1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #define CREATE_TRACE_POINTS
10 #include <trace/events/sched.h>
11 #undef CREATE_TRACE_POINTS
15 #include <linux/nospec.h>
17 #include <linux/kcov.h>
18 #include <linux/scs.h>
20 #include <asm/switch_to.h>
23 #include "../workqueue_internal.h"
24 #include "../../fs/io-wq.h"
25 #include "../smpboot.h"
31 * Export tracepoints that act as a bare tracehook (ie: have no trace event
32 * associated with them) to allow external modules to probe them.
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
38 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
39 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
47 #ifdef CONFIG_SCHED_DEBUG
49 * Debugging: various feature bits
51 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
52 * sysctl_sched_features, defined in sched.h, to allow constants propagation
53 * at compile time and compiler optimization based on features default.
55 #define SCHED_FEAT(name, enabled) \
56 (1UL << __SCHED_FEAT_##name) * enabled |
57 const_debug unsigned int sysctl_sched_features =
64 * Number of tasks to iterate in a single balance run.
65 * Limited because this is done with IRQs disabled.
67 const_debug unsigned int sysctl_sched_nr_migrate = 32;
70 * period over which we measure -rt task CPU usage in us.
73 unsigned int sysctl_sched_rt_period = 1000000;
75 __read_mostly int scheduler_running;
78 * part of the period that we allow rt tasks to run in us.
81 int sysctl_sched_rt_runtime = 950000;
85 * Serialization rules:
91 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
94 * rq2->lock where: rq1 < rq2
98 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
99 * local CPU's rq->lock, it optionally removes the task from the runqueue and
100 * always looks at the local rq data structures to find the most eligible task
103 * Task enqueue is also under rq->lock, possibly taken from another CPU.
104 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
105 * the local CPU to avoid bouncing the runqueue state around [ see
106 * ttwu_queue_wakelist() ]
108 * Task wakeup, specifically wakeups that involve migration, are horribly
109 * complicated to avoid having to take two rq->locks.
113 * System-calls and anything external will use task_rq_lock() which acquires
114 * both p->pi_lock and rq->lock. As a consequence the state they change is
115 * stable while holding either lock:
117 * - sched_setaffinity()/
118 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
119 * - set_user_nice(): p->se.load, p->*prio
120 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
121 * p->se.load, p->rt_priority,
122 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
123 * - sched_setnuma(): p->numa_preferred_nid
124 * - sched_move_task()/
125 * cpu_cgroup_fork(): p->sched_task_group
126 * - uclamp_update_active() p->uclamp*
128 * p->state <- TASK_*:
130 * is changed locklessly using set_current_state(), __set_current_state() or
131 * set_special_state(), see their respective comments, or by
132 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
135 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
137 * is set by activate_task() and cleared by deactivate_task(), under
138 * rq->lock. Non-zero indicates the task is runnable, the special
139 * ON_RQ_MIGRATING state is used for migration without holding both
140 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
142 * p->on_cpu <- { 0, 1 }:
144 * is set by prepare_task() and cleared by finish_task() such that it will be
145 * set before p is scheduled-in and cleared after p is scheduled-out, both
146 * under rq->lock. Non-zero indicates the task is running on its CPU.
148 * [ The astute reader will observe that it is possible for two tasks on one
149 * CPU to have ->on_cpu = 1 at the same time. ]
151 * task_cpu(p): is changed by set_task_cpu(), the rules are:
153 * - Don't call set_task_cpu() on a blocked task:
155 * We don't care what CPU we're not running on, this simplifies hotplug,
156 * the CPU assignment of blocked tasks isn't required to be valid.
158 * - for try_to_wake_up(), called under p->pi_lock:
160 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
162 * - for migration called under rq->lock:
163 * [ see task_on_rq_migrating() in task_rq_lock() ]
165 * o move_queued_task()
168 * - for migration called under double_rq_lock():
170 * o __migrate_swap_task()
171 * o push_rt_task() / pull_rt_task()
172 * o push_dl_task() / pull_dl_task()
173 * o dl_task_offline_migration()
178 * __task_rq_lock - lock the rq @p resides on.
180 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
185 lockdep_assert_held(&p->pi_lock);
189 raw_spin_lock(&rq->lock);
190 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
194 raw_spin_unlock(&rq->lock);
196 while (unlikely(task_on_rq_migrating(p)))
202 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
204 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
205 __acquires(p->pi_lock)
211 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
213 raw_spin_lock(&rq->lock);
215 * move_queued_task() task_rq_lock()
218 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
219 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
220 * [S] ->cpu = new_cpu [L] task_rq()
224 * If we observe the old CPU in task_rq_lock(), the acquire of
225 * the old rq->lock will fully serialize against the stores.
227 * If we observe the new CPU in task_rq_lock(), the address
228 * dependency headed by '[L] rq = task_rq()' and the acquire
229 * will pair with the WMB to ensure we then also see migrating.
231 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
235 raw_spin_unlock(&rq->lock);
236 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
238 while (unlikely(task_on_rq_migrating(p)))
244 * RQ-clock updating methods:
247 static void update_rq_clock_task(struct rq *rq, s64 delta)
250 * In theory, the compile should just see 0 here, and optimize out the call
251 * to sched_rt_avg_update. But I don't trust it...
253 s64 __maybe_unused steal = 0, irq_delta = 0;
255 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
256 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
259 * Since irq_time is only updated on {soft,}irq_exit, we might run into
260 * this case when a previous update_rq_clock() happened inside a
263 * When this happens, we stop ->clock_task and only update the
264 * prev_irq_time stamp to account for the part that fit, so that a next
265 * update will consume the rest. This ensures ->clock_task is
268 * It does however cause some slight miss-attribution of {soft,}irq
269 * time, a more accurate solution would be to update the irq_time using
270 * the current rq->clock timestamp, except that would require using
273 if (irq_delta > delta)
276 rq->prev_irq_time += irq_delta;
279 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
280 if (static_key_false((¶virt_steal_rq_enabled))) {
281 steal = paravirt_steal_clock(cpu_of(rq));
282 steal -= rq->prev_steal_time_rq;
284 if (unlikely(steal > delta))
287 rq->prev_steal_time_rq += steal;
292 rq->clock_task += delta;
294 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
295 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
296 update_irq_load_avg(rq, irq_delta + steal);
298 update_rq_clock_pelt(rq, delta);
301 void update_rq_clock(struct rq *rq)
305 lockdep_assert_held(&rq->lock);
307 if (rq->clock_update_flags & RQCF_ACT_SKIP)
310 #ifdef CONFIG_SCHED_DEBUG
311 if (sched_feat(WARN_DOUBLE_CLOCK))
312 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
313 rq->clock_update_flags |= RQCF_UPDATED;
316 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
320 update_rq_clock_task(rq, delta);
323 #ifdef CONFIG_SCHED_HRTICK
325 * Use HR-timers to deliver accurate preemption points.
328 static void hrtick_clear(struct rq *rq)
330 if (hrtimer_active(&rq->hrtick_timer))
331 hrtimer_cancel(&rq->hrtick_timer);
335 * High-resolution timer tick.
336 * Runs from hardirq context with interrupts disabled.
338 static enum hrtimer_restart hrtick(struct hrtimer *timer)
340 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
343 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
347 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
350 return HRTIMER_NORESTART;
355 static void __hrtick_restart(struct rq *rq)
357 struct hrtimer *timer = &rq->hrtick_timer;
358 ktime_t time = rq->hrtick_time;
360 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
364 * called from hardirq (IPI) context
366 static void __hrtick_start(void *arg)
372 __hrtick_restart(rq);
377 * Called to set the hrtick timer state.
379 * called with rq->lock held and irqs disabled
381 void hrtick_start(struct rq *rq, u64 delay)
383 struct hrtimer *timer = &rq->hrtick_timer;
387 * Don't schedule slices shorter than 10000ns, that just
388 * doesn't make sense and can cause timer DoS.
390 delta = max_t(s64, delay, 10000LL);
391 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
394 __hrtick_restart(rq);
396 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
401 * Called to set the hrtick timer state.
403 * called with rq->lock held and irqs disabled
405 void hrtick_start(struct rq *rq, u64 delay)
408 * Don't schedule slices shorter than 10000ns, that just
409 * doesn't make sense. Rely on vruntime for fairness.
411 delay = max_t(u64, delay, 10000LL);
412 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
413 HRTIMER_MODE_REL_PINNED_HARD);
416 #endif /* CONFIG_SMP */
418 static void hrtick_rq_init(struct rq *rq)
421 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
423 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
424 rq->hrtick_timer.function = hrtick;
426 #else /* CONFIG_SCHED_HRTICK */
427 static inline void hrtick_clear(struct rq *rq)
431 static inline void hrtick_rq_init(struct rq *rq)
434 #endif /* CONFIG_SCHED_HRTICK */
437 * cmpxchg based fetch_or, macro so it works for different integer types
439 #define fetch_or(ptr, mask) \
441 typeof(ptr) _ptr = (ptr); \
442 typeof(mask) _mask = (mask); \
443 typeof(*_ptr) _old, _val = *_ptr; \
446 _old = cmpxchg(_ptr, _val, _val | _mask); \
454 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
456 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
457 * this avoids any races wrt polling state changes and thereby avoids
460 static bool set_nr_and_not_polling(struct task_struct *p)
462 struct thread_info *ti = task_thread_info(p);
463 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
467 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
469 * If this returns true, then the idle task promises to call
470 * sched_ttwu_pending() and reschedule soon.
472 static bool set_nr_if_polling(struct task_struct *p)
474 struct thread_info *ti = task_thread_info(p);
475 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
478 if (!(val & _TIF_POLLING_NRFLAG))
480 if (val & _TIF_NEED_RESCHED)
482 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
491 static bool set_nr_and_not_polling(struct task_struct *p)
493 set_tsk_need_resched(p);
498 static bool set_nr_if_polling(struct task_struct *p)
505 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
507 struct wake_q_node *node = &task->wake_q;
510 * Atomically grab the task, if ->wake_q is !nil already it means
511 * it's already queued (either by us or someone else) and will get the
512 * wakeup due to that.
514 * In order to ensure that a pending wakeup will observe our pending
515 * state, even in the failed case, an explicit smp_mb() must be used.
517 smp_mb__before_atomic();
518 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
522 * The head is context local, there can be no concurrency.
525 head->lastp = &node->next;
530 * wake_q_add() - queue a wakeup for 'later' waking.
531 * @head: the wake_q_head to add @task to
532 * @task: the task to queue for 'later' wakeup
534 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
535 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
538 * This function must be used as-if it were wake_up_process(); IOW the task
539 * must be ready to be woken at this location.
541 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
543 if (__wake_q_add(head, task))
544 get_task_struct(task);
548 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
549 * @head: the wake_q_head to add @task to
550 * @task: the task to queue for 'later' wakeup
552 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
553 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
556 * This function must be used as-if it were wake_up_process(); IOW the task
557 * must be ready to be woken at this location.
559 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
560 * that already hold reference to @task can call the 'safe' version and trust
561 * wake_q to do the right thing depending whether or not the @task is already
564 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
566 if (!__wake_q_add(head, task))
567 put_task_struct(task);
570 void wake_up_q(struct wake_q_head *head)
572 struct wake_q_node *node = head->first;
574 while (node != WAKE_Q_TAIL) {
575 struct task_struct *task;
577 task = container_of(node, struct task_struct, wake_q);
579 /* Task can safely be re-inserted now: */
581 task->wake_q.next = NULL;
584 * wake_up_process() executes a full barrier, which pairs with
585 * the queueing in wake_q_add() so as not to miss wakeups.
587 wake_up_process(task);
588 put_task_struct(task);
593 * resched_curr - mark rq's current task 'to be rescheduled now'.
595 * On UP this means the setting of the need_resched flag, on SMP it
596 * might also involve a cross-CPU call to trigger the scheduler on
599 void resched_curr(struct rq *rq)
601 struct task_struct *curr = rq->curr;
604 lockdep_assert_held(&rq->lock);
606 if (test_tsk_need_resched(curr))
611 if (cpu == smp_processor_id()) {
612 set_tsk_need_resched(curr);
613 set_preempt_need_resched();
617 if (set_nr_and_not_polling(curr))
618 smp_send_reschedule(cpu);
620 trace_sched_wake_idle_without_ipi(cpu);
623 void resched_cpu(int cpu)
625 struct rq *rq = cpu_rq(cpu);
628 raw_spin_lock_irqsave(&rq->lock, flags);
629 if (cpu_online(cpu) || cpu == smp_processor_id())
631 raw_spin_unlock_irqrestore(&rq->lock, flags);
635 #ifdef CONFIG_NO_HZ_COMMON
637 * In the semi idle case, use the nearest busy CPU for migrating timers
638 * from an idle CPU. This is good for power-savings.
640 * We don't do similar optimization for completely idle system, as
641 * selecting an idle CPU will add more delays to the timers than intended
642 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
644 int get_nohz_timer_target(void)
646 int i, cpu = smp_processor_id(), default_cpu = -1;
647 struct sched_domain *sd;
649 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
656 for_each_domain(cpu, sd) {
657 for_each_cpu_and(i, sched_domain_span(sd),
658 housekeeping_cpumask(HK_FLAG_TIMER)) {
669 if (default_cpu == -1)
670 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
678 * When add_timer_on() enqueues a timer into the timer wheel of an
679 * idle CPU then this timer might expire before the next timer event
680 * which is scheduled to wake up that CPU. In case of a completely
681 * idle system the next event might even be infinite time into the
682 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
683 * leaves the inner idle loop so the newly added timer is taken into
684 * account when the CPU goes back to idle and evaluates the timer
685 * wheel for the next timer event.
687 static void wake_up_idle_cpu(int cpu)
689 struct rq *rq = cpu_rq(cpu);
691 if (cpu == smp_processor_id())
694 if (set_nr_and_not_polling(rq->idle))
695 smp_send_reschedule(cpu);
697 trace_sched_wake_idle_without_ipi(cpu);
700 static bool wake_up_full_nohz_cpu(int cpu)
703 * We just need the target to call irq_exit() and re-evaluate
704 * the next tick. The nohz full kick at least implies that.
705 * If needed we can still optimize that later with an
708 if (cpu_is_offline(cpu))
709 return true; /* Don't try to wake offline CPUs. */
710 if (tick_nohz_full_cpu(cpu)) {
711 if (cpu != smp_processor_id() ||
712 tick_nohz_tick_stopped())
713 tick_nohz_full_kick_cpu(cpu);
721 * Wake up the specified CPU. If the CPU is going offline, it is the
722 * caller's responsibility to deal with the lost wakeup, for example,
723 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
725 void wake_up_nohz_cpu(int cpu)
727 if (!wake_up_full_nohz_cpu(cpu))
728 wake_up_idle_cpu(cpu);
731 static void nohz_csd_func(void *info)
733 struct rq *rq = info;
734 int cpu = cpu_of(rq);
738 * Release the rq::nohz_csd.
740 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
741 WARN_ON(!(flags & NOHZ_KICK_MASK));
743 rq->idle_balance = idle_cpu(cpu);
744 if (rq->idle_balance && !need_resched()) {
745 rq->nohz_idle_balance = flags;
746 raise_softirq_irqoff(SCHED_SOFTIRQ);
750 #endif /* CONFIG_NO_HZ_COMMON */
752 #ifdef CONFIG_NO_HZ_FULL
753 bool sched_can_stop_tick(struct rq *rq)
757 /* Deadline tasks, even if single, need the tick */
758 if (rq->dl.dl_nr_running)
762 * If there are more than one RR tasks, we need the tick to affect the
763 * actual RR behaviour.
765 if (rq->rt.rr_nr_running) {
766 if (rq->rt.rr_nr_running == 1)
773 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
774 * forced preemption between FIFO tasks.
776 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
781 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
782 * if there's more than one we need the tick for involuntary
785 if (rq->nr_running > 1)
790 #endif /* CONFIG_NO_HZ_FULL */
791 #endif /* CONFIG_SMP */
793 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
794 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
796 * Iterate task_group tree rooted at *from, calling @down when first entering a
797 * node and @up when leaving it for the final time.
799 * Caller must hold rcu_lock or sufficient equivalent.
801 int walk_tg_tree_from(struct task_group *from,
802 tg_visitor down, tg_visitor up, void *data)
804 struct task_group *parent, *child;
810 ret = (*down)(parent, data);
813 list_for_each_entry_rcu(child, &parent->children, siblings) {
820 ret = (*up)(parent, data);
821 if (ret || parent == from)
825 parent = parent->parent;
832 int tg_nop(struct task_group *tg, void *data)
838 static void set_load_weight(struct task_struct *p, bool update_load)
840 int prio = p->static_prio - MAX_RT_PRIO;
841 struct load_weight *load = &p->se.load;
844 * SCHED_IDLE tasks get minimal weight:
846 if (task_has_idle_policy(p)) {
847 load->weight = scale_load(WEIGHT_IDLEPRIO);
848 load->inv_weight = WMULT_IDLEPRIO;
853 * SCHED_OTHER tasks have to update their load when changing their
856 if (update_load && p->sched_class == &fair_sched_class) {
857 reweight_task(p, prio);
859 load->weight = scale_load(sched_prio_to_weight[prio]);
860 load->inv_weight = sched_prio_to_wmult[prio];
864 #ifdef CONFIG_UCLAMP_TASK
866 * Serializes updates of utilization clamp values
868 * The (slow-path) user-space triggers utilization clamp value updates which
869 * can require updates on (fast-path) scheduler's data structures used to
870 * support enqueue/dequeue operations.
871 * While the per-CPU rq lock protects fast-path update operations, user-space
872 * requests are serialized using a mutex to reduce the risk of conflicting
873 * updates or API abuses.
875 static DEFINE_MUTEX(uclamp_mutex);
877 /* Max allowed minimum utilization */
878 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
880 /* Max allowed maximum utilization */
881 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
884 * By default RT tasks run at the maximum performance point/capacity of the
885 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
886 * SCHED_CAPACITY_SCALE.
888 * This knob allows admins to change the default behavior when uclamp is being
889 * used. In battery powered devices, particularly, running at the maximum
890 * capacity and frequency will increase energy consumption and shorten the
893 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
895 * This knob will not override the system default sched_util_clamp_min defined
898 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
900 /* All clamps are required to be less or equal than these values */
901 static struct uclamp_se uclamp_default[UCLAMP_CNT];
904 * This static key is used to reduce the uclamp overhead in the fast path. It
905 * primarily disables the call to uclamp_rq_{inc, dec}() in
906 * enqueue/dequeue_task().
908 * This allows users to continue to enable uclamp in their kernel config with
909 * minimum uclamp overhead in the fast path.
911 * As soon as userspace modifies any of the uclamp knobs, the static key is
912 * enabled, since we have an actual users that make use of uclamp
915 * The knobs that would enable this static key are:
917 * * A task modifying its uclamp value with sched_setattr().
918 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
919 * * An admin modifying the cgroup cpu.uclamp.{min, max}
921 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
923 /* Integer rounded range for each bucket */
924 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
926 #define for_each_clamp_id(clamp_id) \
927 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
929 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
931 return clamp_value / UCLAMP_BUCKET_DELTA;
934 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
936 if (clamp_id == UCLAMP_MIN)
938 return SCHED_CAPACITY_SCALE;
941 static inline void uclamp_se_set(struct uclamp_se *uc_se,
942 unsigned int value, bool user_defined)
944 uc_se->value = value;
945 uc_se->bucket_id = uclamp_bucket_id(value);
946 uc_se->user_defined = user_defined;
949 static inline unsigned int
950 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
951 unsigned int clamp_value)
954 * Avoid blocked utilization pushing up the frequency when we go
955 * idle (which drops the max-clamp) by retaining the last known
958 if (clamp_id == UCLAMP_MAX) {
959 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
963 return uclamp_none(UCLAMP_MIN);
966 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
967 unsigned int clamp_value)
969 /* Reset max-clamp retention only on idle exit */
970 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
973 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
977 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
978 unsigned int clamp_value)
980 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
981 int bucket_id = UCLAMP_BUCKETS - 1;
984 * Since both min and max clamps are max aggregated, find the
985 * top most bucket with tasks in.
987 for ( ; bucket_id >= 0; bucket_id--) {
988 if (!bucket[bucket_id].tasks)
990 return bucket[bucket_id].value;
993 /* No tasks -- default clamp values */
994 return uclamp_idle_value(rq, clamp_id, clamp_value);
997 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
999 unsigned int default_util_min;
1000 struct uclamp_se *uc_se;
1002 lockdep_assert_held(&p->pi_lock);
1004 uc_se = &p->uclamp_req[UCLAMP_MIN];
1006 /* Only sync if user didn't override the default */
1007 if (uc_se->user_defined)
1010 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1011 uclamp_se_set(uc_se, default_util_min, false);
1014 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1022 /* Protect updates to p->uclamp_* */
1023 rq = task_rq_lock(p, &rf);
1024 __uclamp_update_util_min_rt_default(p);
1025 task_rq_unlock(rq, p, &rf);
1028 static void uclamp_sync_util_min_rt_default(void)
1030 struct task_struct *g, *p;
1033 * copy_process() sysctl_uclamp
1034 * uclamp_min_rt = X;
1035 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1036 * // link thread smp_mb__after_spinlock()
1037 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1038 * sched_post_fork() for_each_process_thread()
1039 * __uclamp_sync_rt() __uclamp_sync_rt()
1041 * Ensures that either sched_post_fork() will observe the new
1042 * uclamp_min_rt or for_each_process_thread() will observe the new
1045 read_lock(&tasklist_lock);
1046 smp_mb__after_spinlock();
1047 read_unlock(&tasklist_lock);
1050 for_each_process_thread(g, p)
1051 uclamp_update_util_min_rt_default(p);
1055 static inline struct uclamp_se
1056 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1058 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1059 #ifdef CONFIG_UCLAMP_TASK_GROUP
1060 struct uclamp_se uc_max;
1063 * Tasks in autogroups or root task group will be
1064 * restricted by system defaults.
1066 if (task_group_is_autogroup(task_group(p)))
1068 if (task_group(p) == &root_task_group)
1071 uc_max = task_group(p)->uclamp[clamp_id];
1072 if (uc_req.value > uc_max.value || !uc_req.user_defined)
1080 * The effective clamp bucket index of a task depends on, by increasing
1082 * - the task specific clamp value, when explicitly requested from userspace
1083 * - the task group effective clamp value, for tasks not either in the root
1084 * group or in an autogroup
1085 * - the system default clamp value, defined by the sysadmin
1087 static inline struct uclamp_se
1088 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1090 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1091 struct uclamp_se uc_max = uclamp_default[clamp_id];
1093 /* System default restrictions always apply */
1094 if (unlikely(uc_req.value > uc_max.value))
1100 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1102 struct uclamp_se uc_eff;
1104 /* Task currently refcounted: use back-annotated (effective) value */
1105 if (p->uclamp[clamp_id].active)
1106 return (unsigned long)p->uclamp[clamp_id].value;
1108 uc_eff = uclamp_eff_get(p, clamp_id);
1110 return (unsigned long)uc_eff.value;
1114 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1115 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1116 * updates the rq's clamp value if required.
1118 * Tasks can have a task-specific value requested from user-space, track
1119 * within each bucket the maximum value for tasks refcounted in it.
1120 * This "local max aggregation" allows to track the exact "requested" value
1121 * for each bucket when all its RUNNABLE tasks require the same clamp.
1123 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1124 enum uclamp_id clamp_id)
1126 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1127 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1128 struct uclamp_bucket *bucket;
1130 lockdep_assert_held(&rq->lock);
1132 /* Update task effective clamp */
1133 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1135 bucket = &uc_rq->bucket[uc_se->bucket_id];
1137 uc_se->active = true;
1139 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1142 * Local max aggregation: rq buckets always track the max
1143 * "requested" clamp value of its RUNNABLE tasks.
1145 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1146 bucket->value = uc_se->value;
1148 if (uc_se->value > READ_ONCE(uc_rq->value))
1149 WRITE_ONCE(uc_rq->value, uc_se->value);
1153 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1154 * is released. If this is the last task reference counting the rq's max
1155 * active clamp value, then the rq's clamp value is updated.
1157 * Both refcounted tasks and rq's cached clamp values are expected to be
1158 * always valid. If it's detected they are not, as defensive programming,
1159 * enforce the expected state and warn.
1161 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1162 enum uclamp_id clamp_id)
1164 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1165 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1166 struct uclamp_bucket *bucket;
1167 unsigned int bkt_clamp;
1168 unsigned int rq_clamp;
1170 lockdep_assert_held(&rq->lock);
1173 * If sched_uclamp_used was enabled after task @p was enqueued,
1174 * we could end up with unbalanced call to uclamp_rq_dec_id().
1176 * In this case the uc_se->active flag should be false since no uclamp
1177 * accounting was performed at enqueue time and we can just return
1180 * Need to be careful of the following enqueue/dequeue ordering
1184 * // sched_uclamp_used gets enabled
1187 * // Must not decrement bucket->tasks here
1190 * where we could end up with stale data in uc_se and
1191 * bucket[uc_se->bucket_id].
1193 * The following check here eliminates the possibility of such race.
1195 if (unlikely(!uc_se->active))
1198 bucket = &uc_rq->bucket[uc_se->bucket_id];
1200 SCHED_WARN_ON(!bucket->tasks);
1201 if (likely(bucket->tasks))
1204 uc_se->active = false;
1207 * Keep "local max aggregation" simple and accept to (possibly)
1208 * overboost some RUNNABLE tasks in the same bucket.
1209 * The rq clamp bucket value is reset to its base value whenever
1210 * there are no more RUNNABLE tasks refcounting it.
1212 if (likely(bucket->tasks))
1215 rq_clamp = READ_ONCE(uc_rq->value);
1217 * Defensive programming: this should never happen. If it happens,
1218 * e.g. due to future modification, warn and fixup the expected value.
1220 SCHED_WARN_ON(bucket->value > rq_clamp);
1221 if (bucket->value >= rq_clamp) {
1222 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1223 WRITE_ONCE(uc_rq->value, bkt_clamp);
1227 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1229 enum uclamp_id clamp_id;
1232 * Avoid any overhead until uclamp is actually used by the userspace.
1234 * The condition is constructed such that a NOP is generated when
1235 * sched_uclamp_used is disabled.
1237 if (!static_branch_unlikely(&sched_uclamp_used))
1240 if (unlikely(!p->sched_class->uclamp_enabled))
1243 for_each_clamp_id(clamp_id)
1244 uclamp_rq_inc_id(rq, p, clamp_id);
1246 /* Reset clamp idle holding when there is one RUNNABLE task */
1247 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1248 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1251 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1253 enum uclamp_id clamp_id;
1256 * Avoid any overhead until uclamp is actually used by the userspace.
1258 * The condition is constructed such that a NOP is generated when
1259 * sched_uclamp_used is disabled.
