1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #include <linux/highmem.h>
10 #include <linux/hrtimer_api.h>
11 #include <linux/ktime_api.h>
12 #include <linux/sched/signal.h>
13 #include <linux/syscalls_api.h>
14 #include <linux/debug_locks.h>
15 #include <linux/prefetch.h>
16 #include <linux/capability.h>
17 #include <linux/pgtable_api.h>
18 #include <linux/wait_bit.h>
19 #include <linux/jiffies.h>
20 #include <linux/spinlock_api.h>
21 #include <linux/cpumask_api.h>
22 #include <linux/lockdep_api.h>
23 #include <linux/hardirq.h>
24 #include <linux/softirq.h>
25 #include <linux/refcount_api.h>
26 #include <linux/topology.h>
27 #include <linux/sched/clock.h>
28 #include <linux/sched/cond_resched.h>
29 #include <linux/sched/cputime.h>
30 #include <linux/sched/debug.h>
31 #include <linux/sched/hotplug.h>
32 #include <linux/sched/init.h>
33 #include <linux/sched/isolation.h>
34 #include <linux/sched/loadavg.h>
35 #include <linux/sched/mm.h>
36 #include <linux/sched/nohz.h>
37 #include <linux/sched/rseq_api.h>
38 #include <linux/sched/rt.h>
40 #include <linux/blkdev.h>
41 #include <linux/context_tracking.h>
42 #include <linux/cpuset.h>
43 #include <linux/delayacct.h>
44 #include <linux/init_task.h>
45 #include <linux/interrupt.h>
46 #include <linux/ioprio.h>
47 #include <linux/kallsyms.h>
48 #include <linux/kcov.h>
49 #include <linux/kprobes.h>
50 #include <linux/llist_api.h>
51 #include <linux/mmu_context.h>
52 #include <linux/mmzone.h>
53 #include <linux/mutex_api.h>
54 #include <linux/nmi.h>
55 #include <linux/nospec.h>
56 #include <linux/perf_event_api.h>
57 #include <linux/profile.h>
58 #include <linux/psi.h>
59 #include <linux/rcuwait_api.h>
60 #include <linux/sched/wake_q.h>
61 #include <linux/scs.h>
62 #include <linux/slab.h>
63 #include <linux/syscalls.h>
64 #include <linux/vtime.h>
65 #include <linux/wait_api.h>
66 #include <linux/workqueue_api.h>
68 #ifdef CONFIG_PREEMPT_DYNAMIC
69 # ifdef CONFIG_GENERIC_ENTRY
70 # include <linux/entry-common.h>
74 #include <uapi/linux/sched/types.h>
76 #include <asm/irq_regs.h>
77 #include <asm/switch_to.h>
80 #define CREATE_TRACE_POINTS
81 #include <linux/sched/rseq_api.h>
82 #include <trace/events/sched.h>
83 #include <trace/events/ipi.h>
84 #undef CREATE_TRACE_POINTS
88 #include "autogroup.h"
90 #include "autogroup.h"
95 #include "../workqueue_internal.h"
96 #include "../../io_uring/io-wq.h"
97 #include "../smpboot.h"
99 EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpu);
100 EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpumask);
103 * Export tracepoints that act as a bare tracehook (ie: have no trace event
104 * associated with them) to allow external modules to probe them.
106 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
107 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
108 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
109 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
110 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
111 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
112 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
113 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
114 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
115 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
116 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
118 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
120 #ifdef CONFIG_SCHED_DEBUG
122 * Debugging: various feature bits
124 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
125 * sysctl_sched_features, defined in sched.h, to allow constants propagation
126 * at compile time and compiler optimization based on features default.
128 #define SCHED_FEAT(name, enabled) \
129 (1UL << __SCHED_FEAT_##name) * enabled |
130 const_debug unsigned int sysctl_sched_features =
131 #include "features.h"
136 * Print a warning if need_resched is set for the given duration (if
137 * LATENCY_WARN is enabled).
139 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
142 __read_mostly int sysctl_resched_latency_warn_ms = 100;
143 __read_mostly int sysctl_resched_latency_warn_once = 1;
144 #endif /* CONFIG_SCHED_DEBUG */
147 * Number of tasks to iterate in a single balance run.
148 * Limited because this is done with IRQs disabled.
150 const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
152 __read_mostly int scheduler_running;
154 #ifdef CONFIG_SCHED_CORE
156 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
158 /* kernel prio, less is more */
159 static inline int __task_prio(const struct task_struct *p)
161 if (p->sched_class == &stop_sched_class) /* trumps deadline */
164 if (rt_prio(p->prio)) /* includes deadline */
165 return p->prio; /* [-1, 99] */
167 if (p->sched_class == &idle_sched_class)
168 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
170 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
180 /* real prio, less is less */
181 static inline bool prio_less(const struct task_struct *a,
182 const struct task_struct *b, bool in_fi)
185 int pa = __task_prio(a), pb = __task_prio(b);
193 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
194 return !dl_time_before(a->dl.deadline, b->dl.deadline);
196 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
197 return cfs_prio_less(a, b, in_fi);
202 static inline bool __sched_core_less(const struct task_struct *a,
203 const struct task_struct *b)
205 if (a->core_cookie < b->core_cookie)
208 if (a->core_cookie > b->core_cookie)
211 /* flip prio, so high prio is leftmost */
212 if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
218 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
220 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
222 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
225 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
227 const struct task_struct *p = __node_2_sc(node);
228 unsigned long cookie = (unsigned long)key;
230 if (cookie < p->core_cookie)
233 if (cookie > p->core_cookie)
239 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
241 rq->core->core_task_seq++;
246 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
249 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
251 rq->core->core_task_seq++;
253 if (sched_core_enqueued(p)) {
254 rb_erase(&p->core_node, &rq->core_tree);
255 RB_CLEAR_NODE(&p->core_node);
259 * Migrating the last task off the cpu, with the cpu in forced idle
260 * state. Reschedule to create an accounting edge for forced idle,
261 * and re-examine whether the core is still in forced idle state.
263 if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
264 rq->core->core_forceidle_count && rq->curr == rq->idle)
268 static int sched_task_is_throttled(struct task_struct *p, int cpu)
270 if (p->sched_class->task_is_throttled)
271 return p->sched_class->task_is_throttled(p, cpu);
276 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
278 struct rb_node *node = &p->core_node;
279 int cpu = task_cpu(p);
282 node = rb_next(node);
286 p = __node_2_sc(node);
287 if (p->core_cookie != cookie)
290 } while (sched_task_is_throttled(p, cpu));
296 * Find left-most (aka, highest priority) and unthrottled task matching @cookie.
297 * If no suitable task is found, NULL will be returned.
299 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
301 struct task_struct *p;
302 struct rb_node *node;
304 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
308 p = __node_2_sc(node);
309 if (!sched_task_is_throttled(p, rq->cpu))
312 return sched_core_next(p, cookie);
316 * Magic required such that:
318 * raw_spin_rq_lock(rq);
320 * raw_spin_rq_unlock(rq);
322 * ends up locking and unlocking the _same_ lock, and all CPUs
323 * always agree on what rq has what lock.
325 * XXX entirely possible to selectively enable cores, don't bother for now.
328 static DEFINE_MUTEX(sched_core_mutex);
329 static atomic_t sched_core_count;
330 static struct cpumask sched_core_mask;
332 static void sched_core_lock(int cpu, unsigned long *flags)
334 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
337 local_irq_save(*flags);
338 for_each_cpu(t, smt_mask)
339 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
342 static void sched_core_unlock(int cpu, unsigned long *flags)
344 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
347 for_each_cpu(t, smt_mask)
348 raw_spin_unlock(&cpu_rq(t)->__lock);
349 local_irq_restore(*flags);
352 static void __sched_core_flip(bool enabled)
360 * Toggle the online cores, one by one.
362 cpumask_copy(&sched_core_mask, cpu_online_mask);
363 for_each_cpu(cpu, &sched_core_mask) {
364 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
366 sched_core_lock(cpu, &flags);
368 for_each_cpu(t, smt_mask)
369 cpu_rq(t)->core_enabled = enabled;
371 cpu_rq(cpu)->core->core_forceidle_start = 0;
373 sched_core_unlock(cpu, &flags);
375 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
379 * Toggle the offline CPUs.
381 for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
382 cpu_rq(cpu)->core_enabled = enabled;
387 static void sched_core_assert_empty(void)
391 for_each_possible_cpu(cpu)
392 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
395 static void __sched_core_enable(void)
397 static_branch_enable(&__sched_core_enabled);
399 * Ensure all previous instances of raw_spin_rq_*lock() have finished
400 * and future ones will observe !sched_core_disabled().
403 __sched_core_flip(true);
404 sched_core_assert_empty();
407 static void __sched_core_disable(void)
409 sched_core_assert_empty();
410 __sched_core_flip(false);
411 static_branch_disable(&__sched_core_enabled);
414 void sched_core_get(void)
416 if (atomic_inc_not_zero(&sched_core_count))
419 mutex_lock(&sched_core_mutex);
420 if (!atomic_read(&sched_core_count))
421 __sched_core_enable();
423 smp_mb__before_atomic();
424 atomic_inc(&sched_core_count);
425 mutex_unlock(&sched_core_mutex);
428 static void __sched_core_put(struct work_struct *work)
430 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
431 __sched_core_disable();
432 mutex_unlock(&sched_core_mutex);
436 void sched_core_put(void)
438 static DECLARE_WORK(_work, __sched_core_put);
441 * "There can be only one"
443 * Either this is the last one, or we don't actually need to do any
444 * 'work'. If it is the last *again*, we rely on
445 * WORK_STRUCT_PENDING_BIT.
447 if (!atomic_add_unless(&sched_core_count, -1, 1))
448 schedule_work(&_work);
451 #else /* !CONFIG_SCHED_CORE */
453 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
455 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
457 #endif /* CONFIG_SCHED_CORE */
460 * Serialization rules:
466 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
469 * rq2->lock where: rq1 < rq2
473 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
474 * local CPU's rq->lock, it optionally removes the task from the runqueue and
475 * always looks at the local rq data structures to find the most eligible task
478 * Task enqueue is also under rq->lock, possibly taken from another CPU.
479 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
480 * the local CPU to avoid bouncing the runqueue state around [ see
481 * ttwu_queue_wakelist() ]
483 * Task wakeup, specifically wakeups that involve migration, are horribly
484 * complicated to avoid having to take two rq->locks.
488 * System-calls and anything external will use task_rq_lock() which acquires
489 * both p->pi_lock and rq->lock. As a consequence the state they change is
490 * stable while holding either lock:
492 * - sched_setaffinity()/
493 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
494 * - set_user_nice(): p->se.load, p->*prio
495 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
496 * p->se.load, p->rt_priority,
497 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
498 * - sched_setnuma(): p->numa_preferred_nid
499 * - sched_move_task(): p->sched_task_group
500 * - uclamp_update_active() p->uclamp*
502 * p->state <- TASK_*:
504 * is changed locklessly using set_current_state(), __set_current_state() or
505 * set_special_state(), see their respective comments, or by
506 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
509 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
511 * is set by activate_task() and cleared by deactivate_task(), under
512 * rq->lock. Non-zero indicates the task is runnable, the special
513 * ON_RQ_MIGRATING state is used for migration without holding both
514 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
516 * p->on_cpu <- { 0, 1 }:
518 * is set by prepare_task() and cleared by finish_task() such that it will be
519 * set before p is scheduled-in and cleared after p is scheduled-out, both
520 * under rq->lock. Non-zero indicates the task is running on its CPU.
522 * [ The astute reader will observe that it is possible for two tasks on one
523 * CPU to have ->on_cpu = 1 at the same time. ]
525 * task_cpu(p): is changed by set_task_cpu(), the rules are:
527 * - Don't call set_task_cpu() on a blocked task:
529 * We don't care what CPU we're not running on, this simplifies hotplug,
530 * the CPU assignment of blocked tasks isn't required to be valid.
532 * - for try_to_wake_up(), called under p->pi_lock:
534 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
536 * - for migration called under rq->lock:
537 * [ see task_on_rq_migrating() in task_rq_lock() ]
539 * o move_queued_task()
542 * - for migration called under double_rq_lock():
544 * o __migrate_swap_task()
545 * o push_rt_task() / pull_rt_task()
546 * o push_dl_task() / pull_dl_task()
547 * o dl_task_offline_migration()
551 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
553 raw_spinlock_t *lock;
555 /* Matches synchronize_rcu() in __sched_core_enable() */
557 if (sched_core_disabled()) {
558 raw_spin_lock_nested(&rq->__lock, subclass);
559 /* preempt_count *MUST* be > 1 */
560 preempt_enable_no_resched();
565 lock = __rq_lockp(rq);
566 raw_spin_lock_nested(lock, subclass);
567 if (likely(lock == __rq_lockp(rq))) {
568 /* preempt_count *MUST* be > 1 */
569 preempt_enable_no_resched();
572 raw_spin_unlock(lock);
576 bool raw_spin_rq_trylock(struct rq *rq)
578 raw_spinlock_t *lock;
581 /* Matches synchronize_rcu() in __sched_core_enable() */
583 if (sched_core_disabled()) {
584 ret = raw_spin_trylock(&rq->__lock);
590 lock = __rq_lockp(rq);
591 ret = raw_spin_trylock(lock);
592 if (!ret || (likely(lock == __rq_lockp(rq)))) {
596 raw_spin_unlock(lock);
600 void raw_spin_rq_unlock(struct rq *rq)
602 raw_spin_unlock(rq_lockp(rq));
607 * double_rq_lock - safely lock two runqueues
609 void double_rq_lock(struct rq *rq1, struct rq *rq2)
611 lockdep_assert_irqs_disabled();
613 if (rq_order_less(rq2, rq1))
616 raw_spin_rq_lock(rq1);
617 if (__rq_lockp(rq1) != __rq_lockp(rq2))
618 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
620 double_rq_clock_clear_update(rq1, rq2);
625 * __task_rq_lock - lock the rq @p resides on.
627 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
632 lockdep_assert_held(&p->pi_lock);
636 raw_spin_rq_lock(rq);
637 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
641 raw_spin_rq_unlock(rq);
643 while (unlikely(task_on_rq_migrating(p)))
649 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
651 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
652 __acquires(p->pi_lock)
658 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
660 raw_spin_rq_lock(rq);
662 * move_queued_task() task_rq_lock()
665 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
666 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
667 * [S] ->cpu = new_cpu [L] task_rq()
671 * If we observe the old CPU in task_rq_lock(), the acquire of
672 * the old rq->lock will fully serialize against the stores.
674 * If we observe the new CPU in task_rq_lock(), the address
675 * dependency headed by '[L] rq = task_rq()' and the acquire
676 * will pair with the WMB to ensure we then also see migrating.
678 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
682 raw_spin_rq_unlock(rq);
683 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
685 while (unlikely(task_on_rq_migrating(p)))
691 * RQ-clock updating methods:
694 static void update_rq_clock_task(struct rq *rq, s64 delta)
697 * In theory, the compile should just see 0 here, and optimize out the call
698 * to sched_rt_avg_update. But I don't trust it...
700 s64 __maybe_unused steal = 0, irq_delta = 0;
702 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
703 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
706 * Since irq_time is only updated on {soft,}irq_exit, we might run into
707 * this case when a previous update_rq_clock() happened inside a
710 * When this happens, we stop ->clock_task and only update the
711 * prev_irq_time stamp to account for the part that fit, so that a next
712 * update will consume the rest. This ensures ->clock_task is
715 * It does however cause some slight miss-attribution of {soft,}irq
716 * time, a more accurate solution would be to update the irq_time using
717 * the current rq->clock timestamp, except that would require using
720 if (irq_delta > delta)
723 rq->prev_irq_time += irq_delta;
725 psi_account_irqtime(rq->curr, irq_delta);
726 delayacct_irq(rq->curr, irq_delta);
728 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
729 if (static_key_false((¶virt_steal_rq_enabled))) {
730 steal = paravirt_steal_clock(cpu_of(rq));
731 steal -= rq->prev_steal_time_rq;
733 if (unlikely(steal > delta))
736 rq->prev_steal_time_rq += steal;
741 rq->clock_task += delta;
743 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
744 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
745 update_irq_load_avg(rq, irq_delta + steal);
747 update_rq_clock_pelt(rq, delta);
750 void update_rq_clock(struct rq *rq)
754 lockdep_assert_rq_held(rq);
756 if (rq->clock_update_flags & RQCF_ACT_SKIP)
759 #ifdef CONFIG_SCHED_DEBUG
760 if (sched_feat(WARN_DOUBLE_CLOCK))
761 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
762 rq->clock_update_flags |= RQCF_UPDATED;
765 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
769 update_rq_clock_task(rq, delta);
772 #ifdef CONFIG_SCHED_HRTICK
774 * Use HR-timers to deliver accurate preemption points.
777 static void hrtick_clear(struct rq *rq)
779 if (hrtimer_active(&rq->hrtick_timer))
780 hrtimer_cancel(&rq->hrtick_timer);
784 * High-resolution timer tick.
785 * Runs from hardirq context with interrupts disabled.
787 static enum hrtimer_restart hrtick(struct hrtimer *timer)
789 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
792 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
796 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
799 return HRTIMER_NORESTART;
804 static void __hrtick_restart(struct rq *rq)
806 struct hrtimer *timer = &rq->hrtick_timer;
807 ktime_t time = rq->hrtick_time;
809 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
813 * called from hardirq (IPI) context
815 static void __hrtick_start(void *arg)
821 __hrtick_restart(rq);
826 * Called to set the hrtick timer state.
828 * called with rq->lock held and irqs disabled
830 void hrtick_start(struct rq *rq, u64 delay)
832 struct hrtimer *timer = &rq->hrtick_timer;
836 * Don't schedule slices shorter than 10000ns, that just
837 * doesn't make sense and can cause timer DoS.
839 delta = max_t(s64, delay, 10000LL);
840 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
843 __hrtick_restart(rq);
845 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
850 * Called to set the hrtick timer state.
852 * called with rq->lock held and irqs disabled
854 void hrtick_start(struct rq *rq, u64 delay)
857 * Don't schedule slices shorter than 10000ns, that just
858 * doesn't make sense. Rely on vruntime for fairness.
860 delay = max_t(u64, delay, 10000LL);
861 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
862 HRTIMER_MODE_REL_PINNED_HARD);
865 #endif /* CONFIG_SMP */
867 static void hrtick_rq_init(struct rq *rq)
870 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
872 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
873 rq->hrtick_timer.function = hrtick;
875 #else /* CONFIG_SCHED_HRTICK */
876 static inline void hrtick_clear(struct rq *rq)
880 static inline void hrtick_rq_init(struct rq *rq)
883 #endif /* CONFIG_SCHED_HRTICK */
886 * cmpxchg based fetch_or, macro so it works for different integer types
888 #define fetch_or(ptr, mask) \
890 typeof(ptr) _ptr = (ptr); \
891 typeof(mask) _mask = (mask); \
892 typeof(*_ptr) _val = *_ptr; \
895 } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
899 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
901 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
902 * this avoids any races wrt polling state changes and thereby avoids
905 static inline bool set_nr_and_not_polling(struct task_struct *p)
907 struct thread_info *ti = task_thread_info(p);
908 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
912 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
914 * If this returns true, then the idle task promises to call
915 * sched_ttwu_pending() and reschedule soon.
917 static bool set_nr_if_polling(struct task_struct *p)
919 struct thread_info *ti = task_thread_info(p);
920 typeof(ti->flags) val = READ_ONCE(ti->flags);
923 if (!(val & _TIF_POLLING_NRFLAG))
925 if (val & _TIF_NEED_RESCHED)
927 if (try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED))
934 static inline bool set_nr_and_not_polling(struct task_struct *p)
936 set_tsk_need_resched(p);
941 static inline bool set_nr_if_polling(struct task_struct *p)
948 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
950 struct wake_q_node *node = &task->wake_q;
953 * Atomically grab the task, if ->wake_q is !nil already it means
954 * it's already queued (either by us or someone else) and will get the
955 * wakeup due to that.
957 * In order to ensure that a pending wakeup will observe our pending
958 * state, even in the failed case, an explicit smp_mb() must be used.
960 smp_mb__before_atomic();
961 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
965 * The head is context local, there can be no concurrency.
968 head->lastp = &node->next;
973 * wake_q_add() - queue a wakeup for 'later' waking.
974 * @head: the wake_q_head to add @task to
975 * @task: the task to queue for 'later' wakeup
977 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
978 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
981 * This function must be used as-if it were wake_up_process(); IOW the task
982 * must be ready to be woken at this location.
984 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
986 if (__wake_q_add(head, task))
987 get_task_struct(task);
991 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
992 * @head: the wake_q_head to add @task to
993 * @task: the task to queue for 'later' wakeup
995 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
996 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
999 * This function must be used as-if it were wake_up_process(); IOW the task
1000 * must be ready to be woken at this location.
1002 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
1003 * that already hold reference to @task can call the 'safe' version and trust
1004 * wake_q to do the right thing depending whether or not the @task is already
1005 * queued for wakeup.
1007 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
1009 if (!__wake_q_add(head, task))
1010 put_task_struct(task);
1013 void wake_up_q(struct wake_q_head *head)
1015 struct wake_q_node *node = head->first;
1017 while (node != WAKE_Q_TAIL) {
1018 struct task_struct *task;
1020 task = container_of(node, struct task_struct, wake_q);
1021 /* Task can safely be re-inserted now: */
1023 task->wake_q.next = NULL;
1026 * wake_up_process() executes a full barrier, which pairs with
1027 * the queueing in wake_q_add() so as not to miss wakeups.
1029 wake_up_process(task);
1030 put_task_struct(task);
1035 * resched_curr - mark rq's current task 'to be rescheduled now'.
1037 * On UP this means the setting of the need_resched flag, on SMP it
1038 * might also involve a cross-CPU call to trigger the scheduler on
1041 void resched_curr(struct rq *rq)
1043 struct task_struct *curr = rq->curr;
1046 lockdep_assert_rq_held(rq);
1048 if (test_tsk_need_resched(curr))
1053 if (cpu == smp_processor_id()) {
1054 set_tsk_need_resched(curr);
1055 set_preempt_need_resched();
1059 if (set_nr_and_not_polling(curr))
1060 smp_send_reschedule(cpu);
1062 trace_sched_wake_idle_without_ipi(cpu);
1065 void resched_cpu(int cpu)
1067 struct rq *rq = cpu_rq(cpu);
1068 unsigned long flags;
1070 raw_spin_rq_lock_irqsave(rq, flags);
1071 if (cpu_online(cpu) || cpu == smp_processor_id())
1073 raw_spin_rq_unlock_irqrestore(rq, flags);
1077 #ifdef CONFIG_NO_HZ_COMMON
1079 * In the semi idle case, use the nearest busy CPU for migrating timers
1080 * from an idle CPU. This is good for power-savings.
1082 * We don't do similar optimization for completely idle system, as
1083 * selecting an idle CPU will add more delays to the timers than intended
1084 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1086 int get_nohz_timer_target(void)
1088 int i, cpu = smp_processor_id(), default_cpu = -1;
1089 struct sched_domain *sd;
1090 const struct cpumask *hk_mask;
1092 if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1098 hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1102 for_each_domain(cpu, sd) {
1103 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1112 if (default_cpu == -1)
1113 default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1119 * When add_timer_on() enqueues a timer into the timer wheel of an
1120 * idle CPU then this timer might expire before the next timer event
1121 * which is scheduled to wake up that CPU. In case of a completely
1122 * idle system the next event might even be infinite time into the
1123 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1124 * leaves the inner idle loop so the newly added timer is taken into
1125 * account when the CPU goes back to idle and evaluates the timer
1126 * wheel for the next timer event.
1128 static void wake_up_idle_cpu(int cpu)
1130 struct rq *rq = cpu_rq(cpu);
1132 if (cpu == smp_processor_id())
1135 if (set_nr_and_not_polling(rq->idle))
1136 smp_send_reschedule(cpu);
1138 trace_sched_wake_idle_without_ipi(cpu);
1141 static bool wake_up_full_nohz_cpu(int cpu)
1144 * We just need the target to call irq_exit() and re-evaluate
1145 * the next tick. The nohz full kick at least implies that.
1146 * If needed we can still optimize that later with an
1149 if (cpu_is_offline(cpu))
1150 return true; /* Don't try to wake offline CPUs. */
1151 if (tick_nohz_full_cpu(cpu)) {
1152 if (cpu != smp_processor_id() ||
1153 tick_nohz_tick_stopped())
1154 tick_nohz_full_kick_cpu(cpu);
1162 * Wake up the specified CPU. If the CPU is going offline, it is the
1163 * caller's responsibility to deal with the lost wakeup, for example,
1164 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1166 void wake_up_nohz_cpu(int cpu)
1168 if (!wake_up_full_nohz_cpu(cpu))
1169 wake_up_idle_cpu(cpu);
1172 static void nohz_csd_func(void *info)
1174 struct rq *rq = info;
1175 int cpu = cpu_of(rq);
1179 * Release the rq::nohz_csd.
1181 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1182 WARN_ON(!(flags & NOHZ_KICK_MASK));
1184 rq->idle_balance = idle_cpu(cpu);
1185 if (rq->idle_balance && !need_resched()) {
1186 rq->nohz_idle_balance = flags;
1187 raise_softirq_irqoff(SCHED_SOFTIRQ);
1191 #endif /* CONFIG_NO_HZ_COMMON */
1193 #ifdef CONFIG_NO_HZ_FULL
1194 static inline bool __need_bw_check(struct rq *rq, struct task_struct *p)
1196 if (rq->nr_running != 1)
1199 if (p->sched_class != &fair_sched_class)
1202 if (!task_on_rq_queued(p))
1208 bool sched_can_stop_tick(struct rq *rq)
1210 int fifo_nr_running;
1212 /* Deadline tasks, even if single, need the tick */
1213 if (rq->dl.dl_nr_running)
1217 * If there are more than one RR tasks, we need the tick to affect the
1218 * actual RR behaviour.
1220 if (rq->rt.rr_nr_running) {
1221 if (rq->rt.rr_nr_running == 1)
1228 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1229 * forced preemption between FIFO tasks.
1231 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1232 if (fifo_nr_running)
1236 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1237 * if there's more than one we need the tick for involuntary
1240 if (rq->nr_running > 1)
1244 * If there is one task and it has CFS runtime bandwidth constraints
1245 * and it's on the cpu now we don't want to stop the tick.
1246 * This check prevents clearing the bit if a newly enqueued task here is
1247 * dequeued by migrating while the constrained task continues to run.
1248 * E.g. going from 2->1 without going through pick_next_task().
1250 if (sched_feat(HZ_BW) && __need_bw_check(rq, rq->curr)) {
1251 if (cfs_task_bw_constrained(rq->curr))
1257 #endif /* CONFIG_NO_HZ_FULL */
1258 #endif /* CONFIG_SMP */
1260 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1261 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1263 * Iterate task_group tree rooted at *from, calling @down when first entering a
1264 * node and @up when leaving it for the final time.
1266 * Caller must hold rcu_lock or sufficient equivalent.
1268 int walk_tg_tree_from(struct task_group *from,
1269 tg_visitor down, tg_visitor up, void *data)
1271 struct task_group *parent, *child;
1277 ret = (*down)(parent, data);
1280 list_for_each_entry_rcu(child, &parent->children, siblings) {
1287 ret = (*up)(parent, data);
1288 if (ret || parent == from)
1292 parent = parent->parent;
1299 int tg_nop(struct task_group *tg, void *data)
1305 static void set_load_weight(struct task_struct *p, bool update_load)
1307 int prio = p->static_prio - MAX_RT_PRIO;
1308 struct load_weight *load = &p->se.load;
1311 * SCHED_IDLE tasks get minimal weight:
1313 if (task_has_idle_policy(p)) {
1314 load->weight = scale_load(WEIGHT_IDLEPRIO);
1315 load->inv_weight = WMULT_IDLEPRIO;
1320 * SCHED_OTHER tasks have to update their load when changing their
1323 if (update_load && p->sched_class == &fair_sched_class) {
1324 reweight_task(p, prio);
1326 load->weight = scale_load(sched_prio_to_weight[prio]);
1327 load->inv_weight = sched_prio_to_wmult[prio];
1331 #ifdef CONFIG_UCLAMP_TASK
1333 * Serializes updates of utilization clamp values
1335 * The (slow-path) user-space triggers utilization clamp value updates which
1336 * can require updates on (fast-path) scheduler's data structures used to
1337 * support enqueue/dequeue operations.
1338 * While the per-CPU rq lock protects fast-path update operations, user-space
1339 * requests are serialized using a mutex to reduce the risk of conflicting
1340 * updates or API abuses.
1342 static DEFINE_MUTEX(uclamp_mutex);
1344 /* Max allowed minimum utilization */
1345 static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1347 /* Max allowed maximum utilization */
1348 static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1351 * By default RT tasks run at the maximum performance point/capacity of the
1352 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1353 * SCHED_CAPACITY_SCALE.
1355 * This knob allows admins to change the default behavior when uclamp is being
1356 * used. In battery powered devices, particularly, running at the maximum
1357 * capacity and frequency will increase energy consumption and shorten the
1360 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1362 * This knob will not override the system default sched_util_clamp_min defined
1365 static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1367 /* All clamps are required to be less or equal than these values */
1368 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1371 * This static key is used to reduce the uclamp overhead in the fast path. It
1372 * primarily disables the call to uclamp_rq_{inc, dec}() in
1373 * enqueue/dequeue_task().
1375 * This allows users to continue to enable uclamp in their kernel config with
1376 * minimum uclamp overhead in the fast path.
1378 * As soon as userspace modifies any of the uclamp knobs, the static key is
1379 * enabled, since we have an actual users that make use of uclamp
1382 * The knobs that would enable this static key are:
1384 * * A task modifying its uclamp value with sched_setattr().
1385 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1386 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1388 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1390 /* Integer rounded range for each bucket */
1391 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1393 #define for_each_clamp_id(clamp_id) \
1394 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1396 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1398 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1401 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1403 if (clamp_id == UCLAMP_MIN)
1405 return SCHED_CAPACITY_SCALE;
1408 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1409 unsigned int value, bool user_defined)
1411 uc_se->value = value;
1412 uc_se->bucket_id = uclamp_bucket_id(value);
1413 uc_se->user_defined = user_defined;
1416 static inline unsigned int
1417 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1418 unsigned int clamp_value)
1421 * Avoid blocked utilization pushing up the frequency when we go
1422 * idle (which drops the max-clamp) by retaining the last known
1425 if (clamp_id == UCLAMP_MAX) {
1426 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1430 return uclamp_none(UCLAMP_MIN);
1433 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1434 unsigned int clamp_value)
1436 /* Reset max-clamp retention only on idle exit */
1437 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1440 uclamp_rq_set(rq, clamp_id, clamp_value);
1444 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1445 unsigned int clamp_value)
1447 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1448 int bucket_id = UCLAMP_BUCKETS - 1;
1451 * Since both min and max clamps are max aggregated, find the
1452 * top most bucket with tasks in.
1454 for ( ; bucket_id >= 0; bucket_id--) {
1455 if (!bucket[bucket_id].tasks)
1457 return bucket[bucket_id].value;
1460 /* No tasks -- default clamp values */
1461 return uclamp_idle_value(rq, clamp_id, clamp_value);
1464 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1466 unsigned int default_util_min;
1467 struct uclamp_se *uc_se;
1469 lockdep_assert_held(&p->pi_lock);
1471 uc_se = &p->uclamp_req[UCLAMP_MIN];
1473 /* Only sync if user didn't override the default */
1474 if (uc_se->user_defined)
1477 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1478 uclamp_se_set(uc_se, default_util_min, false);
1481 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1489 /* Protect updates to p->uclamp_* */
1490 rq = task_rq_lock(p, &rf);
1491 __uclamp_update_util_min_rt_default(p);
1492 task_rq_unlock(rq, p, &rf);
1495 static inline struct uclamp_se
1496 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1498 /* Copy by value as we could modify it */
1499 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1500 #ifdef CONFIG_UCLAMP_TASK_GROUP
1501 unsigned int tg_min, tg_max, value;
1504 * Tasks in autogroups or root task group will be
1505 * restricted by system defaults.
1507 if (task_group_is_autogroup(task_group(p)))
1509 if (task_group(p) == &root_task_group)
1512 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1513 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1514 value = uc_req.value;
1515 value = clamp(value, tg_min, tg_max);
1516 uclamp_se_set(&uc_req, value, false);
1523 * The effective clamp bucket index of a task depends on, by increasing
1525 * - the task specific clamp value, when explicitly requested from userspace
1526 * - the task group effective clamp value, for tasks not either in the root
1527 * group or in an autogroup
1528 * - the system default clamp value, defined by the sysadmin
1530 static inline struct uclamp_se
1531 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1533 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1534 struct uclamp_se uc_max = uclamp_default[clamp_id];
1536 /* System default restrictions always apply */
1537 if (unlikely(uc_req.value > uc_max.value))
1543 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1545 struct uclamp_se uc_eff;
1547 /* Task currently refcounted: use back-annotated (effective) value */
1548 if (p->uclamp[clamp_id].active)
1549 return (unsigned long)p->uclamp[clamp_id].value;
1551 uc_eff = uclamp_eff_get(p, clamp_id);
1553 return (unsigned long)uc_eff.value;
1557 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1558 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1559 * updates the rq's clamp value if required.
1561 * Tasks can have a task-specific value requested from user-space, track
1562 * within each bucket the maximum value for tasks refcounted in it.
1563 * This "local max aggregation" allows to track the exact "requested" value
1564 * for each bucket when all its RUNNABLE tasks require the same clamp.
1566 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1567 enum uclamp_id clamp_id)
1569 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1570 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1571 struct uclamp_bucket *bucket;
1573 lockdep_assert_rq_held(rq);
1575 /* Update task effective clamp */
1576 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1578 bucket = &uc_rq->bucket[uc_se->bucket_id];
1580 uc_se->active = true;
1582 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1585 * Local max aggregation: rq buckets always track the max
1586 * "requested" clamp value of its RUNNABLE tasks.
1588 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1589 bucket->value = uc_se->value;
1591 if (uc_se->value > uclamp_rq_get(rq, clamp_id))
1592 uclamp_rq_set(rq, clamp_id, uc_se->value);
1596 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1597 * is released. If this is the last task reference counting the rq's max
1598 * active clamp value, then the rq's clamp value is updated.
1600 * Both refcounted tasks and rq's cached clamp values are expected to be
1601 * always valid. If it's detected they are not, as defensive programming,
1602 * enforce the expected state and warn.
