 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6442 DEFINE_LOCK_GUARD_1(core_lock, int,
6443 sched_core_lock(*_T->lock, &_T->flags),
6444 sched_core_unlock(*_T->lock, &_T->flags),
6445 unsigned long flags)
6447 static void sched_core_cpu_starting(unsigned int cpu)
6449 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6450 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6453 guard(core_lock)(&cpu);
6455 WARN_ON_ONCE(rq->core != rq);
6457 /* if we're the first, we'll be our own leader */
6458 if (cpumask_weight(smt_mask) == 1)
6461 /* find the leader */
6462 for_each_cpu(t, smt_mask) {
6466 if (rq->core == rq) {
6472 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6475 /* install and validate core_rq */
6476 for_each_cpu(t, smt_mask) {
6482 WARN_ON_ONCE(rq->core != core_rq);
6486 static void sched_core_cpu_deactivate(unsigned int cpu)
6488 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6489 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6492 guard(core_lock)(&cpu);
6494 /* if we're the last man standing, nothing to do */
6495 if (cpumask_weight(smt_mask) == 1) {
6496 WARN_ON_ONCE(rq->core != rq);
6500 /* if we're not the leader, nothing to do */
6504 /* find a new leader */
6505 for_each_cpu(t, smt_mask) {
6508 core_rq = cpu_rq(t);
6512 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6515 /* copy the shared state to the new leader */
6516 core_rq->core_task_seq = rq->core_task_seq;
6517 core_rq->core_pick_seq = rq->core_pick_seq;
6518 core_rq->core_cookie = rq->core_cookie;
6519 core_rq->core_forceidle_count = rq->core_forceidle_count;
6520 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6521 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6524 * Accounting edge for forced idle is handled in pick_next_task().
6525 * Don't need another one here, since the hotplug thread shouldn't
6528 core_rq->core_forceidle_start = 0;
6530 /* install new leader */
6531 for_each_cpu(t, smt_mask) {
6537 static inline void sched_core_cpu_dying(unsigned int cpu)
6539 struct rq *rq = cpu_rq(cpu);
6545 #else /* !CONFIG_SCHED_CORE */
6547 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6548 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6549 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6551 static struct task_struct *
6552 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6554 return __pick_next_task(rq, prev, rf);
6557 #endif /* CONFIG_SCHED_CORE */
6560 * Constants for the sched_mode argument of __schedule().
6562 * The mode argument allows RT enabled kernels to differentiate a
6563 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6564 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6565 * optimize the AND operation out and just check for zero.
6568 #define SM_PREEMPT 0x1
6569 #define SM_RTLOCK_WAIT 0x2
6571 #ifndef CONFIG_PREEMPT_RT
6572 # define SM_MASK_PREEMPT (~0U)
6574 # define SM_MASK_PREEMPT SM_PREEMPT
6578 * __schedule() is the main scheduler function.
6580 * The main means of driving the scheduler and thus entering this function are:
6582 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6584 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6585 * paths. For example, see arch/x86/entry_64.S.
6587 * To drive preemption between tasks, the scheduler sets the flag in timer
6588 * interrupt handler scheduler_tick().
6590 * 3. Wakeups don't really cause entry into schedule(). They add a
6591 * task to the run-queue and that's it.
6593 * Now, if the new task added to the run-queue preempts the current
6594 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6595 * called on the nearest possible occasion:
6597 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6599 * - in syscall or exception context, at the next outmost
6600 * preempt_enable(). (this might be as soon as the wake_up()'s
6603 * - in IRQ context, return from interrupt-handler to
6604 * preemptible context
6606 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6609 * - cond_resched() call
6610 * - explicit schedule() call
6611 * - return from syscall or exception to user-space
6612 * - return from interrupt-handler to user-space
6614 * WARNING: must be called with preemption disabled!
6616 static void __sched notrace __schedule(unsigned int sched_mode)
6618 struct task_struct *prev, *next;
6619 unsigned long *switch_count;
6620 unsigned long prev_state;
6625 cpu = smp_processor_id();
6629 schedule_debug(prev, !!sched_mode);
6631 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6634 local_irq_disable();
6635 rcu_note_context_switch(!!sched_mode);
6638 * Make sure that signal_pending_state()->signal_pending() below
6639 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6640 * done by the caller to avoid the race with signal_wake_up():
6642 * __set_current_state(@state) signal_wake_up()
6643 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6644 * wake_up_state(p, state)
6645 * LOCK rq->lock LOCK p->pi_state
6646 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6647 * if (signal_pending_state()) if (p->state & @state)
6649 * Also, the membarrier system call requires a full memory barrier
6650 * after coming from user-space, before storing to rq->curr; this
6651 * barrier matches a full barrier in the proximity of the membarrier
6655 smp_mb__after_spinlock();
6657 /* Promote REQ to ACT */
6658 rq->clock_update_flags <<= 1;
6659 update_rq_clock(rq);
6660 rq->clock_update_flags = RQCF_UPDATED;
6662 switch_count = &prev->nivcsw;
6665 * We must load prev->state once (task_struct::state is volatile), such
6666 * that we form a control dependency vs deactivate_task() below.
6668 prev_state = READ_ONCE(prev->__state);
6669 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6670 if (signal_pending_state(prev_state, prev)) {
6671 WRITE_ONCE(prev->__state, TASK_RUNNING);
6673 prev->sched_contributes_to_load =
6674 (prev_state & TASK_UNINTERRUPTIBLE) &&
6675 !(prev_state & TASK_NOLOAD) &&
6676 !(prev_state & TASK_FROZEN);
6678 if (prev->sched_contributes_to_load)
6679 rq->nr_uninterruptible++;
6682 * __schedule() ttwu()
6683 * prev_state = prev->state; if (p->on_rq && ...)
6684 * if (prev_state) goto out;
6685 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6686 * p->state = TASK_WAKING
6688 * Where __schedule() and ttwu() have matching control dependencies.
6690 * After this, schedule() must not care about p->state any more.
6692 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6694 if (prev->in_iowait) {
6695 atomic_inc(&rq->nr_iowait);
6696 delayacct_blkio_start();
6699 switch_count = &prev->nvcsw;
6702 next = pick_next_task(rq, prev, &rf);
6703 clear_tsk_need_resched(prev);
6704 clear_preempt_need_resched();
6705 #ifdef CONFIG_SCHED_DEBUG
6706 rq->last_seen_need_resched_ns = 0;
6709 if (likely(prev != next)) {
6712 * RCU users of rcu_dereference(rq->curr) may not see
6713 * changes to task_struct made by pick_next_task().
6715 RCU_INIT_POINTER(rq->curr, next);
6717 * The membarrier system call requires each architecture
6718 * to have a full memory barrier after updating
6719 * rq->curr, before returning to user-space.
6721 * Here are the schemes providing that barrier on the
6722 * various architectures:
6723 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC,
6724 * RISC-V. switch_mm() relies on membarrier_arch_switch_mm()
6725 * on PowerPC and on RISC-V.
6726 * - finish_lock_switch() for weakly-ordered
6727 * architectures where spin_unlock is a full barrier,
6728 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6729 * is a RELEASE barrier),
6731 * The barrier matches a full barrier in the proximity of
6732 * the membarrier system call entry.
6734 * On RISC-V, this barrier pairing is also needed for the
6735 * SYNC_CORE command when switching between processes, cf.
6736 * the inline comments in membarrier_arch_switch_mm().
6740 migrate_disable_switch(rq, prev);
6741 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6743 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6745 /* Also unlocks the rq: */
6746 rq = context_switch(rq, prev, next, &rf);
6748 rq_unpin_lock(rq, &rf);
6749 __balance_callbacks(rq);
6750 raw_spin_rq_unlock_irq(rq);
6754 void __noreturn do_task_dead(void)
6756 /* Causes final put_task_struct in finish_task_switch(): */
6757 set_special_state(TASK_DEAD);
6759 /* Tell freezer to ignore us: */
6760 current->flags |= PF_NOFREEZE;
6762 __schedule(SM_NONE);
6765 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6770 static inline void sched_submit_work(struct task_struct *tsk)
6772 static DEFINE_WAIT_OVERRIDE_MAP(sched_map, LD_WAIT_CONFIG);
6773 unsigned int task_flags;
6776 * Establish LD_WAIT_CONFIG context to ensure none of the code called
6777 * will use a blocking primitive -- which would lead to recursion.
6779 lock_map_acquire_try(&sched_map);
6781 task_flags = tsk->flags;
6783 * If a worker goes to sleep, notify and ask workqueue whether it
6784 * wants to wake up a task to maintain concurrency.
6786 if (task_flags & PF_WQ_WORKER)
6787 wq_worker_sleeping(tsk);
6788 else if (task_flags & PF_IO_WORKER)
6789 io_wq_worker_sleeping(tsk);
6792 * spinlock and rwlock must not flush block requests. This will
6793 * deadlock if the callback attempts to acquire a lock which is
6796 SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6799 * If we are going to sleep and we have plugged IO queued,
6800 * make sure to submit it to avoid deadlocks.
6802 blk_flush_plug(tsk->plug, true);
6804 lock_map_release(&sched_map);
6807 static void sched_update_worker(struct task_struct *tsk)
6809 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER | PF_BLOCK_TS)) {
6810 if (tsk->flags & PF_BLOCK_TS)
6811 blk_plug_invalidate_ts(tsk);
6812 if (tsk->flags & PF_WQ_WORKER)
6813 wq_worker_running(tsk);
6814 else if (tsk->flags & PF_IO_WORKER)
6815 io_wq_worker_running(tsk);
6819 static __always_inline void __schedule_loop(unsigned int sched_mode)
6823 __schedule(sched_mode);
6824 sched_preempt_enable_no_resched();
6825 } while (need_resched());
6828 asmlinkage __visible void __sched schedule(void)
6830 struct task_struct *tsk = current;
6832 #ifdef CONFIG_RT_MUTEXES
6833 lockdep_assert(!tsk->sched_rt_mutex);
6836 if (!task_is_running(tsk))
6837 sched_submit_work(tsk);
6838 __schedule_loop(SM_NONE);
6839 sched_update_worker(tsk);
6841 EXPORT_SYMBOL(schedule);
6844 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6845 * state (have scheduled out non-voluntarily) by making sure that all
6846 * tasks have either left the run queue or have gone into user space.
6847 * As idle tasks do not do either, they must not ever be preempted
6848 * (schedule out non-voluntarily).
6850 * schedule_idle() is similar to schedule_preempt_disable() except that it
6851 * never enables preemption because it does not call sched_submit_work().
6853 void __sched schedule_idle(void)
6856 * As this skips calling sched_submit_work(), which the idle task does
6857 * regardless because that function is a nop when the task is in a
6858 * TASK_RUNNING state, make sure this isn't used someplace that the
6859 * current task can be in any other state. Note, idle is always in the
6860 * TASK_RUNNING state.
6862 WARN_ON_ONCE(current->__state);
6864 __schedule(SM_NONE);
6865 } while (need_resched());
6868 #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6869 asmlinkage __visible void __sched schedule_user(void)
6872 * If we come here after a random call to set_need_resched(),
6873 * or we have been woken up remotely but the IPI has not yet arrived,
6874 * we haven't yet exited the RCU idle mode. Do it here manually until
6875 * we find a better solution.
6877 * NB: There are buggy callers of this function. Ideally we
6878 * should warn if prev_state != CONTEXT_USER, but that will trigger
6879 * too frequently to make sense yet.
6881 enum ctx_state prev_state = exception_enter();
6883 exception_exit(prev_state);
6888 * schedule_preempt_disabled - called with preemption disabled
6890 * Returns with preemption disabled. Note: preempt_count must be 1
6892 void __sched schedule_preempt_disabled(void)
6894 sched_preempt_enable_no_resched();
6899 #ifdef CONFIG_PREEMPT_RT
6900 void __sched notrace schedule_rtlock(void)
6902 __schedule_loop(SM_RTLOCK_WAIT);
6904 NOKPROBE_SYMBOL(schedule_rtlock);
6907 static void __sched notrace preempt_schedule_common(void)
6911 * Because the function tracer can trace preempt_count_sub()
6912 * and it also uses preempt_enable/disable_notrace(), if
6913 * NEED_RESCHED is set, the preempt_enable_notrace() called
6914 * by the function tracer will call this function again and
6915 * cause infinite recursion.
6917 * Preemption must be disabled here before the function
6918 * tracer can trace. Break up preempt_disable() into two
6919 * calls. One to disable preemption without fear of being
6920 * traced. The other to still record the preemption latency,
6921 * which can also be traced by the function tracer.
6923 preempt_disable_notrace();
6924 preempt_latency_start(1);
6925 __schedule(SM_PREEMPT);
6926 preempt_latency_stop(1);
6927 preempt_enable_no_resched_notrace();
6930 * Check again in case we missed a preemption opportunity
6931 * between schedule and now.
6933 } while (need_resched());
6936 #ifdef CONFIG_PREEMPTION
6938 * This is the entry point to schedule() from in-kernel preemption
6939 * off of preempt_enable.
6941 asmlinkage __visible void __sched notrace preempt_schedule(void)
6944 * If there is a non-zero preempt_count or interrupts are disabled,
6945 * we do not want to preempt the current task. Just return..
6947 if (likely(!preemptible()))
6949 preempt_schedule_common();
6951 NOKPROBE_SYMBOL(preempt_schedule);
6952 EXPORT_SYMBOL(preempt_schedule);
6954 #ifdef CONFIG_PREEMPT_DYNAMIC
6955 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6956 #ifndef preempt_schedule_dynamic_enabled
6957 #define preempt_schedule_dynamic_enabled preempt_schedule
6958 #define preempt_schedule_dynamic_disabled NULL
6960 DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6961 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6962 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6963 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6964 void __sched notrace dynamic_preempt_schedule(void)
6966 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6970 NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6971 EXPORT_SYMBOL(dynamic_preempt_schedule);
6976 * preempt_schedule_notrace - preempt_schedule called by tracing
6978 * The tracing infrastructure uses preempt_enable_notrace to prevent
6979 * recursion and tracing preempt enabling caused by the tracing
6980 * infrastructure itself. But as tracing can happen in areas coming
6981 * from userspace or just about to enter userspace, a preempt enable
6982 * can occur before user_exit() is called. This will cause the scheduler
6983 * to be called when the system is still in usermode.
6985 * To prevent this, the preempt_enable_notrace will use this function
6986 * instead of preempt_schedule() to exit user context if needed before
6987 * calling the scheduler.
6989 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6991 enum ctx_state prev_ctx;
6993 if (likely(!preemptible()))
6998 * Because the function tracer can trace preempt_count_sub()
6999 * and it also uses preempt_enable/disable_notrace(), if
7000 * NEED_RESCHED is set, the preempt_enable_notrace() called
7001 * by the function tracer will call this function again and
7002 * cause infinite recursion.
7004 * Preemption must be disabled here before the function
7005 * tracer can trace. Break up preempt_disable() into two
7006 * calls. One to disable preemption without fear of being
7007 * traced. The other to still record the preemption latency,
7008 * which can also be traced by the function tracer.
7010 preempt_disable_notrace();
7011 preempt_latency_start(1);
7013 * Needs preempt disabled in case user_exit() is traced
7014 * and the tracer calls preempt_enable_notrace() causing
7015 * an infinite recursion.
7017 prev_ctx = exception_enter();
7018 __schedule(SM_PREEMPT);
7019 exception_exit(prev_ctx);
7021 preempt_latency_stop(1);
7022 preempt_enable_no_resched_notrace();
7023 } while (need_resched());
7025 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
7027 #ifdef CONFIG_PREEMPT_DYNAMIC
7028 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
7029 #ifndef preempt_schedule_notrace_dynamic_enabled
7030 #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
7031 #define preempt_schedule_notrace_dynamic_disabled NULL
7033 DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
7034 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
7035 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
7036 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
7037 void __sched notrace dynamic_preempt_schedule_notrace(void)
7039 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
7041 preempt_schedule_notrace();
7043 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
7044 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
7048 #endif /* CONFIG_PREEMPTION */
7051 * This is the entry point to schedule() from kernel preemption
7052 * off of irq context.
7053 * Note, that this is called and return with irqs disabled. This will
7054 * protect us against recursive calling from irq.
