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 #undef CREATE_TRACE_POINTS
87 #include "autogroup.h"
89 #include "autogroup.h"
94 #include "../workqueue_internal.h"
95 #include "../../io_uring/io-wq.h"
96 #include "../smpboot.h"
99 * Export tracepoints that act as a bare tracehook (ie: have no trace event
100 * associated with them) to allow external modules to probe them.
102 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
103 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
104 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
105 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
106 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
107 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
108 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
109 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
110 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
111 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
112 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
116 #ifdef CONFIG_SCHED_DEBUG
118 * Debugging: various feature bits
120 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
121 * sysctl_sched_features, defined in sched.h, to allow constants propagation
122 * at compile time and compiler optimization based on features default.
124 #define SCHED_FEAT(name, enabled) \
125 (1UL << __SCHED_FEAT_##name) * enabled |
126 const_debug unsigned int sysctl_sched_features =
127 #include "features.h"
132 * Print a warning if need_resched is set for the given duration (if
133 * LATENCY_WARN is enabled).
135 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
138 __read_mostly int sysctl_resched_latency_warn_ms = 100;
139 __read_mostly int sysctl_resched_latency_warn_once = 1;
140 #endif /* CONFIG_SCHED_DEBUG */
143 * Number of tasks to iterate in a single balance run.
144 * Limited because this is done with IRQs disabled.
146 #ifdef CONFIG_PREEMPT_RT
147 const_debug unsigned int sysctl_sched_nr_migrate = 8;
149 const_debug unsigned int sysctl_sched_nr_migrate = 32;
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(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(struct task_struct *a, struct task_struct *b, bool in_fi)
184 int pa = __task_prio(a), pb = __task_prio(b);
192 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
193 return !dl_time_before(a->dl.deadline, b->dl.deadline);
195 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
196 return cfs_prio_less(a, b, in_fi);
201 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
203 if (a->core_cookie < b->core_cookie)
206 if (a->core_cookie > b->core_cookie)
209 /* flip prio, so high prio is leftmost */
210 if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
216 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
218 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
220 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
223 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
225 const struct task_struct *p = __node_2_sc(node);
226 unsigned long cookie = (unsigned long)key;
228 if (cookie < p->core_cookie)
231 if (cookie > p->core_cookie)
237 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
239 rq->core->core_task_seq++;
244 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
247 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
249 rq->core->core_task_seq++;
251 if (sched_core_enqueued(p)) {
252 rb_erase(&p->core_node, &rq->core_tree);
253 RB_CLEAR_NODE(&p->core_node);
257 * Migrating the last task off the cpu, with the cpu in forced idle
258 * state. Reschedule to create an accounting edge for forced idle,
259 * and re-examine whether the core is still in forced idle state.
261 if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
262 rq->core->core_forceidle_count && rq->curr == rq->idle)
267 * Find left-most (aka, highest priority) task matching @cookie.
269 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
271 struct rb_node *node;
273 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
275 * The idle task always matches any cookie!
278 return idle_sched_class.pick_task(rq);
280 return __node_2_sc(node);
283 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
285 struct rb_node *node = &p->core_node;
287 node = rb_next(node);
291 p = container_of(node, struct task_struct, core_node);
292 if (p->core_cookie != cookie)
299 * Magic required such that:
301 * raw_spin_rq_lock(rq);
303 * raw_spin_rq_unlock(rq);
305 * ends up locking and unlocking the _same_ lock, and all CPUs
306 * always agree on what rq has what lock.
308 * XXX entirely possible to selectively enable cores, don't bother for now.
311 static DEFINE_MUTEX(sched_core_mutex);
312 static atomic_t sched_core_count;
313 static struct cpumask sched_core_mask;
315 static void sched_core_lock(int cpu, unsigned long *flags)
317 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
320 local_irq_save(*flags);
321 for_each_cpu(t, smt_mask)
322 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
325 static void sched_core_unlock(int cpu, unsigned long *flags)
327 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
330 for_each_cpu(t, smt_mask)
331 raw_spin_unlock(&cpu_rq(t)->__lock);
332 local_irq_restore(*flags);
335 static void __sched_core_flip(bool enabled)
343 * Toggle the online cores, one by one.
345 cpumask_copy(&sched_core_mask, cpu_online_mask);
346 for_each_cpu(cpu, &sched_core_mask) {
347 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
349 sched_core_lock(cpu, &flags);
351 for_each_cpu(t, smt_mask)
352 cpu_rq(t)->core_enabled = enabled;
354 cpu_rq(cpu)->core->core_forceidle_start = 0;
356 sched_core_unlock(cpu, &flags);
358 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
362 * Toggle the offline CPUs.
364 cpumask_copy(&sched_core_mask, cpu_possible_mask);
365 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
367 for_each_cpu(cpu, &sched_core_mask)
368 cpu_rq(cpu)->core_enabled = enabled;
373 static void sched_core_assert_empty(void)
377 for_each_possible_cpu(cpu)
378 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
381 static void __sched_core_enable(void)
383 static_branch_enable(&__sched_core_enabled);
385 * Ensure all previous instances of raw_spin_rq_*lock() have finished
386 * and future ones will observe !sched_core_disabled().
389 __sched_core_flip(true);
390 sched_core_assert_empty();
393 static void __sched_core_disable(void)
395 sched_core_assert_empty();
396 __sched_core_flip(false);
397 static_branch_disable(&__sched_core_enabled);
400 void sched_core_get(void)
402 if (atomic_inc_not_zero(&sched_core_count))
405 mutex_lock(&sched_core_mutex);
406 if (!atomic_read(&sched_core_count))
407 __sched_core_enable();
409 smp_mb__before_atomic();
410 atomic_inc(&sched_core_count);
411 mutex_unlock(&sched_core_mutex);
414 static void __sched_core_put(struct work_struct *work)
416 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
417 __sched_core_disable();
418 mutex_unlock(&sched_core_mutex);
422 void sched_core_put(void)
424 static DECLARE_WORK(_work, __sched_core_put);
427 * "There can be only one"
429 * Either this is the last one, or we don't actually need to do any
430 * 'work'. If it is the last *again*, we rely on
431 * WORK_STRUCT_PENDING_BIT.
433 if (!atomic_add_unless(&sched_core_count, -1, 1))
434 schedule_work(&_work);
437 #else /* !CONFIG_SCHED_CORE */
439 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
441 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
443 #endif /* CONFIG_SCHED_CORE */
446 * Serialization rules:
452 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
455 * rq2->lock where: rq1 < rq2
459 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
460 * local CPU's rq->lock, it optionally removes the task from the runqueue and
461 * always looks at the local rq data structures to find the most eligible task
464 * Task enqueue is also under rq->lock, possibly taken from another CPU.
465 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
466 * the local CPU to avoid bouncing the runqueue state around [ see
467 * ttwu_queue_wakelist() ]
469 * Task wakeup, specifically wakeups that involve migration, are horribly
470 * complicated to avoid having to take two rq->locks.
474 * System-calls and anything external will use task_rq_lock() which acquires
475 * both p->pi_lock and rq->lock. As a consequence the state they change is
476 * stable while holding either lock:
478 * - sched_setaffinity()/
479 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
480 * - set_user_nice(): p->se.load, p->*prio
481 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
482 * p->se.load, p->rt_priority,
483 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
484 * - sched_setnuma(): p->numa_preferred_nid
485 * - sched_move_task()/
486 * cpu_cgroup_fork(): p->sched_task_group
487 * - uclamp_update_active() p->uclamp*
489 * p->state <- TASK_*:
491 * is changed locklessly using set_current_state(), __set_current_state() or
492 * set_special_state(), see their respective comments, or by
493 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
496 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
498 * is set by activate_task() and cleared by deactivate_task(), under
499 * rq->lock. Non-zero indicates the task is runnable, the special
500 * ON_RQ_MIGRATING state is used for migration without holding both
501 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
503 * p->on_cpu <- { 0, 1 }:
505 * is set by prepare_task() and cleared by finish_task() such that it will be
506 * set before p is scheduled-in and cleared after p is scheduled-out, both
507 * under rq->lock. Non-zero indicates the task is running on its CPU.
509 * [ The astute reader will observe that it is possible for two tasks on one
510 * CPU to have ->on_cpu = 1 at the same time. ]
512 * task_cpu(p): is changed by set_task_cpu(), the rules are:
514 * - Don't call set_task_cpu() on a blocked task:
516 * We don't care what CPU we're not running on, this simplifies hotplug,
517 * the CPU assignment of blocked tasks isn't required to be valid.
519 * - for try_to_wake_up(), called under p->pi_lock:
521 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
523 * - for migration called under rq->lock:
524 * [ see task_on_rq_migrating() in task_rq_lock() ]
526 * o move_queued_task()
529 * - for migration called under double_rq_lock():
531 * o __migrate_swap_task()
532 * o push_rt_task() / pull_rt_task()
533 * o push_dl_task() / pull_dl_task()
534 * o dl_task_offline_migration()
538 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
540 raw_spinlock_t *lock;
542 /* Matches synchronize_rcu() in __sched_core_enable() */
544 if (sched_core_disabled()) {
545 raw_spin_lock_nested(&rq->__lock, subclass);
546 /* preempt_count *MUST* be > 1 */
547 preempt_enable_no_resched();
552 lock = __rq_lockp(rq);
553 raw_spin_lock_nested(lock, subclass);
554 if (likely(lock == __rq_lockp(rq))) {
555 /* preempt_count *MUST* be > 1 */
556 preempt_enable_no_resched();
559 raw_spin_unlock(lock);
563 bool raw_spin_rq_trylock(struct rq *rq)
565 raw_spinlock_t *lock;
568 /* Matches synchronize_rcu() in __sched_core_enable() */
570 if (sched_core_disabled()) {
571 ret = raw_spin_trylock(&rq->__lock);
577 lock = __rq_lockp(rq);
578 ret = raw_spin_trylock(lock);
579 if (!ret || (likely(lock == __rq_lockp(rq)))) {
583 raw_spin_unlock(lock);
587 void raw_spin_rq_unlock(struct rq *rq)
589 raw_spin_unlock(rq_lockp(rq));
594 * double_rq_lock - safely lock two runqueues
596 void double_rq_lock(struct rq *rq1, struct rq *rq2)
598 lockdep_assert_irqs_disabled();
600 if (rq_order_less(rq2, rq1))
603 raw_spin_rq_lock(rq1);
604 if (__rq_lockp(rq1) != __rq_lockp(rq2))
605 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
607 double_rq_clock_clear_update(rq1, rq2);
612 * __task_rq_lock - lock the rq @p resides on.
614 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
619 lockdep_assert_held(&p->pi_lock);
623 raw_spin_rq_lock(rq);
624 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
628 raw_spin_rq_unlock(rq);
630 while (unlikely(task_on_rq_migrating(p)))
636 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
638 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
639 __acquires(p->pi_lock)
645 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
647 raw_spin_rq_lock(rq);
649 * move_queued_task() task_rq_lock()
652 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
653 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
654 * [S] ->cpu = new_cpu [L] task_rq()
658 * If we observe the old CPU in task_rq_lock(), the acquire of
659 * the old rq->lock will fully serialize against the stores.
661 * If we observe the new CPU in task_rq_lock(), the address
662 * dependency headed by '[L] rq = task_rq()' and the acquire
663 * will pair with the WMB to ensure we then also see migrating.
665 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
669 raw_spin_rq_unlock(rq);
670 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
672 while (unlikely(task_on_rq_migrating(p)))
678 * RQ-clock updating methods:
681 static void update_rq_clock_task(struct rq *rq, s64 delta)
684 * In theory, the compile should just see 0 here, and optimize out the call
685 * to sched_rt_avg_update. But I don't trust it...