1604 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1605 enum uclamp_id clamp_id)
1607 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1608 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1609 struct uclamp_bucket *bucket;
1610 unsigned int bkt_clamp;
1611 unsigned int rq_clamp;
1613 lockdep_assert_rq_held(rq);
1616 * If sched_uclamp_used was enabled after task @p was enqueued,
1617 * we could end up with unbalanced call to uclamp_rq_dec_id().
1619 * In this case the uc_se->active flag should be false since no uclamp
1620 * accounting was performed at enqueue time and we can just return
1623 * Need to be careful of the following enqueue/dequeue ordering
1627 * // sched_uclamp_used gets enabled
1630 * // Must not decrement bucket->tasks here
1633 * where we could end up with stale data in uc_se and
1634 * bucket[uc_se->bucket_id].
1636 * The following check here eliminates the possibility of such race.
1638 if (unlikely(!uc_se->active))
1641 bucket = &uc_rq->bucket[uc_se->bucket_id];
1643 SCHED_WARN_ON(!bucket->tasks);
1644 if (likely(bucket->tasks))
1647 uc_se->active = false;
1650 * Keep "local max aggregation" simple and accept to (possibly)
1651 * overboost some RUNNABLE tasks in the same bucket.
1652 * The rq clamp bucket value is reset to its base value whenever
1653 * there are no more RUNNABLE tasks refcounting it.
1655 if (likely(bucket->tasks))
1658 rq_clamp = uclamp_rq_get(rq, clamp_id);
1660 * Defensive programming: this should never happen. If it happens,
1661 * e.g. due to future modification, warn and fixup the expected value.
1663 SCHED_WARN_ON(bucket->value > rq_clamp);
1664 if (bucket->value >= rq_clamp) {
1665 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1666 uclamp_rq_set(rq, clamp_id, bkt_clamp);
1670 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1672 enum uclamp_id clamp_id;
1675 * Avoid any overhead until uclamp is actually used by the userspace.
1677 * The condition is constructed such that a NOP is generated when
1678 * sched_uclamp_used is disabled.
1680 if (!static_branch_unlikely(&sched_uclamp_used))
1683 if (unlikely(!p->sched_class->uclamp_enabled))
1686 for_each_clamp_id(clamp_id)
1687 uclamp_rq_inc_id(rq, p, clamp_id);
1689 /* Reset clamp idle holding when there is one RUNNABLE task */
1690 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1691 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1694 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1696 enum uclamp_id clamp_id;
1699 * Avoid any overhead until uclamp is actually used by the userspace.
1701 * The condition is constructed such that a NOP is generated when
1702 * sched_uclamp_used is disabled.
1704 if (!static_branch_unlikely(&sched_uclamp_used))
1707 if (unlikely(!p->sched_class->uclamp_enabled))
1710 for_each_clamp_id(clamp_id)
1711 uclamp_rq_dec_id(rq, p, clamp_id);
1714 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1715 enum uclamp_id clamp_id)
1717 if (!p->uclamp[clamp_id].active)
1720 uclamp_rq_dec_id(rq, p, clamp_id);
1721 uclamp_rq_inc_id(rq, p, clamp_id);
1724 * Make sure to clear the idle flag if we've transiently reached 0
1725 * active tasks on rq.
1727 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1728 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1732 uclamp_update_active(struct task_struct *p)
1734 enum uclamp_id clamp_id;
1739 * Lock the task and the rq where the task is (or was) queued.
1741 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1742 * price to pay to safely serialize util_{min,max} updates with
1743 * enqueues, dequeues and migration operations.
1744 * This is the same locking schema used by __set_cpus_allowed_ptr().
1746 rq = task_rq_lock(p, &rf);
1749 * Setting the clamp bucket is serialized by task_rq_lock().
1750 * If the task is not yet RUNNABLE and its task_struct is not
1751 * affecting a valid clamp bucket, the next time it's enqueued,
1752 * it will already see the updated clamp bucket value.
1754 for_each_clamp_id(clamp_id)
1755 uclamp_rq_reinc_id(rq, p, clamp_id);
1757 task_rq_unlock(rq, p, &rf);
1760 #ifdef CONFIG_UCLAMP_TASK_GROUP
1762 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1764 struct css_task_iter it;
1765 struct task_struct *p;
1767 css_task_iter_start(css, 0, &it);
1768 while ((p = css_task_iter_next(&it)))
1769 uclamp_update_active(p);
1770 css_task_iter_end(&it);
1773 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1776 #ifdef CONFIG_SYSCTL
1777 #ifdef CONFIG_UCLAMP_TASK
1778 #ifdef CONFIG_UCLAMP_TASK_GROUP
1779 static void uclamp_update_root_tg(void)
1781 struct task_group *tg = &root_task_group;
1783 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1784 sysctl_sched_uclamp_util_min, false);
1785 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1786 sysctl_sched_uclamp_util_max, false);
1789 cpu_util_update_eff(&root_task_group.css);
1793 static void uclamp_update_root_tg(void) { }
1796 static void uclamp_sync_util_min_rt_default(void)
1798 struct task_struct *g, *p;
1801 * copy_process() sysctl_uclamp
1802 * uclamp_min_rt = X;
1803 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1804 * // link thread smp_mb__after_spinlock()
1805 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1806 * sched_post_fork() for_each_process_thread()
1807 * __uclamp_sync_rt() __uclamp_sync_rt()
1809 * Ensures that either sched_post_fork() will observe the new
1810 * uclamp_min_rt or for_each_process_thread() will observe the new
1813 read_lock(&tasklist_lock);
1814 smp_mb__after_spinlock();
1815 read_unlock(&tasklist_lock);
1818 for_each_process_thread(g, p)
1819 uclamp_update_util_min_rt_default(p);
1823 static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1824 void *buffer, size_t *lenp, loff_t *ppos)
1826 bool update_root_tg = false;
1827 int old_min, old_max, old_min_rt;
1830 guard(mutex)(&uclamp_mutex);
1832 old_min = sysctl_sched_uclamp_util_min;
1833 old_max = sysctl_sched_uclamp_util_max;
1834 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1836 result = proc_dointvec(table, write, buffer, lenp, ppos);
1842 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1843 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1844 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1850 if (old_min != sysctl_sched_uclamp_util_min) {
1851 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1852 sysctl_sched_uclamp_util_min, false);
1853 update_root_tg = true;
1855 if (old_max != sysctl_sched_uclamp_util_max) {
1856 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1857 sysctl_sched_uclamp_util_max, false);
1858 update_root_tg = true;
1861 if (update_root_tg) {
1862 static_branch_enable(&sched_uclamp_used);
1863 uclamp_update_root_tg();
1866 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1867 static_branch_enable(&sched_uclamp_used);
1868 uclamp_sync_util_min_rt_default();
1872 * We update all RUNNABLE tasks only when task groups are in use.
1873 * Otherwise, keep it simple and do just a lazy update at each next
1874 * task enqueue time.
1879 sysctl_sched_uclamp_util_min = old_min;
1880 sysctl_sched_uclamp_util_max = old_max;
1881 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1887 static int uclamp_validate(struct task_struct *p,
1888 const struct sched_attr *attr)
1890 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1891 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1893 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1894 util_min = attr->sched_util_min;
1896 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1900 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1901 util_max = attr->sched_util_max;
1903 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1907 if (util_min != -1 && util_max != -1 && util_min > util_max)
1911 * We have valid uclamp attributes; make sure uclamp is enabled.
1913 * We need to do that here, because enabling static branches is a
1914 * blocking operation which obviously cannot be done while holding
1917 static_branch_enable(&sched_uclamp_used);
1922 static bool uclamp_reset(const struct sched_attr *attr,
1923 enum uclamp_id clamp_id,
1924 struct uclamp_se *uc_se)
1926 /* Reset on sched class change for a non user-defined clamp value. */
1927 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1928 !uc_se->user_defined)
1931 /* Reset on sched_util_{min,max} == -1. */
1932 if (clamp_id == UCLAMP_MIN &&
1933 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1934 attr->sched_util_min == -1) {
1938 if (clamp_id == UCLAMP_MAX &&
1939 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1940 attr->sched_util_max == -1) {
1947 static void __setscheduler_uclamp(struct task_struct *p,
1948 const struct sched_attr *attr)
1950 enum uclamp_id clamp_id;
1952 for_each_clamp_id(clamp_id) {
1953 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1956 if (!uclamp_reset(attr, clamp_id, uc_se))
1960 * RT by default have a 100% boost value that could be modified
1963 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1964 value = sysctl_sched_uclamp_util_min_rt_default;
1966 value = uclamp_none(clamp_id);
1968 uclamp_se_set(uc_se, value, false);
1972 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1975 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1976 attr->sched_util_min != -1) {
1977 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1978 attr->sched_util_min, true);
1981 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1982 attr->sched_util_max != -1) {
1983 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1984 attr->sched_util_max, true);
1988 static void uclamp_fork(struct task_struct *p)
1990 enum uclamp_id clamp_id;
1993 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1994 * as the task is still at its early fork stages.
1996 for_each_clamp_id(clamp_id)
1997 p->uclamp[clamp_id].active = false;
1999 if (likely(!p->sched_reset_on_fork))
2002 for_each_clamp_id(clamp_id) {
2003 uclamp_se_set(&p->uclamp_req[clamp_id],
2004 uclamp_none(clamp_id), false);
2008 static void uclamp_post_fork(struct task_struct *p)
2010 uclamp_update_util_min_rt_default(p);
2013 static void __init init_uclamp_rq(struct rq *rq)
2015 enum uclamp_id clamp_id;
2016 struct uclamp_rq *uc_rq = rq->uclamp;
2018 for_each_clamp_id(clamp_id) {
2019 uc_rq[clamp_id] = (struct uclamp_rq) {
2020 .value = uclamp_none(clamp_id)
2024 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
2027 static void __init init_uclamp(void)
2029 struct uclamp_se uc_max = {};
2030 enum uclamp_id clamp_id;
2033 for_each_possible_cpu(cpu)
2034 init_uclamp_rq(cpu_rq(cpu));
2036 for_each_clamp_id(clamp_id) {
2037 uclamp_se_set(&init_task.uclamp_req[clamp_id],
2038 uclamp_none(clamp_id), false);
2041 /* System defaults allow max clamp values for both indexes */
2042 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2043 for_each_clamp_id(clamp_id) {
2044 uclamp_default[clamp_id] = uc_max;
2045 #ifdef CONFIG_UCLAMP_TASK_GROUP
2046 root_task_group.uclamp_req[clamp_id] = uc_max;
2047 root_task_group.uclamp[clamp_id] = uc_max;
2052 #else /* CONFIG_UCLAMP_TASK */
2053 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2054 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
2055 static inline int uclamp_validate(struct task_struct *p,
2056 const struct sched_attr *attr)
2060 static void __setscheduler_uclamp(struct task_struct *p,
2061 const struct sched_attr *attr) { }
2062 static inline void uclamp_fork(struct task_struct *p) { }
2063 static inline void uclamp_post_fork(struct task_struct *p) { }
2064 static inline void init_uclamp(void) { }
2065 #endif /* CONFIG_UCLAMP_TASK */
2067 bool sched_task_on_rq(struct task_struct *p)
2069 return task_on_rq_queued(p);
2072 unsigned long get_wchan(struct task_struct *p)
2074 unsigned long ip = 0;
2077 if (!p || p == current)
2080 /* Only get wchan if task is blocked and we can keep it that way. */
2081 raw_spin_lock_irq(&p->pi_lock);
2082 state = READ_ONCE(p->__state);
2083 smp_rmb(); /* see try_to_wake_up() */
2084 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2085 ip = __get_wchan(p);
2086 raw_spin_unlock_irq(&p->pi_lock);
2091 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2093 if (!(flags & ENQUEUE_NOCLOCK))
2094 update_rq_clock(rq);
2096 if (!(flags & ENQUEUE_RESTORE)) {
2097 sched_info_enqueue(rq, p);
2098 psi_enqueue(p, (flags & ENQUEUE_WAKEUP) && !(flags & ENQUEUE_MIGRATED));
2101 uclamp_rq_inc(rq, p);
2102 p->sched_class->enqueue_task(rq, p, flags);
2104 if (sched_core_enabled(rq))
2105 sched_core_enqueue(rq, p);
2108 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2110 if (sched_core_enabled(rq))
2111 sched_core_dequeue(rq, p, flags);
2113 if (!(flags & DEQUEUE_NOCLOCK))
2114 update_rq_clock(rq);
2116 if (!(flags & DEQUEUE_SAVE)) {
2117 sched_info_dequeue(rq, p);
2118 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2121 uclamp_rq_dec(rq, p);
2122 p->sched_class->dequeue_task(rq, p, flags);
2125 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2127 if (task_on_rq_migrating(p))
2128 flags |= ENQUEUE_MIGRATED;
2129 if (flags & ENQUEUE_MIGRATED)
2130 sched_mm_cid_migrate_to(rq, p);
2132 enqueue_task(rq, p, flags);
2134 p->on_rq = TASK_ON_RQ_QUEUED;
2137 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2139 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2141 dequeue_task(rq, p, flags);
2144 static inline int __normal_prio(int policy, int rt_prio, int nice)
2148 if (dl_policy(policy))
2149 prio = MAX_DL_PRIO - 1;
2150 else if (rt_policy(policy))
2151 prio = MAX_RT_PRIO - 1 - rt_prio;
2153 prio = NICE_TO_PRIO(nice);
2159 * Calculate the expected normal priority: i.e. priority
2160 * without taking RT-inheritance into account. Might be
2161 * boosted by interactivity modifiers. Changes upon fork,
2162 * setprio syscalls, and whenever the interactivity
2163 * estimator recalculates.
2165 static inline int normal_prio(struct task_struct *p)
2167 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2171 * Calculate the current priority, i.e. the priority
2172 * taken into account by the scheduler. This value might
2173 * be boosted by RT tasks, or might be boosted by
2174 * interactivity modifiers. Will be RT if the task got
2175 * RT-boosted. If not then it returns p->normal_prio.
2177 static int effective_prio(struct task_struct *p)
2179 p->normal_prio = normal_prio(p);
2181 * If we are RT tasks or we were boosted to RT priority,
2182 * keep the priority unchanged. Otherwise, update priority
2183 * to the normal priority:
2185 if (!rt_prio(p->prio))
2186 return p->normal_prio;
2191 * task_curr - is this task currently executing on a CPU?
2192 * @p: the task in question.
2194 * Return: 1 if the task is currently executing. 0 otherwise.
2196 inline int task_curr(const struct task_struct *p)
2198 return cpu_curr(task_cpu(p)) == p;
2202 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2203 * use the balance_callback list if you want balancing.
2205 * this means any call to check_class_changed() must be followed by a call to
2206 * balance_callback().
2208 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2209 const struct sched_class *prev_class,
2212 if (prev_class != p->sched_class) {
2213 if (prev_class->switched_from)
2214 prev_class->switched_from(rq, p);
2216 p->sched_class->switched_to(rq, p);
2217 } else if (oldprio != p->prio || dl_task(p))
2218 p->sched_class->prio_changed(rq, p, oldprio);
2221 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2223 if (p->sched_class == rq->curr->sched_class)
2224 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2225 else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2229 * A queue event has occurred, and we're going to schedule. In
2230 * this case, we can save a useless back to back clock update.
2232 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2233 rq_clock_skip_update(rq);
2236 static __always_inline
2237 int __task_state_match(struct task_struct *p, unsigned int state)
2239 if (READ_ONCE(p->__state) & state)
2242 #ifdef CONFIG_PREEMPT_RT
2243 if (READ_ONCE(p->saved_state) & state)
2249 static __always_inline
2250 int task_state_match(struct task_struct *p, unsigned int state)
2252 #ifdef CONFIG_PREEMPT_RT
2256 * Serialize against current_save_and_set_rtlock_wait_state() and
2257 * current_restore_rtlock_saved_state().
2259 raw_spin_lock_irq(&p->pi_lock);
2260 match = __task_state_match(p, state);
2261 raw_spin_unlock_irq(&p->pi_lock);
2265 return __task_state_match(p, state);
2270 * wait_task_inactive - wait for a thread to unschedule.
2272 * Wait for the thread to block in any of the states set in @match_state.
2273 * If it changes, i.e. @p might have woken up, then return zero. When we
2274 * succeed in waiting for @p to be off its CPU, we return a positive number
2275 * (its total switch count). If a second call a short while later returns the
2276 * same number, the caller can be sure that @p has remained unscheduled the
2279 * The caller must ensure that the task *will* unschedule sometime soon,
2280 * else this function might spin for a *long* time. This function can't
2281 * be called with interrupts off, or it may introduce deadlock with
2282 * smp_call_function() if an IPI is sent by the same process we are
2283 * waiting to become inactive.
2285 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
2287 int running, queued, match;
2294 * We do the initial early heuristics without holding
2295 * any task-queue locks at all. We'll only try to get
2296 * the runqueue lock when things look like they will
2302 * If the task is actively running on another CPU
2303 * still, just relax and busy-wait without holding
2306 * NOTE! Since we don't hold any locks, it's not
2307 * even sure that "rq" stays as the right runqueue!
2308 * But we don't care, since "task_on_cpu()" will
2309 * return false if the runqueue has changed and p
2310 * is actually now running somewhere else!
2312 while (task_on_cpu(rq, p)) {
2313 if (!task_state_match(p, match_state))
2319 * Ok, time to look more closely! We need the rq
2320 * lock now, to be *sure*. If we're wrong, we'll
2321 * just go back and repeat.
2323 rq = task_rq_lock(p, &rf);
2324 trace_sched_wait_task(p);
2325 running = task_on_cpu(rq, p);
2326 queued = task_on_rq_queued(p);
2328 if ((match = __task_state_match(p, match_state))) {
2330 * When matching on p->saved_state, consider this task
2331 * still queued so it will wait.
2335 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2337 task_rq_unlock(rq, p, &rf);
2340 * If it changed from the expected state, bail out now.
2342 if (unlikely(!ncsw))
2346 * Was it really running after all now that we
2347 * checked with the proper locks actually held?
2349 * Oops. Go back and try again..
2351 if (unlikely(running)) {
2357 * It's not enough that it's not actively running,
2358 * it must be off the runqueue _entirely_, and not
2361 * So if it was still runnable (but just not actively
2362 * running right now), it's preempted, and we should
2363 * yield - it could be a while.
2365 if (unlikely(queued)) {
2366 ktime_t to = NSEC_PER_SEC / HZ;
2368 set_current_state(TASK_UNINTERRUPTIBLE);
2369 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
2374 * Ahh, all good. It wasn't running, and it wasn't
2375 * runnable, which means that it will never become
2376 * running in the future either. We're all done!
2387 __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
2389 static int __set_cpus_allowed_ptr(struct task_struct *p,
2390 struct affinity_context *ctx);
2392 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2394 struct affinity_context ac = {
2395 .new_mask = cpumask_of(rq->cpu),
2396 .flags = SCA_MIGRATE_DISABLE,
2399 if (likely(!p->migration_disabled))
2402 if (p->cpus_ptr != &p->cpus_mask)
2406 * Violates locking rules! see comment in __do_set_cpus_allowed().
2408 __do_set_cpus_allowed(p, &ac);
2411 void migrate_disable(void)
2413 struct task_struct *p = current;
2415 if (p->migration_disabled) {
2416 p->migration_disabled++;
2421 this_rq()->nr_pinned++;
2422 p->migration_disabled = 1;
2425 EXPORT_SYMBOL_GPL(migrate_disable);
2427 void migrate_enable(void)
2429 struct task_struct *p = current;
2430 struct affinity_context ac = {
2431 .new_mask = &p->cpus_mask,
2432 .flags = SCA_MIGRATE_ENABLE,
2435 if (p->migration_disabled > 1) {
2436 p->migration_disabled--;
2440 if (WARN_ON_ONCE(!p->migration_disabled))
2444 * Ensure stop_task runs either before or after this, and that
2445 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2448 if (p->cpus_ptr != &p->cpus_mask)
2449 __set_cpus_allowed_ptr(p, &ac);
2451 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2452 * regular cpus_mask, otherwise things that race (eg.
2453 * select_fallback_rq) get confused.
2456 p->migration_disabled = 0;
2457 this_rq()->nr_pinned--;
2460 EXPORT_SYMBOL_GPL(migrate_enable);
2462 static inline bool rq_has_pinned_tasks(struct rq *rq)
2464 return rq->nr_pinned;
2468 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2469 * __set_cpus_allowed_ptr() and select_fallback_rq().
2471 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2473 /* When not in the task's cpumask, no point in looking further. */
2474 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2477 /* migrate_disabled() must be allowed to finish. */
2478 if (is_migration_disabled(p))
2479 return cpu_online(cpu);
2481 /* Non kernel threads are not allowed during either online or offline. */
2482 if (!(p->flags & PF_KTHREAD))
2483 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2485 /* KTHREAD_IS_PER_CPU is always allowed. */
2486 if (kthread_is_per_cpu(p))
2487 return cpu_online(cpu);
2489 /* Regular kernel threads don't get to stay during offline. */
2493 /* But are allowed during online. */
2494 return cpu_online(cpu);
2498 * This is how migration works:
2500 * 1) we invoke migration_cpu_stop() on the target CPU using
2502 * 2) stopper starts to run (implicitly forcing the migrated thread
2504 * 3) it checks whether the migrated task is still in the wrong runqueue.
2505 * 4) if it's in the wrong runqueue then the migration thread removes
2506 * it and puts it into the right queue.
2507 * 5) stopper completes and stop_one_cpu() returns and the migration
2512 * move_queued_task - move a queued task to new rq.
2514 * Returns (locked) new rq. Old rq's lock is released.
2516 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2517 struct task_struct *p, int new_cpu)
2519 lockdep_assert_rq_held(rq);
2521 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2522 set_task_cpu(p, new_cpu);
2525 rq = cpu_rq(new_cpu);
2528 WARN_ON_ONCE(task_cpu(p) != new_cpu);
2529 activate_task(rq, p, 0);
2530 check_preempt_curr(rq, p, 0);
2535 struct migration_arg {
2536 struct task_struct *task;
2538 struct set_affinity_pending *pending;
2542 * @refs: number of wait_for_completion()
2543 * @stop_pending: is @stop_work in use
2545 struct set_affinity_pending {
2547 unsigned int stop_pending;
2548 struct completion done;
2549 struct cpu_stop_work stop_work;
2550 struct migration_arg arg;
2554 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2555 * this because either it can't run here any more (set_cpus_allowed()
2556 * away from this CPU, or CPU going down), or because we're
2557 * attempting to rebalance this task on exec (sched_exec).
2559 * So we race with normal scheduler movements, but that's OK, as long
2560 * as the task is no longer on this CPU.
2562 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2563 struct task_struct *p, int dest_cpu)
2565 /* Affinity changed (again). */
2566 if (!is_cpu_allowed(p, dest_cpu))
2569 rq = move_queued_task(rq, rf, p, dest_cpu);
2575 * migration_cpu_stop - this will be executed by a highprio stopper thread
2576 * and performs thread migration by bumping thread off CPU then
2577 * 'pushing' onto another runqueue.
2579 static int migration_cpu_stop(void *data)
2581 struct migration_arg *arg = data;
2582 struct set_affinity_pending *pending = arg->pending;
2583 struct task_struct *p = arg->task;
2584 struct rq *rq = this_rq();
2585 bool complete = false;
2589 * The original target CPU might have gone down and we might
2590 * be on another CPU but it doesn't matter.
2592 local_irq_save(rf.flags);
2594 * We need to explicitly wake pending tasks before running
2595 * __migrate_task() such that we will not miss enforcing cpus_ptr
2596 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2598 flush_smp_call_function_queue();
2600 raw_spin_lock(&p->pi_lock);
2604 * If we were passed a pending, then ->stop_pending was set, thus
2605 * p->migration_pending must have remained stable.
2607 WARN_ON_ONCE(pending && pending != p->migration_pending);
2610 * If task_rq(p) != rq, it cannot be migrated here, because we're
2611 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2612 * we're holding p->pi_lock.
2614 if (task_rq(p) == rq) {
2615 if (is_migration_disabled(p))
2619 p->migration_pending = NULL;
2622 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2626 if (task_on_rq_queued(p)) {
2627 update_rq_clock(rq);
2628 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2630 p->wake_cpu = arg->dest_cpu;
2634 * XXX __migrate_task() can fail, at which point we might end
2635 * up running on a dodgy CPU, AFAICT this can only happen
2636 * during CPU hotplug, at which point we'll get pushed out
2637 * anyway, so it's probably not a big deal.
2640 } else if (pending) {
2642 * This happens when we get migrated between migrate_enable()'s
2643 * preempt_enable() and scheduling the stopper task. At that
2644 * point we're a regular task again and not current anymore.
2646 * A !PREEMPT kernel has a giant hole here, which makes it far
2651 * The task moved before the stopper got to run. We're holding
2652 * ->pi_lock, so the allowed mask is stable - if it got
2653 * somewhere allowed, we're done.
2655 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2656 p->migration_pending = NULL;
2662 * When migrate_enable() hits a rq mis-match we can't reliably
2663 * determine is_migration_disabled() and so have to chase after
2666 WARN_ON_ONCE(!pending->stop_pending);
2667 task_rq_unlock(rq, p, &rf);
2668 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2669 &pending->arg, &pending->stop_work);
2674 pending->stop_pending = false;
2675 task_rq_unlock(rq, p, &rf);
2678 complete_all(&pending->done);
2683 int push_cpu_stop(void *arg)
2685 struct rq *lowest_rq = NULL, *rq = this_rq();
2686 struct task_struct *p = arg;
2688 raw_spin_lock_irq(&p->pi_lock);
2689 raw_spin_rq_lock(rq);
2691 if (task_rq(p) != rq)
2694 if (is_migration_disabled(p)) {
2695 p->migration_flags |= MDF_PUSH;
2699 p->migration_flags &= ~MDF_PUSH;
2701 if (p->sched_class->find_lock_rq)
2702 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2707 // XXX validate p is still the highest prio task
2708 if (task_rq(p) == rq) {
2709 deactivate_task(rq, p, 0);
2710 set_task_cpu(p, lowest_rq->cpu);
2711 activate_task(lowest_rq, p, 0);
2712 resched_curr(lowest_rq);
2715 double_unlock_balance(rq, lowest_rq);
2718 rq->push_busy = false;
2719 raw_spin_rq_unlock(rq);
2720 raw_spin_unlock_irq(&p->pi_lock);
2727 * sched_class::set_cpus_allowed must do the below, but is not required to
2728 * actually call this function.
2730 void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
2732 if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2733 p->cpus_ptr = ctx->new_mask;
2737 cpumask_copy(&p->cpus_mask, ctx->new_mask);
2738 p->nr_cpus_allowed = cpumask_weight(ctx->new_mask);
2741 * Swap in a new user_cpus_ptr if SCA_USER flag set
2743 if (ctx->flags & SCA_USER)
2744 swap(p->user_cpus_ptr, ctx->user_mask);
2748 __do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
2750 struct rq *rq = task_rq(p);
2751 bool queued, running;
2754 * This here violates the locking rules for affinity, since we're only
2755 * supposed to change these variables while holding both rq->lock and
2758 * HOWEVER, it magically works, because ttwu() is the only code that
2759 * accesses these variables under p->pi_lock and only does so after
2760 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2761 * before finish_task().
2763 * XXX do further audits, this smells like something putrid.
2765 if (ctx->flags & SCA_MIGRATE_DISABLE)
2766 SCHED_WARN_ON(!p->on_cpu);
2768 lockdep_assert_held(&p->pi_lock);
2770 queued = task_on_rq_queued(p);
2771 running = task_current(rq, p);
2775 * Because __kthread_bind() calls this on blocked tasks without
2778 lockdep_assert_rq_held(rq);
2779 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2782 put_prev_task(rq, p);
2784 p->sched_class->set_cpus_allowed(p, ctx);
2787 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2789 set_next_task(rq, p);
2793 * Used for kthread_bind() and select_fallback_rq(), in both cases the user
2794 * affinity (if any) should be destroyed too.
2796 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2798 struct affinity_context ac = {
2799 .new_mask = new_mask,
2801 .flags = SCA_USER, /* clear the user requested mask */
2803 union cpumask_rcuhead {
2805 struct rcu_head rcu;
2808 __do_set_cpus_allowed(p, &ac);
2811 * Because this is called with p->pi_lock held, it is not possible
2812 * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
2815 kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
2818 static cpumask_t *alloc_user_cpus_ptr(int node)
2821 * See do_set_cpus_allowed() above for the rcu_head usage.
2823 int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
2825 return kmalloc_node(size, GFP_KERNEL, node);
2828 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2831 cpumask_t *user_mask;
2832 unsigned long flags;
2835 * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2836 * may differ by now due to racing.
2838 dst->user_cpus_ptr = NULL;
2841 * This check is racy and losing the race is a valid situation.
2842 * It is not worth the extra overhead of taking the pi_lock on
2845 if (data_race(!src->user_cpus_ptr))
2848 user_mask = alloc_user_cpus_ptr(node);
2853 * Use pi_lock to protect content of user_cpus_ptr
2855 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2856 * do_set_cpus_allowed().
2858 raw_spin_lock_irqsave(&src->pi_lock, flags);
2859 if (src->user_cpus_ptr) {
2860 swap(dst->user_cpus_ptr, user_mask);
2861 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2863 raw_spin_unlock_irqrestore(&src->pi_lock, flags);
2865 if (unlikely(user_mask))
2871 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2873 struct cpumask *user_mask = NULL;
2875 swap(p->user_cpus_ptr, user_mask);
2880 void release_user_cpus_ptr(struct task_struct *p)
2882 kfree(clear_user_cpus_ptr(p));
2886 * This function is wildly self concurrent; here be dragons.
2889 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2890 * designated task is enqueued on an allowed CPU. If that task is currently
2891 * running, we have to kick it out using the CPU stopper.
2893 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2896 * Initial conditions: P0->cpus_mask = [0, 1]
2900 * migrate_disable();
2902 * set_cpus_allowed_ptr(P0, [1]);
2904 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2905 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2906 * This means we need the following scheme:
2910 * migrate_disable();
2912 * set_cpus_allowed_ptr(P0, [1]);
2916 * __set_cpus_allowed_ptr();
2917 * <wakes local stopper>
2918 * `--> <woken on migration completion>
2920 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2921 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2922 * task p are serialized by p->pi_lock, which we can leverage: the one that
2923 * should come into effect at the end of the Migrate-Disable region is the last
2924 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2925 * but we still need to properly signal those waiting tasks at the appropriate
2928 * This is implemented using struct set_affinity_pending. The first
2929 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2930 * setup an instance of that struct and install it on the targeted task_struct.
2931 * Any and all further callers will reuse that instance. Those then wait for
2932 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2933 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2936 * (1) In the cases covered above. There is one more where the completion is
2937 * signaled within affine_move_task() itself: when a subsequent affinity request
2938 * occurs after the stopper bailed out due to the targeted task still being
2939 * Migrate-Disable. Consider:
2941 * Initial conditions: P0->cpus_mask = [0, 1]
2945 * migrate_disable();
2947 * set_cpus_allowed_ptr(P0, [1]);
2950 * migration_cpu_stop()
2951 * is_migration_disabled()
2953 * set_cpus_allowed_ptr(P0, [0, 1]);
2954 * <signal completion>
2957 * Note that the above is safe vs a concurrent migrate_enable(), as any
2958 * pending affinity completion is preceded by an uninstallation of
2959 * p->migration_pending done with p->pi_lock held.
2961 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2962 int dest_cpu, unsigned int flags)
2963 __releases(rq->lock)
2964 __releases(p->pi_lock)
2966 struct set_affinity_pending my_pending = { }, *pending = NULL;
2967 bool stop_pending, complete = false;
2969 /* Can the task run on the task's current CPU? If so, we're done */
2970 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2971 struct task_struct *push_task = NULL;
2973 if ((flags & SCA_MIGRATE_ENABLE) &&
2974 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2975 rq->push_busy = true;
2976 push_task = get_task_struct(p);
2980 * If there are pending waiters, but no pending stop_work,
2981 * then complete now.
2983 pending = p->migration_pending;
2984 if (pending && !pending->stop_pending) {
2985 p->migration_pending = NULL;
2989 task_rq_unlock(rq, p, rf);
2992 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2997 complete_all(&pending->done);
3002 if (!(flags & SCA_MIGRATE_ENABLE)) {
3003 /* serialized by p->pi_lock */
3004 if (!p->migration_pending) {
3005 /* Install the request */
3006 refcount_set(&my_pending.refs, 1);
3007 init_completion(&my_pending.done);
3008 my_pending.arg = (struct migration_arg) {
3010 .dest_cpu = dest_cpu,
3011 .pending = &my_pending,
3014 p->migration_pending = &my_pending;
3016 pending = p->migration_pending;
3017 refcount_inc(&pending->refs);
3019 * Affinity has changed, but we've already installed a
3020 * pending. migration_cpu_stop() *must* see this, else
3021 * we risk a completion of the pending despite having a
3022 * task on a disallowed CPU.
3024 * Serialized by p->pi_lock, so this is safe.
3026 pending->arg.dest_cpu = dest_cpu;
3029 pending = p->migration_pending;
3031 * - !MIGRATE_ENABLE:
3032 * we'll have installed a pending if there wasn't one already.
3035 * we're here because the current CPU isn't matching anymore,
3036 * the only way that can happen is because of a concurrent
3037 * set_cpus_allowed_ptr() call, which should then still be
3038 * pending completion.
3040 * Either way, we really should have a @pending here.
3042 if (WARN_ON_ONCE(!pending)) {
3043 task_rq_unlock(rq, p, rf);
3047 if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
3049 * MIGRATE_ENABLE gets here because 'p == current', but for
3050 * anything else we cannot do is_migration_disabled(), punt
3051 * and have the stopper function handle it all race-free.
3053 stop_pending = pending->stop_pending;
3055 pending->stop_pending = true;
3057 if (flags & SCA_MIGRATE_ENABLE)
3058 p->migration_flags &= ~MDF_PUSH;
3060 task_rq_unlock(rq, p, rf);
3062 if (!stop_pending) {
3063 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
3064 &pending->arg, &pending->stop_work);
3067 if (flags & SCA_MIGRATE_ENABLE)
3071 if (!is_migration_disabled(p)) {
3072 if (task_on_rq_queued(p))
3073 rq = move_queued_task(rq, rf, p, dest_cpu);
3075 if (!pending->stop_pending) {
3076 p->migration_pending = NULL;
3080 task_rq_unlock(rq, p, rf);
3083 complete_all(&pending->done);
3086 wait_for_completion(&pending->done);
3088 if (refcount_dec_and_test(&pending->refs))
3089 wake_up_var(&pending->refs); /* No UaF, just an address */
3092 * Block the original owner of &pending until all subsequent callers
3093 * have seen the completion and decremented the refcount
3095 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
3098 WARN_ON_ONCE(my_pending.stop_pending);
3104 * Called with both p->pi_lock and rq->lock held; drops both before returning.