7056 asmlinkage __visible void __sched preempt_schedule_irq(void)
7058 enum ctx_state prev_state;
7060 /* Catch callers which need to be fixed */
7061 BUG_ON(preempt_count() || !irqs_disabled());
7063 prev_state = exception_enter();
7068 __schedule(SM_PREEMPT);
7069 local_irq_disable();
7070 sched_preempt_enable_no_resched();
7071 } while (need_resched());
7073 exception_exit(prev_state);
7076 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
7079 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~(WF_SYNC|WF_CURRENT_CPU));
7080 return try_to_wake_up(curr->private, mode, wake_flags);
7082 EXPORT_SYMBOL(default_wake_function);
7084 static void __setscheduler_prio(struct task_struct *p, int prio)
7087 p->sched_class = &dl_sched_class;
7088 else if (rt_prio(prio))
7089 p->sched_class = &rt_sched_class;
7091 p->sched_class = &fair_sched_class;
7096 #ifdef CONFIG_RT_MUTEXES
7099 * Would be more useful with typeof()/auto_type but they don't mix with
7100 * bit-fields. Since it's a local thing, use int. Keep the generic sounding
7101 * name such that if someone were to implement this function we get to compare
7104 #define fetch_and_set(x, v) ({ int _x = (x); (x) = (v); _x; })
7106 void rt_mutex_pre_schedule(void)
7108 lockdep_assert(!fetch_and_set(current->sched_rt_mutex, 1));
7109 sched_submit_work(current);
7112 void rt_mutex_schedule(void)
7114 lockdep_assert(current->sched_rt_mutex);
7115 __schedule_loop(SM_NONE);
7118 void rt_mutex_post_schedule(void)
7120 sched_update_worker(current);
7121 lockdep_assert(fetch_and_set(current->sched_rt_mutex, 0));
7124 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
7127 prio = min(prio, pi_task->prio);
7132 static inline int rt_effective_prio(struct task_struct *p, int prio)
7134 struct task_struct *pi_task = rt_mutex_get_top_task(p);
7136 return __rt_effective_prio(pi_task, prio);
7140 * rt_mutex_setprio - set the current priority of a task
7142 * @pi_task: donor task
7144 * This function changes the 'effective' priority of a task. It does
7145 * not touch ->normal_prio like __setscheduler().
7147 * Used by the rt_mutex code to implement priority inheritance
7148 * logic. Call site only calls if the priority of the task changed.
7150 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
7152 int prio, oldprio, queued, running, queue_flag =
7153 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7154 const struct sched_class *prev_class;
7158 /* XXX used to be waiter->prio, not waiter->task->prio */
7159 prio = __rt_effective_prio(pi_task, p->normal_prio);
7162 * If nothing changed; bail early.
7164 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
7167 rq = __task_rq_lock(p, &rf);
7168 update_rq_clock(rq);
7170 * Set under pi_lock && rq->lock, such that the value can be used under
7173 * Note that there is loads of tricky to make this pointer cache work
7174 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
7175 * ensure a task is de-boosted (pi_task is set to NULL) before the
7176 * task is allowed to run again (and can exit). This ensures the pointer
7177 * points to a blocked task -- which guarantees the task is present.
7179 p->pi_top_task = pi_task;
7182 * For FIFO/RR we only need to set prio, if that matches we're done.
7184 if (prio == p->prio && !dl_prio(prio))
7188 * Idle task boosting is a nono in general. There is one
7189 * exception, when PREEMPT_RT and NOHZ is active:
7191 * The idle task calls get_next_timer_interrupt() and holds
7192 * the timer wheel base->lock on the CPU and another CPU wants
7193 * to access the timer (probably to cancel it). We can safely
7194 * ignore the boosting request, as the idle CPU runs this code
7195 * with interrupts disabled and will complete the lock
7196 * protected section without being interrupted. So there is no
7197 * real need to boost.
7199 if (unlikely(p == rq->idle)) {
7200 WARN_ON(p != rq->curr);
7201 WARN_ON(p->pi_blocked_on);
7205 trace_sched_pi_setprio(p, pi_task);
7208 if (oldprio == prio)
7209 queue_flag &= ~DEQUEUE_MOVE;
7211 prev_class = p->sched_class;
7212 queued = task_on_rq_queued(p);
7213 running = task_current(rq, p);
7215 dequeue_task(rq, p, queue_flag);
7217 put_prev_task(rq, p);
7220 * Boosting condition are:
7221 * 1. -rt task is running and holds mutex A
7222 * --> -dl task blocks on mutex A
7224 * 2. -dl task is running and holds mutex A
7225 * --> -dl task blocks on mutex A and could preempt the
7228 if (dl_prio(prio)) {
7229 if (!dl_prio(p->normal_prio) ||
7230 (pi_task && dl_prio(pi_task->prio) &&
7231 dl_entity_preempt(&pi_task->dl, &p->dl))) {
7232 p->dl.pi_se = pi_task->dl.pi_se;
7233 queue_flag |= ENQUEUE_REPLENISH;
7235 p->dl.pi_se = &p->dl;
7237 } else if (rt_prio(prio)) {
7238 if (dl_prio(oldprio))
7239 p->dl.pi_se = &p->dl;
7241 queue_flag |= ENQUEUE_HEAD;
7243 if (dl_prio(oldprio))
7244 p->dl.pi_se = &p->dl;
7245 if (rt_prio(oldprio))
7249 __setscheduler_prio(p, prio);
7252 enqueue_task(rq, p, queue_flag);
7254 set_next_task(rq, p);
7256 check_class_changed(rq, p, prev_class, oldprio);
7258 /* Avoid rq from going away on us: */
7261 rq_unpin_lock(rq, &rf);
7262 __balance_callbacks(rq);
7263 raw_spin_rq_unlock(rq);
7268 static inline int rt_effective_prio(struct task_struct *p, int prio)
7274 void set_user_nice(struct task_struct *p, long nice)
7276 bool queued, running;
7280 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
7283 * We have to be careful, if called from sys_setpriority(),
7284 * the task might be in the middle of scheduling on another CPU.
7286 CLASS(task_rq_lock, rq_guard)(p);
7289 update_rq_clock(rq);
7292 * The RT priorities are set via sched_setscheduler(), but we still
7293 * allow the 'normal' nice value to be set - but as expected
7294 * it won't have any effect on scheduling until the task is
7295 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7297 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7298 p->static_prio = NICE_TO_PRIO(nice);
7302 queued = task_on_rq_queued(p);
7303 running = task_current(rq, p);
7305 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7307 put_prev_task(rq, p);
7309 p->static_prio = NICE_TO_PRIO(nice);
7310 set_load_weight(p, true);
7312 p->prio = effective_prio(p);
7315 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7317 set_next_task(rq, p);
7320 * If the task increased its priority or is running and
7321 * lowered its priority, then reschedule its CPU:
7323 p->sched_class->prio_changed(rq, p, old_prio);
7325 EXPORT_SYMBOL(set_user_nice);
7328 * is_nice_reduction - check if nice value is an actual reduction
7330 * Similar to can_nice() but does not perform a capability check.
7335 static bool is_nice_reduction(const struct task_struct *p, const int nice)
7337 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7338 int nice_rlim = nice_to_rlimit(nice);
7340 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7344 * can_nice - check if a task can reduce its nice value
7348 int can_nice(const struct task_struct *p, const int nice)
7350 return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7353 #ifdef __ARCH_WANT_SYS_NICE
7356 * sys_nice - change the priority of the current process.
7357 * @increment: priority increment
7359 * sys_setpriority is a more generic, but much slower function that
7360 * does similar things.
7362 SYSCALL_DEFINE1(nice, int, increment)
7367 * Setpriority might change our priority at the same moment.
7368 * We don't have to worry. Conceptually one call occurs first
7369 * and we have a single winner.
7371 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7372 nice = task_nice(current) + increment;
7374 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7375 if (increment < 0 && !can_nice(current, nice))
7378 retval = security_task_setnice(current, nice);
7382 set_user_nice(current, nice);
7389 * task_prio - return the priority value of a given task.
7390 * @p: the task in question.
7392 * Return: The priority value as seen by users in /proc.
7394 * sched policy return value kernel prio user prio/nice
7396 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7397 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7398 * deadline -101 -1 0
7400 int task_prio(const struct task_struct *p)
7402 return p->prio - MAX_RT_PRIO;
7406 * idle_cpu - is a given CPU idle currently?
7407 * @cpu: the processor in question.
7409 * Return: 1 if the CPU is currently idle. 0 otherwise.
7411 int idle_cpu(int cpu)
7413 struct rq *rq = cpu_rq(cpu);
7415 if (rq->curr != rq->idle)
7422 if (rq->ttwu_pending)
7430 * available_idle_cpu - is a given CPU idle for enqueuing work.
7431 * @cpu: the CPU in question.
7433 * Return: 1 if the CPU is currently idle. 0 otherwise.
7435 int available_idle_cpu(int cpu)
7440 if (vcpu_is_preempted(cpu))
7447 * idle_task - return the idle task for a given CPU.
7448 * @cpu: the processor in question.
7450 * Return: The idle task for the CPU @cpu.
7452 struct task_struct *idle_task(int cpu)
7454 return cpu_rq(cpu)->idle;
7457 #ifdef CONFIG_SCHED_CORE
7458 int sched_core_idle_cpu(int cpu)
7460 struct rq *rq = cpu_rq(cpu);
7462 if (sched_core_enabled(rq) && rq->curr == rq->idle)
7465 return idle_cpu(cpu);
7472 * This function computes an effective utilization for the given CPU, to be
7473 * used for frequency selection given the linear relation: f = u * f_max.
7475 * The scheduler tracks the following metrics:
7477 * cpu_util_{cfs,rt,dl,irq}()
7480 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7481 * synchronized windows and are thus directly comparable.
7483 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7484 * which excludes things like IRQ and steal-time. These latter are then accrued
7485 * in the irq utilization.
7487 * The DL bandwidth number otoh is not a measured metric but a value computed
7488 * based on the task model parameters and gives the minimal utilization
7489 * required to meet deadlines.
7491 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7495 unsigned long util, irq, scale;
7496 struct rq *rq = cpu_rq(cpu);
7498 scale = arch_scale_cpu_capacity(cpu);
7501 * Early check to see if IRQ/steal time saturates the CPU, can be
7502 * because of inaccuracies in how we track these -- see
7503 * update_irq_load_avg().
7505 irq = cpu_util_irq(rq);
7506 if (unlikely(irq >= scale)) {
7516 * The minimum utilization returns the highest level between:
7517 * - the computed DL bandwidth needed with the IRQ pressure which
7518 * steals time to the deadline task.
7519 * - The minimum performance requirement for CFS and/or RT.
7521 *min = max(irq + cpu_bw_dl(rq), uclamp_rq_get(rq, UCLAMP_MIN));
7524 * When an RT task is runnable and uclamp is not used, we must
7525 * ensure that the task will run at maximum compute capacity.
7527 if (!uclamp_is_used() && rt_rq_is_runnable(&rq->rt))
7528 *min = max(*min, scale);
7532 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7533 * CFS tasks and we use the same metric to track the effective
7534 * utilization (PELT windows are synchronized) we can directly add them
7535 * to obtain the CPU's actual utilization.
7537 util = util_cfs + cpu_util_rt(rq);
7538 util += cpu_util_dl(rq);
7541 * The maximum hint is a soft bandwidth requirement, which can be lower
7542 * than the actual utilization because of uclamp_max requirements.
7545 *max = min(scale, uclamp_rq_get(rq, UCLAMP_MAX));
7551 * There is still idle time; further improve the number by using the
7552 * irq metric. Because IRQ/steal time is hidden from the task clock we
7553 * need to scale the task numbers:
7556 * U' = irq + --------- * U
7559 util = scale_irq_capacity(util, irq, scale);
7562 return min(scale, util);
7565 unsigned long sched_cpu_util(int cpu)
7567 return effective_cpu_util(cpu, cpu_util_cfs(cpu), NULL, NULL);
7569 #endif /* CONFIG_SMP */
7572 * find_process_by_pid - find a process with a matching PID value.
7573 * @pid: the pid in question.
7575 * The task of @pid, if found. %NULL otherwise.
7577 static struct task_struct *find_process_by_pid(pid_t pid)
7579 return pid ? find_task_by_vpid(pid) : current;
7582 static struct task_struct *find_get_task(pid_t pid)
7584 struct task_struct *p;
7587 p = find_process_by_pid(pid);
7594 DEFINE_CLASS(find_get_task, struct task_struct *, if (_T) put_task_struct(_T),
7595 find_get_task(pid), pid_t pid)
7598 * sched_setparam() passes in -1 for its policy, to let the functions
7599 * it calls know not to change it.
7601 #define SETPARAM_POLICY -1
7603 static void __setscheduler_params(struct task_struct *p,
7604 const struct sched_attr *attr)
7606 int policy = attr->sched_policy;
7608 if (policy == SETPARAM_POLICY)
7613 if (dl_policy(policy))
7614 __setparam_dl(p, attr);
7615 else if (fair_policy(policy))
7616 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7619 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7620 * !rt_policy. Always setting this ensures that things like
7621 * getparam()/getattr() don't report silly values for !rt tasks.
7623 p->rt_priority = attr->sched_priority;
7624 p->normal_prio = normal_prio(p);
7625 set_load_weight(p, true);
7629 * Check the target process has a UID that matches the current process's:
7631 static bool check_same_owner(struct task_struct *p)
7633 const struct cred *cred = current_cred(), *pcred;
7636 pcred = __task_cred(p);
7637 return (uid_eq(cred->euid, pcred->euid) ||
7638 uid_eq(cred->euid, pcred->uid));
7642 * Allow unprivileged RT tasks to decrease priority.
7643 * Only issue a capable test if needed and only once to avoid an audit
7644 * event on permitted non-privileged operations:
7646 static int user_check_sched_setscheduler(struct task_struct *p,
7647 const struct sched_attr *attr,
7648 int policy, int reset_on_fork)
7650 if (fair_policy(policy)) {
7651 if (attr->sched_nice < task_nice(p) &&
7652 !is_nice_reduction(p, attr->sched_nice))
7656 if (rt_policy(policy)) {
7657 unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7659 /* Can't set/change the rt policy: */
7660 if (policy != p->policy && !rlim_rtprio)
7663 /* Can't increase priority: */
7664 if (attr->sched_priority > p->rt_priority &&
7665 attr->sched_priority > rlim_rtprio)
7670 * Can't set/change SCHED_DEADLINE policy at all for now
7671 * (safest behavior); in the future we would like to allow
7672 * unprivileged DL tasks to increase their relative deadline
7673 * or reduce their runtime (both ways reducing utilization)
7675 if (dl_policy(policy))
7679 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7680 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7682 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7683 if (!is_nice_reduction(p, task_nice(p)))
7687 /* Can't change other user's priorities: */
7688 if (!check_same_owner(p))
7691 /* Normal users shall not reset the sched_reset_on_fork flag: */
7692 if (p->sched_reset_on_fork && !reset_on_fork)
7698 if (!capable(CAP_SYS_NICE))
7704 static int __sched_setscheduler(struct task_struct *p,
7705 const struct sched_attr *attr,
7708 int oldpolicy = -1, policy = attr->sched_policy;
7709 int retval, oldprio, newprio, queued, running;
7710 const struct sched_class *prev_class;
7711 struct balance_callback *head;
7714 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7716 bool cpuset_locked = false;
7718 /* The pi code expects interrupts enabled */
7719 BUG_ON(pi && in_interrupt());
7721 /* Double check policy once rq lock held: */
7723 reset_on_fork = p->sched_reset_on_fork;
7724 policy = oldpolicy = p->policy;
7726 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7728 if (!valid_policy(policy))
7732 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7736 * Valid priorities for SCHED_FIFO and SCHED_RR are
7737 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7738 * SCHED_BATCH and SCHED_IDLE is 0.
7740 if (attr->sched_priority > MAX_RT_PRIO-1)
7742 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7743 (rt_policy(policy) != (attr->sched_priority != 0)))
7747 retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7751 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7754 retval = security_task_setscheduler(p);
7759 /* Update task specific "requested" clamps */
7760 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7761 retval = uclamp_validate(p, attr);
7767 * SCHED_DEADLINE bandwidth accounting relies on stable cpusets
7770 if (dl_policy(policy) || dl_policy(p->policy)) {
7771 cpuset_locked = true;
7776 * Make sure no PI-waiters arrive (or leave) while we are
7777 * changing the priority of the task:
7779 * To be able to change p->policy safely, the appropriate
7780 * runqueue lock must be held.
7782 rq = task_rq_lock(p, &rf);
7783 update_rq_clock(rq);
7786 * Changing the policy of the stop threads its a very bad idea:
7788 if (p == rq->stop) {
7794 * If not changing anything there's no need to proceed further,
7795 * but store a possible modification of reset_on_fork.
7797 if (unlikely(policy == p->policy)) {
7798 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7800 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7802 if (dl_policy(policy) && dl_param_changed(p, attr))
7804 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7807 p->sched_reset_on_fork = reset_on_fork;
7814 #ifdef CONFIG_RT_GROUP_SCHED
7816 * Do not allow realtime tasks into groups that have no runtime
7819 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7820 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7821 !task_group_is_autogroup(task_group(p))) {
7827 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7828 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7829 cpumask_t *span = rq->rd->span;
7832 * Don't allow tasks with an affinity mask smaller than
7833 * the entire root_domain to become SCHED_DEADLINE. We
7834 * will also fail if there's no bandwidth available.
