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 "../../fs/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) _old, _val = *_ptr; \
879 _old = cmpxchg(_ptr, _val, _val | _mask); \
887 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
889 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
890 * this avoids any races wrt polling state changes and thereby avoids
893 static bool set_nr_and_not_polling(struct task_struct *p)
895 struct thread_info *ti = task_thread_info(p);
896 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
900 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
902 * If this returns true, then the idle task promises to call
903 * sched_ttwu_pending() and reschedule soon.
905 static bool set_nr_if_polling(struct task_struct *p)
907 struct thread_info *ti = task_thread_info(p);
908 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
911 if (!(val & _TIF_POLLING_NRFLAG))
913 if (val & _TIF_NEED_RESCHED)
915 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
924 static bool set_nr_and_not_polling(struct task_struct *p)
926 set_tsk_need_resched(p);
931 static bool set_nr_if_polling(struct task_struct *p)
938 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
940 struct wake_q_node *node = &task->wake_q;
943 * Atomically grab the task, if ->wake_q is !nil already it means
944 * it's already queued (either by us or someone else) and will get the
945 * wakeup due to that.
947 * In order to ensure that a pending wakeup will observe our pending
948 * state, even in the failed case, an explicit smp_mb() must be used.
950 smp_mb__before_atomic();
951 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
955 * The head is context local, there can be no concurrency.
958 head->lastp = &node->next;
963 * wake_q_add() - queue a wakeup for 'later' waking.
964 * @head: the wake_q_head to add @task to
965 * @task: the task to queue for 'later' wakeup
967 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
968 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
971 * This function must be used as-if it were wake_up_process(); IOW the task
972 * must be ready to be woken at this location.
974 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
976 if (__wake_q_add(head, task))
977 get_task_struct(task);
981 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
982 * @head: the wake_q_head to add @task to
983 * @task: the task to queue for 'later' wakeup
985 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
986 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
989 * This function must be used as-if it were wake_up_process(); IOW the task
990 * must be ready to be woken at this location.
992 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
993 * that already hold reference to @task can call the 'safe' version and trust
994 * wake_q to do the right thing depending whether or not the @task is already
997 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
999 if (!__wake_q_add(head, task))
1000 put_task_struct(task);
1003 void wake_up_q(struct wake_q_head *head)
1005 struct wake_q_node *node = head->first;
1007 while (node != WAKE_Q_TAIL) {
1008 struct task_struct *task;
1010 task = container_of(node, struct task_struct, wake_q);
1011 /* Task can safely be re-inserted now: */
1013 task->wake_q.next = NULL;
1016 * wake_up_process() executes a full barrier, which pairs with
1017 * the queueing in wake_q_add() so as not to miss wakeups.
1019 wake_up_process(task);
1020 put_task_struct(task);
1025 * resched_curr - mark rq's current task 'to be rescheduled now'.
1027 * On UP this means the setting of the need_resched flag, on SMP it
1028 * might also involve a cross-CPU call to trigger the scheduler on
1031 void resched_curr(struct rq *rq)
1033 struct task_struct *curr = rq->curr;
1036 lockdep_assert_rq_held(rq);
1038 if (test_tsk_need_resched(curr))
1043 if (cpu == smp_processor_id()) {
1044 set_tsk_need_resched(curr);
1045 set_preempt_need_resched();
1049 if (set_nr_and_not_polling(curr))
1050 smp_send_reschedule(cpu);
1052 trace_sched_wake_idle_without_ipi(cpu);
1055 void resched_cpu(int cpu)
1057 struct rq *rq = cpu_rq(cpu);
1058 unsigned long flags;
1060 raw_spin_rq_lock_irqsave(rq, flags);
1061 if (cpu_online(cpu) || cpu == smp_processor_id())
1063 raw_spin_rq_unlock_irqrestore(rq, flags);
1067 #ifdef CONFIG_NO_HZ_COMMON
1069 * In the semi idle case, use the nearest busy CPU for migrating timers
1070 * from an idle CPU. This is good for power-savings.
1072 * We don't do similar optimization for completely idle system, as
1073 * selecting an idle CPU will add more delays to the timers than intended
1074 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1076 int get_nohz_timer_target(void)
1078 int i, cpu = smp_processor_id(), default_cpu = -1;
1079 struct sched_domain *sd;
1080 const struct cpumask *hk_mask;
1082 if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1088 hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1091 for_each_domain(cpu, sd) {
1092 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1103 if (default_cpu == -1)
1104 default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1112 * When add_timer_on() enqueues a timer into the timer wheel of an
1113 * idle CPU then this timer might expire before the next timer event
1114 * which is scheduled to wake up that CPU. In case of a completely
1115 * idle system the next event might even be infinite time into the
1116 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1117 * leaves the inner idle loop so the newly added timer is taken into
1118 * account when the CPU goes back to idle and evaluates the timer
1119 * wheel for the next timer event.
1121 static void wake_up_idle_cpu(int cpu)
1123 struct rq *rq = cpu_rq(cpu);
1125 if (cpu == smp_processor_id())
1128 if (set_nr_and_not_polling(rq->idle))
1129 smp_send_reschedule(cpu);
1131 trace_sched_wake_idle_without_ipi(cpu);
1134 static bool wake_up_full_nohz_cpu(int cpu)
1137 * We just need the target to call irq_exit() and re-evaluate
1138 * the next tick. The nohz full kick at least implies that.
1139 * If needed we can still optimize that later with an
1142 if (cpu_is_offline(cpu))
1143 return true; /* Don't try to wake offline CPUs. */
1144 if (tick_nohz_full_cpu(cpu)) {
1145 if (cpu != smp_processor_id() ||
1146 tick_nohz_tick_stopped())
1147 tick_nohz_full_kick_cpu(cpu);
1155 * Wake up the specified CPU. If the CPU is going offline, it is the
1156 * caller's responsibility to deal with the lost wakeup, for example,
1157 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1159 void wake_up_nohz_cpu(int cpu)
1161 if (!wake_up_full_nohz_cpu(cpu))
1162 wake_up_idle_cpu(cpu);
1165 static void nohz_csd_func(void *info)
1167 struct rq *rq = info;
1168 int cpu = cpu_of(rq);
1172 * Release the rq::nohz_csd.
1174 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1175 WARN_ON(!(flags & NOHZ_KICK_MASK));
1177 rq->idle_balance = idle_cpu(cpu);
1178 if (rq->idle_balance && !need_resched()) {
1179 rq->nohz_idle_balance = flags;
1180 raise_softirq_irqoff(SCHED_SOFTIRQ);
1184 #endif /* CONFIG_NO_HZ_COMMON */
1186 #ifdef CONFIG_NO_HZ_FULL
1187 bool sched_can_stop_tick(struct rq *rq)
1189 int fifo_nr_running;
1191 /* Deadline tasks, even if single, need the tick */
1192 if (rq->dl.dl_nr_running)
1196 * If there are more than one RR tasks, we need the tick to affect the
1197 * actual RR behaviour.
1199 if (rq->rt.rr_nr_running) {
1200 if (rq->rt.rr_nr_running == 1)
1207 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1208 * forced preemption between FIFO tasks.
1210 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1211 if (fifo_nr_running)
1215 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1216 * if there's more than one we need the tick for involuntary
1219 if (rq->nr_running > 1)
1224 #endif /* CONFIG_NO_HZ_FULL */
1225 #endif /* CONFIG_SMP */
1227 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1228 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1230 * Iterate task_group tree rooted at *from, calling @down when first entering a
1231 * node and @up when leaving it for the final time.
1233 * Caller must hold rcu_lock or sufficient equivalent.
1235 int walk_tg_tree_from(struct task_group *from,
1236 tg_visitor down, tg_visitor up, void *data)
1238 struct task_group *parent, *child;
1244 ret = (*down)(parent, data);
1247 list_for_each_entry_rcu(child, &parent->children, siblings) {
1254 ret = (*up)(parent, data);
1255 if (ret || parent == from)
1259 parent = parent->parent;
1266 int tg_nop(struct task_group *tg, void *data)
1272 static void set_load_weight(struct task_struct *p, bool update_load)
1274 int prio = p->static_prio - MAX_RT_PRIO;
1275 struct load_weight *load = &p->se.load;
1278 * SCHED_IDLE tasks get minimal weight:
1280 if (task_has_idle_policy(p)) {
1281 load->weight = scale_load(WEIGHT_IDLEPRIO);
1282 load->inv_weight = WMULT_IDLEPRIO;
1287 * SCHED_OTHER tasks have to update their load when changing their
1290 if (update_load && p->sched_class == &fair_sched_class) {
1291 reweight_task(p, prio);
1293 load->weight = scale_load(sched_prio_to_weight[prio]);
1294 load->inv_weight = sched_prio_to_wmult[prio];
1298 #ifdef CONFIG_UCLAMP_TASK
1300 * Serializes updates of utilization clamp values
1302 * The (slow-path) user-space triggers utilization clamp value updates which
1303 * can require updates on (fast-path) scheduler's data structures used to
1304 * support enqueue/dequeue operations.
1305 * While the per-CPU rq lock protects fast-path update operations, user-space
1306 * requests are serialized using a mutex to reduce the risk of conflicting
1307 * updates or API abuses.
1309 static DEFINE_MUTEX(uclamp_mutex);
1311 /* Max allowed minimum utilization */
1312 static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1314 /* Max allowed maximum utilization */
1315 static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1318 * By default RT tasks run at the maximum performance point/capacity of the
1319 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1320 * SCHED_CAPACITY_SCALE.
1322 * This knob allows admins to change the default behavior when uclamp is being
1323 * used. In battery powered devices, particularly, running at the maximum
1324 * capacity and frequency will increase energy consumption and shorten the
1327 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1329 * This knob will not override the system default sched_util_clamp_min defined
1332 static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1334 /* All clamps are required to be less or equal than these values */
1335 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1338 * This static key is used to reduce the uclamp overhead in the fast path. It
1339 * primarily disables the call to uclamp_rq_{inc, dec}() in
1340 * enqueue/dequeue_task().
1342 * This allows users to continue to enable uclamp in their kernel config with
1343 * minimum uclamp overhead in the fast path.
1345 * As soon as userspace modifies any of the uclamp knobs, the static key is
1346 * enabled, since we have an actual users that make use of uclamp
1349 * The knobs that would enable this static key are:
1351 * * A task modifying its uclamp value with sched_setattr().
1352 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1353 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1355 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1357 /* Integer rounded range for each bucket */
1358 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1360 #define for_each_clamp_id(clamp_id) \
1361 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1363 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1365 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1368 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1370 if (clamp_id == UCLAMP_MIN)
1372 return SCHED_CAPACITY_SCALE;
1375 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1376 unsigned int value, bool user_defined)
1378 uc_se->value = value;
1379 uc_se->bucket_id = uclamp_bucket_id(value);
1380 uc_se->user_defined = user_defined;
1383 static inline unsigned int
1384 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1385 unsigned int clamp_value)
1388 * Avoid blocked utilization pushing up the frequency when we go
1389 * idle (which drops the max-clamp) by retaining the last known
1392 if (clamp_id == UCLAMP_MAX) {
1393 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1397 return uclamp_none(UCLAMP_MIN);
1400 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1401 unsigned int clamp_value)
1403 /* Reset max-clamp retention only on idle exit */
1404 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1407 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1411 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1412 unsigned int clamp_value)
1414 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1415 int bucket_id = UCLAMP_BUCKETS - 1;
1418 * Since both min and max clamps are max aggregated, find the
1419 * top most bucket with tasks in.
1421 for ( ; bucket_id >= 0; bucket_id--) {
1422 if (!bucket[bucket_id].tasks)
1424 return bucket[bucket_id].value;
1427 /* No tasks -- default clamp values */
1428 return uclamp_idle_value(rq, clamp_id, clamp_value);
1431 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1433 unsigned int default_util_min;
1434 struct uclamp_se *uc_se;
1436 lockdep_assert_held(&p->pi_lock);
1438 uc_se = &p->uclamp_req[UCLAMP_MIN];
1440 /* Only sync if user didn't override the default */
1441 if (uc_se->user_defined)
1444 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1445 uclamp_se_set(uc_se, default_util_min, false);
1448 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1456 /* Protect updates to p->uclamp_* */
1457 rq = task_rq_lock(p, &rf);
1458 __uclamp_update_util_min_rt_default(p);
1459 task_rq_unlock(rq, p, &rf);
1462 static inline struct uclamp_se
1463 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1465 /* Copy by value as we could modify it */
1466 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1467 #ifdef CONFIG_UCLAMP_TASK_GROUP
1468 unsigned int tg_min, tg_max, value;
1471 * Tasks in autogroups or root task group will be
1472 * restricted by system defaults.
1474 if (task_group_is_autogroup(task_group(p)))
1476 if (task_group(p) == &root_task_group)
1479 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1480 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1481 value = uc_req.value;
1482 value = clamp(value, tg_min, tg_max);
1483 uclamp_se_set(&uc_req, value, false);
1490 * The effective clamp bucket index of a task depends on, by increasing
1492 * - the task specific clamp value, when explicitly requested from userspace
1493 * - the task group effective clamp value, for tasks not either in the root
1494 * group or in an autogroup
1495 * - the system default clamp value, defined by the sysadmin
1497 static inline struct uclamp_se
1498 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1500 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1501 struct uclamp_se uc_max = uclamp_default[clamp_id];
1503 /* System default restrictions always apply */
1504 if (unlikely(uc_req.value > uc_max.value))
1510 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1512 struct uclamp_se uc_eff;
1514 /* Task currently refcounted: use back-annotated (effective) value */
1515 if (p->uclamp[clamp_id].active)
1516 return (unsigned long)p->uclamp[clamp_id].value;
1518 uc_eff = uclamp_eff_get(p, clamp_id);
1520 return (unsigned long)uc_eff.value;
1524 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1525 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1526 * updates the rq's clamp value if required.
1528 * Tasks can have a task-specific value requested from user-space, track
1529 * within each bucket the maximum value for tasks refcounted in it.
1530 * This "local max aggregation" allows to track the exact "requested" value
1531 * for each bucket when all its RUNNABLE tasks require the same clamp.
1533 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1534 enum uclamp_id clamp_id)
1536 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1537 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1538 struct uclamp_bucket *bucket;
1540 lockdep_assert_rq_held(rq);
1542 /* Update task effective clamp */
1543 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1545 bucket = &uc_rq->bucket[uc_se->bucket_id];
1547 uc_se->active = true;
1549 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1552 * Local max aggregation: rq buckets always track the max
1553 * "requested" clamp value of its RUNNABLE tasks.
1555 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1556 bucket->value = uc_se->value;
1558 if (uc_se->value > READ_ONCE(uc_rq->value))
1559 WRITE_ONCE(uc_rq->value, uc_se->value);
1563 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1564 * is released. If this is the last task reference counting the rq's max
1565 * active clamp value, then the rq's clamp value is updated.
1567 * Both refcounted tasks and rq's cached clamp values are expected to be
1568 * always valid. If it's detected they are not, as defensive programming,
1569 * enforce the expected state and warn.
1571 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1572 enum uclamp_id clamp_id)
1574 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1575 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1576 struct uclamp_bucket *bucket;
1577 unsigned int bkt_clamp;
1578 unsigned int rq_clamp;
1580 lockdep_assert_rq_held(rq);
1583 * If sched_uclamp_used was enabled after task @p was enqueued,
1584 * we could end up with unbalanced call to uclamp_rq_dec_id().
1586 * In this case the uc_se->active flag should be false since no uclamp
1587 * accounting was performed at enqueue time and we can just return
1590 * Need to be careful of the following enqueue/dequeue ordering
1594 * // sched_uclamp_used gets enabled
1597 * // Must not decrement bucket->tasks here
1600 * where we could end up with stale data in uc_se and
1601 * bucket[uc_se->bucket_id].
1603 * The following check here eliminates the possibility of such race.
1605 if (unlikely(!uc_se->active))
1608 bucket = &uc_rq->bucket[uc_se->bucket_id];
1610 SCHED_WARN_ON(!bucket->tasks);
1611 if (likely(bucket->tasks))
1614 uc_se->active = false;
1617 * Keep "local max aggregation" simple and accept to (possibly)
1618 * overboost some RUNNABLE tasks in the same bucket.
1619 * The rq clamp bucket value is reset to its base value whenever
1620 * there are no more RUNNABLE tasks refcounting it.
1622 if (likely(bucket->tasks))
1625 rq_clamp = READ_ONCE(uc_rq->value);
1627 * Defensive programming: this should never happen. If it happens,
1628 * e.g. due to future modification, warn and fixup the expected value.
1630 SCHED_WARN_ON(bucket->value > rq_clamp);
1631 if (bucket->value >= rq_clamp) {
1632 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1633 WRITE_ONCE(uc_rq->value, bkt_clamp);
1637 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1639 enum uclamp_id clamp_id;
1642 * Avoid any overhead until uclamp is actually used by the userspace.
1644 * The condition is constructed such that a NOP is generated when
1645 * sched_uclamp_used is disabled.
1647 if (!static_branch_unlikely(&sched_uclamp_used))
1650 if (unlikely(!p->sched_class->uclamp_enabled))
1653 for_each_clamp_id(clamp_id)
1654 uclamp_rq_inc_id(rq, p, clamp_id);
1656 /* Reset clamp idle holding when there is one RUNNABLE task */
1657 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1658 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1661 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1663 enum uclamp_id clamp_id;
1666 * Avoid any overhead until uclamp is actually used by the userspace.
1668 * The condition is constructed such that a NOP is generated when
1669 * sched_uclamp_used is disabled.
1671 if (!static_branch_unlikely(&sched_uclamp_used))
1674 if (unlikely(!p->sched_class->uclamp_enabled))
1677 for_each_clamp_id(clamp_id)
1678 uclamp_rq_dec_id(rq, p, clamp_id);
1681 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1682 enum uclamp_id clamp_id)
1684 if (!p->uclamp[clamp_id].active)
1687 uclamp_rq_dec_id(rq, p, clamp_id);
1688 uclamp_rq_inc_id(rq, p, clamp_id);
1691 * Make sure to clear the idle flag if we've transiently reached 0
1692 * active tasks on rq.
1694 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1695 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1699 uclamp_update_active(struct task_struct *p)
1701 enum uclamp_id clamp_id;
1706 * Lock the task and the rq where the task is (or was) queued.
1708 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1709 * price to pay to safely serialize util_{min,max} updates with
1710 * enqueues, dequeues and migration operations.
1711 * This is the same locking schema used by __set_cpus_allowed_ptr().
1713 rq = task_rq_lock(p, &rf);
1716 * Setting the clamp bucket is serialized by task_rq_lock().
1717 * If the task is not yet RUNNABLE and its task_struct is not
1718 * affecting a valid clamp bucket, the next time it's enqueued,
1719 * it will already see the updated clamp bucket value.
1721 for_each_clamp_id(clamp_id)
1722 uclamp_rq_reinc_id(rq, p, clamp_id);
1724 task_rq_unlock(rq, p, &rf);
1727 #ifdef CONFIG_UCLAMP_TASK_GROUP
1729 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1731 struct css_task_iter it;
1732 struct task_struct *p;
1734 css_task_iter_start(css, 0, &it);
1735 while ((p = css_task_iter_next(&it)))
1736 uclamp_update_active(p);
1737 css_task_iter_end(&it);
1740 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1743 #ifdef CONFIG_SYSCTL
1744 #ifdef CONFIG_UCLAMP_TASK
1745 #ifdef CONFIG_UCLAMP_TASK_GROUP
1746 static void uclamp_update_root_tg(void)
1748 struct task_group *tg = &root_task_group;
1750 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1751 sysctl_sched_uclamp_util_min, false);
1752 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1753 sysctl_sched_uclamp_util_max, false);
1756 cpu_util_update_eff(&root_task_group.css);
1760 static void uclamp_update_root_tg(void) { }
1763 static void uclamp_sync_util_min_rt_default(void)
1765 struct task_struct *g, *p;
1768 * copy_process() sysctl_uclamp
1769 * uclamp_min_rt = X;
1770 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1771 * // link thread smp_mb__after_spinlock()
1772 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1773 * sched_post_fork() for_each_process_thread()
1774 * __uclamp_sync_rt() __uclamp_sync_rt()
1776 * Ensures that either sched_post_fork() will observe the new
1777 * uclamp_min_rt or for_each_process_thread() will observe the new
1780 read_lock(&tasklist_lock);
1781 smp_mb__after_spinlock();
1782 read_unlock(&tasklist_lock);
1785 for_each_process_thread(g, p)
1786 uclamp_update_util_min_rt_default(p);
1790 static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1791 void *buffer, size_t *lenp, loff_t *ppos)
1793 bool update_root_tg = false;
1794 int old_min, old_max, old_min_rt;
1797 mutex_lock(&uclamp_mutex);
1798 old_min = sysctl_sched_uclamp_util_min;
1799 old_max = sysctl_sched_uclamp_util_max;
1800 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1802 result = proc_dointvec(table, write, buffer, lenp, ppos);
1808 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1809 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1810 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1816 if (old_min != sysctl_sched_uclamp_util_min) {
1817 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1818 sysctl_sched_uclamp_util_min, false);
1819 update_root_tg = true;
1821 if (old_max != sysctl_sched_uclamp_util_max) {
1822 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1823 sysctl_sched_uclamp_util_max, false);
1824 update_root_tg = true;
1827 if (update_root_tg) {
1828 static_branch_enable(&sched_uclamp_used);
1829 uclamp_update_root_tg();
1832 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1833 static_branch_enable(&sched_uclamp_used);
1834 uclamp_sync_util_min_rt_default();
1838 * We update all RUNNABLE tasks only when task groups are in use.
1839 * Otherwise, keep it simple and do just a lazy update at each next
1840 * task enqueue time.
1846 sysctl_sched_uclamp_util_min = old_min;
1847 sysctl_sched_uclamp_util_max = old_max;
1848 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1850 mutex_unlock(&uclamp_mutex);
1857 static int uclamp_validate(struct task_struct *p,
1858 const struct sched_attr *attr)
1860 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1861 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1863 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1864 util_min = attr->sched_util_min;
1866 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1870 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1871 util_max = attr->sched_util_max;
1873 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1877 if (util_min != -1 && util_max != -1 && util_min > util_max)
1881 * We have valid uclamp attributes; make sure uclamp is enabled.
1883 * We need to do that here, because enabling static branches is a
1884 * blocking operation which obviously cannot be done while holding
1887 static_branch_enable(&sched_uclamp_used);
1892 static bool uclamp_reset(const struct sched_attr *attr,
1893 enum uclamp_id clamp_id,
1894 struct uclamp_se *uc_se)
1896 /* Reset on sched class change for a non user-defined clamp value. */
1897 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1898 !uc_se->user_defined)
1901 /* Reset on sched_util_{min,max} == -1. */
1902 if (clamp_id == UCLAMP_MIN &&
1903 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1904 attr->sched_util_min == -1) {
1908 if (clamp_id == UCLAMP_MAX &&
1909 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1910 attr->sched_util_max == -1) {
1917 static void __setscheduler_uclamp(struct task_struct *p,
1918 const struct sched_attr *attr)
1920 enum uclamp_id clamp_id;
1922 for_each_clamp_id(clamp_id) {
1923 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1926 if (!uclamp_reset(attr, clamp_id, uc_se))
1930 * RT by default have a 100% boost value that could be modified
1933 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1934 value = sysctl_sched_uclamp_util_min_rt_default;
1936 value = uclamp_none(clamp_id);
1938 uclamp_se_set(uc_se, value, false);
1942 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1945 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1946 attr->sched_util_min != -1) {
1947 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1948 attr->sched_util_min, true);
1951 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1952 attr->sched_util_max != -1) {
1953 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1954 attr->sched_util_max, true);
1958 static void uclamp_fork(struct task_struct *p)
1960 enum uclamp_id clamp_id;
1963 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1964 * as the task is still at its early fork stages.
1966 for_each_clamp_id(clamp_id)
1967 p->uclamp[clamp_id].active = false;
1969 if (likely(!p->sched_reset_on_fork))
1972 for_each_clamp_id(clamp_id) {
1973 uclamp_se_set(&p->uclamp_req[clamp_id],
1974 uclamp_none(clamp_id), false);
1978 static void uclamp_post_fork(struct task_struct *p)
1980 uclamp_update_util_min_rt_default(p);
1983 static void __init init_uclamp_rq(struct rq *rq)
1985 enum uclamp_id clamp_id;
1986 struct uclamp_rq *uc_rq = rq->uclamp;
1988 for_each_clamp_id(clamp_id) {
1989 uc_rq[clamp_id] = (struct uclamp_rq) {
1990 .value = uclamp_none(clamp_id)
1994 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1997 static void __init init_uclamp(void)
1999 struct uclamp_se uc_max = {};
2000 enum uclamp_id clamp_id;
2003 for_each_possible_cpu(cpu)
2004 init_uclamp_rq(cpu_rq(cpu));
2006 for_each_clamp_id(clamp_id) {
2007 uclamp_se_set(&init_task.uclamp_req[clamp_id],
2008 uclamp_none(clamp_id), false);
2011 /* System defaults allow max clamp values for both indexes */
2012 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2013 for_each_clamp_id(clamp_id) {
2014 uclamp_default[clamp_id] = uc_max;
2015 #ifdef CONFIG_UCLAMP_TASK_GROUP
2016 root_task_group.uclamp_req[clamp_id] = uc_max;
2017 root_task_group.uclamp[clamp_id] = uc_max;
2022 #else /* CONFIG_UCLAMP_TASK */
2023 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2024 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
2025 static inline int uclamp_validate(struct task_struct *p,
2026 const struct sched_attr *attr)
2030 static void __setscheduler_uclamp(struct task_struct *p,
2031 const struct sched_attr *attr) { }
2032 static inline void uclamp_fork(struct task_struct *p) { }
2033 static inline void uclamp_post_fork(struct task_struct *p) { }
2034 static inline void init_uclamp(void) { }
2035 #endif /* CONFIG_UCLAMP_TASK */
2037 bool sched_task_on_rq(struct task_struct *p)
2039 return task_on_rq_queued(p);
2042 unsigned long get_wchan(struct task_struct *p)
2044 unsigned long ip = 0;
2047 if (!p || p == current)
2050 /* Only get wchan if task is blocked and we can keep it that way. */
2051 raw_spin_lock_irq(&p->pi_lock);
2052 state = READ_ONCE(p->__state);
2053 smp_rmb(); /* see try_to_wake_up() */
2054 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2055 ip = __get_wchan(p);
2056 raw_spin_unlock_irq(&p->pi_lock);
2061 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2063 if (!(flags & ENQUEUE_NOCLOCK))
2064 update_rq_clock(rq);
2066 if (!(flags & ENQUEUE_RESTORE)) {
2067 sched_info_enqueue(rq, p);
2068 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
2071 uclamp_rq_inc(rq, p);
2072 p->sched_class->enqueue_task(rq, p, flags);
2074 if (sched_core_enabled(rq))
2075 sched_core_enqueue(rq, p);
2078 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2080 if (sched_core_enabled(rq))
2081 sched_core_dequeue(rq, p, flags);
2083 if (!(flags & DEQUEUE_NOCLOCK))
2084 update_rq_clock(rq);
2086 if (!(flags & DEQUEUE_SAVE)) {
2087 sched_info_dequeue(rq, p);
2088 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2091 uclamp_rq_dec(rq, p);
2092 p->sched_class->dequeue_task(rq, p, flags);
2095 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2097 enqueue_task(rq, p, flags);
2099 p->on_rq = TASK_ON_RQ_QUEUED;
2102 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2104 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2106 dequeue_task(rq, p, flags);
2109 static inline int __normal_prio(int policy, int rt_prio, int nice)
2113 if (dl_policy(policy))
2114 prio = MAX_DL_PRIO - 1;
2115 else if (rt_policy(policy))
2116 prio = MAX_RT_PRIO - 1 - rt_prio;
2118 prio = NICE_TO_PRIO(nice);
2124 * Calculate the expected normal priority: i.e. priority
2125 * without taking RT-inheritance into account. Might be
2126 * boosted by interactivity modifiers. Changes upon fork,
2127 * setprio syscalls, and whenever the interactivity
2128 * estimator recalculates.
2130 static inline int normal_prio(struct task_struct *p)
2132 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2136 * Calculate the current priority, i.e. the priority
2137 * taken into account by the scheduler. This value might
2138 * be boosted by RT tasks, or might be boosted by
2139 * interactivity modifiers. Will be RT if the task got
2140 * RT-boosted. If not then it returns p->normal_prio.