1261 if (!static_branch_unlikely(&sched_uclamp_used))
1264 if (unlikely(!p->sched_class->uclamp_enabled))
1267 for_each_clamp_id(clamp_id)
1268 uclamp_rq_dec_id(rq, p, clamp_id);
1272 uclamp_update_active(struct task_struct *p, enum uclamp_id clamp_id)
1278 * Lock the task and the rq where the task is (or was) queued.
1280 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1281 * price to pay to safely serialize util_{min,max} updates with
1282 * enqueues, dequeues and migration operations.
1283 * This is the same locking schema used by __set_cpus_allowed_ptr().
1285 rq = task_rq_lock(p, &rf);
1288 * Setting the clamp bucket is serialized by task_rq_lock().
1289 * If the task is not yet RUNNABLE and its task_struct is not
1290 * affecting a valid clamp bucket, the next time it's enqueued,
1291 * it will already see the updated clamp bucket value.
1293 if (p->uclamp[clamp_id].active) {
1294 uclamp_rq_dec_id(rq, p, clamp_id);
1295 uclamp_rq_inc_id(rq, p, clamp_id);
1298 task_rq_unlock(rq, p, &rf);
1301 #ifdef CONFIG_UCLAMP_TASK_GROUP
1303 uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1304 unsigned int clamps)
1306 enum uclamp_id clamp_id;
1307 struct css_task_iter it;
1308 struct task_struct *p;
1310 css_task_iter_start(css, 0, &it);
1311 while ((p = css_task_iter_next(&it))) {
1312 for_each_clamp_id(clamp_id) {
1313 if ((0x1 << clamp_id) & clamps)
1314 uclamp_update_active(p, clamp_id);
1317 css_task_iter_end(&it);
1320 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1321 static void uclamp_update_root_tg(void)
1323 struct task_group *tg = &root_task_group;
1325 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1326 sysctl_sched_uclamp_util_min, false);
1327 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1328 sysctl_sched_uclamp_util_max, false);
1331 cpu_util_update_eff(&root_task_group.css);
1335 static void uclamp_update_root_tg(void) { }
1338 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1339 void *buffer, size_t *lenp, loff_t *ppos)
1341 bool update_root_tg = false;
1342 int old_min, old_max, old_min_rt;
1345 mutex_lock(&uclamp_mutex);
1346 old_min = sysctl_sched_uclamp_util_min;
1347 old_max = sysctl_sched_uclamp_util_max;
1348 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1350 result = proc_dointvec(table, write, buffer, lenp, ppos);
1356 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1357 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1358 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1364 if (old_min != sysctl_sched_uclamp_util_min) {
1365 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1366 sysctl_sched_uclamp_util_min, false);
1367 update_root_tg = true;
1369 if (old_max != sysctl_sched_uclamp_util_max) {
1370 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1371 sysctl_sched_uclamp_util_max, false);
1372 update_root_tg = true;
1375 if (update_root_tg) {
1376 static_branch_enable(&sched_uclamp_used);
1377 uclamp_update_root_tg();
1380 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1381 static_branch_enable(&sched_uclamp_used);
1382 uclamp_sync_util_min_rt_default();
1386 * We update all RUNNABLE tasks only when task groups are in use.
1387 * Otherwise, keep it simple and do just a lazy update at each next
1388 * task enqueue time.
1394 sysctl_sched_uclamp_util_min = old_min;
1395 sysctl_sched_uclamp_util_max = old_max;
1396 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1398 mutex_unlock(&uclamp_mutex);
1403 static int uclamp_validate(struct task_struct *p,
1404 const struct sched_attr *attr)
1406 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1407 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1409 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1410 util_min = attr->sched_util_min;
1412 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1416 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1417 util_max = attr->sched_util_max;
1419 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1423 if (util_min != -1 && util_max != -1 && util_min > util_max)
1427 * We have valid uclamp attributes; make sure uclamp is enabled.
1429 * We need to do that here, because enabling static branches is a
1430 * blocking operation which obviously cannot be done while holding
1433 static_branch_enable(&sched_uclamp_used);
1438 static bool uclamp_reset(const struct sched_attr *attr,
1439 enum uclamp_id clamp_id,
1440 struct uclamp_se *uc_se)
1442 /* Reset on sched class change for a non user-defined clamp value. */
1443 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1444 !uc_se->user_defined)
1447 /* Reset on sched_util_{min,max} == -1. */
1448 if (clamp_id == UCLAMP_MIN &&
1449 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1450 attr->sched_util_min == -1) {
1454 if (clamp_id == UCLAMP_MAX &&
1455 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1456 attr->sched_util_max == -1) {
1463 static void __setscheduler_uclamp(struct task_struct *p,
1464 const struct sched_attr *attr)
1466 enum uclamp_id clamp_id;
1468 for_each_clamp_id(clamp_id) {
1469 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1472 if (!uclamp_reset(attr, clamp_id, uc_se))
1476 * RT by default have a 100% boost value that could be modified
1479 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1480 value = sysctl_sched_uclamp_util_min_rt_default;
1482 value = uclamp_none(clamp_id);
1484 uclamp_se_set(uc_se, value, false);
1488 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1491 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1492 attr->sched_util_min != -1) {
1493 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1494 attr->sched_util_min, true);
1497 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1498 attr->sched_util_max != -1) {
1499 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1500 attr->sched_util_max, true);
1504 static void uclamp_fork(struct task_struct *p)
1506 enum uclamp_id clamp_id;
1509 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1510 * as the task is still at its early fork stages.
1512 for_each_clamp_id(clamp_id)
1513 p->uclamp[clamp_id].active = false;
1515 if (likely(!p->sched_reset_on_fork))
1518 for_each_clamp_id(clamp_id) {
1519 uclamp_se_set(&p->uclamp_req[clamp_id],
1520 uclamp_none(clamp_id), false);
1524 static void uclamp_post_fork(struct task_struct *p)
1526 uclamp_update_util_min_rt_default(p);
1529 static void __init init_uclamp_rq(struct rq *rq)
1531 enum uclamp_id clamp_id;
1532 struct uclamp_rq *uc_rq = rq->uclamp;
1534 for_each_clamp_id(clamp_id) {
1535 uc_rq[clamp_id] = (struct uclamp_rq) {
1536 .value = uclamp_none(clamp_id)
1540 rq->uclamp_flags = 0;
1543 static void __init init_uclamp(void)
1545 struct uclamp_se uc_max = {};
1546 enum uclamp_id clamp_id;
1549 for_each_possible_cpu(cpu)
1550 init_uclamp_rq(cpu_rq(cpu));
1552 for_each_clamp_id(clamp_id) {
1553 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1554 uclamp_none(clamp_id), false);
1557 /* System defaults allow max clamp values for both indexes */
1558 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1559 for_each_clamp_id(clamp_id) {
1560 uclamp_default[clamp_id] = uc_max;
1561 #ifdef CONFIG_UCLAMP_TASK_GROUP
1562 root_task_group.uclamp_req[clamp_id] = uc_max;
1563 root_task_group.uclamp[clamp_id] = uc_max;
1568 #else /* CONFIG_UCLAMP_TASK */
1569 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1570 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1571 static inline int uclamp_validate(struct task_struct *p,
1572 const struct sched_attr *attr)
1576 static void __setscheduler_uclamp(struct task_struct *p,
1577 const struct sched_attr *attr) { }
1578 static inline void uclamp_fork(struct task_struct *p) { }
1579 static inline void uclamp_post_fork(struct task_struct *p) { }
1580 static inline void init_uclamp(void) { }
1581 #endif /* CONFIG_UCLAMP_TASK */
1583 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1585 if (!(flags & ENQUEUE_NOCLOCK))
1586 update_rq_clock(rq);
1588 if (!(flags & ENQUEUE_RESTORE)) {
1589 sched_info_queued(rq, p);
1590 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1593 uclamp_rq_inc(rq, p);
1594 p->sched_class->enqueue_task(rq, p, flags);
1597 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1599 if (!(flags & DEQUEUE_NOCLOCK))
1600 update_rq_clock(rq);
1602 if (!(flags & DEQUEUE_SAVE)) {
1603 sched_info_dequeued(rq, p);
1604 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1607 uclamp_rq_dec(rq, p);
1608 p->sched_class->dequeue_task(rq, p, flags);
1611 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1613 enqueue_task(rq, p, flags);
1615 p->on_rq = TASK_ON_RQ_QUEUED;
1618 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1620 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1622 dequeue_task(rq, p, flags);
1626 * __normal_prio - return the priority that is based on the static prio
1628 static inline int __normal_prio(struct task_struct *p)
1630 return p->static_prio;
1634 * Calculate the expected normal priority: i.e. priority
1635 * without taking RT-inheritance into account. Might be
1636 * boosted by interactivity modifiers. Changes upon fork,
1637 * setprio syscalls, and whenever the interactivity
1638 * estimator recalculates.
1640 static inline int normal_prio(struct task_struct *p)
1644 if (task_has_dl_policy(p))
1645 prio = MAX_DL_PRIO-1;
1646 else if (task_has_rt_policy(p))
1647 prio = MAX_RT_PRIO-1 - p->rt_priority;
1649 prio = __normal_prio(p);
1654 * Calculate the current priority, i.e. the priority
1655 * taken into account by the scheduler. This value might
1656 * be boosted by RT tasks, or might be boosted by
1657 * interactivity modifiers. Will be RT if the task got
1658 * RT-boosted. If not then it returns p->normal_prio.
1660 static int effective_prio(struct task_struct *p)
1662 p->normal_prio = normal_prio(p);
1664 * If we are RT tasks or we were boosted to RT priority,
1665 * keep the priority unchanged. Otherwise, update priority
1666 * to the normal priority:
1668 if (!rt_prio(p->prio))
1669 return p->normal_prio;
1674 * task_curr - is this task currently executing on a CPU?
1675 * @p: the task in question.
1677 * Return: 1 if the task is currently executing. 0 otherwise.
1679 inline int task_curr(const struct task_struct *p)
1681 return cpu_curr(task_cpu(p)) == p;
1685 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1686 * use the balance_callback list if you want balancing.
1688 * this means any call to check_class_changed() must be followed by a call to
1689 * balance_callback().
1691 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1692 const struct sched_class *prev_class,
1695 if (prev_class != p->sched_class) {
1696 if (prev_class->switched_from)
1697 prev_class->switched_from(rq, p);
1699 p->sched_class->switched_to(rq, p);
1700 } else if (oldprio != p->prio || dl_task(p))
1701 p->sched_class->prio_changed(rq, p, oldprio);
1704 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1706 if (p->sched_class == rq->curr->sched_class)
1707 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1708 else if (p->sched_class > rq->curr->sched_class)
1712 * A queue event has occurred, and we're going to schedule. In
1713 * this case, we can save a useless back to back clock update.
1715 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1716 rq_clock_skip_update(rq);
1722 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
1724 static int __set_cpus_allowed_ptr(struct task_struct *p,
1725 const struct cpumask *new_mask,
1728 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
1730 if (likely(!p->migration_disabled))
1733 if (p->cpus_ptr != &p->cpus_mask)
1737 * Violates locking rules! see comment in __do_set_cpus_allowed().
1739 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
1742 void migrate_disable(void)
1744 struct task_struct *p = current;
1746 if (p->migration_disabled) {
1747 p->migration_disabled++;
1752 this_rq()->nr_pinned++;
1753 p->migration_disabled = 1;
1756 EXPORT_SYMBOL_GPL(migrate_disable);
1758 void migrate_enable(void)
1760 struct task_struct *p = current;
1762 if (p->migration_disabled > 1) {
1763 p->migration_disabled--;
1768 * Ensure stop_task runs either before or after this, and that
1769 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
1772 if (p->cpus_ptr != &p->cpus_mask)
1773 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
1775 * Mustn't clear migration_disabled() until cpus_ptr points back at the
1776 * regular cpus_mask, otherwise things that race (eg.
1777 * select_fallback_rq) get confused.
1780 p->migration_disabled = 0;
1781 this_rq()->nr_pinned--;
1784 EXPORT_SYMBOL_GPL(migrate_enable);
1786 static inline bool rq_has_pinned_tasks(struct rq *rq)
1788 return rq->nr_pinned;
1792 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1793 * __set_cpus_allowed_ptr() and select_fallback_rq().
1795 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1797 /* When not in the task's cpumask, no point in looking further. */
1798 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1801 /* migrate_disabled() must be allowed to finish. */
1802 if (is_migration_disabled(p))
1803 return cpu_online(cpu);
1805 /* Non kernel threads are not allowed during either online or offline. */
1806 if (!(p->flags & PF_KTHREAD))
1807 return cpu_active(cpu);
1809 /* KTHREAD_IS_PER_CPU is always allowed. */
1810 if (kthread_is_per_cpu(p))
1811 return cpu_online(cpu);
1813 /* Regular kernel threads don't get to stay during offline. */
1814 if (cpu_rq(cpu)->balance_push)
1817 /* But are allowed during online. */
1818 return cpu_online(cpu);
1822 * This is how migration works:
1824 * 1) we invoke migration_cpu_stop() on the target CPU using
1826 * 2) stopper starts to run (implicitly forcing the migrated thread
1828 * 3) it checks whether the migrated task is still in the wrong runqueue.
1829 * 4) if it's in the wrong runqueue then the migration thread removes
1830 * it and puts it into the right queue.
1831 * 5) stopper completes and stop_one_cpu() returns and the migration
1836 * move_queued_task - move a queued task to new rq.
1838 * Returns (locked) new rq. Old rq's lock is released.
1840 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1841 struct task_struct *p, int new_cpu)
1843 lockdep_assert_held(&rq->lock);
1845 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
1846 set_task_cpu(p, new_cpu);
1849 rq = cpu_rq(new_cpu);
1852 BUG_ON(task_cpu(p) != new_cpu);
1853 activate_task(rq, p, 0);
1854 check_preempt_curr(rq, p, 0);
1859 struct migration_arg {
1860 struct task_struct *task;
1862 struct set_affinity_pending *pending;
1866 * @refs: number of wait_for_completion()
1867 * @stop_pending: is @stop_work in use
1869 struct set_affinity_pending {
1871 unsigned int stop_pending;
1872 struct completion done;
1873 struct cpu_stop_work stop_work;
1874 struct migration_arg arg;
1878 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1879 * this because either it can't run here any more (set_cpus_allowed()
1880 * away from this CPU, or CPU going down), or because we're
1881 * attempting to rebalance this task on exec (sched_exec).
1883 * So we race with normal scheduler movements, but that's OK, as long
1884 * as the task is no longer on this CPU.
1886 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1887 struct task_struct *p, int dest_cpu)
1889 /* Affinity changed (again). */
1890 if (!is_cpu_allowed(p, dest_cpu))
1893 update_rq_clock(rq);
1894 rq = move_queued_task(rq, rf, p, dest_cpu);
1900 * migration_cpu_stop - this will be executed by a highprio stopper thread
1901 * and performs thread migration by bumping thread off CPU then
1902 * 'pushing' onto another runqueue.
1904 static int migration_cpu_stop(void *data)
1906 struct migration_arg *arg = data;
1907 struct set_affinity_pending *pending = arg->pending;
1908 struct task_struct *p = arg->task;
1909 int dest_cpu = arg->dest_cpu;
1910 struct rq *rq = this_rq();
1911 bool complete = false;
1915 * The original target CPU might have gone down and we might
1916 * be on another CPU but it doesn't matter.
1918 local_irq_save(rf.flags);
1920 * We need to explicitly wake pending tasks before running
1921 * __migrate_task() such that we will not miss enforcing cpus_ptr
1922 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1924 flush_smp_call_function_from_idle();
1926 raw_spin_lock(&p->pi_lock);
1930 * If task_rq(p) != rq, it cannot be migrated here, because we're
1931 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1932 * we're holding p->pi_lock.
1934 if (task_rq(p) == rq) {
1935 if (is_migration_disabled(p))
1939 if (p->migration_pending == pending)
1940 p->migration_pending = NULL;
1945 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
1948 dest_cpu = cpumask_any_distribute(&p->cpus_mask);
1951 if (task_on_rq_queued(p))
1952 rq = __migrate_task(rq, &rf, p, dest_cpu);
1954 p->wake_cpu = dest_cpu;
1957 * XXX __migrate_task() can fail, at which point we might end
1958 * up running on a dodgy CPU, AFAICT this can only happen
1959 * during CPU hotplug, at which point we'll get pushed out
1960 * anyway, so it's probably not a big deal.
1963 } else if (pending) {
1965 * This happens when we get migrated between migrate_enable()'s
1966 * preempt_enable() and scheduling the stopper task. At that
1967 * point we're a regular task again and not current anymore.
1969 * A !PREEMPT kernel has a giant hole here, which makes it far
1974 * The task moved before the stopper got to run. We're holding
1975 * ->pi_lock, so the allowed mask is stable - if it got
1976 * somewhere allowed, we're done.
1978 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
1979 if (p->migration_pending == pending)
1980 p->migration_pending = NULL;
1986 * When migrate_enable() hits a rq mis-match we can't reliably
1987 * determine is_migration_disabled() and so have to chase after
1990 WARN_ON_ONCE(!pending->stop_pending);
1991 task_rq_unlock(rq, p, &rf);
1992 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
1993 &pending->arg, &pending->stop_work);
1998 pending->stop_pending = false;
1999 task_rq_unlock(rq, p, &rf);
2002 complete_all(&pending->done);
2007 int push_cpu_stop(void *arg)
2009 struct rq *lowest_rq = NULL, *rq = this_rq();
2010 struct task_struct *p = arg;
2012 raw_spin_lock_irq(&p->pi_lock);
2013 raw_spin_lock(&rq->lock);
2015 if (task_rq(p) != rq)
2018 if (is_migration_disabled(p)) {
2019 p->migration_flags |= MDF_PUSH;
2023 p->migration_flags &= ~MDF_PUSH;
2025 if (p->sched_class->find_lock_rq)
2026 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2031 // XXX validate p is still the highest prio task
2032 if (task_rq(p) == rq) {
2033 deactivate_task(rq, p, 0);
2034 set_task_cpu(p, lowest_rq->cpu);
2035 activate_task(lowest_rq, p, 0);
2036 resched_curr(lowest_rq);
2039 double_unlock_balance(rq, lowest_rq);
2042 rq->push_busy = false;
2043 raw_spin_unlock(&rq->lock);
2044 raw_spin_unlock_irq(&p->pi_lock);
2051 * sched_class::set_cpus_allowed must do the below, but is not required to
2052 * actually call this function.
2054 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2056 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2057 p->cpus_ptr = new_mask;
2061 cpumask_copy(&p->cpus_mask, new_mask);
2062 p->nr_cpus_allowed = cpumask_weight(new_mask);
2066 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2068 struct rq *rq = task_rq(p);
2069 bool queued, running;
2072 * This here violates the locking rules for affinity, since we're only
2073 * supposed to change these variables while holding both rq->lock and
2076 * HOWEVER, it magically works, because ttwu() is the only code that
2077 * accesses these variables under p->pi_lock and only does so after
2078 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2079 * before finish_task().
2081 * XXX do further audits, this smells like something putrid.
2083 if (flags & SCA_MIGRATE_DISABLE)
2084 SCHED_WARN_ON(!p->on_cpu);
2086 lockdep_assert_held(&p->pi_lock);
2088 queued = task_on_rq_queued(p);
2089 running = task_current(rq, p);
2093 * Because __kthread_bind() calls this on blocked tasks without
2096 lockdep_assert_held(&rq->lock);
2097 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2100 put_prev_task(rq, p);
2102 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2105 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2107 set_next_task(rq, p);
2110 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2112 __do_set_cpus_allowed(p, new_mask, 0);
2116 * This function is wildly self concurrent; here be dragons.
2119 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2120 * designated task is enqueued on an allowed CPU. If that task is currently
2121 * running, we have to kick it out using the CPU stopper.
2123 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2126 * Initial conditions: P0->cpus_mask = [0, 1]
2130 * migrate_disable();
2132 * set_cpus_allowed_ptr(P0, [1]);
2134 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2135 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2136 * This means we need the following scheme:
2140 * migrate_disable();
2142 * set_cpus_allowed_ptr(P0, [1]);
2146 * __set_cpus_allowed_ptr();
2147 * <wakes local stopper>
2148 * `--> <woken on migration completion>
2150 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2151 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2152 * task p are serialized by p->pi_lock, which we can leverage: the one that
2153 * should come into effect at the end of the Migrate-Disable region is the last
2154 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2155 * but we still need to properly signal those waiting tasks at the appropriate
2158 * This is implemented using struct set_affinity_pending. The first
2159 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2160 * setup an instance of that struct and install it on the targeted task_struct.
2161 * Any and all further callers will reuse that instance. Those then wait for
2162 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2163 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2166 * (1) In the cases covered above. There is one more where the completion is
2167 * signaled within affine_move_task() itself: when a subsequent affinity request
2168 * cancels the need for an active migration. Consider:
2170 * Initial conditions: P0->cpus_mask = [0, 1]
2174 * migrate_disable();
2176 * set_cpus_allowed_ptr(P0, [1]);
2178 * set_cpus_allowed_ptr(P0, [0, 1]);
2179 * <signal completion>
2182 * Note that the above is safe vs a concurrent migrate_enable(), as any
2183 * pending affinity completion is preceded by an uninstallation of
2184 * p->migration_pending done with p->pi_lock held.
2186 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2187 int dest_cpu, unsigned int flags)
2189 struct set_affinity_pending my_pending = { }, *pending = NULL;
2190 bool stop_pending, complete = false;
2192 /* Can the task run on the task's current CPU? If so, we're done */
2193 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2194 struct task_struct *push_task = NULL;
2196 if ((flags & SCA_MIGRATE_ENABLE) &&
2197 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2198 rq->push_busy = true;
2199 push_task = get_task_struct(p);
2203 * If there are pending waiters, but no pending stop_work,
2204 * then complete now.
2206 pending = p->migration_pending;
2207 if (pending && !pending->stop_pending) {
2208 p->migration_pending = NULL;
2212 task_rq_unlock(rq, p, rf);
2215 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2220 complete_all(&pending->done);
2225 if (!(flags & SCA_MIGRATE_ENABLE)) {
2226 /* serialized by p->pi_lock */
2227 if (!p->migration_pending) {
2228 /* Install the request */
2229 refcount_set(&my_pending.refs, 1);
2230 init_completion(&my_pending.done);
2231 my_pending.arg = (struct migration_arg) {
2233 .dest_cpu = -1, /* any */
2234 .pending = &my_pending,
2237 p->migration_pending = &my_pending;
2239 pending = p->migration_pending;
2240 refcount_inc(&pending->refs);
2243 pending = p->migration_pending;
2245 * - !MIGRATE_ENABLE:
2246 * we'll have installed a pending if there wasn't one already.
2249 * we're here because the current CPU isn't matching anymore,
2250 * the only way that can happen is because of a concurrent
2251 * set_cpus_allowed_ptr() call, which should then still be
2252 * pending completion.
2254 * Either way, we really should have a @pending here.
2256 if (WARN_ON_ONCE(!pending)) {
2257 task_rq_unlock(rq, p, rf);
2261 if (task_running(rq, p) || p->state == TASK_WAKING) {
2263 * MIGRATE_ENABLE gets here because 'p == current', but for
2264 * anything else we cannot do is_migration_disabled(), punt
2265 * and have the stopper function handle it all race-free.
2267 stop_pending = pending->stop_pending;
2269 pending->stop_pending = true;
2271 if (flags & SCA_MIGRATE_ENABLE)
2272 p->migration_flags &= ~MDF_PUSH;
2274 task_rq_unlock(rq, p, rf);
2276 if (!stop_pending) {
2277 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2278 &pending->arg, &pending->stop_work);
2281 if (flags & SCA_MIGRATE_ENABLE)
2285 if (!is_migration_disabled(p)) {
2286 if (task_on_rq_queued(p))
2287 rq = move_queued_task(rq, rf, p, dest_cpu);
2289 if (!pending->stop_pending) {
2290 p->migration_pending = NULL;
2294 task_rq_unlock(rq, p, rf);
2297 complete_all(&pending->done);
2300 wait_for_completion(&pending->done);
2302 if (refcount_dec_and_test(&pending->refs))
2303 wake_up_var(&pending->refs); /* No UaF, just an address */
2306 * Block the original owner of &pending until all subsequent callers
2307 * have seen the completion and decremented the refcount
2309 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2312 WARN_ON_ONCE(my_pending.stop_pending);
2318 * Change a given task's CPU affinity. Migrate the thread to a
2319 * proper CPU and schedule it away if the CPU it's executing on
2320 * is removed from the allowed bitmask.
2322 * NOTE: the caller must have a valid reference to the task, the
2323 * task must not exit() & deallocate itself prematurely. The
2324 * call is not atomic; no spinlocks may be held.
2326 static int __set_cpus_allowed_ptr(struct task_struct *p,
2327 const struct cpumask *new_mask,
2330 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2331 unsigned int dest_cpu;
2336 rq = task_rq_lock(p, &rf);
2337 update_rq_clock(rq);
2339 if (p->flags & PF_KTHREAD || is_migration_disabled(p)) {
2341 * Kernel threads are allowed on online && !active CPUs,
2342 * however, during cpu-hot-unplug, even these might get pushed
2343 * away if not KTHREAD_IS_PER_CPU.
2345 * Specifically, migration_disabled() tasks must not fail the
2346 * cpumask_any_and_distribute() pick below, esp. so on
2347 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2348 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2350 cpu_valid_mask = cpu_online_mask;
2354 * Must re-check here, to close a race against __kthread_bind(),
2355 * sched_setaffinity() is not guaranteed to observe the flag.
2357 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2362 if (!(flags & SCA_MIGRATE_ENABLE)) {
2363 if (cpumask_equal(&p->cpus_mask, new_mask))
2366 if (WARN_ON_ONCE(p == current &&
2367 is_migration_disabled(p) &&
2368 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2375 * Picking a ~random cpu helps in cases where we are changing affinity
2376 * for groups of tasks (ie. cpuset), so that load balancing is not
2377 * immediately required to distribute the tasks within their new mask.
2379 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2380 if (dest_cpu >= nr_cpu_ids) {
2385 __do_set_cpus_allowed(p, new_mask, flags);
2387 return affine_move_task(rq, p, &rf, dest_cpu, flags);
2390 task_rq_unlock(rq, p, &rf);
2395 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2397 return __set_cpus_allowed_ptr(p, new_mask, 0);
2399 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2401 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2403 #ifdef CONFIG_SCHED_DEBUG
2405 * We should never call set_task_cpu() on a blocked task,
2406 * ttwu() will sort out the placement.
2408 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2412 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
2413 * because schedstat_wait_{start,end} rebase migrating task's wait_start
2414 * time relying on p->on_rq.