3106 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
3107 struct affinity_context *ctx,
3109 struct rq_flags *rf)
3110 __releases(rq->lock)
3111 __releases(p->pi_lock)
3113 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
3114 const struct cpumask *cpu_valid_mask = cpu_active_mask;
3115 bool kthread = p->flags & PF_KTHREAD;
3116 unsigned int dest_cpu;
3119 update_rq_clock(rq);
3121 if (kthread || is_migration_disabled(p)) {
3123 * Kernel threads are allowed on online && !active CPUs,
3124 * however, during cpu-hot-unplug, even these might get pushed
3125 * away if not KTHREAD_IS_PER_CPU.
3127 * Specifically, migration_disabled() tasks must not fail the
3128 * cpumask_any_and_distribute() pick below, esp. so on
3129 * SCA_MIGRATE_ENABLE, otherwise we'll not call
3130 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
3132 cpu_valid_mask = cpu_online_mask;
3135 if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) {
3141 * Must re-check here, to close a race against __kthread_bind(),
3142 * sched_setaffinity() is not guaranteed to observe the flag.
3144 if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
3149 if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
3150 if (cpumask_equal(&p->cpus_mask, ctx->new_mask)) {
3151 if (ctx->flags & SCA_USER)
3152 swap(p->user_cpus_ptr, ctx->user_mask);
3156 if (WARN_ON_ONCE(p == current &&
3157 is_migration_disabled(p) &&
3158 !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
3165 * Picking a ~random cpu helps in cases where we are changing affinity
3166 * for groups of tasks (ie. cpuset), so that load balancing is not
3167 * immediately required to distribute the tasks within their new mask.
3169 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
3170 if (dest_cpu >= nr_cpu_ids) {
3175 __do_set_cpus_allowed(p, ctx);
3177 return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
3180 task_rq_unlock(rq, p, rf);
3186 * Change a given task's CPU affinity. Migrate the thread to a
3187 * proper CPU and schedule it away if the CPU it's executing on
3188 * is removed from the allowed bitmask.
3190 * NOTE: the caller must have a valid reference to the task, the
3191 * task must not exit() & deallocate itself prematurely. The
3192 * call is not atomic; no spinlocks may be held.
3194 static int __set_cpus_allowed_ptr(struct task_struct *p,
3195 struct affinity_context *ctx)
3200 rq = task_rq_lock(p, &rf);
3202 * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
3205 if (p->user_cpus_ptr &&
3206 !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
3207 cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr))
3208 ctx->new_mask = rq->scratch_mask;
3210 return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf);
3213 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3215 struct affinity_context ac = {
3216 .new_mask = new_mask,
3220 return __set_cpus_allowed_ptr(p, &ac);
3222 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
3225 * Change a given task's CPU affinity to the intersection of its current
3226 * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
3227 * If user_cpus_ptr is defined, use it as the basis for restricting CPU
3228 * affinity or use cpu_online_mask instead.
3230 * If the resulting mask is empty, leave the affinity unchanged and return
3233 static int restrict_cpus_allowed_ptr(struct task_struct *p,
3234 struct cpumask *new_mask,
3235 const struct cpumask *subset_mask)
3237 struct affinity_context ac = {
3238 .new_mask = new_mask,
3245 rq = task_rq_lock(p, &rf);
3248 * Forcefully restricting the affinity of a deadline task is
3249 * likely to cause problems, so fail and noisily override the
3252 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
3257 if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) {
3262 return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf);
3265 task_rq_unlock(rq, p, &rf);
3270 * Restrict the CPU affinity of task @p so that it is a subset of
3271 * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
3272 * old affinity mask. If the resulting mask is empty, we warn and walk
3273 * up the cpuset hierarchy until we find a suitable mask.
3275 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3277 cpumask_var_t new_mask;
3278 const struct cpumask *override_mask = task_cpu_possible_mask(p);
3280 alloc_cpumask_var(&new_mask, GFP_KERNEL);
3283 * __migrate_task() can fail silently in the face of concurrent
3284 * offlining of the chosen destination CPU, so take the hotplug
3285 * lock to ensure that the migration succeeds.
3288 if (!cpumask_available(new_mask))
3291 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3295 * We failed to find a valid subset of the affinity mask for the
3296 * task, so override it based on its cpuset hierarchy.
3298 cpuset_cpus_allowed(p, new_mask);
3299 override_mask = new_mask;
3302 if (printk_ratelimit()) {
3303 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3304 task_pid_nr(p), p->comm,
3305 cpumask_pr_args(override_mask));
3308 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3311 free_cpumask_var(new_mask);
3315 __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
3318 * Restore the affinity of a task @p which was previously restricted by a
3319 * call to force_compatible_cpus_allowed_ptr().
3321 * It is the caller's responsibility to serialise this with any calls to
3322 * force_compatible_cpus_allowed_ptr(@p).
3324 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3326 struct affinity_context ac = {
3327 .new_mask = task_user_cpus(p),
3333 * Try to restore the old affinity mask with __sched_setaffinity().
3334 * Cpuset masking will be done there too.
3336 ret = __sched_setaffinity(p, &ac);
3340 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3342 #ifdef CONFIG_SCHED_DEBUG
3343 unsigned int state = READ_ONCE(p->__state);
3346 * We should never call set_task_cpu() on a blocked task,
3347 * ttwu() will sort out the placement.
3349 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3352 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3353 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3354 * time relying on p->on_rq.
3356 WARN_ON_ONCE(state == TASK_RUNNING &&
3357 p->sched_class == &fair_sched_class &&
3358 (p->on_rq && !task_on_rq_migrating(p)));
3360 #ifdef CONFIG_LOCKDEP
3362 * The caller should hold either p->pi_lock or rq->lock, when changing
3363 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3365 * sched_move_task() holds both and thus holding either pins the cgroup,
3368 * Furthermore, all task_rq users should acquire both locks, see
3371 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3372 lockdep_is_held(__rq_lockp(task_rq(p)))));
3375 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3377 WARN_ON_ONCE(!cpu_online(new_cpu));
3379 WARN_ON_ONCE(is_migration_disabled(p));
3382 trace_sched_migrate_task(p, new_cpu);
3384 if (task_cpu(p) != new_cpu) {
3385 if (p->sched_class->migrate_task_rq)
3386 p->sched_class->migrate_task_rq(p, new_cpu);
3387 p->se.nr_migrations++;
3389 sched_mm_cid_migrate_from(p);
3390 perf_event_task_migrate(p);
3393 __set_task_cpu(p, new_cpu);
3396 #ifdef CONFIG_NUMA_BALANCING
3397 static void __migrate_swap_task(struct task_struct *p, int cpu)
3399 if (task_on_rq_queued(p)) {
3400 struct rq *src_rq, *dst_rq;
3401 struct rq_flags srf, drf;
3403 src_rq = task_rq(p);
3404 dst_rq = cpu_rq(cpu);
3406 rq_pin_lock(src_rq, &srf);
3407 rq_pin_lock(dst_rq, &drf);
3409 deactivate_task(src_rq, p, 0);
3410 set_task_cpu(p, cpu);
3411 activate_task(dst_rq, p, 0);
3412 check_preempt_curr(dst_rq, p, 0);
3414 rq_unpin_lock(dst_rq, &drf);
3415 rq_unpin_lock(src_rq, &srf);
3419 * Task isn't running anymore; make it appear like we migrated
3420 * it before it went to sleep. This means on wakeup we make the
3421 * previous CPU our target instead of where it really is.
3427 struct migration_swap_arg {
3428 struct task_struct *src_task, *dst_task;
3429 int src_cpu, dst_cpu;
3432 static int migrate_swap_stop(void *data)
3434 struct migration_swap_arg *arg = data;
3435 struct rq *src_rq, *dst_rq;
3437 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3440 src_rq = cpu_rq(arg->src_cpu);
3441 dst_rq = cpu_rq(arg->dst_cpu);
3443 guard(double_raw_spinlock)(&arg->src_task->pi_lock, &arg->dst_task->pi_lock);
3444 guard(double_rq_lock)(src_rq, dst_rq);
3446 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3449 if (task_cpu(arg->src_task) != arg->src_cpu)
3452 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3455 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3458 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3459 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3465 * Cross migrate two tasks
3467 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3468 int target_cpu, int curr_cpu)
3470 struct migration_swap_arg arg;
3473 arg = (struct migration_swap_arg){
3475 .src_cpu = curr_cpu,
3477 .dst_cpu = target_cpu,
3480 if (arg.src_cpu == arg.dst_cpu)
3484 * These three tests are all lockless; this is OK since all of them
3485 * will be re-checked with proper locks held further down the line.
3487 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3490 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3493 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3496 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3497 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3502 #endif /* CONFIG_NUMA_BALANCING */
3505 * kick_process - kick a running thread to enter/exit the kernel
3506 * @p: the to-be-kicked thread
3508 * Cause a process which is running on another CPU to enter
3509 * kernel-mode, without any delay. (to get signals handled.)
3511 * NOTE: this function doesn't have to take the runqueue lock,
3512 * because all it wants to ensure is that the remote task enters
3513 * the kernel. If the IPI races and the task has been migrated
3514 * to another CPU then no harm is done and the purpose has been
3517 void kick_process(struct task_struct *p)
3523 if ((cpu != smp_processor_id()) && task_curr(p))
3524 smp_send_reschedule(cpu);
3527 EXPORT_SYMBOL_GPL(kick_process);
3530 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3532 * A few notes on cpu_active vs cpu_online:
3534 * - cpu_active must be a subset of cpu_online
3536 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3537 * see __set_cpus_allowed_ptr(). At this point the newly online
3538 * CPU isn't yet part of the sched domains, and balancing will not
3541 * - on CPU-down we clear cpu_active() to mask the sched domains and
3542 * avoid the load balancer to place new tasks on the to be removed
3543 * CPU. Existing tasks will remain running there and will be taken
3546 * This means that fallback selection must not select !active CPUs.
3547 * And can assume that any active CPU must be online. Conversely
3548 * select_task_rq() below may allow selection of !active CPUs in order
3549 * to satisfy the above rules.
3551 static int select_fallback_rq(int cpu, struct task_struct *p)
3553 int nid = cpu_to_node(cpu);
3554 const struct cpumask *nodemask = NULL;
3555 enum { cpuset, possible, fail } state = cpuset;
3559 * If the node that the CPU is on has been offlined, cpu_to_node()
3560 * will return -1. There is no CPU on the node, and we should
3561 * select the CPU on the other node.
3564 nodemask = cpumask_of_node(nid);
3566 /* Look for allowed, online CPU in same node. */
3567 for_each_cpu(dest_cpu, nodemask) {
3568 if (is_cpu_allowed(p, dest_cpu))
3574 /* Any allowed, online CPU? */
3575 for_each_cpu(dest_cpu, p->cpus_ptr) {
3576 if (!is_cpu_allowed(p, dest_cpu))
3582 /* No more Mr. Nice Guy. */
3585 if (cpuset_cpus_allowed_fallback(p)) {
3592 * XXX When called from select_task_rq() we only
3593 * hold p->pi_lock and again violate locking order.
3595 * More yuck to audit.
3597 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3607 if (state != cpuset) {
3609 * Don't tell them about moving exiting tasks or
3610 * kernel threads (both mm NULL), since they never
3613 if (p->mm && printk_ratelimit()) {
3614 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3615 task_pid_nr(p), p->comm, cpu);
3623 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3626 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3628 lockdep_assert_held(&p->pi_lock);
3630 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3631 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3633 cpu = cpumask_any(p->cpus_ptr);
3636 * In order not to call set_task_cpu() on a blocking task we need
3637 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3640 * Since this is common to all placement strategies, this lives here.
3642 * [ this allows ->select_task() to simply return task_cpu(p) and
3643 * not worry about this generic constraint ]
3645 if (unlikely(!is_cpu_allowed(p, cpu)))
3646 cpu = select_fallback_rq(task_cpu(p), p);
3651 void sched_set_stop_task(int cpu, struct task_struct *stop)
3653 static struct lock_class_key stop_pi_lock;
3654 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3655 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3659 * Make it appear like a SCHED_FIFO task, its something
3660 * userspace knows about and won't get confused about.
3662 * Also, it will make PI more or less work without too
3663 * much confusion -- but then, stop work should not
3664 * rely on PI working anyway.
3666 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3668 stop->sched_class = &stop_sched_class;
3671 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3672 * adjust the effective priority of a task. As a result,
3673 * rt_mutex_setprio() can trigger (RT) balancing operations,
3674 * which can then trigger wakeups of the stop thread to push
3675 * around the current task.
3677 * The stop task itself will never be part of the PI-chain, it
3678 * never blocks, therefore that ->pi_lock recursion is safe.
3679 * Tell lockdep about this by placing the stop->pi_lock in its
3682 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3685 cpu_rq(cpu)->stop = stop;
3689 * Reset it back to a normal scheduling class so that
3690 * it can die in pieces.
3692 old_stop->sched_class = &rt_sched_class;
3696 #else /* CONFIG_SMP */
3698 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3699 struct affinity_context *ctx)
3701 return set_cpus_allowed_ptr(p, ctx->new_mask);
3704 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3706 static inline bool rq_has_pinned_tasks(struct rq *rq)
3711 static inline cpumask_t *alloc_user_cpus_ptr(int node)
3716 #endif /* !CONFIG_SMP */
3719 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3723 if (!schedstat_enabled())
3729 if (cpu == rq->cpu) {
3730 __schedstat_inc(rq->ttwu_local);
3731 __schedstat_inc(p->stats.nr_wakeups_local);
3733 struct sched_domain *sd;
3735 __schedstat_inc(p->stats.nr_wakeups_remote);
3738 for_each_domain(rq->cpu, sd) {
3739 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3740 __schedstat_inc(sd->ttwu_wake_remote);
3746 if (wake_flags & WF_MIGRATED)
3747 __schedstat_inc(p->stats.nr_wakeups_migrate);
3748 #endif /* CONFIG_SMP */
3750 __schedstat_inc(rq->ttwu_count);
3751 __schedstat_inc(p->stats.nr_wakeups);
3753 if (wake_flags & WF_SYNC)
3754 __schedstat_inc(p->stats.nr_wakeups_sync);
3758 * Mark the task runnable.
3760 static inline void ttwu_do_wakeup(struct task_struct *p)
3762 WRITE_ONCE(p->__state, TASK_RUNNING);
3763 trace_sched_wakeup(p);
3767 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3768 struct rq_flags *rf)
3770 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3772 lockdep_assert_rq_held(rq);
3774 if (p->sched_contributes_to_load)
3775 rq->nr_uninterruptible--;
3778 if (wake_flags & WF_MIGRATED)
3779 en_flags |= ENQUEUE_MIGRATED;
3783 delayacct_blkio_end(p);
3784 atomic_dec(&task_rq(p)->nr_iowait);
3787 activate_task(rq, p, en_flags);
3788 check_preempt_curr(rq, p, wake_flags);
3793 if (p->sched_class->task_woken) {
3795 * Our task @p is fully woken up and running; so it's safe to
3796 * drop the rq->lock, hereafter rq is only used for statistics.
3798 rq_unpin_lock(rq, rf);
3799 p->sched_class->task_woken(rq, p);
3800 rq_repin_lock(rq, rf);
3803 if (rq->idle_stamp) {
3804 u64 delta = rq_clock(rq) - rq->idle_stamp;
3805 u64 max = 2*rq->max_idle_balance_cost;
3807 update_avg(&rq->avg_idle, delta);
3809 if (rq->avg_idle > max)
3812 rq->wake_stamp = jiffies;
3813 rq->wake_avg_idle = rq->avg_idle / 2;
3821 * Consider @p being inside a wait loop:
3824 * set_current_state(TASK_UNINTERRUPTIBLE);
3831 * __set_current_state(TASK_RUNNING);
3833 * between set_current_state() and schedule(). In this case @p is still
3834 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3837 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3838 * then schedule() must still happen and p->state can be changed to
3839 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3840 * need to do a full wakeup with enqueue.
3842 * Returns: %true when the wakeup is done,
3845 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3851 rq = __task_rq_lock(p, &rf);
3852 if (task_on_rq_queued(p)) {
3853 if (!task_on_cpu(rq, p)) {
3855 * When on_rq && !on_cpu the task is preempted, see if
3856 * it should preempt the task that is current now.
3858 update_rq_clock(rq);
3859 check_preempt_curr(rq, p, wake_flags);
3864 __task_rq_unlock(rq, &rf);
3870 void sched_ttwu_pending(void *arg)
3872 struct llist_node *llist = arg;
3873 struct rq *rq = this_rq();
3874 struct task_struct *p, *t;
3880 rq_lock_irqsave(rq, &rf);
3881 update_rq_clock(rq);
3883 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3884 if (WARN_ON_ONCE(p->on_cpu))
3885 smp_cond_load_acquire(&p->on_cpu, !VAL);
3887 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3888 set_task_cpu(p, cpu_of(rq));
3890 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3894 * Must be after enqueueing at least once task such that
3895 * idle_cpu() does not observe a false-negative -- if it does,
3896 * it is possible for select_idle_siblings() to stack a number
3897 * of tasks on this CPU during that window.
3899 * It is ok to clear ttwu_pending when another task pending.
3900 * We will receive IPI after local irq enabled and then enqueue it.
3901 * Since now nr_running > 0, idle_cpu() will always get correct result.
3903 WRITE_ONCE(rq->ttwu_pending, 0);
3904 rq_unlock_irqrestore(rq, &rf);
3908 * Prepare the scene for sending an IPI for a remote smp_call
3910 * Returns true if the caller can proceed with sending the IPI.
3911 * Returns false otherwise.
3913 bool call_function_single_prep_ipi(int cpu)
3915 if (set_nr_if_polling(cpu_rq(cpu)->idle)) {
3916 trace_sched_wake_idle_without_ipi(cpu);
3924 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3925 * necessary. The wakee CPU on receipt of the IPI will queue the task
3926 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3927 * of the wakeup instead of the waker.
3929 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3931 struct rq *rq = cpu_rq(cpu);
3933 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3935 WRITE_ONCE(rq->ttwu_pending, 1);
3936 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3939 void wake_up_if_idle(int cpu)
3941 struct rq *rq = cpu_rq(cpu);
3944 if (is_idle_task(rcu_dereference(rq->curr))) {
3945 guard(rq_lock_irqsave)(rq);
3946 if (is_idle_task(rq->curr))
3951 bool cpus_share_cache(int this_cpu, int that_cpu)
3953 if (this_cpu == that_cpu)
3956 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3959 static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3962 * Do not complicate things with the async wake_list while the CPU is
3965 if (!cpu_active(cpu))
3968 /* Ensure the task will still be allowed to run on the CPU. */
3969 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3973 * If the CPU does not share cache, then queue the task on the
3974 * remote rqs wakelist to avoid accessing remote data.
3976 if (!cpus_share_cache(smp_processor_id(), cpu))
3979 if (cpu == smp_processor_id())
3983 * If the wakee cpu is idle, or the task is descheduling and the
3984 * only running task on the CPU, then use the wakelist to offload
3985 * the task activation to the idle (or soon-to-be-idle) CPU as
3986 * the current CPU is likely busy. nr_running is checked to
3987 * avoid unnecessary task stacking.
3989 * Note that we can only get here with (wakee) p->on_rq=0,
3990 * p->on_cpu can be whatever, we've done the dequeue, so
3991 * the wakee has been accounted out of ->nr_running.
3993 if (!cpu_rq(cpu)->nr_running)
3999 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4001 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
4002 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
4003 __ttwu_queue_wakelist(p, cpu, wake_flags);
4010 #else /* !CONFIG_SMP */
4012 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
4017 #endif /* CONFIG_SMP */
4019 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
4021 struct rq *rq = cpu_rq(cpu);
4024 if (ttwu_queue_wakelist(p, cpu, wake_flags))
4028 update_rq_clock(rq);
4029 ttwu_do_activate(rq, p, wake_flags, &rf);
4034 * Invoked from try_to_wake_up() to check whether the task can be woken up.
4036 * The caller holds p::pi_lock if p != current or has preemption
4037 * disabled when p == current.
4039 * The rules of PREEMPT_RT saved_state:
4041 * The related locking code always holds p::pi_lock when updating
4042 * p::saved_state, which means the code is fully serialized in both cases.
4044 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
4045 * bits set. This allows to distinguish all wakeup scenarios.
4047 static __always_inline
4048 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
4052 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
4053 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
4054 state != TASK_RTLOCK_WAIT);
4057 *success = !!(match = __task_state_match(p, state));
4059 #ifdef CONFIG_PREEMPT_RT
4061 * Saved state preserves the task state across blocking on
4062 * an RT lock. If the state matches, set p::saved_state to
4063 * TASK_RUNNING, but do not wake the task because it waits
4064 * for a lock wakeup. Also indicate success because from
4065 * the regular waker's point of view this has succeeded.
4067 * After acquiring the lock the task will restore p::__state
4068 * from p::saved_state which ensures that the regular
4069 * wakeup is not lost. The restore will also set
4070 * p::saved_state to TASK_RUNNING so any further tests will
4071 * not result in false positives vs. @success
4074 p->saved_state = TASK_RUNNING;
4080 * Notes on Program-Order guarantees on SMP systems.
4084 * The basic program-order guarantee on SMP systems is that when a task [t]
4085 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
4086 * execution on its new CPU [c1].
4088 * For migration (of runnable tasks) this is provided by the following means:
4090 * A) UNLOCK of the rq(c0)->lock scheduling out task t
4091 * B) migration for t is required to synchronize *both* rq(c0)->lock and
4092 * rq(c1)->lock (if not at the same time, then in that order).
4093 * C) LOCK of the rq(c1)->lock scheduling in task
4095 * Release/acquire chaining guarantees that B happens after A and C after B.
4096 * Note: the CPU doing B need not be c0 or c1
4105 * UNLOCK rq(0)->lock
4107 * LOCK rq(0)->lock // orders against CPU0
4109 * UNLOCK rq(0)->lock
4113 * UNLOCK rq(1)->lock
4115 * LOCK rq(1)->lock // orders against CPU2
4118 * UNLOCK rq(1)->lock
4121 * BLOCKING -- aka. SLEEP + WAKEUP
4123 * For blocking we (obviously) need to provide the same guarantee as for
4124 * migration. However the means are completely different as there is no lock
4125 * chain to provide order. Instead we do:
4127 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
4128 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
4132 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
4134 * LOCK rq(0)->lock LOCK X->pi_lock
4137 * smp_store_release(X->on_cpu, 0);
4139 * smp_cond_load_acquire(&X->on_cpu, !VAL);
4145 * X->state = RUNNING
4146 * UNLOCK rq(2)->lock
4148 * LOCK rq(2)->lock // orders against CPU1
4151 * UNLOCK rq(2)->lock
4154 * UNLOCK rq(0)->lock
4157 * However, for wakeups there is a second guarantee we must provide, namely we
4158 * must ensure that CONDITION=1 done by the caller can not be reordered with
4159 * accesses to the task state; see try_to_wake_up() and set_current_state().
4163 * try_to_wake_up - wake up a thread
4164 * @p: the thread to be awakened
4165 * @state: the mask of task states that can be woken
4166 * @wake_flags: wake modifier flags (WF_*)
4168 * Conceptually does:
4170 * If (@state & @p->state) @p->state = TASK_RUNNING.
4172 * If the task was not queued/runnable, also place it back on a runqueue.
4174 * This function is atomic against schedule() which would dequeue the task.
4176 * It issues a full memory barrier before accessing @p->state, see the comment
4177 * with set_current_state().
4179 * Uses p->pi_lock to serialize against concurrent wake-ups.
4181 * Relies on p->pi_lock stabilizing:
4184 * - p->sched_task_group
4185 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4187 * Tries really hard to only take one task_rq(p)->lock for performance.
4188 * Takes rq->lock in:
4189 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4190 * - ttwu_queue() -- new rq, for enqueue of the task;
4191 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4193 * As a consequence we race really badly with just about everything. See the
4194 * many memory barriers and their comments for details.
4196 * Return: %true if @p->state changes (an actual wakeup was done),
4199 int try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4202 int cpu, success = 0;
4206 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4207 * == smp_processor_id()'. Together this means we can special
4208 * case the whole 'p->on_rq && ttwu_runnable()' case below
4209 * without taking any locks.
4212 * - we rely on Program-Order guarantees for all the ordering,
4213 * - we're serialized against set_special_state() by virtue of
4214 * it disabling IRQs (this allows not taking ->pi_lock).
4216 if (!ttwu_state_match(p, state, &success))
4219 trace_sched_waking(p);
4225 * If we are going to wake up a thread waiting for CONDITION we
4226 * need to ensure that CONDITION=1 done by the caller can not be
4227 * reordered with p->state check below. This pairs with smp_store_mb()
4228 * in set_current_state() that the waiting thread does.
4230 scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
4231 smp_mb__after_spinlock();
4232 if (!ttwu_state_match(p, state, &success))
4235 trace_sched_waking(p);
4238 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4239 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4240 * in smp_cond_load_acquire() below.
4242 * sched_ttwu_pending() try_to_wake_up()
4243 * STORE p->on_rq = 1 LOAD p->state
4246 * __schedule() (switch to task 'p')
4247 * LOCK rq->lock smp_rmb();
4248 * smp_mb__after_spinlock();
4252 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4254 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4255 * __schedule(). See the comment for smp_mb__after_spinlock().
4257 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4260 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4265 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4266 * possible to, falsely, observe p->on_cpu == 0.
4268 * One must be running (->on_cpu == 1) in order to remove oneself
4269 * from the runqueue.
4271 * __schedule() (switch to task 'p') try_to_wake_up()
4272 * STORE p->on_cpu = 1 LOAD p->on_rq
4275 * __schedule() (put 'p' to sleep)
4276 * LOCK rq->lock smp_rmb();
4277 * smp_mb__after_spinlock();
4278 * STORE p->on_rq = 0 LOAD p->on_cpu
4280 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4281 * __schedule(). See the comment for smp_mb__after_spinlock().
4283 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4284 * schedule()'s deactivate_task() has 'happened' and p will no longer
4285 * care about it's own p->state. See the comment in __schedule().
4287 smp_acquire__after_ctrl_dep();
4290 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4291 * == 0), which means we need to do an enqueue, change p->state to
4292 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4293 * enqueue, such as ttwu_queue_wakelist().
4295 WRITE_ONCE(p->__state, TASK_WAKING);
4298 * If the owning (remote) CPU is still in the middle of schedule() with
4299 * this task as prev, considering queueing p on the remote CPUs wake_list
4300 * which potentially sends an IPI instead of spinning on p->on_cpu to
4301 * let the waker make forward progress. This is safe because IRQs are
4302 * disabled and the IPI will deliver after on_cpu is cleared.
4304 * Ensure we load task_cpu(p) after p->on_cpu:
4306 * set_task_cpu(p, cpu);
4307 * STORE p->cpu = @cpu
4308 * __schedule() (switch to task 'p')
4310 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4311 * STORE p->on_cpu = 1 LOAD p->cpu
4313 * to ensure we observe the correct CPU on which the task is currently
4316 if (smp_load_acquire(&p->on_cpu) &&
4317 ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4321 * If the owning (remote) CPU is still in the middle of schedule() with
4322 * this task as prev, wait until it's done referencing the task.
4324 * Pairs with the smp_store_release() in finish_task().
4326 * This ensures that tasks getting woken will be fully ordered against
4327 * their previous state and preserve Program Order.
4329 smp_cond_load_acquire(&p->on_cpu, !VAL);
4331 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4332 if (task_cpu(p) != cpu) {
4334 delayacct_blkio_end(p);
4335 atomic_dec(&task_rq(p)->nr_iowait);
4338 wake_flags |= WF_MIGRATED;
4339 psi_ttwu_dequeue(p);
4340 set_task_cpu(p, cpu);
4344 #endif /* CONFIG_SMP */
4346 ttwu_queue(p, cpu, wake_flags);
4350 ttwu_stat(p, task_cpu(p), wake_flags);
4355 static bool __task_needs_rq_lock(struct task_struct *p)
4357 unsigned int state = READ_ONCE(p->__state);
4360 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4361 * the task is blocked. Make sure to check @state since ttwu() can drop
4362 * locks at the end, see ttwu_queue_wakelist().
4364 if (state == TASK_RUNNING || state == TASK_WAKING)
4368 * Ensure we load p->on_rq after p->__state, otherwise it would be
4369 * possible to, falsely, observe p->on_rq == 0.
4371 * See try_to_wake_up() for a longer comment.
4379 * Ensure the task has finished __schedule() and will not be referenced
4380 * anymore. Again, see try_to_wake_up() for a longer comment.
4383 smp_cond_load_acquire(&p->on_cpu, !VAL);
4390 * task_call_func - Invoke a function on task in fixed state
4391 * @p: Process for which the function is to be invoked, can be @current.
4392 * @func: Function to invoke.
4393 * @arg: Argument to function.
4395 * Fix the task in it's current state by avoiding wakeups and or rq operations
4396 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4397 * to work out what the state is, if required. Given that @func can be invoked
4398 * with a runqueue lock held, it had better be quite lightweight.
4401 * Whatever @func returns
4403 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4405 struct rq *rq = NULL;
4409 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4411 if (__task_needs_rq_lock(p))
4412 rq = __task_rq_lock(p, &rf);
4415 * At this point the task is pinned; either:
4416 * - blocked and we're holding off wakeups (pi->lock)
4417 * - woken, and we're holding off enqueue (rq->lock)
4418 * - queued, and we're holding off schedule (rq->lock)
4419 * - running, and we're holding off de-schedule (rq->lock)
4421 * The called function (@func) can use: task_curr(), p->on_rq and
4422 * p->__state to differentiate between these states.
4429 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4434 * cpu_curr_snapshot - Return a snapshot of the currently running task
4435 * @cpu: The CPU on which to snapshot the task.
4437 * Returns the task_struct pointer of the task "currently" running on
4438 * the specified CPU. If the same task is running on that CPU throughout,
4439 * the return value will be a pointer to that task's task_struct structure.
4440 * If the CPU did any context switches even vaguely concurrently with the
4441 * execution of this function, the return value will be a pointer to the
4442 * task_struct structure of a randomly chosen task that was running on
4443 * that CPU somewhere around the time that this function was executing.
4445 * If the specified CPU was offline, the return value is whatever it
4446 * is, perhaps a pointer to the task_struct structure of that CPU's idle
4447 * task, but there is no guarantee. Callers wishing a useful return
4448 * value must take some action to ensure that the specified CPU remains
4449 * online throughout.
4451 * This function executes full memory barriers before and after fetching
4452 * the pointer, which permits the caller to confine this function's fetch
4453 * with respect to the caller's accesses to other shared variables.
4455 struct task_struct *cpu_curr_snapshot(int cpu)
4457 struct task_struct *t;
4459 smp_mb(); /* Pairing determined by caller's synchronization design. */
4460 t = rcu_dereference(cpu_curr(cpu));
4461 smp_mb(); /* Pairing determined by caller's synchronization design. */
4466 * wake_up_process - Wake up a specific process
4467 * @p: The process to be woken up.
4469 * Attempt to wake up the nominated process and move it to the set of runnable
4472 * Return: 1 if the process was woken up, 0 if it was already running.
4474 * This function executes a full memory barrier before accessing the task state.
4476 int wake_up_process(struct task_struct *p)
4478 return try_to_wake_up(p, TASK_NORMAL, 0);
4480 EXPORT_SYMBOL(wake_up_process);
4482 int wake_up_state(struct task_struct *p, unsigned int state)
4484 return try_to_wake_up(p, state, 0);
4488 * Perform scheduler related setup for a newly forked process p.
4489 * p is forked by current.