2142 static int effective_prio(struct task_struct *p)
2144 p->normal_prio = normal_prio(p);
2146 * If we are RT tasks or we were boosted to RT priority,
2147 * keep the priority unchanged. Otherwise, update priority
2148 * to the normal priority:
2150 if (!rt_prio(p->prio))
2151 return p->normal_prio;
2156 * task_curr - is this task currently executing on a CPU?
2157 * @p: the task in question.
2159 * Return: 1 if the task is currently executing. 0 otherwise.
2161 inline int task_curr(const struct task_struct *p)
2163 return cpu_curr(task_cpu(p)) == p;
2167 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2168 * use the balance_callback list if you want balancing.
2170 * this means any call to check_class_changed() must be followed by a call to
2171 * balance_callback().
2173 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2174 const struct sched_class *prev_class,
2177 if (prev_class != p->sched_class) {
2178 if (prev_class->switched_from)
2179 prev_class->switched_from(rq, p);
2181 p->sched_class->switched_to(rq, p);
2182 } else if (oldprio != p->prio || dl_task(p))
2183 p->sched_class->prio_changed(rq, p, oldprio);
2186 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2188 if (p->sched_class == rq->curr->sched_class)
2189 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2190 else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2194 * A queue event has occurred, and we're going to schedule. In
2195 * this case, we can save a useless back to back clock update.
2197 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2198 rq_clock_skip_update(rq);
2204 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2206 static int __set_cpus_allowed_ptr(struct task_struct *p,
2207 const struct cpumask *new_mask,
2210 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2212 if (likely(!p->migration_disabled))
2215 if (p->cpus_ptr != &p->cpus_mask)
2219 * Violates locking rules! see comment in __do_set_cpus_allowed().
2221 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2224 void migrate_disable(void)
2226 struct task_struct *p = current;
2228 if (p->migration_disabled) {
2229 p->migration_disabled++;
2234 this_rq()->nr_pinned++;
2235 p->migration_disabled = 1;
2238 EXPORT_SYMBOL_GPL(migrate_disable);
2240 void migrate_enable(void)
2242 struct task_struct *p = current;
2244 if (p->migration_disabled > 1) {
2245 p->migration_disabled--;
2249 if (WARN_ON_ONCE(!p->migration_disabled))
2253 * Ensure stop_task runs either before or after this, and that
2254 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2257 if (p->cpus_ptr != &p->cpus_mask)
2258 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2260 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2261 * regular cpus_mask, otherwise things that race (eg.
2262 * select_fallback_rq) get confused.
2265 p->migration_disabled = 0;
2266 this_rq()->nr_pinned--;
2269 EXPORT_SYMBOL_GPL(migrate_enable);
2271 static inline bool rq_has_pinned_tasks(struct rq *rq)
2273 return rq->nr_pinned;
2277 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2278 * __set_cpus_allowed_ptr() and select_fallback_rq().
2280 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2282 /* When not in the task's cpumask, no point in looking further. */
2283 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2286 /* migrate_disabled() must be allowed to finish. */
2287 if (is_migration_disabled(p))
2288 return cpu_online(cpu);
2290 /* Non kernel threads are not allowed during either online or offline. */
2291 if (!(p->flags & PF_KTHREAD))
2292 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2294 /* KTHREAD_IS_PER_CPU is always allowed. */
2295 if (kthread_is_per_cpu(p))
2296 return cpu_online(cpu);
2298 /* Regular kernel threads don't get to stay during offline. */
2302 /* But are allowed during online. */
2303 return cpu_online(cpu);
2307 * This is how migration works:
2309 * 1) we invoke migration_cpu_stop() on the target CPU using
2311 * 2) stopper starts to run (implicitly forcing the migrated thread
2313 * 3) it checks whether the migrated task is still in the wrong runqueue.
2314 * 4) if it's in the wrong runqueue then the migration thread removes
2315 * it and puts it into the right queue.
2316 * 5) stopper completes and stop_one_cpu() returns and the migration
2321 * move_queued_task - move a queued task to new rq.
2323 * Returns (locked) new rq. Old rq's lock is released.
2325 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2326 struct task_struct *p, int new_cpu)
2328 lockdep_assert_rq_held(rq);
2330 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2331 set_task_cpu(p, new_cpu);
2334 rq = cpu_rq(new_cpu);
2337 BUG_ON(task_cpu(p) != new_cpu);
2338 activate_task(rq, p, 0);
2339 check_preempt_curr(rq, p, 0);
2344 struct migration_arg {
2345 struct task_struct *task;
2347 struct set_affinity_pending *pending;
2351 * @refs: number of wait_for_completion()
2352 * @stop_pending: is @stop_work in use
2354 struct set_affinity_pending {
2356 unsigned int stop_pending;
2357 struct completion done;
2358 struct cpu_stop_work stop_work;
2359 struct migration_arg arg;
2363 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2364 * this because either it can't run here any more (set_cpus_allowed()
2365 * away from this CPU, or CPU going down), or because we're
2366 * attempting to rebalance this task on exec (sched_exec).
2368 * So we race with normal scheduler movements, but that's OK, as long
2369 * as the task is no longer on this CPU.
2371 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2372 struct task_struct *p, int dest_cpu)
2374 /* Affinity changed (again). */
2375 if (!is_cpu_allowed(p, dest_cpu))
2378 update_rq_clock(rq);
2379 rq = move_queued_task(rq, rf, p, dest_cpu);
2385 * migration_cpu_stop - this will be executed by a highprio stopper thread
2386 * and performs thread migration by bumping thread off CPU then
2387 * 'pushing' onto another runqueue.
2389 static int migration_cpu_stop(void *data)
2391 struct migration_arg *arg = data;
2392 struct set_affinity_pending *pending = arg->pending;
2393 struct task_struct *p = arg->task;
2394 struct rq *rq = this_rq();
2395 bool complete = false;
2399 * The original target CPU might have gone down and we might
2400 * be on another CPU but it doesn't matter.
2402 local_irq_save(rf.flags);
2404 * We need to explicitly wake pending tasks before running
2405 * __migrate_task() such that we will not miss enforcing cpus_ptr
2406 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2408 flush_smp_call_function_queue();
2410 raw_spin_lock(&p->pi_lock);
2414 * If we were passed a pending, then ->stop_pending was set, thus
2415 * p->migration_pending must have remained stable.
2417 WARN_ON_ONCE(pending && pending != p->migration_pending);
2420 * If task_rq(p) != rq, it cannot be migrated here, because we're
2421 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2422 * we're holding p->pi_lock.
2424 if (task_rq(p) == rq) {
2425 if (is_migration_disabled(p))
2429 p->migration_pending = NULL;
2432 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2436 if (task_on_rq_queued(p))
2437 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2439 p->wake_cpu = arg->dest_cpu;
2442 * XXX __migrate_task() can fail, at which point we might end
2443 * up running on a dodgy CPU, AFAICT this can only happen
2444 * during CPU hotplug, at which point we'll get pushed out
2445 * anyway, so it's probably not a big deal.
2448 } else if (pending) {
2450 * This happens when we get migrated between migrate_enable()'s
2451 * preempt_enable() and scheduling the stopper task. At that
2452 * point we're a regular task again and not current anymore.
2454 * A !PREEMPT kernel has a giant hole here, which makes it far
2459 * The task moved before the stopper got to run. We're holding
2460 * ->pi_lock, so the allowed mask is stable - if it got
2461 * somewhere allowed, we're done.
2463 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2464 p->migration_pending = NULL;
2470 * When migrate_enable() hits a rq mis-match we can't reliably
2471 * determine is_migration_disabled() and so have to chase after
2474 WARN_ON_ONCE(!pending->stop_pending);
2475 task_rq_unlock(rq, p, &rf);
2476 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2477 &pending->arg, &pending->stop_work);
2482 pending->stop_pending = false;
2483 task_rq_unlock(rq, p, &rf);
2486 complete_all(&pending->done);
2491 int push_cpu_stop(void *arg)
2493 struct rq *lowest_rq = NULL, *rq = this_rq();
2494 struct task_struct *p = arg;
2496 raw_spin_lock_irq(&p->pi_lock);
2497 raw_spin_rq_lock(rq);
2499 if (task_rq(p) != rq)
2502 if (is_migration_disabled(p)) {
2503 p->migration_flags |= MDF_PUSH;
2507 p->migration_flags &= ~MDF_PUSH;
2509 if (p->sched_class->find_lock_rq)
2510 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2515 // XXX validate p is still the highest prio task
2516 if (task_rq(p) == rq) {
2517 deactivate_task(rq, p, 0);
2518 set_task_cpu(p, lowest_rq->cpu);
2519 activate_task(lowest_rq, p, 0);
2520 resched_curr(lowest_rq);
2523 double_unlock_balance(rq, lowest_rq);
2526 rq->push_busy = false;
2527 raw_spin_rq_unlock(rq);
2528 raw_spin_unlock_irq(&p->pi_lock);
2535 * sched_class::set_cpus_allowed must do the below, but is not required to
2536 * actually call this function.
2538 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2540 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2541 p->cpus_ptr = new_mask;
2545 cpumask_copy(&p->cpus_mask, new_mask);
2546 p->nr_cpus_allowed = cpumask_weight(new_mask);
2550 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2552 struct rq *rq = task_rq(p);
2553 bool queued, running;
2556 * This here violates the locking rules for affinity, since we're only
2557 * supposed to change these variables while holding both rq->lock and
2560 * HOWEVER, it magically works, because ttwu() is the only code that
2561 * accesses these variables under p->pi_lock and only does so after
2562 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2563 * before finish_task().
2565 * XXX do further audits, this smells like something putrid.
2567 if (flags & SCA_MIGRATE_DISABLE)
2568 SCHED_WARN_ON(!p->on_cpu);
2570 lockdep_assert_held(&p->pi_lock);
2572 queued = task_on_rq_queued(p);
2573 running = task_current(rq, p);
2577 * Because __kthread_bind() calls this on blocked tasks without
2580 lockdep_assert_rq_held(rq);
2581 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2584 put_prev_task(rq, p);
2586 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2589 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2591 set_next_task(rq, p);
2594 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2596 __do_set_cpus_allowed(p, new_mask, 0);
2599 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2602 if (!src->user_cpus_ptr)
2605 dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2606 if (!dst->user_cpus_ptr)
2609 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2613 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2615 struct cpumask *user_mask = NULL;
2617 swap(p->user_cpus_ptr, user_mask);
2622 void release_user_cpus_ptr(struct task_struct *p)
2624 kfree(clear_user_cpus_ptr(p));
2628 * This function is wildly self concurrent; here be dragons.
2631 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2632 * designated task is enqueued on an allowed CPU. If that task is currently
2633 * running, we have to kick it out using the CPU stopper.
2635 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2638 * Initial conditions: P0->cpus_mask = [0, 1]
2642 * migrate_disable();
2644 * set_cpus_allowed_ptr(P0, [1]);
2646 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2647 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2648 * This means we need the following scheme:
2652 * migrate_disable();
2654 * set_cpus_allowed_ptr(P0, [1]);
2658 * __set_cpus_allowed_ptr();
2659 * <wakes local stopper>
2660 * `--> <woken on migration completion>
2662 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2663 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2664 * task p are serialized by p->pi_lock, which we can leverage: the one that
2665 * should come into effect at the end of the Migrate-Disable region is the last
2666 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2667 * but we still need to properly signal those waiting tasks at the appropriate
2670 * This is implemented using struct set_affinity_pending. The first
2671 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2672 * setup an instance of that struct and install it on the targeted task_struct.
2673 * Any and all further callers will reuse that instance. Those then wait for
2674 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2675 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2678 * (1) In the cases covered above. There is one more where the completion is
2679 * signaled within affine_move_task() itself: when a subsequent affinity request
2680 * occurs after the stopper bailed out due to the targeted task still being
2681 * Migrate-Disable. Consider:
2683 * Initial conditions: P0->cpus_mask = [0, 1]
2687 * migrate_disable();
2689 * set_cpus_allowed_ptr(P0, [1]);
2692 * migration_cpu_stop()
2693 * is_migration_disabled()
2695 * set_cpus_allowed_ptr(P0, [0, 1]);
2696 * <signal completion>
2699 * Note that the above is safe vs a concurrent migrate_enable(), as any
2700 * pending affinity completion is preceded by an uninstallation of
2701 * p->migration_pending done with p->pi_lock held.
2703 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2704 int dest_cpu, unsigned int flags)
2706 struct set_affinity_pending my_pending = { }, *pending = NULL;
2707 bool stop_pending, complete = false;
2709 /* Can the task run on the task's current CPU? If so, we're done */
2710 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2711 struct task_struct *push_task = NULL;
2713 if ((flags & SCA_MIGRATE_ENABLE) &&
2714 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2715 rq->push_busy = true;
2716 push_task = get_task_struct(p);
2720 * If there are pending waiters, but no pending stop_work,
2721 * then complete now.
2723 pending = p->migration_pending;
2724 if (pending && !pending->stop_pending) {
2725 p->migration_pending = NULL;
2729 task_rq_unlock(rq, p, rf);
2732 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2737 complete_all(&pending->done);
2742 if (!(flags & SCA_MIGRATE_ENABLE)) {
2743 /* serialized by p->pi_lock */
2744 if (!p->migration_pending) {
2745 /* Install the request */
2746 refcount_set(&my_pending.refs, 1);
2747 init_completion(&my_pending.done);
2748 my_pending.arg = (struct migration_arg) {
2750 .dest_cpu = dest_cpu,
2751 .pending = &my_pending,
2754 p->migration_pending = &my_pending;
2756 pending = p->migration_pending;
2757 refcount_inc(&pending->refs);
2759 * Affinity has changed, but we've already installed a
2760 * pending. migration_cpu_stop() *must* see this, else
2761 * we risk a completion of the pending despite having a
2762 * task on a disallowed CPU.
2764 * Serialized by p->pi_lock, so this is safe.
2766 pending->arg.dest_cpu = dest_cpu;
2769 pending = p->migration_pending;
2771 * - !MIGRATE_ENABLE:
2772 * we'll have installed a pending if there wasn't one already.
2775 * we're here because the current CPU isn't matching anymore,
2776 * the only way that can happen is because of a concurrent
2777 * set_cpus_allowed_ptr() call, which should then still be
2778 * pending completion.
2780 * Either way, we really should have a @pending here.
2782 if (WARN_ON_ONCE(!pending)) {
2783 task_rq_unlock(rq, p, rf);
2787 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2789 * MIGRATE_ENABLE gets here because 'p == current', but for
2790 * anything else we cannot do is_migration_disabled(), punt
2791 * and have the stopper function handle it all race-free.
2793 stop_pending = pending->stop_pending;
2795 pending->stop_pending = true;
2797 if (flags & SCA_MIGRATE_ENABLE)
2798 p->migration_flags &= ~MDF_PUSH;
2800 task_rq_unlock(rq, p, rf);
2802 if (!stop_pending) {
2803 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2804 &pending->arg, &pending->stop_work);
2807 if (flags & SCA_MIGRATE_ENABLE)
2811 if (!is_migration_disabled(p)) {
2812 if (task_on_rq_queued(p))
2813 rq = move_queued_task(rq, rf, p, dest_cpu);
2815 if (!pending->stop_pending) {
2816 p->migration_pending = NULL;
2820 task_rq_unlock(rq, p, rf);
2823 complete_all(&pending->done);
2826 wait_for_completion(&pending->done);
2828 if (refcount_dec_and_test(&pending->refs))
2829 wake_up_var(&pending->refs); /* No UaF, just an address */
2832 * Block the original owner of &pending until all subsequent callers
2833 * have seen the completion and decremented the refcount
2835 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2838 WARN_ON_ONCE(my_pending.stop_pending);
2844 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2846 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2847 const struct cpumask *new_mask,
2850 struct rq_flags *rf)
2851 __releases(rq->lock)
2852 __releases(p->pi_lock)
2854 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2855 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2856 bool kthread = p->flags & PF_KTHREAD;
2857 struct cpumask *user_mask = NULL;
2858 unsigned int dest_cpu;
2861 update_rq_clock(rq);
2863 if (kthread || is_migration_disabled(p)) {
2865 * Kernel threads are allowed on online && !active CPUs,
2866 * however, during cpu-hot-unplug, even these might get pushed
2867 * away if not KTHREAD_IS_PER_CPU.
2869 * Specifically, migration_disabled() tasks must not fail the
2870 * cpumask_any_and_distribute() pick below, esp. so on
2871 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2872 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2874 cpu_valid_mask = cpu_online_mask;
2877 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2883 * Must re-check here, to close a race against __kthread_bind(),
2884 * sched_setaffinity() is not guaranteed to observe the flag.
2886 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2891 if (!(flags & SCA_MIGRATE_ENABLE)) {
2892 if (cpumask_equal(&p->cpus_mask, new_mask))
2895 if (WARN_ON_ONCE(p == current &&
2896 is_migration_disabled(p) &&
2897 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2904 * Picking a ~random cpu helps in cases where we are changing affinity
2905 * for groups of tasks (ie. cpuset), so that load balancing is not
2906 * immediately required to distribute the tasks within their new mask.
2908 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2909 if (dest_cpu >= nr_cpu_ids) {
2914 __do_set_cpus_allowed(p, new_mask, flags);
2916 if (flags & SCA_USER)
2917 user_mask = clear_user_cpus_ptr(p);
2919 ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2926 task_rq_unlock(rq, p, rf);
2932 * Change a given task's CPU affinity. Migrate the thread to a
2933 * proper CPU and schedule it away if the CPU it's executing on
2934 * is removed from the allowed bitmask.
2936 * NOTE: the caller must have a valid reference to the task, the
2937 * task must not exit() & deallocate itself prematurely. The
2938 * call is not atomic; no spinlocks may be held.
2940 static int __set_cpus_allowed_ptr(struct task_struct *p,
2941 const struct cpumask *new_mask, u32 flags)
2946 rq = task_rq_lock(p, &rf);
2947 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2950 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2952 return __set_cpus_allowed_ptr(p, new_mask, 0);
2954 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2957 * Change a given task's CPU affinity to the intersection of its current
2958 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2959 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2960 * If the resulting mask is empty, leave the affinity unchanged and return
2963 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2964 struct cpumask *new_mask,
2965 const struct cpumask *subset_mask)
2967 struct cpumask *user_mask = NULL;
2972 if (!p->user_cpus_ptr) {
2973 user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2978 rq = task_rq_lock(p, &rf);
2981 * Forcefully restricting the affinity of a deadline task is
2982 * likely to cause problems, so fail and noisily override the
2985 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2990 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2996 * We're about to butcher the task affinity, so keep track of what
2997 * the user asked for in case we're able to restore it later on.
3000 cpumask_copy(user_mask, p->cpus_ptr);
3001 p->user_cpus_ptr = user_mask;
3004 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
3007 task_rq_unlock(rq, p, &rf);
3013 * Restrict the CPU affinity of task @p so that it is a subset of
3014 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
3015 * old affinity mask. If the resulting mask is empty, we warn and walk
3016 * up the cpuset hierarchy until we find a suitable mask.
3018 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3020 cpumask_var_t new_mask;
3021 const struct cpumask *override_mask = task_cpu_possible_mask(p);
3023 alloc_cpumask_var(&new_mask, GFP_KERNEL);
3026 * __migrate_task() can fail silently in the face of concurrent
3027 * offlining of the chosen destination CPU, so take the hotplug
3028 * lock to ensure that the migration succeeds.
3031 if (!cpumask_available(new_mask))
3034 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3038 * We failed to find a valid subset of the affinity mask for the
3039 * task, so override it based on its cpuset hierarchy.
3041 cpuset_cpus_allowed(p, new_mask);
3042 override_mask = new_mask;
3045 if (printk_ratelimit()) {
3046 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3047 task_pid_nr(p), p->comm,
3048 cpumask_pr_args(override_mask));
3051 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3054 free_cpumask_var(new_mask);
3058 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
3061 * Restore the affinity of a task @p which was previously restricted by a
3062 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
3063 * @p->user_cpus_ptr.
3065 * It is the caller's responsibility to serialise this with any calls to
3066 * force_compatible_cpus_allowed_ptr(@p).
3068 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3070 struct cpumask *user_mask = p->user_cpus_ptr;
3071 unsigned long flags;
3074 * Try to restore the old affinity mask. If this fails, then
3075 * we free the mask explicitly to avoid it being inherited across
3076 * a subsequent fork().
3078 if (!user_mask || !__sched_setaffinity(p, user_mask))
3081 raw_spin_lock_irqsave(&p->pi_lock, flags);
3082 user_mask = clear_user_cpus_ptr(p);
3083 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3088 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3090 #ifdef CONFIG_SCHED_DEBUG
3091 unsigned int state = READ_ONCE(p->__state);
3094 * We should never call set_task_cpu() on a blocked task,
3095 * ttwu() will sort out the placement.
3097 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3100 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3101 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3102 * time relying on p->on_rq.
3104 WARN_ON_ONCE(state == TASK_RUNNING &&
3105 p->sched_class == &fair_sched_class &&
3106 (p->on_rq && !task_on_rq_migrating(p)));
3108 #ifdef CONFIG_LOCKDEP
3110 * The caller should hold either p->pi_lock or rq->lock, when changing
3111 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3113 * sched_move_task() holds both and thus holding either pins the cgroup,
3116 * Furthermore, all task_rq users should acquire both locks, see
3119 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3120 lockdep_is_held(__rq_lockp(task_rq(p)))));
3123 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3125 WARN_ON_ONCE(!cpu_online(new_cpu));
3127 WARN_ON_ONCE(is_migration_disabled(p));
3130 trace_sched_migrate_task(p, new_cpu);
3132 if (task_cpu(p) != new_cpu) {
3133 if (p->sched_class->migrate_task_rq)
3134 p->sched_class->migrate_task_rq(p, new_cpu);
3135 p->se.nr_migrations++;
3137 perf_event_task_migrate(p);
3140 __set_task_cpu(p, new_cpu);
3143 #ifdef CONFIG_NUMA_BALANCING
3144 static void __migrate_swap_task(struct task_struct *p, int cpu)
3146 if (task_on_rq_queued(p)) {
3147 struct rq *src_rq, *dst_rq;
3148 struct rq_flags srf, drf;
3150 src_rq = task_rq(p);
3151 dst_rq = cpu_rq(cpu);
3153 rq_pin_lock(src_rq, &srf);
3154 rq_pin_lock(dst_rq, &drf);
3156 deactivate_task(src_rq, p, 0);
3157 set_task_cpu(p, cpu);
3158 activate_task(dst_rq, p, 0);
3159 check_preempt_curr(dst_rq, p, 0);
3161 rq_unpin_lock(dst_rq, &drf);
3162 rq_unpin_lock(src_rq, &srf);
3166 * Task isn't running anymore; make it appear like we migrated
3167 * it before it went to sleep. This means on wakeup we make the
3168 * previous CPU our target instead of where it really is.
3174 struct migration_swap_arg {
3175 struct task_struct *src_task, *dst_task;
3176 int src_cpu, dst_cpu;
3179 static int migrate_swap_stop(void *data)
3181 struct migration_swap_arg *arg = data;
3182 struct rq *src_rq, *dst_rq;
3185 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3188 src_rq = cpu_rq(arg->src_cpu);
3189 dst_rq = cpu_rq(arg->dst_cpu);
3191 double_raw_lock(&arg->src_task->pi_lock,
3192 &arg->dst_task->pi_lock);
3193 double_rq_lock(src_rq, dst_rq);
3195 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3198 if (task_cpu(arg->src_task) != arg->src_cpu)
3201 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3204 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3207 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3208 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3213 double_rq_unlock(src_rq, dst_rq);
3214 raw_spin_unlock(&arg->dst_task->pi_lock);
3215 raw_spin_unlock(&arg->src_task->pi_lock);
3221 * Cross migrate two tasks
3223 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3224 int target_cpu, int curr_cpu)
3226 struct migration_swap_arg arg;
3229 arg = (struct migration_swap_arg){
3231 .src_cpu = curr_cpu,
3233 .dst_cpu = target_cpu,
3236 if (arg.src_cpu == arg.dst_cpu)
3240 * These three tests are all lockless; this is OK since all of them
3241 * will be re-checked with proper locks held further down the line.
3243 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3246 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3249 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3252 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3253 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3258 #endif /* CONFIG_NUMA_BALANCING */
3261 * wait_task_inactive - wait for a thread to unschedule.
3263 * If @match_state is nonzero, it's the @p->state value just checked and
3264 * not expected to change. If it changes, i.e. @p might have woken up,
3265 * then return zero. When we succeed in waiting for @p to be off its CPU,
3266 * we return a positive number (its total switch count). If a second call
3267 * a short while later returns the same number, the caller can be sure that
3268 * @p has remained unscheduled the whole time.
3270 * The caller must ensure that the task *will* unschedule sometime soon,
3271 * else this function might spin for a *long* time. This function can't
3272 * be called with interrupts off, or it may introduce deadlock with
3273 * smp_call_function() if an IPI is sent by the same process we are
3274 * waiting to become inactive.
3276 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3278 int running, queued;
3285 * We do the initial early heuristics without holding
3286 * any task-queue locks at all. We'll only try to get
3287 * the runqueue lock when things look like they will
3293 * If the task is actively running on another CPU
3294 * still, just relax and busy-wait without holding
3297 * NOTE! Since we don't hold any locks, it's not
3298 * even sure that "rq" stays as the right runqueue!
3299 * But we don't care, since "task_running()" will
3300 * return false if the runqueue has changed and p
3301 * is actually now running somewhere else!
3303 while (task_running(rq, p)) {
3304 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3310 * Ok, time to look more closely! We need the rq
3311 * lock now, to be *sure*. If we're wrong, we'll
3312 * just go back and repeat.
3314 rq = task_rq_lock(p, &rf);
3315 trace_sched_wait_task(p);
3316 running = task_running(rq, p);
3317 queued = task_on_rq_queued(p);
3319 if (!match_state || READ_ONCE(p->__state) == match_state)
3320 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3321 task_rq_unlock(rq, p, &rf);
3324 * If it changed from the expected state, bail out now.
3326 if (unlikely(!ncsw))
3330 * Was it really running after all now that we
3331 * checked with the proper locks actually held?
3333 * Oops. Go back and try again..
3335 if (unlikely(running)) {
3341 * It's not enough that it's not actively running,
3342 * it must be off the runqueue _entirely_, and not
3345 * So if it was still runnable (but just not actively
3346 * running right now), it's preempted, and we should
3347 * yield - it could be a while.
3349 if (unlikely(queued)) {
3350 ktime_t to = NSEC_PER_SEC / HZ;
3352 set_current_state(TASK_UNINTERRUPTIBLE);
3353 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3358 * Ahh, all good. It wasn't running, and it wasn't
3359 * runnable, which means that it will never become
3360 * running in the future either. We're all done!
3369 * kick_process - kick a running thread to enter/exit the kernel
3370 * @p: the to-be-kicked thread
3372 * Cause a process which is running on another CPU to enter
3373 * kernel-mode, without any delay. (to get signals handled.)
3375 * NOTE: this function doesn't have to take the runqueue lock,
3376 * because all it wants to ensure is that the remote task enters
3377 * the kernel. If the IPI races and the task has been migrated
3378 * to another CPU then no harm is done and the purpose has been
3381 void kick_process(struct task_struct *p)
3387 if ((cpu != smp_processor_id()) && task_curr(p))
3388 smp_send_reschedule(cpu);
3391 EXPORT_SYMBOL_GPL(kick_process);
3394 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3396 * A few notes on cpu_active vs cpu_online:
3398 * - cpu_active must be a subset of cpu_online
3400 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3401 * see __set_cpus_allowed_ptr(). At this point the newly online
3402 * CPU isn't yet part of the sched domains, and balancing will not
3405 * - on CPU-down we clear cpu_active() to mask the sched domains and
3406 * avoid the load balancer to place new tasks on the to be removed
3407 * CPU. Existing tasks will remain running there and will be taken
3410 * This means that fallback selection must not select !active CPUs.
3411 * And can assume that any active CPU must be online. Conversely
3412 * select_task_rq() below may allow selection of !active CPUs in order
3413 * to satisfy the above rules.
3415 static int select_fallback_rq(int cpu, struct task_struct *p)
3417 int nid = cpu_to_node(cpu);
3418 const struct cpumask *nodemask = NULL;
3419 enum { cpuset, possible, fail } state = cpuset;
3423 * If the node that the CPU is on has been offlined, cpu_to_node()
3424 * will return -1. There is no CPU on the node, and we should
3425 * select the CPU on the other node.
