1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #include <linux/highmem.h>
10 #include <linux/hrtimer_api.h>
11 #include <linux/ktime_api.h>
12 #include <linux/sched/signal.h>
13 #include <linux/syscalls_api.h>
14 #include <linux/debug_locks.h>
15 #include <linux/prefetch.h>
16 #include <linux/capability.h>
17 #include <linux/pgtable_api.h>
18 #include <linux/wait_bit.h>
19 #include <linux/jiffies.h>
20 #include <linux/spinlock_api.h>
21 #include <linux/cpumask_api.h>
22 #include <linux/lockdep_api.h>
23 #include <linux/hardirq.h>
24 #include <linux/softirq.h>
25 #include <linux/refcount_api.h>
26 #include <linux/topology.h>
27 #include <linux/sched/clock.h>
28 #include <linux/sched/cond_resched.h>
29 #include <linux/sched/cputime.h>
30 #include <linux/sched/debug.h>
31 #include <linux/sched/hotplug.h>
32 #include <linux/sched/init.h>
33 #include <linux/sched/isolation.h>
34 #include <linux/sched/loadavg.h>
35 #include <linux/sched/mm.h>
36 #include <linux/sched/nohz.h>
37 #include <linux/sched/rseq_api.h>
38 #include <linux/sched/rt.h>
40 #include <linux/blkdev.h>
41 #include <linux/context_tracking.h>
42 #include <linux/cpuset.h>
43 #include <linux/delayacct.h>
44 #include <linux/init_task.h>
45 #include <linux/interrupt.h>
46 #include <linux/ioprio.h>
47 #include <linux/kallsyms.h>
48 #include <linux/kcov.h>
49 #include <linux/kprobes.h>
50 #include <linux/llist_api.h>
51 #include <linux/mmu_context.h>
52 #include <linux/mmzone.h>
53 #include <linux/mutex_api.h>
54 #include <linux/nmi.h>
55 #include <linux/nospec.h>
56 #include <linux/perf_event_api.h>
57 #include <linux/profile.h>
58 #include <linux/psi.h>
59 #include <linux/rcuwait_api.h>
60 #include <linux/sched/wake_q.h>
61 #include <linux/scs.h>
62 #include <linux/slab.h>
63 #include <linux/syscalls.h>
64 #include <linux/vtime.h>
65 #include <linux/wait_api.h>
66 #include <linux/workqueue_api.h>
68 #ifdef CONFIG_PREEMPT_DYNAMIC
69 # ifdef CONFIG_GENERIC_ENTRY
70 # include <linux/entry-common.h>
74 #include <uapi/linux/sched/types.h>
76 #include <asm/switch_to.h>
79 #define CREATE_TRACE_POINTS
80 #include <linux/sched/rseq_api.h>
81 #include <trace/events/sched.h>
82 #undef CREATE_TRACE_POINTS
86 #include "autogroup.h"
88 #include "autogroup.h"
93 #include "../workqueue_internal.h"
94 #include "../../io_uring/io-wq.h"
95 #include "../smpboot.h"
98 * Export tracepoints that act as a bare tracehook (ie: have no trace event
99 * associated with them) to allow external modules to probe them.
101 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
102 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
103 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
104 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
105 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
106 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
107 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
108 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
109 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
110 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
111 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 #ifdef CONFIG_SCHED_DEBUG
117 * Debugging: various feature bits
119 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
120 * sysctl_sched_features, defined in sched.h, to allow constants propagation
121 * at compile time and compiler optimization based on features default.
123 #define SCHED_FEAT(name, enabled) \
124 (1UL << __SCHED_FEAT_##name) * enabled |
125 const_debug unsigned int sysctl_sched_features =
126 #include "features.h"
131 * Print a warning if need_resched is set for the given duration (if
132 * LATENCY_WARN is enabled).
134 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
137 __read_mostly int sysctl_resched_latency_warn_ms = 100;
138 __read_mostly int sysctl_resched_latency_warn_once = 1;
139 #endif /* CONFIG_SCHED_DEBUG */
142 * Number of tasks to iterate in a single balance run.
143 * Limited because this is done with IRQs disabled.
145 #ifdef CONFIG_PREEMPT_RT
146 const_debug unsigned int sysctl_sched_nr_migrate = 8;
148 const_debug unsigned int sysctl_sched_nr_migrate = 32;
151 __read_mostly int scheduler_running;
153 #ifdef CONFIG_SCHED_CORE
155 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
157 /* kernel prio, less is more */
158 static inline int __task_prio(struct task_struct *p)
160 if (p->sched_class == &stop_sched_class) /* trumps deadline */
163 if (rt_prio(p->prio)) /* includes deadline */
164 return p->prio; /* [-1, 99] */
166 if (p->sched_class == &idle_sched_class)
167 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
169 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
179 /* real prio, less is less */
180 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
183 int pa = __task_prio(a), pb = __task_prio(b);
191 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
192 return !dl_time_before(a->dl.deadline, b->dl.deadline);
194 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
195 return cfs_prio_less(a, b, in_fi);
200 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
202 if (a->core_cookie < b->core_cookie)
205 if (a->core_cookie > b->core_cookie)
208 /* flip prio, so high prio is leftmost */
209 if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
215 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
217 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
219 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
222 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
224 const struct task_struct *p = __node_2_sc(node);
225 unsigned long cookie = (unsigned long)key;
227 if (cookie < p->core_cookie)
230 if (cookie > p->core_cookie)
236 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
238 rq->core->core_task_seq++;
243 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
246 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
248 rq->core->core_task_seq++;
250 if (sched_core_enqueued(p)) {
251 rb_erase(&p->core_node, &rq->core_tree);
252 RB_CLEAR_NODE(&p->core_node);
256 * Migrating the last task off the cpu, with the cpu in forced idle
257 * state. Reschedule to create an accounting edge for forced idle,
258 * and re-examine whether the core is still in forced idle state.
260 if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
261 rq->core->core_forceidle_count && rq->curr == rq->idle)
266 * Find left-most (aka, highest priority) task matching @cookie.
268 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
270 struct rb_node *node;
272 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
274 * The idle task always matches any cookie!
277 return idle_sched_class.pick_task(rq);
279 return __node_2_sc(node);
282 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
284 struct rb_node *node = &p->core_node;
286 node = rb_next(node);
290 p = container_of(node, struct task_struct, core_node);
291 if (p->core_cookie != cookie)
298 * Magic required such that:
300 * raw_spin_rq_lock(rq);
302 * raw_spin_rq_unlock(rq);
304 * ends up locking and unlocking the _same_ lock, and all CPUs
305 * always agree on what rq has what lock.
307 * XXX entirely possible to selectively enable cores, don't bother for now.
310 static DEFINE_MUTEX(sched_core_mutex);
311 static atomic_t sched_core_count;
312 static struct cpumask sched_core_mask;
314 static void sched_core_lock(int cpu, unsigned long *flags)
316 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
319 local_irq_save(*flags);
320 for_each_cpu(t, smt_mask)
321 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
324 static void sched_core_unlock(int cpu, unsigned long *flags)
326 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
329 for_each_cpu(t, smt_mask)
330 raw_spin_unlock(&cpu_rq(t)->__lock);
331 local_irq_restore(*flags);
334 static void __sched_core_flip(bool enabled)
342 * Toggle the online cores, one by one.
344 cpumask_copy(&sched_core_mask, cpu_online_mask);
345 for_each_cpu(cpu, &sched_core_mask) {
346 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
348 sched_core_lock(cpu, &flags);
350 for_each_cpu(t, smt_mask)
351 cpu_rq(t)->core_enabled = enabled;
353 cpu_rq(cpu)->core->core_forceidle_start = 0;
355 sched_core_unlock(cpu, &flags);
357 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
361 * Toggle the offline CPUs.
363 cpumask_copy(&sched_core_mask, cpu_possible_mask);
364 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
366 for_each_cpu(cpu, &sched_core_mask)
367 cpu_rq(cpu)->core_enabled = enabled;
372 static void sched_core_assert_empty(void)
376 for_each_possible_cpu(cpu)
377 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
380 static void __sched_core_enable(void)
382 static_branch_enable(&__sched_core_enabled);
384 * Ensure all previous instances of raw_spin_rq_*lock() have finished
385 * and future ones will observe !sched_core_disabled().
388 __sched_core_flip(true);
389 sched_core_assert_empty();
392 static void __sched_core_disable(void)
394 sched_core_assert_empty();
395 __sched_core_flip(false);
396 static_branch_disable(&__sched_core_enabled);
399 void sched_core_get(void)
401 if (atomic_inc_not_zero(&sched_core_count))
404 mutex_lock(&sched_core_mutex);
405 if (!atomic_read(&sched_core_count))
406 __sched_core_enable();
408 smp_mb__before_atomic();
409 atomic_inc(&sched_core_count);
410 mutex_unlock(&sched_core_mutex);
413 static void __sched_core_put(struct work_struct *work)
415 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
416 __sched_core_disable();
417 mutex_unlock(&sched_core_mutex);
421 void sched_core_put(void)
423 static DECLARE_WORK(_work, __sched_core_put);
426 * "There can be only one"
428 * Either this is the last one, or we don't actually need to do any
429 * 'work'. If it is the last *again*, we rely on
430 * WORK_STRUCT_PENDING_BIT.
432 if (!atomic_add_unless(&sched_core_count, -1, 1))
433 schedule_work(&_work);
436 #else /* !CONFIG_SCHED_CORE */
438 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
440 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
442 #endif /* CONFIG_SCHED_CORE */
445 * Serialization rules:
451 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
454 * rq2->lock where: rq1 < rq2
458 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
459 * local CPU's rq->lock, it optionally removes the task from the runqueue and
460 * always looks at the local rq data structures to find the most eligible task
463 * Task enqueue is also under rq->lock, possibly taken from another CPU.
464 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
465 * the local CPU to avoid bouncing the runqueue state around [ see
466 * ttwu_queue_wakelist() ]
468 * Task wakeup, specifically wakeups that involve migration, are horribly
469 * complicated to avoid having to take two rq->locks.
473 * System-calls and anything external will use task_rq_lock() which acquires
474 * both p->pi_lock and rq->lock. As a consequence the state they change is
475 * stable while holding either lock:
477 * - sched_setaffinity()/
478 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
479 * - set_user_nice(): p->se.load, p->*prio
480 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
481 * p->se.load, p->rt_priority,
482 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
483 * - sched_setnuma(): p->numa_preferred_nid
484 * - sched_move_task()/
485 * cpu_cgroup_fork(): p->sched_task_group
486 * - uclamp_update_active() p->uclamp*
488 * p->state <- TASK_*:
490 * is changed locklessly using set_current_state(), __set_current_state() or
491 * set_special_state(), see their respective comments, or by
492 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
495 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
497 * is set by activate_task() and cleared by deactivate_task(), under
498 * rq->lock. Non-zero indicates the task is runnable, the special
499 * ON_RQ_MIGRATING state is used for migration without holding both
500 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
502 * p->on_cpu <- { 0, 1 }:
504 * is set by prepare_task() and cleared by finish_task() such that it will be
505 * set before p is scheduled-in and cleared after p is scheduled-out, both
506 * under rq->lock. Non-zero indicates the task is running on its CPU.
508 * [ The astute reader will observe that it is possible for two tasks on one
509 * CPU to have ->on_cpu = 1 at the same time. ]
511 * task_cpu(p): is changed by set_task_cpu(), the rules are:
513 * - Don't call set_task_cpu() on a blocked task:
515 * We don't care what CPU we're not running on, this simplifies hotplug,
516 * the CPU assignment of blocked tasks isn't required to be valid.
518 * - for try_to_wake_up(), called under p->pi_lock:
520 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
522 * - for migration called under rq->lock:
523 * [ see task_on_rq_migrating() in task_rq_lock() ]
525 * o move_queued_task()
528 * - for migration called under double_rq_lock():
530 * o __migrate_swap_task()
531 * o push_rt_task() / pull_rt_task()
532 * o push_dl_task() / pull_dl_task()
533 * o dl_task_offline_migration()
537 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
539 raw_spinlock_t *lock;
541 /* Matches synchronize_rcu() in __sched_core_enable() */
543 if (sched_core_disabled()) {
544 raw_spin_lock_nested(&rq->__lock, subclass);
545 /* preempt_count *MUST* be > 1 */
546 preempt_enable_no_resched();
551 lock = __rq_lockp(rq);
552 raw_spin_lock_nested(lock, subclass);
553 if (likely(lock == __rq_lockp(rq))) {
554 /* preempt_count *MUST* be > 1 */
555 preempt_enable_no_resched();
558 raw_spin_unlock(lock);
562 bool raw_spin_rq_trylock(struct rq *rq)
564 raw_spinlock_t *lock;
567 /* Matches synchronize_rcu() in __sched_core_enable() */
569 if (sched_core_disabled()) {
570 ret = raw_spin_trylock(&rq->__lock);
576 lock = __rq_lockp(rq);
577 ret = raw_spin_trylock(lock);
578 if (!ret || (likely(lock == __rq_lockp(rq)))) {
582 raw_spin_unlock(lock);
586 void raw_spin_rq_unlock(struct rq *rq)
588 raw_spin_unlock(rq_lockp(rq));
593 * double_rq_lock - safely lock two runqueues
595 void double_rq_lock(struct rq *rq1, struct rq *rq2)
597 lockdep_assert_irqs_disabled();
599 if (rq_order_less(rq2, rq1))
602 raw_spin_rq_lock(rq1);
603 if (__rq_lockp(rq1) != __rq_lockp(rq2))
604 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
606 double_rq_clock_clear_update(rq1, rq2);
611 * __task_rq_lock - lock the rq @p resides on.
613 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
618 lockdep_assert_held(&p->pi_lock);
622 raw_spin_rq_lock(rq);
623 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
627 raw_spin_rq_unlock(rq);
629 while (unlikely(task_on_rq_migrating(p)))
635 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
637 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
638 __acquires(p->pi_lock)
644 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
646 raw_spin_rq_lock(rq);
648 * move_queued_task() task_rq_lock()
651 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
652 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
653 * [S] ->cpu = new_cpu [L] task_rq()
657 * If we observe the old CPU in task_rq_lock(), the acquire of
658 * the old rq->lock will fully serialize against the stores.
660 * If we observe the new CPU in task_rq_lock(), the address
661 * dependency headed by '[L] rq = task_rq()' and the acquire
662 * will pair with the WMB to ensure we then also see migrating.
664 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
668 raw_spin_rq_unlock(rq);
669 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
671 while (unlikely(task_on_rq_migrating(p)))
677 * RQ-clock updating methods:
680 static void update_rq_clock_task(struct rq *rq, s64 delta)
683 * In theory, the compile should just see 0 here, and optimize out the call
684 * to sched_rt_avg_update. But I don't trust it...
686 s64 __maybe_unused steal = 0, irq_delta = 0;
688 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
689 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
692 * Since irq_time is only updated on {soft,}irq_exit, we might run into
693 * this case when a previous update_rq_clock() happened inside a
696 * When this happens, we stop ->clock_task and only update the
697 * prev_irq_time stamp to account for the part that fit, so that a next
698 * update will consume the rest. This ensures ->clock_task is
701 * It does however cause some slight miss-attribution of {soft,}irq
702 * time, a more accurate solution would be to update the irq_time using
703 * the current rq->clock timestamp, except that would require using
706 if (irq_delta > delta)
709 rq->prev_irq_time += irq_delta;
712 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
713 if (static_key_false((¶virt_steal_rq_enabled))) {
714 steal = paravirt_steal_clock(cpu_of(rq));
715 steal -= rq->prev_steal_time_rq;
717 if (unlikely(steal > delta))
720 rq->prev_steal_time_rq += steal;
725 rq->clock_task += delta;
727 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
728 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
729 update_irq_load_avg(rq, irq_delta + steal);
731 update_rq_clock_pelt(rq, delta);
734 void update_rq_clock(struct rq *rq)
738 lockdep_assert_rq_held(rq);
740 if (rq->clock_update_flags & RQCF_ACT_SKIP)
743 #ifdef CONFIG_SCHED_DEBUG
744 if (sched_feat(WARN_DOUBLE_CLOCK))
745 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
746 rq->clock_update_flags |= RQCF_UPDATED;
749 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
753 update_rq_clock_task(rq, delta);
756 #ifdef CONFIG_SCHED_HRTICK
758 * Use HR-timers to deliver accurate preemption points.
761 static void hrtick_clear(struct rq *rq)
763 if (hrtimer_active(&rq->hrtick_timer))
764 hrtimer_cancel(&rq->hrtick_timer);
768 * High-resolution timer tick.
769 * Runs from hardirq context with interrupts disabled.
771 static enum hrtimer_restart hrtick(struct hrtimer *timer)
773 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
776 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
780 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
783 return HRTIMER_NORESTART;
788 static void __hrtick_restart(struct rq *rq)
790 struct hrtimer *timer = &rq->hrtick_timer;
791 ktime_t time = rq->hrtick_time;
793 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
797 * called from hardirq (IPI) context
799 static void __hrtick_start(void *arg)
805 __hrtick_restart(rq);
810 * Called to set the hrtick timer state.
812 * called with rq->lock held and irqs disabled
814 void hrtick_start(struct rq *rq, u64 delay)
816 struct hrtimer *timer = &rq->hrtick_timer;
820 * Don't schedule slices shorter than 10000ns, that just
821 * doesn't make sense and can cause timer DoS.
823 delta = max_t(s64, delay, 10000LL);
824 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
827 __hrtick_restart(rq);
829 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
834 * Called to set the hrtick timer state.
836 * called with rq->lock held and irqs disabled
838 void hrtick_start(struct rq *rq, u64 delay)
841 * Don't schedule slices shorter than 10000ns, that just
842 * doesn't make sense. Rely on vruntime for fairness.
844 delay = max_t(u64, delay, 10000LL);
845 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
846 HRTIMER_MODE_REL_PINNED_HARD);
849 #endif /* CONFIG_SMP */
851 static void hrtick_rq_init(struct rq *rq)
854 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
856 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
857 rq->hrtick_timer.function = hrtick;
859 #else /* CONFIG_SCHED_HRTICK */
860 static inline void hrtick_clear(struct rq *rq)
864 static inline void hrtick_rq_init(struct rq *rq)
867 #endif /* CONFIG_SCHED_HRTICK */
870 * cmpxchg based fetch_or, macro so it works for different integer types
872 #define fetch_or(ptr, mask) \
874 typeof(ptr) _ptr = (ptr); \
875 typeof(mask) _mask = (mask); \
876 typeof(*_ptr) _val = *_ptr; \
879 } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
883 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
885 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
886 * this avoids any races wrt polling state changes and thereby avoids
889 static inline bool set_nr_and_not_polling(struct task_struct *p)
891 struct thread_info *ti = task_thread_info(p);
892 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
896 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
898 * If this returns true, then the idle task promises to call
899 * sched_ttwu_pending() and reschedule soon.
901 static bool set_nr_if_polling(struct task_struct *p)
903 struct thread_info *ti = task_thread_info(p);
904 typeof(ti->flags) val = READ_ONCE(ti->flags);
907 if (!(val & _TIF_POLLING_NRFLAG))
909 if (val & _TIF_NEED_RESCHED)
911 if (try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED))
918 static inline bool set_nr_and_not_polling(struct task_struct *p)
920 set_tsk_need_resched(p);
925 static inline bool set_nr_if_polling(struct task_struct *p)
932 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
934 struct wake_q_node *node = &task->wake_q;
937 * Atomically grab the task, if ->wake_q is !nil already it means
938 * it's already queued (either by us or someone else) and will get the
939 * wakeup due to that.
941 * In order to ensure that a pending wakeup will observe our pending
942 * state, even in the failed case, an explicit smp_mb() must be used.
944 smp_mb__before_atomic();
945 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
949 * The head is context local, there can be no concurrency.
952 head->lastp = &node->next;
957 * wake_q_add() - queue a wakeup for 'later' waking.
958 * @head: the wake_q_head to add @task to
959 * @task: the task to queue for 'later' wakeup
961 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
962 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
965 * This function must be used as-if it were wake_up_process(); IOW the task
966 * must be ready to be woken at this location.
968 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
970 if (__wake_q_add(head, task))
971 get_task_struct(task);
975 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
976 * @head: the wake_q_head to add @task to
977 * @task: the task to queue for 'later' wakeup
979 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
980 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
983 * This function must be used as-if it were wake_up_process(); IOW the task
984 * must be ready to be woken at this location.
986 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
987 * that already hold reference to @task can call the 'safe' version and trust
988 * wake_q to do the right thing depending whether or not the @task is already
991 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
993 if (!__wake_q_add(head, task))
994 put_task_struct(task);
997 void wake_up_q(struct wake_q_head *head)
999 struct wake_q_node *node = head->first;
1001 while (node != WAKE_Q_TAIL) {
1002 struct task_struct *task;
1004 task = container_of(node, struct task_struct, wake_q);
1005 /* Task can safely be re-inserted now: */
1007 task->wake_q.next = NULL;
1010 * wake_up_process() executes a full barrier, which pairs with
1011 * the queueing in wake_q_add() so as not to miss wakeups.
1013 wake_up_process(task);
1014 put_task_struct(task);
1019 * resched_curr - mark rq's current task 'to be rescheduled now'.
1021 * On UP this means the setting of the need_resched flag, on SMP it
1022 * might also involve a cross-CPU call to trigger the scheduler on
1025 void resched_curr(struct rq *rq)
1027 struct task_struct *curr = rq->curr;
1030 lockdep_assert_rq_held(rq);
1032 if (test_tsk_need_resched(curr))
1037 if (cpu == smp_processor_id()) {
1038 set_tsk_need_resched(curr);
1039 set_preempt_need_resched();
1043 if (set_nr_and_not_polling(curr))
1044 smp_send_reschedule(cpu);
1046 trace_sched_wake_idle_without_ipi(cpu);
1049 void resched_cpu(int cpu)
1051 struct rq *rq = cpu_rq(cpu);
1052 unsigned long flags;
1054 raw_spin_rq_lock_irqsave(rq, flags);
1055 if (cpu_online(cpu) || cpu == smp_processor_id())
1057 raw_spin_rq_unlock_irqrestore(rq, flags);
1061 #ifdef CONFIG_NO_HZ_COMMON
1063 * In the semi idle case, use the nearest busy CPU for migrating timers
1064 * from an idle CPU. This is good for power-savings.
1066 * We don't do similar optimization for completely idle system, as
1067 * selecting an idle CPU will add more delays to the timers than intended
1068 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1070 int get_nohz_timer_target(void)
1072 int i, cpu = smp_processor_id(), default_cpu = -1;
1073 struct sched_domain *sd;
1074 const struct cpumask *hk_mask;
1076 if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1082 hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1085 for_each_domain(cpu, sd) {
1086 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1097 if (default_cpu == -1)
1098 default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1106 * When add_timer_on() enqueues a timer into the timer wheel of an
1107 * idle CPU then this timer might expire before the next timer event
1108 * which is scheduled to wake up that CPU. In case of a completely
1109 * idle system the next event might even be infinite time into the
1110 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1111 * leaves the inner idle loop so the newly added timer is taken into
1112 * account when the CPU goes back to idle and evaluates the timer
1113 * wheel for the next timer event.
1115 static void wake_up_idle_cpu(int cpu)
1117 struct rq *rq = cpu_rq(cpu);
1119 if (cpu == smp_processor_id())
1122 if (set_nr_and_not_polling(rq->idle))
1123 smp_send_reschedule(cpu);
1125 trace_sched_wake_idle_without_ipi(cpu);
1128 static bool wake_up_full_nohz_cpu(int cpu)
1131 * We just need the target to call irq_exit() and re-evaluate
1132 * the next tick. The nohz full kick at least implies that.
1133 * If needed we can still optimize that later with an
1136 if (cpu_is_offline(cpu))
1137 return true; /* Don't try to wake offline CPUs. */
1138 if (tick_nohz_full_cpu(cpu)) {
1139 if (cpu != smp_processor_id() ||
1140 tick_nohz_tick_stopped())
1141 tick_nohz_full_kick_cpu(cpu);
1149 * Wake up the specified CPU. If the CPU is going offline, it is the
1150 * caller's responsibility to deal with the lost wakeup, for example,
1151 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1153 void wake_up_nohz_cpu(int cpu)
1155 if (!wake_up_full_nohz_cpu(cpu))
1156 wake_up_idle_cpu(cpu);
1159 static void nohz_csd_func(void *info)
1161 struct rq *rq = info;
1162 int cpu = cpu_of(rq);
1166 * Release the rq::nohz_csd.
1168 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1169 WARN_ON(!(flags & NOHZ_KICK_MASK));
1171 rq->idle_balance = idle_cpu(cpu);
1172 if (rq->idle_balance && !need_resched()) {
1173 rq->nohz_idle_balance = flags;
1174 raise_softirq_irqoff(SCHED_SOFTIRQ);
1178 #endif /* CONFIG_NO_HZ_COMMON */
1180 #ifdef CONFIG_NO_HZ_FULL
1181 bool sched_can_stop_tick(struct rq *rq)
1183 int fifo_nr_running;
1185 /* Deadline tasks, even if single, need the tick */
1186 if (rq->dl.dl_nr_running)
1190 * If there are more than one RR tasks, we need the tick to affect the
1191 * actual RR behaviour.
1193 if (rq->rt.rr_nr_running) {
1194 if (rq->rt.rr_nr_running == 1)
1201 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1202 * forced preemption between FIFO tasks.
1204 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1205 if (fifo_nr_running)
1209 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1210 * if there's more than one we need the tick for involuntary
1213 if (rq->nr_running > 1)
1218 #endif /* CONFIG_NO_HZ_FULL */
1219 #endif /* CONFIG_SMP */
1221 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1222 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1224 * Iterate task_group tree rooted at *from, calling @down when first entering a
1225 * node and @up when leaving it for the final time.
1227 * Caller must hold rcu_lock or sufficient equivalent.
1229 int walk_tg_tree_from(struct task_group *from,
1230 tg_visitor down, tg_visitor up, void *data)
1232 struct task_group *parent, *child;
1238 ret = (*down)(parent, data);
1241 list_for_each_entry_rcu(child, &parent->children, siblings) {
1248 ret = (*up)(parent, data);
1249 if (ret || parent == from)
1253 parent = parent->parent;
1260 int tg_nop(struct task_group *tg, void *data)
1266 static void set_load_weight(struct task_struct *p, bool update_load)
1268 int prio = p->static_prio - MAX_RT_PRIO;
1269 struct load_weight *load = &p->se.load;
1272 * SCHED_IDLE tasks get minimal weight:
1274 if (task_has_idle_policy(p)) {
1275 load->weight = scale_load(WEIGHT_IDLEPRIO);
1276 load->inv_weight = WMULT_IDLEPRIO;
1281 * SCHED_OTHER tasks have to update their load when changing their
1284 if (update_load && p->sched_class == &fair_sched_class) {
1285 reweight_task(p, prio);
1287 load->weight = scale_load(sched_prio_to_weight[prio]);
1288 load->inv_weight = sched_prio_to_wmult[prio];
1292 #ifdef CONFIG_UCLAMP_TASK
1294 * Serializes updates of utilization clamp values
1296 * The (slow-path) user-space triggers utilization clamp value updates which
1297 * can require updates on (fast-path) scheduler's data structures used to
1298 * support enqueue/dequeue operations.
1299 * While the per-CPU rq lock protects fast-path update operations, user-space
1300 * requests are serialized using a mutex to reduce the risk of conflicting
1301 * updates or API abuses.
1303 static DEFINE_MUTEX(uclamp_mutex);
1305 /* Max allowed minimum utilization */
1306 static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1308 /* Max allowed maximum utilization */
1309 static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1312 * By default RT tasks run at the maximum performance point/capacity of the
1313 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1314 * SCHED_CAPACITY_SCALE.
1316 * This knob allows admins to change the default behavior when uclamp is being
1317 * used. In battery powered devices, particularly, running at the maximum
1318 * capacity and frequency will increase energy consumption and shorten the
1321 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1323 * This knob will not override the system default sched_util_clamp_min defined
1326 static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1328 /* All clamps are required to be less or equal than these values */
1329 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1332 * This static key is used to reduce the uclamp overhead in the fast path. It
1333 * primarily disables the call to uclamp_rq_{inc, dec}() in
1334 * enqueue/dequeue_task().
1336 * This allows users to continue to enable uclamp in their kernel config with
1337 * minimum uclamp overhead in the fast path.
1339 * As soon as userspace modifies any of the uclamp knobs, the static key is
1340 * enabled, since we have an actual users that make use of uclamp
1343 * The knobs that would enable this static key are:
1345 * * A task modifying its uclamp value with sched_setattr().
1346 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1347 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1349 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1351 /* Integer rounded range for each bucket */
1352 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1354 #define for_each_clamp_id(clamp_id) \
1355 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1357 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1359 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1362 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1364 if (clamp_id == UCLAMP_MIN)
1366 return SCHED_CAPACITY_SCALE;
1369 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1370 unsigned int value, bool user_defined)
1372 uc_se->value = value;
1373 uc_se->bucket_id = uclamp_bucket_id(value);
1374 uc_se->user_defined = user_defined;
1377 static inline unsigned int
1378 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1379 unsigned int clamp_value)
1382 * Avoid blocked utilization pushing up the frequency when we go
1383 * idle (which drops the max-clamp) by retaining the last known
1386 if (clamp_id == UCLAMP_MAX) {
1387 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1391 return uclamp_none(UCLAMP_MIN);
1394 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1395 unsigned int clamp_value)
1397 /* Reset max-clamp retention only on idle exit */
1398 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1401 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1405 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1406 unsigned int clamp_value)
1408 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1409 int bucket_id = UCLAMP_BUCKETS - 1;
1412 * Since both min and max clamps are max aggregated, find the
1413 * top most bucket with tasks in.
1415 for ( ; bucket_id >= 0; bucket_id--) {
1416 if (!bucket[bucket_id].tasks)
1418 return bucket[bucket_id].value;
1421 /* No tasks -- default clamp values */
1422 return uclamp_idle_value(rq, clamp_id, clamp_value);
1425 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1427 unsigned int default_util_min;
1428 struct uclamp_se *uc_se;
1430 lockdep_assert_held(&p->pi_lock);
1432 uc_se = &p->uclamp_req[UCLAMP_MIN];
1434 /* Only sync if user didn't override the default */
1435 if (uc_se->user_defined)
1438 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1439 uclamp_se_set(uc_se, default_util_min, false);
1442 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1450 /* Protect updates to p->uclamp_* */
1451 rq = task_rq_lock(p, &rf);
1452 __uclamp_update_util_min_rt_default(p);
1453 task_rq_unlock(rq, p, &rf);
1456 static inline struct uclamp_se
1457 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1459 /* Copy by value as we could modify it */
1460 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1461 #ifdef CONFIG_UCLAMP_TASK_GROUP
1462 unsigned int tg_min, tg_max, value;
1465 * Tasks in autogroups or root task group will be
1466 * restricted by system defaults.