687 s64 __maybe_unused steal = 0, irq_delta = 0;
689 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
690 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
693 * Since irq_time is only updated on {soft,}irq_exit, we might run into
694 * this case when a previous update_rq_clock() happened inside a
697 * When this happens, we stop ->clock_task and only update the
698 * prev_irq_time stamp to account for the part that fit, so that a next
699 * update will consume the rest. This ensures ->clock_task is
702 * It does however cause some slight miss-attribution of {soft,}irq
703 * time, a more accurate solution would be to update the irq_time using
704 * the current rq->clock timestamp, except that would require using
707 if (irq_delta > delta)
710 rq->prev_irq_time += irq_delta;
713 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
714 if (static_key_false((¶virt_steal_rq_enabled))) {
715 steal = paravirt_steal_clock(cpu_of(rq));
716 steal -= rq->prev_steal_time_rq;
718 if (unlikely(steal > delta))
721 rq->prev_steal_time_rq += steal;
726 rq->clock_task += delta;
728 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
729 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
730 update_irq_load_avg(rq, irq_delta + steal);
732 update_rq_clock_pelt(rq, delta);
735 void update_rq_clock(struct rq *rq)
739 lockdep_assert_rq_held(rq);
741 if (rq->clock_update_flags & RQCF_ACT_SKIP)
744 #ifdef CONFIG_SCHED_DEBUG
745 if (sched_feat(WARN_DOUBLE_CLOCK))
746 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
747 rq->clock_update_flags |= RQCF_UPDATED;
750 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
754 update_rq_clock_task(rq, delta);
757 #ifdef CONFIG_SCHED_HRTICK
759 * Use HR-timers to deliver accurate preemption points.
762 static void hrtick_clear(struct rq *rq)
764 if (hrtimer_active(&rq->hrtick_timer))
765 hrtimer_cancel(&rq->hrtick_timer);
769 * High-resolution timer tick.
770 * Runs from hardirq context with interrupts disabled.
772 static enum hrtimer_restart hrtick(struct hrtimer *timer)
774 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
777 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
781 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
784 return HRTIMER_NORESTART;
789 static void __hrtick_restart(struct rq *rq)
791 struct hrtimer *timer = &rq->hrtick_timer;
792 ktime_t time = rq->hrtick_time;
794 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
798 * called from hardirq (IPI) context
800 static void __hrtick_start(void *arg)
806 __hrtick_restart(rq);
811 * Called to set the hrtick timer state.
813 * called with rq->lock held and irqs disabled
815 void hrtick_start(struct rq *rq, u64 delay)
817 struct hrtimer *timer = &rq->hrtick_timer;
821 * Don't schedule slices shorter than 10000ns, that just
822 * doesn't make sense and can cause timer DoS.
824 delta = max_t(s64, delay, 10000LL);
825 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
828 __hrtick_restart(rq);
830 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
835 * Called to set the hrtick timer state.
837 * called with rq->lock held and irqs disabled
839 void hrtick_start(struct rq *rq, u64 delay)
842 * Don't schedule slices shorter than 10000ns, that just
843 * doesn't make sense. Rely on vruntime for fairness.
845 delay = max_t(u64, delay, 10000LL);
846 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
847 HRTIMER_MODE_REL_PINNED_HARD);
850 #endif /* CONFIG_SMP */
852 static void hrtick_rq_init(struct rq *rq)
855 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
857 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
858 rq->hrtick_timer.function = hrtick;
860 #else /* CONFIG_SCHED_HRTICK */
861 static inline void hrtick_clear(struct rq *rq)
865 static inline void hrtick_rq_init(struct rq *rq)
868 #endif /* CONFIG_SCHED_HRTICK */
871 * cmpxchg based fetch_or, macro so it works for different integer types
873 #define fetch_or(ptr, mask) \
875 typeof(ptr) _ptr = (ptr); \
876 typeof(mask) _mask = (mask); \
877 typeof(*_ptr) _val = *_ptr; \
880 } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
884 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
886 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
887 * this avoids any races wrt polling state changes and thereby avoids
890 static inline bool set_nr_and_not_polling(struct task_struct *p)
892 struct thread_info *ti = task_thread_info(p);
893 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
897 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
899 * If this returns true, then the idle task promises to call
900 * sched_ttwu_pending() and reschedule soon.
902 static bool set_nr_if_polling(struct task_struct *p)
904 struct thread_info *ti = task_thread_info(p);
905 typeof(ti->flags) val = READ_ONCE(ti->flags);
908 if (!(val & _TIF_POLLING_NRFLAG))
910 if (val & _TIF_NEED_RESCHED)
912 if (try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED))
919 static inline bool set_nr_and_not_polling(struct task_struct *p)
921 set_tsk_need_resched(p);
926 static inline bool set_nr_if_polling(struct task_struct *p)
933 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
935 struct wake_q_node *node = &task->wake_q;
938 * Atomically grab the task, if ->wake_q is !nil already it means
939 * it's already queued (either by us or someone else) and will get the
940 * wakeup due to that.
942 * In order to ensure that a pending wakeup will observe our pending
943 * state, even in the failed case, an explicit smp_mb() must be used.
945 smp_mb__before_atomic();
946 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
950 * The head is context local, there can be no concurrency.
953 head->lastp = &node->next;
958 * wake_q_add() - queue a wakeup for 'later' waking.
959 * @head: the wake_q_head to add @task to
960 * @task: the task to queue for 'later' wakeup
962 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
963 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
966 * This function must be used as-if it were wake_up_process(); IOW the task
967 * must be ready to be woken at this location.
969 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
971 if (__wake_q_add(head, task))
972 get_task_struct(task);
976 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
977 * @head: the wake_q_head to add @task to
978 * @task: the task to queue for 'later' wakeup
980 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
981 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
984 * This function must be used as-if it were wake_up_process(); IOW the task
985 * must be ready to be woken at this location.
987 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
988 * that already hold reference to @task can call the 'safe' version and trust
989 * wake_q to do the right thing depending whether or not the @task is already
992 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
994 if (!__wake_q_add(head, task))
995 put_task_struct(task);
998 void wake_up_q(struct wake_q_head *head)
1000 struct wake_q_node *node = head->first;
1002 while (node != WAKE_Q_TAIL) {
1003 struct task_struct *task;
1005 task = container_of(node, struct task_struct, wake_q);
1006 /* Task can safely be re-inserted now: */
1008 task->wake_q.next = NULL;
1011 * wake_up_process() executes a full barrier, which pairs with
1012 * the queueing in wake_q_add() so as not to miss wakeups.
1014 wake_up_process(task);
1015 put_task_struct(task);
1020 * resched_curr - mark rq's current task 'to be rescheduled now'.
1022 * On UP this means the setting of the need_resched flag, on SMP it
1023 * might also involve a cross-CPU call to trigger the scheduler on
1026 void resched_curr(struct rq *rq)
1028 struct task_struct *curr = rq->curr;
1031 lockdep_assert_rq_held(rq);
1033 if (test_tsk_need_resched(curr))
1038 if (cpu == smp_processor_id()) {
1039 set_tsk_need_resched(curr);
1040 set_preempt_need_resched();
1044 if (set_nr_and_not_polling(curr))
1045 smp_send_reschedule(cpu);
1047 trace_sched_wake_idle_without_ipi(cpu);
1050 void resched_cpu(int cpu)
1052 struct rq *rq = cpu_rq(cpu);
1053 unsigned long flags;
1055 raw_spin_rq_lock_irqsave(rq, flags);
1056 if (cpu_online(cpu) || cpu == smp_processor_id())
1058 raw_spin_rq_unlock_irqrestore(rq, flags);
1062 #ifdef CONFIG_NO_HZ_COMMON
1064 * In the semi idle case, use the nearest busy CPU for migrating timers
1065 * from an idle CPU. This is good for power-savings.
1067 * We don't do similar optimization for completely idle system, as
1068 * selecting an idle CPU will add more delays to the timers than intended
1069 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1071 int get_nohz_timer_target(void)
1073 int i, cpu = smp_processor_id(), default_cpu = -1;
1074 struct sched_domain *sd;
1075 const struct cpumask *hk_mask;
1077 if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
1083 hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
1086 for_each_domain(cpu, sd) {
1087 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1098 if (default_cpu == -1)
1099 default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
1107 * When add_timer_on() enqueues a timer into the timer wheel of an
1108 * idle CPU then this timer might expire before the next timer event
1109 * which is scheduled to wake up that CPU. In case of a completely
1110 * idle system the next event might even be infinite time into the
1111 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1112 * leaves the inner idle loop so the newly added timer is taken into
1113 * account when the CPU goes back to idle and evaluates the timer
1114 * wheel for the next timer event.
1116 static void wake_up_idle_cpu(int cpu)
1118 struct rq *rq = cpu_rq(cpu);
1120 if (cpu == smp_processor_id())
1123 if (set_nr_and_not_polling(rq->idle))
1124 smp_send_reschedule(cpu);
1126 trace_sched_wake_idle_without_ipi(cpu);
1129 static bool wake_up_full_nohz_cpu(int cpu)
1132 * We just need the target to call irq_exit() and re-evaluate
1133 * the next tick. The nohz full kick at least implies that.
1134 * If needed we can still optimize that later with an
1137 if (cpu_is_offline(cpu))
1138 return true; /* Don't try to wake offline CPUs. */
1139 if (tick_nohz_full_cpu(cpu)) {
1140 if (cpu != smp_processor_id() ||
1141 tick_nohz_tick_stopped())
1142 tick_nohz_full_kick_cpu(cpu);
1150 * Wake up the specified CPU. If the CPU is going offline, it is the
1151 * caller's responsibility to deal with the lost wakeup, for example,
1152 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1154 void wake_up_nohz_cpu(int cpu)
1156 if (!wake_up_full_nohz_cpu(cpu))
1157 wake_up_idle_cpu(cpu);
1160 static void nohz_csd_func(void *info)
1162 struct rq *rq = info;
1163 int cpu = cpu_of(rq);
1167 * Release the rq::nohz_csd.
1169 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1170 WARN_ON(!(flags & NOHZ_KICK_MASK));
1172 rq->idle_balance = idle_cpu(cpu);
1173 if (rq->idle_balance && !need_resched()) {
1174 rq->nohz_idle_balance = flags;
1175 raise_softirq_irqoff(SCHED_SOFTIRQ);
1179 #endif /* CONFIG_NO_HZ_COMMON */
1181 #ifdef CONFIG_NO_HZ_FULL
1182 bool sched_can_stop_tick(struct rq *rq)
1184 int fifo_nr_running;
1186 /* Deadline tasks, even if single, need the tick */
1187 if (rq->dl.dl_nr_running)
1191 * If there are more than one RR tasks, we need the tick to affect the
1192 * actual RR behaviour.
1194 if (rq->rt.rr_nr_running) {
1195 if (rq->rt.rr_nr_running == 1)
1202 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1203 * forced preemption between FIFO tasks.
1205 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1206 if (fifo_nr_running)
1210 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1211 * if there's more than one we need the tick for involuntary
1214 if (rq->nr_running > 1)
1219 #endif /* CONFIG_NO_HZ_FULL */
1220 #endif /* CONFIG_SMP */
1222 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1223 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1225 * Iterate task_group tree rooted at *from, calling @down when first entering a
1226 * node and @up when leaving it for the final time.
1228 * Caller must hold rcu_lock or sufficient equivalent.
1230 int walk_tg_tree_from(struct task_group *from,
1231 tg_visitor down, tg_visitor up, void *data)
1233 struct task_group *parent, *child;
1239 ret = (*down)(parent, data);
1242 list_for_each_entry_rcu(child, &parent->children, siblings) {
1249 ret = (*up)(parent, data);
1250 if (ret || parent == from)
1254 parent = parent->parent;
1261 int tg_nop(struct task_group *tg, void *data)
1267 static void set_load_weight(struct task_struct *p, bool update_load)
1269 int prio = p->static_prio - MAX_RT_PRIO;
1270 struct load_weight *load = &p->se.load;
1273 * SCHED_IDLE tasks get minimal weight:
1275 if (task_has_idle_policy(p)) {
1276 load->weight = scale_load(WEIGHT_IDLEPRIO);
1277 load->inv_weight = WMULT_IDLEPRIO;
1282 * SCHED_OTHER tasks have to update their load when changing their
1285 if (update_load && p->sched_class == &fair_sched_class) {
1286 reweight_task(p, prio);
1288 load->weight = scale_load(sched_prio_to_weight[prio]);
1289 load->inv_weight = sched_prio_to_wmult[prio];
1293 #ifdef CONFIG_UCLAMP_TASK
1295 * Serializes updates of utilization clamp values
1297 * The (slow-path) user-space triggers utilization clamp value updates which
1298 * can require updates on (fast-path) scheduler's data structures used to
1299 * support enqueue/dequeue operations.
1300 * While the per-CPU rq lock protects fast-path update operations, user-space
1301 * requests are serialized using a mutex to reduce the risk of conflicting
1302 * updates or API abuses.
1304 static DEFINE_MUTEX(uclamp_mutex);
1306 /* Max allowed minimum utilization */
1307 static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1309 /* Max allowed maximum utilization */
1310 static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1313 * By default RT tasks run at the maximum performance point/capacity of the
1314 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1315 * SCHED_CAPACITY_SCALE.