3428 nodemask = cpumask_of_node(nid);
3430 /* Look for allowed, online CPU in same node. */
3431 for_each_cpu(dest_cpu, nodemask) {
3432 if (is_cpu_allowed(p, dest_cpu))
3438 /* Any allowed, online CPU? */
3439 for_each_cpu(dest_cpu, p->cpus_ptr) {
3440 if (!is_cpu_allowed(p, dest_cpu))
3446 /* No more Mr. Nice Guy. */
3449 if (cpuset_cpus_allowed_fallback(p)) {
3456 * XXX When called from select_task_rq() we only
3457 * hold p->pi_lock and again violate locking order.
3459 * More yuck to audit.
3461 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3471 if (state != cpuset) {
3473 * Don't tell them about moving exiting tasks or
3474 * kernel threads (both mm NULL), since they never
3477 if (p->mm && printk_ratelimit()) {
3478 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3479 task_pid_nr(p), p->comm, cpu);
3487 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3490 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3492 lockdep_assert_held(&p->pi_lock);
3494 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3495 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3497 cpu = cpumask_any(p->cpus_ptr);
3500 * In order not to call set_task_cpu() on a blocking task we need
3501 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3504 * Since this is common to all placement strategies, this lives here.
3506 * [ this allows ->select_task() to simply return task_cpu(p) and
3507 * not worry about this generic constraint ]
3509 if (unlikely(!is_cpu_allowed(p, cpu)))
3510 cpu = select_fallback_rq(task_cpu(p), p);
3515 void sched_set_stop_task(int cpu, struct task_struct *stop)
3517 static struct lock_class_key stop_pi_lock;
3518 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3519 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3523 * Make it appear like a SCHED_FIFO task, its something
3524 * userspace knows about and won't get confused about.
3526 * Also, it will make PI more or less work without too
3527 * much confusion -- but then, stop work should not
3528 * rely on PI working anyway.
3530 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3532 stop->sched_class = &stop_sched_class;
3535 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3536 * adjust the effective priority of a task. As a result,
3537 * rt_mutex_setprio() can trigger (RT) balancing operations,
3538 * which can then trigger wakeups of the stop thread to push
3539 * around the current task.
3541 * The stop task itself will never be part of the PI-chain, it
3542 * never blocks, therefore that ->pi_lock recursion is safe.
3543 * Tell lockdep about this by placing the stop->pi_lock in its
3546 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3549 cpu_rq(cpu)->stop = stop;
3553 * Reset it back to a normal scheduling class so that
3554 * it can die in pieces.
3556 old_stop->sched_class = &rt_sched_class;
3560 #else /* CONFIG_SMP */
3562 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3563 const struct cpumask *new_mask,
3566 return set_cpus_allowed_ptr(p, new_mask);
3569 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3571 static inline bool rq_has_pinned_tasks(struct rq *rq)
3576 #endif /* !CONFIG_SMP */
3579 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3583 if (!schedstat_enabled())
3589 if (cpu == rq->cpu) {
3590 __schedstat_inc(rq->ttwu_local);
3591 __schedstat_inc(p->stats.nr_wakeups_local);
3593 struct sched_domain *sd;
3595 __schedstat_inc(p->stats.nr_wakeups_remote);
3597 for_each_domain(rq->cpu, sd) {
3598 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3599 __schedstat_inc(sd->ttwu_wake_remote);
3606 if (wake_flags & WF_MIGRATED)
3607 __schedstat_inc(p->stats.nr_wakeups_migrate);
3608 #endif /* CONFIG_SMP */
3610 __schedstat_inc(rq->ttwu_count);
3611 __schedstat_inc(p->stats.nr_wakeups);
3613 if (wake_flags & WF_SYNC)
3614 __schedstat_inc(p->stats.nr_wakeups_sync);
3618 * Mark the task runnable and perform wakeup-preemption.
3620 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3621 struct rq_flags *rf)
3623 check_preempt_curr(rq, p, wake_flags);
3624 WRITE_ONCE(p->__state, TASK_RUNNING);
3625 trace_sched_wakeup(p);
3628 if (p->sched_class->task_woken) {
3630 * Our task @p is fully woken up and running; so it's safe to
3631 * drop the rq->lock, hereafter rq is only used for statistics.
3633 rq_unpin_lock(rq, rf);
3634 p->sched_class->task_woken(rq, p);
3635 rq_repin_lock(rq, rf);
3638 if (rq->idle_stamp) {
3639 u64 delta = rq_clock(rq) - rq->idle_stamp;
3640 u64 max = 2*rq->max_idle_balance_cost;
3642 update_avg(&rq->avg_idle, delta);
3644 if (rq->avg_idle > max)
3647 rq->wake_stamp = jiffies;
3648 rq->wake_avg_idle = rq->avg_idle / 2;
3656 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3657 struct rq_flags *rf)
3659 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3661 lockdep_assert_rq_held(rq);
3663 if (p->sched_contributes_to_load)
3664 rq->nr_uninterruptible--;
3667 if (wake_flags & WF_MIGRATED)
3668 en_flags |= ENQUEUE_MIGRATED;
3672 delayacct_blkio_end(p);
3673 atomic_dec(&task_rq(p)->nr_iowait);
3676 activate_task(rq, p, en_flags);
3677 ttwu_do_wakeup(rq, p, wake_flags, rf);
3681 * Consider @p being inside a wait loop:
3684 * set_current_state(TASK_UNINTERRUPTIBLE);
3691 * __set_current_state(TASK_RUNNING);
3693 * between set_current_state() and schedule(). In this case @p is still
3694 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3697 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3698 * then schedule() must still happen and p->state can be changed to
3699 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3700 * need to do a full wakeup with enqueue.
3702 * Returns: %true when the wakeup is done,
3705 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3711 rq = __task_rq_lock(p, &rf);
3712 if (task_on_rq_queued(p)) {
3713 /* check_preempt_curr() may use rq clock */
3714 update_rq_clock(rq);
3715 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3718 __task_rq_unlock(rq, &rf);
3724 void sched_ttwu_pending(void *arg)
3726 struct llist_node *llist = arg;
3727 struct rq *rq = this_rq();
3728 struct task_struct *p, *t;
3735 * rq::ttwu_pending racy indication of out-standing wakeups.
3736 * Races such that false-negatives are possible, since they
3737 * are shorter lived that false-positives would be.
3739 WRITE_ONCE(rq->ttwu_pending, 0);
3741 rq_lock_irqsave(rq, &rf);
3742 update_rq_clock(rq);
3744 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3745 if (WARN_ON_ONCE(p->on_cpu))
3746 smp_cond_load_acquire(&p->on_cpu, !VAL);
3748 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3749 set_task_cpu(p, cpu_of(rq));
3751 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3754 rq_unlock_irqrestore(rq, &rf);
3757 void send_call_function_single_ipi(int cpu)
3759 struct rq *rq = cpu_rq(cpu);
3761 if (!set_nr_if_polling(rq->idle))
3762 arch_send_call_function_single_ipi(cpu);
3764 trace_sched_wake_idle_without_ipi(cpu);
3768 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3769 * necessary. The wakee CPU on receipt of the IPI will queue the task
3770 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3771 * of the wakeup instead of the waker.
3773 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3775 struct rq *rq = cpu_rq(cpu);
3777 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3779 WRITE_ONCE(rq->ttwu_pending, 1);
3780 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3783 void wake_up_if_idle(int cpu)
3785 struct rq *rq = cpu_rq(cpu);
3790 if (!is_idle_task(rcu_dereference(rq->curr)))
3793 rq_lock_irqsave(rq, &rf);
3794 if (is_idle_task(rq->curr))
3796 /* Else CPU is not idle, do nothing here: */
3797 rq_unlock_irqrestore(rq, &rf);
3803 bool cpus_share_cache(int this_cpu, int that_cpu)
3805 if (this_cpu == that_cpu)
3808 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3811 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3814 * Do not complicate things with the async wake_list while the CPU is
3817 if (!cpu_active(cpu))
3821 * If the CPU does not share cache, then queue the task on the
3822 * remote rqs wakelist to avoid accessing remote data.
3824 if (!cpus_share_cache(smp_processor_id(), cpu))
3828 * If the task is descheduling and the only running task on the
3829 * CPU then use the wakelist to offload the task activation to
3830 * the soon-to-be-idle CPU as the current CPU is likely busy.
3831 * nr_running is checked to avoid unnecessary task stacking.
3833 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3839 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3841 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3842 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3845 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3846 __ttwu_queue_wakelist(p, cpu, wake_flags);
3853 #else /* !CONFIG_SMP */
3855 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3860 #endif /* CONFIG_SMP */
3862 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3864 struct rq *rq = cpu_rq(cpu);
3867 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3871 update_rq_clock(rq);
3872 ttwu_do_activate(rq, p, wake_flags, &rf);
3877 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3879 * The caller holds p::pi_lock if p != current or has preemption
3880 * disabled when p == current.
3882 * The rules of PREEMPT_RT saved_state:
3884 * The related locking code always holds p::pi_lock when updating
3885 * p::saved_state, which means the code is fully serialized in both cases.
3887 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3888 * bits set. This allows to distinguish all wakeup scenarios.
3890 static __always_inline
3891 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3893 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3894 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3895 state != TASK_RTLOCK_WAIT);
3898 if (READ_ONCE(p->__state) & state) {
3903 #ifdef CONFIG_PREEMPT_RT
3905 * Saved state preserves the task state across blocking on
3906 * an RT lock. If the state matches, set p::saved_state to
3907 * TASK_RUNNING, but do not wake the task because it waits
3908 * for a lock wakeup. Also indicate success because from
3909 * the regular waker's point of view this has succeeded.
3911 * After acquiring the lock the task will restore p::__state
3912 * from p::saved_state which ensures that the regular
3913 * wakeup is not lost. The restore will also set
3914 * p::saved_state to TASK_RUNNING so any further tests will
3915 * not result in false positives vs. @success
3917 if (p->saved_state & state) {
3918 p->saved_state = TASK_RUNNING;
3926 * Notes on Program-Order guarantees on SMP systems.
3930 * The basic program-order guarantee on SMP systems is that when a task [t]
3931 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3932 * execution on its new CPU [c1].
3934 * For migration (of runnable tasks) this is provided by the following means:
3936 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3937 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3938 * rq(c1)->lock (if not at the same time, then in that order).
3939 * C) LOCK of the rq(c1)->lock scheduling in task
3941 * Release/acquire chaining guarantees that B happens after A and C after B.
3942 * Note: the CPU doing B need not be c0 or c1
3951 * UNLOCK rq(0)->lock
3953 * LOCK rq(0)->lock // orders against CPU0
3955 * UNLOCK rq(0)->lock
3959 * UNLOCK rq(1)->lock
3961 * LOCK rq(1)->lock // orders against CPU2
3964 * UNLOCK rq(1)->lock
3967 * BLOCKING -- aka. SLEEP + WAKEUP
3969 * For blocking we (obviously) need to provide the same guarantee as for
3970 * migration. However the means are completely different as there is no lock
3971 * chain to provide order. Instead we do:
3973 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3974 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3978 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3980 * LOCK rq(0)->lock LOCK X->pi_lock
3983 * smp_store_release(X->on_cpu, 0);
3985 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3991 * X->state = RUNNING
3992 * UNLOCK rq(2)->lock
3994 * LOCK rq(2)->lock // orders against CPU1
3997 * UNLOCK rq(2)->lock
4000 * UNLOCK rq(0)->lock
4003 * However, for wakeups there is a second guarantee we must provide, namely we
4004 * must ensure that CONDITION=1 done by the caller can not be reordered with
4005 * accesses to the task state; see try_to_wake_up() and set_current_state().
4009 * try_to_wake_up - wake up a thread
4010 * @p: the thread to be awakened
4011 * @state: the mask of task states that can be woken
4012 * @wake_flags: wake modifier flags (WF_*)
4014 * Conceptually does:
4016 * If (@state & @p->state) @p->state = TASK_RUNNING.
4018 * If the task was not queued/runnable, also place it back on a runqueue.
4020 * This function is atomic against schedule() which would dequeue the task.
4022 * It issues a full memory barrier before accessing @p->state, see the comment
4023 * with set_current_state().
4025 * Uses p->pi_lock to serialize against concurrent wake-ups.
4027 * Relies on p->pi_lock stabilizing:
4030 * - p->sched_task_group
4031 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4033 * Tries really hard to only take one task_rq(p)->lock for performance.
4034 * Takes rq->lock in:
4035 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4036 * - ttwu_queue() -- new rq, for enqueue of the task;
4037 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4039 * As a consequence we race really badly with just about everything. See the
4040 * many memory barriers and their comments for details.
4042 * Return: %true if @p->state changes (an actual wakeup was done),
4046 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4048 unsigned long flags;
4049 int cpu, success = 0;
4054 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4055 * == smp_processor_id()'. Together this means we can special
4056 * case the whole 'p->on_rq && ttwu_runnable()' case below
4057 * without taking any locks.
4060 * - we rely on Program-Order guarantees for all the ordering,
4061 * - we're serialized against set_special_state() by virtue of
4062 * it disabling IRQs (this allows not taking ->pi_lock).
4064 if (!ttwu_state_match(p, state, &success))
4067 trace_sched_waking(p);
4068 WRITE_ONCE(p->__state, TASK_RUNNING);
4069 trace_sched_wakeup(p);
4074 * If we are going to wake up a thread waiting for CONDITION we
4075 * need to ensure that CONDITION=1 done by the caller can not be
4076 * reordered with p->state check below. This pairs with smp_store_mb()
4077 * in set_current_state() that the waiting thread does.
4079 raw_spin_lock_irqsave(&p->pi_lock, flags);
4080 smp_mb__after_spinlock();
4081 if (!ttwu_state_match(p, state, &success))
4084 trace_sched_waking(p);
4087 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4088 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4089 * in smp_cond_load_acquire() below.
4091 * sched_ttwu_pending() try_to_wake_up()
4092 * STORE p->on_rq = 1 LOAD p->state
4095 * __schedule() (switch to task 'p')
4096 * LOCK rq->lock smp_rmb();
4097 * smp_mb__after_spinlock();
4101 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4103 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4104 * __schedule(). See the comment for smp_mb__after_spinlock().
4106 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4109 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4114 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4115 * possible to, falsely, observe p->on_cpu == 0.
4117 * One must be running (->on_cpu == 1) in order to remove oneself
4118 * from the runqueue.
4120 * __schedule() (switch to task 'p') try_to_wake_up()
4121 * STORE p->on_cpu = 1 LOAD p->on_rq
4124 * __schedule() (put 'p' to sleep)
4125 * LOCK rq->lock smp_rmb();
4126 * smp_mb__after_spinlock();
4127 * STORE p->on_rq = 0 LOAD p->on_cpu
4129 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4130 * __schedule(). See the comment for smp_mb__after_spinlock().
4132 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4133 * schedule()'s deactivate_task() has 'happened' and p will no longer
4134 * care about it's own p->state. See the comment in __schedule().
4136 smp_acquire__after_ctrl_dep();
4139 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4140 * == 0), which means we need to do an enqueue, change p->state to
4141 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4142 * enqueue, such as ttwu_queue_wakelist().
4144 WRITE_ONCE(p->__state, TASK_WAKING);
4147 * If the owning (remote) CPU is still in the middle of schedule() with
4148 * this task as prev, considering queueing p on the remote CPUs wake_list
4149 * which potentially sends an IPI instead of spinning on p->on_cpu to
4150 * let the waker make forward progress. This is safe because IRQs are
4151 * disabled and the IPI will deliver after on_cpu is cleared.
4153 * Ensure we load task_cpu(p) after p->on_cpu:
4155 * set_task_cpu(p, cpu);
4156 * STORE p->cpu = @cpu
4157 * __schedule() (switch to task 'p')
4159 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4160 * STORE p->on_cpu = 1 LOAD p->cpu
4162 * to ensure we observe the correct CPU on which the task is currently
4165 if (smp_load_acquire(&p->on_cpu) &&
4166 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
4170 * If the owning (remote) CPU is still in the middle of schedule() with
4171 * this task as prev, wait until it's done referencing the task.
4173 * Pairs with the smp_store_release() in finish_task().
4175 * This ensures that tasks getting woken will be fully ordered against
4176 * their previous state and preserve Program Order.
4178 smp_cond_load_acquire(&p->on_cpu, !VAL);
4180 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4181 if (task_cpu(p) != cpu) {
4183 delayacct_blkio_end(p);
4184 atomic_dec(&task_rq(p)->nr_iowait);
4187 wake_flags |= WF_MIGRATED;
4188 psi_ttwu_dequeue(p);
4189 set_task_cpu(p, cpu);
4193 #endif /* CONFIG_SMP */
4195 ttwu_queue(p, cpu, wake_flags);
4197 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4200 ttwu_stat(p, task_cpu(p), wake_flags);
4207 * task_call_func - Invoke a function on task in fixed state
4208 * @p: Process for which the function is to be invoked, can be @current.
4209 * @func: Function to invoke.
4210 * @arg: Argument to function.
4212 * Fix the task in it's current state by avoiding wakeups and or rq operations
4213 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4214 * to work out what the state is, if required. Given that @func can be invoked
4215 * with a runqueue lock held, it had better be quite lightweight.
4218 * Whatever @func returns
4220 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4222 struct rq *rq = NULL;
4227 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4229 state = READ_ONCE(p->__state);
4232 * Ensure we load p->on_rq after p->__state, otherwise it would be
4233 * possible to, falsely, observe p->on_rq == 0.
4235 * See try_to_wake_up() for a longer comment.
4240 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4241 * the task is blocked. Make sure to check @state since ttwu() can drop
4242 * locks at the end, see ttwu_queue_wakelist().
4244 if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4245 rq = __task_rq_lock(p, &rf);
4248 * At this point the task is pinned; either:
4249 * - blocked and we're holding off wakeups (pi->lock)
4250 * - woken, and we're holding off enqueue (rq->lock)
4251 * - queued, and we're holding off schedule (rq->lock)
4252 * - running, and we're holding off de-schedule (rq->lock)
4254 * The called function (@func) can use: task_curr(), p->on_rq and
4255 * p->__state to differentiate between these states.
4262 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4267 * wake_up_process - Wake up a specific process
4268 * @p: The process to be woken up.
4270 * Attempt to wake up the nominated process and move it to the set of runnable
4273 * Return: 1 if the process was woken up, 0 if it was already running.
4275 * This function executes a full memory barrier before accessing the task state.
4277 int wake_up_process(struct task_struct *p)
4279 return try_to_wake_up(p, TASK_NORMAL, 0);
4281 EXPORT_SYMBOL(wake_up_process);
4283 int wake_up_state(struct task_struct *p, unsigned int state)
4285 return try_to_wake_up(p, state, 0);
4289 * Perform scheduler related setup for a newly forked process p.
4290 * p is forked by current.
4292 * __sched_fork() is basic setup used by init_idle() too:
4294 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4299 p->se.exec_start = 0;
4300 p->se.sum_exec_runtime = 0;
4301 p->se.prev_sum_exec_runtime = 0;
4302 p->se.nr_migrations = 0;
4304 INIT_LIST_HEAD(&p->se.group_node);
4306 #ifdef CONFIG_FAIR_GROUP_SCHED
4307 p->se.cfs_rq = NULL;
4310 #ifdef CONFIG_SCHEDSTATS
4311 /* Even if schedstat is disabled, there should not be garbage */
4312 memset(&p->stats, 0, sizeof(p->stats));
4315 RB_CLEAR_NODE(&p->dl.rb_node);
4316 init_dl_task_timer(&p->dl);
4317 init_dl_inactive_task_timer(&p->dl);
4318 __dl_clear_params(p);
4320 INIT_LIST_HEAD(&p->rt.run_list);
4322 p->rt.time_slice = sched_rr_timeslice;
4326 #ifdef CONFIG_PREEMPT_NOTIFIERS
4327 INIT_HLIST_HEAD(&p->preempt_notifiers);
4330 #ifdef CONFIG_COMPACTION
4331 p->capture_control = NULL;
4333 init_numa_balancing(clone_flags, p);
4335 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4336 p->migration_pending = NULL;
4340 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4342 #ifdef CONFIG_NUMA_BALANCING
4344 int sysctl_numa_balancing_mode;
4346 static void __set_numabalancing_state(bool enabled)
4349 static_branch_enable(&sched_numa_balancing);
4351 static_branch_disable(&sched_numa_balancing);
4354 void set_numabalancing_state(bool enabled)
4357 sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4359 sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4360 __set_numabalancing_state(enabled);
4363 #ifdef CONFIG_PROC_SYSCTL
4364 int sysctl_numa_balancing(struct ctl_table *table, int write,
4365 void *buffer, size_t *lenp, loff_t *ppos)
4369 int state = sysctl_numa_balancing_mode;
4371 if (write && !capable(CAP_SYS_ADMIN))
4376 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4380 sysctl_numa_balancing_mode = state;
4381 __set_numabalancing_state(state);
4388 #ifdef CONFIG_SCHEDSTATS
4390 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4392 static void set_schedstats(bool enabled)
4395 static_branch_enable(&sched_schedstats);
4397 static_branch_disable(&sched_schedstats);
4400 void force_schedstat_enabled(void)
4402 if (!schedstat_enabled()) {
4403 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4404 static_branch_enable(&sched_schedstats);
4408 static int __init setup_schedstats(char *str)
4414 if (!strcmp(str, "enable")) {
4415 set_schedstats(true);
4417 } else if (!strcmp(str, "disable")) {
4418 set_schedstats(false);
4423 pr_warn("Unable to parse schedstats=\n");
4427 __setup("schedstats=", setup_schedstats);
4429 #ifdef CONFIG_PROC_SYSCTL
4430 static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4431 size_t *lenp, loff_t *ppos)
4435 int state = static_branch_likely(&sched_schedstats);
4437 if (write && !capable(CAP_SYS_ADMIN))
4442 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4446 set_schedstats(state);
4449 #endif /* CONFIG_PROC_SYSCTL */
4450 #endif /* CONFIG_SCHEDSTATS */
4452 #ifdef CONFIG_SYSCTL
4453 static struct ctl_table sched_core_sysctls[] = {
4454 #ifdef CONFIG_SCHEDSTATS
4456 .procname = "sched_schedstats",
4458 .maxlen = sizeof(unsigned int),
4460 .proc_handler = sysctl_schedstats,
4461 .extra1 = SYSCTL_ZERO,
4462 .extra2 = SYSCTL_ONE,
4464 #endif /* CONFIG_SCHEDSTATS */
4465 #ifdef CONFIG_UCLAMP_TASK
4467 .procname = "sched_util_clamp_min",
4468 .data = &sysctl_sched_uclamp_util_min,
4469 .maxlen = sizeof(unsigned int),
4471 .proc_handler = sysctl_sched_uclamp_handler,
4474 .procname = "sched_util_clamp_max",
4475 .data = &sysctl_sched_uclamp_util_max,
4476 .maxlen = sizeof(unsigned int),
4478 .proc_handler = sysctl_sched_uclamp_handler,
4481 .procname = "sched_util_clamp_min_rt_default",
4482 .data = &sysctl_sched_uclamp_util_min_rt_default,
4483 .maxlen = sizeof(unsigned int),
4485 .proc_handler = sysctl_sched_uclamp_handler,
4487 #endif /* CONFIG_UCLAMP_TASK */
4490 static int __init sched_core_sysctl_init(void)
4492 register_sysctl_init("kernel", sched_core_sysctls);
4495 late_initcall(sched_core_sysctl_init);
4496 #endif /* CONFIG_SYSCTL */
4499 * fork()/clone()-time setup:
4501 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4503 __sched_fork(clone_flags, p);
4505 * We mark the process as NEW here. This guarantees that
4506 * nobody will actually run it, and a signal or other external
4507 * event cannot wake it up and insert it on the runqueue either.
4509 p->__state = TASK_NEW;
4512 * Make sure we do not leak PI boosting priority to the child.
4514 p->prio = current->normal_prio;
4519 * Revert to default priority/policy on fork if requested.
4521 if (unlikely(p->sched_reset_on_fork)) {
4522 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4523 p->policy = SCHED_NORMAL;
4524 p->static_prio = NICE_TO_PRIO(0);
4526 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4527 p->static_prio = NICE_TO_PRIO(0);
4529 p->prio = p->normal_prio = p->static_prio;
4530 set_load_weight(p, false);
4533 * We don't need the reset flag anymore after the fork. It has
4534 * fulfilled its duty:
4536 p->sched_reset_on_fork = 0;
4539 if (dl_prio(p->prio))
4541 else if (rt_prio(p->prio))
4542 p->sched_class = &rt_sched_class;
4544 p->sched_class = &fair_sched_class;
4546 init_entity_runnable_average(&p->se);
4549 #ifdef CONFIG_SCHED_INFO
4550 if (likely(sched_info_on()))
4551 memset(&p->sched_info, 0, sizeof(p->sched_info));
4553 #if defined(CONFIG_SMP)
4556 init_task_preempt_count(p);
4558 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4559 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4564 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4566 unsigned long flags;
4569 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4570 * required yet, but lockdep gets upset if rules are violated.
4572 raw_spin_lock_irqsave(&p->pi_lock, flags);
4573 #ifdef CONFIG_CGROUP_SCHED
4575 struct task_group *tg;
4576 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4577 struct task_group, css);
4578 tg = autogroup_task_group(p, tg);
4579 p->sched_task_group = tg;
4584 * We're setting the CPU for the first time, we don't migrate,
4585 * so use __set_task_cpu().
4587 __set_task_cpu(p, smp_processor_id());
4588 if (p->sched_class->task_fork)
4589 p->sched_class->task_fork(p);
4590 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4593 void sched_post_fork(struct task_struct *p)
4595 uclamp_post_fork(p);
4598 unsigned long to_ratio(u64 period, u64 runtime)
4600 if (runtime == RUNTIME_INF)
4604 * Doing this here saves a lot of checks in all
4605 * the calling paths, and returning zero seems
4606 * safe for them anyway.
4611 return div64_u64(runtime << BW_SHIFT, period);
4615 * wake_up_new_task - wake up a newly created task for the first time.
4617 * This function will do some initial scheduler statistics housekeeping
4618 * that must be done for every newly created context, then puts the task
4619 * on the runqueue and wakes it.
4621 void wake_up_new_task(struct task_struct *p)
4626 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4627 WRITE_ONCE(p->__state, TASK_RUNNING);
4630 * Fork balancing, do it here and not earlier because:
4631 * - cpus_ptr can change in the fork path
4632 * - any previously selected CPU might disappear through hotplug
4634 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4635 * as we're not fully set-up yet.
4637 p->recent_used_cpu = task_cpu(p);
4639 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4641 rq = __task_rq_lock(p, &rf);
4642 update_rq_clock(rq);
4643 post_init_entity_util_avg(p);
4645 activate_task(rq, p, ENQUEUE_NOCLOCK);
4646 trace_sched_wakeup_new(p);
4647 check_preempt_curr(rq, p, WF_FORK);
4649 if (p->sched_class->task_woken) {
4651 * Nothing relies on rq->lock after this, so it's fine to
4654 rq_unpin_lock(rq, &rf);
4655 p->sched_class->task_woken(rq, p);
4656 rq_repin_lock(rq, &rf);
4659 task_rq_unlock(rq, p, &rf);
4662 #ifdef CONFIG_PREEMPT_NOTIFIERS
4664 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4666 void preempt_notifier_inc(void)
4668 static_branch_inc(&preempt_notifier_key);
4670 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4672 void preempt_notifier_dec(void)
4674 static_branch_dec(&preempt_notifier_key);
4676 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4679 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4680 * @notifier: notifier struct to register
4682 void preempt_notifier_register(struct preempt_notifier *notifier)
4684 if (!static_branch_unlikely(&preempt_notifier_key))
4685 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4687 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4689 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4692 * preempt_notifier_unregister - no longer interested in preemption notifications
4693 * @notifier: notifier struct to unregister
4695 * This is *not* safe to call from within a preemption notifier.
4697 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4699 hlist_del(¬ifier->link);
4701 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4703 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4705 struct preempt_notifier *notifier;
4707 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4708 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4711 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4713 if (static_branch_unlikely(&preempt_notifier_key))
4714 __fire_sched_in_preempt_notifiers(curr);
4718 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4719 struct task_struct *next)
4721 struct preempt_notifier *notifier;
4723 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4724 notifier->ops->sched_out(notifier, next);
4727 static __always_inline void
4728 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4729 struct task_struct *next)
4731 if (static_branch_unlikely(&preempt_notifier_key))
4732 __fire_sched_out_preempt_notifiers(curr, next);
4735 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4737 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4742 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4743 struct task_struct *next)
4747 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4749 static inline void prepare_task(struct task_struct *next)
4753 * Claim the task as running, we do this before switching to it
4754 * such that any running task will have this set.