1468 if (task_group_is_autogroup(task_group(p)))
1470 if (task_group(p) == &root_task_group)
1473 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1474 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1475 value = uc_req.value;
1476 value = clamp(value, tg_min, tg_max);
1477 uclamp_se_set(&uc_req, value, false);
1484 * The effective clamp bucket index of a task depends on, by increasing
1486 * - the task specific clamp value, when explicitly requested from userspace
1487 * - the task group effective clamp value, for tasks not either in the root
1488 * group or in an autogroup
1489 * - the system default clamp value, defined by the sysadmin
1491 static inline struct uclamp_se
1492 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1494 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1495 struct uclamp_se uc_max = uclamp_default[clamp_id];
1497 /* System default restrictions always apply */
1498 if (unlikely(uc_req.value > uc_max.value))
1504 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1506 struct uclamp_se uc_eff;
1508 /* Task currently refcounted: use back-annotated (effective) value */
1509 if (p->uclamp[clamp_id].active)
1510 return (unsigned long)p->uclamp[clamp_id].value;
1512 uc_eff = uclamp_eff_get(p, clamp_id);
1514 return (unsigned long)uc_eff.value;
1518 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1519 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1520 * updates the rq's clamp value if required.
1522 * Tasks can have a task-specific value requested from user-space, track
1523 * within each bucket the maximum value for tasks refcounted in it.
1524 * This "local max aggregation" allows to track the exact "requested" value
1525 * for each bucket when all its RUNNABLE tasks require the same clamp.
1527 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1528 enum uclamp_id clamp_id)
1530 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1531 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1532 struct uclamp_bucket *bucket;
1534 lockdep_assert_rq_held(rq);
1536 /* Update task effective clamp */
1537 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1539 bucket = &uc_rq->bucket[uc_se->bucket_id];
1541 uc_se->active = true;
1543 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1546 * Local max aggregation: rq buckets always track the max
1547 * "requested" clamp value of its RUNNABLE tasks.
1549 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1550 bucket->value = uc_se->value;
1552 if (uc_se->value > READ_ONCE(uc_rq->value))
1553 WRITE_ONCE(uc_rq->value, uc_se->value);
1557 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1558 * is released. If this is the last task reference counting the rq's max
1559 * active clamp value, then the rq's clamp value is updated.
1561 * Both refcounted tasks and rq's cached clamp values are expected to be
1562 * always valid. If it's detected they are not, as defensive programming,
1563 * enforce the expected state and warn.
1565 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1566 enum uclamp_id clamp_id)
1568 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1569 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1570 struct uclamp_bucket *bucket;
1571 unsigned int bkt_clamp;
1572 unsigned int rq_clamp;
1574 lockdep_assert_rq_held(rq);
1577 * If sched_uclamp_used was enabled after task @p was enqueued,
1578 * we could end up with unbalanced call to uclamp_rq_dec_id().
1580 * In this case the uc_se->active flag should be false since no uclamp
1581 * accounting was performed at enqueue time and we can just return
1584 * Need to be careful of the following enqueue/dequeue ordering
1588 * // sched_uclamp_used gets enabled
1591 * // Must not decrement bucket->tasks here
1594 * where we could end up with stale data in uc_se and
1595 * bucket[uc_se->bucket_id].
1597 * The following check here eliminates the possibility of such race.
1599 if (unlikely(!uc_se->active))
1602 bucket = &uc_rq->bucket[uc_se->bucket_id];
1604 SCHED_WARN_ON(!bucket->tasks);
1605 if (likely(bucket->tasks))
1608 uc_se->active = false;
1611 * Keep "local max aggregation" simple and accept to (possibly)
1612 * overboost some RUNNABLE tasks in the same bucket.
1613 * The rq clamp bucket value is reset to its base value whenever
1614 * there are no more RUNNABLE tasks refcounting it.
1616 if (likely(bucket->tasks))
1619 rq_clamp = READ_ONCE(uc_rq->value);
1621 * Defensive programming: this should never happen. If it happens,
1622 * e.g. due to future modification, warn and fixup the expected value.
1624 SCHED_WARN_ON(bucket->value > rq_clamp);
1625 if (bucket->value >= rq_clamp) {
1626 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1627 WRITE_ONCE(uc_rq->value, bkt_clamp);
1631 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1633 enum uclamp_id clamp_id;
1636 * Avoid any overhead until uclamp is actually used by the userspace.
1638 * The condition is constructed such that a NOP is generated when
1639 * sched_uclamp_used is disabled.
1641 if (!static_branch_unlikely(&sched_uclamp_used))
1644 if (unlikely(!p->sched_class->uclamp_enabled))
1647 for_each_clamp_id(clamp_id)
1648 uclamp_rq_inc_id(rq, p, clamp_id);
1650 /* Reset clamp idle holding when there is one RUNNABLE task */
1651 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1652 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1655 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1657 enum uclamp_id clamp_id;
1660 * Avoid any overhead until uclamp is actually used by the userspace.
1662 * The condition is constructed such that a NOP is generated when
1663 * sched_uclamp_used is disabled.
1665 if (!static_branch_unlikely(&sched_uclamp_used))
1668 if (unlikely(!p->sched_class->uclamp_enabled))
1671 for_each_clamp_id(clamp_id)
1672 uclamp_rq_dec_id(rq, p, clamp_id);
1675 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1676 enum uclamp_id clamp_id)
1678 if (!p->uclamp[clamp_id].active)
1681 uclamp_rq_dec_id(rq, p, clamp_id);
1682 uclamp_rq_inc_id(rq, p, clamp_id);
1685 * Make sure to clear the idle flag if we've transiently reached 0
1686 * active tasks on rq.
1688 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1689 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1693 uclamp_update_active(struct task_struct *p)
1695 enum uclamp_id clamp_id;
1700 * Lock the task and the rq where the task is (or was) queued.
1702 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1703 * price to pay to safely serialize util_{min,max} updates with
1704 * enqueues, dequeues and migration operations.
1705 * This is the same locking schema used by __set_cpus_allowed_ptr().
1707 rq = task_rq_lock(p, &rf);
1710 * Setting the clamp bucket is serialized by task_rq_lock().
1711 * If the task is not yet RUNNABLE and its task_struct is not
1712 * affecting a valid clamp bucket, the next time it's enqueued,
1713 * it will already see the updated clamp bucket value.
1715 for_each_clamp_id(clamp_id)
1716 uclamp_rq_reinc_id(rq, p, clamp_id);
1718 task_rq_unlock(rq, p, &rf);
1721 #ifdef CONFIG_UCLAMP_TASK_GROUP
1723 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1725 struct css_task_iter it;
1726 struct task_struct *p;
1728 css_task_iter_start(css, 0, &it);
1729 while ((p = css_task_iter_next(&it)))
1730 uclamp_update_active(p);
1731 css_task_iter_end(&it);
1734 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1737 #ifdef CONFIG_SYSCTL
1738 #ifdef CONFIG_UCLAMP_TASK
1739 #ifdef CONFIG_UCLAMP_TASK_GROUP
1740 static void uclamp_update_root_tg(void)
1742 struct task_group *tg = &root_task_group;
1744 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1745 sysctl_sched_uclamp_util_min, false);
1746 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1747 sysctl_sched_uclamp_util_max, false);
1750 cpu_util_update_eff(&root_task_group.css);
1754 static void uclamp_update_root_tg(void) { }
1757 static void uclamp_sync_util_min_rt_default(void)
1759 struct task_struct *g, *p;
1762 * copy_process() sysctl_uclamp
1763 * uclamp_min_rt = X;
1764 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1765 * // link thread smp_mb__after_spinlock()
1766 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1767 * sched_post_fork() for_each_process_thread()
1768 * __uclamp_sync_rt() __uclamp_sync_rt()
1770 * Ensures that either sched_post_fork() will observe the new
1771 * uclamp_min_rt or for_each_process_thread() will observe the new
1774 read_lock(&tasklist_lock);
1775 smp_mb__after_spinlock();
1776 read_unlock(&tasklist_lock);
1779 for_each_process_thread(g, p)
1780 uclamp_update_util_min_rt_default(p);
1784 static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1785 void *buffer, size_t *lenp, loff_t *ppos)
1787 bool update_root_tg = false;
1788 int old_min, old_max, old_min_rt;
1791 mutex_lock(&uclamp_mutex);
1792 old_min = sysctl_sched_uclamp_util_min;
1793 old_max = sysctl_sched_uclamp_util_max;
1794 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1796 result = proc_dointvec(table, write, buffer, lenp, ppos);
1802 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1803 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1804 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1810 if (old_min != sysctl_sched_uclamp_util_min) {
1811 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1812 sysctl_sched_uclamp_util_min, false);
1813 update_root_tg = true;
1815 if (old_max != sysctl_sched_uclamp_util_max) {
1816 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1817 sysctl_sched_uclamp_util_max, false);
1818 update_root_tg = true;
1821 if (update_root_tg) {
1822 static_branch_enable(&sched_uclamp_used);
1823 uclamp_update_root_tg();
1826 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1827 static_branch_enable(&sched_uclamp_used);
1828 uclamp_sync_util_min_rt_default();
1832 * We update all RUNNABLE tasks only when task groups are in use.
1833 * Otherwise, keep it simple and do just a lazy update at each next
1834 * task enqueue time.
1840 sysctl_sched_uclamp_util_min = old_min;
1841 sysctl_sched_uclamp_util_max = old_max;
1842 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1844 mutex_unlock(&uclamp_mutex);
1851 static int uclamp_validate(struct task_struct *p,
1852 const struct sched_attr *attr)
1854 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1855 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1857 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1858 util_min = attr->sched_util_min;
1860 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1864 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1865 util_max = attr->sched_util_max;
1867 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1871 if (util_min != -1 && util_max != -1 && util_min > util_max)
1875 * We have valid uclamp attributes; make sure uclamp is enabled.
1877 * We need to do that here, because enabling static branches is a
1878 * blocking operation which obviously cannot be done while holding
1881 static_branch_enable(&sched_uclamp_used);
1886 static bool uclamp_reset(const struct sched_attr *attr,
1887 enum uclamp_id clamp_id,
1888 struct uclamp_se *uc_se)
1890 /* Reset on sched class change for a non user-defined clamp value. */
1891 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1892 !uc_se->user_defined)
1895 /* Reset on sched_util_{min,max} == -1. */
1896 if (clamp_id == UCLAMP_MIN &&
1897 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1898 attr->sched_util_min == -1) {
1902 if (clamp_id == UCLAMP_MAX &&
1903 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1904 attr->sched_util_max == -1) {
1911 static void __setscheduler_uclamp(struct task_struct *p,
1912 const struct sched_attr *attr)
1914 enum uclamp_id clamp_id;
1916 for_each_clamp_id(clamp_id) {
1917 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1920 if (!uclamp_reset(attr, clamp_id, uc_se))
1924 * RT by default have a 100% boost value that could be modified
1927 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1928 value = sysctl_sched_uclamp_util_min_rt_default;
1930 value = uclamp_none(clamp_id);
1932 uclamp_se_set(uc_se, value, false);
1936 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1939 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1940 attr->sched_util_min != -1) {
1941 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1942 attr->sched_util_min, true);
1945 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1946 attr->sched_util_max != -1) {
1947 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1948 attr->sched_util_max, true);
1952 static void uclamp_fork(struct task_struct *p)
1954 enum uclamp_id clamp_id;
1957 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1958 * as the task is still at its early fork stages.
1960 for_each_clamp_id(clamp_id)
1961 p->uclamp[clamp_id].active = false;
1963 if (likely(!p->sched_reset_on_fork))
1966 for_each_clamp_id(clamp_id) {
1967 uclamp_se_set(&p->uclamp_req[clamp_id],
1968 uclamp_none(clamp_id), false);
1972 static void uclamp_post_fork(struct task_struct *p)
1974 uclamp_update_util_min_rt_default(p);
1977 static void __init init_uclamp_rq(struct rq *rq)
1979 enum uclamp_id clamp_id;
1980 struct uclamp_rq *uc_rq = rq->uclamp;
1982 for_each_clamp_id(clamp_id) {
1983 uc_rq[clamp_id] = (struct uclamp_rq) {
1984 .value = uclamp_none(clamp_id)
1988 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1991 static void __init init_uclamp(void)
1993 struct uclamp_se uc_max = {};
1994 enum uclamp_id clamp_id;
1997 for_each_possible_cpu(cpu)
1998 init_uclamp_rq(cpu_rq(cpu));
2000 for_each_clamp_id(clamp_id) {
2001 uclamp_se_set(&init_task.uclamp_req[clamp_id],
2002 uclamp_none(clamp_id), false);
2005 /* System defaults allow max clamp values for both indexes */
2006 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2007 for_each_clamp_id(clamp_id) {
2008 uclamp_default[clamp_id] = uc_max;
2009 #ifdef CONFIG_UCLAMP_TASK_GROUP
2010 root_task_group.uclamp_req[clamp_id] = uc_max;
2011 root_task_group.uclamp[clamp_id] = uc_max;
2016 #else /* CONFIG_UCLAMP_TASK */
2017 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2018 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
2019 static inline int uclamp_validate(struct task_struct *p,
2020 const struct sched_attr *attr)
2024 static void __setscheduler_uclamp(struct task_struct *p,
2025 const struct sched_attr *attr) { }
2026 static inline void uclamp_fork(struct task_struct *p) { }
2027 static inline void uclamp_post_fork(struct task_struct *p) { }
2028 static inline void init_uclamp(void) { }
2029 #endif /* CONFIG_UCLAMP_TASK */
2031 bool sched_task_on_rq(struct task_struct *p)
2033 return task_on_rq_queued(p);
2036 unsigned long get_wchan(struct task_struct *p)
2038 unsigned long ip = 0;
2041 if (!p || p == current)
2044 /* Only get wchan if task is blocked and we can keep it that way. */
2045 raw_spin_lock_irq(&p->pi_lock);
2046 state = READ_ONCE(p->__state);
2047 smp_rmb(); /* see try_to_wake_up() */
2048 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2049 ip = __get_wchan(p);
2050 raw_spin_unlock_irq(&p->pi_lock);
2055 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2057 if (!(flags & ENQUEUE_NOCLOCK))
2058 update_rq_clock(rq);
2060 if (!(flags & ENQUEUE_RESTORE)) {
2061 sched_info_enqueue(rq, p);
2062 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
2065 uclamp_rq_inc(rq, p);
2066 p->sched_class->enqueue_task(rq, p, flags);
2068 if (sched_core_enabled(rq))
2069 sched_core_enqueue(rq, p);
2072 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2074 if (sched_core_enabled(rq))
2075 sched_core_dequeue(rq, p, flags);
2077 if (!(flags & DEQUEUE_NOCLOCK))
2078 update_rq_clock(rq);
2080 if (!(flags & DEQUEUE_SAVE)) {
2081 sched_info_dequeue(rq, p);
2082 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2085 uclamp_rq_dec(rq, p);
2086 p->sched_class->dequeue_task(rq, p, flags);
2089 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2091 enqueue_task(rq, p, flags);
2093 p->on_rq = TASK_ON_RQ_QUEUED;
2096 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2098 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2100 dequeue_task(rq, p, flags);
2103 static inline int __normal_prio(int policy, int rt_prio, int nice)
2107 if (dl_policy(policy))
2108 prio = MAX_DL_PRIO - 1;
2109 else if (rt_policy(policy))
2110 prio = MAX_RT_PRIO - 1 - rt_prio;
2112 prio = NICE_TO_PRIO(nice);
2118 * Calculate the expected normal priority: i.e. priority
2119 * without taking RT-inheritance into account. Might be
2120 * boosted by interactivity modifiers. Changes upon fork,
2121 * setprio syscalls, and whenever the interactivity
2122 * estimator recalculates.
2124 static inline int normal_prio(struct task_struct *p)
2126 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2130 * Calculate the current priority, i.e. the priority
2131 * taken into account by the scheduler. This value might
2132 * be boosted by RT tasks, or might be boosted by
2133 * interactivity modifiers. Will be RT if the task got
2134 * RT-boosted. If not then it returns p->normal_prio.
2136 static int effective_prio(struct task_struct *p)
2138 p->normal_prio = normal_prio(p);
2140 * If we are RT tasks or we were boosted to RT priority,
2141 * keep the priority unchanged. Otherwise, update priority
2142 * to the normal priority:
2144 if (!rt_prio(p->prio))
2145 return p->normal_prio;
2150 * task_curr - is this task currently executing on a CPU?
2151 * @p: the task in question.
2153 * Return: 1 if the task is currently executing. 0 otherwise.
2155 inline int task_curr(const struct task_struct *p)
2157 return cpu_curr(task_cpu(p)) == p;
2161 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2162 * use the balance_callback list if you want balancing.
2164 * this means any call to check_class_changed() must be followed by a call to
2165 * balance_callback().
2167 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2168 const struct sched_class *prev_class,
2171 if (prev_class != p->sched_class) {
2172 if (prev_class->switched_from)
2173 prev_class->switched_from(rq, p);
2175 p->sched_class->switched_to(rq, p);
2176 } else if (oldprio != p->prio || dl_task(p))
2177 p->sched_class->prio_changed(rq, p, oldprio);
2180 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2182 if (p->sched_class == rq->curr->sched_class)
2183 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2184 else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2188 * A queue event has occurred, and we're going to schedule. In
2189 * this case, we can save a useless back to back clock update.
2191 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2192 rq_clock_skip_update(rq);
2198 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2200 static int __set_cpus_allowed_ptr(struct task_struct *p,
2201 const struct cpumask *new_mask,
2204 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2206 if (likely(!p->migration_disabled))
2209 if (p->cpus_ptr != &p->cpus_mask)
2213 * Violates locking rules! see comment in __do_set_cpus_allowed().
2215 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2218 void migrate_disable(void)
2220 struct task_struct *p = current;
2222 if (p->migration_disabled) {
2223 p->migration_disabled++;
2228 this_rq()->nr_pinned++;
2229 p->migration_disabled = 1;
2232 EXPORT_SYMBOL_GPL(migrate_disable);
2234 void migrate_enable(void)
2236 struct task_struct *p = current;
2238 if (p->migration_disabled > 1) {
2239 p->migration_disabled--;
2243 if (WARN_ON_ONCE(!p->migration_disabled))
2247 * Ensure stop_task runs either before or after this, and that
2248 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2251 if (p->cpus_ptr != &p->cpus_mask)
2252 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2254 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2255 * regular cpus_mask, otherwise things that race (eg.
2256 * select_fallback_rq) get confused.
2259 p->migration_disabled = 0;
2260 this_rq()->nr_pinned--;
2263 EXPORT_SYMBOL_GPL(migrate_enable);
2265 static inline bool rq_has_pinned_tasks(struct rq *rq)
2267 return rq->nr_pinned;
2271 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2272 * __set_cpus_allowed_ptr() and select_fallback_rq().
2274 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2276 /* When not in the task's cpumask, no point in looking further. */
2277 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2280 /* migrate_disabled() must be allowed to finish. */
2281 if (is_migration_disabled(p))
2282 return cpu_online(cpu);
2284 /* Non kernel threads are not allowed during either online or offline. */
2285 if (!(p->flags & PF_KTHREAD))
2286 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2288 /* KTHREAD_IS_PER_CPU is always allowed. */
2289 if (kthread_is_per_cpu(p))
2290 return cpu_online(cpu);
2292 /* Regular kernel threads don't get to stay during offline. */
2296 /* But are allowed during online. */
2297 return cpu_online(cpu);
2301 * This is how migration works:
2303 * 1) we invoke migration_cpu_stop() on the target CPU using
2305 * 2) stopper starts to run (implicitly forcing the migrated thread
2307 * 3) it checks whether the migrated task is still in the wrong runqueue.
2308 * 4) if it's in the wrong runqueue then the migration thread removes
2309 * it and puts it into the right queue.
2310 * 5) stopper completes and stop_one_cpu() returns and the migration
2315 * move_queued_task - move a queued task to new rq.
2317 * Returns (locked) new rq. Old rq's lock is released.
2319 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2320 struct task_struct *p, int new_cpu)
2322 lockdep_assert_rq_held(rq);
2324 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2325 set_task_cpu(p, new_cpu);
2328 rq = cpu_rq(new_cpu);
2331 BUG_ON(task_cpu(p) != new_cpu);
2332 activate_task(rq, p, 0);
2333 check_preempt_curr(rq, p, 0);
2338 struct migration_arg {
2339 struct task_struct *task;
2341 struct set_affinity_pending *pending;
2345 * @refs: number of wait_for_completion()
2346 * @stop_pending: is @stop_work in use
2348 struct set_affinity_pending {
2350 unsigned int stop_pending;
2351 struct completion done;
2352 struct cpu_stop_work stop_work;
2353 struct migration_arg arg;
2357 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2358 * this because either it can't run here any more (set_cpus_allowed()
2359 * away from this CPU, or CPU going down), or because we're
2360 * attempting to rebalance this task on exec (sched_exec).
2362 * So we race with normal scheduler movements, but that's OK, as long
2363 * as the task is no longer on this CPU.
2365 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2366 struct task_struct *p, int dest_cpu)
2368 /* Affinity changed (again). */
2369 if (!is_cpu_allowed(p, dest_cpu))
2372 update_rq_clock(rq);
2373 rq = move_queued_task(rq, rf, p, dest_cpu);
2379 * migration_cpu_stop - this will be executed by a highprio stopper thread
2380 * and performs thread migration by bumping thread off CPU then
2381 * 'pushing' onto another runqueue.
2383 static int migration_cpu_stop(void *data)
2385 struct migration_arg *arg = data;
2386 struct set_affinity_pending *pending = arg->pending;
2387 struct task_struct *p = arg->task;
2388 struct rq *rq = this_rq();
2389 bool complete = false;
2393 * The original target CPU might have gone down and we might
2394 * be on another CPU but it doesn't matter.
2396 local_irq_save(rf.flags);
2398 * We need to explicitly wake pending tasks before running
2399 * __migrate_task() such that we will not miss enforcing cpus_ptr
2400 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2402 flush_smp_call_function_queue();
2404 raw_spin_lock(&p->pi_lock);
2408 * If we were passed a pending, then ->stop_pending was set, thus
2409 * p->migration_pending must have remained stable.
2411 WARN_ON_ONCE(pending && pending != p->migration_pending);
2414 * If task_rq(p) != rq, it cannot be migrated here, because we're
2415 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2416 * we're holding p->pi_lock.
2418 if (task_rq(p) == rq) {
2419 if (is_migration_disabled(p))
2423 p->migration_pending = NULL;
2426 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2430 if (task_on_rq_queued(p))
2431 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2433 p->wake_cpu = arg->dest_cpu;
2436 * XXX __migrate_task() can fail, at which point we might end
2437 * up running on a dodgy CPU, AFAICT this can only happen
2438 * during CPU hotplug, at which point we'll get pushed out
2439 * anyway, so it's probably not a big deal.
2442 } else if (pending) {
2444 * This happens when we get migrated between migrate_enable()'s
2445 * preempt_enable() and scheduling the stopper task. At that
2446 * point we're a regular task again and not current anymore.
2448 * A !PREEMPT kernel has a giant hole here, which makes it far
2453 * The task moved before the stopper got to run. We're holding
2454 * ->pi_lock, so the allowed mask is stable - if it got
2455 * somewhere allowed, we're done.
2457 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2458 p->migration_pending = NULL;
2464 * When migrate_enable() hits a rq mis-match we can't reliably
2465 * determine is_migration_disabled() and so have to chase after
2468 WARN_ON_ONCE(!pending->stop_pending);
2469 task_rq_unlock(rq, p, &rf);
2470 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2471 &pending->arg, &pending->stop_work);
2476 pending->stop_pending = false;
2477 task_rq_unlock(rq, p, &rf);
2480 complete_all(&pending->done);
2485 int push_cpu_stop(void *arg)
2487 struct rq *lowest_rq = NULL, *rq = this_rq();
2488 struct task_struct *p = arg;
2490 raw_spin_lock_irq(&p->pi_lock);
2491 raw_spin_rq_lock(rq);
2493 if (task_rq(p) != rq)
2496 if (is_migration_disabled(p)) {
2497 p->migration_flags |= MDF_PUSH;
2501 p->migration_flags &= ~MDF_PUSH;
2503 if (p->sched_class->find_lock_rq)
2504 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2509 // XXX validate p is still the highest prio task
2510 if (task_rq(p) == rq) {
2511 deactivate_task(rq, p, 0);
2512 set_task_cpu(p, lowest_rq->cpu);
2513 activate_task(lowest_rq, p, 0);
2514 resched_curr(lowest_rq);
2517 double_unlock_balance(rq, lowest_rq);
2520 rq->push_busy = false;
2521 raw_spin_rq_unlock(rq);
2522 raw_spin_unlock_irq(&p->pi_lock);
2529 * sched_class::set_cpus_allowed must do the below, but is not required to
2530 * actually call this function.
2532 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2534 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2535 p->cpus_ptr = new_mask;
2539 cpumask_copy(&p->cpus_mask, new_mask);
2540 p->nr_cpus_allowed = cpumask_weight(new_mask);
2544 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2546 struct rq *rq = task_rq(p);
2547 bool queued, running;
2550 * This here violates the locking rules for affinity, since we're only
2551 * supposed to change these variables while holding both rq->lock and
2554 * HOWEVER, it magically works, because ttwu() is the only code that
2555 * accesses these variables under p->pi_lock and only does so after
2556 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2557 * before finish_task().
2559 * XXX do further audits, this smells like something putrid.
2561 if (flags & SCA_MIGRATE_DISABLE)
2562 SCHED_WARN_ON(!p->on_cpu);
2564 lockdep_assert_held(&p->pi_lock);
2566 queued = task_on_rq_queued(p);
2567 running = task_current(rq, p);
2571 * Because __kthread_bind() calls this on blocked tasks without
2574 lockdep_assert_rq_held(rq);
2575 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2578 put_prev_task(rq, p);
2580 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2583 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2585 set_next_task(rq, p);
2588 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2590 __do_set_cpus_allowed(p, new_mask, 0);
2593 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2596 if (!src->user_cpus_ptr)
2599 dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2600 if (!dst->user_cpus_ptr)
2603 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2607 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2609 struct cpumask *user_mask = NULL;
2611 swap(p->user_cpus_ptr, user_mask);
2616 void release_user_cpus_ptr(struct task_struct *p)
2618 kfree(clear_user_cpus_ptr(p));
2622 * This function is wildly self concurrent; here be dragons.
2625 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2626 * designated task is enqueued on an allowed CPU. If that task is currently
2627 * running, we have to kick it out using the CPU stopper.
2629 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2632 * Initial conditions: P0->cpus_mask = [0, 1]
2636 * migrate_disable();
2638 * set_cpus_allowed_ptr(P0, [1]);
2640 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2641 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2642 * This means we need the following scheme:
2646 * migrate_disable();
2648 * set_cpus_allowed_ptr(P0, [1]);
2652 * __set_cpus_allowed_ptr();
2653 * <wakes local stopper>
2654 * `--> <woken on migration completion>
2656 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2657 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2658 * task p are serialized by p->pi_lock, which we can leverage: the one that
2659 * should come into effect at the end of the Migrate-Disable region is the last
2660 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2661 * but we still need to properly signal those waiting tasks at the appropriate
2664 * This is implemented using struct set_affinity_pending. The first
2665 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2666 * setup an instance of that struct and install it on the targeted task_struct.
2667 * Any and all further callers will reuse that instance. Those then wait for
2668 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2669 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2672 * (1) In the cases covered above. There is one more where the completion is
2673 * signaled within affine_move_task() itself: when a subsequent affinity request
2674 * occurs after the stopper bailed out due to the targeted task still being
2675 * Migrate-Disable. Consider:
2677 * Initial conditions: P0->cpus_mask = [0, 1]
2681 * migrate_disable();
2683 * set_cpus_allowed_ptr(P0, [1]);
2686 * migration_cpu_stop()
2687 * is_migration_disabled()
2689 * set_cpus_allowed_ptr(P0, [0, 1]);
2690 * <signal completion>
2693 * Note that the above is safe vs a concurrent migrate_enable(), as any
2694 * pending affinity completion is preceded by an uninstallation of
2695 * p->migration_pending done with p->pi_lock held.
2697 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2698 int dest_cpu, unsigned int flags)
2700 struct set_affinity_pending my_pending = { }, *pending = NULL;
2701 bool stop_pending, complete = false;
2703 /* Can the task run on the task's current CPU? If so, we're done */
2704 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2705 struct task_struct *push_task = NULL;
2707 if ((flags & SCA_MIGRATE_ENABLE) &&
2708 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2709 rq->push_busy = true;
2710 push_task = get_task_struct(p);
2714 * If there are pending waiters, but no pending stop_work,
2715 * then complete now.
2717 pending = p->migration_pending;
2718 if (pending && !pending->stop_pending) {
2719 p->migration_pending = NULL;
2723 task_rq_unlock(rq, p, rf);
2726 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2731 complete_all(&pending->done);
2736 if (!(flags & SCA_MIGRATE_ENABLE)) {
2737 /* serialized by p->pi_lock */
2738 if (!p->migration_pending) {
2739 /* Install the request */
2740 refcount_set(&my_pending.refs, 1);
2741 init_completion(&my_pending.done);
2742 my_pending.arg = (struct migration_arg) {
2744 .dest_cpu = dest_cpu,
2745 .pending = &my_pending,
2748 p->migration_pending = &my_pending;
2750 pending = p->migration_pending;
2751 refcount_inc(&pending->refs);
2753 * Affinity has changed, but we've already installed a
2754 * pending. migration_cpu_stop() *must* see this, else
2755 * we risk a completion of the pending despite having a
2756 * task on a disallowed CPU.
2758 * Serialized by p->pi_lock, so this is safe.
2760 pending->arg.dest_cpu = dest_cpu;
2763 pending = p->migration_pending;
2765 * - !MIGRATE_ENABLE:
2766 * we'll have installed a pending if there wasn't one already.
2769 * we're here because the current CPU isn't matching anymore,
2770 * the only way that can happen is because of a concurrent
2771 * set_cpus_allowed_ptr() call, which should then still be
2772 * pending completion.
2774 * Either way, we really should have a @pending here.
2776 if (WARN_ON_ONCE(!pending)) {
2777 task_rq_unlock(rq, p, rf);
2781 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2783 * MIGRATE_ENABLE gets here because 'p == current', but for
2784 * anything else we cannot do is_migration_disabled(), punt
2785 * and have the stopper function handle it all race-free.
2787 stop_pending = pending->stop_pending;
2789 pending->stop_pending = true;
2791 if (flags & SCA_MIGRATE_ENABLE)
2792 p->migration_flags &= ~MDF_PUSH;
2794 task_rq_unlock(rq, p, rf);
2796 if (!stop_pending) {
2797 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2798 &pending->arg, &pending->stop_work);
2801 if (flags & SCA_MIGRATE_ENABLE)
2805 if (!is_migration_disabled(p)) {
2806 if (task_on_rq_queued(p))
2807 rq = move_queued_task(rq, rf, p, dest_cpu);
2809 if (!pending->stop_pending) {
2810 p->migration_pending = NULL;
2814 task_rq_unlock(rq, p, rf);
2817 complete_all(&pending->done);
2820 wait_for_completion(&pending->done);
2822 if (refcount_dec_and_test(&pending->refs))
2823 wake_up_var(&pending->refs); /* No UaF, just an address */
2826 * Block the original owner of &pending until all subsequent callers
2827 * have seen the completion and decremented the refcount
2829 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2832 WARN_ON_ONCE(my_pending.stop_pending);
2838 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2840 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2841 const struct cpumask *new_mask,
2844 struct rq_flags *rf)
2845 __releases(rq->lock)
2846 __releases(p->pi_lock)
2848 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2849 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2850 bool kthread = p->flags & PF_KTHREAD;
2851 struct cpumask *user_mask = NULL;
2852 unsigned int dest_cpu;
2855 update_rq_clock(rq);
2857 if (kthread || is_migration_disabled(p)) {
2859 * Kernel threads are allowed on online && !active CPUs,
2860 * however, during cpu-hot-unplug, even these might get pushed
2861 * away if not KTHREAD_IS_PER_CPU.