1317 * This knob allows admins to change the default behavior when uclamp is being
1318 * used. In battery powered devices, particularly, running at the maximum
1319 * capacity and frequency will increase energy consumption and shorten the
1322 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1324 * This knob will not override the system default sched_util_clamp_min defined
1327 static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1329 /* All clamps are required to be less or equal than these values */
1330 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1333 * This static key is used to reduce the uclamp overhead in the fast path. It
1334 * primarily disables the call to uclamp_rq_{inc, dec}() in
1335 * enqueue/dequeue_task().
1337 * This allows users to continue to enable uclamp in their kernel config with
1338 * minimum uclamp overhead in the fast path.
1340 * As soon as userspace modifies any of the uclamp knobs, the static key is
1341 * enabled, since we have an actual users that make use of uclamp
1344 * The knobs that would enable this static key are:
1346 * * A task modifying its uclamp value with sched_setattr().
1347 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1348 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1350 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1352 /* Integer rounded range for each bucket */
1353 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1355 #define for_each_clamp_id(clamp_id) \
1356 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1358 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1360 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1363 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1365 if (clamp_id == UCLAMP_MIN)
1367 return SCHED_CAPACITY_SCALE;
1370 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1371 unsigned int value, bool user_defined)
1373 uc_se->value = value;
1374 uc_se->bucket_id = uclamp_bucket_id(value);
1375 uc_se->user_defined = user_defined;
1378 static inline unsigned int
1379 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1380 unsigned int clamp_value)
1383 * Avoid blocked utilization pushing up the frequency when we go
1384 * idle (which drops the max-clamp) by retaining the last known
1387 if (clamp_id == UCLAMP_MAX) {
1388 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1392 return uclamp_none(UCLAMP_MIN);
1395 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1396 unsigned int clamp_value)
1398 /* Reset max-clamp retention only on idle exit */
1399 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1402 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1406 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1407 unsigned int clamp_value)
1409 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1410 int bucket_id = UCLAMP_BUCKETS - 1;
1413 * Since both min and max clamps are max aggregated, find the
1414 * top most bucket with tasks in.
1416 for ( ; bucket_id >= 0; bucket_id--) {
1417 if (!bucket[bucket_id].tasks)
1419 return bucket[bucket_id].value;
1422 /* No tasks -- default clamp values */
1423 return uclamp_idle_value(rq, clamp_id, clamp_value);
1426 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1428 unsigned int default_util_min;
1429 struct uclamp_se *uc_se;
1431 lockdep_assert_held(&p->pi_lock);
1433 uc_se = &p->uclamp_req[UCLAMP_MIN];
1435 /* Only sync if user didn't override the default */
1436 if (uc_se->user_defined)
1439 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1440 uclamp_se_set(uc_se, default_util_min, false);
1443 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1451 /* Protect updates to p->uclamp_* */
1452 rq = task_rq_lock(p, &rf);
1453 __uclamp_update_util_min_rt_default(p);
1454 task_rq_unlock(rq, p, &rf);
1457 static inline struct uclamp_se
1458 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1460 /* Copy by value as we could modify it */
1461 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1462 #ifdef CONFIG_UCLAMP_TASK_GROUP
1463 unsigned int tg_min, tg_max, value;
1466 * Tasks in autogroups or root task group will be
1467 * restricted by system defaults.
1469 if (task_group_is_autogroup(task_group(p)))
1471 if (task_group(p) == &root_task_group)
1474 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1475 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1476 value = uc_req.value;
1477 value = clamp(value, tg_min, tg_max);
1478 uclamp_se_set(&uc_req, value, false);
1485 * The effective clamp bucket index of a task depends on, by increasing
1487 * - the task specific clamp value, when explicitly requested from userspace
1488 * - the task group effective clamp value, for tasks not either in the root
1489 * group or in an autogroup
1490 * - the system default clamp value, defined by the sysadmin
1492 static inline struct uclamp_se
1493 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1495 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1496 struct uclamp_se uc_max = uclamp_default[clamp_id];
1498 /* System default restrictions always apply */
1499 if (unlikely(uc_req.value > uc_max.value))
1505 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1507 struct uclamp_se uc_eff;
1509 /* Task currently refcounted: use back-annotated (effective) value */
1510 if (p->uclamp[clamp_id].active)
1511 return (unsigned long)p->uclamp[clamp_id].value;
1513 uc_eff = uclamp_eff_get(p, clamp_id);
1515 return (unsigned long)uc_eff.value;
1519 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1520 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1521 * updates the rq's clamp value if required.
1523 * Tasks can have a task-specific value requested from user-space, track
1524 * within each bucket the maximum value for tasks refcounted in it.
1525 * This "local max aggregation" allows to track the exact "requested" value
1526 * for each bucket when all its RUNNABLE tasks require the same clamp.
1528 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1529 enum uclamp_id clamp_id)
1531 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1532 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1533 struct uclamp_bucket *bucket;
1535 lockdep_assert_rq_held(rq);
1537 /* Update task effective clamp */
1538 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1540 bucket = &uc_rq->bucket[uc_se->bucket_id];
1542 uc_se->active = true;
1544 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1547 * Local max aggregation: rq buckets always track the max
1548 * "requested" clamp value of its RUNNABLE tasks.
1550 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1551 bucket->value = uc_se->value;
1553 if (uc_se->value > READ_ONCE(uc_rq->value))
1554 WRITE_ONCE(uc_rq->value, uc_se->value);
1558 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1559 * is released. If this is the last task reference counting the rq's max
1560 * active clamp value, then the rq's clamp value is updated.
1562 * Both refcounted tasks and rq's cached clamp values are expected to be
1563 * always valid. If it's detected they are not, as defensive programming,
1564 * enforce the expected state and warn.
1566 static inline void uclamp_rq_dec_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;
1572 unsigned int bkt_clamp;
1573 unsigned int rq_clamp;
1575 lockdep_assert_rq_held(rq);
1578 * If sched_uclamp_used was enabled after task @p was enqueued,
1579 * we could end up with unbalanced call to uclamp_rq_dec_id().
1581 * In this case the uc_se->active flag should be false since no uclamp
1582 * accounting was performed at enqueue time and we can just return
1585 * Need to be careful of the following enqueue/dequeue ordering
1589 * // sched_uclamp_used gets enabled
1592 * // Must not decrement bucket->tasks here
1595 * where we could end up with stale data in uc_se and
1596 * bucket[uc_se->bucket_id].
1598 * The following check here eliminates the possibility of such race.
1600 if (unlikely(!uc_se->active))
1603 bucket = &uc_rq->bucket[uc_se->bucket_id];
1605 SCHED_WARN_ON(!bucket->tasks);
1606 if (likely(bucket->tasks))
1609 uc_se->active = false;
1612 * Keep "local max aggregation" simple and accept to (possibly)
1613 * overboost some RUNNABLE tasks in the same bucket.
1614 * The rq clamp bucket value is reset to its base value whenever
1615 * there are no more RUNNABLE tasks refcounting it.
1617 if (likely(bucket->tasks))
1620 rq_clamp = READ_ONCE(uc_rq->value);
1622 * Defensive programming: this should never happen. If it happens,
1623 * e.g. due to future modification, warn and fixup the expected value.
1625 SCHED_WARN_ON(bucket->value > rq_clamp);
1626 if (bucket->value >= rq_clamp) {
1627 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1628 WRITE_ONCE(uc_rq->value, bkt_clamp);
1632 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1634 enum uclamp_id clamp_id;
1637 * Avoid any overhead until uclamp is actually used by the userspace.
1639 * The condition is constructed such that a NOP is generated when
1640 * sched_uclamp_used is disabled.
1642 if (!static_branch_unlikely(&sched_uclamp_used))
1645 if (unlikely(!p->sched_class->uclamp_enabled))
1648 for_each_clamp_id(clamp_id)
1649 uclamp_rq_inc_id(rq, p, clamp_id);
1651 /* Reset clamp idle holding when there is one RUNNABLE task */
1652 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1653 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1656 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1658 enum uclamp_id clamp_id;
1661 * Avoid any overhead until uclamp is actually used by the userspace.
1663 * The condition is constructed such that a NOP is generated when
1664 * sched_uclamp_used is disabled.
1666 if (!static_branch_unlikely(&sched_uclamp_used))
1669 if (unlikely(!p->sched_class->uclamp_enabled))
1672 for_each_clamp_id(clamp_id)
1673 uclamp_rq_dec_id(rq, p, clamp_id);
1676 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1677 enum uclamp_id clamp_id)
1679 if (!p->uclamp[clamp_id].active)
1682 uclamp_rq_dec_id(rq, p, clamp_id);
1683 uclamp_rq_inc_id(rq, p, clamp_id);
1686 * Make sure to clear the idle flag if we've transiently reached 0
1687 * active tasks on rq.
1689 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1690 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1694 uclamp_update_active(struct task_struct *p)
1696 enum uclamp_id clamp_id;
1701 * Lock the task and the rq where the task is (or was) queued.
1703 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1704 * price to pay to safely serialize util_{min,max} updates with
1705 * enqueues, dequeues and migration operations.
1706 * This is the same locking schema used by __set_cpus_allowed_ptr().
1708 rq = task_rq_lock(p, &rf);
1711 * Setting the clamp bucket is serialized by task_rq_lock().
1712 * If the task is not yet RUNNABLE and its task_struct is not
1713 * affecting a valid clamp bucket, the next time it's enqueued,
1714 * it will already see the updated clamp bucket value.
1716 for_each_clamp_id(clamp_id)
1717 uclamp_rq_reinc_id(rq, p, clamp_id);
1719 task_rq_unlock(rq, p, &rf);
1722 #ifdef CONFIG_UCLAMP_TASK_GROUP
1724 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1726 struct css_task_iter it;
1727 struct task_struct *p;
1729 css_task_iter_start(css, 0, &it);
1730 while ((p = css_task_iter_next(&it)))
1731 uclamp_update_active(p);
1732 css_task_iter_end(&it);
1735 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1738 #ifdef CONFIG_SYSCTL
1739 #ifdef CONFIG_UCLAMP_TASK
1740 #ifdef CONFIG_UCLAMP_TASK_GROUP
1741 static void uclamp_update_root_tg(void)
1743 struct task_group *tg = &root_task_group;
1745 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1746 sysctl_sched_uclamp_util_min, false);
1747 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1748 sysctl_sched_uclamp_util_max, false);
1751 cpu_util_update_eff(&root_task_group.css);
1755 static void uclamp_update_root_tg(void) { }
1758 static void uclamp_sync_util_min_rt_default(void)
1760 struct task_struct *g, *p;
1763 * copy_process() sysctl_uclamp
1764 * uclamp_min_rt = X;
1765 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1766 * // link thread smp_mb__after_spinlock()
1767 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1768 * sched_post_fork() for_each_process_thread()
1769 * __uclamp_sync_rt() __uclamp_sync_rt()
1771 * Ensures that either sched_post_fork() will observe the new
1772 * uclamp_min_rt or for_each_process_thread() will observe the new
1775 read_lock(&tasklist_lock);
1776 smp_mb__after_spinlock();
1777 read_unlock(&tasklist_lock);
1780 for_each_process_thread(g, p)
1781 uclamp_update_util_min_rt_default(p);
1785 static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1786 void *buffer, size_t *lenp, loff_t *ppos)
1788 bool update_root_tg = false;
1789 int old_min, old_max, old_min_rt;
1792 mutex_lock(&uclamp_mutex);
1793 old_min = sysctl_sched_uclamp_util_min;
1794 old_max = sysctl_sched_uclamp_util_max;
1795 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1797 result = proc_dointvec(table, write, buffer, lenp, ppos);
1803 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1804 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1805 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1811 if (old_min != sysctl_sched_uclamp_util_min) {
1812 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1813 sysctl_sched_uclamp_util_min, false);
1814 update_root_tg = true;
1816 if (old_max != sysctl_sched_uclamp_util_max) {
1817 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1818 sysctl_sched_uclamp_util_max, false);
1819 update_root_tg = true;
1822 if (update_root_tg) {
1823 static_branch_enable(&sched_uclamp_used);
1824 uclamp_update_root_tg();
1827 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1828 static_branch_enable(&sched_uclamp_used);
1829 uclamp_sync_util_min_rt_default();
1833 * We update all RUNNABLE tasks only when task groups are in use.