4756 * See the ttwu() WF_ON_CPU case and its ordering comment.
4758 WRITE_ONCE(next->on_cpu, 1);
4762 static inline void finish_task(struct task_struct *prev)
4766 * This must be the very last reference to @prev from this CPU. After
4767 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4768 * must ensure this doesn't happen until the switch is completely
4771 * In particular, the load of prev->state in finish_task_switch() must
4772 * happen before this.
4774 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4776 smp_store_release(&prev->on_cpu, 0);
4782 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4784 void (*func)(struct rq *rq);
4785 struct callback_head *next;
4787 lockdep_assert_rq_held(rq);
4790 func = (void (*)(struct rq *))head->func;
4799 static void balance_push(struct rq *rq);
4802 * balance_push_callback is a right abuse of the callback interface and plays
4803 * by significantly different rules.
4805 * Where the normal balance_callback's purpose is to be ran in the same context
4806 * that queued it (only later, when it's safe to drop rq->lock again),
4807 * balance_push_callback is specifically targeted at __schedule().
4809 * This abuse is tolerated because it places all the unlikely/odd cases behind
4810 * a single test, namely: rq->balance_callback == NULL.
4812 struct callback_head balance_push_callback = {
4814 .func = (void (*)(struct callback_head *))balance_push,
4817 static inline struct callback_head *
4818 __splice_balance_callbacks(struct rq *rq, bool split)
4820 struct callback_head *head = rq->balance_callback;
4825 lockdep_assert_rq_held(rq);
4827 * Must not take balance_push_callback off the list when
4828 * splice_balance_callbacks() and balance_callbacks() are not
4829 * in the same rq->lock section.
4831 * In that case it would be possible for __schedule() to interleave
4832 * and observe the list empty.
4834 if (split && head == &balance_push_callback)
4837 rq->balance_callback = NULL;
4842 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4844 return __splice_balance_callbacks(rq, true);
4847 static void __balance_callbacks(struct rq *rq)
4849 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
4852 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4854 unsigned long flags;
4856 if (unlikely(head)) {
4857 raw_spin_rq_lock_irqsave(rq, flags);
4858 do_balance_callbacks(rq, head);
4859 raw_spin_rq_unlock_irqrestore(rq, flags);
4865 static inline void __balance_callbacks(struct rq *rq)
4869 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4874 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4881 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4884 * Since the runqueue lock will be released by the next
4885 * task (which is an invalid locking op but in the case
4886 * of the scheduler it's an obvious special-case), so we
4887 * do an early lockdep release here:
4889 rq_unpin_lock(rq, rf);
4890 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4891 #ifdef CONFIG_DEBUG_SPINLOCK
4892 /* this is a valid case when another task releases the spinlock */
4893 rq_lockp(rq)->owner = next;
4897 static inline void finish_lock_switch(struct rq *rq)
4900 * If we are tracking spinlock dependencies then we have to
4901 * fix up the runqueue lock - which gets 'carried over' from
4902 * prev into current:
4904 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4905 __balance_callbacks(rq);
4906 raw_spin_rq_unlock_irq(rq);
4910 * NOP if the arch has not defined these:
4913 #ifndef prepare_arch_switch
4914 # define prepare_arch_switch(next) do { } while (0)
4917 #ifndef finish_arch_post_lock_switch
4918 # define finish_arch_post_lock_switch() do { } while (0)
4921 static inline void kmap_local_sched_out(void)
4923 #ifdef CONFIG_KMAP_LOCAL
4924 if (unlikely(current->kmap_ctrl.idx))
4925 __kmap_local_sched_out();
4929 static inline void kmap_local_sched_in(void)
4931 #ifdef CONFIG_KMAP_LOCAL
4932 if (unlikely(current->kmap_ctrl.idx))
4933 __kmap_local_sched_in();
4938 * prepare_task_switch - prepare to switch tasks
4939 * @rq: the runqueue preparing to switch
4940 * @prev: the current task that is being switched out
4941 * @next: the task we are going to switch to.
4943 * This is called with the rq lock held and interrupts off. It must
4944 * be paired with a subsequent finish_task_switch after the context
4947 * prepare_task_switch sets up locking and calls architecture specific
4951 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4952 struct task_struct *next)
4954 kcov_prepare_switch(prev);
4955 sched_info_switch(rq, prev, next);
4956 perf_event_task_sched_out(prev, next);
4958 fire_sched_out_preempt_notifiers(prev, next);
4959 kmap_local_sched_out();
4961 prepare_arch_switch(next);
4965 * finish_task_switch - clean up after a task-switch
4966 * @prev: the thread we just switched away from.
4968 * finish_task_switch must be called after the context switch, paired
4969 * with a prepare_task_switch call before the context switch.
4970 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4971 * and do any other architecture-specific cleanup actions.
4973 * Note that we may have delayed dropping an mm in context_switch(). If
4974 * so, we finish that here outside of the runqueue lock. (Doing it
4975 * with the lock held can cause deadlocks; see schedule() for
4978 * The context switch have flipped the stack from under us and restored the
4979 * local variables which were saved when this task called schedule() in the
4980 * past. prev == current is still correct but we need to recalculate this_rq
4981 * because prev may have moved to another CPU.
4983 static struct rq *finish_task_switch(struct task_struct *prev)
4984 __releases(rq->lock)
4986 struct rq *rq = this_rq();
4987 struct mm_struct *mm = rq->prev_mm;
4988 unsigned int prev_state;
4991 * The previous task will have left us with a preempt_count of 2
4992 * because it left us after:
4995 * preempt_disable(); // 1
4997 * raw_spin_lock_irq(&rq->lock) // 2
4999 * Also, see FORK_PREEMPT_COUNT.
5001 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5002 "corrupted preempt_count: %s/%d/0x%x\n",
5003 current->comm, current->pid, preempt_count()))
5004 preempt_count_set(FORK_PREEMPT_COUNT);
5009 * A task struct has one reference for the use as "current".
5010 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5011 * schedule one last time. The schedule call will never return, and
5012 * the scheduled task must drop that reference.
5014 * We must observe prev->state before clearing prev->on_cpu (in
5015 * finish_task), otherwise a concurrent wakeup can get prev
5016 * running on another CPU and we could rave with its RUNNING -> DEAD
5017 * transition, resulting in a double drop.
5019 prev_state = READ_ONCE(prev->__state);
5020 vtime_task_switch(prev);
5021 perf_event_task_sched_in(prev, current);
5023 tick_nohz_task_switch();
5024 finish_lock_switch(rq);
5025 finish_arch_post_lock_switch();
5026 kcov_finish_switch(current);
5028 * kmap_local_sched_out() is invoked with rq::lock held and
5029 * interrupts disabled. There is no requirement for that, but the
5030 * sched out code does not have an interrupt enabled section.
5031 * Restoring the maps on sched in does not require interrupts being
5034 kmap_local_sched_in();
5036 fire_sched_in_preempt_notifiers(current);
5038 * When switching through a kernel thread, the loop in
5039 * membarrier_{private,global}_expedited() may have observed that
5040 * kernel thread and not issued an IPI. It is therefore possible to
5041 * schedule between user->kernel->user threads without passing though
5042 * switch_mm(). Membarrier requires a barrier after storing to
5043 * rq->curr, before returning to userspace, so provide them here:
5045 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5046 * provided by mmdrop(),
5047 * - a sync_core for SYNC_CORE.
5050 membarrier_mm_sync_core_before_usermode(mm);
5053 if (unlikely(prev_state == TASK_DEAD)) {
5054 if (prev->sched_class->task_dead)
5055 prev->sched_class->task_dead(prev);
5057 /* Task is done with its stack. */
5058 put_task_stack(prev);
5060 put_task_struct_rcu_user(prev);
5067 * schedule_tail - first thing a freshly forked thread must call.
5068 * @prev: the thread we just switched away from.
5070 asmlinkage __visible void schedule_tail(struct task_struct *prev)
5071 __releases(rq->lock)
5074 * New tasks start with FORK_PREEMPT_COUNT, see there and
5075 * finish_task_switch() for details.
5077 * finish_task_switch() will drop rq->lock() and lower preempt_count
5078 * and the preempt_enable() will end up enabling preemption (on
5079 * PREEMPT_COUNT kernels).
5082 finish_task_switch(prev);
5085 if (current->set_child_tid)
5086 put_user(task_pid_vnr(current), current->set_child_tid);
5088 calculate_sigpending();
5092 * context_switch - switch to the new MM and the new thread's register state.
5094 static __always_inline struct rq *
5095 context_switch(struct rq *rq, struct task_struct *prev,
5096 struct task_struct *next, struct rq_flags *rf)
5098 prepare_task_switch(rq, prev, next);
5101 * For paravirt, this is coupled with an exit in switch_to to
5102 * combine the page table reload and the switch backend into
5105 arch_start_context_switch(prev);
5108 * kernel -> kernel lazy + transfer active
5109 * user -> kernel lazy + mmgrab() active
5111 * kernel -> user switch + mmdrop() active
5112 * user -> user switch
5114 if (!next->mm) { // to kernel
5115 enter_lazy_tlb(prev->active_mm, next);
5117 next->active_mm = prev->active_mm;
5118 if (prev->mm) // from user
5119 mmgrab(prev->active_mm);
5121 prev->active_mm = NULL;
5123 membarrier_switch_mm(rq, prev->active_mm, next->mm);
5125 * sys_membarrier() requires an smp_mb() between setting
5126 * rq->curr / membarrier_switch_mm() and returning to userspace.
5128 * The below provides this either through switch_mm(), or in
5129 * case 'prev->active_mm == next->mm' through
5130 * finish_task_switch()'s mmdrop().
5132 switch_mm_irqs_off(prev->active_mm, next->mm, next);
5134 if (!prev->mm) { // from kernel
5135 /* will mmdrop() in finish_task_switch(). */
5136 rq->prev_mm = prev->active_mm;
5137 prev->active_mm = NULL;
5141 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5143 prepare_lock_switch(rq, next, rf);
5145 /* Here we just switch the register state and the stack. */
5146 switch_to(prev, next, prev);
5149 return finish_task_switch(prev);
5153 * nr_running and nr_context_switches:
5155 * externally visible scheduler statistics: current number of runnable
5156 * threads, total number of context switches performed since bootup.
5158 unsigned int nr_running(void)
5160 unsigned int i, sum = 0;
5162 for_each_online_cpu(i)
5163 sum += cpu_rq(i)->nr_running;
5169 * Check if only the current task is running on the CPU.
5171 * Caution: this function does not check that the caller has disabled
5172 * preemption, thus the result might have a time-of-check-to-time-of-use
5173 * race. The caller is responsible to use it correctly, for example:
5175 * - from a non-preemptible section (of course)
5177 * - from a thread that is bound to a single CPU
5179 * - in a loop with very short iterations (e.g. a polling loop)
5181 bool single_task_running(void)
5183 return raw_rq()->nr_running == 1;
5185 EXPORT_SYMBOL(single_task_running);
5187 unsigned long long nr_context_switches(void)
5190 unsigned long long sum = 0;
5192 for_each_possible_cpu(i)
5193 sum += cpu_rq(i)->nr_switches;
5199 * Consumers of these two interfaces, like for example the cpuidle menu
5200 * governor, are using nonsensical data. Preferring shallow idle state selection
5201 * for a CPU that has IO-wait which might not even end up running the task when
5202 * it does become runnable.
5205 unsigned int nr_iowait_cpu(int cpu)
5207 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5211 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5213 * The idea behind IO-wait account is to account the idle time that we could
5214 * have spend running if it were not for IO. That is, if we were to improve the
5215 * storage performance, we'd have a proportional reduction in IO-wait time.
5217 * This all works nicely on UP, where, when a task blocks on IO, we account
5218 * idle time as IO-wait, because if the storage were faster, it could've been
5219 * running and we'd not be idle.
5221 * This has been extended to SMP, by doing the same for each CPU. This however
5224 * Imagine for instance the case where two tasks block on one CPU, only the one
5225 * CPU will have IO-wait accounted, while the other has regular idle. Even
5226 * though, if the storage were faster, both could've ran at the same time,
5227 * utilising both CPUs.
5229 * This means, that when looking globally, the current IO-wait accounting on
5230 * SMP is a lower bound, by reason of under accounting.
5232 * Worse, since the numbers are provided per CPU, they are sometimes
5233 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5234 * associated with any one particular CPU, it can wake to another CPU than it
5235 * blocked on. This means the per CPU IO-wait number is meaningless.
5237 * Task CPU affinities can make all that even more 'interesting'.
5240 unsigned int nr_iowait(void)
5242 unsigned int i, sum = 0;
5244 for_each_possible_cpu(i)
5245 sum += nr_iowait_cpu(i);
5253 * sched_exec - execve() is a valuable balancing opportunity, because at
5254 * this point the task has the smallest effective memory and cache footprint.
5256 void sched_exec(void)
5258 struct task_struct *p = current;
5259 unsigned long flags;
5262 raw_spin_lock_irqsave(&p->pi_lock, flags);
5263 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5264 if (dest_cpu == smp_processor_id())
5267 if (likely(cpu_active(dest_cpu))) {
5268 struct migration_arg arg = { p, dest_cpu };
5270 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5271 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5275 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5280 DEFINE_PER_CPU(struct kernel_stat, kstat);
5281 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5283 EXPORT_PER_CPU_SYMBOL(kstat);
5284 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5287 * The function fair_sched_class.update_curr accesses the struct curr
5288 * and its field curr->exec_start; when called from task_sched_runtime(),
5289 * we observe a high rate of cache misses in practice.
5290 * Prefetching this data results in improved performance.
5292 static inline void prefetch_curr_exec_start(struct task_struct *p)
5294 #ifdef CONFIG_FAIR_GROUP_SCHED
5295 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5297 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5300 prefetch(&curr->exec_start);
5304 * Return accounted runtime for the task.
5305 * In case the task is currently running, return the runtime plus current's
5306 * pending runtime that have not been accounted yet.
5308 unsigned long long task_sched_runtime(struct task_struct *p)
5314 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5316 * 64-bit doesn't need locks to atomically read a 64-bit value.
5317 * So we have a optimization chance when the task's delta_exec is 0.
5318 * Reading ->on_cpu is racy, but this is ok.
5320 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5321 * If we race with it entering CPU, unaccounted time is 0. This is
5322 * indistinguishable from the read occurring a few cycles earlier.
5323 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5324 * been accounted, so we're correct here as well.
5326 if (!p->on_cpu || !task_on_rq_queued(p))
5327 return p->se.sum_exec_runtime;
5330 rq = task_rq_lock(p, &rf);
5332 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5333 * project cycles that may never be accounted to this
5334 * thread, breaking clock_gettime().
5336 if (task_current(rq, p) && task_on_rq_queued(p)) {
5337 prefetch_curr_exec_start(p);
5338 update_rq_clock(rq);
5339 p->sched_class->update_curr(rq);
5341 ns = p->se.sum_exec_runtime;
5342 task_rq_unlock(rq, p, &rf);
5347 #ifdef CONFIG_SCHED_DEBUG
5348 static u64 cpu_resched_latency(struct rq *rq)
5350 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5351 u64 resched_latency, now = rq_clock(rq);
5352 static bool warned_once;
5354 if (sysctl_resched_latency_warn_once && warned_once)
5357 if (!need_resched() || !latency_warn_ms)
5360 if (system_state == SYSTEM_BOOTING)
5363 if (!rq->last_seen_need_resched_ns) {
5364 rq->last_seen_need_resched_ns = now;
5365 rq->ticks_without_resched = 0;
5369 rq->ticks_without_resched++;
5370 resched_latency = now - rq->last_seen_need_resched_ns;
5371 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5376 return resched_latency;
5379 static int __init setup_resched_latency_warn_ms(char *str)
5383 if ((kstrtol(str, 0, &val))) {
5384 pr_warn("Unable to set resched_latency_warn_ms\n");
5388 sysctl_resched_latency_warn_ms = val;
5391 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5393 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5394 #endif /* CONFIG_SCHED_DEBUG */
5397 * This function gets called by the timer code, with HZ frequency.
5398 * We call it with interrupts disabled.
5400 void scheduler_tick(void)
5402 int cpu = smp_processor_id();
5403 struct rq *rq = cpu_rq(cpu);
5404 struct task_struct *curr = rq->curr;
5406 unsigned long thermal_pressure;
5407 u64 resched_latency;
5409 arch_scale_freq_tick();
5414 update_rq_clock(rq);
5415 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5416 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5417 curr->sched_class->task_tick(rq, curr, 0);
5418 if (sched_feat(LATENCY_WARN))
5419 resched_latency = cpu_resched_latency(rq);
5420 calc_global_load_tick(rq);
5421 sched_core_tick(rq);
5425 if (sched_feat(LATENCY_WARN) && resched_latency)
5426 resched_latency_warn(cpu, resched_latency);
5428 perf_event_task_tick();
5431 rq->idle_balance = idle_cpu(cpu);
5432 trigger_load_balance(rq);
5436 #ifdef CONFIG_NO_HZ_FULL
5441 struct delayed_work work;
5443 /* Values for ->state, see diagram below. */
5444 #define TICK_SCHED_REMOTE_OFFLINE 0
5445 #define TICK_SCHED_REMOTE_OFFLINING 1
5446 #define TICK_SCHED_REMOTE_RUNNING 2
5449 * State diagram for ->state:
5452 * TICK_SCHED_REMOTE_OFFLINE
5455 * | | sched_tick_remote()
5458 * +--TICK_SCHED_REMOTE_OFFLINING
5461 * sched_tick_start() | | sched_tick_stop()
5464 * TICK_SCHED_REMOTE_RUNNING
5467 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5468 * and sched_tick_start() are happy to leave the state in RUNNING.
5471 static struct tick_work __percpu *tick_work_cpu;
5473 static void sched_tick_remote(struct work_struct *work)
5475 struct delayed_work *dwork = to_delayed_work(work);
5476 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5477 int cpu = twork->cpu;
5478 struct rq *rq = cpu_rq(cpu);
5479 struct task_struct *curr;
5485 * Handle the tick only if it appears the remote CPU is running in full
5486 * dynticks mode. The check is racy by nature, but missing a tick or
5487 * having one too much is no big deal because the scheduler tick updates
5488 * statistics and checks timeslices in a time-independent way, regardless
5489 * of when exactly it is running.
5491 if (!tick_nohz_tick_stopped_cpu(cpu))
5494 rq_lock_irq(rq, &rf);
5496 if (cpu_is_offline(cpu))
5499 update_rq_clock(rq);
5501 if (!is_idle_task(curr)) {
5503 * Make sure the next tick runs within a reasonable
5506 delta = rq_clock_task(rq) - curr->se.exec_start;
5507 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5509 curr->sched_class->task_tick(rq, curr, 0);
5511 calc_load_nohz_remote(rq);
5513 rq_unlock_irq(rq, &rf);
5517 * Run the remote tick once per second (1Hz). This arbitrary
5518 * frequency is large enough to avoid overload but short enough
5519 * to keep scheduler internal stats reasonably up to date. But
5520 * first update state to reflect hotplug activity if required.
5522 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5523 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5524 if (os == TICK_SCHED_REMOTE_RUNNING)
5525 queue_delayed_work(system_unbound_wq, dwork, HZ);
5528 static void sched_tick_start(int cpu)
5531 struct tick_work *twork;
5533 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5536 WARN_ON_ONCE(!tick_work_cpu);
5538 twork = per_cpu_ptr(tick_work_cpu, cpu);
5539 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5540 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5541 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5543 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5544 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5548 #ifdef CONFIG_HOTPLUG_CPU
5549 static void sched_tick_stop(int cpu)
5551 struct tick_work *twork;
5554 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5557 WARN_ON_ONCE(!tick_work_cpu);
5559 twork = per_cpu_ptr(tick_work_cpu, cpu);
5560 /* There cannot be competing actions, but don't rely on stop-machine. */
5561 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5562 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5563 /* Don't cancel, as this would mess up the state machine. */
5565 #endif /* CONFIG_HOTPLUG_CPU */
5567 int __init sched_tick_offload_init(void)
5569 tick_work_cpu = alloc_percpu(struct tick_work);
5570 BUG_ON(!tick_work_cpu);
5574 #else /* !CONFIG_NO_HZ_FULL */
5575 static inline void sched_tick_start(int cpu) { }
5576 static inline void sched_tick_stop(int cpu) { }
5579 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5580 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5582 * If the value passed in is equal to the current preempt count
5583 * then we just disabled preemption. Start timing the latency.
5585 static inline void preempt_latency_start(int val)
5587 if (preempt_count() == val) {
5588 unsigned long ip = get_lock_parent_ip();
5589 #ifdef CONFIG_DEBUG_PREEMPT
5590 current->preempt_disable_ip = ip;
5592 trace_preempt_off(CALLER_ADDR0, ip);
5596 void preempt_count_add(int val)
5598 #ifdef CONFIG_DEBUG_PREEMPT
5602 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5605 __preempt_count_add(val);
5606 #ifdef CONFIG_DEBUG_PREEMPT
5608 * Spinlock count overflowing soon?
5610 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5613 preempt_latency_start(val);
5615 EXPORT_SYMBOL(preempt_count_add);
5616 NOKPROBE_SYMBOL(preempt_count_add);
5619 * If the value passed in equals to the current preempt count
5620 * then we just enabled preemption. Stop timing the latency.
5622 static inline void preempt_latency_stop(int val)
5624 if (preempt_count() == val)
5625 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5628 void preempt_count_sub(int val)
5630 #ifdef CONFIG_DEBUG_PREEMPT
5634 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5637 * Is the spinlock portion underflowing?
5639 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5640 !(preempt_count() & PREEMPT_MASK)))
5644 preempt_latency_stop(val);
5645 __preempt_count_sub(val);
5647 EXPORT_SYMBOL(preempt_count_sub);
5648 NOKPROBE_SYMBOL(preempt_count_sub);
5651 static inline void preempt_latency_start(int val) { }
5652 static inline void preempt_latency_stop(int val) { }
5655 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5657 #ifdef CONFIG_DEBUG_PREEMPT
5658 return p->preempt_disable_ip;
5665 * Print scheduling while atomic bug:
5667 static noinline void __schedule_bug(struct task_struct *prev)
5669 /* Save this before calling printk(), since that will clobber it */
5670 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5672 if (oops_in_progress)
5675 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5676 prev->comm, prev->pid, preempt_count());
5678 debug_show_held_locks(prev);
5680 if (irqs_disabled())
5681 print_irqtrace_events(prev);
5682 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5683 && in_atomic_preempt_off()) {
5684 pr_err("Preemption disabled at:");
5685 print_ip_sym(KERN_ERR, preempt_disable_ip);
5688 panic("scheduling while atomic\n");
5691 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5695 * Various schedule()-time debugging checks and statistics:
5697 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5699 #ifdef CONFIG_SCHED_STACK_END_CHECK
5700 if (task_stack_end_corrupted(prev))
5701 panic("corrupted stack end detected inside scheduler\n");
5703 if (task_scs_end_corrupted(prev))
5704 panic("corrupted shadow stack detected inside scheduler\n");
5707 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5708 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5709 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5710 prev->comm, prev->pid, prev->non_block_count);
5712 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5716 if (unlikely(in_atomic_preempt_off())) {
5717 __schedule_bug(prev);
5718 preempt_count_set(PREEMPT_DISABLED);
5721 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5723 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5725 schedstat_inc(this_rq()->sched_count);
5728 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5729 struct rq_flags *rf)
5732 const struct sched_class *class;
5734 * We must do the balancing pass before put_prev_task(), such
5735 * that when we release the rq->lock the task is in the same
5736 * state as before we took rq->lock.
5738 * We can terminate the balance pass as soon as we know there is
5739 * a runnable task of @class priority or higher.
5741 for_class_range(class, prev->sched_class, &idle_sched_class) {
5742 if (class->balance(rq, prev, rf))
5747 put_prev_task(rq, prev);
5751 * Pick up the highest-prio task:
5753 static inline struct task_struct *
5754 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5756 const struct sched_class *class;
5757 struct task_struct *p;
5760 * Optimization: we know that if all tasks are in the fair class we can
5761 * call that function directly, but only if the @prev task wasn't of a
5762 * higher scheduling class, because otherwise those lose the
5763 * opportunity to pull in more work from other CPUs.
5765 if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
5766 rq->nr_running == rq->cfs.h_nr_running)) {
5768 p = pick_next_task_fair(rq, prev, rf);
5769 if (unlikely(p == RETRY_TASK))
5772 /* Assume the next prioritized class is idle_sched_class */
5774 put_prev_task(rq, prev);
5775 p = pick_next_task_idle(rq);
5782 put_prev_task_balance(rq, prev, rf);
5784 for_each_class(class) {
5785 p = class->pick_next_task(rq);
5790 BUG(); /* The idle class should always have a runnable task. */
5793 #ifdef CONFIG_SCHED_CORE
5794 static inline bool is_task_rq_idle(struct task_struct *t)
5796 return (task_rq(t)->idle == t);
5799 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5801 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5804 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5806 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5809 return a->core_cookie == b->core_cookie;
5812 static inline struct task_struct *pick_task(struct rq *rq)
5814 const struct sched_class *class;
5815 struct task_struct *p;
5817 for_each_class(class) {
5818 p = class->pick_task(rq);
5823 BUG(); /* The idle class should always have a runnable task. */
5826 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5828 static void queue_core_balance(struct rq *rq);
5830 static struct task_struct *
5831 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5833 struct task_struct *next, *p, *max = NULL;
5834 const struct cpumask *smt_mask;
5835 bool fi_before = false;
5836 bool core_clock_updated = (rq == rq->core);
5837 unsigned long cookie;
5838 int i, cpu, occ = 0;
5842 if (!sched_core_enabled(rq))
5843 return __pick_next_task(rq, prev, rf);
5847 /* Stopper task is switching into idle, no need core-wide selection. */
5848 if (cpu_is_offline(cpu)) {
5850 * Reset core_pick so that we don't enter the fastpath when
5851 * coming online. core_pick would already be migrated to
5852 * another cpu during offline.
5854 rq->core_pick = NULL;
5855 return __pick_next_task(rq, prev, rf);
5859 * If there were no {en,de}queues since we picked (IOW, the task
5860 * pointers are all still valid), and we haven't scheduled the last
5861 * pick yet, do so now.
5863 * rq->core_pick can be NULL if no selection was made for a CPU because
5864 * it was either offline or went offline during a sibling's core-wide
5865 * selection. In this case, do a core-wide selection.
5867 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5868 rq->core->core_pick_seq != rq->core_sched_seq &&
5870 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5872 next = rq->core_pick;
5874 put_prev_task(rq, prev);
5875 set_next_task(rq, next);
5878 rq->core_pick = NULL;
5882 put_prev_task_balance(rq, prev, rf);
5884 smt_mask = cpu_smt_mask(cpu);
5885 need_sync = !!rq->core->core_cookie;
5888 rq->core->core_cookie = 0UL;
5889 if (rq->core->core_forceidle_count) {
5890 if (!core_clock_updated) {
5891 update_rq_clock(rq->core);
5892 core_clock_updated = true;
5894 sched_core_account_forceidle(rq);
5895 /* reset after accounting force idle */
5896 rq->core->core_forceidle_start = 0;
5897 rq->core->core_forceidle_count = 0;
5898 rq->core->core_forceidle_occupation = 0;
5904 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5906 * @task_seq guards the task state ({en,de}queues)
5907 * @pick_seq is the @task_seq we did a selection on
5908 * @sched_seq is the @pick_seq we scheduled
5910 * However, preemptions can cause multiple picks on the same task set.
5911 * 'Fix' this by also increasing @task_seq for every pick.
5913 rq->core->core_task_seq++;
5916 * Optimize for common case where this CPU has no cookies
5917 * and there are no cookied tasks running on siblings.
5920 next = pick_task(rq);
5921 if (!next->core_cookie) {
5922 rq->core_pick = NULL;
5924 * For robustness, update the min_vruntime_fi for
5925 * unconstrained picks as well.
5927 WARN_ON_ONCE(fi_before);
5928 task_vruntime_update(rq, next, false);
5934 * For each thread: do the regular task pick and find the max prio task
5937 * Tie-break prio towards the current CPU
5939 for_each_cpu_wrap(i, smt_mask, cpu) {
5943 * Current cpu always has its clock updated on entrance to
5944 * pick_next_task(). If the current cpu is not the core,
5945 * the core may also have been updated above.