7836 if (!cpumask_subset(span, p->cpus_ptr) ||
7837 rq->rd->dl_bw.bw == 0) {
7845 /* Re-check policy now with rq lock held: */
7846 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7847 policy = oldpolicy = -1;
7848 task_rq_unlock(rq, p, &rf);
7855 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7856 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7859 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7864 p->sched_reset_on_fork = reset_on_fork;
7867 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7870 * Take priority boosted tasks into account. If the new
7871 * effective priority is unchanged, we just store the new
7872 * normal parameters and do not touch the scheduler class and
7873 * the runqueue. This will be done when the task deboost
7876 newprio = rt_effective_prio(p, newprio);
7877 if (newprio == oldprio)
7878 queue_flags &= ~DEQUEUE_MOVE;
7881 queued = task_on_rq_queued(p);
7882 running = task_current(rq, p);
7884 dequeue_task(rq, p, queue_flags);
7886 put_prev_task(rq, p);
7888 prev_class = p->sched_class;
7890 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7891 __setscheduler_params(p, attr);
7892 __setscheduler_prio(p, newprio);
7894 __setscheduler_uclamp(p, attr);
7898 * We enqueue to tail when the priority of a task is
7899 * increased (user space view).
7901 if (oldprio < p->prio)
7902 queue_flags |= ENQUEUE_HEAD;
7904 enqueue_task(rq, p, queue_flags);
7907 set_next_task(rq, p);
7909 check_class_changed(rq, p, prev_class, oldprio);
7911 /* Avoid rq from going away on us: */
7913 head = splice_balance_callbacks(rq);
7914 task_rq_unlock(rq, p, &rf);
7919 rt_mutex_adjust_pi(p);
7922 /* Run balance callbacks after we've adjusted the PI chain: */
7923 balance_callbacks(rq, head);
7929 task_rq_unlock(rq, p, &rf);
7935 static int _sched_setscheduler(struct task_struct *p, int policy,
7936 const struct sched_param *param, bool check)
7938 struct sched_attr attr = {
7939 .sched_policy = policy,
7940 .sched_priority = param->sched_priority,
7941 .sched_nice = PRIO_TO_NICE(p->static_prio),
7944 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7945 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7946 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7947 policy &= ~SCHED_RESET_ON_FORK;
7948 attr.sched_policy = policy;
7951 return __sched_setscheduler(p, &attr, check, true);
7954 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7955 * @p: the task in question.
7956 * @policy: new policy.
7957 * @param: structure containing the new RT priority.
7959 * Use sched_set_fifo(), read its comment.
7961 * Return: 0 on success. An error code otherwise.
7963 * NOTE that the task may be already dead.
7965 int sched_setscheduler(struct task_struct *p, int policy,
7966 const struct sched_param *param)
7968 return _sched_setscheduler(p, policy, param, true);
7971 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7973 return __sched_setscheduler(p, attr, true, true);
7976 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7978 return __sched_setscheduler(p, attr, false, true);
7980 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7983 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7984 * @p: the task in question.
7985 * @policy: new policy.
7986 * @param: structure containing the new RT priority.
7988 * Just like sched_setscheduler, only don't bother checking if the
7989 * current context has permission. For example, this is needed in
7990 * stop_machine(): we create temporary high priority worker threads,
7991 * but our caller might not have that capability.
7993 * Return: 0 on success. An error code otherwise.
7995 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7996 const struct sched_param *param)
7998 return _sched_setscheduler(p, policy, param, false);
8002 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
8003 * incapable of resource management, which is the one thing an OS really should
8006 * This is of course the reason it is limited to privileged users only.
8008 * Worse still; it is fundamentally impossible to compose static priority
8009 * workloads. You cannot take two correctly working static prio workloads
8010 * and smash them together and still expect them to work.
8012 * For this reason 'all' FIFO tasks the kernel creates are basically at:
8016 * The administrator _MUST_ configure the system, the kernel simply doesn't
8017 * know enough information to make a sensible choice.
8019 void sched_set_fifo(struct task_struct *p)
8021 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
8022 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
8024 EXPORT_SYMBOL_GPL(sched_set_fifo);
8027 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
8029 void sched_set_fifo_low(struct task_struct *p)
8031 struct sched_param sp = { .sched_priority = 1 };
8032 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
8034 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
8036 void sched_set_normal(struct task_struct *p, int nice)
8038 struct sched_attr attr = {
8039 .sched_policy = SCHED_NORMAL,
8042 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
8044 EXPORT_SYMBOL_GPL(sched_set_normal);
8047 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
8049 struct sched_param lparam;
8051 if (!param || pid < 0)
8053 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
8056 CLASS(find_get_task, p)(pid);
8060 return sched_setscheduler(p, policy, &lparam);
8064 * Mimics kernel/events/core.c perf_copy_attr().
8066 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
8071 /* Zero the full structure, so that a short copy will be nice: */
8072 memset(attr, 0, sizeof(*attr));
8074 ret = get_user(size, &uattr->size);
8078 /* ABI compatibility quirk: */
8080 size = SCHED_ATTR_SIZE_VER0;
8081 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
8084 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
8091 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
8092 size < SCHED_ATTR_SIZE_VER1)
8096 * XXX: Do we want to be lenient like existing syscalls; or do we want
8097 * to be strict and return an error on out-of-bounds values?
8099 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
8104 put_user(sizeof(*attr), &uattr->size);
8108 static void get_params(struct task_struct *p, struct sched_attr *attr)
8110 if (task_has_dl_policy(p))
8111 __getparam_dl(p, attr);
8112 else if (task_has_rt_policy(p))
8113 attr->sched_priority = p->rt_priority;
8115 attr->sched_nice = task_nice(p);
8119 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
8120 * @pid: the pid in question.
8121 * @policy: new policy.
8122 * @param: structure containing the new RT priority.
8124 * Return: 0 on success. An error code otherwise.
8126 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
8131 return do_sched_setscheduler(pid, policy, param);
8135 * sys_sched_setparam - set/change the RT priority of a thread
8136 * @pid: the pid in question.
8137 * @param: structure containing the new RT priority.
8139 * Return: 0 on success. An error code otherwise.
8141 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
8143 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
8147 * sys_sched_setattr - same as above, but with extended sched_attr
8148 * @pid: the pid in question.
8149 * @uattr: structure containing the extended parameters.
8150 * @flags: for future extension.
8152 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
8153 unsigned int, flags)
8155 struct sched_attr attr;
8158 if (!uattr || pid < 0 || flags)
8161 retval = sched_copy_attr(uattr, &attr);
8165 if ((int)attr.sched_policy < 0)
8167 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
8168 attr.sched_policy = SETPARAM_POLICY;
8170 CLASS(find_get_task, p)(pid);
8174 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
8175 get_params(p, &attr);
8177 return sched_setattr(p, &attr);
8181 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
8182 * @pid: the pid in question.
8184 * Return: On success, the policy of the thread. Otherwise, a negative error
8187 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
8189 struct task_struct *p;
8196 p = find_process_by_pid(pid);
8200 retval = security_task_getscheduler(p);
8203 if (p->sched_reset_on_fork)
8204 retval |= SCHED_RESET_ON_FORK;
8210 * sys_sched_getparam - get the RT priority of a thread
8211 * @pid: the pid in question.
8212 * @param: structure containing the RT priority.
8214 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
8217 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
8219 struct sched_param lp = { .sched_priority = 0 };
8220 struct task_struct *p;
8223 if (!param || pid < 0)
8226 scoped_guard (rcu) {
8227 p = find_process_by_pid(pid);
8231 retval = security_task_getscheduler(p);
8235 if (task_has_rt_policy(p))
8236 lp.sched_priority = p->rt_priority;
8240 * This one might sleep, we cannot do it with a spinlock held ...
8242 return copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
8246 * Copy the kernel size attribute structure (which might be larger
8247 * than what user-space knows about) to user-space.
8249 * Note that all cases are valid: user-space buffer can be larger or
8250 * smaller than the kernel-space buffer. The usual case is that both
8251 * have the same size.
8254 sched_attr_copy_to_user(struct sched_attr __user *uattr,
8255 struct sched_attr *kattr,
8258 unsigned int ksize = sizeof(*kattr);
8260 if (!access_ok(uattr, usize))
8264 * sched_getattr() ABI forwards and backwards compatibility:
8266 * If usize == ksize then we just copy everything to user-space and all is good.
8268 * If usize < ksize then we only copy as much as user-space has space for,
8269 * this keeps ABI compatibility as well. We skip the rest.
8271 * If usize > ksize then user-space is using a newer version of the ABI,
8272 * which part the kernel doesn't know about. Just ignore it - tooling can
8273 * detect the kernel's knowledge of attributes from the attr->size value
8274 * which is set to ksize in this case.
8276 kattr->size = min(usize, ksize);
8278 if (copy_to_user(uattr, kattr, kattr->size))
8285 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8286 * @pid: the pid in question.
8287 * @uattr: structure containing the extended parameters.
8288 * @usize: sizeof(attr) for fwd/bwd comp.
8289 * @flags: for future extension.
8291 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8292 unsigned int, usize, unsigned int, flags)
8294 struct sched_attr kattr = { };
8295 struct task_struct *p;
8298 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8299 usize < SCHED_ATTR_SIZE_VER0 || flags)
8302 scoped_guard (rcu) {
8303 p = find_process_by_pid(pid);
8307 retval = security_task_getscheduler(p);
8311 kattr.sched_policy = p->policy;
8312 if (p->sched_reset_on_fork)
8313 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8314 get_params(p, &kattr);
8315 kattr.sched_flags &= SCHED_FLAG_ALL;
8317 #ifdef CONFIG_UCLAMP_TASK
8319 * This could race with another potential updater, but this is fine
8320 * because it'll correctly read the old or the new value. We don't need
8321 * to guarantee who wins the race as long as it doesn't return garbage.
8323 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8324 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8328 return sched_attr_copy_to_user(uattr, &kattr, usize);
8332 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8335 * If the task isn't a deadline task or admission control is
8336 * disabled then we don't care about affinity changes.
8338 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8342 * Since bandwidth control happens on root_domain basis,
8343 * if admission test is enabled, we only admit -deadline
8344 * tasks allowed to run on all the CPUs in the task's
8348 if (!cpumask_subset(task_rq(p)->rd->span, mask))
8356 __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx)
8359 cpumask_var_t cpus_allowed, new_mask;
8361 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8364 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8366 goto out_free_cpus_allowed;
8369 cpuset_cpus_allowed(p, cpus_allowed);
8370 cpumask_and(new_mask, ctx->new_mask, cpus_allowed);
8372 ctx->new_mask = new_mask;
8373 ctx->flags |= SCA_CHECK;
8375 retval = dl_task_check_affinity(p, new_mask);
8377 goto out_free_new_mask;
8379 retval = __set_cpus_allowed_ptr(p, ctx);
8381 goto out_free_new_mask;
8383 cpuset_cpus_allowed(p, cpus_allowed);
8384 if (!cpumask_subset(new_mask, cpus_allowed)) {
8386 * We must have raced with a concurrent cpuset update.
8387 * Just reset the cpumask to the cpuset's cpus_allowed.
8389 cpumask_copy(new_mask, cpus_allowed);
8392 * If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
8393 * will restore the previous user_cpus_ptr value.
8395 * In the unlikely event a previous user_cpus_ptr exists,
8396 * we need to further restrict the mask to what is allowed
8397 * by that old user_cpus_ptr.
8399 if (unlikely((ctx->flags & SCA_USER) && ctx->user_mask)) {
8400 bool empty = !cpumask_and(new_mask, new_mask,
8403 if (WARN_ON_ONCE(empty))
8404 cpumask_copy(new_mask, cpus_allowed);
8406 __set_cpus_allowed_ptr(p, ctx);
8411 free_cpumask_var(new_mask);
8412 out_free_cpus_allowed:
8413 free_cpumask_var(cpus_allowed);
8417 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8419 struct affinity_context ac;
8420 struct cpumask *user_mask;
8423 CLASS(find_get_task, p)(pid);
8427 if (p->flags & PF_NO_SETAFFINITY)
8430 if (!check_same_owner(p)) {
8432 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE))
8436 retval = security_task_setscheduler(p);
8441 * With non-SMP configs, user_cpus_ptr/user_mask isn't used and
8442 * alloc_user_cpus_ptr() returns NULL.
8444 user_mask = alloc_user_cpus_ptr(NUMA_NO_NODE);
8446 cpumask_copy(user_mask, in_mask);
8447 } else if (IS_ENABLED(CONFIG_SMP)) {
8451 ac = (struct affinity_context){
8452 .new_mask = in_mask,
8453 .user_mask = user_mask,
8457 retval = __sched_setaffinity(p, &ac);
8458 kfree(ac.user_mask);
8463 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8464 struct cpumask *new_mask)
8466 if (len < cpumask_size())
8467 cpumask_clear(new_mask);
8468 else if (len > cpumask_size())
8469 len = cpumask_size();
8471 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8475 * sys_sched_setaffinity - set the CPU affinity of a process
8476 * @pid: pid of the process
8477 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8478 * @user_mask_ptr: user-space pointer to the new CPU mask
8480 * Return: 0 on success. An error code otherwise.
8482 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8483 unsigned long __user *, user_mask_ptr)
8485 cpumask_var_t new_mask;
8488 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8491 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8493 retval = sched_setaffinity(pid, new_mask);
8494 free_cpumask_var(new_mask);
8498 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8500 struct task_struct *p;
8504 p = find_process_by_pid(pid);
8508 retval = security_task_getscheduler(p);
8512 guard(raw_spinlock_irqsave)(&p->pi_lock);
8513 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8519 * sys_sched_getaffinity - get the CPU affinity of a process
8520 * @pid: pid of the process
8521 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8522 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8524 * Return: size of CPU mask copied to user_mask_ptr on success. An
8525 * error code otherwise.
8527 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8528 unsigned long __user *, user_mask_ptr)
8533 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8535 if (len & (sizeof(unsigned long)-1))
8538 if (!zalloc_cpumask_var(&mask, GFP_KERNEL))
8541 ret = sched_getaffinity(pid, mask);
8543 unsigned int retlen = min(len, cpumask_size());
8545 if (copy_to_user(user_mask_ptr, cpumask_bits(mask), retlen))
8550 free_cpumask_var(mask);
8555 static void do_sched_yield(void)
8560 rq = this_rq_lock_irq(&rf);
8562 schedstat_inc(rq->yld_count);
8563 current->sched_class->yield_task(rq);
8566 rq_unlock_irq(rq, &rf);
8567 sched_preempt_enable_no_resched();
8573 * sys_sched_yield - yield the current processor to other threads.
8575 * This function yields the current CPU to other tasks. If there are no
8576 * other threads running on this CPU then this function will return.
8580 SYSCALL_DEFINE0(sched_yield)
8586 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8587 int __sched __cond_resched(void)
8589 if (should_resched(0)) {
8590 preempt_schedule_common();
8594 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8595 * whether the current CPU is in an RCU read-side critical section,
8596 * so the tick can report quiescent states even for CPUs looping
8597 * in kernel context. In contrast, in non-preemptible kernels,
8598 * RCU readers leave no in-memory hints, which means that CPU-bound
8599 * processes executing in kernel context might never report an
8600 * RCU quiescent state. Therefore, the following code causes
8601 * cond_resched() to report a quiescent state, but only when RCU
8602 * is in urgent need of one.
8604 #ifndef CONFIG_PREEMPT_RCU
8609 EXPORT_SYMBOL(__cond_resched);
8612 #ifdef CONFIG_PREEMPT_DYNAMIC
8613 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8614 #define cond_resched_dynamic_enabled __cond_resched
8615 #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8616 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8617 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8619 #define might_resched_dynamic_enabled __cond_resched
8620 #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8621 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8622 EXPORT_STATIC_CALL_TRAMP(might_resched);
8623 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8624 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8625 int __sched dynamic_cond_resched(void)
8627 klp_sched_try_switch();
8628 if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8630 return __cond_resched();
8632 EXPORT_SYMBOL(dynamic_cond_resched);
8634 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8635 int __sched dynamic_might_resched(void)
8637 if (!static_branch_unlikely(&sk_dynamic_might_resched))
8639 return __cond_resched();
8641 EXPORT_SYMBOL(dynamic_might_resched);
8646 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8647 * call schedule, and on return reacquire the lock.
8649 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8650 * operations here to prevent schedule() from being called twice (once via
8651 * spin_unlock(), once by hand).
