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(struct task_struct *p, int cpu)
3808 * Do not complicate things with the async wake_list while the CPU is
3811 if (!cpu_active(cpu))
3814 /* Ensure the task will still be allowed to run on the CPU. */
3815 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3819 * If the CPU does not share cache, then queue the task on the
3820 * remote rqs wakelist to avoid accessing remote data.
3822 if (!cpus_share_cache(smp_processor_id(), cpu))
3825 if (cpu == smp_processor_id())
3829 * If the wakee cpu is idle, or the task is descheduling and the
3830 * only running task on the CPU, then use the wakelist to offload
3831 * the task activation to the idle (or soon-to-be-idle) CPU as
3832 * the current CPU is likely busy. nr_running is checked to
3833 * avoid unnecessary task stacking.
3835 * Note that we can only get here with (wakee) p->on_rq=0,
3836 * p->on_cpu can be whatever, we've done the dequeue, so
3837 * the wakee has been accounted out of ->nr_running.
3839 if (!cpu_rq(cpu)->nr_running)
3845 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3847 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
3848 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3849 __ttwu_queue_wakelist(p, cpu, wake_flags);
3856 #else /* !CONFIG_SMP */
3858 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3863 #endif /* CONFIG_SMP */
3865 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3867 struct rq *rq = cpu_rq(cpu);
3870 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3874 update_rq_clock(rq);
3875 ttwu_do_activate(rq, p, wake_flags, &rf);
3880 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3882 * The caller holds p::pi_lock if p != current or has preemption
3883 * disabled when p == current.
3885 * The rules of PREEMPT_RT saved_state:
3887 * The related locking code always holds p::pi_lock when updating
3888 * p::saved_state, which means the code is fully serialized in both cases.
3890 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3891 * bits set. This allows to distinguish all wakeup scenarios.
3893 static __always_inline
3894 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3896 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3897 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3898 state != TASK_RTLOCK_WAIT);
3901 if (READ_ONCE(p->__state) & state) {
3906 #ifdef CONFIG_PREEMPT_RT
3908 * Saved state preserves the task state across blocking on
3909 * an RT lock. If the state matches, set p::saved_state to
3910 * TASK_RUNNING, but do not wake the task because it waits
3911 * for a lock wakeup. Also indicate success because from
3912 * the regular waker's point of view this has succeeded.
3914 * After acquiring the lock the task will restore p::__state
3915 * from p::saved_state which ensures that the regular
3916 * wakeup is not lost. The restore will also set
3917 * p::saved_state to TASK_RUNNING so any further tests will
3918 * not result in false positives vs. @success
3920 if (p->saved_state & state) {
3921 p->saved_state = TASK_RUNNING;
3929 * Notes on Program-Order guarantees on SMP systems.
3933 * The basic program-order guarantee on SMP systems is that when a task [t]
3934 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3935 * execution on its new CPU [c1].
3937 * For migration (of runnable tasks) this is provided by the following means:
3939 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3940 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3941 * rq(c1)->lock (if not at the same time, then in that order).
3942 * C) LOCK of the rq(c1)->lock scheduling in task
3944 * Release/acquire chaining guarantees that B happens after A and C after B.
3945 * Note: the CPU doing B need not be c0 or c1
3954 * UNLOCK rq(0)->lock
3956 * LOCK rq(0)->lock // orders against CPU0
3958 * UNLOCK rq(0)->lock
3962 * UNLOCK rq(1)->lock
3964 * LOCK rq(1)->lock // orders against CPU2
3967 * UNLOCK rq(1)->lock
3970 * BLOCKING -- aka. SLEEP + WAKEUP
3972 * For blocking we (obviously) need to provide the same guarantee as for
3973 * migration. However the means are completely different as there is no lock
3974 * chain to provide order. Instead we do:
3976 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3977 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3981 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3983 * LOCK rq(0)->lock LOCK X->pi_lock
3986 * smp_store_release(X->on_cpu, 0);
3988 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3994 * X->state = RUNNING
3995 * UNLOCK rq(2)->lock
3997 * LOCK rq(2)->lock // orders against CPU1
4000 * UNLOCK rq(2)->lock
4003 * UNLOCK rq(0)->lock
4006 * However, for wakeups there is a second guarantee we must provide, namely we
4007 * must ensure that CONDITION=1 done by the caller can not be reordered with
4008 * accesses to the task state; see try_to_wake_up() and set_current_state().
4012 * try_to_wake_up - wake up a thread
4013 * @p: the thread to be awakened
4014 * @state: the mask of task states that can be woken
4015 * @wake_flags: wake modifier flags (WF_*)
4017 * Conceptually does:
4019 * If (@state & @p->state) @p->state = TASK_RUNNING.
4021 * If the task was not queued/runnable, also place it back on a runqueue.
4023 * This function is atomic against schedule() which would dequeue the task.
4025 * It issues a full memory barrier before accessing @p->state, see the comment
4026 * with set_current_state().
4028 * Uses p->pi_lock to serialize against concurrent wake-ups.
4030 * Relies on p->pi_lock stabilizing:
4033 * - p->sched_task_group
4034 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4036 * Tries really hard to only take one task_rq(p)->lock for performance.
4037 * Takes rq->lock in:
4038 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4039 * - ttwu_queue() -- new rq, for enqueue of the task;
4040 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4042 * As a consequence we race really badly with just about everything. See the
4043 * many memory barriers and their comments for details.
4045 * Return: %true if @p->state changes (an actual wakeup was done),
4049 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4051 unsigned long flags;
4052 int cpu, success = 0;
4057 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4058 * == smp_processor_id()'. Together this means we can special
4059 * case the whole 'p->on_rq && ttwu_runnable()' case below
4060 * without taking any locks.
4063 * - we rely on Program-Order guarantees for all the ordering,
4064 * - we're serialized against set_special_state() by virtue of
4065 * it disabling IRQs (this allows not taking ->pi_lock).
4067 if (!ttwu_state_match(p, state, &success))
4070 trace_sched_waking(p);
4071 WRITE_ONCE(p->__state, TASK_RUNNING);
4072 trace_sched_wakeup(p);
4077 * If we are going to wake up a thread waiting for CONDITION we
4078 * need to ensure that CONDITION=1 done by the caller can not be
4079 * reordered with p->state check below. This pairs with smp_store_mb()
4080 * in set_current_state() that the waiting thread does.
4082 raw_spin_lock_irqsave(&p->pi_lock, flags);
4083 smp_mb__after_spinlock();
4084 if (!ttwu_state_match(p, state, &success))
4087 trace_sched_waking(p);
4090 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4091 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4092 * in smp_cond_load_acquire() below.
4094 * sched_ttwu_pending() try_to_wake_up()
4095 * STORE p->on_rq = 1 LOAD p->state
4098 * __schedule() (switch to task 'p')
4099 * LOCK rq->lock smp_rmb();
4100 * smp_mb__after_spinlock();
4104 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4106 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4107 * __schedule(). See the comment for smp_mb__after_spinlock().
4109 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4112 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4117 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4118 * possible to, falsely, observe p->on_cpu == 0.
4120 * One must be running (->on_cpu == 1) in order to remove oneself
4121 * from the runqueue.
4123 * __schedule() (switch to task 'p') try_to_wake_up()
4124 * STORE p->on_cpu = 1 LOAD p->on_rq
4127 * __schedule() (put 'p' to sleep)
4128 * LOCK rq->lock smp_rmb();
4129 * smp_mb__after_spinlock();
4130 * STORE p->on_rq = 0 LOAD p->on_cpu
4132 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4133 * __schedule(). See the comment for smp_mb__after_spinlock().
4135 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4136 * schedule()'s deactivate_task() has 'happened' and p will no longer
4137 * care about it's own p->state. See the comment in __schedule().
4139 smp_acquire__after_ctrl_dep();
4142 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4143 * == 0), which means we need to do an enqueue, change p->state to
4144 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4145 * enqueue, such as ttwu_queue_wakelist().
4147 WRITE_ONCE(p->__state, TASK_WAKING);
4150 * If the owning (remote) CPU is still in the middle of schedule() with
4151 * this task as prev, considering queueing p on the remote CPUs wake_list
4152 * which potentially sends an IPI instead of spinning on p->on_cpu to
4153 * let the waker make forward progress. This is safe because IRQs are
4154 * disabled and the IPI will deliver after on_cpu is cleared.
4156 * Ensure we load task_cpu(p) after p->on_cpu:
4158 * set_task_cpu(p, cpu);
4159 * STORE p->cpu = @cpu
4160 * __schedule() (switch to task 'p')
4162 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4163 * STORE p->on_cpu = 1 LOAD p->cpu
4165 * to ensure we observe the correct CPU on which the task is currently
4168 if (smp_load_acquire(&p->on_cpu) &&
4169 ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4173 * If the owning (remote) CPU is still in the middle of schedule() with
4174 * this task as prev, wait until it's done referencing the task.
4176 * Pairs with the smp_store_release() in finish_task().
4178 * This ensures that tasks getting woken will be fully ordered against
4179 * their previous state and preserve Program Order.
4181 smp_cond_load_acquire(&p->on_cpu, !VAL);
4183 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4184 if (task_cpu(p) != cpu) {
4186 delayacct_blkio_end(p);
4187 atomic_dec(&task_rq(p)->nr_iowait);
4190 wake_flags |= WF_MIGRATED;
4191 psi_ttwu_dequeue(p);
4192 set_task_cpu(p, cpu);
4196 #endif /* CONFIG_SMP */
4198 ttwu_queue(p, cpu, wake_flags);
4200 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4203 ttwu_stat(p, task_cpu(p), wake_flags);
4210 * task_call_func - Invoke a function on task in fixed state
4211 * @p: Process for which the function is to be invoked, can be @current.
4212 * @func: Function to invoke.
4213 * @arg: Argument to function.
4215 * Fix the task in it's current state by avoiding wakeups and or rq operations
4216 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4217 * to work out what the state is, if required. Given that @func can be invoked
4218 * with a runqueue lock held, it had better be quite lightweight.
4221 * Whatever @func returns
4223 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4225 struct rq *rq = NULL;
4230 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4232 state = READ_ONCE(p->__state);
4235 * Ensure we load p->on_rq after p->__state, otherwise it would be
4236 * possible to, falsely, observe p->on_rq == 0.
4238 * See try_to_wake_up() for a longer comment.
4243 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4244 * the task is blocked. Make sure to check @state since ttwu() can drop
4245 * locks at the end, see ttwu_queue_wakelist().
4247 if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4248 rq = __task_rq_lock(p, &rf);
4251 * At this point the task is pinned; either:
4252 * - blocked and we're holding off wakeups (pi->lock)
4253 * - woken, and we're holding off enqueue (rq->lock)
4254 * - queued, and we're holding off schedule (rq->lock)
4255 * - running, and we're holding off de-schedule (rq->lock)
4257 * The called function (@func) can use: task_curr(), p->on_rq and
4258 * p->__state to differentiate between these states.
4265 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4270 * cpu_curr_snapshot - Return a snapshot of the currently running task
4271 * @cpu: The CPU on which to snapshot the task.
4273 * Returns the task_struct pointer of the task "currently" running on
4274 * the specified CPU. If the same task is running on that CPU throughout,
4275 * the return value will be a pointer to that task's task_struct structure.
4276 * If the CPU did any context switches even vaguely concurrently with the
4277 * execution of this function, the return value will be a pointer to the
4278 * task_struct structure of a randomly chosen task that was running on
4279 * that CPU somewhere around the time that this function was executing.
4281 * If the specified CPU was offline, the return value is whatever it
4282 * is, perhaps a pointer to the task_struct structure of that CPU's idle
4283 * task, but there is no guarantee. Callers wishing a useful return
4284 * value must take some action to ensure that the specified CPU remains
4285 * online throughout.
4287 * This function executes full memory barriers before and after fetching
4288 * the pointer, which permits the caller to confine this function's fetch
4289 * with respect to the caller's accesses to other shared variables.
4291 struct task_struct *cpu_curr_snapshot(int cpu)
4293 struct task_struct *t;
4295 smp_mb(); /* Pairing determined by caller's synchronization design. */
4296 t = rcu_dereference(cpu_curr(cpu));
4297 smp_mb(); /* Pairing determined by caller's synchronization design. */
4302 * wake_up_process - Wake up a specific process
4303 * @p: The process to be woken up.
4305 * Attempt to wake up the nominated process and move it to the set of runnable
4308 * Return: 1 if the process was woken up, 0 if it was already running.
4310 * This function executes a full memory barrier before accessing the task state.
4312 int wake_up_process(struct task_struct *p)
4314 return try_to_wake_up(p, TASK_NORMAL, 0);
4316 EXPORT_SYMBOL(wake_up_process);
4318 int wake_up_state(struct task_struct *p, unsigned int state)
4320 return try_to_wake_up(p, state, 0);
4324 * Perform scheduler related setup for a newly forked process p.
4325 * p is forked by current.
4327 * __sched_fork() is basic setup used by init_idle() too:
4329 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4334 p->se.exec_start = 0;
4335 p->se.sum_exec_runtime = 0;
4336 p->se.prev_sum_exec_runtime = 0;
4337 p->se.nr_migrations = 0;
4339 INIT_LIST_HEAD(&p->se.group_node);
4341 #ifdef CONFIG_FAIR_GROUP_SCHED
4342 p->se.cfs_rq = NULL;
4345 #ifdef CONFIG_SCHEDSTATS
4346 /* Even if schedstat is disabled, there should not be garbage */
4347 memset(&p->stats, 0, sizeof(p->stats));
4350 RB_CLEAR_NODE(&p->dl.rb_node);
4351 init_dl_task_timer(&p->dl);
4352 init_dl_inactive_task_timer(&p->dl);
4353 __dl_clear_params(p);
4355 INIT_LIST_HEAD(&p->rt.run_list);
4357 p->rt.time_slice = sched_rr_timeslice;
4361 #ifdef CONFIG_PREEMPT_NOTIFIERS
4362 INIT_HLIST_HEAD(&p->preempt_notifiers);
4365 #ifdef CONFIG_COMPACTION
4366 p->capture_control = NULL;
4368 init_numa_balancing(clone_flags, p);
4370 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4371 p->migration_pending = NULL;
4375 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4377 #ifdef CONFIG_NUMA_BALANCING
4379 int sysctl_numa_balancing_mode;
4381 static void __set_numabalancing_state(bool enabled)
4384 static_branch_enable(&sched_numa_balancing);
4386 static_branch_disable(&sched_numa_balancing);
4389 void set_numabalancing_state(bool enabled)
4392 sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4394 sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4395 __set_numabalancing_state(enabled);
4398 #ifdef CONFIG_PROC_SYSCTL
4399 int sysctl_numa_balancing(struct ctl_table *table, int write,
4400 void *buffer, size_t *lenp, loff_t *ppos)
4404 int state = sysctl_numa_balancing_mode;
4406 if (write && !capable(CAP_SYS_ADMIN))
4411 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4415 sysctl_numa_balancing_mode = state;
4416 __set_numabalancing_state(state);
4423 #ifdef CONFIG_SCHEDSTATS
4425 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4427 static void set_schedstats(bool enabled)
4430 static_branch_enable(&sched_schedstats);
4432 static_branch_disable(&sched_schedstats);
4435 void force_schedstat_enabled(void)
4437 if (!schedstat_enabled()) {
4438 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4439 static_branch_enable(&sched_schedstats);
4443 static int __init setup_schedstats(char *str)
4449 if (!strcmp(str, "enable")) {
4450 set_schedstats(true);
4452 } else if (!strcmp(str, "disable")) {
4453 set_schedstats(false);
4458 pr_warn("Unable to parse schedstats=\n");
4462 __setup("schedstats=", setup_schedstats);
4464 #ifdef CONFIG_PROC_SYSCTL
4465 static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4466 size_t *lenp, loff_t *ppos)
4470 int state = static_branch_likely(&sched_schedstats);
4472 if (write && !capable(CAP_SYS_ADMIN))
4477 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4481 set_schedstats(state);
4484 #endif /* CONFIG_PROC_SYSCTL */
4485 #endif /* CONFIG_SCHEDSTATS */
4487 #ifdef CONFIG_SYSCTL
4488 static struct ctl_table sched_core_sysctls[] = {
4489 #ifdef CONFIG_SCHEDSTATS
4491 .procname = "sched_schedstats",
4493 .maxlen = sizeof(unsigned int),
4495 .proc_handler = sysctl_schedstats,
4496 .extra1 = SYSCTL_ZERO,
4497 .extra2 = SYSCTL_ONE,
4499 #endif /* CONFIG_SCHEDSTATS */
4500 #ifdef CONFIG_UCLAMP_TASK
4502 .procname = "sched_util_clamp_min",
4503 .data = &sysctl_sched_uclamp_util_min,
4504 .maxlen = sizeof(unsigned int),
4506 .proc_handler = sysctl_sched_uclamp_handler,
4509 .procname = "sched_util_clamp_max",
4510 .data = &sysctl_sched_uclamp_util_max,
4511 .maxlen = sizeof(unsigned int),
4513 .proc_handler = sysctl_sched_uclamp_handler,
4516 .procname = "sched_util_clamp_min_rt_default",
4517 .data = &sysctl_sched_uclamp_util_min_rt_default,
4518 .maxlen = sizeof(unsigned int),
4520 .proc_handler = sysctl_sched_uclamp_handler,
4522 #endif /* CONFIG_UCLAMP_TASK */
4525 static int __init sched_core_sysctl_init(void)
4527 register_sysctl_init("kernel", sched_core_sysctls);
4530 late_initcall(sched_core_sysctl_init);
4531 #endif /* CONFIG_SYSCTL */
4534 * fork()/clone()-time setup:
4536 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4538 __sched_fork(clone_flags, p);
4540 * We mark the process as NEW here. This guarantees that
4541 * nobody will actually run it, and a signal or other external
4542 * event cannot wake it up and insert it on the runqueue either.
4544 p->__state = TASK_NEW;
4547 * Make sure we do not leak PI boosting priority to the child.
4549 p->prio = current->normal_prio;
4554 * Revert to default priority/policy on fork if requested.
4556 if (unlikely(p->sched_reset_on_fork)) {
4557 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4558 p->policy = SCHED_NORMAL;
4559 p->static_prio = NICE_TO_PRIO(0);
4561 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4562 p->static_prio = NICE_TO_PRIO(0);
4564 p->prio = p->normal_prio = p->static_prio;
4565 set_load_weight(p, false);
4568 * We don't need the reset flag anymore after the fork. It has
4569 * fulfilled its duty:
4571 p->sched_reset_on_fork = 0;
4574 if (dl_prio(p->prio))
4576 else if (rt_prio(p->prio))
4577 p->sched_class = &rt_sched_class;
4579 p->sched_class = &fair_sched_class;
4581 init_entity_runnable_average(&p->se);
4584 #ifdef CONFIG_SCHED_INFO
4585 if (likely(sched_info_on()))
4586 memset(&p->sched_info, 0, sizeof(p->sched_info));
4588 #if defined(CONFIG_SMP)
4591 init_task_preempt_count(p);
4593 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4594 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4599 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4601 unsigned long flags;
4604 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4605 * required yet, but lockdep gets upset if rules are violated.
4607 raw_spin_lock_irqsave(&p->pi_lock, flags);
4608 #ifdef CONFIG_CGROUP_SCHED
4610 struct task_group *tg;
4611 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4612 struct task_group, css);
4613 tg = autogroup_task_group(p, tg);
4614 p->sched_task_group = tg;
4619 * We're setting the CPU for the first time, we don't migrate,
4620 * so use __set_task_cpu().
4622 __set_task_cpu(p, smp_processor_id());
4623 if (p->sched_class->task_fork)
4624 p->sched_class->task_fork(p);
4625 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4628 void sched_post_fork(struct task_struct *p)
4630 uclamp_post_fork(p);
4633 unsigned long to_ratio(u64 period, u64 runtime)
4635 if (runtime == RUNTIME_INF)
4639 * Doing this here saves a lot of checks in all
4640 * the calling paths, and returning zero seems
4641 * safe for them anyway.
4646 return div64_u64(runtime << BW_SHIFT, period);
4650 * wake_up_new_task - wake up a newly created task for the first time.
4652 * This function will do some initial scheduler statistics housekeeping
4653 * that must be done for every newly created context, then puts the task
4654 * on the runqueue and wakes it.
4656 void wake_up_new_task(struct task_struct *p)
4661 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4662 WRITE_ONCE(p->__state, TASK_RUNNING);
4665 * Fork balancing, do it here and not earlier because:
4666 * - cpus_ptr can change in the fork path
4667 * - any previously selected CPU might disappear through hotplug
4669 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4670 * as we're not fully set-up yet.