5947 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
5948 update_rq_clock(rq_i);
5950 p = rq_i->core_pick = pick_task(rq_i);
5951 if (!max || prio_less(max, p, fi_before))
5955 cookie = rq->core->core_cookie = max->core_cookie;
5958 * For each thread: try and find a runnable task that matches @max or
5961 for_each_cpu(i, smt_mask) {
5963 p = rq_i->core_pick;
5965 if (!cookie_equals(p, cookie)) {
5968 p = sched_core_find(rq_i, cookie);
5970 p = idle_sched_class.pick_task(rq_i);
5973 rq_i->core_pick = p;
5975 if (p == rq_i->idle) {
5976 if (rq_i->nr_running) {
5977 rq->core->core_forceidle_count++;
5979 rq->core->core_forceidle_seq++;
5986 if (schedstat_enabled() && rq->core->core_forceidle_count) {
5987 rq->core->core_forceidle_start = rq_clock(rq->core);
5988 rq->core->core_forceidle_occupation = occ;
5991 rq->core->core_pick_seq = rq->core->core_task_seq;
5992 next = rq->core_pick;
5993 rq->core_sched_seq = rq->core->core_pick_seq;
5995 /* Something should have been selected for current CPU */
5996 WARN_ON_ONCE(!next);
5999 * Reschedule siblings
6001 * NOTE: L1TF -- at this point we're no longer running the old task and
6002 * sending an IPI (below) ensures the sibling will no longer be running
6003 * their task. This ensures there is no inter-sibling overlap between
6004 * non-matching user state.
6006 for_each_cpu(i, smt_mask) {
6010 * An online sibling might have gone offline before a task
6011 * could be picked for it, or it might be offline but later
6012 * happen to come online, but its too late and nothing was
6013 * picked for it. That's Ok - it will pick tasks for itself,
6016 if (!rq_i->core_pick)
6020 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6021 * fi_before fi update?
6027 if (!(fi_before && rq->core->core_forceidle_count))
6028 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6030 rq_i->core_pick->core_occupation = occ;
6033 rq_i->core_pick = NULL;
6037 /* Did we break L1TF mitigation requirements? */
6038 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6040 if (rq_i->curr == rq_i->core_pick) {
6041 rq_i->core_pick = NULL;
6049 set_next_task(rq, next);
6051 if (rq->core->core_forceidle_count && next == rq->idle)
6052 queue_core_balance(rq);
6057 static bool try_steal_cookie(int this, int that)
6059 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6060 struct task_struct *p;
6061 unsigned long cookie;
6062 bool success = false;
6064 local_irq_disable();
6065 double_rq_lock(dst, src);
6067 cookie = dst->core->core_cookie;
6071 if (dst->curr != dst->idle)
6074 p = sched_core_find(src, cookie);
6079 if (p == src->core_pick || p == src->curr)
6082 if (!is_cpu_allowed(p, this))
6085 if (p->core_occupation > dst->idle->core_occupation)
6088 deactivate_task(src, p, 0);
6089 set_task_cpu(p, this);
6090 activate_task(dst, p, 0);
6098 p = sched_core_next(p, cookie);
6102 double_rq_unlock(dst, src);
6108 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6112 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
6119 if (try_steal_cookie(cpu, i))
6126 static void sched_core_balance(struct rq *rq)
6128 struct sched_domain *sd;
6129 int cpu = cpu_of(rq);
6133 raw_spin_rq_unlock_irq(rq);
6134 for_each_domain(cpu, sd) {
6138 if (steal_cookie_task(cpu, sd))
6141 raw_spin_rq_lock_irq(rq);
6146 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
6148 static void queue_core_balance(struct rq *rq)
6150 if (!sched_core_enabled(rq))
6153 if (!rq->core->core_cookie)
6156 if (!rq->nr_running) /* not forced idle */
6159 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6162 static void sched_core_cpu_starting(unsigned int cpu)
6164 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6165 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6166 unsigned long flags;
6169 sched_core_lock(cpu, &flags);
6171 WARN_ON_ONCE(rq->core != rq);
6173 /* if we're the first, we'll be our own leader */
6174 if (cpumask_weight(smt_mask) == 1)
6177 /* find the leader */
6178 for_each_cpu(t, smt_mask) {
6182 if (rq->core == rq) {
6188 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6191 /* install and validate core_rq */
6192 for_each_cpu(t, smt_mask) {
6198 WARN_ON_ONCE(rq->core != core_rq);
6202 sched_core_unlock(cpu, &flags);
6205 static void sched_core_cpu_deactivate(unsigned int cpu)
6207 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6208 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6209 unsigned long flags;
6212 sched_core_lock(cpu, &flags);
6214 /* if we're the last man standing, nothing to do */
6215 if (cpumask_weight(smt_mask) == 1) {
6216 WARN_ON_ONCE(rq->core != rq);
6220 /* if we're not the leader, nothing to do */
6224 /* find a new leader */
6225 for_each_cpu(t, smt_mask) {
6228 core_rq = cpu_rq(t);
6232 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6235 /* copy the shared state to the new leader */
6236 core_rq->core_task_seq = rq->core_task_seq;
6237 core_rq->core_pick_seq = rq->core_pick_seq;
6238 core_rq->core_cookie = rq->core_cookie;
6239 core_rq->core_forceidle_count = rq->core_forceidle_count;
6240 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6241 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6244 * Accounting edge for forced idle is handled in pick_next_task().
6245 * Don't need another one here, since the hotplug thread shouldn't
6248 core_rq->core_forceidle_start = 0;
6250 /* install new leader */
6251 for_each_cpu(t, smt_mask) {
6257 sched_core_unlock(cpu, &flags);
6260 static inline void sched_core_cpu_dying(unsigned int cpu)
6262 struct rq *rq = cpu_rq(cpu);
6268 #else /* !CONFIG_SCHED_CORE */
6270 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6271 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6272 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6274 static struct task_struct *
6275 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6277 return __pick_next_task(rq, prev, rf);
6280 #endif /* CONFIG_SCHED_CORE */
6283 * Constants for the sched_mode argument of __schedule().
6285 * The mode argument allows RT enabled kernels to differentiate a
6286 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6287 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6288 * optimize the AND operation out and just check for zero.
6291 #define SM_PREEMPT 0x1
6292 #define SM_RTLOCK_WAIT 0x2
6294 #ifndef CONFIG_PREEMPT_RT
6295 # define SM_MASK_PREEMPT (~0U)
6297 # define SM_MASK_PREEMPT SM_PREEMPT
6301 * __schedule() is the main scheduler function.
6303 * The main means of driving the scheduler and thus entering this function are:
6305 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6307 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6308 * paths. For example, see arch/x86/entry_64.S.
6310 * To drive preemption between tasks, the scheduler sets the flag in timer
6311 * interrupt handler scheduler_tick().
6313 * 3. Wakeups don't really cause entry into schedule(). They add a
6314 * task to the run-queue and that's it.
6316 * Now, if the new task added to the run-queue preempts the current
6317 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6318 * called on the nearest possible occasion:
6320 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6322 * - in syscall or exception context, at the next outmost
6323 * preempt_enable(). (this might be as soon as the wake_up()'s
6326 * - in IRQ context, return from interrupt-handler to
6327 * preemptible context
6329 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6332 * - cond_resched() call
6333 * - explicit schedule() call
6334 * - return from syscall or exception to user-space
6335 * - return from interrupt-handler to user-space
6337 * WARNING: must be called with preemption disabled!
6339 static void __sched notrace __schedule(unsigned int sched_mode)
6341 struct task_struct *prev, *next;
6342 unsigned long *switch_count;
6343 unsigned long prev_state;
6348 cpu = smp_processor_id();
6352 schedule_debug(prev, !!sched_mode);
6354 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6357 local_irq_disable();
6358 rcu_note_context_switch(!!sched_mode);
6361 * Make sure that signal_pending_state()->signal_pending() below
6362 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6363 * done by the caller to avoid the race with signal_wake_up():
6365 * __set_current_state(@state) signal_wake_up()
6366 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6367 * wake_up_state(p, state)
6368 * LOCK rq->lock LOCK p->pi_state
6369 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6370 * if (signal_pending_state()) if (p->state & @state)
6372 * Also, the membarrier system call requires a full memory barrier
6373 * after coming from user-space, before storing to rq->curr.
6376 smp_mb__after_spinlock();
6378 /* Promote REQ to ACT */
6379 rq->clock_update_flags <<= 1;
6380 update_rq_clock(rq);
6382 switch_count = &prev->nivcsw;
6385 * We must load prev->state once (task_struct::state is volatile), such
6386 * that we form a control dependency vs deactivate_task() below.
6388 prev_state = READ_ONCE(prev->__state);
6389 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6390 if (signal_pending_state(prev_state, prev)) {
6391 WRITE_ONCE(prev->__state, TASK_RUNNING);
6393 prev->sched_contributes_to_load =
6394 (prev_state & TASK_UNINTERRUPTIBLE) &&
6395 !(prev_state & TASK_NOLOAD) &&
6396 !(prev->flags & PF_FROZEN);
6398 if (prev->sched_contributes_to_load)
6399 rq->nr_uninterruptible++;
6402 * __schedule() ttwu()
6403 * prev_state = prev->state; if (p->on_rq && ...)
6404 * if (prev_state) goto out;
6405 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6406 * p->state = TASK_WAKING
6408 * Where __schedule() and ttwu() have matching control dependencies.
6410 * After this, schedule() must not care about p->state any more.
6412 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6414 if (prev->in_iowait) {
6415 atomic_inc(&rq->nr_iowait);
6416 delayacct_blkio_start();
6419 switch_count = &prev->nvcsw;
6422 next = pick_next_task(rq, prev, &rf);
6423 clear_tsk_need_resched(prev);
6424 clear_preempt_need_resched();
6425 #ifdef CONFIG_SCHED_DEBUG
6426 rq->last_seen_need_resched_ns = 0;
6429 if (likely(prev != next)) {
6432 * RCU users of rcu_dereference(rq->curr) may not see
6433 * changes to task_struct made by pick_next_task().
6435 RCU_INIT_POINTER(rq->curr, next);
6437 * The membarrier system call requires each architecture
6438 * to have a full memory barrier after updating
6439 * rq->curr, before returning to user-space.
6441 * Here are the schemes providing that barrier on the
6442 * various architectures:
6443 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6444 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6445 * - finish_lock_switch() for weakly-ordered
6446 * architectures where spin_unlock is a full barrier,
6447 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6448 * is a RELEASE barrier),
6452 migrate_disable_switch(rq, prev);
6453 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6455 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6457 /* Also unlocks the rq: */
6458 rq = context_switch(rq, prev, next, &rf);
6460 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6462 rq_unpin_lock(rq, &rf);
6463 __balance_callbacks(rq);
6464 raw_spin_rq_unlock_irq(rq);
6468 void __noreturn do_task_dead(void)
6470 /* Causes final put_task_struct in finish_task_switch(): */
6471 set_special_state(TASK_DEAD);
6473 /* Tell freezer to ignore us: */
6474 current->flags |= PF_NOFREEZE;
6476 __schedule(SM_NONE);
6479 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6484 static inline void sched_submit_work(struct task_struct *tsk)
6486 unsigned int task_flags;
6488 if (task_is_running(tsk))
6491 task_flags = tsk->flags;
6493 * If a worker goes to sleep, notify and ask workqueue whether it
6494 * wants to wake up a task to maintain concurrency.
6496 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6497 if (task_flags & PF_WQ_WORKER)
6498 wq_worker_sleeping(tsk);
6500 io_wq_worker_sleeping(tsk);
6503 if (tsk_is_pi_blocked(tsk))
6507 * If we are going to sleep and we have plugged IO queued,
6508 * make sure to submit it to avoid deadlocks.
6510 blk_flush_plug(tsk->plug, true);
6513 static void sched_update_worker(struct task_struct *tsk)
6515 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6516 if (tsk->flags & PF_WQ_WORKER)
6517 wq_worker_running(tsk);
6519 io_wq_worker_running(tsk);
6523 asmlinkage __visible void __sched schedule(void)
6525 struct task_struct *tsk = current;
6527 sched_submit_work(tsk);
6530 __schedule(SM_NONE);
6531 sched_preempt_enable_no_resched();
6532 } while (need_resched());
6533 sched_update_worker(tsk);
6535 EXPORT_SYMBOL(schedule);
6538 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6539 * state (have scheduled out non-voluntarily) by making sure that all
6540 * tasks have either left the run queue or have gone into user space.
6541 * As idle tasks do not do either, they must not ever be preempted
6542 * (schedule out non-voluntarily).
6544 * schedule_idle() is similar to schedule_preempt_disable() except that it
6545 * never enables preemption because it does not call sched_submit_work().
6547 void __sched schedule_idle(void)
6550 * As this skips calling sched_submit_work(), which the idle task does
6551 * regardless because that function is a nop when the task is in a
6552 * TASK_RUNNING state, make sure this isn't used someplace that the
6553 * current task can be in any other state. Note, idle is always in the
6554 * TASK_RUNNING state.
6556 WARN_ON_ONCE(current->__state);
6558 __schedule(SM_NONE);
6559 } while (need_resched());
6562 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6563 asmlinkage __visible void __sched schedule_user(void)
6566 * If we come here after a random call to set_need_resched(),
6567 * or we have been woken up remotely but the IPI has not yet arrived,
6568 * we haven't yet exited the RCU idle mode. Do it here manually until
6569 * we find a better solution.
6571 * NB: There are buggy callers of this function. Ideally we
6572 * should warn if prev_state != CONTEXT_USER, but that will trigger
6573 * too frequently to make sense yet.
6575 enum ctx_state prev_state = exception_enter();
6577 exception_exit(prev_state);
6582 * schedule_preempt_disabled - called with preemption disabled
6584 * Returns with preemption disabled. Note: preempt_count must be 1
6586 void __sched schedule_preempt_disabled(void)
6588 sched_preempt_enable_no_resched();
6593 #ifdef CONFIG_PREEMPT_RT
6594 void __sched notrace schedule_rtlock(void)
6598 __schedule(SM_RTLOCK_WAIT);
6599 sched_preempt_enable_no_resched();
6600 } while (need_resched());
6602 NOKPROBE_SYMBOL(schedule_rtlock);
6605 static void __sched notrace preempt_schedule_common(void)
6609 * Because the function tracer can trace preempt_count_sub()
6610 * and it also uses preempt_enable/disable_notrace(), if
6611 * NEED_RESCHED is set, the preempt_enable_notrace() called
6612 * by the function tracer will call this function again and
6613 * cause infinite recursion.
6615 * Preemption must be disabled here before the function
6616 * tracer can trace. Break up preempt_disable() into two
6617 * calls. One to disable preemption without fear of being
6618 * traced. The other to still record the preemption latency,
6619 * which can also be traced by the function tracer.
6621 preempt_disable_notrace();
6622 preempt_latency_start(1);
6623 __schedule(SM_PREEMPT);
6624 preempt_latency_stop(1);
6625 preempt_enable_no_resched_notrace();
6628 * Check again in case we missed a preemption opportunity
6629 * between schedule and now.
6631 } while (need_resched());
6634 #ifdef CONFIG_PREEMPTION
6636 * This is the entry point to schedule() from in-kernel preemption
6637 * off of preempt_enable.
6639 asmlinkage __visible void __sched notrace preempt_schedule(void)
6642 * If there is a non-zero preempt_count or interrupts are disabled,
6643 * we do not want to preempt the current task. Just return..
6645 if (likely(!preemptible()))
6647 preempt_schedule_common();
6649 NOKPROBE_SYMBOL(preempt_schedule);
6650 EXPORT_SYMBOL(preempt_schedule);
6652 #ifdef CONFIG_PREEMPT_DYNAMIC
6653 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6654 #ifndef preempt_schedule_dynamic_enabled
6655 #define preempt_schedule_dynamic_enabled preempt_schedule
6656 #define preempt_schedule_dynamic_disabled NULL
6658 DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6659 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6660 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6661 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6662 void __sched notrace dynamic_preempt_schedule(void)
6664 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6668 NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6669 EXPORT_SYMBOL(dynamic_preempt_schedule);
6674 * preempt_schedule_notrace - preempt_schedule called by tracing
6676 * The tracing infrastructure uses preempt_enable_notrace to prevent
6677 * recursion and tracing preempt enabling caused by the tracing
6678 * infrastructure itself. But as tracing can happen in areas coming
6679 * from userspace or just about to enter userspace, a preempt enable
6680 * can occur before user_exit() is called. This will cause the scheduler
6681 * to be called when the system is still in usermode.
6683 * To prevent this, the preempt_enable_notrace will use this function
6684 * instead of preempt_schedule() to exit user context if needed before
6685 * calling the scheduler.
6687 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6689 enum ctx_state prev_ctx;
6691 if (likely(!preemptible()))
6696 * Because the function tracer can trace preempt_count_sub()
6697 * and it also uses preempt_enable/disable_notrace(), if
6698 * NEED_RESCHED is set, the preempt_enable_notrace() called
6699 * by the function tracer will call this function again and
6700 * cause infinite recursion.
6702 * Preemption must be disabled here before the function
6703 * tracer can trace. Break up preempt_disable() into two
6704 * calls. One to disable preemption without fear of being
6705 * traced. The other to still record the preemption latency,
6706 * which can also be traced by the function tracer.
6708 preempt_disable_notrace();
6709 preempt_latency_start(1);
6711 * Needs preempt disabled in case user_exit() is traced
6712 * and the tracer calls preempt_enable_notrace() causing
6713 * an infinite recursion.
6715 prev_ctx = exception_enter();
6716 __schedule(SM_PREEMPT);
6717 exception_exit(prev_ctx);
6719 preempt_latency_stop(1);
6720 preempt_enable_no_resched_notrace();
6721 } while (need_resched());
6723 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6725 #ifdef CONFIG_PREEMPT_DYNAMIC
6726 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6727 #ifndef preempt_schedule_notrace_dynamic_enabled
6728 #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6729 #define preempt_schedule_notrace_dynamic_disabled NULL
6731 DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6732 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6733 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6734 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6735 void __sched notrace dynamic_preempt_schedule_notrace(void)
6737 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6739 preempt_schedule_notrace();
6741 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6742 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6746 #endif /* CONFIG_PREEMPTION */
6749 * This is the entry point to schedule() from kernel preemption
6750 * off of irq context.
6751 * Note, that this is called and return with irqs disabled. This will
6752 * protect us against recursive calling from irq.
6754 asmlinkage __visible void __sched preempt_schedule_irq(void)
6756 enum ctx_state prev_state;
6758 /* Catch callers which need to be fixed */
6759 BUG_ON(preempt_count() || !irqs_disabled());
6761 prev_state = exception_enter();
6766 __schedule(SM_PREEMPT);
6767 local_irq_disable();
6768 sched_preempt_enable_no_resched();
6769 } while (need_resched());
6771 exception_exit(prev_state);
6774 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6777 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6778 return try_to_wake_up(curr->private, mode, wake_flags);
6780 EXPORT_SYMBOL(default_wake_function);
6782 static void __setscheduler_prio(struct task_struct *p, int prio)
6785 p->sched_class = &dl_sched_class;
6786 else if (rt_prio(prio))
6787 p->sched_class = &rt_sched_class;
6789 p->sched_class = &fair_sched_class;
6794 #ifdef CONFIG_RT_MUTEXES
6796 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6799 prio = min(prio, pi_task->prio);
6804 static inline int rt_effective_prio(struct task_struct *p, int prio)
6806 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6808 return __rt_effective_prio(pi_task, prio);
6812 * rt_mutex_setprio - set the current priority of a task
6814 * @pi_task: donor task
6816 * This function changes the 'effective' priority of a task. It does
6817 * not touch ->normal_prio like __setscheduler().
6819 * Used by the rt_mutex code to implement priority inheritance
6820 * logic. Call site only calls if the priority of the task changed.
6822 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6824 int prio, oldprio, queued, running, queue_flag =
6825 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6826 const struct sched_class *prev_class;
6830 /* XXX used to be waiter->prio, not waiter->task->prio */
6831 prio = __rt_effective_prio(pi_task, p->normal_prio);
6834 * If nothing changed; bail early.
6836 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6839 rq = __task_rq_lock(p, &rf);
6840 update_rq_clock(rq);
6842 * Set under pi_lock && rq->lock, such that the value can be used under
6845 * Note that there is loads of tricky to make this pointer cache work
6846 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6847 * ensure a task is de-boosted (pi_task is set to NULL) before the
6848 * task is allowed to run again (and can exit). This ensures the pointer
6849 * points to a blocked task -- which guarantees the task is present.
6851 p->pi_top_task = pi_task;
6854 * For FIFO/RR we only need to set prio, if that matches we're done.
6856 if (prio == p->prio && !dl_prio(prio))
6860 * Idle task boosting is a nono in general. There is one
6861 * exception, when PREEMPT_RT and NOHZ is active:
6863 * The idle task calls get_next_timer_interrupt() and holds
6864 * the timer wheel base->lock on the CPU and another CPU wants
6865 * to access the timer (probably to cancel it). We can safely
6866 * ignore the boosting request, as the idle CPU runs this code
6867 * with interrupts disabled and will complete the lock
6868 * protected section without being interrupted. So there is no
6869 * real need to boost.
6871 if (unlikely(p == rq->idle)) {
6872 WARN_ON(p != rq->curr);
6873 WARN_ON(p->pi_blocked_on);
6877 trace_sched_pi_setprio(p, pi_task);
6880 if (oldprio == prio)
6881 queue_flag &= ~DEQUEUE_MOVE;
6883 prev_class = p->sched_class;
6884 queued = task_on_rq_queued(p);
6885 running = task_current(rq, p);
6887 dequeue_task(rq, p, queue_flag);
6889 put_prev_task(rq, p);
6892 * Boosting condition are:
6893 * 1. -rt task is running and holds mutex A
6894 * --> -dl task blocks on mutex A
6896 * 2. -dl task is running and holds mutex A
6897 * --> -dl task blocks on mutex A and could preempt the
6900 if (dl_prio(prio)) {
6901 if (!dl_prio(p->normal_prio) ||
6902 (pi_task && dl_prio(pi_task->prio) &&
6903 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6904 p->dl.pi_se = pi_task->dl.pi_se;
6905 queue_flag |= ENQUEUE_REPLENISH;
6907 p->dl.pi_se = &p->dl;
6909 } else if (rt_prio(prio)) {
6910 if (dl_prio(oldprio))
6911 p->dl.pi_se = &p->dl;
6913 queue_flag |= ENQUEUE_HEAD;
6915 if (dl_prio(oldprio))
6916 p->dl.pi_se = &p->dl;
6917 if (rt_prio(oldprio))
6921 __setscheduler_prio(p, prio);
6924 enqueue_task(rq, p, queue_flag);
6926 set_next_task(rq, p);
6928 check_class_changed(rq, p, prev_class, oldprio);
6930 /* Avoid rq from going away on us: */
6933 rq_unpin_lock(rq, &rf);
6934 __balance_callbacks(rq);
6935 raw_spin_rq_unlock(rq);
6940 static inline int rt_effective_prio(struct task_struct *p, int prio)
6946 void set_user_nice(struct task_struct *p, long nice)
6948 bool queued, running;
6953 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6956 * We have to be careful, if called from sys_setpriority(),
6957 * the task might be in the middle of scheduling on another CPU.
6959 rq = task_rq_lock(p, &rf);
6960 update_rq_clock(rq);
6963 * The RT priorities are set via sched_setscheduler(), but we still
6964 * allow the 'normal' nice value to be set - but as expected
6965 * it won't have any effect on scheduling until the task is
6966 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6968 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6969 p->static_prio = NICE_TO_PRIO(nice);
6972 queued = task_on_rq_queued(p);
6973 running = task_current(rq, p);
6975 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6977 put_prev_task(rq, p);
6979 p->static_prio = NICE_TO_PRIO(nice);
6980 set_load_weight(p, true);
6982 p->prio = effective_prio(p);
6985 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6987 set_next_task(rq, p);
6990 * If the task increased its priority or is running and
6991 * lowered its priority, then reschedule its CPU:
6993 p->sched_class->prio_changed(rq, p, old_prio);
6996 task_rq_unlock(rq, p, &rf);
6998 EXPORT_SYMBOL(set_user_nice);
7001 * can_nice - check if a task can reduce its nice value
7005 int can_nice(const struct task_struct *p, const int nice)
7007 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7008 int nice_rlim = nice_to_rlimit(nice);
7010 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
7011 capable(CAP_SYS_NICE));
7014 #ifdef __ARCH_WANT_SYS_NICE
7017 * sys_nice - change the priority of the current process.
7018 * @increment: priority increment
7020 * sys_setpriority is a more generic, but much slower function that
7021 * does similar things.
7023 SYSCALL_DEFINE1(nice, int, increment)
7028 * Setpriority might change our priority at the same moment.
7029 * We don't have to worry. Conceptually one call occurs first
7030 * and we have a single winner.
7032 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7033 nice = task_nice(current) + increment;
7035 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7036 if (increment < 0 && !can_nice(current, nice))
7039 retval = security_task_setnice(current, nice);
7043 set_user_nice(current, nice);
7050 * task_prio - return the priority value of a given task.
7051 * @p: the task in question.
7053 * Return: The priority value as seen by users in /proc.
7055 * sched policy return value kernel prio user prio/nice
7057 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7058 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7059 * deadline -101 -1 0
7061 int task_prio(const struct task_struct *p)
7063 return p->prio - MAX_RT_PRIO;
7067 * idle_cpu - is a given CPU idle currently?
7068 * @cpu: the processor in question.
7070 * Return: 1 if the CPU is currently idle. 0 otherwise.
7072 int idle_cpu(int cpu)
7074 struct rq *rq = cpu_rq(cpu);
7076 if (rq->curr != rq->idle)
7083 if (rq->ttwu_pending)
7091 * available_idle_cpu - is a given CPU idle for enqueuing work.
7092 * @cpu: the CPU in question.
7094 * Return: 1 if the CPU is currently idle. 0 otherwise.
7096 int available_idle_cpu(int cpu)
7101 if (vcpu_is_preempted(cpu))
7108 * idle_task - return the idle task for a given CPU.
7109 * @cpu: the processor in question.
7111 * Return: The idle task for the CPU @cpu.
7113 struct task_struct *idle_task(int cpu)
7115 return cpu_rq(cpu)->idle;
7120 * This function computes an effective utilization for the given CPU, to be
7121 * used for frequency selection given the linear relation: f = u * f_max.
7123 * The scheduler tracks the following metrics:
7125 * cpu_util_{cfs,rt,dl,irq}()
7128 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7129 * synchronized windows and are thus directly comparable.
7131 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7132 * which excludes things like IRQ and steal-time. These latter are then accrued
7133 * in the irq utilization.
7135 * The DL bandwidth number otoh is not a measured metric but a value computed
7136 * based on the task model parameters and gives the minimal utilization
7137 * required to meet deadlines.
7139 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7140 unsigned long max, enum cpu_util_type type,
7141 struct task_struct *p)
7143 unsigned long dl_util, util, irq;
7144 struct rq *rq = cpu_rq(cpu);
7146 if (!uclamp_is_used() &&
7147 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7152 * Early check to see if IRQ/steal time saturates the CPU, can be
7153 * because of inaccuracies in how we track these -- see
7154 * update_irq_load_avg().
7156 irq = cpu_util_irq(rq);
7157 if (unlikely(irq >= max))
7161 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7162 * CFS tasks and we use the same metric to track the effective
7163 * utilization (PELT windows are synchronized) we can directly add them
7164 * to obtain the CPU's actual utilization.
7166 * CFS and RT utilization can be boosted or capped, depending on
7167 * utilization clamp constraints requested by currently RUNNABLE
7169 * When there are no CFS RUNNABLE tasks, clamps are released and
7170 * frequency will be gracefully reduced with the utilization decay.
7172 util = util_cfs + cpu_util_rt(rq);
7173 if (type == FREQUENCY_UTIL)
7174 util = uclamp_rq_util_with(rq, util, p);
7176 dl_util = cpu_util_dl(rq);
7179 * For frequency selection we do not make cpu_util_dl() a permanent part
7180 * of this sum because we want to use cpu_bw_dl() later on, but we need
7181 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7182 * that we select f_max when there is no idle time.
7184 * NOTE: numerical errors or stop class might cause us to not quite hit
7185 * saturation when we should -- something for later.
7187 if (util + dl_util >= max)
7191 * OTOH, for energy computation we need the estimated running time, so
7192 * include util_dl and ignore dl_bw.
7194 if (type == ENERGY_UTIL)
7198 * There is still idle time; further improve the number by using the
7199 * irq metric. Because IRQ/steal time is hidden from the task clock we
7200 * need to scale the task numbers:
7203 * U' = irq + --------- * U
7206 util = scale_irq_capacity(util, irq, max);
7210 * Bandwidth required by DEADLINE must always be granted while, for
7211 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7212 * to gracefully reduce the frequency when no tasks show up for longer
7215 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7216 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7217 * an interface. So, we only do the latter for now.