2863 * Specifically, migration_disabled() tasks must not fail the
2864 * cpumask_any_and_distribute() pick below, esp. so on
2865 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2866 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2868 cpu_valid_mask = cpu_online_mask;
2871 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2877 * Must re-check here, to close a race against __kthread_bind(),
2878 * sched_setaffinity() is not guaranteed to observe the flag.
2880 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2885 if (!(flags & SCA_MIGRATE_ENABLE)) {
2886 if (cpumask_equal(&p->cpus_mask, new_mask))
2889 if (WARN_ON_ONCE(p == current &&
2890 is_migration_disabled(p) &&
2891 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2898 * Picking a ~random cpu helps in cases where we are changing affinity
2899 * for groups of tasks (ie. cpuset), so that load balancing is not
2900 * immediately required to distribute the tasks within their new mask.
2902 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2903 if (dest_cpu >= nr_cpu_ids) {
2908 __do_set_cpus_allowed(p, new_mask, flags);
2910 if (flags & SCA_USER)
2911 user_mask = clear_user_cpus_ptr(p);
2913 ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2920 task_rq_unlock(rq, p, rf);
2926 * Change a given task's CPU affinity. Migrate the thread to a
2927 * proper CPU and schedule it away if the CPU it's executing on
2928 * is removed from the allowed bitmask.
2930 * NOTE: the caller must have a valid reference to the task, the
2931 * task must not exit() & deallocate itself prematurely. The
2932 * call is not atomic; no spinlocks may be held.
2934 static int __set_cpus_allowed_ptr(struct task_struct *p,
2935 const struct cpumask *new_mask, u32 flags)
2940 rq = task_rq_lock(p, &rf);
2941 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2944 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2946 return __set_cpus_allowed_ptr(p, new_mask, 0);
2948 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2951 * Change a given task's CPU affinity to the intersection of its current
2952 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2953 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2954 * If the resulting mask is empty, leave the affinity unchanged and return
2957 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2958 struct cpumask *new_mask,
2959 const struct cpumask *subset_mask)
2961 struct cpumask *user_mask = NULL;
2966 if (!p->user_cpus_ptr) {
2967 user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2972 rq = task_rq_lock(p, &rf);
2975 * Forcefully restricting the affinity of a deadline task is
2976 * likely to cause problems, so fail and noisily override the
2979 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2984 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2990 * We're about to butcher the task affinity, so keep track of what
2991 * the user asked for in case we're able to restore it later on.
2994 cpumask_copy(user_mask, p->cpus_ptr);
2995 p->user_cpus_ptr = user_mask;
2998 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
3001 task_rq_unlock(rq, p, &rf);
3007 * Restrict the CPU affinity of task @p so that it is a subset of
3008 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
3009 * old affinity mask. If the resulting mask is empty, we warn and walk
3010 * up the cpuset hierarchy until we find a suitable mask.
3012 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3014 cpumask_var_t new_mask;
3015 const struct cpumask *override_mask = task_cpu_possible_mask(p);
3017 alloc_cpumask_var(&new_mask, GFP_KERNEL);
3020 * __migrate_task() can fail silently in the face of concurrent
3021 * offlining of the chosen destination CPU, so take the hotplug
3022 * lock to ensure that the migration succeeds.
3025 if (!cpumask_available(new_mask))
3028 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3032 * We failed to find a valid subset of the affinity mask for the
3033 * task, so override it based on its cpuset hierarchy.
3035 cpuset_cpus_allowed(p, new_mask);
3036 override_mask = new_mask;
3039 if (printk_ratelimit()) {
3040 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3041 task_pid_nr(p), p->comm,
3042 cpumask_pr_args(override_mask));
3045 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3048 free_cpumask_var(new_mask);
3052 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
3055 * Restore the affinity of a task @p which was previously restricted by a
3056 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
3057 * @p->user_cpus_ptr.
3059 * It is the caller's responsibility to serialise this with any calls to
3060 * force_compatible_cpus_allowed_ptr(@p).
3062 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3064 struct cpumask *user_mask = p->user_cpus_ptr;
3065 unsigned long flags;
3068 * Try to restore the old affinity mask. If this fails, then
3069 * we free the mask explicitly to avoid it being inherited across
3070 * a subsequent fork().
3072 if (!user_mask || !__sched_setaffinity(p, user_mask))
3075 raw_spin_lock_irqsave(&p->pi_lock, flags);
3076 user_mask = clear_user_cpus_ptr(p);
3077 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3082 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3084 #ifdef CONFIG_SCHED_DEBUG
3085 unsigned int state = READ_ONCE(p->__state);
3088 * We should never call set_task_cpu() on a blocked task,
3089 * ttwu() will sort out the placement.
3091 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3094 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3095 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3096 * time relying on p->on_rq.
3098 WARN_ON_ONCE(state == TASK_RUNNING &&
3099 p->sched_class == &fair_sched_class &&
3100 (p->on_rq && !task_on_rq_migrating(p)));
3102 #ifdef CONFIG_LOCKDEP
3104 * The caller should hold either p->pi_lock or rq->lock, when changing
3105 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3107 * sched_move_task() holds both and thus holding either pins the cgroup,
3110 * Furthermore, all task_rq users should acquire both locks, see
3113 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3114 lockdep_is_held(__rq_lockp(task_rq(p)))));
3117 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3119 WARN_ON_ONCE(!cpu_online(new_cpu));
3121 WARN_ON_ONCE(is_migration_disabled(p));
3124 trace_sched_migrate_task(p, new_cpu);
3126 if (task_cpu(p) != new_cpu) {
3127 if (p->sched_class->migrate_task_rq)
3128 p->sched_class->migrate_task_rq(p, new_cpu);
3129 p->se.nr_migrations++;
3131 perf_event_task_migrate(p);
3134 __set_task_cpu(p, new_cpu);
3137 #ifdef CONFIG_NUMA_BALANCING
3138 static void __migrate_swap_task(struct task_struct *p, int cpu)
3140 if (task_on_rq_queued(p)) {
3141 struct rq *src_rq, *dst_rq;
3142 struct rq_flags srf, drf;
3144 src_rq = task_rq(p);
3145 dst_rq = cpu_rq(cpu);
3147 rq_pin_lock(src_rq, &srf);
3148 rq_pin_lock(dst_rq, &drf);
3150 deactivate_task(src_rq, p, 0);
3151 set_task_cpu(p, cpu);
3152 activate_task(dst_rq, p, 0);
3153 check_preempt_curr(dst_rq, p, 0);
3155 rq_unpin_lock(dst_rq, &drf);
3156 rq_unpin_lock(src_rq, &srf);
3160 * Task isn't running anymore; make it appear like we migrated
3161 * it before it went to sleep. This means on wakeup we make the
3162 * previous CPU our target instead of where it really is.
3168 struct migration_swap_arg {
3169 struct task_struct *src_task, *dst_task;
3170 int src_cpu, dst_cpu;
3173 static int migrate_swap_stop(void *data)
3175 struct migration_swap_arg *arg = data;
3176 struct rq *src_rq, *dst_rq;
3179 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3182 src_rq = cpu_rq(arg->src_cpu);
3183 dst_rq = cpu_rq(arg->dst_cpu);
3185 double_raw_lock(&arg->src_task->pi_lock,
3186 &arg->dst_task->pi_lock);
3187 double_rq_lock(src_rq, dst_rq);
3189 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3192 if (task_cpu(arg->src_task) != arg->src_cpu)
3195 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3198 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3201 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3202 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3207 double_rq_unlock(src_rq, dst_rq);
3208 raw_spin_unlock(&arg->dst_task->pi_lock);
3209 raw_spin_unlock(&arg->src_task->pi_lock);
3215 * Cross migrate two tasks
3217 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3218 int target_cpu, int curr_cpu)
3220 struct migration_swap_arg arg;
3223 arg = (struct migration_swap_arg){
3225 .src_cpu = curr_cpu,
3227 .dst_cpu = target_cpu,
3230 if (arg.src_cpu == arg.dst_cpu)
3234 * These three tests are all lockless; this is OK since all of them
3235 * will be re-checked with proper locks held further down the line.
3237 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3240 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3243 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3246 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3247 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3252 #endif /* CONFIG_NUMA_BALANCING */
3255 * wait_task_inactive - wait for a thread to unschedule.
3257 * If @match_state is nonzero, it's the @p->state value just checked and
3258 * not expected to change. If it changes, i.e. @p might have woken up,
3259 * then return zero. When we succeed in waiting for @p to be off its CPU,
3260 * we return a positive number (its total switch count). If a second call
3261 * a short while later returns the same number, the caller can be sure that
3262 * @p has remained unscheduled the whole time.
3264 * The caller must ensure that the task *will* unschedule sometime soon,
3265 * else this function might spin for a *long* time. This function can't
3266 * be called with interrupts off, or it may introduce deadlock with
3267 * smp_call_function() if an IPI is sent by the same process we are
3268 * waiting to become inactive.
3270 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3272 int running, queued;
3279 * We do the initial early heuristics without holding
3280 * any task-queue locks at all. We'll only try to get
3281 * the runqueue lock when things look like they will
3287 * If the task is actively running on another CPU
3288 * still, just relax and busy-wait without holding
3291 * NOTE! Since we don't hold any locks, it's not
3292 * even sure that "rq" stays as the right runqueue!
3293 * But we don't care, since "task_running()" will
3294 * return false if the runqueue has changed and p
3295 * is actually now running somewhere else!
3297 while (task_running(rq, p)) {
3298 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3304 * Ok, time to look more closely! We need the rq
3305 * lock now, to be *sure*. If we're wrong, we'll
3306 * just go back and repeat.
3308 rq = task_rq_lock(p, &rf);
3309 trace_sched_wait_task(p);
3310 running = task_running(rq, p);
3311 queued = task_on_rq_queued(p);
3313 if (!match_state || READ_ONCE(p->__state) == match_state)
3314 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3315 task_rq_unlock(rq, p, &rf);
3318 * If it changed from the expected state, bail out now.
3320 if (unlikely(!ncsw))
3324 * Was it really running after all now that we
3325 * checked with the proper locks actually held?
3327 * Oops. Go back and try again..
3329 if (unlikely(running)) {
3335 * It's not enough that it's not actively running,
3336 * it must be off the runqueue _entirely_, and not
3339 * So if it was still runnable (but just not actively
3340 * running right now), it's preempted, and we should
3341 * yield - it could be a while.
3343 if (unlikely(queued)) {
3344 ktime_t to = NSEC_PER_SEC / HZ;
3346 set_current_state(TASK_UNINTERRUPTIBLE);
3347 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3352 * Ahh, all good. It wasn't running, and it wasn't
3353 * runnable, which means that it will never become
3354 * running in the future either. We're all done!
3363 * kick_process - kick a running thread to enter/exit the kernel
3364 * @p: the to-be-kicked thread
3366 * Cause a process which is running on another CPU to enter
3367 * kernel-mode, without any delay. (to get signals handled.)
3369 * NOTE: this function doesn't have to take the runqueue lock,
3370 * because all it wants to ensure is that the remote task enters
3371 * the kernel. If the IPI races and the task has been migrated
3372 * to another CPU then no harm is done and the purpose has been
3375 void kick_process(struct task_struct *p)
3381 if ((cpu != smp_processor_id()) && task_curr(p))
3382 smp_send_reschedule(cpu);
3385 EXPORT_SYMBOL_GPL(kick_process);
3388 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3390 * A few notes on cpu_active vs cpu_online:
3392 * - cpu_active must be a subset of cpu_online
3394 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3395 * see __set_cpus_allowed_ptr(). At this point the newly online
3396 * CPU isn't yet part of the sched domains, and balancing will not
3399 * - on CPU-down we clear cpu_active() to mask the sched domains and
3400 * avoid the load balancer to place new tasks on the to be removed
3401 * CPU. Existing tasks will remain running there and will be taken
3404 * This means that fallback selection must not select !active CPUs.
3405 * And can assume that any active CPU must be online. Conversely
3406 * select_task_rq() below may allow selection of !active CPUs in order
3407 * to satisfy the above rules.
3409 static int select_fallback_rq(int cpu, struct task_struct *p)
3411 int nid = cpu_to_node(cpu);
3412 const struct cpumask *nodemask = NULL;
3413 enum { cpuset, possible, fail } state = cpuset;
3417 * If the node that the CPU is on has been offlined, cpu_to_node()
3418 * will return -1. There is no CPU on the node, and we should
3419 * select the CPU on the other node.
3422 nodemask = cpumask_of_node(nid);
3424 /* Look for allowed, online CPU in same node. */
3425 for_each_cpu(dest_cpu, nodemask) {
3426 if (is_cpu_allowed(p, dest_cpu))
3432 /* Any allowed, online CPU? */
3433 for_each_cpu(dest_cpu, p->cpus_ptr) {
3434 if (!is_cpu_allowed(p, dest_cpu))
3440 /* No more Mr. Nice Guy. */
3443 if (cpuset_cpus_allowed_fallback(p)) {
3450 * XXX When called from select_task_rq() we only
3451 * hold p->pi_lock and again violate locking order.
3453 * More yuck to audit.
3455 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3465 if (state != cpuset) {
3467 * Don't tell them about moving exiting tasks or
3468 * kernel threads (both mm NULL), since they never
3471 if (p->mm && printk_ratelimit()) {
3472 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3473 task_pid_nr(p), p->comm, cpu);
3481 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3484 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3486 lockdep_assert_held(&p->pi_lock);
3488 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3489 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3491 cpu = cpumask_any(p->cpus_ptr);
3494 * In order not to call set_task_cpu() on a blocking task we need
3495 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3498 * Since this is common to all placement strategies, this lives here.
3500 * [ this allows ->select_task() to simply return task_cpu(p) and
3501 * not worry about this generic constraint ]
3503 if (unlikely(!is_cpu_allowed(p, cpu)))
3504 cpu = select_fallback_rq(task_cpu(p), p);
3509 void sched_set_stop_task(int cpu, struct task_struct *stop)
3511 static struct lock_class_key stop_pi_lock;
3512 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3513 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3517 * Make it appear like a SCHED_FIFO task, its something
3518 * userspace knows about and won't get confused about.
3520 * Also, it will make PI more or less work without too
3521 * much confusion -- but then, stop work should not
3522 * rely on PI working anyway.
3524 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3526 stop->sched_class = &stop_sched_class;
3529 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3530 * adjust the effective priority of a task. As a result,
3531 * rt_mutex_setprio() can trigger (RT) balancing operations,
3532 * which can then trigger wakeups of the stop thread to push
3533 * around the current task.
3535 * The stop task itself will never be part of the PI-chain, it
3536 * never blocks, therefore that ->pi_lock recursion is safe.
3537 * Tell lockdep about this by placing the stop->pi_lock in its
3540 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3543 cpu_rq(cpu)->stop = stop;
3547 * Reset it back to a normal scheduling class so that
3548 * it can die in pieces.
3550 old_stop->sched_class = &rt_sched_class;
3554 #else /* CONFIG_SMP */
3556 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3557 const struct cpumask *new_mask,
3560 return set_cpus_allowed_ptr(p, new_mask);
3563 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3565 static inline bool rq_has_pinned_tasks(struct rq *rq)
3570 #endif /* !CONFIG_SMP */
3573 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3577 if (!schedstat_enabled())
3583 if (cpu == rq->cpu) {
3584 __schedstat_inc(rq->ttwu_local);
3585 __schedstat_inc(p->stats.nr_wakeups_local);
3587 struct sched_domain *sd;
3589 __schedstat_inc(p->stats.nr_wakeups_remote);
3591 for_each_domain(rq->cpu, sd) {
3592 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3593 __schedstat_inc(sd->ttwu_wake_remote);
3600 if (wake_flags & WF_MIGRATED)
3601 __schedstat_inc(p->stats.nr_wakeups_migrate);
3602 #endif /* CONFIG_SMP */
3604 __schedstat_inc(rq->ttwu_count);
3605 __schedstat_inc(p->stats.nr_wakeups);
3607 if (wake_flags & WF_SYNC)
3608 __schedstat_inc(p->stats.nr_wakeups_sync);
3612 * Mark the task runnable and perform wakeup-preemption.
3614 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3615 struct rq_flags *rf)
3617 check_preempt_curr(rq, p, wake_flags);
3618 WRITE_ONCE(p->__state, TASK_RUNNING);
3619 trace_sched_wakeup(p);
3622 if (p->sched_class->task_woken) {
3624 * Our task @p is fully woken up and running; so it's safe to
3625 * drop the rq->lock, hereafter rq is only used for statistics.
3627 rq_unpin_lock(rq, rf);
3628 p->sched_class->task_woken(rq, p);
3629 rq_repin_lock(rq, rf);
3632 if (rq->idle_stamp) {
3633 u64 delta = rq_clock(rq) - rq->idle_stamp;
3634 u64 max = 2*rq->max_idle_balance_cost;
3636 update_avg(&rq->avg_idle, delta);
3638 if (rq->avg_idle > max)
3641 rq->wake_stamp = jiffies;
3642 rq->wake_avg_idle = rq->avg_idle / 2;
3650 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3651 struct rq_flags *rf)
3653 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3655 lockdep_assert_rq_held(rq);
3657 if (p->sched_contributes_to_load)
3658 rq->nr_uninterruptible--;
3661 if (wake_flags & WF_MIGRATED)
3662 en_flags |= ENQUEUE_MIGRATED;
3666 delayacct_blkio_end(p);
3667 atomic_dec(&task_rq(p)->nr_iowait);
3670 activate_task(rq, p, en_flags);
3671 ttwu_do_wakeup(rq, p, wake_flags, rf);
3675 * Consider @p being inside a wait loop:
3678 * set_current_state(TASK_UNINTERRUPTIBLE);
3685 * __set_current_state(TASK_RUNNING);
3687 * between set_current_state() and schedule(). In this case @p is still
3688 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3691 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3692 * then schedule() must still happen and p->state can be changed to
3693 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3694 * need to do a full wakeup with enqueue.
3696 * Returns: %true when the wakeup is done,
3699 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3705 rq = __task_rq_lock(p, &rf);
3706 if (task_on_rq_queued(p)) {
3707 /* check_preempt_curr() may use rq clock */
3708 update_rq_clock(rq);
3709 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3712 __task_rq_unlock(rq, &rf);
3718 void sched_ttwu_pending(void *arg)
3720 struct llist_node *llist = arg;
3721 struct rq *rq = this_rq();
3722 struct task_struct *p, *t;
3729 * rq::ttwu_pending racy indication of out-standing wakeups.
3730 * Races such that false-negatives are possible, since they
3731 * are shorter lived that false-positives would be.
3733 WRITE_ONCE(rq->ttwu_pending, 0);
3735 rq_lock_irqsave(rq, &rf);
3736 update_rq_clock(rq);
3738 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3739 if (WARN_ON_ONCE(p->on_cpu))
3740 smp_cond_load_acquire(&p->on_cpu, !VAL);
3742 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3743 set_task_cpu(p, cpu_of(rq));
3745 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3748 rq_unlock_irqrestore(rq, &rf);
3751 void send_call_function_single_ipi(int cpu)
3753 struct rq *rq = cpu_rq(cpu);
3755 if (!set_nr_if_polling(rq->idle))
3756 arch_send_call_function_single_ipi(cpu);
3758 trace_sched_wake_idle_without_ipi(cpu);
3762 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3763 * necessary. The wakee CPU on receipt of the IPI will queue the task
3764 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3765 * of the wakeup instead of the waker.
3767 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3769 struct rq *rq = cpu_rq(cpu);
3771 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3773 WRITE_ONCE(rq->ttwu_pending, 1);
3774 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3777 void wake_up_if_idle(int cpu)
3779 struct rq *rq = cpu_rq(cpu);
3784 if (!is_idle_task(rcu_dereference(rq->curr)))
3787 rq_lock_irqsave(rq, &rf);
3788 if (is_idle_task(rq->curr))
3790 /* Else CPU is not idle, do nothing here: */
3791 rq_unlock_irqrestore(rq, &rf);
3797 bool cpus_share_cache(int this_cpu, int that_cpu)
3799 if (this_cpu == that_cpu)
3802 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3805 static inline bool ttwu_queue_cond(int cpu)
3808 * Do not complicate things with the async wake_list while the CPU is
3811 if (!cpu_active(cpu))
3815 * If the CPU does not share cache, then queue the task on the
3816 * remote rqs wakelist to avoid accessing remote data.
3818 if (!cpus_share_cache(smp_processor_id(), cpu))
3821 if (cpu == smp_processor_id())
3825 * If the wakee cpu is idle, or the task is descheduling and the
3826 * only running task on the CPU, then use the wakelist to offload
3827 * the task activation to the idle (or soon-to-be-idle) CPU as
3828 * the current CPU is likely busy. nr_running is checked to
3829 * avoid unnecessary task stacking.
3831 * Note that we can only get here with (wakee) p->on_rq=0,
3832 * p->on_cpu can be whatever, we've done the dequeue, so
3833 * the wakee has been accounted out of ->nr_running.
3835 if (!cpu_rq(cpu)->nr_running)
3841 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3843 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu)) {
3844 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3845 __ttwu_queue_wakelist(p, cpu, wake_flags);
3852 #else /* !CONFIG_SMP */
3854 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3859 #endif /* CONFIG_SMP */
3861 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3863 struct rq *rq = cpu_rq(cpu);
3866 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3870 update_rq_clock(rq);
3871 ttwu_do_activate(rq, p, wake_flags, &rf);
3876 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3878 * The caller holds p::pi_lock if p != current or has preemption
3879 * disabled when p == current.
3881 * The rules of PREEMPT_RT saved_state:
3883 * The related locking code always holds p::pi_lock when updating
3884 * p::saved_state, which means the code is fully serialized in both cases.
3886 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3887 * bits set. This allows to distinguish all wakeup scenarios.
3889 static __always_inline
3890 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3892 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3893 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3894 state != TASK_RTLOCK_WAIT);
3897 if (READ_ONCE(p->__state) & state) {
3902 #ifdef CONFIG_PREEMPT_RT
3904 * Saved state preserves the task state across blocking on
3905 * an RT lock. If the state matches, set p::saved_state to
3906 * TASK_RUNNING, but do not wake the task because it waits
3907 * for a lock wakeup. Also indicate success because from
3908 * the regular waker's point of view this has succeeded.
3910 * After acquiring the lock the task will restore p::__state
3911 * from p::saved_state which ensures that the regular
3912 * wakeup is not lost. The restore will also set
3913 * p::saved_state to TASK_RUNNING so any further tests will
3914 * not result in false positives vs. @success
3916 if (p->saved_state & state) {
3917 p->saved_state = TASK_RUNNING;
3925 * Notes on Program-Order guarantees on SMP systems.
3929 * The basic program-order guarantee on SMP systems is that when a task [t]
3930 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3931 * execution on its new CPU [c1].
3933 * For migration (of runnable tasks) this is provided by the following means:
3935 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3936 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3937 * rq(c1)->lock (if not at the same time, then in that order).
3938 * C) LOCK of the rq(c1)->lock scheduling in task
3940 * Release/acquire chaining guarantees that B happens after A and C after B.
3941 * Note: the CPU doing B need not be c0 or c1
3950 * UNLOCK rq(0)->lock
3952 * LOCK rq(0)->lock // orders against CPU0
3954 * UNLOCK rq(0)->lock
3958 * UNLOCK rq(1)->lock
3960 * LOCK rq(1)->lock // orders against CPU2
3963 * UNLOCK rq(1)->lock
3966 * BLOCKING -- aka. SLEEP + WAKEUP
3968 * For blocking we (obviously) need to provide the same guarantee as for
3969 * migration. However the means are completely different as there is no lock
3970 * chain to provide order. Instead we do:
3972 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3973 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3977 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3979 * LOCK rq(0)->lock LOCK X->pi_lock
3982 * smp_store_release(X->on_cpu, 0);
3984 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3990 * X->state = RUNNING
3991 * UNLOCK rq(2)->lock
3993 * LOCK rq(2)->lock // orders against CPU1
3996 * UNLOCK rq(2)->lock
3999 * UNLOCK rq(0)->lock
4002 * However, for wakeups there is a second guarantee we must provide, namely we
4003 * must ensure that CONDITION=1 done by the caller can not be reordered with
4004 * accesses to the task state; see try_to_wake_up() and set_current_state().
4008 * try_to_wake_up - wake up a thread
4009 * @p: the thread to be awakened
4010 * @state: the mask of task states that can be woken
4011 * @wake_flags: wake modifier flags (WF_*)
4013 * Conceptually does:
4015 * If (@state & @p->state) @p->state = TASK_RUNNING.
4017 * If the task was not queued/runnable, also place it back on a runqueue.
4019 * This function is atomic against schedule() which would dequeue the task.
4021 * It issues a full memory barrier before accessing @p->state, see the comment
4022 * with set_current_state().
4024 * Uses p->pi_lock to serialize against concurrent wake-ups.
4026 * Relies on p->pi_lock stabilizing:
4029 * - p->sched_task_group
4030 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4032 * Tries really hard to only take one task_rq(p)->lock for performance.
4033 * Takes rq->lock in:
4034 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4035 * - ttwu_queue() -- new rq, for enqueue of the task;
4036 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4038 * As a consequence we race really badly with just about everything. See the
4039 * many memory barriers and their comments for details.
4041 * Return: %true if @p->state changes (an actual wakeup was done),
4045 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4047 unsigned long flags;
4048 int cpu, success = 0;
4053 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4054 * == smp_processor_id()'. Together this means we can special
4055 * case the whole 'p->on_rq && ttwu_runnable()' case below
4056 * without taking any locks.
4059 * - we rely on Program-Order guarantees for all the ordering,
4060 * - we're serialized against set_special_state() by virtue of
4061 * it disabling IRQs (this allows not taking ->pi_lock).
4063 if (!ttwu_state_match(p, state, &success))
4066 trace_sched_waking(p);
4067 WRITE_ONCE(p->__state, TASK_RUNNING);
4068 trace_sched_wakeup(p);
4073 * If we are going to wake up a thread waiting for CONDITION we
4074 * need to ensure that CONDITION=1 done by the caller can not be
4075 * reordered with p->state check below. This pairs with smp_store_mb()
4076 * in set_current_state() that the waiting thread does.
4078 raw_spin_lock_irqsave(&p->pi_lock, flags);
4079 smp_mb__after_spinlock();
4080 if (!ttwu_state_match(p, state, &success))
4083 trace_sched_waking(p);
4086 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4087 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4088 * in smp_cond_load_acquire() below.
4090 * sched_ttwu_pending() try_to_wake_up()
4091 * STORE p->on_rq = 1 LOAD p->state
4094 * __schedule() (switch to task 'p')
4095 * LOCK rq->lock smp_rmb();
4096 * smp_mb__after_spinlock();
4100 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4102 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4103 * __schedule(). See the comment for smp_mb__after_spinlock().
4105 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4108 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4113 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4114 * possible to, falsely, observe p->on_cpu == 0.
4116 * One must be running (->on_cpu == 1) in order to remove oneself
4117 * from the runqueue.
4119 * __schedule() (switch to task 'p') try_to_wake_up()
4120 * STORE p->on_cpu = 1 LOAD p->on_rq
4123 * __schedule() (put 'p' to sleep)
4124 * LOCK rq->lock smp_rmb();
4125 * smp_mb__after_spinlock();
4126 * STORE p->on_rq = 0 LOAD p->on_cpu
4128 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4129 * __schedule(). See the comment for smp_mb__after_spinlock().
4131 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4132 * schedule()'s deactivate_task() has 'happened' and p will no longer
4133 * care about it's own p->state. See the comment in __schedule().
4135 smp_acquire__after_ctrl_dep();
4138 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4139 * == 0), which means we need to do an enqueue, change p->state to
4140 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4141 * enqueue, such as ttwu_queue_wakelist().
4143 WRITE_ONCE(p->__state, TASK_WAKING);
4146 * If the owning (remote) CPU is still in the middle of schedule() with
4147 * this task as prev, considering queueing p on the remote CPUs wake_list
4148 * which potentially sends an IPI instead of spinning on p->on_cpu to
4149 * let the waker make forward progress. This is safe because IRQs are
4150 * disabled and the IPI will deliver after on_cpu is cleared.
4152 * Ensure we load task_cpu(p) after p->on_cpu:
4154 * set_task_cpu(p, cpu);
4155 * STORE p->cpu = @cpu
4156 * __schedule() (switch to task 'p')
4158 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4159 * STORE p->on_cpu = 1 LOAD p->cpu
4161 * to ensure we observe the correct CPU on which the task is currently
4164 if (smp_load_acquire(&p->on_cpu) &&
4165 ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4169 * If the owning (remote) CPU is still in the middle of schedule() with
4170 * this task as prev, wait until it's done referencing the task.
4172 * Pairs with the smp_store_release() in finish_task().
4174 * This ensures that tasks getting woken will be fully ordered against
4175 * their previous state and preserve Program Order.
4177 smp_cond_load_acquire(&p->on_cpu, !VAL);
4179 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4180 if (task_cpu(p) != cpu) {
4182 delayacct_blkio_end(p);
4183 atomic_dec(&task_rq(p)->nr_iowait);
4186 wake_flags |= WF_MIGRATED;
4187 psi_ttwu_dequeue(p);
4188 set_task_cpu(p, cpu);
4192 #endif /* CONFIG_SMP */
4194 ttwu_queue(p, cpu, wake_flags);
4196 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4199 ttwu_stat(p, task_cpu(p), wake_flags);
4206 * task_call_func - Invoke a function on task in fixed state
4207 * @p: Process for which the function is to be invoked, can be @current.
4208 * @func: Function to invoke.
4209 * @arg: Argument to function.
4211 * Fix the task in it's current state by avoiding wakeups and or rq operations
4212 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4213 * to work out what the state is, if required. Given that @func can be invoked
4214 * with a runqueue lock held, it had better be quite lightweight.
4217 * Whatever @func returns
4219 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4221 struct rq *rq = NULL;
4226 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4228 state = READ_ONCE(p->__state);
4231 * Ensure we load p->on_rq after p->__state, otherwise it would be
4232 * possible to, falsely, observe p->on_rq == 0.
4234 * See try_to_wake_up() for a longer comment.
4239 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4240 * the task is blocked. Make sure to check @state since ttwu() can drop
4241 * locks at the end, see ttwu_queue_wakelist().