2416 WARN_ON_ONCE(p->state == TASK_RUNNING &&
2417 p->sched_class == &fair_sched_class &&
2418 (p->on_rq && !task_on_rq_migrating(p)));
2420 #ifdef CONFIG_LOCKDEP
2422 * The caller should hold either p->pi_lock or rq->lock, when changing
2423 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2425 * sched_move_task() holds both and thus holding either pins the cgroup,
2428 * Furthermore, all task_rq users should acquire both locks, see
2431 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2432 lockdep_is_held(&task_rq(p)->lock)));
2435 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
2437 WARN_ON_ONCE(!cpu_online(new_cpu));
2439 WARN_ON_ONCE(is_migration_disabled(p));
2442 trace_sched_migrate_task(p, new_cpu);
2444 if (task_cpu(p) != new_cpu) {
2445 if (p->sched_class->migrate_task_rq)
2446 p->sched_class->migrate_task_rq(p, new_cpu);
2447 p->se.nr_migrations++;
2449 perf_event_task_migrate(p);
2452 __set_task_cpu(p, new_cpu);
2455 #ifdef CONFIG_NUMA_BALANCING
2456 static void __migrate_swap_task(struct task_struct *p, int cpu)
2458 if (task_on_rq_queued(p)) {
2459 struct rq *src_rq, *dst_rq;
2460 struct rq_flags srf, drf;
2462 src_rq = task_rq(p);
2463 dst_rq = cpu_rq(cpu);
2465 rq_pin_lock(src_rq, &srf);
2466 rq_pin_lock(dst_rq, &drf);
2468 deactivate_task(src_rq, p, 0);
2469 set_task_cpu(p, cpu);
2470 activate_task(dst_rq, p, 0);
2471 check_preempt_curr(dst_rq, p, 0);
2473 rq_unpin_lock(dst_rq, &drf);
2474 rq_unpin_lock(src_rq, &srf);
2478 * Task isn't running anymore; make it appear like we migrated
2479 * it before it went to sleep. This means on wakeup we make the
2480 * previous CPU our target instead of where it really is.
2486 struct migration_swap_arg {
2487 struct task_struct *src_task, *dst_task;
2488 int src_cpu, dst_cpu;
2491 static int migrate_swap_stop(void *data)
2493 struct migration_swap_arg *arg = data;
2494 struct rq *src_rq, *dst_rq;
2497 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
2500 src_rq = cpu_rq(arg->src_cpu);
2501 dst_rq = cpu_rq(arg->dst_cpu);
2503 double_raw_lock(&arg->src_task->pi_lock,
2504 &arg->dst_task->pi_lock);
2505 double_rq_lock(src_rq, dst_rq);
2507 if (task_cpu(arg->dst_task) != arg->dst_cpu)
2510 if (task_cpu(arg->src_task) != arg->src_cpu)
2513 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
2516 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
2519 __migrate_swap_task(arg->src_task, arg->dst_cpu);
2520 __migrate_swap_task(arg->dst_task, arg->src_cpu);
2525 double_rq_unlock(src_rq, dst_rq);
2526 raw_spin_unlock(&arg->dst_task->pi_lock);
2527 raw_spin_unlock(&arg->src_task->pi_lock);
2533 * Cross migrate two tasks
2535 int migrate_swap(struct task_struct *cur, struct task_struct *p,
2536 int target_cpu, int curr_cpu)
2538 struct migration_swap_arg arg;
2541 arg = (struct migration_swap_arg){
2543 .src_cpu = curr_cpu,
2545 .dst_cpu = target_cpu,
2548 if (arg.src_cpu == arg.dst_cpu)
2552 * These three tests are all lockless; this is OK since all of them
2553 * will be re-checked with proper locks held further down the line.
2555 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
2558 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
2561 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
2564 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
2565 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
2570 #endif /* CONFIG_NUMA_BALANCING */
2573 * wait_task_inactive - wait for a thread to unschedule.
2575 * If @match_state is nonzero, it's the @p->state value just checked and
2576 * not expected to change. If it changes, i.e. @p might have woken up,
2577 * then return zero. When we succeed in waiting for @p to be off its CPU,
2578 * we return a positive number (its total switch count). If a second call
2579 * a short while later returns the same number, the caller can be sure that
2580 * @p has remained unscheduled the whole time.
2582 * The caller must ensure that the task *will* unschedule sometime soon,
2583 * else this function might spin for a *long* time. This function can't
2584 * be called with interrupts off, or it may introduce deadlock with
2585 * smp_call_function() if an IPI is sent by the same process we are
2586 * waiting to become inactive.
2588 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2590 int running, queued;
2597 * We do the initial early heuristics without holding
2598 * any task-queue locks at all. We'll only try to get
2599 * the runqueue lock when things look like they will
2605 * If the task is actively running on another CPU
2606 * still, just relax and busy-wait without holding
2609 * NOTE! Since we don't hold any locks, it's not
2610 * even sure that "rq" stays as the right runqueue!
2611 * But we don't care, since "task_running()" will
2612 * return false if the runqueue has changed and p
2613 * is actually now running somewhere else!
2615 while (task_running(rq, p)) {
2616 if (match_state && unlikely(p->state != match_state))
2622 * Ok, time to look more closely! We need the rq
2623 * lock now, to be *sure*. If we're wrong, we'll
2624 * just go back and repeat.
2626 rq = task_rq_lock(p, &rf);
2627 trace_sched_wait_task(p);
2628 running = task_running(rq, p);
2629 queued = task_on_rq_queued(p);
2631 if (!match_state || p->state == match_state)
2632 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2633 task_rq_unlock(rq, p, &rf);
2636 * If it changed from the expected state, bail out now.
2638 if (unlikely(!ncsw))
2642 * Was it really running after all now that we
2643 * checked with the proper locks actually held?
2645 * Oops. Go back and try again..
2647 if (unlikely(running)) {
2653 * It's not enough that it's not actively running,
2654 * it must be off the runqueue _entirely_, and not
2657 * So if it was still runnable (but just not actively
2658 * running right now), it's preempted, and we should
2659 * yield - it could be a while.
2661 if (unlikely(queued)) {
2662 ktime_t to = NSEC_PER_SEC / HZ;
2664 set_current_state(TASK_UNINTERRUPTIBLE);
2665 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2670 * Ahh, all good. It wasn't running, and it wasn't
2671 * runnable, which means that it will never become
2672 * running in the future either. We're all done!
2681 * kick_process - kick a running thread to enter/exit the kernel
2682 * @p: the to-be-kicked thread
2684 * Cause a process which is running on another CPU to enter
2685 * kernel-mode, without any delay. (to get signals handled.)
2687 * NOTE: this function doesn't have to take the runqueue lock,
2688 * because all it wants to ensure is that the remote task enters
2689 * the kernel. If the IPI races and the task has been migrated
2690 * to another CPU then no harm is done and the purpose has been
2693 void kick_process(struct task_struct *p)
2699 if ((cpu != smp_processor_id()) && task_curr(p))
2700 smp_send_reschedule(cpu);
2703 EXPORT_SYMBOL_GPL(kick_process);
2706 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2708 * A few notes on cpu_active vs cpu_online:
2710 * - cpu_active must be a subset of cpu_online
2712 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2713 * see __set_cpus_allowed_ptr(). At this point the newly online
2714 * CPU isn't yet part of the sched domains, and balancing will not
2717 * - on CPU-down we clear cpu_active() to mask the sched domains and
2718 * avoid the load balancer to place new tasks on the to be removed
2719 * CPU. Existing tasks will remain running there and will be taken
2722 * This means that fallback selection must not select !active CPUs.
2723 * And can assume that any active CPU must be online. Conversely
2724 * select_task_rq() below may allow selection of !active CPUs in order
2725 * to satisfy the above rules.
2727 static int select_fallback_rq(int cpu, struct task_struct *p)
2729 int nid = cpu_to_node(cpu);
2730 const struct cpumask *nodemask = NULL;
2731 enum { cpuset, possible, fail } state = cpuset;
2735 * If the node that the CPU is on has been offlined, cpu_to_node()
2736 * will return -1. There is no CPU on the node, and we should
2737 * select the CPU on the other node.
2740 nodemask = cpumask_of_node(nid);
2742 /* Look for allowed, online CPU in same node. */
2743 for_each_cpu(dest_cpu, nodemask) {
2744 if (!cpu_active(dest_cpu))
2746 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2752 /* Any allowed, online CPU? */
2753 for_each_cpu(dest_cpu, p->cpus_ptr) {
2754 if (!is_cpu_allowed(p, dest_cpu))
2760 /* No more Mr. Nice Guy. */
2763 if (IS_ENABLED(CONFIG_CPUSETS)) {
2764 cpuset_cpus_allowed_fallback(p);
2771 * XXX When called from select_task_rq() we only
2772 * hold p->pi_lock and again violate locking order.
2774 * More yuck to audit.
2776 do_set_cpus_allowed(p, cpu_possible_mask);
2787 if (state != cpuset) {
2789 * Don't tell them about moving exiting tasks or
2790 * kernel threads (both mm NULL), since they never
2793 if (p->mm && printk_ratelimit()) {
2794 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2795 task_pid_nr(p), p->comm, cpu);
2803 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2806 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
2808 lockdep_assert_held(&p->pi_lock);
2810 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
2811 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
2813 cpu = cpumask_any(p->cpus_ptr);
2816 * In order not to call set_task_cpu() on a blocking task we need
2817 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2820 * Since this is common to all placement strategies, this lives here.
2822 * [ this allows ->select_task() to simply return task_cpu(p) and
2823 * not worry about this generic constraint ]
2825 if (unlikely(!is_cpu_allowed(p, cpu)))
2826 cpu = select_fallback_rq(task_cpu(p), p);
2831 void sched_set_stop_task(int cpu, struct task_struct *stop)
2833 static struct lock_class_key stop_pi_lock;
2834 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2835 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2839 * Make it appear like a SCHED_FIFO task, its something
2840 * userspace knows about and won't get confused about.
2842 * Also, it will make PI more or less work without too
2843 * much confusion -- but then, stop work should not
2844 * rely on PI working anyway.
2846 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2848 stop->sched_class = &stop_sched_class;
2851 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
2852 * adjust the effective priority of a task. As a result,
2853 * rt_mutex_setprio() can trigger (RT) balancing operations,
2854 * which can then trigger wakeups of the stop thread to push
2855 * around the current task.
2857 * The stop task itself will never be part of the PI-chain, it
2858 * never blocks, therefore that ->pi_lock recursion is safe.
2859 * Tell lockdep about this by placing the stop->pi_lock in its
2862 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
2865 cpu_rq(cpu)->stop = stop;
2869 * Reset it back to a normal scheduling class so that
2870 * it can die in pieces.
2872 old_stop->sched_class = &rt_sched_class;
2876 #else /* CONFIG_SMP */
2878 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2879 const struct cpumask *new_mask,
2882 return set_cpus_allowed_ptr(p, new_mask);
2885 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
2887 static inline bool rq_has_pinned_tasks(struct rq *rq)
2892 #endif /* !CONFIG_SMP */
2895 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2899 if (!schedstat_enabled())
2905 if (cpu == rq->cpu) {
2906 __schedstat_inc(rq->ttwu_local);
2907 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2909 struct sched_domain *sd;
2911 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2913 for_each_domain(rq->cpu, sd) {
2914 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2915 __schedstat_inc(sd->ttwu_wake_remote);
2922 if (wake_flags & WF_MIGRATED)
2923 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2924 #endif /* CONFIG_SMP */
2926 __schedstat_inc(rq->ttwu_count);
2927 __schedstat_inc(p->se.statistics.nr_wakeups);
2929 if (wake_flags & WF_SYNC)
2930 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2934 * Mark the task runnable and perform wakeup-preemption.
2936 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2937 struct rq_flags *rf)
2939 check_preempt_curr(rq, p, wake_flags);
2940 p->state = TASK_RUNNING;
2941 trace_sched_wakeup(p);
2944 if (p->sched_class->task_woken) {
2946 * Our task @p is fully woken up and running; so it's safe to
2947 * drop the rq->lock, hereafter rq is only used for statistics.
2949 rq_unpin_lock(rq, rf);
2950 p->sched_class->task_woken(rq, p);
2951 rq_repin_lock(rq, rf);
2954 if (rq->idle_stamp) {
2955 u64 delta = rq_clock(rq) - rq->idle_stamp;
2956 u64 max = 2*rq->max_idle_balance_cost;
2958 update_avg(&rq->avg_idle, delta);
2960 if (rq->avg_idle > max)
2969 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2970 struct rq_flags *rf)
2972 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2974 lockdep_assert_held(&rq->lock);
2976 if (p->sched_contributes_to_load)
2977 rq->nr_uninterruptible--;
2980 if (wake_flags & WF_MIGRATED)
2981 en_flags |= ENQUEUE_MIGRATED;
2985 delayacct_blkio_end(p);
2986 atomic_dec(&task_rq(p)->nr_iowait);
2989 activate_task(rq, p, en_flags);
2990 ttwu_do_wakeup(rq, p, wake_flags, rf);
2994 * Consider @p being inside a wait loop:
2997 * set_current_state(TASK_UNINTERRUPTIBLE);
3004 * __set_current_state(TASK_RUNNING);
3006 * between set_current_state() and schedule(). In this case @p is still
3007 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3010 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3011 * then schedule() must still happen and p->state can be changed to
3012 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3013 * need to do a full wakeup with enqueue.
3015 * Returns: %true when the wakeup is done,
3018 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3024 rq = __task_rq_lock(p, &rf);
3025 if (task_on_rq_queued(p)) {
3026 /* check_preempt_curr() may use rq clock */
3027 update_rq_clock(rq);
3028 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3031 __task_rq_unlock(rq, &rf);
3037 void sched_ttwu_pending(void *arg)
3039 struct llist_node *llist = arg;
3040 struct rq *rq = this_rq();
3041 struct task_struct *p, *t;
3048 * rq::ttwu_pending racy indication of out-standing wakeups.
3049 * Races such that false-negatives are possible, since they
3050 * are shorter lived that false-positives would be.
3052 WRITE_ONCE(rq->ttwu_pending, 0);
3054 rq_lock_irqsave(rq, &rf);
3055 update_rq_clock(rq);
3057 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3058 if (WARN_ON_ONCE(p->on_cpu))
3059 smp_cond_load_acquire(&p->on_cpu, !VAL);
3061 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3062 set_task_cpu(p, cpu_of(rq));
3064 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3067 rq_unlock_irqrestore(rq, &rf);
3070 void send_call_function_single_ipi(int cpu)
3072 struct rq *rq = cpu_rq(cpu);
3074 if (!set_nr_if_polling(rq->idle))
3075 arch_send_call_function_single_ipi(cpu);
3077 trace_sched_wake_idle_without_ipi(cpu);
3081 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3082 * necessary. The wakee CPU on receipt of the IPI will queue the task
3083 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3084 * of the wakeup instead of the waker.
3086 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3088 struct rq *rq = cpu_rq(cpu);
3090 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3092 WRITE_ONCE(rq->ttwu_pending, 1);
3093 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3096 void wake_up_if_idle(int cpu)
3098 struct rq *rq = cpu_rq(cpu);
3103 if (!is_idle_task(rcu_dereference(rq->curr)))
3106 if (set_nr_if_polling(rq->idle)) {
3107 trace_sched_wake_idle_without_ipi(cpu);
3109 rq_lock_irqsave(rq, &rf);
3110 if (is_idle_task(rq->curr))
3111 smp_send_reschedule(cpu);
3112 /* Else CPU is not idle, do nothing here: */
3113 rq_unlock_irqrestore(rq, &rf);
3120 bool cpus_share_cache(int this_cpu, int that_cpu)
3122 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3125 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3128 * Do not complicate things with the async wake_list while the CPU is
3131 if (!cpu_active(cpu))
3135 * If the CPU does not share cache, then queue the task on the
3136 * remote rqs wakelist to avoid accessing remote data.
3138 if (!cpus_share_cache(smp_processor_id(), cpu))
3142 * If the task is descheduling and the only running task on the
3143 * CPU then use the wakelist to offload the task activation to
3144 * the soon-to-be-idle CPU as the current CPU is likely busy.
3145 * nr_running is checked to avoid unnecessary task stacking.
3147 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3153 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3155 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3156 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3159 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3160 __ttwu_queue_wakelist(p, cpu, wake_flags);
3167 #else /* !CONFIG_SMP */
3169 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3174 #endif /* CONFIG_SMP */
3176 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3178 struct rq *rq = cpu_rq(cpu);
3181 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3185 update_rq_clock(rq);
3186 ttwu_do_activate(rq, p, wake_flags, &rf);
3191 * Notes on Program-Order guarantees on SMP systems.
3195 * The basic program-order guarantee on SMP systems is that when a task [t]
3196 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3197 * execution on its new CPU [c1].
3199 * For migration (of runnable tasks) this is provided by the following means:
3201 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3202 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3203 * rq(c1)->lock (if not at the same time, then in that order).
3204 * C) LOCK of the rq(c1)->lock scheduling in task
3206 * Release/acquire chaining guarantees that B happens after A and C after B.
3207 * Note: the CPU doing B need not be c0 or c1
3216 * UNLOCK rq(0)->lock
3218 * LOCK rq(0)->lock // orders against CPU0
3220 * UNLOCK rq(0)->lock
3224 * UNLOCK rq(1)->lock
3226 * LOCK rq(1)->lock // orders against CPU2
3229 * UNLOCK rq(1)->lock
3232 * BLOCKING -- aka. SLEEP + WAKEUP
3234 * For blocking we (obviously) need to provide the same guarantee as for
3235 * migration. However the means are completely different as there is no lock
3236 * chain to provide order. Instead we do:
3238 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3239 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3243 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3245 * LOCK rq(0)->lock LOCK X->pi_lock
3248 * smp_store_release(X->on_cpu, 0);
3250 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3256 * X->state = RUNNING
3257 * UNLOCK rq(2)->lock
3259 * LOCK rq(2)->lock // orders against CPU1
3262 * UNLOCK rq(2)->lock
3265 * UNLOCK rq(0)->lock
3268 * However, for wakeups there is a second guarantee we must provide, namely we
3269 * must ensure that CONDITION=1 done by the caller can not be reordered with
3270 * accesses to the task state; see try_to_wake_up() and set_current_state().
3274 * try_to_wake_up - wake up a thread
3275 * @p: the thread to be awakened
3276 * @state: the mask of task states that can be woken
3277 * @wake_flags: wake modifier flags (WF_*)
3279 * Conceptually does:
3281 * If (@state & @p->state) @p->state = TASK_RUNNING.
3283 * If the task was not queued/runnable, also place it back on a runqueue.
3285 * This function is atomic against schedule() which would dequeue the task.
3287 * It issues a full memory barrier before accessing @p->state, see the comment
3288 * with set_current_state().
3290 * Uses p->pi_lock to serialize against concurrent wake-ups.
3292 * Relies on p->pi_lock stabilizing:
3295 * - p->sched_task_group
3296 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3298 * Tries really hard to only take one task_rq(p)->lock for performance.
3299 * Takes rq->lock in:
3300 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3301 * - ttwu_queue() -- new rq, for enqueue of the task;
3302 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3304 * As a consequence we race really badly with just about everything. See the
3305 * many memory barriers and their comments for details.
3307 * Return: %true if @p->state changes (an actual wakeup was done),
3311 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3313 unsigned long flags;
3314 int cpu, success = 0;
3319 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3320 * == smp_processor_id()'. Together this means we can special
3321 * case the whole 'p->on_rq && ttwu_runnable()' case below
3322 * without taking any locks.
3325 * - we rely on Program-Order guarantees for all the ordering,
3326 * - we're serialized against set_special_state() by virtue of
3327 * it disabling IRQs (this allows not taking ->pi_lock).
3329 if (!(p->state & state))
3333 trace_sched_waking(p);
3334 p->state = TASK_RUNNING;
3335 trace_sched_wakeup(p);
3340 * If we are going to wake up a thread waiting for CONDITION we
3341 * need to ensure that CONDITION=1 done by the caller can not be
3342 * reordered with p->state check below. This pairs with smp_store_mb()
3343 * in set_current_state() that the waiting thread does.
3345 raw_spin_lock_irqsave(&p->pi_lock, flags);
3346 smp_mb__after_spinlock();
3347 if (!(p->state & state))
3350 trace_sched_waking(p);
3352 /* We're going to change ->state: */
3356 * Ensure we load p->on_rq _after_ p->state, otherwise it would
3357 * be possible to, falsely, observe p->on_rq == 0 and get stuck
3358 * in smp_cond_load_acquire() below.
3360 * sched_ttwu_pending() try_to_wake_up()
3361 * STORE p->on_rq = 1 LOAD p->state
3364 * __schedule() (switch to task 'p')
3365 * LOCK rq->lock smp_rmb();
3366 * smp_mb__after_spinlock();
3370 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3372 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3373 * __schedule(). See the comment for smp_mb__after_spinlock().
3375 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3378 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
3383 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3384 * possible to, falsely, observe p->on_cpu == 0.
3386 * One must be running (->on_cpu == 1) in order to remove oneself
3387 * from the runqueue.
3389 * __schedule() (switch to task 'p') try_to_wake_up()
3390 * STORE p->on_cpu = 1 LOAD p->on_rq
3393 * __schedule() (put 'p' to sleep)
3394 * LOCK rq->lock smp_rmb();
3395 * smp_mb__after_spinlock();
3396 * STORE p->on_rq = 0 LOAD p->on_cpu
3398 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3399 * __schedule(). See the comment for smp_mb__after_spinlock().
3401 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3402 * schedule()'s deactivate_task() has 'happened' and p will no longer
3403 * care about it's own p->state. See the comment in __schedule().
3405 smp_acquire__after_ctrl_dep();
3408 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3409 * == 0), which means we need to do an enqueue, change p->state to
3410 * TASK_WAKING such that we can unlock p->pi_lock before doing the
3411 * enqueue, such as ttwu_queue_wakelist().
3413 p->state = TASK_WAKING;
3416 * If the owning (remote) CPU is still in the middle of schedule() with
3417 * this task as prev, considering queueing p on the remote CPUs wake_list
3418 * which potentially sends an IPI instead of spinning on p->on_cpu to
3419 * let the waker make forward progress. This is safe because IRQs are
3420 * disabled and the IPI will deliver after on_cpu is cleared.
3422 * Ensure we load task_cpu(p) after p->on_cpu:
3424 * set_task_cpu(p, cpu);
3425 * STORE p->cpu = @cpu
3426 * __schedule() (switch to task 'p')
3428 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
3429 * STORE p->on_cpu = 1 LOAD p->cpu
3431 * to ensure we observe the correct CPU on which the task is currently
3434 if (smp_load_acquire(&p->on_cpu) &&
3435 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
3439 * If the owning (remote) CPU is still in the middle of schedule() with
3440 * this task as prev, wait until it's done referencing the task.
3442 * Pairs with the smp_store_release() in finish_task().
3444 * This ensures that tasks getting woken will be fully ordered against
3445 * their previous state and preserve Program Order.
3447 smp_cond_load_acquire(&p->on_cpu, !VAL);
3449 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
3450 if (task_cpu(p) != cpu) {
3452 delayacct_blkio_end(p);
3453 atomic_dec(&task_rq(p)->nr_iowait);
3456 wake_flags |= WF_MIGRATED;
3457 psi_ttwu_dequeue(p);
3458 set_task_cpu(p, cpu);
3462 #endif /* CONFIG_SMP */
3464 ttwu_queue(p, cpu, wake_flags);
3466 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3469 ttwu_stat(p, task_cpu(p), wake_flags);
3476 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
3477 * @p: Process for which the function is to be invoked, can be @current.
3478 * @func: Function to invoke.
3479 * @arg: Argument to function.
3481 * If the specified task can be quickly locked into a definite state
3482 * (either sleeping or on a given runqueue), arrange to keep it in that
3483 * state while invoking @func(@arg). This function can use ->on_rq and
3484 * task_curr() to work out what the state is, if required. Given that
3485 * @func can be invoked with a runqueue lock held, it had better be quite
3489 * @false if the task slipped out from under the locks.
3490 * @true if the task was locked onto a runqueue or is sleeping.
3491 * However, @func can override this by returning @false.
3493 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
3499 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3501 rq = __task_rq_lock(p, &rf);
3502 if (task_rq(p) == rq)
3511 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3516 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
3521 * wake_up_process - Wake up a specific process
3522 * @p: The process to be woken up.
3524 * Attempt to wake up the nominated process and move it to the set of runnable
3527 * Return: 1 if the process was woken up, 0 if it was already running.
3529 * This function executes a full memory barrier before accessing the task state.
3531 int wake_up_process(struct task_struct *p)
3533 return try_to_wake_up(p, TASK_NORMAL, 0);
3535 EXPORT_SYMBOL(wake_up_process);
3537 int wake_up_state(struct task_struct *p, unsigned int state)
3539 return try_to_wake_up(p, state, 0);
3543 * Perform scheduler related setup for a newly forked process p.
3544 * p is forked by current.
3546 * __sched_fork() is basic setup used by init_idle() too:
3548 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
3553 p->se.exec_start = 0;
3554 p->se.sum_exec_runtime = 0;
3555 p->se.prev_sum_exec_runtime = 0;
3556 p->se.nr_migrations = 0;
3558 INIT_LIST_HEAD(&p->se.group_node);
3560 #ifdef CONFIG_FAIR_GROUP_SCHED
3561 p->se.cfs_rq = NULL;
3564 #ifdef CONFIG_SCHEDSTATS
3565 /* Even if schedstat is disabled, there should not be garbage */
3566 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
3569 RB_CLEAR_NODE(&p->dl.rb_node);
3570 init_dl_task_timer(&p->dl);
3571 init_dl_inactive_task_timer(&p->dl);
3572 __dl_clear_params(p);
3574 INIT_LIST_HEAD(&p->rt.run_list);
3576 p->rt.time_slice = sched_rr_timeslice;
3580 #ifdef CONFIG_PREEMPT_NOTIFIERS
3581 INIT_HLIST_HEAD(&p->preempt_notifiers);
3584 #ifdef CONFIG_COMPACTION
3585 p->capture_control = NULL;
3587 init_numa_balancing(clone_flags, p);
3589 p->wake_entry.u_flags = CSD_TYPE_TTWU;
3590 p->migration_pending = NULL;
3594 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
3596 #ifdef CONFIG_NUMA_BALANCING
3598 void set_numabalancing_state(bool enabled)
3601 static_branch_enable(&sched_numa_balancing);
3603 static_branch_disable(&sched_numa_balancing);
3606 #ifdef CONFIG_PROC_SYSCTL
3607 int sysctl_numa_balancing(struct ctl_table *table, int write,
3608 void *buffer, size_t *lenp, loff_t *ppos)
3612 int state = static_branch_likely(&sched_numa_balancing);
3614 if (write && !capable(CAP_SYS_ADMIN))
3619 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3623 set_numabalancing_state(state);
3629 #ifdef CONFIG_SCHEDSTATS
3631 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
3632 static bool __initdata __sched_schedstats = false;
3634 static void set_schedstats(bool enabled)
3637 static_branch_enable(&sched_schedstats);
3639 static_branch_disable(&sched_schedstats);
3642 void force_schedstat_enabled(void)
3644 if (!schedstat_enabled()) {
3645 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
3646 static_branch_enable(&sched_schedstats);
3650 static int __init setup_schedstats(char *str)
3657 * This code is called before jump labels have been set up, so we can't
3658 * change the static branch directly just yet. Instead set a temporary
3659 * variable so init_schedstats() can do it later.