8653 int __cond_resched_lock(spinlock_t *lock)
8655 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8658 lockdep_assert_held(lock);
8660 if (spin_needbreak(lock) || resched) {
8662 if (!_cond_resched())
8669 EXPORT_SYMBOL(__cond_resched_lock);
8671 int __cond_resched_rwlock_read(rwlock_t *lock)
8673 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8676 lockdep_assert_held_read(lock);
8678 if (rwlock_needbreak(lock) || resched) {
8680 if (!_cond_resched())
8687 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8689 int __cond_resched_rwlock_write(rwlock_t *lock)
8691 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8694 lockdep_assert_held_write(lock);
8696 if (rwlock_needbreak(lock) || resched) {
8698 if (!_cond_resched())
8705 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8707 #ifdef CONFIG_PREEMPT_DYNAMIC
8709 #ifdef CONFIG_GENERIC_ENTRY
8710 #include <linux/entry-common.h>
8716 * SC:preempt_schedule
8717 * SC:preempt_schedule_notrace
8718 * SC:irqentry_exit_cond_resched
8722 * cond_resched <- __cond_resched
8723 * might_resched <- RET0
8724 * preempt_schedule <- NOP
8725 * preempt_schedule_notrace <- NOP
8726 * irqentry_exit_cond_resched <- NOP
8729 * cond_resched <- __cond_resched
8730 * might_resched <- __cond_resched
8731 * preempt_schedule <- NOP
8732 * preempt_schedule_notrace <- NOP
8733 * irqentry_exit_cond_resched <- NOP
8736 * cond_resched <- RET0
8737 * might_resched <- RET0
8738 * preempt_schedule <- preempt_schedule
8739 * preempt_schedule_notrace <- preempt_schedule_notrace
8740 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8744 preempt_dynamic_undefined = -1,
8745 preempt_dynamic_none,
8746 preempt_dynamic_voluntary,
8747 preempt_dynamic_full,
8750 int preempt_dynamic_mode = preempt_dynamic_undefined;
8752 int sched_dynamic_mode(const char *str)
8754 if (!strcmp(str, "none"))
8755 return preempt_dynamic_none;
8757 if (!strcmp(str, "voluntary"))
8758 return preempt_dynamic_voluntary;
8760 if (!strcmp(str, "full"))
8761 return preempt_dynamic_full;
8766 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8767 #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8768 #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8769 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8770 #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8771 #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8773 #error "Unsupported PREEMPT_DYNAMIC mechanism"
8776 static DEFINE_MUTEX(sched_dynamic_mutex);
8777 static bool klp_override;
8779 static void __sched_dynamic_update(int mode)
8782 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8783 * the ZERO state, which is invalid.
8786 preempt_dynamic_enable(cond_resched);
8787 preempt_dynamic_enable(might_resched);
8788 preempt_dynamic_enable(preempt_schedule);
8789 preempt_dynamic_enable(preempt_schedule_notrace);
8790 preempt_dynamic_enable(irqentry_exit_cond_resched);
8793 case preempt_dynamic_none:
8795 preempt_dynamic_enable(cond_resched);
8796 preempt_dynamic_disable(might_resched);
8797 preempt_dynamic_disable(preempt_schedule);
8798 preempt_dynamic_disable(preempt_schedule_notrace);
8799 preempt_dynamic_disable(irqentry_exit_cond_resched);
8800 if (mode != preempt_dynamic_mode)
8801 pr_info("Dynamic Preempt: none\n");
8804 case preempt_dynamic_voluntary:
8806 preempt_dynamic_enable(cond_resched);
8807 preempt_dynamic_enable(might_resched);
8808 preempt_dynamic_disable(preempt_schedule);
8809 preempt_dynamic_disable(preempt_schedule_notrace);
8810 preempt_dynamic_disable(irqentry_exit_cond_resched);
8811 if (mode != preempt_dynamic_mode)
8812 pr_info("Dynamic Preempt: voluntary\n");
8815 case preempt_dynamic_full:
8817 preempt_dynamic_disable(cond_resched);
8818 preempt_dynamic_disable(might_resched);
8819 preempt_dynamic_enable(preempt_schedule);
8820 preempt_dynamic_enable(preempt_schedule_notrace);
8821 preempt_dynamic_enable(irqentry_exit_cond_resched);
8822 if (mode != preempt_dynamic_mode)
8823 pr_info("Dynamic Preempt: full\n");
8827 preempt_dynamic_mode = mode;
8830 void sched_dynamic_update(int mode)
8832 mutex_lock(&sched_dynamic_mutex);
8833 __sched_dynamic_update(mode);
8834 mutex_unlock(&sched_dynamic_mutex);
8837 #ifdef CONFIG_HAVE_PREEMPT_DYNAMIC_CALL
8839 static int klp_cond_resched(void)
8841 __klp_sched_try_switch();
8842 return __cond_resched();
8845 void sched_dynamic_klp_enable(void)
8847 mutex_lock(&sched_dynamic_mutex);
8849 klp_override = true;
8850 static_call_update(cond_resched, klp_cond_resched);
8852 mutex_unlock(&sched_dynamic_mutex);
8855 void sched_dynamic_klp_disable(void)
8857 mutex_lock(&sched_dynamic_mutex);
8859 klp_override = false;
8860 __sched_dynamic_update(preempt_dynamic_mode);
8862 mutex_unlock(&sched_dynamic_mutex);
8865 #endif /* CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
8867 static int __init setup_preempt_mode(char *str)
8869 int mode = sched_dynamic_mode(str);
8871 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8875 sched_dynamic_update(mode);
8878 __setup("preempt=", setup_preempt_mode);
8880 static void __init preempt_dynamic_init(void)
8882 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8883 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8884 sched_dynamic_update(preempt_dynamic_none);
8885 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8886 sched_dynamic_update(preempt_dynamic_voluntary);
8888 /* Default static call setting, nothing to do */
8889 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8890 preempt_dynamic_mode = preempt_dynamic_full;
8891 pr_info("Dynamic Preempt: full\n");
8896 #define PREEMPT_MODEL_ACCESSOR(mode) \
8897 bool preempt_model_##mode(void) \
8899 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8900 return preempt_dynamic_mode == preempt_dynamic_##mode; \
8902 EXPORT_SYMBOL_GPL(preempt_model_##mode)
8904 PREEMPT_MODEL_ACCESSOR(none);
8905 PREEMPT_MODEL_ACCESSOR(voluntary);
8906 PREEMPT_MODEL_ACCESSOR(full);
8908 #else /* !CONFIG_PREEMPT_DYNAMIC */
8910 static inline void preempt_dynamic_init(void) { }
8912 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8915 * yield - yield the current processor to other threads.
8917 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8919 * The scheduler is at all times free to pick the calling task as the most
8920 * eligible task to run, if removing the yield() call from your code breaks
8921 * it, it's already broken.
8923 * Typical broken usage is:
8928 * where one assumes that yield() will let 'the other' process run that will
8929 * make event true. If the current task is a SCHED_FIFO task that will never
8930 * happen. Never use yield() as a progress guarantee!!
8932 * If you want to use yield() to wait for something, use wait_event().
8933 * If you want to use yield() to be 'nice' for others, use cond_resched().
8934 * If you still want to use yield(), do not!
8936 void __sched yield(void)
8938 set_current_state(TASK_RUNNING);
8941 EXPORT_SYMBOL(yield);
8944 * yield_to - yield the current processor to another thread in
8945 * your thread group, or accelerate that thread toward the
8946 * processor it's on.
8948 * @preempt: whether task preemption is allowed or not
8950 * It's the caller's job to ensure that the target task struct
8951 * can't go away on us before we can do any checks.
8954 * true (>0) if we indeed boosted the target task.
8955 * false (0) if we failed to boost the target.
8956 * -ESRCH if there's no task to yield to.
8958 int __sched yield_to(struct task_struct *p, bool preempt)
8960 struct task_struct *curr = current;
8961 struct rq *rq, *p_rq;
8964 scoped_guard (irqsave) {
8970 * If we're the only runnable task on the rq and target rq also
8971 * has only one task, there's absolutely no point in yielding.
8973 if (rq->nr_running == 1 && p_rq->nr_running == 1)
8976 guard(double_rq_lock)(rq, p_rq);
8977 if (task_rq(p) != p_rq)
8980 if (!curr->sched_class->yield_to_task)
8983 if (curr->sched_class != p->sched_class)
8986 if (task_on_cpu(p_rq, p) || !task_is_running(p))
8989 yielded = curr->sched_class->yield_to_task(rq, p);
8991 schedstat_inc(rq->yld_count);
8993 * Make p's CPU reschedule; pick_next_entity
8994 * takes care of fairness.
8996 if (preempt && rq != p_rq)
9006 EXPORT_SYMBOL_GPL(yield_to);
9008 int io_schedule_prepare(void)
9010 int old_iowait = current->in_iowait;
9012 current->in_iowait = 1;
9013 blk_flush_plug(current->plug, true);
9017 void io_schedule_finish(int token)
9019 current->in_iowait = token;
9023 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
9024 * that process accounting knows that this is a task in IO wait state.
9026 long __sched io_schedule_timeout(long timeout)
9031 token = io_schedule_prepare();
9032 ret = schedule_timeout(timeout);
9033 io_schedule_finish(token);
9037 EXPORT_SYMBOL(io_schedule_timeout);
9039 void __sched io_schedule(void)
9043 token = io_schedule_prepare();
9045 io_schedule_finish(token);
9047 EXPORT_SYMBOL(io_schedule);
9050 * sys_sched_get_priority_max - return maximum RT priority.
9051 * @policy: scheduling class.
9053 * Return: On success, this syscall returns the maximum
9054 * rt_priority that can be used by a given scheduling class.
9055 * On failure, a negative error code is returned.
9057 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
9064 ret = MAX_RT_PRIO-1;
9066 case SCHED_DEADLINE:
9077 * sys_sched_get_priority_min - return minimum RT priority.
9078 * @policy: scheduling class.
9080 * Return: On success, this syscall returns the minimum
9081 * rt_priority that can be used by a given scheduling class.
9082 * On failure, a negative error code is returned.
9084 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
9093 case SCHED_DEADLINE:
9102 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
9104 unsigned int time_slice = 0;
9110 scoped_guard (rcu) {
9111 struct task_struct *p = find_process_by_pid(pid);
9115 retval = security_task_getscheduler(p);
9119 scoped_guard (task_rq_lock, p) {
9120 struct rq *rq = scope.rq;
9121 if (p->sched_class->get_rr_interval)
9122 time_slice = p->sched_class->get_rr_interval(rq, p);
9126 jiffies_to_timespec64(time_slice, t);
9131 * sys_sched_rr_get_interval - return the default timeslice of a process.
9132 * @pid: pid of the process.
9133 * @interval: userspace pointer to the timeslice value.
9135 * this syscall writes the default timeslice value of a given process
9136 * into the user-space timespec buffer. A value of '0' means infinity.
9138 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
9141 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
9142 struct __kernel_timespec __user *, interval)
9144 struct timespec64 t;
9145 int retval = sched_rr_get_interval(pid, &t);
9148 retval = put_timespec64(&t, interval);
9153 #ifdef CONFIG_COMPAT_32BIT_TIME
9154 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
9155 struct old_timespec32 __user *, interval)
9157 struct timespec64 t;
9158 int retval = sched_rr_get_interval(pid, &t);
9161 retval = put_old_timespec32(&t, interval);
9166 void sched_show_task(struct task_struct *p)
9168 unsigned long free = 0;
9171 if (!try_get_task_stack(p))
9174 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
9176 if (task_is_running(p))
9177 pr_cont(" running task ");
9178 #ifdef CONFIG_DEBUG_STACK_USAGE
9179 free = stack_not_used(p);
9184 ppid = task_pid_nr(rcu_dereference(p->real_parent));
9186 pr_cont(" stack:%-5lu pid:%-5d tgid:%-5d ppid:%-6d flags:0x%08lx\n",
9187 free, task_pid_nr(p), task_tgid_nr(p),
9188 ppid, read_task_thread_flags(p));
9190 print_worker_info(KERN_INFO, p);
9191 print_stop_info(KERN_INFO, p);
9192 show_stack(p, NULL, KERN_INFO);
9195 EXPORT_SYMBOL_GPL(sched_show_task);
9198 state_filter_match(unsigned long state_filter, struct task_struct *p)
9200 unsigned int state = READ_ONCE(p->__state);
9202 /* no filter, everything matches */
9206 /* filter, but doesn't match */
9207 if (!(state & state_filter))
9211 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
9214 if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
9221 void show_state_filter(unsigned int state_filter)
9223 struct task_struct *g, *p;
9226 for_each_process_thread(g, p) {
9228 * reset the NMI-timeout, listing all files on a slow
9229 * console might take a lot of time:
9230 * Also, reset softlockup watchdogs on all CPUs, because
9231 * another CPU might be blocked waiting for us to process
9234 touch_nmi_watchdog();
9235 touch_all_softlockup_watchdogs();
9236 if (state_filter_match(state_filter, p))
9240 #ifdef CONFIG_SCHED_DEBUG
9242 sysrq_sched_debug_show();
9246 * Only show locks if all tasks are dumped:
9249 debug_show_all_locks();
9253 * init_idle - set up an idle thread for a given CPU
9254 * @idle: task in question
9255 * @cpu: CPU the idle task belongs to
9257 * NOTE: this function does not set the idle thread's NEED_RESCHED
9258 * flag, to make booting more robust.
9260 void __init init_idle(struct task_struct *idle, int cpu)
9263 struct affinity_context ac = (struct affinity_context) {
9264 .new_mask = cpumask_of(cpu),
9268 struct rq *rq = cpu_rq(cpu);
9269 unsigned long flags;
9271 __sched_fork(0, idle);
9273 raw_spin_lock_irqsave(&idle->pi_lock, flags);
9274 raw_spin_rq_lock(rq);
9276 idle->__state = TASK_RUNNING;
9277 idle->se.exec_start = sched_clock();
9279 * PF_KTHREAD should already be set at this point; regardless, make it
9280 * look like a proper per-CPU kthread.
9282 idle->flags |= PF_KTHREAD | PF_NO_SETAFFINITY;
9283 kthread_set_per_cpu(idle, cpu);
9287 * It's possible that init_idle() gets called multiple times on a task,
9288 * in that case do_set_cpus_allowed() will not do the right thing.
9290 * And since this is boot we can forgo the serialization.
9292 set_cpus_allowed_common(idle, &ac);
9295 * We're having a chicken and egg problem, even though we are
9296 * holding rq->lock, the CPU isn't yet set to this CPU so the
9297 * lockdep check in task_group() will fail.
9299 * Similar case to sched_fork(). / Alternatively we could
9300 * use task_rq_lock() here and obtain the other rq->lock.
9305 __set_task_cpu(idle, cpu);
9309 rcu_assign_pointer(rq->curr, idle);
9310 idle->on_rq = TASK_ON_RQ_QUEUED;
9314 raw_spin_rq_unlock(rq);
9315 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
9317 /* Set the preempt count _outside_ the spinlocks! */
9318 init_idle_preempt_count(idle, cpu);
9321 * The idle tasks have their own, simple scheduling class:
9323 idle->sched_class = &idle_sched_class;
9324 ftrace_graph_init_idle_task(idle, cpu);
9325 vtime_init_idle(idle, cpu);
9327 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9333 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9334 const struct cpumask *trial)
9338 if (cpumask_empty(cur))
9341 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9346 int task_can_attach(struct task_struct *p)
9351 * Kthreads which disallow setaffinity shouldn't be moved
9352 * to a new cpuset; we don't want to change their CPU
9353 * affinity and isolating such threads by their set of
9354 * allowed nodes is unnecessary. Thus, cpusets are not
9355 * applicable for such threads. This prevents checking for
9356 * success of set_cpus_allowed_ptr() on all attached tasks
9357 * before cpus_mask may be changed.