4243 if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4244 rq = __task_rq_lock(p, &rf);
4247 * At this point the task is pinned; either:
4248 * - blocked and we're holding off wakeups (pi->lock)
4249 * - woken, and we're holding off enqueue (rq->lock)
4250 * - queued, and we're holding off schedule (rq->lock)
4251 * - running, and we're holding off de-schedule (rq->lock)
4253 * The called function (@func) can use: task_curr(), p->on_rq and
4254 * p->__state to differentiate between these states.
4261 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4266 * cpu_curr_snapshot - Return a snapshot of the currently running task
4267 * @cpu: The CPU on which to snapshot the task.
4269 * Returns the task_struct pointer of the task "currently" running on
4270 * the specified CPU. If the same task is running on that CPU throughout,
4271 * the return value will be a pointer to that task's task_struct structure.
4272 * If the CPU did any context switches even vaguely concurrently with the
4273 * execution of this function, the return value will be a pointer to the
4274 * task_struct structure of a randomly chosen task that was running on
4275 * that CPU somewhere around the time that this function was executing.
4277 * If the specified CPU was offline, the return value is whatever it
4278 * is, perhaps a pointer to the task_struct structure of that CPU's idle
4279 * task, but there is no guarantee. Callers wishing a useful return
4280 * value must take some action to ensure that the specified CPU remains
4281 * online throughout.
4283 * This function executes full memory barriers before and after fetching
4284 * the pointer, which permits the caller to confine this function's fetch
4285 * with respect to the caller's accesses to other shared variables.
4287 struct task_struct *cpu_curr_snapshot(int cpu)
4289 struct task_struct *t;
4291 smp_mb(); /* Pairing determined by caller's synchronization design. */
4292 t = rcu_dereference(cpu_curr(cpu));
4293 smp_mb(); /* Pairing determined by caller's synchronization design. */
4298 * wake_up_process - Wake up a specific process
4299 * @p: The process to be woken up.
4301 * Attempt to wake up the nominated process and move it to the set of runnable
4304 * Return: 1 if the process was woken up, 0 if it was already running.
4306 * This function executes a full memory barrier before accessing the task state.
4308 int wake_up_process(struct task_struct *p)
4310 return try_to_wake_up(p, TASK_NORMAL, 0);
4312 EXPORT_SYMBOL(wake_up_process);
4314 int wake_up_state(struct task_struct *p, unsigned int state)
4316 return try_to_wake_up(p, state, 0);
4320 * Perform scheduler related setup for a newly forked process p.
4321 * p is forked by current.
4323 * __sched_fork() is basic setup used by init_idle() too:
4325 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4330 p->se.exec_start = 0;
4331 p->se.sum_exec_runtime = 0;
4332 p->se.prev_sum_exec_runtime = 0;
4333 p->se.nr_migrations = 0;
4335 INIT_LIST_HEAD(&p->se.group_node);
4337 #ifdef CONFIG_FAIR_GROUP_SCHED
4338 p->se.cfs_rq = NULL;
4341 #ifdef CONFIG_SCHEDSTATS
4342 /* Even if schedstat is disabled, there should not be garbage */
4343 memset(&p->stats, 0, sizeof(p->stats));
4346 RB_CLEAR_NODE(&p->dl.rb_node);
4347 init_dl_task_timer(&p->dl);
4348 init_dl_inactive_task_timer(&p->dl);
4349 __dl_clear_params(p);
4351 INIT_LIST_HEAD(&p->rt.run_list);
4353 p->rt.time_slice = sched_rr_timeslice;
4357 #ifdef CONFIG_PREEMPT_NOTIFIERS
4358 INIT_HLIST_HEAD(&p->preempt_notifiers);
4361 #ifdef CONFIG_COMPACTION
4362 p->capture_control = NULL;
4364 init_numa_balancing(clone_flags, p);
4366 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4367 p->migration_pending = NULL;
4371 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4373 #ifdef CONFIG_NUMA_BALANCING
4375 int sysctl_numa_balancing_mode;
4377 static void __set_numabalancing_state(bool enabled)
4380 static_branch_enable(&sched_numa_balancing);
4382 static_branch_disable(&sched_numa_balancing);
4385 void set_numabalancing_state(bool enabled)
4388 sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4390 sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4391 __set_numabalancing_state(enabled);
4394 #ifdef CONFIG_PROC_SYSCTL
4395 int sysctl_numa_balancing(struct ctl_table *table, int write,
4396 void *buffer, size_t *lenp, loff_t *ppos)
4400 int state = sysctl_numa_balancing_mode;
4402 if (write && !capable(CAP_SYS_ADMIN))
4407 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4411 sysctl_numa_balancing_mode = state;
4412 __set_numabalancing_state(state);
4419 #ifdef CONFIG_SCHEDSTATS
4421 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4423 static void set_schedstats(bool enabled)
4426 static_branch_enable(&sched_schedstats);
4428 static_branch_disable(&sched_schedstats);
4431 void force_schedstat_enabled(void)
4433 if (!schedstat_enabled()) {
4434 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4435 static_branch_enable(&sched_schedstats);
4439 static int __init setup_schedstats(char *str)
4445 if (!strcmp(str, "enable")) {
4446 set_schedstats(true);
4448 } else if (!strcmp(str, "disable")) {
4449 set_schedstats(false);
4454 pr_warn("Unable to parse schedstats=\n");
4458 __setup("schedstats=", setup_schedstats);
4460 #ifdef CONFIG_PROC_SYSCTL
4461 static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4462 size_t *lenp, loff_t *ppos)
4466 int state = static_branch_likely(&sched_schedstats);
4468 if (write && !capable(CAP_SYS_ADMIN))
4473 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4477 set_schedstats(state);
4480 #endif /* CONFIG_PROC_SYSCTL */
4481 #endif /* CONFIG_SCHEDSTATS */
4483 #ifdef CONFIG_SYSCTL
4484 static struct ctl_table sched_core_sysctls[] = {
4485 #ifdef CONFIG_SCHEDSTATS
4487 .procname = "sched_schedstats",
4489 .maxlen = sizeof(unsigned int),
4491 .proc_handler = sysctl_schedstats,
4492 .extra1 = SYSCTL_ZERO,
4493 .extra2 = SYSCTL_ONE,
4495 #endif /* CONFIG_SCHEDSTATS */
4496 #ifdef CONFIG_UCLAMP_TASK
4498 .procname = "sched_util_clamp_min",
4499 .data = &sysctl_sched_uclamp_util_min,
4500 .maxlen = sizeof(unsigned int),
4502 .proc_handler = sysctl_sched_uclamp_handler,
4505 .procname = "sched_util_clamp_max",
4506 .data = &sysctl_sched_uclamp_util_max,
4507 .maxlen = sizeof(unsigned int),
4509 .proc_handler = sysctl_sched_uclamp_handler,
4512 .procname = "sched_util_clamp_min_rt_default",
4513 .data = &sysctl_sched_uclamp_util_min_rt_default,
4514 .maxlen = sizeof(unsigned int),
4516 .proc_handler = sysctl_sched_uclamp_handler,
4518 #endif /* CONFIG_UCLAMP_TASK */
4521 static int __init sched_core_sysctl_init(void)
4523 register_sysctl_init("kernel", sched_core_sysctls);
4526 late_initcall(sched_core_sysctl_init);
4527 #endif /* CONFIG_SYSCTL */
4530 * fork()/clone()-time setup:
4532 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4534 __sched_fork(clone_flags, p);
4536 * We mark the process as NEW here. This guarantees that
4537 * nobody will actually run it, and a signal or other external
4538 * event cannot wake it up and insert it on the runqueue either.
4540 p->__state = TASK_NEW;
4543 * Make sure we do not leak PI boosting priority to the child.
4545 p->prio = current->normal_prio;
4550 * Revert to default priority/policy on fork if requested.
4552 if (unlikely(p->sched_reset_on_fork)) {
4553 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4554 p->policy = SCHED_NORMAL;
4555 p->static_prio = NICE_TO_PRIO(0);
4557 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4558 p->static_prio = NICE_TO_PRIO(0);
4560 p->prio = p->normal_prio = p->static_prio;
4561 set_load_weight(p, false);
4564 * We don't need the reset flag anymore after the fork. It has
4565 * fulfilled its duty:
4567 p->sched_reset_on_fork = 0;
4570 if (dl_prio(p->prio))
4572 else if (rt_prio(p->prio))
4573 p->sched_class = &rt_sched_class;
4575 p->sched_class = &fair_sched_class;
4577 init_entity_runnable_average(&p->se);
4580 #ifdef CONFIG_SCHED_INFO
4581 if (likely(sched_info_on()))
4582 memset(&p->sched_info, 0, sizeof(p->sched_info));
4584 #if defined(CONFIG_SMP)
4587 init_task_preempt_count(p);
4589 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4590 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4595 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4597 unsigned long flags;
4600 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4601 * required yet, but lockdep gets upset if rules are violated.
4603 raw_spin_lock_irqsave(&p->pi_lock, flags);
4604 #ifdef CONFIG_CGROUP_SCHED
4606 struct task_group *tg;
4607 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4608 struct task_group, css);
4609 tg = autogroup_task_group(p, tg);
4610 p->sched_task_group = tg;
4615 * We're setting the CPU for the first time, we don't migrate,
4616 * so use __set_task_cpu().
4618 __set_task_cpu(p, smp_processor_id());
4619 if (p->sched_class->task_fork)
4620 p->sched_class->task_fork(p);
4621 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4624 void sched_post_fork(struct task_struct *p)
4626 uclamp_post_fork(p);
4629 unsigned long to_ratio(u64 period, u64 runtime)
4631 if (runtime == RUNTIME_INF)
4635 * Doing this here saves a lot of checks in all
4636 * the calling paths, and returning zero seems
4637 * safe for them anyway.
4642 return div64_u64(runtime << BW_SHIFT, period);
4646 * wake_up_new_task - wake up a newly created task for the first time.
4648 * This function will do some initial scheduler statistics housekeeping
4649 * that must be done for every newly created context, then puts the task
4650 * on the runqueue and wakes it.
4652 void wake_up_new_task(struct task_struct *p)
4657 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4658 WRITE_ONCE(p->__state, TASK_RUNNING);
4661 * Fork balancing, do it here and not earlier because:
4662 * - cpus_ptr can change in the fork path
4663 * - any previously selected CPU might disappear through hotplug
4665 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4666 * as we're not fully set-up yet.
4668 p->recent_used_cpu = task_cpu(p);
4670 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4672 rq = __task_rq_lock(p, &rf);
4673 update_rq_clock(rq);
4674 post_init_entity_util_avg(p);
4676 activate_task(rq, p, ENQUEUE_NOCLOCK);
4677 trace_sched_wakeup_new(p);
4678 check_preempt_curr(rq, p, WF_FORK);
4680 if (p->sched_class->task_woken) {
4682 * Nothing relies on rq->lock after this, so it's fine to
4685 rq_unpin_lock(rq, &rf);
4686 p->sched_class->task_woken(rq, p);
4687 rq_repin_lock(rq, &rf);
4690 task_rq_unlock(rq, p, &rf);
4693 #ifdef CONFIG_PREEMPT_NOTIFIERS
4695 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4697 void preempt_notifier_inc(void)
4699 static_branch_inc(&preempt_notifier_key);
4701 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4703 void preempt_notifier_dec(void)
4705 static_branch_dec(&preempt_notifier_key);
4707 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4710 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4711 * @notifier: notifier struct to register
4713 void preempt_notifier_register(struct preempt_notifier *notifier)
4715 if (!static_branch_unlikely(&preempt_notifier_key))
4716 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4718 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4720 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4723 * preempt_notifier_unregister - no longer interested in preemption notifications
4724 * @notifier: notifier struct to unregister
4726 * This is *not* safe to call from within a preemption notifier.
4728 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4730 hlist_del(¬ifier->link);
4732 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4734 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4736 struct preempt_notifier *notifier;
4738 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4739 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4742 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4744 if (static_branch_unlikely(&preempt_notifier_key))
4745 __fire_sched_in_preempt_notifiers(curr);
4749 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4750 struct task_struct *next)
4752 struct preempt_notifier *notifier;
4754 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4755 notifier->ops->sched_out(notifier, next);
4758 static __always_inline void
4759 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4760 struct task_struct *next)
4762 if (static_branch_unlikely(&preempt_notifier_key))
4763 __fire_sched_out_preempt_notifiers(curr, next);
4766 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4768 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4773 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4774 struct task_struct *next)
4778 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4780 static inline void prepare_task(struct task_struct *next)
4784 * Claim the task as running, we do this before switching to it
4785 * such that any running task will have this set.
4787 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4788 * its ordering comment.
4790 WRITE_ONCE(next->on_cpu, 1);
4794 static inline void finish_task(struct task_struct *prev)
4798 * This must be the very last reference to @prev from this CPU. After
4799 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4800 * must ensure this doesn't happen until the switch is completely
4803 * In particular, the load of prev->state in finish_task_switch() must
4804 * happen before this.
4806 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4808 smp_store_release(&prev->on_cpu, 0);
4814 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4816 void (*func)(struct rq *rq);
4817 struct callback_head *next;
4819 lockdep_assert_rq_held(rq);
4822 func = (void (*)(struct rq *))head->func;
4831 static void balance_push(struct rq *rq);
4834 * balance_push_callback is a right abuse of the callback interface and plays
4835 * by significantly different rules.
4837 * Where the normal balance_callback's purpose is to be ran in the same context
4838 * that queued it (only later, when it's safe to drop rq->lock again),
4839 * balance_push_callback is specifically targeted at __schedule().
4841 * This abuse is tolerated because it places all the unlikely/odd cases behind
4842 * a single test, namely: rq->balance_callback == NULL.
4844 struct callback_head balance_push_callback = {
4846 .func = (void (*)(struct callback_head *))balance_push,
4849 static inline struct callback_head *
4850 __splice_balance_callbacks(struct rq *rq, bool split)
4852 struct callback_head *head = rq->balance_callback;
4857 lockdep_assert_rq_held(rq);
4859 * Must not take balance_push_callback off the list when
4860 * splice_balance_callbacks() and balance_callbacks() are not
4861 * in the same rq->lock section.
4863 * In that case it would be possible for __schedule() to interleave
4864 * and observe the list empty.
4866 if (split && head == &balance_push_callback)
4869 rq->balance_callback = NULL;
4874 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4876 return __splice_balance_callbacks(rq, true);
4879 static void __balance_callbacks(struct rq *rq)
4881 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
4884 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4886 unsigned long flags;
4888 if (unlikely(head)) {
4889 raw_spin_rq_lock_irqsave(rq, flags);
4890 do_balance_callbacks(rq, head);
4891 raw_spin_rq_unlock_irqrestore(rq, flags);
4897 static inline void __balance_callbacks(struct rq *rq)
4901 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4906 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4913 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4916 * Since the runqueue lock will be released by the next
4917 * task (which is an invalid locking op but in the case
4918 * of the scheduler it's an obvious special-case), so we
4919 * do an early lockdep release here:
4921 rq_unpin_lock(rq, rf);
4922 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4923 #ifdef CONFIG_DEBUG_SPINLOCK
4924 /* this is a valid case when another task releases the spinlock */
4925 rq_lockp(rq)->owner = next;
4929 static inline void finish_lock_switch(struct rq *rq)
4932 * If we are tracking spinlock dependencies then we have to
4933 * fix up the runqueue lock - which gets 'carried over' from
4934 * prev into current:
4936 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4937 __balance_callbacks(rq);
4938 raw_spin_rq_unlock_irq(rq);
4942 * NOP if the arch has not defined these:
4945 #ifndef prepare_arch_switch
4946 # define prepare_arch_switch(next) do { } while (0)
4949 #ifndef finish_arch_post_lock_switch
4950 # define finish_arch_post_lock_switch() do { } while (0)
4953 static inline void kmap_local_sched_out(void)
4955 #ifdef CONFIG_KMAP_LOCAL
4956 if (unlikely(current->kmap_ctrl.idx))
4957 __kmap_local_sched_out();
4961 static inline void kmap_local_sched_in(void)
4963 #ifdef CONFIG_KMAP_LOCAL
4964 if (unlikely(current->kmap_ctrl.idx))
4965 __kmap_local_sched_in();
4970 * prepare_task_switch - prepare to switch tasks
4971 * @rq: the runqueue preparing to switch
4972 * @prev: the current task that is being switched out
4973 * @next: the task we are going to switch to.
4975 * This is called with the rq lock held and interrupts off. It must
4976 * be paired with a subsequent finish_task_switch after the context
4979 * prepare_task_switch sets up locking and calls architecture specific
4983 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4984 struct task_struct *next)
4986 kcov_prepare_switch(prev);
4987 sched_info_switch(rq, prev, next);
4988 perf_event_task_sched_out(prev, next);
4990 fire_sched_out_preempt_notifiers(prev, next);
4991 kmap_local_sched_out();
4993 prepare_arch_switch(next);
4997 * finish_task_switch - clean up after a task-switch
4998 * @prev: the thread we just switched away from.
5000 * finish_task_switch must be called after the context switch, paired
5001 * with a prepare_task_switch call before the context switch.
5002 * finish_task_switch will reconcile locking set up by prepare_task_switch,
5003 * and do any other architecture-specific cleanup actions.
5005 * Note that we may have delayed dropping an mm in context_switch(). If
5006 * so, we finish that here outside of the runqueue lock. (Doing it
5007 * with the lock held can cause deadlocks; see schedule() for
5010 * The context switch have flipped the stack from under us and restored the
5011 * local variables which were saved when this task called schedule() in the
5012 * past. prev == current is still correct but we need to recalculate this_rq
5013 * because prev may have moved to another CPU.
5015 static struct rq *finish_task_switch(struct task_struct *prev)
5016 __releases(rq->lock)
5018 struct rq *rq = this_rq();
5019 struct mm_struct *mm = rq->prev_mm;
5020 unsigned int prev_state;
5023 * The previous task will have left us with a preempt_count of 2
5024 * because it left us after:
5027 * preempt_disable(); // 1
5029 * raw_spin_lock_irq(&rq->lock) // 2
5031 * Also, see FORK_PREEMPT_COUNT.
5033 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5034 "corrupted preempt_count: %s/%d/0x%x\n",
5035 current->comm, current->pid, preempt_count()))
5036 preempt_count_set(FORK_PREEMPT_COUNT);
5041 * A task struct has one reference for the use as "current".
5042 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5043 * schedule one last time. The schedule call will never return, and
5044 * the scheduled task must drop that reference.
5046 * We must observe prev->state before clearing prev->on_cpu (in
5047 * finish_task), otherwise a concurrent wakeup can get prev
5048 * running on another CPU and we could rave with its RUNNING -> DEAD
5049 * transition, resulting in a double drop.
5051 prev_state = READ_ONCE(prev->__state);
5052 vtime_task_switch(prev);
5053 perf_event_task_sched_in(prev, current);
5055 tick_nohz_task_switch();
5056 finish_lock_switch(rq);
5057 finish_arch_post_lock_switch();
5058 kcov_finish_switch(current);
5060 * kmap_local_sched_out() is invoked with rq::lock held and
5061 * interrupts disabled. There is no requirement for that, but the
5062 * sched out code does not have an interrupt enabled section.
5063 * Restoring the maps on sched in does not require interrupts being
5066 kmap_local_sched_in();
5068 fire_sched_in_preempt_notifiers(current);
5070 * When switching through a kernel thread, the loop in
5071 * membarrier_{private,global}_expedited() may have observed that
5072 * kernel thread and not issued an IPI. It is therefore possible to
5073 * schedule between user->kernel->user threads without passing though
5074 * switch_mm(). Membarrier requires a barrier after storing to
5075 * rq->curr, before returning to userspace, so provide them here:
5077 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5078 * provided by mmdrop(),
5079 * - a sync_core for SYNC_CORE.
5082 membarrier_mm_sync_core_before_usermode(mm);
5085 if (unlikely(prev_state == TASK_DEAD)) {
5086 if (prev->sched_class->task_dead)
5087 prev->sched_class->task_dead(prev);
5089 /* Task is done with its stack. */
5090 put_task_stack(prev);
5092 put_task_struct_rcu_user(prev);
5099 * schedule_tail - first thing a freshly forked thread must call.
5100 * @prev: the thread we just switched away from.
5102 asmlinkage __visible void schedule_tail(struct task_struct *prev)
5103 __releases(rq->lock)
5106 * New tasks start with FORK_PREEMPT_COUNT, see there and
5107 * finish_task_switch() for details.
5109 * finish_task_switch() will drop rq->lock() and lower preempt_count
5110 * and the preempt_enable() will end up enabling preemption (on
5111 * PREEMPT_COUNT kernels).
5114 finish_task_switch(prev);
5117 if (current->set_child_tid)
5118 put_user(task_pid_vnr(current), current->set_child_tid);
5120 calculate_sigpending();
5124 * context_switch - switch to the new MM and the new thread's register state.
5126 static __always_inline struct rq *
5127 context_switch(struct rq *rq, struct task_struct *prev,
5128 struct task_struct *next, struct rq_flags *rf)
5130 prepare_task_switch(rq, prev, next);
5133 * For paravirt, this is coupled with an exit in switch_to to
5134 * combine the page table reload and the switch backend into
5137 arch_start_context_switch(prev);
5140 * kernel -> kernel lazy + transfer active
5141 * user -> kernel lazy + mmgrab() active
5143 * kernel -> user switch + mmdrop() active
5144 * user -> user switch
5146 if (!next->mm) { // to kernel
5147 enter_lazy_tlb(prev->active_mm, next);
5149 next->active_mm = prev->active_mm;
5150 if (prev->mm) // from user
5151 mmgrab(prev->active_mm);
5153 prev->active_mm = NULL;
5155 membarrier_switch_mm(rq, prev->active_mm, next->mm);
5157 * sys_membarrier() requires an smp_mb() between setting
5158 * rq->curr / membarrier_switch_mm() and returning to userspace.
5160 * The below provides this either through switch_mm(), or in
5161 * case 'prev->active_mm == next->mm' through
5162 * finish_task_switch()'s mmdrop().
5164 switch_mm_irqs_off(prev->active_mm, next->mm, next);
5166 if (!prev->mm) { // from kernel
5167 /* will mmdrop() in finish_task_switch(). */
5168 rq->prev_mm = prev->active_mm;
5169 prev->active_mm = NULL;
5173 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5175 prepare_lock_switch(rq, next, rf);
5177 /* Here we just switch the register state and the stack. */
5178 switch_to(prev, next, prev);
5181 return finish_task_switch(prev);
5185 * nr_running and nr_context_switches:
5187 * externally visible scheduler statistics: current number of runnable
5188 * threads, total number of context switches performed since bootup.
5190 unsigned int nr_running(void)
5192 unsigned int i, sum = 0;
5194 for_each_online_cpu(i)
5195 sum += cpu_rq(i)->nr_running;
5201 * Check if only the current task is running on the CPU.
5203 * Caution: this function does not check that the caller has disabled
5204 * preemption, thus the result might have a time-of-check-to-time-of-use
5205 * race. The caller is responsible to use it correctly, for example:
5207 * - from a non-preemptible section (of course)
5209 * - from a thread that is bound to a single CPU
5211 * - in a loop with very short iterations (e.g. a polling loop)
5213 bool single_task_running(void)
5215 return raw_rq()->nr_running == 1;
5217 EXPORT_SYMBOL(single_task_running);
5219 unsigned long long nr_context_switches(void)
5222 unsigned long long sum = 0;
5224 for_each_possible_cpu(i)
5225 sum += cpu_rq(i)->nr_switches;
5231 * Consumers of these two interfaces, like for example the cpuidle menu
5232 * governor, are using nonsensical data. Preferring shallow idle state selection
5233 * for a CPU that has IO-wait which might not even end up running the task when
5234 * it does become runnable.
5237 unsigned int nr_iowait_cpu(int cpu)
5239 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5243 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5245 * The idea behind IO-wait account is to account the idle time that we could
5246 * have spend running if it were not for IO. That is, if we were to improve the
5247 * storage performance, we'd have a proportional reduction in IO-wait time.
5249 * This all works nicely on UP, where, when a task blocks on IO, we account
5250 * idle time as IO-wait, because if the storage were faster, it could've been
5251 * running and we'd not be idle.
5253 * This has been extended to SMP, by doing the same for each CPU. This however
5256 * Imagine for instance the case where two tasks block on one CPU, only the one
5257 * CPU will have IO-wait accounted, while the other has regular idle. Even
5258 * though, if the storage were faster, both could've ran at the same time,
5259 * utilising both CPUs.
5261 * This means, that when looking globally, the current IO-wait accounting on
5262 * SMP is a lower bound, by reason of under accounting.
5264 * Worse, since the numbers are provided per CPU, they are sometimes
5265 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5266 * associated with any one particular CPU, it can wake to another CPU than it
5267 * blocked on. This means the per CPU IO-wait number is meaningless.
5269 * Task CPU affinities can make all that even more 'interesting'.
5272 unsigned int nr_iowait(void)
5274 unsigned int i, sum = 0;
5276 for_each_possible_cpu(i)
5277 sum += nr_iowait_cpu(i);
5285 * sched_exec - execve() is a valuable balancing opportunity, because at
5286 * this point the task has the smallest effective memory and cache footprint.
5288 void sched_exec(void)
5290 struct task_struct *p = current;
5291 unsigned long flags;
5294 raw_spin_lock_irqsave(&p->pi_lock, flags);
5295 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5296 if (dest_cpu == smp_processor_id())
5299 if (likely(cpu_active(dest_cpu))) {
5300 struct migration_arg arg = { p, dest_cpu };
5302 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5303 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5307 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5312 DEFINE_PER_CPU(struct kernel_stat, kstat);
5313 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5315 EXPORT_PER_CPU_SYMBOL(kstat);
5316 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5319 * The function fair_sched_class.update_curr accesses the struct curr
5320 * and its field curr->exec_start; when called from task_sched_runtime(),
5321 * we observe a high rate of cache misses in practice.
5322 * Prefetching this data results in improved performance.
5324 static inline void prefetch_curr_exec_start(struct task_struct *p)
5326 #ifdef CONFIG_FAIR_GROUP_SCHED
5327 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5329 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5332 prefetch(&curr->exec_start);
5336 * Return accounted runtime for the task.
5337 * In case the task is currently running, return the runtime plus current's
5338 * pending runtime that have not been accounted yet.
5340 unsigned long long task_sched_runtime(struct task_struct *p)
5346 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5348 * 64-bit doesn't need locks to atomically read a 64-bit value.
5349 * So we have a optimization chance when the task's delta_exec is 0.
5350 * Reading ->on_cpu is racy, but this is ok.
5352 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5353 * If we race with it entering CPU, unaccounted time is 0. This is
5354 * indistinguishable from the read occurring a few cycles earlier.
5355 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5356 * been accounted, so we're correct here as well.
5358 if (!p->on_cpu || !task_on_rq_queued(p))
5359 return p->se.sum_exec_runtime;
5362 rq = task_rq_lock(p, &rf);
5364 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5365 * project cycles that may never be accounted to this
5366 * thread, breaking clock_gettime().
5368 if (task_current(rq, p) && task_on_rq_queued(p)) {
5369 prefetch_curr_exec_start(p);
5370 update_rq_clock(rq);
5371 p->sched_class->update_curr(rq);
5373 ns = p->se.sum_exec_runtime;
5374 task_rq_unlock(rq, p, &rf);
5379 #ifdef CONFIG_SCHED_DEBUG
5380 static u64 cpu_resched_latency(struct rq *rq)
5382 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5383 u64 resched_latency, now = rq_clock(rq);
5384 static bool warned_once;
5386 if (sysctl_resched_latency_warn_once && warned_once)
5389 if (!need_resched() || !latency_warn_ms)
5392 if (system_state == SYSTEM_BOOTING)
5395 if (!rq->last_seen_need_resched_ns) {
5396 rq->last_seen_need_resched_ns = now;
5397 rq->ticks_without_resched = 0;
5401 rq->ticks_without_resched++;
5402 resched_latency = now - rq->last_seen_need_resched_ns;
5403 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5408 return resched_latency;
5411 static int __init setup_resched_latency_warn_ms(char *str)
5415 if ((kstrtol(str, 0, &val))) {
5416 pr_warn("Unable to set resched_latency_warn_ms\n");
5420 sysctl_resched_latency_warn_ms = val;
5423 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5425 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5426 #endif /* CONFIG_SCHED_DEBUG */
5429 * This function gets called by the timer code, with HZ frequency.
5430 * We call it with interrupts disabled.
5432 void scheduler_tick(void)
5434 int cpu = smp_processor_id();
5435 struct rq *rq = cpu_rq(cpu);
5436 struct task_struct *curr = rq->curr;
5438 unsigned long thermal_pressure;
5439 u64 resched_latency;
5441 arch_scale_freq_tick();
5446 update_rq_clock(rq);
5447 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5448 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5449 curr->sched_class->task_tick(rq, curr, 0);
5450 if (sched_feat(LATENCY_WARN))
5451 resched_latency = cpu_resched_latency(rq);
5452 calc_global_load_tick(rq);
5453 sched_core_tick(rq);
5457 if (sched_feat(LATENCY_WARN) && resched_latency)
5458 resched_latency_warn(cpu, resched_latency);
5460 perf_event_task_tick();
5463 rq->idle_balance = idle_cpu(cpu);
5464 trigger_load_balance(rq);
5468 #ifdef CONFIG_NO_HZ_FULL
5473 struct delayed_work work;
5475 /* Values for ->state, see diagram below. */
5476 #define TICK_SCHED_REMOTE_OFFLINE 0
5477 #define TICK_SCHED_REMOTE_OFFLINING 1
5478 #define TICK_SCHED_REMOTE_RUNNING 2
5481 * State diagram for ->state:
5484 * TICK_SCHED_REMOTE_OFFLINE
5487 * | | sched_tick_remote()
5490 * +--TICK_SCHED_REMOTE_OFFLINING
5493 * sched_tick_start() | | sched_tick_stop()
5496 * TICK_SCHED_REMOTE_RUNNING
5499 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5500 * and sched_tick_start() are happy to leave the state in RUNNING.
5503 static struct tick_work __percpu *tick_work_cpu;
5505 static void sched_tick_remote(struct work_struct *work)
5507 struct delayed_work *dwork = to_delayed_work(work);
5508 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5509 int cpu = twork->cpu;
5510 struct rq *rq = cpu_rq(cpu);
5511 struct task_struct *curr;
5517 * Handle the tick only if it appears the remote CPU is running in full
5518 * dynticks mode. The check is racy by nature, but missing a tick or
5519 * having one too much is no big deal because the scheduler tick updates
5520 * statistics and checks timeslices in a time-independent way, regardless
5521 * of when exactly it is running.
5523 if (!tick_nohz_tick_stopped_cpu(cpu))
5526 rq_lock_irq(rq, &rf);
5528 if (cpu_is_offline(cpu))
5531 update_rq_clock(rq);
5533 if (!is_idle_task(curr)) {
5535 * Make sure the next tick runs within a reasonable
5538 delta = rq_clock_task(rq) - curr->se.exec_start;
5539 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5541 curr->sched_class->task_tick(rq, curr, 0);
5543 calc_load_nohz_remote(rq);
5545 rq_unlock_irq(rq, &rf);
5549 * Run the remote tick once per second (1Hz). This arbitrary
5550 * frequency is large enough to avoid overload but short enough
5551 * to keep scheduler internal stats reasonably up to date. But
5552 * first update state to reflect hotplug activity if required.