4672 p->recent_used_cpu = task_cpu(p);
4674 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4676 rq = __task_rq_lock(p, &rf);
4677 update_rq_clock(rq);
4678 post_init_entity_util_avg(p);
4680 activate_task(rq, p, ENQUEUE_NOCLOCK);
4681 trace_sched_wakeup_new(p);
4682 check_preempt_curr(rq, p, WF_FORK);
4684 if (p->sched_class->task_woken) {
4686 * Nothing relies on rq->lock after this, so it's fine to
4689 rq_unpin_lock(rq, &rf);
4690 p->sched_class->task_woken(rq, p);
4691 rq_repin_lock(rq, &rf);
4694 task_rq_unlock(rq, p, &rf);
4697 #ifdef CONFIG_PREEMPT_NOTIFIERS
4699 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4701 void preempt_notifier_inc(void)
4703 static_branch_inc(&preempt_notifier_key);
4705 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4707 void preempt_notifier_dec(void)
4709 static_branch_dec(&preempt_notifier_key);
4711 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4714 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4715 * @notifier: notifier struct to register
4717 void preempt_notifier_register(struct preempt_notifier *notifier)
4719 if (!static_branch_unlikely(&preempt_notifier_key))
4720 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4722 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4724 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4727 * preempt_notifier_unregister - no longer interested in preemption notifications
4728 * @notifier: notifier struct to unregister
4730 * This is *not* safe to call from within a preemption notifier.
4732 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4734 hlist_del(¬ifier->link);
4736 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4738 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4740 struct preempt_notifier *notifier;
4742 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4743 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4746 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4748 if (static_branch_unlikely(&preempt_notifier_key))
4749 __fire_sched_in_preempt_notifiers(curr);
4753 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4754 struct task_struct *next)
4756 struct preempt_notifier *notifier;
4758 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4759 notifier->ops->sched_out(notifier, next);
4762 static __always_inline void
4763 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4764 struct task_struct *next)
4766 if (static_branch_unlikely(&preempt_notifier_key))
4767 __fire_sched_out_preempt_notifiers(curr, next);
4770 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4772 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4777 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4778 struct task_struct *next)
4782 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4784 static inline void prepare_task(struct task_struct *next)
4788 * Claim the task as running, we do this before switching to it
4789 * such that any running task will have this set.
4791 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4792 * its ordering comment.
4794 WRITE_ONCE(next->on_cpu, 1);
4798 static inline void finish_task(struct task_struct *prev)
4802 * This must be the very last reference to @prev from this CPU. After
4803 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4804 * must ensure this doesn't happen until the switch is completely
4807 * In particular, the load of prev->state in finish_task_switch() must
4808 * happen before this.
4810 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4812 smp_store_release(&prev->on_cpu, 0);
4818 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4820 void (*func)(struct rq *rq);
4821 struct callback_head *next;
4823 lockdep_assert_rq_held(rq);
4826 func = (void (*)(struct rq *))head->func;
4835 static void balance_push(struct rq *rq);
4838 * balance_push_callback is a right abuse of the callback interface and plays
4839 * by significantly different rules.
4841 * Where the normal balance_callback's purpose is to be ran in the same context
4842 * that queued it (only later, when it's safe to drop rq->lock again),
4843 * balance_push_callback is specifically targeted at __schedule().
4845 * This abuse is tolerated because it places all the unlikely/odd cases behind
4846 * a single test, namely: rq->balance_callback == NULL.
4848 struct callback_head balance_push_callback = {
4850 .func = (void (*)(struct callback_head *))balance_push,
4853 static inline struct callback_head *
4854 __splice_balance_callbacks(struct rq *rq, bool split)
4856 struct callback_head *head = rq->balance_callback;
4861 lockdep_assert_rq_held(rq);
4863 * Must not take balance_push_callback off the list when
4864 * splice_balance_callbacks() and balance_callbacks() are not
4865 * in the same rq->lock section.
4867 * In that case it would be possible for __schedule() to interleave
4868 * and observe the list empty.
4870 if (split && head == &balance_push_callback)
4873 rq->balance_callback = NULL;
4878 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4880 return __splice_balance_callbacks(rq, true);
4883 static void __balance_callbacks(struct rq *rq)
4885 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
4888 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4890 unsigned long flags;
4892 if (unlikely(head)) {
4893 raw_spin_rq_lock_irqsave(rq, flags);
4894 do_balance_callbacks(rq, head);
4895 raw_spin_rq_unlock_irqrestore(rq, flags);
4901 static inline void __balance_callbacks(struct rq *rq)
4905 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4910 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4917 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4920 * Since the runqueue lock will be released by the next
4921 * task (which is an invalid locking op but in the case
4922 * of the scheduler it's an obvious special-case), so we
4923 * do an early lockdep release here:
4925 rq_unpin_lock(rq, rf);
4926 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4927 #ifdef CONFIG_DEBUG_SPINLOCK
4928 /* this is a valid case when another task releases the spinlock */
4929 rq_lockp(rq)->owner = next;
4933 static inline void finish_lock_switch(struct rq *rq)
4936 * If we are tracking spinlock dependencies then we have to
4937 * fix up the runqueue lock - which gets 'carried over' from
4938 * prev into current:
4940 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4941 __balance_callbacks(rq);
4942 raw_spin_rq_unlock_irq(rq);
4946 * NOP if the arch has not defined these:
4949 #ifndef prepare_arch_switch
4950 # define prepare_arch_switch(next) do { } while (0)
4953 #ifndef finish_arch_post_lock_switch
4954 # define finish_arch_post_lock_switch() do { } while (0)
4957 static inline void kmap_local_sched_out(void)
4959 #ifdef CONFIG_KMAP_LOCAL
4960 if (unlikely(current->kmap_ctrl.idx))
4961 __kmap_local_sched_out();
4965 static inline void kmap_local_sched_in(void)
4967 #ifdef CONFIG_KMAP_LOCAL
4968 if (unlikely(current->kmap_ctrl.idx))
4969 __kmap_local_sched_in();
4974 * prepare_task_switch - prepare to switch tasks
4975 * @rq: the runqueue preparing to switch
4976 * @prev: the current task that is being switched out
4977 * @next: the task we are going to switch to.
4979 * This is called with the rq lock held and interrupts off. It must
4980 * be paired with a subsequent finish_task_switch after the context
4983 * prepare_task_switch sets up locking and calls architecture specific
4987 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4988 struct task_struct *next)
4990 kcov_prepare_switch(prev);
4991 sched_info_switch(rq, prev, next);
4992 perf_event_task_sched_out(prev, next);
4994 fire_sched_out_preempt_notifiers(prev, next);
4995 kmap_local_sched_out();
4997 prepare_arch_switch(next);
5001 * finish_task_switch - clean up after a task-switch
5002 * @prev: the thread we just switched away from.
5004 * finish_task_switch must be called after the context switch, paired
5005 * with a prepare_task_switch call before the context switch.
5006 * finish_task_switch will reconcile locking set up by prepare_task_switch,
5007 * and do any other architecture-specific cleanup actions.
5009 * Note that we may have delayed dropping an mm in context_switch(). If
5010 * so, we finish that here outside of the runqueue lock. (Doing it
5011 * with the lock held can cause deadlocks; see schedule() for
5014 * The context switch have flipped the stack from under us and restored the
5015 * local variables which were saved when this task called schedule() in the
5016 * past. prev == current is still correct but we need to recalculate this_rq
5017 * because prev may have moved to another CPU.
5019 static struct rq *finish_task_switch(struct task_struct *prev)
5020 __releases(rq->lock)
5022 struct rq *rq = this_rq();
5023 struct mm_struct *mm = rq->prev_mm;
5024 unsigned int prev_state;
5027 * The previous task will have left us with a preempt_count of 2
5028 * because it left us after:
5031 * preempt_disable(); // 1
5033 * raw_spin_lock_irq(&rq->lock) // 2
5035 * Also, see FORK_PREEMPT_COUNT.
5037 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5038 "corrupted preempt_count: %s/%d/0x%x\n",
5039 current->comm, current->pid, preempt_count()))
5040 preempt_count_set(FORK_PREEMPT_COUNT);
5045 * A task struct has one reference for the use as "current".
5046 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5047 * schedule one last time. The schedule call will never return, and
5048 * the scheduled task must drop that reference.
5050 * We must observe prev->state before clearing prev->on_cpu (in
5051 * finish_task), otherwise a concurrent wakeup can get prev
5052 * running on another CPU and we could rave with its RUNNING -> DEAD
5053 * transition, resulting in a double drop.
5055 prev_state = READ_ONCE(prev->__state);
5056 vtime_task_switch(prev);
5057 perf_event_task_sched_in(prev, current);
5059 tick_nohz_task_switch();
5060 finish_lock_switch(rq);
5061 finish_arch_post_lock_switch();
5062 kcov_finish_switch(current);
5064 * kmap_local_sched_out() is invoked with rq::lock held and
5065 * interrupts disabled. There is no requirement for that, but the
5066 * sched out code does not have an interrupt enabled section.
5067 * Restoring the maps on sched in does not require interrupts being
5070 kmap_local_sched_in();
5072 fire_sched_in_preempt_notifiers(current);
5074 * When switching through a kernel thread, the loop in
5075 * membarrier_{private,global}_expedited() may have observed that
5076 * kernel thread and not issued an IPI. It is therefore possible to
5077 * schedule between user->kernel->user threads without passing though
5078 * switch_mm(). Membarrier requires a barrier after storing to
5079 * rq->curr, before returning to userspace, so provide them here:
5081 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5082 * provided by mmdrop(),
5083 * - a sync_core for SYNC_CORE.
5086 membarrier_mm_sync_core_before_usermode(mm);
5089 if (unlikely(prev_state == TASK_DEAD)) {
5090 if (prev->sched_class->task_dead)
5091 prev->sched_class->task_dead(prev);
5093 /* Task is done with its stack. */
5094 put_task_stack(prev);
5096 put_task_struct_rcu_user(prev);
5103 * schedule_tail - first thing a freshly forked thread must call.
5104 * @prev: the thread we just switched away from.
5106 asmlinkage __visible void schedule_tail(struct task_struct *prev)
5107 __releases(rq->lock)
5110 * New tasks start with FORK_PREEMPT_COUNT, see there and
5111 * finish_task_switch() for details.
5113 * finish_task_switch() will drop rq->lock() and lower preempt_count
5114 * and the preempt_enable() will end up enabling preemption (on
5115 * PREEMPT_COUNT kernels).
5118 finish_task_switch(prev);
5121 if (current->set_child_tid)
5122 put_user(task_pid_vnr(current), current->set_child_tid);
5124 calculate_sigpending();
5128 * context_switch - switch to the new MM and the new thread's register state.
5130 static __always_inline struct rq *
5131 context_switch(struct rq *rq, struct task_struct *prev,
5132 struct task_struct *next, struct rq_flags *rf)
5134 prepare_task_switch(rq, prev, next);
5137 * For paravirt, this is coupled with an exit in switch_to to
5138 * combine the page table reload and the switch backend into
5141 arch_start_context_switch(prev);
5144 * kernel -> kernel lazy + transfer active
5145 * user -> kernel lazy + mmgrab() active
5147 * kernel -> user switch + mmdrop() active
5148 * user -> user switch
5150 if (!next->mm) { // to kernel
5151 enter_lazy_tlb(prev->active_mm, next);
5153 next->active_mm = prev->active_mm;
5154 if (prev->mm) // from user
5155 mmgrab(prev->active_mm);
5157 prev->active_mm = NULL;
5159 membarrier_switch_mm(rq, prev->active_mm, next->mm);
5161 * sys_membarrier() requires an smp_mb() between setting
5162 * rq->curr / membarrier_switch_mm() and returning to userspace.
5164 * The below provides this either through switch_mm(), or in
5165 * case 'prev->active_mm == next->mm' through
5166 * finish_task_switch()'s mmdrop().
5168 switch_mm_irqs_off(prev->active_mm, next->mm, next);
5170 if (!prev->mm) { // from kernel
5171 /* will mmdrop() in finish_task_switch(). */
5172 rq->prev_mm = prev->active_mm;
5173 prev->active_mm = NULL;
5177 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5179 prepare_lock_switch(rq, next, rf);
5181 /* Here we just switch the register state and the stack. */
5182 switch_to(prev, next, prev);
5185 return finish_task_switch(prev);
5189 * nr_running and nr_context_switches:
5191 * externally visible scheduler statistics: current number of runnable
5192 * threads, total number of context switches performed since bootup.
5194 unsigned int nr_running(void)
5196 unsigned int i, sum = 0;
5198 for_each_online_cpu(i)
5199 sum += cpu_rq(i)->nr_running;
5205 * Check if only the current task is running on the CPU.
5207 * Caution: this function does not check that the caller has disabled
5208 * preemption, thus the result might have a time-of-check-to-time-of-use
5209 * race. The caller is responsible to use it correctly, for example:
5211 * - from a non-preemptible section (of course)
5213 * - from a thread that is bound to a single CPU
5215 * - in a loop with very short iterations (e.g. a polling loop)
5217 bool single_task_running(void)
5219 return raw_rq()->nr_running == 1;
5221 EXPORT_SYMBOL(single_task_running);
5223 unsigned long long nr_context_switches(void)
5226 unsigned long long sum = 0;
5228 for_each_possible_cpu(i)
5229 sum += cpu_rq(i)->nr_switches;
5235 * Consumers of these two interfaces, like for example the cpuidle menu
5236 * governor, are using nonsensical data. Preferring shallow idle state selection
5237 * for a CPU that has IO-wait which might not even end up running the task when
5238 * it does become runnable.
5241 unsigned int nr_iowait_cpu(int cpu)
5243 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5247 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5249 * The idea behind IO-wait account is to account the idle time that we could
5250 * have spend running if it were not for IO. That is, if we were to improve the
5251 * storage performance, we'd have a proportional reduction in IO-wait time.
5253 * This all works nicely on UP, where, when a task blocks on IO, we account
5254 * idle time as IO-wait, because if the storage were faster, it could've been
5255 * running and we'd not be idle.
5257 * This has been extended to SMP, by doing the same for each CPU. This however
5260 * Imagine for instance the case where two tasks block on one CPU, only the one
5261 * CPU will have IO-wait accounted, while the other has regular idle. Even
5262 * though, if the storage were faster, both could've ran at the same time,
5263 * utilising both CPUs.
5265 * This means, that when looking globally, the current IO-wait accounting on
5266 * SMP is a lower bound, by reason of under accounting.
5268 * Worse, since the numbers are provided per CPU, they are sometimes
5269 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5270 * associated with any one particular CPU, it can wake to another CPU than it
5271 * blocked on. This means the per CPU IO-wait number is meaningless.
5273 * Task CPU affinities can make all that even more 'interesting'.
5276 unsigned int nr_iowait(void)
5278 unsigned int i, sum = 0;
5280 for_each_possible_cpu(i)
5281 sum += nr_iowait_cpu(i);
5289 * sched_exec - execve() is a valuable balancing opportunity, because at
5290 * this point the task has the smallest effective memory and cache footprint.
5292 void sched_exec(void)
5294 struct task_struct *p = current;
5295 unsigned long flags;
5298 raw_spin_lock_irqsave(&p->pi_lock, flags);
5299 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5300 if (dest_cpu == smp_processor_id())
5303 if (likely(cpu_active(dest_cpu))) {
5304 struct migration_arg arg = { p, dest_cpu };
5306 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5307 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5311 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5316 DEFINE_PER_CPU(struct kernel_stat, kstat);
5317 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5319 EXPORT_PER_CPU_SYMBOL(kstat);
5320 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5323 * The function fair_sched_class.update_curr accesses the struct curr
5324 * and its field curr->exec_start; when called from task_sched_runtime(),
5325 * we observe a high rate of cache misses in practice.
5326 * Prefetching this data results in improved performance.
5328 static inline void prefetch_curr_exec_start(struct task_struct *p)
5330 #ifdef CONFIG_FAIR_GROUP_SCHED
5331 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5333 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5336 prefetch(&curr->exec_start);
5340 * Return accounted runtime for the task.
5341 * In case the task is currently running, return the runtime plus current's
5342 * pending runtime that have not been accounted yet.
5344 unsigned long long task_sched_runtime(struct task_struct *p)
5350 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5352 * 64-bit doesn't need locks to atomically read a 64-bit value.
5353 * So we have a optimization chance when the task's delta_exec is 0.
5354 * Reading ->on_cpu is racy, but this is ok.
5356 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5357 * If we race with it entering CPU, unaccounted time is 0. This is
5358 * indistinguishable from the read occurring a few cycles earlier.
5359 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5360 * been accounted, so we're correct here as well.
5362 if (!p->on_cpu || !task_on_rq_queued(p))
5363 return p->se.sum_exec_runtime;
5366 rq = task_rq_lock(p, &rf);
5368 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5369 * project cycles that may never be accounted to this
5370 * thread, breaking clock_gettime().
5372 if (task_current(rq, p) && task_on_rq_queued(p)) {
5373 prefetch_curr_exec_start(p);
5374 update_rq_clock(rq);
5375 p->sched_class->update_curr(rq);
5377 ns = p->se.sum_exec_runtime;
5378 task_rq_unlock(rq, p, &rf);
5383 #ifdef CONFIG_SCHED_DEBUG
5384 static u64 cpu_resched_latency(struct rq *rq)
5386 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5387 u64 resched_latency, now = rq_clock(rq);
5388 static bool warned_once;
5390 if (sysctl_resched_latency_warn_once && warned_once)
5393 if (!need_resched() || !latency_warn_ms)
5396 if (system_state == SYSTEM_BOOTING)
5399 if (!rq->last_seen_need_resched_ns) {
5400 rq->last_seen_need_resched_ns = now;
5401 rq->ticks_without_resched = 0;
5405 rq->ticks_without_resched++;
5406 resched_latency = now - rq->last_seen_need_resched_ns;
5407 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5412 return resched_latency;
5415 static int __init setup_resched_latency_warn_ms(char *str)
5419 if ((kstrtol(str, 0, &val))) {
5420 pr_warn("Unable to set resched_latency_warn_ms\n");
5424 sysctl_resched_latency_warn_ms = val;
5427 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5429 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5430 #endif /* CONFIG_SCHED_DEBUG */
5433 * This function gets called by the timer code, with HZ frequency.
5434 * We call it with interrupts disabled.
5436 void scheduler_tick(void)
5438 int cpu = smp_processor_id();
5439 struct rq *rq = cpu_rq(cpu);
5440 struct task_struct *curr = rq->curr;
5442 unsigned long thermal_pressure;
5443 u64 resched_latency;
5445 arch_scale_freq_tick();
5450 update_rq_clock(rq);
5451 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5452 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5453 curr->sched_class->task_tick(rq, curr, 0);
5454 if (sched_feat(LATENCY_WARN))
5455 resched_latency = cpu_resched_latency(rq);
5456 calc_global_load_tick(rq);
5457 sched_core_tick(rq);
5461 if (sched_feat(LATENCY_WARN) && resched_latency)
5462 resched_latency_warn(cpu, resched_latency);
5464 perf_event_task_tick();
5467 rq->idle_balance = idle_cpu(cpu);
5468 trigger_load_balance(rq);
5472 #ifdef CONFIG_NO_HZ_FULL
5477 struct delayed_work work;
5479 /* Values for ->state, see diagram below. */
5480 #define TICK_SCHED_REMOTE_OFFLINE 0
5481 #define TICK_SCHED_REMOTE_OFFLINING 1
5482 #define TICK_SCHED_REMOTE_RUNNING 2
5485 * State diagram for ->state:
5488 * TICK_SCHED_REMOTE_OFFLINE
5491 * | | sched_tick_remote()
5494 * +--TICK_SCHED_REMOTE_OFFLINING
5497 * sched_tick_start() | | sched_tick_stop()
5500 * TICK_SCHED_REMOTE_RUNNING
5503 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5504 * and sched_tick_start() are happy to leave the state in RUNNING.
5507 static struct tick_work __percpu *tick_work_cpu;
5509 static void sched_tick_remote(struct work_struct *work)
5511 struct delayed_work *dwork = to_delayed_work(work);
5512 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5513 int cpu = twork->cpu;
5514 struct rq *rq = cpu_rq(cpu);
5515 struct task_struct *curr;
5521 * Handle the tick only if it appears the remote CPU is running in full
5522 * dynticks mode. The check is racy by nature, but missing a tick or
5523 * having one too much is no big deal because the scheduler tick updates
5524 * statistics and checks timeslices in a time-independent way, regardless
5525 * of when exactly it is running.
5527 if (!tick_nohz_tick_stopped_cpu(cpu))
5530 rq_lock_irq(rq, &rf);
5532 if (cpu_is_offline(cpu))
5535 update_rq_clock(rq);
5537 if (!is_idle_task(curr)) {
5539 * Make sure the next tick runs within a reasonable
5542 delta = rq_clock_task(rq) - curr->se.exec_start;
5543 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5545 curr->sched_class->task_tick(rq, curr, 0);
5547 calc_load_nohz_remote(rq);
5549 rq_unlock_irq(rq, &rf);
5553 * Run the remote tick once per second (1Hz). This arbitrary
5554 * frequency is large enough to avoid overload but short enough
5555 * to keep scheduler internal stats reasonably up to date. But
5556 * first update state to reflect hotplug activity if required.