9359 if (p->flags & PF_NO_SETAFFINITY)
9365 bool sched_smp_initialized __read_mostly;
9367 #ifdef CONFIG_NUMA_BALANCING
9368 /* Migrate current task p to target_cpu */
9369 int migrate_task_to(struct task_struct *p, int target_cpu)
9371 struct migration_arg arg = { p, target_cpu };
9372 int curr_cpu = task_cpu(p);
9374 if (curr_cpu == target_cpu)
9377 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9380 /* TODO: This is not properly updating schedstats */
9382 trace_sched_move_numa(p, curr_cpu, target_cpu);
9383 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9387 * Requeue a task on a given node and accurately track the number of NUMA
9388 * tasks on the runqueues
9390 void sched_setnuma(struct task_struct *p, int nid)
9392 bool queued, running;
9396 rq = task_rq_lock(p, &rf);
9397 queued = task_on_rq_queued(p);
9398 running = task_current(rq, p);
9401 dequeue_task(rq, p, DEQUEUE_SAVE);
9403 put_prev_task(rq, p);
9405 p->numa_preferred_nid = nid;
9408 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9410 set_next_task(rq, p);
9411 task_rq_unlock(rq, p, &rf);
9413 #endif /* CONFIG_NUMA_BALANCING */
9415 #ifdef CONFIG_HOTPLUG_CPU
9417 * Ensure that the idle task is using init_mm right before its CPU goes
9420 void idle_task_exit(void)
9422 struct mm_struct *mm = current->active_mm;
9424 BUG_ON(cpu_online(smp_processor_id()));
9425 BUG_ON(current != this_rq()->idle);
9427 if (mm != &init_mm) {
9428 switch_mm(mm, &init_mm, current);
9429 finish_arch_post_lock_switch();
9432 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9435 static int __balance_push_cpu_stop(void *arg)
9437 struct task_struct *p = arg;
9438 struct rq *rq = this_rq();
9442 raw_spin_lock_irq(&p->pi_lock);
9445 update_rq_clock(rq);
9447 if (task_rq(p) == rq && task_on_rq_queued(p)) {
9448 cpu = select_fallback_rq(rq->cpu, p);
9449 rq = __migrate_task(rq, &rf, p, cpu);
9453 raw_spin_unlock_irq(&p->pi_lock);
9460 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9463 * Ensure we only run per-cpu kthreads once the CPU goes !active.
9465 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9466 * effective when the hotplug motion is down.
9468 static void balance_push(struct rq *rq)
9470 struct task_struct *push_task = rq->curr;
9472 lockdep_assert_rq_held(rq);
9475 * Ensure the thing is persistent until balance_push_set(.on = false);
9477 rq->balance_callback = &balance_push_callback;
9480 * Only active while going offline and when invoked on the outgoing
9483 if (!cpu_dying(rq->cpu) || rq != this_rq())
9487 * Both the cpu-hotplug and stop task are in this case and are
9488 * required to complete the hotplug process.
9490 if (kthread_is_per_cpu(push_task) ||
9491 is_migration_disabled(push_task)) {
9494 * If this is the idle task on the outgoing CPU try to wake
9495 * up the hotplug control thread which might wait for the
9496 * last task to vanish. The rcuwait_active() check is
9497 * accurate here because the waiter is pinned on this CPU
9498 * and can't obviously be running in parallel.
9500 * On RT kernels this also has to check whether there are
9501 * pinned and scheduled out tasks on the runqueue. They
9502 * need to leave the migrate disabled section first.
9504 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9505 rcuwait_active(&rq->hotplug_wait)) {
9506 raw_spin_rq_unlock(rq);
9507 rcuwait_wake_up(&rq->hotplug_wait);
9508 raw_spin_rq_lock(rq);
9513 get_task_struct(push_task);
9515 * Temporarily drop rq->lock such that we can wake-up the stop task.
9516 * Both preemption and IRQs are still disabled.
9519 raw_spin_rq_unlock(rq);
9520 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9521 this_cpu_ptr(&push_work));
9524 * At this point need_resched() is true and we'll take the loop in
9525 * schedule(). The next pick is obviously going to be the stop task
9526 * which kthread_is_per_cpu() and will push this task away.
9528 raw_spin_rq_lock(rq);
9531 static void balance_push_set(int cpu, bool on)
9533 struct rq *rq = cpu_rq(cpu);
9536 rq_lock_irqsave(rq, &rf);
9538 WARN_ON_ONCE(rq->balance_callback);
9539 rq->balance_callback = &balance_push_callback;
9540 } else if (rq->balance_callback == &balance_push_callback) {
9541 rq->balance_callback = NULL;
9543 rq_unlock_irqrestore(rq, &rf);
9547 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9548 * inactive. All tasks which are not per CPU kernel threads are either
9549 * pushed off this CPU now via balance_push() or placed on a different CPU
9550 * during wakeup. Wait until the CPU is quiescent.
9552 static void balance_hotplug_wait(void)
9554 struct rq *rq = this_rq();
9556 rcuwait_wait_event(&rq->hotplug_wait,
9557 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9558 TASK_UNINTERRUPTIBLE);
9563 static inline void balance_push(struct rq *rq)
9567 static inline void balance_push_set(int cpu, bool on)
9571 static inline void balance_hotplug_wait(void)
9575 #endif /* CONFIG_HOTPLUG_CPU */
9577 void set_rq_online(struct rq *rq)
9580 const struct sched_class *class;
9582 cpumask_set_cpu(rq->cpu, rq->rd->online);
9585 for_each_class(class) {
9586 if (class->rq_online)
9587 class->rq_online(rq);
9592 void set_rq_offline(struct rq *rq)
9595 const struct sched_class *class;
9597 update_rq_clock(rq);
9598 for_each_class(class) {
9599 if (class->rq_offline)
9600 class->rq_offline(rq);
9603 cpumask_clear_cpu(rq->cpu, rq->rd->online);
9609 * used to mark begin/end of suspend/resume:
9611 static int num_cpus_frozen;
9614 * Update cpusets according to cpu_active mask. If cpusets are
9615 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9616 * around partition_sched_domains().
9618 * If we come here as part of a suspend/resume, don't touch cpusets because we
9619 * want to restore it back to its original state upon resume anyway.
9621 static void cpuset_cpu_active(void)
9623 if (cpuhp_tasks_frozen) {
9625 * num_cpus_frozen tracks how many CPUs are involved in suspend
9626 * resume sequence. As long as this is not the last online
9627 * operation in the resume sequence, just build a single sched
9628 * domain, ignoring cpusets.
9630 partition_sched_domains(1, NULL, NULL);
9631 if (--num_cpus_frozen)
9634 * This is the last CPU online operation. So fall through and
9635 * restore the original sched domains by considering the
9636 * cpuset configurations.
9638 cpuset_force_rebuild();
9640 cpuset_update_active_cpus();
9643 static int cpuset_cpu_inactive(unsigned int cpu)
9645 if (!cpuhp_tasks_frozen) {
9646 int ret = dl_bw_check_overflow(cpu);
9650 cpuset_update_active_cpus();
9653 partition_sched_domains(1, NULL, NULL);
9658 int sched_cpu_activate(unsigned int cpu)
9660 struct rq *rq = cpu_rq(cpu);
9664 * Clear the balance_push callback and prepare to schedule
9667 balance_push_set(cpu, false);
9669 #ifdef CONFIG_SCHED_SMT
9671 * When going up, increment the number of cores with SMT present.
9673 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9674 static_branch_inc_cpuslocked(&sched_smt_present);
9676 set_cpu_active(cpu, true);
9678 if (sched_smp_initialized) {
9679 sched_update_numa(cpu, true);
9680 sched_domains_numa_masks_set(cpu);
9681 cpuset_cpu_active();
9685 * Put the rq online, if not already. This happens:
9687 * 1) In the early boot process, because we build the real domains
9688 * after all CPUs have been brought up.
9690 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9693 rq_lock_irqsave(rq, &rf);
9695 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9698 rq_unlock_irqrestore(rq, &rf);
9703 int sched_cpu_deactivate(unsigned int cpu)
9705 struct rq *rq = cpu_rq(cpu);
9710 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9711 * load balancing when not active
9713 nohz_balance_exit_idle(rq);
9715 set_cpu_active(cpu, false);
9718 * From this point forward, this CPU will refuse to run any task that
9719 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9720 * push those tasks away until this gets cleared, see
9721 * sched_cpu_dying().
9723 balance_push_set(cpu, true);
9726 * We've cleared cpu_active_mask / set balance_push, wait for all
9727 * preempt-disabled and RCU users of this state to go away such that
9728 * all new such users will observe it.
9730 * Specifically, we rely on ttwu to no longer target this CPU, see
9731 * ttwu_queue_cond() and is_cpu_allowed().
9733 * Do sync before park smpboot threads to take care the rcu boost case.
9737 rq_lock_irqsave(rq, &rf);
9739 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9742 rq_unlock_irqrestore(rq, &rf);
9744 #ifdef CONFIG_SCHED_SMT
9746 * When going down, decrement the number of cores with SMT present.
9748 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9749 static_branch_dec_cpuslocked(&sched_smt_present);
9751 sched_core_cpu_deactivate(cpu);
9754 if (!sched_smp_initialized)
9757 sched_update_numa(cpu, false);
9758 ret = cpuset_cpu_inactive(cpu);
9760 balance_push_set(cpu, false);
9761 set_cpu_active(cpu, true);
9762 sched_update_numa(cpu, true);
9765 sched_domains_numa_masks_clear(cpu);
9769 static void sched_rq_cpu_starting(unsigned int cpu)
9771 struct rq *rq = cpu_rq(cpu);
9773 rq->calc_load_update = calc_load_update;
9774 update_max_interval();
9777 int sched_cpu_starting(unsigned int cpu)
9779 sched_core_cpu_starting(cpu);
9780 sched_rq_cpu_starting(cpu);
9781 sched_tick_start(cpu);
9785 #ifdef CONFIG_HOTPLUG_CPU
9788 * Invoked immediately before the stopper thread is invoked to bring the
9789 * CPU down completely. At this point all per CPU kthreads except the
9790 * hotplug thread (current) and the stopper thread (inactive) have been
9791 * either parked or have been unbound from the outgoing CPU. Ensure that
9792 * any of those which might be on the way out are gone.
9794 * If after this point a bound task is being woken on this CPU then the
9795 * responsible hotplug callback has failed to do it's job.
9796 * sched_cpu_dying() will catch it with the appropriate fireworks.
9798 int sched_cpu_wait_empty(unsigned int cpu)
9800 balance_hotplug_wait();
9805 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9806 * might have. Called from the CPU stopper task after ensuring that the
9807 * stopper is the last running task on the CPU, so nr_active count is
9808 * stable. We need to take the teardown thread which is calling this into
9809 * account, so we hand in adjust = 1 to the load calculation.
9811 * Also see the comment "Global load-average calculations".
9813 static void calc_load_migrate(struct rq *rq)
9815 long delta = calc_load_fold_active(rq, 1);
9818 atomic_long_add(delta, &calc_load_tasks);
9821 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9823 struct task_struct *g, *p;
9824 int cpu = cpu_of(rq);
9826 lockdep_assert_rq_held(rq);
9828 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9829 for_each_process_thread(g, p) {
9830 if (task_cpu(p) != cpu)
9833 if (!task_on_rq_queued(p))
9836 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9840 int sched_cpu_dying(unsigned int cpu)
9842 struct rq *rq = cpu_rq(cpu);
9845 /* Handle pending wakeups and then migrate everything off */
9846 sched_tick_stop(cpu);
9848 rq_lock_irqsave(rq, &rf);
9849 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9850 WARN(true, "Dying CPU not properly vacated!");
9851 dump_rq_tasks(rq, KERN_WARNING);
9853 rq_unlock_irqrestore(rq, &rf);
9855 calc_load_migrate(rq);
9856 update_max_interval();
9858 sched_core_cpu_dying(cpu);
9863 void __init sched_init_smp(void)
9865 sched_init_numa(NUMA_NO_NODE);
9868 * There's no userspace yet to cause hotplug operations; hence all the
9869 * CPU masks are stable and all blatant races in the below code cannot
9872 mutex_lock(&sched_domains_mutex);
9873 sched_init_domains(cpu_active_mask);
9874 mutex_unlock(&sched_domains_mutex);
9876 /* Move init over to a non-isolated CPU */
9877 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9879 current->flags &= ~PF_NO_SETAFFINITY;
9880 sched_init_granularity();
9882 init_sched_rt_class();
9883 init_sched_dl_class();
9885 sched_smp_initialized = true;
9888 static int __init migration_init(void)
9890 sched_cpu_starting(smp_processor_id());
9893 early_initcall(migration_init);
9896 void __init sched_init_smp(void)
9898 sched_init_granularity();
9900 #endif /* CONFIG_SMP */
9902 int in_sched_functions(unsigned long addr)
9904 return in_lock_functions(addr) ||
9905 (addr >= (unsigned long)__sched_text_start
9906 && addr < (unsigned long)__sched_text_end);
9909 #ifdef CONFIG_CGROUP_SCHED
9911 * Default task group.
9912 * Every task in system belongs to this group at bootup.
9914 struct task_group root_task_group;
9915 LIST_HEAD(task_groups);
9917 /* Cacheline aligned slab cache for task_group */
9918 static struct kmem_cache *task_group_cache __ro_after_init;
9921 void __init sched_init(void)
9923 unsigned long ptr = 0;
9926 /* Make sure the linker didn't screw up */
9927 BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9928 &fair_sched_class != &rt_sched_class + 1 ||
9929 &rt_sched_class != &dl_sched_class + 1);
9931 BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9936 #ifdef CONFIG_FAIR_GROUP_SCHED
9937 ptr += 2 * nr_cpu_ids * sizeof(void **);
9939 #ifdef CONFIG_RT_GROUP_SCHED
9940 ptr += 2 * nr_cpu_ids * sizeof(void **);
9943 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9945 #ifdef CONFIG_FAIR_GROUP_SCHED
9946 root_task_group.se = (struct sched_entity **)ptr;
9947 ptr += nr_cpu_ids * sizeof(void **);
9949 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9950 ptr += nr_cpu_ids * sizeof(void **);
9952 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9953 init_cfs_bandwidth(&root_task_group.cfs_bandwidth, NULL);
9954 #endif /* CONFIG_FAIR_GROUP_SCHED */
9955 #ifdef CONFIG_RT_GROUP_SCHED
9956 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9957 ptr += nr_cpu_ids * sizeof(void **);
9959 root_task_group.rt_rq = (struct rt_rq **)ptr;
9960 ptr += nr_cpu_ids * sizeof(void **);
9962 #endif /* CONFIG_RT_GROUP_SCHED */
9965 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9968 init_defrootdomain();
9971 #ifdef CONFIG_RT_GROUP_SCHED
9972 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9973 global_rt_period(), global_rt_runtime());
9974 #endif /* CONFIG_RT_GROUP_SCHED */
9976 #ifdef CONFIG_CGROUP_SCHED
9977 task_group_cache = KMEM_CACHE(task_group, 0);
9979 list_add(&root_task_group.list, &task_groups);
9980 INIT_LIST_HEAD(&root_task_group.children);
9981 INIT_LIST_HEAD(&root_task_group.siblings);
9982 autogroup_init(&init_task);
9983 #endif /* CONFIG_CGROUP_SCHED */
9985 for_each_possible_cpu(i) {
9989 raw_spin_lock_init(&rq->__lock);
9991 rq->calc_load_active = 0;
9992 rq->calc_load_update = jiffies + LOAD_FREQ;
9993 init_cfs_rq(&rq->cfs);
9994 init_rt_rq(&rq->rt);
9995 init_dl_rq(&rq->dl);
9996 #ifdef CONFIG_FAIR_GROUP_SCHED
9997 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9998 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
10000 * How much CPU bandwidth does root_task_group get?
10002 * In case of task-groups formed thr' the cgroup filesystem, it
10003 * gets 100% of the CPU resources in the system. This overall
10004 * system CPU resource is divided among the tasks of
10005 * root_task_group and its child task-groups in a fair manner,
10006 * based on each entity's (task or task-group's) weight
10007 * (se->load.weight).
10009 * In other words, if root_task_group has 10 tasks of weight
10010 * 1024) and two child groups A0 and A1 (of weight 1024 each),
10011 * then A0's share of the CPU resource is:
10013 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
10015 * We achieve this by letting root_task_group's tasks sit
10016 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
10018 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
10019 #endif /* CONFIG_FAIR_GROUP_SCHED */
10021 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
10022 #ifdef CONFIG_RT_GROUP_SCHED
10023 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
10028 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
10029 rq->balance_callback = &balance_push_callback;
10030 rq->active_balance = 0;
10031 rq->next_balance = jiffies;
10035 rq->idle_stamp = 0;
10036 rq->avg_idle = 2*sysctl_sched_migration_cost;
10037 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
10039 INIT_LIST_HEAD(&rq->cfs_tasks);
10041 rq_attach_root(rq, &def_root_domain);
10042 #ifdef CONFIG_NO_HZ_COMMON
10043 rq->last_blocked_load_update_tick = jiffies;
10044 atomic_set(&rq->nohz_flags, 0);
10046 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
10048 #ifdef CONFIG_HOTPLUG_CPU
10049 rcuwait_init(&rq->hotplug_wait);
10051 #endif /* CONFIG_SMP */
10052 hrtick_rq_init(rq);
10053 atomic_set(&rq->nr_iowait, 0);
10055 #ifdef CONFIG_SCHED_CORE
10057 rq->core_pick = NULL;
10058 rq->core_enabled = 0;
10059 rq->core_tree = RB_ROOT;
10060 rq->core_forceidle_count = 0;
10061 rq->core_forceidle_occupation = 0;
10062 rq->core_forceidle_start = 0;
10064 rq->core_cookie = 0UL;
10066 zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i));
10069 set_load_weight(&init_task, false);
10072 * The boot idle thread does lazy MMU switching as well:
10074 mmgrab_lazy_tlb(&init_mm);
10075 enter_lazy_tlb(&init_mm, current);
10078 * The idle task doesn't need the kthread struct to function, but it
10079 * is dressed up as a per-CPU kthread and thus needs to play the part
10080 * if we want to avoid special-casing it in code that deals with per-CPU
10083 WARN_ON(!set_kthread_struct(current));
10086 * Make us the idle thread. Technically, schedule() should not be
10087 * called from this thread, however somewhere below it might be,
10088 * but because we are the idle thread, we just pick up running again
10089 * when this runqueue becomes "idle".