1834 * Otherwise, keep it simple and do just a lazy update at each next
1835 * task enqueue time.
1841 sysctl_sched_uclamp_util_min = old_min;
1842 sysctl_sched_uclamp_util_max = old_max;
1843 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1845 mutex_unlock(&uclamp_mutex);
1852 static int uclamp_validate(struct task_struct *p,
1853 const struct sched_attr *attr)
1855 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1856 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1858 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1859 util_min = attr->sched_util_min;
1861 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1865 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1866 util_max = attr->sched_util_max;
1868 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1872 if (util_min != -1 && util_max != -1 && util_min > util_max)
1876 * We have valid uclamp attributes; make sure uclamp is enabled.
1878 * We need to do that here, because enabling static branches is a
1879 * blocking operation which obviously cannot be done while holding
1882 static_branch_enable(&sched_uclamp_used);
1887 static bool uclamp_reset(const struct sched_attr *attr,
1888 enum uclamp_id clamp_id,
1889 struct uclamp_se *uc_se)
1891 /* Reset on sched class change for a non user-defined clamp value. */
1892 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1893 !uc_se->user_defined)
1896 /* Reset on sched_util_{min,max} == -1. */
1897 if (clamp_id == UCLAMP_MIN &&
1898 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1899 attr->sched_util_min == -1) {
1903 if (clamp_id == UCLAMP_MAX &&
1904 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1905 attr->sched_util_max == -1) {
1912 static void __setscheduler_uclamp(struct task_struct *p,
1913 const struct sched_attr *attr)
1915 enum uclamp_id clamp_id;
1917 for_each_clamp_id(clamp_id) {
1918 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1921 if (!uclamp_reset(attr, clamp_id, uc_se))
1925 * RT by default have a 100% boost value that could be modified
1928 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1929 value = sysctl_sched_uclamp_util_min_rt_default;
1931 value = uclamp_none(clamp_id);
1933 uclamp_se_set(uc_se, value, false);
1937 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1940 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1941 attr->sched_util_min != -1) {
1942 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1943 attr->sched_util_min, true);
1946 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1947 attr->sched_util_max != -1) {
1948 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1949 attr->sched_util_max, true);
1953 static void uclamp_fork(struct task_struct *p)
1955 enum uclamp_id clamp_id;
1958 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1959 * as the task is still at its early fork stages.
1961 for_each_clamp_id(clamp_id)
1962 p->uclamp[clamp_id].active = false;
1964 if (likely(!p->sched_reset_on_fork))
1967 for_each_clamp_id(clamp_id) {
1968 uclamp_se_set(&p->uclamp_req[clamp_id],
1969 uclamp_none(clamp_id), false);
1973 static void uclamp_post_fork(struct task_struct *p)
1975 uclamp_update_util_min_rt_default(p);
1978 static void __init init_uclamp_rq(struct rq *rq)
1980 enum uclamp_id clamp_id;
1981 struct uclamp_rq *uc_rq = rq->uclamp;
1983 for_each_clamp_id(clamp_id) {
1984 uc_rq[clamp_id] = (struct uclamp_rq) {
1985 .value = uclamp_none(clamp_id)
1989 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1992 static void __init init_uclamp(void)
1994 struct uclamp_se uc_max = {};
1995 enum uclamp_id clamp_id;
1998 for_each_possible_cpu(cpu)
1999 init_uclamp_rq(cpu_rq(cpu));
2001 for_each_clamp_id(clamp_id) {
2002 uclamp_se_set(&init_task.uclamp_req[clamp_id],
2003 uclamp_none(clamp_id), false);
2006 /* System defaults allow max clamp values for both indexes */
2007 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
2008 for_each_clamp_id(clamp_id) {
2009 uclamp_default[clamp_id] = uc_max;
2010 #ifdef CONFIG_UCLAMP_TASK_GROUP
2011 root_task_group.uclamp_req[clamp_id] = uc_max;
2012 root_task_group.uclamp[clamp_id] = uc_max;
2017 #else /* CONFIG_UCLAMP_TASK */
2018 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2019 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
2020 static inline int uclamp_validate(struct task_struct *p,
2021 const struct sched_attr *attr)
2025 static void __setscheduler_uclamp(struct task_struct *p,
2026 const struct sched_attr *attr) { }
2027 static inline void uclamp_fork(struct task_struct *p) { }
2028 static inline void uclamp_post_fork(struct task_struct *p) { }
2029 static inline void init_uclamp(void) { }
2030 #endif /* CONFIG_UCLAMP_TASK */
2032 bool sched_task_on_rq(struct task_struct *p)
2034 return task_on_rq_queued(p);
2037 unsigned long get_wchan(struct task_struct *p)
2039 unsigned long ip = 0;
2042 if (!p || p == current)
2045 /* Only get wchan if task is blocked and we can keep it that way. */
2046 raw_spin_lock_irq(&p->pi_lock);
2047 state = READ_ONCE(p->__state);
2048 smp_rmb(); /* see try_to_wake_up() */
2049 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2050 ip = __get_wchan(p);
2051 raw_spin_unlock_irq(&p->pi_lock);
2056 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2058 if (!(flags & ENQUEUE_NOCLOCK))
2059 update_rq_clock(rq);
2061 if (!(flags & ENQUEUE_RESTORE)) {
2062 sched_info_enqueue(rq, p);
2063 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
2066 uclamp_rq_inc(rq, p);
2067 p->sched_class->enqueue_task(rq, p, flags);
2069 if (sched_core_enabled(rq))
2070 sched_core_enqueue(rq, p);
2073 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2075 if (sched_core_enabled(rq))
2076 sched_core_dequeue(rq, p, flags);
2078 if (!(flags & DEQUEUE_NOCLOCK))
2079 update_rq_clock(rq);
2081 if (!(flags & DEQUEUE_SAVE)) {
2082 sched_info_dequeue(rq, p);
2083 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2086 uclamp_rq_dec(rq, p);
2087 p->sched_class->dequeue_task(rq, p, flags);
2090 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2092 enqueue_task(rq, p, flags);
2094 p->on_rq = TASK_ON_RQ_QUEUED;
2097 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2099 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2101 dequeue_task(rq, p, flags);
2104 static inline int __normal_prio(int policy, int rt_prio, int nice)
2108 if (dl_policy(policy))
2109 prio = MAX_DL_PRIO - 1;
2110 else if (rt_policy(policy))
2111 prio = MAX_RT_PRIO - 1 - rt_prio;
2113 prio = NICE_TO_PRIO(nice);
2119 * Calculate the expected normal priority: i.e. priority
2120 * without taking RT-inheritance into account. Might be
2121 * boosted by interactivity modifiers. Changes upon fork,
2122 * setprio syscalls, and whenever the interactivity
2123 * estimator recalculates.
2125 static inline int normal_prio(struct task_struct *p)
2127 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2131 * Calculate the current priority, i.e. the priority
2132 * taken into account by the scheduler. This value might
2133 * be boosted by RT tasks, or might be boosted by
2134 * interactivity modifiers. Will be RT if the task got
2135 * RT-boosted. If not then it returns p->normal_prio.
2137 static int effective_prio(struct task_struct *p)
2139 p->normal_prio = normal_prio(p);
2141 * If we are RT tasks or we were boosted to RT priority,
2142 * keep the priority unchanged. Otherwise, update priority
2143 * to the normal priority:
2145 if (!rt_prio(p->prio))
2146 return p->normal_prio;
2151 * task_curr - is this task currently executing on a CPU?
2152 * @p: the task in question.
2154 * Return: 1 if the task is currently executing. 0 otherwise.
2156 inline int task_curr(const struct task_struct *p)
2158 return cpu_curr(task_cpu(p)) == p;
2162 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2163 * use the balance_callback list if you want balancing.
2165 * this means any call to check_class_changed() must be followed by a call to
2166 * balance_callback().
2168 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2169 const struct sched_class *prev_class,
2172 if (prev_class != p->sched_class) {
2173 if (prev_class->switched_from)
2174 prev_class->switched_from(rq, p);
2176 p->sched_class->switched_to(rq, p);
2177 } else if (oldprio != p->prio || dl_task(p))
2178 p->sched_class->prio_changed(rq, p, oldprio);
2181 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2183 if (p->sched_class == rq->curr->sched_class)
2184 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2185 else if (sched_class_above(p->sched_class, rq->curr->sched_class))
2189 * A queue event has occurred, and we're going to schedule. In
2190 * this case, we can save a useless back to back clock update.
2192 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2193 rq_clock_skip_update(rq);
2199 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2201 static int __set_cpus_allowed_ptr(struct task_struct *p,
2202 const struct cpumask *new_mask,
2205 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2207 if (likely(!p->migration_disabled))
2210 if (p->cpus_ptr != &p->cpus_mask)
2214 * Violates locking rules! see comment in __do_set_cpus_allowed().
2216 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2219 void migrate_disable(void)
2221 struct task_struct *p = current;
2223 if (p->migration_disabled) {
2224 p->migration_disabled++;
2229 this_rq()->nr_pinned++;
2230 p->migration_disabled = 1;
2233 EXPORT_SYMBOL_GPL(migrate_disable);
2235 void migrate_enable(void)
2237 struct task_struct *p = current;
2239 if (p->migration_disabled > 1) {
2240 p->migration_disabled--;
2244 if (WARN_ON_ONCE(!p->migration_disabled))
2248 * Ensure stop_task runs either before or after this, and that
2249 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2252 if (p->cpus_ptr != &p->cpus_mask)
2253 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2255 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2256 * regular cpus_mask, otherwise things that race (eg.
2257 * select_fallback_rq) get confused.
2260 p->migration_disabled = 0;
2261 this_rq()->nr_pinned--;
2264 EXPORT_SYMBOL_GPL(migrate_enable);
2266 static inline bool rq_has_pinned_tasks(struct rq *rq)
2268 return rq->nr_pinned;
2272 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2273 * __set_cpus_allowed_ptr() and select_fallback_rq().
2275 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2277 /* When not in the task's cpumask, no point in looking further. */
2278 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2281 /* migrate_disabled() must be allowed to finish. */
2282 if (is_migration_disabled(p))
2283 return cpu_online(cpu);
2285 /* Non kernel threads are not allowed during either online or offline. */
2286 if (!(p->flags & PF_KTHREAD))
2287 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2289 /* KTHREAD_IS_PER_CPU is always allowed. */
2290 if (kthread_is_per_cpu(p))
2291 return cpu_online(cpu);
2293 /* Regular kernel threads don't get to stay during offline. */
2297 /* But are allowed during online. */
2298 return cpu_online(cpu);
2302 * This is how migration works:
2304 * 1) we invoke migration_cpu_stop() on the target CPU using
2306 * 2) stopper starts to run (implicitly forcing the migrated thread
2308 * 3) it checks whether the migrated task is still in the wrong runqueue.
2309 * 4) if it's in the wrong runqueue then the migration thread removes
2310 * it and puts it into the right queue.
2311 * 5) stopper completes and stop_one_cpu() returns and the migration
2316 * move_queued_task - move a queued task to new rq.
2318 * Returns (locked) new rq. Old rq's lock is released.
2320 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2321 struct task_struct *p, int new_cpu)
2323 lockdep_assert_rq_held(rq);
2325 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2326 set_task_cpu(p, new_cpu);
2329 rq = cpu_rq(new_cpu);
2332 BUG_ON(task_cpu(p) != new_cpu);
2333 activate_task(rq, p, 0);
2334 check_preempt_curr(rq, p, 0);
2339 struct migration_arg {
2340 struct task_struct *task;
2342 struct set_affinity_pending *pending;
2346 * @refs: number of wait_for_completion()
2347 * @stop_pending: is @stop_work in use
2349 struct set_affinity_pending {
2351 unsigned int stop_pending;
2352 struct completion done;
2353 struct cpu_stop_work stop_work;
2354 struct migration_arg arg;
2358 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2359 * this because either it can't run here any more (set_cpus_allowed()
2360 * away from this CPU, or CPU going down), or because we're
2361 * attempting to rebalance this task on exec (sched_exec).
2363 * So we race with normal scheduler movements, but that's OK, as long
2364 * as the task is no longer on this CPU.
2366 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2367 struct task_struct *p, int dest_cpu)
2369 /* Affinity changed (again). */
2370 if (!is_cpu_allowed(p, dest_cpu))
2373 update_rq_clock(rq);
2374 rq = move_queued_task(rq, rf, p, dest_cpu);
2380 * migration_cpu_stop - this will be executed by a highprio stopper thread
2381 * and performs thread migration by bumping thread off CPU then
2382 * 'pushing' onto another runqueue.