5554 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5555 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5556 if (os == TICK_SCHED_REMOTE_RUNNING)
5557 queue_delayed_work(system_unbound_wq, dwork, HZ);
5560 static void sched_tick_start(int cpu)
5563 struct tick_work *twork;
5565 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5568 WARN_ON_ONCE(!tick_work_cpu);
5570 twork = per_cpu_ptr(tick_work_cpu, cpu);
5571 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5572 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5573 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5575 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5576 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5580 #ifdef CONFIG_HOTPLUG_CPU
5581 static void sched_tick_stop(int cpu)
5583 struct tick_work *twork;
5586 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5589 WARN_ON_ONCE(!tick_work_cpu);
5591 twork = per_cpu_ptr(tick_work_cpu, cpu);
5592 /* There cannot be competing actions, but don't rely on stop-machine. */
5593 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5594 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5595 /* Don't cancel, as this would mess up the state machine. */
5597 #endif /* CONFIG_HOTPLUG_CPU */
5599 int __init sched_tick_offload_init(void)
5601 tick_work_cpu = alloc_percpu(struct tick_work);
5602 BUG_ON(!tick_work_cpu);
5606 #else /* !CONFIG_NO_HZ_FULL */
5607 static inline void sched_tick_start(int cpu) { }
5608 static inline void sched_tick_stop(int cpu) { }
5611 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5612 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5614 * If the value passed in is equal to the current preempt count
5615 * then we just disabled preemption. Start timing the latency.
5617 static inline void preempt_latency_start(int val)
5619 if (preempt_count() == val) {
5620 unsigned long ip = get_lock_parent_ip();
5621 #ifdef CONFIG_DEBUG_PREEMPT
5622 current->preempt_disable_ip = ip;
5624 trace_preempt_off(CALLER_ADDR0, ip);
5628 void preempt_count_add(int val)
5630 #ifdef CONFIG_DEBUG_PREEMPT
5634 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5637 __preempt_count_add(val);
5638 #ifdef CONFIG_DEBUG_PREEMPT
5640 * Spinlock count overflowing soon?
5642 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5645 preempt_latency_start(val);
5647 EXPORT_SYMBOL(preempt_count_add);
5648 NOKPROBE_SYMBOL(preempt_count_add);
5651 * If the value passed in equals to the current preempt count
5652 * then we just enabled preemption. Stop timing the latency.
5654 static inline void preempt_latency_stop(int val)
5656 if (preempt_count() == val)
5657 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5660 void preempt_count_sub(int val)
5662 #ifdef CONFIG_DEBUG_PREEMPT
5666 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5669 * Is the spinlock portion underflowing?
5671 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5672 !(preempt_count() & PREEMPT_MASK)))
5676 preempt_latency_stop(val);
5677 __preempt_count_sub(val);
5679 EXPORT_SYMBOL(preempt_count_sub);
5680 NOKPROBE_SYMBOL(preempt_count_sub);
5683 static inline void preempt_latency_start(int val) { }
5684 static inline void preempt_latency_stop(int val) { }
5687 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5689 #ifdef CONFIG_DEBUG_PREEMPT
5690 return p->preempt_disable_ip;
5697 * Print scheduling while atomic bug:
5699 static noinline void __schedule_bug(struct task_struct *prev)
5701 /* Save this before calling printk(), since that will clobber it */
5702 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5704 if (oops_in_progress)
5707 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5708 prev->comm, prev->pid, preempt_count());
5710 debug_show_held_locks(prev);
5712 if (irqs_disabled())
5713 print_irqtrace_events(prev);
5714 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5715 && in_atomic_preempt_off()) {
5716 pr_err("Preemption disabled at:");
5717 print_ip_sym(KERN_ERR, preempt_disable_ip);
5720 panic("scheduling while atomic\n");
5723 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5727 * Various schedule()-time debugging checks and statistics:
5729 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5731 #ifdef CONFIG_SCHED_STACK_END_CHECK
5732 if (task_stack_end_corrupted(prev))
5733 panic("corrupted stack end detected inside scheduler\n");
5735 if (task_scs_end_corrupted(prev))
5736 panic("corrupted shadow stack detected inside scheduler\n");
5739 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5740 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5741 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5742 prev->comm, prev->pid, prev->non_block_count);
5744 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5748 if (unlikely(in_atomic_preempt_off())) {
5749 __schedule_bug(prev);
5750 preempt_count_set(PREEMPT_DISABLED);
5753 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5755 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5757 schedstat_inc(this_rq()->sched_count);
5760 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5761 struct rq_flags *rf)
5764 const struct sched_class *class;
5766 * We must do the balancing pass before put_prev_task(), such
5767 * that when we release the rq->lock the task is in the same
5768 * state as before we took rq->lock.
5770 * We can terminate the balance pass as soon as we know there is
5771 * a runnable task of @class priority or higher.
5773 for_class_range(class, prev->sched_class, &idle_sched_class) {
5774 if (class->balance(rq, prev, rf))
5779 put_prev_task(rq, prev);
5783 * Pick up the highest-prio task:
5785 static inline struct task_struct *
5786 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5788 const struct sched_class *class;
5789 struct task_struct *p;
5792 * Optimization: we know that if all tasks are in the fair class we can
5793 * call that function directly, but only if the @prev task wasn't of a
5794 * higher scheduling class, because otherwise those lose the
5795 * opportunity to pull in more work from other CPUs.
5797 if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
5798 rq->nr_running == rq->cfs.h_nr_running)) {
5800 p = pick_next_task_fair(rq, prev, rf);
5801 if (unlikely(p == RETRY_TASK))
5804 /* Assume the next prioritized class is idle_sched_class */
5806 put_prev_task(rq, prev);
5807 p = pick_next_task_idle(rq);
5814 put_prev_task_balance(rq, prev, rf);
5816 for_each_class(class) {
5817 p = class->pick_next_task(rq);
5822 BUG(); /* The idle class should always have a runnable task. */
5825 #ifdef CONFIG_SCHED_CORE
5826 static inline bool is_task_rq_idle(struct task_struct *t)
5828 return (task_rq(t)->idle == t);
5831 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5833 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5836 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5838 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5841 return a->core_cookie == b->core_cookie;
5844 static inline struct task_struct *pick_task(struct rq *rq)
5846 const struct sched_class *class;
5847 struct task_struct *p;
5849 for_each_class(class) {
5850 p = class->pick_task(rq);
5855 BUG(); /* The idle class should always have a runnable task. */
5858 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5860 static void queue_core_balance(struct rq *rq);
5862 static struct task_struct *
5863 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5865 struct task_struct *next, *p, *max = NULL;
5866 const struct cpumask *smt_mask;
5867 bool fi_before = false;
5868 bool core_clock_updated = (rq == rq->core);
5869 unsigned long cookie;
5870 int i, cpu, occ = 0;
5874 if (!sched_core_enabled(rq))
5875 return __pick_next_task(rq, prev, rf);
5879 /* Stopper task is switching into idle, no need core-wide selection. */
5880 if (cpu_is_offline(cpu)) {
5882 * Reset core_pick so that we don't enter the fastpath when
5883 * coming online. core_pick would already be migrated to
5884 * another cpu during offline.
5886 rq->core_pick = NULL;
5887 return __pick_next_task(rq, prev, rf);
5891 * If there were no {en,de}queues since we picked (IOW, the task
5892 * pointers are all still valid), and we haven't scheduled the last
5893 * pick yet, do so now.
5895 * rq->core_pick can be NULL if no selection was made for a CPU because
5896 * it was either offline or went offline during a sibling's core-wide
5897 * selection. In this case, do a core-wide selection.
5899 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5900 rq->core->core_pick_seq != rq->core_sched_seq &&
5902 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5904 next = rq->core_pick;
5906 put_prev_task(rq, prev);
5907 set_next_task(rq, next);
5910 rq->core_pick = NULL;
5914 put_prev_task_balance(rq, prev, rf);
5916 smt_mask = cpu_smt_mask(cpu);
5917 need_sync = !!rq->core->core_cookie;
5920 rq->core->core_cookie = 0UL;
5921 if (rq->core->core_forceidle_count) {
5922 if (!core_clock_updated) {
5923 update_rq_clock(rq->core);
5924 core_clock_updated = true;
5926 sched_core_account_forceidle(rq);
5927 /* reset after accounting force idle */
5928 rq->core->core_forceidle_start = 0;
5929 rq->core->core_forceidle_count = 0;
5930 rq->core->core_forceidle_occupation = 0;
5936 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5938 * @task_seq guards the task state ({en,de}queues)
5939 * @pick_seq is the @task_seq we did a selection on
5940 * @sched_seq is the @pick_seq we scheduled
5942 * However, preemptions can cause multiple picks on the same task set.
5943 * 'Fix' this by also increasing @task_seq for every pick.
5945 rq->core->core_task_seq++;
5948 * Optimize for common case where this CPU has no cookies
5949 * and there are no cookied tasks running on siblings.
5952 next = pick_task(rq);
5953 if (!next->core_cookie) {
5954 rq->core_pick = NULL;
5956 * For robustness, update the min_vruntime_fi for
5957 * unconstrained picks as well.
5959 WARN_ON_ONCE(fi_before);
5960 task_vruntime_update(rq, next, false);
5966 * For each thread: do the regular task pick and find the max prio task
5969 * Tie-break prio towards the current CPU
5971 for_each_cpu_wrap(i, smt_mask, cpu) {
5975 * Current cpu always has its clock updated on entrance to
5976 * pick_next_task(). If the current cpu is not the core,
5977 * the core may also have been updated above.
5979 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
5980 update_rq_clock(rq_i);
5982 p = rq_i->core_pick = pick_task(rq_i);
5983 if (!max || prio_less(max, p, fi_before))
5987 cookie = rq->core->core_cookie = max->core_cookie;
5990 * For each thread: try and find a runnable task that matches @max or
5993 for_each_cpu(i, smt_mask) {
5995 p = rq_i->core_pick;
5997 if (!cookie_equals(p, cookie)) {
6000 p = sched_core_find(rq_i, cookie);
6002 p = idle_sched_class.pick_task(rq_i);
6005 rq_i->core_pick = p;
6007 if (p == rq_i->idle) {
6008 if (rq_i->nr_running) {
6009 rq->core->core_forceidle_count++;
6011 rq->core->core_forceidle_seq++;
6018 if (schedstat_enabled() && rq->core->core_forceidle_count) {
6019 rq->core->core_forceidle_start = rq_clock(rq->core);
6020 rq->core->core_forceidle_occupation = occ;
6023 rq->core->core_pick_seq = rq->core->core_task_seq;
6024 next = rq->core_pick;
6025 rq->core_sched_seq = rq->core->core_pick_seq;
6027 /* Something should have been selected for current CPU */
6028 WARN_ON_ONCE(!next);
6031 * Reschedule siblings
6033 * NOTE: L1TF -- at this point we're no longer running the old task and
6034 * sending an IPI (below) ensures the sibling will no longer be running
6035 * their task. This ensures there is no inter-sibling overlap between
6036 * non-matching user state.
6038 for_each_cpu(i, smt_mask) {
6042 * An online sibling might have gone offline before a task
6043 * could be picked for it, or it might be offline but later
6044 * happen to come online, but its too late and nothing was
6045 * picked for it. That's Ok - it will pick tasks for itself,
6048 if (!rq_i->core_pick)
6052 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6053 * fi_before fi update?
6059 if (!(fi_before && rq->core->core_forceidle_count))
6060 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6062 rq_i->core_pick->core_occupation = occ;
6065 rq_i->core_pick = NULL;
6069 /* Did we break L1TF mitigation requirements? */
6070 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6072 if (rq_i->curr == rq_i->core_pick) {
6073 rq_i->core_pick = NULL;
6081 set_next_task(rq, next);
6083 if (rq->core->core_forceidle_count && next == rq->idle)
6084 queue_core_balance(rq);
6089 static bool try_steal_cookie(int this, int that)
6091 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6092 struct task_struct *p;
6093 unsigned long cookie;
6094 bool success = false;
6096 local_irq_disable();
6097 double_rq_lock(dst, src);
6099 cookie = dst->core->core_cookie;
6103 if (dst->curr != dst->idle)
6106 p = sched_core_find(src, cookie);
6111 if (p == src->core_pick || p == src->curr)
6114 if (!is_cpu_allowed(p, this))
6117 if (p->core_occupation > dst->idle->core_occupation)
6120 deactivate_task(src, p, 0);
6121 set_task_cpu(p, this);
6122 activate_task(dst, p, 0);
6130 p = sched_core_next(p, cookie);
6134 double_rq_unlock(dst, src);
6140 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6144 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
6151 if (try_steal_cookie(cpu, i))
6158 static void sched_core_balance(struct rq *rq)
6160 struct sched_domain *sd;
6161 int cpu = cpu_of(rq);
6165 raw_spin_rq_unlock_irq(rq);
6166 for_each_domain(cpu, sd) {
6170 if (steal_cookie_task(cpu, sd))
6173 raw_spin_rq_lock_irq(rq);
6178 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
6180 static void queue_core_balance(struct rq *rq)
6182 if (!sched_core_enabled(rq))
6185 if (!rq->core->core_cookie)
6188 if (!rq->nr_running) /* not forced idle */
6191 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6194 static void sched_core_cpu_starting(unsigned int cpu)
6196 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6197 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6198 unsigned long flags;
6201 sched_core_lock(cpu, &flags);
6203 WARN_ON_ONCE(rq->core != rq);
6205 /* if we're the first, we'll be our own leader */
6206 if (cpumask_weight(smt_mask) == 1)
6209 /* find the leader */
6210 for_each_cpu(t, smt_mask) {
6214 if (rq->core == rq) {
6220 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6223 /* install and validate core_rq */
6224 for_each_cpu(t, smt_mask) {
6230 WARN_ON_ONCE(rq->core != core_rq);
6234 sched_core_unlock(cpu, &flags);
6237 static void sched_core_cpu_deactivate(unsigned int cpu)
6239 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6240 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6241 unsigned long flags;
6244 sched_core_lock(cpu, &flags);
6246 /* if we're the last man standing, nothing to do */
6247 if (cpumask_weight(smt_mask) == 1) {
6248 WARN_ON_ONCE(rq->core != rq);
6252 /* if we're not the leader, nothing to do */
6256 /* find a new leader */
6257 for_each_cpu(t, smt_mask) {
6260 core_rq = cpu_rq(t);
6264 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6267 /* copy the shared state to the new leader */
6268 core_rq->core_task_seq = rq->core_task_seq;
6269 core_rq->core_pick_seq = rq->core_pick_seq;
6270 core_rq->core_cookie = rq->core_cookie;
6271 core_rq->core_forceidle_count = rq->core_forceidle_count;
6272 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6273 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6276 * Accounting edge for forced idle is handled in pick_next_task().
6277 * Don't need another one here, since the hotplug thread shouldn't
6280 core_rq->core_forceidle_start = 0;
6282 /* install new leader */
6283 for_each_cpu(t, smt_mask) {
6289 sched_core_unlock(cpu, &flags);
6292 static inline void sched_core_cpu_dying(unsigned int cpu)
6294 struct rq *rq = cpu_rq(cpu);
6300 #else /* !CONFIG_SCHED_CORE */
6302 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6303 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6304 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6306 static struct task_struct *
6307 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6309 return __pick_next_task(rq, prev, rf);
6312 #endif /* CONFIG_SCHED_CORE */
6315 * Constants for the sched_mode argument of __schedule().
6317 * The mode argument allows RT enabled kernels to differentiate a
6318 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6319 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6320 * optimize the AND operation out and just check for zero.
6323 #define SM_PREEMPT 0x1
6324 #define SM_RTLOCK_WAIT 0x2
6326 #ifndef CONFIG_PREEMPT_RT
6327 # define SM_MASK_PREEMPT (~0U)
6329 # define SM_MASK_PREEMPT SM_PREEMPT
6333 * __schedule() is the main scheduler function.
6335 * The main means of driving the scheduler and thus entering this function are:
6337 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6339 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6340 * paths. For example, see arch/x86/entry_64.S.
6342 * To drive preemption between tasks, the scheduler sets the flag in timer
6343 * interrupt handler scheduler_tick().
6345 * 3. Wakeups don't really cause entry into schedule(). They add a
6346 * task to the run-queue and that's it.
6348 * Now, if the new task added to the run-queue preempts the current
6349 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6350 * called on the nearest possible occasion:
6352 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6354 * - in syscall or exception context, at the next outmost
6355 * preempt_enable(). (this might be as soon as the wake_up()'s
6358 * - in IRQ context, return from interrupt-handler to
6359 * preemptible context
6361 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6364 * - cond_resched() call
6365 * - explicit schedule() call
6366 * - return from syscall or exception to user-space
6367 * - return from interrupt-handler to user-space
6369 * WARNING: must be called with preemption disabled!
6371 static void __sched notrace __schedule(unsigned int sched_mode)
6373 struct task_struct *prev, *next;
6374 unsigned long *switch_count;
6375 unsigned long prev_state;
6380 cpu = smp_processor_id();
6384 schedule_debug(prev, !!sched_mode);
6386 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6389 local_irq_disable();
6390 rcu_note_context_switch(!!sched_mode);
6393 * Make sure that signal_pending_state()->signal_pending() below
6394 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6395 * done by the caller to avoid the race with signal_wake_up():
6397 * __set_current_state(@state) signal_wake_up()
6398 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6399 * wake_up_state(p, state)
6400 * LOCK rq->lock LOCK p->pi_state
6401 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6402 * if (signal_pending_state()) if (p->state & @state)
6404 * Also, the membarrier system call requires a full memory barrier
6405 * after coming from user-space, before storing to rq->curr.
6408 smp_mb__after_spinlock();
6410 /* Promote REQ to ACT */
6411 rq->clock_update_flags <<= 1;
6412 update_rq_clock(rq);
6414 switch_count = &prev->nivcsw;
6417 * We must load prev->state once (task_struct::state is volatile), such
6418 * that we form a control dependency vs deactivate_task() below.
6420 prev_state = READ_ONCE(prev->__state);
6421 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6422 if (signal_pending_state(prev_state, prev)) {
6423 WRITE_ONCE(prev->__state, TASK_RUNNING);
6425 prev->sched_contributes_to_load =
6426 (prev_state & TASK_UNINTERRUPTIBLE) &&
6427 !(prev_state & TASK_NOLOAD) &&
6428 !(prev->flags & PF_FROZEN);
6430 if (prev->sched_contributes_to_load)
6431 rq->nr_uninterruptible++;
6434 * __schedule() ttwu()
6435 * prev_state = prev->state; if (p->on_rq && ...)
6436 * if (prev_state) goto out;
6437 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6438 * p->state = TASK_WAKING
6440 * Where __schedule() and ttwu() have matching control dependencies.
6442 * After this, schedule() must not care about p->state any more.
6444 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6446 if (prev->in_iowait) {
6447 atomic_inc(&rq->nr_iowait);
6448 delayacct_blkio_start();
6451 switch_count = &prev->nvcsw;
6454 next = pick_next_task(rq, prev, &rf);
6455 clear_tsk_need_resched(prev);
6456 clear_preempt_need_resched();
6457 #ifdef CONFIG_SCHED_DEBUG
6458 rq->last_seen_need_resched_ns = 0;
6461 if (likely(prev != next)) {
6464 * RCU users of rcu_dereference(rq->curr) may not see
6465 * changes to task_struct made by pick_next_task().
6467 RCU_INIT_POINTER(rq->curr, next);
6469 * The membarrier system call requires each architecture
6470 * to have a full memory barrier after updating
6471 * rq->curr, before returning to user-space.
6473 * Here are the schemes providing that barrier on the
6474 * various architectures:
6475 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6476 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6477 * - finish_lock_switch() for weakly-ordered
6478 * architectures where spin_unlock is a full barrier,
6479 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6480 * is a RELEASE barrier),
6484 migrate_disable_switch(rq, prev);
6485 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6487 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6489 /* Also unlocks the rq: */
6490 rq = context_switch(rq, prev, next, &rf);
6492 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6494 rq_unpin_lock(rq, &rf);
6495 __balance_callbacks(rq);
6496 raw_spin_rq_unlock_irq(rq);
6500 void __noreturn do_task_dead(void)
6502 /* Causes final put_task_struct in finish_task_switch(): */
6503 set_special_state(TASK_DEAD);
6505 /* Tell freezer to ignore us: */
6506 current->flags |= PF_NOFREEZE;
6508 __schedule(SM_NONE);
6511 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6516 static inline void sched_submit_work(struct task_struct *tsk)
6518 unsigned int task_flags;
6520 if (task_is_running(tsk))
6523 task_flags = tsk->flags;
6525 * If a worker goes to sleep, notify and ask workqueue whether it
6526 * wants to wake up a task to maintain concurrency.
6528 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6529 if (task_flags & PF_WQ_WORKER)
6530 wq_worker_sleeping(tsk);
6532 io_wq_worker_sleeping(tsk);
6536 * spinlock and rwlock must not flush block requests. This will
6537 * deadlock if the callback attempts to acquire a lock which is
6540 SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6543 * If we are going to sleep and we have plugged IO queued,
6544 * make sure to submit it to avoid deadlocks.
6546 blk_flush_plug(tsk->plug, true);
6549 static void sched_update_worker(struct task_struct *tsk)
6551 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6552 if (tsk->flags & PF_WQ_WORKER)
6553 wq_worker_running(tsk);
6555 io_wq_worker_running(tsk);
6559 asmlinkage __visible void __sched schedule(void)
6561 struct task_struct *tsk = current;
6563 sched_submit_work(tsk);
6566 __schedule(SM_NONE);
6567 sched_preempt_enable_no_resched();
6568 } while (need_resched());
6569 sched_update_worker(tsk);
6571 EXPORT_SYMBOL(schedule);
6574 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6575 * state (have scheduled out non-voluntarily) by making sure that all
6576 * tasks have either left the run queue or have gone into user space.
6577 * As idle tasks do not do either, they must not ever be preempted
6578 * (schedule out non-voluntarily).
6580 * schedule_idle() is similar to schedule_preempt_disable() except that it
6581 * never enables preemption because it does not call sched_submit_work().
6583 void __sched schedule_idle(void)
6586 * As this skips calling sched_submit_work(), which the idle task does
6587 * regardless because that function is a nop when the task is in a
6588 * TASK_RUNNING state, make sure this isn't used someplace that the
6589 * current task can be in any other state. Note, idle is always in the
6590 * TASK_RUNNING state.
6592 WARN_ON_ONCE(current->__state);
6594 __schedule(SM_NONE);
6595 } while (need_resched());
6598 #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6599 asmlinkage __visible void __sched schedule_user(void)
6602 * If we come here after a random call to set_need_resched(),
6603 * or we have been woken up remotely but the IPI has not yet arrived,
6604 * we haven't yet exited the RCU idle mode. Do it here manually until
6605 * we find a better solution.
6607 * NB: There are buggy callers of this function. Ideally we
6608 * should warn if prev_state != CONTEXT_USER, but that will trigger
6609 * too frequently to make sense yet.
6611 enum ctx_state prev_state = exception_enter();
6613 exception_exit(prev_state);
6618 * schedule_preempt_disabled - called with preemption disabled
6620 * Returns with preemption disabled. Note: preempt_count must be 1
6622 void __sched schedule_preempt_disabled(void)
6624 sched_preempt_enable_no_resched();
6629 #ifdef CONFIG_PREEMPT_RT
6630 void __sched notrace schedule_rtlock(void)
6634 __schedule(SM_RTLOCK_WAIT);
6635 sched_preempt_enable_no_resched();
6636 } while (need_resched());
6638 NOKPROBE_SYMBOL(schedule_rtlock);
6641 static void __sched notrace preempt_schedule_common(void)
6645 * Because the function tracer can trace preempt_count_sub()
6646 * and it also uses preempt_enable/disable_notrace(), if
6647 * NEED_RESCHED is set, the preempt_enable_notrace() called
6648 * by the function tracer will call this function again and
6649 * cause infinite recursion.
6651 * Preemption must be disabled here before the function
6652 * tracer can trace. Break up preempt_disable() into two
6653 * calls. One to disable preemption without fear of being
6654 * traced. The other to still record the preemption latency,
6655 * which can also be traced by the function tracer.
6657 preempt_disable_notrace();
6658 preempt_latency_start(1);
6659 __schedule(SM_PREEMPT);
6660 preempt_latency_stop(1);
6661 preempt_enable_no_resched_notrace();
6664 * Check again in case we missed a preemption opportunity
6665 * between schedule and now.
6667 } while (need_resched());
6670 #ifdef CONFIG_PREEMPTION
6672 * This is the entry point to schedule() from in-kernel preemption
6673 * off of preempt_enable.
6675 asmlinkage __visible void __sched notrace preempt_schedule(void)
6678 * If there is a non-zero preempt_count or interrupts are disabled,
6679 * we do not want to preempt the current task. Just return..
6681 if (likely(!preemptible()))
6683 preempt_schedule_common();
6685 NOKPROBE_SYMBOL(preempt_schedule);
6686 EXPORT_SYMBOL(preempt_schedule);
6688 #ifdef CONFIG_PREEMPT_DYNAMIC
6689 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6690 #ifndef preempt_schedule_dynamic_enabled
6691 #define preempt_schedule_dynamic_enabled preempt_schedule
6692 #define preempt_schedule_dynamic_disabled NULL
6694 DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6695 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6696 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6697 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6698 void __sched notrace dynamic_preempt_schedule(void)
6700 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6704 NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6705 EXPORT_SYMBOL(dynamic_preempt_schedule);
6710 * preempt_schedule_notrace - preempt_schedule called by tracing
6712 * The tracing infrastructure uses preempt_enable_notrace to prevent
6713 * recursion and tracing preempt enabling caused by the tracing
6714 * infrastructure itself. But as tracing can happen in areas coming
6715 * from userspace or just about to enter userspace, a preempt enable
6716 * can occur before user_exit() is called. This will cause the scheduler
6717 * to be called when the system is still in usermode.
6719 * To prevent this, the preempt_enable_notrace will use this function
6720 * instead of preempt_schedule() to exit user context if needed before
6721 * calling the scheduler.
6723 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6725 enum ctx_state prev_ctx;
6727 if (likely(!preemptible()))
6732 * Because the function tracer can trace preempt_count_sub()
6733 * and it also uses preempt_enable/disable_notrace(), if
6734 * NEED_RESCHED is set, the preempt_enable_notrace() called
6735 * by the function tracer will call this function again and
6736 * cause infinite recursion.
6738 * Preemption must be disabled here before the function
6739 * tracer can trace. Break up preempt_disable() into two
6740 * calls. One to disable preemption without fear of being
6741 * traced. The other to still record the preemption latency,
6742 * which can also be traced by the function tracer.
6744 preempt_disable_notrace();
6745 preempt_latency_start(1);
6747 * Needs preempt disabled in case user_exit() is traced
6748 * and the tracer calls preempt_enable_notrace() causing
6749 * an infinite recursion.
6751 prev_ctx = exception_enter();
6752 __schedule(SM_PREEMPT);
6753 exception_exit(prev_ctx);
6755 preempt_latency_stop(1);
6756 preempt_enable_no_resched_notrace();
6757 } while (need_resched());
6759 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6761 #ifdef CONFIG_PREEMPT_DYNAMIC
6762 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6763 #ifndef preempt_schedule_notrace_dynamic_enabled
6764 #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6765 #define preempt_schedule_notrace_dynamic_disabled NULL
6767 DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6768 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6769 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6770 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6771 void __sched notrace dynamic_preempt_schedule_notrace(void)
6773 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6775 preempt_schedule_notrace();
6777 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6778 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6782 #endif /* CONFIG_PREEMPTION */
6785 * This is the entry point to schedule() from kernel preemption
6786 * off of irq context.
6787 * Note, that this is called and return with irqs disabled. This will
6788 * protect us against recursive calling from irq.
6790 asmlinkage __visible void __sched preempt_schedule_irq(void)
6792 enum ctx_state prev_state;
6794 /* Catch callers which need to be fixed */
6795 BUG_ON(preempt_count() || !irqs_disabled());
6797 prev_state = exception_enter();
6802 __schedule(SM_PREEMPT);
6803 local_irq_disable();
6804 sched_preempt_enable_no_resched();
6805 } while (need_resched());
6807 exception_exit(prev_state);
6810 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6813 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6814 return try_to_wake_up(curr->private, mode, wake_flags);
6816 EXPORT_SYMBOL(default_wake_function);
6818 static void __setscheduler_prio(struct task_struct *p, int prio)
6821 p->sched_class = &dl_sched_class;
6822 else if (rt_prio(prio))
6823 p->sched_class = &rt_sched_class;
6825 p->sched_class = &fair_sched_class;
6830 #ifdef CONFIG_RT_MUTEXES
6832 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6835 prio = min(prio, pi_task->prio);
6840 static inline int rt_effective_prio(struct task_struct *p, int prio)
6842 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6844 return __rt_effective_prio(pi_task, prio);
6848 * rt_mutex_setprio - set the current priority of a task
6850 * @pi_task: donor task
6852 * This function changes the 'effective' priority of a task. It does
6853 * not touch ->normal_prio like __setscheduler().
6855 * Used by the rt_mutex code to implement priority inheritance
6856 * logic. Call site only calls if the priority of the task changed.
6858 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6860 int prio, oldprio, queued, running, queue_flag =
6861 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6862 const struct sched_class *prev_class;
6866 /* XXX used to be waiter->prio, not waiter->task->prio */
6867 prio = __rt_effective_prio(pi_task, p->normal_prio);
6870 * If nothing changed; bail early.
6872 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6875 rq = __task_rq_lock(p, &rf);
6876 update_rq_clock(rq);
6878 * Set under pi_lock && rq->lock, such that the value can be used under
6881 * Note that there is loads of tricky to make this pointer cache work
6882 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6883 * ensure a task is de-boosted (pi_task is set to NULL) before the
6884 * task is allowed to run again (and can exit). This ensures the pointer
6885 * points to a blocked task -- which guarantees the task is present.
6887 p->pi_top_task = pi_task;
6890 * For FIFO/RR we only need to set prio, if that matches we're done.
6892 if (prio == p->prio && !dl_prio(prio))
6896 * Idle task boosting is a nono in general. There is one
6897 * exception, when PREEMPT_RT and NOHZ is active:
6899 * The idle task calls get_next_timer_interrupt() and holds
6900 * the timer wheel base->lock on the CPU and another CPU wants
6901 * to access the timer (probably to cancel it). We can safely
6902 * ignore the boosting request, as the idle CPU runs this code
6903 * with interrupts disabled and will complete the lock
6904 * protected section without being interrupted. So there is no
6905 * real need to boost.