4491 * __sched_fork() is basic setup used by init_idle() too:
4493 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4498 p->se.exec_start = 0;
4499 p->se.sum_exec_runtime = 0;
4500 p->se.prev_sum_exec_runtime = 0;
4501 p->se.nr_migrations = 0;
4504 p->se.slice = sysctl_sched_base_slice;
4505 INIT_LIST_HEAD(&p->se.group_node);
4507 #ifdef CONFIG_FAIR_GROUP_SCHED
4508 p->se.cfs_rq = NULL;
4511 #ifdef CONFIG_SCHEDSTATS
4512 /* Even if schedstat is disabled, there should not be garbage */
4513 memset(&p->stats, 0, sizeof(p->stats));
4516 RB_CLEAR_NODE(&p->dl.rb_node);
4517 init_dl_task_timer(&p->dl);
4518 init_dl_inactive_task_timer(&p->dl);
4519 __dl_clear_params(p);
4521 INIT_LIST_HEAD(&p->rt.run_list);
4523 p->rt.time_slice = sched_rr_timeslice;
4527 #ifdef CONFIG_PREEMPT_NOTIFIERS
4528 INIT_HLIST_HEAD(&p->preempt_notifiers);
4531 #ifdef CONFIG_COMPACTION
4532 p->capture_control = NULL;
4534 init_numa_balancing(clone_flags, p);
4536 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4537 p->migration_pending = NULL;
4539 init_sched_mm_cid(p);
4542 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4544 #ifdef CONFIG_NUMA_BALANCING
4546 int sysctl_numa_balancing_mode;
4548 static void __set_numabalancing_state(bool enabled)
4551 static_branch_enable(&sched_numa_balancing);
4553 static_branch_disable(&sched_numa_balancing);
4556 void set_numabalancing_state(bool enabled)
4559 sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4561 sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4562 __set_numabalancing_state(enabled);
4565 #ifdef CONFIG_PROC_SYSCTL
4566 static void reset_memory_tiering(void)
4568 struct pglist_data *pgdat;
4570 for_each_online_pgdat(pgdat) {
4571 pgdat->nbp_threshold = 0;
4572 pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
4573 pgdat->nbp_th_start = jiffies_to_msecs(jiffies);
4577 static int sysctl_numa_balancing(struct ctl_table *table, int write,
4578 void *buffer, size_t *lenp, loff_t *ppos)
4582 int state = sysctl_numa_balancing_mode;
4584 if (write && !capable(CAP_SYS_ADMIN))
4589 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4593 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4594 (state & NUMA_BALANCING_MEMORY_TIERING))
4595 reset_memory_tiering();
4596 sysctl_numa_balancing_mode = state;
4597 __set_numabalancing_state(state);
4604 #ifdef CONFIG_SCHEDSTATS
4606 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4608 static void set_schedstats(bool enabled)
4611 static_branch_enable(&sched_schedstats);
4613 static_branch_disable(&sched_schedstats);
4616 void force_schedstat_enabled(void)
4618 if (!schedstat_enabled()) {
4619 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4620 static_branch_enable(&sched_schedstats);
4624 static int __init setup_schedstats(char *str)
4630 if (!strcmp(str, "enable")) {
4631 set_schedstats(true);
4633 } else if (!strcmp(str, "disable")) {
4634 set_schedstats(false);
4639 pr_warn("Unable to parse schedstats=\n");
4643 __setup("schedstats=", setup_schedstats);
4645 #ifdef CONFIG_PROC_SYSCTL
4646 static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4647 size_t *lenp, loff_t *ppos)
4651 int state = static_branch_likely(&sched_schedstats);
4653 if (write && !capable(CAP_SYS_ADMIN))
4658 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4662 set_schedstats(state);
4665 #endif /* CONFIG_PROC_SYSCTL */
4666 #endif /* CONFIG_SCHEDSTATS */
4668 #ifdef CONFIG_SYSCTL
4669 static struct ctl_table sched_core_sysctls[] = {
4670 #ifdef CONFIG_SCHEDSTATS
4672 .procname = "sched_schedstats",
4674 .maxlen = sizeof(unsigned int),
4676 .proc_handler = sysctl_schedstats,
4677 .extra1 = SYSCTL_ZERO,
4678 .extra2 = SYSCTL_ONE,
4680 #endif /* CONFIG_SCHEDSTATS */
4681 #ifdef CONFIG_UCLAMP_TASK
4683 .procname = "sched_util_clamp_min",
4684 .data = &sysctl_sched_uclamp_util_min,
4685 .maxlen = sizeof(unsigned int),
4687 .proc_handler = sysctl_sched_uclamp_handler,
4690 .procname = "sched_util_clamp_max",
4691 .data = &sysctl_sched_uclamp_util_max,
4692 .maxlen = sizeof(unsigned int),
4694 .proc_handler = sysctl_sched_uclamp_handler,
4697 .procname = "sched_util_clamp_min_rt_default",
4698 .data = &sysctl_sched_uclamp_util_min_rt_default,
4699 .maxlen = sizeof(unsigned int),
4701 .proc_handler = sysctl_sched_uclamp_handler,
4703 #endif /* CONFIG_UCLAMP_TASK */
4704 #ifdef CONFIG_NUMA_BALANCING
4706 .procname = "numa_balancing",
4707 .data = NULL, /* filled in by handler */
4708 .maxlen = sizeof(unsigned int),
4710 .proc_handler = sysctl_numa_balancing,
4711 .extra1 = SYSCTL_ZERO,
4712 .extra2 = SYSCTL_FOUR,
4714 #endif /* CONFIG_NUMA_BALANCING */
4717 static int __init sched_core_sysctl_init(void)
4719 register_sysctl_init("kernel", sched_core_sysctls);
4722 late_initcall(sched_core_sysctl_init);
4723 #endif /* CONFIG_SYSCTL */
4726 * fork()/clone()-time setup:
4728 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4730 __sched_fork(clone_flags, p);
4732 * We mark the process as NEW here. This guarantees that
4733 * nobody will actually run it, and a signal or other external
4734 * event cannot wake it up and insert it on the runqueue either.
4736 p->__state = TASK_NEW;
4739 * Make sure we do not leak PI boosting priority to the child.
4741 p->prio = current->normal_prio;
4746 * Revert to default priority/policy on fork if requested.
4748 if (unlikely(p->sched_reset_on_fork)) {
4749 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4750 p->policy = SCHED_NORMAL;
4751 p->static_prio = NICE_TO_PRIO(0);
4753 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4754 p->static_prio = NICE_TO_PRIO(0);
4756 p->prio = p->normal_prio = p->static_prio;
4757 set_load_weight(p, false);
4760 * We don't need the reset flag anymore after the fork. It has
4761 * fulfilled its duty:
4763 p->sched_reset_on_fork = 0;
4766 if (dl_prio(p->prio))
4768 else if (rt_prio(p->prio))
4769 p->sched_class = &rt_sched_class;
4771 p->sched_class = &fair_sched_class;
4773 init_entity_runnable_average(&p->se);
4776 #ifdef CONFIG_SCHED_INFO
4777 if (likely(sched_info_on()))
4778 memset(&p->sched_info, 0, sizeof(p->sched_info));
4780 #if defined(CONFIG_SMP)
4783 init_task_preempt_count(p);
4785 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4786 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4791 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4793 unsigned long flags;
4796 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4797 * required yet, but lockdep gets upset if rules are violated.
4799 raw_spin_lock_irqsave(&p->pi_lock, flags);
4800 #ifdef CONFIG_CGROUP_SCHED
4802 struct task_group *tg;
4803 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4804 struct task_group, css);
4805 tg = autogroup_task_group(p, tg);
4806 p->sched_task_group = tg;
4811 * We're setting the CPU for the first time, we don't migrate,
4812 * so use __set_task_cpu().
4814 __set_task_cpu(p, smp_processor_id());
4815 if (p->sched_class->task_fork)
4816 p->sched_class->task_fork(p);
4817 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4820 void sched_post_fork(struct task_struct *p)
4822 uclamp_post_fork(p);
4825 unsigned long to_ratio(u64 period, u64 runtime)
4827 if (runtime == RUNTIME_INF)
4831 * Doing this here saves a lot of checks in all
4832 * the calling paths, and returning zero seems
4833 * safe for them anyway.
4838 return div64_u64(runtime << BW_SHIFT, period);
4842 * wake_up_new_task - wake up a newly created task for the first time.
4844 * This function will do some initial scheduler statistics housekeeping
4845 * that must be done for every newly created context, then puts the task
4846 * on the runqueue and wakes it.
4848 void wake_up_new_task(struct task_struct *p)
4853 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4854 WRITE_ONCE(p->__state, TASK_RUNNING);
4857 * Fork balancing, do it here and not earlier because:
4858 * - cpus_ptr can change in the fork path
4859 * - any previously selected CPU might disappear through hotplug
4861 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4862 * as we're not fully set-up yet.
4864 p->recent_used_cpu = task_cpu(p);
4866 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4868 rq = __task_rq_lock(p, &rf);
4869 update_rq_clock(rq);
4870 post_init_entity_util_avg(p);
4872 activate_task(rq, p, ENQUEUE_NOCLOCK);
4873 trace_sched_wakeup_new(p);
4874 check_preempt_curr(rq, p, WF_FORK);
4876 if (p->sched_class->task_woken) {
4878 * Nothing relies on rq->lock after this, so it's fine to
4881 rq_unpin_lock(rq, &rf);
4882 p->sched_class->task_woken(rq, p);
4883 rq_repin_lock(rq, &rf);
4886 task_rq_unlock(rq, p, &rf);
4889 #ifdef CONFIG_PREEMPT_NOTIFIERS
4891 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4893 void preempt_notifier_inc(void)
4895 static_branch_inc(&preempt_notifier_key);
4897 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4899 void preempt_notifier_dec(void)
4901 static_branch_dec(&preempt_notifier_key);
4903 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4906 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4907 * @notifier: notifier struct to register
4909 void preempt_notifier_register(struct preempt_notifier *notifier)
4911 if (!static_branch_unlikely(&preempt_notifier_key))
4912 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4914 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4916 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4919 * preempt_notifier_unregister - no longer interested in preemption notifications
4920 * @notifier: notifier struct to unregister
4922 * This is *not* safe to call from within a preemption notifier.
4924 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4926 hlist_del(¬ifier->link);
4928 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4930 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4932 struct preempt_notifier *notifier;
4934 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4935 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4938 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4940 if (static_branch_unlikely(&preempt_notifier_key))
4941 __fire_sched_in_preempt_notifiers(curr);
4945 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4946 struct task_struct *next)
4948 struct preempt_notifier *notifier;
4950 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4951 notifier->ops->sched_out(notifier, next);
4954 static __always_inline void
4955 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4956 struct task_struct *next)
4958 if (static_branch_unlikely(&preempt_notifier_key))
4959 __fire_sched_out_preempt_notifiers(curr, next);
4962 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4964 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4969 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4970 struct task_struct *next)
4974 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4976 static inline void prepare_task(struct task_struct *next)
4980 * Claim the task as running, we do this before switching to it
4981 * such that any running task will have this set.
4983 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4984 * its ordering comment.
4986 WRITE_ONCE(next->on_cpu, 1);
4990 static inline void finish_task(struct task_struct *prev)
4994 * This must be the very last reference to @prev from this CPU. After
4995 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4996 * must ensure this doesn't happen until the switch is completely
4999 * In particular, the load of prev->state in finish_task_switch() must
5000 * happen before this.
5002 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
5004 smp_store_release(&prev->on_cpu, 0);
5010 static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
5012 void (*func)(struct rq *rq);
5013 struct balance_callback *next;
5015 lockdep_assert_rq_held(rq);
5018 func = (void (*)(struct rq *))head->func;
5027 static void balance_push(struct rq *rq);
5030 * balance_push_callback is a right abuse of the callback interface and plays
5031 * by significantly different rules.
5033 * Where the normal balance_callback's purpose is to be ran in the same context
5034 * that queued it (only later, when it's safe to drop rq->lock again),
5035 * balance_push_callback is specifically targeted at __schedule().
5037 * This abuse is tolerated because it places all the unlikely/odd cases behind
5038 * a single test, namely: rq->balance_callback == NULL.
5040 struct balance_callback balance_push_callback = {
5042 .func = balance_push,
5045 static inline struct balance_callback *
5046 __splice_balance_callbacks(struct rq *rq, bool split)
5048 struct balance_callback *head = rq->balance_callback;
5053 lockdep_assert_rq_held(rq);
5055 * Must not take balance_push_callback off the list when
5056 * splice_balance_callbacks() and balance_callbacks() are not
5057 * in the same rq->lock section.
5059 * In that case it would be possible for __schedule() to interleave
5060 * and observe the list empty.
5062 if (split && head == &balance_push_callback)
5065 rq->balance_callback = NULL;
5070 static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5072 return __splice_balance_callbacks(rq, true);
5075 static void __balance_callbacks(struct rq *rq)
5077 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
5080 static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5082 unsigned long flags;
5084 if (unlikely(head)) {
5085 raw_spin_rq_lock_irqsave(rq, flags);
5086 do_balance_callbacks(rq, head);
5087 raw_spin_rq_unlock_irqrestore(rq, flags);
5093 static inline void __balance_callbacks(struct rq *rq)
5097 static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
5102 static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
5109 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
5112 * Since the runqueue lock will be released by the next
5113 * task (which is an invalid locking op but in the case
5114 * of the scheduler it's an obvious special-case), so we
5115 * do an early lockdep release here:
5117 rq_unpin_lock(rq, rf);
5118 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
5119 #ifdef CONFIG_DEBUG_SPINLOCK
5120 /* this is a valid case when another task releases the spinlock */
5121 rq_lockp(rq)->owner = next;
5125 static inline void finish_lock_switch(struct rq *rq)
5128 * If we are tracking spinlock dependencies then we have to
5129 * fix up the runqueue lock - which gets 'carried over' from
5130 * prev into current:
5132 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
5133 __balance_callbacks(rq);
5134 raw_spin_rq_unlock_irq(rq);
5138 * NOP if the arch has not defined these:
5141 #ifndef prepare_arch_switch
5142 # define prepare_arch_switch(next) do { } while (0)
5145 #ifndef finish_arch_post_lock_switch
5146 # define finish_arch_post_lock_switch() do { } while (0)
5149 static inline void kmap_local_sched_out(void)
5151 #ifdef CONFIG_KMAP_LOCAL
5152 if (unlikely(current->kmap_ctrl.idx))
5153 __kmap_local_sched_out();
5157 static inline void kmap_local_sched_in(void)
5159 #ifdef CONFIG_KMAP_LOCAL
5160 if (unlikely(current->kmap_ctrl.idx))
5161 __kmap_local_sched_in();
5166 * prepare_task_switch - prepare to switch tasks
5167 * @rq: the runqueue preparing to switch
5168 * @prev: the current task that is being switched out
5169 * @next: the task we are going to switch to.
5171 * This is called with the rq lock held and interrupts off. It must
5172 * be paired with a subsequent finish_task_switch after the context
5175 * prepare_task_switch sets up locking and calls architecture specific
5179 prepare_task_switch(struct rq *rq, struct task_struct *prev,
5180 struct task_struct *next)
5182 kcov_prepare_switch(prev);
5183 sched_info_switch(rq, prev, next);
5184 perf_event_task_sched_out(prev, next);
5186 fire_sched_out_preempt_notifiers(prev, next);
5187 kmap_local_sched_out();
5189 prepare_arch_switch(next);
5193 * finish_task_switch - clean up after a task-switch
5194 * @prev: the thread we just switched away from.
5196 * finish_task_switch must be called after the context switch, paired
5197 * with a prepare_task_switch call before the context switch.
5198 * finish_task_switch will reconcile locking set up by prepare_task_switch,
5199 * and do any other architecture-specific cleanup actions.
5201 * Note that we may have delayed dropping an mm in context_switch(). If
5202 * so, we finish that here outside of the runqueue lock. (Doing it
5203 * with the lock held can cause deadlocks; see schedule() for
5206 * The context switch have flipped the stack from under us and restored the
5207 * local variables which were saved when this task called schedule() in the
5208 * past. prev == current is still correct but we need to recalculate this_rq
5209 * because prev may have moved to another CPU.
5211 static struct rq *finish_task_switch(struct task_struct *prev)
5212 __releases(rq->lock)
5214 struct rq *rq = this_rq();
5215 struct mm_struct *mm = rq->prev_mm;
5216 unsigned int prev_state;
5219 * The previous task will have left us with a preempt_count of 2
5220 * because it left us after:
5223 * preempt_disable(); // 1
5225 * raw_spin_lock_irq(&rq->lock) // 2
5227 * Also, see FORK_PREEMPT_COUNT.
5229 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5230 "corrupted preempt_count: %s/%d/0x%x\n",
5231 current->comm, current->pid, preempt_count()))
5232 preempt_count_set(FORK_PREEMPT_COUNT);
5237 * A task struct has one reference for the use as "current".
5238 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5239 * schedule one last time. The schedule call will never return, and
5240 * the scheduled task must drop that reference.
5242 * We must observe prev->state before clearing prev->on_cpu (in
5243 * finish_task), otherwise a concurrent wakeup can get prev
5244 * running on another CPU and we could rave with its RUNNING -> DEAD
5245 * transition, resulting in a double drop.
5247 prev_state = READ_ONCE(prev->__state);
5248 vtime_task_switch(prev);
5249 perf_event_task_sched_in(prev, current);
5251 tick_nohz_task_switch();
5252 finish_lock_switch(rq);
5253 finish_arch_post_lock_switch();
5254 kcov_finish_switch(current);
5256 * kmap_local_sched_out() is invoked with rq::lock held and
5257 * interrupts disabled. There is no requirement for that, but the
5258 * sched out code does not have an interrupt enabled section.
5259 * Restoring the maps on sched in does not require interrupts being
5262 kmap_local_sched_in();
5264 fire_sched_in_preempt_notifiers(current);
5266 * When switching through a kernel thread, the loop in
5267 * membarrier_{private,global}_expedited() may have observed that
5268 * kernel thread and not issued an IPI. It is therefore possible to
5269 * schedule between user->kernel->user threads without passing though
5270 * switch_mm(). Membarrier requires a barrier after storing to
5271 * rq->curr, before returning to userspace, so provide them here:
5273 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5274 * provided by mmdrop_lazy_tlb(),
5275 * - a sync_core for SYNC_CORE.
5278 membarrier_mm_sync_core_before_usermode(mm);
5279 mmdrop_lazy_tlb_sched(mm);
5282 if (unlikely(prev_state == TASK_DEAD)) {
5283 if (prev->sched_class->task_dead)
5284 prev->sched_class->task_dead(prev);
5286 /* Task is done with its stack. */
5287 put_task_stack(prev);
5289 put_task_struct_rcu_user(prev);
5296 * schedule_tail - first thing a freshly forked thread must call.
5297 * @prev: the thread we just switched away from.
5299 asmlinkage __visible void schedule_tail(struct task_struct *prev)
5300 __releases(rq->lock)
5303 * New tasks start with FORK_PREEMPT_COUNT, see there and
5304 * finish_task_switch() for details.
5306 * finish_task_switch() will drop rq->lock() and lower preempt_count
5307 * and the preempt_enable() will end up enabling preemption (on
5308 * PREEMPT_COUNT kernels).
5311 finish_task_switch(prev);
5314 if (current->set_child_tid)
5315 put_user(task_pid_vnr(current), current->set_child_tid);
5317 calculate_sigpending();
5321 * context_switch - switch to the new MM and the new thread's register state.
5323 static __always_inline struct rq *
5324 context_switch(struct rq *rq, struct task_struct *prev,
5325 struct task_struct *next, struct rq_flags *rf)
5327 prepare_task_switch(rq, prev, next);
5330 * For paravirt, this is coupled with an exit in switch_to to
5331 * combine the page table reload and the switch backend into
5334 arch_start_context_switch(prev);
5337 * kernel -> kernel lazy + transfer active
5338 * user -> kernel lazy + mmgrab_lazy_tlb() active
5340 * kernel -> user switch + mmdrop_lazy_tlb() active
5341 * user -> user switch
5343 * switch_mm_cid() needs to be updated if the barriers provided
5344 * by context_switch() are modified.
5346 if (!next->mm) { // to kernel
5347 enter_lazy_tlb(prev->active_mm, next);
5349 next->active_mm = prev->active_mm;
5350 if (prev->mm) // from user
5351 mmgrab_lazy_tlb(prev->active_mm);
5353 prev->active_mm = NULL;
5355 membarrier_switch_mm(rq, prev->active_mm, next->mm);
5357 * sys_membarrier() requires an smp_mb() between setting
5358 * rq->curr / membarrier_switch_mm() and returning to userspace.
5360 * The below provides this either through switch_mm(), or in
5361 * case 'prev->active_mm == next->mm' through
5362 * finish_task_switch()'s mmdrop().
5364 switch_mm_irqs_off(prev->active_mm, next->mm, next);
5365 lru_gen_use_mm(next->mm);
5367 if (!prev->mm) { // from kernel
5368 /* will mmdrop_lazy_tlb() in finish_task_switch(). */
5369 rq->prev_mm = prev->active_mm;
5370 prev->active_mm = NULL;
5374 /* switch_mm_cid() requires the memory barriers above. */
5375 switch_mm_cid(rq, prev, next);
5377 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5379 prepare_lock_switch(rq, next, rf);
5381 /* Here we just switch the register state and the stack. */
5382 switch_to(prev, next, prev);
5385 return finish_task_switch(prev);
5389 * nr_running and nr_context_switches:
5391 * externally visible scheduler statistics: current number of runnable
5392 * threads, total number of context switches performed since bootup.
5394 unsigned int nr_running(void)
5396 unsigned int i, sum = 0;
5398 for_each_online_cpu(i)
5399 sum += cpu_rq(i)->nr_running;
5405 * Check if only the current task is running on the CPU.
5407 * Caution: this function does not check that the caller has disabled
5408 * preemption, thus the result might have a time-of-check-to-time-of-use
5409 * race. The caller is responsible to use it correctly, for example:
5411 * - from a non-preemptible section (of course)
5413 * - from a thread that is bound to a single CPU
5415 * - in a loop with very short iterations (e.g. a polling loop)
5417 bool single_task_running(void)
5419 return raw_rq()->nr_running == 1;
5421 EXPORT_SYMBOL(single_task_running);
5423 unsigned long long nr_context_switches_cpu(int cpu)
5425 return cpu_rq(cpu)->nr_switches;
5428 unsigned long long nr_context_switches(void)
5431 unsigned long long sum = 0;
5433 for_each_possible_cpu(i)
5434 sum += cpu_rq(i)->nr_switches;
5440 * Consumers of these two interfaces, like for example the cpuidle menu
5441 * governor, are using nonsensical data. Preferring shallow idle state selection
5442 * for a CPU that has IO-wait which might not even end up running the task when
5443 * it does become runnable.
5446 unsigned int nr_iowait_cpu(int cpu)
5448 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5452 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5454 * The idea behind IO-wait account is to account the idle time that we could
5455 * have spend running if it were not for IO. That is, if we were to improve the
5456 * storage performance, we'd have a proportional reduction in IO-wait time.
5458 * This all works nicely on UP, where, when a task blocks on IO, we account
5459 * idle time as IO-wait, because if the storage were faster, it could've been
5460 * running and we'd not be idle.
5462 * This has been extended to SMP, by doing the same for each CPU. This however
5465 * Imagine for instance the case where two tasks block on one CPU, only the one
5466 * CPU will have IO-wait accounted, while the other has regular idle. Even
5467 * though, if the storage were faster, both could've ran at the same time,
5468 * utilising both CPUs.
5470 * This means, that when looking globally, the current IO-wait accounting on
5471 * SMP is a lower bound, by reason of under accounting.
5473 * Worse, since the numbers are provided per CPU, they are sometimes
5474 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5475 * associated with any one particular CPU, it can wake to another CPU than it
5476 * blocked on. This means the per CPU IO-wait number is meaningless.
5478 * Task CPU affinities can make all that even more 'interesting'.
5481 unsigned int nr_iowait(void)
5483 unsigned int i, sum = 0;
5485 for_each_possible_cpu(i)
5486 sum += nr_iowait_cpu(i);
5494 * sched_exec - execve() is a valuable balancing opportunity, because at
5495 * this point the task has the smallest effective memory and cache footprint.
5497 void sched_exec(void)
5499 struct task_struct *p = current;
5500 struct migration_arg arg;
5503 scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
5504 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5505 if (dest_cpu == smp_processor_id())
5508 if (unlikely(!cpu_active(dest_cpu)))
5511 arg = (struct migration_arg){ p, dest_cpu };
5513 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5518 DEFINE_PER_CPU(struct kernel_stat, kstat);
5519 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5521 EXPORT_PER_CPU_SYMBOL(kstat);
5522 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5525 * The function fair_sched_class.update_curr accesses the struct curr
5526 * and its field curr->exec_start; when called from task_sched_runtime(),
5527 * we observe a high rate of cache misses in practice.
5528 * Prefetching this data results in improved performance.
5530 static inline void prefetch_curr_exec_start(struct task_struct *p)
5532 #ifdef CONFIG_FAIR_GROUP_SCHED
5533 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5535 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5538 prefetch(&curr->exec_start);
5542 * Return accounted runtime for the task.
5543 * In case the task is currently running, return the runtime plus current's
5544 * pending runtime that have not been accounted yet.
5546 unsigned long long task_sched_runtime(struct task_struct *p)
5552 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5554 * 64-bit doesn't need locks to atomically read a 64-bit value.
5555 * So we have a optimization chance when the task's delta_exec is 0.
5556 * Reading ->on_cpu is racy, but this is ok.
5558 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5559 * If we race with it entering CPU, unaccounted time is 0. This is
5560 * indistinguishable from the read occurring a few cycles earlier.
5561 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5562 * been accounted, so we're correct here as well.
5564 if (!p->on_cpu || !task_on_rq_queued(p))
5565 return p->se.sum_exec_runtime;
5568 rq = task_rq_lock(p, &rf);
5570 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5571 * project cycles that may never be accounted to this
5572 * thread, breaking clock_gettime().
5574 if (task_current(rq, p) && task_on_rq_queued(p)) {
5575 prefetch_curr_exec_start(p);
5576 update_rq_clock(rq);
5577 p->sched_class->update_curr(rq);
5579 ns = p->se.sum_exec_runtime;
5580 task_rq_unlock(rq, p, &rf);
5585 #ifdef CONFIG_SCHED_DEBUG
5586 static u64 cpu_resched_latency(struct rq *rq)
5588 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5589 u64 resched_latency, now = rq_clock(rq);
5590 static bool warned_once;
5592 if (sysctl_resched_latency_warn_once && warned_once)
5595 if (!need_resched() || !latency_warn_ms)
5598 if (system_state == SYSTEM_BOOTING)
5601 if (!rq->last_seen_need_resched_ns) {
5602 rq->last_seen_need_resched_ns = now;
5603 rq->ticks_without_resched = 0;
5607 rq->ticks_without_resched++;
5608 resched_latency = now - rq->last_seen_need_resched_ns;
5609 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5614 return resched_latency;
5617 static int __init setup_resched_latency_warn_ms(char *str)
5621 if ((kstrtol(str, 0, &val))) {
5622 pr_warn("Unable to set resched_latency_warn_ms\n");
5626 sysctl_resched_latency_warn_ms = val;
5629 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5631 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5632 #endif /* CONFIG_SCHED_DEBUG */
5635 * This function gets called by the timer code, with HZ frequency.
5636 * We call it with interrupts disabled.
5638 void scheduler_tick(void)
5640 int cpu = smp_processor_id();
5641 struct rq *rq = cpu_rq(cpu);
5642 struct task_struct *curr = rq->curr;
5644 unsigned long thermal_pressure;
5645 u64 resched_latency;
5647 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5648 arch_scale_freq_tick();
5654 update_rq_clock(rq);
5655 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5656 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5657 curr->sched_class->task_tick(rq, curr, 0);
5658 if (sched_feat(LATENCY_WARN))
5659 resched_latency = cpu_resched_latency(rq);
5660 calc_global_load_tick(rq);
5661 sched_core_tick(rq);
5662 task_tick_mm_cid(rq, curr);
5666 if (sched_feat(LATENCY_WARN) && resched_latency)
5667 resched_latency_warn(cpu, resched_latency);
5669 perf_event_task_tick();
5671 if (curr->flags & PF_WQ_WORKER)
5672 wq_worker_tick(curr);
5675 rq->idle_balance = idle_cpu(cpu);
5676 trigger_load_balance(rq);
5680 #ifdef CONFIG_NO_HZ_FULL
5685 struct delayed_work work;
5687 /* Values for ->state, see diagram below. */
5688 #define TICK_SCHED_REMOTE_OFFLINE 0
5689 #define TICK_SCHED_REMOTE_OFFLINING 1
5690 #define TICK_SCHED_REMOTE_RUNNING 2
5693 * State diagram for ->state:
5696 * TICK_SCHED_REMOTE_OFFLINE
5699 * | | sched_tick_remote()
5702 * +--TICK_SCHED_REMOTE_OFFLINING
5705 * sched_tick_start() | | sched_tick_stop()
5708 * TICK_SCHED_REMOTE_RUNNING
5711 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5712 * and sched_tick_start() are happy to leave the state in RUNNING.
5715 static struct tick_work __percpu *tick_work_cpu;
5717 static void sched_tick_remote(struct work_struct *work)
5719 struct delayed_work *dwork = to_delayed_work(work);
5720 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5721 int cpu = twork->cpu;
5722 struct rq *rq = cpu_rq(cpu);
5726 * Handle the tick only if it appears the remote CPU is running in full
5727 * dynticks mode. The check is racy by nature, but missing a tick or
5728 * having one too much is no big deal because the scheduler tick updates
5729 * statistics and checks timeslices in a time-independent way, regardless
5730 * of when exactly it is running.
5732 if (tick_nohz_tick_stopped_cpu(cpu)) {
5733 guard(rq_lock_irq)(rq);
5734 struct task_struct *curr = rq->curr;
5736 if (cpu_online(cpu)) {
5737 update_rq_clock(rq);
5739 if (!is_idle_task(curr)) {
5741 * Make sure the next tick runs within a
5742 * reasonable amount of time.
5744 u64 delta = rq_clock_task(rq) - curr->se.exec_start;
5745 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5747 curr->sched_class->task_tick(rq, curr, 0);
5749 calc_load_nohz_remote(rq);
5754 * Run the remote tick once per second (1Hz). This arbitrary
5755 * frequency is large enough to avoid overload but short enough
5756 * to keep scheduler internal stats reasonably up to date. But
5757 * first update state to reflect hotplug activity if required.
5759 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5760 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5761 if (os == TICK_SCHED_REMOTE_RUNNING)
5762 queue_delayed_work(system_unbound_wq, dwork, HZ);
5765 static void sched_tick_start(int cpu)
5768 struct tick_work *twork;
5770 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5773 WARN_ON_ONCE(!tick_work_cpu);
5775 twork = per_cpu_ptr(tick_work_cpu, cpu);
5776 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5777 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5778 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5780 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5781 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5785 #ifdef CONFIG_HOTPLUG_CPU
5786 static void sched_tick_stop(int cpu)
5788 struct tick_work *twork;
5791 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5794 WARN_ON_ONCE(!tick_work_cpu);
5796 twork = per_cpu_ptr(tick_work_cpu, cpu);
5797 /* There cannot be competing actions, but don't rely on stop-machine. */
5798 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5799 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5800 /* Don't cancel, as this would mess up the state machine. */
5802 #endif /* CONFIG_HOTPLUG_CPU */
5804 int __init sched_tick_offload_init(void)
5806 tick_work_cpu = alloc_percpu(struct tick_work);
5807 BUG_ON(!tick_work_cpu);
5811 #else /* !CONFIG_NO_HZ_FULL */
5812 static inline void sched_tick_start(int cpu) { }
5813 static inline void sched_tick_stop(int cpu) { }
5816 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5817 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5819 * If the value passed in is equal to the current preempt count
5820 * then we just disabled preemption. Start timing the latency.
5822 static inline void preempt_latency_start(int val)
5824 if (preempt_count() == val) {
5825 unsigned long ip = get_lock_parent_ip();
5826 #ifdef CONFIG_DEBUG_PREEMPT
5827 current->preempt_disable_ip = ip;
5829 trace_preempt_off(CALLER_ADDR0, ip);
5833 void preempt_count_add(int val)
5835 #ifdef CONFIG_DEBUG_PREEMPT
5839 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5842 __preempt_count_add(val);
5843 #ifdef CONFIG_DEBUG_PREEMPT
5845 * Spinlock count overflowing soon?
5847 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5850 preempt_latency_start(val);
5852 EXPORT_SYMBOL(preempt_count_add);
5853 NOKPROBE_SYMBOL(preempt_count_add);
5856 * If the value passed in equals to the current preempt count
5857 * then we just enabled preemption. Stop timing the latency.
5859 static inline void preempt_latency_stop(int val)
5861 if (preempt_count() == val)
5862 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5865 void preempt_count_sub(int val)
5867 #ifdef CONFIG_DEBUG_PREEMPT
5871 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5874 * Is the spinlock portion underflowing?
5876 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5877 !(preempt_count() & PREEMPT_MASK)))
5881 preempt_latency_stop(val);
5882 __preempt_count_sub(val);
5884 EXPORT_SYMBOL(preempt_count_sub);
5885 NOKPROBE_SYMBOL(preempt_count_sub);
5888 static inline void preempt_latency_start(int val) { }
5889 static inline void preempt_latency_stop(int val) { }
5892 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5894 #ifdef CONFIG_DEBUG_PREEMPT
5895 return p->preempt_disable_ip;
5902 * Print scheduling while atomic bug:
5904 static noinline void __schedule_bug(struct task_struct *prev)
5906 /* Save this before calling printk(), since that will clobber it */
5907 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5909 if (oops_in_progress)
5912 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5913 prev->comm, prev->pid, preempt_count());
5915 debug_show_held_locks(prev);
5917 if (irqs_disabled())
5918 print_irqtrace_events(prev);
5919 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5920 && in_atomic_preempt_off()) {
5921 pr_err("Preemption disabled at:");
5922 print_ip_sym(KERN_ERR, preempt_disable_ip);
5924 check_panic_on_warn("scheduling while atomic");
5927 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5931 * Various schedule()-time debugging checks and statistics:
5933 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5935 #ifdef CONFIG_SCHED_STACK_END_CHECK
5936 if (task_stack_end_corrupted(prev))
5937 panic("corrupted stack end detected inside scheduler\n");
5939 if (task_scs_end_corrupted(prev))
5940 panic("corrupted shadow stack detected inside scheduler\n");
5943 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5944 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5945 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5946 prev->comm, prev->pid, prev->non_block_count);
5948 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5952 if (unlikely(in_atomic_preempt_off())) {
5953 __schedule_bug(prev);
5954 preempt_count_set(PREEMPT_DISABLED);
5957 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5959 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5961 schedstat_inc(this_rq()->sched_count);
5964 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5965 struct rq_flags *rf)
5968 const struct sched_class *class;
5970 * We must do the balancing pass before put_prev_task(), such
5971 * that when we release the rq->lock the task is in the same
5972 * state as before we took rq->lock.
5974 * We can terminate the balance pass as soon as we know there is
5975 * a runnable task of @class priority or higher.
5977 for_class_range(class, prev->sched_class, &idle_sched_class) {
5978 if (class->balance(rq, prev, rf))
5983 put_prev_task(rq, prev);
5987 * Pick up the highest-prio task:
5989 static inline struct task_struct *
5990 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5992 const struct sched_class *class;
5993 struct task_struct *p;
5996 * Optimization: we know that if all tasks are in the fair class we can
5997 * call that function directly, but only if the @prev task wasn't of a
5998 * higher scheduling class, because otherwise those lose the
5999 * opportunity to pull in more work from other CPUs.
6001 if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
6002 rq->nr_running == rq->cfs.h_nr_running)) {
6004 p = pick_next_task_fair(rq, prev, rf);
6005 if (unlikely(p == RETRY_TASK))
6008 /* Assume the next prioritized class is idle_sched_class */
6010 put_prev_task(rq, prev);
6011 p = pick_next_task_idle(rq);
6018 put_prev_task_balance(rq, prev, rf);
6020 for_each_class(class) {
6021 p = class->pick_next_task(rq);
6026 BUG(); /* The idle class should always have a runnable task. */
6029 #ifdef CONFIG_SCHED_CORE
6030 static inline bool is_task_rq_idle(struct task_struct *t)
6032 return (task_rq(t)->idle == t);
6035 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
6037 return is_task_rq_idle(a) || (a->core_cookie == cookie);
6040 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
6042 if (is_task_rq_idle(a) || is_task_rq_idle(b))
6045 return a->core_cookie == b->core_cookie;
6048 static inline struct task_struct *pick_task(struct rq *rq)
6050 const struct sched_class *class;
6051 struct task_struct *p;
6053 for_each_class(class) {
6054 p = class->pick_task(rq);
6059 BUG(); /* The idle class should always have a runnable task. */
6062 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
6064 static void queue_core_balance(struct rq *rq);
6066 static struct task_struct *
6067 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6069 struct task_struct *next, *p, *max = NULL;
6070 const struct cpumask *smt_mask;
6071 bool fi_before = false;
6072 bool core_clock_updated = (rq == rq->core);
6073 unsigned long cookie;
6074 int i, cpu, occ = 0;
6078 if (!sched_core_enabled(rq))
6079 return __pick_next_task(rq, prev, rf);
6083 /* Stopper task is switching into idle, no need core-wide selection. */
6084 if (cpu_is_offline(cpu)) {
6086 * Reset core_pick so that we don't enter the fastpath when
6087 * coming online. core_pick would already be migrated to
6088 * another cpu during offline.