3661 if (!strcmp(str, "enable")) {
3662 __sched_schedstats = true;
3664 } else if (!strcmp(str, "disable")) {
3665 __sched_schedstats = false;
3670 pr_warn("Unable to parse schedstats=\n");
3674 __setup("schedstats=", setup_schedstats);
3676 static void __init init_schedstats(void)
3678 set_schedstats(__sched_schedstats);
3681 #ifdef CONFIG_PROC_SYSCTL
3682 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
3683 size_t *lenp, loff_t *ppos)
3687 int state = static_branch_likely(&sched_schedstats);
3689 if (write && !capable(CAP_SYS_ADMIN))
3694 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3698 set_schedstats(state);
3701 #endif /* CONFIG_PROC_SYSCTL */
3702 #else /* !CONFIG_SCHEDSTATS */
3703 static inline void init_schedstats(void) {}
3704 #endif /* CONFIG_SCHEDSTATS */
3707 * fork()/clone()-time setup:
3709 int sched_fork(unsigned long clone_flags, struct task_struct *p)
3711 unsigned long flags;
3713 __sched_fork(clone_flags, p);
3715 * We mark the process as NEW here. This guarantees that
3716 * nobody will actually run it, and a signal or other external
3717 * event cannot wake it up and insert it on the runqueue either.
3719 p->state = TASK_NEW;
3722 * Make sure we do not leak PI boosting priority to the child.
3724 p->prio = current->normal_prio;
3729 * Revert to default priority/policy on fork if requested.
3731 if (unlikely(p->sched_reset_on_fork)) {
3732 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3733 p->policy = SCHED_NORMAL;
3734 p->static_prio = NICE_TO_PRIO(0);
3736 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3737 p->static_prio = NICE_TO_PRIO(0);
3739 p->prio = p->normal_prio = __normal_prio(p);
3740 set_load_weight(p, false);
3743 * We don't need the reset flag anymore after the fork. It has
3744 * fulfilled its duty:
3746 p->sched_reset_on_fork = 0;
3749 if (dl_prio(p->prio))
3751 else if (rt_prio(p->prio))
3752 p->sched_class = &rt_sched_class;
3754 p->sched_class = &fair_sched_class;
3756 init_entity_runnable_average(&p->se);
3759 * The child is not yet in the pid-hash so no cgroup attach races,
3760 * and the cgroup is pinned to this child due to cgroup_fork()
3761 * is ran before sched_fork().
3763 * Silence PROVE_RCU.
3765 raw_spin_lock_irqsave(&p->pi_lock, flags);
3768 * We're setting the CPU for the first time, we don't migrate,
3769 * so use __set_task_cpu().
3771 __set_task_cpu(p, smp_processor_id());
3772 if (p->sched_class->task_fork)
3773 p->sched_class->task_fork(p);
3774 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3776 #ifdef CONFIG_SCHED_INFO
3777 if (likely(sched_info_on()))
3778 memset(&p->sched_info, 0, sizeof(p->sched_info));
3780 #if defined(CONFIG_SMP)
3783 init_task_preempt_count(p);
3785 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3786 RB_CLEAR_NODE(&p->pushable_dl_tasks);
3791 void sched_post_fork(struct task_struct *p)
3793 uclamp_post_fork(p);
3796 unsigned long to_ratio(u64 period, u64 runtime)
3798 if (runtime == RUNTIME_INF)
3802 * Doing this here saves a lot of checks in all
3803 * the calling paths, and returning zero seems
3804 * safe for them anyway.
3809 return div64_u64(runtime << BW_SHIFT, period);
3813 * wake_up_new_task - wake up a newly created task for the first time.
3815 * This function will do some initial scheduler statistics housekeeping
3816 * that must be done for every newly created context, then puts the task
3817 * on the runqueue and wakes it.
3819 void wake_up_new_task(struct task_struct *p)
3824 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3825 p->state = TASK_RUNNING;
3828 * Fork balancing, do it here and not earlier because:
3829 * - cpus_ptr can change in the fork path
3830 * - any previously selected CPU might disappear through hotplug
3832 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3833 * as we're not fully set-up yet.
3835 p->recent_used_cpu = task_cpu(p);
3837 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
3839 rq = __task_rq_lock(p, &rf);
3840 update_rq_clock(rq);
3841 post_init_entity_util_avg(p);
3843 activate_task(rq, p, ENQUEUE_NOCLOCK);
3844 trace_sched_wakeup_new(p);
3845 check_preempt_curr(rq, p, WF_FORK);
3847 if (p->sched_class->task_woken) {
3849 * Nothing relies on rq->lock after this, so it's fine to
3852 rq_unpin_lock(rq, &rf);
3853 p->sched_class->task_woken(rq, p);
3854 rq_repin_lock(rq, &rf);
3857 task_rq_unlock(rq, p, &rf);
3860 #ifdef CONFIG_PREEMPT_NOTIFIERS
3862 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
3864 void preempt_notifier_inc(void)
3866 static_branch_inc(&preempt_notifier_key);
3868 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
3870 void preempt_notifier_dec(void)
3872 static_branch_dec(&preempt_notifier_key);
3874 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
3877 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3878 * @notifier: notifier struct to register
3880 void preempt_notifier_register(struct preempt_notifier *notifier)
3882 if (!static_branch_unlikely(&preempt_notifier_key))
3883 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3885 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3887 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3890 * preempt_notifier_unregister - no longer interested in preemption notifications
3891 * @notifier: notifier struct to unregister
3893 * This is *not* safe to call from within a preemption notifier.
3895 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3897 hlist_del(¬ifier->link);
3899 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3901 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3903 struct preempt_notifier *notifier;
3905 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3906 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3909 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3911 if (static_branch_unlikely(&preempt_notifier_key))
3912 __fire_sched_in_preempt_notifiers(curr);
3916 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
3917 struct task_struct *next)
3919 struct preempt_notifier *notifier;
3921 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3922 notifier->ops->sched_out(notifier, next);
3925 static __always_inline void
3926 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3927 struct task_struct *next)
3929 if (static_branch_unlikely(&preempt_notifier_key))
3930 __fire_sched_out_preempt_notifiers(curr, next);
3933 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3935 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3940 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3941 struct task_struct *next)
3945 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3947 static inline void prepare_task(struct task_struct *next)
3951 * Claim the task as running, we do this before switching to it
3952 * such that any running task will have this set.
3954 * See the ttwu() WF_ON_CPU case and its ordering comment.
3956 WRITE_ONCE(next->on_cpu, 1);
3960 static inline void finish_task(struct task_struct *prev)
3964 * This must be the very last reference to @prev from this CPU. After
3965 * p->on_cpu is cleared, the task can be moved to a different CPU. We
3966 * must ensure this doesn't happen until the switch is completely
3969 * In particular, the load of prev->state in finish_task_switch() must
3970 * happen before this.
3972 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3974 smp_store_release(&prev->on_cpu, 0);
3980 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
3982 void (*func)(struct rq *rq);
3983 struct callback_head *next;
3985 lockdep_assert_held(&rq->lock);
3988 func = (void (*)(struct rq *))head->func;
3997 static void balance_push(struct rq *rq);
3999 struct callback_head balance_push_callback = {
4001 .func = (void (*)(struct callback_head *))balance_push,
4004 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4006 struct callback_head *head = rq->balance_callback;
4008 lockdep_assert_held(&rq->lock);
4010 rq->balance_callback = NULL;
4015 static void __balance_callbacks(struct rq *rq)
4017 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4020 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4022 unsigned long flags;
4024 if (unlikely(head)) {
4025 raw_spin_lock_irqsave(&rq->lock, flags);
4026 do_balance_callbacks(rq, head);
4027 raw_spin_unlock_irqrestore(&rq->lock, flags);
4033 static inline void __balance_callbacks(struct rq *rq)
4037 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4042 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4049 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4052 * Since the runqueue lock will be released by the next
4053 * task (which is an invalid locking op but in the case
4054 * of the scheduler it's an obvious special-case), so we
4055 * do an early lockdep release here:
4057 rq_unpin_lock(rq, rf);
4058 spin_release(&rq->lock.dep_map, _THIS_IP_);
4059 #ifdef CONFIG_DEBUG_SPINLOCK
4060 /* this is a valid case when another task releases the spinlock */
4061 rq->lock.owner = next;
4065 static inline void finish_lock_switch(struct rq *rq)
4068 * If we are tracking spinlock dependencies then we have to
4069 * fix up the runqueue lock - which gets 'carried over' from
4070 * prev into current:
4072 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
4073 __balance_callbacks(rq);
4074 raw_spin_unlock_irq(&rq->lock);
4078 * NOP if the arch has not defined these:
4081 #ifndef prepare_arch_switch
4082 # define prepare_arch_switch(next) do { } while (0)
4085 #ifndef finish_arch_post_lock_switch
4086 # define finish_arch_post_lock_switch() do { } while (0)
4089 static inline void kmap_local_sched_out(void)
4091 #ifdef CONFIG_KMAP_LOCAL
4092 if (unlikely(current->kmap_ctrl.idx))
4093 __kmap_local_sched_out();
4097 static inline void kmap_local_sched_in(void)
4099 #ifdef CONFIG_KMAP_LOCAL
4100 if (unlikely(current->kmap_ctrl.idx))
4101 __kmap_local_sched_in();
4106 * prepare_task_switch - prepare to switch tasks
4107 * @rq: the runqueue preparing to switch
4108 * @prev: the current task that is being switched out
4109 * @next: the task we are going to switch to.
4111 * This is called with the rq lock held and interrupts off. It must
4112 * be paired with a subsequent finish_task_switch after the context
4115 * prepare_task_switch sets up locking and calls architecture specific
4119 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4120 struct task_struct *next)
4122 kcov_prepare_switch(prev);
4123 sched_info_switch(rq, prev, next);
4124 perf_event_task_sched_out(prev, next);
4126 fire_sched_out_preempt_notifiers(prev, next);
4127 kmap_local_sched_out();
4129 prepare_arch_switch(next);
4133 * finish_task_switch - clean up after a task-switch
4134 * @prev: the thread we just switched away from.
4136 * finish_task_switch must be called after the context switch, paired
4137 * with a prepare_task_switch call before the context switch.
4138 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4139 * and do any other architecture-specific cleanup actions.
4141 * Note that we may have delayed dropping an mm in context_switch(). If
4142 * so, we finish that here outside of the runqueue lock. (Doing it
4143 * with the lock held can cause deadlocks; see schedule() for
4146 * The context switch have flipped the stack from under us and restored the
4147 * local variables which were saved when this task called schedule() in the
4148 * past. prev == current is still correct but we need to recalculate this_rq
4149 * because prev may have moved to another CPU.
4151 static struct rq *finish_task_switch(struct task_struct *prev)
4152 __releases(rq->lock)
4154 struct rq *rq = this_rq();
4155 struct mm_struct *mm = rq->prev_mm;
4159 * The previous task will have left us with a preempt_count of 2
4160 * because it left us after:
4163 * preempt_disable(); // 1
4165 * raw_spin_lock_irq(&rq->lock) // 2
4167 * Also, see FORK_PREEMPT_COUNT.
4169 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4170 "corrupted preempt_count: %s/%d/0x%x\n",
4171 current->comm, current->pid, preempt_count()))
4172 preempt_count_set(FORK_PREEMPT_COUNT);
4177 * A task struct has one reference for the use as "current".
4178 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4179 * schedule one last time. The schedule call will never return, and
4180 * the scheduled task must drop that reference.
4182 * We must observe prev->state before clearing prev->on_cpu (in
4183 * finish_task), otherwise a concurrent wakeup can get prev
4184 * running on another CPU and we could rave with its RUNNING -> DEAD
4185 * transition, resulting in a double drop.
4187 prev_state = prev->state;
4188 vtime_task_switch(prev);
4189 perf_event_task_sched_in(prev, current);
4191 finish_lock_switch(rq);
4192 finish_arch_post_lock_switch();
4193 kcov_finish_switch(current);
4195 * kmap_local_sched_out() is invoked with rq::lock held and
4196 * interrupts disabled. There is no requirement for that, but the
4197 * sched out code does not have an interrupt enabled section.
4198 * Restoring the maps on sched in does not require interrupts being
4201 kmap_local_sched_in();
4203 fire_sched_in_preempt_notifiers(current);
4205 * When switching through a kernel thread, the loop in
4206 * membarrier_{private,global}_expedited() may have observed that
4207 * kernel thread and not issued an IPI. It is therefore possible to
4208 * schedule between user->kernel->user threads without passing though
4209 * switch_mm(). Membarrier requires a barrier after storing to
4210 * rq->curr, before returning to userspace, so provide them here:
4212 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4213 * provided by mmdrop(),
4214 * - a sync_core for SYNC_CORE.
4217 membarrier_mm_sync_core_before_usermode(mm);
4220 if (unlikely(prev_state == TASK_DEAD)) {
4221 if (prev->sched_class->task_dead)
4222 prev->sched_class->task_dead(prev);
4225 * Remove function-return probe instances associated with this
4226 * task and put them back on the free list.
4228 kprobe_flush_task(prev);
4230 /* Task is done with its stack. */
4231 put_task_stack(prev);
4233 put_task_struct_rcu_user(prev);
4236 tick_nohz_task_switch();
4241 * schedule_tail - first thing a freshly forked thread must call.
4242 * @prev: the thread we just switched away from.
4244 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4245 __releases(rq->lock)
4250 * New tasks start with FORK_PREEMPT_COUNT, see there and
4251 * finish_task_switch() for details.
4253 * finish_task_switch() will drop rq->lock() and lower preempt_count
4254 * and the preempt_enable() will end up enabling preemption (on
4255 * PREEMPT_COUNT kernels).
4258 rq = finish_task_switch(prev);
4261 if (current->set_child_tid)
4262 put_user(task_pid_vnr(current), current->set_child_tid);
4264 calculate_sigpending();
4268 * context_switch - switch to the new MM and the new thread's register state.
4270 static __always_inline struct rq *
4271 context_switch(struct rq *rq, struct task_struct *prev,
4272 struct task_struct *next, struct rq_flags *rf)
4274 prepare_task_switch(rq, prev, next);
4277 * For paravirt, this is coupled with an exit in switch_to to
4278 * combine the page table reload and the switch backend into
4281 arch_start_context_switch(prev);
4284 * kernel -> kernel lazy + transfer active
4285 * user -> kernel lazy + mmgrab() active
4287 * kernel -> user switch + mmdrop() active
4288 * user -> user switch
4290 if (!next->mm) { // to kernel
4291 enter_lazy_tlb(prev->active_mm, next);
4293 next->active_mm = prev->active_mm;
4294 if (prev->mm) // from user
4295 mmgrab(prev->active_mm);
4297 prev->active_mm = NULL;
4299 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4301 * sys_membarrier() requires an smp_mb() between setting
4302 * rq->curr / membarrier_switch_mm() and returning to userspace.
4304 * The below provides this either through switch_mm(), or in
4305 * case 'prev->active_mm == next->mm' through
4306 * finish_task_switch()'s mmdrop().
4308 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4310 if (!prev->mm) { // from kernel
4311 /* will mmdrop() in finish_task_switch(). */
4312 rq->prev_mm = prev->active_mm;
4313 prev->active_mm = NULL;
4317 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4319 prepare_lock_switch(rq, next, rf);
4321 /* Here we just switch the register state and the stack. */
4322 switch_to(prev, next, prev);
4325 return finish_task_switch(prev);
4329 * nr_running and nr_context_switches:
4331 * externally visible scheduler statistics: current number of runnable
4332 * threads, total number of context switches performed since bootup.
4334 unsigned long nr_running(void)
4336 unsigned long i, sum = 0;
4338 for_each_online_cpu(i)
4339 sum += cpu_rq(i)->nr_running;
4345 * Check if only the current task is running on the CPU.
4347 * Caution: this function does not check that the caller has disabled
4348 * preemption, thus the result might have a time-of-check-to-time-of-use
4349 * race. The caller is responsible to use it correctly, for example:
4351 * - from a non-preemptible section (of course)
4353 * - from a thread that is bound to a single CPU
4355 * - in a loop with very short iterations (e.g. a polling loop)
4357 bool single_task_running(void)
4359 return raw_rq()->nr_running == 1;
4361 EXPORT_SYMBOL(single_task_running);
4363 unsigned long long nr_context_switches(void)
4366 unsigned long long sum = 0;
4368 for_each_possible_cpu(i)
4369 sum += cpu_rq(i)->nr_switches;
4375 * Consumers of these two interfaces, like for example the cpuidle menu
4376 * governor, are using nonsensical data. Preferring shallow idle state selection
4377 * for a CPU that has IO-wait which might not even end up running the task when
4378 * it does become runnable.
4381 unsigned long nr_iowait_cpu(int cpu)
4383 return atomic_read(&cpu_rq(cpu)->nr_iowait);
4387 * IO-wait accounting, and how it's mostly bollocks (on SMP).
4389 * The idea behind IO-wait account is to account the idle time that we could
4390 * have spend running if it were not for IO. That is, if we were to improve the
4391 * storage performance, we'd have a proportional reduction in IO-wait time.
4393 * This all works nicely on UP, where, when a task blocks on IO, we account
4394 * idle time as IO-wait, because if the storage were faster, it could've been
4395 * running and we'd not be idle.
4397 * This has been extended to SMP, by doing the same for each CPU. This however
4400 * Imagine for instance the case where two tasks block on one CPU, only the one
4401 * CPU will have IO-wait accounted, while the other has regular idle. Even
4402 * though, if the storage were faster, both could've ran at the same time,
4403 * utilising both CPUs.
4405 * This means, that when looking globally, the current IO-wait accounting on
4406 * SMP is a lower bound, by reason of under accounting.
4408 * Worse, since the numbers are provided per CPU, they are sometimes
4409 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4410 * associated with any one particular CPU, it can wake to another CPU than it
4411 * blocked on. This means the per CPU IO-wait number is meaningless.
4413 * Task CPU affinities can make all that even more 'interesting'.
4416 unsigned long nr_iowait(void)
4418 unsigned long i, sum = 0;
4420 for_each_possible_cpu(i)
4421 sum += nr_iowait_cpu(i);
4429 * sched_exec - execve() is a valuable balancing opportunity, because at
4430 * this point the task has the smallest effective memory and cache footprint.
4432 void sched_exec(void)
4434 struct task_struct *p = current;
4435 unsigned long flags;
4438 raw_spin_lock_irqsave(&p->pi_lock, flags);
4439 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
4440 if (dest_cpu == smp_processor_id())
4443 if (likely(cpu_active(dest_cpu))) {
4444 struct migration_arg arg = { p, dest_cpu };
4446 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4447 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
4451 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4456 DEFINE_PER_CPU(struct kernel_stat, kstat);
4457 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
4459 EXPORT_PER_CPU_SYMBOL(kstat);
4460 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
4463 * The function fair_sched_class.update_curr accesses the struct curr
4464 * and its field curr->exec_start; when called from task_sched_runtime(),
4465 * we observe a high rate of cache misses in practice.
4466 * Prefetching this data results in improved performance.
4468 static inline void prefetch_curr_exec_start(struct task_struct *p)
4470 #ifdef CONFIG_FAIR_GROUP_SCHED
4471 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
4473 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
4476 prefetch(&curr->exec_start);
4480 * Return accounted runtime for the task.
4481 * In case the task is currently running, return the runtime plus current's
4482 * pending runtime that have not been accounted yet.
4484 unsigned long long task_sched_runtime(struct task_struct *p)
4490 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
4492 * 64-bit doesn't need locks to atomically read a 64-bit value.
4493 * So we have a optimization chance when the task's delta_exec is 0.
4494 * Reading ->on_cpu is racy, but this is ok.
4496 * If we race with it leaving CPU, we'll take a lock. So we're correct.
4497 * If we race with it entering CPU, unaccounted time is 0. This is
4498 * indistinguishable from the read occurring a few cycles earlier.
4499 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
4500 * been accounted, so we're correct here as well.
4502 if (!p->on_cpu || !task_on_rq_queued(p))
4503 return p->se.sum_exec_runtime;
4506 rq = task_rq_lock(p, &rf);
4508 * Must be ->curr _and_ ->on_rq. If dequeued, we would
4509 * project cycles that may never be accounted to this
4510 * thread, breaking clock_gettime().
4512 if (task_current(rq, p) && task_on_rq_queued(p)) {
4513 prefetch_curr_exec_start(p);
4514 update_rq_clock(rq);
4515 p->sched_class->update_curr(rq);
4517 ns = p->se.sum_exec_runtime;
4518 task_rq_unlock(rq, p, &rf);
4524 * This function gets called by the timer code, with HZ frequency.
4525 * We call it with interrupts disabled.
4527 void scheduler_tick(void)
4529 int cpu = smp_processor_id();
4530 struct rq *rq = cpu_rq(cpu);
4531 struct task_struct *curr = rq->curr;
4533 unsigned long thermal_pressure;
4535 arch_scale_freq_tick();
4540 update_rq_clock(rq);
4541 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
4542 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
4543 curr->sched_class->task_tick(rq, curr, 0);
4544 calc_global_load_tick(rq);
4549 perf_event_task_tick();
4552 rq->idle_balance = idle_cpu(cpu);
4553 trigger_load_balance(rq);
4557 #ifdef CONFIG_NO_HZ_FULL
4562 struct delayed_work work;
4564 /* Values for ->state, see diagram below. */
4565 #define TICK_SCHED_REMOTE_OFFLINE 0
4566 #define TICK_SCHED_REMOTE_OFFLINING 1
4567 #define TICK_SCHED_REMOTE_RUNNING 2
4570 * State diagram for ->state:
4573 * TICK_SCHED_REMOTE_OFFLINE
4576 * | | sched_tick_remote()
4579 * +--TICK_SCHED_REMOTE_OFFLINING
4582 * sched_tick_start() | | sched_tick_stop()
4585 * TICK_SCHED_REMOTE_RUNNING
4588 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
4589 * and sched_tick_start() are happy to leave the state in RUNNING.
4592 static struct tick_work __percpu *tick_work_cpu;
4594 static void sched_tick_remote(struct work_struct *work)
4596 struct delayed_work *dwork = to_delayed_work(work);
4597 struct tick_work *twork = container_of(dwork, struct tick_work, work);
4598 int cpu = twork->cpu;
4599 struct rq *rq = cpu_rq(cpu);
4600 struct task_struct *curr;
4606 * Handle the tick only if it appears the remote CPU is running in full
4607 * dynticks mode. The check is racy by nature, but missing a tick or
4608 * having one too much is no big deal because the scheduler tick updates
4609 * statistics and checks timeslices in a time-independent way, regardless
4610 * of when exactly it is running.
4612 if (!tick_nohz_tick_stopped_cpu(cpu))
4615 rq_lock_irq(rq, &rf);
4617 if (cpu_is_offline(cpu))
4620 update_rq_clock(rq);
4622 if (!is_idle_task(curr)) {
4624 * Make sure the next tick runs within a reasonable
4627 delta = rq_clock_task(rq) - curr->se.exec_start;
4628 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
4630 curr->sched_class->task_tick(rq, curr, 0);
4632 calc_load_nohz_remote(rq);
4634 rq_unlock_irq(rq, &rf);
4638 * Run the remote tick once per second (1Hz). This arbitrary
4639 * frequency is large enough to avoid overload but short enough
4640 * to keep scheduler internal stats reasonably up to date. But
4641 * first update state to reflect hotplug activity if required.
4643 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
4644 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
4645 if (os == TICK_SCHED_REMOTE_RUNNING)
4646 queue_delayed_work(system_unbound_wq, dwork, HZ);
4649 static void sched_tick_start(int cpu)
4652 struct tick_work *twork;
4654 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4657 WARN_ON_ONCE(!tick_work_cpu);
4659 twork = per_cpu_ptr(tick_work_cpu, cpu);
4660 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
4661 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
4662 if (os == TICK_SCHED_REMOTE_OFFLINE) {
4664 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
4665 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
4669 #ifdef CONFIG_HOTPLUG_CPU
4670 static void sched_tick_stop(int cpu)
4672 struct tick_work *twork;
4675 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4678 WARN_ON_ONCE(!tick_work_cpu);
4680 twork = per_cpu_ptr(tick_work_cpu, cpu);
4681 /* There cannot be competing actions, but don't rely on stop-machine. */
4682 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
4683 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
4684 /* Don't cancel, as this would mess up the state machine. */
4686 #endif /* CONFIG_HOTPLUG_CPU */
4688 int __init sched_tick_offload_init(void)
4690 tick_work_cpu = alloc_percpu(struct tick_work);
4691 BUG_ON(!tick_work_cpu);
4695 #else /* !CONFIG_NO_HZ_FULL */
4696 static inline void sched_tick_start(int cpu) { }
4697 static inline void sched_tick_stop(int cpu) { }
4700 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
4701 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
4703 * If the value passed in is equal to the current preempt count
4704 * then we just disabled preemption. Start timing the latency.
4706 static inline void preempt_latency_start(int val)
4708 if (preempt_count() == val) {
4709 unsigned long ip = get_lock_parent_ip();
4710 #ifdef CONFIG_DEBUG_PREEMPT
4711 current->preempt_disable_ip = ip;
4713 trace_preempt_off(CALLER_ADDR0, ip);
4717 void preempt_count_add(int val)
4719 #ifdef CONFIG_DEBUG_PREEMPT
4723 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4726 __preempt_count_add(val);
4727 #ifdef CONFIG_DEBUG_PREEMPT
4729 * Spinlock count overflowing soon?
4731 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4734 preempt_latency_start(val);
4736 EXPORT_SYMBOL(preempt_count_add);
4737 NOKPROBE_SYMBOL(preempt_count_add);
4740 * If the value passed in equals to the current preempt count
4741 * then we just enabled preemption. Stop timing the latency.
4743 static inline void preempt_latency_stop(int val)
4745 if (preempt_count() == val)
4746 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
4749 void preempt_count_sub(int val)
4751 #ifdef CONFIG_DEBUG_PREEMPT
4755 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4758 * Is the spinlock portion underflowing?