5558 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5559 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5560 if (os == TICK_SCHED_REMOTE_RUNNING)
5561 queue_delayed_work(system_unbound_wq, dwork, HZ);
5564 static void sched_tick_start(int cpu)
5567 struct tick_work *twork;
5569 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5572 WARN_ON_ONCE(!tick_work_cpu);
5574 twork = per_cpu_ptr(tick_work_cpu, cpu);
5575 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5576 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5577 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5579 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5580 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5584 #ifdef CONFIG_HOTPLUG_CPU
5585 static void sched_tick_stop(int cpu)
5587 struct tick_work *twork;
5590 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5593 WARN_ON_ONCE(!tick_work_cpu);
5595 twork = per_cpu_ptr(tick_work_cpu, cpu);
5596 /* There cannot be competing actions, but don't rely on stop-machine. */
5597 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5598 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5599 /* Don't cancel, as this would mess up the state machine. */
5601 #endif /* CONFIG_HOTPLUG_CPU */
5603 int __init sched_tick_offload_init(void)
5605 tick_work_cpu = alloc_percpu(struct tick_work);
5606 BUG_ON(!tick_work_cpu);
5610 #else /* !CONFIG_NO_HZ_FULL */
5611 static inline void sched_tick_start(int cpu) { }
5612 static inline void sched_tick_stop(int cpu) { }
5615 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5616 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5618 * If the value passed in is equal to the current preempt count
5619 * then we just disabled preemption. Start timing the latency.
5621 static inline void preempt_latency_start(int val)
5623 if (preempt_count() == val) {
5624 unsigned long ip = get_lock_parent_ip();
5625 #ifdef CONFIG_DEBUG_PREEMPT
5626 current->preempt_disable_ip = ip;
5628 trace_preempt_off(CALLER_ADDR0, ip);
5632 void preempt_count_add(int val)
5634 #ifdef CONFIG_DEBUG_PREEMPT
5638 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5641 __preempt_count_add(val);
5642 #ifdef CONFIG_DEBUG_PREEMPT
5644 * Spinlock count overflowing soon?
5646 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5649 preempt_latency_start(val);
5651 EXPORT_SYMBOL(preempt_count_add);
5652 NOKPROBE_SYMBOL(preempt_count_add);
5655 * If the value passed in equals to the current preempt count
5656 * then we just enabled preemption. Stop timing the latency.
5658 static inline void preempt_latency_stop(int val)
5660 if (preempt_count() == val)
5661 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5664 void preempt_count_sub(int val)
5666 #ifdef CONFIG_DEBUG_PREEMPT
5670 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5673 * Is the spinlock portion underflowing?
5675 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5676 !(preempt_count() & PREEMPT_MASK)))
5680 preempt_latency_stop(val);
5681 __preempt_count_sub(val);
5683 EXPORT_SYMBOL(preempt_count_sub);
5684 NOKPROBE_SYMBOL(preempt_count_sub);
5687 static inline void preempt_latency_start(int val) { }
5688 static inline void preempt_latency_stop(int val) { }
5691 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5693 #ifdef CONFIG_DEBUG_PREEMPT
5694 return p->preempt_disable_ip;
5701 * Print scheduling while atomic bug:
5703 static noinline void __schedule_bug(struct task_struct *prev)
5705 /* Save this before calling printk(), since that will clobber it */
5706 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5708 if (oops_in_progress)
5711 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5712 prev->comm, prev->pid, preempt_count());
5714 debug_show_held_locks(prev);
5716 if (irqs_disabled())
5717 print_irqtrace_events(prev);
5718 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5719 && in_atomic_preempt_off()) {
5720 pr_err("Preemption disabled at:");
5721 print_ip_sym(KERN_ERR, preempt_disable_ip);
5724 panic("scheduling while atomic\n");
5727 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5731 * Various schedule()-time debugging checks and statistics:
5733 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5735 #ifdef CONFIG_SCHED_STACK_END_CHECK
5736 if (task_stack_end_corrupted(prev))
5737 panic("corrupted stack end detected inside scheduler\n");
5739 if (task_scs_end_corrupted(prev))
5740 panic("corrupted shadow stack detected inside scheduler\n");
5743 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5744 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5745 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5746 prev->comm, prev->pid, prev->non_block_count);
5748 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5752 if (unlikely(in_atomic_preempt_off())) {
5753 __schedule_bug(prev);
5754 preempt_count_set(PREEMPT_DISABLED);
5757 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5759 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5761 schedstat_inc(this_rq()->sched_count);
5764 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5765 struct rq_flags *rf)
5768 const struct sched_class *class;
5770 * We must do the balancing pass before put_prev_task(), such
5771 * that when we release the rq->lock the task is in the same
5772 * state as before we took rq->lock.
5774 * We can terminate the balance pass as soon as we know there is
5775 * a runnable task of @class priority or higher.
5777 for_class_range(class, prev->sched_class, &idle_sched_class) {
5778 if (class->balance(rq, prev, rf))
5783 put_prev_task(rq, prev);
5787 * Pick up the highest-prio task:
5789 static inline struct task_struct *
5790 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5792 const struct sched_class *class;
5793 struct task_struct *p;
5796 * Optimization: we know that if all tasks are in the fair class we can
5797 * call that function directly, but only if the @prev task wasn't of a
5798 * higher scheduling class, because otherwise those lose the
5799 * opportunity to pull in more work from other CPUs.
5801 if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
5802 rq->nr_running == rq->cfs.h_nr_running)) {
5804 p = pick_next_task_fair(rq, prev, rf);
5805 if (unlikely(p == RETRY_TASK))
5808 /* Assume the next prioritized class is idle_sched_class */
5810 put_prev_task(rq, prev);
5811 p = pick_next_task_idle(rq);
5818 put_prev_task_balance(rq, prev, rf);
5820 for_each_class(class) {
5821 p = class->pick_next_task(rq);
5826 BUG(); /* The idle class should always have a runnable task. */
5829 #ifdef CONFIG_SCHED_CORE
5830 static inline bool is_task_rq_idle(struct task_struct *t)
5832 return (task_rq(t)->idle == t);
5835 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5837 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5840 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5842 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5845 return a->core_cookie == b->core_cookie;
5848 static inline struct task_struct *pick_task(struct rq *rq)
5850 const struct sched_class *class;
5851 struct task_struct *p;
5853 for_each_class(class) {
5854 p = class->pick_task(rq);
5859 BUG(); /* The idle class should always have a runnable task. */
5862 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5864 static void queue_core_balance(struct rq *rq);
5866 static struct task_struct *
5867 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5869 struct task_struct *next, *p, *max = NULL;
5870 const struct cpumask *smt_mask;
5871 bool fi_before = false;
5872 bool core_clock_updated = (rq == rq->core);
5873 unsigned long cookie;
5874 int i, cpu, occ = 0;
5878 if (!sched_core_enabled(rq))
5879 return __pick_next_task(rq, prev, rf);
5883 /* Stopper task is switching into idle, no need core-wide selection. */
5884 if (cpu_is_offline(cpu)) {
5886 * Reset core_pick so that we don't enter the fastpath when
5887 * coming online. core_pick would already be migrated to
5888 * another cpu during offline.
5890 rq->core_pick = NULL;
5891 return __pick_next_task(rq, prev, rf);
5895 * If there were no {en,de}queues since we picked (IOW, the task
5896 * pointers are all still valid), and we haven't scheduled the last
5897 * pick yet, do so now.
5899 * rq->core_pick can be NULL if no selection was made for a CPU because
5900 * it was either offline or went offline during a sibling's core-wide
5901 * selection. In this case, do a core-wide selection.
5903 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5904 rq->core->core_pick_seq != rq->core_sched_seq &&
5906 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5908 next = rq->core_pick;
5910 put_prev_task(rq, prev);
5911 set_next_task(rq, next);
5914 rq->core_pick = NULL;
5918 put_prev_task_balance(rq, prev, rf);
5920 smt_mask = cpu_smt_mask(cpu);
5921 need_sync = !!rq->core->core_cookie;
5924 rq->core->core_cookie = 0UL;
5925 if (rq->core->core_forceidle_count) {
5926 if (!core_clock_updated) {
5927 update_rq_clock(rq->core);
5928 core_clock_updated = true;
5930 sched_core_account_forceidle(rq);
5931 /* reset after accounting force idle */
5932 rq->core->core_forceidle_start = 0;
5933 rq->core->core_forceidle_count = 0;
5934 rq->core->core_forceidle_occupation = 0;
5940 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5942 * @task_seq guards the task state ({en,de}queues)
5943 * @pick_seq is the @task_seq we did a selection on
5944 * @sched_seq is the @pick_seq we scheduled
5946 * However, preemptions can cause multiple picks on the same task set.
5947 * 'Fix' this by also increasing @task_seq for every pick.
5949 rq->core->core_task_seq++;
5952 * Optimize for common case where this CPU has no cookies
5953 * and there are no cookied tasks running on siblings.
5956 next = pick_task(rq);
5957 if (!next->core_cookie) {
5958 rq->core_pick = NULL;
5960 * For robustness, update the min_vruntime_fi for
5961 * unconstrained picks as well.
5963 WARN_ON_ONCE(fi_before);
5964 task_vruntime_update(rq, next, false);
5970 * For each thread: do the regular task pick and find the max prio task
5973 * Tie-break prio towards the current CPU
5975 for_each_cpu_wrap(i, smt_mask, cpu) {
5979 * Current cpu always has its clock updated on entrance to
5980 * pick_next_task(). If the current cpu is not the core,
5981 * the core may also have been updated above.
5983 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
5984 update_rq_clock(rq_i);
5986 p = rq_i->core_pick = pick_task(rq_i);
5987 if (!max || prio_less(max, p, fi_before))
5991 cookie = rq->core->core_cookie = max->core_cookie;
5994 * For each thread: try and find a runnable task that matches @max or
5997 for_each_cpu(i, smt_mask) {
5999 p = rq_i->core_pick;
6001 if (!cookie_equals(p, cookie)) {
6004 p = sched_core_find(rq_i, cookie);
6006 p = idle_sched_class.pick_task(rq_i);
6009 rq_i->core_pick = p;
6011 if (p == rq_i->idle) {
6012 if (rq_i->nr_running) {
6013 rq->core->core_forceidle_count++;
6015 rq->core->core_forceidle_seq++;
6022 if (schedstat_enabled() && rq->core->core_forceidle_count) {
6023 rq->core->core_forceidle_start = rq_clock(rq->core);
6024 rq->core->core_forceidle_occupation = occ;
6027 rq->core->core_pick_seq = rq->core->core_task_seq;
6028 next = rq->core_pick;
6029 rq->core_sched_seq = rq->core->core_pick_seq;
6031 /* Something should have been selected for current CPU */
6032 WARN_ON_ONCE(!next);
6035 * Reschedule siblings
6037 * NOTE: L1TF -- at this point we're no longer running the old task and
6038 * sending an IPI (below) ensures the sibling will no longer be running
6039 * their task. This ensures there is no inter-sibling overlap between
6040 * non-matching user state.
6042 for_each_cpu(i, smt_mask) {
6046 * An online sibling might have gone offline before a task
6047 * could be picked for it, or it might be offline but later
6048 * happen to come online, but its too late and nothing was
6049 * picked for it. That's Ok - it will pick tasks for itself,
6052 if (!rq_i->core_pick)
6056 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6057 * fi_before fi update?
6063 if (!(fi_before && rq->core->core_forceidle_count))
6064 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6066 rq_i->core_pick->core_occupation = occ;
6069 rq_i->core_pick = NULL;
6073 /* Did we break L1TF mitigation requirements? */
6074 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6076 if (rq_i->curr == rq_i->core_pick) {
6077 rq_i->core_pick = NULL;
6085 set_next_task(rq, next);
6087 if (rq->core->core_forceidle_count && next == rq->idle)
6088 queue_core_balance(rq);
6093 static bool try_steal_cookie(int this, int that)
6095 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6096 struct task_struct *p;
6097 unsigned long cookie;
6098 bool success = false;
6100 local_irq_disable();
6101 double_rq_lock(dst, src);
6103 cookie = dst->core->core_cookie;
6107 if (dst->curr != dst->idle)
6110 p = sched_core_find(src, cookie);
6115 if (p == src->core_pick || p == src->curr)
6118 if (!is_cpu_allowed(p, this))
6121 if (p->core_occupation > dst->idle->core_occupation)
6124 deactivate_task(src, p, 0);
6125 set_task_cpu(p, this);
6126 activate_task(dst, p, 0);
6134 p = sched_core_next(p, cookie);
6138 double_rq_unlock(dst, src);
6144 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6148 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
6155 if (try_steal_cookie(cpu, i))
6162 static void sched_core_balance(struct rq *rq)
6164 struct sched_domain *sd;
6165 int cpu = cpu_of(rq);
6169 raw_spin_rq_unlock_irq(rq);
6170 for_each_domain(cpu, sd) {
6174 if (steal_cookie_task(cpu, sd))
6177 raw_spin_rq_lock_irq(rq);
6182 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
6184 static void queue_core_balance(struct rq *rq)
6186 if (!sched_core_enabled(rq))
6189 if (!rq->core->core_cookie)
6192 if (!rq->nr_running) /* not forced idle */
6195 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6198 static void sched_core_cpu_starting(unsigned int cpu)
6200 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6201 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6202 unsigned long flags;
6205 sched_core_lock(cpu, &flags);
6207 WARN_ON_ONCE(rq->core != rq);
6209 /* if we're the first, we'll be our own leader */
6210 if (cpumask_weight(smt_mask) == 1)
6213 /* find the leader */
6214 for_each_cpu(t, smt_mask) {
6218 if (rq->core == rq) {
6224 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6227 /* install and validate core_rq */
6228 for_each_cpu(t, smt_mask) {
6234 WARN_ON_ONCE(rq->core != core_rq);
6238 sched_core_unlock(cpu, &flags);
6241 static void sched_core_cpu_deactivate(unsigned int cpu)
6243 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6244 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6245 unsigned long flags;
6248 sched_core_lock(cpu, &flags);
6250 /* if we're the last man standing, nothing to do */
6251 if (cpumask_weight(smt_mask) == 1) {
6252 WARN_ON_ONCE(rq->core != rq);
6256 /* if we're not the leader, nothing to do */
6260 /* find a new leader */
6261 for_each_cpu(t, smt_mask) {
6264 core_rq = cpu_rq(t);
6268 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6271 /* copy the shared state to the new leader */
6272 core_rq->core_task_seq = rq->core_task_seq;
6273 core_rq->core_pick_seq = rq->core_pick_seq;
6274 core_rq->core_cookie = rq->core_cookie;
6275 core_rq->core_forceidle_count = rq->core_forceidle_count;
6276 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6277 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6280 * Accounting edge for forced idle is handled in pick_next_task().
6281 * Don't need another one here, since the hotplug thread shouldn't
6284 core_rq->core_forceidle_start = 0;
6286 /* install new leader */
6287 for_each_cpu(t, smt_mask) {
6293 sched_core_unlock(cpu, &flags);
6296 static inline void sched_core_cpu_dying(unsigned int cpu)
6298 struct rq *rq = cpu_rq(cpu);
6304 #else /* !CONFIG_SCHED_CORE */
6306 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6307 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6308 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6310 static struct task_struct *
6311 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6313 return __pick_next_task(rq, prev, rf);
6316 #endif /* CONFIG_SCHED_CORE */
6319 * Constants for the sched_mode argument of __schedule().
6321 * The mode argument allows RT enabled kernels to differentiate a
6322 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6323 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6324 * optimize the AND operation out and just check for zero.
6327 #define SM_PREEMPT 0x1
6328 #define SM_RTLOCK_WAIT 0x2
6330 #ifndef CONFIG_PREEMPT_RT
6331 # define SM_MASK_PREEMPT (~0U)
6333 # define SM_MASK_PREEMPT SM_PREEMPT
6337 * __schedule() is the main scheduler function.
6339 * The main means of driving the scheduler and thus entering this function are:
6341 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6343 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6344 * paths. For example, see arch/x86/entry_64.S.
6346 * To drive preemption between tasks, the scheduler sets the flag in timer
6347 * interrupt handler scheduler_tick().
6349 * 3. Wakeups don't really cause entry into schedule(). They add a
6350 * task to the run-queue and that's it.
6352 * Now, if the new task added to the run-queue preempts the current
6353 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6354 * called on the nearest possible occasion:
6356 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6358 * - in syscall or exception context, at the next outmost
6359 * preempt_enable(). (this might be as soon as the wake_up()'s
6362 * - in IRQ context, return from interrupt-handler to
6363 * preemptible context
6365 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6368 * - cond_resched() call
6369 * - explicit schedule() call
6370 * - return from syscall or exception to user-space
6371 * - return from interrupt-handler to user-space
6373 * WARNING: must be called with preemption disabled!
6375 static void __sched notrace __schedule(unsigned int sched_mode)
6377 struct task_struct *prev, *next;
6378 unsigned long *switch_count;
6379 unsigned long prev_state;
6384 cpu = smp_processor_id();
6388 schedule_debug(prev, !!sched_mode);
6390 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6393 local_irq_disable();
6394 rcu_note_context_switch(!!sched_mode);
6397 * Make sure that signal_pending_state()->signal_pending() below
6398 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6399 * done by the caller to avoid the race with signal_wake_up():
6401 * __set_current_state(@state) signal_wake_up()
6402 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6403 * wake_up_state(p, state)
6404 * LOCK rq->lock LOCK p->pi_state
6405 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6406 * if (signal_pending_state()) if (p->state & @state)
6408 * Also, the membarrier system call requires a full memory barrier
6409 * after coming from user-space, before storing to rq->curr.
6412 smp_mb__after_spinlock();
6414 /* Promote REQ to ACT */
6415 rq->clock_update_flags <<= 1;
6416 update_rq_clock(rq);
6418 switch_count = &prev->nivcsw;
6421 * We must load prev->state once (task_struct::state is volatile), such
6422 * that we form a control dependency vs deactivate_task() below.
6424 prev_state = READ_ONCE(prev->__state);
6425 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6426 if (signal_pending_state(prev_state, prev)) {
6427 WRITE_ONCE(prev->__state, TASK_RUNNING);
6429 prev->sched_contributes_to_load =
6430 (prev_state & TASK_UNINTERRUPTIBLE) &&
6431 !(prev_state & TASK_NOLOAD) &&
6432 !(prev->flags & PF_FROZEN);
6434 if (prev->sched_contributes_to_load)
6435 rq->nr_uninterruptible++;
6438 * __schedule() ttwu()
6439 * prev_state = prev->state; if (p->on_rq && ...)
6440 * if (prev_state) goto out;
6441 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6442 * p->state = TASK_WAKING
6444 * Where __schedule() and ttwu() have matching control dependencies.
6446 * After this, schedule() must not care about p->state any more.
6448 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6450 if (prev->in_iowait) {
6451 atomic_inc(&rq->nr_iowait);
6452 delayacct_blkio_start();
6455 switch_count = &prev->nvcsw;
6458 next = pick_next_task(rq, prev, &rf);
6459 clear_tsk_need_resched(prev);
6460 clear_preempt_need_resched();
6461 #ifdef CONFIG_SCHED_DEBUG
6462 rq->last_seen_need_resched_ns = 0;
6465 if (likely(prev != next)) {
6468 * RCU users of rcu_dereference(rq->curr) may not see
6469 * changes to task_struct made by pick_next_task().
6471 RCU_INIT_POINTER(rq->curr, next);
6473 * The membarrier system call requires each architecture
6474 * to have a full memory barrier after updating
6475 * rq->curr, before returning to user-space.
6477 * Here are the schemes providing that barrier on the
6478 * various architectures:
6479 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6480 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6481 * - finish_lock_switch() for weakly-ordered
6482 * architectures where spin_unlock is a full barrier,
6483 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6484 * is a RELEASE barrier),
6488 migrate_disable_switch(rq, prev);
6489 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6491 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6493 /* Also unlocks the rq: */
6494 rq = context_switch(rq, prev, next, &rf);
6496 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6498 rq_unpin_lock(rq, &rf);
6499 __balance_callbacks(rq);
6500 raw_spin_rq_unlock_irq(rq);
6504 void __noreturn do_task_dead(void)
6506 /* Causes final put_task_struct in finish_task_switch(): */
6507 set_special_state(TASK_DEAD);
6509 /* Tell freezer to ignore us: */
6510 current->flags |= PF_NOFREEZE;
6512 __schedule(SM_NONE);
6515 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6520 static inline void sched_submit_work(struct task_struct *tsk)
6522 unsigned int task_flags;
6524 if (task_is_running(tsk))
6527 task_flags = tsk->flags;
6529 * If a worker goes to sleep, notify and ask workqueue whether it
6530 * wants to wake up a task to maintain concurrency.