10091 init_idle(current, smp_processor_id());
10093 calc_load_update = jiffies + LOAD_FREQ;
10096 idle_thread_set_boot_cpu();
10097 balance_push_set(smp_processor_id(), false);
10099 init_sched_fair_class();
10105 preempt_dynamic_init();
10107 scheduler_running = 1;
10110 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
10112 void __might_sleep(const char *file, int line)
10114 unsigned int state = get_current_state();
10116 * Blocking primitives will set (and therefore destroy) current->state,
10117 * since we will exit with TASK_RUNNING make sure we enter with it,
10118 * otherwise we will destroy state.
10120 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
10121 "do not call blocking ops when !TASK_RUNNING; "
10122 "state=%x set at [<%p>] %pS\n", state,
10123 (void *)current->task_state_change,
10124 (void *)current->task_state_change);
10126 __might_resched(file, line, 0);
10128 EXPORT_SYMBOL(__might_sleep);
10130 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
10132 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
10135 if (preempt_count() == preempt_offset)
10138 pr_err("Preemption disabled at:");
10139 print_ip_sym(KERN_ERR, ip);
10142 static inline bool resched_offsets_ok(unsigned int offsets)
10144 unsigned int nested = preempt_count();
10146 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
10148 return nested == offsets;
10151 void __might_resched(const char *file, int line, unsigned int offsets)
10153 /* Ratelimiting timestamp: */
10154 static unsigned long prev_jiffy;
10156 unsigned long preempt_disable_ip;
10158 /* WARN_ON_ONCE() by default, no rate limit required: */
10161 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
10162 !is_idle_task(current) && !current->non_block_count) ||
10163 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
10167 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10169 prev_jiffy = jiffies;
10171 /* Save this before calling printk(), since that will clobber it: */
10172 preempt_disable_ip = get_preempt_disable_ip(current);
10174 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10176 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10177 in_atomic(), irqs_disabled(), current->non_block_count,
10178 current->pid, current->comm);
10179 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
10180 offsets & MIGHT_RESCHED_PREEMPT_MASK);
10182 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
10183 pr_err("RCU nest depth: %d, expected: %u\n",
10184 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
10187 if (task_stack_end_corrupted(current))
10188 pr_emerg("Thread overran stack, or stack corrupted\n");
10190 debug_show_held_locks(current);
10191 if (irqs_disabled())
10192 print_irqtrace_events(current);
10194 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
10195 preempt_disable_ip);
10198 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10200 EXPORT_SYMBOL(__might_resched);
10202 void __cant_sleep(const char *file, int line, int preempt_offset)
10204 static unsigned long prev_jiffy;
10206 if (irqs_disabled())
10209 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10212 if (preempt_count() > preempt_offset)
10215 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10217 prev_jiffy = jiffies;
10219 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
10220 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10221 in_atomic(), irqs_disabled(),
10222 current->pid, current->comm);
10224 debug_show_held_locks(current);
10226 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10228 EXPORT_SYMBOL_GPL(__cant_sleep);
10231 void __cant_migrate(const char *file, int line)
10233 static unsigned long prev_jiffy;
10235 if (irqs_disabled())
10238 if (is_migration_disabled(current))
10241 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10244 if (preempt_count() > 0)
10247 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10249 prev_jiffy = jiffies;
10251 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
10252 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10253 in_atomic(), irqs_disabled(), is_migration_disabled(current),
10254 current->pid, current->comm);
10256 debug_show_held_locks(current);
10258 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10260 EXPORT_SYMBOL_GPL(__cant_migrate);
10264 #ifdef CONFIG_MAGIC_SYSRQ
10265 void normalize_rt_tasks(void)
10267 struct task_struct *g, *p;
10268 struct sched_attr attr = {
10269 .sched_policy = SCHED_NORMAL,
10272 read_lock(&tasklist_lock);
10273 for_each_process_thread(g, p) {
10275 * Only normalize user tasks:
10277 if (p->flags & PF_KTHREAD)
10280 p->se.exec_start = 0;
10281 schedstat_set(p->stats.wait_start, 0);
10282 schedstat_set(p->stats.sleep_start, 0);
10283 schedstat_set(p->stats.block_start, 0);
10285 if (!dl_task(p) && !rt_task(p)) {
10287 * Renice negative nice level userspace
10290 if (task_nice(p) < 0)
10291 set_user_nice(p, 0);
10295 __sched_setscheduler(p, &attr, false, false);
10297 read_unlock(&tasklist_lock);
10300 #endif /* CONFIG_MAGIC_SYSRQ */
10302 #if defined(CONFIG_KGDB_KDB)
10304 * These functions are only useful for kdb.
10306 * They can only be called when the whole system has been
10307 * stopped - every CPU needs to be quiescent, and no scheduling
10308 * activity can take place. Using them for anything else would
10309 * be a serious bug, and as a result, they aren't even visible
10310 * under any other configuration.
10314 * curr_task - return the current task for a given CPU.
10315 * @cpu: the processor in question.
10317 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10319 * Return: The current task for @cpu.
10321 struct task_struct *curr_task(int cpu)
10323 return cpu_curr(cpu);
10326 #endif /* defined(CONFIG_KGDB_KDB) */
10328 #ifdef CONFIG_CGROUP_SCHED
10329 /* task_group_lock serializes the addition/removal of task groups */
10330 static DEFINE_SPINLOCK(task_group_lock);
10332 static inline void alloc_uclamp_sched_group(struct task_group *tg,
10333 struct task_group *parent)
10335 #ifdef CONFIG_UCLAMP_TASK_GROUP
10336 enum uclamp_id clamp_id;
10338 for_each_clamp_id(clamp_id) {
10339 uclamp_se_set(&tg->uclamp_req[clamp_id],
10340 uclamp_none(clamp_id), false);
10341 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10346 static void sched_free_group(struct task_group *tg)
10348 free_fair_sched_group(tg);
10349 free_rt_sched_group(tg);
10350 autogroup_free(tg);
10351 kmem_cache_free(task_group_cache, tg);
10354 static void sched_free_group_rcu(struct rcu_head *rcu)
10356 sched_free_group(container_of(rcu, struct task_group, rcu));
10359 static void sched_unregister_group(struct task_group *tg)
10361 unregister_fair_sched_group(tg);
10362 unregister_rt_sched_group(tg);
10364 * We have to wait for yet another RCU grace period to expire, as
10365 * print_cfs_stats() might run concurrently.
10367 call_rcu(&tg->rcu, sched_free_group_rcu);
10370 /* allocate runqueue etc for a new task group */
10371 struct task_group *sched_create_group(struct task_group *parent)
10373 struct task_group *tg;
10375 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10377 return ERR_PTR(-ENOMEM);
10379 if (!alloc_fair_sched_group(tg, parent))
10382 if (!alloc_rt_sched_group(tg, parent))
10385 alloc_uclamp_sched_group(tg, parent);
10390 sched_free_group(tg);
10391 return ERR_PTR(-ENOMEM);
10394 void sched_online_group(struct task_group *tg, struct task_group *parent)
10396 unsigned long flags;
10398 spin_lock_irqsave(&task_group_lock, flags);
10399 list_add_rcu(&tg->list, &task_groups);
10401 /* Root should already exist: */
10404 tg->parent = parent;
10405 INIT_LIST_HEAD(&tg->children);
10406 list_add_rcu(&tg->siblings, &parent->children);
10407 spin_unlock_irqrestore(&task_group_lock, flags);
10409 online_fair_sched_group(tg);
10412 /* rcu callback to free various structures associated with a task group */
10413 static void sched_unregister_group_rcu(struct rcu_head *rhp)
10415 /* Now it should be safe to free those cfs_rqs: */
10416 sched_unregister_group(container_of(rhp, struct task_group, rcu));
10419 void sched_destroy_group(struct task_group *tg)
10421 /* Wait for possible concurrent references to cfs_rqs complete: */
10422 call_rcu(&tg->rcu, sched_unregister_group_rcu);
10425 void sched_release_group(struct task_group *tg)
10427 unsigned long flags;
10430 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10431 * sched_cfs_period_timer()).
10433 * For this to be effective, we have to wait for all pending users of
10434 * this task group to leave their RCU critical section to ensure no new
10435 * user will see our dying task group any more. Specifically ensure
10436 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10438 * We therefore defer calling unregister_fair_sched_group() to
10439 * sched_unregister_group() which is guarantied to get called only after the
10440 * current RCU grace period has expired.
10442 spin_lock_irqsave(&task_group_lock, flags);
10443 list_del_rcu(&tg->list);
10444 list_del_rcu(&tg->siblings);
10445 spin_unlock_irqrestore(&task_group_lock, flags);
10448 static struct task_group *sched_get_task_group(struct task_struct *tsk)
10450 struct task_group *tg;
10453 * All callers are synchronized by task_rq_lock(); we do not use RCU
10454 * which is pointless here. Thus, we pass "true" to task_css_check()
10455 * to prevent lockdep warnings.
10457 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10458 struct task_group, css);
10459 tg = autogroup_task_group(tsk, tg);
10464 static void sched_change_group(struct task_struct *tsk, struct task_group *group)
10466 tsk->sched_task_group = group;
10468 #ifdef CONFIG_FAIR_GROUP_SCHED
10469 if (tsk->sched_class->task_change_group)
10470 tsk->sched_class->task_change_group(tsk);
10473 set_task_rq(tsk, task_cpu(tsk));
10477 * Change task's runqueue when it moves between groups.
10479 * The caller of this function should have put the task in its new group by
10480 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10483 void sched_move_task(struct task_struct *tsk)
10485 int queued, running, queue_flags =
10486 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10487 struct task_group *group;
10490 CLASS(task_rq_lock, rq_guard)(tsk);
10494 * Esp. with SCHED_AUTOGROUP enabled it is possible to get superfluous
10497 group = sched_get_task_group(tsk);
10498 if (group == tsk->sched_task_group)
10501 update_rq_clock(rq);
10503 running = task_current(rq, tsk);
10504 queued = task_on_rq_queued(tsk);
10507 dequeue_task(rq, tsk, queue_flags);
10509 put_prev_task(rq, tsk);
10511 sched_change_group(tsk, group);
10514 enqueue_task(rq, tsk, queue_flags);
10516 set_next_task(rq, tsk);
10518 * After changing group, the running task may have joined a
10519 * throttled one but it's still the running task. Trigger a
10520 * resched to make sure that task can still run.
10526 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10528 return css ? container_of(css, struct task_group, css) : NULL;
10531 static struct cgroup_subsys_state *
10532 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10534 struct task_group *parent = css_tg(parent_css);
10535 struct task_group *tg;
10538 /* This is early initialization for the top cgroup */
10539 return &root_task_group.css;
10542 tg = sched_create_group(parent);
10544 return ERR_PTR(-ENOMEM);
10549 /* Expose task group only after completing cgroup initialization */
10550 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10552 struct task_group *tg = css_tg(css);
10553 struct task_group *parent = css_tg(css->parent);
10556 sched_online_group(tg, parent);
10558 #ifdef CONFIG_UCLAMP_TASK_GROUP
10559 /* Propagate the effective uclamp value for the new group */
10560 guard(mutex)(&uclamp_mutex);
10562 cpu_util_update_eff(css);
10568 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10570 struct task_group *tg = css_tg(css);
10572 sched_release_group(tg);
10575 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10577 struct task_group *tg = css_tg(css);
10580 * Relies on the RCU grace period between css_released() and this.
10582 sched_unregister_group(tg);
10585 #ifdef CONFIG_RT_GROUP_SCHED
10586 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10588 struct task_struct *task;
10589 struct cgroup_subsys_state *css;
10591 cgroup_taskset_for_each(task, css, tset) {
10592 if (!sched_rt_can_attach(css_tg(css), task))
10599 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10601 struct task_struct *task;
10602 struct cgroup_subsys_state *css;
10604 cgroup_taskset_for_each(task, css, tset)
10605 sched_move_task(task);
10608 #ifdef CONFIG_UCLAMP_TASK_GROUP
10609 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10611 struct cgroup_subsys_state *top_css = css;
10612 struct uclamp_se *uc_parent = NULL;
10613 struct uclamp_se *uc_se = NULL;
10614 unsigned int eff[UCLAMP_CNT];
10615 enum uclamp_id clamp_id;
10616 unsigned int clamps;
10618 lockdep_assert_held(&uclamp_mutex);
10619 SCHED_WARN_ON(!rcu_read_lock_held());
10621 css_for_each_descendant_pre(css, top_css) {
10622 uc_parent = css_tg(css)->parent
10623 ? css_tg(css)->parent->uclamp : NULL;
10625 for_each_clamp_id(clamp_id) {
10626 /* Assume effective clamps matches requested clamps */
10627 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10628 /* Cap effective clamps with parent's effective clamps */
10630 eff[clamp_id] > uc_parent[clamp_id].value) {
10631 eff[clamp_id] = uc_parent[clamp_id].value;
10634 /* Ensure protection is always capped by limit */
10635 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10637 /* Propagate most restrictive effective clamps */
10639 uc_se = css_tg(css)->uclamp;
10640 for_each_clamp_id(clamp_id) {
10641 if (eff[clamp_id] == uc_se[clamp_id].value)
10643 uc_se[clamp_id].value = eff[clamp_id];
10644 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10645 clamps |= (0x1 << clamp_id);
10648 css = css_rightmost_descendant(css);
10652 /* Immediately update descendants RUNNABLE tasks */
10653 uclamp_update_active_tasks(css);
10658 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10659 * C expression. Since there is no way to convert a macro argument (N) into a
10660 * character constant, use two levels of macros.
10662 #define _POW10(exp) ((unsigned int)1e##exp)
10663 #define POW10(exp) _POW10(exp)
10665 struct uclamp_request {
10666 #define UCLAMP_PERCENT_SHIFT 2
10667 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10673 static inline struct uclamp_request
10674 capacity_from_percent(char *buf)
10676 struct uclamp_request req = {
10677 .percent = UCLAMP_PERCENT_SCALE,
10678 .util = SCHED_CAPACITY_SCALE,
10683 if (strcmp(buf, "max")) {
10684 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10688 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10693 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10694 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10700 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10701 size_t nbytes, loff_t off,
10702 enum uclamp_id clamp_id)
10704 struct uclamp_request req;
10705 struct task_group *tg;
10707 req = capacity_from_percent(buf);
10711 static_branch_enable(&sched_uclamp_used);
10713 guard(mutex)(&uclamp_mutex);
10716 tg = css_tg(of_css(of));
10717 if (tg->uclamp_req[clamp_id].value != req.util)
10718 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10721 * Because of not recoverable conversion rounding we keep track of the
10722 * exact requested value
10724 tg->uclamp_pct[clamp_id] = req.percent;
10726 /* Update effective clamps to track the most restrictive value */
10727 cpu_util_update_eff(of_css(of));
10732 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10733 char *buf, size_t nbytes,
10736 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10739 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10740 char *buf, size_t nbytes,
10743 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10746 static inline void cpu_uclamp_print(struct seq_file *sf,
10747 enum uclamp_id clamp_id)
10749 struct task_group *tg;
10754 scoped_guard (rcu) {
10755 tg = css_tg(seq_css(sf));
10756 util_clamp = tg->uclamp_req[clamp_id].value;
10759 if (util_clamp == SCHED_CAPACITY_SCALE) {
10760 seq_puts(sf, "max\n");
10764 percent = tg->uclamp_pct[clamp_id];
10765 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10766 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10769 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10771 cpu_uclamp_print(sf, UCLAMP_MIN);
10775 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10777 cpu_uclamp_print(sf, UCLAMP_MAX);
10780 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10782 #ifdef CONFIG_FAIR_GROUP_SCHED
10783 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10784 struct cftype *cftype, u64 shareval)
10786 if (shareval > scale_load_down(ULONG_MAX))
10787 shareval = MAX_SHARES;
10788 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10791 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10792 struct cftype *cft)
10794 struct task_group *tg = css_tg(css);
10796 return (u64) scale_load_down(tg->shares);
10799 #ifdef CONFIG_CFS_BANDWIDTH
10800 static DEFINE_MUTEX(cfs_constraints_mutex);
10802 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10803 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10804 /* More than 203 days if BW_SHIFT equals 20. */
10805 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10807 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10809 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10812 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10813 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10815 if (tg == &root_task_group)
10819 * Ensure we have at some amount of bandwidth every period. This is
10820 * to prevent reaching a state of large arrears when throttled via
10821 * entity_tick() resulting in prolonged exit starvation.