6090 rq->core_pick = NULL;
6091 return __pick_next_task(rq, prev, rf);
6095 * If there were no {en,de}queues since we picked (IOW, the task
6096 * pointers are all still valid), and we haven't scheduled the last
6097 * pick yet, do so now.
6099 * rq->core_pick can be NULL if no selection was made for a CPU because
6100 * it was either offline or went offline during a sibling's core-wide
6101 * selection. In this case, do a core-wide selection.
6103 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
6104 rq->core->core_pick_seq != rq->core_sched_seq &&
6106 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
6108 next = rq->core_pick;
6110 put_prev_task(rq, prev);
6111 set_next_task(rq, next);
6114 rq->core_pick = NULL;
6118 put_prev_task_balance(rq, prev, rf);
6120 smt_mask = cpu_smt_mask(cpu);
6121 need_sync = !!rq->core->core_cookie;
6124 rq->core->core_cookie = 0UL;
6125 if (rq->core->core_forceidle_count) {
6126 if (!core_clock_updated) {
6127 update_rq_clock(rq->core);
6128 core_clock_updated = true;
6130 sched_core_account_forceidle(rq);
6131 /* reset after accounting force idle */
6132 rq->core->core_forceidle_start = 0;
6133 rq->core->core_forceidle_count = 0;
6134 rq->core->core_forceidle_occupation = 0;
6140 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
6142 * @task_seq guards the task state ({en,de}queues)
6143 * @pick_seq is the @task_seq we did a selection on
6144 * @sched_seq is the @pick_seq we scheduled
6146 * However, preemptions can cause multiple picks on the same task set.
6147 * 'Fix' this by also increasing @task_seq for every pick.
6149 rq->core->core_task_seq++;
6152 * Optimize for common case where this CPU has no cookies
6153 * and there are no cookied tasks running on siblings.
6156 next = pick_task(rq);
6157 if (!next->core_cookie) {
6158 rq->core_pick = NULL;
6160 * For robustness, update the min_vruntime_fi for
6161 * unconstrained picks as well.
6163 WARN_ON_ONCE(fi_before);
6164 task_vruntime_update(rq, next, false);
6170 * For each thread: do the regular task pick and find the max prio task
6173 * Tie-break prio towards the current CPU
6175 for_each_cpu_wrap(i, smt_mask, cpu) {
6179 * Current cpu always has its clock updated on entrance to
6180 * pick_next_task(). If the current cpu is not the core,
6181 * the core may also have been updated above.
6183 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
6184 update_rq_clock(rq_i);
6186 p = rq_i->core_pick = pick_task(rq_i);
6187 if (!max || prio_less(max, p, fi_before))
6191 cookie = rq->core->core_cookie = max->core_cookie;
6194 * For each thread: try and find a runnable task that matches @max or
6197 for_each_cpu(i, smt_mask) {
6199 p = rq_i->core_pick;
6201 if (!cookie_equals(p, cookie)) {
6204 p = sched_core_find(rq_i, cookie);
6206 p = idle_sched_class.pick_task(rq_i);
6209 rq_i->core_pick = p;
6211 if (p == rq_i->idle) {
6212 if (rq_i->nr_running) {
6213 rq->core->core_forceidle_count++;
6215 rq->core->core_forceidle_seq++;
6222 if (schedstat_enabled() && rq->core->core_forceidle_count) {
6223 rq->core->core_forceidle_start = rq_clock(rq->core);
6224 rq->core->core_forceidle_occupation = occ;
6227 rq->core->core_pick_seq = rq->core->core_task_seq;
6228 next = rq->core_pick;
6229 rq->core_sched_seq = rq->core->core_pick_seq;
6231 /* Something should have been selected for current CPU */
6232 WARN_ON_ONCE(!next);
6235 * Reschedule siblings
6237 * NOTE: L1TF -- at this point we're no longer running the old task and
6238 * sending an IPI (below) ensures the sibling will no longer be running
6239 * their task. This ensures there is no inter-sibling overlap between
6240 * non-matching user state.
6242 for_each_cpu(i, smt_mask) {
6246 * An online sibling might have gone offline before a task
6247 * could be picked for it, or it might be offline but later
6248 * happen to come online, but its too late and nothing was
6249 * picked for it. That's Ok - it will pick tasks for itself,
6252 if (!rq_i->core_pick)
6256 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6257 * fi_before fi update?
6263 if (!(fi_before && rq->core->core_forceidle_count))
6264 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6266 rq_i->core_pick->core_occupation = occ;
6269 rq_i->core_pick = NULL;
6273 /* Did we break L1TF mitigation requirements? */
6274 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6276 if (rq_i->curr == rq_i->core_pick) {
6277 rq_i->core_pick = NULL;
6285 set_next_task(rq, next);
6287 if (rq->core->core_forceidle_count && next == rq->idle)
6288 queue_core_balance(rq);
6293 static bool try_steal_cookie(int this, int that)
6295 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6296 struct task_struct *p;
6297 unsigned long cookie;
6298 bool success = false;
6301 guard(double_rq_lock)(dst, src);
6303 cookie = dst->core->core_cookie;
6307 if (dst->curr != dst->idle)
6310 p = sched_core_find(src, cookie);
6315 if (p == src->core_pick || p == src->curr)
6318 if (!is_cpu_allowed(p, this))
6321 if (p->core_occupation > dst->idle->core_occupation)
6324 * sched_core_find() and sched_core_next() will ensure
6325 * that task @p is not throttled now, we also need to
6326 * check whether the runqueue of the destination CPU is
6329 if (sched_task_is_throttled(p, this))
6332 deactivate_task(src, p, 0);
6333 set_task_cpu(p, this);
6334 activate_task(dst, p, 0);
6342 p = sched_core_next(p, cookie);
6348 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6352 for_each_cpu_wrap(i, sched_domain_span(sd), cpu + 1) {
6359 if (try_steal_cookie(cpu, i))
6366 static void sched_core_balance(struct rq *rq)
6368 struct sched_domain *sd;
6369 int cpu = cpu_of(rq);
6373 raw_spin_rq_unlock_irq(rq);
6374 for_each_domain(cpu, sd) {
6378 if (steal_cookie_task(cpu, sd))
6381 raw_spin_rq_lock_irq(rq);
6386 static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
6388 static void queue_core_balance(struct rq *rq)
6390 if (!sched_core_enabled(rq))
6393 if (!rq->core->core_cookie)
6396 if (!rq->nr_running) /* not forced idle */
6399 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6402 DEFINE_LOCK_GUARD_1(core_lock, int,
6403 sched_core_lock(*_T->lock, &_T->flags),
6404 sched_core_unlock(*_T->lock, &_T->flags),
6405 unsigned long flags)
6407 static void sched_core_cpu_starting(unsigned int cpu)
6409 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6410 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6413 guard(core_lock)(&cpu);
6415 WARN_ON_ONCE(rq->core != rq);
6417 /* if we're the first, we'll be our own leader */
6418 if (cpumask_weight(smt_mask) == 1)
6421 /* find the leader */
6422 for_each_cpu(t, smt_mask) {
6426 if (rq->core == rq) {
6432 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6435 /* install and validate core_rq */
6436 for_each_cpu(t, smt_mask) {
6442 WARN_ON_ONCE(rq->core != core_rq);
6446 static void sched_core_cpu_deactivate(unsigned int cpu)
6448 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6449 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6452 guard(core_lock)(&cpu);
6454 /* if we're the last man standing, nothing to do */
6455 if (cpumask_weight(smt_mask) == 1) {
6456 WARN_ON_ONCE(rq->core != rq);
6460 /* if we're not the leader, nothing to do */
6464 /* find a new leader */
6465 for_each_cpu(t, smt_mask) {
6468 core_rq = cpu_rq(t);
6472 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6475 /* copy the shared state to the new leader */
6476 core_rq->core_task_seq = rq->core_task_seq;
6477 core_rq->core_pick_seq = rq->core_pick_seq;
6478 core_rq->core_cookie = rq->core_cookie;
6479 core_rq->core_forceidle_count = rq->core_forceidle_count;
6480 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6481 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6484 * Accounting edge for forced idle is handled in pick_next_task().
6485 * Don't need another one here, since the hotplug thread shouldn't
6488 core_rq->core_forceidle_start = 0;
6490 /* install new leader */
6491 for_each_cpu(t, smt_mask) {
6497 static inline void sched_core_cpu_dying(unsigned int cpu)
6499 struct rq *rq = cpu_rq(cpu);
6505 #else /* !CONFIG_SCHED_CORE */
6507 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6508 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6509 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6511 static struct task_struct *
6512 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6514 return __pick_next_task(rq, prev, rf);
6517 #endif /* CONFIG_SCHED_CORE */
6520 * Constants for the sched_mode argument of __schedule().
6522 * The mode argument allows RT enabled kernels to differentiate a
6523 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6524 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6525 * optimize the AND operation out and just check for zero.
6528 #define SM_PREEMPT 0x1
6529 #define SM_RTLOCK_WAIT 0x2
6531 #ifndef CONFIG_PREEMPT_RT
6532 # define SM_MASK_PREEMPT (~0U)
6534 # define SM_MASK_PREEMPT SM_PREEMPT
6538 * __schedule() is the main scheduler function.
6540 * The main means of driving the scheduler and thus entering this function are:
6542 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6544 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6545 * paths. For example, see arch/x86/entry_64.S.
6547 * To drive preemption between tasks, the scheduler sets the flag in timer
6548 * interrupt handler scheduler_tick().
6550 * 3. Wakeups don't really cause entry into schedule(). They add a
6551 * task to the run-queue and that's it.
6553 * Now, if the new task added to the run-queue preempts the current
6554 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6555 * called on the nearest possible occasion:
6557 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6559 * - in syscall or exception context, at the next outmost
6560 * preempt_enable(). (this might be as soon as the wake_up()'s
6563 * - in IRQ context, return from interrupt-handler to
6564 * preemptible context
6566 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6569 * - cond_resched() call
6570 * - explicit schedule() call
6571 * - return from syscall or exception to user-space
6572 * - return from interrupt-handler to user-space
6574 * WARNING: must be called with preemption disabled!
6576 static void __sched notrace __schedule(unsigned int sched_mode)
6578 struct task_struct *prev, *next;
6579 unsigned long *switch_count;
6580 unsigned long prev_state;
6585 cpu = smp_processor_id();
6589 schedule_debug(prev, !!sched_mode);
6591 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6594 local_irq_disable();
6595 rcu_note_context_switch(!!sched_mode);
6598 * Make sure that signal_pending_state()->signal_pending() below
6599 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6600 * done by the caller to avoid the race with signal_wake_up():
6602 * __set_current_state(@state) signal_wake_up()
6603 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6604 * wake_up_state(p, state)
6605 * LOCK rq->lock LOCK p->pi_state
6606 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6607 * if (signal_pending_state()) if (p->state & @state)
6609 * Also, the membarrier system call requires a full memory barrier
6610 * after coming from user-space, before storing to rq->curr.
6613 smp_mb__after_spinlock();
6615 /* Promote REQ to ACT */
6616 rq->clock_update_flags <<= 1;
6617 update_rq_clock(rq);
6619 switch_count = &prev->nivcsw;
6622 * We must load prev->state once (task_struct::state is volatile), such
6623 * that we form a control dependency vs deactivate_task() below.
6625 prev_state = READ_ONCE(prev->__state);
6626 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6627 if (signal_pending_state(prev_state, prev)) {
6628 WRITE_ONCE(prev->__state, TASK_RUNNING);
6630 prev->sched_contributes_to_load =
6631 (prev_state & TASK_UNINTERRUPTIBLE) &&
6632 !(prev_state & TASK_NOLOAD) &&
6633 !(prev_state & TASK_FROZEN);
6635 if (prev->sched_contributes_to_load)
6636 rq->nr_uninterruptible++;
6639 * __schedule() ttwu()
6640 * prev_state = prev->state; if (p->on_rq && ...)
6641 * if (prev_state) goto out;
6642 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6643 * p->state = TASK_WAKING
6645 * Where __schedule() and ttwu() have matching control dependencies.
6647 * After this, schedule() must not care about p->state any more.
6649 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6651 if (prev->in_iowait) {
6652 atomic_inc(&rq->nr_iowait);
6653 delayacct_blkio_start();
6656 switch_count = &prev->nvcsw;
6659 next = pick_next_task(rq, prev, &rf);
6660 clear_tsk_need_resched(prev);
6661 clear_preempt_need_resched();
6662 #ifdef CONFIG_SCHED_DEBUG
6663 rq->last_seen_need_resched_ns = 0;
6666 if (likely(prev != next)) {
6669 * RCU users of rcu_dereference(rq->curr) may not see
6670 * changes to task_struct made by pick_next_task().
6672 RCU_INIT_POINTER(rq->curr, next);
6674 * The membarrier system call requires each architecture
6675 * to have a full memory barrier after updating
6676 * rq->curr, before returning to user-space.
6678 * Here are the schemes providing that barrier on the
6679 * various architectures:
6680 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6681 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6682 * - finish_lock_switch() for weakly-ordered
6683 * architectures where spin_unlock is a full barrier,
6684 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6685 * is a RELEASE barrier),
6689 migrate_disable_switch(rq, prev);
6690 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6692 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6694 /* Also unlocks the rq: */
6695 rq = context_switch(rq, prev, next, &rf);
6697 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6699 rq_unpin_lock(rq, &rf);
6700 __balance_callbacks(rq);
6701 raw_spin_rq_unlock_irq(rq);
6705 void __noreturn do_task_dead(void)
6707 /* Causes final put_task_struct in finish_task_switch(): */
6708 set_special_state(TASK_DEAD);
6710 /* Tell freezer to ignore us: */
6711 current->flags |= PF_NOFREEZE;
6713 __schedule(SM_NONE);
6716 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6721 static inline void sched_submit_work(struct task_struct *tsk)
6723 unsigned int task_flags;
6725 if (task_is_running(tsk))
6728 task_flags = tsk->flags;
6730 * If a worker goes to sleep, notify and ask workqueue whether it
6731 * wants to wake up a task to maintain concurrency.
6733 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6734 if (task_flags & PF_WQ_WORKER)
6735 wq_worker_sleeping(tsk);
6737 io_wq_worker_sleeping(tsk);
6741 * spinlock and rwlock must not flush block requests. This will
6742 * deadlock if the callback attempts to acquire a lock which is
6745 SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6748 * If we are going to sleep and we have plugged IO queued,
6749 * make sure to submit it to avoid deadlocks.
6751 blk_flush_plug(tsk->plug, true);
6754 static void sched_update_worker(struct task_struct *tsk)
6756 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6757 if (tsk->flags & PF_WQ_WORKER)
6758 wq_worker_running(tsk);
6760 io_wq_worker_running(tsk);
6764 asmlinkage __visible void __sched schedule(void)
6766 struct task_struct *tsk = current;
6768 sched_submit_work(tsk);
6771 __schedule(SM_NONE);
6772 sched_preempt_enable_no_resched();
6773 } while (need_resched());
6774 sched_update_worker(tsk);
6776 EXPORT_SYMBOL(schedule);
6779 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6780 * state (have scheduled out non-voluntarily) by making sure that all
6781 * tasks have either left the run queue or have gone into user space.
6782 * As idle tasks do not do either, they must not ever be preempted
6783 * (schedule out non-voluntarily).
6785 * schedule_idle() is similar to schedule_preempt_disable() except that it
6786 * never enables preemption because it does not call sched_submit_work().
6788 void __sched schedule_idle(void)
6791 * As this skips calling sched_submit_work(), which the idle task does
6792 * regardless because that function is a nop when the task is in a
6793 * TASK_RUNNING state, make sure this isn't used someplace that the
6794 * current task can be in any other state. Note, idle is always in the
6795 * TASK_RUNNING state.
6797 WARN_ON_ONCE(current->__state);
6799 __schedule(SM_NONE);
6800 } while (need_resched());
6803 #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6804 asmlinkage __visible void __sched schedule_user(void)
6807 * If we come here after a random call to set_need_resched(),
6808 * or we have been woken up remotely but the IPI has not yet arrived,
6809 * we haven't yet exited the RCU idle mode. Do it here manually until
6810 * we find a better solution.
6812 * NB: There are buggy callers of this function. Ideally we
6813 * should warn if prev_state != CONTEXT_USER, but that will trigger
6814 * too frequently to make sense yet.
6816 enum ctx_state prev_state = exception_enter();
6818 exception_exit(prev_state);
6823 * schedule_preempt_disabled - called with preemption disabled
6825 * Returns with preemption disabled. Note: preempt_count must be 1
6827 void __sched schedule_preempt_disabled(void)
6829 sched_preempt_enable_no_resched();
6834 #ifdef CONFIG_PREEMPT_RT
6835 void __sched notrace schedule_rtlock(void)
6839 __schedule(SM_RTLOCK_WAIT);
6840 sched_preempt_enable_no_resched();
6841 } while (need_resched());
6843 NOKPROBE_SYMBOL(schedule_rtlock);
6846 static void __sched notrace preempt_schedule_common(void)
6850 * Because the function tracer can trace preempt_count_sub()
6851 * and it also uses preempt_enable/disable_notrace(), if
6852 * NEED_RESCHED is set, the preempt_enable_notrace() called
6853 * by the function tracer will call this function again and
6854 * cause infinite recursion.
6856 * Preemption must be disabled here before the function
6857 * tracer can trace. Break up preempt_disable() into two
6858 * calls. One to disable preemption without fear of being
6859 * traced. The other to still record the preemption latency,
6860 * which can also be traced by the function tracer.
6862 preempt_disable_notrace();
6863 preempt_latency_start(1);
6864 __schedule(SM_PREEMPT);
6865 preempt_latency_stop(1);
6866 preempt_enable_no_resched_notrace();
6869 * Check again in case we missed a preemption opportunity
6870 * between schedule and now.
6872 } while (need_resched());
6875 #ifdef CONFIG_PREEMPTION
6877 * This is the entry point to schedule() from in-kernel preemption
6878 * off of preempt_enable.
6880 asmlinkage __visible void __sched notrace preempt_schedule(void)
6883 * If there is a non-zero preempt_count or interrupts are disabled,
6884 * we do not want to preempt the current task. Just return..
6886 if (likely(!preemptible()))
6888 preempt_schedule_common();
6890 NOKPROBE_SYMBOL(preempt_schedule);
6891 EXPORT_SYMBOL(preempt_schedule);
6893 #ifdef CONFIG_PREEMPT_DYNAMIC
6894 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6895 #ifndef preempt_schedule_dynamic_enabled
6896 #define preempt_schedule_dynamic_enabled preempt_schedule
6897 #define preempt_schedule_dynamic_disabled NULL
6899 DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6900 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6901 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6902 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6903 void __sched notrace dynamic_preempt_schedule(void)
6905 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6909 NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6910 EXPORT_SYMBOL(dynamic_preempt_schedule);
6915 * preempt_schedule_notrace - preempt_schedule called by tracing
6917 * The tracing infrastructure uses preempt_enable_notrace to prevent
6918 * recursion and tracing preempt enabling caused by the tracing
6919 * infrastructure itself. But as tracing can happen in areas coming
6920 * from userspace or just about to enter userspace, a preempt enable
6921 * can occur before user_exit() is called. This will cause the scheduler
6922 * to be called when the system is still in usermode.
6924 * To prevent this, the preempt_enable_notrace will use this function
6925 * instead of preempt_schedule() to exit user context if needed before
6926 * calling the scheduler.
6928 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6930 enum ctx_state prev_ctx;
6932 if (likely(!preemptible()))
6937 * Because the function tracer can trace preempt_count_sub()
6938 * and it also uses preempt_enable/disable_notrace(), if
6939 * NEED_RESCHED is set, the preempt_enable_notrace() called
6940 * by the function tracer will call this function again and
6941 * cause infinite recursion.
6943 * Preemption must be disabled here before the function
6944 * tracer can trace. Break up preempt_disable() into two
6945 * calls. One to disable preemption without fear of being
6946 * traced. The other to still record the preemption latency,
6947 * which can also be traced by the function tracer.
6949 preempt_disable_notrace();
6950 preempt_latency_start(1);
6952 * Needs preempt disabled in case user_exit() is traced
6953 * and the tracer calls preempt_enable_notrace() causing
6954 * an infinite recursion.
6956 prev_ctx = exception_enter();
6957 __schedule(SM_PREEMPT);
6958 exception_exit(prev_ctx);
6960 preempt_latency_stop(1);
6961 preempt_enable_no_resched_notrace();
6962 } while (need_resched());
6964 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6966 #ifdef CONFIG_PREEMPT_DYNAMIC
6967 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6968 #ifndef preempt_schedule_notrace_dynamic_enabled
6969 #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6970 #define preempt_schedule_notrace_dynamic_disabled NULL
6972 DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6973 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6974 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6975 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6976 void __sched notrace dynamic_preempt_schedule_notrace(void)
6978 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6980 preempt_schedule_notrace();
6982 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6983 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6987 #endif /* CONFIG_PREEMPTION */
6990 * This is the entry point to schedule() from kernel preemption
6991 * off of irq context.
6992 * Note, that this is called and return with irqs disabled. This will
6993 * protect us against recursive calling from irq.
6995 asmlinkage __visible void __sched preempt_schedule_irq(void)
6997 enum ctx_state prev_state;
6999 /* Catch callers which need to be fixed */
7000 BUG_ON(preempt_count() || !irqs_disabled());
7002 prev_state = exception_enter();
7007 __schedule(SM_PREEMPT);
7008 local_irq_disable();
7009 sched_preempt_enable_no_resched();
7010 } while (need_resched());
7012 exception_exit(prev_state);
7015 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
7018 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~(WF_SYNC|WF_CURRENT_CPU));
7019 return try_to_wake_up(curr->private, mode, wake_flags);
7021 EXPORT_SYMBOL(default_wake_function);
7023 static void __setscheduler_prio(struct task_struct *p, int prio)
7026 p->sched_class = &dl_sched_class;
7027 else if (rt_prio(prio))
7028 p->sched_class = &rt_sched_class;
7030 p->sched_class = &fair_sched_class;
7035 #ifdef CONFIG_RT_MUTEXES
7037 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
7040 prio = min(prio, pi_task->prio);
7045 static inline int rt_effective_prio(struct task_struct *p, int prio)
7047 struct task_struct *pi_task = rt_mutex_get_top_task(p);
7049 return __rt_effective_prio(pi_task, prio);
7053 * rt_mutex_setprio - set the current priority of a task
7055 * @pi_task: donor task
7057 * This function changes the 'effective' priority of a task. It does
7058 * not touch ->normal_prio like __setscheduler().
7060 * Used by the rt_mutex code to implement priority inheritance
7061 * logic. Call site only calls if the priority of the task changed.
7063 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
7065 int prio, oldprio, queued, running, queue_flag =
7066 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7067 const struct sched_class *prev_class;
7071 /* XXX used to be waiter->prio, not waiter->task->prio */
7072 prio = __rt_effective_prio(pi_task, p->normal_prio);
7075 * If nothing changed; bail early.
7077 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
7080 rq = __task_rq_lock(p, &rf);
7081 update_rq_clock(rq);
7083 * Set under pi_lock && rq->lock, such that the value can be used under
7086 * Note that there is loads of tricky to make this pointer cache work
7087 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
7088 * ensure a task is de-boosted (pi_task is set to NULL) before the
7089 * task is allowed to run again (and can exit). This ensures the pointer
7090 * points to a blocked task -- which guarantees the task is present.
7092 p->pi_top_task = pi_task;
7095 * For FIFO/RR we only need to set prio, if that matches we're done.
7097 if (prio == p->prio && !dl_prio(prio))
7101 * Idle task boosting is a nono in general. There is one
7102 * exception, when PREEMPT_RT and NOHZ is active:
7104 * The idle task calls get_next_timer_interrupt() and holds
7105 * the timer wheel base->lock on the CPU and another CPU wants
7106 * to access the timer (probably to cancel it). We can safely
7107 * ignore the boosting request, as the idle CPU runs this code
7108 * with interrupts disabled and will complete the lock
7109 * protected section without being interrupted. So there is no
7110 * real need to boost.
7112 if (unlikely(p == rq->idle)) {
7113 WARN_ON(p != rq->curr);
7114 WARN_ON(p->pi_blocked_on);
7118 trace_sched_pi_setprio(p, pi_task);
7121 if (oldprio == prio)
7122 queue_flag &= ~DEQUEUE_MOVE;
7124 prev_class = p->sched_class;
7125 queued = task_on_rq_queued(p);
7126 running = task_current(rq, p);
7128 dequeue_task(rq, p, queue_flag);
7130 put_prev_task(rq, p);
7133 * Boosting condition are:
7134 * 1. -rt task is running and holds mutex A
7135 * --> -dl task blocks on mutex A
7137 * 2. -dl task is running and holds mutex A
7138 * --> -dl task blocks on mutex A and could preempt the
7141 if (dl_prio(prio)) {
7142 if (!dl_prio(p->normal_prio) ||
7143 (pi_task && dl_prio(pi_task->prio) &&
7144 dl_entity_preempt(&pi_task->dl, &p->dl))) {
7145 p->dl.pi_se = pi_task->dl.pi_se;
7146 queue_flag |= ENQUEUE_REPLENISH;
7148 p->dl.pi_se = &p->dl;
7150 } else if (rt_prio(prio)) {
7151 if (dl_prio(oldprio))
7152 p->dl.pi_se = &p->dl;
7154 queue_flag |= ENQUEUE_HEAD;
7156 if (dl_prio(oldprio))
7157 p->dl.pi_se = &p->dl;
7158 if (rt_prio(oldprio))
7162 __setscheduler_prio(p, prio);
7165 enqueue_task(rq, p, queue_flag);
7167 set_next_task(rq, p);
7169 check_class_changed(rq, p, prev_class, oldprio);
7171 /* Avoid rq from going away on us: */
7174 rq_unpin_lock(rq, &rf);
7175 __balance_callbacks(rq);
7176 raw_spin_rq_unlock(rq);
7181 static inline int rt_effective_prio(struct task_struct *p, int prio)
7187 void set_user_nice(struct task_struct *p, long nice)
7189 bool queued, running;
7194 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
7197 * We have to be careful, if called from sys_setpriority(),
7198 * the task might be in the middle of scheduling on another CPU.
7200 rq = task_rq_lock(p, &rf);
7201 update_rq_clock(rq);
7204 * The RT priorities are set via sched_setscheduler(), but we still
7205 * allow the 'normal' nice value to be set - but as expected
7206 * it won't have any effect on scheduling until the task is
7207 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7209 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7210 p->static_prio = NICE_TO_PRIO(nice);
7213 queued = task_on_rq_queued(p);
7214 running = task_current(rq, p);
7216 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7218 put_prev_task(rq, p);
7220 p->static_prio = NICE_TO_PRIO(nice);
7221 set_load_weight(p, true);
7223 p->prio = effective_prio(p);
7226 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7228 set_next_task(rq, p);
7231 * If the task increased its priority or is running and
7232 * lowered its priority, then reschedule its CPU:
7234 p->sched_class->prio_changed(rq, p, old_prio);
7237 task_rq_unlock(rq, p, &rf);
7239 EXPORT_SYMBOL(set_user_nice);
7242 * is_nice_reduction - check if nice value is an actual reduction
7244 * Similar to can_nice() but does not perform a capability check.
7249 static bool is_nice_reduction(const struct task_struct *p, const int nice)
7251 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7252 int nice_rlim = nice_to_rlimit(nice);
7254 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7258 * can_nice - check if a task can reduce its nice value
7262 int can_nice(const struct task_struct *p, const int nice)
7264 return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7267 #ifdef __ARCH_WANT_SYS_NICE
7270 * sys_nice - change the priority of the current process.
7271 * @increment: priority increment
7273 * sys_setpriority is a more generic, but much slower function that
7274 * does similar things.
7276 SYSCALL_DEFINE1(nice, int, increment)
7281 * Setpriority might change our priority at the same moment.
7282 * We don't have to worry. Conceptually one call occurs first
7283 * and we have a single winner.
7285 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7286 nice = task_nice(current) + increment;
7288 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7289 if (increment < 0 && !can_nice(current, nice))
7292 retval = security_task_setnice(current, nice);
7296 set_user_nice(current, nice);
7303 * task_prio - return the priority value of a given task.
7304 * @p: the task in question.
7306 * Return: The priority value as seen by users in /proc.
7308 * sched policy return value kernel prio user prio/nice
7310 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7311 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7312 * deadline -101 -1 0
7314 int task_prio(const struct task_struct *p)
7316 return p->prio - MAX_RT_PRIO;
7320 * idle_cpu - is a given CPU idle currently?
7321 * @cpu: the processor in question.
7323 * Return: 1 if the CPU is currently idle. 0 otherwise.
7325 int idle_cpu(int cpu)
7327 struct rq *rq = cpu_rq(cpu);
7329 if (rq->curr != rq->idle)
7336 if (rq->ttwu_pending)
7344 * available_idle_cpu - is a given CPU idle for enqueuing work.
7345 * @cpu: the CPU in question.
7347 * Return: 1 if the CPU is currently idle. 0 otherwise.
7349 int available_idle_cpu(int cpu)
7354 if (vcpu_is_preempted(cpu))
7361 * idle_task - return the idle task for a given CPU.
7362 * @cpu: the processor in question.
7364 * Return: The idle task for the CPU @cpu.
7366 struct task_struct *idle_task(int cpu)
7368 return cpu_rq(cpu)->idle;
7371 #ifdef CONFIG_SCHED_CORE
7372 int sched_core_idle_cpu(int cpu)
7374 struct rq *rq = cpu_rq(cpu);
7376 if (sched_core_enabled(rq) && rq->curr == rq->idle)
7379 return idle_cpu(cpu);
7386 * This function computes an effective utilization for the given CPU, to be
7387 * used for frequency selection given the linear relation: f = u * f_max.
7389 * The scheduler tracks the following metrics:
7391 * cpu_util_{cfs,rt,dl,irq}()
7394 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7395 * synchronized windows and are thus directly comparable.
7397 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7398 * which excludes things like IRQ and steal-time. These latter are then accrued
7399 * in the irq utilization.
7401 * The DL bandwidth number otoh is not a measured metric but a value computed
7402 * based on the task model parameters and gives the minimal utilization
7403 * required to meet deadlines.
7405 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7406 enum cpu_util_type type,
7407 struct task_struct *p)
7409 unsigned long dl_util, util, irq, max;
7410 struct rq *rq = cpu_rq(cpu);
7412 max = arch_scale_cpu_capacity(cpu);
7414 if (!uclamp_is_used() &&
7415 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7420 * Early check to see if IRQ/steal time saturates the CPU, can be
7421 * because of inaccuracies in how we track these -- see
7422 * update_irq_load_avg().
7424 irq = cpu_util_irq(rq);
7425 if (unlikely(irq >= max))
7429 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7430 * CFS tasks and we use the same metric to track the effective
7431 * utilization (PELT windows are synchronized) we can directly add them
7432 * to obtain the CPU's actual utilization.
7434 * CFS and RT utilization can be boosted or capped, depending on
7435 * utilization clamp constraints requested by currently RUNNABLE
7437 * When there are no CFS RUNNABLE tasks, clamps are released and
7438 * frequency will be gracefully reduced with the utilization decay.
7440 util = util_cfs + cpu_util_rt(rq);
7441 if (type == FREQUENCY_UTIL)
7442 util = uclamp_rq_util_with(rq, util, p);
7444 dl_util = cpu_util_dl(rq);
7447 * For frequency selection we do not make cpu_util_dl() a permanent part
7448 * of this sum because we want to use cpu_bw_dl() later on, but we need
7449 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7450 * that we select f_max when there is no idle time.
7452 * NOTE: numerical errors or stop class might cause us to not quite hit
7453 * saturation when we should -- something for later.
7455 if (util + dl_util >= max)
7459 * OTOH, for energy computation we need the estimated running time, so
7460 * include util_dl and ignore dl_bw.
7462 if (type == ENERGY_UTIL)
7466 * There is still idle time; further improve the number by using the
7467 * irq metric. Because IRQ/steal time is hidden from the task clock we
7468 * need to scale the task numbers:
7471 * U' = irq + --------- * U
7474 util = scale_irq_capacity(util, irq, max);
7478 * Bandwidth required by DEADLINE must always be granted while, for
7479 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7480 * to gracefully reduce the frequency when no tasks show up for longer
7483 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7484 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7485 * an interface. So, we only do the latter for now.
7487 if (type == FREQUENCY_UTIL)
7488 util += cpu_bw_dl(rq);
7490 return min(max, util);
7493 unsigned long sched_cpu_util(int cpu)
7495 return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
7497 #endif /* CONFIG_SMP */
7500 * find_process_by_pid - find a process with a matching PID value.
7501 * @pid: the pid in question.
7503 * The task of @pid, if found. %NULL otherwise.
7505 static struct task_struct *find_process_by_pid(pid_t pid)
7507 return pid ? find_task_by_vpid(pid) : current;
7511 * sched_setparam() passes in -1 for its policy, to let the functions
7512 * it calls know not to change it.
7514 #define SETPARAM_POLICY -1
7516 static void __setscheduler_params(struct task_struct *p,
7517 const struct sched_attr *attr)
7519 int policy = attr->sched_policy;
7521 if (policy == SETPARAM_POLICY)
7526 if (dl_policy(policy))
7527 __setparam_dl(p, attr);
7528 else if (fair_policy(policy))
7529 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7532 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7533 * !rt_policy. Always setting this ensures that things like
7534 * getparam()/getattr() don't report silly values for !rt tasks.
7536 p->rt_priority = attr->sched_priority;
7537 p->normal_prio = normal_prio(p);
7538 set_load_weight(p, true);
7542 * Check the target process has a UID that matches the current process's:
7544 static bool check_same_owner(struct task_struct *p)
7546 const struct cred *cred = current_cred(), *pcred;
7550 pcred = __task_cred(p);
7551 match = (uid_eq(cred->euid, pcred->euid) ||
7552 uid_eq(cred->euid, pcred->uid));
7558 * Allow unprivileged RT tasks to decrease priority.