4760 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4761 !(preempt_count() & PREEMPT_MASK)))
4765 preempt_latency_stop(val);
4766 __preempt_count_sub(val);
4768 EXPORT_SYMBOL(preempt_count_sub);
4769 NOKPROBE_SYMBOL(preempt_count_sub);
4772 static inline void preempt_latency_start(int val) { }
4773 static inline void preempt_latency_stop(int val) { }
4776 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
4778 #ifdef CONFIG_DEBUG_PREEMPT
4779 return p->preempt_disable_ip;
4786 * Print scheduling while atomic bug:
4788 static noinline void __schedule_bug(struct task_struct *prev)
4790 /* Save this before calling printk(), since that will clobber it */
4791 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
4793 if (oops_in_progress)
4796 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4797 prev->comm, prev->pid, preempt_count());
4799 debug_show_held_locks(prev);
4801 if (irqs_disabled())
4802 print_irqtrace_events(prev);
4803 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
4804 && in_atomic_preempt_off()) {
4805 pr_err("Preemption disabled at:");
4806 print_ip_sym(KERN_ERR, preempt_disable_ip);
4809 panic("scheduling while atomic\n");
4812 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4816 * Various schedule()-time debugging checks and statistics:
4818 static inline void schedule_debug(struct task_struct *prev, bool preempt)
4820 #ifdef CONFIG_SCHED_STACK_END_CHECK
4821 if (task_stack_end_corrupted(prev))
4822 panic("corrupted stack end detected inside scheduler\n");
4824 if (task_scs_end_corrupted(prev))
4825 panic("corrupted shadow stack detected inside scheduler\n");
4828 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
4829 if (!preempt && prev->state && prev->non_block_count) {
4830 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
4831 prev->comm, prev->pid, prev->non_block_count);
4833 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4837 if (unlikely(in_atomic_preempt_off())) {
4838 __schedule_bug(prev);
4839 preempt_count_set(PREEMPT_DISABLED);
4842 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
4844 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4846 schedstat_inc(this_rq()->sched_count);
4849 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
4850 struct rq_flags *rf)
4853 const struct sched_class *class;
4855 * We must do the balancing pass before put_prev_task(), such
4856 * that when we release the rq->lock the task is in the same
4857 * state as before we took rq->lock.
4859 * We can terminate the balance pass as soon as we know there is
4860 * a runnable task of @class priority or higher.
4862 for_class_range(class, prev->sched_class, &idle_sched_class) {
4863 if (class->balance(rq, prev, rf))
4868 put_prev_task(rq, prev);
4872 * Pick up the highest-prio task:
4874 static inline struct task_struct *
4875 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
4877 const struct sched_class *class;
4878 struct task_struct *p;
4881 * Optimization: we know that if all tasks are in the fair class we can
4882 * call that function directly, but only if the @prev task wasn't of a
4883 * higher scheduling class, because otherwise those lose the
4884 * opportunity to pull in more work from other CPUs.
4886 if (likely(prev->sched_class <= &fair_sched_class &&
4887 rq->nr_running == rq->cfs.h_nr_running)) {
4889 p = pick_next_task_fair(rq, prev, rf);
4890 if (unlikely(p == RETRY_TASK))
4893 /* Assumes fair_sched_class->next == idle_sched_class */
4895 put_prev_task(rq, prev);
4896 p = pick_next_task_idle(rq);
4903 put_prev_task_balance(rq, prev, rf);
4905 for_each_class(class) {
4906 p = class->pick_next_task(rq);
4911 /* The idle class should always have a runnable task: */
4916 * __schedule() is the main scheduler function.
4918 * The main means of driving the scheduler and thus entering this function are:
4920 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4922 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4923 * paths. For example, see arch/x86/entry_64.S.
4925 * To drive preemption between tasks, the scheduler sets the flag in timer
4926 * interrupt handler scheduler_tick().
4928 * 3. Wakeups don't really cause entry into schedule(). They add a
4929 * task to the run-queue and that's it.
4931 * Now, if the new task added to the run-queue preempts the current
4932 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4933 * called on the nearest possible occasion:
4935 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4937 * - in syscall or exception context, at the next outmost
4938 * preempt_enable(). (this might be as soon as the wake_up()'s
4941 * - in IRQ context, return from interrupt-handler to
4942 * preemptible context
4944 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4947 * - cond_resched() call
4948 * - explicit schedule() call
4949 * - return from syscall or exception to user-space
4950 * - return from interrupt-handler to user-space
4952 * WARNING: must be called with preemption disabled!
4954 static void __sched notrace __schedule(bool preempt)
4956 struct task_struct *prev, *next;
4957 unsigned long *switch_count;
4958 unsigned long prev_state;
4963 cpu = smp_processor_id();
4967 schedule_debug(prev, preempt);
4969 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
4972 local_irq_disable();
4973 rcu_note_context_switch(preempt);
4976 * Make sure that signal_pending_state()->signal_pending() below
4977 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4978 * done by the caller to avoid the race with signal_wake_up():
4980 * __set_current_state(@state) signal_wake_up()
4981 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
4982 * wake_up_state(p, state)
4983 * LOCK rq->lock LOCK p->pi_state
4984 * smp_mb__after_spinlock() smp_mb__after_spinlock()
4985 * if (signal_pending_state()) if (p->state & @state)
4987 * Also, the membarrier system call requires a full memory barrier
4988 * after coming from user-space, before storing to rq->curr.
4991 smp_mb__after_spinlock();
4993 /* Promote REQ to ACT */
4994 rq->clock_update_flags <<= 1;
4995 update_rq_clock(rq);
4997 switch_count = &prev->nivcsw;
5000 * We must load prev->state once (task_struct::state is volatile), such
5003 * - we form a control dependency vs deactivate_task() below.
5004 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
5006 prev_state = prev->state;
5007 if (!preempt && prev_state) {
5008 if (signal_pending_state(prev_state, prev)) {
5009 prev->state = TASK_RUNNING;
5011 prev->sched_contributes_to_load =
5012 (prev_state & TASK_UNINTERRUPTIBLE) &&
5013 !(prev_state & TASK_NOLOAD) &&
5014 !(prev->flags & PF_FROZEN);
5016 if (prev->sched_contributes_to_load)
5017 rq->nr_uninterruptible++;
5020 * __schedule() ttwu()
5021 * prev_state = prev->state; if (p->on_rq && ...)
5022 * if (prev_state) goto out;
5023 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
5024 * p->state = TASK_WAKING
5026 * Where __schedule() and ttwu() have matching control dependencies.
5028 * After this, schedule() must not care about p->state any more.
5030 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
5032 if (prev->in_iowait) {
5033 atomic_inc(&rq->nr_iowait);
5034 delayacct_blkio_start();
5037 switch_count = &prev->nvcsw;
5040 next = pick_next_task(rq, prev, &rf);
5041 clear_tsk_need_resched(prev);
5042 clear_preempt_need_resched();
5044 if (likely(prev != next)) {
5047 * RCU users of rcu_dereference(rq->curr) may not see
5048 * changes to task_struct made by pick_next_task().
5050 RCU_INIT_POINTER(rq->curr, next);
5052 * The membarrier system call requires each architecture
5053 * to have a full memory barrier after updating
5054 * rq->curr, before returning to user-space.
5056 * Here are the schemes providing that barrier on the
5057 * various architectures:
5058 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
5059 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
5060 * - finish_lock_switch() for weakly-ordered
5061 * architectures where spin_unlock is a full barrier,
5062 * - switch_to() for arm64 (weakly-ordered, spin_unlock
5063 * is a RELEASE barrier),
5067 migrate_disable_switch(rq, prev);
5068 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
5070 trace_sched_switch(preempt, prev, next);
5072 /* Also unlocks the rq: */
5073 rq = context_switch(rq, prev, next, &rf);
5075 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5077 rq_unpin_lock(rq, &rf);
5078 __balance_callbacks(rq);
5079 raw_spin_unlock_irq(&rq->lock);
5083 void __noreturn do_task_dead(void)
5085 /* Causes final put_task_struct in finish_task_switch(): */
5086 set_special_state(TASK_DEAD);
5088 /* Tell freezer to ignore us: */
5089 current->flags |= PF_NOFREEZE;
5094 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
5099 static inline void sched_submit_work(struct task_struct *tsk)
5101 unsigned int task_flags;
5106 task_flags = tsk->flags;
5108 * If a worker went to sleep, notify and ask workqueue whether
5109 * it wants to wake up a task to maintain concurrency.
5110 * As this function is called inside the schedule() context,
5111 * we disable preemption to avoid it calling schedule() again
5112 * in the possible wakeup of a kworker and because wq_worker_sleeping()
5115 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
5117 if (task_flags & PF_WQ_WORKER)
5118 wq_worker_sleeping(tsk);
5120 io_wq_worker_sleeping(tsk);
5121 preempt_enable_no_resched();
5124 if (tsk_is_pi_blocked(tsk))
5128 * If we are going to sleep and we have plugged IO queued,
5129 * make sure to submit it to avoid deadlocks.
5131 if (blk_needs_flush_plug(tsk))
5132 blk_schedule_flush_plug(tsk);
5135 static void sched_update_worker(struct task_struct *tsk)
5137 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
5138 if (tsk->flags & PF_WQ_WORKER)
5139 wq_worker_running(tsk);
5141 io_wq_worker_running(tsk);
5145 asmlinkage __visible void __sched schedule(void)
5147 struct task_struct *tsk = current;
5149 sched_submit_work(tsk);
5153 sched_preempt_enable_no_resched();
5154 } while (need_resched());
5155 sched_update_worker(tsk);
5157 EXPORT_SYMBOL(schedule);
5160 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
5161 * state (have scheduled out non-voluntarily) by making sure that all
5162 * tasks have either left the run queue or have gone into user space.
5163 * As idle tasks do not do either, they must not ever be preempted
5164 * (schedule out non-voluntarily).
5166 * schedule_idle() is similar to schedule_preempt_disable() except that it
5167 * never enables preemption because it does not call sched_submit_work().
5169 void __sched schedule_idle(void)
5172 * As this skips calling sched_submit_work(), which the idle task does
5173 * regardless because that function is a nop when the task is in a
5174 * TASK_RUNNING state, make sure this isn't used someplace that the
5175 * current task can be in any other state. Note, idle is always in the
5176 * TASK_RUNNING state.
5178 WARN_ON_ONCE(current->state);
5181 } while (need_resched());
5184 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
5185 asmlinkage __visible void __sched schedule_user(void)
5188 * If we come here after a random call to set_need_resched(),
5189 * or we have been woken up remotely but the IPI has not yet arrived,
5190 * we haven't yet exited the RCU idle mode. Do it here manually until
5191 * we find a better solution.
5193 * NB: There are buggy callers of this function. Ideally we
5194 * should warn if prev_state != CONTEXT_USER, but that will trigger
5195 * too frequently to make sense yet.
5197 enum ctx_state prev_state = exception_enter();
5199 exception_exit(prev_state);
5204 * schedule_preempt_disabled - called with preemption disabled
5206 * Returns with preemption disabled. Note: preempt_count must be 1
5208 void __sched schedule_preempt_disabled(void)
5210 sched_preempt_enable_no_resched();
5215 static void __sched notrace preempt_schedule_common(void)
5219 * Because the function tracer can trace preempt_count_sub()
5220 * and it also uses preempt_enable/disable_notrace(), if
5221 * NEED_RESCHED is set, the preempt_enable_notrace() called
5222 * by the function tracer will call this function again and
5223 * cause infinite recursion.
5225 * Preemption must be disabled here before the function
5226 * tracer can trace. Break up preempt_disable() into two
5227 * calls. One to disable preemption without fear of being
5228 * traced. The other to still record the preemption latency,
5229 * which can also be traced by the function tracer.
5231 preempt_disable_notrace();
5232 preempt_latency_start(1);
5234 preempt_latency_stop(1);
5235 preempt_enable_no_resched_notrace();
5238 * Check again in case we missed a preemption opportunity
5239 * between schedule and now.
5241 } while (need_resched());
5244 #ifdef CONFIG_PREEMPTION
5246 * This is the entry point to schedule() from in-kernel preemption
5247 * off of preempt_enable.
5249 asmlinkage __visible void __sched notrace preempt_schedule(void)
5252 * If there is a non-zero preempt_count or interrupts are disabled,
5253 * we do not want to preempt the current task. Just return..
5255 if (likely(!preemptible()))
5258 preempt_schedule_common();
5260 NOKPROBE_SYMBOL(preempt_schedule);
5261 EXPORT_SYMBOL(preempt_schedule);
5263 #ifdef CONFIG_PREEMPT_DYNAMIC
5264 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
5265 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
5270 * preempt_schedule_notrace - preempt_schedule called by tracing
5272 * The tracing infrastructure uses preempt_enable_notrace to prevent
5273 * recursion and tracing preempt enabling caused by the tracing
5274 * infrastructure itself. But as tracing can happen in areas coming
5275 * from userspace or just about to enter userspace, a preempt enable
5276 * can occur before user_exit() is called. This will cause the scheduler
5277 * to be called when the system is still in usermode.
5279 * To prevent this, the preempt_enable_notrace will use this function
5280 * instead of preempt_schedule() to exit user context if needed before
5281 * calling the scheduler.
5283 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
5285 enum ctx_state prev_ctx;
5287 if (likely(!preemptible()))
5292 * Because the function tracer can trace preempt_count_sub()
5293 * and it also uses preempt_enable/disable_notrace(), if
5294 * NEED_RESCHED is set, the preempt_enable_notrace() called
5295 * by the function tracer will call this function again and
5296 * cause infinite recursion.
5298 * Preemption must be disabled here before the function
5299 * tracer can trace. Break up preempt_disable() into two
5300 * calls. One to disable preemption without fear of being
5301 * traced. The other to still record the preemption latency,
5302 * which can also be traced by the function tracer.
5304 preempt_disable_notrace();
5305 preempt_latency_start(1);
5307 * Needs preempt disabled in case user_exit() is traced
5308 * and the tracer calls preempt_enable_notrace() causing
5309 * an infinite recursion.
5311 prev_ctx = exception_enter();
5313 exception_exit(prev_ctx);
5315 preempt_latency_stop(1);
5316 preempt_enable_no_resched_notrace();
5317 } while (need_resched());
5319 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
5321 #ifdef CONFIG_PREEMPT_DYNAMIC
5322 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
5323 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
5326 #endif /* CONFIG_PREEMPTION */
5328 #ifdef CONFIG_PREEMPT_DYNAMIC
5330 #include <linux/entry-common.h>
5335 * SC:preempt_schedule
5336 * SC:preempt_schedule_notrace
5337 * SC:irqentry_exit_cond_resched
5341 * cond_resched <- __cond_resched
5342 * might_resched <- RET0
5343 * preempt_schedule <- NOP
5344 * preempt_schedule_notrace <- NOP
5345 * irqentry_exit_cond_resched <- NOP
5348 * cond_resched <- __cond_resched
5349 * might_resched <- __cond_resched
5350 * preempt_schedule <- NOP
5351 * preempt_schedule_notrace <- NOP
5352 * irqentry_exit_cond_resched <- NOP
5355 * cond_resched <- RET0
5356 * might_resched <- RET0
5357 * preempt_schedule <- preempt_schedule
5358 * preempt_schedule_notrace <- preempt_schedule_notrace
5359 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
5363 preempt_dynamic_none = 0,
5364 preempt_dynamic_voluntary,
5365 preempt_dynamic_full,
5368 static int preempt_dynamic_mode = preempt_dynamic_full;
5370 static int sched_dynamic_mode(const char *str)
5372 if (!strcmp(str, "none"))
5375 if (!strcmp(str, "voluntary"))
5378 if (!strcmp(str, "full"))
5384 static void sched_dynamic_update(int mode)
5387 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
5388 * the ZERO state, which is invalid.
5390 static_call_update(cond_resched, __cond_resched);
5391 static_call_update(might_resched, __cond_resched);
5392 static_call_update(preempt_schedule, __preempt_schedule_func);
5393 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
5394 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
5397 case preempt_dynamic_none:
5398 static_call_update(cond_resched, __cond_resched);
5399 static_call_update(might_resched, (typeof(&__cond_resched)) __static_call_return0);
5400 static_call_update(preempt_schedule, (typeof(&preempt_schedule)) NULL);
5401 static_call_update(preempt_schedule_notrace, (typeof(&preempt_schedule_notrace)) NULL);
5402 static_call_update(irqentry_exit_cond_resched, (typeof(&irqentry_exit_cond_resched)) NULL);
5403 pr_info("Dynamic Preempt: none\n");
5406 case preempt_dynamic_voluntary:
5407 static_call_update(cond_resched, __cond_resched);
5408 static_call_update(might_resched, __cond_resched);
5409 static_call_update(preempt_schedule, (typeof(&preempt_schedule)) NULL);
5410 static_call_update(preempt_schedule_notrace, (typeof(&preempt_schedule_notrace)) NULL);
5411 static_call_update(irqentry_exit_cond_resched, (typeof(&irqentry_exit_cond_resched)) NULL);
5412 pr_info("Dynamic Preempt: voluntary\n");
5415 case preempt_dynamic_full:
5416 static_call_update(cond_resched, (typeof(&__cond_resched)) __static_call_return0);
5417 static_call_update(might_resched, (typeof(&__cond_resched)) __static_call_return0);
5418 static_call_update(preempt_schedule, __preempt_schedule_func);
5419 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
5420 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
5421 pr_info("Dynamic Preempt: full\n");
5425 preempt_dynamic_mode = mode;
5428 static int __init setup_preempt_mode(char *str)
5430 int mode = sched_dynamic_mode(str);
5432 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
5436 sched_dynamic_update(mode);
5439 __setup("preempt=", setup_preempt_mode);
5441 #ifdef CONFIG_SCHED_DEBUG
5443 static ssize_t sched_dynamic_write(struct file *filp, const char __user *ubuf,
5444 size_t cnt, loff_t *ppos)
5452 if (copy_from_user(&buf, ubuf, cnt))
5456 mode = sched_dynamic_mode(strstrip(buf));
5460 sched_dynamic_update(mode);
5467 static int sched_dynamic_show(struct seq_file *m, void *v)
5469 static const char * preempt_modes[] = {
5470 "none", "voluntary", "full"
5474 for (i = 0; i < ARRAY_SIZE(preempt_modes); i++) {
5475 if (preempt_dynamic_mode == i)
5477 seq_puts(m, preempt_modes[i]);
5478 if (preempt_dynamic_mode == i)
5488 static int sched_dynamic_open(struct inode *inode, struct file *filp)
5490 return single_open(filp, sched_dynamic_show, NULL);
5493 static const struct file_operations sched_dynamic_fops = {
5494 .open = sched_dynamic_open,
5495 .write = sched_dynamic_write,
5497 .llseek = seq_lseek,
5498 .release = single_release,
5501 static __init int sched_init_debug_dynamic(void)
5503 debugfs_create_file("sched_preempt", 0644, NULL, NULL, &sched_dynamic_fops);
5506 late_initcall(sched_init_debug_dynamic);
5508 #endif /* CONFIG_SCHED_DEBUG */
5509 #endif /* CONFIG_PREEMPT_DYNAMIC */
5513 * This is the entry point to schedule() from kernel preemption
5514 * off of irq context.
5515 * Note, that this is called and return with irqs disabled. This will
5516 * protect us against recursive calling from irq.
5518 asmlinkage __visible void __sched preempt_schedule_irq(void)
5520 enum ctx_state prev_state;
5522 /* Catch callers which need to be fixed */
5523 BUG_ON(preempt_count() || !irqs_disabled());
5525 prev_state = exception_enter();
5531 local_irq_disable();
5532 sched_preempt_enable_no_resched();
5533 } while (need_resched());
5535 exception_exit(prev_state);
5538 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
5541 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
5542 return try_to_wake_up(curr->private, mode, wake_flags);
5544 EXPORT_SYMBOL(default_wake_function);
5546 #ifdef CONFIG_RT_MUTEXES
5548 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
5551 prio = min(prio, pi_task->prio);
5556 static inline int rt_effective_prio(struct task_struct *p, int prio)
5558 struct task_struct *pi_task = rt_mutex_get_top_task(p);
5560 return __rt_effective_prio(pi_task, prio);
5564 * rt_mutex_setprio - set the current priority of a task
5566 * @pi_task: donor task
5568 * This function changes the 'effective' priority of a task. It does
5569 * not touch ->normal_prio like __setscheduler().
5571 * Used by the rt_mutex code to implement priority inheritance
5572 * logic. Call site only calls if the priority of the task changed.
5574 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
5576 int prio, oldprio, queued, running, queue_flag =
5577 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
5578 const struct sched_class *prev_class;
5582 /* XXX used to be waiter->prio, not waiter->task->prio */
5583 prio = __rt_effective_prio(pi_task, p->normal_prio);
5586 * If nothing changed; bail early.
5588 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
5591 rq = __task_rq_lock(p, &rf);
5592 update_rq_clock(rq);
5594 * Set under pi_lock && rq->lock, such that the value can be used under
5597 * Note that there is loads of tricky to make this pointer cache work
5598 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
5599 * ensure a task is de-boosted (pi_task is set to NULL) before the
5600 * task is allowed to run again (and can exit). This ensures the pointer
5601 * points to a blocked task -- which guarantees the task is present.
5603 p->pi_top_task = pi_task;
5606 * For FIFO/RR we only need to set prio, if that matches we're done.
5608 if (prio == p->prio && !dl_prio(prio))
5612 * Idle task boosting is a nono in general. There is one
5613 * exception, when PREEMPT_RT and NOHZ is active:
5615 * The idle task calls get_next_timer_interrupt() and holds
5616 * the timer wheel base->lock on the CPU and another CPU wants
5617 * to access the timer (probably to cancel it). We can safely
5618 * ignore the boosting request, as the idle CPU runs this code
5619 * with interrupts disabled and will complete the lock
5620 * protected section without being interrupted. So there is no
5621 * real need to boost.
5623 if (unlikely(p == rq->idle)) {
5624 WARN_ON(p != rq->curr);
5625 WARN_ON(p->pi_blocked_on);
5629 trace_sched_pi_setprio(p, pi_task);
5632 if (oldprio == prio)
5633 queue_flag &= ~DEQUEUE_MOVE;
5635 prev_class = p->sched_class;
5636 queued = task_on_rq_queued(p);
5637 running = task_current(rq, p);
5639 dequeue_task(rq, p, queue_flag);
5641 put_prev_task(rq, p);
5644 * Boosting condition are:
5645 * 1. -rt task is running and holds mutex A
5646 * --> -dl task blocks on mutex A
5648 * 2. -dl task is running and holds mutex A
5649 * --> -dl task blocks on mutex A and could preempt the
5652 if (dl_prio(prio)) {
5653 if (!dl_prio(p->normal_prio) ||
5654 (pi_task && dl_prio(pi_task->prio) &&
5655 dl_entity_preempt(&pi_task->dl, &p->dl))) {
5656 p->dl.pi_se = pi_task->dl.pi_se;
5657 queue_flag |= ENQUEUE_REPLENISH;
5659 p->dl.pi_se = &p->dl;
5661 p->sched_class = &dl_sched_class;
5662 } else if (rt_prio(prio)) {
5663 if (dl_prio(oldprio))
5664 p->dl.pi_se = &p->dl;
5666 queue_flag |= ENQUEUE_HEAD;
5667 p->sched_class = &rt_sched_class;
5669 if (dl_prio(oldprio))
5670 p->dl.pi_se = &p->dl;
5671 if (rt_prio(oldprio))
5673 p->sched_class = &fair_sched_class;
5679 enqueue_task(rq, p, queue_flag);
5681 set_next_task(rq, p);
5683 check_class_changed(rq, p, prev_class, oldprio);
5685 /* Avoid rq from going away on us: */
5688 rq_unpin_lock(rq, &rf);
5689 __balance_callbacks(rq);
5690 raw_spin_unlock(&rq->lock);
5695 static inline int rt_effective_prio(struct task_struct *p, int prio)
5701 void set_user_nice(struct task_struct *p, long nice)
5703 bool queued, running;
5708 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
5711 * We have to be careful, if called from sys_setpriority(),
5712 * the task might be in the middle of scheduling on another CPU.
5714 rq = task_rq_lock(p, &rf);
5715 update_rq_clock(rq);
5718 * The RT priorities are set via sched_setscheduler(), but we still
5719 * allow the 'normal' nice value to be set - but as expected
5720 * it won't have any effect on scheduling until the task is
5721 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
5723 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
5724 p->static_prio = NICE_TO_PRIO(nice);
5727 queued = task_on_rq_queued(p);
5728 running = task_current(rq, p);
5730 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
5732 put_prev_task(rq, p);
5734 p->static_prio = NICE_TO_PRIO(nice);
5735 set_load_weight(p, true);
5737 p->prio = effective_prio(p);
5740 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5742 set_next_task(rq, p);
5745 * If the task increased its priority or is running and
5746 * lowered its priority, then reschedule its CPU:
5748 p->sched_class->prio_changed(rq, p, old_prio);
5751 task_rq_unlock(rq, p, &rf);
5753 EXPORT_SYMBOL(set_user_nice);
5756 * can_nice - check if a task can reduce its nice value
5760 int can_nice(const struct task_struct *p, const int nice)
5762 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
5763 int nice_rlim = nice_to_rlimit(nice);
5765 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5766 capable(CAP_SYS_NICE));
5769 #ifdef __ARCH_WANT_SYS_NICE
5772 * sys_nice - change the priority of the current process.
5773 * @increment: priority increment
5775 * sys_setpriority is a more generic, but much slower function that
5776 * does similar things.
5778 SYSCALL_DEFINE1(nice, int, increment)
5783 * Setpriority might change our priority at the same moment.
5784 * We don't have to worry. Conceptually one call occurs first
5785 * and we have a single winner.
5787 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
5788 nice = task_nice(current) + increment;
5790 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
5791 if (increment < 0 && !can_nice(current, nice))
5794 retval = security_task_setnice(current, nice);
5798 set_user_nice(current, nice);
5805 * task_prio - return the priority value of a given task.
5806 * @p: the task in question.
5808 * Return: The priority value as seen by users in /proc.
5810 * sched policy return value kernel prio user prio/nice
5812 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
5813 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
5814 * deadline -101 -1 0
5816 int task_prio(const struct task_struct *p)
5818 return p->prio - MAX_RT_PRIO;
5822 * idle_cpu - is a given CPU idle currently?
5823 * @cpu: the processor in question.
5825 * Return: 1 if the CPU is currently idle. 0 otherwise.
5827 int idle_cpu(int cpu)
5829 struct rq *rq = cpu_rq(cpu);
5831 if (rq->curr != rq->idle)
5838 if (rq->ttwu_pending)
5846 * available_idle_cpu - is a given CPU idle for enqueuing work.
5847 * @cpu: the CPU in question.
5849 * Return: 1 if the CPU is currently idle. 0 otherwise.
5851 int available_idle_cpu(int cpu)
5856 if (vcpu_is_preempted(cpu))
5863 * idle_task - return the idle task for a given CPU.
5864 * @cpu: the processor in question.
5866 * Return: The idle task for the CPU @cpu.
5868 struct task_struct *idle_task(int cpu)
5870 return cpu_rq(cpu)->idle;
5875 * This function computes an effective utilization for the given CPU, to be
5876 * used for frequency selection given the linear relation: f = u * f_max.