6907 if (unlikely(p == rq->idle)) {
6908 WARN_ON(p != rq->curr);
6909 WARN_ON(p->pi_blocked_on);
6913 trace_sched_pi_setprio(p, pi_task);
6916 if (oldprio == prio)
6917 queue_flag &= ~DEQUEUE_MOVE;
6919 prev_class = p->sched_class;
6920 queued = task_on_rq_queued(p);
6921 running = task_current(rq, p);
6923 dequeue_task(rq, p, queue_flag);
6925 put_prev_task(rq, p);
6928 * Boosting condition are:
6929 * 1. -rt task is running and holds mutex A
6930 * --> -dl task blocks on mutex A
6932 * 2. -dl task is running and holds mutex A
6933 * --> -dl task blocks on mutex A and could preempt the
6936 if (dl_prio(prio)) {
6937 if (!dl_prio(p->normal_prio) ||
6938 (pi_task && dl_prio(pi_task->prio) &&
6939 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6940 p->dl.pi_se = pi_task->dl.pi_se;
6941 queue_flag |= ENQUEUE_REPLENISH;
6943 p->dl.pi_se = &p->dl;
6945 } else if (rt_prio(prio)) {
6946 if (dl_prio(oldprio))
6947 p->dl.pi_se = &p->dl;
6949 queue_flag |= ENQUEUE_HEAD;
6951 if (dl_prio(oldprio))
6952 p->dl.pi_se = &p->dl;
6953 if (rt_prio(oldprio))
6957 __setscheduler_prio(p, prio);
6960 enqueue_task(rq, p, queue_flag);
6962 set_next_task(rq, p);
6964 check_class_changed(rq, p, prev_class, oldprio);
6966 /* Avoid rq from going away on us: */
6969 rq_unpin_lock(rq, &rf);
6970 __balance_callbacks(rq);
6971 raw_spin_rq_unlock(rq);
6976 static inline int rt_effective_prio(struct task_struct *p, int prio)
6982 void set_user_nice(struct task_struct *p, long nice)
6984 bool queued, running;
6989 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6992 * We have to be careful, if called from sys_setpriority(),
6993 * the task might be in the middle of scheduling on another CPU.
6995 rq = task_rq_lock(p, &rf);
6996 update_rq_clock(rq);
6999 * The RT priorities are set via sched_setscheduler(), but we still
7000 * allow the 'normal' nice value to be set - but as expected
7001 * it won't have any effect on scheduling until the task is
7002 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7004 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7005 p->static_prio = NICE_TO_PRIO(nice);
7008 queued = task_on_rq_queued(p);
7009 running = task_current(rq, p);
7011 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7013 put_prev_task(rq, p);
7015 p->static_prio = NICE_TO_PRIO(nice);
7016 set_load_weight(p, true);
7018 p->prio = effective_prio(p);
7021 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7023 set_next_task(rq, p);
7026 * If the task increased its priority or is running and
7027 * lowered its priority, then reschedule its CPU:
7029 p->sched_class->prio_changed(rq, p, old_prio);
7032 task_rq_unlock(rq, p, &rf);
7034 EXPORT_SYMBOL(set_user_nice);
7037 * is_nice_reduction - check if nice value is an actual reduction
7039 * Similar to can_nice() but does not perform a capability check.
7044 static bool is_nice_reduction(const struct task_struct *p, const int nice)
7046 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7047 int nice_rlim = nice_to_rlimit(nice);
7049 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7053 * can_nice - check if a task can reduce its nice value
7057 int can_nice(const struct task_struct *p, const int nice)
7059 return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7062 #ifdef __ARCH_WANT_SYS_NICE
7065 * sys_nice - change the priority of the current process.
7066 * @increment: priority increment
7068 * sys_setpriority is a more generic, but much slower function that
7069 * does similar things.
7071 SYSCALL_DEFINE1(nice, int, increment)
7076 * Setpriority might change our priority at the same moment.
7077 * We don't have to worry. Conceptually one call occurs first
7078 * and we have a single winner.
7080 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7081 nice = task_nice(current) + increment;
7083 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7084 if (increment < 0 && !can_nice(current, nice))
7087 retval = security_task_setnice(current, nice);
7091 set_user_nice(current, nice);
7098 * task_prio - return the priority value of a given task.
7099 * @p: the task in question.
7101 * Return: The priority value as seen by users in /proc.
7103 * sched policy return value kernel prio user prio/nice
7105 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7106 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7107 * deadline -101 -1 0
7109 int task_prio(const struct task_struct *p)
7111 return p->prio - MAX_RT_PRIO;
7115 * idle_cpu - is a given CPU idle currently?
7116 * @cpu: the processor in question.
7118 * Return: 1 if the CPU is currently idle. 0 otherwise.
7120 int idle_cpu(int cpu)
7122 struct rq *rq = cpu_rq(cpu);
7124 if (rq->curr != rq->idle)
7131 if (rq->ttwu_pending)
7139 * available_idle_cpu - is a given CPU idle for enqueuing work.
7140 * @cpu: the CPU in question.
7142 * Return: 1 if the CPU is currently idle. 0 otherwise.
7144 int available_idle_cpu(int cpu)
7149 if (vcpu_is_preempted(cpu))
7156 * idle_task - return the idle task for a given CPU.
7157 * @cpu: the processor in question.
7159 * Return: The idle task for the CPU @cpu.
7161 struct task_struct *idle_task(int cpu)
7163 return cpu_rq(cpu)->idle;
7168 * This function computes an effective utilization for the given CPU, to be
7169 * used for frequency selection given the linear relation: f = u * f_max.
7171 * The scheduler tracks the following metrics:
7173 * cpu_util_{cfs,rt,dl,irq}()
7176 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7177 * synchronized windows and are thus directly comparable.
7179 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7180 * which excludes things like IRQ and steal-time. These latter are then accrued
7181 * in the irq utilization.
7183 * The DL bandwidth number otoh is not a measured metric but a value computed
7184 * based on the task model parameters and gives the minimal utilization
7185 * required to meet deadlines.
7187 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7188 enum cpu_util_type type,
7189 struct task_struct *p)
7191 unsigned long dl_util, util, irq, max;
7192 struct rq *rq = cpu_rq(cpu);
7194 max = arch_scale_cpu_capacity(cpu);
7196 if (!uclamp_is_used() &&
7197 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7202 * Early check to see if IRQ/steal time saturates the CPU, can be
7203 * because of inaccuracies in how we track these -- see
7204 * update_irq_load_avg().
7206 irq = cpu_util_irq(rq);
7207 if (unlikely(irq >= max))
7211 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7212 * CFS tasks and we use the same metric to track the effective
7213 * utilization (PELT windows are synchronized) we can directly add them
7214 * to obtain the CPU's actual utilization.
7216 * CFS and RT utilization can be boosted or capped, depending on
7217 * utilization clamp constraints requested by currently RUNNABLE
7219 * When there are no CFS RUNNABLE tasks, clamps are released and
7220 * frequency will be gracefully reduced with the utilization decay.
7222 util = util_cfs + cpu_util_rt(rq);
7223 if (type == FREQUENCY_UTIL)
7224 util = uclamp_rq_util_with(rq, util, p);
7226 dl_util = cpu_util_dl(rq);
7229 * For frequency selection we do not make cpu_util_dl() a permanent part
7230 * of this sum because we want to use cpu_bw_dl() later on, but we need
7231 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7232 * that we select f_max when there is no idle time.
7234 * NOTE: numerical errors or stop class might cause us to not quite hit
7235 * saturation when we should -- something for later.
7237 if (util + dl_util >= max)
7241 * OTOH, for energy computation we need the estimated running time, so
7242 * include util_dl and ignore dl_bw.
7244 if (type == ENERGY_UTIL)
7248 * There is still idle time; further improve the number by using the
7249 * irq metric. Because IRQ/steal time is hidden from the task clock we
7250 * need to scale the task numbers:
7253 * U' = irq + --------- * U
7256 util = scale_irq_capacity(util, irq, max);
7260 * Bandwidth required by DEADLINE must always be granted while, for
7261 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7262 * to gracefully reduce the frequency when no tasks show up for longer
7265 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7266 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7267 * an interface. So, we only do the latter for now.
7269 if (type == FREQUENCY_UTIL)
7270 util += cpu_bw_dl(rq);
7272 return min(max, util);
7275 unsigned long sched_cpu_util(int cpu)
7277 return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
7279 #endif /* CONFIG_SMP */
7282 * find_process_by_pid - find a process with a matching PID value.
7283 * @pid: the pid in question.
7285 * The task of @pid, if found. %NULL otherwise.
7287 static struct task_struct *find_process_by_pid(pid_t pid)
7289 return pid ? find_task_by_vpid(pid) : current;
7293 * sched_setparam() passes in -1 for its policy, to let the functions
7294 * it calls know not to change it.
7296 #define SETPARAM_POLICY -1
7298 static void __setscheduler_params(struct task_struct *p,
7299 const struct sched_attr *attr)
7301 int policy = attr->sched_policy;
7303 if (policy == SETPARAM_POLICY)
7308 if (dl_policy(policy))
7309 __setparam_dl(p, attr);
7310 else if (fair_policy(policy))
7311 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7314 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7315 * !rt_policy. Always setting this ensures that things like
7316 * getparam()/getattr() don't report silly values for !rt tasks.
7318 p->rt_priority = attr->sched_priority;
7319 p->normal_prio = normal_prio(p);
7320 set_load_weight(p, true);
7324 * Check the target process has a UID that matches the current process's:
7326 static bool check_same_owner(struct task_struct *p)
7328 const struct cred *cred = current_cred(), *pcred;
7332 pcred = __task_cred(p);
7333 match = (uid_eq(cred->euid, pcred->euid) ||
7334 uid_eq(cred->euid, pcred->uid));
7340 * Allow unprivileged RT tasks to decrease priority.
7341 * Only issue a capable test if needed and only once to avoid an audit
7342 * event on permitted non-privileged operations:
7344 static int user_check_sched_setscheduler(struct task_struct *p,
7345 const struct sched_attr *attr,
7346 int policy, int reset_on_fork)
7348 if (fair_policy(policy)) {
7349 if (attr->sched_nice < task_nice(p) &&
7350 !is_nice_reduction(p, attr->sched_nice))
7354 if (rt_policy(policy)) {
7355 unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7357 /* Can't set/change the rt policy: */
7358 if (policy != p->policy && !rlim_rtprio)
7361 /* Can't increase priority: */
7362 if (attr->sched_priority > p->rt_priority &&
7363 attr->sched_priority > rlim_rtprio)
7368 * Can't set/change SCHED_DEADLINE policy at all for now
7369 * (safest behavior); in the future we would like to allow
7370 * unprivileged DL tasks to increase their relative deadline
7371 * or reduce their runtime (both ways reducing utilization)
7373 if (dl_policy(policy))
7377 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7378 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7380 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7381 if (!is_nice_reduction(p, task_nice(p)))
7385 /* Can't change other user's priorities: */
7386 if (!check_same_owner(p))
7389 /* Normal users shall not reset the sched_reset_on_fork flag: */
7390 if (p->sched_reset_on_fork && !reset_on_fork)
7396 if (!capable(CAP_SYS_NICE))
7402 static int __sched_setscheduler(struct task_struct *p,
7403 const struct sched_attr *attr,
7406 int oldpolicy = -1, policy = attr->sched_policy;
7407 int retval, oldprio, newprio, queued, running;
7408 const struct sched_class *prev_class;
7409 struct callback_head *head;
7412 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7415 /* The pi code expects interrupts enabled */
7416 BUG_ON(pi && in_interrupt());
7418 /* Double check policy once rq lock held: */
7420 reset_on_fork = p->sched_reset_on_fork;
7421 policy = oldpolicy = p->policy;
7423 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7425 if (!valid_policy(policy))
7429 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7433 * Valid priorities for SCHED_FIFO and SCHED_RR are
7434 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7435 * SCHED_BATCH and SCHED_IDLE is 0.
7437 if (attr->sched_priority > MAX_RT_PRIO-1)
7439 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7440 (rt_policy(policy) != (attr->sched_priority != 0)))
7444 retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7448 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7451 retval = security_task_setscheduler(p);
7456 /* Update task specific "requested" clamps */
7457 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7458 retval = uclamp_validate(p, attr);
7467 * Make sure no PI-waiters arrive (or leave) while we are
7468 * changing the priority of the task:
7470 * To be able to change p->policy safely, the appropriate
7471 * runqueue lock must be held.
7473 rq = task_rq_lock(p, &rf);
7474 update_rq_clock(rq);
7477 * Changing the policy of the stop threads its a very bad idea:
7479 if (p == rq->stop) {
7485 * If not changing anything there's no need to proceed further,
7486 * but store a possible modification of reset_on_fork.
7488 if (unlikely(policy == p->policy)) {
7489 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7491 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7493 if (dl_policy(policy) && dl_param_changed(p, attr))
7495 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7498 p->sched_reset_on_fork = reset_on_fork;
7505 #ifdef CONFIG_RT_GROUP_SCHED
7507 * Do not allow realtime tasks into groups that have no runtime
7510 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7511 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7512 !task_group_is_autogroup(task_group(p))) {
7518 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7519 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7520 cpumask_t *span = rq->rd->span;
7523 * Don't allow tasks with an affinity mask smaller than
7524 * the entire root_domain to become SCHED_DEADLINE. We
7525 * will also fail if there's no bandwidth available.
7527 if (!cpumask_subset(span, p->cpus_ptr) ||
7528 rq->rd->dl_bw.bw == 0) {
7536 /* Re-check policy now with rq lock held: */
7537 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7538 policy = oldpolicy = -1;
7539 task_rq_unlock(rq, p, &rf);
7541 cpuset_read_unlock();
7546 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7547 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7550 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7555 p->sched_reset_on_fork = reset_on_fork;
7558 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7561 * Take priority boosted tasks into account. If the new
7562 * effective priority is unchanged, we just store the new
7563 * normal parameters and do not touch the scheduler class and
7564 * the runqueue. This will be done when the task deboost
7567 newprio = rt_effective_prio(p, newprio);
7568 if (newprio == oldprio)
7569 queue_flags &= ~DEQUEUE_MOVE;
7572 queued = task_on_rq_queued(p);
7573 running = task_current(rq, p);
7575 dequeue_task(rq, p, queue_flags);
7577 put_prev_task(rq, p);
7579 prev_class = p->sched_class;
7581 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7582 __setscheduler_params(p, attr);
7583 __setscheduler_prio(p, newprio);
7585 __setscheduler_uclamp(p, attr);
7589 * We enqueue to tail when the priority of a task is
7590 * increased (user space view).
7592 if (oldprio < p->prio)
7593 queue_flags |= ENQUEUE_HEAD;
7595 enqueue_task(rq, p, queue_flags);
7598 set_next_task(rq, p);
7600 check_class_changed(rq, p, prev_class, oldprio);
7602 /* Avoid rq from going away on us: */
7604 head = splice_balance_callbacks(rq);
7605 task_rq_unlock(rq, p, &rf);
7608 cpuset_read_unlock();
7609 rt_mutex_adjust_pi(p);
7612 /* Run balance callbacks after we've adjusted the PI chain: */
7613 balance_callbacks(rq, head);
7619 task_rq_unlock(rq, p, &rf);
7621 cpuset_read_unlock();
7625 static int _sched_setscheduler(struct task_struct *p, int policy,
7626 const struct sched_param *param, bool check)
7628 struct sched_attr attr = {
7629 .sched_policy = policy,
7630 .sched_priority = param->sched_priority,
7631 .sched_nice = PRIO_TO_NICE(p->static_prio),
7634 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7635 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7636 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7637 policy &= ~SCHED_RESET_ON_FORK;
7638 attr.sched_policy = policy;
7641 return __sched_setscheduler(p, &attr, check, true);
7644 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7645 * @p: the task in question.
7646 * @policy: new policy.
7647 * @param: structure containing the new RT priority.
7649 * Use sched_set_fifo(), read its comment.
7651 * Return: 0 on success. An error code otherwise.
7653 * NOTE that the task may be already dead.
7655 int sched_setscheduler(struct task_struct *p, int policy,
7656 const struct sched_param *param)
7658 return _sched_setscheduler(p, policy, param, true);
7661 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7663 return __sched_setscheduler(p, attr, true, true);
7666 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7668 return __sched_setscheduler(p, attr, false, true);
7670 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7673 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7674 * @p: the task in question.
7675 * @policy: new policy.
7676 * @param: structure containing the new RT priority.
7678 * Just like sched_setscheduler, only don't bother checking if the
7679 * current context has permission. For example, this is needed in
7680 * stop_machine(): we create temporary high priority worker threads,
7681 * but our caller might not have that capability.
7683 * Return: 0 on success. An error code otherwise.
7685 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7686 const struct sched_param *param)
7688 return _sched_setscheduler(p, policy, param, false);
7692 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7693 * incapable of resource management, which is the one thing an OS really should
7696 * This is of course the reason it is limited to privileged users only.
7698 * Worse still; it is fundamentally impossible to compose static priority
7699 * workloads. You cannot take two correctly working static prio workloads
7700 * and smash them together and still expect them to work.
7702 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7706 * The administrator _MUST_ configure the system, the kernel simply doesn't
7707 * know enough information to make a sensible choice.
7709 void sched_set_fifo(struct task_struct *p)
7711 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7712 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7714 EXPORT_SYMBOL_GPL(sched_set_fifo);
7717 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7719 void sched_set_fifo_low(struct task_struct *p)
7721 struct sched_param sp = { .sched_priority = 1 };
7722 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7724 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7726 void sched_set_normal(struct task_struct *p, int nice)
7728 struct sched_attr attr = {
7729 .sched_policy = SCHED_NORMAL,
7732 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7734 EXPORT_SYMBOL_GPL(sched_set_normal);
7737 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7739 struct sched_param lparam;
7740 struct task_struct *p;
7743 if (!param || pid < 0)
7745 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7750 p = find_process_by_pid(pid);
7756 retval = sched_setscheduler(p, policy, &lparam);
7764 * Mimics kernel/events/core.c perf_copy_attr().
7766 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7771 /* Zero the full structure, so that a short copy will be nice: */
7772 memset(attr, 0, sizeof(*attr));
7774 ret = get_user(size, &uattr->size);
7778 /* ABI compatibility quirk: */
7780 size = SCHED_ATTR_SIZE_VER0;
7781 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7784 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7791 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7792 size < SCHED_ATTR_SIZE_VER1)
7796 * XXX: Do we want to be lenient like existing syscalls; or do we want
7797 * to be strict and return an error on out-of-bounds values?
7799 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7804 put_user(sizeof(*attr), &uattr->size);
7808 static void get_params(struct task_struct *p, struct sched_attr *attr)
7810 if (task_has_dl_policy(p))
7811 __getparam_dl(p, attr);
7812 else if (task_has_rt_policy(p))
7813 attr->sched_priority = p->rt_priority;
7815 attr->sched_nice = task_nice(p);
7819 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7820 * @pid: the pid in question.
7821 * @policy: new policy.
7822 * @param: structure containing the new RT priority.
7824 * Return: 0 on success. An error code otherwise.
7826 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7831 return do_sched_setscheduler(pid, policy, param);
7835 * sys_sched_setparam - set/change the RT priority of a thread
7836 * @pid: the pid in question.
7837 * @param: structure containing the new RT priority.
7839 * Return: 0 on success. An error code otherwise.
7841 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7843 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7847 * sys_sched_setattr - same as above, but with extended sched_attr
7848 * @pid: the pid in question.
7849 * @uattr: structure containing the extended parameters.
7850 * @flags: for future extension.
7852 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7853 unsigned int, flags)
7855 struct sched_attr attr;
7856 struct task_struct *p;
7859 if (!uattr || pid < 0 || flags)
7862 retval = sched_copy_attr(uattr, &attr);
7866 if ((int)attr.sched_policy < 0)
7868 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7869 attr.sched_policy = SETPARAM_POLICY;
7873 p = find_process_by_pid(pid);
7879 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7880 get_params(p, &attr);
7881 retval = sched_setattr(p, &attr);
7889 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7890 * @pid: the pid in question.
7892 * Return: On success, the policy of the thread. Otherwise, a negative error
7895 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7897 struct task_struct *p;
7905 p = find_process_by_pid(pid);
7907 retval = security_task_getscheduler(p);
7910 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7917 * sys_sched_getparam - get the RT priority of a thread
7918 * @pid: the pid in question.
7919 * @param: structure containing the RT priority.
7921 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7924 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7926 struct sched_param lp = { .sched_priority = 0 };
7927 struct task_struct *p;
7930 if (!param || pid < 0)
7934 p = find_process_by_pid(pid);
7939 retval = security_task_getscheduler(p);
7943 if (task_has_rt_policy(p))
7944 lp.sched_priority = p->rt_priority;
7948 * This one might sleep, we cannot do it with a spinlock held ...
7950 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7960 * Copy the kernel size attribute structure (which might be larger
7961 * than what user-space knows about) to user-space.
7963 * Note that all cases are valid: user-space buffer can be larger or
7964 * smaller than the kernel-space buffer. The usual case is that both
7965 * have the same size.
7968 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7969 struct sched_attr *kattr,
7972 unsigned int ksize = sizeof(*kattr);
7974 if (!access_ok(uattr, usize))
7978 * sched_getattr() ABI forwards and backwards compatibility:
7980 * If usize == ksize then we just copy everything to user-space and all is good.
7982 * If usize < ksize then we only copy as much as user-space has space for,
7983 * this keeps ABI compatibility as well. We skip the rest.
7985 * If usize > ksize then user-space is using a newer version of the ABI,
7986 * which part the kernel doesn't know about. Just ignore it - tooling can
7987 * detect the kernel's knowledge of attributes from the attr->size value
7988 * which is set to ksize in this case.
7990 kattr->size = min(usize, ksize);
7992 if (copy_to_user(uattr, kattr, kattr->size))
7999 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8000 * @pid: the pid in question.
8001 * @uattr: structure containing the extended parameters.
8002 * @usize: sizeof(attr) for fwd/bwd comp.
8003 * @flags: for future extension.
8005 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8006 unsigned int, usize, unsigned int, flags)
8008 struct sched_attr kattr = { };
8009 struct task_struct *p;
8012 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8013 usize < SCHED_ATTR_SIZE_VER0 || flags)
8017 p = find_process_by_pid(pid);
8022 retval = security_task_getscheduler(p);
8026 kattr.sched_policy = p->policy;
8027 if (p->sched_reset_on_fork)
8028 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8029 get_params(p, &kattr);
8030 kattr.sched_flags &= SCHED_FLAG_ALL;
8032 #ifdef CONFIG_UCLAMP_TASK
8034 * This could race with another potential updater, but this is fine
8035 * because it'll correctly read the old or the new value. We don't need
8036 * to guarantee who wins the race as long as it doesn't return garbage.
8038 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8039 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8044 return sched_attr_copy_to_user(uattr, &kattr, usize);
8052 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8057 * If the task isn't a deadline task or admission control is
8058 * disabled then we don't care about affinity changes.
8060 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8064 * Since bandwidth control happens on root_domain basis,
8065 * if admission test is enabled, we only admit -deadline
8066 * tasks allowed to run on all the CPUs in the task's
8070 if (!cpumask_subset(task_rq(p)->rd->span, mask))
8078 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
8081 cpumask_var_t cpus_allowed, new_mask;
8083 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8086 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8088 goto out_free_cpus_allowed;
8091 cpuset_cpus_allowed(p, cpus_allowed);
8092 cpumask_and(new_mask, mask, cpus_allowed);
8094 retval = dl_task_check_affinity(p, new_mask);
8096 goto out_free_new_mask;
8098 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
8100 goto out_free_new_mask;
8102 cpuset_cpus_allowed(p, cpus_allowed);
8103 if (!cpumask_subset(new_mask, cpus_allowed)) {
8105 * We must have raced with a concurrent cpuset update.
8106 * Just reset the cpumask to the cpuset's cpus_allowed.
8108 cpumask_copy(new_mask, cpus_allowed);
8113 free_cpumask_var(new_mask);
8114 out_free_cpus_allowed:
8115 free_cpumask_var(cpus_allowed);
8119 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8121 struct task_struct *p;
8126 p = find_process_by_pid(pid);
8132 /* Prevent p going away */
8136 if (p->flags & PF_NO_SETAFFINITY) {
8141 if (!check_same_owner(p)) {
8143 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8151 retval = security_task_setscheduler(p);
8155 retval = __sched_setaffinity(p, in_mask);
8161 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8162 struct cpumask *new_mask)
8164 if (len < cpumask_size())
8165 cpumask_clear(new_mask);
8166 else if (len > cpumask_size())
8167 len = cpumask_size();
8169 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8173 * sys_sched_setaffinity - set the CPU affinity of a process
8174 * @pid: pid of the process
8175 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8176 * @user_mask_ptr: user-space pointer to the new CPU mask
8178 * Return: 0 on success. An error code otherwise.
8180 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8181 unsigned long __user *, user_mask_ptr)
8183 cpumask_var_t new_mask;
8186 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8189 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8191 retval = sched_setaffinity(pid, new_mask);
8192 free_cpumask_var(new_mask);
8196 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8198 struct task_struct *p;
8199 unsigned long flags;
8205 p = find_process_by_pid(pid);
8209 retval = security_task_getscheduler(p);
8213 raw_spin_lock_irqsave(&p->pi_lock, flags);
8214 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8215 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8224 * sys_sched_getaffinity - get the CPU affinity of a process
8225 * @pid: pid of the process
8226 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8227 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8229 * Return: size of CPU mask copied to user_mask_ptr on success. An
8230 * error code otherwise.
8232 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8233 unsigned long __user *, user_mask_ptr)
8238 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8240 if (len & (sizeof(unsigned long)-1))
8243 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8246 ret = sched_getaffinity(pid, mask);
8248 unsigned int retlen = min(len, cpumask_size());
8250 if (copy_to_user(user_mask_ptr, mask, retlen))
8255 free_cpumask_var(mask);
8260 static void do_sched_yield(void)
8265 rq = this_rq_lock_irq(&rf);
8267 schedstat_inc(rq->yld_count);
8268 current->sched_class->yield_task(rq);
8271 rq_unlock_irq(rq, &rf);
8272 sched_preempt_enable_no_resched();
8278 * sys_sched_yield - yield the current processor to other threads.
8280 * This function yields the current CPU to other tasks. If there are no
8281 * other threads running on this CPU then this function will return.
8285 SYSCALL_DEFINE0(sched_yield)
8291 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8292 int __sched __cond_resched(void)
8294 if (should_resched(0)) {
8295 preempt_schedule_common();
8299 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8300 * whether the current CPU is in an RCU read-side critical section,
8301 * so the tick can report quiescent states even for CPUs looping
8302 * in kernel context. In contrast, in non-preemptible kernels,
8303 * RCU readers leave no in-memory hints, which means that CPU-bound
8304 * processes executing in kernel context might never report an
8305 * RCU quiescent state. Therefore, the following code causes
8306 * cond_resched() to report a quiescent state, but only when RCU
8307 * is in urgent need of one.
8309 #ifndef CONFIG_PREEMPT_RCU
8314 EXPORT_SYMBOL(__cond_resched);
8317 #ifdef CONFIG_PREEMPT_DYNAMIC
8318 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8319 #define cond_resched_dynamic_enabled __cond_resched
8320 #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8321 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8322 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8324 #define might_resched_dynamic_enabled __cond_resched
8325 #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8326 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8327 EXPORT_STATIC_CALL_TRAMP(might_resched);
8328 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8329 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8330 int __sched dynamic_cond_resched(void)
8332 if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8334 return __cond_resched();
8336 EXPORT_SYMBOL(dynamic_cond_resched);
8338 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8339 int __sched dynamic_might_resched(void)
8341 if (!static_branch_unlikely(&sk_dynamic_might_resched))
8343 return __cond_resched();
8345 EXPORT_SYMBOL(dynamic_might_resched);
8350 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8351 * call schedule, and on return reacquire the lock.
8353 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8354 * operations here to prevent schedule() from being called twice (once via
8355 * spin_unlock(), once by hand).
8357 int __cond_resched_lock(spinlock_t *lock)
8359 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8362 lockdep_assert_held(lock);
8364 if (spin_needbreak(lock) || resched) {
8366 if (!_cond_resched())
8373 EXPORT_SYMBOL(__cond_resched_lock);
8375 int __cond_resched_rwlock_read(rwlock_t *lock)
8377 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8380 lockdep_assert_held_read(lock);
8382 if (rwlock_needbreak(lock) || resched) {
8384 if (!_cond_resched())
8391 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8393 int __cond_resched_rwlock_write(rwlock_t *lock)
8395 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8398 lockdep_assert_held_write(lock);
8400 if (rwlock_needbreak(lock) || resched) {
8402 if (!_cond_resched())
8409 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8411 #ifdef CONFIG_PREEMPT_DYNAMIC
8413 #ifdef CONFIG_GENERIC_ENTRY
8414 #include <linux/entry-common.h>
8420 * SC:preempt_schedule
8421 * SC:preempt_schedule_notrace
8422 * SC:irqentry_exit_cond_resched
8426 * cond_resched <- __cond_resched
8427 * might_resched <- RET0
8428 * preempt_schedule <- NOP
8429 * preempt_schedule_notrace <- NOP
8430 * irqentry_exit_cond_resched <- NOP
8433 * cond_resched <- __cond_resched
8434 * might_resched <- __cond_resched
8435 * preempt_schedule <- NOP
8436 * preempt_schedule_notrace <- NOP
8437 * irqentry_exit_cond_resched <- NOP
8440 * cond_resched <- RET0
8441 * might_resched <- RET0
8442 * preempt_schedule <- preempt_schedule
8443 * preempt_schedule_notrace <- preempt_schedule_notrace
8444 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8448 preempt_dynamic_undefined = -1,
8449 preempt_dynamic_none,
8450 preempt_dynamic_voluntary,
8451 preempt_dynamic_full,
8454 int preempt_dynamic_mode = preempt_dynamic_undefined;
8456 int sched_dynamic_mode(const char *str)
8458 if (!strcmp(str, "none"))
8459 return preempt_dynamic_none;
8461 if (!strcmp(str, "voluntary"))
8462 return preempt_dynamic_voluntary;
8464 if (!strcmp(str, "full"))
8465 return preempt_dynamic_full;
8470 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8471 #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8472 #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8473 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8474 #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8475 #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8477 #error "Unsupported PREEMPT_DYNAMIC mechanism"
8480 void sched_dynamic_update(int mode)
8483 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8484 * the ZERO state, which is invalid.