6532 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6533 if (task_flags & PF_WQ_WORKER)
6534 wq_worker_sleeping(tsk);
6536 io_wq_worker_sleeping(tsk);
6540 * spinlock and rwlock must not flush block requests. This will
6541 * deadlock if the callback attempts to acquire a lock which is
6544 SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6547 * If we are going to sleep and we have plugged IO queued,
6548 * make sure to submit it to avoid deadlocks.
6550 blk_flush_plug(tsk->plug, true);
6553 static void sched_update_worker(struct task_struct *tsk)
6555 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6556 if (tsk->flags & PF_WQ_WORKER)
6557 wq_worker_running(tsk);
6559 io_wq_worker_running(tsk);
6563 asmlinkage __visible void __sched schedule(void)
6565 struct task_struct *tsk = current;
6567 sched_submit_work(tsk);
6570 __schedule(SM_NONE);
6571 sched_preempt_enable_no_resched();
6572 } while (need_resched());
6573 sched_update_worker(tsk);
6575 EXPORT_SYMBOL(schedule);
6578 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6579 * state (have scheduled out non-voluntarily) by making sure that all
6580 * tasks have either left the run queue or have gone into user space.
6581 * As idle tasks do not do either, they must not ever be preempted
6582 * (schedule out non-voluntarily).
6584 * schedule_idle() is similar to schedule_preempt_disable() except that it
6585 * never enables preemption because it does not call sched_submit_work().
6587 void __sched schedule_idle(void)
6590 * As this skips calling sched_submit_work(), which the idle task does
6591 * regardless because that function is a nop when the task is in a
6592 * TASK_RUNNING state, make sure this isn't used someplace that the
6593 * current task can be in any other state. Note, idle is always in the
6594 * TASK_RUNNING state.
6596 WARN_ON_ONCE(current->__state);
6598 __schedule(SM_NONE);
6599 } while (need_resched());
6602 #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6603 asmlinkage __visible void __sched schedule_user(void)
6606 * If we come here after a random call to set_need_resched(),
6607 * or we have been woken up remotely but the IPI has not yet arrived,
6608 * we haven't yet exited the RCU idle mode. Do it here manually until
6609 * we find a better solution.
6611 * NB: There are buggy callers of this function. Ideally we
6612 * should warn if prev_state != CONTEXT_USER, but that will trigger
6613 * too frequently to make sense yet.
6615 enum ctx_state prev_state = exception_enter();
6617 exception_exit(prev_state);
6622 * schedule_preempt_disabled - called with preemption disabled
6624 * Returns with preemption disabled. Note: preempt_count must be 1
6626 void __sched schedule_preempt_disabled(void)
6628 sched_preempt_enable_no_resched();
6633 #ifdef CONFIG_PREEMPT_RT
6634 void __sched notrace schedule_rtlock(void)
6638 __schedule(SM_RTLOCK_WAIT);
6639 sched_preempt_enable_no_resched();
6640 } while (need_resched());
6642 NOKPROBE_SYMBOL(schedule_rtlock);
6645 static void __sched notrace preempt_schedule_common(void)
6649 * Because the function tracer can trace preempt_count_sub()
6650 * and it also uses preempt_enable/disable_notrace(), if
6651 * NEED_RESCHED is set, the preempt_enable_notrace() called
6652 * by the function tracer will call this function again and
6653 * cause infinite recursion.
6655 * Preemption must be disabled here before the function
6656 * tracer can trace. Break up preempt_disable() into two
6657 * calls. One to disable preemption without fear of being
6658 * traced. The other to still record the preemption latency,
6659 * which can also be traced by the function tracer.
6661 preempt_disable_notrace();
6662 preempt_latency_start(1);
6663 __schedule(SM_PREEMPT);
6664 preempt_latency_stop(1);
6665 preempt_enable_no_resched_notrace();
6668 * Check again in case we missed a preemption opportunity
6669 * between schedule and now.
6671 } while (need_resched());
6674 #ifdef CONFIG_PREEMPTION
6676 * This is the entry point to schedule() from in-kernel preemption
6677 * off of preempt_enable.
6679 asmlinkage __visible void __sched notrace preempt_schedule(void)
6682 * If there is a non-zero preempt_count or interrupts are disabled,
6683 * we do not want to preempt the current task. Just return..
6685 if (likely(!preemptible()))
6687 preempt_schedule_common();
6689 NOKPROBE_SYMBOL(preempt_schedule);
6690 EXPORT_SYMBOL(preempt_schedule);
6692 #ifdef CONFIG_PREEMPT_DYNAMIC
6693 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6694 #ifndef preempt_schedule_dynamic_enabled
6695 #define preempt_schedule_dynamic_enabled preempt_schedule
6696 #define preempt_schedule_dynamic_disabled NULL
6698 DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6699 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6700 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6701 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6702 void __sched notrace dynamic_preempt_schedule(void)
6704 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6708 NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6709 EXPORT_SYMBOL(dynamic_preempt_schedule);
6714 * preempt_schedule_notrace - preempt_schedule called by tracing
6716 * The tracing infrastructure uses preempt_enable_notrace to prevent
6717 * recursion and tracing preempt enabling caused by the tracing
6718 * infrastructure itself. But as tracing can happen in areas coming
6719 * from userspace or just about to enter userspace, a preempt enable
6720 * can occur before user_exit() is called. This will cause the scheduler
6721 * to be called when the system is still in usermode.
6723 * To prevent this, the preempt_enable_notrace will use this function
6724 * instead of preempt_schedule() to exit user context if needed before
6725 * calling the scheduler.
6727 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6729 enum ctx_state prev_ctx;
6731 if (likely(!preemptible()))
6736 * Because the function tracer can trace preempt_count_sub()
6737 * and it also uses preempt_enable/disable_notrace(), if
6738 * NEED_RESCHED is set, the preempt_enable_notrace() called
6739 * by the function tracer will call this function again and
6740 * cause infinite recursion.
6742 * Preemption must be disabled here before the function
6743 * tracer can trace. Break up preempt_disable() into two
6744 * calls. One to disable preemption without fear of being
6745 * traced. The other to still record the preemption latency,
6746 * which can also be traced by the function tracer.
6748 preempt_disable_notrace();
6749 preempt_latency_start(1);
6751 * Needs preempt disabled in case user_exit() is traced
6752 * and the tracer calls preempt_enable_notrace() causing
6753 * an infinite recursion.
6755 prev_ctx = exception_enter();
6756 __schedule(SM_PREEMPT);
6757 exception_exit(prev_ctx);
6759 preempt_latency_stop(1);
6760 preempt_enable_no_resched_notrace();
6761 } while (need_resched());
6763 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6765 #ifdef CONFIG_PREEMPT_DYNAMIC
6766 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6767 #ifndef preempt_schedule_notrace_dynamic_enabled
6768 #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6769 #define preempt_schedule_notrace_dynamic_disabled NULL
6771 DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6772 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6773 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6774 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6775 void __sched notrace dynamic_preempt_schedule_notrace(void)
6777 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6779 preempt_schedule_notrace();
6781 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6782 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6786 #endif /* CONFIG_PREEMPTION */
6789 * This is the entry point to schedule() from kernel preemption
6790 * off of irq context.
6791 * Note, that this is called and return with irqs disabled. This will
6792 * protect us against recursive calling from irq.
6794 asmlinkage __visible void __sched preempt_schedule_irq(void)
6796 enum ctx_state prev_state;
6798 /* Catch callers which need to be fixed */
6799 BUG_ON(preempt_count() || !irqs_disabled());
6801 prev_state = exception_enter();
6806 __schedule(SM_PREEMPT);
6807 local_irq_disable();
6808 sched_preempt_enable_no_resched();
6809 } while (need_resched());
6811 exception_exit(prev_state);
6814 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6817 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6818 return try_to_wake_up(curr->private, mode, wake_flags);
6820 EXPORT_SYMBOL(default_wake_function);
6822 static void __setscheduler_prio(struct task_struct *p, int prio)
6825 p->sched_class = &dl_sched_class;
6826 else if (rt_prio(prio))
6827 p->sched_class = &rt_sched_class;
6829 p->sched_class = &fair_sched_class;
6834 #ifdef CONFIG_RT_MUTEXES
6836 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6839 prio = min(prio, pi_task->prio);
6844 static inline int rt_effective_prio(struct task_struct *p, int prio)
6846 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6848 return __rt_effective_prio(pi_task, prio);
6852 * rt_mutex_setprio - set the current priority of a task
6854 * @pi_task: donor task
6856 * This function changes the 'effective' priority of a task. It does
6857 * not touch ->normal_prio like __setscheduler().
6859 * Used by the rt_mutex code to implement priority inheritance
6860 * logic. Call site only calls if the priority of the task changed.
6862 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6864 int prio, oldprio, queued, running, queue_flag =
6865 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6866 const struct sched_class *prev_class;
6870 /* XXX used to be waiter->prio, not waiter->task->prio */
6871 prio = __rt_effective_prio(pi_task, p->normal_prio);
6874 * If nothing changed; bail early.
6876 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6879 rq = __task_rq_lock(p, &rf);
6880 update_rq_clock(rq);
6882 * Set under pi_lock && rq->lock, such that the value can be used under
6885 * Note that there is loads of tricky to make this pointer cache work
6886 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6887 * ensure a task is de-boosted (pi_task is set to NULL) before the
6888 * task is allowed to run again (and can exit). This ensures the pointer
6889 * points to a blocked task -- which guarantees the task is present.
6891 p->pi_top_task = pi_task;
6894 * For FIFO/RR we only need to set prio, if that matches we're done.
6896 if (prio == p->prio && !dl_prio(prio))
6900 * Idle task boosting is a nono in general. There is one
6901 * exception, when PREEMPT_RT and NOHZ is active:
6903 * The idle task calls get_next_timer_interrupt() and holds
6904 * the timer wheel base->lock on the CPU and another CPU wants
6905 * to access the timer (probably to cancel it). We can safely
6906 * ignore the boosting request, as the idle CPU runs this code
6907 * with interrupts disabled and will complete the lock
6908 * protected section without being interrupted. So there is no
6909 * real need to boost.
6911 if (unlikely(p == rq->idle)) {
6912 WARN_ON(p != rq->curr);
6913 WARN_ON(p->pi_blocked_on);
6917 trace_sched_pi_setprio(p, pi_task);
6920 if (oldprio == prio)
6921 queue_flag &= ~DEQUEUE_MOVE;
6923 prev_class = p->sched_class;
6924 queued = task_on_rq_queued(p);
6925 running = task_current(rq, p);
6927 dequeue_task(rq, p, queue_flag);
6929 put_prev_task(rq, p);
6932 * Boosting condition are:
6933 * 1. -rt task is running and holds mutex A
6934 * --> -dl task blocks on mutex A
6936 * 2. -dl task is running and holds mutex A
6937 * --> -dl task blocks on mutex A and could preempt the
6940 if (dl_prio(prio)) {
6941 if (!dl_prio(p->normal_prio) ||
6942 (pi_task && dl_prio(pi_task->prio) &&
6943 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6944 p->dl.pi_se = pi_task->dl.pi_se;
6945 queue_flag |= ENQUEUE_REPLENISH;
6947 p->dl.pi_se = &p->dl;
6949 } else if (rt_prio(prio)) {
6950 if (dl_prio(oldprio))
6951 p->dl.pi_se = &p->dl;
6953 queue_flag |= ENQUEUE_HEAD;
6955 if (dl_prio(oldprio))
6956 p->dl.pi_se = &p->dl;
6957 if (rt_prio(oldprio))
6961 __setscheduler_prio(p, prio);
6964 enqueue_task(rq, p, queue_flag);
6966 set_next_task(rq, p);
6968 check_class_changed(rq, p, prev_class, oldprio);
6970 /* Avoid rq from going away on us: */
6973 rq_unpin_lock(rq, &rf);
6974 __balance_callbacks(rq);
6975 raw_spin_rq_unlock(rq);
6980 static inline int rt_effective_prio(struct task_struct *p, int prio)
6986 void set_user_nice(struct task_struct *p, long nice)
6988 bool queued, running;
6993 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6996 * We have to be careful, if called from sys_setpriority(),
6997 * the task might be in the middle of scheduling on another CPU.
6999 rq = task_rq_lock(p, &rf);
7000 update_rq_clock(rq);
7003 * The RT priorities are set via sched_setscheduler(), but we still
7004 * allow the 'normal' nice value to be set - but as expected
7005 * it won't have any effect on scheduling until the task is
7006 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7008 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7009 p->static_prio = NICE_TO_PRIO(nice);
7012 queued = task_on_rq_queued(p);
7013 running = task_current(rq, p);
7015 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7017 put_prev_task(rq, p);
7019 p->static_prio = NICE_TO_PRIO(nice);
7020 set_load_weight(p, true);
7022 p->prio = effective_prio(p);
7025 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7027 set_next_task(rq, p);
7030 * If the task increased its priority or is running and
7031 * lowered its priority, then reschedule its CPU:
7033 p->sched_class->prio_changed(rq, p, old_prio);
7036 task_rq_unlock(rq, p, &rf);
7038 EXPORT_SYMBOL(set_user_nice);
7041 * is_nice_reduction - check if nice value is an actual reduction
7043 * Similar to can_nice() but does not perform a capability check.
7048 static bool is_nice_reduction(const struct task_struct *p, const int nice)
7050 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7051 int nice_rlim = nice_to_rlimit(nice);
7053 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7057 * can_nice - check if a task can reduce its nice value
7061 int can_nice(const struct task_struct *p, const int nice)
7063 return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7066 #ifdef __ARCH_WANT_SYS_NICE
7069 * sys_nice - change the priority of the current process.
7070 * @increment: priority increment
7072 * sys_setpriority is a more generic, but much slower function that
7073 * does similar things.
7075 SYSCALL_DEFINE1(nice, int, increment)
7080 * Setpriority might change our priority at the same moment.
7081 * We don't have to worry. Conceptually one call occurs first
7082 * and we have a single winner.
7084 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7085 nice = task_nice(current) + increment;
7087 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7088 if (increment < 0 && !can_nice(current, nice))
7091 retval = security_task_setnice(current, nice);
7095 set_user_nice(current, nice);
7102 * task_prio - return the priority value of a given task.
7103 * @p: the task in question.
7105 * Return: The priority value as seen by users in /proc.
7107 * sched policy return value kernel prio user prio/nice
7109 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7110 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7111 * deadline -101 -1 0
7113 int task_prio(const struct task_struct *p)
7115 return p->prio - MAX_RT_PRIO;
7119 * idle_cpu - is a given CPU idle currently?
7120 * @cpu: the processor in question.
7122 * Return: 1 if the CPU is currently idle. 0 otherwise.
7124 int idle_cpu(int cpu)
7126 struct rq *rq = cpu_rq(cpu);
7128 if (rq->curr != rq->idle)
7135 if (rq->ttwu_pending)
7143 * available_idle_cpu - is a given CPU idle for enqueuing work.
7144 * @cpu: the CPU in question.
7146 * Return: 1 if the CPU is currently idle. 0 otherwise.
7148 int available_idle_cpu(int cpu)
7153 if (vcpu_is_preempted(cpu))
7160 * idle_task - return the idle task for a given CPU.
7161 * @cpu: the processor in question.
7163 * Return: The idle task for the CPU @cpu.
7165 struct task_struct *idle_task(int cpu)
7167 return cpu_rq(cpu)->idle;
7172 * This function computes an effective utilization for the given CPU, to be
7173 * used for frequency selection given the linear relation: f = u * f_max.
7175 * The scheduler tracks the following metrics:
7177 * cpu_util_{cfs,rt,dl,irq}()
7180 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7181 * synchronized windows and are thus directly comparable.
7183 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7184 * which excludes things like IRQ and steal-time. These latter are then accrued
7185 * in the irq utilization.
7187 * The DL bandwidth number otoh is not a measured metric but a value computed
7188 * based on the task model parameters and gives the minimal utilization
7189 * required to meet deadlines.
7191 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7192 enum cpu_util_type type,
7193 struct task_struct *p)
7195 unsigned long dl_util, util, irq, max;
7196 struct rq *rq = cpu_rq(cpu);
7198 max = arch_scale_cpu_capacity(cpu);
7200 if (!uclamp_is_used() &&
7201 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7206 * Early check to see if IRQ/steal time saturates the CPU, can be
7207 * because of inaccuracies in how we track these -- see
7208 * update_irq_load_avg().
7210 irq = cpu_util_irq(rq);
7211 if (unlikely(irq >= max))
7215 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7216 * CFS tasks and we use the same metric to track the effective
7217 * utilization (PELT windows are synchronized) we can directly add them
7218 * to obtain the CPU's actual utilization.
7220 * CFS and RT utilization can be boosted or capped, depending on
7221 * utilization clamp constraints requested by currently RUNNABLE
7223 * When there are no CFS RUNNABLE tasks, clamps are released and
7224 * frequency will be gracefully reduced with the utilization decay.
7226 util = util_cfs + cpu_util_rt(rq);
7227 if (type == FREQUENCY_UTIL)
7228 util = uclamp_rq_util_with(rq, util, p);
7230 dl_util = cpu_util_dl(rq);
7233 * For frequency selection we do not make cpu_util_dl() a permanent part
7234 * of this sum because we want to use cpu_bw_dl() later on, but we need
7235 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7236 * that we select f_max when there is no idle time.
7238 * NOTE: numerical errors or stop class might cause us to not quite hit
7239 * saturation when we should -- something for later.
7241 if (util + dl_util >= max)
7245 * OTOH, for energy computation we need the estimated running time, so
7246 * include util_dl and ignore dl_bw.
7248 if (type == ENERGY_UTIL)
7252 * There is still idle time; further improve the number by using the
7253 * irq metric. Because IRQ/steal time is hidden from the task clock we
7254 * need to scale the task numbers:
7257 * U' = irq + --------- * U
7260 util = scale_irq_capacity(util, irq, max);
7264 * Bandwidth required by DEADLINE must always be granted while, for
7265 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7266 * to gracefully reduce the frequency when no tasks show up for longer
7269 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7270 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7271 * an interface. So, we only do the latter for now.
7273 if (type == FREQUENCY_UTIL)
7274 util += cpu_bw_dl(rq);
7276 return min(max, util);
7279 unsigned long sched_cpu_util(int cpu)
7281 return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
7283 #endif /* CONFIG_SMP */
7286 * find_process_by_pid - find a process with a matching PID value.
7287 * @pid: the pid in question.
7289 * The task of @pid, if found. %NULL otherwise.
7291 static struct task_struct *find_process_by_pid(pid_t pid)
7293 return pid ? find_task_by_vpid(pid) : current;
7297 * sched_setparam() passes in -1 for its policy, to let the functions
7298 * it calls know not to change it.
7300 #define SETPARAM_POLICY -1
7302 static void __setscheduler_params(struct task_struct *p,
7303 const struct sched_attr *attr)
7305 int policy = attr->sched_policy;
7307 if (policy == SETPARAM_POLICY)
7312 if (dl_policy(policy))
7313 __setparam_dl(p, attr);
7314 else if (fair_policy(policy))
7315 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7318 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7319 * !rt_policy. Always setting this ensures that things like
7320 * getparam()/getattr() don't report silly values for !rt tasks.
7322 p->rt_priority = attr->sched_priority;
7323 p->normal_prio = normal_prio(p);
7324 set_load_weight(p, true);
7328 * Check the target process has a UID that matches the current process's:
7330 static bool check_same_owner(struct task_struct *p)
7332 const struct cred *cred = current_cred(), *pcred;
7336 pcred = __task_cred(p);
7337 match = (uid_eq(cred->euid, pcred->euid) ||
7338 uid_eq(cred->euid, pcred->uid));
7344 * Allow unprivileged RT tasks to decrease priority.
7345 * Only issue a capable test if needed and only once to avoid an audit
7346 * event on permitted non-privileged operations:
7348 static int user_check_sched_setscheduler(struct task_struct *p,
7349 const struct sched_attr *attr,
7350 int policy, int reset_on_fork)
7352 if (fair_policy(policy)) {
7353 if (attr->sched_nice < task_nice(p) &&
7354 !is_nice_reduction(p, attr->sched_nice))
7358 if (rt_policy(policy)) {
7359 unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7361 /* Can't set/change the rt policy: */
7362 if (policy != p->policy && !rlim_rtprio)
7365 /* Can't increase priority: */
7366 if (attr->sched_priority > p->rt_priority &&
7367 attr->sched_priority > rlim_rtprio)
7372 * Can't set/change SCHED_DEADLINE policy at all for now
7373 * (safest behavior); in the future we would like to allow
7374 * unprivileged DL tasks to increase their relative deadline
7375 * or reduce their runtime (both ways reducing utilization)
7377 if (dl_policy(policy))
7381 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7382 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7384 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7385 if (!is_nice_reduction(p, task_nice(p)))
7389 /* Can't change other user's priorities: */
7390 if (!check_same_owner(p))
7393 /* Normal users shall not reset the sched_reset_on_fork flag: */
7394 if (p->sched_reset_on_fork && !reset_on_fork)
7400 if (!capable(CAP_SYS_NICE))
7406 static int __sched_setscheduler(struct task_struct *p,
7407 const struct sched_attr *attr,
7410 int oldpolicy = -1, policy = attr->sched_policy;
7411 int retval, oldprio, newprio, queued, running;
7412 const struct sched_class *prev_class;
7413 struct callback_head *head;
7416 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7419 /* The pi code expects interrupts enabled */
7420 BUG_ON(pi && in_interrupt());
7422 /* Double check policy once rq lock held: */
7424 reset_on_fork = p->sched_reset_on_fork;
7425 policy = oldpolicy = p->policy;
7427 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7429 if (!valid_policy(policy))
7433 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7437 * Valid priorities for SCHED_FIFO and SCHED_RR are
7438 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7439 * SCHED_BATCH and SCHED_IDLE is 0.