7559 * Only issue a capable test if needed and only once to avoid an audit
7560 * event on permitted non-privileged operations:
7562 static int user_check_sched_setscheduler(struct task_struct *p,
7563 const struct sched_attr *attr,
7564 int policy, int reset_on_fork)
7566 if (fair_policy(policy)) {
7567 if (attr->sched_nice < task_nice(p) &&
7568 !is_nice_reduction(p, attr->sched_nice))
7572 if (rt_policy(policy)) {
7573 unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7575 /* Can't set/change the rt policy: */
7576 if (policy != p->policy && !rlim_rtprio)
7579 /* Can't increase priority: */
7580 if (attr->sched_priority > p->rt_priority &&
7581 attr->sched_priority > rlim_rtprio)
7586 * Can't set/change SCHED_DEADLINE policy at all for now
7587 * (safest behavior); in the future we would like to allow
7588 * unprivileged DL tasks to increase their relative deadline
7589 * or reduce their runtime (both ways reducing utilization)
7591 if (dl_policy(policy))
7595 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7596 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7598 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7599 if (!is_nice_reduction(p, task_nice(p)))
7603 /* Can't change other user's priorities: */
7604 if (!check_same_owner(p))
7607 /* Normal users shall not reset the sched_reset_on_fork flag: */
7608 if (p->sched_reset_on_fork && !reset_on_fork)
7614 if (!capable(CAP_SYS_NICE))
7620 static int __sched_setscheduler(struct task_struct *p,
7621 const struct sched_attr *attr,
7624 int oldpolicy = -1, policy = attr->sched_policy;
7625 int retval, oldprio, newprio, queued, running;
7626 const struct sched_class *prev_class;
7627 struct balance_callback *head;
7630 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7632 bool cpuset_locked = false;
7634 /* The pi code expects interrupts enabled */
7635 BUG_ON(pi && in_interrupt());
7637 /* Double check policy once rq lock held: */
7639 reset_on_fork = p->sched_reset_on_fork;
7640 policy = oldpolicy = p->policy;
7642 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7644 if (!valid_policy(policy))
7648 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7652 * Valid priorities for SCHED_FIFO and SCHED_RR are
7653 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7654 * SCHED_BATCH and SCHED_IDLE is 0.
7656 if (attr->sched_priority > MAX_RT_PRIO-1)
7658 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7659 (rt_policy(policy) != (attr->sched_priority != 0)))
7663 retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7667 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7670 retval = security_task_setscheduler(p);
7675 /* Update task specific "requested" clamps */
7676 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7677 retval = uclamp_validate(p, attr);
7683 * SCHED_DEADLINE bandwidth accounting relies on stable cpusets
7686 if (dl_policy(policy) || dl_policy(p->policy)) {
7687 cpuset_locked = true;
7692 * Make sure no PI-waiters arrive (or leave) while we are
7693 * changing the priority of the task:
7695 * To be able to change p->policy safely, the appropriate
7696 * runqueue lock must be held.
7698 rq = task_rq_lock(p, &rf);
7699 update_rq_clock(rq);
7702 * Changing the policy of the stop threads its a very bad idea:
7704 if (p == rq->stop) {
7710 * If not changing anything there's no need to proceed further,
7711 * but store a possible modification of reset_on_fork.
7713 if (unlikely(policy == p->policy)) {
7714 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7716 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7718 if (dl_policy(policy) && dl_param_changed(p, attr))
7720 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7723 p->sched_reset_on_fork = reset_on_fork;
7730 #ifdef CONFIG_RT_GROUP_SCHED
7732 * Do not allow realtime tasks into groups that have no runtime
7735 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7736 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7737 !task_group_is_autogroup(task_group(p))) {
7743 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7744 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7745 cpumask_t *span = rq->rd->span;
7748 * Don't allow tasks with an affinity mask smaller than
7749 * the entire root_domain to become SCHED_DEADLINE. We
7750 * will also fail if there's no bandwidth available.
7752 if (!cpumask_subset(span, p->cpus_ptr) ||
7753 rq->rd->dl_bw.bw == 0) {
7761 /* Re-check policy now with rq lock held: */
7762 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7763 policy = oldpolicy = -1;
7764 task_rq_unlock(rq, p, &rf);
7771 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7772 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7775 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7780 p->sched_reset_on_fork = reset_on_fork;
7783 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7786 * Take priority boosted tasks into account. If the new
7787 * effective priority is unchanged, we just store the new
7788 * normal parameters and do not touch the scheduler class and
7789 * the runqueue. This will be done when the task deboost
7792 newprio = rt_effective_prio(p, newprio);
7793 if (newprio == oldprio)
7794 queue_flags &= ~DEQUEUE_MOVE;
7797 queued = task_on_rq_queued(p);
7798 running = task_current(rq, p);
7800 dequeue_task(rq, p, queue_flags);
7802 put_prev_task(rq, p);
7804 prev_class = p->sched_class;
7806 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7807 __setscheduler_params(p, attr);
7808 __setscheduler_prio(p, newprio);
7810 __setscheduler_uclamp(p, attr);
7814 * We enqueue to tail when the priority of a task is
7815 * increased (user space view).
7817 if (oldprio < p->prio)
7818 queue_flags |= ENQUEUE_HEAD;
7820 enqueue_task(rq, p, queue_flags);
7823 set_next_task(rq, p);
7825 check_class_changed(rq, p, prev_class, oldprio);
7827 /* Avoid rq from going away on us: */
7829 head = splice_balance_callbacks(rq);
7830 task_rq_unlock(rq, p, &rf);
7835 rt_mutex_adjust_pi(p);
7838 /* Run balance callbacks after we've adjusted the PI chain: */
7839 balance_callbacks(rq, head);
7845 task_rq_unlock(rq, p, &rf);
7851 static int _sched_setscheduler(struct task_struct *p, int policy,
7852 const struct sched_param *param, bool check)
7854 struct sched_attr attr = {
7855 .sched_policy = policy,
7856 .sched_priority = param->sched_priority,
7857 .sched_nice = PRIO_TO_NICE(p->static_prio),
7860 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7861 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7862 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7863 policy &= ~SCHED_RESET_ON_FORK;
7864 attr.sched_policy = policy;
7867 return __sched_setscheduler(p, &attr, check, true);
7870 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7871 * @p: the task in question.
7872 * @policy: new policy.
7873 * @param: structure containing the new RT priority.
7875 * Use sched_set_fifo(), read its comment.
7877 * Return: 0 on success. An error code otherwise.
7879 * NOTE that the task may be already dead.
7881 int sched_setscheduler(struct task_struct *p, int policy,
7882 const struct sched_param *param)
7884 return _sched_setscheduler(p, policy, param, true);
7887 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7889 return __sched_setscheduler(p, attr, true, true);
7892 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7894 return __sched_setscheduler(p, attr, false, true);
7896 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7899 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7900 * @p: the task in question.
7901 * @policy: new policy.
7902 * @param: structure containing the new RT priority.
7904 * Just like sched_setscheduler, only don't bother checking if the
7905 * current context has permission. For example, this is needed in
7906 * stop_machine(): we create temporary high priority worker threads,
7907 * but our caller might not have that capability.
7909 * Return: 0 on success. An error code otherwise.
7911 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7912 const struct sched_param *param)
7914 return _sched_setscheduler(p, policy, param, false);
7918 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7919 * incapable of resource management, which is the one thing an OS really should
7922 * This is of course the reason it is limited to privileged users only.
7924 * Worse still; it is fundamentally impossible to compose static priority
7925 * workloads. You cannot take two correctly working static prio workloads
7926 * and smash them together and still expect them to work.
7928 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7932 * The administrator _MUST_ configure the system, the kernel simply doesn't
7933 * know enough information to make a sensible choice.
7935 void sched_set_fifo(struct task_struct *p)
7937 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7938 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7940 EXPORT_SYMBOL_GPL(sched_set_fifo);
7943 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7945 void sched_set_fifo_low(struct task_struct *p)
7947 struct sched_param sp = { .sched_priority = 1 };
7948 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7950 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7952 void sched_set_normal(struct task_struct *p, int nice)
7954 struct sched_attr attr = {
7955 .sched_policy = SCHED_NORMAL,
7958 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7960 EXPORT_SYMBOL_GPL(sched_set_normal);
7963 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7965 struct sched_param lparam;
7966 struct task_struct *p;
7969 if (!param || pid < 0)
7971 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7976 p = find_process_by_pid(pid);
7982 retval = sched_setscheduler(p, policy, &lparam);
7990 * Mimics kernel/events/core.c perf_copy_attr().
7992 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7997 /* Zero the full structure, so that a short copy will be nice: */
7998 memset(attr, 0, sizeof(*attr));
8000 ret = get_user(size, &uattr->size);
8004 /* ABI compatibility quirk: */
8006 size = SCHED_ATTR_SIZE_VER0;
8007 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
8010 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
8017 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
8018 size < SCHED_ATTR_SIZE_VER1)
8022 * XXX: Do we want to be lenient like existing syscalls; or do we want
8023 * to be strict and return an error on out-of-bounds values?
8025 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
8030 put_user(sizeof(*attr), &uattr->size);
8034 static void get_params(struct task_struct *p, struct sched_attr *attr)
8036 if (task_has_dl_policy(p))
8037 __getparam_dl(p, attr);
8038 else if (task_has_rt_policy(p))
8039 attr->sched_priority = p->rt_priority;
8041 attr->sched_nice = task_nice(p);
8045 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
8046 * @pid: the pid in question.
8047 * @policy: new policy.
8048 * @param: structure containing the new RT priority.
8050 * Return: 0 on success. An error code otherwise.
8052 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
8057 return do_sched_setscheduler(pid, policy, param);
8061 * sys_sched_setparam - set/change the RT priority of a thread
8062 * @pid: the pid in question.
8063 * @param: structure containing the new RT priority.
8065 * Return: 0 on success. An error code otherwise.
8067 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
8069 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
8073 * sys_sched_setattr - same as above, but with extended sched_attr
8074 * @pid: the pid in question.
8075 * @uattr: structure containing the extended parameters.
8076 * @flags: for future extension.
8078 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
8079 unsigned int, flags)
8081 struct sched_attr attr;
8082 struct task_struct *p;
8085 if (!uattr || pid < 0 || flags)
8088 retval = sched_copy_attr(uattr, &attr);
8092 if ((int)attr.sched_policy < 0)
8094 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
8095 attr.sched_policy = SETPARAM_POLICY;
8099 p = find_process_by_pid(pid);
8105 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
8106 get_params(p, &attr);
8107 retval = sched_setattr(p, &attr);
8115 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
8116 * @pid: the pid in question.
8118 * Return: On success, the policy of the thread. Otherwise, a negative error
8121 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
8123 struct task_struct *p;
8131 p = find_process_by_pid(pid);
8133 retval = security_task_getscheduler(p);
8136 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
8143 * sys_sched_getparam - get the RT priority of a thread
8144 * @pid: the pid in question.
8145 * @param: structure containing the RT priority.
8147 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
8150 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
8152 struct sched_param lp = { .sched_priority = 0 };
8153 struct task_struct *p;
8156 if (!param || pid < 0)
8160 p = find_process_by_pid(pid);
8165 retval = security_task_getscheduler(p);
8169 if (task_has_rt_policy(p))
8170 lp.sched_priority = p->rt_priority;
8174 * This one might sleep, we cannot do it with a spinlock held ...
8176 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
8186 * Copy the kernel size attribute structure (which might be larger
8187 * than what user-space knows about) to user-space.
8189 * Note that all cases are valid: user-space buffer can be larger or
8190 * smaller than the kernel-space buffer. The usual case is that both
8191 * have the same size.
8194 sched_attr_copy_to_user(struct sched_attr __user *uattr,
8195 struct sched_attr *kattr,
8198 unsigned int ksize = sizeof(*kattr);
8200 if (!access_ok(uattr, usize))
8204 * sched_getattr() ABI forwards and backwards compatibility:
8206 * If usize == ksize then we just copy everything to user-space and all is good.
8208 * If usize < ksize then we only copy as much as user-space has space for,
8209 * this keeps ABI compatibility as well. We skip the rest.
8211 * If usize > ksize then user-space is using a newer version of the ABI,
8212 * which part the kernel doesn't know about. Just ignore it - tooling can
8213 * detect the kernel's knowledge of attributes from the attr->size value
8214 * which is set to ksize in this case.
8216 kattr->size = min(usize, ksize);
8218 if (copy_to_user(uattr, kattr, kattr->size))
8225 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8226 * @pid: the pid in question.
8227 * @uattr: structure containing the extended parameters.
8228 * @usize: sizeof(attr) for fwd/bwd comp.
8229 * @flags: for future extension.
8231 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8232 unsigned int, usize, unsigned int, flags)
8234 struct sched_attr kattr = { };
8235 struct task_struct *p;
8238 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8239 usize < SCHED_ATTR_SIZE_VER0 || flags)
8243 p = find_process_by_pid(pid);
8248 retval = security_task_getscheduler(p);
8252 kattr.sched_policy = p->policy;
8253 if (p->sched_reset_on_fork)
8254 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8255 get_params(p, &kattr);
8256 kattr.sched_flags &= SCHED_FLAG_ALL;
8258 #ifdef CONFIG_UCLAMP_TASK
8260 * This could race with another potential updater, but this is fine
8261 * because it'll correctly read the old or the new value. We don't need
8262 * to guarantee who wins the race as long as it doesn't return garbage.
8264 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8265 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8270 return sched_attr_copy_to_user(uattr, &kattr, usize);
8278 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8283 * If the task isn't a deadline task or admission control is
8284 * disabled then we don't care about affinity changes.
8286 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8290 * Since bandwidth control happens on root_domain basis,
8291 * if admission test is enabled, we only admit -deadline
8292 * tasks allowed to run on all the CPUs in the task's
8296 if (!cpumask_subset(task_rq(p)->rd->span, mask))
8304 __sched_setaffinity(struct task_struct *p, struct affinity_context *ctx)
8307 cpumask_var_t cpus_allowed, new_mask;
8309 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8312 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8314 goto out_free_cpus_allowed;
8317 cpuset_cpus_allowed(p, cpus_allowed);
8318 cpumask_and(new_mask, ctx->new_mask, cpus_allowed);
8320 ctx->new_mask = new_mask;
8321 ctx->flags |= SCA_CHECK;
8323 retval = dl_task_check_affinity(p, new_mask);
8325 goto out_free_new_mask;
8327 retval = __set_cpus_allowed_ptr(p, ctx);
8329 goto out_free_new_mask;
8331 cpuset_cpus_allowed(p, cpus_allowed);
8332 if (!cpumask_subset(new_mask, cpus_allowed)) {
8334 * We must have raced with a concurrent cpuset update.
8335 * Just reset the cpumask to the cpuset's cpus_allowed.
8337 cpumask_copy(new_mask, cpus_allowed);
8340 * If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
8341 * will restore the previous user_cpus_ptr value.
8343 * In the unlikely event a previous user_cpus_ptr exists,
8344 * we need to further restrict the mask to what is allowed
8345 * by that old user_cpus_ptr.
8347 if (unlikely((ctx->flags & SCA_USER) && ctx->user_mask)) {
8348 bool empty = !cpumask_and(new_mask, new_mask,
8351 if (WARN_ON_ONCE(empty))
8352 cpumask_copy(new_mask, cpus_allowed);
8354 __set_cpus_allowed_ptr(p, ctx);
8359 free_cpumask_var(new_mask);
8360 out_free_cpus_allowed:
8361 free_cpumask_var(cpus_allowed);
8365 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8367 struct affinity_context ac;
8368 struct cpumask *user_mask;
8369 struct task_struct *p;
8374 p = find_process_by_pid(pid);
8380 /* Prevent p going away */
8384 if (p->flags & PF_NO_SETAFFINITY) {
8389 if (!check_same_owner(p)) {
8391 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8399 retval = security_task_setscheduler(p);
8404 * With non-SMP configs, user_cpus_ptr/user_mask isn't used and
8405 * alloc_user_cpus_ptr() returns NULL.
8407 user_mask = alloc_user_cpus_ptr(NUMA_NO_NODE);
8409 cpumask_copy(user_mask, in_mask);
8410 } else if (IS_ENABLED(CONFIG_SMP)) {
8415 ac = (struct affinity_context){
8416 .new_mask = in_mask,
8417 .user_mask = user_mask,
8421 retval = __sched_setaffinity(p, &ac);
8422 kfree(ac.user_mask);
8429 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8430 struct cpumask *new_mask)
8432 if (len < cpumask_size())
8433 cpumask_clear(new_mask);
8434 else if (len > cpumask_size())
8435 len = cpumask_size();
8437 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8441 * sys_sched_setaffinity - set the CPU affinity of a process
8442 * @pid: pid of the process
8443 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8444 * @user_mask_ptr: user-space pointer to the new CPU mask
8446 * Return: 0 on success. An error code otherwise.
8448 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8449 unsigned long __user *, user_mask_ptr)
8451 cpumask_var_t new_mask;
8454 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8457 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8459 retval = sched_setaffinity(pid, new_mask);
8460 free_cpumask_var(new_mask);
8464 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8466 struct task_struct *p;
8467 unsigned long flags;
8473 p = find_process_by_pid(pid);
8477 retval = security_task_getscheduler(p);
8481 raw_spin_lock_irqsave(&p->pi_lock, flags);
8482 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8483 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8492 * sys_sched_getaffinity - get the CPU affinity of a process
8493 * @pid: pid of the process
8494 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8495 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8497 * Return: size of CPU mask copied to user_mask_ptr on success. An
8498 * error code otherwise.
8500 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8501 unsigned long __user *, user_mask_ptr)
8506 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8508 if (len & (sizeof(unsigned long)-1))
8511 if (!zalloc_cpumask_var(&mask, GFP_KERNEL))
8514 ret = sched_getaffinity(pid, mask);
8516 unsigned int retlen = min(len, cpumask_size());
8518 if (copy_to_user(user_mask_ptr, cpumask_bits(mask), retlen))
8523 free_cpumask_var(mask);
8528 static void do_sched_yield(void)
8533 rq = this_rq_lock_irq(&rf);
8535 schedstat_inc(rq->yld_count);
8536 current->sched_class->yield_task(rq);
8539 rq_unlock_irq(rq, &rf);
8540 sched_preempt_enable_no_resched();
8546 * sys_sched_yield - yield the current processor to other threads.
8548 * This function yields the current CPU to other tasks. If there are no
8549 * other threads running on this CPU then this function will return.
8553 SYSCALL_DEFINE0(sched_yield)
8559 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8560 int __sched __cond_resched(void)
8562 if (should_resched(0)) {
8563 preempt_schedule_common();
8567 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8568 * whether the current CPU is in an RCU read-side critical section,
8569 * so the tick can report quiescent states even for CPUs looping
8570 * in kernel context. In contrast, in non-preemptible kernels,
8571 * RCU readers leave no in-memory hints, which means that CPU-bound
8572 * processes executing in kernel context might never report an
8573 * RCU quiescent state. Therefore, the following code causes
8574 * cond_resched() to report a quiescent state, but only when RCU
8575 * is in urgent need of one.
8577 #ifndef CONFIG_PREEMPT_RCU
8582 EXPORT_SYMBOL(__cond_resched);
8585 #ifdef CONFIG_PREEMPT_DYNAMIC
8586 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8587 #define cond_resched_dynamic_enabled __cond_resched
8588 #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8589 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8590 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8592 #define might_resched_dynamic_enabled __cond_resched
8593 #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8594 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8595 EXPORT_STATIC_CALL_TRAMP(might_resched);
8596 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8597 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8598 int __sched dynamic_cond_resched(void)
8600 klp_sched_try_switch();
8601 if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8603 return __cond_resched();
8605 EXPORT_SYMBOL(dynamic_cond_resched);
8607 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8608 int __sched dynamic_might_resched(void)
8610 if (!static_branch_unlikely(&sk_dynamic_might_resched))
8612 return __cond_resched();
8614 EXPORT_SYMBOL(dynamic_might_resched);
8619 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8620 * call schedule, and on return reacquire the lock.
8622 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8623 * operations here to prevent schedule() from being called twice (once via
8624 * spin_unlock(), once by hand).
8626 int __cond_resched_lock(spinlock_t *lock)
8628 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8631 lockdep_assert_held(lock);
8633 if (spin_needbreak(lock) || resched) {
8635 if (!_cond_resched())
8642 EXPORT_SYMBOL(__cond_resched_lock);
8644 int __cond_resched_rwlock_read(rwlock_t *lock)
8646 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8649 lockdep_assert_held_read(lock);
8651 if (rwlock_needbreak(lock) || resched) {
8653 if (!_cond_resched())
8660 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8662 int __cond_resched_rwlock_write(rwlock_t *lock)
8664 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8667 lockdep_assert_held_write(lock);
8669 if (rwlock_needbreak(lock) || resched) {
8671 if (!_cond_resched())
8678 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8680 #ifdef CONFIG_PREEMPT_DYNAMIC
8682 #ifdef CONFIG_GENERIC_ENTRY
8683 #include <linux/entry-common.h>
8689 * SC:preempt_schedule
8690 * SC:preempt_schedule_notrace
8691 * SC:irqentry_exit_cond_resched
8695 * cond_resched <- __cond_resched
8696 * might_resched <- RET0
8697 * preempt_schedule <- NOP
8698 * preempt_schedule_notrace <- NOP
8699 * irqentry_exit_cond_resched <- NOP
8702 * cond_resched <- __cond_resched
8703 * might_resched <- __cond_resched
8704 * preempt_schedule <- NOP
8705 * preempt_schedule_notrace <- NOP
8706 * irqentry_exit_cond_resched <- NOP
8709 * cond_resched <- RET0
8710 * might_resched <- RET0
8711 * preempt_schedule <- preempt_schedule
8712 * preempt_schedule_notrace <- preempt_schedule_notrace
8713 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8717 preempt_dynamic_undefined = -1,
8718 preempt_dynamic_none,
8719 preempt_dynamic_voluntary,
8720 preempt_dynamic_full,
8723 int preempt_dynamic_mode = preempt_dynamic_undefined;
8725 int sched_dynamic_mode(const char *str)
8727 if (!strcmp(str, "none"))
8728 return preempt_dynamic_none;
8730 if (!strcmp(str, "voluntary"))
8731 return preempt_dynamic_voluntary;
8733 if (!strcmp(str, "full"))
8734 return preempt_dynamic_full;
8739 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8740 #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8741 #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8742 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8743 #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8744 #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8746 #error "Unsupported PREEMPT_DYNAMIC mechanism"
8749 static DEFINE_MUTEX(sched_dynamic_mutex);
8750 static bool klp_override;
8752 static void __sched_dynamic_update(int mode)
8755 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8756 * the ZERO state, which is invalid.
8759 preempt_dynamic_enable(cond_resched);
8760 preempt_dynamic_enable(might_resched);
8761 preempt_dynamic_enable(preempt_schedule);
8762 preempt_dynamic_enable(preempt_schedule_notrace);
8763 preempt_dynamic_enable(irqentry_exit_cond_resched);
8766 case preempt_dynamic_none:
8768 preempt_dynamic_enable(cond_resched);
8769 preempt_dynamic_disable(might_resched);
8770 preempt_dynamic_disable(preempt_schedule);
8771 preempt_dynamic_disable(preempt_schedule_notrace);
8772 preempt_dynamic_disable(irqentry_exit_cond_resched);
8773 if (mode != preempt_dynamic_mode)
8774 pr_info("Dynamic Preempt: none\n");
8777 case preempt_dynamic_voluntary:
8779 preempt_dynamic_enable(cond_resched);
8780 preempt_dynamic_enable(might_resched);
8781 preempt_dynamic_disable(preempt_schedule);
8782 preempt_dynamic_disable(preempt_schedule_notrace);
8783 preempt_dynamic_disable(irqentry_exit_cond_resched);
8784 if (mode != preempt_dynamic_mode)
8785 pr_info("Dynamic Preempt: voluntary\n");
8788 case preempt_dynamic_full:
8790 preempt_dynamic_disable(cond_resched);
8791 preempt_dynamic_disable(might_resched);
8792 preempt_dynamic_enable(preempt_schedule);
8793 preempt_dynamic_enable(preempt_schedule_notrace);
8794 preempt_dynamic_enable(irqentry_exit_cond_resched);
8795 if (mode != preempt_dynamic_mode)
8796 pr_info("Dynamic Preempt: full\n");
8800 preempt_dynamic_mode = mode;
8803 void sched_dynamic_update(int mode)
8805 mutex_lock(&sched_dynamic_mutex);
8806 __sched_dynamic_update(mode);
8807 mutex_unlock(&sched_dynamic_mutex);
8810 #ifdef CONFIG_HAVE_PREEMPT_DYNAMIC_CALL
8812 static int klp_cond_resched(void)
8814 __klp_sched_try_switch();
8815 return __cond_resched();
8818 void sched_dynamic_klp_enable(void)
8820 mutex_lock(&sched_dynamic_mutex);
8822 klp_override = true;
8823 static_call_update(cond_resched, klp_cond_resched);
8825 mutex_unlock(&sched_dynamic_mutex);
8828 void sched_dynamic_klp_disable(void)
8830 mutex_lock(&sched_dynamic_mutex);
8832 klp_override = false;
8833 __sched_dynamic_update(preempt_dynamic_mode);
8835 mutex_unlock(&sched_dynamic_mutex);
8838 #endif /* CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
8840 static int __init setup_preempt_mode(char *str)
8842 int mode = sched_dynamic_mode(str);
8844 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8848 sched_dynamic_update(mode);
8851 __setup("preempt=", setup_preempt_mode);
8853 static void __init preempt_dynamic_init(void)
8855 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8856 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8857 sched_dynamic_update(preempt_dynamic_none);
8858 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8859 sched_dynamic_update(preempt_dynamic_voluntary);
8861 /* Default static call setting, nothing to do */
8862 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8863 preempt_dynamic_mode = preempt_dynamic_full;
8864 pr_info("Dynamic Preempt: full\n");
8869 #define PREEMPT_MODEL_ACCESSOR(mode) \
8870 bool preempt_model_##mode(void) \
8872 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8873 return preempt_dynamic_mode == preempt_dynamic_##mode; \
8875 EXPORT_SYMBOL_GPL(preempt_model_##mode)
8877 PREEMPT_MODEL_ACCESSOR(none);
8878 PREEMPT_MODEL_ACCESSOR(voluntary);
8879 PREEMPT_MODEL_ACCESSOR(full);
8881 #else /* !CONFIG_PREEMPT_DYNAMIC */
8883 static inline void preempt_dynamic_init(void) { }
8885 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8888 * yield - yield the current processor to other threads.
8890 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8892 * The scheduler is at all times free to pick the calling task as the most
8893 * eligible task to run, if removing the yield() call from your code breaks
8894 * it, it's already broken.
8896 * Typical broken usage is:
8901 * where one assumes that yield() will let 'the other' process run that will
8902 * make event true. If the current task is a SCHED_FIFO task that will never
8903 * happen. Never use yield() as a progress guarantee!!
8905 * If you want to use yield() to wait for something, use wait_event().
8906 * If you want to use yield() to be 'nice' for others, use cond_resched().
8907 * If you still want to use yield(), do not!
8909 void __sched yield(void)
8911 set_current_state(TASK_RUNNING);
8914 EXPORT_SYMBOL(yield);
8917 * yield_to - yield the current processor to another thread in
8918 * your thread group, or accelerate that thread toward the
8919 * processor it's on.
8921 * @preempt: whether task preemption is allowed or not
8923 * It's the caller's job to ensure that the target task struct
8924 * can't go away on us before we can do any checks.
8927 * true (>0) if we indeed boosted the target task.
8928 * false (0) if we failed to boost the target.
8929 * -ESRCH if there's no task to yield to.
8931 int __sched yield_to(struct task_struct *p, bool preempt)
8933 struct task_struct *curr = current;
8934 struct rq *rq, *p_rq;
8935 unsigned long flags;
8938 local_irq_save(flags);
8944 * If we're the only runnable task on the rq and target rq also
8945 * has only one task, there's absolutely no point in yielding.
8947 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8952 double_rq_lock(rq, p_rq);
8953 if (task_rq(p) != p_rq) {
8954 double_rq_unlock(rq, p_rq);
8958 if (!curr->sched_class->yield_to_task)
8961 if (curr->sched_class != p->sched_class)
8964 if (task_on_cpu(p_rq, p) || !task_is_running(p))
8967 yielded = curr->sched_class->yield_to_task(rq, p);
8969 schedstat_inc(rq->yld_count);
8971 * Make p's CPU reschedule; pick_next_entity takes care of
8974 if (preempt && rq != p_rq)
8979 double_rq_unlock(rq, p_rq);
8981 local_irq_restore(flags);
8988 EXPORT_SYMBOL_GPL(yield_to);
8990 int io_schedule_prepare(void)
8992 int old_iowait = current->in_iowait;
8994 current->in_iowait = 1;
8995 blk_flush_plug(current->plug, true);
8999 void io_schedule_finish(int token)
9001 current->in_iowait = token;
9005 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
9006 * that process accounting knows that this is a task in IO wait state.
9008 long __sched io_schedule_timeout(long timeout)
9013 token = io_schedule_prepare();
9014 ret = schedule_timeout(timeout);
9015 io_schedule_finish(token);
9019 EXPORT_SYMBOL(io_schedule_timeout);
9021 void __sched io_schedule(void)
9025 token = io_schedule_prepare();
9027 io_schedule_finish(token);
9029 EXPORT_SYMBOL(io_schedule);
9032 * sys_sched_get_priority_max - return maximum RT priority.
9033 * @policy: scheduling class.
9035 * Return: On success, this syscall returns the maximum
9036 * rt_priority that can be used by a given scheduling class.
9037 * On failure, a negative error code is returned.
9039 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
9046 ret = MAX_RT_PRIO-1;
9048 case SCHED_DEADLINE:
9059 * sys_sched_get_priority_min - return minimum RT priority.
9060 * @policy: scheduling class.
9062 * Return: On success, this syscall returns the minimum
9063 * rt_priority that can be used by a given scheduling class.
9064 * On failure, a negative error code is returned.
9066 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
9075 case SCHED_DEADLINE:
9084 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
9086 struct task_struct *p;
9087 unsigned int time_slice;
9097 p = find_process_by_pid(pid);
9101 retval = security_task_getscheduler(p);
9105 rq = task_rq_lock(p, &rf);
9107 if (p->sched_class->get_rr_interval)
9108 time_slice = p->sched_class->get_rr_interval(rq, p);
9109 task_rq_unlock(rq, p, &rf);
9112 jiffies_to_timespec64(time_slice, t);
9121 * sys_sched_rr_get_interval - return the default timeslice of a process.
9122 * @pid: pid of the process.
9123 * @interval: userspace pointer to the timeslice value.
9125 * this syscall writes the default timeslice value of a given process
9126 * into the user-space timespec buffer. A value of '0' means infinity.
9128 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
9131 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
9132 struct __kernel_timespec __user *, interval)
9134 struct timespec64 t;
9135 int retval = sched_rr_get_interval(pid, &t);
9138 retval = put_timespec64(&t, interval);
9143 #ifdef CONFIG_COMPAT_32BIT_TIME
9144 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
9145 struct old_timespec32 __user *, interval)
9147 struct timespec64 t;
9148 int retval = sched_rr_get_interval(pid, &t);
9151 retval = put_old_timespec32(&t, interval);
9156 void sched_show_task(struct task_struct *p)
9158 unsigned long free = 0;
9161 if (!try_get_task_stack(p))
9164 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
9166 if (task_is_running(p))
9167 pr_cont(" running task ");
9168 #ifdef CONFIG_DEBUG_STACK_USAGE
9169 free = stack_not_used(p);
9174 ppid = task_pid_nr(rcu_dereference(p->real_parent));
9176 pr_cont(" stack:%-5lu pid:%-5d ppid:%-6d flags:0x%08lx\n",
9177 free, task_pid_nr(p), ppid,
9178 read_task_thread_flags(p));
9180 print_worker_info(KERN_INFO, p);
9181 print_stop_info(KERN_INFO, p);
9182 show_stack(p, NULL, KERN_INFO);
9185 EXPORT_SYMBOL_GPL(sched_show_task);
9188 state_filter_match(unsigned long state_filter, struct task_struct *p)
9190 unsigned int state = READ_ONCE(p->__state);
9192 /* no filter, everything matches */
9196 /* filter, but doesn't match */
9197 if (!(state & state_filter))
9201 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
9204 if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
9211 void show_state_filter(unsigned int state_filter)
9213 struct task_struct *g, *p;
9216 for_each_process_thread(g, p) {
9218 * reset the NMI-timeout, listing all files on a slow
9219 * console might take a lot of time:
9220 * Also, reset softlockup watchdogs on all CPUs, because
9221 * another CPU might be blocked waiting for us to process
9224 touch_nmi_watchdog();
9225 touch_all_softlockup_watchdogs();
9226 if (state_filter_match(state_filter, p))
9230 #ifdef CONFIG_SCHED_DEBUG
9232 sysrq_sched_debug_show();
9236 * Only show locks if all tasks are dumped:
9239 debug_show_all_locks();
9243 * init_idle - set up an idle thread for a given CPU
9244 * @idle: task in question
9245 * @cpu: CPU the idle task belongs to
9247 * NOTE: this function does not set the idle thread's NEED_RESCHED
9248 * flag, to make booting more robust.
9250 void __init init_idle(struct task_struct *idle, int cpu)
9253 struct affinity_context ac = (struct affinity_context) {
9254 .new_mask = cpumask_of(cpu),
9258 struct rq *rq = cpu_rq(cpu);
9259 unsigned long flags;
9261 __sched_fork(0, idle);
9263 raw_spin_lock_irqsave(&idle->pi_lock, flags);
9264 raw_spin_rq_lock(rq);
9266 idle->__state = TASK_RUNNING;
9267 idle->se.exec_start = sched_clock();
9269 * PF_KTHREAD should already be set at this point; regardless, make it
9270 * look like a proper per-CPU kthread.
9272 idle->flags |= PF_KTHREAD | PF_NO_SETAFFINITY;
9273 kthread_set_per_cpu(idle, cpu);
9277 * It's possible that init_idle() gets called multiple times on a task,
9278 * in that case do_set_cpus_allowed() will not do the right thing.
9280 * And since this is boot we can forgo the serialization.
9282 set_cpus_allowed_common(idle, &ac);
9285 * We're having a chicken and egg problem, even though we are
9286 * holding rq->lock, the CPU isn't yet set to this CPU so the
9287 * lockdep check in task_group() will fail.
9289 * Similar case to sched_fork(). / Alternatively we could
9290 * use task_rq_lock() here and obtain the other rq->lock.