2384 static int migration_cpu_stop(void *data)
2386 struct migration_arg *arg = data;
2387 struct set_affinity_pending *pending = arg->pending;
2388 struct task_struct *p = arg->task;
2389 struct rq *rq = this_rq();
2390 bool complete = false;
2394 * The original target CPU might have gone down and we might
2395 * be on another CPU but it doesn't matter.
2397 local_irq_save(rf.flags);
2399 * We need to explicitly wake pending tasks before running
2400 * __migrate_task() such that we will not miss enforcing cpus_ptr
2401 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2403 flush_smp_call_function_queue();
2405 raw_spin_lock(&p->pi_lock);
2409 * If we were passed a pending, then ->stop_pending was set, thus
2410 * p->migration_pending must have remained stable.
2412 WARN_ON_ONCE(pending && pending != p->migration_pending);
2415 * If task_rq(p) != rq, it cannot be migrated here, because we're
2416 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2417 * we're holding p->pi_lock.
2419 if (task_rq(p) == rq) {
2420 if (is_migration_disabled(p))
2424 p->migration_pending = NULL;
2427 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2431 if (task_on_rq_queued(p))
2432 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2434 p->wake_cpu = arg->dest_cpu;
2437 * XXX __migrate_task() can fail, at which point we might end
2438 * up running on a dodgy CPU, AFAICT this can only happen
2439 * during CPU hotplug, at which point we'll get pushed out
2440 * anyway, so it's probably not a big deal.
2443 } else if (pending) {
2445 * This happens when we get migrated between migrate_enable()'s
2446 * preempt_enable() and scheduling the stopper task. At that
2447 * point we're a regular task again and not current anymore.
2449 * A !PREEMPT kernel has a giant hole here, which makes it far
2454 * The task moved before the stopper got to run. We're holding
2455 * ->pi_lock, so the allowed mask is stable - if it got
2456 * somewhere allowed, we're done.
2458 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2459 p->migration_pending = NULL;
2465 * When migrate_enable() hits a rq mis-match we can't reliably
2466 * determine is_migration_disabled() and so have to chase after
2469 WARN_ON_ONCE(!pending->stop_pending);
2470 task_rq_unlock(rq, p, &rf);
2471 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2472 &pending->arg, &pending->stop_work);
2477 pending->stop_pending = false;
2478 task_rq_unlock(rq, p, &rf);
2481 complete_all(&pending->done);
2486 int push_cpu_stop(void *arg)
2488 struct rq *lowest_rq = NULL, *rq = this_rq();
2489 struct task_struct *p = arg;
2491 raw_spin_lock_irq(&p->pi_lock);
2492 raw_spin_rq_lock(rq);
2494 if (task_rq(p) != rq)
2497 if (is_migration_disabled(p)) {
2498 p->migration_flags |= MDF_PUSH;
2502 p->migration_flags &= ~MDF_PUSH;
2504 if (p->sched_class->find_lock_rq)
2505 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2510 // XXX validate p is still the highest prio task
2511 if (task_rq(p) == rq) {
2512 deactivate_task(rq, p, 0);
2513 set_task_cpu(p, lowest_rq->cpu);
2514 activate_task(lowest_rq, p, 0);
2515 resched_curr(lowest_rq);
2518 double_unlock_balance(rq, lowest_rq);
2521 rq->push_busy = false;
2522 raw_spin_rq_unlock(rq);
2523 raw_spin_unlock_irq(&p->pi_lock);
2530 * sched_class::set_cpus_allowed must do the below, but is not required to
2531 * actually call this function.
2533 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2535 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2536 p->cpus_ptr = new_mask;
2540 cpumask_copy(&p->cpus_mask, new_mask);
2541 p->nr_cpus_allowed = cpumask_weight(new_mask);
2545 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2547 struct rq *rq = task_rq(p);
2548 bool queued, running;
2551 * This here violates the locking rules for affinity, since we're only
2552 * supposed to change these variables while holding both rq->lock and
2555 * HOWEVER, it magically works, because ttwu() is the only code that
2556 * accesses these variables under p->pi_lock and only does so after
2557 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2558 * before finish_task().
2560 * XXX do further audits, this smells like something putrid.
2562 if (flags & SCA_MIGRATE_DISABLE)
2563 SCHED_WARN_ON(!p->on_cpu);
2565 lockdep_assert_held(&p->pi_lock);
2567 queued = task_on_rq_queued(p);
2568 running = task_current(rq, p);
2572 * Because __kthread_bind() calls this on blocked tasks without
2575 lockdep_assert_rq_held(rq);
2576 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2579 put_prev_task(rq, p);
2581 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2584 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2586 set_next_task(rq, p);
2589 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2591 __do_set_cpus_allowed(p, new_mask, 0);
2594 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2597 if (!src->user_cpus_ptr)
2600 dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2601 if (!dst->user_cpus_ptr)
2604 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2608 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2610 struct cpumask *user_mask = NULL;
2612 swap(p->user_cpus_ptr, user_mask);
2617 void release_user_cpus_ptr(struct task_struct *p)
2619 kfree(clear_user_cpus_ptr(p));
2623 * This function is wildly self concurrent; here be dragons.
2626 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2627 * designated task is enqueued on an allowed CPU. If that task is currently
2628 * running, we have to kick it out using the CPU stopper.
2630 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2633 * Initial conditions: P0->cpus_mask = [0, 1]
2637 * migrate_disable();
2639 * set_cpus_allowed_ptr(P0, [1]);
2641 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2642 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2643 * This means we need the following scheme:
2647 * migrate_disable();
2649 * set_cpus_allowed_ptr(P0, [1]);
2653 * __set_cpus_allowed_ptr();
2654 * <wakes local stopper>
2655 * `--> <woken on migration completion>
2657 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2658 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2659 * task p are serialized by p->pi_lock, which we can leverage: the one that
2660 * should come into effect at the end of the Migrate-Disable region is the last
2661 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2662 * but we still need to properly signal those waiting tasks at the appropriate
2665 * This is implemented using struct set_affinity_pending. The first
2666 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2667 * setup an instance of that struct and install it on the targeted task_struct.
2668 * Any and all further callers will reuse that instance. Those then wait for
2669 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2670 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2673 * (1) In the cases covered above. There is one more where the completion is
2674 * signaled within affine_move_task() itself: when a subsequent affinity request
2675 * occurs after the stopper bailed out due to the targeted task still being
2676 * Migrate-Disable. Consider:
2678 * Initial conditions: P0->cpus_mask = [0, 1]
2682 * migrate_disable();
2684 * set_cpus_allowed_ptr(P0, [1]);
2687 * migration_cpu_stop()
2688 * is_migration_disabled()
2690 * set_cpus_allowed_ptr(P0, [0, 1]);
2691 * <signal completion>
2694 * Note that the above is safe vs a concurrent migrate_enable(), as any
2695 * pending affinity completion is preceded by an uninstallation of
2696 * p->migration_pending done with p->pi_lock held.
2698 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2699 int dest_cpu, unsigned int flags)
2701 struct set_affinity_pending my_pending = { }, *pending = NULL;
2702 bool stop_pending, complete = false;
2704 /* Can the task run on the task's current CPU? If so, we're done */
2705 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2706 struct task_struct *push_task = NULL;
2708 if ((flags & SCA_MIGRATE_ENABLE) &&
2709 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2710 rq->push_busy = true;
2711 push_task = get_task_struct(p);
2715 * If there are pending waiters, but no pending stop_work,
2716 * then complete now.
2718 pending = p->migration_pending;
2719 if (pending && !pending->stop_pending) {
2720 p->migration_pending = NULL;
2724 task_rq_unlock(rq, p, rf);
2727 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2732 complete_all(&pending->done);
2737 if (!(flags & SCA_MIGRATE_ENABLE)) {
2738 /* serialized by p->pi_lock */
2739 if (!p->migration_pending) {
2740 /* Install the request */
2741 refcount_set(&my_pending.refs, 1);
2742 init_completion(&my_pending.done);
2743 my_pending.arg = (struct migration_arg) {
2745 .dest_cpu = dest_cpu,
2746 .pending = &my_pending,
2749 p->migration_pending = &my_pending;
2751 pending = p->migration_pending;
2752 refcount_inc(&pending->refs);
2754 * Affinity has changed, but we've already installed a
2755 * pending. migration_cpu_stop() *must* see this, else
2756 * we risk a completion of the pending despite having a
2757 * task on a disallowed CPU.
2759 * Serialized by p->pi_lock, so this is safe.
2761 pending->arg.dest_cpu = dest_cpu;
2764 pending = p->migration_pending;
2766 * - !MIGRATE_ENABLE:
2767 * we'll have installed a pending if there wasn't one already.
2770 * we're here because the current CPU isn't matching anymore,
2771 * the only way that can happen is because of a concurrent
2772 * set_cpus_allowed_ptr() call, which should then still be
2773 * pending completion.
2775 * Either way, we really should have a @pending here.
2777 if (WARN_ON_ONCE(!pending)) {
2778 task_rq_unlock(rq, p, rf);
2782 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2784 * MIGRATE_ENABLE gets here because 'p == current', but for
2785 * anything else we cannot do is_migration_disabled(), punt
2786 * and have the stopper function handle it all race-free.
2788 stop_pending = pending->stop_pending;
2790 pending->stop_pending = true;
2792 if (flags & SCA_MIGRATE_ENABLE)
2793 p->migration_flags &= ~MDF_PUSH;
2795 task_rq_unlock(rq, p, rf);
2797 if (!stop_pending) {
2798 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2799 &pending->arg, &pending->stop_work);
2802 if (flags & SCA_MIGRATE_ENABLE)
2806 if (!is_migration_disabled(p)) {
2807 if (task_on_rq_queued(p))
2808 rq = move_queued_task(rq, rf, p, dest_cpu);
2810 if (!pending->stop_pending) {
2811 p->migration_pending = NULL;
2815 task_rq_unlock(rq, p, rf);
2818 complete_all(&pending->done);
2821 wait_for_completion(&pending->done);
2823 if (refcount_dec_and_test(&pending->refs))
2824 wake_up_var(&pending->refs); /* No UaF, just an address */
2827 * Block the original owner of &pending until all subsequent callers
2828 * have seen the completion and decremented the refcount
2830 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2833 WARN_ON_ONCE(my_pending.stop_pending);
2839 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2841 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2842 const struct cpumask *new_mask,
2845 struct rq_flags *rf)
2846 __releases(rq->lock)
2847 __releases(p->pi_lock)
2849 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2850 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2851 bool kthread = p->flags & PF_KTHREAD;
2852 struct cpumask *user_mask = NULL;
2853 unsigned int dest_cpu;
2856 update_rq_clock(rq);
2858 if (kthread || is_migration_disabled(p)) {
2860 * Kernel threads are allowed on online && !active CPUs,
2861 * however, during cpu-hot-unplug, even these might get pushed
2862 * away if not KTHREAD_IS_PER_CPU.
2864 * Specifically, migration_disabled() tasks must not fail the
2865 * cpumask_any_and_distribute() pick below, esp. so on
2866 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2867 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2869 cpu_valid_mask = cpu_online_mask;
2872 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2878 * Must re-check here, to close a race against __kthread_bind(),
2879 * sched_setaffinity() is not guaranteed to observe the flag.
2881 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2886 if (!(flags & SCA_MIGRATE_ENABLE)) {
2887 if (cpumask_equal(&p->cpus_mask, new_mask))
2890 if (WARN_ON_ONCE(p == current &&
2891 is_migration_disabled(p) &&
2892 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2899 * Picking a ~random cpu helps in cases where we are changing affinity
2900 * for groups of tasks (ie. cpuset), so that load balancing is not
2901 * immediately required to distribute the tasks within their new mask.
2903 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2904 if (dest_cpu >= nr_cpu_ids) {
2909 __do_set_cpus_allowed(p, new_mask, flags);
2911 if (flags & SCA_USER)
2912 user_mask = clear_user_cpus_ptr(p);
2914 ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2921 task_rq_unlock(rq, p, rf);
2927 * Change a given task's CPU affinity. Migrate the thread to a
2928 * proper CPU and schedule it away if the CPU it's executing on
2929 * is removed from the allowed bitmask.