5878 * The scheduler tracks the following metrics:
5880 * cpu_util_{cfs,rt,dl,irq}()
5883 * Where the cfs,rt and dl util numbers are tracked with the same metric and
5884 * synchronized windows and are thus directly comparable.
5886 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
5887 * which excludes things like IRQ and steal-time. These latter are then accrued
5888 * in the irq utilization.
5890 * The DL bandwidth number otoh is not a measured metric but a value computed
5891 * based on the task model parameters and gives the minimal utilization
5892 * required to meet deadlines.
5894 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
5895 unsigned long max, enum cpu_util_type type,
5896 struct task_struct *p)
5898 unsigned long dl_util, util, irq;
5899 struct rq *rq = cpu_rq(cpu);
5901 if (!uclamp_is_used() &&
5902 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
5907 * Early check to see if IRQ/steal time saturates the CPU, can be
5908 * because of inaccuracies in how we track these -- see
5909 * update_irq_load_avg().
5911 irq = cpu_util_irq(rq);
5912 if (unlikely(irq >= max))
5916 * Because the time spend on RT/DL tasks is visible as 'lost' time to
5917 * CFS tasks and we use the same metric to track the effective
5918 * utilization (PELT windows are synchronized) we can directly add them
5919 * to obtain the CPU's actual utilization.
5921 * CFS and RT utilization can be boosted or capped, depending on
5922 * utilization clamp constraints requested by currently RUNNABLE
5924 * When there are no CFS RUNNABLE tasks, clamps are released and
5925 * frequency will be gracefully reduced with the utilization decay.
5927 util = util_cfs + cpu_util_rt(rq);
5928 if (type == FREQUENCY_UTIL)
5929 util = uclamp_rq_util_with(rq, util, p);
5931 dl_util = cpu_util_dl(rq);
5934 * For frequency selection we do not make cpu_util_dl() a permanent part
5935 * of this sum because we want to use cpu_bw_dl() later on, but we need
5936 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
5937 * that we select f_max when there is no idle time.
5939 * NOTE: numerical errors or stop class might cause us to not quite hit
5940 * saturation when we should -- something for later.
5942 if (util + dl_util >= max)
5946 * OTOH, for energy computation we need the estimated running time, so
5947 * include util_dl and ignore dl_bw.
5949 if (type == ENERGY_UTIL)
5953 * There is still idle time; further improve the number by using the
5954 * irq metric. Because IRQ/steal time is hidden from the task clock we
5955 * need to scale the task numbers:
5958 * U' = irq + --------- * U
5961 util = scale_irq_capacity(util, irq, max);
5965 * Bandwidth required by DEADLINE must always be granted while, for
5966 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
5967 * to gracefully reduce the frequency when no tasks show up for longer
5970 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
5971 * bw_dl as requested freq. However, cpufreq is not yet ready for such
5972 * an interface. So, we only do the latter for now.
5974 if (type == FREQUENCY_UTIL)
5975 util += cpu_bw_dl(rq);
5977 return min(max, util);
5980 unsigned long sched_cpu_util(int cpu, unsigned long max)
5982 return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max,
5985 #endif /* CONFIG_SMP */
5988 * find_process_by_pid - find a process with a matching PID value.
5989 * @pid: the pid in question.
5991 * The task of @pid, if found. %NULL otherwise.
5993 static struct task_struct *find_process_by_pid(pid_t pid)
5995 return pid ? find_task_by_vpid(pid) : current;
5999 * sched_setparam() passes in -1 for its policy, to let the functions
6000 * it calls know not to change it.
6002 #define SETPARAM_POLICY -1
6004 static void __setscheduler_params(struct task_struct *p,
6005 const struct sched_attr *attr)
6007 int policy = attr->sched_policy;
6009 if (policy == SETPARAM_POLICY)
6014 if (dl_policy(policy))
6015 __setparam_dl(p, attr);
6016 else if (fair_policy(policy))
6017 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
6020 * __sched_setscheduler() ensures attr->sched_priority == 0 when
6021 * !rt_policy. Always setting this ensures that things like
6022 * getparam()/getattr() don't report silly values for !rt tasks.
6024 p->rt_priority = attr->sched_priority;
6025 p->normal_prio = normal_prio(p);
6026 set_load_weight(p, true);
6029 /* Actually do priority change: must hold pi & rq lock. */
6030 static void __setscheduler(struct rq *rq, struct task_struct *p,
6031 const struct sched_attr *attr, bool keep_boost)
6034 * If params can't change scheduling class changes aren't allowed
6037 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
6040 __setscheduler_params(p, attr);
6043 * Keep a potential priority boosting if called from
6044 * sched_setscheduler().
6046 p->prio = normal_prio(p);
6048 p->prio = rt_effective_prio(p, p->prio);
6050 if (dl_prio(p->prio))
6051 p->sched_class = &dl_sched_class;
6052 else if (rt_prio(p->prio))
6053 p->sched_class = &rt_sched_class;
6055 p->sched_class = &fair_sched_class;
6059 * Check the target process has a UID that matches the current process's:
6061 static bool check_same_owner(struct task_struct *p)
6063 const struct cred *cred = current_cred(), *pcred;
6067 pcred = __task_cred(p);
6068 match = (uid_eq(cred->euid, pcred->euid) ||
6069 uid_eq(cred->euid, pcred->uid));
6074 static int __sched_setscheduler(struct task_struct *p,
6075 const struct sched_attr *attr,
6078 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
6079 MAX_RT_PRIO - 1 - attr->sched_priority;
6080 int retval, oldprio, oldpolicy = -1, queued, running;
6081 int new_effective_prio, policy = attr->sched_policy;
6082 const struct sched_class *prev_class;
6083 struct callback_head *head;
6086 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6089 /* The pi code expects interrupts enabled */
6090 BUG_ON(pi && in_interrupt());
6092 /* Double check policy once rq lock held: */
6094 reset_on_fork = p->sched_reset_on_fork;
6095 policy = oldpolicy = p->policy;
6097 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
6099 if (!valid_policy(policy))
6103 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
6107 * Valid priorities for SCHED_FIFO and SCHED_RR are
6108 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
6109 * SCHED_BATCH and SCHED_IDLE is 0.
6111 if (attr->sched_priority > MAX_RT_PRIO-1)
6113 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
6114 (rt_policy(policy) != (attr->sched_priority != 0)))
6118 * Allow unprivileged RT tasks to decrease priority:
6120 if (user && !capable(CAP_SYS_NICE)) {
6121 if (fair_policy(policy)) {
6122 if (attr->sched_nice < task_nice(p) &&
6123 !can_nice(p, attr->sched_nice))
6127 if (rt_policy(policy)) {
6128 unsigned long rlim_rtprio =
6129 task_rlimit(p, RLIMIT_RTPRIO);
6131 /* Can't set/change the rt policy: */
6132 if (policy != p->policy && !rlim_rtprio)
6135 /* Can't increase priority: */
6136 if (attr->sched_priority > p->rt_priority &&
6137 attr->sched_priority > rlim_rtprio)
6142 * Can't set/change SCHED_DEADLINE policy at all for now
6143 * (safest behavior); in the future we would like to allow
6144 * unprivileged DL tasks to increase their relative deadline
6145 * or reduce their runtime (both ways reducing utilization)
6147 if (dl_policy(policy))
6151 * Treat SCHED_IDLE as nice 20. Only allow a switch to
6152 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
6154 if (task_has_idle_policy(p) && !idle_policy(policy)) {
6155 if (!can_nice(p, task_nice(p)))
6159 /* Can't change other user's priorities: */
6160 if (!check_same_owner(p))
6163 /* Normal users shall not reset the sched_reset_on_fork flag: */
6164 if (p->sched_reset_on_fork && !reset_on_fork)
6169 if (attr->sched_flags & SCHED_FLAG_SUGOV)
6172 retval = security_task_setscheduler(p);
6177 /* Update task specific "requested" clamps */
6178 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
6179 retval = uclamp_validate(p, attr);
6188 * Make sure no PI-waiters arrive (or leave) while we are
6189 * changing the priority of the task:
6191 * To be able to change p->policy safely, the appropriate
6192 * runqueue lock must be held.
6194 rq = task_rq_lock(p, &rf);
6195 update_rq_clock(rq);
6198 * Changing the policy of the stop threads its a very bad idea:
6200 if (p == rq->stop) {
6206 * If not changing anything there's no need to proceed further,
6207 * but store a possible modification of reset_on_fork.
6209 if (unlikely(policy == p->policy)) {
6210 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
6212 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
6214 if (dl_policy(policy) && dl_param_changed(p, attr))
6216 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
6219 p->sched_reset_on_fork = reset_on_fork;
6226 #ifdef CONFIG_RT_GROUP_SCHED
6228 * Do not allow realtime tasks into groups that have no runtime
6231 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6232 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
6233 !task_group_is_autogroup(task_group(p))) {
6239 if (dl_bandwidth_enabled() && dl_policy(policy) &&
6240 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
6241 cpumask_t *span = rq->rd->span;
6244 * Don't allow tasks with an affinity mask smaller than
6245 * the entire root_domain to become SCHED_DEADLINE. We
6246 * will also fail if there's no bandwidth available.
6248 if (!cpumask_subset(span, p->cpus_ptr) ||
6249 rq->rd->dl_bw.bw == 0) {
6257 /* Re-check policy now with rq lock held: */
6258 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6259 policy = oldpolicy = -1;
6260 task_rq_unlock(rq, p, &rf);
6262 cpuset_read_unlock();
6267 * If setscheduling to SCHED_DEADLINE (or changing the parameters
6268 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
6271 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
6276 p->sched_reset_on_fork = reset_on_fork;
6281 * Take priority boosted tasks into account. If the new
6282 * effective priority is unchanged, we just store the new
6283 * normal parameters and do not touch the scheduler class and
6284 * the runqueue. This will be done when the task deboost
6287 new_effective_prio = rt_effective_prio(p, newprio);
6288 if (new_effective_prio == oldprio)
6289 queue_flags &= ~DEQUEUE_MOVE;
6292 queued = task_on_rq_queued(p);
6293 running = task_current(rq, p);
6295 dequeue_task(rq, p, queue_flags);
6297 put_prev_task(rq, p);
6299 prev_class = p->sched_class;
6301 __setscheduler(rq, p, attr, pi);
6302 __setscheduler_uclamp(p, attr);
6306 * We enqueue to tail when the priority of a task is
6307 * increased (user space view).
6309 if (oldprio < p->prio)
6310 queue_flags |= ENQUEUE_HEAD;
6312 enqueue_task(rq, p, queue_flags);
6315 set_next_task(rq, p);
6317 check_class_changed(rq, p, prev_class, oldprio);
6319 /* Avoid rq from going away on us: */
6321 head = splice_balance_callbacks(rq);
6322 task_rq_unlock(rq, p, &rf);
6325 cpuset_read_unlock();
6326 rt_mutex_adjust_pi(p);
6329 /* Run balance callbacks after we've adjusted the PI chain: */
6330 balance_callbacks(rq, head);
6336 task_rq_unlock(rq, p, &rf);
6338 cpuset_read_unlock();
6342 static int _sched_setscheduler(struct task_struct *p, int policy,
6343 const struct sched_param *param, bool check)
6345 struct sched_attr attr = {
6346 .sched_policy = policy,
6347 .sched_priority = param->sched_priority,
6348 .sched_nice = PRIO_TO_NICE(p->static_prio),
6351 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
6352 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
6353 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
6354 policy &= ~SCHED_RESET_ON_FORK;
6355 attr.sched_policy = policy;
6358 return __sched_setscheduler(p, &attr, check, true);
6361 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6362 * @p: the task in question.
6363 * @policy: new policy.
6364 * @param: structure containing the new RT priority.
6366 * Use sched_set_fifo(), read its comment.
6368 * Return: 0 on success. An error code otherwise.
6370 * NOTE that the task may be already dead.
6372 int sched_setscheduler(struct task_struct *p, int policy,
6373 const struct sched_param *param)
6375 return _sched_setscheduler(p, policy, param, true);
6378 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
6380 return __sched_setscheduler(p, attr, true, true);
6383 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
6385 return __sched_setscheduler(p, attr, false, true);
6389 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6390 * @p: the task in question.
6391 * @policy: new policy.
6392 * @param: structure containing the new RT priority.
6394 * Just like sched_setscheduler, only don't bother checking if the
6395 * current context has permission. For example, this is needed in
6396 * stop_machine(): we create temporary high priority worker threads,
6397 * but our caller might not have that capability.
6399 * Return: 0 on success. An error code otherwise.
6401 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6402 const struct sched_param *param)
6404 return _sched_setscheduler(p, policy, param, false);
6408 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
6409 * incapable of resource management, which is the one thing an OS really should
6412 * This is of course the reason it is limited to privileged users only.
6414 * Worse still; it is fundamentally impossible to compose static priority
6415 * workloads. You cannot take two correctly working static prio workloads
6416 * and smash them together and still expect them to work.
6418 * For this reason 'all' FIFO tasks the kernel creates are basically at:
6422 * The administrator _MUST_ configure the system, the kernel simply doesn't
6423 * know enough information to make a sensible choice.
6425 void sched_set_fifo(struct task_struct *p)
6427 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
6428 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
6430 EXPORT_SYMBOL_GPL(sched_set_fifo);
6433 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
6435 void sched_set_fifo_low(struct task_struct *p)
6437 struct sched_param sp = { .sched_priority = 1 };
6438 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
6440 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
6442 void sched_set_normal(struct task_struct *p, int nice)
6444 struct sched_attr attr = {
6445 .sched_policy = SCHED_NORMAL,
6448 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
6450 EXPORT_SYMBOL_GPL(sched_set_normal);
6453 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6455 struct sched_param lparam;
6456 struct task_struct *p;
6459 if (!param || pid < 0)
6461 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6466 p = find_process_by_pid(pid);
6472 retval = sched_setscheduler(p, policy, &lparam);
6480 * Mimics kernel/events/core.c perf_copy_attr().
6482 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
6487 /* Zero the full structure, so that a short copy will be nice: */
6488 memset(attr, 0, sizeof(*attr));
6490 ret = get_user(size, &uattr->size);
6494 /* ABI compatibility quirk: */
6496 size = SCHED_ATTR_SIZE_VER0;
6497 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
6500 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
6507 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
6508 size < SCHED_ATTR_SIZE_VER1)
6512 * XXX: Do we want to be lenient like existing syscalls; or do we want
6513 * to be strict and return an error on out-of-bounds values?
6515 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
6520 put_user(sizeof(*attr), &uattr->size);
6525 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6526 * @pid: the pid in question.
6527 * @policy: new policy.
6528 * @param: structure containing the new RT priority.
6530 * Return: 0 on success. An error code otherwise.
6532 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
6537 return do_sched_setscheduler(pid, policy, param);
6541 * sys_sched_setparam - set/change the RT priority of a thread
6542 * @pid: the pid in question.
6543 * @param: structure containing the new RT priority.
6545 * Return: 0 on success. An error code otherwise.
6547 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6549 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
6553 * sys_sched_setattr - same as above, but with extended sched_attr
6554 * @pid: the pid in question.
6555 * @uattr: structure containing the extended parameters.
6556 * @flags: for future extension.
6558 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
6559 unsigned int, flags)
6561 struct sched_attr attr;
6562 struct task_struct *p;
6565 if (!uattr || pid < 0 || flags)
6568 retval = sched_copy_attr(uattr, &attr);
6572 if ((int)attr.sched_policy < 0)
6574 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
6575 attr.sched_policy = SETPARAM_POLICY;
6579 p = find_process_by_pid(pid);
6585 retval = sched_setattr(p, &attr);
6593 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6594 * @pid: the pid in question.
6596 * Return: On success, the policy of the thread. Otherwise, a negative error
6599 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6601 struct task_struct *p;
6609 p = find_process_by_pid(pid);
6611 retval = security_task_getscheduler(p);
6614 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6621 * sys_sched_getparam - get the RT priority of a thread
6622 * @pid: the pid in question.
6623 * @param: structure containing the RT priority.
6625 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
6628 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6630 struct sched_param lp = { .sched_priority = 0 };
6631 struct task_struct *p;
6634 if (!param || pid < 0)
6638 p = find_process_by_pid(pid);
6643 retval = security_task_getscheduler(p);
6647 if (task_has_rt_policy(p))
6648 lp.sched_priority = p->rt_priority;
6652 * This one might sleep, we cannot do it with a spinlock held ...
6654 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6664 * Copy the kernel size attribute structure (which might be larger
6665 * than what user-space knows about) to user-space.
6667 * Note that all cases are valid: user-space buffer can be larger or
6668 * smaller than the kernel-space buffer. The usual case is that both
6669 * have the same size.
6672 sched_attr_copy_to_user(struct sched_attr __user *uattr,
6673 struct sched_attr *kattr,
6676 unsigned int ksize = sizeof(*kattr);
6678 if (!access_ok(uattr, usize))
6682 * sched_getattr() ABI forwards and backwards compatibility:
6684 * If usize == ksize then we just copy everything to user-space and all is good.
6686 * If usize < ksize then we only copy as much as user-space has space for,
6687 * this keeps ABI compatibility as well. We skip the rest.
6689 * If usize > ksize then user-space is using a newer version of the ABI,
6690 * which part the kernel doesn't know about. Just ignore it - tooling can
6691 * detect the kernel's knowledge of attributes from the attr->size value
6692 * which is set to ksize in this case.
6694 kattr->size = min(usize, ksize);
6696 if (copy_to_user(uattr, kattr, kattr->size))
6703 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
6704 * @pid: the pid in question.
6705 * @uattr: structure containing the extended parameters.
6706 * @usize: sizeof(attr) for fwd/bwd comp.
6707 * @flags: for future extension.
6709 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
6710 unsigned int, usize, unsigned int, flags)
6712 struct sched_attr kattr = { };
6713 struct task_struct *p;
6716 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
6717 usize < SCHED_ATTR_SIZE_VER0 || flags)
6721 p = find_process_by_pid(pid);
6726 retval = security_task_getscheduler(p);
6730 kattr.sched_policy = p->policy;
6731 if (p->sched_reset_on_fork)
6732 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
6733 if (task_has_dl_policy(p))
6734 __getparam_dl(p, &kattr);
6735 else if (task_has_rt_policy(p))
6736 kattr.sched_priority = p->rt_priority;
6738 kattr.sched_nice = task_nice(p);
6740 #ifdef CONFIG_UCLAMP_TASK
6742 * This could race with another potential updater, but this is fine
6743 * because it'll correctly read the old or the new value. We don't need
6744 * to guarantee who wins the race as long as it doesn't return garbage.
6746 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
6747 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
6752 return sched_attr_copy_to_user(uattr, &kattr, usize);
6759 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6761 cpumask_var_t cpus_allowed, new_mask;
6762 struct task_struct *p;
6767 p = find_process_by_pid(pid);
6773 /* Prevent p going away */
6777 if (p->flags & PF_NO_SETAFFINITY) {
6781 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6785 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6787 goto out_free_cpus_allowed;
6790 if (!check_same_owner(p)) {
6792 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
6794 goto out_free_new_mask;
6799 retval = security_task_setscheduler(p);
6801 goto out_free_new_mask;
6804 cpuset_cpus_allowed(p, cpus_allowed);
6805 cpumask_and(new_mask, in_mask, cpus_allowed);
6808 * Since bandwidth control happens on root_domain basis,
6809 * if admission test is enabled, we only admit -deadline
6810 * tasks allowed to run on all the CPUs in the task's
6814 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
6816 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
6819 goto out_free_new_mask;
6825 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK);
6828 cpuset_cpus_allowed(p, cpus_allowed);
6829 if (!cpumask_subset(new_mask, cpus_allowed)) {
6831 * We must have raced with a concurrent cpuset
6832 * update. Just reset the cpus_allowed to the
6833 * cpuset's cpus_allowed
6835 cpumask_copy(new_mask, cpus_allowed);
6840 free_cpumask_var(new_mask);
6841 out_free_cpus_allowed:
6842 free_cpumask_var(cpus_allowed);
6848 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6849 struct cpumask *new_mask)
6851 if (len < cpumask_size())
6852 cpumask_clear(new_mask);
6853 else if (len > cpumask_size())
6854 len = cpumask_size();
6856 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6860 * sys_sched_setaffinity - set the CPU affinity of a process
6861 * @pid: pid of the process
6862 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6863 * @user_mask_ptr: user-space pointer to the new CPU mask
6865 * Return: 0 on success. An error code otherwise.
6867 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6868 unsigned long __user *, user_mask_ptr)
6870 cpumask_var_t new_mask;
6873 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6876 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6878 retval = sched_setaffinity(pid, new_mask);
6879 free_cpumask_var(new_mask);
6883 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6885 struct task_struct *p;
6886 unsigned long flags;
6892 p = find_process_by_pid(pid);
6896 retval = security_task_getscheduler(p);
6900 raw_spin_lock_irqsave(&p->pi_lock, flags);
6901 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
6902 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6911 * sys_sched_getaffinity - get the CPU affinity of a process
6912 * @pid: pid of the process
6913 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6914 * @user_mask_ptr: user-space pointer to hold the current CPU mask
6916 * Return: size of CPU mask copied to user_mask_ptr on success. An
6917 * error code otherwise.
6919 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6920 unsigned long __user *, user_mask_ptr)
6925 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
6927 if (len & (sizeof(unsigned long)-1))
6930 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6933 ret = sched_getaffinity(pid, mask);
6935 unsigned int retlen = min(len, cpumask_size());
6937 if (copy_to_user(user_mask_ptr, mask, retlen))
6942 free_cpumask_var(mask);
6947 static void do_sched_yield(void)
6952 rq = this_rq_lock_irq(&rf);
6954 schedstat_inc(rq->yld_count);
6955 current->sched_class->yield_task(rq);
6958 rq_unlock_irq(rq, &rf);
6959 sched_preempt_enable_no_resched();
6965 * sys_sched_yield - yield the current processor to other threads.
6967 * This function yields the current CPU to other tasks. If there are no
6968 * other threads running on this CPU then this function will return.
6972 SYSCALL_DEFINE0(sched_yield)
6978 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
6979 int __sched __cond_resched(void)
6981 if (should_resched(0)) {
6982 preempt_schedule_common();
6985 #ifndef CONFIG_PREEMPT_RCU
6990 EXPORT_SYMBOL(__cond_resched);
6993 #ifdef CONFIG_PREEMPT_DYNAMIC
6994 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
6995 EXPORT_STATIC_CALL_TRAMP(cond_resched);
6997 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
6998 EXPORT_STATIC_CALL_TRAMP(might_resched);
7002 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
7003 * call schedule, and on return reacquire the lock.
7005 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
7006 * operations here to prevent schedule() from being called twice (once via
7007 * spin_unlock(), once by hand).
7009 int __cond_resched_lock(spinlock_t *lock)
7011 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7014 lockdep_assert_held(lock);
7016 if (spin_needbreak(lock) || resched) {
7019 preempt_schedule_common();
7027 EXPORT_SYMBOL(__cond_resched_lock);
7029 int __cond_resched_rwlock_read(rwlock_t *lock)
7031 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7034 lockdep_assert_held_read(lock);
7036 if (rwlock_needbreak(lock) || resched) {
7039 preempt_schedule_common();
7047 EXPORT_SYMBOL(__cond_resched_rwlock_read);
7049 int __cond_resched_rwlock_write(rwlock_t *lock)
7051 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7054 lockdep_assert_held_write(lock);
7056 if (rwlock_needbreak(lock) || resched) {
7059 preempt_schedule_common();
7067 EXPORT_SYMBOL(__cond_resched_rwlock_write);
7070 * yield - yield the current processor to other threads.
7072 * Do not ever use this function, there's a 99% chance you're doing it wrong.
7074 * The scheduler is at all times free to pick the calling task as the most
7075 * eligible task to run, if removing the yield() call from your code breaks
7076 * it, it's already broken.
7078 * Typical broken usage is:
7083 * where one assumes that yield() will let 'the other' process run that will
7084 * make event true. If the current task is a SCHED_FIFO task that will never
7085 * happen. Never use yield() as a progress guarantee!!
7087 * If you want to use yield() to wait for something, use wait_event().
7088 * If you want to use yield() to be 'nice' for others, use cond_resched().
7089 * If you still want to use yield(), do not!
7091 void __sched yield(void)
7093 set_current_state(TASK_RUNNING);
7096 EXPORT_SYMBOL(yield);
7099 * yield_to - yield the current processor to another thread in
7100 * your thread group, or accelerate that thread toward the
7101 * processor it's on.
7103 * @preempt: whether task preemption is allowed or not
7105 * It's the caller's job to ensure that the target task struct
7106 * can't go away on us before we can do any checks.
7109 * true (>0) if we indeed boosted the target task.
7110 * false (0) if we failed to boost the target.
7111 * -ESRCH if there's no task to yield to.
7113 int __sched yield_to(struct task_struct *p, bool preempt)
7115 struct task_struct *curr = current;
7116 struct rq *rq, *p_rq;
7117 unsigned long flags;
7120 local_irq_save(flags);
7126 * If we're the only runnable task on the rq and target rq also
7127 * has only one task, there's absolutely no point in yielding.
7129 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
7134 double_rq_lock(rq, p_rq);
7135 if (task_rq(p) != p_rq) {
7136 double_rq_unlock(rq, p_rq);
7140 if (!curr->sched_class->yield_to_task)
7143 if (curr->sched_class != p->sched_class)
7146 if (task_running(p_rq, p) || p->state)
7149 yielded = curr->sched_class->yield_to_task(rq, p);
7151 schedstat_inc(rq->yld_count);
7153 * Make p's CPU reschedule; pick_next_entity takes care of
7156 if (preempt && rq != p_rq)
7161 double_rq_unlock(rq, p_rq);
7163 local_irq_restore(flags);
7170 EXPORT_SYMBOL_GPL(yield_to);
7172 int io_schedule_prepare(void)
7174 int old_iowait = current->in_iowait;
7176 current->in_iowait = 1;
7177 blk_schedule_flush_plug(current);
7182 void io_schedule_finish(int token)
7184 current->in_iowait = token;
7188 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
7189 * that process accounting knows that this is a task in IO wait state.
7191 long __sched io_schedule_timeout(long timeout)
7196 token = io_schedule_prepare();
7197 ret = schedule_timeout(timeout);
7198 io_schedule_finish(token);
7202 EXPORT_SYMBOL(io_schedule_timeout);
7204 void __sched io_schedule(void)
7208 token = io_schedule_prepare();
7210 io_schedule_finish(token);
7212 EXPORT_SYMBOL(io_schedule);
7215 * sys_sched_get_priority_max - return maximum RT priority.
7216 * @policy: scheduling class.
7218 * Return: On success, this syscall returns the maximum
7219 * rt_priority that can be used by a given scheduling class.
7220 * On failure, a negative error code is returned.