7219 if (type == FREQUENCY_UTIL)
7220 util += cpu_bw_dl(rq);
7222 return min(max, util);
7225 unsigned long sched_cpu_util(int cpu, unsigned long max)
7227 return effective_cpu_util(cpu, cpu_util_cfs(cpu), max,
7230 #endif /* CONFIG_SMP */
7233 * find_process_by_pid - find a process with a matching PID value.
7234 * @pid: the pid in question.
7236 * The task of @pid, if found. %NULL otherwise.
7238 static struct task_struct *find_process_by_pid(pid_t pid)
7240 return pid ? find_task_by_vpid(pid) : current;
7244 * sched_setparam() passes in -1 for its policy, to let the functions
7245 * it calls know not to change it.
7247 #define SETPARAM_POLICY -1
7249 static void __setscheduler_params(struct task_struct *p,
7250 const struct sched_attr *attr)
7252 int policy = attr->sched_policy;
7254 if (policy == SETPARAM_POLICY)
7259 if (dl_policy(policy))
7260 __setparam_dl(p, attr);
7261 else if (fair_policy(policy))
7262 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7265 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7266 * !rt_policy. Always setting this ensures that things like
7267 * getparam()/getattr() don't report silly values for !rt tasks.
7269 p->rt_priority = attr->sched_priority;
7270 p->normal_prio = normal_prio(p);
7271 set_load_weight(p, true);
7275 * Check the target process has a UID that matches the current process's:
7277 static bool check_same_owner(struct task_struct *p)
7279 const struct cred *cred = current_cred(), *pcred;
7283 pcred = __task_cred(p);
7284 match = (uid_eq(cred->euid, pcred->euid) ||
7285 uid_eq(cred->euid, pcred->uid));
7290 static int __sched_setscheduler(struct task_struct *p,
7291 const struct sched_attr *attr,
7294 int oldpolicy = -1, policy = attr->sched_policy;
7295 int retval, oldprio, newprio, queued, running;
7296 const struct sched_class *prev_class;
7297 struct callback_head *head;
7300 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7303 /* The pi code expects interrupts enabled */
7304 BUG_ON(pi && in_interrupt());
7306 /* Double check policy once rq lock held: */
7308 reset_on_fork = p->sched_reset_on_fork;
7309 policy = oldpolicy = p->policy;
7311 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7313 if (!valid_policy(policy))
7317 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7321 * Valid priorities for SCHED_FIFO and SCHED_RR are
7322 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7323 * SCHED_BATCH and SCHED_IDLE is 0.
7325 if (attr->sched_priority > MAX_RT_PRIO-1)
7327 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7328 (rt_policy(policy) != (attr->sched_priority != 0)))
7332 * Allow unprivileged RT tasks to decrease priority:
7334 if (user && !capable(CAP_SYS_NICE)) {
7335 if (fair_policy(policy)) {
7336 if (attr->sched_nice < task_nice(p) &&
7337 !can_nice(p, attr->sched_nice))
7341 if (rt_policy(policy)) {
7342 unsigned long rlim_rtprio =
7343 task_rlimit(p, RLIMIT_RTPRIO);
7345 /* Can't set/change the rt policy: */
7346 if (policy != p->policy && !rlim_rtprio)
7349 /* Can't increase priority: */
7350 if (attr->sched_priority > p->rt_priority &&
7351 attr->sched_priority > rlim_rtprio)
7356 * Can't set/change SCHED_DEADLINE policy at all for now
7357 * (safest behavior); in the future we would like to allow
7358 * unprivileged DL tasks to increase their relative deadline
7359 * or reduce their runtime (both ways reducing utilization)
7361 if (dl_policy(policy))
7365 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7366 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7368 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7369 if (!can_nice(p, task_nice(p)))
7373 /* Can't change other user's priorities: */
7374 if (!check_same_owner(p))
7377 /* Normal users shall not reset the sched_reset_on_fork flag: */
7378 if (p->sched_reset_on_fork && !reset_on_fork)
7383 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7386 retval = security_task_setscheduler(p);
7391 /* Update task specific "requested" clamps */
7392 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7393 retval = uclamp_validate(p, attr);
7402 * Make sure no PI-waiters arrive (or leave) while we are
7403 * changing the priority of the task:
7405 * To be able to change p->policy safely, the appropriate
7406 * runqueue lock must be held.
7408 rq = task_rq_lock(p, &rf);
7409 update_rq_clock(rq);
7412 * Changing the policy of the stop threads its a very bad idea:
7414 if (p == rq->stop) {
7420 * If not changing anything there's no need to proceed further,
7421 * but store a possible modification of reset_on_fork.
7423 if (unlikely(policy == p->policy)) {
7424 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7426 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7428 if (dl_policy(policy) && dl_param_changed(p, attr))
7430 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7433 p->sched_reset_on_fork = reset_on_fork;
7440 #ifdef CONFIG_RT_GROUP_SCHED
7442 * Do not allow realtime tasks into groups that have no runtime
7445 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7446 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7447 !task_group_is_autogroup(task_group(p))) {
7453 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7454 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7455 cpumask_t *span = rq->rd->span;
7458 * Don't allow tasks with an affinity mask smaller than
7459 * the entire root_domain to become SCHED_DEADLINE. We
7460 * will also fail if there's no bandwidth available.
7462 if (!cpumask_subset(span, p->cpus_ptr) ||
7463 rq->rd->dl_bw.bw == 0) {
7471 /* Re-check policy now with rq lock held: */
7472 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7473 policy = oldpolicy = -1;
7474 task_rq_unlock(rq, p, &rf);
7476 cpuset_read_unlock();
7481 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7482 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7485 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7490 p->sched_reset_on_fork = reset_on_fork;
7493 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7496 * Take priority boosted tasks into account. If the new
7497 * effective priority is unchanged, we just store the new
7498 * normal parameters and do not touch the scheduler class and
7499 * the runqueue. This will be done when the task deboost
7502 newprio = rt_effective_prio(p, newprio);
7503 if (newprio == oldprio)
7504 queue_flags &= ~DEQUEUE_MOVE;
7507 queued = task_on_rq_queued(p);
7508 running = task_current(rq, p);
7510 dequeue_task(rq, p, queue_flags);
7512 put_prev_task(rq, p);
7514 prev_class = p->sched_class;
7516 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7517 __setscheduler_params(p, attr);
7518 __setscheduler_prio(p, newprio);
7520 __setscheduler_uclamp(p, attr);
7524 * We enqueue to tail when the priority of a task is
7525 * increased (user space view).
7527 if (oldprio < p->prio)
7528 queue_flags |= ENQUEUE_HEAD;
7530 enqueue_task(rq, p, queue_flags);
7533 set_next_task(rq, p);
7535 check_class_changed(rq, p, prev_class, oldprio);
7537 /* Avoid rq from going away on us: */
7539 head = splice_balance_callbacks(rq);
7540 task_rq_unlock(rq, p, &rf);
7543 cpuset_read_unlock();
7544 rt_mutex_adjust_pi(p);
7547 /* Run balance callbacks after we've adjusted the PI chain: */
7548 balance_callbacks(rq, head);
7554 task_rq_unlock(rq, p, &rf);
7556 cpuset_read_unlock();
7560 static int _sched_setscheduler(struct task_struct *p, int policy,
7561 const struct sched_param *param, bool check)
7563 struct sched_attr attr = {
7564 .sched_policy = policy,
7565 .sched_priority = param->sched_priority,
7566 .sched_nice = PRIO_TO_NICE(p->static_prio),
7569 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7570 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7571 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7572 policy &= ~SCHED_RESET_ON_FORK;
7573 attr.sched_policy = policy;
7576 return __sched_setscheduler(p, &attr, check, true);
7579 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7580 * @p: the task in question.
7581 * @policy: new policy.
7582 * @param: structure containing the new RT priority.
7584 * Use sched_set_fifo(), read its comment.
7586 * Return: 0 on success. An error code otherwise.
7588 * NOTE that the task may be already dead.
7590 int sched_setscheduler(struct task_struct *p, int policy,
7591 const struct sched_param *param)
7593 return _sched_setscheduler(p, policy, param, true);
7596 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7598 return __sched_setscheduler(p, attr, true, true);
7601 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7603 return __sched_setscheduler(p, attr, false, true);
7605 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7608 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7609 * @p: the task in question.
7610 * @policy: new policy.
7611 * @param: structure containing the new RT priority.
7613 * Just like sched_setscheduler, only don't bother checking if the
7614 * current context has permission. For example, this is needed in
7615 * stop_machine(): we create temporary high priority worker threads,
7616 * but our caller might not have that capability.
7618 * Return: 0 on success. An error code otherwise.
7620 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7621 const struct sched_param *param)
7623 return _sched_setscheduler(p, policy, param, false);
7627 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7628 * incapable of resource management, which is the one thing an OS really should
7631 * This is of course the reason it is limited to privileged users only.
7633 * Worse still; it is fundamentally impossible to compose static priority
7634 * workloads. You cannot take two correctly working static prio workloads
7635 * and smash them together and still expect them to work.
7637 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7641 * The administrator _MUST_ configure the system, the kernel simply doesn't
7642 * know enough information to make a sensible choice.
7644 void sched_set_fifo(struct task_struct *p)
7646 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7647 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7649 EXPORT_SYMBOL_GPL(sched_set_fifo);
7652 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7654 void sched_set_fifo_low(struct task_struct *p)
7656 struct sched_param sp = { .sched_priority = 1 };
7657 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7659 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7661 void sched_set_normal(struct task_struct *p, int nice)
7663 struct sched_attr attr = {
7664 .sched_policy = SCHED_NORMAL,
7667 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7669 EXPORT_SYMBOL_GPL(sched_set_normal);
7672 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7674 struct sched_param lparam;
7675 struct task_struct *p;
7678 if (!param || pid < 0)
7680 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7685 p = find_process_by_pid(pid);
7691 retval = sched_setscheduler(p, policy, &lparam);
7699 * Mimics kernel/events/core.c perf_copy_attr().
7701 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7706 /* Zero the full structure, so that a short copy will be nice: */
7707 memset(attr, 0, sizeof(*attr));
7709 ret = get_user(size, &uattr->size);
7713 /* ABI compatibility quirk: */
7715 size = SCHED_ATTR_SIZE_VER0;
7716 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7719 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7726 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7727 size < SCHED_ATTR_SIZE_VER1)
7731 * XXX: Do we want to be lenient like existing syscalls; or do we want
7732 * to be strict and return an error on out-of-bounds values?
7734 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7739 put_user(sizeof(*attr), &uattr->size);
7743 static void get_params(struct task_struct *p, struct sched_attr *attr)
7745 if (task_has_dl_policy(p))
7746 __getparam_dl(p, attr);
7747 else if (task_has_rt_policy(p))
7748 attr->sched_priority = p->rt_priority;
7750 attr->sched_nice = task_nice(p);
7754 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7755 * @pid: the pid in question.
7756 * @policy: new policy.
7757 * @param: structure containing the new RT priority.
7759 * Return: 0 on success. An error code otherwise.
7761 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7766 return do_sched_setscheduler(pid, policy, param);
7770 * sys_sched_setparam - set/change the RT priority of a thread
7771 * @pid: the pid in question.
7772 * @param: structure containing the new RT priority.
7774 * Return: 0 on success. An error code otherwise.
7776 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7778 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7782 * sys_sched_setattr - same as above, but with extended sched_attr
7783 * @pid: the pid in question.
7784 * @uattr: structure containing the extended parameters.
7785 * @flags: for future extension.
7787 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7788 unsigned int, flags)
7790 struct sched_attr attr;
7791 struct task_struct *p;
7794 if (!uattr || pid < 0 || flags)
7797 retval = sched_copy_attr(uattr, &attr);
7801 if ((int)attr.sched_policy < 0)
7803 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7804 attr.sched_policy = SETPARAM_POLICY;
7808 p = find_process_by_pid(pid);
7814 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7815 get_params(p, &attr);
7816 retval = sched_setattr(p, &attr);
7824 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7825 * @pid: the pid in question.
7827 * Return: On success, the policy of the thread. Otherwise, a negative error
7830 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7832 struct task_struct *p;
7840 p = find_process_by_pid(pid);
7842 retval = security_task_getscheduler(p);
7845 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7852 * sys_sched_getparam - get the RT priority of a thread
7853 * @pid: the pid in question.
7854 * @param: structure containing the RT priority.
7856 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7859 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7861 struct sched_param lp = { .sched_priority = 0 };
7862 struct task_struct *p;
7865 if (!param || pid < 0)
7869 p = find_process_by_pid(pid);
7874 retval = security_task_getscheduler(p);
7878 if (task_has_rt_policy(p))
7879 lp.sched_priority = p->rt_priority;
7883 * This one might sleep, we cannot do it with a spinlock held ...
7885 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7895 * Copy the kernel size attribute structure (which might be larger
7896 * than what user-space knows about) to user-space.
7898 * Note that all cases are valid: user-space buffer can be larger or
7899 * smaller than the kernel-space buffer. The usual case is that both
7900 * have the same size.
7903 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7904 struct sched_attr *kattr,
7907 unsigned int ksize = sizeof(*kattr);
7909 if (!access_ok(uattr, usize))
7913 * sched_getattr() ABI forwards and backwards compatibility:
7915 * If usize == ksize then we just copy everything to user-space and all is good.
7917 * If usize < ksize then we only copy as much as user-space has space for,
7918 * this keeps ABI compatibility as well. We skip the rest.
7920 * If usize > ksize then user-space is using a newer version of the ABI,
7921 * which part the kernel doesn't know about. Just ignore it - tooling can
7922 * detect the kernel's knowledge of attributes from the attr->size value
7923 * which is set to ksize in this case.
7925 kattr->size = min(usize, ksize);
7927 if (copy_to_user(uattr, kattr, kattr->size))
7934 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7935 * @pid: the pid in question.
7936 * @uattr: structure containing the extended parameters.
7937 * @usize: sizeof(attr) for fwd/bwd comp.
7938 * @flags: for future extension.
7940 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7941 unsigned int, usize, unsigned int, flags)
7943 struct sched_attr kattr = { };
7944 struct task_struct *p;
7947 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7948 usize < SCHED_ATTR_SIZE_VER0 || flags)
7952 p = find_process_by_pid(pid);
7957 retval = security_task_getscheduler(p);
7961 kattr.sched_policy = p->policy;
7962 if (p->sched_reset_on_fork)
7963 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7964 get_params(p, &kattr);
7965 kattr.sched_flags &= SCHED_FLAG_ALL;
7967 #ifdef CONFIG_UCLAMP_TASK
7969 * This could race with another potential updater, but this is fine
7970 * because it'll correctly read the old or the new value. We don't need
7971 * to guarantee who wins the race as long as it doesn't return garbage.
7973 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
7974 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
7979 return sched_attr_copy_to_user(uattr, &kattr, usize);
7987 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
7992 * If the task isn't a deadline task or admission control is
7993 * disabled then we don't care about affinity changes.
7995 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
7999 * Since bandwidth control happens on root_domain basis,
8000 * if admission test is enabled, we only admit -deadline
8001 * tasks allowed to run on all the CPUs in the task's
8005 if (!cpumask_subset(task_rq(p)->rd->span, mask))
8013 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
8016 cpumask_var_t cpus_allowed, new_mask;
8018 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8021 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8023 goto out_free_cpus_allowed;
8026 cpuset_cpus_allowed(p, cpus_allowed);
8027 cpumask_and(new_mask, mask, cpus_allowed);
8029 retval = dl_task_check_affinity(p, new_mask);
8031 goto out_free_new_mask;
8033 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
8035 goto out_free_new_mask;
8037 cpuset_cpus_allowed(p, cpus_allowed);
8038 if (!cpumask_subset(new_mask, cpus_allowed)) {
8040 * We must have raced with a concurrent cpuset update.
8041 * Just reset the cpumask to the cpuset's cpus_allowed.
8043 cpumask_copy(new_mask, cpus_allowed);
8048 free_cpumask_var(new_mask);
8049 out_free_cpus_allowed:
8050 free_cpumask_var(cpus_allowed);
8054 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8056 struct task_struct *p;
8061 p = find_process_by_pid(pid);
8067 /* Prevent p going away */
8071 if (p->flags & PF_NO_SETAFFINITY) {
8076 if (!check_same_owner(p)) {
8078 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8086 retval = security_task_setscheduler(p);
8090 retval = __sched_setaffinity(p, in_mask);
8096 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8097 struct cpumask *new_mask)
8099 if (len < cpumask_size())
8100 cpumask_clear(new_mask);
8101 else if (len > cpumask_size())
8102 len = cpumask_size();
8104 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8108 * sys_sched_setaffinity - set the CPU affinity of a process
8109 * @pid: pid of the process
8110 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8111 * @user_mask_ptr: user-space pointer to the new CPU mask
8113 * Return: 0 on success. An error code otherwise.
8115 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8116 unsigned long __user *, user_mask_ptr)
8118 cpumask_var_t new_mask;
8121 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8124 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8126 retval = sched_setaffinity(pid, new_mask);
8127 free_cpumask_var(new_mask);
8131 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8133 struct task_struct *p;
8134 unsigned long flags;
8140 p = find_process_by_pid(pid);
8144 retval = security_task_getscheduler(p);
8148 raw_spin_lock_irqsave(&p->pi_lock, flags);
8149 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8150 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8159 * sys_sched_getaffinity - get the CPU affinity of a process
8160 * @pid: pid of the process
8161 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8162 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8164 * Return: size of CPU mask copied to user_mask_ptr on success. An
8165 * error code otherwise.
8167 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8168 unsigned long __user *, user_mask_ptr)
8173 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8175 if (len & (sizeof(unsigned long)-1))
8178 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8181 ret = sched_getaffinity(pid, mask);
8183 unsigned int retlen = min(len, cpumask_size());
8185 if (copy_to_user(user_mask_ptr, mask, retlen))
8190 free_cpumask_var(mask);
8195 static void do_sched_yield(void)
8200 rq = this_rq_lock_irq(&rf);
8202 schedstat_inc(rq->yld_count);
8203 current->sched_class->yield_task(rq);
8206 rq_unlock_irq(rq, &rf);
8207 sched_preempt_enable_no_resched();
8213 * sys_sched_yield - yield the current processor to other threads.
8215 * This function yields the current CPU to other tasks. If there are no
8216 * other threads running on this CPU then this function will return.
8220 SYSCALL_DEFINE0(sched_yield)
8226 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8227 int __sched __cond_resched(void)
8229 if (should_resched(0)) {
8230 preempt_schedule_common();
8234 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8235 * whether the current CPU is in an RCU read-side critical section,
8236 * so the tick can report quiescent states even for CPUs looping
8237 * in kernel context. In contrast, in non-preemptible kernels,
8238 * RCU readers leave no in-memory hints, which means that CPU-bound
8239 * processes executing in kernel context might never report an
8240 * RCU quiescent state. Therefore, the following code causes
8241 * cond_resched() to report a quiescent state, but only when RCU
8242 * is in urgent need of one.
8244 #ifndef CONFIG_PREEMPT_RCU
8249 EXPORT_SYMBOL(__cond_resched);
8252 #ifdef CONFIG_PREEMPT_DYNAMIC
8253 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8254 #define cond_resched_dynamic_enabled __cond_resched
8255 #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8256 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8257 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8259 #define might_resched_dynamic_enabled __cond_resched
8260 #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8261 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8262 EXPORT_STATIC_CALL_TRAMP(might_resched);
8263 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8264 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8265 int __sched dynamic_cond_resched(void)
8267 if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8269 return __cond_resched();
8271 EXPORT_SYMBOL(dynamic_cond_resched);
8273 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8274 int __sched dynamic_might_resched(void)
8276 if (!static_branch_unlikely(&sk_dynamic_might_resched))
8278 return __cond_resched();
8280 EXPORT_SYMBOL(dynamic_might_resched);
8285 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8286 * call schedule, and on return reacquire the lock.
8288 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8289 * operations here to prevent schedule() from being called twice (once via
8290 * spin_unlock(), once by hand).
8292 int __cond_resched_lock(spinlock_t *lock)
8294 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8297 lockdep_assert_held(lock);
8299 if (spin_needbreak(lock) || resched) {
8301 if (!_cond_resched())
8308 EXPORT_SYMBOL(__cond_resched_lock);
8310 int __cond_resched_rwlock_read(rwlock_t *lock)
8312 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8315 lockdep_assert_held_read(lock);
8317 if (rwlock_needbreak(lock) || resched) {
8319 if (!_cond_resched())
8326 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8328 int __cond_resched_rwlock_write(rwlock_t *lock)
8330 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8333 lockdep_assert_held_write(lock);
8335 if (rwlock_needbreak(lock) || resched) {
8337 if (!_cond_resched())
8344 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8346 #ifdef CONFIG_PREEMPT_DYNAMIC
8348 #ifdef CONFIG_GENERIC_ENTRY
8349 #include <linux/entry-common.h>
8355 * SC:preempt_schedule
8356 * SC:preempt_schedule_notrace
8357 * SC:irqentry_exit_cond_resched
8361 * cond_resched <- __cond_resched
8362 * might_resched <- RET0
8363 * preempt_schedule <- NOP
8364 * preempt_schedule_notrace <- NOP
8365 * irqentry_exit_cond_resched <- NOP
8368 * cond_resched <- __cond_resched
8369 * might_resched <- __cond_resched
8370 * preempt_schedule <- NOP
8371 * preempt_schedule_notrace <- NOP
8372 * irqentry_exit_cond_resched <- NOP
8375 * cond_resched <- RET0
8376 * might_resched <- RET0
8377 * preempt_schedule <- preempt_schedule
8378 * preempt_schedule_notrace <- preempt_schedule_notrace
8379 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8383 preempt_dynamic_undefined = -1,
8384 preempt_dynamic_none,
8385 preempt_dynamic_voluntary,
8386 preempt_dynamic_full,
8389 int preempt_dynamic_mode = preempt_dynamic_undefined;
8391 int sched_dynamic_mode(const char *str)
8393 if (!strcmp(str, "none"))
8394 return preempt_dynamic_none;
8396 if (!strcmp(str, "voluntary"))
8397 return preempt_dynamic_voluntary;
8399 if (!strcmp(str, "full"))
8400 return preempt_dynamic_full;
8405 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8406 #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8407 #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8408 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8409 #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8410 #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8412 #error "Unsupported PREEMPT_DYNAMIC mechanism"
8415 void sched_dynamic_update(int mode)
8418 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8419 * the ZERO state, which is invalid.
8421 preempt_dynamic_enable(cond_resched);
8422 preempt_dynamic_enable(might_resched);
8423 preempt_dynamic_enable(preempt_schedule);
8424 preempt_dynamic_enable(preempt_schedule_notrace);
8425 preempt_dynamic_enable(irqentry_exit_cond_resched);
8428 case preempt_dynamic_none:
8429 preempt_dynamic_enable(cond_resched);
8430 preempt_dynamic_disable(might_resched);
8431 preempt_dynamic_disable(preempt_schedule);
8432 preempt_dynamic_disable(preempt_schedule_notrace);
8433 preempt_dynamic_disable(irqentry_exit_cond_resched);
8434 pr_info("Dynamic Preempt: none\n");
8437 case preempt_dynamic_voluntary:
8438 preempt_dynamic_enable(cond_resched);
8439 preempt_dynamic_enable(might_resched);
8440 preempt_dynamic_disable(preempt_schedule);
8441 preempt_dynamic_disable(preempt_schedule_notrace);
8442 preempt_dynamic_disable(irqentry_exit_cond_resched);
8443 pr_info("Dynamic Preempt: voluntary\n");
8446 case preempt_dynamic_full:
8447 preempt_dynamic_disable(cond_resched);
8448 preempt_dynamic_disable(might_resched);
8449 preempt_dynamic_enable(preempt_schedule);
8450 preempt_dynamic_enable(preempt_schedule_notrace);
8451 preempt_dynamic_enable(irqentry_exit_cond_resched);
8452 pr_info("Dynamic Preempt: full\n");
8456 preempt_dynamic_mode = mode;
8459 static int __init setup_preempt_mode(char *str)
8461 int mode = sched_dynamic_mode(str);
8463 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8467 sched_dynamic_update(mode);
8470 __setup("preempt=", setup_preempt_mode);
8472 static void __init preempt_dynamic_init(void)
8474 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8475 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8476 sched_dynamic_update(preempt_dynamic_none);
8477 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8478 sched_dynamic_update(preempt_dynamic_voluntary);
8480 /* Default static call setting, nothing to do */
8481 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8482 preempt_dynamic_mode = preempt_dynamic_full;
8483 pr_info("Dynamic Preempt: full\n");
8488 #define PREEMPT_MODEL_ACCESSOR(mode) \
8489 bool preempt_model_##mode(void) \
8491 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8492 return preempt_dynamic_mode == preempt_dynamic_##mode; \
8494 EXPORT_SYMBOL_GPL(preempt_model_##mode)
8496 PREEMPT_MODEL_ACCESSOR(none);
8497 PREEMPT_MODEL_ACCESSOR(voluntary);
8498 PREEMPT_MODEL_ACCESSOR(full);
8500 #else /* !CONFIG_PREEMPT_DYNAMIC */
8502 static inline void preempt_dynamic_init(void) { }
8504 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8507 * yield - yield the current processor to other threads.
8509 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8511 * The scheduler is at all times free to pick the calling task as the most
8512 * eligible task to run, if removing the yield() call from your code breaks
8513 * it, it's already broken.
8515 * Typical broken usage is:
8520 * where one assumes that yield() will let 'the other' process run that will
8521 * make event true. If the current task is a SCHED_FIFO task that will never
8522 * happen. Never use yield() as a progress guarantee!!
8524 * If you want to use yield() to wait for something, use wait_event().
8525 * If you want to use yield() to be 'nice' for others, use cond_resched().
8526 * If you still want to use yield(), do not!
8528 void __sched yield(void)
8530 set_current_state(TASK_RUNNING);
8533 EXPORT_SYMBOL(yield);
8536 * yield_to - yield the current processor to another thread in
8537 * your thread group, or accelerate that thread toward the
8538 * processor it's on.
8540 * @preempt: whether task preemption is allowed or not
8542 * It's the caller's job to ensure that the target task struct
8543 * can't go away on us before we can do any checks.
8546 * true (>0) if we indeed boosted the target task.
8547 * false (0) if we failed to boost the target.
8548 * -ESRCH if there's no task to yield to.
8550 int __sched yield_to(struct task_struct *p, bool preempt)
8552 struct task_struct *curr = current;
8553 struct rq *rq, *p_rq;
8554 unsigned long flags;
8557 local_irq_save(flags);
8563 * If we're the only runnable task on the rq and target rq also
8564 * has only one task, there's absolutely no point in yielding.
8566 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8571 double_rq_lock(rq, p_rq);
8572 if (task_rq(p) != p_rq) {
8573 double_rq_unlock(rq, p_rq);
8577 if (!curr->sched_class->yield_to_task)
8580 if (curr->sched_class != p->sched_class)
8583 if (task_running(p_rq, p) || !task_is_running(p))
8586 yielded = curr->sched_class->yield_to_task(rq, p);
8588 schedstat_inc(rq->yld_count);
8590 * Make p's CPU reschedule; pick_next_entity takes care of
8593 if (preempt && rq != p_rq)
8598 double_rq_unlock(rq, p_rq);
8600 local_irq_restore(flags);
8607 EXPORT_SYMBOL_GPL(yield_to);
8609 int io_schedule_prepare(void)
8611 int old_iowait = current->in_iowait;
8613 current->in_iowait = 1;
8614 blk_flush_plug(current->plug, true);
8618 void io_schedule_finish(int token)
8620 current->in_iowait = token;
8624 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8625 * that process accounting knows that this is a task in IO wait state.
8627 long __sched io_schedule_timeout(long timeout)
8632 token = io_schedule_prepare();
8633 ret = schedule_timeout(timeout);
8634 io_schedule_finish(token);
8638 EXPORT_SYMBOL(io_schedule_timeout);
8640 void __sched io_schedule(void)
8644 token = io_schedule_prepare();
8646 io_schedule_finish(token);
8648 EXPORT_SYMBOL(io_schedule);
8651 * sys_sched_get_priority_max - return maximum RT priority.
8652 * @policy: scheduling class.
8654 * Return: On success, this syscall returns the maximum
8655 * rt_priority that can be used by a given scheduling class.
8656 * On failure, a negative error code is returned.
8658 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8665 ret = MAX_RT_PRIO-1;
8667 case SCHED_DEADLINE:
8678 * sys_sched_get_priority_min - return minimum RT priority.