8486 preempt_dynamic_enable(cond_resched);
8487 preempt_dynamic_enable(might_resched);
8488 preempt_dynamic_enable(preempt_schedule);
8489 preempt_dynamic_enable(preempt_schedule_notrace);
8490 preempt_dynamic_enable(irqentry_exit_cond_resched);
8493 case preempt_dynamic_none:
8494 preempt_dynamic_enable(cond_resched);
8495 preempt_dynamic_disable(might_resched);
8496 preempt_dynamic_disable(preempt_schedule);
8497 preempt_dynamic_disable(preempt_schedule_notrace);
8498 preempt_dynamic_disable(irqentry_exit_cond_resched);
8499 pr_info("Dynamic Preempt: none\n");
8502 case preempt_dynamic_voluntary:
8503 preempt_dynamic_enable(cond_resched);
8504 preempt_dynamic_enable(might_resched);
8505 preempt_dynamic_disable(preempt_schedule);
8506 preempt_dynamic_disable(preempt_schedule_notrace);
8507 preempt_dynamic_disable(irqentry_exit_cond_resched);
8508 pr_info("Dynamic Preempt: voluntary\n");
8511 case preempt_dynamic_full:
8512 preempt_dynamic_disable(cond_resched);
8513 preempt_dynamic_disable(might_resched);
8514 preempt_dynamic_enable(preempt_schedule);
8515 preempt_dynamic_enable(preempt_schedule_notrace);
8516 preempt_dynamic_enable(irqentry_exit_cond_resched);
8517 pr_info("Dynamic Preempt: full\n");
8521 preempt_dynamic_mode = mode;
8524 static int __init setup_preempt_mode(char *str)
8526 int mode = sched_dynamic_mode(str);
8528 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8532 sched_dynamic_update(mode);
8535 __setup("preempt=", setup_preempt_mode);
8537 static void __init preempt_dynamic_init(void)
8539 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8540 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8541 sched_dynamic_update(preempt_dynamic_none);
8542 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8543 sched_dynamic_update(preempt_dynamic_voluntary);
8545 /* Default static call setting, nothing to do */
8546 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8547 preempt_dynamic_mode = preempt_dynamic_full;
8548 pr_info("Dynamic Preempt: full\n");
8553 #define PREEMPT_MODEL_ACCESSOR(mode) \
8554 bool preempt_model_##mode(void) \
8556 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8557 return preempt_dynamic_mode == preempt_dynamic_##mode; \
8559 EXPORT_SYMBOL_GPL(preempt_model_##mode)
8561 PREEMPT_MODEL_ACCESSOR(none);
8562 PREEMPT_MODEL_ACCESSOR(voluntary);
8563 PREEMPT_MODEL_ACCESSOR(full);
8565 #else /* !CONFIG_PREEMPT_DYNAMIC */
8567 static inline void preempt_dynamic_init(void) { }
8569 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8572 * yield - yield the current processor to other threads.
8574 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8576 * The scheduler is at all times free to pick the calling task as the most
8577 * eligible task to run, if removing the yield() call from your code breaks
8578 * it, it's already broken.
8580 * Typical broken usage is:
8585 * where one assumes that yield() will let 'the other' process run that will
8586 * make event true. If the current task is a SCHED_FIFO task that will never
8587 * happen. Never use yield() as a progress guarantee!!
8589 * If you want to use yield() to wait for something, use wait_event().
8590 * If you want to use yield() to be 'nice' for others, use cond_resched().
8591 * If you still want to use yield(), do not!
8593 void __sched yield(void)
8595 set_current_state(TASK_RUNNING);
8598 EXPORT_SYMBOL(yield);
8601 * yield_to - yield the current processor to another thread in
8602 * your thread group, or accelerate that thread toward the
8603 * processor it's on.
8605 * @preempt: whether task preemption is allowed or not
8607 * It's the caller's job to ensure that the target task struct
8608 * can't go away on us before we can do any checks.
8611 * true (>0) if we indeed boosted the target task.
8612 * false (0) if we failed to boost the target.
8613 * -ESRCH if there's no task to yield to.
8615 int __sched yield_to(struct task_struct *p, bool preempt)
8617 struct task_struct *curr = current;
8618 struct rq *rq, *p_rq;
8619 unsigned long flags;
8622 local_irq_save(flags);
8628 * If we're the only runnable task on the rq and target rq also
8629 * has only one task, there's absolutely no point in yielding.
8631 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8636 double_rq_lock(rq, p_rq);
8637 if (task_rq(p) != p_rq) {
8638 double_rq_unlock(rq, p_rq);
8642 if (!curr->sched_class->yield_to_task)
8645 if (curr->sched_class != p->sched_class)
8648 if (task_running(p_rq, p) || !task_is_running(p))
8651 yielded = curr->sched_class->yield_to_task(rq, p);
8653 schedstat_inc(rq->yld_count);
8655 * Make p's CPU reschedule; pick_next_entity takes care of
8658 if (preempt && rq != p_rq)
8663 double_rq_unlock(rq, p_rq);
8665 local_irq_restore(flags);
8672 EXPORT_SYMBOL_GPL(yield_to);
8674 int io_schedule_prepare(void)
8676 int old_iowait = current->in_iowait;
8678 current->in_iowait = 1;
8679 blk_flush_plug(current->plug, true);
8683 void io_schedule_finish(int token)
8685 current->in_iowait = token;
8689 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8690 * that process accounting knows that this is a task in IO wait state.
8692 long __sched io_schedule_timeout(long timeout)
8697 token = io_schedule_prepare();
8698 ret = schedule_timeout(timeout);
8699 io_schedule_finish(token);
8703 EXPORT_SYMBOL(io_schedule_timeout);
8705 void __sched io_schedule(void)
8709 token = io_schedule_prepare();
8711 io_schedule_finish(token);
8713 EXPORT_SYMBOL(io_schedule);
8716 * sys_sched_get_priority_max - return maximum RT priority.
8717 * @policy: scheduling class.
8719 * Return: On success, this syscall returns the maximum
8720 * rt_priority that can be used by a given scheduling class.
8721 * On failure, a negative error code is returned.
8723 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8730 ret = MAX_RT_PRIO-1;
8732 case SCHED_DEADLINE:
8743 * sys_sched_get_priority_min - return minimum RT priority.
8744 * @policy: scheduling class.
8746 * Return: On success, this syscall returns the minimum
8747 * rt_priority that can be used by a given scheduling class.
8748 * On failure, a negative error code is returned.
8750 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8759 case SCHED_DEADLINE:
8768 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8770 struct task_struct *p;
8771 unsigned int time_slice;
8781 p = find_process_by_pid(pid);
8785 retval = security_task_getscheduler(p);
8789 rq = task_rq_lock(p, &rf);
8791 if (p->sched_class->get_rr_interval)
8792 time_slice = p->sched_class->get_rr_interval(rq, p);
8793 task_rq_unlock(rq, p, &rf);
8796 jiffies_to_timespec64(time_slice, t);
8805 * sys_sched_rr_get_interval - return the default timeslice of a process.
8806 * @pid: pid of the process.
8807 * @interval: userspace pointer to the timeslice value.
8809 * this syscall writes the default timeslice value of a given process
8810 * into the user-space timespec buffer. A value of '0' means infinity.
8812 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8815 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8816 struct __kernel_timespec __user *, interval)
8818 struct timespec64 t;
8819 int retval = sched_rr_get_interval(pid, &t);
8822 retval = put_timespec64(&t, interval);
8827 #ifdef CONFIG_COMPAT_32BIT_TIME
8828 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8829 struct old_timespec32 __user *, interval)
8831 struct timespec64 t;
8832 int retval = sched_rr_get_interval(pid, &t);
8835 retval = put_old_timespec32(&t, interval);
8840 void sched_show_task(struct task_struct *p)
8842 unsigned long free = 0;
8845 if (!try_get_task_stack(p))
8848 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8850 if (task_is_running(p))
8851 pr_cont(" running task ");
8852 #ifdef CONFIG_DEBUG_STACK_USAGE
8853 free = stack_not_used(p);
8858 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8860 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8861 free, task_pid_nr(p), ppid,
8862 read_task_thread_flags(p));
8864 print_worker_info(KERN_INFO, p);
8865 print_stop_info(KERN_INFO, p);
8866 show_stack(p, NULL, KERN_INFO);
8869 EXPORT_SYMBOL_GPL(sched_show_task);
8872 state_filter_match(unsigned long state_filter, struct task_struct *p)
8874 unsigned int state = READ_ONCE(p->__state);
8876 /* no filter, everything matches */
8880 /* filter, but doesn't match */
8881 if (!(state & state_filter))
8885 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8888 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8895 void show_state_filter(unsigned int state_filter)
8897 struct task_struct *g, *p;
8900 for_each_process_thread(g, p) {
8902 * reset the NMI-timeout, listing all files on a slow
8903 * console might take a lot of time:
8904 * Also, reset softlockup watchdogs on all CPUs, because
8905 * another CPU might be blocked waiting for us to process
8908 touch_nmi_watchdog();
8909 touch_all_softlockup_watchdogs();
8910 if (state_filter_match(state_filter, p))
8914 #ifdef CONFIG_SCHED_DEBUG
8916 sysrq_sched_debug_show();
8920 * Only show locks if all tasks are dumped:
8923 debug_show_all_locks();
8927 * init_idle - set up an idle thread for a given CPU
8928 * @idle: task in question
8929 * @cpu: CPU the idle task belongs to
8931 * NOTE: this function does not set the idle thread's NEED_RESCHED
8932 * flag, to make booting more robust.
8934 void __init init_idle(struct task_struct *idle, int cpu)
8936 struct rq *rq = cpu_rq(cpu);
8937 unsigned long flags;
8939 __sched_fork(0, idle);
8941 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8942 raw_spin_rq_lock(rq);
8944 idle->__state = TASK_RUNNING;
8945 idle->se.exec_start = sched_clock();
8947 * PF_KTHREAD should already be set at this point; regardless, make it
8948 * look like a proper per-CPU kthread.
8950 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8951 kthread_set_per_cpu(idle, cpu);
8955 * It's possible that init_idle() gets called multiple times on a task,
8956 * in that case do_set_cpus_allowed() will not do the right thing.
8958 * And since this is boot we can forgo the serialization.
8960 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8963 * We're having a chicken and egg problem, even though we are
8964 * holding rq->lock, the CPU isn't yet set to this CPU so the
8965 * lockdep check in task_group() will fail.
8967 * Similar case to sched_fork(). / Alternatively we could
8968 * use task_rq_lock() here and obtain the other rq->lock.
8973 __set_task_cpu(idle, cpu);
8977 rcu_assign_pointer(rq->curr, idle);
8978 idle->on_rq = TASK_ON_RQ_QUEUED;
8982 raw_spin_rq_unlock(rq);
8983 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8985 /* Set the preempt count _outside_ the spinlocks! */
8986 init_idle_preempt_count(idle, cpu);
8989 * The idle tasks have their own, simple scheduling class:
8991 idle->sched_class = &idle_sched_class;
8992 ftrace_graph_init_idle_task(idle, cpu);
8993 vtime_init_idle(idle, cpu);
8995 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9001 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9002 const struct cpumask *trial)
9006 if (cpumask_empty(cur))
9009 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9014 int task_can_attach(struct task_struct *p,
9015 const struct cpumask *cs_cpus_allowed)
9020 * Kthreads which disallow setaffinity shouldn't be moved
9021 * to a new cpuset; we don't want to change their CPU
9022 * affinity and isolating such threads by their set of
9023 * allowed nodes is unnecessary. Thus, cpusets are not
9024 * applicable for such threads. This prevents checking for
9025 * success of set_cpus_allowed_ptr() on all attached tasks
9026 * before cpus_mask may be changed.
9028 if (p->flags & PF_NO_SETAFFINITY) {
9033 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
9035 int cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed);
9037 ret = dl_cpu_busy(cpu, p);
9044 bool sched_smp_initialized __read_mostly;
9046 #ifdef CONFIG_NUMA_BALANCING
9047 /* Migrate current task p to target_cpu */
9048 int migrate_task_to(struct task_struct *p, int target_cpu)
9050 struct migration_arg arg = { p, target_cpu };
9051 int curr_cpu = task_cpu(p);
9053 if (curr_cpu == target_cpu)
9056 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9059 /* TODO: This is not properly updating schedstats */
9061 trace_sched_move_numa(p, curr_cpu, target_cpu);
9062 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9066 * Requeue a task on a given node and accurately track the number of NUMA
9067 * tasks on the runqueues
9069 void sched_setnuma(struct task_struct *p, int nid)
9071 bool queued, running;
9075 rq = task_rq_lock(p, &rf);
9076 queued = task_on_rq_queued(p);
9077 running = task_current(rq, p);
9080 dequeue_task(rq, p, DEQUEUE_SAVE);
9082 put_prev_task(rq, p);
9084 p->numa_preferred_nid = nid;
9087 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9089 set_next_task(rq, p);
9090 task_rq_unlock(rq, p, &rf);
9092 #endif /* CONFIG_NUMA_BALANCING */
9094 #ifdef CONFIG_HOTPLUG_CPU
9096 * Ensure that the idle task is using init_mm right before its CPU goes
9099 void idle_task_exit(void)
9101 struct mm_struct *mm = current->active_mm;
9103 BUG_ON(cpu_online(smp_processor_id()));
9104 BUG_ON(current != this_rq()->idle);
9106 if (mm != &init_mm) {
9107 switch_mm(mm, &init_mm, current);
9108 finish_arch_post_lock_switch();
9111 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9114 static int __balance_push_cpu_stop(void *arg)
9116 struct task_struct *p = arg;
9117 struct rq *rq = this_rq();
9121 raw_spin_lock_irq(&p->pi_lock);
9124 update_rq_clock(rq);
9126 if (task_rq(p) == rq && task_on_rq_queued(p)) {
9127 cpu = select_fallback_rq(rq->cpu, p);
9128 rq = __migrate_task(rq, &rf, p, cpu);
9132 raw_spin_unlock_irq(&p->pi_lock);
9139 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9142 * Ensure we only run per-cpu kthreads once the CPU goes !active.
9144 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9145 * effective when the hotplug motion is down.
9147 static void balance_push(struct rq *rq)
9149 struct task_struct *push_task = rq->curr;
9151 lockdep_assert_rq_held(rq);
9154 * Ensure the thing is persistent until balance_push_set(.on = false);
9156 rq->balance_callback = &balance_push_callback;
9159 * Only active while going offline and when invoked on the outgoing
9162 if (!cpu_dying(rq->cpu) || rq != this_rq())
9166 * Both the cpu-hotplug and stop task are in this case and are
9167 * required to complete the hotplug process.
9169 if (kthread_is_per_cpu(push_task) ||
9170 is_migration_disabled(push_task)) {
9173 * If this is the idle task on the outgoing CPU try to wake
9174 * up the hotplug control thread which might wait for the
9175 * last task to vanish. The rcuwait_active() check is
9176 * accurate here because the waiter is pinned on this CPU
9177 * and can't obviously be running in parallel.
9179 * On RT kernels this also has to check whether there are
9180 * pinned and scheduled out tasks on the runqueue. They
9181 * need to leave the migrate disabled section first.
9183 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9184 rcuwait_active(&rq->hotplug_wait)) {
9185 raw_spin_rq_unlock(rq);
9186 rcuwait_wake_up(&rq->hotplug_wait);
9187 raw_spin_rq_lock(rq);
9192 get_task_struct(push_task);
9194 * Temporarily drop rq->lock such that we can wake-up the stop task.
9195 * Both preemption and IRQs are still disabled.
9197 raw_spin_rq_unlock(rq);
9198 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9199 this_cpu_ptr(&push_work));
9201 * At this point need_resched() is true and we'll take the loop in
9202 * schedule(). The next pick is obviously going to be the stop task
9203 * which kthread_is_per_cpu() and will push this task away.
9205 raw_spin_rq_lock(rq);
9208 static void balance_push_set(int cpu, bool on)
9210 struct rq *rq = cpu_rq(cpu);
9213 rq_lock_irqsave(rq, &rf);
9215 WARN_ON_ONCE(rq->balance_callback);
9216 rq->balance_callback = &balance_push_callback;
9217 } else if (rq->balance_callback == &balance_push_callback) {
9218 rq->balance_callback = NULL;
9220 rq_unlock_irqrestore(rq, &rf);
9224 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9225 * inactive. All tasks which are not per CPU kernel threads are either
9226 * pushed off this CPU now via balance_push() or placed on a different CPU
9227 * during wakeup. Wait until the CPU is quiescent.
9229 static void balance_hotplug_wait(void)
9231 struct rq *rq = this_rq();
9233 rcuwait_wait_event(&rq->hotplug_wait,
9234 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9235 TASK_UNINTERRUPTIBLE);
9240 static inline void balance_push(struct rq *rq)
9244 static inline void balance_push_set(int cpu, bool on)
9248 static inline void balance_hotplug_wait(void)
9252 #endif /* CONFIG_HOTPLUG_CPU */
9254 void set_rq_online(struct rq *rq)
9257 const struct sched_class *class;
9259 cpumask_set_cpu(rq->cpu, rq->rd->online);
9262 for_each_class(class) {
9263 if (class->rq_online)
9264 class->rq_online(rq);
9269 void set_rq_offline(struct rq *rq)
9272 const struct sched_class *class;
9274 for_each_class(class) {
9275 if (class->rq_offline)
9276 class->rq_offline(rq);
9279 cpumask_clear_cpu(rq->cpu, rq->rd->online);
9285 * used to mark begin/end of suspend/resume:
9287 static int num_cpus_frozen;
9290 * Update cpusets according to cpu_active mask. If cpusets are
9291 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9292 * around partition_sched_domains().
9294 * If we come here as part of a suspend/resume, don't touch cpusets because we
9295 * want to restore it back to its original state upon resume anyway.
9297 static void cpuset_cpu_active(void)
9299 if (cpuhp_tasks_frozen) {
9301 * num_cpus_frozen tracks how many CPUs are involved in suspend
9302 * resume sequence. As long as this is not the last online
9303 * operation in the resume sequence, just build a single sched
9304 * domain, ignoring cpusets.
9306 partition_sched_domains(1, NULL, NULL);
9307 if (--num_cpus_frozen)
9310 * This is the last CPU online operation. So fall through and
9311 * restore the original sched domains by considering the
9312 * cpuset configurations.
9314 cpuset_force_rebuild();
9316 cpuset_update_active_cpus();
9319 static int cpuset_cpu_inactive(unsigned int cpu)
9321 if (!cpuhp_tasks_frozen) {
9322 int ret = dl_cpu_busy(cpu, NULL);
9326 cpuset_update_active_cpus();
9329 partition_sched_domains(1, NULL, NULL);
9334 int sched_cpu_activate(unsigned int cpu)
9336 struct rq *rq = cpu_rq(cpu);
9340 * Clear the balance_push callback and prepare to schedule
9343 balance_push_set(cpu, false);
9345 #ifdef CONFIG_SCHED_SMT
9347 * When going up, increment the number of cores with SMT present.
9349 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9350 static_branch_inc_cpuslocked(&sched_smt_present);
9352 set_cpu_active(cpu, true);
9354 if (sched_smp_initialized) {
9355 sched_update_numa(cpu, true);
9356 sched_domains_numa_masks_set(cpu);
9357 cpuset_cpu_active();
9361 * Put the rq online, if not already. This happens:
9363 * 1) In the early boot process, because we build the real domains
9364 * after all CPUs have been brought up.
9366 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9369 rq_lock_irqsave(rq, &rf);
9371 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9374 rq_unlock_irqrestore(rq, &rf);
9379 int sched_cpu_deactivate(unsigned int cpu)
9381 struct rq *rq = cpu_rq(cpu);
9386 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9387 * load balancing when not active
9389 nohz_balance_exit_idle(rq);
9391 set_cpu_active(cpu, false);
9394 * From this point forward, this CPU will refuse to run any task that
9395 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9396 * push those tasks away until this gets cleared, see
9397 * sched_cpu_dying().
9399 balance_push_set(cpu, true);
9402 * We've cleared cpu_active_mask / set balance_push, wait for all
9403 * preempt-disabled and RCU users of this state to go away such that
9404 * all new such users will observe it.
9406 * Specifically, we rely on ttwu to no longer target this CPU, see
9407 * ttwu_queue_cond() and is_cpu_allowed().
9409 * Do sync before park smpboot threads to take care the rcu boost case.
9413 rq_lock_irqsave(rq, &rf);
9415 update_rq_clock(rq);
9416 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9419 rq_unlock_irqrestore(rq, &rf);
9421 #ifdef CONFIG_SCHED_SMT
9423 * When going down, decrement the number of cores with SMT present.
9425 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9426 static_branch_dec_cpuslocked(&sched_smt_present);
9428 sched_core_cpu_deactivate(cpu);
9431 if (!sched_smp_initialized)
9434 sched_update_numa(cpu, false);
9435 ret = cpuset_cpu_inactive(cpu);
9437 balance_push_set(cpu, false);
9438 set_cpu_active(cpu, true);
9439 sched_update_numa(cpu, true);
9442 sched_domains_numa_masks_clear(cpu);
9446 static void sched_rq_cpu_starting(unsigned int cpu)
9448 struct rq *rq = cpu_rq(cpu);
9450 rq->calc_load_update = calc_load_update;
9451 update_max_interval();
9454 int sched_cpu_starting(unsigned int cpu)
9456 sched_core_cpu_starting(cpu);
9457 sched_rq_cpu_starting(cpu);
9458 sched_tick_start(cpu);
9462 #ifdef CONFIG_HOTPLUG_CPU
9465 * Invoked immediately before the stopper thread is invoked to bring the
9466 * CPU down completely. At this point all per CPU kthreads except the
9467 * hotplug thread (current) and the stopper thread (inactive) have been
9468 * either parked or have been unbound from the outgoing CPU. Ensure that
9469 * any of those which might be on the way out are gone.
9471 * If after this point a bound task is being woken on this CPU then the
9472 * responsible hotplug callback has failed to do it's job.
9473 * sched_cpu_dying() will catch it with the appropriate fireworks.
9475 int sched_cpu_wait_empty(unsigned int cpu)
9477 balance_hotplug_wait();
9482 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9483 * might have. Called from the CPU stopper task after ensuring that the
9484 * stopper is the last running task on the CPU, so nr_active count is
9485 * stable. We need to take the teardown thread which is calling this into
9486 * account, so we hand in adjust = 1 to the load calculation.
9488 * Also see the comment "Global load-average calculations".
9490 static void calc_load_migrate(struct rq *rq)
9492 long delta = calc_load_fold_active(rq, 1);
9495 atomic_long_add(delta, &calc_load_tasks);
9498 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9500 struct task_struct *g, *p;
9501 int cpu = cpu_of(rq);
9503 lockdep_assert_rq_held(rq);
9505 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9506 for_each_process_thread(g, p) {
9507 if (task_cpu(p) != cpu)
9510 if (!task_on_rq_queued(p))
9513 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9517 int sched_cpu_dying(unsigned int cpu)
9519 struct rq *rq = cpu_rq(cpu);
9522 /* Handle pending wakeups and then migrate everything off */
9523 sched_tick_stop(cpu);
9525 rq_lock_irqsave(rq, &rf);
9526 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9527 WARN(true, "Dying CPU not properly vacated!");
9528 dump_rq_tasks(rq, KERN_WARNING);
9530 rq_unlock_irqrestore(rq, &rf);
9532 calc_load_migrate(rq);
9533 update_max_interval();
9535 sched_core_cpu_dying(cpu);
9540 void __init sched_init_smp(void)
9542 sched_init_numa(NUMA_NO_NODE);
9545 * There's no userspace yet to cause hotplug operations; hence all the
9546 * CPU masks are stable and all blatant races in the below code cannot
9549 mutex_lock(&sched_domains_mutex);
9550 sched_init_domains(cpu_active_mask);
9551 mutex_unlock(&sched_domains_mutex);
9553 /* Move init over to a non-isolated CPU */
9554 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9556 current->flags &= ~PF_NO_SETAFFINITY;
9557 sched_init_granularity();
9559 init_sched_rt_class();
9560 init_sched_dl_class();
9562 sched_smp_initialized = true;
9565 static int __init migration_init(void)
9567 sched_cpu_starting(smp_processor_id());
9570 early_initcall(migration_init);
9573 void __init sched_init_smp(void)
9575 sched_init_granularity();
9577 #endif /* CONFIG_SMP */
9579 int in_sched_functions(unsigned long addr)
9581 return in_lock_functions(addr) ||
9582 (addr >= (unsigned long)__sched_text_start
9583 && addr < (unsigned long)__sched_text_end);
9586 #ifdef CONFIG_CGROUP_SCHED
9588 * Default task group.
9589 * Every task in system belongs to this group at bootup.
9591 struct task_group root_task_group;
9592 LIST_HEAD(task_groups);
9594 /* Cacheline aligned slab cache for task_group */
9595 static struct kmem_cache *task_group_cache __read_mostly;
9598 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9599 DECLARE_PER_CPU(cpumask_var_t, select_rq_mask);
9601 void __init sched_init(void)
9603 unsigned long ptr = 0;
9606 /* Make sure the linker didn't screw up */
9607 BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9608 &fair_sched_class != &rt_sched_class + 1 ||
9609 &rt_sched_class != &dl_sched_class + 1);
9611 BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9616 #ifdef CONFIG_FAIR_GROUP_SCHED
9617 ptr += 2 * nr_cpu_ids * sizeof(void **);
9619 #ifdef CONFIG_RT_GROUP_SCHED
9620 ptr += 2 * nr_cpu_ids * sizeof(void **);
9623 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9625 #ifdef CONFIG_FAIR_GROUP_SCHED
9626 root_task_group.se = (struct sched_entity **)ptr;
9627 ptr += nr_cpu_ids * sizeof(void **);
9629 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9630 ptr += nr_cpu_ids * sizeof(void **);
9632 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9633 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9634 #endif /* CONFIG_FAIR_GROUP_SCHED */
9635 #ifdef CONFIG_RT_GROUP_SCHED
9636 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9637 ptr += nr_cpu_ids * sizeof(void **);
9639 root_task_group.rt_rq = (struct rt_rq **)ptr;
9640 ptr += nr_cpu_ids * sizeof(void **);
9642 #endif /* CONFIG_RT_GROUP_SCHED */
9644 #ifdef CONFIG_CPUMASK_OFFSTACK
9645 for_each_possible_cpu(i) {
9646 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9647 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9648 per_cpu(select_rq_mask, i) = (cpumask_var_t)kzalloc_node(
9649 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9651 #endif /* CONFIG_CPUMASK_OFFSTACK */
9653 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9656 init_defrootdomain();
9659 #ifdef CONFIG_RT_GROUP_SCHED
9660 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9661 global_rt_period(), global_rt_runtime());
9662 #endif /* CONFIG_RT_GROUP_SCHED */
9664 #ifdef CONFIG_CGROUP_SCHED
9665 task_group_cache = KMEM_CACHE(task_group, 0);
9667 list_add(&root_task_group.list, &task_groups);
9668 INIT_LIST_HEAD(&root_task_group.children);
9669 INIT_LIST_HEAD(&root_task_group.siblings);
9670 autogroup_init(&init_task);
9671 #endif /* CONFIG_CGROUP_SCHED */
9673 for_each_possible_cpu(i) {
9677 raw_spin_lock_init(&rq->__lock);
9679 rq->calc_load_active = 0;
9680 rq->calc_load_update = jiffies + LOAD_FREQ;
9681 init_cfs_rq(&rq->cfs);
9682 init_rt_rq(&rq->rt);
9683 init_dl_rq(&rq->dl);
9684 #ifdef CONFIG_FAIR_GROUP_SCHED
9685 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9686 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9688 * How much CPU bandwidth does root_task_group get?
9690 * In case of task-groups formed thr' the cgroup filesystem, it
9691 * gets 100% of the CPU resources in the system. This overall
9692 * system CPU resource is divided among the tasks of
9693 * root_task_group and its child task-groups in a fair manner,
9694 * based on each entity's (task or task-group's) weight
9695 * (se->load.weight).
9697 * In other words, if root_task_group has 10 tasks of weight
9698 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9699 * then A0's share of the CPU resource is:
9701 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9703 * We achieve this by letting root_task_group's tasks sit
9704 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9706 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9707 #endif /* CONFIG_FAIR_GROUP_SCHED */
9709 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9710 #ifdef CONFIG_RT_GROUP_SCHED
9711 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9716 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9717 rq->balance_callback = &balance_push_callback;
9718 rq->active_balance = 0;
9719 rq->next_balance = jiffies;
9724 rq->avg_idle = 2*sysctl_sched_migration_cost;
9725 rq->wake_stamp = jiffies;
9726 rq->wake_avg_idle = rq->avg_idle;
9727 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9729 INIT_LIST_HEAD(&rq->cfs_tasks);
9731 rq_attach_root(rq, &def_root_domain);
9732 #ifdef CONFIG_NO_HZ_COMMON
9733 rq->last_blocked_load_update_tick = jiffies;
9734 atomic_set(&rq->nohz_flags, 0);
9736 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9738 #ifdef CONFIG_HOTPLUG_CPU
9739 rcuwait_init(&rq->hotplug_wait);
9741 #endif /* CONFIG_SMP */
9743 atomic_set(&rq->nr_iowait, 0);
9745 #ifdef CONFIG_SCHED_CORE
9747 rq->core_pick = NULL;
9748 rq->core_enabled = 0;
9749 rq->core_tree = RB_ROOT;
9750 rq->core_forceidle_count = 0;
9751 rq->core_forceidle_occupation = 0;
9752 rq->core_forceidle_start = 0;
9754 rq->core_cookie = 0UL;
9758 set_load_weight(&init_task, false);
9761 * The boot idle thread does lazy MMU switching as well:
9764 enter_lazy_tlb(&init_mm, current);
9767 * The idle task doesn't need the kthread struct to function, but it
9768 * is dressed up as a per-CPU kthread and thus needs to play the part
9769 * if we want to avoid special-casing it in code that deals with per-CPU
9772 WARN_ON(!set_kthread_struct(current));
9775 * Make us the idle thread. Technically, schedule() should not be
9776 * called from this thread, however somewhere below it might be,
9777 * but because we are the idle thread, we just pick up running again
9778 * when this runqueue becomes "idle".
9780 init_idle(current, smp_processor_id());
9782 calc_load_update = jiffies + LOAD_FREQ;
9785 idle_thread_set_boot_cpu();
9786 balance_push_set(smp_processor_id(), false);
9788 init_sched_fair_class();
9794 preempt_dynamic_init();
9796 scheduler_running = 1;
9799 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9801 void __might_sleep(const char *file, int line)
9803 unsigned int state = get_current_state();
9805 * Blocking primitives will set (and therefore destroy) current->state,
9806 * since we will exit with TASK_RUNNING make sure we enter with it,
9807 * otherwise we will destroy state.