10823 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10827 * Likewise, bound things on the other side by preventing insane quota
10828 * periods. This also allows us to normalize in computing quota
10831 if (period > max_cfs_quota_period)
10835 * Bound quota to defend quota against overflow during bandwidth shift.
10837 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10840 if (quota != RUNTIME_INF && (burst > quota ||
10841 burst + quota > max_cfs_runtime))
10845 * Prevent race between setting of cfs_rq->runtime_enabled and
10846 * unthrottle_offline_cfs_rqs().
10848 guard(cpus_read_lock)();
10849 guard(mutex)(&cfs_constraints_mutex);
10851 ret = __cfs_schedulable(tg, period, quota);
10855 runtime_enabled = quota != RUNTIME_INF;
10856 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10858 * If we need to toggle cfs_bandwidth_used, off->on must occur
10859 * before making related changes, and on->off must occur afterwards
10861 if (runtime_enabled && !runtime_was_enabled)
10862 cfs_bandwidth_usage_inc();
10864 scoped_guard (raw_spinlock_irq, &cfs_b->lock) {
10865 cfs_b->period = ns_to_ktime(period);
10866 cfs_b->quota = quota;
10867 cfs_b->burst = burst;
10869 __refill_cfs_bandwidth_runtime(cfs_b);
10872 * Restart the period timer (if active) to handle new
10875 if (runtime_enabled)
10876 start_cfs_bandwidth(cfs_b);
10879 for_each_online_cpu(i) {
10880 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10881 struct rq *rq = cfs_rq->rq;
10883 guard(rq_lock_irq)(rq);
10884 cfs_rq->runtime_enabled = runtime_enabled;
10885 cfs_rq->runtime_remaining = 0;
10887 if (cfs_rq->throttled)
10888 unthrottle_cfs_rq(cfs_rq);
10891 if (runtime_was_enabled && !runtime_enabled)
10892 cfs_bandwidth_usage_dec();
10897 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10899 u64 quota, period, burst;
10901 period = ktime_to_ns(tg->cfs_bandwidth.period);
10902 burst = tg->cfs_bandwidth.burst;
10903 if (cfs_quota_us < 0)
10904 quota = RUNTIME_INF;
10905 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10906 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10910 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10913 static long tg_get_cfs_quota(struct task_group *tg)
10917 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10920 quota_us = tg->cfs_bandwidth.quota;
10921 do_div(quota_us, NSEC_PER_USEC);
10926 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10928 u64 quota, period, burst;
10930 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10933 period = (u64)cfs_period_us * NSEC_PER_USEC;
10934 quota = tg->cfs_bandwidth.quota;
10935 burst = tg->cfs_bandwidth.burst;
10937 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10940 static long tg_get_cfs_period(struct task_group *tg)
10944 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10945 do_div(cfs_period_us, NSEC_PER_USEC);
10947 return cfs_period_us;
10950 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10952 u64 quota, period, burst;
10954 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10957 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10958 period = ktime_to_ns(tg->cfs_bandwidth.period);
10959 quota = tg->cfs_bandwidth.quota;
10961 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10964 static long tg_get_cfs_burst(struct task_group *tg)
10968 burst_us = tg->cfs_bandwidth.burst;
10969 do_div(burst_us, NSEC_PER_USEC);
10974 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10975 struct cftype *cft)
10977 return tg_get_cfs_quota(css_tg(css));
10980 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10981 struct cftype *cftype, s64 cfs_quota_us)
10983 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10986 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10987 struct cftype *cft)
10989 return tg_get_cfs_period(css_tg(css));
10992 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10993 struct cftype *cftype, u64 cfs_period_us)
10995 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10998 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10999 struct cftype *cft)
11001 return tg_get_cfs_burst(css_tg(css));
11004 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
11005 struct cftype *cftype, u64 cfs_burst_us)
11007 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
11010 struct cfs_schedulable_data {
11011 struct task_group *tg;
11016 * normalize group quota/period to be quota/max_period
11017 * note: units are usecs
11019 static u64 normalize_cfs_quota(struct task_group *tg,
11020 struct cfs_schedulable_data *d)
11025 period = d->period;
11028 period = tg_get_cfs_period(tg);
11029 quota = tg_get_cfs_quota(tg);
11032 /* note: these should typically be equivalent */
11033 if (quota == RUNTIME_INF || quota == -1)
11034 return RUNTIME_INF;
11036 return to_ratio(period, quota);
11039 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
11041 struct cfs_schedulable_data *d = data;
11042 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11043 s64 quota = 0, parent_quota = -1;
11046 quota = RUNTIME_INF;
11048 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
11050 quota = normalize_cfs_quota(tg, d);
11051 parent_quota = parent_b->hierarchical_quota;
11054 * Ensure max(child_quota) <= parent_quota. On cgroup2,
11055 * always take the non-RUNTIME_INF min. On cgroup1, only
11056 * inherit when no limit is set. In both cases this is used
11057 * by the scheduler to determine if a given CFS task has a
11058 * bandwidth constraint at some higher level.
11060 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
11061 if (quota == RUNTIME_INF)
11062 quota = parent_quota;
11063 else if (parent_quota != RUNTIME_INF)
11064 quota = min(quota, parent_quota);
11066 if (quota == RUNTIME_INF)
11067 quota = parent_quota;
11068 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
11072 cfs_b->hierarchical_quota = quota;
11077 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
11079 struct cfs_schedulable_data data = {
11085 if (quota != RUNTIME_INF) {
11086 do_div(data.period, NSEC_PER_USEC);
11087 do_div(data.quota, NSEC_PER_USEC);
11091 return walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
11094 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
11096 struct task_group *tg = css_tg(seq_css(sf));
11097 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11099 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
11100 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
11101 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
11103 if (schedstat_enabled() && tg != &root_task_group) {
11104 struct sched_statistics *stats;
11108 for_each_possible_cpu(i) {
11109 stats = __schedstats_from_se(tg->se[i]);
11110 ws += schedstat_val(stats->wait_sum);
11113 seq_printf(sf, "wait_sum %llu\n", ws);
11116 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
11117 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
11122 static u64 throttled_time_self(struct task_group *tg)
11127 for_each_possible_cpu(i) {
11128 total += READ_ONCE(tg->cfs_rq[i]->throttled_clock_self_time);
11134 static int cpu_cfs_local_stat_show(struct seq_file *sf, void *v)
11136 struct task_group *tg = css_tg(seq_css(sf));
11138 seq_printf(sf, "throttled_time %llu\n", throttled_time_self(tg));
11142 #endif /* CONFIG_CFS_BANDWIDTH */
11143 #endif /* CONFIG_FAIR_GROUP_SCHED */
11145 #ifdef CONFIG_RT_GROUP_SCHED
11146 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
11147 struct cftype *cft, s64 val)
11149 return sched_group_set_rt_runtime(css_tg(css), val);
11152 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
11153 struct cftype *cft)
11155 return sched_group_rt_runtime(css_tg(css));
11158 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
11159 struct cftype *cftype, u64 rt_period_us)
11161 return sched_group_set_rt_period(css_tg(css), rt_period_us);
11164 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
11165 struct cftype *cft)
11167 return sched_group_rt_period(css_tg(css));
11169 #endif /* CONFIG_RT_GROUP_SCHED */
11171 #ifdef CONFIG_FAIR_GROUP_SCHED
11172 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
11173 struct cftype *cft)
11175 return css_tg(css)->idle;
11178 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
11179 struct cftype *cft, s64 idle)
11181 return sched_group_set_idle(css_tg(css), idle);
11185 static struct cftype cpu_legacy_files[] = {
11186 #ifdef CONFIG_FAIR_GROUP_SCHED
11189 .read_u64 = cpu_shares_read_u64,
11190 .write_u64 = cpu_shares_write_u64,
11194 .read_s64 = cpu_idle_read_s64,
11195 .write_s64 = cpu_idle_write_s64,
11198 #ifdef CONFIG_CFS_BANDWIDTH
11200 .name = "cfs_quota_us",
11201 .read_s64 = cpu_cfs_quota_read_s64,
11202 .write_s64 = cpu_cfs_quota_write_s64,
11205 .name = "cfs_period_us",
11206 .read_u64 = cpu_cfs_period_read_u64,
11207 .write_u64 = cpu_cfs_period_write_u64,
11210 .name = "cfs_burst_us",
11211 .read_u64 = cpu_cfs_burst_read_u64,
11212 .write_u64 = cpu_cfs_burst_write_u64,
11216 .seq_show = cpu_cfs_stat_show,
11219 .name = "stat.local",
11220 .seq_show = cpu_cfs_local_stat_show,
11223 #ifdef CONFIG_RT_GROUP_SCHED
11225 .name = "rt_runtime_us",
11226 .read_s64 = cpu_rt_runtime_read,
11227 .write_s64 = cpu_rt_runtime_write,
11230 .name = "rt_period_us",
11231 .read_u64 = cpu_rt_period_read_uint,
11232 .write_u64 = cpu_rt_period_write_uint,
11235 #ifdef CONFIG_UCLAMP_TASK_GROUP
11237 .name = "uclamp.min",
11238 .flags = CFTYPE_NOT_ON_ROOT,
11239 .seq_show = cpu_uclamp_min_show,
11240 .write = cpu_uclamp_min_write,
11243 .name = "uclamp.max",
11244 .flags = CFTYPE_NOT_ON_ROOT,
11245 .seq_show = cpu_uclamp_max_show,
11246 .write = cpu_uclamp_max_write,
11249 { } /* Terminate */
11252 static int cpu_extra_stat_show(struct seq_file *sf,
11253 struct cgroup_subsys_state *css)
11255 #ifdef CONFIG_CFS_BANDWIDTH
11257 struct task_group *tg = css_tg(css);
11258 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11259 u64 throttled_usec, burst_usec;
11261 throttled_usec = cfs_b->throttled_time;
11262 do_div(throttled_usec, NSEC_PER_USEC);
11263 burst_usec = cfs_b->burst_time;
11264 do_div(burst_usec, NSEC_PER_USEC);
11266 seq_printf(sf, "nr_periods %d\n"
11267 "nr_throttled %d\n"
11268 "throttled_usec %llu\n"
11270 "burst_usec %llu\n",
11271 cfs_b->nr_periods, cfs_b->nr_throttled,
11272 throttled_usec, cfs_b->nr_burst, burst_usec);
11278 static int cpu_local_stat_show(struct seq_file *sf,
11279 struct cgroup_subsys_state *css)
11281 #ifdef CONFIG_CFS_BANDWIDTH
11283 struct task_group *tg = css_tg(css);
11284 u64 throttled_self_usec;
11286 throttled_self_usec = throttled_time_self(tg);
11287 do_div(throttled_self_usec, NSEC_PER_USEC);
11289 seq_printf(sf, "throttled_usec %llu\n",
11290 throttled_self_usec);
11296 #ifdef CONFIG_FAIR_GROUP_SCHED
11297 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
11298 struct cftype *cft)
11300 struct task_group *tg = css_tg(css);
11301 u64 weight = scale_load_down(tg->shares);
11303 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11306 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11307 struct cftype *cft, u64 weight)
11310 * cgroup weight knobs should use the common MIN, DFL and MAX
11311 * values which are 1, 100 and 10000 respectively. While it loses
11312 * a bit of range on both ends, it maps pretty well onto the shares
11313 * value used by scheduler and the round-trip conversions preserve
11314 * the original value over the entire range.
11316 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11319 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11321 return sched_group_set_shares(css_tg(css), scale_load(weight));
11324 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11325 struct cftype *cft)
11327 unsigned long weight = scale_load_down(css_tg(css)->shares);
11328 int last_delta = INT_MAX;
11331 /* find the closest nice value to the current weight */
11332 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11333 delta = abs(sched_prio_to_weight[prio] - weight);
11334 if (delta >= last_delta)
11336 last_delta = delta;
11339 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11342 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11343 struct cftype *cft, s64 nice)
11345 unsigned long weight;
11348 if (nice < MIN_NICE || nice > MAX_NICE)
11351 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11352 idx = array_index_nospec(idx, 40);
11353 weight = sched_prio_to_weight[idx];
11355 return sched_group_set_shares(css_tg(css), scale_load(weight));
11359 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11360 long period, long quota)
11363 seq_puts(sf, "max");
11365 seq_printf(sf, "%ld", quota);
11367 seq_printf(sf, " %ld\n", period);
11370 /* caller should put the current value in *@periodp before calling */
11371 static int __maybe_unused cpu_period_quota_parse(char *buf,
11372 u64 *periodp, u64 *quotap)
11374 char tok[21]; /* U64_MAX */
11376 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11379 *periodp *= NSEC_PER_USEC;
11381 if (sscanf(tok, "%llu", quotap))
11382 *quotap *= NSEC_PER_USEC;
11383 else if (!strcmp(tok, "max"))
11384 *quotap = RUNTIME_INF;
11391 #ifdef CONFIG_CFS_BANDWIDTH
11392 static int cpu_max_show(struct seq_file *sf, void *v)
11394 struct task_group *tg = css_tg(seq_css(sf));
11396 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11400 static ssize_t cpu_max_write(struct kernfs_open_file *of,
11401 char *buf, size_t nbytes, loff_t off)
11403 struct task_group *tg = css_tg(of_css(of));
11404 u64 period = tg_get_cfs_period(tg);
11405 u64 burst = tg_get_cfs_burst(tg);
11409 ret = cpu_period_quota_parse(buf, &period, "a);
11411 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11412 return ret ?: nbytes;
11416 static struct cftype cpu_files[] = {
11417 #ifdef CONFIG_FAIR_GROUP_SCHED
11420 .flags = CFTYPE_NOT_ON_ROOT,
11421 .read_u64 = cpu_weight_read_u64,
11422 .write_u64 = cpu_weight_write_u64,
11425 .name = "weight.nice",
11426 .flags = CFTYPE_NOT_ON_ROOT,
11427 .read_s64 = cpu_weight_nice_read_s64,
11428 .write_s64 = cpu_weight_nice_write_s64,
11432 .flags = CFTYPE_NOT_ON_ROOT,
11433 .read_s64 = cpu_idle_read_s64,
11434 .write_s64 = cpu_idle_write_s64,
11437 #ifdef CONFIG_CFS_BANDWIDTH
11440 .flags = CFTYPE_NOT_ON_ROOT,
11441 .seq_show = cpu_max_show,
11442 .write = cpu_max_write,
11445 .name = "max.burst",
11446 .flags = CFTYPE_NOT_ON_ROOT,
11447 .read_u64 = cpu_cfs_burst_read_u64,
11448 .write_u64 = cpu_cfs_burst_write_u64,
11451 #ifdef CONFIG_UCLAMP_TASK_GROUP
11453 .name = "uclamp.min",
11454 .flags = CFTYPE_NOT_ON_ROOT,
11455 .seq_show = cpu_uclamp_min_show,
11456 .write = cpu_uclamp_min_write,
11459 .name = "uclamp.max",
11460 .flags = CFTYPE_NOT_ON_ROOT,
11461 .seq_show = cpu_uclamp_max_show,
11462 .write = cpu_uclamp_max_write,
11465 { } /* terminate */
11468 struct cgroup_subsys cpu_cgrp_subsys = {
11469 .css_alloc = cpu_cgroup_css_alloc,
11470 .css_online = cpu_cgroup_css_online,
11471 .css_released = cpu_cgroup_css_released,
11472 .css_free = cpu_cgroup_css_free,
11473 .css_extra_stat_show = cpu_extra_stat_show,
11474 .css_local_stat_show = cpu_local_stat_show,
11475 #ifdef CONFIG_RT_GROUP_SCHED
11476 .can_attach = cpu_cgroup_can_attach,
11478 .attach = cpu_cgroup_attach,
11479 .legacy_cftypes = cpu_legacy_files,
11480 .dfl_cftypes = cpu_files,
11481 .early_init = true,
11485 #endif /* CONFIG_CGROUP_SCHED */
11487 void dump_cpu_task(int cpu)
11489 if (cpu == smp_processor_id() && in_hardirq()) {
11490 struct pt_regs *regs;
11492 regs = get_irq_regs();
11499 if (trigger_single_cpu_backtrace(cpu))
11502 pr_info("Task dump for CPU %d:\n", cpu);
11503 sched_show_task(cpu_curr(cpu));
11507 * Nice levels are multiplicative, with a gentle 10% change for every
11508 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11509 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11510 * that remained on nice 0.