7441 if (attr->sched_priority > MAX_RT_PRIO-1)
7443 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7444 (rt_policy(policy) != (attr->sched_priority != 0)))
7448 retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7452 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7455 retval = security_task_setscheduler(p);
7460 /* Update task specific "requested" clamps */
7461 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7462 retval = uclamp_validate(p, attr);
7471 * Make sure no PI-waiters arrive (or leave) while we are
7472 * changing the priority of the task:
7474 * To be able to change p->policy safely, the appropriate
7475 * runqueue lock must be held.
7477 rq = task_rq_lock(p, &rf);
7478 update_rq_clock(rq);
7481 * Changing the policy of the stop threads its a very bad idea:
7483 if (p == rq->stop) {
7489 * If not changing anything there's no need to proceed further,
7490 * but store a possible modification of reset_on_fork.
7492 if (unlikely(policy == p->policy)) {
7493 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7495 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7497 if (dl_policy(policy) && dl_param_changed(p, attr))
7499 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7502 p->sched_reset_on_fork = reset_on_fork;
7509 #ifdef CONFIG_RT_GROUP_SCHED
7511 * Do not allow realtime tasks into groups that have no runtime
7514 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7515 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7516 !task_group_is_autogroup(task_group(p))) {
7522 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7523 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7524 cpumask_t *span = rq->rd->span;
7527 * Don't allow tasks with an affinity mask smaller than
7528 * the entire root_domain to become SCHED_DEADLINE. We
7529 * will also fail if there's no bandwidth available.
7531 if (!cpumask_subset(span, p->cpus_ptr) ||
7532 rq->rd->dl_bw.bw == 0) {
7540 /* Re-check policy now with rq lock held: */
7541 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7542 policy = oldpolicy = -1;
7543 task_rq_unlock(rq, p, &rf);
7545 cpuset_read_unlock();
7550 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7551 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7554 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7559 p->sched_reset_on_fork = reset_on_fork;
7562 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7565 * Take priority boosted tasks into account. If the new
7566 * effective priority is unchanged, we just store the new
7567 * normal parameters and do not touch the scheduler class and
7568 * the runqueue. This will be done when the task deboost
7571 newprio = rt_effective_prio(p, newprio);
7572 if (newprio == oldprio)
7573 queue_flags &= ~DEQUEUE_MOVE;
7576 queued = task_on_rq_queued(p);
7577 running = task_current(rq, p);
7579 dequeue_task(rq, p, queue_flags);
7581 put_prev_task(rq, p);
7583 prev_class = p->sched_class;
7585 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7586 __setscheduler_params(p, attr);
7587 __setscheduler_prio(p, newprio);
7589 __setscheduler_uclamp(p, attr);
7593 * We enqueue to tail when the priority of a task is
7594 * increased (user space view).
7596 if (oldprio < p->prio)
7597 queue_flags |= ENQUEUE_HEAD;
7599 enqueue_task(rq, p, queue_flags);
7602 set_next_task(rq, p);
7604 check_class_changed(rq, p, prev_class, oldprio);
7606 /* Avoid rq from going away on us: */
7608 head = splice_balance_callbacks(rq);
7609 task_rq_unlock(rq, p, &rf);
7612 cpuset_read_unlock();
7613 rt_mutex_adjust_pi(p);
7616 /* Run balance callbacks after we've adjusted the PI chain: */
7617 balance_callbacks(rq, head);
7623 task_rq_unlock(rq, p, &rf);
7625 cpuset_read_unlock();
7629 static int _sched_setscheduler(struct task_struct *p, int policy,
7630 const struct sched_param *param, bool check)
7632 struct sched_attr attr = {
7633 .sched_policy = policy,
7634 .sched_priority = param->sched_priority,
7635 .sched_nice = PRIO_TO_NICE(p->static_prio),
7638 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7639 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7640 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7641 policy &= ~SCHED_RESET_ON_FORK;
7642 attr.sched_policy = policy;
7645 return __sched_setscheduler(p, &attr, check, true);
7648 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7649 * @p: the task in question.
7650 * @policy: new policy.
7651 * @param: structure containing the new RT priority.
7653 * Use sched_set_fifo(), read its comment.
7655 * Return: 0 on success. An error code otherwise.
7657 * NOTE that the task may be already dead.
7659 int sched_setscheduler(struct task_struct *p, int policy,
7660 const struct sched_param *param)
7662 return _sched_setscheduler(p, policy, param, true);
7665 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7667 return __sched_setscheduler(p, attr, true, true);
7670 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7672 return __sched_setscheduler(p, attr, false, true);
7674 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7677 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7678 * @p: the task in question.
7679 * @policy: new policy.
7680 * @param: structure containing the new RT priority.
7682 * Just like sched_setscheduler, only don't bother checking if the
7683 * current context has permission. For example, this is needed in
7684 * stop_machine(): we create temporary high priority worker threads,
7685 * but our caller might not have that capability.
7687 * Return: 0 on success. An error code otherwise.
7689 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7690 const struct sched_param *param)
7692 return _sched_setscheduler(p, policy, param, false);
7696 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7697 * incapable of resource management, which is the one thing an OS really should
7700 * This is of course the reason it is limited to privileged users only.
7702 * Worse still; it is fundamentally impossible to compose static priority
7703 * workloads. You cannot take two correctly working static prio workloads
7704 * and smash them together and still expect them to work.
7706 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7710 * The administrator _MUST_ configure the system, the kernel simply doesn't
7711 * know enough information to make a sensible choice.
7713 void sched_set_fifo(struct task_struct *p)
7715 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7716 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7718 EXPORT_SYMBOL_GPL(sched_set_fifo);
7721 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7723 void sched_set_fifo_low(struct task_struct *p)
7725 struct sched_param sp = { .sched_priority = 1 };
7726 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7728 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7730 void sched_set_normal(struct task_struct *p, int nice)
7732 struct sched_attr attr = {
7733 .sched_policy = SCHED_NORMAL,
7736 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7738 EXPORT_SYMBOL_GPL(sched_set_normal);
7741 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7743 struct sched_param lparam;
7744 struct task_struct *p;
7747 if (!param || pid < 0)
7749 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7754 p = find_process_by_pid(pid);
7760 retval = sched_setscheduler(p, policy, &lparam);
7768 * Mimics kernel/events/core.c perf_copy_attr().
7770 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7775 /* Zero the full structure, so that a short copy will be nice: */
7776 memset(attr, 0, sizeof(*attr));
7778 ret = get_user(size, &uattr->size);
7782 /* ABI compatibility quirk: */
7784 size = SCHED_ATTR_SIZE_VER0;
7785 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7788 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7795 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7796 size < SCHED_ATTR_SIZE_VER1)
7800 * XXX: Do we want to be lenient like existing syscalls; or do we want
7801 * to be strict and return an error on out-of-bounds values?
7803 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7808 put_user(sizeof(*attr), &uattr->size);
7812 static void get_params(struct task_struct *p, struct sched_attr *attr)
7814 if (task_has_dl_policy(p))
7815 __getparam_dl(p, attr);
7816 else if (task_has_rt_policy(p))
7817 attr->sched_priority = p->rt_priority;
7819 attr->sched_nice = task_nice(p);
7823 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7824 * @pid: the pid in question.
7825 * @policy: new policy.
7826 * @param: structure containing the new RT priority.
7828 * Return: 0 on success. An error code otherwise.
7830 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7835 return do_sched_setscheduler(pid, policy, param);
7839 * sys_sched_setparam - set/change the RT priority of a thread
7840 * @pid: the pid in question.
7841 * @param: structure containing the new RT priority.
7843 * Return: 0 on success. An error code otherwise.
7845 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7847 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7851 * sys_sched_setattr - same as above, but with extended sched_attr
7852 * @pid: the pid in question.
7853 * @uattr: structure containing the extended parameters.
7854 * @flags: for future extension.
7856 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7857 unsigned int, flags)
7859 struct sched_attr attr;
7860 struct task_struct *p;
7863 if (!uattr || pid < 0 || flags)
7866 retval = sched_copy_attr(uattr, &attr);
7870 if ((int)attr.sched_policy < 0)
7872 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7873 attr.sched_policy = SETPARAM_POLICY;
7877 p = find_process_by_pid(pid);
7883 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7884 get_params(p, &attr);
7885 retval = sched_setattr(p, &attr);
7893 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7894 * @pid: the pid in question.
7896 * Return: On success, the policy of the thread. Otherwise, a negative error
7899 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7901 struct task_struct *p;
7909 p = find_process_by_pid(pid);
7911 retval = security_task_getscheduler(p);
7914 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7921 * sys_sched_getparam - get the RT priority of a thread
7922 * @pid: the pid in question.
7923 * @param: structure containing the RT priority.
7925 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7928 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7930 struct sched_param lp = { .sched_priority = 0 };
7931 struct task_struct *p;
7934 if (!param || pid < 0)
7938 p = find_process_by_pid(pid);
7943 retval = security_task_getscheduler(p);
7947 if (task_has_rt_policy(p))
7948 lp.sched_priority = p->rt_priority;
7952 * This one might sleep, we cannot do it with a spinlock held ...
7954 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7964 * Copy the kernel size attribute structure (which might be larger
7965 * than what user-space knows about) to user-space.
7967 * Note that all cases are valid: user-space buffer can be larger or
7968 * smaller than the kernel-space buffer. The usual case is that both
7969 * have the same size.
7972 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7973 struct sched_attr *kattr,
7976 unsigned int ksize = sizeof(*kattr);
7978 if (!access_ok(uattr, usize))
7982 * sched_getattr() ABI forwards and backwards compatibility:
7984 * If usize == ksize then we just copy everything to user-space and all is good.
7986 * If usize < ksize then we only copy as much as user-space has space for,
7987 * this keeps ABI compatibility as well. We skip the rest.
7989 * If usize > ksize then user-space is using a newer version of the ABI,
7990 * which part the kernel doesn't know about. Just ignore it - tooling can
7991 * detect the kernel's knowledge of attributes from the attr->size value
7992 * which is set to ksize in this case.
7994 kattr->size = min(usize, ksize);
7996 if (copy_to_user(uattr, kattr, kattr->size))
8003 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8004 * @pid: the pid in question.
8005 * @uattr: structure containing the extended parameters.
8006 * @usize: sizeof(attr) for fwd/bwd comp.
8007 * @flags: for future extension.
8009 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8010 unsigned int, usize, unsigned int, flags)
8012 struct sched_attr kattr = { };
8013 struct task_struct *p;
8016 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8017 usize < SCHED_ATTR_SIZE_VER0 || flags)
8021 p = find_process_by_pid(pid);
8026 retval = security_task_getscheduler(p);
8030 kattr.sched_policy = p->policy;
8031 if (p->sched_reset_on_fork)
8032 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8033 get_params(p, &kattr);
8034 kattr.sched_flags &= SCHED_FLAG_ALL;
8036 #ifdef CONFIG_UCLAMP_TASK
8038 * This could race with another potential updater, but this is fine
8039 * because it'll correctly read the old or the new value. We don't need
8040 * to guarantee who wins the race as long as it doesn't return garbage.
8042 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8043 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8048 return sched_attr_copy_to_user(uattr, &kattr, usize);
8056 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8061 * If the task isn't a deadline task or admission control is
8062 * disabled then we don't care about affinity changes.
8064 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8068 * Since bandwidth control happens on root_domain basis,
8069 * if admission test is enabled, we only admit -deadline
8070 * tasks allowed to run on all the CPUs in the task's
8074 if (!cpumask_subset(task_rq(p)->rd->span, mask))
8082 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
8085 cpumask_var_t cpus_allowed, new_mask;
8087 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8090 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8092 goto out_free_cpus_allowed;
8095 cpuset_cpus_allowed(p, cpus_allowed);
8096 cpumask_and(new_mask, mask, cpus_allowed);
8098 retval = dl_task_check_affinity(p, new_mask);
8100 goto out_free_new_mask;
8102 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
8104 goto out_free_new_mask;
8106 cpuset_cpus_allowed(p, cpus_allowed);
8107 if (!cpumask_subset(new_mask, cpus_allowed)) {
8109 * We must have raced with a concurrent cpuset update.
8110 * Just reset the cpumask to the cpuset's cpus_allowed.
8112 cpumask_copy(new_mask, cpus_allowed);
8117 free_cpumask_var(new_mask);
8118 out_free_cpus_allowed:
8119 free_cpumask_var(cpus_allowed);
8123 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8125 struct task_struct *p;
8130 p = find_process_by_pid(pid);
8136 /* Prevent p going away */
8140 if (p->flags & PF_NO_SETAFFINITY) {
8145 if (!check_same_owner(p)) {
8147 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8155 retval = security_task_setscheduler(p);
8159 retval = __sched_setaffinity(p, in_mask);
8165 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8166 struct cpumask *new_mask)
8168 if (len < cpumask_size())
8169 cpumask_clear(new_mask);
8170 else if (len > cpumask_size())
8171 len = cpumask_size();
8173 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8177 * sys_sched_setaffinity - set the CPU affinity of a process
8178 * @pid: pid of the process
8179 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8180 * @user_mask_ptr: user-space pointer to the new CPU mask
8182 * Return: 0 on success. An error code otherwise.
8184 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8185 unsigned long __user *, user_mask_ptr)
8187 cpumask_var_t new_mask;
8190 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8193 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8195 retval = sched_setaffinity(pid, new_mask);
8196 free_cpumask_var(new_mask);
8200 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8202 struct task_struct *p;
8203 unsigned long flags;
8209 p = find_process_by_pid(pid);
8213 retval = security_task_getscheduler(p);
8217 raw_spin_lock_irqsave(&p->pi_lock, flags);
8218 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8219 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8228 * sys_sched_getaffinity - get the CPU affinity of a process
8229 * @pid: pid of the process
8230 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8231 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8233 * Return: size of CPU mask copied to user_mask_ptr on success. An
8234 * error code otherwise.
8236 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8237 unsigned long __user *, user_mask_ptr)
8242 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8244 if (len & (sizeof(unsigned long)-1))
8247 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8250 ret = sched_getaffinity(pid, mask);
8252 unsigned int retlen = min(len, cpumask_size());
8254 if (copy_to_user(user_mask_ptr, mask, retlen))
8259 free_cpumask_var(mask);
8264 static void do_sched_yield(void)
8269 rq = this_rq_lock_irq(&rf);
8271 schedstat_inc(rq->yld_count);
8272 current->sched_class->yield_task(rq);
8275 rq_unlock_irq(rq, &rf);
8276 sched_preempt_enable_no_resched();
8282 * sys_sched_yield - yield the current processor to other threads.
8284 * This function yields the current CPU to other tasks. If there are no
8285 * other threads running on this CPU then this function will return.
8289 SYSCALL_DEFINE0(sched_yield)
8295 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8296 int __sched __cond_resched(void)
8298 if (should_resched(0)) {
8299 preempt_schedule_common();
8303 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8304 * whether the current CPU is in an RCU read-side critical section,
8305 * so the tick can report quiescent states even for CPUs looping
8306 * in kernel context. In contrast, in non-preemptible kernels,
8307 * RCU readers leave no in-memory hints, which means that CPU-bound
8308 * processes executing in kernel context might never report an
8309 * RCU quiescent state. Therefore, the following code causes
8310 * cond_resched() to report a quiescent state, but only when RCU
8311 * is in urgent need of one.
8313 #ifndef CONFIG_PREEMPT_RCU
8318 EXPORT_SYMBOL(__cond_resched);
8321 #ifdef CONFIG_PREEMPT_DYNAMIC
8322 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8323 #define cond_resched_dynamic_enabled __cond_resched
8324 #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8325 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8326 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8328 #define might_resched_dynamic_enabled __cond_resched
8329 #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8330 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8331 EXPORT_STATIC_CALL_TRAMP(might_resched);
8332 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8333 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8334 int __sched dynamic_cond_resched(void)
8336 if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8338 return __cond_resched();
8340 EXPORT_SYMBOL(dynamic_cond_resched);
8342 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8343 int __sched dynamic_might_resched(void)
8345 if (!static_branch_unlikely(&sk_dynamic_might_resched))
8347 return __cond_resched();
8349 EXPORT_SYMBOL(dynamic_might_resched);
8354 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8355 * call schedule, and on return reacquire the lock.
8357 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8358 * operations here to prevent schedule() from being called twice (once via
8359 * spin_unlock(), once by hand).
8361 int __cond_resched_lock(spinlock_t *lock)
8363 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8366 lockdep_assert_held(lock);
8368 if (spin_needbreak(lock) || resched) {
8370 if (!_cond_resched())
8377 EXPORT_SYMBOL(__cond_resched_lock);
8379 int __cond_resched_rwlock_read(rwlock_t *lock)
8381 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8384 lockdep_assert_held_read(lock);
8386 if (rwlock_needbreak(lock) || resched) {
8388 if (!_cond_resched())
8395 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8397 int __cond_resched_rwlock_write(rwlock_t *lock)
8399 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8402 lockdep_assert_held_write(lock);
8404 if (rwlock_needbreak(lock) || resched) {
8406 if (!_cond_resched())
8413 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8415 #ifdef CONFIG_PREEMPT_DYNAMIC
8417 #ifdef CONFIG_GENERIC_ENTRY
8418 #include <linux/entry-common.h>
8424 * SC:preempt_schedule
8425 * SC:preempt_schedule_notrace
8426 * SC:irqentry_exit_cond_resched
8430 * cond_resched <- __cond_resched
8431 * might_resched <- RET0
8432 * preempt_schedule <- NOP
8433 * preempt_schedule_notrace <- NOP
8434 * irqentry_exit_cond_resched <- NOP
8437 * cond_resched <- __cond_resched
8438 * might_resched <- __cond_resched
8439 * preempt_schedule <- NOP
8440 * preempt_schedule_notrace <- NOP
8441 * irqentry_exit_cond_resched <- NOP
8444 * cond_resched <- RET0
8445 * might_resched <- RET0
8446 * preempt_schedule <- preempt_schedule
8447 * preempt_schedule_notrace <- preempt_schedule_notrace
8448 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8452 preempt_dynamic_undefined = -1,
8453 preempt_dynamic_none,
8454 preempt_dynamic_voluntary,
8455 preempt_dynamic_full,
8458 int preempt_dynamic_mode = preempt_dynamic_undefined;
8460 int sched_dynamic_mode(const char *str)
8462 if (!strcmp(str, "none"))
8463 return preempt_dynamic_none;
8465 if (!strcmp(str, "voluntary"))
8466 return preempt_dynamic_voluntary;
8468 if (!strcmp(str, "full"))
8469 return preempt_dynamic_full;
8474 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8475 #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8476 #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8477 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8478 #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8479 #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8481 #error "Unsupported PREEMPT_DYNAMIC mechanism"
8484 void sched_dynamic_update(int mode)
8487 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8488 * the ZERO state, which is invalid.