7222 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
7229 ret = MAX_RT_PRIO-1;
7231 case SCHED_DEADLINE:
7242 * sys_sched_get_priority_min - return minimum RT priority.
7243 * @policy: scheduling class.
7245 * Return: On success, this syscall returns the minimum
7246 * rt_priority that can be used by a given scheduling class.
7247 * On failure, a negative error code is returned.
7249 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
7258 case SCHED_DEADLINE:
7267 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
7269 struct task_struct *p;
7270 unsigned int time_slice;
7280 p = find_process_by_pid(pid);
7284 retval = security_task_getscheduler(p);
7288 rq = task_rq_lock(p, &rf);
7290 if (p->sched_class->get_rr_interval)
7291 time_slice = p->sched_class->get_rr_interval(rq, p);
7292 task_rq_unlock(rq, p, &rf);
7295 jiffies_to_timespec64(time_slice, t);
7304 * sys_sched_rr_get_interval - return the default timeslice of a process.
7305 * @pid: pid of the process.
7306 * @interval: userspace pointer to the timeslice value.
7308 * this syscall writes the default timeslice value of a given process
7309 * into the user-space timespec buffer. A value of '0' means infinity.
7311 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
7314 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
7315 struct __kernel_timespec __user *, interval)
7317 struct timespec64 t;
7318 int retval = sched_rr_get_interval(pid, &t);
7321 retval = put_timespec64(&t, interval);
7326 #ifdef CONFIG_COMPAT_32BIT_TIME
7327 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
7328 struct old_timespec32 __user *, interval)
7330 struct timespec64 t;
7331 int retval = sched_rr_get_interval(pid, &t);
7334 retval = put_old_timespec32(&t, interval);
7339 void sched_show_task(struct task_struct *p)
7341 unsigned long free = 0;
7344 if (!try_get_task_stack(p))
7347 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
7349 if (p->state == TASK_RUNNING)
7350 pr_cont(" running task ");
7351 #ifdef CONFIG_DEBUG_STACK_USAGE
7352 free = stack_not_used(p);
7357 ppid = task_pid_nr(rcu_dereference(p->real_parent));
7359 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
7360 free, task_pid_nr(p), ppid,
7361 (unsigned long)task_thread_info(p)->flags);
7363 print_worker_info(KERN_INFO, p);
7364 print_stop_info(KERN_INFO, p);
7365 show_stack(p, NULL, KERN_INFO);
7368 EXPORT_SYMBOL_GPL(sched_show_task);
7371 state_filter_match(unsigned long state_filter, struct task_struct *p)
7373 /* no filter, everything matches */
7377 /* filter, but doesn't match */
7378 if (!(p->state & state_filter))
7382 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
7385 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
7392 void show_state_filter(unsigned long state_filter)
7394 struct task_struct *g, *p;
7397 for_each_process_thread(g, p) {
7399 * reset the NMI-timeout, listing all files on a slow
7400 * console might take a lot of time:
7401 * Also, reset softlockup watchdogs on all CPUs, because
7402 * another CPU might be blocked waiting for us to process
7405 touch_nmi_watchdog();
7406 touch_all_softlockup_watchdogs();
7407 if (state_filter_match(state_filter, p))
7411 #ifdef CONFIG_SCHED_DEBUG
7413 sysrq_sched_debug_show();
7417 * Only show locks if all tasks are dumped:
7420 debug_show_all_locks();
7424 * init_idle - set up an idle thread for a given CPU
7425 * @idle: task in question
7426 * @cpu: CPU the idle task belongs to
7428 * NOTE: this function does not set the idle thread's NEED_RESCHED
7429 * flag, to make booting more robust.
7431 void init_idle(struct task_struct *idle, int cpu)
7433 struct rq *rq = cpu_rq(cpu);
7434 unsigned long flags;
7436 __sched_fork(0, idle);
7438 raw_spin_lock_irqsave(&idle->pi_lock, flags);
7439 raw_spin_lock(&rq->lock);
7441 idle->state = TASK_RUNNING;
7442 idle->se.exec_start = sched_clock();
7443 idle->flags |= PF_IDLE;
7445 scs_task_reset(idle);
7446 kasan_unpoison_task_stack(idle);
7450 * It's possible that init_idle() gets called multiple times on a task,
7451 * in that case do_set_cpus_allowed() will not do the right thing.
7453 * And since this is boot we can forgo the serialization.
7455 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
7458 * We're having a chicken and egg problem, even though we are
7459 * holding rq->lock, the CPU isn't yet set to this CPU so the
7460 * lockdep check in task_group() will fail.
7462 * Similar case to sched_fork(). / Alternatively we could
7463 * use task_rq_lock() here and obtain the other rq->lock.
7468 __set_task_cpu(idle, cpu);
7472 rcu_assign_pointer(rq->curr, idle);
7473 idle->on_rq = TASK_ON_RQ_QUEUED;
7477 raw_spin_unlock(&rq->lock);
7478 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
7480 /* Set the preempt count _outside_ the spinlocks! */
7481 init_idle_preempt_count(idle, cpu);
7484 * The idle tasks have their own, simple scheduling class:
7486 idle->sched_class = &idle_sched_class;
7487 ftrace_graph_init_idle_task(idle, cpu);
7488 vtime_init_idle(idle, cpu);
7490 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
7496 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
7497 const struct cpumask *trial)
7501 if (!cpumask_weight(cur))
7504 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
7509 int task_can_attach(struct task_struct *p,
7510 const struct cpumask *cs_cpus_allowed)
7515 * Kthreads which disallow setaffinity shouldn't be moved
7516 * to a new cpuset; we don't want to change their CPU
7517 * affinity and isolating such threads by their set of
7518 * allowed nodes is unnecessary. Thus, cpusets are not
7519 * applicable for such threads. This prevents checking for
7520 * success of set_cpus_allowed_ptr() on all attached tasks
7521 * before cpus_mask may be changed.
7523 if (p->flags & PF_NO_SETAFFINITY) {
7528 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
7530 ret = dl_task_can_attach(p, cs_cpus_allowed);
7536 bool sched_smp_initialized __read_mostly;
7538 #ifdef CONFIG_NUMA_BALANCING
7539 /* Migrate current task p to target_cpu */
7540 int migrate_task_to(struct task_struct *p, int target_cpu)
7542 struct migration_arg arg = { p, target_cpu };
7543 int curr_cpu = task_cpu(p);
7545 if (curr_cpu == target_cpu)
7548 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
7551 /* TODO: This is not properly updating schedstats */
7553 trace_sched_move_numa(p, curr_cpu, target_cpu);
7554 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
7558 * Requeue a task on a given node and accurately track the number of NUMA
7559 * tasks on the runqueues
7561 void sched_setnuma(struct task_struct *p, int nid)
7563 bool queued, running;
7567 rq = task_rq_lock(p, &rf);
7568 queued = task_on_rq_queued(p);
7569 running = task_current(rq, p);
7572 dequeue_task(rq, p, DEQUEUE_SAVE);
7574 put_prev_task(rq, p);
7576 p->numa_preferred_nid = nid;
7579 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7581 set_next_task(rq, p);
7582 task_rq_unlock(rq, p, &rf);
7584 #endif /* CONFIG_NUMA_BALANCING */
7586 #ifdef CONFIG_HOTPLUG_CPU
7588 * Ensure that the idle task is using init_mm right before its CPU goes
7591 void idle_task_exit(void)
7593 struct mm_struct *mm = current->active_mm;
7595 BUG_ON(cpu_online(smp_processor_id()));
7596 BUG_ON(current != this_rq()->idle);
7598 if (mm != &init_mm) {
7599 switch_mm(mm, &init_mm, current);
7600 finish_arch_post_lock_switch();
7603 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
7606 static int __balance_push_cpu_stop(void *arg)
7608 struct task_struct *p = arg;
7609 struct rq *rq = this_rq();
7613 raw_spin_lock_irq(&p->pi_lock);
7616 update_rq_clock(rq);
7618 if (task_rq(p) == rq && task_on_rq_queued(p)) {
7619 cpu = select_fallback_rq(rq->cpu, p);
7620 rq = __migrate_task(rq, &rf, p, cpu);
7624 raw_spin_unlock_irq(&p->pi_lock);
7631 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
7634 * Ensure we only run per-cpu kthreads once the CPU goes !active.
7636 static void balance_push(struct rq *rq)
7638 struct task_struct *push_task = rq->curr;
7640 lockdep_assert_held(&rq->lock);
7641 SCHED_WARN_ON(rq->cpu != smp_processor_id());
7643 * Ensure the thing is persistent until balance_push_set(.on = false);
7645 rq->balance_callback = &balance_push_callback;
7648 * Both the cpu-hotplug and stop task are in this case and are
7649 * required to complete the hotplug process.
7651 * XXX: the idle task does not match kthread_is_per_cpu() due to
7652 * histerical raisins.
7654 if (rq->idle == push_task ||
7655 ((push_task->flags & PF_KTHREAD) && kthread_is_per_cpu(push_task)) ||
7656 is_migration_disabled(push_task)) {
7659 * If this is the idle task on the outgoing CPU try to wake
7660 * up the hotplug control thread which might wait for the
7661 * last task to vanish. The rcuwait_active() check is
7662 * accurate here because the waiter is pinned on this CPU
7663 * and can't obviously be running in parallel.
7665 * On RT kernels this also has to check whether there are
7666 * pinned and scheduled out tasks on the runqueue. They
7667 * need to leave the migrate disabled section first.
7669 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
7670 rcuwait_active(&rq->hotplug_wait)) {
7671 raw_spin_unlock(&rq->lock);
7672 rcuwait_wake_up(&rq->hotplug_wait);
7673 raw_spin_lock(&rq->lock);
7678 get_task_struct(push_task);
7680 * Temporarily drop rq->lock such that we can wake-up the stop task.
7681 * Both preemption and IRQs are still disabled.
7683 raw_spin_unlock(&rq->lock);
7684 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
7685 this_cpu_ptr(&push_work));
7687 * At this point need_resched() is true and we'll take the loop in
7688 * schedule(). The next pick is obviously going to be the stop task
7689 * which kthread_is_per_cpu() and will push this task away.
7691 raw_spin_lock(&rq->lock);
7694 static void balance_push_set(int cpu, bool on)
7696 struct rq *rq = cpu_rq(cpu);
7699 rq_lock_irqsave(rq, &rf);
7700 rq->balance_push = on;
7702 WARN_ON_ONCE(rq->balance_callback);
7703 rq->balance_callback = &balance_push_callback;
7704 } else if (rq->balance_callback == &balance_push_callback) {
7705 rq->balance_callback = NULL;
7707 rq_unlock_irqrestore(rq, &rf);
7711 * Invoked from a CPUs hotplug control thread after the CPU has been marked
7712 * inactive. All tasks which are not per CPU kernel threads are either
7713 * pushed off this CPU now via balance_push() or placed on a different CPU
7714 * during wakeup. Wait until the CPU is quiescent.
7716 static void balance_hotplug_wait(void)
7718 struct rq *rq = this_rq();
7720 rcuwait_wait_event(&rq->hotplug_wait,
7721 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
7722 TASK_UNINTERRUPTIBLE);
7727 static inline void balance_push(struct rq *rq)
7731 static inline void balance_push_set(int cpu, bool on)
7735 static inline void balance_hotplug_wait(void)
7739 #endif /* CONFIG_HOTPLUG_CPU */
7741 void set_rq_online(struct rq *rq)
7744 const struct sched_class *class;
7746 cpumask_set_cpu(rq->cpu, rq->rd->online);
7749 for_each_class(class) {
7750 if (class->rq_online)
7751 class->rq_online(rq);
7756 void set_rq_offline(struct rq *rq)
7759 const struct sched_class *class;
7761 for_each_class(class) {
7762 if (class->rq_offline)
7763 class->rq_offline(rq);
7766 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7772 * used to mark begin/end of suspend/resume:
7774 static int num_cpus_frozen;
7777 * Update cpusets according to cpu_active mask. If cpusets are
7778 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7779 * around partition_sched_domains().
7781 * If we come here as part of a suspend/resume, don't touch cpusets because we
7782 * want to restore it back to its original state upon resume anyway.
7784 static void cpuset_cpu_active(void)
7786 if (cpuhp_tasks_frozen) {
7788 * num_cpus_frozen tracks how many CPUs are involved in suspend
7789 * resume sequence. As long as this is not the last online
7790 * operation in the resume sequence, just build a single sched
7791 * domain, ignoring cpusets.
7793 partition_sched_domains(1, NULL, NULL);
7794 if (--num_cpus_frozen)
7797 * This is the last CPU online operation. So fall through and
7798 * restore the original sched domains by considering the
7799 * cpuset configurations.
7801 cpuset_force_rebuild();
7803 cpuset_update_active_cpus();
7806 static int cpuset_cpu_inactive(unsigned int cpu)
7808 if (!cpuhp_tasks_frozen) {
7809 if (dl_cpu_busy(cpu))
7811 cpuset_update_active_cpus();
7814 partition_sched_domains(1, NULL, NULL);
7819 int sched_cpu_activate(unsigned int cpu)
7821 struct rq *rq = cpu_rq(cpu);
7825 * Make sure that when the hotplug state machine does a roll-back
7826 * we clear balance_push. Ideally that would happen earlier...
7828 balance_push_set(cpu, false);
7830 #ifdef CONFIG_SCHED_SMT
7832 * When going up, increment the number of cores with SMT present.
7834 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
7835 static_branch_inc_cpuslocked(&sched_smt_present);
7837 set_cpu_active(cpu, true);
7839 if (sched_smp_initialized) {
7840 sched_domains_numa_masks_set(cpu);
7841 cpuset_cpu_active();
7845 * Put the rq online, if not already. This happens:
7847 * 1) In the early boot process, because we build the real domains
7848 * after all CPUs have been brought up.
7850 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7853 rq_lock_irqsave(rq, &rf);
7855 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7858 rq_unlock_irqrestore(rq, &rf);
7863 int sched_cpu_deactivate(unsigned int cpu)
7865 struct rq *rq = cpu_rq(cpu);
7870 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
7871 * load balancing when not active
7873 nohz_balance_exit_idle(rq);
7875 set_cpu_active(cpu, false);
7878 * From this point forward, this CPU will refuse to run any task that
7879 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
7880 * push those tasks away until this gets cleared, see
7881 * sched_cpu_dying().
7883 balance_push_set(cpu, true);
7886 * We've cleared cpu_active_mask / set balance_push, wait for all
7887 * preempt-disabled and RCU users of this state to go away such that
7888 * all new such users will observe it.
7890 * Specifically, we rely on ttwu to no longer target this CPU, see
7891 * ttwu_queue_cond() and is_cpu_allowed().
7893 * Do sync before park smpboot threads to take care the rcu boost case.
7897 rq_lock_irqsave(rq, &rf);
7899 update_rq_clock(rq);
7900 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7903 rq_unlock_irqrestore(rq, &rf);
7905 #ifdef CONFIG_SCHED_SMT
7907 * When going down, decrement the number of cores with SMT present.
7909 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
7910 static_branch_dec_cpuslocked(&sched_smt_present);
7913 if (!sched_smp_initialized)
7916 ret = cpuset_cpu_inactive(cpu);
7918 balance_push_set(cpu, false);
7919 set_cpu_active(cpu, true);
7922 sched_domains_numa_masks_clear(cpu);
7926 static void sched_rq_cpu_starting(unsigned int cpu)
7928 struct rq *rq = cpu_rq(cpu);
7930 rq->calc_load_update = calc_load_update;
7931 update_max_interval();
7934 int sched_cpu_starting(unsigned int cpu)
7936 sched_rq_cpu_starting(cpu);
7937 sched_tick_start(cpu);
7941 #ifdef CONFIG_HOTPLUG_CPU
7944 * Invoked immediately before the stopper thread is invoked to bring the
7945 * CPU down completely. At this point all per CPU kthreads except the
7946 * hotplug thread (current) and the stopper thread (inactive) have been
7947 * either parked or have been unbound from the outgoing CPU. Ensure that
7948 * any of those which might be on the way out are gone.
7950 * If after this point a bound task is being woken on this CPU then the
7951 * responsible hotplug callback has failed to do it's job.
7952 * sched_cpu_dying() will catch it with the appropriate fireworks.
7954 int sched_cpu_wait_empty(unsigned int cpu)
7956 balance_hotplug_wait();
7961 * Since this CPU is going 'away' for a while, fold any nr_active delta we
7962 * might have. Called from the CPU stopper task after ensuring that the
7963 * stopper is the last running task on the CPU, so nr_active count is
7964 * stable. We need to take the teardown thread which is calling this into
7965 * account, so we hand in adjust = 1 to the load calculation.
7967 * Also see the comment "Global load-average calculations".
7969 static void calc_load_migrate(struct rq *rq)
7971 long delta = calc_load_fold_active(rq, 1);
7974 atomic_long_add(delta, &calc_load_tasks);
7977 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
7979 struct task_struct *g, *p;
7980 int cpu = cpu_of(rq);
7982 lockdep_assert_held(&rq->lock);
7984 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
7985 for_each_process_thread(g, p) {
7986 if (task_cpu(p) != cpu)
7989 if (!task_on_rq_queued(p))
7992 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
7996 int sched_cpu_dying(unsigned int cpu)
7998 struct rq *rq = cpu_rq(cpu);
8001 /* Handle pending wakeups and then migrate everything off */
8002 sched_tick_stop(cpu);
8004 rq_lock_irqsave(rq, &rf);
8005 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
8006 WARN(true, "Dying CPU not properly vacated!");
8007 dump_rq_tasks(rq, KERN_WARNING);
8009 rq_unlock_irqrestore(rq, &rf);
8012 * Now that the CPU is offline, make sure we're welcome
8013 * to new tasks once we come back up.
8015 balance_push_set(cpu, false);
8017 calc_load_migrate(rq);
8018 update_max_interval();
8024 void __init sched_init_smp(void)
8029 * There's no userspace yet to cause hotplug operations; hence all the
8030 * CPU masks are stable and all blatant races in the below code cannot
8033 mutex_lock(&sched_domains_mutex);
8034 sched_init_domains(cpu_active_mask);
8035 mutex_unlock(&sched_domains_mutex);
8037 /* Move init over to a non-isolated CPU */
8038 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
8040 sched_init_granularity();
8042 init_sched_rt_class();
8043 init_sched_dl_class();
8045 sched_smp_initialized = true;
8048 static int __init migration_init(void)
8050 sched_cpu_starting(smp_processor_id());
8053 early_initcall(migration_init);
8056 void __init sched_init_smp(void)
8058 sched_init_granularity();
8060 #endif /* CONFIG_SMP */
8062 int in_sched_functions(unsigned long addr)
8064 return in_lock_functions(addr) ||
8065 (addr >= (unsigned long)__sched_text_start
8066 && addr < (unsigned long)__sched_text_end);
8069 #ifdef CONFIG_CGROUP_SCHED
8071 * Default task group.
8072 * Every task in system belongs to this group at bootup.
8074 struct task_group root_task_group;
8075 LIST_HEAD(task_groups);
8077 /* Cacheline aligned slab cache for task_group */
8078 static struct kmem_cache *task_group_cache __read_mostly;
8081 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
8082 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
8084 void __init sched_init(void)
8086 unsigned long ptr = 0;
8089 /* Make sure the linker didn't screw up */
8090 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
8091 &fair_sched_class + 1 != &rt_sched_class ||
8092 &rt_sched_class + 1 != &dl_sched_class);
8094 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
8099 #ifdef CONFIG_FAIR_GROUP_SCHED
8100 ptr += 2 * nr_cpu_ids * sizeof(void **);
8102 #ifdef CONFIG_RT_GROUP_SCHED
8103 ptr += 2 * nr_cpu_ids * sizeof(void **);
8106 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
8108 #ifdef CONFIG_FAIR_GROUP_SCHED
8109 root_task_group.se = (struct sched_entity **)ptr;
8110 ptr += nr_cpu_ids * sizeof(void **);
8112 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8113 ptr += nr_cpu_ids * sizeof(void **);
8115 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
8116 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8117 #endif /* CONFIG_FAIR_GROUP_SCHED */
8118 #ifdef CONFIG_RT_GROUP_SCHED
8119 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8120 ptr += nr_cpu_ids * sizeof(void **);
8122 root_task_group.rt_rq = (struct rt_rq **)ptr;
8123 ptr += nr_cpu_ids * sizeof(void **);
8125 #endif /* CONFIG_RT_GROUP_SCHED */
8127 #ifdef CONFIG_CPUMASK_OFFSTACK
8128 for_each_possible_cpu(i) {
8129 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
8130 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8131 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
8132 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8134 #endif /* CONFIG_CPUMASK_OFFSTACK */
8136 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
8137 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
8140 init_defrootdomain();
8143 #ifdef CONFIG_RT_GROUP_SCHED
8144 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8145 global_rt_period(), global_rt_runtime());
8146 #endif /* CONFIG_RT_GROUP_SCHED */
8148 #ifdef CONFIG_CGROUP_SCHED
8149 task_group_cache = KMEM_CACHE(task_group, 0);
8151 list_add(&root_task_group.list, &task_groups);
8152 INIT_LIST_HEAD(&root_task_group.children);
8153 INIT_LIST_HEAD(&root_task_group.siblings);
8154 autogroup_init(&init_task);
8155 #endif /* CONFIG_CGROUP_SCHED */
8157 for_each_possible_cpu(i) {
8161 raw_spin_lock_init(&rq->lock);
8163 rq->calc_load_active = 0;
8164 rq->calc_load_update = jiffies + LOAD_FREQ;
8165 init_cfs_rq(&rq->cfs);
8166 init_rt_rq(&rq->rt);
8167 init_dl_rq(&rq->dl);
8168 #ifdef CONFIG_FAIR_GROUP_SCHED
8169 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8170 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
8172 * How much CPU bandwidth does root_task_group get?
8174 * In case of task-groups formed thr' the cgroup filesystem, it
8175 * gets 100% of the CPU resources in the system. This overall
8176 * system CPU resource is divided among the tasks of
8177 * root_task_group and its child task-groups in a fair manner,
8178 * based on each entity's (task or task-group's) weight
8179 * (se->load.weight).
8181 * In other words, if root_task_group has 10 tasks of weight
8182 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8183 * then A0's share of the CPU resource is:
8185 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8187 * We achieve this by letting root_task_group's tasks sit
8188 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8190 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8191 #endif /* CONFIG_FAIR_GROUP_SCHED */
8193 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8194 #ifdef CONFIG_RT_GROUP_SCHED
8195 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8200 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
8201 rq->balance_callback = NULL;
8202 rq->active_balance = 0;
8203 rq->next_balance = jiffies;
8208 rq->avg_idle = 2*sysctl_sched_migration_cost;
8209 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
8211 INIT_LIST_HEAD(&rq->cfs_tasks);
8213 rq_attach_root(rq, &def_root_domain);
8214 #ifdef CONFIG_NO_HZ_COMMON
8215 rq->last_blocked_load_update_tick = jiffies;
8216 atomic_set(&rq->nohz_flags, 0);
8218 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
8220 #ifdef CONFIG_HOTPLUG_CPU
8221 rcuwait_init(&rq->hotplug_wait);
8223 #endif /* CONFIG_SMP */
8225 atomic_set(&rq->nr_iowait, 0);
8228 set_load_weight(&init_task, false);
8231 * The boot idle thread does lazy MMU switching as well:
8234 enter_lazy_tlb(&init_mm, current);
8237 * Make us the idle thread. Technically, schedule() should not be
8238 * called from this thread, however somewhere below it might be,
8239 * but because we are the idle thread, we just pick up running again
8240 * when this runqueue becomes "idle".
8242 init_idle(current, smp_processor_id());
8244 calc_load_update = jiffies + LOAD_FREQ;
8247 idle_thread_set_boot_cpu();
8249 init_sched_fair_class();
8257 scheduler_running = 1;
8260 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8261 static inline int preempt_count_equals(int preempt_offset)
8263 int nested = preempt_count() + rcu_preempt_depth();
8265 return (nested == preempt_offset);
8268 void __might_sleep(const char *file, int line, int preempt_offset)
8271 * Blocking primitives will set (and therefore destroy) current->state,
8272 * since we will exit with TASK_RUNNING make sure we enter with it,
8273 * otherwise we will destroy state.
8275 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
8276 "do not call blocking ops when !TASK_RUNNING; "
8277 "state=%lx set at [<%p>] %pS\n",
8279 (void *)current->task_state_change,
8280 (void *)current->task_state_change);
8282 ___might_sleep(file, line, preempt_offset);
8284 EXPORT_SYMBOL(__might_sleep);
8286 void ___might_sleep(const char *file, int line, int preempt_offset)
8288 /* Ratelimiting timestamp: */
8289 static unsigned long prev_jiffy;
8291 unsigned long preempt_disable_ip;
8293 /* WARN_ON_ONCE() by default, no rate limit required: */
8296 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
8297 !is_idle_task(current) && !current->non_block_count) ||
8298 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
8302 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8304 prev_jiffy = jiffies;
8306 /* Save this before calling printk(), since that will clobber it: */
8307 preempt_disable_ip = get_preempt_disable_ip(current);
8310 "BUG: sleeping function called from invalid context at %s:%d\n",
8313 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
8314 in_atomic(), irqs_disabled(), current->non_block_count,
8315 current->pid, current->comm);
8317 if (task_stack_end_corrupted(current))
8318 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
8320 debug_show_held_locks(current);
8321 if (irqs_disabled())
8322 print_irqtrace_events(current);
8323 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
8324 && !preempt_count_equals(preempt_offset)) {
8325 pr_err("Preemption disabled at:");
8326 print_ip_sym(KERN_ERR, preempt_disable_ip);
8329 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
8331 EXPORT_SYMBOL(___might_sleep);
8333 void __cant_sleep(const char *file, int line, int preempt_offset)
8335 static unsigned long prev_jiffy;
8337 if (irqs_disabled())
8340 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
8343 if (preempt_count() > preempt_offset)
8346 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8348 prev_jiffy = jiffies;
8350 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
8351 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8352 in_atomic(), irqs_disabled(),
8353 current->pid, current->comm);
8355 debug_show_held_locks(current);
8357 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
8359 EXPORT_SYMBOL_GPL(__cant_sleep);
8362 void __cant_migrate(const char *file, int line)
8364 static unsigned long prev_jiffy;
8366 if (irqs_disabled())
8369 if (is_migration_disabled(current))
8372 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
8375 if (preempt_count() > 0)
8378 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8380 prev_jiffy = jiffies;
8382 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
8383 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
8384 in_atomic(), irqs_disabled(), is_migration_disabled(current),
8385 current->pid, current->comm);
8387 debug_show_held_locks(current);
8389 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
8391 EXPORT_SYMBOL_GPL(__cant_migrate);
8395 #ifdef CONFIG_MAGIC_SYSRQ
8396 void normalize_rt_tasks(void)
8398 struct task_struct *g, *p;
8399 struct sched_attr attr = {
8400 .sched_policy = SCHED_NORMAL,
8403 read_lock(&tasklist_lock);
8404 for_each_process_thread(g, p) {
8406 * Only normalize user tasks:
8408 if (p->flags & PF_KTHREAD)
8411 p->se.exec_start = 0;
8412 schedstat_set(p->se.statistics.wait_start, 0);
8413 schedstat_set(p->se.statistics.sleep_start, 0);
8414 schedstat_set(p->se.statistics.block_start, 0);
8416 if (!dl_task(p) && !rt_task(p)) {
8418 * Renice negative nice level userspace
8421 if (task_nice(p) < 0)
8422 set_user_nice(p, 0);
8426 __sched_setscheduler(p, &attr, false, false);
8428 read_unlock(&tasklist_lock);
8431 #endif /* CONFIG_MAGIC_SYSRQ */
8433 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8435 * These functions are only useful for the IA64 MCA handling, or kdb.