2931 * NOTE: the caller must have a valid reference to the task, the
2932 * task must not exit() & deallocate itself prematurely. The
2933 * call is not atomic; no spinlocks may be held.
2935 static int __set_cpus_allowed_ptr(struct task_struct *p,
2936 const struct cpumask *new_mask, u32 flags)
2941 rq = task_rq_lock(p, &rf);
2942 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2945 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2947 return __set_cpus_allowed_ptr(p, new_mask, 0);
2949 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2952 * Change a given task's CPU affinity to the intersection of its current
2953 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2954 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2955 * If the resulting mask is empty, leave the affinity unchanged and return
2958 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2959 struct cpumask *new_mask,
2960 const struct cpumask *subset_mask)
2962 struct cpumask *user_mask = NULL;
2967 if (!p->user_cpus_ptr) {
2968 user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2973 rq = task_rq_lock(p, &rf);
2976 * Forcefully restricting the affinity of a deadline task is
2977 * likely to cause problems, so fail and noisily override the
2980 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2985 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2991 * We're about to butcher the task affinity, so keep track of what
2992 * the user asked for in case we're able to restore it later on.
2995 cpumask_copy(user_mask, p->cpus_ptr);
2996 p->user_cpus_ptr = user_mask;
2999 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
3002 task_rq_unlock(rq, p, &rf);
3008 * Restrict the CPU affinity of task @p so that it is a subset of
3009 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
3010 * old affinity mask. If the resulting mask is empty, we warn and walk
3011 * up the cpuset hierarchy until we find a suitable mask.
3013 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3015 cpumask_var_t new_mask;
3016 const struct cpumask *override_mask = task_cpu_possible_mask(p);
3018 alloc_cpumask_var(&new_mask, GFP_KERNEL);
3021 * __migrate_task() can fail silently in the face of concurrent
3022 * offlining of the chosen destination CPU, so take the hotplug
3023 * lock to ensure that the migration succeeds.
3026 if (!cpumask_available(new_mask))
3029 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3033 * We failed to find a valid subset of the affinity mask for the
3034 * task, so override it based on its cpuset hierarchy.
3036 cpuset_cpus_allowed(p, new_mask);
3037 override_mask = new_mask;
3040 if (printk_ratelimit()) {
3041 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3042 task_pid_nr(p), p->comm,
3043 cpumask_pr_args(override_mask));
3046 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3049 free_cpumask_var(new_mask);
3053 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
3056 * Restore the affinity of a task @p which was previously restricted by a
3057 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
3058 * @p->user_cpus_ptr.
3060 * It is the caller's responsibility to serialise this with any calls to
3061 * force_compatible_cpus_allowed_ptr(@p).
3063 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3065 struct cpumask *user_mask = p->user_cpus_ptr;
3066 unsigned long flags;
3069 * Try to restore the old affinity mask. If this fails, then
3070 * we free the mask explicitly to avoid it being inherited across
3071 * a subsequent fork().
3073 if (!user_mask || !__sched_setaffinity(p, user_mask))
3076 raw_spin_lock_irqsave(&p->pi_lock, flags);
3077 user_mask = clear_user_cpus_ptr(p);
3078 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3083 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3085 #ifdef CONFIG_SCHED_DEBUG
3086 unsigned int state = READ_ONCE(p->__state);
3089 * We should never call set_task_cpu() on a blocked task,
3090 * ttwu() will sort out the placement.
3092 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3095 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3096 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3097 * time relying on p->on_rq.
3099 WARN_ON_ONCE(state == TASK_RUNNING &&
3100 p->sched_class == &fair_sched_class &&
3101 (p->on_rq && !task_on_rq_migrating(p)));
3103 #ifdef CONFIG_LOCKDEP
3105 * The caller should hold either p->pi_lock or rq->lock, when changing
3106 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3108 * sched_move_task() holds both and thus holding either pins the cgroup,
3111 * Furthermore, all task_rq users should acquire both locks, see
3114 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3115 lockdep_is_held(__rq_lockp(task_rq(p)))));
3118 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3120 WARN_ON_ONCE(!cpu_online(new_cpu));
3122 WARN_ON_ONCE(is_migration_disabled(p));
3125 trace_sched_migrate_task(p, new_cpu);
3127 if (task_cpu(p) != new_cpu) {
3128 if (p->sched_class->migrate_task_rq)
3129 p->sched_class->migrate_task_rq(p, new_cpu);
3130 p->se.nr_migrations++;
3132 perf_event_task_migrate(p);
3135 __set_task_cpu(p, new_cpu);
3138 #ifdef CONFIG_NUMA_BALANCING
3139 static void __migrate_swap_task(struct task_struct *p, int cpu)
3141 if (task_on_rq_queued(p)) {
3142 struct rq *src_rq, *dst_rq;
3143 struct rq_flags srf, drf;
3145 src_rq = task_rq(p);
3146 dst_rq = cpu_rq(cpu);
3148 rq_pin_lock(src_rq, &srf);
3149 rq_pin_lock(dst_rq, &drf);
3151 deactivate_task(src_rq, p, 0);
3152 set_task_cpu(p, cpu);
3153 activate_task(dst_rq, p, 0);
3154 check_preempt_curr(dst_rq, p, 0);
3156 rq_unpin_lock(dst_rq, &drf);
3157 rq_unpin_lock(src_rq, &srf);
3161 * Task isn't running anymore; make it appear like we migrated
3162 * it before it went to sleep. This means on wakeup we make the
3163 * previous CPU our target instead of where it really is.
3169 struct migration_swap_arg {
3170 struct task_struct *src_task, *dst_task;
3171 int src_cpu, dst_cpu;
3174 static int migrate_swap_stop(void *data)
3176 struct migration_swap_arg *arg = data;
3177 struct rq *src_rq, *dst_rq;
3180 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3183 src_rq = cpu_rq(arg->src_cpu);
3184 dst_rq = cpu_rq(arg->dst_cpu);
3186 double_raw_lock(&arg->src_task->pi_lock,
3187 &arg->dst_task->pi_lock);
3188 double_rq_lock(src_rq, dst_rq);
3190 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3193 if (task_cpu(arg->src_task) != arg->src_cpu)
3196 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3199 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3202 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3203 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3208 double_rq_unlock(src_rq, dst_rq);
3209 raw_spin_unlock(&arg->dst_task->pi_lock);
3210 raw_spin_unlock(&arg->src_task->pi_lock);
3216 * Cross migrate two tasks
3218 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3219 int target_cpu, int curr_cpu)
3221 struct migration_swap_arg arg;
3224 arg = (struct migration_swap_arg){
3226 .src_cpu = curr_cpu,
3228 .dst_cpu = target_cpu,
3231 if (arg.src_cpu == arg.dst_cpu)
3235 * These three tests are all lockless; this is OK since all of them
3236 * will be re-checked with proper locks held further down the line.
3238 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3241 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3244 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3247 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3248 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3253 #endif /* CONFIG_NUMA_BALANCING */
3256 * wait_task_inactive - wait for a thread to unschedule.
3258 * If @match_state is nonzero, it's the @p->state value just checked and
3259 * not expected to change. If it changes, i.e. @p might have woken up,
3260 * then return zero. When we succeed in waiting for @p to be off its CPU,
3261 * we return a positive number (its total switch count). If a second call
3262 * a short while later returns the same number, the caller can be sure that
3263 * @p has remained unscheduled the whole time.
3265 * The caller must ensure that the task *will* unschedule sometime soon,
3266 * else this function might spin for a *long* time. This function can't
3267 * be called with interrupts off, or it may introduce deadlock with
3268 * smp_call_function() if an IPI is sent by the same process we are
3269 * waiting to become inactive.
3271 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3273 int running, queued;
3280 * We do the initial early heuristics without holding
3281 * any task-queue locks at all. We'll only try to get
3282 * the runqueue lock when things look like they will
3288 * If the task is actively running on another CPU
3289 * still, just relax and busy-wait without holding
3292 * NOTE! Since we don't hold any locks, it's not
3293 * even sure that "rq" stays as the right runqueue!
3294 * But we don't care, since "task_running()" will
3295 * return false if the runqueue has changed and p
3296 * is actually now running somewhere else!
3298 while (task_running(rq, p)) {
3299 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3305 * Ok, time to look more closely! We need the rq
3306 * lock now, to be *sure*. If we're wrong, we'll
3307 * just go back and repeat.
3309 rq = task_rq_lock(p, &rf);
3310 trace_sched_wait_task(p);
3311 running = task_running(rq, p);
3312 queued = task_on_rq_queued(p);
3314 if (!match_state || READ_ONCE(p->__state) == match_state)
3315 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3316 task_rq_unlock(rq, p, &rf);
3319 * If it changed from the expected state, bail out now.
3321 if (unlikely(!ncsw))
3325 * Was it really running after all now that we
3326 * checked with the proper locks actually held?
3328 * Oops. Go back and try again..
3330 if (unlikely(running)) {
3336 * It's not enough that it's not actively running,
3337 * it must be off the runqueue _entirely_, and not
3340 * So if it was still runnable (but just not actively
3341 * running right now), it's preempted, and we should
3342 * yield - it could be a while.
3344 if (unlikely(queued)) {
3345 ktime_t to = NSEC_PER_SEC / HZ;
3347 set_current_state(TASK_UNINTERRUPTIBLE);
3348 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3353 * Ahh, all good. It wasn't running, and it wasn't
3354 * runnable, which means that it will never become
3355 * running in the future either. We're all done!
3364 * kick_process - kick a running thread to enter/exit the kernel
3365 * @p: the to-be-kicked thread
3367 * Cause a process which is running on another CPU to enter
3368 * kernel-mode, without any delay. (to get signals handled.)
3370 * NOTE: this function doesn't have to take the runqueue lock,
3371 * because all it wants to ensure is that the remote task enters
3372 * the kernel. If the IPI races and the task has been migrated
3373 * to another CPU then no harm is done and the purpose has been
3376 void kick_process(struct task_struct *p)
3382 if ((cpu != smp_processor_id()) && task_curr(p))
3383 smp_send_reschedule(cpu);
3386 EXPORT_SYMBOL_GPL(kick_process);
3389 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3391 * A few notes on cpu_active vs cpu_online:
3393 * - cpu_active must be a subset of cpu_online
3395 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3396 * see __set_cpus_allowed_ptr(). At this point the newly online
3397 * CPU isn't yet part of the sched domains, and balancing will not
3400 * - on CPU-down we clear cpu_active() to mask the sched domains and
3401 * avoid the load balancer to place new tasks on the to be removed
3402 * CPU. Existing tasks will remain running there and will be taken
3405 * This means that fallback selection must not select !active CPUs.
3406 * And can assume that any active CPU must be online. Conversely
3407 * select_task_rq() below may allow selection of !active CPUs in order
3408 * to satisfy the above rules.
3410 static int select_fallback_rq(int cpu, struct task_struct *p)
3412 int nid = cpu_to_node(cpu);
3413 const struct cpumask *nodemask = NULL;
3414 enum { cpuset, possible, fail } state = cpuset;
3418 * If the node that the CPU is on has been offlined, cpu_to_node()
3419 * will return -1. There is no CPU on the node, and we should
3420 * select the CPU on the other node.
3423 nodemask = cpumask_of_node(nid);
3425 /* Look for allowed, online CPU in same node. */
3426 for_each_cpu(dest_cpu, nodemask) {
3427 if (is_cpu_allowed(p, dest_cpu))
3433 /* Any allowed, online CPU? */
3434 for_each_cpu(dest_cpu, p->cpus_ptr) {
3435 if (!is_cpu_allowed(p, dest_cpu))
3441 /* No more Mr. Nice Guy. */
3444 if (cpuset_cpus_allowed_fallback(p)) {
3451 * XXX When called from select_task_rq() we only
3452 * hold p->pi_lock and again violate locking order.
3454 * More yuck to audit.
3456 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3466 if (state != cpuset) {
3468 * Don't tell them about moving exiting tasks or
3469 * kernel threads (both mm NULL), since they never
3472 if (p->mm && printk_ratelimit()) {
3473 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3474 task_pid_nr(p), p->comm, cpu);
3482 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3485 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3487 lockdep_assert_held(&p->pi_lock);
3489 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3490 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3492 cpu = cpumask_any(p->cpus_ptr);
3495 * In order not to call set_task_cpu() on a blocking task we need
3496 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3499 * Since this is common to all placement strategies, this lives here.