8490 preempt_dynamic_enable(cond_resched);
8491 preempt_dynamic_enable(might_resched);
8492 preempt_dynamic_enable(preempt_schedule);
8493 preempt_dynamic_enable(preempt_schedule_notrace);
8494 preempt_dynamic_enable(irqentry_exit_cond_resched);
8497 case preempt_dynamic_none:
8498 preempt_dynamic_enable(cond_resched);
8499 preempt_dynamic_disable(might_resched);
8500 preempt_dynamic_disable(preempt_schedule);
8501 preempt_dynamic_disable(preempt_schedule_notrace);
8502 preempt_dynamic_disable(irqentry_exit_cond_resched);
8503 pr_info("Dynamic Preempt: none\n");
8506 case preempt_dynamic_voluntary:
8507 preempt_dynamic_enable(cond_resched);
8508 preempt_dynamic_enable(might_resched);
8509 preempt_dynamic_disable(preempt_schedule);
8510 preempt_dynamic_disable(preempt_schedule_notrace);
8511 preempt_dynamic_disable(irqentry_exit_cond_resched);
8512 pr_info("Dynamic Preempt: voluntary\n");
8515 case preempt_dynamic_full:
8516 preempt_dynamic_disable(cond_resched);
8517 preempt_dynamic_disable(might_resched);
8518 preempt_dynamic_enable(preempt_schedule);
8519 preempt_dynamic_enable(preempt_schedule_notrace);
8520 preempt_dynamic_enable(irqentry_exit_cond_resched);
8521 pr_info("Dynamic Preempt: full\n");
8525 preempt_dynamic_mode = mode;
8528 static int __init setup_preempt_mode(char *str)
8530 int mode = sched_dynamic_mode(str);
8532 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8536 sched_dynamic_update(mode);
8539 __setup("preempt=", setup_preempt_mode);
8541 static void __init preempt_dynamic_init(void)
8543 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8544 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8545 sched_dynamic_update(preempt_dynamic_none);
8546 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8547 sched_dynamic_update(preempt_dynamic_voluntary);
8549 /* Default static call setting, nothing to do */
8550 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8551 preempt_dynamic_mode = preempt_dynamic_full;
8552 pr_info("Dynamic Preempt: full\n");
8557 #define PREEMPT_MODEL_ACCESSOR(mode) \
8558 bool preempt_model_##mode(void) \
8560 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8561 return preempt_dynamic_mode == preempt_dynamic_##mode; \
8563 EXPORT_SYMBOL_GPL(preempt_model_##mode)
8565 PREEMPT_MODEL_ACCESSOR(none);
8566 PREEMPT_MODEL_ACCESSOR(voluntary);
8567 PREEMPT_MODEL_ACCESSOR(full);
8569 #else /* !CONFIG_PREEMPT_DYNAMIC */
8571 static inline void preempt_dynamic_init(void) { }
8573 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8576 * yield - yield the current processor to other threads.
8578 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8580 * The scheduler is at all times free to pick the calling task as the most
8581 * eligible task to run, if removing the yield() call from your code breaks
8582 * it, it's already broken.
8584 * Typical broken usage is:
8589 * where one assumes that yield() will let 'the other' process run that will
8590 * make event true. If the current task is a SCHED_FIFO task that will never
8591 * happen. Never use yield() as a progress guarantee!!
8593 * If you want to use yield() to wait for something, use wait_event().
8594 * If you want to use yield() to be 'nice' for others, use cond_resched().
8595 * If you still want to use yield(), do not!
8597 void __sched yield(void)
8599 set_current_state(TASK_RUNNING);
8602 EXPORT_SYMBOL(yield);
8605 * yield_to - yield the current processor to another thread in
8606 * your thread group, or accelerate that thread toward the
8607 * processor it's on.
8609 * @preempt: whether task preemption is allowed or not
8611 * It's the caller's job to ensure that the target task struct
8612 * can't go away on us before we can do any checks.
8615 * true (>0) if we indeed boosted the target task.
8616 * false (0) if we failed to boost the target.
8617 * -ESRCH if there's no task to yield to.
8619 int __sched yield_to(struct task_struct *p, bool preempt)
8621 struct task_struct *curr = current;
8622 struct rq *rq, *p_rq;
8623 unsigned long flags;
8626 local_irq_save(flags);
8632 * If we're the only runnable task on the rq and target rq also
8633 * has only one task, there's absolutely no point in yielding.
8635 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8640 double_rq_lock(rq, p_rq);
8641 if (task_rq(p) != p_rq) {
8642 double_rq_unlock(rq, p_rq);
8646 if (!curr->sched_class->yield_to_task)
8649 if (curr->sched_class != p->sched_class)
8652 if (task_running(p_rq, p) || !task_is_running(p))
8655 yielded = curr->sched_class->yield_to_task(rq, p);
8657 schedstat_inc(rq->yld_count);
8659 * Make p's CPU reschedule; pick_next_entity takes care of
8662 if (preempt && rq != p_rq)
8667 double_rq_unlock(rq, p_rq);
8669 local_irq_restore(flags);
8676 EXPORT_SYMBOL_GPL(yield_to);
8678 int io_schedule_prepare(void)
8680 int old_iowait = current->in_iowait;
8682 current->in_iowait = 1;
8683 blk_flush_plug(current->plug, true);
8687 void io_schedule_finish(int token)
8689 current->in_iowait = token;
8693 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8694 * that process accounting knows that this is a task in IO wait state.
8696 long __sched io_schedule_timeout(long timeout)
8701 token = io_schedule_prepare();
8702 ret = schedule_timeout(timeout);
8703 io_schedule_finish(token);
8707 EXPORT_SYMBOL(io_schedule_timeout);
8709 void __sched io_schedule(void)
8713 token = io_schedule_prepare();
8715 io_schedule_finish(token);
8717 EXPORT_SYMBOL(io_schedule);
8720 * sys_sched_get_priority_max - return maximum RT priority.
8721 * @policy: scheduling class.
8723 * Return: On success, this syscall returns the maximum
8724 * rt_priority that can be used by a given scheduling class.
8725 * On failure, a negative error code is returned.
8727 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8734 ret = MAX_RT_PRIO-1;
8736 case SCHED_DEADLINE:
8747 * sys_sched_get_priority_min - return minimum RT priority.
8748 * @policy: scheduling class.
8750 * Return: On success, this syscall returns the minimum
8751 * rt_priority that can be used by a given scheduling class.
8752 * On failure, a negative error code is returned.
8754 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8763 case SCHED_DEADLINE:
8772 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8774 struct task_struct *p;
8775 unsigned int time_slice;
8785 p = find_process_by_pid(pid);
8789 retval = security_task_getscheduler(p);
8793 rq = task_rq_lock(p, &rf);
8795 if (p->sched_class->get_rr_interval)
8796 time_slice = p->sched_class->get_rr_interval(rq, p);
8797 task_rq_unlock(rq, p, &rf);
8800 jiffies_to_timespec64(time_slice, t);
8809 * sys_sched_rr_get_interval - return the default timeslice of a process.
8810 * @pid: pid of the process.
8811 * @interval: userspace pointer to the timeslice value.
8813 * this syscall writes the default timeslice value of a given process
8814 * into the user-space timespec buffer. A value of '0' means infinity.
8816 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8819 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8820 struct __kernel_timespec __user *, interval)
8822 struct timespec64 t;
8823 int retval = sched_rr_get_interval(pid, &t);
8826 retval = put_timespec64(&t, interval);
8831 #ifdef CONFIG_COMPAT_32BIT_TIME
8832 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8833 struct old_timespec32 __user *, interval)
8835 struct timespec64 t;
8836 int retval = sched_rr_get_interval(pid, &t);
8839 retval = put_old_timespec32(&t, interval);
8844 void sched_show_task(struct task_struct *p)
8846 unsigned long free = 0;
8849 if (!try_get_task_stack(p))
8852 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8854 if (task_is_running(p))
8855 pr_cont(" running task ");
8856 #ifdef CONFIG_DEBUG_STACK_USAGE
8857 free = stack_not_used(p);
8862 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8864 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8865 free, task_pid_nr(p), ppid,
8866 read_task_thread_flags(p));
8868 print_worker_info(KERN_INFO, p);
8869 print_stop_info(KERN_INFO, p);
8870 show_stack(p, NULL, KERN_INFO);
8873 EXPORT_SYMBOL_GPL(sched_show_task);
8876 state_filter_match(unsigned long state_filter, struct task_struct *p)
8878 unsigned int state = READ_ONCE(p->__state);
8880 /* no filter, everything matches */
8884 /* filter, but doesn't match */
8885 if (!(state & state_filter))
8889 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8892 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8899 void show_state_filter(unsigned int state_filter)
8901 struct task_struct *g, *p;
8904 for_each_process_thread(g, p) {
8906 * reset the NMI-timeout, listing all files on a slow
8907 * console might take a lot of time:
8908 * Also, reset softlockup watchdogs on all CPUs, because
8909 * another CPU might be blocked waiting for us to process
8912 touch_nmi_watchdog();
8913 touch_all_softlockup_watchdogs();
8914 if (state_filter_match(state_filter, p))
8918 #ifdef CONFIG_SCHED_DEBUG
8920 sysrq_sched_debug_show();
8924 * Only show locks if all tasks are dumped:
8927 debug_show_all_locks();
8931 * init_idle - set up an idle thread for a given CPU
8932 * @idle: task in question
8933 * @cpu: CPU the idle task belongs to
8935 * NOTE: this function does not set the idle thread's NEED_RESCHED
8936 * flag, to make booting more robust.
8938 void __init init_idle(struct task_struct *idle, int cpu)
8940 struct rq *rq = cpu_rq(cpu);
8941 unsigned long flags;
8943 __sched_fork(0, idle);
8945 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8946 raw_spin_rq_lock(rq);
8948 idle->__state = TASK_RUNNING;
8949 idle->se.exec_start = sched_clock();
8951 * PF_KTHREAD should already be set at this point; regardless, make it
8952 * look like a proper per-CPU kthread.
8954 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8955 kthread_set_per_cpu(idle, cpu);
8959 * It's possible that init_idle() gets called multiple times on a task,
8960 * in that case do_set_cpus_allowed() will not do the right thing.
8962 * And since this is boot we can forgo the serialization.
8964 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8967 * We're having a chicken and egg problem, even though we are
8968 * holding rq->lock, the CPU isn't yet set to this CPU so the
8969 * lockdep check in task_group() will fail.
8971 * Similar case to sched_fork(). / Alternatively we could
8972 * use task_rq_lock() here and obtain the other rq->lock.
8977 __set_task_cpu(idle, cpu);
8981 rcu_assign_pointer(rq->curr, idle);
8982 idle->on_rq = TASK_ON_RQ_QUEUED;
8986 raw_spin_rq_unlock(rq);
8987 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8989 /* Set the preempt count _outside_ the spinlocks! */
8990 init_idle_preempt_count(idle, cpu);
8993 * The idle tasks have their own, simple scheduling class:
8995 idle->sched_class = &idle_sched_class;
8996 ftrace_graph_init_idle_task(idle, cpu);
8997 vtime_init_idle(idle, cpu);
8999 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9005 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9006 const struct cpumask *trial)
9010 if (cpumask_empty(cur))
9013 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9018 int task_can_attach(struct task_struct *p,
9019 const struct cpumask *cs_effective_cpus)
9024 * Kthreads which disallow setaffinity shouldn't be moved
9025 * to a new cpuset; we don't want to change their CPU
9026 * affinity and isolating such threads by their set of
9027 * allowed nodes is unnecessary. Thus, cpusets are not
9028 * applicable for such threads. This prevents checking for
9029 * success of set_cpus_allowed_ptr() on all attached tasks
9030 * before cpus_mask may be changed.
9032 if (p->flags & PF_NO_SETAFFINITY) {
9037 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
9038 cs_effective_cpus)) {
9039 int cpu = cpumask_any_and(cpu_active_mask, cs_effective_cpus);
9041 if (unlikely(cpu >= nr_cpu_ids))
9043 ret = dl_cpu_busy(cpu, p);
9050 bool sched_smp_initialized __read_mostly;
9052 #ifdef CONFIG_NUMA_BALANCING
9053 /* Migrate current task p to target_cpu */
9054 int migrate_task_to(struct task_struct *p, int target_cpu)
9056 struct migration_arg arg = { p, target_cpu };
9057 int curr_cpu = task_cpu(p);
9059 if (curr_cpu == target_cpu)
9062 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9065 /* TODO: This is not properly updating schedstats */
9067 trace_sched_move_numa(p, curr_cpu, target_cpu);
9068 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9072 * Requeue a task on a given node and accurately track the number of NUMA
9073 * tasks on the runqueues
9075 void sched_setnuma(struct task_struct *p, int nid)
9077 bool queued, running;
9081 rq = task_rq_lock(p, &rf);
9082 queued = task_on_rq_queued(p);
9083 running = task_current(rq, p);
9086 dequeue_task(rq, p, DEQUEUE_SAVE);
9088 put_prev_task(rq, p);
9090 p->numa_preferred_nid = nid;
9093 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9095 set_next_task(rq, p);
9096 task_rq_unlock(rq, p, &rf);
9098 #endif /* CONFIG_NUMA_BALANCING */
9100 #ifdef CONFIG_HOTPLUG_CPU
9102 * Ensure that the idle task is using init_mm right before its CPU goes
9105 void idle_task_exit(void)
9107 struct mm_struct *mm = current->active_mm;
9109 BUG_ON(cpu_online(smp_processor_id()));
9110 BUG_ON(current != this_rq()->idle);
9112 if (mm != &init_mm) {
9113 switch_mm(mm, &init_mm, current);
9114 finish_arch_post_lock_switch();
9117 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9120 static int __balance_push_cpu_stop(void *arg)
9122 struct task_struct *p = arg;
9123 struct rq *rq = this_rq();
9127 raw_spin_lock_irq(&p->pi_lock);
9130 update_rq_clock(rq);
9132 if (task_rq(p) == rq && task_on_rq_queued(p)) {
9133 cpu = select_fallback_rq(rq->cpu, p);
9134 rq = __migrate_task(rq, &rf, p, cpu);
9138 raw_spin_unlock_irq(&p->pi_lock);
9145 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9148 * Ensure we only run per-cpu kthreads once the CPU goes !active.
9150 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9151 * effective when the hotplug motion is down.
9153 static void balance_push(struct rq *rq)
9155 struct task_struct *push_task = rq->curr;
9157 lockdep_assert_rq_held(rq);
9160 * Ensure the thing is persistent until balance_push_set(.on = false);
9162 rq->balance_callback = &balance_push_callback;
9165 * Only active while going offline and when invoked on the outgoing
9168 if (!cpu_dying(rq->cpu) || rq != this_rq())
9172 * Both the cpu-hotplug and stop task are in this case and are
9173 * required to complete the hotplug process.
9175 if (kthread_is_per_cpu(push_task) ||
9176 is_migration_disabled(push_task)) {
9179 * If this is the idle task on the outgoing CPU try to wake
9180 * up the hotplug control thread which might wait for the
9181 * last task to vanish. The rcuwait_active() check is
9182 * accurate here because the waiter is pinned on this CPU
9183 * and can't obviously be running in parallel.
9185 * On RT kernels this also has to check whether there are
9186 * pinned and scheduled out tasks on the runqueue. They
9187 * need to leave the migrate disabled section first.
9189 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9190 rcuwait_active(&rq->hotplug_wait)) {
9191 raw_spin_rq_unlock(rq);
9192 rcuwait_wake_up(&rq->hotplug_wait);
9193 raw_spin_rq_lock(rq);
9198 get_task_struct(push_task);
9200 * Temporarily drop rq->lock such that we can wake-up the stop task.
9201 * Both preemption and IRQs are still disabled.
9203 raw_spin_rq_unlock(rq);
9204 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9205 this_cpu_ptr(&push_work));
9207 * At this point need_resched() is true and we'll take the loop in
9208 * schedule(). The next pick is obviously going to be the stop task
9209 * which kthread_is_per_cpu() and will push this task away.
9211 raw_spin_rq_lock(rq);
9214 static void balance_push_set(int cpu, bool on)
9216 struct rq *rq = cpu_rq(cpu);
9219 rq_lock_irqsave(rq, &rf);
9221 WARN_ON_ONCE(rq->balance_callback);
9222 rq->balance_callback = &balance_push_callback;
9223 } else if (rq->balance_callback == &balance_push_callback) {
9224 rq->balance_callback = NULL;
9226 rq_unlock_irqrestore(rq, &rf);
9230 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9231 * inactive. All tasks which are not per CPU kernel threads are either
9232 * pushed off this CPU now via balance_push() or placed on a different CPU
9233 * during wakeup. Wait until the CPU is quiescent.
9235 static void balance_hotplug_wait(void)
9237 struct rq *rq = this_rq();
9239 rcuwait_wait_event(&rq->hotplug_wait,
9240 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9241 TASK_UNINTERRUPTIBLE);
9246 static inline void balance_push(struct rq *rq)
9250 static inline void balance_push_set(int cpu, bool on)
9254 static inline void balance_hotplug_wait(void)
9258 #endif /* CONFIG_HOTPLUG_CPU */
9260 void set_rq_online(struct rq *rq)
9263 const struct sched_class *class;
9265 cpumask_set_cpu(rq->cpu, rq->rd->online);
9268 for_each_class(class) {
9269 if (class->rq_online)
9270 class->rq_online(rq);
9275 void set_rq_offline(struct rq *rq)
9278 const struct sched_class *class;
9280 for_each_class(class) {
9281 if (class->rq_offline)
9282 class->rq_offline(rq);
9285 cpumask_clear_cpu(rq->cpu, rq->rd->online);
9291 * used to mark begin/end of suspend/resume:
9293 static int num_cpus_frozen;
9296 * Update cpusets according to cpu_active mask. If cpusets are
9297 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9298 * around partition_sched_domains().
9300 * If we come here as part of a suspend/resume, don't touch cpusets because we
9301 * want to restore it back to its original state upon resume anyway.
9303 static void cpuset_cpu_active(void)
9305 if (cpuhp_tasks_frozen) {
9307 * num_cpus_frozen tracks how many CPUs are involved in suspend
9308 * resume sequence. As long as this is not the last online
9309 * operation in the resume sequence, just build a single sched
9310 * domain, ignoring cpusets.
9312 partition_sched_domains(1, NULL, NULL);
9313 if (--num_cpus_frozen)
9316 * This is the last CPU online operation. So fall through and
9317 * restore the original sched domains by considering the
9318 * cpuset configurations.
9320 cpuset_force_rebuild();
9322 cpuset_update_active_cpus();
9325 static int cpuset_cpu_inactive(unsigned int cpu)
9327 if (!cpuhp_tasks_frozen) {
9328 int ret = dl_cpu_busy(cpu, NULL);
9332 cpuset_update_active_cpus();
9335 partition_sched_domains(1, NULL, NULL);
9340 int sched_cpu_activate(unsigned int cpu)
9342 struct rq *rq = cpu_rq(cpu);
9346 * Clear the balance_push callback and prepare to schedule
9349 balance_push_set(cpu, false);
9351 #ifdef CONFIG_SCHED_SMT
9353 * When going up, increment the number of cores with SMT present.
9355 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9356 static_branch_inc_cpuslocked(&sched_smt_present);
9358 set_cpu_active(cpu, true);
9360 if (sched_smp_initialized) {
9361 sched_update_numa(cpu, true);
9362 sched_domains_numa_masks_set(cpu);
9363 cpuset_cpu_active();
9367 * Put the rq online, if not already. This happens:
9369 * 1) In the early boot process, because we build the real domains
9370 * after all CPUs have been brought up.
9372 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9375 rq_lock_irqsave(rq, &rf);
9377 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9380 rq_unlock_irqrestore(rq, &rf);
9385 int sched_cpu_deactivate(unsigned int cpu)
9387 struct rq *rq = cpu_rq(cpu);
9392 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9393 * load balancing when not active
9395 nohz_balance_exit_idle(rq);
9397 set_cpu_active(cpu, false);
9400 * From this point forward, this CPU will refuse to run any task that
9401 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9402 * push those tasks away until this gets cleared, see
9403 * sched_cpu_dying().
9405 balance_push_set(cpu, true);
9408 * We've cleared cpu_active_mask / set balance_push, wait for all
9409 * preempt-disabled and RCU users of this state to go away such that
9410 * all new such users will observe it.
9412 * Specifically, we rely on ttwu to no longer target this CPU, see
9413 * ttwu_queue_cond() and is_cpu_allowed().
9415 * Do sync before park smpboot threads to take care the rcu boost case.
9419 rq_lock_irqsave(rq, &rf);
9421 update_rq_clock(rq);
9422 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9425 rq_unlock_irqrestore(rq, &rf);
9427 #ifdef CONFIG_SCHED_SMT
9429 * When going down, decrement the number of cores with SMT present.
9431 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9432 static_branch_dec_cpuslocked(&sched_smt_present);
9434 sched_core_cpu_deactivate(cpu);
9437 if (!sched_smp_initialized)
9440 sched_update_numa(cpu, false);
9441 ret = cpuset_cpu_inactive(cpu);
9443 balance_push_set(cpu, false);
9444 set_cpu_active(cpu, true);
9445 sched_update_numa(cpu, true);
9448 sched_domains_numa_masks_clear(cpu);
9452 static void sched_rq_cpu_starting(unsigned int cpu)
9454 struct rq *rq = cpu_rq(cpu);
9456 rq->calc_load_update = calc_load_update;
9457 update_max_interval();
9460 int sched_cpu_starting(unsigned int cpu)
9462 sched_core_cpu_starting(cpu);
9463 sched_rq_cpu_starting(cpu);
9464 sched_tick_start(cpu);
9468 #ifdef CONFIG_HOTPLUG_CPU
9471 * Invoked immediately before the stopper thread is invoked to bring the
9472 * CPU down completely. At this point all per CPU kthreads except the
9473 * hotplug thread (current) and the stopper thread (inactive) have been
9474 * either parked or have been unbound from the outgoing CPU. Ensure that
9475 * any of those which might be on the way out are gone.