9295 __set_task_cpu(idle, cpu);
9299 rcu_assign_pointer(rq->curr, idle);
9300 idle->on_rq = TASK_ON_RQ_QUEUED;
9304 raw_spin_rq_unlock(rq);
9305 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
9307 /* Set the preempt count _outside_ the spinlocks! */
9308 init_idle_preempt_count(idle, cpu);
9311 * The idle tasks have their own, simple scheduling class:
9313 idle->sched_class = &idle_sched_class;
9314 ftrace_graph_init_idle_task(idle, cpu);
9315 vtime_init_idle(idle, cpu);
9317 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9323 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9324 const struct cpumask *trial)
9328 if (cpumask_empty(cur))
9331 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9336 int task_can_attach(struct task_struct *p)
9341 * Kthreads which disallow setaffinity shouldn't be moved
9342 * to a new cpuset; we don't want to change their CPU
9343 * affinity and isolating such threads by their set of
9344 * allowed nodes is unnecessary. Thus, cpusets are not
9345 * applicable for such threads. This prevents checking for
9346 * success of set_cpus_allowed_ptr() on all attached tasks
9347 * before cpus_mask may be changed.
9349 if (p->flags & PF_NO_SETAFFINITY)
9355 bool sched_smp_initialized __read_mostly;
9357 #ifdef CONFIG_NUMA_BALANCING
9358 /* Migrate current task p to target_cpu */
9359 int migrate_task_to(struct task_struct *p, int target_cpu)
9361 struct migration_arg arg = { p, target_cpu };
9362 int curr_cpu = task_cpu(p);
9364 if (curr_cpu == target_cpu)
9367 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9370 /* TODO: This is not properly updating schedstats */
9372 trace_sched_move_numa(p, curr_cpu, target_cpu);
9373 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9377 * Requeue a task on a given node and accurately track the number of NUMA
9378 * tasks on the runqueues
9380 void sched_setnuma(struct task_struct *p, int nid)
9382 bool queued, running;
9386 rq = task_rq_lock(p, &rf);
9387 queued = task_on_rq_queued(p);
9388 running = task_current(rq, p);
9391 dequeue_task(rq, p, DEQUEUE_SAVE);
9393 put_prev_task(rq, p);
9395 p->numa_preferred_nid = nid;
9398 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9400 set_next_task(rq, p);
9401 task_rq_unlock(rq, p, &rf);
9403 #endif /* CONFIG_NUMA_BALANCING */
9405 #ifdef CONFIG_HOTPLUG_CPU
9407 * Ensure that the idle task is using init_mm right before its CPU goes
9410 void idle_task_exit(void)
9412 struct mm_struct *mm = current->active_mm;
9414 BUG_ON(cpu_online(smp_processor_id()));
9415 BUG_ON(current != this_rq()->idle);
9417 if (mm != &init_mm) {
9418 switch_mm(mm, &init_mm, current);
9419 finish_arch_post_lock_switch();
9422 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9425 static int __balance_push_cpu_stop(void *arg)
9427 struct task_struct *p = arg;
9428 struct rq *rq = this_rq();
9432 raw_spin_lock_irq(&p->pi_lock);
9435 update_rq_clock(rq);
9437 if (task_rq(p) == rq && task_on_rq_queued(p)) {
9438 cpu = select_fallback_rq(rq->cpu, p);
9439 rq = __migrate_task(rq, &rf, p, cpu);
9443 raw_spin_unlock_irq(&p->pi_lock);
9450 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9453 * Ensure we only run per-cpu kthreads once the CPU goes !active.
9455 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9456 * effective when the hotplug motion is down.
9458 static void balance_push(struct rq *rq)
9460 struct task_struct *push_task = rq->curr;
9462 lockdep_assert_rq_held(rq);
9465 * Ensure the thing is persistent until balance_push_set(.on = false);
9467 rq->balance_callback = &balance_push_callback;
9470 * Only active while going offline and when invoked on the outgoing
9473 if (!cpu_dying(rq->cpu) || rq != this_rq())
9477 * Both the cpu-hotplug and stop task are in this case and are
9478 * required to complete the hotplug process.
9480 if (kthread_is_per_cpu(push_task) ||
9481 is_migration_disabled(push_task)) {
9484 * If this is the idle task on the outgoing CPU try to wake
9485 * up the hotplug control thread which might wait for the
9486 * last task to vanish. The rcuwait_active() check is
9487 * accurate here because the waiter is pinned on this CPU
9488 * and can't obviously be running in parallel.
9490 * On RT kernels this also has to check whether there are
9491 * pinned and scheduled out tasks on the runqueue. They
9492 * need to leave the migrate disabled section first.
9494 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9495 rcuwait_active(&rq->hotplug_wait)) {
9496 raw_spin_rq_unlock(rq);
9497 rcuwait_wake_up(&rq->hotplug_wait);
9498 raw_spin_rq_lock(rq);
9503 get_task_struct(push_task);
9505 * Temporarily drop rq->lock such that we can wake-up the stop task.
9506 * Both preemption and IRQs are still disabled.
9508 raw_spin_rq_unlock(rq);
9509 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9510 this_cpu_ptr(&push_work));
9512 * At this point need_resched() is true and we'll take the loop in
9513 * schedule(). The next pick is obviously going to be the stop task
9514 * which kthread_is_per_cpu() and will push this task away.
9516 raw_spin_rq_lock(rq);
9519 static void balance_push_set(int cpu, bool on)
9521 struct rq *rq = cpu_rq(cpu);
9524 rq_lock_irqsave(rq, &rf);
9526 WARN_ON_ONCE(rq->balance_callback);
9527 rq->balance_callback = &balance_push_callback;
9528 } else if (rq->balance_callback == &balance_push_callback) {
9529 rq->balance_callback = NULL;
9531 rq_unlock_irqrestore(rq, &rf);
9535 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9536 * inactive. All tasks which are not per CPU kernel threads are either
9537 * pushed off this CPU now via balance_push() or placed on a different CPU
9538 * during wakeup. Wait until the CPU is quiescent.
9540 static void balance_hotplug_wait(void)
9542 struct rq *rq = this_rq();
9544 rcuwait_wait_event(&rq->hotplug_wait,
9545 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9546 TASK_UNINTERRUPTIBLE);
9551 static inline void balance_push(struct rq *rq)
9555 static inline void balance_push_set(int cpu, bool on)
9559 static inline void balance_hotplug_wait(void)
9563 #endif /* CONFIG_HOTPLUG_CPU */
9565 void set_rq_online(struct rq *rq)
9568 const struct sched_class *class;
9570 cpumask_set_cpu(rq->cpu, rq->rd->online);
9573 for_each_class(class) {
9574 if (class->rq_online)
9575 class->rq_online(rq);
9580 void set_rq_offline(struct rq *rq)
9583 const struct sched_class *class;
9585 update_rq_clock(rq);
9586 for_each_class(class) {
9587 if (class->rq_offline)
9588 class->rq_offline(rq);
9591 cpumask_clear_cpu(rq->cpu, rq->rd->online);
9597 * used to mark begin/end of suspend/resume:
9599 static int num_cpus_frozen;
9602 * Update cpusets according to cpu_active mask. If cpusets are
9603 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9604 * around partition_sched_domains().
9606 * If we come here as part of a suspend/resume, don't touch cpusets because we
9607 * want to restore it back to its original state upon resume anyway.
9609 static void cpuset_cpu_active(void)
9611 if (cpuhp_tasks_frozen) {
9613 * num_cpus_frozen tracks how many CPUs are involved in suspend
9614 * resume sequence. As long as this is not the last online
9615 * operation in the resume sequence, just build a single sched
9616 * domain, ignoring cpusets.
9618 partition_sched_domains(1, NULL, NULL);
9619 if (--num_cpus_frozen)
9622 * This is the last CPU online operation. So fall through and
9623 * restore the original sched domains by considering the
9624 * cpuset configurations.
9626 cpuset_force_rebuild();
9628 cpuset_update_active_cpus();
9631 static int cpuset_cpu_inactive(unsigned int cpu)
9633 if (!cpuhp_tasks_frozen) {
9634 int ret = dl_bw_check_overflow(cpu);
9638 cpuset_update_active_cpus();
9641 partition_sched_domains(1, NULL, NULL);
9646 int sched_cpu_activate(unsigned int cpu)
9648 struct rq *rq = cpu_rq(cpu);
9652 * Clear the balance_push callback and prepare to schedule
9655 balance_push_set(cpu, false);
9657 #ifdef CONFIG_SCHED_SMT
9659 * When going up, increment the number of cores with SMT present.
9661 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9662 static_branch_inc_cpuslocked(&sched_smt_present);
9664 set_cpu_active(cpu, true);
9666 if (sched_smp_initialized) {
9667 sched_update_numa(cpu, true);
9668 sched_domains_numa_masks_set(cpu);
9669 cpuset_cpu_active();
9673 * Put the rq online, if not already. This happens:
9675 * 1) In the early boot process, because we build the real domains
9676 * after all CPUs have been brought up.
9678 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9681 rq_lock_irqsave(rq, &rf);
9683 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9686 rq_unlock_irqrestore(rq, &rf);
9691 int sched_cpu_deactivate(unsigned int cpu)
9693 struct rq *rq = cpu_rq(cpu);
9698 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9699 * load balancing when not active
9701 nohz_balance_exit_idle(rq);
9703 set_cpu_active(cpu, false);
9706 * From this point forward, this CPU will refuse to run any task that
9707 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9708 * push those tasks away until this gets cleared, see
9709 * sched_cpu_dying().
9711 balance_push_set(cpu, true);
9714 * We've cleared cpu_active_mask / set balance_push, wait for all
9715 * preempt-disabled and RCU users of this state to go away such that
9716 * all new such users will observe it.
9718 * Specifically, we rely on ttwu to no longer target this CPU, see
9719 * ttwu_queue_cond() and is_cpu_allowed().
9721 * Do sync before park smpboot threads to take care the rcu boost case.
9725 rq_lock_irqsave(rq, &rf);
9727 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9730 rq_unlock_irqrestore(rq, &rf);
9732 #ifdef CONFIG_SCHED_SMT
9734 * When going down, decrement the number of cores with SMT present.
9736 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9737 static_branch_dec_cpuslocked(&sched_smt_present);
9739 sched_core_cpu_deactivate(cpu);
9742 if (!sched_smp_initialized)
9745 sched_update_numa(cpu, false);
9746 ret = cpuset_cpu_inactive(cpu);
9748 balance_push_set(cpu, false);
9749 set_cpu_active(cpu, true);
9750 sched_update_numa(cpu, true);
9753 sched_domains_numa_masks_clear(cpu);
9757 static void sched_rq_cpu_starting(unsigned int cpu)
9759 struct rq *rq = cpu_rq(cpu);
9761 rq->calc_load_update = calc_load_update;
9762 update_max_interval();
9765 int sched_cpu_starting(unsigned int cpu)
9767 sched_core_cpu_starting(cpu);
9768 sched_rq_cpu_starting(cpu);
9769 sched_tick_start(cpu);
9773 #ifdef CONFIG_HOTPLUG_CPU
9776 * Invoked immediately before the stopper thread is invoked to bring the
9777 * CPU down completely. At this point all per CPU kthreads except the
9778 * hotplug thread (current) and the stopper thread (inactive) have been
9779 * either parked or have been unbound from the outgoing CPU. Ensure that
9780 * any of those which might be on the way out are gone.
9782 * If after this point a bound task is being woken on this CPU then the
9783 * responsible hotplug callback has failed to do it's job.
9784 * sched_cpu_dying() will catch it with the appropriate fireworks.
9786 int sched_cpu_wait_empty(unsigned int cpu)
9788 balance_hotplug_wait();
9793 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9794 * might have. Called from the CPU stopper task after ensuring that the
9795 * stopper is the last running task on the CPU, so nr_active count is
9796 * stable. We need to take the teardown thread which is calling this into
9797 * account, so we hand in adjust = 1 to the load calculation.
9799 * Also see the comment "Global load-average calculations".
9801 static void calc_load_migrate(struct rq *rq)
9803 long delta = calc_load_fold_active(rq, 1);
9806 atomic_long_add(delta, &calc_load_tasks);
9809 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9811 struct task_struct *g, *p;
9812 int cpu = cpu_of(rq);
9814 lockdep_assert_rq_held(rq);
9816 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9817 for_each_process_thread(g, p) {
9818 if (task_cpu(p) != cpu)
9821 if (!task_on_rq_queued(p))
9824 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9828 int sched_cpu_dying(unsigned int cpu)
9830 struct rq *rq = cpu_rq(cpu);
9833 /* Handle pending wakeups and then migrate everything off */
9834 sched_tick_stop(cpu);
9836 rq_lock_irqsave(rq, &rf);
9837 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9838 WARN(true, "Dying CPU not properly vacated!");
9839 dump_rq_tasks(rq, KERN_WARNING);
9841 rq_unlock_irqrestore(rq, &rf);
9843 calc_load_migrate(rq);
9844 update_max_interval();
9846 sched_core_cpu_dying(cpu);
9851 void __init sched_init_smp(void)
9853 sched_init_numa(NUMA_NO_NODE);
9856 * There's no userspace yet to cause hotplug operations; hence all the
9857 * CPU masks are stable and all blatant races in the below code cannot
9860 mutex_lock(&sched_domains_mutex);
9861 sched_init_domains(cpu_active_mask);
9862 mutex_unlock(&sched_domains_mutex);
9864 /* Move init over to a non-isolated CPU */
9865 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9867 current->flags &= ~PF_NO_SETAFFINITY;
9868 sched_init_granularity();
9870 init_sched_rt_class();
9871 init_sched_dl_class();
9873 sched_smp_initialized = true;
9876 static int __init migration_init(void)
9878 sched_cpu_starting(smp_processor_id());
9881 early_initcall(migration_init);
9884 void __init sched_init_smp(void)
9886 sched_init_granularity();
9888 #endif /* CONFIG_SMP */
9890 int in_sched_functions(unsigned long addr)
9892 return in_lock_functions(addr) ||
9893 (addr >= (unsigned long)__sched_text_start
9894 && addr < (unsigned long)__sched_text_end);
9897 #ifdef CONFIG_CGROUP_SCHED
9899 * Default task group.
9900 * Every task in system belongs to this group at bootup.
9902 struct task_group root_task_group;
9903 LIST_HEAD(task_groups);
9905 /* Cacheline aligned slab cache for task_group */
9906 static struct kmem_cache *task_group_cache __read_mostly;
9909 void __init sched_init(void)
9911 unsigned long ptr = 0;
9914 /* Make sure the linker didn't screw up */
9915 BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9916 &fair_sched_class != &rt_sched_class + 1 ||
9917 &rt_sched_class != &dl_sched_class + 1);
9919 BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9924 #ifdef CONFIG_FAIR_GROUP_SCHED
9925 ptr += 2 * nr_cpu_ids * sizeof(void **);
9927 #ifdef CONFIG_RT_GROUP_SCHED
9928 ptr += 2 * nr_cpu_ids * sizeof(void **);
9931 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9933 #ifdef CONFIG_FAIR_GROUP_SCHED
9934 root_task_group.se = (struct sched_entity **)ptr;
9935 ptr += nr_cpu_ids * sizeof(void **);
9937 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9938 ptr += nr_cpu_ids * sizeof(void **);
9940 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9941 init_cfs_bandwidth(&root_task_group.cfs_bandwidth, NULL);
9942 #endif /* CONFIG_FAIR_GROUP_SCHED */
9943 #ifdef CONFIG_RT_GROUP_SCHED
9944 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9945 ptr += nr_cpu_ids * sizeof(void **);
9947 root_task_group.rt_rq = (struct rt_rq **)ptr;
9948 ptr += nr_cpu_ids * sizeof(void **);
9950 #endif /* CONFIG_RT_GROUP_SCHED */
9953 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9956 init_defrootdomain();
9959 #ifdef CONFIG_RT_GROUP_SCHED
9960 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9961 global_rt_period(), global_rt_runtime());
9962 #endif /* CONFIG_RT_GROUP_SCHED */
9964 #ifdef CONFIG_CGROUP_SCHED
9965 task_group_cache = KMEM_CACHE(task_group, 0);
9967 list_add(&root_task_group.list, &task_groups);
9968 INIT_LIST_HEAD(&root_task_group.children);
9969 INIT_LIST_HEAD(&root_task_group.siblings);
9970 autogroup_init(&init_task);
9971 #endif /* CONFIG_CGROUP_SCHED */
9973 for_each_possible_cpu(i) {
9977 raw_spin_lock_init(&rq->__lock);
9979 rq->calc_load_active = 0;
9980 rq->calc_load_update = jiffies + LOAD_FREQ;
9981 init_cfs_rq(&rq->cfs);
9982 init_rt_rq(&rq->rt);
9983 init_dl_rq(&rq->dl);
9984 #ifdef CONFIG_FAIR_GROUP_SCHED
9985 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9986 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9988 * How much CPU bandwidth does root_task_group get?
9990 * In case of task-groups formed thr' the cgroup filesystem, it
9991 * gets 100% of the CPU resources in the system. This overall
9992 * system CPU resource is divided among the tasks of
9993 * root_task_group and its child task-groups in a fair manner,
9994 * based on each entity's (task or task-group's) weight
9995 * (se->load.weight).
9997 * In other words, if root_task_group has 10 tasks of weight
9998 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9999 * then A0's share of the CPU resource is:
10001 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
10003 * We achieve this by letting root_task_group's tasks sit
10004 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
10006 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
10007 #endif /* CONFIG_FAIR_GROUP_SCHED */
10009 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
10010 #ifdef CONFIG_RT_GROUP_SCHED
10011 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
10016 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
10017 rq->balance_callback = &balance_push_callback;
10018 rq->active_balance = 0;
10019 rq->next_balance = jiffies;
10023 rq->idle_stamp = 0;
10024 rq->avg_idle = 2*sysctl_sched_migration_cost;
10025 rq->wake_stamp = jiffies;
10026 rq->wake_avg_idle = rq->avg_idle;
10027 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
10029 INIT_LIST_HEAD(&rq->cfs_tasks);
10031 rq_attach_root(rq, &def_root_domain);
10032 #ifdef CONFIG_NO_HZ_COMMON
10033 rq->last_blocked_load_update_tick = jiffies;
10034 atomic_set(&rq->nohz_flags, 0);
10036 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
10038 #ifdef CONFIG_HOTPLUG_CPU
10039 rcuwait_init(&rq->hotplug_wait);
10041 #endif /* CONFIG_SMP */
10042 hrtick_rq_init(rq);
10043 atomic_set(&rq->nr_iowait, 0);
10045 #ifdef CONFIG_SCHED_CORE
10047 rq->core_pick = NULL;
10048 rq->core_enabled = 0;
10049 rq->core_tree = RB_ROOT;
10050 rq->core_forceidle_count = 0;
10051 rq->core_forceidle_occupation = 0;
10052 rq->core_forceidle_start = 0;
10054 rq->core_cookie = 0UL;
10056 zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i));
10059 set_load_weight(&init_task, false);
10062 * The boot idle thread does lazy MMU switching as well:
10064 mmgrab_lazy_tlb(&init_mm);
10065 enter_lazy_tlb(&init_mm, current);
10068 * The idle task doesn't need the kthread struct to function, but it
10069 * is dressed up as a per-CPU kthread and thus needs to play the part
10070 * if we want to avoid special-casing it in code that deals with per-CPU
10073 WARN_ON(!set_kthread_struct(current));
10076 * Make us the idle thread. Technically, schedule() should not be
10077 * called from this thread, however somewhere below it might be,
10078 * but because we are the idle thread, we just pick up running again
10079 * when this runqueue becomes "idle".
10081 init_idle(current, smp_processor_id());
10083 calc_load_update = jiffies + LOAD_FREQ;
10086 idle_thread_set_boot_cpu();
10087 balance_push_set(smp_processor_id(), false);
10089 init_sched_fair_class();
10095 preempt_dynamic_init();
10097 scheduler_running = 1;
10100 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
10102 void __might_sleep(const char *file, int line)
10104 unsigned int state = get_current_state();
10106 * Blocking primitives will set (and therefore destroy) current->state,
10107 * since we will exit with TASK_RUNNING make sure we enter with it,
10108 * otherwise we will destroy state.
10110 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
10111 "do not call blocking ops when !TASK_RUNNING; "
10112 "state=%x set at [<%p>] %pS\n", state,
10113 (void *)current->task_state_change,
10114 (void *)current->task_state_change);
10116 __might_resched(file, line, 0);
10118 EXPORT_SYMBOL(__might_sleep);
10120 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
10122 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
10125 if (preempt_count() == preempt_offset)
10128 pr_err("Preemption disabled at:");
10129 print_ip_sym(KERN_ERR, ip);
10132 static inline bool resched_offsets_ok(unsigned int offsets)
10134 unsigned int nested = preempt_count();
10136 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
10138 return nested == offsets;
10141 void __might_resched(const char *file, int line, unsigned int offsets)
10143 /* Ratelimiting timestamp: */
10144 static unsigned long prev_jiffy;
10146 unsigned long preempt_disable_ip;
10148 /* WARN_ON_ONCE() by default, no rate limit required: */
10151 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
10152 !is_idle_task(current) && !current->non_block_count) ||
10153 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
10157 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10159 prev_jiffy = jiffies;
10161 /* Save this before calling printk(), since that will clobber it: */
10162 preempt_disable_ip = get_preempt_disable_ip(current);
10164 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10166 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10167 in_atomic(), irqs_disabled(), current->non_block_count,
10168 current->pid, current->comm);
10169 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
10170 offsets & MIGHT_RESCHED_PREEMPT_MASK);
10172 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
10173 pr_err("RCU nest depth: %d, expected: %u\n",
10174 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
10177 if (task_stack_end_corrupted(current))
10178 pr_emerg("Thread overran stack, or stack corrupted\n");
10180 debug_show_held_locks(current);
10181 if (irqs_disabled())
10182 print_irqtrace_events(current);
10184 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
10185 preempt_disable_ip);
10188 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10190 EXPORT_SYMBOL(__might_resched);
10192 void __cant_sleep(const char *file, int line, int preempt_offset)
10194 static unsigned long prev_jiffy;
10196 if (irqs_disabled())
10199 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10202 if (preempt_count() > preempt_offset)
10205 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10207 prev_jiffy = jiffies;
10209 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
10210 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10211 in_atomic(), irqs_disabled(),
10212 current->pid, current->comm);
10214 debug_show_held_locks(current);
10216 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10218 EXPORT_SYMBOL_GPL(__cant_sleep);
10221 void __cant_migrate(const char *file, int line)
10223 static unsigned long prev_jiffy;
10225 if (irqs_disabled())
10228 if (is_migration_disabled(current))
10231 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10234 if (preempt_count() > 0)
10237 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10239 prev_jiffy = jiffies;
10241 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
10242 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10243 in_atomic(), irqs_disabled(), is_migration_disabled(current),
10244 current->pid, current->comm);
10246 debug_show_held_locks(current);
10248 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10250 EXPORT_SYMBOL_GPL(__cant_migrate);
10254 #ifdef CONFIG_MAGIC_SYSRQ
10255 void normalize_rt_tasks(void)
10257 struct task_struct *g, *p;
10258 struct sched_attr attr = {
10259 .sched_policy = SCHED_NORMAL,
10262 read_lock(&tasklist_lock);
10263 for_each_process_thread(g, p) {
10265 * Only normalize user tasks:
10267 if (p->flags & PF_KTHREAD)
10270 p->se.exec_start = 0;
10271 schedstat_set(p->stats.wait_start, 0);
10272 schedstat_set(p->stats.sleep_start, 0);
10273 schedstat_set(p->stats.block_start, 0);
10275 if (!dl_task(p) && !rt_task(p)) {
10277 * Renice negative nice level userspace
10280 if (task_nice(p) < 0)
10281 set_user_nice(p, 0);
10285 __sched_setscheduler(p, &attr, false, false);
10287 read_unlock(&tasklist_lock);
10290 #endif /* CONFIG_MAGIC_SYSRQ */
10292 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
10294 * These functions are only useful for the IA64 MCA handling, or kdb.
10296 * They can only be called when the whole system has been
10297 * stopped - every CPU needs to be quiescent, and no scheduling
10298 * activity can take place. Using them for anything else would
10299 * be a serious bug, and as a result, they aren't even visible
10300 * under any other configuration.
10304 * curr_task - return the current task for a given CPU.
10305 * @cpu: the processor in question.
10307 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10309 * Return: The current task for @cpu.
10311 struct task_struct *curr_task(int cpu)
10313 return cpu_curr(cpu);
10316 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
10320 * ia64_set_curr_task - set the current task for a given CPU.
10321 * @cpu: the processor in question.
10322 * @p: the task pointer to set.
10324 * Description: This function must only be used when non-maskable interrupts
10325 * are serviced on a separate stack. It allows the architecture to switch the
10326 * notion of the current task on a CPU in a non-blocking manner. This function
10327 * must be called with all CPU's synchronized, and interrupts disabled, the
10328 * and caller must save the original value of the current task (see
10329 * curr_task() above) and restore that value before reenabling interrupts and
10330 * re-starting the system.
10332 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10334 void ia64_set_curr_task(int cpu, struct task_struct *p)
10341 #ifdef CONFIG_CGROUP_SCHED
10342 /* task_group_lock serializes the addition/removal of task groups */
10343 static DEFINE_SPINLOCK(task_group_lock);
10345 static inline void alloc_uclamp_sched_group(struct task_group *tg,
10346 struct task_group *parent)
10348 #ifdef CONFIG_UCLAMP_TASK_GROUP
10349 enum uclamp_id clamp_id;
10351 for_each_clamp_id(clamp_id) {
10352 uclamp_se_set(&tg->uclamp_req[clamp_id],
10353 uclamp_none(clamp_id), false);
10354 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10359 static void sched_free_group(struct task_group *tg)
10361 free_fair_sched_group(tg);
10362 free_rt_sched_group(tg);
10363 autogroup_free(tg);
10364 kmem_cache_free(task_group_cache, tg);
10367 static void sched_free_group_rcu(struct rcu_head *rcu)
10369 sched_free_group(container_of(rcu, struct task_group, rcu));
10372 static void sched_unregister_group(struct task_group *tg)
10374 unregister_fair_sched_group(tg);
10375 unregister_rt_sched_group(tg);
10377 * We have to wait for yet another RCU grace period to expire, as
10378 * print_cfs_stats() might run concurrently.
10380 call_rcu(&tg->rcu, sched_free_group_rcu);
10383 /* allocate runqueue etc for a new task group */
10384 struct task_group *sched_create_group(struct task_group *parent)
10386 struct task_group *tg;
10388 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10390 return ERR_PTR(-ENOMEM);
10392 if (!alloc_fair_sched_group(tg, parent))
10395 if (!alloc_rt_sched_group(tg, parent))
10398 alloc_uclamp_sched_group(tg, parent);
10403 sched_free_group(tg);
10404 return ERR_PTR(-ENOMEM);
10407 void sched_online_group(struct task_group *tg, struct task_group *parent)
10409 unsigned long flags;
10411 spin_lock_irqsave(&task_group_lock, flags);
10412 list_add_rcu(&tg->list, &task_groups);
10414 /* Root should already exist: */
10417 tg->parent = parent;
10418 INIT_LIST_HEAD(&tg->children);
10419 list_add_rcu(&tg->siblings, &parent->children);
10420 spin_unlock_irqrestore(&task_group_lock, flags);
10422 online_fair_sched_group(tg);
10425 /* rcu callback to free various structures associated with a task group */
10426 static void sched_unregister_group_rcu(struct rcu_head *rhp)
10428 /* Now it should be safe to free those cfs_rqs: */
10429 sched_unregister_group(container_of(rhp, struct task_group, rcu));
10432 void sched_destroy_group(struct task_group *tg)
10434 /* Wait for possible concurrent references to cfs_rqs complete: */
10435 call_rcu(&tg->rcu, sched_unregister_group_rcu);
10438 void sched_release_group(struct task_group *tg)
10440 unsigned long flags;
10443 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10444 * sched_cfs_period_timer()).
10446 * For this to be effective, we have to wait for all pending users of
10447 * this task group to leave their RCU critical section to ensure no new
10448 * user will see our dying task group any more. Specifically ensure
10449 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10451 * We therefore defer calling unregister_fair_sched_group() to
10452 * sched_unregister_group() which is guarantied to get called only after the
10453 * current RCU grace period has expired.
10455 spin_lock_irqsave(&task_group_lock, flags);
10456 list_del_rcu(&tg->list);
10457 list_del_rcu(&tg->siblings);
10458 spin_unlock_irqrestore(&task_group_lock, flags);
10461 static struct task_group *sched_get_task_group(struct task_struct *tsk)
10463 struct task_group *tg;
10466 * All callers are synchronized by task_rq_lock(); we do not use RCU
10467 * which is pointless here. Thus, we pass "true" to task_css_check()
10468 * to prevent lockdep warnings.
10470 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10471 struct task_group, css);
10472 tg = autogroup_task_group(tsk, tg);
10477 static void sched_change_group(struct task_struct *tsk, struct task_group *group)
10479 tsk->sched_task_group = group;
10481 #ifdef CONFIG_FAIR_GROUP_SCHED
10482 if (tsk->sched_class->task_change_group)
10483 tsk->sched_class->task_change_group(tsk);
10486 set_task_rq(tsk, task_cpu(tsk));
10490 * Change task's runqueue when it moves between groups.
10492 * The caller of this function should have put the task in its new group by
10493 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10496 void sched_move_task(struct task_struct *tsk)
10498 int queued, running, queue_flags =
10499 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10500 struct task_group *group;
10501 struct rq_flags rf;
10504 rq = task_rq_lock(tsk, &rf);
10506 * Esp. with SCHED_AUTOGROUP enabled it is possible to get superfluous
10509 group = sched_get_task_group(tsk);
10510 if (group == tsk->sched_task_group)
10513 update_rq_clock(rq);
10515 running = task_current(rq, tsk);
10516 queued = task_on_rq_queued(tsk);
10519 dequeue_task(rq, tsk, queue_flags);
10521 put_prev_task(rq, tsk);
10523 sched_change_group(tsk, group);
10526 enqueue_task(rq, tsk, queue_flags);
10528 set_next_task(rq, tsk);
10530 * After changing group, the running task may have joined a
10531 * throttled one but it's still the running task. Trigger a
10532 * resched to make sure that task can still run.
10538 task_rq_unlock(rq, tsk, &rf);
10541 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10543 return css ? container_of(css, struct task_group, css) : NULL;
10546 static struct cgroup_subsys_state *
10547 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10549 struct task_group *parent = css_tg(parent_css);
10550 struct task_group *tg;
10553 /* This is early initialization for the top cgroup */
10554 return &root_task_group.css;
10557 tg = sched_create_group(parent);
10559 return ERR_PTR(-ENOMEM);
10564 /* Expose task group only after completing cgroup initialization */
10565 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10567 struct task_group *tg = css_tg(css);
10568 struct task_group *parent = css_tg(css->parent);
10571 sched_online_group(tg, parent);
10573 #ifdef CONFIG_UCLAMP_TASK_GROUP
10574 /* Propagate the effective uclamp value for the new group */
10575 mutex_lock(&uclamp_mutex);
10577 cpu_util_update_eff(css);
10579 mutex_unlock(&uclamp_mutex);
10585 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10587 struct task_group *tg = css_tg(css);
10589 sched_release_group(tg);
10592 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10594 struct task_group *tg = css_tg(css);
10597 * Relies on the RCU grace period between css_released() and this.
10599 sched_unregister_group(tg);
10602 #ifdef CONFIG_RT_GROUP_SCHED
10603 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10605 struct task_struct *task;
10606 struct cgroup_subsys_state *css;
10608 cgroup_taskset_for_each(task, css, tset) {
10609 if (!sched_rt_can_attach(css_tg(css), task))
10616 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10618 struct task_struct *task;
10619 struct cgroup_subsys_state *css;
10621 cgroup_taskset_for_each(task, css, tset)
10622 sched_move_task(task);
10625 #ifdef CONFIG_UCLAMP_TASK_GROUP
10626 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10628 struct cgroup_subsys_state *top_css = css;
10629 struct uclamp_se *uc_parent = NULL;
10630 struct uclamp_se *uc_se = NULL;
10631 unsigned int eff[UCLAMP_CNT];
10632 enum uclamp_id clamp_id;
10633 unsigned int clamps;
10635 lockdep_assert_held(&uclamp_mutex);
10636 SCHED_WARN_ON(!rcu_read_lock_held());
10638 css_for_each_descendant_pre(css, top_css) {
10639 uc_parent = css_tg(css)->parent
10640 ? css_tg(css)->parent->uclamp : NULL;
10642 for_each_clamp_id(clamp_id) {
10643 /* Assume effective clamps matches requested clamps */
10644 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10645 /* Cap effective clamps with parent's effective clamps */
10647 eff[clamp_id] > uc_parent[clamp_id].value) {
10648 eff[clamp_id] = uc_parent[clamp_id].value;
10651 /* Ensure protection is always capped by limit */
10652 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10654 /* Propagate most restrictive effective clamps */
10656 uc_se = css_tg(css)->uclamp;
10657 for_each_clamp_id(clamp_id) {
10658 if (eff[clamp_id] == uc_se[clamp_id].value)
10660 uc_se[clamp_id].value = eff[clamp_id];
10661 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10662 clamps |= (0x1 << clamp_id);
10665 css = css_rightmost_descendant(css);
10669 /* Immediately update descendants RUNNABLE tasks */
10670 uclamp_update_active_tasks(css);
10675 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10676 * C expression. Since there is no way to convert a macro argument (N) into a
10677 * character constant, use two levels of macros.