3501 * [ this allows ->select_task() to simply return task_cpu(p) and
3502 * not worry about this generic constraint ]
3504 if (unlikely(!is_cpu_allowed(p, cpu)))
3505 cpu = select_fallback_rq(task_cpu(p), p);
3510 void sched_set_stop_task(int cpu, struct task_struct *stop)
3512 static struct lock_class_key stop_pi_lock;
3513 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3514 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3518 * Make it appear like a SCHED_FIFO task, its something
3519 * userspace knows about and won't get confused about.
3521 * Also, it will make PI more or less work without too
3522 * much confusion -- but then, stop work should not
3523 * rely on PI working anyway.
3525 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3527 stop->sched_class = &stop_sched_class;
3530 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3531 * adjust the effective priority of a task. As a result,
3532 * rt_mutex_setprio() can trigger (RT) balancing operations,
3533 * which can then trigger wakeups of the stop thread to push
3534 * around the current task.
3536 * The stop task itself will never be part of the PI-chain, it
3537 * never blocks, therefore that ->pi_lock recursion is safe.
3538 * Tell lockdep about this by placing the stop->pi_lock in its
3541 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3544 cpu_rq(cpu)->stop = stop;
3548 * Reset it back to a normal scheduling class so that
3549 * it can die in pieces.
3551 old_stop->sched_class = &rt_sched_class;
3555 #else /* CONFIG_SMP */
3557 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3558 const struct cpumask *new_mask,
3561 return set_cpus_allowed_ptr(p, new_mask);
3564 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3566 static inline bool rq_has_pinned_tasks(struct rq *rq)
3571 #endif /* !CONFIG_SMP */
3574 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3578 if (!schedstat_enabled())
3584 if (cpu == rq->cpu) {
3585 __schedstat_inc(rq->ttwu_local);
3586 __schedstat_inc(p->stats.nr_wakeups_local);
3588 struct sched_domain *sd;
3590 __schedstat_inc(p->stats.nr_wakeups_remote);
3592 for_each_domain(rq->cpu, sd) {
3593 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3594 __schedstat_inc(sd->ttwu_wake_remote);
3601 if (wake_flags & WF_MIGRATED)
3602 __schedstat_inc(p->stats.nr_wakeups_migrate);
3603 #endif /* CONFIG_SMP */
3605 __schedstat_inc(rq->ttwu_count);
3606 __schedstat_inc(p->stats.nr_wakeups);
3608 if (wake_flags & WF_SYNC)
3609 __schedstat_inc(p->stats.nr_wakeups_sync);
3613 * Mark the task runnable and perform wakeup-preemption.
3615 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3616 struct rq_flags *rf)
3618 check_preempt_curr(rq, p, wake_flags);
3619 WRITE_ONCE(p->__state, TASK_RUNNING);
3620 trace_sched_wakeup(p);
3623 if (p->sched_class->task_woken) {
3625 * Our task @p is fully woken up and running; so it's safe to
3626 * drop the rq->lock, hereafter rq is only used for statistics.
3628 rq_unpin_lock(rq, rf);
3629 p->sched_class->task_woken(rq, p);
3630 rq_repin_lock(rq, rf);
3633 if (rq->idle_stamp) {
3634 u64 delta = rq_clock(rq) - rq->idle_stamp;
3635 u64 max = 2*rq->max_idle_balance_cost;
3637 update_avg(&rq->avg_idle, delta);
3639 if (rq->avg_idle > max)
3642 rq->wake_stamp = jiffies;
3643 rq->wake_avg_idle = rq->avg_idle / 2;
3651 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3652 struct rq_flags *rf)
3654 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3656 lockdep_assert_rq_held(rq);
3658 if (p->sched_contributes_to_load)
3659 rq->nr_uninterruptible--;
3662 if (wake_flags & WF_MIGRATED)
3663 en_flags |= ENQUEUE_MIGRATED;
3667 delayacct_blkio_end(p);
3668 atomic_dec(&task_rq(p)->nr_iowait);
3671 activate_task(rq, p, en_flags);
3672 ttwu_do_wakeup(rq, p, wake_flags, rf);
3676 * Consider @p being inside a wait loop:
3679 * set_current_state(TASK_UNINTERRUPTIBLE);
3686 * __set_current_state(TASK_RUNNING);
3688 * between set_current_state() and schedule(). In this case @p is still
3689 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3692 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3693 * then schedule() must still happen and p->state can be changed to
3694 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3695 * need to do a full wakeup with enqueue.
3697 * Returns: %true when the wakeup is done,
3700 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3706 rq = __task_rq_lock(p, &rf);
3707 if (task_on_rq_queued(p)) {
3708 /* check_preempt_curr() may use rq clock */
3709 update_rq_clock(rq);
3710 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3713 __task_rq_unlock(rq, &rf);
3719 void sched_ttwu_pending(void *arg)
3721 struct llist_node *llist = arg;
3722 struct rq *rq = this_rq();
3723 struct task_struct *p, *t;
3730 * rq::ttwu_pending racy indication of out-standing wakeups.
3731 * Races such that false-negatives are possible, since they
3732 * are shorter lived that false-positives would be.
3734 WRITE_ONCE(rq->ttwu_pending, 0);
3736 rq_lock_irqsave(rq, &rf);
3737 update_rq_clock(rq);
3739 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3740 if (WARN_ON_ONCE(p->on_cpu))
3741 smp_cond_load_acquire(&p->on_cpu, !VAL);
3743 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3744 set_task_cpu(p, cpu_of(rq));
3746 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3749 rq_unlock_irqrestore(rq, &rf);
3752 void send_call_function_single_ipi(int cpu)
3754 struct rq *rq = cpu_rq(cpu);
3756 if (!set_nr_if_polling(rq->idle))
3757 arch_send_call_function_single_ipi(cpu);
3759 trace_sched_wake_idle_without_ipi(cpu);
3763 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3764 * necessary. The wakee CPU on receipt of the IPI will queue the task
3765 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3766 * of the wakeup instead of the waker.
3768 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3770 struct rq *rq = cpu_rq(cpu);
3772 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3774 WRITE_ONCE(rq->ttwu_pending, 1);
3775 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3778 void wake_up_if_idle(int cpu)
3780 struct rq *rq = cpu_rq(cpu);
3785 if (!is_idle_task(rcu_dereference(rq->curr)))
3788 rq_lock_irqsave(rq, &rf);
3789 if (is_idle_task(rq->curr))
3791 /* Else CPU is not idle, do nothing here: */
3792 rq_unlock_irqrestore(rq, &rf);
3798 bool cpus_share_cache(int this_cpu, int that_cpu)
3800 if (this_cpu == that_cpu)
3803 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3806 static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3809 * Do not complicate things with the async wake_list while the CPU is
3812 if (!cpu_active(cpu))
3815 /* Ensure the task will still be allowed to run on the CPU. */
3816 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3820 * If the CPU does not share cache, then queue the task on the
3821 * remote rqs wakelist to avoid accessing remote data.
3823 if (!cpus_share_cache(smp_processor_id(), cpu))
3826 if (cpu == smp_processor_id())
3830 * If the wakee cpu is idle, or the task is descheduling and the
3831 * only running task on the CPU, then use the wakelist to offload
3832 * the task activation to the idle (or soon-to-be-idle) CPU as
3833 * the current CPU is likely busy. nr_running is checked to
3834 * avoid unnecessary task stacking.
3836 * Note that we can only get here with (wakee) p->on_rq=0,
3837 * p->on_cpu can be whatever, we've done the dequeue, so
3838 * the wakee has been accounted out of ->nr_running.
3840 if (!cpu_rq(cpu)->nr_running)
3846 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3848 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
3849 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3850 __ttwu_queue_wakelist(p, cpu, wake_flags);
3857 #else /* !CONFIG_SMP */
3859 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3864 #endif /* CONFIG_SMP */
3866 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3868 struct rq *rq = cpu_rq(cpu);
3871 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3875 update_rq_clock(rq);
3876 ttwu_do_activate(rq, p, wake_flags, &rf);
3881 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3883 * The caller holds p::pi_lock if p != current or has preemption
3884 * disabled when p == current.
3886 * The rules of PREEMPT_RT saved_state:
3888 * The related locking code always holds p::pi_lock when updating
3889 * p::saved_state, which means the code is fully serialized in both cases.
3891 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3892 * bits set. This allows to distinguish all wakeup scenarios.
3894 static __always_inline
3895 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3897 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3898 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3899 state != TASK_RTLOCK_WAIT);
3902 if (READ_ONCE(p->__state) & state) {
3907 #ifdef CONFIG_PREEMPT_RT
3909 * Saved state preserves the task state across blocking on
3910 * an RT lock. If the state matches, set p::saved_state to
3911 * TASK_RUNNING, but do not wake the task because it waits
3912 * for a lock wakeup. Also indicate success because from
3913 * the regular waker's point of view this has succeeded.
3915 * After acquiring the lock the task will restore p::__state
3916 * from p::saved_state which ensures that the regular
3917 * wakeup is not lost. The restore will also set
3918 * p::saved_state to TASK_RUNNING so any further tests will
3919 * not result in false positives vs. @success
3921 if (p->saved_state & state) {
3922 p->saved_state = TASK_RUNNING;
3930 * Notes on Program-Order guarantees on SMP systems.
3934 * The basic program-order guarantee on SMP systems is that when a task [t]
3935 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3936 * execution on its new CPU [c1].
3938 * For migration (of runnable tasks) this is provided by the following means:
3940 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3941 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3942 * rq(c1)->lock (if not at the same time, then in that order).
3943 * C) LOCK of the rq(c1)->lock scheduling in task
3945 * Release/acquire chaining guarantees that B happens after A and C after B.
3946 * Note: the CPU doing B need not be c0 or c1
3955 * UNLOCK rq(0)->lock
3957 * LOCK rq(0)->lock // orders against CPU0
3959 * UNLOCK rq(0)->lock
3963 * UNLOCK rq(1)->lock
3965 * LOCK rq(1)->lock // orders against CPU2
3968 * UNLOCK rq(1)->lock
3971 * BLOCKING -- aka. SLEEP + WAKEUP
3973 * For blocking we (obviously) need to provide the same guarantee as for
3974 * migration. However the means are completely different as there is no lock
3975 * chain to provide order. Instead we do:
3977 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3978 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3982 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3984 * LOCK rq(0)->lock LOCK X->pi_lock
3987 * smp_store_release(X->on_cpu, 0);
3989 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3995 * X->state = RUNNING
3996 * UNLOCK rq(2)->lock
3998 * LOCK rq(2)->lock // orders against CPU1
4001 * UNLOCK rq(2)->lock
4004 * UNLOCK rq(0)->lock
4007 * However, for wakeups there is a second guarantee we must provide, namely we
4008 * must ensure that CONDITION=1 done by the caller can not be reordered with
4009 * accesses to the task state; see try_to_wake_up() and set_current_state().
4013 * try_to_wake_up - wake up a thread
4014 * @p: the thread to be awakened
4015 * @state: the mask of task states that can be woken
4016 * @wake_flags: wake modifier flags (WF_*)
4018 * Conceptually does:
4020 * If (@state & @p->state) @p->state = TASK_RUNNING.
4022 * If the task was not queued/runnable, also place it back on a runqueue.
4024 * This function is atomic against schedule() which would dequeue the task.
4026 * It issues a full memory barrier before accessing @p->state, see the comment
4027 * with set_current_state().
4029 * Uses p->pi_lock to serialize against concurrent wake-ups.
4031 * Relies on p->pi_lock stabilizing:
4034 * - p->sched_task_group
4035 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4037 * Tries really hard to only take one task_rq(p)->lock for performance.
4038 * Takes rq->lock in:
4039 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4040 * - ttwu_queue() -- new rq, for enqueue of the task;
4041 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4043 * As a consequence we race really badly with just about everything. See the
4044 * many memory barriers and their comments for details.
4046 * Return: %true if @p->state changes (an actual wakeup was done),
4050 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4052 unsigned long flags;
4053 int cpu, success = 0;
4058 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4059 * == smp_processor_id()'. Together this means we can special
4060 * case the whole 'p->on_rq && ttwu_runnable()' case below
4061 * without taking any locks.
4064 * - we rely on Program-Order guarantees for all the ordering,
4065 * - we're serialized against set_special_state() by virtue of
4066 * it disabling IRQs (this allows not taking ->pi_lock).