9477 * If after this point a bound task is being woken on this CPU then the
9478 * responsible hotplug callback has failed to do it's job.
9479 * sched_cpu_dying() will catch it with the appropriate fireworks.
9481 int sched_cpu_wait_empty(unsigned int cpu)
9483 balance_hotplug_wait();
9488 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9489 * might have. Called from the CPU stopper task after ensuring that the
9490 * stopper is the last running task on the CPU, so nr_active count is
9491 * stable. We need to take the teardown thread which is calling this into
9492 * account, so we hand in adjust = 1 to the load calculation.
9494 * Also see the comment "Global load-average calculations".
9496 static void calc_load_migrate(struct rq *rq)
9498 long delta = calc_load_fold_active(rq, 1);
9501 atomic_long_add(delta, &calc_load_tasks);
9504 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9506 struct task_struct *g, *p;
9507 int cpu = cpu_of(rq);
9509 lockdep_assert_rq_held(rq);
9511 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9512 for_each_process_thread(g, p) {
9513 if (task_cpu(p) != cpu)
9516 if (!task_on_rq_queued(p))
9519 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9523 int sched_cpu_dying(unsigned int cpu)
9525 struct rq *rq = cpu_rq(cpu);
9528 /* Handle pending wakeups and then migrate everything off */
9529 sched_tick_stop(cpu);
9531 rq_lock_irqsave(rq, &rf);
9532 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9533 WARN(true, "Dying CPU not properly vacated!");
9534 dump_rq_tasks(rq, KERN_WARNING);
9536 rq_unlock_irqrestore(rq, &rf);
9538 calc_load_migrate(rq);
9539 update_max_interval();
9541 sched_core_cpu_dying(cpu);
9546 void __init sched_init_smp(void)
9548 sched_init_numa(NUMA_NO_NODE);
9551 * There's no userspace yet to cause hotplug operations; hence all the
9552 * CPU masks are stable and all blatant races in the below code cannot
9555 mutex_lock(&sched_domains_mutex);
9556 sched_init_domains(cpu_active_mask);
9557 mutex_unlock(&sched_domains_mutex);
9559 /* Move init over to a non-isolated CPU */
9560 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9562 current->flags &= ~PF_NO_SETAFFINITY;
9563 sched_init_granularity();
9565 init_sched_rt_class();
9566 init_sched_dl_class();
9568 sched_smp_initialized = true;
9571 static int __init migration_init(void)
9573 sched_cpu_starting(smp_processor_id());
9576 early_initcall(migration_init);
9579 void __init sched_init_smp(void)
9581 sched_init_granularity();
9583 #endif /* CONFIG_SMP */
9585 int in_sched_functions(unsigned long addr)
9587 return in_lock_functions(addr) ||
9588 (addr >= (unsigned long)__sched_text_start
9589 && addr < (unsigned long)__sched_text_end);
9592 #ifdef CONFIG_CGROUP_SCHED
9594 * Default task group.
9595 * Every task in system belongs to this group at bootup.
9597 struct task_group root_task_group;
9598 LIST_HEAD(task_groups);
9600 /* Cacheline aligned slab cache for task_group */
9601 static struct kmem_cache *task_group_cache __read_mostly;
9604 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9605 DECLARE_PER_CPU(cpumask_var_t, select_rq_mask);
9607 void __init sched_init(void)
9609 unsigned long ptr = 0;
9612 /* Make sure the linker didn't screw up */
9613 BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9614 &fair_sched_class != &rt_sched_class + 1 ||
9615 &rt_sched_class != &dl_sched_class + 1);
9617 BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9622 #ifdef CONFIG_FAIR_GROUP_SCHED
9623 ptr += 2 * nr_cpu_ids * sizeof(void **);
9625 #ifdef CONFIG_RT_GROUP_SCHED
9626 ptr += 2 * nr_cpu_ids * sizeof(void **);
9629 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9631 #ifdef CONFIG_FAIR_GROUP_SCHED
9632 root_task_group.se = (struct sched_entity **)ptr;
9633 ptr += nr_cpu_ids * sizeof(void **);
9635 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9636 ptr += nr_cpu_ids * sizeof(void **);
9638 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9639 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9640 #endif /* CONFIG_FAIR_GROUP_SCHED */
9641 #ifdef CONFIG_RT_GROUP_SCHED
9642 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9643 ptr += nr_cpu_ids * sizeof(void **);
9645 root_task_group.rt_rq = (struct rt_rq **)ptr;
9646 ptr += nr_cpu_ids * sizeof(void **);
9648 #endif /* CONFIG_RT_GROUP_SCHED */
9650 #ifdef CONFIG_CPUMASK_OFFSTACK
9651 for_each_possible_cpu(i) {
9652 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9653 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9654 per_cpu(select_rq_mask, i) = (cpumask_var_t)kzalloc_node(
9655 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9657 #endif /* CONFIG_CPUMASK_OFFSTACK */
9659 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9662 init_defrootdomain();
9665 #ifdef CONFIG_RT_GROUP_SCHED
9666 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9667 global_rt_period(), global_rt_runtime());
9668 #endif /* CONFIG_RT_GROUP_SCHED */
9670 #ifdef CONFIG_CGROUP_SCHED
9671 task_group_cache = KMEM_CACHE(task_group, 0);
9673 list_add(&root_task_group.list, &task_groups);
9674 INIT_LIST_HEAD(&root_task_group.children);
9675 INIT_LIST_HEAD(&root_task_group.siblings);
9676 autogroup_init(&init_task);
9677 #endif /* CONFIG_CGROUP_SCHED */
9679 for_each_possible_cpu(i) {
9683 raw_spin_lock_init(&rq->__lock);
9685 rq->calc_load_active = 0;
9686 rq->calc_load_update = jiffies + LOAD_FREQ;
9687 init_cfs_rq(&rq->cfs);
9688 init_rt_rq(&rq->rt);
9689 init_dl_rq(&rq->dl);
9690 #ifdef CONFIG_FAIR_GROUP_SCHED
9691 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9692 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9694 * How much CPU bandwidth does root_task_group get?
9696 * In case of task-groups formed thr' the cgroup filesystem, it
9697 * gets 100% of the CPU resources in the system. This overall
9698 * system CPU resource is divided among the tasks of
9699 * root_task_group and its child task-groups in a fair manner,
9700 * based on each entity's (task or task-group's) weight
9701 * (se->load.weight).
9703 * In other words, if root_task_group has 10 tasks of weight
9704 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9705 * then A0's share of the CPU resource is:
9707 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9709 * We achieve this by letting root_task_group's tasks sit
9710 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9712 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9713 #endif /* CONFIG_FAIR_GROUP_SCHED */
9715 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9716 #ifdef CONFIG_RT_GROUP_SCHED
9717 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9722 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9723 rq->balance_callback = &balance_push_callback;
9724 rq->active_balance = 0;
9725 rq->next_balance = jiffies;
9730 rq->avg_idle = 2*sysctl_sched_migration_cost;
9731 rq->wake_stamp = jiffies;
9732 rq->wake_avg_idle = rq->avg_idle;
9733 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9735 INIT_LIST_HEAD(&rq->cfs_tasks);
9737 rq_attach_root(rq, &def_root_domain);
9738 #ifdef CONFIG_NO_HZ_COMMON
9739 rq->last_blocked_load_update_tick = jiffies;
9740 atomic_set(&rq->nohz_flags, 0);
9742 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9744 #ifdef CONFIG_HOTPLUG_CPU
9745 rcuwait_init(&rq->hotplug_wait);
9747 #endif /* CONFIG_SMP */
9749 atomic_set(&rq->nr_iowait, 0);
9751 #ifdef CONFIG_SCHED_CORE
9753 rq->core_pick = NULL;
9754 rq->core_enabled = 0;
9755 rq->core_tree = RB_ROOT;
9756 rq->core_forceidle_count = 0;
9757 rq->core_forceidle_occupation = 0;
9758 rq->core_forceidle_start = 0;
9760 rq->core_cookie = 0UL;
9764 set_load_weight(&init_task, false);
9767 * The boot idle thread does lazy MMU switching as well:
9770 enter_lazy_tlb(&init_mm, current);
9773 * The idle task doesn't need the kthread struct to function, but it
9774 * is dressed up as a per-CPU kthread and thus needs to play the part
9775 * if we want to avoid special-casing it in code that deals with per-CPU
9778 WARN_ON(!set_kthread_struct(current));
9781 * Make us the idle thread. Technically, schedule() should not be
9782 * called from this thread, however somewhere below it might be,
9783 * but because we are the idle thread, we just pick up running again
9784 * when this runqueue becomes "idle".
9786 init_idle(current, smp_processor_id());
9788 calc_load_update = jiffies + LOAD_FREQ;
9791 idle_thread_set_boot_cpu();
9792 balance_push_set(smp_processor_id(), false);
9794 init_sched_fair_class();
9800 preempt_dynamic_init();
9802 scheduler_running = 1;
9805 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9807 void __might_sleep(const char *file, int line)
9809 unsigned int state = get_current_state();
9811 * Blocking primitives will set (and therefore destroy) current->state,
9812 * since we will exit with TASK_RUNNING make sure we enter with it,
9813 * otherwise we will destroy state.
9815 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9816 "do not call blocking ops when !TASK_RUNNING; "
9817 "state=%x set at [<%p>] %pS\n", state,
9818 (void *)current->task_state_change,
9819 (void *)current->task_state_change);
9821 __might_resched(file, line, 0);
9823 EXPORT_SYMBOL(__might_sleep);
9825 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9827 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9830 if (preempt_count() == preempt_offset)
9833 pr_err("Preemption disabled at:");
9834 print_ip_sym(KERN_ERR, ip);
9837 static inline bool resched_offsets_ok(unsigned int offsets)
9839 unsigned int nested = preempt_count();
9841 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9843 return nested == offsets;
9846 void __might_resched(const char *file, int line, unsigned int offsets)
9848 /* Ratelimiting timestamp: */
9849 static unsigned long prev_jiffy;
9851 unsigned long preempt_disable_ip;
9853 /* WARN_ON_ONCE() by default, no rate limit required: */
9856 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9857 !is_idle_task(current) && !current->non_block_count) ||
9858 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9862 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9864 prev_jiffy = jiffies;
9866 /* Save this before calling printk(), since that will clobber it: */
9867 preempt_disable_ip = get_preempt_disable_ip(current);
9869 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9871 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9872 in_atomic(), irqs_disabled(), current->non_block_count,
9873 current->pid, current->comm);
9874 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9875 offsets & MIGHT_RESCHED_PREEMPT_MASK);
9877 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9878 pr_err("RCU nest depth: %d, expected: %u\n",
9879 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9882 if (task_stack_end_corrupted(current))
9883 pr_emerg("Thread overran stack, or stack corrupted\n");
9885 debug_show_held_locks(current);
9886 if (irqs_disabled())
9887 print_irqtrace_events(current);
9889 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9890 preempt_disable_ip);
9893 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9895 EXPORT_SYMBOL(__might_resched);
9897 void __cant_sleep(const char *file, int line, int preempt_offset)
9899 static unsigned long prev_jiffy;
9901 if (irqs_disabled())
9904 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9907 if (preempt_count() > preempt_offset)
9910 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9912 prev_jiffy = jiffies;
9914 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9915 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9916 in_atomic(), irqs_disabled(),
9917 current->pid, current->comm);
9919 debug_show_held_locks(current);
9921 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9923 EXPORT_SYMBOL_GPL(__cant_sleep);
9926 void __cant_migrate(const char *file, int line)
9928 static unsigned long prev_jiffy;
9930 if (irqs_disabled())
9933 if (is_migration_disabled(current))
9936 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9939 if (preempt_count() > 0)
9942 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9944 prev_jiffy = jiffies;
9946 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9947 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9948 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9949 current->pid, current->comm);
9951 debug_show_held_locks(current);
9953 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9955 EXPORT_SYMBOL_GPL(__cant_migrate);
9959 #ifdef CONFIG_MAGIC_SYSRQ
9960 void normalize_rt_tasks(void)
9962 struct task_struct *g, *p;
9963 struct sched_attr attr = {
9964 .sched_policy = SCHED_NORMAL,
9967 read_lock(&tasklist_lock);
9968 for_each_process_thread(g, p) {
9970 * Only normalize user tasks:
9972 if (p->flags & PF_KTHREAD)
9975 p->se.exec_start = 0;
9976 schedstat_set(p->stats.wait_start, 0);
9977 schedstat_set(p->stats.sleep_start, 0);
9978 schedstat_set(p->stats.block_start, 0);
9980 if (!dl_task(p) && !rt_task(p)) {
9982 * Renice negative nice level userspace
9985 if (task_nice(p) < 0)
9986 set_user_nice(p, 0);
9990 __sched_setscheduler(p, &attr, false, false);
9992 read_unlock(&tasklist_lock);
9995 #endif /* CONFIG_MAGIC_SYSRQ */
9997 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9999 * These functions are only useful for the IA64 MCA handling, or kdb.
10001 * They can only be called when the whole system has been
10002 * stopped - every CPU needs to be quiescent, and no scheduling
10003 * activity can take place. Using them for anything else would
10004 * be a serious bug, and as a result, they aren't even visible
10005 * under any other configuration.
10009 * curr_task - return the current task for a given CPU.
10010 * @cpu: the processor in question.
10012 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10014 * Return: The current task for @cpu.
10016 struct task_struct *curr_task(int cpu)
10018 return cpu_curr(cpu);
10021 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
10025 * ia64_set_curr_task - set the current task for a given CPU.
10026 * @cpu: the processor in question.
10027 * @p: the task pointer to set.
10029 * Description: This function must only be used when non-maskable interrupts
10030 * are serviced on a separate stack. It allows the architecture to switch the
10031 * notion of the current task on a CPU in a non-blocking manner. This function
10032 * must be called with all CPU's synchronized, and interrupts disabled, the
10033 * and caller must save the original value of the current task (see
10034 * curr_task() above) and restore that value before reenabling interrupts and
10035 * re-starting the system.
10037 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10039 void ia64_set_curr_task(int cpu, struct task_struct *p)
10046 #ifdef CONFIG_CGROUP_SCHED
10047 /* task_group_lock serializes the addition/removal of task groups */
10048 static DEFINE_SPINLOCK(task_group_lock);
10050 static inline void alloc_uclamp_sched_group(struct task_group *tg,
10051 struct task_group *parent)
10053 #ifdef CONFIG_UCLAMP_TASK_GROUP
10054 enum uclamp_id clamp_id;
10056 for_each_clamp_id(clamp_id) {
10057 uclamp_se_set(&tg->uclamp_req[clamp_id],
10058 uclamp_none(clamp_id), false);
10059 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10064 static void sched_free_group(struct task_group *tg)
10066 free_fair_sched_group(tg);
10067 free_rt_sched_group(tg);
10068 autogroup_free(tg);
10069 kmem_cache_free(task_group_cache, tg);
10072 static void sched_free_group_rcu(struct rcu_head *rcu)
10074 sched_free_group(container_of(rcu, struct task_group, rcu));
10077 static void sched_unregister_group(struct task_group *tg)
10079 unregister_fair_sched_group(tg);
10080 unregister_rt_sched_group(tg);
10082 * We have to wait for yet another RCU grace period to expire, as
10083 * print_cfs_stats() might run concurrently.
10085 call_rcu(&tg->rcu, sched_free_group_rcu);
10088 /* allocate runqueue etc for a new task group */
10089 struct task_group *sched_create_group(struct task_group *parent)
10091 struct task_group *tg;
10093 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10095 return ERR_PTR(-ENOMEM);
10097 if (!alloc_fair_sched_group(tg, parent))
10100 if (!alloc_rt_sched_group(tg, parent))
10103 alloc_uclamp_sched_group(tg, parent);
10108 sched_free_group(tg);
10109 return ERR_PTR(-ENOMEM);
10112 void sched_online_group(struct task_group *tg, struct task_group *parent)
10114 unsigned long flags;
10116 spin_lock_irqsave(&task_group_lock, flags);
10117 list_add_rcu(&tg->list, &task_groups);
10119 /* Root should already exist: */
10122 tg->parent = parent;
10123 INIT_LIST_HEAD(&tg->children);
10124 list_add_rcu(&tg->siblings, &parent->children);
10125 spin_unlock_irqrestore(&task_group_lock, flags);
10127 online_fair_sched_group(tg);
10130 /* rcu callback to free various structures associated with a task group */
10131 static void sched_unregister_group_rcu(struct rcu_head *rhp)
10133 /* Now it should be safe to free those cfs_rqs: */
10134 sched_unregister_group(container_of(rhp, struct task_group, rcu));
10137 void sched_destroy_group(struct task_group *tg)
10139 /* Wait for possible concurrent references to cfs_rqs complete: */
10140 call_rcu(&tg->rcu, sched_unregister_group_rcu);
10143 void sched_release_group(struct task_group *tg)
10145 unsigned long flags;
10148 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10149 * sched_cfs_period_timer()).
10151 * For this to be effective, we have to wait for all pending users of
10152 * this task group to leave their RCU critical section to ensure no new
10153 * user will see our dying task group any more. Specifically ensure
10154 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10156 * We therefore defer calling unregister_fair_sched_group() to
10157 * sched_unregister_group() which is guarantied to get called only after the
10158 * current RCU grace period has expired.
10160 spin_lock_irqsave(&task_group_lock, flags);
10161 list_del_rcu(&tg->list);
10162 list_del_rcu(&tg->siblings);
10163 spin_unlock_irqrestore(&task_group_lock, flags);
10166 static void sched_change_group(struct task_struct *tsk, int type)
10168 struct task_group *tg;
10171 * All callers are synchronized by task_rq_lock(); we do not use RCU
10172 * which is pointless here. Thus, we pass "true" to task_css_check()
10173 * to prevent lockdep warnings.
10175 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10176 struct task_group, css);
10177 tg = autogroup_task_group(tsk, tg);
10178 tsk->sched_task_group = tg;
10180 #ifdef CONFIG_FAIR_GROUP_SCHED
10181 if (tsk->sched_class->task_change_group)
10182 tsk->sched_class->task_change_group(tsk, type);
10185 set_task_rq(tsk, task_cpu(tsk));
10189 * Change task's runqueue when it moves between groups.
10191 * The caller of this function should have put the task in its new group by
10192 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10195 void sched_move_task(struct task_struct *tsk)
10197 int queued, running, queue_flags =
10198 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10199 struct rq_flags rf;
10202 rq = task_rq_lock(tsk, &rf);
10203 update_rq_clock(rq);
10205 running = task_current(rq, tsk);
10206 queued = task_on_rq_queued(tsk);
10209 dequeue_task(rq, tsk, queue_flags);
10211 put_prev_task(rq, tsk);
10213 sched_change_group(tsk, TASK_MOVE_GROUP);
10216 enqueue_task(rq, tsk, queue_flags);
10218 set_next_task(rq, tsk);
10220 * After changing group, the running task may have joined a
10221 * throttled one but it's still the running task. Trigger a
10222 * resched to make sure that task can still run.
10227 task_rq_unlock(rq, tsk, &rf);
10230 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10232 return css ? container_of(css, struct task_group, css) : NULL;
10235 static struct cgroup_subsys_state *
10236 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10238 struct task_group *parent = css_tg(parent_css);
10239 struct task_group *tg;
10242 /* This is early initialization for the top cgroup */
10243 return &root_task_group.css;
10246 tg = sched_create_group(parent);
10248 return ERR_PTR(-ENOMEM);
10253 /* Expose task group only after completing cgroup initialization */
10254 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10256 struct task_group *tg = css_tg(css);
10257 struct task_group *parent = css_tg(css->parent);
10260 sched_online_group(tg, parent);
10262 #ifdef CONFIG_UCLAMP_TASK_GROUP
10263 /* Propagate the effective uclamp value for the new group */
10264 mutex_lock(&uclamp_mutex);
10266 cpu_util_update_eff(css);
10268 mutex_unlock(&uclamp_mutex);
10274 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10276 struct task_group *tg = css_tg(css);
10278 sched_release_group(tg);
10281 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10283 struct task_group *tg = css_tg(css);
10286 * Relies on the RCU grace period between css_released() and this.
10288 sched_unregister_group(tg);
10292 * This is called before wake_up_new_task(), therefore we really only
10293 * have to set its group bits, all the other stuff does not apply.
10295 static void cpu_cgroup_fork(struct task_struct *task)
10297 struct rq_flags rf;
10300 rq = task_rq_lock(task, &rf);
10302 update_rq_clock(rq);
10303 sched_change_group(task, TASK_SET_GROUP);
10305 task_rq_unlock(rq, task, &rf);
10308 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10310 struct task_struct *task;
10311 struct cgroup_subsys_state *css;
10314 cgroup_taskset_for_each(task, css, tset) {
10315 #ifdef CONFIG_RT_GROUP_SCHED
10316 if (!sched_rt_can_attach(css_tg(css), task))
10320 * Serialize against wake_up_new_task() such that if it's
10321 * running, we're sure to observe its full state.