8437 * They can only be called when the whole system has been
8438 * stopped - every CPU needs to be quiescent, and no scheduling
8439 * activity can take place. Using them for anything else would
8440 * be a serious bug, and as a result, they aren't even visible
8441 * under any other configuration.
8445 * curr_task - return the current task for a given CPU.
8446 * @cpu: the processor in question.
8448 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8450 * Return: The current task for @cpu.
8452 struct task_struct *curr_task(int cpu)
8454 return cpu_curr(cpu);
8457 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8461 * ia64_set_curr_task - set the current task for a given CPU.
8462 * @cpu: the processor in question.
8463 * @p: the task pointer to set.
8465 * Description: This function must only be used when non-maskable interrupts
8466 * are serviced on a separate stack. It allows the architecture to switch the
8467 * notion of the current task on a CPU in a non-blocking manner. This function
8468 * must be called with all CPU's synchronized, and interrupts disabled, the
8469 * and caller must save the original value of the current task (see
8470 * curr_task() above) and restore that value before reenabling interrupts and
8471 * re-starting the system.
8473 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8475 void ia64_set_curr_task(int cpu, struct task_struct *p)
8482 #ifdef CONFIG_CGROUP_SCHED
8483 /* task_group_lock serializes the addition/removal of task groups */
8484 static DEFINE_SPINLOCK(task_group_lock);
8486 static inline void alloc_uclamp_sched_group(struct task_group *tg,
8487 struct task_group *parent)
8489 #ifdef CONFIG_UCLAMP_TASK_GROUP
8490 enum uclamp_id clamp_id;
8492 for_each_clamp_id(clamp_id) {
8493 uclamp_se_set(&tg->uclamp_req[clamp_id],
8494 uclamp_none(clamp_id), false);
8495 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
8500 static void sched_free_group(struct task_group *tg)
8502 free_fair_sched_group(tg);
8503 free_rt_sched_group(tg);
8505 kmem_cache_free(task_group_cache, tg);
8508 /* allocate runqueue etc for a new task group */
8509 struct task_group *sched_create_group(struct task_group *parent)
8511 struct task_group *tg;
8513 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
8515 return ERR_PTR(-ENOMEM);
8517 if (!alloc_fair_sched_group(tg, parent))
8520 if (!alloc_rt_sched_group(tg, parent))
8523 alloc_uclamp_sched_group(tg, parent);
8528 sched_free_group(tg);
8529 return ERR_PTR(-ENOMEM);
8532 void sched_online_group(struct task_group *tg, struct task_group *parent)
8534 unsigned long flags;
8536 spin_lock_irqsave(&task_group_lock, flags);
8537 list_add_rcu(&tg->list, &task_groups);
8539 /* Root should already exist: */
8542 tg->parent = parent;
8543 INIT_LIST_HEAD(&tg->children);
8544 list_add_rcu(&tg->siblings, &parent->children);
8545 spin_unlock_irqrestore(&task_group_lock, flags);
8547 online_fair_sched_group(tg);
8550 /* rcu callback to free various structures associated with a task group */
8551 static void sched_free_group_rcu(struct rcu_head *rhp)
8553 /* Now it should be safe to free those cfs_rqs: */
8554 sched_free_group(container_of(rhp, struct task_group, rcu));
8557 void sched_destroy_group(struct task_group *tg)
8559 /* Wait for possible concurrent references to cfs_rqs complete: */
8560 call_rcu(&tg->rcu, sched_free_group_rcu);
8563 void sched_offline_group(struct task_group *tg)
8565 unsigned long flags;
8567 /* End participation in shares distribution: */
8568 unregister_fair_sched_group(tg);
8570 spin_lock_irqsave(&task_group_lock, flags);
8571 list_del_rcu(&tg->list);
8572 list_del_rcu(&tg->siblings);
8573 spin_unlock_irqrestore(&task_group_lock, flags);
8576 static void sched_change_group(struct task_struct *tsk, int type)
8578 struct task_group *tg;
8581 * All callers are synchronized by task_rq_lock(); we do not use RCU
8582 * which is pointless here. Thus, we pass "true" to task_css_check()
8583 * to prevent lockdep warnings.
8585 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8586 struct task_group, css);
8587 tg = autogroup_task_group(tsk, tg);
8588 tsk->sched_task_group = tg;
8590 #ifdef CONFIG_FAIR_GROUP_SCHED
8591 if (tsk->sched_class->task_change_group)
8592 tsk->sched_class->task_change_group(tsk, type);
8595 set_task_rq(tsk, task_cpu(tsk));
8599 * Change task's runqueue when it moves between groups.
8601 * The caller of this function should have put the task in its new group by
8602 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8605 void sched_move_task(struct task_struct *tsk)
8607 int queued, running, queue_flags =
8608 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
8612 rq = task_rq_lock(tsk, &rf);
8613 update_rq_clock(rq);
8615 running = task_current(rq, tsk);
8616 queued = task_on_rq_queued(tsk);
8619 dequeue_task(rq, tsk, queue_flags);
8621 put_prev_task(rq, tsk);
8623 sched_change_group(tsk, TASK_MOVE_GROUP);
8626 enqueue_task(rq, tsk, queue_flags);
8628 set_next_task(rq, tsk);
8630 * After changing group, the running task may have joined a
8631 * throttled one but it's still the running task. Trigger a
8632 * resched to make sure that task can still run.
8637 task_rq_unlock(rq, tsk, &rf);
8640 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8642 return css ? container_of(css, struct task_group, css) : NULL;
8645 static struct cgroup_subsys_state *
8646 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8648 struct task_group *parent = css_tg(parent_css);
8649 struct task_group *tg;
8652 /* This is early initialization for the top cgroup */
8653 return &root_task_group.css;
8656 tg = sched_create_group(parent);
8658 return ERR_PTR(-ENOMEM);
8663 /* Expose task group only after completing cgroup initialization */
8664 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8666 struct task_group *tg = css_tg(css);
8667 struct task_group *parent = css_tg(css->parent);
8670 sched_online_group(tg, parent);
8672 #ifdef CONFIG_UCLAMP_TASK_GROUP
8673 /* Propagate the effective uclamp value for the new group */
8674 cpu_util_update_eff(css);
8680 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8682 struct task_group *tg = css_tg(css);
8684 sched_offline_group(tg);
8687 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8689 struct task_group *tg = css_tg(css);
8692 * Relies on the RCU grace period between css_released() and this.
8694 sched_free_group(tg);
8698 * This is called before wake_up_new_task(), therefore we really only
8699 * have to set its group bits, all the other stuff does not apply.
8701 static void cpu_cgroup_fork(struct task_struct *task)
8706 rq = task_rq_lock(task, &rf);
8708 update_rq_clock(rq);
8709 sched_change_group(task, TASK_SET_GROUP);
8711 task_rq_unlock(rq, task, &rf);
8714 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8716 struct task_struct *task;
8717 struct cgroup_subsys_state *css;
8720 cgroup_taskset_for_each(task, css, tset) {
8721 #ifdef CONFIG_RT_GROUP_SCHED
8722 if (!sched_rt_can_attach(css_tg(css), task))
8726 * Serialize against wake_up_new_task() such that if it's
8727 * running, we're sure to observe its full state.
8729 raw_spin_lock_irq(&task->pi_lock);
8731 * Avoid calling sched_move_task() before wake_up_new_task()
8732 * has happened. This would lead to problems with PELT, due to
8733 * move wanting to detach+attach while we're not attached yet.
8735 if (task->state == TASK_NEW)
8737 raw_spin_unlock_irq(&task->pi_lock);
8745 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8747 struct task_struct *task;
8748 struct cgroup_subsys_state *css;
8750 cgroup_taskset_for_each(task, css, tset)
8751 sched_move_task(task);
8754 #ifdef CONFIG_UCLAMP_TASK_GROUP
8755 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
8757 struct cgroup_subsys_state *top_css = css;
8758 struct uclamp_se *uc_parent = NULL;
8759 struct uclamp_se *uc_se = NULL;
8760 unsigned int eff[UCLAMP_CNT];
8761 enum uclamp_id clamp_id;
8762 unsigned int clamps;
8764 css_for_each_descendant_pre(css, top_css) {
8765 uc_parent = css_tg(css)->parent
8766 ? css_tg(css)->parent->uclamp : NULL;
8768 for_each_clamp_id(clamp_id) {
8769 /* Assume effective clamps matches requested clamps */
8770 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
8771 /* Cap effective clamps with parent's effective clamps */
8773 eff[clamp_id] > uc_parent[clamp_id].value) {
8774 eff[clamp_id] = uc_parent[clamp_id].value;
8777 /* Ensure protection is always capped by limit */
8778 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
8780 /* Propagate most restrictive effective clamps */
8782 uc_se = css_tg(css)->uclamp;
8783 for_each_clamp_id(clamp_id) {
8784 if (eff[clamp_id] == uc_se[clamp_id].value)
8786 uc_se[clamp_id].value = eff[clamp_id];
8787 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
8788 clamps |= (0x1 << clamp_id);
8791 css = css_rightmost_descendant(css);
8795 /* Immediately update descendants RUNNABLE tasks */
8796 uclamp_update_active_tasks(css, clamps);
8801 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
8802 * C expression. Since there is no way to convert a macro argument (N) into a
8803 * character constant, use two levels of macros.
8805 #define _POW10(exp) ((unsigned int)1e##exp)
8806 #define POW10(exp) _POW10(exp)
8808 struct uclamp_request {
8809 #define UCLAMP_PERCENT_SHIFT 2
8810 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
8816 static inline struct uclamp_request
8817 capacity_from_percent(char *buf)
8819 struct uclamp_request req = {
8820 .percent = UCLAMP_PERCENT_SCALE,
8821 .util = SCHED_CAPACITY_SCALE,
8826 if (strcmp(buf, "max")) {
8827 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
8831 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
8836 req.util = req.percent << SCHED_CAPACITY_SHIFT;
8837 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
8843 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
8844 size_t nbytes, loff_t off,
8845 enum uclamp_id clamp_id)
8847 struct uclamp_request req;
8848 struct task_group *tg;
8850 req = capacity_from_percent(buf);
8854 static_branch_enable(&sched_uclamp_used);
8856 mutex_lock(&uclamp_mutex);
8859 tg = css_tg(of_css(of));
8860 if (tg->uclamp_req[clamp_id].value != req.util)
8861 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
8864 * Because of not recoverable conversion rounding we keep track of the
8865 * exact requested value
8867 tg->uclamp_pct[clamp_id] = req.percent;
8869 /* Update effective clamps to track the most restrictive value */
8870 cpu_util_update_eff(of_css(of));
8873 mutex_unlock(&uclamp_mutex);
8878 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
8879 char *buf, size_t nbytes,
8882 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
8885 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
8886 char *buf, size_t nbytes,
8889 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
8892 static inline void cpu_uclamp_print(struct seq_file *sf,
8893 enum uclamp_id clamp_id)
8895 struct task_group *tg;
8901 tg = css_tg(seq_css(sf));
8902 util_clamp = tg->uclamp_req[clamp_id].value;
8905 if (util_clamp == SCHED_CAPACITY_SCALE) {
8906 seq_puts(sf, "max\n");
8910 percent = tg->uclamp_pct[clamp_id];
8911 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
8912 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
8915 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
8917 cpu_uclamp_print(sf, UCLAMP_MIN);
8921 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
8923 cpu_uclamp_print(sf, UCLAMP_MAX);
8926 #endif /* CONFIG_UCLAMP_TASK_GROUP */
8928 #ifdef CONFIG_FAIR_GROUP_SCHED
8929 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8930 struct cftype *cftype, u64 shareval)
8932 if (shareval > scale_load_down(ULONG_MAX))
8933 shareval = MAX_SHARES;
8934 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8937 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8940 struct task_group *tg = css_tg(css);
8942 return (u64) scale_load_down(tg->shares);
8945 #ifdef CONFIG_CFS_BANDWIDTH
8946 static DEFINE_MUTEX(cfs_constraints_mutex);
8948 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8949 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8950 /* More than 203 days if BW_SHIFT equals 20. */
8951 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
8953 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8955 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8957 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8958 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8960 if (tg == &root_task_group)
8964 * Ensure we have at some amount of bandwidth every period. This is
8965 * to prevent reaching a state of large arrears when throttled via
8966 * entity_tick() resulting in prolonged exit starvation.
8968 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8972 * Likewise, bound things on the otherside by preventing insane quota
8973 * periods. This also allows us to normalize in computing quota
8976 if (period > max_cfs_quota_period)
8980 * Bound quota to defend quota against overflow during bandwidth shift.
8982 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
8986 * Prevent race between setting of cfs_rq->runtime_enabled and
8987 * unthrottle_offline_cfs_rqs().
8990 mutex_lock(&cfs_constraints_mutex);
8991 ret = __cfs_schedulable(tg, period, quota);
8995 runtime_enabled = quota != RUNTIME_INF;
8996 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8998 * If we need to toggle cfs_bandwidth_used, off->on must occur
8999 * before making related changes, and on->off must occur afterwards
9001 if (runtime_enabled && !runtime_was_enabled)
9002 cfs_bandwidth_usage_inc();
9003 raw_spin_lock_irq(&cfs_b->lock);
9004 cfs_b->period = ns_to_ktime(period);
9005 cfs_b->quota = quota;
9007 __refill_cfs_bandwidth_runtime(cfs_b);
9009 /* Restart the period timer (if active) to handle new period expiry: */
9010 if (runtime_enabled)
9011 start_cfs_bandwidth(cfs_b);
9013 raw_spin_unlock_irq(&cfs_b->lock);
9015 for_each_online_cpu(i) {
9016 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9017 struct rq *rq = cfs_rq->rq;
9020 rq_lock_irq(rq, &rf);
9021 cfs_rq->runtime_enabled = runtime_enabled;
9022 cfs_rq->runtime_remaining = 0;
9024 if (cfs_rq->throttled)
9025 unthrottle_cfs_rq(cfs_rq);
9026 rq_unlock_irq(rq, &rf);
9028 if (runtime_was_enabled && !runtime_enabled)
9029 cfs_bandwidth_usage_dec();
9031 mutex_unlock(&cfs_constraints_mutex);
9037 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9041 period = ktime_to_ns(tg->cfs_bandwidth.period);
9042 if (cfs_quota_us < 0)
9043 quota = RUNTIME_INF;
9044 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
9045 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9049 return tg_set_cfs_bandwidth(tg, period, quota);
9052 static long tg_get_cfs_quota(struct task_group *tg)
9056 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
9059 quota_us = tg->cfs_bandwidth.quota;
9060 do_div(quota_us, NSEC_PER_USEC);
9065 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9069 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
9072 period = (u64)cfs_period_us * NSEC_PER_USEC;
9073 quota = tg->cfs_bandwidth.quota;
9075 return tg_set_cfs_bandwidth(tg, period, quota);
9078 static long tg_get_cfs_period(struct task_group *tg)
9082 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
9083 do_div(cfs_period_us, NSEC_PER_USEC);
9085 return cfs_period_us;
9088 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
9091 return tg_get_cfs_quota(css_tg(css));
9094 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
9095 struct cftype *cftype, s64 cfs_quota_us)
9097 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
9100 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
9103 return tg_get_cfs_period(css_tg(css));
9106 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
9107 struct cftype *cftype, u64 cfs_period_us)
9109 return tg_set_cfs_period(css_tg(css), cfs_period_us);
9112 struct cfs_schedulable_data {
9113 struct task_group *tg;
9118 * normalize group quota/period to be quota/max_period
9119 * note: units are usecs
9121 static u64 normalize_cfs_quota(struct task_group *tg,
9122 struct cfs_schedulable_data *d)
9130 period = tg_get_cfs_period(tg);
9131 quota = tg_get_cfs_quota(tg);
9134 /* note: these should typically be equivalent */
9135 if (quota == RUNTIME_INF || quota == -1)
9138 return to_ratio(period, quota);
9141 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9143 struct cfs_schedulable_data *d = data;
9144 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9145 s64 quota = 0, parent_quota = -1;
9148 quota = RUNTIME_INF;
9150 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
9152 quota = normalize_cfs_quota(tg, d);
9153 parent_quota = parent_b->hierarchical_quota;
9156 * Ensure max(child_quota) <= parent_quota. On cgroup2,
9157 * always take the min. On cgroup1, only inherit when no
9160 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
9161 quota = min(quota, parent_quota);
9163 if (quota == RUNTIME_INF)
9164 quota = parent_quota;
9165 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9169 cfs_b->hierarchical_quota = quota;
9174 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9177 struct cfs_schedulable_data data = {
9183 if (quota != RUNTIME_INF) {
9184 do_div(data.period, NSEC_PER_USEC);
9185 do_div(data.quota, NSEC_PER_USEC);
9189 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9195 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
9197 struct task_group *tg = css_tg(seq_css(sf));
9198 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9200 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
9201 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
9202 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
9204 if (schedstat_enabled() && tg != &root_task_group) {
9208 for_each_possible_cpu(i)
9209 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
9211 seq_printf(sf, "wait_sum %llu\n", ws);
9216 #endif /* CONFIG_CFS_BANDWIDTH */
9217 #endif /* CONFIG_FAIR_GROUP_SCHED */
9219 #ifdef CONFIG_RT_GROUP_SCHED
9220 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
9221 struct cftype *cft, s64 val)
9223 return sched_group_set_rt_runtime(css_tg(css), val);
9226 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
9229 return sched_group_rt_runtime(css_tg(css));
9232 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
9233 struct cftype *cftype, u64 rt_period_us)
9235 return sched_group_set_rt_period(css_tg(css), rt_period_us);
9238 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
9241 return sched_group_rt_period(css_tg(css));
9243 #endif /* CONFIG_RT_GROUP_SCHED */
9245 static struct cftype cpu_legacy_files[] = {
9246 #ifdef CONFIG_FAIR_GROUP_SCHED
9249 .read_u64 = cpu_shares_read_u64,
9250 .write_u64 = cpu_shares_write_u64,
9253 #ifdef CONFIG_CFS_BANDWIDTH
9255 .name = "cfs_quota_us",
9256 .read_s64 = cpu_cfs_quota_read_s64,
9257 .write_s64 = cpu_cfs_quota_write_s64,
9260 .name = "cfs_period_us",
9261 .read_u64 = cpu_cfs_period_read_u64,
9262 .write_u64 = cpu_cfs_period_write_u64,
9266 .seq_show = cpu_cfs_stat_show,
9269 #ifdef CONFIG_RT_GROUP_SCHED
9271 .name = "rt_runtime_us",
9272 .read_s64 = cpu_rt_runtime_read,
9273 .write_s64 = cpu_rt_runtime_write,
9276 .name = "rt_period_us",
9277 .read_u64 = cpu_rt_period_read_uint,
9278 .write_u64 = cpu_rt_period_write_uint,
9281 #ifdef CONFIG_UCLAMP_TASK_GROUP
9283 .name = "uclamp.min",
9284 .flags = CFTYPE_NOT_ON_ROOT,
9285 .seq_show = cpu_uclamp_min_show,
9286 .write = cpu_uclamp_min_write,
9289 .name = "uclamp.max",
9290 .flags = CFTYPE_NOT_ON_ROOT,
9291 .seq_show = cpu_uclamp_max_show,
9292 .write = cpu_uclamp_max_write,
9298 static int cpu_extra_stat_show(struct seq_file *sf,
9299 struct cgroup_subsys_state *css)
9301 #ifdef CONFIG_CFS_BANDWIDTH
9303 struct task_group *tg = css_tg(css);
9304 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9307 throttled_usec = cfs_b->throttled_time;
9308 do_div(throttled_usec, NSEC_PER_USEC);
9310 seq_printf(sf, "nr_periods %d\n"
9312 "throttled_usec %llu\n",
9313 cfs_b->nr_periods, cfs_b->nr_throttled,
9320 #ifdef CONFIG_FAIR_GROUP_SCHED
9321 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
9324 struct task_group *tg = css_tg(css);
9325 u64 weight = scale_load_down(tg->shares);
9327 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
9330 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
9331 struct cftype *cft, u64 weight)
9334 * cgroup weight knobs should use the common MIN, DFL and MAX
9335 * values which are 1, 100 and 10000 respectively. While it loses
9336 * a bit of range on both ends, it maps pretty well onto the shares
9337 * value used by scheduler and the round-trip conversions preserve
9338 * the original value over the entire range.
9340 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
9343 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
9345 return sched_group_set_shares(css_tg(css), scale_load(weight));
9348 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
9351 unsigned long weight = scale_load_down(css_tg(css)->shares);
9352 int last_delta = INT_MAX;
9355 /* find the closest nice value to the current weight */
9356 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
9357 delta = abs(sched_prio_to_weight[prio] - weight);
9358 if (delta >= last_delta)
9363 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
9366 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
9367 struct cftype *cft, s64 nice)
9369 unsigned long weight;
9372 if (nice < MIN_NICE || nice > MAX_NICE)
9375 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
9376 idx = array_index_nospec(idx, 40);
9377 weight = sched_prio_to_weight[idx];
9379 return sched_group_set_shares(css_tg(css), scale_load(weight));
9383 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
9384 long period, long quota)
9387 seq_puts(sf, "max");
9389 seq_printf(sf, "%ld", quota);
9391 seq_printf(sf, " %ld\n", period);
9394 /* caller should put the current value in *@periodp before calling */
9395 static int __maybe_unused cpu_period_quota_parse(char *buf,
9396 u64 *periodp, u64 *quotap)
9398 char tok[21]; /* U64_MAX */
9400 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
9403 *periodp *= NSEC_PER_USEC;
9405 if (sscanf(tok, "%llu", quotap))
9406 *quotap *= NSEC_PER_USEC;
9407 else if (!strcmp(tok, "max"))
9408 *quotap = RUNTIME_INF;
9415 #ifdef CONFIG_CFS_BANDWIDTH
9416 static int cpu_max_show(struct seq_file *sf, void *v)
9418 struct task_group *tg = css_tg(seq_css(sf));
9420 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
9424 static ssize_t cpu_max_write(struct kernfs_open_file *of,
9425 char *buf, size_t nbytes, loff_t off)
9427 struct task_group *tg = css_tg(of_css(of));
9428 u64 period = tg_get_cfs_period(tg);
9432 ret = cpu_period_quota_parse(buf, &period, "a);
9434 ret = tg_set_cfs_bandwidth(tg, period, quota);
9435 return ret ?: nbytes;
9439 static struct cftype cpu_files[] = {
9440 #ifdef CONFIG_FAIR_GROUP_SCHED
9443 .flags = CFTYPE_NOT_ON_ROOT,
9444 .read_u64 = cpu_weight_read_u64,
9445 .write_u64 = cpu_weight_write_u64,
9448 .name = "weight.nice",
9449 .flags = CFTYPE_NOT_ON_ROOT,
9450 .read_s64 = cpu_weight_nice_read_s64,
9451 .write_s64 = cpu_weight_nice_write_s64,
9454 #ifdef CONFIG_CFS_BANDWIDTH
9457 .flags = CFTYPE_NOT_ON_ROOT,
9458 .seq_show = cpu_max_show,
9459 .write = cpu_max_write,
9462 #ifdef CONFIG_UCLAMP_TASK_GROUP
9464 .name = "uclamp.min",
9465 .flags = CFTYPE_NOT_ON_ROOT,
9466 .seq_show = cpu_uclamp_min_show,
9467 .write = cpu_uclamp_min_write,
9470 .name = "uclamp.max",
9471 .flags = CFTYPE_NOT_ON_ROOT,
9472 .seq_show = cpu_uclamp_max_show,
9473 .write = cpu_uclamp_max_write,
9479 struct cgroup_subsys cpu_cgrp_subsys = {
9480 .css_alloc = cpu_cgroup_css_alloc,
9481 .css_online = cpu_cgroup_css_online,
9482 .css_released = cpu_cgroup_css_released,
9483 .css_free = cpu_cgroup_css_free,
9484 .css_extra_stat_show = cpu_extra_stat_show,
9485 .fork = cpu_cgroup_fork,
9486 .can_attach = cpu_cgroup_can_attach,
9487 .attach = cpu_cgroup_attach,
9488 .legacy_cftypes = cpu_legacy_files,
9489 .dfl_cftypes = cpu_files,
9494 #endif /* CONFIG_CGROUP_SCHED */
9496 void dump_cpu_task(int cpu)
9498 pr_info("Task dump for CPU %d:\n", cpu);
9499 sched_show_task(cpu_curr(cpu));
9503 * Nice levels are multiplicative, with a gentle 10% change for every
9504 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
9505 * nice 1, it will get ~10% less CPU time than another CPU-bound task
9506 * that remained on nice 0.
9508 * The "10% effect" is relative and cumulative: from _any_ nice level,
9509 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
9510 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
9511 * If a task goes up by ~10% and another task goes down by ~10% then
9512 * the relative distance between them is ~25%.)
9514 const int sched_prio_to_weight[40] = {
9515 /* -20 */ 88761, 71755, 56483, 46273, 36291,
9516 /* -15 */ 29154, 23254, 18705, 14949, 11916,
9517 /* -10 */ 9548, 7620, 6100, 4904, 3906,
9518 /* -5 */ 3121, 2501, 1991, 1586, 1277,
9519 /* 0 */ 1024, 820, 655, 526, 423,
9520 /* 5 */ 335, 272, 215, 172, 137,
9521 /* 10 */ 110, 87, 70, 56, 45,
9522 /* 15 */ 36, 29, 23, 18, 15,
9526 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
9528 * In cases where the weight does not change often, we can use the
9529 * precalculated inverse to speed up arithmetics by turning divisions
9530 * into multiplications:
9532 const u32 sched_prio_to_wmult[40] = {
9533 /* -20 */ 48388, 59856, 76040, 92818, 118348,
9534 /* -15 */ 147320, 184698, 229616, 287308, 360437,
9535 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
9536 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
9537 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
9538 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
9539 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
9540 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
9543 void call_trace_sched_update_nr_running(struct rq *rq, int count)
9545 trace_sched_update_nr_running_tp(rq, count);