8679 * @policy: scheduling class.
8681 * Return: On success, this syscall returns the minimum
8682 * rt_priority that can be used by a given scheduling class.
8683 * On failure, a negative error code is returned.
8685 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8694 case SCHED_DEADLINE:
8703 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8705 struct task_struct *p;
8706 unsigned int time_slice;
8716 p = find_process_by_pid(pid);
8720 retval = security_task_getscheduler(p);
8724 rq = task_rq_lock(p, &rf);
8726 if (p->sched_class->get_rr_interval)
8727 time_slice = p->sched_class->get_rr_interval(rq, p);
8728 task_rq_unlock(rq, p, &rf);
8731 jiffies_to_timespec64(time_slice, t);
8740 * sys_sched_rr_get_interval - return the default timeslice of a process.
8741 * @pid: pid of the process.
8742 * @interval: userspace pointer to the timeslice value.
8744 * this syscall writes the default timeslice value of a given process
8745 * into the user-space timespec buffer. A value of '0' means infinity.
8747 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8750 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8751 struct __kernel_timespec __user *, interval)
8753 struct timespec64 t;
8754 int retval = sched_rr_get_interval(pid, &t);
8757 retval = put_timespec64(&t, interval);
8762 #ifdef CONFIG_COMPAT_32BIT_TIME
8763 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8764 struct old_timespec32 __user *, interval)
8766 struct timespec64 t;
8767 int retval = sched_rr_get_interval(pid, &t);
8770 retval = put_old_timespec32(&t, interval);
8775 void sched_show_task(struct task_struct *p)
8777 unsigned long free = 0;
8780 if (!try_get_task_stack(p))
8783 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8785 if (task_is_running(p))
8786 pr_cont(" running task ");
8787 #ifdef CONFIG_DEBUG_STACK_USAGE
8788 free = stack_not_used(p);
8793 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8795 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8796 free, task_pid_nr(p), ppid,
8797 read_task_thread_flags(p));
8799 print_worker_info(KERN_INFO, p);
8800 print_stop_info(KERN_INFO, p);
8801 show_stack(p, NULL, KERN_INFO);
8804 EXPORT_SYMBOL_GPL(sched_show_task);
8807 state_filter_match(unsigned long state_filter, struct task_struct *p)
8809 unsigned int state = READ_ONCE(p->__state);
8811 /* no filter, everything matches */
8815 /* filter, but doesn't match */
8816 if (!(state & state_filter))
8820 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8823 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8830 void show_state_filter(unsigned int state_filter)
8832 struct task_struct *g, *p;
8835 for_each_process_thread(g, p) {
8837 * reset the NMI-timeout, listing all files on a slow
8838 * console might take a lot of time:
8839 * Also, reset softlockup watchdogs on all CPUs, because
8840 * another CPU might be blocked waiting for us to process
8843 touch_nmi_watchdog();
8844 touch_all_softlockup_watchdogs();
8845 if (state_filter_match(state_filter, p))
8849 #ifdef CONFIG_SCHED_DEBUG
8851 sysrq_sched_debug_show();
8855 * Only show locks if all tasks are dumped:
8858 debug_show_all_locks();
8862 * init_idle - set up an idle thread for a given CPU
8863 * @idle: task in question
8864 * @cpu: CPU the idle task belongs to
8866 * NOTE: this function does not set the idle thread's NEED_RESCHED
8867 * flag, to make booting more robust.
8869 void __init init_idle(struct task_struct *idle, int cpu)
8871 struct rq *rq = cpu_rq(cpu);
8872 unsigned long flags;
8874 __sched_fork(0, idle);
8876 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8877 raw_spin_rq_lock(rq);
8879 idle->__state = TASK_RUNNING;
8880 idle->se.exec_start = sched_clock();
8882 * PF_KTHREAD should already be set at this point; regardless, make it
8883 * look like a proper per-CPU kthread.
8885 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8886 kthread_set_per_cpu(idle, cpu);
8890 * It's possible that init_idle() gets called multiple times on a task,
8891 * in that case do_set_cpus_allowed() will not do the right thing.
8893 * And since this is boot we can forgo the serialization.
8895 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8898 * We're having a chicken and egg problem, even though we are
8899 * holding rq->lock, the CPU isn't yet set to this CPU so the
8900 * lockdep check in task_group() will fail.
8902 * Similar case to sched_fork(). / Alternatively we could
8903 * use task_rq_lock() here and obtain the other rq->lock.
8908 __set_task_cpu(idle, cpu);
8912 rcu_assign_pointer(rq->curr, idle);
8913 idle->on_rq = TASK_ON_RQ_QUEUED;
8917 raw_spin_rq_unlock(rq);
8918 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8920 /* Set the preempt count _outside_ the spinlocks! */
8921 init_idle_preempt_count(idle, cpu);
8924 * The idle tasks have their own, simple scheduling class:
8926 idle->sched_class = &idle_sched_class;
8927 ftrace_graph_init_idle_task(idle, cpu);
8928 vtime_init_idle(idle, cpu);
8930 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8936 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8937 const struct cpumask *trial)
8941 if (cpumask_empty(cur))
8944 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8949 int task_can_attach(struct task_struct *p,
8950 const struct cpumask *cs_cpus_allowed)
8955 * Kthreads which disallow setaffinity shouldn't be moved
8956 * to a new cpuset; we don't want to change their CPU
8957 * affinity and isolating such threads by their set of
8958 * allowed nodes is unnecessary. Thus, cpusets are not
8959 * applicable for such threads. This prevents checking for
8960 * success of set_cpus_allowed_ptr() on all attached tasks
8961 * before cpus_mask may be changed.
8963 if (p->flags & PF_NO_SETAFFINITY) {
8968 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8970 int cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed);
8972 ret = dl_cpu_busy(cpu, p);
8979 bool sched_smp_initialized __read_mostly;
8981 #ifdef CONFIG_NUMA_BALANCING
8982 /* Migrate current task p to target_cpu */
8983 int migrate_task_to(struct task_struct *p, int target_cpu)
8985 struct migration_arg arg = { p, target_cpu };
8986 int curr_cpu = task_cpu(p);
8988 if (curr_cpu == target_cpu)
8991 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8994 /* TODO: This is not properly updating schedstats */
8996 trace_sched_move_numa(p, curr_cpu, target_cpu);
8997 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9001 * Requeue a task on a given node and accurately track the number of NUMA
9002 * tasks on the runqueues
9004 void sched_setnuma(struct task_struct *p, int nid)
9006 bool queued, running;
9010 rq = task_rq_lock(p, &rf);
9011 queued = task_on_rq_queued(p);
9012 running = task_current(rq, p);
9015 dequeue_task(rq, p, DEQUEUE_SAVE);
9017 put_prev_task(rq, p);
9019 p->numa_preferred_nid = nid;
9022 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9024 set_next_task(rq, p);
9025 task_rq_unlock(rq, p, &rf);
9027 #endif /* CONFIG_NUMA_BALANCING */
9029 #ifdef CONFIG_HOTPLUG_CPU
9031 * Ensure that the idle task is using init_mm right before its CPU goes
9034 void idle_task_exit(void)
9036 struct mm_struct *mm = current->active_mm;
9038 BUG_ON(cpu_online(smp_processor_id()));
9039 BUG_ON(current != this_rq()->idle);
9041 if (mm != &init_mm) {
9042 switch_mm(mm, &init_mm, current);
9043 finish_arch_post_lock_switch();
9046 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9049 static int __balance_push_cpu_stop(void *arg)
9051 struct task_struct *p = arg;
9052 struct rq *rq = this_rq();
9056 raw_spin_lock_irq(&p->pi_lock);
9059 update_rq_clock(rq);
9061 if (task_rq(p) == rq && task_on_rq_queued(p)) {
9062 cpu = select_fallback_rq(rq->cpu, p);
9063 rq = __migrate_task(rq, &rf, p, cpu);
9067 raw_spin_unlock_irq(&p->pi_lock);
9074 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9077 * Ensure we only run per-cpu kthreads once the CPU goes !active.
9079 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9080 * effective when the hotplug motion is down.
9082 static void balance_push(struct rq *rq)
9084 struct task_struct *push_task = rq->curr;
9086 lockdep_assert_rq_held(rq);
9089 * Ensure the thing is persistent until balance_push_set(.on = false);
9091 rq->balance_callback = &balance_push_callback;
9094 * Only active while going offline and when invoked on the outgoing
9097 if (!cpu_dying(rq->cpu) || rq != this_rq())
9101 * Both the cpu-hotplug and stop task are in this case and are
9102 * required to complete the hotplug process.
9104 if (kthread_is_per_cpu(push_task) ||
9105 is_migration_disabled(push_task)) {
9108 * If this is the idle task on the outgoing CPU try to wake
9109 * up the hotplug control thread which might wait for the
9110 * last task to vanish. The rcuwait_active() check is
9111 * accurate here because the waiter is pinned on this CPU
9112 * and can't obviously be running in parallel.
9114 * On RT kernels this also has to check whether there are
9115 * pinned and scheduled out tasks on the runqueue. They
9116 * need to leave the migrate disabled section first.
9118 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9119 rcuwait_active(&rq->hotplug_wait)) {
9120 raw_spin_rq_unlock(rq);
9121 rcuwait_wake_up(&rq->hotplug_wait);
9122 raw_spin_rq_lock(rq);
9127 get_task_struct(push_task);
9129 * Temporarily drop rq->lock such that we can wake-up the stop task.
9130 * Both preemption and IRQs are still disabled.
9132 raw_spin_rq_unlock(rq);
9133 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9134 this_cpu_ptr(&push_work));
9136 * At this point need_resched() is true and we'll take the loop in
9137 * schedule(). The next pick is obviously going to be the stop task
9138 * which kthread_is_per_cpu() and will push this task away.
9140 raw_spin_rq_lock(rq);
9143 static void balance_push_set(int cpu, bool on)
9145 struct rq *rq = cpu_rq(cpu);
9148 rq_lock_irqsave(rq, &rf);
9150 WARN_ON_ONCE(rq->balance_callback);
9151 rq->balance_callback = &balance_push_callback;
9152 } else if (rq->balance_callback == &balance_push_callback) {
9153 rq->balance_callback = NULL;
9155 rq_unlock_irqrestore(rq, &rf);
9159 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9160 * inactive. All tasks which are not per CPU kernel threads are either
9161 * pushed off this CPU now via balance_push() or placed on a different CPU
9162 * during wakeup. Wait until the CPU is quiescent.
9164 static void balance_hotplug_wait(void)
9166 struct rq *rq = this_rq();
9168 rcuwait_wait_event(&rq->hotplug_wait,
9169 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9170 TASK_UNINTERRUPTIBLE);
9175 static inline void balance_push(struct rq *rq)
9179 static inline void balance_push_set(int cpu, bool on)
9183 static inline void balance_hotplug_wait(void)
9187 #endif /* CONFIG_HOTPLUG_CPU */
9189 void set_rq_online(struct rq *rq)
9192 const struct sched_class *class;
9194 cpumask_set_cpu(rq->cpu, rq->rd->online);
9197 for_each_class(class) {
9198 if (class->rq_online)
9199 class->rq_online(rq);
9204 void set_rq_offline(struct rq *rq)
9207 const struct sched_class *class;
9209 for_each_class(class) {
9210 if (class->rq_offline)
9211 class->rq_offline(rq);
9214 cpumask_clear_cpu(rq->cpu, rq->rd->online);
9220 * used to mark begin/end of suspend/resume:
9222 static int num_cpus_frozen;
9225 * Update cpusets according to cpu_active mask. If cpusets are
9226 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9227 * around partition_sched_domains().
9229 * If we come here as part of a suspend/resume, don't touch cpusets because we
9230 * want to restore it back to its original state upon resume anyway.
9232 static void cpuset_cpu_active(void)
9234 if (cpuhp_tasks_frozen) {
9236 * num_cpus_frozen tracks how many CPUs are involved in suspend
9237 * resume sequence. As long as this is not the last online
9238 * operation in the resume sequence, just build a single sched
9239 * domain, ignoring cpusets.
9241 partition_sched_domains(1, NULL, NULL);
9242 if (--num_cpus_frozen)
9245 * This is the last CPU online operation. So fall through and
9246 * restore the original sched domains by considering the
9247 * cpuset configurations.
9249 cpuset_force_rebuild();
9251 cpuset_update_active_cpus();
9254 static int cpuset_cpu_inactive(unsigned int cpu)
9256 if (!cpuhp_tasks_frozen) {
9257 int ret = dl_cpu_busy(cpu, NULL);
9261 cpuset_update_active_cpus();
9264 partition_sched_domains(1, NULL, NULL);
9269 int sched_cpu_activate(unsigned int cpu)
9271 struct rq *rq = cpu_rq(cpu);
9275 * Clear the balance_push callback and prepare to schedule
9278 balance_push_set(cpu, false);
9280 #ifdef CONFIG_SCHED_SMT
9282 * When going up, increment the number of cores with SMT present.
9284 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9285 static_branch_inc_cpuslocked(&sched_smt_present);
9287 set_cpu_active(cpu, true);
9289 if (sched_smp_initialized) {
9290 sched_update_numa(cpu, true);
9291 sched_domains_numa_masks_set(cpu);
9292 cpuset_cpu_active();
9296 * Put the rq online, if not already. This happens:
9298 * 1) In the early boot process, because we build the real domains
9299 * after all CPUs have been brought up.
9301 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9304 rq_lock_irqsave(rq, &rf);
9306 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9309 rq_unlock_irqrestore(rq, &rf);
9314 int sched_cpu_deactivate(unsigned int cpu)
9316 struct rq *rq = cpu_rq(cpu);
9321 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9322 * load balancing when not active
9324 nohz_balance_exit_idle(rq);
9326 set_cpu_active(cpu, false);
9329 * From this point forward, this CPU will refuse to run any task that
9330 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9331 * push those tasks away until this gets cleared, see
9332 * sched_cpu_dying().
9334 balance_push_set(cpu, true);
9337 * We've cleared cpu_active_mask / set balance_push, wait for all
9338 * preempt-disabled and RCU users of this state to go away such that
9339 * all new such users will observe it.
9341 * Specifically, we rely on ttwu to no longer target this CPU, see
9342 * ttwu_queue_cond() and is_cpu_allowed().
9344 * Do sync before park smpboot threads to take care the rcu boost case.
9348 rq_lock_irqsave(rq, &rf);
9350 update_rq_clock(rq);
9351 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9354 rq_unlock_irqrestore(rq, &rf);
9356 #ifdef CONFIG_SCHED_SMT
9358 * When going down, decrement the number of cores with SMT present.
9360 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9361 static_branch_dec_cpuslocked(&sched_smt_present);
9363 sched_core_cpu_deactivate(cpu);
9366 if (!sched_smp_initialized)
9369 sched_update_numa(cpu, false);
9370 ret = cpuset_cpu_inactive(cpu);
9372 balance_push_set(cpu, false);
9373 set_cpu_active(cpu, true);
9374 sched_update_numa(cpu, true);
9377 sched_domains_numa_masks_clear(cpu);
9381 static void sched_rq_cpu_starting(unsigned int cpu)
9383 struct rq *rq = cpu_rq(cpu);
9385 rq->calc_load_update = calc_load_update;
9386 update_max_interval();
9389 int sched_cpu_starting(unsigned int cpu)
9391 sched_core_cpu_starting(cpu);
9392 sched_rq_cpu_starting(cpu);
9393 sched_tick_start(cpu);
9397 #ifdef CONFIG_HOTPLUG_CPU
9400 * Invoked immediately before the stopper thread is invoked to bring the
9401 * CPU down completely. At this point all per CPU kthreads except the
9402 * hotplug thread (current) and the stopper thread (inactive) have been
9403 * either parked or have been unbound from the outgoing CPU. Ensure that
9404 * any of those which might be on the way out are gone.
9406 * If after this point a bound task is being woken on this CPU then the
9407 * responsible hotplug callback has failed to do it's job.
9408 * sched_cpu_dying() will catch it with the appropriate fireworks.
9410 int sched_cpu_wait_empty(unsigned int cpu)
9412 balance_hotplug_wait();
9417 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9418 * might have. Called from the CPU stopper task after ensuring that the
9419 * stopper is the last running task on the CPU, so nr_active count is
9420 * stable. We need to take the teardown thread which is calling this into
9421 * account, so we hand in adjust = 1 to the load calculation.
9423 * Also see the comment "Global load-average calculations".
9425 static void calc_load_migrate(struct rq *rq)
9427 long delta = calc_load_fold_active(rq, 1);
9430 atomic_long_add(delta, &calc_load_tasks);
9433 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9435 struct task_struct *g, *p;
9436 int cpu = cpu_of(rq);
9438 lockdep_assert_rq_held(rq);
9440 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9441 for_each_process_thread(g, p) {
9442 if (task_cpu(p) != cpu)
9445 if (!task_on_rq_queued(p))
9448 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9452 int sched_cpu_dying(unsigned int cpu)
9454 struct rq *rq = cpu_rq(cpu);
9457 /* Handle pending wakeups and then migrate everything off */
9458 sched_tick_stop(cpu);
9460 rq_lock_irqsave(rq, &rf);
9461 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9462 WARN(true, "Dying CPU not properly vacated!");
9463 dump_rq_tasks(rq, KERN_WARNING);
9465 rq_unlock_irqrestore(rq, &rf);
9467 calc_load_migrate(rq);
9468 update_max_interval();
9470 sched_core_cpu_dying(cpu);
9475 void __init sched_init_smp(void)
9477 sched_init_numa(NUMA_NO_NODE);
9480 * There's no userspace yet to cause hotplug operations; hence all the
9481 * CPU masks are stable and all blatant races in the below code cannot
9484 mutex_lock(&sched_domains_mutex);
9485 sched_init_domains(cpu_active_mask);
9486 mutex_unlock(&sched_domains_mutex);
9488 /* Move init over to a non-isolated CPU */
9489 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9491 current->flags &= ~PF_NO_SETAFFINITY;
9492 sched_init_granularity();
9494 init_sched_rt_class();
9495 init_sched_dl_class();
9497 sched_smp_initialized = true;
9500 static int __init migration_init(void)
9502 sched_cpu_starting(smp_processor_id());
9505 early_initcall(migration_init);
9508 void __init sched_init_smp(void)
9510 sched_init_granularity();
9512 #endif /* CONFIG_SMP */
9514 int in_sched_functions(unsigned long addr)
9516 return in_lock_functions(addr) ||
9517 (addr >= (unsigned long)__sched_text_start
9518 && addr < (unsigned long)__sched_text_end);
9521 #ifdef CONFIG_CGROUP_SCHED
9523 * Default task group.
9524 * Every task in system belongs to this group at bootup.
9526 struct task_group root_task_group;
9527 LIST_HEAD(task_groups);
9529 /* Cacheline aligned slab cache for task_group */
9530 static struct kmem_cache *task_group_cache __read_mostly;
9533 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9534 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
9536 void __init sched_init(void)
9538 unsigned long ptr = 0;
9541 /* Make sure the linker didn't screw up */
9542 BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9543 &fair_sched_class != &rt_sched_class + 1 ||
9544 &rt_sched_class != &dl_sched_class + 1);
9546 BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9551 #ifdef CONFIG_FAIR_GROUP_SCHED
9552 ptr += 2 * nr_cpu_ids * sizeof(void **);
9554 #ifdef CONFIG_RT_GROUP_SCHED
9555 ptr += 2 * nr_cpu_ids * sizeof(void **);
9558 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9560 #ifdef CONFIG_FAIR_GROUP_SCHED
9561 root_task_group.se = (struct sched_entity **)ptr;
9562 ptr += nr_cpu_ids * sizeof(void **);
9564 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9565 ptr += nr_cpu_ids * sizeof(void **);
9567 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9568 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9569 #endif /* CONFIG_FAIR_GROUP_SCHED */
9570 #ifdef CONFIG_RT_GROUP_SCHED
9571 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9572 ptr += nr_cpu_ids * sizeof(void **);
9574 root_task_group.rt_rq = (struct rt_rq **)ptr;
9575 ptr += nr_cpu_ids * sizeof(void **);
9577 #endif /* CONFIG_RT_GROUP_SCHED */
9579 #ifdef CONFIG_CPUMASK_OFFSTACK
9580 for_each_possible_cpu(i) {
9581 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9582 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9583 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
9584 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9586 #endif /* CONFIG_CPUMASK_OFFSTACK */
9588 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9591 init_defrootdomain();
9594 #ifdef CONFIG_RT_GROUP_SCHED
9595 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9596 global_rt_period(), global_rt_runtime());
9597 #endif /* CONFIG_RT_GROUP_SCHED */
9599 #ifdef CONFIG_CGROUP_SCHED
9600 task_group_cache = KMEM_CACHE(task_group, 0);
9602 list_add(&root_task_group.list, &task_groups);
9603 INIT_LIST_HEAD(&root_task_group.children);
9604 INIT_LIST_HEAD(&root_task_group.siblings);
9605 autogroup_init(&init_task);
9606 #endif /* CONFIG_CGROUP_SCHED */
9608 for_each_possible_cpu(i) {
9612 raw_spin_lock_init(&rq->__lock);
9614 rq->calc_load_active = 0;
9615 rq->calc_load_update = jiffies + LOAD_FREQ;
9616 init_cfs_rq(&rq->cfs);
9617 init_rt_rq(&rq->rt);
9618 init_dl_rq(&rq->dl);
9619 #ifdef CONFIG_FAIR_GROUP_SCHED
9620 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9621 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9623 * How much CPU bandwidth does root_task_group get?
9625 * In case of task-groups formed thr' the cgroup filesystem, it
9626 * gets 100% of the CPU resources in the system. This overall
9627 * system CPU resource is divided among the tasks of
9628 * root_task_group and its child task-groups in a fair manner,
9629 * based on each entity's (task or task-group's) weight
9630 * (se->load.weight).
9632 * In other words, if root_task_group has 10 tasks of weight
9633 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9634 * then A0's share of the CPU resource is:
9636 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9638 * We achieve this by letting root_task_group's tasks sit
9639 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9641 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9642 #endif /* CONFIG_FAIR_GROUP_SCHED */
9644 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9645 #ifdef CONFIG_RT_GROUP_SCHED
9646 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9651 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9652 rq->balance_callback = &balance_push_callback;
9653 rq->active_balance = 0;
9654 rq->next_balance = jiffies;
9659 rq->avg_idle = 2*sysctl_sched_migration_cost;
9660 rq->wake_stamp = jiffies;
9661 rq->wake_avg_idle = rq->avg_idle;
9662 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9664 INIT_LIST_HEAD(&rq->cfs_tasks);
9666 rq_attach_root(rq, &def_root_domain);
9667 #ifdef CONFIG_NO_HZ_COMMON
9668 rq->last_blocked_load_update_tick = jiffies;
9669 atomic_set(&rq->nohz_flags, 0);
9671 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9673 #ifdef CONFIG_HOTPLUG_CPU
9674 rcuwait_init(&rq->hotplug_wait);
9676 #endif /* CONFIG_SMP */
9678 atomic_set(&rq->nr_iowait, 0);
9680 #ifdef CONFIG_SCHED_CORE
9682 rq->core_pick = NULL;
9683 rq->core_enabled = 0;
9684 rq->core_tree = RB_ROOT;
9685 rq->core_forceidle_count = 0;
9686 rq->core_forceidle_occupation = 0;
9687 rq->core_forceidle_start = 0;
9689 rq->core_cookie = 0UL;
9693 set_load_weight(&init_task, false);
9696 * The boot idle thread does lazy MMU switching as well:
9699 enter_lazy_tlb(&init_mm, current);
9702 * The idle task doesn't need the kthread struct to function, but it
9703 * is dressed up as a per-CPU kthread and thus needs to play the part
9704 * if we want to avoid special-casing it in code that deals with per-CPU
9707 WARN_ON(!set_kthread_struct(current));
9710 * Make us the idle thread. Technically, schedule() should not be
9711 * called from this thread, however somewhere below it might be,
9712 * but because we are the idle thread, we just pick up running again
9713 * when this runqueue becomes "idle".
9715 init_idle(current, smp_processor_id());
9717 calc_load_update = jiffies + LOAD_FREQ;
9720 idle_thread_set_boot_cpu();
9721 balance_push_set(smp_processor_id(), false);
9723 init_sched_fair_class();
9729 preempt_dynamic_init();
9731 scheduler_running = 1;
9734 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9736 void __might_sleep(const char *file, int line)
9738 unsigned int state = get_current_state();
9740 * Blocking primitives will set (and therefore destroy) current->state,
9741 * since we will exit with TASK_RUNNING make sure we enter with it,
9742 * otherwise we will destroy state.
9744 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9745 "do not call blocking ops when !TASK_RUNNING; "
9746 "state=%x set at [<%p>] %pS\n", state,
9747 (void *)current->task_state_change,
9748 (void *)current->task_state_change);
9750 __might_resched(file, line, 0);
9752 EXPORT_SYMBOL(__might_sleep);
9754 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9756 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9759 if (preempt_count() == preempt_offset)
9762 pr_err("Preemption disabled at:");
9763 print_ip_sym(KERN_ERR, ip);
9766 static inline bool resched_offsets_ok(unsigned int offsets)
9768 unsigned int nested = preempt_count();
9770 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9772 return nested == offsets;
9775 void __might_resched(const char *file, int line, unsigned int offsets)
9777 /* Ratelimiting timestamp: */
9778 static unsigned long prev_jiffy;
9780 unsigned long preempt_disable_ip;
9782 /* WARN_ON_ONCE() by default, no rate limit required: */
9785 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9786 !is_idle_task(current) && !current->non_block_count) ||
9787 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9791 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9793 prev_jiffy = jiffies;
9795 /* Save this before calling printk(), since that will clobber it: */
9796 preempt_disable_ip = get_preempt_disable_ip(current);
9798 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9800 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9801 in_atomic(), irqs_disabled(), current->non_block_count,
9802 current->pid, current->comm);
9803 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9804 offsets & MIGHT_RESCHED_PREEMPT_MASK);
9806 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9807 pr_err("RCU nest depth: %d, expected: %u\n",
9808 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9811 if (task_stack_end_corrupted(current))
9812 pr_emerg("Thread overran stack, or stack corrupted\n");
9814 debug_show_held_locks(current);
9815 if (irqs_disabled())
9816 print_irqtrace_events(current);
9818 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9819 preempt_disable_ip);
9822 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9824 EXPORT_SYMBOL(__might_resched);
9826 void __cant_sleep(const char *file, int line, int preempt_offset)
9828 static unsigned long prev_jiffy;
9830 if (irqs_disabled())
9833 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9836 if (preempt_count() > preempt_offset)
9839 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9841 prev_jiffy = jiffies;
9843 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9844 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9845 in_atomic(), irqs_disabled(),
9846 current->pid, current->comm);
9848 debug_show_held_locks(current);
9850 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9852 EXPORT_SYMBOL_GPL(__cant_sleep);
9855 void __cant_migrate(const char *file, int line)
9857 static unsigned long prev_jiffy;
9859 if (irqs_disabled())
9862 if (is_migration_disabled(current))
9865 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9868 if (preempt_count() > 0)
9871 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9873 prev_jiffy = jiffies;
9875 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9876 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9877 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9878 current->pid, current->comm);
9880 debug_show_held_locks(current);
9882 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9884 EXPORT_SYMBOL_GPL(__cant_migrate);
9888 #ifdef CONFIG_MAGIC_SYSRQ
9889 void normalize_rt_tasks(void)
9891 struct task_struct *g, *p;
9892 struct sched_attr attr = {
9893 .sched_policy = SCHED_NORMAL,
9896 read_lock(&tasklist_lock);
9897 for_each_process_thread(g, p) {
9899 * Only normalize user tasks:
9901 if (p->flags & PF_KTHREAD)
9904 p->se.exec_start = 0;
9905 schedstat_set(p->stats.wait_start, 0);
9906 schedstat_set(p->stats.sleep_start, 0);
9907 schedstat_set(p->stats.block_start, 0);
9909 if (!dl_task(p) && !rt_task(p)) {
9911 * Renice negative nice level userspace
9914 if (task_nice(p) < 0)
9915 set_user_nice(p, 0);
9919 __sched_setscheduler(p, &attr, false, false);
9921 read_unlock(&tasklist_lock);
9924 #endif /* CONFIG_MAGIC_SYSRQ */
9926 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9928 * These functions are only useful for the IA64 MCA handling, or kdb.
9930 * They can only be called when the whole system has been
9931 * stopped - every CPU needs to be quiescent, and no scheduling
9932 * activity can take place. Using them for anything else would
9933 * be a serious bug, and as a result, they aren't even visible
9934 * under any other configuration.