10323 raw_spin_lock_irq(&task->pi_lock);
10325 * Avoid calling sched_move_task() before wake_up_new_task()
10326 * has happened. This would lead to problems with PELT, due to
10327 * move wanting to detach+attach while we're not attached yet.
10329 if (READ_ONCE(task->__state) == TASK_NEW)
10331 raw_spin_unlock_irq(&task->pi_lock);
10339 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10341 struct task_struct *task;
10342 struct cgroup_subsys_state *css;
10344 cgroup_taskset_for_each(task, css, tset)
10345 sched_move_task(task);
10348 #ifdef CONFIG_UCLAMP_TASK_GROUP
10349 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10351 struct cgroup_subsys_state *top_css = css;
10352 struct uclamp_se *uc_parent = NULL;
10353 struct uclamp_se *uc_se = NULL;
10354 unsigned int eff[UCLAMP_CNT];
10355 enum uclamp_id clamp_id;
10356 unsigned int clamps;
10358 lockdep_assert_held(&uclamp_mutex);
10359 SCHED_WARN_ON(!rcu_read_lock_held());
10361 css_for_each_descendant_pre(css, top_css) {
10362 uc_parent = css_tg(css)->parent
10363 ? css_tg(css)->parent->uclamp : NULL;
10365 for_each_clamp_id(clamp_id) {
10366 /* Assume effective clamps matches requested clamps */
10367 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10368 /* Cap effective clamps with parent's effective clamps */
10370 eff[clamp_id] > uc_parent[clamp_id].value) {
10371 eff[clamp_id] = uc_parent[clamp_id].value;
10374 /* Ensure protection is always capped by limit */
10375 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10377 /* Propagate most restrictive effective clamps */
10379 uc_se = css_tg(css)->uclamp;
10380 for_each_clamp_id(clamp_id) {
10381 if (eff[clamp_id] == uc_se[clamp_id].value)
10383 uc_se[clamp_id].value = eff[clamp_id];
10384 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10385 clamps |= (0x1 << clamp_id);
10388 css = css_rightmost_descendant(css);
10392 /* Immediately update descendants RUNNABLE tasks */
10393 uclamp_update_active_tasks(css);
10398 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10399 * C expression. Since there is no way to convert a macro argument (N) into a
10400 * character constant, use two levels of macros.
10402 #define _POW10(exp) ((unsigned int)1e##exp)
10403 #define POW10(exp) _POW10(exp)
10405 struct uclamp_request {
10406 #define UCLAMP_PERCENT_SHIFT 2
10407 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10413 static inline struct uclamp_request
10414 capacity_from_percent(char *buf)
10416 struct uclamp_request req = {
10417 .percent = UCLAMP_PERCENT_SCALE,
10418 .util = SCHED_CAPACITY_SCALE,
10423 if (strcmp(buf, "max")) {
10424 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10428 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10433 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10434 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10440 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10441 size_t nbytes, loff_t off,
10442 enum uclamp_id clamp_id)
10444 struct uclamp_request req;
10445 struct task_group *tg;
10447 req = capacity_from_percent(buf);
10451 static_branch_enable(&sched_uclamp_used);
10453 mutex_lock(&uclamp_mutex);
10456 tg = css_tg(of_css(of));
10457 if (tg->uclamp_req[clamp_id].value != req.util)
10458 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10461 * Because of not recoverable conversion rounding we keep track of the
10462 * exact requested value
10464 tg->uclamp_pct[clamp_id] = req.percent;
10466 /* Update effective clamps to track the most restrictive value */
10467 cpu_util_update_eff(of_css(of));
10470 mutex_unlock(&uclamp_mutex);
10475 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10476 char *buf, size_t nbytes,
10479 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10482 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10483 char *buf, size_t nbytes,
10486 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10489 static inline void cpu_uclamp_print(struct seq_file *sf,
10490 enum uclamp_id clamp_id)
10492 struct task_group *tg;
10498 tg = css_tg(seq_css(sf));
10499 util_clamp = tg->uclamp_req[clamp_id].value;
10502 if (util_clamp == SCHED_CAPACITY_SCALE) {
10503 seq_puts(sf, "max\n");
10507 percent = tg->uclamp_pct[clamp_id];
10508 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10509 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10512 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10514 cpu_uclamp_print(sf, UCLAMP_MIN);
10518 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10520 cpu_uclamp_print(sf, UCLAMP_MAX);
10523 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10525 #ifdef CONFIG_FAIR_GROUP_SCHED
10526 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10527 struct cftype *cftype, u64 shareval)
10529 if (shareval > scale_load_down(ULONG_MAX))
10530 shareval = MAX_SHARES;
10531 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10534 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10535 struct cftype *cft)
10537 struct task_group *tg = css_tg(css);
10539 return (u64) scale_load_down(tg->shares);
10542 #ifdef CONFIG_CFS_BANDWIDTH
10543 static DEFINE_MUTEX(cfs_constraints_mutex);
10545 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10546 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10547 /* More than 203 days if BW_SHIFT equals 20. */
10548 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10550 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10552 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10555 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10556 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10558 if (tg == &root_task_group)
10562 * Ensure we have at some amount of bandwidth every period. This is
10563 * to prevent reaching a state of large arrears when throttled via
10564 * entity_tick() resulting in prolonged exit starvation.
10566 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10570 * Likewise, bound things on the other side by preventing insane quota
10571 * periods. This also allows us to normalize in computing quota
10574 if (period > max_cfs_quota_period)
10578 * Bound quota to defend quota against overflow during bandwidth shift.
10580 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10583 if (quota != RUNTIME_INF && (burst > quota ||
10584 burst + quota > max_cfs_runtime))
10588 * Prevent race between setting of cfs_rq->runtime_enabled and
10589 * unthrottle_offline_cfs_rqs().
10592 mutex_lock(&cfs_constraints_mutex);
10593 ret = __cfs_schedulable(tg, period, quota);
10597 runtime_enabled = quota != RUNTIME_INF;
10598 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10600 * If we need to toggle cfs_bandwidth_used, off->on must occur
10601 * before making related changes, and on->off must occur afterwards
10603 if (runtime_enabled && !runtime_was_enabled)
10604 cfs_bandwidth_usage_inc();
10605 raw_spin_lock_irq(&cfs_b->lock);
10606 cfs_b->period = ns_to_ktime(period);
10607 cfs_b->quota = quota;
10608 cfs_b->burst = burst;
10610 __refill_cfs_bandwidth_runtime(cfs_b);
10612 /* Restart the period timer (if active) to handle new period expiry: */
10613 if (runtime_enabled)
10614 start_cfs_bandwidth(cfs_b);
10616 raw_spin_unlock_irq(&cfs_b->lock);
10618 for_each_online_cpu(i) {
10619 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10620 struct rq *rq = cfs_rq->rq;
10621 struct rq_flags rf;
10623 rq_lock_irq(rq, &rf);
10624 cfs_rq->runtime_enabled = runtime_enabled;
10625 cfs_rq->runtime_remaining = 0;
10627 if (cfs_rq->throttled)
10628 unthrottle_cfs_rq(cfs_rq);
10629 rq_unlock_irq(rq, &rf);
10631 if (runtime_was_enabled && !runtime_enabled)
10632 cfs_bandwidth_usage_dec();
10634 mutex_unlock(&cfs_constraints_mutex);
10635 cpus_read_unlock();
10640 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10642 u64 quota, period, burst;
10644 period = ktime_to_ns(tg->cfs_bandwidth.period);
10645 burst = tg->cfs_bandwidth.burst;
10646 if (cfs_quota_us < 0)
10647 quota = RUNTIME_INF;
10648 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10649 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10653 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10656 static long tg_get_cfs_quota(struct task_group *tg)
10660 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10663 quota_us = tg->cfs_bandwidth.quota;
10664 do_div(quota_us, NSEC_PER_USEC);
10669 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10671 u64 quota, period, burst;
10673 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10676 period = (u64)cfs_period_us * NSEC_PER_USEC;
10677 quota = tg->cfs_bandwidth.quota;
10678 burst = tg->cfs_bandwidth.burst;
10680 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10683 static long tg_get_cfs_period(struct task_group *tg)
10687 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10688 do_div(cfs_period_us, NSEC_PER_USEC);
10690 return cfs_period_us;
10693 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10695 u64 quota, period, burst;
10697 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10700 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10701 period = ktime_to_ns(tg->cfs_bandwidth.period);
10702 quota = tg->cfs_bandwidth.quota;
10704 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10707 static long tg_get_cfs_burst(struct task_group *tg)
10711 burst_us = tg->cfs_bandwidth.burst;
10712 do_div(burst_us, NSEC_PER_USEC);
10717 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10718 struct cftype *cft)
10720 return tg_get_cfs_quota(css_tg(css));
10723 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10724 struct cftype *cftype, s64 cfs_quota_us)
10726 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10729 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10730 struct cftype *cft)
10732 return tg_get_cfs_period(css_tg(css));
10735 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10736 struct cftype *cftype, u64 cfs_period_us)
10738 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10741 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10742 struct cftype *cft)
10744 return tg_get_cfs_burst(css_tg(css));
10747 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10748 struct cftype *cftype, u64 cfs_burst_us)
10750 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10753 struct cfs_schedulable_data {
10754 struct task_group *tg;
10759 * normalize group quota/period to be quota/max_period
10760 * note: units are usecs
10762 static u64 normalize_cfs_quota(struct task_group *tg,
10763 struct cfs_schedulable_data *d)
10768 period = d->period;
10771 period = tg_get_cfs_period(tg);
10772 quota = tg_get_cfs_quota(tg);
10775 /* note: these should typically be equivalent */
10776 if (quota == RUNTIME_INF || quota == -1)
10777 return RUNTIME_INF;
10779 return to_ratio(period, quota);
10782 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10784 struct cfs_schedulable_data *d = data;
10785 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10786 s64 quota = 0, parent_quota = -1;
10789 quota = RUNTIME_INF;
10791 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10793 quota = normalize_cfs_quota(tg, d);
10794 parent_quota = parent_b->hierarchical_quota;
10797 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10798 * always take the min. On cgroup1, only inherit when no
10801 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10802 quota = min(quota, parent_quota);
10804 if (quota == RUNTIME_INF)
10805 quota = parent_quota;
10806 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10810 cfs_b->hierarchical_quota = quota;
10815 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10818 struct cfs_schedulable_data data = {
10824 if (quota != RUNTIME_INF) {
10825 do_div(data.period, NSEC_PER_USEC);
10826 do_div(data.quota, NSEC_PER_USEC);
10830 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10836 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10838 struct task_group *tg = css_tg(seq_css(sf));
10839 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10841 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10842 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10843 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10845 if (schedstat_enabled() && tg != &root_task_group) {
10846 struct sched_statistics *stats;
10850 for_each_possible_cpu(i) {
10851 stats = __schedstats_from_se(tg->se[i]);
10852 ws += schedstat_val(stats->wait_sum);
10855 seq_printf(sf, "wait_sum %llu\n", ws);
10858 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10859 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10863 #endif /* CONFIG_CFS_BANDWIDTH */
10864 #endif /* CONFIG_FAIR_GROUP_SCHED */
10866 #ifdef CONFIG_RT_GROUP_SCHED
10867 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10868 struct cftype *cft, s64 val)
10870 return sched_group_set_rt_runtime(css_tg(css), val);
10873 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10874 struct cftype *cft)
10876 return sched_group_rt_runtime(css_tg(css));
10879 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10880 struct cftype *cftype, u64 rt_period_us)
10882 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10885 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10886 struct cftype *cft)
10888 return sched_group_rt_period(css_tg(css));
10890 #endif /* CONFIG_RT_GROUP_SCHED */
10892 #ifdef CONFIG_FAIR_GROUP_SCHED
10893 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10894 struct cftype *cft)
10896 return css_tg(css)->idle;
10899 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10900 struct cftype *cft, s64 idle)
10902 return sched_group_set_idle(css_tg(css), idle);
10906 static struct cftype cpu_legacy_files[] = {
10907 #ifdef CONFIG_FAIR_GROUP_SCHED
10910 .read_u64 = cpu_shares_read_u64,
10911 .write_u64 = cpu_shares_write_u64,
10915 .read_s64 = cpu_idle_read_s64,
10916 .write_s64 = cpu_idle_write_s64,
10919 #ifdef CONFIG_CFS_BANDWIDTH
10921 .name = "cfs_quota_us",
10922 .read_s64 = cpu_cfs_quota_read_s64,
10923 .write_s64 = cpu_cfs_quota_write_s64,
10926 .name = "cfs_period_us",
10927 .read_u64 = cpu_cfs_period_read_u64,
10928 .write_u64 = cpu_cfs_period_write_u64,
10931 .name = "cfs_burst_us",
10932 .read_u64 = cpu_cfs_burst_read_u64,
10933 .write_u64 = cpu_cfs_burst_write_u64,
10937 .seq_show = cpu_cfs_stat_show,
10940 #ifdef CONFIG_RT_GROUP_SCHED
10942 .name = "rt_runtime_us",
10943 .read_s64 = cpu_rt_runtime_read,
10944 .write_s64 = cpu_rt_runtime_write,
10947 .name = "rt_period_us",
10948 .read_u64 = cpu_rt_period_read_uint,
10949 .write_u64 = cpu_rt_period_write_uint,
10952 #ifdef CONFIG_UCLAMP_TASK_GROUP
10954 .name = "uclamp.min",
10955 .flags = CFTYPE_NOT_ON_ROOT,
10956 .seq_show = cpu_uclamp_min_show,
10957 .write = cpu_uclamp_min_write,
10960 .name = "uclamp.max",
10961 .flags = CFTYPE_NOT_ON_ROOT,
10962 .seq_show = cpu_uclamp_max_show,
10963 .write = cpu_uclamp_max_write,
10966 { } /* Terminate */
10969 static int cpu_extra_stat_show(struct seq_file *sf,
10970 struct cgroup_subsys_state *css)
10972 #ifdef CONFIG_CFS_BANDWIDTH
10974 struct task_group *tg = css_tg(css);
10975 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10976 u64 throttled_usec, burst_usec;
10978 throttled_usec = cfs_b->throttled_time;
10979 do_div(throttled_usec, NSEC_PER_USEC);
10980 burst_usec = cfs_b->burst_time;
10981 do_div(burst_usec, NSEC_PER_USEC);
10983 seq_printf(sf, "nr_periods %d\n"
10984 "nr_throttled %d\n"
10985 "throttled_usec %llu\n"
10987 "burst_usec %llu\n",
10988 cfs_b->nr_periods, cfs_b->nr_throttled,
10989 throttled_usec, cfs_b->nr_burst, burst_usec);
10995 #ifdef CONFIG_FAIR_GROUP_SCHED
10996 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10997 struct cftype *cft)
10999 struct task_group *tg = css_tg(css);
11000 u64 weight = scale_load_down(tg->shares);
11002 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11005 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11006 struct cftype *cft, u64 weight)
11009 * cgroup weight knobs should use the common MIN, DFL and MAX
11010 * values which are 1, 100 and 10000 respectively. While it loses
11011 * a bit of range on both ends, it maps pretty well onto the shares
11012 * value used by scheduler and the round-trip conversions preserve
11013 * the original value over the entire range.
11015 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11018 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11020 return sched_group_set_shares(css_tg(css), scale_load(weight));
11023 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11024 struct cftype *cft)
11026 unsigned long weight = scale_load_down(css_tg(css)->shares);
11027 int last_delta = INT_MAX;
11030 /* find the closest nice value to the current weight */
11031 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11032 delta = abs(sched_prio_to_weight[prio] - weight);
11033 if (delta >= last_delta)
11035 last_delta = delta;
11038 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11041 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11042 struct cftype *cft, s64 nice)
11044 unsigned long weight;
11047 if (nice < MIN_NICE || nice > MAX_NICE)
11050 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11051 idx = array_index_nospec(idx, 40);
11052 weight = sched_prio_to_weight[idx];
11054 return sched_group_set_shares(css_tg(css), scale_load(weight));
11058 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11059 long period, long quota)
11062 seq_puts(sf, "max");
11064 seq_printf(sf, "%ld", quota);
11066 seq_printf(sf, " %ld\n", period);
11069 /* caller should put the current value in *@periodp before calling */
11070 static int __maybe_unused cpu_period_quota_parse(char *buf,
11071 u64 *periodp, u64 *quotap)
11073 char tok[21]; /* U64_MAX */
11075 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11078 *periodp *= NSEC_PER_USEC;
11080 if (sscanf(tok, "%llu", quotap))
11081 *quotap *= NSEC_PER_USEC;
11082 else if (!strcmp(tok, "max"))
11083 *quotap = RUNTIME_INF;
11090 #ifdef CONFIG_CFS_BANDWIDTH
11091 static int cpu_max_show(struct seq_file *sf, void *v)
11093 struct task_group *tg = css_tg(seq_css(sf));
11095 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11099 static ssize_t cpu_max_write(struct kernfs_open_file *of,
11100 char *buf, size_t nbytes, loff_t off)
11102 struct task_group *tg = css_tg(of_css(of));
11103 u64 period = tg_get_cfs_period(tg);
11104 u64 burst = tg_get_cfs_burst(tg);
11108 ret = cpu_period_quota_parse(buf, &period, "a);
11110 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11111 return ret ?: nbytes;
11115 static struct cftype cpu_files[] = {
11116 #ifdef CONFIG_FAIR_GROUP_SCHED
11119 .flags = CFTYPE_NOT_ON_ROOT,
11120 .read_u64 = cpu_weight_read_u64,
11121 .write_u64 = cpu_weight_write_u64,
11124 .name = "weight.nice",
11125 .flags = CFTYPE_NOT_ON_ROOT,
11126 .read_s64 = cpu_weight_nice_read_s64,
11127 .write_s64 = cpu_weight_nice_write_s64,
11131 .flags = CFTYPE_NOT_ON_ROOT,
11132 .read_s64 = cpu_idle_read_s64,
11133 .write_s64 = cpu_idle_write_s64,
11136 #ifdef CONFIG_CFS_BANDWIDTH
11139 .flags = CFTYPE_NOT_ON_ROOT,
11140 .seq_show = cpu_max_show,
11141 .write = cpu_max_write,
11144 .name = "max.burst",
11145 .flags = CFTYPE_NOT_ON_ROOT,
11146 .read_u64 = cpu_cfs_burst_read_u64,
11147 .write_u64 = cpu_cfs_burst_write_u64,
11150 #ifdef CONFIG_UCLAMP_TASK_GROUP
11152 .name = "uclamp.min",
11153 .flags = CFTYPE_NOT_ON_ROOT,
11154 .seq_show = cpu_uclamp_min_show,
11155 .write = cpu_uclamp_min_write,
11158 .name = "uclamp.max",
11159 .flags = CFTYPE_NOT_ON_ROOT,
11160 .seq_show = cpu_uclamp_max_show,
11161 .write = cpu_uclamp_max_write,
11164 { } /* terminate */
11167 struct cgroup_subsys cpu_cgrp_subsys = {
11168 .css_alloc = cpu_cgroup_css_alloc,
11169 .css_online = cpu_cgroup_css_online,
11170 .css_released = cpu_cgroup_css_released,
11171 .css_free = cpu_cgroup_css_free,
11172 .css_extra_stat_show = cpu_extra_stat_show,
11173 .fork = cpu_cgroup_fork,
11174 .can_attach = cpu_cgroup_can_attach,
11175 .attach = cpu_cgroup_attach,
11176 .legacy_cftypes = cpu_legacy_files,
11177 .dfl_cftypes = cpu_files,
11178 .early_init = true,
11182 #endif /* CONFIG_CGROUP_SCHED */
11184 void dump_cpu_task(int cpu)
11186 pr_info("Task dump for CPU %d:\n", cpu);
11187 sched_show_task(cpu_curr(cpu));
11191 * Nice levels are multiplicative, with a gentle 10% change for every
11192 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11193 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11194 * that remained on nice 0.
11196 * The "10% effect" is relative and cumulative: from _any_ nice level,
11197 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11198 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11199 * If a task goes up by ~10% and another task goes down by ~10% then
11200 * the relative distance between them is ~25%.)
11202 const int sched_prio_to_weight[40] = {
11203 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11204 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11205 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11206 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11207 /* 0 */ 1024, 820, 655, 526, 423,
11208 /* 5 */ 335, 272, 215, 172, 137,
11209 /* 10 */ 110, 87, 70, 56, 45,
11210 /* 15 */ 36, 29, 23, 18, 15,
11214 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11216 * In cases where the weight does not change often, we can use the
11217 * precalculated inverse to speed up arithmetics by turning divisions
11218 * into multiplications:
11220 const u32 sched_prio_to_wmult[40] = {
11221 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11222 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11223 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11224 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11225 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11226 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11227 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11228 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11231 void call_trace_sched_update_nr_running(struct rq *rq, int count)
11233 trace_sched_update_nr_running_tp(rq, count);