9809 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9810 "do not call blocking ops when !TASK_RUNNING; "
9811 "state=%x set at [<%p>] %pS\n", state,
9812 (void *)current->task_state_change,
9813 (void *)current->task_state_change);
9815 __might_resched(file, line, 0);
9817 EXPORT_SYMBOL(__might_sleep);
9819 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9821 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9824 if (preempt_count() == preempt_offset)
9827 pr_err("Preemption disabled at:");
9828 print_ip_sym(KERN_ERR, ip);
9831 static inline bool resched_offsets_ok(unsigned int offsets)
9833 unsigned int nested = preempt_count();
9835 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9837 return nested == offsets;
9840 void __might_resched(const char *file, int line, unsigned int offsets)
9842 /* Ratelimiting timestamp: */
9843 static unsigned long prev_jiffy;
9845 unsigned long preempt_disable_ip;
9847 /* WARN_ON_ONCE() by default, no rate limit required: */
9850 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9851 !is_idle_task(current) && !current->non_block_count) ||
9852 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9856 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9858 prev_jiffy = jiffies;
9860 /* Save this before calling printk(), since that will clobber it: */
9861 preempt_disable_ip = get_preempt_disable_ip(current);
9863 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9865 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9866 in_atomic(), irqs_disabled(), current->non_block_count,
9867 current->pid, current->comm);
9868 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9869 offsets & MIGHT_RESCHED_PREEMPT_MASK);
9871 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9872 pr_err("RCU nest depth: %d, expected: %u\n",
9873 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9876 if (task_stack_end_corrupted(current))
9877 pr_emerg("Thread overran stack, or stack corrupted\n");
9879 debug_show_held_locks(current);
9880 if (irqs_disabled())
9881 print_irqtrace_events(current);
9883 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9884 preempt_disable_ip);
9887 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9889 EXPORT_SYMBOL(__might_resched);
9891 void __cant_sleep(const char *file, int line, int preempt_offset)
9893 static unsigned long prev_jiffy;
9895 if (irqs_disabled())
9898 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9901 if (preempt_count() > preempt_offset)
9904 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9906 prev_jiffy = jiffies;
9908 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9909 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9910 in_atomic(), irqs_disabled(),
9911 current->pid, current->comm);
9913 debug_show_held_locks(current);
9915 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9917 EXPORT_SYMBOL_GPL(__cant_sleep);
9920 void __cant_migrate(const char *file, int line)
9922 static unsigned long prev_jiffy;
9924 if (irqs_disabled())
9927 if (is_migration_disabled(current))
9930 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9933 if (preempt_count() > 0)
9936 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9938 prev_jiffy = jiffies;
9940 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9941 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9942 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9943 current->pid, current->comm);
9945 debug_show_held_locks(current);
9947 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9949 EXPORT_SYMBOL_GPL(__cant_migrate);
9953 #ifdef CONFIG_MAGIC_SYSRQ
9954 void normalize_rt_tasks(void)
9956 struct task_struct *g, *p;
9957 struct sched_attr attr = {
9958 .sched_policy = SCHED_NORMAL,
9961 read_lock(&tasklist_lock);
9962 for_each_process_thread(g, p) {
9964 * Only normalize user tasks:
9966 if (p->flags & PF_KTHREAD)
9969 p->se.exec_start = 0;
9970 schedstat_set(p->stats.wait_start, 0);
9971 schedstat_set(p->stats.sleep_start, 0);
9972 schedstat_set(p->stats.block_start, 0);
9974 if (!dl_task(p) && !rt_task(p)) {
9976 * Renice negative nice level userspace
9979 if (task_nice(p) < 0)
9980 set_user_nice(p, 0);
9984 __sched_setscheduler(p, &attr, false, false);
9986 read_unlock(&tasklist_lock);
9989 #endif /* CONFIG_MAGIC_SYSRQ */
9991 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9993 * These functions are only useful for the IA64 MCA handling, or kdb.
9995 * They can only be called when the whole system has been
9996 * stopped - every CPU needs to be quiescent, and no scheduling
9997 * activity can take place. Using them for anything else would
9998 * be a serious bug, and as a result, they aren't even visible
9999 * under any other configuration.
10003 * curr_task - return the current task for a given CPU.
10004 * @cpu: the processor in question.
10006 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10008 * Return: The current task for @cpu.
10010 struct task_struct *curr_task(int cpu)
10012 return cpu_curr(cpu);
10015 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
10019 * ia64_set_curr_task - set the current task for a given CPU.
10020 * @cpu: the processor in question.
10021 * @p: the task pointer to set.
10023 * Description: This function must only be used when non-maskable interrupts
10024 * are serviced on a separate stack. It allows the architecture to switch the
10025 * notion of the current task on a CPU in a non-blocking manner. This function
10026 * must be called with all CPU's synchronized, and interrupts disabled, the
10027 * and caller must save the original value of the current task (see
10028 * curr_task() above) and restore that value before reenabling interrupts and
10029 * re-starting the system.
10031 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10033 void ia64_set_curr_task(int cpu, struct task_struct *p)
10040 #ifdef CONFIG_CGROUP_SCHED
10041 /* task_group_lock serializes the addition/removal of task groups */
10042 static DEFINE_SPINLOCK(task_group_lock);
10044 static inline void alloc_uclamp_sched_group(struct task_group *tg,
10045 struct task_group *parent)
10047 #ifdef CONFIG_UCLAMP_TASK_GROUP
10048 enum uclamp_id clamp_id;
10050 for_each_clamp_id(clamp_id) {
10051 uclamp_se_set(&tg->uclamp_req[clamp_id],
10052 uclamp_none(clamp_id), false);
10053 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10058 static void sched_free_group(struct task_group *tg)
10060 free_fair_sched_group(tg);
10061 free_rt_sched_group(tg);
10062 autogroup_free(tg);
10063 kmem_cache_free(task_group_cache, tg);
10066 static void sched_free_group_rcu(struct rcu_head *rcu)
10068 sched_free_group(container_of(rcu, struct task_group, rcu));
10071 static void sched_unregister_group(struct task_group *tg)
10073 unregister_fair_sched_group(tg);
10074 unregister_rt_sched_group(tg);
10076 * We have to wait for yet another RCU grace period to expire, as
10077 * print_cfs_stats() might run concurrently.
10079 call_rcu(&tg->rcu, sched_free_group_rcu);
10082 /* allocate runqueue etc for a new task group */
10083 struct task_group *sched_create_group(struct task_group *parent)
10085 struct task_group *tg;
10087 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10089 return ERR_PTR(-ENOMEM);
10091 if (!alloc_fair_sched_group(tg, parent))
10094 if (!alloc_rt_sched_group(tg, parent))
10097 alloc_uclamp_sched_group(tg, parent);
10102 sched_free_group(tg);
10103 return ERR_PTR(-ENOMEM);
10106 void sched_online_group(struct task_group *tg, struct task_group *parent)
10108 unsigned long flags;
10110 spin_lock_irqsave(&task_group_lock, flags);
10111 list_add_rcu(&tg->list, &task_groups);
10113 /* Root should already exist: */
10116 tg->parent = parent;
10117 INIT_LIST_HEAD(&tg->children);
10118 list_add_rcu(&tg->siblings, &parent->children);
10119 spin_unlock_irqrestore(&task_group_lock, flags);
10121 online_fair_sched_group(tg);
10124 /* rcu callback to free various structures associated with a task group */
10125 static void sched_unregister_group_rcu(struct rcu_head *rhp)
10127 /* Now it should be safe to free those cfs_rqs: */
10128 sched_unregister_group(container_of(rhp, struct task_group, rcu));
10131 void sched_destroy_group(struct task_group *tg)
10133 /* Wait for possible concurrent references to cfs_rqs complete: */
10134 call_rcu(&tg->rcu, sched_unregister_group_rcu);
10137 void sched_release_group(struct task_group *tg)
10139 unsigned long flags;
10142 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10143 * sched_cfs_period_timer()).
10145 * For this to be effective, we have to wait for all pending users of
10146 * this task group to leave their RCU critical section to ensure no new
10147 * user will see our dying task group any more. Specifically ensure
10148 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10150 * We therefore defer calling unregister_fair_sched_group() to
10151 * sched_unregister_group() which is guarantied to get called only after the
10152 * current RCU grace period has expired.
10154 spin_lock_irqsave(&task_group_lock, flags);
10155 list_del_rcu(&tg->list);
10156 list_del_rcu(&tg->siblings);
10157 spin_unlock_irqrestore(&task_group_lock, flags);
10160 static void sched_change_group(struct task_struct *tsk, int type)
10162 struct task_group *tg;
10165 * All callers are synchronized by task_rq_lock(); we do not use RCU
10166 * which is pointless here. Thus, we pass "true" to task_css_check()
10167 * to prevent lockdep warnings.
10169 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10170 struct task_group, css);
10171 tg = autogroup_task_group(tsk, tg);
10172 tsk->sched_task_group = tg;
10174 #ifdef CONFIG_FAIR_GROUP_SCHED
10175 if (tsk->sched_class->task_change_group)
10176 tsk->sched_class->task_change_group(tsk, type);
10179 set_task_rq(tsk, task_cpu(tsk));
10183 * Change task's runqueue when it moves between groups.
10185 * The caller of this function should have put the task in its new group by
10186 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10189 void sched_move_task(struct task_struct *tsk)
10191 int queued, running, queue_flags =
10192 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10193 struct rq_flags rf;
10196 rq = task_rq_lock(tsk, &rf);
10197 update_rq_clock(rq);
10199 running = task_current(rq, tsk);
10200 queued = task_on_rq_queued(tsk);
10203 dequeue_task(rq, tsk, queue_flags);
10205 put_prev_task(rq, tsk);
10207 sched_change_group(tsk, TASK_MOVE_GROUP);
10210 enqueue_task(rq, tsk, queue_flags);
10212 set_next_task(rq, tsk);
10214 * After changing group, the running task may have joined a
10215 * throttled one but it's still the running task. Trigger a
10216 * resched to make sure that task can still run.
10221 task_rq_unlock(rq, tsk, &rf);
10224 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10226 return css ? container_of(css, struct task_group, css) : NULL;
10229 static struct cgroup_subsys_state *
10230 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10232 struct task_group *parent = css_tg(parent_css);
10233 struct task_group *tg;
10236 /* This is early initialization for the top cgroup */
10237 return &root_task_group.css;
10240 tg = sched_create_group(parent);
10242 return ERR_PTR(-ENOMEM);
10247 /* Expose task group only after completing cgroup initialization */
10248 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10250 struct task_group *tg = css_tg(css);
10251 struct task_group *parent = css_tg(css->parent);
10254 sched_online_group(tg, parent);
10256 #ifdef CONFIG_UCLAMP_TASK_GROUP
10257 /* Propagate the effective uclamp value for the new group */
10258 mutex_lock(&uclamp_mutex);
10260 cpu_util_update_eff(css);
10262 mutex_unlock(&uclamp_mutex);
10268 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10270 struct task_group *tg = css_tg(css);
10272 sched_release_group(tg);
10275 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10277 struct task_group *tg = css_tg(css);
10280 * Relies on the RCU grace period between css_released() and this.
10282 sched_unregister_group(tg);
10286 * This is called before wake_up_new_task(), therefore we really only
10287 * have to set its group bits, all the other stuff does not apply.
10289 static void cpu_cgroup_fork(struct task_struct *task)
10291 struct rq_flags rf;
10294 rq = task_rq_lock(task, &rf);
10296 update_rq_clock(rq);
10297 sched_change_group(task, TASK_SET_GROUP);
10299 task_rq_unlock(rq, task, &rf);
10302 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10304 struct task_struct *task;
10305 struct cgroup_subsys_state *css;
10308 cgroup_taskset_for_each(task, css, tset) {
10309 #ifdef CONFIG_RT_GROUP_SCHED
10310 if (!sched_rt_can_attach(css_tg(css), task))
10314 * Serialize against wake_up_new_task() such that if it's
10315 * running, we're sure to observe its full state.
10317 raw_spin_lock_irq(&task->pi_lock);
10319 * Avoid calling sched_move_task() before wake_up_new_task()
10320 * has happened. This would lead to problems with PELT, due to
10321 * move wanting to detach+attach while we're not attached yet.
10323 if (READ_ONCE(task->__state) == TASK_NEW)
10325 raw_spin_unlock_irq(&task->pi_lock);
10333 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10335 struct task_struct *task;
10336 struct cgroup_subsys_state *css;
10338 cgroup_taskset_for_each(task, css, tset)
10339 sched_move_task(task);
10342 #ifdef CONFIG_UCLAMP_TASK_GROUP
10343 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10345 struct cgroup_subsys_state *top_css = css;
10346 struct uclamp_se *uc_parent = NULL;
10347 struct uclamp_se *uc_se = NULL;
10348 unsigned int eff[UCLAMP_CNT];
10349 enum uclamp_id clamp_id;
10350 unsigned int clamps;
10352 lockdep_assert_held(&uclamp_mutex);
10353 SCHED_WARN_ON(!rcu_read_lock_held());
10355 css_for_each_descendant_pre(css, top_css) {
10356 uc_parent = css_tg(css)->parent
10357 ? css_tg(css)->parent->uclamp : NULL;
10359 for_each_clamp_id(clamp_id) {
10360 /* Assume effective clamps matches requested clamps */
10361 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10362 /* Cap effective clamps with parent's effective clamps */
10364 eff[clamp_id] > uc_parent[clamp_id].value) {
10365 eff[clamp_id] = uc_parent[clamp_id].value;
10368 /* Ensure protection is always capped by limit */
10369 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10371 /* Propagate most restrictive effective clamps */
10373 uc_se = css_tg(css)->uclamp;
10374 for_each_clamp_id(clamp_id) {
10375 if (eff[clamp_id] == uc_se[clamp_id].value)
10377 uc_se[clamp_id].value = eff[clamp_id];
10378 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10379 clamps |= (0x1 << clamp_id);
10382 css = css_rightmost_descendant(css);
10386 /* Immediately update descendants RUNNABLE tasks */
10387 uclamp_update_active_tasks(css);
10392 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10393 * C expression. Since there is no way to convert a macro argument (N) into a
10394 * character constant, use two levels of macros.
10396 #define _POW10(exp) ((unsigned int)1e##exp)
10397 #define POW10(exp) _POW10(exp)
10399 struct uclamp_request {
10400 #define UCLAMP_PERCENT_SHIFT 2
10401 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10407 static inline struct uclamp_request
10408 capacity_from_percent(char *buf)
10410 struct uclamp_request req = {
10411 .percent = UCLAMP_PERCENT_SCALE,
10412 .util = SCHED_CAPACITY_SCALE,
10417 if (strcmp(buf, "max")) {
10418 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10422 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10427 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10428 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10434 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10435 size_t nbytes, loff_t off,
10436 enum uclamp_id clamp_id)
10438 struct uclamp_request req;
10439 struct task_group *tg;
10441 req = capacity_from_percent(buf);
10445 static_branch_enable(&sched_uclamp_used);
10447 mutex_lock(&uclamp_mutex);
10450 tg = css_tg(of_css(of));
10451 if (tg->uclamp_req[clamp_id].value != req.util)
10452 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10455 * Because of not recoverable conversion rounding we keep track of the
10456 * exact requested value
10458 tg->uclamp_pct[clamp_id] = req.percent;
10460 /* Update effective clamps to track the most restrictive value */
10461 cpu_util_update_eff(of_css(of));
10464 mutex_unlock(&uclamp_mutex);
10469 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10470 char *buf, size_t nbytes,
10473 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10476 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10477 char *buf, size_t nbytes,
10480 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10483 static inline void cpu_uclamp_print(struct seq_file *sf,
10484 enum uclamp_id clamp_id)
10486 struct task_group *tg;
10492 tg = css_tg(seq_css(sf));
10493 util_clamp = tg->uclamp_req[clamp_id].value;
10496 if (util_clamp == SCHED_CAPACITY_SCALE) {
10497 seq_puts(sf, "max\n");
10501 percent = tg->uclamp_pct[clamp_id];
10502 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10503 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10506 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10508 cpu_uclamp_print(sf, UCLAMP_MIN);
10512 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10514 cpu_uclamp_print(sf, UCLAMP_MAX);
10517 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10519 #ifdef CONFIG_FAIR_GROUP_SCHED
10520 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10521 struct cftype *cftype, u64 shareval)
10523 if (shareval > scale_load_down(ULONG_MAX))
10524 shareval = MAX_SHARES;
10525 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10528 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10529 struct cftype *cft)
10531 struct task_group *tg = css_tg(css);
10533 return (u64) scale_load_down(tg->shares);
10536 #ifdef CONFIG_CFS_BANDWIDTH
10537 static DEFINE_MUTEX(cfs_constraints_mutex);
10539 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10540 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10541 /* More than 203 days if BW_SHIFT equals 20. */
10542 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10544 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10546 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10549 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10550 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10552 if (tg == &root_task_group)
10556 * Ensure we have at some amount of bandwidth every period. This is
10557 * to prevent reaching a state of large arrears when throttled via
10558 * entity_tick() resulting in prolonged exit starvation.
10560 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10564 * Likewise, bound things on the other side by preventing insane quota
10565 * periods. This also allows us to normalize in computing quota
10568 if (period > max_cfs_quota_period)
10572 * Bound quota to defend quota against overflow during bandwidth shift.
10574 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10577 if (quota != RUNTIME_INF && (burst > quota ||
10578 burst + quota > max_cfs_runtime))
10582 * Prevent race between setting of cfs_rq->runtime_enabled and
10583 * unthrottle_offline_cfs_rqs().
10586 mutex_lock(&cfs_constraints_mutex);
10587 ret = __cfs_schedulable(tg, period, quota);
10591 runtime_enabled = quota != RUNTIME_INF;
10592 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10594 * If we need to toggle cfs_bandwidth_used, off->on must occur
10595 * before making related changes, and on->off must occur afterwards
10597 if (runtime_enabled && !runtime_was_enabled)
10598 cfs_bandwidth_usage_inc();
10599 raw_spin_lock_irq(&cfs_b->lock);
10600 cfs_b->period = ns_to_ktime(period);
10601 cfs_b->quota = quota;
10602 cfs_b->burst = burst;
10604 __refill_cfs_bandwidth_runtime(cfs_b);
10606 /* Restart the period timer (if active) to handle new period expiry: */
10607 if (runtime_enabled)
10608 start_cfs_bandwidth(cfs_b);
10610 raw_spin_unlock_irq(&cfs_b->lock);
10612 for_each_online_cpu(i) {
10613 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10614 struct rq *rq = cfs_rq->rq;
10615 struct rq_flags rf;
10617 rq_lock_irq(rq, &rf);
10618 cfs_rq->runtime_enabled = runtime_enabled;
10619 cfs_rq->runtime_remaining = 0;
10621 if (cfs_rq->throttled)
10622 unthrottle_cfs_rq(cfs_rq);
10623 rq_unlock_irq(rq, &rf);
10625 if (runtime_was_enabled && !runtime_enabled)
10626 cfs_bandwidth_usage_dec();
10628 mutex_unlock(&cfs_constraints_mutex);
10629 cpus_read_unlock();
10634 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10636 u64 quota, period, burst;
10638 period = ktime_to_ns(tg->cfs_bandwidth.period);
10639 burst = tg->cfs_bandwidth.burst;
10640 if (cfs_quota_us < 0)
10641 quota = RUNTIME_INF;
10642 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10643 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10647 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10650 static long tg_get_cfs_quota(struct task_group *tg)
10654 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10657 quota_us = tg->cfs_bandwidth.quota;
10658 do_div(quota_us, NSEC_PER_USEC);
10663 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10665 u64 quota, period, burst;
10667 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10670 period = (u64)cfs_period_us * NSEC_PER_USEC;
10671 quota = tg->cfs_bandwidth.quota;
10672 burst = tg->cfs_bandwidth.burst;
10674 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10677 static long tg_get_cfs_period(struct task_group *tg)
10681 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10682 do_div(cfs_period_us, NSEC_PER_USEC);
10684 return cfs_period_us;
10687 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10689 u64 quota, period, burst;
10691 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10694 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10695 period = ktime_to_ns(tg->cfs_bandwidth.period);
10696 quota = tg->cfs_bandwidth.quota;
10698 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10701 static long tg_get_cfs_burst(struct task_group *tg)
10705 burst_us = tg->cfs_bandwidth.burst;
10706 do_div(burst_us, NSEC_PER_USEC);
10711 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10712 struct cftype *cft)
10714 return tg_get_cfs_quota(css_tg(css));
10717 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10718 struct cftype *cftype, s64 cfs_quota_us)
10720 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10723 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10724 struct cftype *cft)
10726 return tg_get_cfs_period(css_tg(css));
10729 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10730 struct cftype *cftype, u64 cfs_period_us)
10732 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10735 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10736 struct cftype *cft)
10738 return tg_get_cfs_burst(css_tg(css));
10741 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10742 struct cftype *cftype, u64 cfs_burst_us)
10744 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10747 struct cfs_schedulable_data {
10748 struct task_group *tg;
10753 * normalize group quota/period to be quota/max_period
10754 * note: units are usecs
10756 static u64 normalize_cfs_quota(struct task_group *tg,
10757 struct cfs_schedulable_data *d)
10762 period = d->period;
10765 period = tg_get_cfs_period(tg);
10766 quota = tg_get_cfs_quota(tg);
10769 /* note: these should typically be equivalent */
10770 if (quota == RUNTIME_INF || quota == -1)
10771 return RUNTIME_INF;
10773 return to_ratio(period, quota);
10776 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10778 struct cfs_schedulable_data *d = data;
10779 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10780 s64 quota = 0, parent_quota = -1;
10783 quota = RUNTIME_INF;
10785 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10787 quota = normalize_cfs_quota(tg, d);
10788 parent_quota = parent_b->hierarchical_quota;
10791 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10792 * always take the min. On cgroup1, only inherit when no
10795 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10796 quota = min(quota, parent_quota);
10798 if (quota == RUNTIME_INF)
10799 quota = parent_quota;
10800 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10804 cfs_b->hierarchical_quota = quota;
10809 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10812 struct cfs_schedulable_data data = {
10818 if (quota != RUNTIME_INF) {
10819 do_div(data.period, NSEC_PER_USEC);
10820 do_div(data.quota, NSEC_PER_USEC);
10824 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10830 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10832 struct task_group *tg = css_tg(seq_css(sf));
10833 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10835 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10836 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10837 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10839 if (schedstat_enabled() && tg != &root_task_group) {
10840 struct sched_statistics *stats;
10844 for_each_possible_cpu(i) {
10845 stats = __schedstats_from_se(tg->se[i]);
10846 ws += schedstat_val(stats->wait_sum);
10849 seq_printf(sf, "wait_sum %llu\n", ws);
10852 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10853 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10857 #endif /* CONFIG_CFS_BANDWIDTH */
10858 #endif /* CONFIG_FAIR_GROUP_SCHED */
10860 #ifdef CONFIG_RT_GROUP_SCHED
10861 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10862 struct cftype *cft, s64 val)
10864 return sched_group_set_rt_runtime(css_tg(css), val);
10867 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10868 struct cftype *cft)
10870 return sched_group_rt_runtime(css_tg(css));
10873 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10874 struct cftype *cftype, u64 rt_period_us)
10876 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10879 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10880 struct cftype *cft)
10882 return sched_group_rt_period(css_tg(css));
10884 #endif /* CONFIG_RT_GROUP_SCHED */
10886 #ifdef CONFIG_FAIR_GROUP_SCHED
10887 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10888 struct cftype *cft)
10890 return css_tg(css)->idle;
10893 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10894 struct cftype *cft, s64 idle)
10896 return sched_group_set_idle(css_tg(css), idle);
10900 static struct cftype cpu_legacy_files[] = {
10901 #ifdef CONFIG_FAIR_GROUP_SCHED
10904 .read_u64 = cpu_shares_read_u64,
10905 .write_u64 = cpu_shares_write_u64,
10909 .read_s64 = cpu_idle_read_s64,
10910 .write_s64 = cpu_idle_write_s64,
10913 #ifdef CONFIG_CFS_BANDWIDTH
10915 .name = "cfs_quota_us",
10916 .read_s64 = cpu_cfs_quota_read_s64,
10917 .write_s64 = cpu_cfs_quota_write_s64,
10920 .name = "cfs_period_us",
10921 .read_u64 = cpu_cfs_period_read_u64,
10922 .write_u64 = cpu_cfs_period_write_u64,
10925 .name = "cfs_burst_us",
10926 .read_u64 = cpu_cfs_burst_read_u64,
10927 .write_u64 = cpu_cfs_burst_write_u64,
10931 .seq_show = cpu_cfs_stat_show,
10934 #ifdef CONFIG_RT_GROUP_SCHED
10936 .name = "rt_runtime_us",
10937 .read_s64 = cpu_rt_runtime_read,
10938 .write_s64 = cpu_rt_runtime_write,
10941 .name = "rt_period_us",
10942 .read_u64 = cpu_rt_period_read_uint,
10943 .write_u64 = cpu_rt_period_write_uint,
10946 #ifdef CONFIG_UCLAMP_TASK_GROUP
10948 .name = "uclamp.min",
10949 .flags = CFTYPE_NOT_ON_ROOT,
10950 .seq_show = cpu_uclamp_min_show,
10951 .write = cpu_uclamp_min_write,
10954 .name = "uclamp.max",
10955 .flags = CFTYPE_NOT_ON_ROOT,
10956 .seq_show = cpu_uclamp_max_show,
10957 .write = cpu_uclamp_max_write,
10960 { } /* Terminate */
10963 static int cpu_extra_stat_show(struct seq_file *sf,
10964 struct cgroup_subsys_state *css)
10966 #ifdef CONFIG_CFS_BANDWIDTH
10968 struct task_group *tg = css_tg(css);
10969 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10970 u64 throttled_usec, burst_usec;
10972 throttled_usec = cfs_b->throttled_time;
10973 do_div(throttled_usec, NSEC_PER_USEC);
10974 burst_usec = cfs_b->burst_time;
10975 do_div(burst_usec, NSEC_PER_USEC);
10977 seq_printf(sf, "nr_periods %d\n"
10978 "nr_throttled %d\n"
10979 "throttled_usec %llu\n"
10981 "burst_usec %llu\n",
10982 cfs_b->nr_periods, cfs_b->nr_throttled,
10983 throttled_usec, cfs_b->nr_burst, burst_usec);
10989 #ifdef CONFIG_FAIR_GROUP_SCHED
10990 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10991 struct cftype *cft)
10993 struct task_group *tg = css_tg(css);
10994 u64 weight = scale_load_down(tg->shares);
10996 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10999 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11000 struct cftype *cft, u64 weight)
11003 * cgroup weight knobs should use the common MIN, DFL and MAX
11004 * values which are 1, 100 and 10000 respectively. While it loses
11005 * a bit of range on both ends, it maps pretty well onto the shares
11006 * value used by scheduler and the round-trip conversions preserve
11007 * the original value over the entire range.
11009 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11012 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11014 return sched_group_set_shares(css_tg(css), scale_load(weight));
11017 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11018 struct cftype *cft)
11020 unsigned long weight = scale_load_down(css_tg(css)->shares);
11021 int last_delta = INT_MAX;
11024 /* find the closest nice value to the current weight */
11025 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11026 delta = abs(sched_prio_to_weight[prio] - weight);
11027 if (delta >= last_delta)
11029 last_delta = delta;
11032 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11035 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11036 struct cftype *cft, s64 nice)
11038 unsigned long weight;
11041 if (nice < MIN_NICE || nice > MAX_NICE)
11044 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11045 idx = array_index_nospec(idx, 40);
11046 weight = sched_prio_to_weight[idx];
11048 return sched_group_set_shares(css_tg(css), scale_load(weight));
11052 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11053 long period, long quota)
11056 seq_puts(sf, "max");
11058 seq_printf(sf, "%ld", quota);
11060 seq_printf(sf, " %ld\n", period);
11063 /* caller should put the current value in *@periodp before calling */
11064 static int __maybe_unused cpu_period_quota_parse(char *buf,
11065 u64 *periodp, u64 *quotap)
11067 char tok[21]; /* U64_MAX */
11069 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11072 *periodp *= NSEC_PER_USEC;
11074 if (sscanf(tok, "%llu", quotap))
11075 *quotap *= NSEC_PER_USEC;
11076 else if (!strcmp(tok, "max"))
11077 *quotap = RUNTIME_INF;
11084 #ifdef CONFIG_CFS_BANDWIDTH
11085 static int cpu_max_show(struct seq_file *sf, void *v)
11087 struct task_group *tg = css_tg(seq_css(sf));
11089 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11093 static ssize_t cpu_max_write(struct kernfs_open_file *of,
11094 char *buf, size_t nbytes, loff_t off)
11096 struct task_group *tg = css_tg(of_css(of));
11097 u64 period = tg_get_cfs_period(tg);
11098 u64 burst = tg_get_cfs_burst(tg);
11102 ret = cpu_period_quota_parse(buf, &period, "a);
11104 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11105 return ret ?: nbytes;
11109 static struct cftype cpu_files[] = {
11110 #ifdef CONFIG_FAIR_GROUP_SCHED
11113 .flags = CFTYPE_NOT_ON_ROOT,
11114 .read_u64 = cpu_weight_read_u64,
11115 .write_u64 = cpu_weight_write_u64,
11118 .name = "weight.nice",
11119 .flags = CFTYPE_NOT_ON_ROOT,
11120 .read_s64 = cpu_weight_nice_read_s64,
11121 .write_s64 = cpu_weight_nice_write_s64,
11125 .flags = CFTYPE_NOT_ON_ROOT,
11126 .read_s64 = cpu_idle_read_s64,
11127 .write_s64 = cpu_idle_write_s64,
11130 #ifdef CONFIG_CFS_BANDWIDTH
11133 .flags = CFTYPE_NOT_ON_ROOT,
11134 .seq_show = cpu_max_show,
11135 .write = cpu_max_write,
11138 .name = "max.burst",
11139 .flags = CFTYPE_NOT_ON_ROOT,
11140 .read_u64 = cpu_cfs_burst_read_u64,
11141 .write_u64 = cpu_cfs_burst_write_u64,
11144 #ifdef CONFIG_UCLAMP_TASK_GROUP
11146 .name = "uclamp.min",
11147 .flags = CFTYPE_NOT_ON_ROOT,
11148 .seq_show = cpu_uclamp_min_show,
11149 .write = cpu_uclamp_min_write,
11152 .name = "uclamp.max",
11153 .flags = CFTYPE_NOT_ON_ROOT,
11154 .seq_show = cpu_uclamp_max_show,
11155 .write = cpu_uclamp_max_write,
11158 { } /* terminate */
11161 struct cgroup_subsys cpu_cgrp_subsys = {
11162 .css_alloc = cpu_cgroup_css_alloc,
11163 .css_online = cpu_cgroup_css_online,
11164 .css_released = cpu_cgroup_css_released,
11165 .css_free = cpu_cgroup_css_free,
11166 .css_extra_stat_show = cpu_extra_stat_show,
11167 .fork = cpu_cgroup_fork,
11168 .can_attach = cpu_cgroup_can_attach,
11169 .attach = cpu_cgroup_attach,
11170 .legacy_cftypes = cpu_legacy_files,
11171 .dfl_cftypes = cpu_files,
11172 .early_init = true,
11176 #endif /* CONFIG_CGROUP_SCHED */
11178 void dump_cpu_task(int cpu)
11180 pr_info("Task dump for CPU %d:\n", cpu);
11181 sched_show_task(cpu_curr(cpu));
11185 * Nice levels are multiplicative, with a gentle 10% change for every
11186 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11187 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11188 * that remained on nice 0.
11190 * The "10% effect" is relative and cumulative: from _any_ nice level,
11191 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11192 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11193 * If a task goes up by ~10% and another task goes down by ~10% then
11194 * the relative distance between them is ~25%.)
11196 const int sched_prio_to_weight[40] = {
11197 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11198 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11199 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11200 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11201 /* 0 */ 1024, 820, 655, 526, 423,
11202 /* 5 */ 335, 272, 215, 172, 137,
11203 /* 10 */ 110, 87, 70, 56, 45,
11204 /* 15 */ 36, 29, 23, 18, 15,
11208 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11210 * In cases where the weight does not change often, we can use the
11211 * precalculated inverse to speed up arithmetics by turning divisions
11212 * into multiplications:
11214 const u32 sched_prio_to_wmult[40] = {
11215 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11216 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11217 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11218 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11219 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11220 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11221 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11222 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11225 void call_trace_sched_update_nr_running(struct rq *rq, int count)
11227 trace_sched_update_nr_running_tp(rq, count);