10679 #define _POW10(exp) ((unsigned int)1e##exp)
10680 #define POW10(exp) _POW10(exp)
10682 struct uclamp_request {
10683 #define UCLAMP_PERCENT_SHIFT 2
10684 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10690 static inline struct uclamp_request
10691 capacity_from_percent(char *buf)
10693 struct uclamp_request req = {
10694 .percent = UCLAMP_PERCENT_SCALE,
10695 .util = SCHED_CAPACITY_SCALE,
10700 if (strcmp(buf, "max")) {
10701 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10705 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10710 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10711 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10717 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10718 size_t nbytes, loff_t off,
10719 enum uclamp_id clamp_id)
10721 struct uclamp_request req;
10722 struct task_group *tg;
10724 req = capacity_from_percent(buf);
10728 static_branch_enable(&sched_uclamp_used);
10730 mutex_lock(&uclamp_mutex);
10733 tg = css_tg(of_css(of));
10734 if (tg->uclamp_req[clamp_id].value != req.util)
10735 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10738 * Because of not recoverable conversion rounding we keep track of the
10739 * exact requested value
10741 tg->uclamp_pct[clamp_id] = req.percent;
10743 /* Update effective clamps to track the most restrictive value */
10744 cpu_util_update_eff(of_css(of));
10747 mutex_unlock(&uclamp_mutex);
10752 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10753 char *buf, size_t nbytes,
10756 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10759 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10760 char *buf, size_t nbytes,
10763 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10766 static inline void cpu_uclamp_print(struct seq_file *sf,
10767 enum uclamp_id clamp_id)
10769 struct task_group *tg;
10775 tg = css_tg(seq_css(sf));
10776 util_clamp = tg->uclamp_req[clamp_id].value;
10779 if (util_clamp == SCHED_CAPACITY_SCALE) {
10780 seq_puts(sf, "max\n");
10784 percent = tg->uclamp_pct[clamp_id];
10785 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10786 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10789 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10791 cpu_uclamp_print(sf, UCLAMP_MIN);
10795 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10797 cpu_uclamp_print(sf, UCLAMP_MAX);
10800 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10802 #ifdef CONFIG_FAIR_GROUP_SCHED
10803 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10804 struct cftype *cftype, u64 shareval)
10806 if (shareval > scale_load_down(ULONG_MAX))
10807 shareval = MAX_SHARES;
10808 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10811 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10812 struct cftype *cft)
10814 struct task_group *tg = css_tg(css);
10816 return (u64) scale_load_down(tg->shares);
10819 #ifdef CONFIG_CFS_BANDWIDTH
10820 static DEFINE_MUTEX(cfs_constraints_mutex);
10822 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10823 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10824 /* More than 203 days if BW_SHIFT equals 20. */
10825 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10827 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10829 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10832 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10833 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10835 if (tg == &root_task_group)
10839 * Ensure we have at some amount of bandwidth every period. This is
10840 * to prevent reaching a state of large arrears when throttled via
10841 * entity_tick() resulting in prolonged exit starvation.
10843 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10847 * Likewise, bound things on the other side by preventing insane quota
10848 * periods. This also allows us to normalize in computing quota
10851 if (period > max_cfs_quota_period)
10855 * Bound quota to defend quota against overflow during bandwidth shift.
10857 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10860 if (quota != RUNTIME_INF && (burst > quota ||
10861 burst + quota > max_cfs_runtime))
10865 * Prevent race between setting of cfs_rq->runtime_enabled and
10866 * unthrottle_offline_cfs_rqs().
10869 mutex_lock(&cfs_constraints_mutex);
10870 ret = __cfs_schedulable(tg, period, quota);
10874 runtime_enabled = quota != RUNTIME_INF;
10875 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10877 * If we need to toggle cfs_bandwidth_used, off->on must occur
10878 * before making related changes, and on->off must occur afterwards
10880 if (runtime_enabled && !runtime_was_enabled)
10881 cfs_bandwidth_usage_inc();
10882 raw_spin_lock_irq(&cfs_b->lock);
10883 cfs_b->period = ns_to_ktime(period);
10884 cfs_b->quota = quota;
10885 cfs_b->burst = burst;
10887 __refill_cfs_bandwidth_runtime(cfs_b);
10889 /* Restart the period timer (if active) to handle new period expiry: */
10890 if (runtime_enabled)
10891 start_cfs_bandwidth(cfs_b);
10893 raw_spin_unlock_irq(&cfs_b->lock);
10895 for_each_online_cpu(i) {
10896 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10897 struct rq *rq = cfs_rq->rq;
10898 struct rq_flags rf;
10900 rq_lock_irq(rq, &rf);
10901 cfs_rq->runtime_enabled = runtime_enabled;
10902 cfs_rq->runtime_remaining = 0;
10904 if (cfs_rq->throttled)
10905 unthrottle_cfs_rq(cfs_rq);
10906 rq_unlock_irq(rq, &rf);
10908 if (runtime_was_enabled && !runtime_enabled)
10909 cfs_bandwidth_usage_dec();
10911 mutex_unlock(&cfs_constraints_mutex);
10912 cpus_read_unlock();
10917 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10919 u64 quota, period, burst;
10921 period = ktime_to_ns(tg->cfs_bandwidth.period);
10922 burst = tg->cfs_bandwidth.burst;
10923 if (cfs_quota_us < 0)
10924 quota = RUNTIME_INF;
10925 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10926 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10930 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10933 static long tg_get_cfs_quota(struct task_group *tg)
10937 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10940 quota_us = tg->cfs_bandwidth.quota;
10941 do_div(quota_us, NSEC_PER_USEC);
10946 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10948 u64 quota, period, burst;
10950 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10953 period = (u64)cfs_period_us * NSEC_PER_USEC;
10954 quota = tg->cfs_bandwidth.quota;
10955 burst = tg->cfs_bandwidth.burst;
10957 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10960 static long tg_get_cfs_period(struct task_group *tg)
10964 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10965 do_div(cfs_period_us, NSEC_PER_USEC);
10967 return cfs_period_us;
10970 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10972 u64 quota, period, burst;
10974 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10977 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10978 period = ktime_to_ns(tg->cfs_bandwidth.period);
10979 quota = tg->cfs_bandwidth.quota;
10981 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10984 static long tg_get_cfs_burst(struct task_group *tg)
10988 burst_us = tg->cfs_bandwidth.burst;
10989 do_div(burst_us, NSEC_PER_USEC);
10994 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10995 struct cftype *cft)
10997 return tg_get_cfs_quota(css_tg(css));
11000 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
11001 struct cftype *cftype, s64 cfs_quota_us)
11003 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
11006 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
11007 struct cftype *cft)
11009 return tg_get_cfs_period(css_tg(css));
11012 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
11013 struct cftype *cftype, u64 cfs_period_us)
11015 return tg_set_cfs_period(css_tg(css), cfs_period_us);
11018 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
11019 struct cftype *cft)
11021 return tg_get_cfs_burst(css_tg(css));
11024 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
11025 struct cftype *cftype, u64 cfs_burst_us)
11027 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
11030 struct cfs_schedulable_data {
11031 struct task_group *tg;
11036 * normalize group quota/period to be quota/max_period
11037 * note: units are usecs
11039 static u64 normalize_cfs_quota(struct task_group *tg,
11040 struct cfs_schedulable_data *d)
11045 period = d->period;
11048 period = tg_get_cfs_period(tg);
11049 quota = tg_get_cfs_quota(tg);
11052 /* note: these should typically be equivalent */
11053 if (quota == RUNTIME_INF || quota == -1)
11054 return RUNTIME_INF;
11056 return to_ratio(period, quota);
11059 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
11061 struct cfs_schedulable_data *d = data;
11062 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11063 s64 quota = 0, parent_quota = -1;
11066 quota = RUNTIME_INF;
11068 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
11070 quota = normalize_cfs_quota(tg, d);
11071 parent_quota = parent_b->hierarchical_quota;
11074 * Ensure max(child_quota) <= parent_quota. On cgroup2,
11075 * always take the non-RUNTIME_INF min. On cgroup1, only
11076 * inherit when no limit is set. In both cases this is used
11077 * by the scheduler to determine if a given CFS task has a
11078 * bandwidth constraint at some higher level.
11080 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
11081 if (quota == RUNTIME_INF)
11082 quota = parent_quota;
11083 else if (parent_quota != RUNTIME_INF)
11084 quota = min(quota, parent_quota);
11086 if (quota == RUNTIME_INF)
11087 quota = parent_quota;
11088 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
11092 cfs_b->hierarchical_quota = quota;
11097 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
11100 struct cfs_schedulable_data data = {
11106 if (quota != RUNTIME_INF) {
11107 do_div(data.period, NSEC_PER_USEC);
11108 do_div(data.quota, NSEC_PER_USEC);
11112 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
11118 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
11120 struct task_group *tg = css_tg(seq_css(sf));
11121 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11123 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
11124 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
11125 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
11127 if (schedstat_enabled() && tg != &root_task_group) {
11128 struct sched_statistics *stats;
11132 for_each_possible_cpu(i) {
11133 stats = __schedstats_from_se(tg->se[i]);
11134 ws += schedstat_val(stats->wait_sum);
11137 seq_printf(sf, "wait_sum %llu\n", ws);
11140 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
11141 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
11146 static u64 throttled_time_self(struct task_group *tg)
11151 for_each_possible_cpu(i) {
11152 total += READ_ONCE(tg->cfs_rq[i]->throttled_clock_self_time);
11158 static int cpu_cfs_local_stat_show(struct seq_file *sf, void *v)
11160 struct task_group *tg = css_tg(seq_css(sf));
11162 seq_printf(sf, "throttled_time %llu\n", throttled_time_self(tg));
11166 #endif /* CONFIG_CFS_BANDWIDTH */
11167 #endif /* CONFIG_FAIR_GROUP_SCHED */
11169 #ifdef CONFIG_RT_GROUP_SCHED
11170 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
11171 struct cftype *cft, s64 val)
11173 return sched_group_set_rt_runtime(css_tg(css), val);
11176 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
11177 struct cftype *cft)
11179 return sched_group_rt_runtime(css_tg(css));
11182 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
11183 struct cftype *cftype, u64 rt_period_us)
11185 return sched_group_set_rt_period(css_tg(css), rt_period_us);
11188 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
11189 struct cftype *cft)
11191 return sched_group_rt_period(css_tg(css));
11193 #endif /* CONFIG_RT_GROUP_SCHED */
11195 #ifdef CONFIG_FAIR_GROUP_SCHED
11196 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
11197 struct cftype *cft)
11199 return css_tg(css)->idle;
11202 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
11203 struct cftype *cft, s64 idle)
11205 return sched_group_set_idle(css_tg(css), idle);
11209 static struct cftype cpu_legacy_files[] = {
11210 #ifdef CONFIG_FAIR_GROUP_SCHED
11213 .read_u64 = cpu_shares_read_u64,
11214 .write_u64 = cpu_shares_write_u64,
11218 .read_s64 = cpu_idle_read_s64,
11219 .write_s64 = cpu_idle_write_s64,
11222 #ifdef CONFIG_CFS_BANDWIDTH
11224 .name = "cfs_quota_us",
11225 .read_s64 = cpu_cfs_quota_read_s64,
11226 .write_s64 = cpu_cfs_quota_write_s64,
11229 .name = "cfs_period_us",
11230 .read_u64 = cpu_cfs_period_read_u64,
11231 .write_u64 = cpu_cfs_period_write_u64,
11234 .name = "cfs_burst_us",
11235 .read_u64 = cpu_cfs_burst_read_u64,
11236 .write_u64 = cpu_cfs_burst_write_u64,
11240 .seq_show = cpu_cfs_stat_show,
11243 .name = "stat.local",
11244 .seq_show = cpu_cfs_local_stat_show,
11247 #ifdef CONFIG_RT_GROUP_SCHED
11249 .name = "rt_runtime_us",
11250 .read_s64 = cpu_rt_runtime_read,
11251 .write_s64 = cpu_rt_runtime_write,
11254 .name = "rt_period_us",
11255 .read_u64 = cpu_rt_period_read_uint,
11256 .write_u64 = cpu_rt_period_write_uint,
11259 #ifdef CONFIG_UCLAMP_TASK_GROUP
11261 .name = "uclamp.min",
11262 .flags = CFTYPE_NOT_ON_ROOT,
11263 .seq_show = cpu_uclamp_min_show,
11264 .write = cpu_uclamp_min_write,
11267 .name = "uclamp.max",
11268 .flags = CFTYPE_NOT_ON_ROOT,
11269 .seq_show = cpu_uclamp_max_show,
11270 .write = cpu_uclamp_max_write,
11273 { } /* Terminate */
11276 static int cpu_extra_stat_show(struct seq_file *sf,
11277 struct cgroup_subsys_state *css)
11279 #ifdef CONFIG_CFS_BANDWIDTH
11281 struct task_group *tg = css_tg(css);
11282 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
11283 u64 throttled_usec, burst_usec;
11285 throttled_usec = cfs_b->throttled_time;
11286 do_div(throttled_usec, NSEC_PER_USEC);
11287 burst_usec = cfs_b->burst_time;
11288 do_div(burst_usec, NSEC_PER_USEC);
11290 seq_printf(sf, "nr_periods %d\n"
11291 "nr_throttled %d\n"
11292 "throttled_usec %llu\n"
11294 "burst_usec %llu\n",
11295 cfs_b->nr_periods, cfs_b->nr_throttled,
11296 throttled_usec, cfs_b->nr_burst, burst_usec);
11302 static int cpu_local_stat_show(struct seq_file *sf,
11303 struct cgroup_subsys_state *css)
11305 #ifdef CONFIG_CFS_BANDWIDTH
11307 struct task_group *tg = css_tg(css);
11308 u64 throttled_self_usec;
11310 throttled_self_usec = throttled_time_self(tg);
11311 do_div(throttled_self_usec, NSEC_PER_USEC);
11313 seq_printf(sf, "throttled_usec %llu\n",
11314 throttled_self_usec);
11320 #ifdef CONFIG_FAIR_GROUP_SCHED
11321 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
11322 struct cftype *cft)
11324 struct task_group *tg = css_tg(css);
11325 u64 weight = scale_load_down(tg->shares);
11327 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11330 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11331 struct cftype *cft, u64 weight)
11334 * cgroup weight knobs should use the common MIN, DFL and MAX
11335 * values which are 1, 100 and 10000 respectively. While it loses
11336 * a bit of range on both ends, it maps pretty well onto the shares
11337 * value used by scheduler and the round-trip conversions preserve
11338 * the original value over the entire range.
11340 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11343 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11345 return sched_group_set_shares(css_tg(css), scale_load(weight));
11348 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11349 struct cftype *cft)
11351 unsigned long weight = scale_load_down(css_tg(css)->shares);
11352 int last_delta = INT_MAX;
11355 /* find the closest nice value to the current weight */
11356 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11357 delta = abs(sched_prio_to_weight[prio] - weight);
11358 if (delta >= last_delta)
11360 last_delta = delta;
11363 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11366 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11367 struct cftype *cft, s64 nice)
11369 unsigned long weight;
11372 if (nice < MIN_NICE || nice > MAX_NICE)
11375 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11376 idx = array_index_nospec(idx, 40);
11377 weight = sched_prio_to_weight[idx];
11379 return sched_group_set_shares(css_tg(css), scale_load(weight));
11383 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11384 long period, long quota)
11387 seq_puts(sf, "max");
11389 seq_printf(sf, "%ld", quota);
11391 seq_printf(sf, " %ld\n", period);
11394 /* caller should put the current value in *@periodp before calling */
11395 static int __maybe_unused cpu_period_quota_parse(char *buf,
11396 u64 *periodp, u64 *quotap)
11398 char tok[21]; /* U64_MAX */
11400 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11403 *periodp *= NSEC_PER_USEC;
11405 if (sscanf(tok, "%llu", quotap))
11406 *quotap *= NSEC_PER_USEC;
11407 else if (!strcmp(tok, "max"))
11408 *quotap = RUNTIME_INF;
11415 #ifdef CONFIG_CFS_BANDWIDTH
11416 static int cpu_max_show(struct seq_file *sf, void *v)
11418 struct task_group *tg = css_tg(seq_css(sf));
11420 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11424 static ssize_t cpu_max_write(struct kernfs_open_file *of,
11425 char *buf, size_t nbytes, loff_t off)
11427 struct task_group *tg = css_tg(of_css(of));
11428 u64 period = tg_get_cfs_period(tg);
11429 u64 burst = tg_get_cfs_burst(tg);
11433 ret = cpu_period_quota_parse(buf, &period, "a);
11435 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11436 return ret ?: nbytes;
11440 static struct cftype cpu_files[] = {
11441 #ifdef CONFIG_FAIR_GROUP_SCHED
11444 .flags = CFTYPE_NOT_ON_ROOT,
11445 .read_u64 = cpu_weight_read_u64,
11446 .write_u64 = cpu_weight_write_u64,
11449 .name = "weight.nice",
11450 .flags = CFTYPE_NOT_ON_ROOT,
11451 .read_s64 = cpu_weight_nice_read_s64,
11452 .write_s64 = cpu_weight_nice_write_s64,
11456 .flags = CFTYPE_NOT_ON_ROOT,
11457 .read_s64 = cpu_idle_read_s64,
11458 .write_s64 = cpu_idle_write_s64,
11461 #ifdef CONFIG_CFS_BANDWIDTH
11464 .flags = CFTYPE_NOT_ON_ROOT,
11465 .seq_show = cpu_max_show,
11466 .write = cpu_max_write,
11469 .name = "max.burst",
11470 .flags = CFTYPE_NOT_ON_ROOT,
11471 .read_u64 = cpu_cfs_burst_read_u64,
11472 .write_u64 = cpu_cfs_burst_write_u64,
11475 #ifdef CONFIG_UCLAMP_TASK_GROUP
11477 .name = "uclamp.min",
11478 .flags = CFTYPE_NOT_ON_ROOT,
11479 .seq_show = cpu_uclamp_min_show,
11480 .write = cpu_uclamp_min_write,
11483 .name = "uclamp.max",
11484 .flags = CFTYPE_NOT_ON_ROOT,
11485 .seq_show = cpu_uclamp_max_show,
11486 .write = cpu_uclamp_max_write,
11489 { } /* terminate */
11492 struct cgroup_subsys cpu_cgrp_subsys = {
11493 .css_alloc = cpu_cgroup_css_alloc,
11494 .css_online = cpu_cgroup_css_online,
11495 .css_released = cpu_cgroup_css_released,
11496 .css_free = cpu_cgroup_css_free,
11497 .css_extra_stat_show = cpu_extra_stat_show,
11498 .css_local_stat_show = cpu_local_stat_show,
11499 #ifdef CONFIG_RT_GROUP_SCHED
11500 .can_attach = cpu_cgroup_can_attach,
11502 .attach = cpu_cgroup_attach,
11503 .legacy_cftypes = cpu_legacy_files,
11504 .dfl_cftypes = cpu_files,
11505 .early_init = true,
11509 #endif /* CONFIG_CGROUP_SCHED */
11511 void dump_cpu_task(int cpu)
11513 if (cpu == smp_processor_id() && in_hardirq()) {
11514 struct pt_regs *regs;
11516 regs = get_irq_regs();
11523 if (trigger_single_cpu_backtrace(cpu))
11526 pr_info("Task dump for CPU %d:\n", cpu);
11527 sched_show_task(cpu_curr(cpu));
11531 * Nice levels are multiplicative, with a gentle 10% change for every
11532 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11533 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11534 * that remained on nice 0.
11536 * The "10% effect" is relative and cumulative: from _any_ nice level,
11537 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11538 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11539 * If a task goes up by ~10% and another task goes down by ~10% then
11540 * the relative distance between them is ~25%.)
11542 const int sched_prio_to_weight[40] = {
11543 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11544 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11545 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11546 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11547 /* 0 */ 1024, 820, 655, 526, 423,
11548 /* 5 */ 335, 272, 215, 172, 137,
11549 /* 10 */ 110, 87, 70, 56, 45,
11550 /* 15 */ 36, 29, 23, 18, 15,
11554 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11556 * In cases where the weight does not change often, we can use the
11557 * precalculated inverse to speed up arithmetics by turning divisions
11558 * into multiplications:
11560 const u32 sched_prio_to_wmult[40] = {
11561 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11562 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11563 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11564 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11565 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11566 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11567 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11568 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11571 void call_trace_sched_update_nr_running(struct rq *rq, int count)
11573 trace_sched_update_nr_running_tp(rq, count);
11576 #ifdef CONFIG_SCHED_MM_CID
11579 * @cid_lock: Guarantee forward-progress of cid allocation.
11581 * Concurrency ID allocation within a bitmap is mostly lock-free. The cid_lock
11582 * is only used when contention is detected by the lock-free allocation so
11583 * forward progress can be guaranteed.
11585 DEFINE_RAW_SPINLOCK(cid_lock);
11588 * @use_cid_lock: Select cid allocation behavior: lock-free vs spinlock.
11590 * When @use_cid_lock is 0, the cid allocation is lock-free. When contention is
11591 * detected, it is set to 1 to ensure that all newly coming allocations are
11592 * serialized by @cid_lock until the allocation which detected contention
11593 * completes and sets @use_cid_lock back to 0. This guarantees forward progress
11594 * of a cid allocation.
11599 * mm_cid remote-clear implements a lock-free algorithm to clear per-mm/cpu cid
11600 * concurrently with respect to the execution of the source runqueue context
11603 * There is one basic properties we want to guarantee here:
11605 * (1) Remote-clear should _never_ mark a per-cpu cid UNSET when it is actively
11606 * used by a task. That would lead to concurrent allocation of the cid and
11607 * userspace corruption.
11609 * Provide this guarantee by introducing a Dekker memory ordering to guarantee
11610 * that a pair of loads observe at least one of a pair of stores, which can be
11619 * Which guarantees that x==0 && y==0 is impossible. But rather than using
11620 * values 0 and 1, this algorithm cares about specific state transitions of the
11621 * runqueue current task (as updated by the scheduler context switch), and the
11622 * per-mm/cpu cid value.
11624 * Let's introduce task (Y) which has task->mm == mm and task (N) which has
11625 * task->mm != mm for the rest of the discussion. There are two scheduler state
11626 * transitions on context switch we care about:
11628 * (TSA) Store to rq->curr with transition from (N) to (Y)
11630 * (TSB) Store to rq->curr with transition from (Y) to (N)
11632 * On the remote-clear side, there is one transition we care about:
11634 * (TMA) cmpxchg to *pcpu_cid to set the LAZY flag
11636 * There is also a transition to UNSET state which can be performed from all
11637 * sides (scheduler, remote-clear). It is always performed with a cmpxchg which
11638 * guarantees that only a single thread will succeed:
11640 * (TMB) cmpxchg to *pcpu_cid to mark UNSET
11642 * Just to be clear, what we do _not_ want to happen is a transition to UNSET
11643 * when a thread is actively using the cid (property (1)).
11645 * Let's looks at the relevant combinations of TSA/TSB, and TMA transitions.
11647 * Scenario A) (TSA)+(TMA) (from next task perspective)
11651 * Context switch CS-1 Remote-clear
11652 * - store to rq->curr: (N)->(Y) (TSA) - cmpxchg to *pcpu_id to LAZY (TMA)
11653 * (implied barrier after cmpxchg)
11654 * - switch_mm_cid()
11655 * - memory barrier (see switch_mm_cid()
11656 * comment explaining how this barrier
11657 * is combined with other scheduler
11659 * - mm_cid_get (next)
11660 * - READ_ONCE(*pcpu_cid) - rcu_dereference(src_rq->curr)
11662 * This Dekker ensures that either task (Y) is observed by the
11663 * rcu_dereference() or the LAZY flag is observed by READ_ONCE(), or both are
11666 * If task (Y) store is observed by rcu_dereference(), it means that there is
11667 * still an active task on the cpu. Remote-clear will therefore not transition
11668 * to UNSET, which fulfills property (1).
11670 * If task (Y) is not observed, but the lazy flag is observed by READ_ONCE(),
11671 * it will move its state to UNSET, which clears the percpu cid perhaps
11672 * uselessly (which is not an issue for correctness). Because task (Y) is not
11673 * observed, CPU1 can move ahead to set the state to UNSET. Because moving
11674 * state to UNSET is done with a cmpxchg expecting that the old state has the
11675 * LAZY flag set, only one thread will successfully UNSET.
11677 * If both states (LAZY flag and task (Y)) are observed, the thread on CPU0
11678 * will observe the LAZY flag and transition to UNSET (perhaps uselessly), and
11679 * CPU1 will observe task (Y) and do nothing more, which is fine.
11681 * What we are effectively preventing with this Dekker is a scenario where
11682 * neither LAZY flag nor store (Y) are observed, which would fail property (1)
11683 * because this would UNSET a cid which is actively used.
11686 void sched_mm_cid_migrate_from(struct task_struct *t)
11688 t->migrate_from_cpu = task_cpu(t);
11692 int __sched_mm_cid_migrate_from_fetch_cid(struct rq *src_rq,
11693 struct task_struct *t,
11694 struct mm_cid *src_pcpu_cid)
11696 struct mm_struct *mm = t->mm;
11697 struct task_struct *src_task;
11698 int src_cid, last_mm_cid;
11703 last_mm_cid = t->last_mm_cid;
11705 * If the migrated task has no last cid, or if the current
11706 * task on src rq uses the cid, it means the source cid does not need
11707 * to be moved to the destination cpu.
11709 if (last_mm_cid == -1)
11711 src_cid = READ_ONCE(src_pcpu_cid->cid);
11712 if (!mm_cid_is_valid(src_cid) || last_mm_cid != src_cid)
11716 * If we observe an active task using the mm on this rq, it means we
11717 * are not the last task to be migrated from this cpu for this mm, so
11718 * there is no need to move src_cid to the destination cpu.
11721 src_task = rcu_dereference(src_rq->curr);
11722 if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11724 t->last_mm_cid = -1;
11733 int __sched_mm_cid_migrate_from_try_steal_cid(struct rq *src_rq,
11734 struct task_struct *t,
11735 struct mm_cid *src_pcpu_cid,
11738 struct task_struct *src_task;
11739 struct mm_struct *mm = t->mm;
11746 * Attempt to clear the source cpu cid to move it to the destination
11749 lazy_cid = mm_cid_set_lazy_put(src_cid);
11750 if (!try_cmpxchg(&src_pcpu_cid->cid, &src_cid, lazy_cid))
11754 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11755 * rq->curr->mm matches the scheduler barrier in context_switch()
11756 * between store to rq->curr and load of prev and next task's
11759 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11760 * rq->curr->mm_cid_active matches the barrier in
11761 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11762 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11763 * load of per-mm/cpu cid.
11767 * If we observe an active task using the mm on this rq after setting
11768 * the lazy-put flag, this task will be responsible for transitioning
11769 * from lazy-put flag set to MM_CID_UNSET.
11772 src_task = rcu_dereference(src_rq->curr);
11773 if (READ_ONCE(src_task->mm_cid_active) && src_task->mm == mm) {
11776 * We observed an active task for this mm, there is therefore
11777 * no point in moving this cid to the destination cpu.
11779 t->last_mm_cid = -1;
11785 * The src_cid is unused, so it can be unset.
11787 if (!try_cmpxchg(&src_pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11793 * Migration to dst cpu. Called with dst_rq lock held.
11794 * Interrupts are disabled, which keeps the window of cid ownership without the
11795 * source rq lock held small.
11797 void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t)
11799 struct mm_cid *src_pcpu_cid, *dst_pcpu_cid;
11800 struct mm_struct *mm = t->mm;
11801 int src_cid, dst_cid, src_cpu;
11804 lockdep_assert_rq_held(dst_rq);
11808 src_cpu = t->migrate_from_cpu;
11809 if (src_cpu == -1) {
11810 t->last_mm_cid = -1;
11814 * Move the src cid if the dst cid is unset. This keeps id
11815 * allocation closest to 0 in cases where few threads migrate around
11818 * If destination cid is already set, we may have to just clear
11819 * the src cid to ensure compactness in frequent migrations
11822 * It is not useful to clear the src cid when the number of threads is
11823 * greater or equal to the number of allowed cpus, because user-space
11824 * can expect that the number of allowed cids can reach the number of
11827 dst_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(dst_rq));
11828 dst_cid = READ_ONCE(dst_pcpu_cid->cid);
11829 if (!mm_cid_is_unset(dst_cid) &&
11830 atomic_read(&mm->mm_users) >= t->nr_cpus_allowed)
11832 src_pcpu_cid = per_cpu_ptr(mm->pcpu_cid, src_cpu);
11833 src_rq = cpu_rq(src_cpu);
11834 src_cid = __sched_mm_cid_migrate_from_fetch_cid(src_rq, t, src_pcpu_cid);
11837 src_cid = __sched_mm_cid_migrate_from_try_steal_cid(src_rq, t, src_pcpu_cid,
11841 if (!mm_cid_is_unset(dst_cid)) {
11842 __mm_cid_put(mm, src_cid);
11845 /* Move src_cid to dst cpu. */
11846 mm_cid_snapshot_time(dst_rq, mm);
11847 WRITE_ONCE(dst_pcpu_cid->cid, src_cid);
11850 static void sched_mm_cid_remote_clear(struct mm_struct *mm, struct mm_cid *pcpu_cid,
11853 struct rq *rq = cpu_rq(cpu);
11854 struct task_struct *t;
11855 unsigned long flags;
11858 cid = READ_ONCE(pcpu_cid->cid);
11859 if (!mm_cid_is_valid(cid))
11863 * Clear the cpu cid if it is set to keep cid allocation compact. If
11864 * there happens to be other tasks left on the source cpu using this
11865 * mm, the next task using this mm will reallocate its cid on context
11868 lazy_cid = mm_cid_set_lazy_put(cid);
11869 if (!try_cmpxchg(&pcpu_cid->cid, &cid, lazy_cid))
11873 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11874 * rq->curr->mm matches the scheduler barrier in context_switch()
11875 * between store to rq->curr and load of prev and next task's
11878 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11879 * rq->curr->mm_cid_active matches the barrier in
11880 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11881 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11882 * load of per-mm/cpu cid.
11886 * If we observe an active task using the mm on this rq after setting
11887 * the lazy-put flag, that task will be responsible for transitioning
11888 * from lazy-put flag set to MM_CID_UNSET.
11891 t = rcu_dereference(rq->curr);
11892 if (READ_ONCE(t->mm_cid_active) && t->mm == mm) {
11899 * The cid is unused, so it can be unset.
11900 * Disable interrupts to keep the window of cid ownership without rq
11903 local_irq_save(flags);
11904 if (try_cmpxchg(&pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
11905 __mm_cid_put(mm, cid);
11906 local_irq_restore(flags);
11909 static void sched_mm_cid_remote_clear_old(struct mm_struct *mm, int cpu)
11911 struct rq *rq = cpu_rq(cpu);
11912 struct mm_cid *pcpu_cid;
11913 struct task_struct *curr;
11917 * rq->clock load is racy on 32-bit but one spurious clear once in a
11918 * while is irrelevant.
11920 rq_clock = READ_ONCE(rq->clock);
11921 pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11924 * In order to take care of infrequently scheduled tasks, bump the time
11925 * snapshot associated with this cid if an active task using the mm is
11926 * observed on this rq.
11929 curr = rcu_dereference(rq->curr);
11930 if (READ_ONCE(curr->mm_cid_active) && curr->mm == mm) {
11931 WRITE_ONCE(pcpu_cid->time, rq_clock);
11937 if (rq_clock < pcpu_cid->time + SCHED_MM_CID_PERIOD_NS)
11939 sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11942 static void sched_mm_cid_remote_clear_weight(struct mm_struct *mm, int cpu,
11945 struct mm_cid *pcpu_cid;
11948 pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu);
11949 cid = READ_ONCE(pcpu_cid->cid);
11950 if (!mm_cid_is_valid(cid) || cid < weight)
11952 sched_mm_cid_remote_clear(mm, pcpu_cid, cpu);
11955 static void task_mm_cid_work(struct callback_head *work)
11957 unsigned long now = jiffies, old_scan, next_scan;
11958 struct task_struct *t = current;
11959 struct cpumask *cidmask;
11960 struct mm_struct *mm;
11963 SCHED_WARN_ON(t != container_of(work, struct task_struct, cid_work));
11965 work->next = work; /* Prevent double-add */
11966 if (t->flags & PF_EXITING)
11971 old_scan = READ_ONCE(mm->mm_cid_next_scan);
11972 next_scan = now + msecs_to_jiffies(MM_CID_SCAN_DELAY);
11976 res = cmpxchg(&mm->mm_cid_next_scan, old_scan, next_scan);
11977 if (res != old_scan)
11980 old_scan = next_scan;
11982 if (time_before(now, old_scan))
11984 if (!try_cmpxchg(&mm->mm_cid_next_scan, &old_scan, next_scan))
11986 cidmask = mm_cidmask(mm);
11987 /* Clear cids that were not recently used. */
11988 for_each_possible_cpu(cpu)
11989 sched_mm_cid_remote_clear_old(mm, cpu);
11990 weight = cpumask_weight(cidmask);
11992 * Clear cids that are greater or equal to the cidmask weight to
11995 for_each_possible_cpu(cpu)
11996 sched_mm_cid_remote_clear_weight(mm, cpu, weight);
11999 void init_sched_mm_cid(struct task_struct *t)
12001 struct mm_struct *mm = t->mm;
12005 mm_users = atomic_read(&mm->mm_users);
12007 mm->mm_cid_next_scan = jiffies + msecs_to_jiffies(MM_CID_SCAN_DELAY);
12009 t->cid_work.next = &t->cid_work; /* Protect against double add */
12010 init_task_work(&t->cid_work, task_mm_cid_work);
12013 void task_tick_mm_cid(struct rq *rq, struct task_struct *curr)
12015 struct callback_head *work = &curr->cid_work;
12016 unsigned long now = jiffies;
12018 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) ||
12019 work->next != work)
12021 if (time_before(now, READ_ONCE(curr->mm->mm_cid_next_scan)))
12023 task_work_add(curr, work, TWA_RESUME);
12026 void sched_mm_cid_exit_signals(struct task_struct *t)
12028 struct mm_struct *mm = t->mm;
12029 struct rq_flags rf;
12037 rq_lock_irqsave(rq, &rf);
12038 preempt_enable_no_resched(); /* holding spinlock */
12039 WRITE_ONCE(t->mm_cid_active, 0);
12041 * Store t->mm_cid_active before loading per-mm/cpu cid.
12042 * Matches barrier in sched_mm_cid_remote_clear_old().
12046 t->last_mm_cid = t->mm_cid = -1;
12047 rq_unlock_irqrestore(rq, &rf);
12050 void sched_mm_cid_before_execve(struct task_struct *t)
12052 struct mm_struct *mm = t->mm;
12053 struct rq_flags rf;
12061 rq_lock_irqsave(rq, &rf);
12062 preempt_enable_no_resched(); /* holding spinlock */
12063 WRITE_ONCE(t->mm_cid_active, 0);
12065 * Store t->mm_cid_active before loading per-mm/cpu cid.
12066 * Matches barrier in sched_mm_cid_remote_clear_old().
12070 t->last_mm_cid = t->mm_cid = -1;
12071 rq_unlock_irqrestore(rq, &rf);
12074 void sched_mm_cid_after_execve(struct task_struct *t)
12076 struct mm_struct *mm = t->mm;
12077 struct rq_flags rf;
12085 rq_lock_irqsave(rq, &rf);
12086 preempt_enable_no_resched(); /* holding spinlock */
12087 WRITE_ONCE(t->mm_cid_active, 1);
12089 * Store t->mm_cid_active before loading per-mm/cpu cid.
12090 * Matches barrier in sched_mm_cid_remote_clear_old().
12093 t->last_mm_cid = t->mm_cid = mm_cid_get(rq, mm);
12094 rq_unlock_irqrestore(rq, &rf);
12095 rseq_set_notify_resume(t);
12098 void sched_mm_cid_fork(struct task_struct *t)
12100 WARN_ON_ONCE(!t->mm || t->mm_cid != -1);
12101 t->mm_cid_active = 1;