11512 * The "10% effect" is relative and cumulative: from _any_ nice level,
11513 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11514 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11515 * If a task goes up by ~10% and another task goes down by ~10% then
11516 * the relative distance between them is ~25%.)
11518 const int sched_prio_to_weight[40] = {
11519 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11520 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11521 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11522 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11523 /* 0 */ 1024, 820, 655, 526, 423,
11524 /* 5 */ 335, 272, 215, 172, 137,
11525 /* 10 */ 110, 87, 70, 56, 45,
11526 /* 15 */ 36, 29, 23, 18, 15,
11530 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11532 * In cases where the weight does not change often, we can use the
11533 * precalculated inverse to speed up arithmetics by turning divisions
11534 * into multiplications:
11536 const u32 sched_prio_to_wmult[40] = {
11537 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11538 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11539 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11540 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11541 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11542 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11543 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11544 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11547 void call_trace_sched_update_nr_running(struct rq *rq, int count)
11549 trace_sched_update_nr_running_tp(rq, count);
11552 #ifdef CONFIG_SCHED_MM_CID
11555 * @cid_lock: Guarantee forward-progress of cid allocation.
11557 * Concurrency ID allocation within a bitmap is mostly lock-free. The cid_lock
11558 * is only used when contention is detected by the lock-free allocation so
11559 * forward progress can be guaranteed.
11561 DEFINE_RAW_SPINLOCK(cid_lock);
11564 * @use_cid_lock: Select cid allocation behavior: lock-free vs spinlock.
11566 * When @use_cid_lock is 0, the cid allocation is lock-free. When contention is
11567 * detected, it is set to 1 to ensure that all newly coming allocations are
11568 * serialized by @cid_lock until the allocation which detected contention
11569 * completes and sets @use_cid_lock back to 0. This guarantees forward progress
11570 * of a cid allocation.
11575 * mm_cid remote-clear implements a lock-free algorithm to clear per-mm/cpu cid
11576 * concurrently with respect to the execution of the source runqueue context
11579 * There is one basic properties we want to guarantee here:
11581 * (1) Remote-clear should _never_ mark a per-cpu cid UNSET when it is actively
11582 * used by a task. That would lead to concurrent allocation of the cid and
11583 * userspace corruption.
11585 * Provide this guarantee by introducing a Dekker memory ordering to guarantee
11586 * that a pair of loads observe at least one of a pair of stores, which can be
11595 * Which guarantees that x==0 && y==0 is impossible. But rather than using
11596 * values 0 and 1, this algorithm cares about specific state transitions of the
11597 * runqueue current task (as updated by the scheduler context switch), and the
11598 * per-mm/cpu cid value.
11600 * Let's introduce task (Y) which has task->mm == mm and task (N) which has
11601 * task->mm != mm for the rest of the discussion. There are two scheduler state
11602 * transitions on context switch we care about:
11604 * (TSA) Store to rq->curr with transition from (N) to (Y)
11606 * (TSB) Store to rq->curr with transition from (Y) to (N)
11608 * On the remote-clear side, there is one transition we care about:
11610 * (TMA) cmpxchg to *pcpu_cid to set the LAZY flag
11612 * There is also a transition to UNSET state which can be performed from all
11613 * sides (scheduler, remote-clear). It is always performed with a cmpxchg which
11614 * guarantees that only a single thread will succeed:
11616 * (TMB) cmpxchg to *pcpu_cid to mark UNSET
11618 * Just to be clear, what we do _not_ want to happen is a transition to UNSET
11619 * when a thread is actively using the cid (property (1)).
11621 * Let's looks at the relevant combinations of TSA/TSB, and TMA transitions.
11623 * Scenario A) (TSA)+(TMA) (from next task perspective)
11627 * Context switch CS-1 Remote-clear
11628 * - store to rq->curr: (N)->(Y) (TSA) - cmpxchg to *pcpu_id to LAZY (TMA)
11629 * (implied barrier after cmpxchg)
11630 * - switch_mm_cid()
11631 * - memory barrier (see switch_mm_cid()
11632 * comment explaining how this barrier
11633 * is combined with other scheduler
11635 * - mm_cid_get (next)
11636 * - READ_ONCE(*pcpu_cid) - rcu_dereference(src_rq->curr)
11638 * This Dekker ensures that either task (Y) is observed by the
11639 * rcu_dereference() or the LAZY flag is observed by READ_ONCE(), or both are
11642 * If task (Y) store is observed by rcu_dereference(), it means that there is
11643 * still an active task on the cpu. Remote-clear will therefore not transition
11644 * to UNSET, which fulfills property (1).
11646 * If task (Y) is not observed, but the lazy flag is observed by READ_ONCE(),
11647 * it will move its state to UNSET, which clears the percpu cid perhaps
11648 * uselessly (which is not an issue for correctness). Because task (Y) is not
11649 * observed, CPU1 can move ahead to set the state to UNSET. Because moving
11650 * state to UNSET is done with a cmpxchg expecting that the old state has the
11651 * LAZY flag set, only one thread will successfully UNSET.
11653 * If both states (LAZY flag and task (Y)) are observed, the thread on CPU0
11654 * will observe the LAZY flag and transition to UNSET (perhaps uselessly), and
11655 * CPU1 will observe task (Y) and do nothing more, which is fine.
11657 * What we are effectively preventing with this Dekker is a scenario where
11658 * neither LAZY flag nor store (Y) are observed, which would fail property (1)
11659 * because this would UNSET a cid which is actively used.
11662 void sched_mm_cid_migrate_from(struct task_struct *t)
11664 t->migrate_from_cpu = task_cpu(t);
11668 int __sched_mm_cid_migrate_from_fetch_cid(struct rq *src_rq,
11669 struct task_struct *t,
11670 struct mm_cid *src_pcpu_cid)
11672 struct mm_struct *mm = t->mm;
11673 struct task_struct *src_task;
11674 int src_cid, last_mm_cid;
11679 last_mm_cid = t->last_mm_cid;
11681 * If the migrated task has no last cid, or if the current
11682 * task on src rq uses the cid, it means the source cid does not need
11683 * to be moved to the destination cpu.
11685 if (last_mm_cid == -1)
11687 src_cid = READ_ONCE(src_pcpu_cid->cid);
11688 if (!mm_cid_is_valid(src_cid) || last_mm_cid != src_cid)
11692 * If we observe an active task using the mm on this rq, it means we
11693 * are not the last task to be migrated from this cpu for this mm, so
11694 * there is no need to move src_cid to the destination cpu.
11697 src_task = rcu_dereference(src_rq->curr);
11698 if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11699 t->last_mm_cid = -1;
11707 int __sched_mm_cid_migrate_from_try_steal_cid(struct rq *src_rq,
11708 struct task_struct *t,
11709 struct mm_cid *src_pcpu_cid,
11712 struct task_struct *src_task;
11713 struct mm_struct *mm = t->mm;
11720 * Attempt to clear the source cpu cid to move it to the destination
11723 lazy_cid = mm_cid_set_lazy_put(src_cid);
11724 if (!try_cmpxchg(&src_pcpu_cid->cid, &src_cid, lazy_cid))
11728 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11729 * rq->curr->mm matches the scheduler barrier in context_switch()
11730 * between store to rq->curr and load of prev and next task's
11733 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11734 * rq->curr->mm_cid_active matches the barrier in
11735 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11736 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11737 * load of per-mm/cpu cid.
11741 * If we observe an active task using the mm on this rq after setting
11742 * the lazy-put flag, this task will be responsible for transitioning
11743 * from lazy-put flag set to MM_CID_UNSET.
11745 scoped_guard (rcu) {
11746 src_task = rcu_dereference(src_rq->curr);
11747 if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11749 * We observed an active task for this mm, there is therefore
11750 * no point in moving this cid to the destination cpu.
11752 t->last_mm_cid = -1;
11758 * The src_cid is unused, so it can be unset.
11760 if (!try_cmpxchg(&src_pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11766 * Migration to dst cpu. Called with dst_rq lock held.
11767 * Interrupts are disabled, which keeps the window of cid ownership without the
11768 * source rq lock held small.
11770 void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t)
11772 struct mm_cid *src_pcpu_cid, *dst_pcpu_cid;
11773 struct mm_struct *mm = t->mm;
11774 int src_cid, dst_cid, src_cpu;
11777 lockdep_assert_rq_held(dst_rq);
11781 src_cpu = t->migrate_from_cpu;
11782 if (src_cpu == -1) {
11783 t->last_mm_cid = -1;
11787 * Move the src cid if the dst cid is unset. This keeps id
11788 * allocation closest to 0 in cases where few threads migrate around
11791 * If destination cid is already set, we may have to just clear
11792 * the src cid to ensure compactness in frequent migrations
11795 * It is not useful to clear the src cid when the number of threads is
11796 * greater or equal to the number of allowed cpus, because user-space
11797 * can expect that the number of allowed cids can reach the number of
11800 dst_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(dst_rq));
11801 dst_cid = READ_ONCE(dst_pcpu_cid->cid);
11802 if (!mm_cid_is_unset(dst_cid) &&
11803 atomic_read(&mm->mm_users) >= t->nr_cpus_allowed)
11805 src_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, src_cpu);
11806 src_rq = cpu_rq(src_cpu);
11807 src_cid = __sched_mm_cid_migrate_from_fetch_cid(src_rq, t, src_pcpu_cid);
11810 src_cid = __sched_mm_cid_migrate_from_try_steal_cid(src_rq, t, src_pcpu_cid,
11814 if (!mm_cid_is_unset(dst_cid)) {
11815 __mm_cid_put(mm, src_cid);
11818 /* Move src_cid to dst cpu. */
11819 mm_cid_snapshot_time(dst_rq, mm);
11820 WRITE_ONCE(dst_pcpu_cid->cid, src_cid);
11823 static void sched_mm_cid_remote_clear(struct mm_struct *mm, struct mm_cid *pcpu_cid,
11826 struct rq *rq = cpu_rq(cpu);
11827 struct task_struct *t;
11830 cid = READ_ONCE(pcpu_cid->cid);
11831 if (!mm_cid_is_valid(cid))
11835 * Clear the cpu cid if it is set to keep cid allocation compact. If
11836 * there happens to be other tasks left on the source cpu using this
11837 * mm, the next task using this mm will reallocate its cid on context
11840 lazy_cid = mm_cid_set_lazy_put(cid);
11841 if (!try_cmpxchg(&pcpu_cid->cid, &cid, lazy_cid))
11845 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11846 * rq->curr->mm matches the scheduler barrier in context_switch()
11847 * between store to rq->curr and load of prev and next task's
11850 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11851 * rq->curr->mm_cid_active matches the barrier in
11852 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11853 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11854 * load of per-mm/cpu cid.
11858 * If we observe an active task using the mm on this rq after setting
11859 * the lazy-put flag, that task will be responsible for transitioning
11860 * from lazy-put flag set to MM_CID_UNSET.
11862 scoped_guard (rcu) {
11863 t = rcu_dereference(rq->curr);
11864 if (READ_ONCE(t->mm_cid_active) && t->mm == mm)
11869 * The cid is unused, so it can be unset.
11870 * Disable interrupts to keep the window of cid ownership without rq
11873 scoped_guard (irqsave) {
11874 if (try_cmpxchg(&pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11875 __mm_cid_put(mm, cid);
11879 static void sched_mm_cid_remote_clear_old(struct mm_struct *mm, int cpu)
11881 struct rq *rq = cpu_rq(cpu);
11882 struct mm_cid *pcpu_cid;
11883 struct task_struct *curr;
11887 * rq->clock load is racy on 32-bit but one spurious clear once in a
11888 * while is irrelevant.
11890 rq_clock = READ_ONCE(rq->clock);
11891 pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11894 * In order to take care of infrequently scheduled tasks, bump the time
11895 * snapshot associated with this cid if an active task using the mm is
11896 * observed on this rq.
11898 scoped_guard (rcu) {
11899 curr = rcu_dereference(rq->curr);
11900 if (READ_ONCE(curr->mm_cid_active) && curr->mm == mm) {
11901 WRITE_ONCE(pcpu_cid->time, rq_clock);
11906 if (rq_clock < pcpu_cid->time + SCHED_MM_CID_PERIOD_NS)
11908 sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11911 static void sched_mm_cid_remote_clear_weight(struct mm_struct *mm, int cpu,
11914 struct mm_cid *pcpu_cid;
11917 pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11918 cid = READ_ONCE(pcpu_cid->cid);
11919 if (!mm_cid_is_valid(cid) || cid < weight)
11921 sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11924 static void task_mm_cid_work(struct callback_head *work)
11926 unsigned long now = jiffies, old_scan, next_scan;
11927 struct task_struct *t = current;
11928 struct cpumask *cidmask;
11929 struct mm_struct *mm;
11932 SCHED_WARN_ON(t != container_of(work, struct task_struct, cid_work));
11934 work->next = work; /* Prevent double-add */
11935 if (t->flags & PF_EXITING)
11940 old_scan = READ_ONCE(mm->mm_cid_next_scan);
11941 next_scan = now + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11945 res = cmpxchg(&mm->mm_cid_next_scan, old_scan, next_scan);
11946 if (res != old_scan)
11949 old_scan = next_scan;
11951 if (time_before(now, old_scan))
11953 if (!try_cmpxchg(&mm->mm_cid_next_scan, &old_scan, next_scan))
11955 cidmask = mm_cidmask(mm);
11956 /* Clear cids that were not recently used. */
11957 for_each_possible_cpu(cpu)
11958 sched_mm_cid_remote_clear_old(mm, cpu);
11959 weight = cpumask_weight(cidmask);
11961 * Clear cids that are greater or equal to the cidmask weight to
11964 for_each_possible_cpu(cpu)
11965 sched_mm_cid_remote_clear_weight(mm, cpu, weight);
11968 void init_sched_mm_cid(struct task_struct *t)
11970 struct mm_struct *mm = t->mm;
11974 mm_users = atomic_read(&mm->mm_users);
11976 mm->mm_cid_next_scan = jiffies + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11978 t->cid_work.next = &t->cid_work; /* Protect against double add */
11979 init_task_work(&t->cid_work, task_mm_cid_work);
11982 void task_tick_mm_cid(struct rq *rq, struct task_struct *curr)
11984 struct callback_head *work = &curr->cid_work;
11985 unsigned long now = jiffies;
11987 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) ||
11988 work->next != work)
11990 if (time_before(now, READ_ONCE(curr->mm->mm_cid_next_scan)))
11992 task_work_add(curr, work, TWA_RESUME);
11995 void sched_mm_cid_exit_signals(struct task_struct *t)
11997 struct mm_struct *mm = t->mm;
12005 guard(rq_lock_irqsave)(rq);
12006 preempt_enable_no_resched(); /* holding spinlock */
12007 WRITE_ONCE(t->mm_cid_active, 0);
12009 * Store t->mm_cid_active before loading per-mm/cpu cid.
12010 * Matches barrier in sched_mm_cid_remote_clear_old().
12014 t->last_mm_cid = t->mm_cid = -1;
12017 void sched_mm_cid_before_execve(struct task_struct *t)
12019 struct mm_struct *mm = t->mm;
12027 guard(rq_lock_irqsave)(rq);
12028 preempt_enable_no_resched(); /* holding spinlock */
12029 WRITE_ONCE(t->mm_cid_active, 0);
12031 * Store t->mm_cid_active before loading per-mm/cpu cid.
12032 * Matches barrier in sched_mm_cid_remote_clear_old().
12036 t->last_mm_cid = t->mm_cid = -1;
12039 void sched_mm_cid_after_execve(struct task_struct *t)
12041 struct mm_struct *mm = t->mm;
12049 scoped_guard (rq_lock_irqsave, rq) {
12050 preempt_enable_no_resched(); /* holding spinlock */
12051 WRITE_ONCE(t->mm_cid_active, 1);
12053 * Store t->mm_cid_active before loading per-mm/cpu cid.
12054 * Matches barrier in sched_mm_cid_remote_clear_old().
12057 t->last_mm_cid = t->mm_cid = mm_cid_get(rq, mm);
12059 rseq_set_notify_resume(t);
12062 void sched_mm_cid_fork(struct task_struct *t)
12064 WARN_ON_ONCE(!t->mm || t->mm_cid != -1);
12065 t->mm_cid_active = 1;