9938 * curr_task - return the current task for a given CPU.
9939 * @cpu: the processor in question.
9941 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9943 * Return: The current task for @cpu.
9945 struct task_struct *curr_task(int cpu)
9947 return cpu_curr(cpu);
9950 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9954 * ia64_set_curr_task - set the current task for a given CPU.
9955 * @cpu: the processor in question.
9956 * @p: the task pointer to set.
9958 * Description: This function must only be used when non-maskable interrupts
9959 * are serviced on a separate stack. It allows the architecture to switch the
9960 * notion of the current task on a CPU in a non-blocking manner. This function
9961 * must be called with all CPU's synchronized, and interrupts disabled, the
9962 * and caller must save the original value of the current task (see
9963 * curr_task() above) and restore that value before reenabling interrupts and
9964 * re-starting the system.
9966 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9968 void ia64_set_curr_task(int cpu, struct task_struct *p)
9975 #ifdef CONFIG_CGROUP_SCHED
9976 /* task_group_lock serializes the addition/removal of task groups */
9977 static DEFINE_SPINLOCK(task_group_lock);
9979 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9980 struct task_group *parent)
9982 #ifdef CONFIG_UCLAMP_TASK_GROUP
9983 enum uclamp_id clamp_id;
9985 for_each_clamp_id(clamp_id) {
9986 uclamp_se_set(&tg->uclamp_req[clamp_id],
9987 uclamp_none(clamp_id), false);
9988 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9993 static void sched_free_group(struct task_group *tg)
9995 free_fair_sched_group(tg);
9996 free_rt_sched_group(tg);
9998 kmem_cache_free(task_group_cache, tg);
10001 static void sched_free_group_rcu(struct rcu_head *rcu)
10003 sched_free_group(container_of(rcu, struct task_group, rcu));
10006 static void sched_unregister_group(struct task_group *tg)
10008 unregister_fair_sched_group(tg);
10009 unregister_rt_sched_group(tg);
10011 * We have to wait for yet another RCU grace period to expire, as
10012 * print_cfs_stats() might run concurrently.
10014 call_rcu(&tg->rcu, sched_free_group_rcu);
10017 /* allocate runqueue etc for a new task group */
10018 struct task_group *sched_create_group(struct task_group *parent)
10020 struct task_group *tg;
10022 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10024 return ERR_PTR(-ENOMEM);
10026 if (!alloc_fair_sched_group(tg, parent))
10029 if (!alloc_rt_sched_group(tg, parent))
10032 alloc_uclamp_sched_group(tg, parent);
10037 sched_free_group(tg);
10038 return ERR_PTR(-ENOMEM);
10041 void sched_online_group(struct task_group *tg, struct task_group *parent)
10043 unsigned long flags;
10045 spin_lock_irqsave(&task_group_lock, flags);
10046 list_add_rcu(&tg->list, &task_groups);
10048 /* Root should already exist: */
10051 tg->parent = parent;
10052 INIT_LIST_HEAD(&tg->children);
10053 list_add_rcu(&tg->siblings, &parent->children);
10054 spin_unlock_irqrestore(&task_group_lock, flags);
10056 online_fair_sched_group(tg);
10059 /* rcu callback to free various structures associated with a task group */
10060 static void sched_unregister_group_rcu(struct rcu_head *rhp)
10062 /* Now it should be safe to free those cfs_rqs: */
10063 sched_unregister_group(container_of(rhp, struct task_group, rcu));
10066 void sched_destroy_group(struct task_group *tg)
10068 /* Wait for possible concurrent references to cfs_rqs complete: */
10069 call_rcu(&tg->rcu, sched_unregister_group_rcu);
10072 void sched_release_group(struct task_group *tg)
10074 unsigned long flags;
10077 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10078 * sched_cfs_period_timer()).
10080 * For this to be effective, we have to wait for all pending users of
10081 * this task group to leave their RCU critical section to ensure no new
10082 * user will see our dying task group any more. Specifically ensure
10083 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10085 * We therefore defer calling unregister_fair_sched_group() to
10086 * sched_unregister_group() which is guarantied to get called only after the
10087 * current RCU grace period has expired.
10089 spin_lock_irqsave(&task_group_lock, flags);
10090 list_del_rcu(&tg->list);
10091 list_del_rcu(&tg->siblings);
10092 spin_unlock_irqrestore(&task_group_lock, flags);
10095 static void sched_change_group(struct task_struct *tsk, int type)
10097 struct task_group *tg;
10100 * All callers are synchronized by task_rq_lock(); we do not use RCU
10101 * which is pointless here. Thus, we pass "true" to task_css_check()
10102 * to prevent lockdep warnings.
10104 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10105 struct task_group, css);
10106 tg = autogroup_task_group(tsk, tg);
10107 tsk->sched_task_group = tg;
10109 #ifdef CONFIG_FAIR_GROUP_SCHED
10110 if (tsk->sched_class->task_change_group)
10111 tsk->sched_class->task_change_group(tsk, type);
10114 set_task_rq(tsk, task_cpu(tsk));
10118 * Change task's runqueue when it moves between groups.
10120 * The caller of this function should have put the task in its new group by
10121 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10124 void sched_move_task(struct task_struct *tsk)
10126 int queued, running, queue_flags =
10127 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10128 struct rq_flags rf;
10131 rq = task_rq_lock(tsk, &rf);
10132 update_rq_clock(rq);
10134 running = task_current(rq, tsk);
10135 queued = task_on_rq_queued(tsk);
10138 dequeue_task(rq, tsk, queue_flags);
10140 put_prev_task(rq, tsk);
10142 sched_change_group(tsk, TASK_MOVE_GROUP);
10145 enqueue_task(rq, tsk, queue_flags);
10147 set_next_task(rq, tsk);
10149 * After changing group, the running task may have joined a
10150 * throttled one but it's still the running task. Trigger a
10151 * resched to make sure that task can still run.
10156 task_rq_unlock(rq, tsk, &rf);
10159 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10161 return css ? container_of(css, struct task_group, css) : NULL;
10164 static struct cgroup_subsys_state *
10165 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10167 struct task_group *parent = css_tg(parent_css);
10168 struct task_group *tg;
10171 /* This is early initialization for the top cgroup */
10172 return &root_task_group.css;
10175 tg = sched_create_group(parent);
10177 return ERR_PTR(-ENOMEM);
10182 /* Expose task group only after completing cgroup initialization */
10183 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10185 struct task_group *tg = css_tg(css);
10186 struct task_group *parent = css_tg(css->parent);
10189 sched_online_group(tg, parent);
10191 #ifdef CONFIG_UCLAMP_TASK_GROUP
10192 /* Propagate the effective uclamp value for the new group */
10193 mutex_lock(&uclamp_mutex);
10195 cpu_util_update_eff(css);
10197 mutex_unlock(&uclamp_mutex);
10203 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10205 struct task_group *tg = css_tg(css);
10207 sched_release_group(tg);
10210 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10212 struct task_group *tg = css_tg(css);
10215 * Relies on the RCU grace period between css_released() and this.
10217 sched_unregister_group(tg);
10221 * This is called before wake_up_new_task(), therefore we really only
10222 * have to set its group bits, all the other stuff does not apply.
10224 static void cpu_cgroup_fork(struct task_struct *task)
10226 struct rq_flags rf;
10229 rq = task_rq_lock(task, &rf);
10231 update_rq_clock(rq);
10232 sched_change_group(task, TASK_SET_GROUP);
10234 task_rq_unlock(rq, task, &rf);
10237 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10239 struct task_struct *task;
10240 struct cgroup_subsys_state *css;
10243 cgroup_taskset_for_each(task, css, tset) {
10244 #ifdef CONFIG_RT_GROUP_SCHED
10245 if (!sched_rt_can_attach(css_tg(css), task))
10249 * Serialize against wake_up_new_task() such that if it's
10250 * running, we're sure to observe its full state.
10252 raw_spin_lock_irq(&task->pi_lock);
10254 * Avoid calling sched_move_task() before wake_up_new_task()
10255 * has happened. This would lead to problems with PELT, due to
10256 * move wanting to detach+attach while we're not attached yet.
10258 if (READ_ONCE(task->__state) == TASK_NEW)
10260 raw_spin_unlock_irq(&task->pi_lock);
10268 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10270 struct task_struct *task;
10271 struct cgroup_subsys_state *css;
10273 cgroup_taskset_for_each(task, css, tset)
10274 sched_move_task(task);
10277 #ifdef CONFIG_UCLAMP_TASK_GROUP
10278 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10280 struct cgroup_subsys_state *top_css = css;
10281 struct uclamp_se *uc_parent = NULL;
10282 struct uclamp_se *uc_se = NULL;
10283 unsigned int eff[UCLAMP_CNT];
10284 enum uclamp_id clamp_id;
10285 unsigned int clamps;
10287 lockdep_assert_held(&uclamp_mutex);
10288 SCHED_WARN_ON(!rcu_read_lock_held());
10290 css_for_each_descendant_pre(css, top_css) {
10291 uc_parent = css_tg(css)->parent
10292 ? css_tg(css)->parent->uclamp : NULL;
10294 for_each_clamp_id(clamp_id) {
10295 /* Assume effective clamps matches requested clamps */
10296 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10297 /* Cap effective clamps with parent's effective clamps */
10299 eff[clamp_id] > uc_parent[clamp_id].value) {
10300 eff[clamp_id] = uc_parent[clamp_id].value;
10303 /* Ensure protection is always capped by limit */
10304 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10306 /* Propagate most restrictive effective clamps */
10308 uc_se = css_tg(css)->uclamp;
10309 for_each_clamp_id(clamp_id) {
10310 if (eff[clamp_id] == uc_se[clamp_id].value)
10312 uc_se[clamp_id].value = eff[clamp_id];
10313 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10314 clamps |= (0x1 << clamp_id);
10317 css = css_rightmost_descendant(css);
10321 /* Immediately update descendants RUNNABLE tasks */
10322 uclamp_update_active_tasks(css);
10327 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10328 * C expression. Since there is no way to convert a macro argument (N) into a
10329 * character constant, use two levels of macros.
10331 #define _POW10(exp) ((unsigned int)1e##exp)
10332 #define POW10(exp) _POW10(exp)
10334 struct uclamp_request {
10335 #define UCLAMP_PERCENT_SHIFT 2
10336 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10342 static inline struct uclamp_request
10343 capacity_from_percent(char *buf)
10345 struct uclamp_request req = {
10346 .percent = UCLAMP_PERCENT_SCALE,
10347 .util = SCHED_CAPACITY_SCALE,
10352 if (strcmp(buf, "max")) {
10353 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10357 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10362 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10363 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10369 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10370 size_t nbytes, loff_t off,
10371 enum uclamp_id clamp_id)
10373 struct uclamp_request req;
10374 struct task_group *tg;
10376 req = capacity_from_percent(buf);
10380 static_branch_enable(&sched_uclamp_used);
10382 mutex_lock(&uclamp_mutex);
10385 tg = css_tg(of_css(of));
10386 if (tg->uclamp_req[clamp_id].value != req.util)
10387 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10390 * Because of not recoverable conversion rounding we keep track of the
10391 * exact requested value
10393 tg->uclamp_pct[clamp_id] = req.percent;
10395 /* Update effective clamps to track the most restrictive value */
10396 cpu_util_update_eff(of_css(of));
10399 mutex_unlock(&uclamp_mutex);
10404 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10405 char *buf, size_t nbytes,
10408 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10411 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10412 char *buf, size_t nbytes,
10415 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10418 static inline void cpu_uclamp_print(struct seq_file *sf,
10419 enum uclamp_id clamp_id)
10421 struct task_group *tg;
10427 tg = css_tg(seq_css(sf));
10428 util_clamp = tg->uclamp_req[clamp_id].value;
10431 if (util_clamp == SCHED_CAPACITY_SCALE) {
10432 seq_puts(sf, "max\n");
10436 percent = tg->uclamp_pct[clamp_id];
10437 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10438 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10441 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10443 cpu_uclamp_print(sf, UCLAMP_MIN);
10447 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10449 cpu_uclamp_print(sf, UCLAMP_MAX);
10452 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10454 #ifdef CONFIG_FAIR_GROUP_SCHED
10455 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10456 struct cftype *cftype, u64 shareval)
10458 if (shareval > scale_load_down(ULONG_MAX))
10459 shareval = MAX_SHARES;
10460 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10463 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10464 struct cftype *cft)
10466 struct task_group *tg = css_tg(css);
10468 return (u64) scale_load_down(tg->shares);
10471 #ifdef CONFIG_CFS_BANDWIDTH
10472 static DEFINE_MUTEX(cfs_constraints_mutex);
10474 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10475 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10476 /* More than 203 days if BW_SHIFT equals 20. */
10477 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10479 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10481 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10484 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10485 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10487 if (tg == &root_task_group)
10491 * Ensure we have at some amount of bandwidth every period. This is
10492 * to prevent reaching a state of large arrears when throttled via
10493 * entity_tick() resulting in prolonged exit starvation.
10495 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10499 * Likewise, bound things on the other side by preventing insane quota
10500 * periods. This also allows us to normalize in computing quota
10503 if (period > max_cfs_quota_period)
10507 * Bound quota to defend quota against overflow during bandwidth shift.
10509 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10512 if (quota != RUNTIME_INF && (burst > quota ||
10513 burst + quota > max_cfs_runtime))
10517 * Prevent race between setting of cfs_rq->runtime_enabled and
10518 * unthrottle_offline_cfs_rqs().
10521 mutex_lock(&cfs_constraints_mutex);
10522 ret = __cfs_schedulable(tg, period, quota);
10526 runtime_enabled = quota != RUNTIME_INF;
10527 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10529 * If we need to toggle cfs_bandwidth_used, off->on must occur
10530 * before making related changes, and on->off must occur afterwards
10532 if (runtime_enabled && !runtime_was_enabled)
10533 cfs_bandwidth_usage_inc();
10534 raw_spin_lock_irq(&cfs_b->lock);
10535 cfs_b->period = ns_to_ktime(period);
10536 cfs_b->quota = quota;
10537 cfs_b->burst = burst;
10539 __refill_cfs_bandwidth_runtime(cfs_b);
10541 /* Restart the period timer (if active) to handle new period expiry: */
10542 if (runtime_enabled)
10543 start_cfs_bandwidth(cfs_b);
10545 raw_spin_unlock_irq(&cfs_b->lock);
10547 for_each_online_cpu(i) {
10548 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10549 struct rq *rq = cfs_rq->rq;
10550 struct rq_flags rf;
10552 rq_lock_irq(rq, &rf);
10553 cfs_rq->runtime_enabled = runtime_enabled;
10554 cfs_rq->runtime_remaining = 0;
10556 if (cfs_rq->throttled)
10557 unthrottle_cfs_rq(cfs_rq);
10558 rq_unlock_irq(rq, &rf);
10560 if (runtime_was_enabled && !runtime_enabled)
10561 cfs_bandwidth_usage_dec();
10563 mutex_unlock(&cfs_constraints_mutex);
10564 cpus_read_unlock();
10569 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10571 u64 quota, period, burst;
10573 period = ktime_to_ns(tg->cfs_bandwidth.period);
10574 burst = tg->cfs_bandwidth.burst;
10575 if (cfs_quota_us < 0)
10576 quota = RUNTIME_INF;
10577 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10578 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10582 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10585 static long tg_get_cfs_quota(struct task_group *tg)
10589 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10592 quota_us = tg->cfs_bandwidth.quota;
10593 do_div(quota_us, NSEC_PER_USEC);
10598 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10600 u64 quota, period, burst;
10602 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10605 period = (u64)cfs_period_us * NSEC_PER_USEC;
10606 quota = tg->cfs_bandwidth.quota;
10607 burst = tg->cfs_bandwidth.burst;
10609 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10612 static long tg_get_cfs_period(struct task_group *tg)
10616 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10617 do_div(cfs_period_us, NSEC_PER_USEC);
10619 return cfs_period_us;
10622 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10624 u64 quota, period, burst;
10626 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10629 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10630 period = ktime_to_ns(tg->cfs_bandwidth.period);
10631 quota = tg->cfs_bandwidth.quota;
10633 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10636 static long tg_get_cfs_burst(struct task_group *tg)
10640 burst_us = tg->cfs_bandwidth.burst;
10641 do_div(burst_us, NSEC_PER_USEC);
10646 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10647 struct cftype *cft)
10649 return tg_get_cfs_quota(css_tg(css));
10652 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10653 struct cftype *cftype, s64 cfs_quota_us)
10655 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10658 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10659 struct cftype *cft)
10661 return tg_get_cfs_period(css_tg(css));
10664 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10665 struct cftype *cftype, u64 cfs_period_us)
10667 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10670 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10671 struct cftype *cft)
10673 return tg_get_cfs_burst(css_tg(css));
10676 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10677 struct cftype *cftype, u64 cfs_burst_us)
10679 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10682 struct cfs_schedulable_data {
10683 struct task_group *tg;
10688 * normalize group quota/period to be quota/max_period
10689 * note: units are usecs
10691 static u64 normalize_cfs_quota(struct task_group *tg,
10692 struct cfs_schedulable_data *d)
10697 period = d->period;
10700 period = tg_get_cfs_period(tg);
10701 quota = tg_get_cfs_quota(tg);
10704 /* note: these should typically be equivalent */
10705 if (quota == RUNTIME_INF || quota == -1)
10706 return RUNTIME_INF;
10708 return to_ratio(period, quota);
10711 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10713 struct cfs_schedulable_data *d = data;
10714 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10715 s64 quota = 0, parent_quota = -1;
10718 quota = RUNTIME_INF;
10720 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10722 quota = normalize_cfs_quota(tg, d);
10723 parent_quota = parent_b->hierarchical_quota;
10726 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10727 * always take the min. On cgroup1, only inherit when no
10730 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10731 quota = min(quota, parent_quota);
10733 if (quota == RUNTIME_INF)
10734 quota = parent_quota;
10735 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10739 cfs_b->hierarchical_quota = quota;
10744 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10747 struct cfs_schedulable_data data = {
10753 if (quota != RUNTIME_INF) {
10754 do_div(data.period, NSEC_PER_USEC);
10755 do_div(data.quota, NSEC_PER_USEC);
10759 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10765 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10767 struct task_group *tg = css_tg(seq_css(sf));
10768 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10770 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10771 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10772 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10774 if (schedstat_enabled() && tg != &root_task_group) {
10775 struct sched_statistics *stats;
10779 for_each_possible_cpu(i) {
10780 stats = __schedstats_from_se(tg->se[i]);
10781 ws += schedstat_val(stats->wait_sum);
10784 seq_printf(sf, "wait_sum %llu\n", ws);
10787 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10788 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10792 #endif /* CONFIG_CFS_BANDWIDTH */
10793 #endif /* CONFIG_FAIR_GROUP_SCHED */
10795 #ifdef CONFIG_RT_GROUP_SCHED
10796 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10797 struct cftype *cft, s64 val)
10799 return sched_group_set_rt_runtime(css_tg(css), val);
10802 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10803 struct cftype *cft)
10805 return sched_group_rt_runtime(css_tg(css));
10808 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10809 struct cftype *cftype, u64 rt_period_us)
10811 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10814 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10815 struct cftype *cft)
10817 return sched_group_rt_period(css_tg(css));
10819 #endif /* CONFIG_RT_GROUP_SCHED */
10821 #ifdef CONFIG_FAIR_GROUP_SCHED
10822 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10823 struct cftype *cft)
10825 return css_tg(css)->idle;
10828 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10829 struct cftype *cft, s64 idle)
10831 return sched_group_set_idle(css_tg(css), idle);
10835 static struct cftype cpu_legacy_files[] = {
10836 #ifdef CONFIG_FAIR_GROUP_SCHED
10839 .read_u64 = cpu_shares_read_u64,
10840 .write_u64 = cpu_shares_write_u64,
10844 .read_s64 = cpu_idle_read_s64,
10845 .write_s64 = cpu_idle_write_s64,
10848 #ifdef CONFIG_CFS_BANDWIDTH
10850 .name = "cfs_quota_us",
10851 .read_s64 = cpu_cfs_quota_read_s64,
10852 .write_s64 = cpu_cfs_quota_write_s64,
10855 .name = "cfs_period_us",
10856 .read_u64 = cpu_cfs_period_read_u64,
10857 .write_u64 = cpu_cfs_period_write_u64,
10860 .name = "cfs_burst_us",
10861 .read_u64 = cpu_cfs_burst_read_u64,
10862 .write_u64 = cpu_cfs_burst_write_u64,
10866 .seq_show = cpu_cfs_stat_show,
10869 #ifdef CONFIG_RT_GROUP_SCHED
10871 .name = "rt_runtime_us",
10872 .read_s64 = cpu_rt_runtime_read,
10873 .write_s64 = cpu_rt_runtime_write,
10876 .name = "rt_period_us",
10877 .read_u64 = cpu_rt_period_read_uint,
10878 .write_u64 = cpu_rt_period_write_uint,
10881 #ifdef CONFIG_UCLAMP_TASK_GROUP
10883 .name = "uclamp.min",
10884 .flags = CFTYPE_NOT_ON_ROOT,
10885 .seq_show = cpu_uclamp_min_show,
10886 .write = cpu_uclamp_min_write,
10889 .name = "uclamp.max",
10890 .flags = CFTYPE_NOT_ON_ROOT,
10891 .seq_show = cpu_uclamp_max_show,
10892 .write = cpu_uclamp_max_write,
10895 { } /* Terminate */
10898 static int cpu_extra_stat_show(struct seq_file *sf,
10899 struct cgroup_subsys_state *css)
10901 #ifdef CONFIG_CFS_BANDWIDTH
10903 struct task_group *tg = css_tg(css);
10904 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10905 u64 throttled_usec, burst_usec;
10907 throttled_usec = cfs_b->throttled_time;
10908 do_div(throttled_usec, NSEC_PER_USEC);
10909 burst_usec = cfs_b->burst_time;
10910 do_div(burst_usec, NSEC_PER_USEC);
10912 seq_printf(sf, "nr_periods %d\n"
10913 "nr_throttled %d\n"
10914 "throttled_usec %llu\n"
10916 "burst_usec %llu\n",
10917 cfs_b->nr_periods, cfs_b->nr_throttled,
10918 throttled_usec, cfs_b->nr_burst, burst_usec);
10924 #ifdef CONFIG_FAIR_GROUP_SCHED
10925 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10926 struct cftype *cft)
10928 struct task_group *tg = css_tg(css);
10929 u64 weight = scale_load_down(tg->shares);
10931 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10934 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10935 struct cftype *cft, u64 weight)
10938 * cgroup weight knobs should use the common MIN, DFL and MAX
10939 * values which are 1, 100 and 10000 respectively. While it loses
10940 * a bit of range on both ends, it maps pretty well onto the shares
10941 * value used by scheduler and the round-trip conversions preserve
10942 * the original value over the entire range.
10944 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10947 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10949 return sched_group_set_shares(css_tg(css), scale_load(weight));
10952 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10953 struct cftype *cft)
10955 unsigned long weight = scale_load_down(css_tg(css)->shares);
10956 int last_delta = INT_MAX;
10959 /* find the closest nice value to the current weight */
10960 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10961 delta = abs(sched_prio_to_weight[prio] - weight);
10962 if (delta >= last_delta)
10964 last_delta = delta;
10967 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10970 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10971 struct cftype *cft, s64 nice)
10973 unsigned long weight;
10976 if (nice < MIN_NICE || nice > MAX_NICE)
10979 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10980 idx = array_index_nospec(idx, 40);
10981 weight = sched_prio_to_weight[idx];
10983 return sched_group_set_shares(css_tg(css), scale_load(weight));
10987 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10988 long period, long quota)
10991 seq_puts(sf, "max");
10993 seq_printf(sf, "%ld", quota);
10995 seq_printf(sf, " %ld\n", period);
10998 /* caller should put the current value in *@periodp before calling */
10999 static int __maybe_unused cpu_period_quota_parse(char *buf,
11000 u64 *periodp, u64 *quotap)
11002 char tok[21]; /* U64_MAX */
11004 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11007 *periodp *= NSEC_PER_USEC;
11009 if (sscanf(tok, "%llu", quotap))
11010 *quotap *= NSEC_PER_USEC;
11011 else if (!strcmp(tok, "max"))
11012 *quotap = RUNTIME_INF;
11019 #ifdef CONFIG_CFS_BANDWIDTH
11020 static int cpu_max_show(struct seq_file *sf, void *v)
11022 struct task_group *tg = css_tg(seq_css(sf));
11024 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11028 static ssize_t cpu_max_write(struct kernfs_open_file *of,
11029 char *buf, size_t nbytes, loff_t off)
11031 struct task_group *tg = css_tg(of_css(of));
11032 u64 period = tg_get_cfs_period(tg);
11033 u64 burst = tg_get_cfs_burst(tg);
11037 ret = cpu_period_quota_parse(buf, &period, "a);
11039 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11040 return ret ?: nbytes;
11044 static struct cftype cpu_files[] = {
11045 #ifdef CONFIG_FAIR_GROUP_SCHED
11048 .flags = CFTYPE_NOT_ON_ROOT,
11049 .read_u64 = cpu_weight_read_u64,
11050 .write_u64 = cpu_weight_write_u64,
11053 .name = "weight.nice",
11054 .flags = CFTYPE_NOT_ON_ROOT,
11055 .read_s64 = cpu_weight_nice_read_s64,
11056 .write_s64 = cpu_weight_nice_write_s64,
11060 .flags = CFTYPE_NOT_ON_ROOT,
11061 .read_s64 = cpu_idle_read_s64,
11062 .write_s64 = cpu_idle_write_s64,
11065 #ifdef CONFIG_CFS_BANDWIDTH
11068 .flags = CFTYPE_NOT_ON_ROOT,
11069 .seq_show = cpu_max_show,
11070 .write = cpu_max_write,
11073 .name = "max.burst",
11074 .flags = CFTYPE_NOT_ON_ROOT,
11075 .read_u64 = cpu_cfs_burst_read_u64,
11076 .write_u64 = cpu_cfs_burst_write_u64,
11079 #ifdef CONFIG_UCLAMP_TASK_GROUP
11081 .name = "uclamp.min",
11082 .flags = CFTYPE_NOT_ON_ROOT,
11083 .seq_show = cpu_uclamp_min_show,
11084 .write = cpu_uclamp_min_write,
11087 .name = "uclamp.max",
11088 .flags = CFTYPE_NOT_ON_ROOT,
11089 .seq_show = cpu_uclamp_max_show,
11090 .write = cpu_uclamp_max_write,
11093 { } /* terminate */
11096 struct cgroup_subsys cpu_cgrp_subsys = {
11097 .css_alloc = cpu_cgroup_css_alloc,
11098 .css_online = cpu_cgroup_css_online,
11099 .css_released = cpu_cgroup_css_released,
11100 .css_free = cpu_cgroup_css_free,
11101 .css_extra_stat_show = cpu_extra_stat_show,
11102 .fork = cpu_cgroup_fork,
11103 .can_attach = cpu_cgroup_can_attach,
11104 .attach = cpu_cgroup_attach,
11105 .legacy_cftypes = cpu_legacy_files,
11106 .dfl_cftypes = cpu_files,
11107 .early_init = true,
11111 #endif /* CONFIG_CGROUP_SCHED */
11113 void dump_cpu_task(int cpu)
11115 pr_info("Task dump for CPU %d:\n", cpu);
11116 sched_show_task(cpu_curr(cpu));
11120 * Nice levels are multiplicative, with a gentle 10% change for every
11121 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11122 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11123 * that remained on nice 0.
11125 * The "10% effect" is relative and cumulative: from _any_ nice level,
11126 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11127 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11128 * If a task goes up by ~10% and another task goes down by ~10% then
11129 * the relative distance between them is ~25%.)
11131 const int sched_prio_to_weight[40] = {
11132 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11133 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11134 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11135 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11136 /* 0 */ 1024, 820, 655, 526, 423,
11137 /* 5 */ 335, 272, 215, 172, 137,
11138 /* 10 */ 110, 87, 70, 56, 45,
11139 /* 15 */ 36, 29, 23, 18, 15,
11143 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11145 * In cases where the weight does not change often, we can use the
11146 * precalculated inverse to speed up arithmetics by turning divisions
11147 * into multiplications:
11149 const u32 sched_prio_to_wmult[40] = {
11150 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11151 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11152 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11153 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11154 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11155 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11156 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11157 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11160 void call_trace_sched_update_nr_running(struct rq *rq, int count)
11162 trace_sched_update_nr_running_tp(rq, count);