4068 if (!ttwu_state_match(p, state, &success))
4071 trace_sched_waking(p);
4072 WRITE_ONCE(p->__state, TASK_RUNNING);
4073 trace_sched_wakeup(p);
4078 * If we are going to wake up a thread waiting for CONDITION we
4079 * need to ensure that CONDITION=1 done by the caller can not be
4080 * reordered with p->state check below. This pairs with smp_store_mb()
4081 * in set_current_state() that the waiting thread does.
4083 raw_spin_lock_irqsave(&p->pi_lock, flags);
4084 smp_mb__after_spinlock();
4085 if (!ttwu_state_match(p, state, &success))
4088 trace_sched_waking(p);
4091 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4092 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4093 * in smp_cond_load_acquire() below.
4095 * sched_ttwu_pending() try_to_wake_up()
4096 * STORE p->on_rq = 1 LOAD p->state
4099 * __schedule() (switch to task 'p')
4100 * LOCK rq->lock smp_rmb();
4101 * smp_mb__after_spinlock();
4105 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4107 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4108 * __schedule(). See the comment for smp_mb__after_spinlock().
4110 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4113 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4118 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4119 * possible to, falsely, observe p->on_cpu == 0.
4121 * One must be running (->on_cpu == 1) in order to remove oneself
4122 * from the runqueue.
4124 * __schedule() (switch to task 'p') try_to_wake_up()
4125 * STORE p->on_cpu = 1 LOAD p->on_rq
4128 * __schedule() (put 'p' to sleep)
4129 * LOCK rq->lock smp_rmb();
4130 * smp_mb__after_spinlock();
4131 * STORE p->on_rq = 0 LOAD p->on_cpu
4133 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4134 * __schedule(). See the comment for smp_mb__after_spinlock().
4136 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4137 * schedule()'s deactivate_task() has 'happened' and p will no longer
4138 * care about it's own p->state. See the comment in __schedule().
4140 smp_acquire__after_ctrl_dep();
4143 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4144 * == 0), which means we need to do an enqueue, change p->state to
4145 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4146 * enqueue, such as ttwu_queue_wakelist().
4148 WRITE_ONCE(p->__state, TASK_WAKING);
4151 * If the owning (remote) CPU is still in the middle of schedule() with
4152 * this task as prev, considering queueing p on the remote CPUs wake_list
4153 * which potentially sends an IPI instead of spinning on p->on_cpu to
4154 * let the waker make forward progress. This is safe because IRQs are
4155 * disabled and the IPI will deliver after on_cpu is cleared.
4157 * Ensure we load task_cpu(p) after p->on_cpu:
4159 * set_task_cpu(p, cpu);
4160 * STORE p->cpu = @cpu
4161 * __schedule() (switch to task 'p')
4163 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4164 * STORE p->on_cpu = 1 LOAD p->cpu
4166 * to ensure we observe the correct CPU on which the task is currently
4169 if (smp_load_acquire(&p->on_cpu) &&
4170 ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4174 * If the owning (remote) CPU is still in the middle of schedule() with
4175 * this task as prev, wait until it's done referencing the task.
4177 * Pairs with the smp_store_release() in finish_task().
4179 * This ensures that tasks getting woken will be fully ordered against
4180 * their previous state and preserve Program Order.
4182 smp_cond_load_acquire(&p->on_cpu, !VAL);
4184 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4185 if (task_cpu(p) != cpu) {
4187 delayacct_blkio_end(p);
4188 atomic_dec(&task_rq(p)->nr_iowait);
4191 wake_flags |= WF_MIGRATED;
4192 psi_ttwu_dequeue(p);
4193 set_task_cpu(p, cpu);
4197 #endif /* CONFIG_SMP */
4199 ttwu_queue(p, cpu, wake_flags);
4201 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4204 ttwu_stat(p, task_cpu(p), wake_flags);
4211 * task_call_func - Invoke a function on task in fixed state
4212 * @p: Process for which the function is to be invoked, can be @current.
4213 * @func: Function to invoke.
4214 * @arg: Argument to function.
4216 * Fix the task in it's current state by avoiding wakeups and or rq operations
4217 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4218 * to work out what the state is, if required. Given that @func can be invoked
4219 * with a runqueue lock held, it had better be quite lightweight.
4222 * Whatever @func returns
4224 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4226 struct rq *rq = NULL;
4231 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4233 state = READ_ONCE(p->__state);
4236 * Ensure we load p->on_rq after p->__state, otherwise it would be
4237 * possible to, falsely, observe p->on_rq == 0.
4239 * See try_to_wake_up() for a longer comment.
4244 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4245 * the task is blocked. Make sure to check @state since ttwu() can drop
4246 * locks at the end, see ttwu_queue_wakelist().
4248 if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4249 rq = __task_rq_lock(p, &rf);
4252 * At this point the task is pinned; either:
4253 * - blocked and we're holding off wakeups (pi->lock)
4254 * - woken, and we're holding off enqueue (rq->lock)
4255 * - queued, and we're holding off schedule (rq->lock)
4256 * - running, and we're holding off de-schedule (rq->lock)
4258 * The called function (@func) can use: task_curr(), p->on_rq and
4259 * p->__state to differentiate between these states.
4266 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4271 * cpu_curr_snapshot - Return a snapshot of the currently running task
4272 * @cpu: The CPU on which to snapshot the task.
4274 * Returns the task_struct pointer of the task "currently" running on
4275 * the specified CPU. If the same task is running on that CPU throughout,
4276 * the return value will be a pointer to that task's task_struct structure.
4277 * If the CPU did any context switches even vaguely concurrently with the
4278 * execution of this function, the return value will be a pointer to the
4279 * task_struct structure of a randomly chosen task that was running on
4280 * that CPU somewhere around the time that this function was executing.
4282 * If the specified CPU was offline, the return value is whatever it
4283 * is, perhaps a pointer to the task_struct structure of that CPU's idle
4284 * task, but there is no guarantee. Callers wishing a useful return
4285 * value must take some action to ensure that the specified CPU remains
4286 * online throughout.
4288 * This function executes full memory barriers before and after fetching
4289 * the pointer, which permits the caller to confine this function's fetch
4290 * with respect to the caller's accesses to other shared variables.
4292 struct task_struct *cpu_curr_snapshot(int cpu)
4294 struct task_struct *t;
4296 smp_mb(); /* Pairing determined by caller's synchronization design. */
4297 t = rcu_dereference(cpu_curr(cpu));
4298 smp_mb(); /* Pairing determined by caller's synchronization design. */
4303 * wake_up_process - Wake up a specific process
4304 * @p: The process to be woken up.
4306 * Attempt to wake up the nominated process and move it to the set of runnable
4309 * Return: 1 if the process was woken up, 0 if it was already running.
4311 * This function executes a full memory barrier before accessing the task state.
4313 int wake_up_process(struct task_struct *p)
4315 return try_to_wake_up(p, TASK_NORMAL, 0);
4317 EXPORT_SYMBOL(wake_up_process);
4319 int wake_up_state(struct task_struct *p, unsigned int state)
4321 return try_to_wake_up(p, state, 0);
4325 * Perform scheduler related setup for a newly forked process p.
4326 * p is forked by current.
4328 * __sched_fork() is basic setup used by init_idle() too:
4330 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4335 p->se.exec_start = 0;
4336 p->se.sum_exec_runtime = 0;
4337 p->se.prev_sum_exec_runtime = 0;
4338 p->se.nr_migrations = 0;
4340 INIT_LIST_HEAD(&p->se.group_node);
4342 #ifdef CONFIG_FAIR_GROUP_SCHED
4343 p->se.cfs_rq = NULL;
4346 #ifdef CONFIG_SCHEDSTATS
4347 /* Even if schedstat is disabled, there should not be garbage */
4348 memset(&p->stats, 0, sizeof(p->stats));
4351 RB_CLEAR_NODE(&p->dl.rb_node);
4352 init_dl_task_timer(&p->dl);
4353 init_dl_inactive_task_timer(&p->dl);
4354 __dl_clear_params(p);
4356 INIT_LIST_HEAD(&p->rt.run_list);
4358 p->rt.time_slice = sched_rr_timeslice;
4362 #ifdef CONFIG_PREEMPT_NOTIFIERS
4363 INIT_HLIST_HEAD(&p->preempt_notifiers);
4366 #ifdef CONFIG_COMPACTION
4367 p->capture_control = NULL;
4369 init_numa_balancing(clone_flags, p);
4371 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4372 p->migration_pending = NULL;
4376 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4378 #ifdef CONFIG_NUMA_BALANCING
4380 int sysctl_numa_balancing_mode;
4382 static void __set_numabalancing_state(bool enabled)
4385 static_branch_enable(&sched_numa_balancing);
4387 static_branch_disable(&sched_numa_balancing);
4390 void set_numabalancing_state(bool enabled)
4393 sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4395 sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4396 __set_numabalancing_state(enabled);
4399 #ifdef CONFIG_PROC_SYSCTL
4400 int sysctl_numa_balancing(struct ctl_table *table, int write,
4401 void *buffer, size_t *lenp, loff_t *ppos)
4405 int state = sysctl_numa_balancing_mode;
4407 if (write && !capable(CAP_SYS_ADMIN))
4412 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4416 sysctl_numa_balancing_mode = state;
4417 __set_numabalancing_state(state);
4424 #ifdef CONFIG_SCHEDSTATS
4426 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4428 static void set_schedstats(bool enabled)
4431 static_branch_enable(&sched_schedstats);
4433 static_branch_disable(&sched_schedstats);
4436 void force_schedstat_enabled(void)
4438 if (!schedstat_enabled()) {
4439 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4440 static_branch_enable(&sched_schedstats);
4444 static int __init setup_schedstats(char *str)
4450 if (!strcmp(str, "enable")) {
4451 set_schedstats(true);
4453 } else if (!strcmp(str, "disable")) {
4454 set_schedstats(false);
4459 pr_warn("Unable to parse schedstats=\n");
4463 __setup("schedstats=", setup_schedstats);
4465 #ifdef CONFIG_PROC_SYSCTL
4466 static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4467 size_t *lenp, loff_t *ppos)
4471 int state = static_branch_likely(&sched_schedstats);
4473 if (write && !capable(CAP_SYS_ADMIN))
4478 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4482 set_schedstats(state);
4485 #endif /* CONFIG_PROC_SYSCTL */
4486 #endif /* CONFIG_SCHEDSTATS */
4488 #ifdef CONFIG_SYSCTL
4489 static struct ctl_table sched_core_sysctls[] = {
4490 #ifdef CONFIG_SCHEDSTATS
4492 .procname = "sched_schedstats",
4494 .maxlen = sizeof(unsigned int),
4496 .proc_handler = sysctl_schedstats,
4497 .extra1 = SYSCTL_ZERO,
4498 .extra2 = SYSCTL_ONE,
4500 #endif /* CONFIG_SCHEDSTATS */
4501 #ifdef CONFIG_UCLAMP_TASK
4503 .procname = "sched_util_clamp_min",
4504 .data = &sysctl_sched_uclamp_util_min,
4505 .maxlen = sizeof(unsigned int),
4507 .proc_handler = sysctl_sched_uclamp_handler,
4510 .procname = "sched_util_clamp_max",
4511 .data = &sysctl_sched_uclamp_util_max,
4512 .maxlen = sizeof(unsigned int),
4514 .proc_handler = sysctl_sched_uclamp_handler,
4517 .procname = "sched_util_clamp_min_rt_default",
4518 .data = &sysctl_sched_uclamp_util_min_rt_default,
4519 .maxlen = sizeof(unsigned int),
4521 .proc_handler = sysctl_sched_uclamp_handler,
4523 #endif /* CONFIG_UCLAMP_TASK */
4526 static int __init sched_core_sysctl_init(void)
4528 register_sysctl_init("kernel", sched_core_sysctls);
4531 late_initcall(sched_core_sysctl_init);
4532 #endif /* CONFIG_SYSCTL */
4535 * fork()/clone()-time setup:
4537 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4539 __sched_fork(clone_flags, p);
4541 * We mark the process as NEW here. This guarantees that
4542 * nobody will actually run it, and a signal or other external
4543 * event cannot wake it up and insert it on the runqueue either.
4545 p->__state = TASK_NEW;
4548 * Make sure we do not leak PI boosting priority to the child.
4550 p->prio = current->normal_prio;
4555 * Revert to default priority/policy on fork if requested.
4557 if (unlikely(p->sched_reset_on_fork)) {
4558 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4559 p->policy = SCHED_NORMAL;
4560 p->static_prio = NICE_TO_PRIO(0);
4562 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4563 p->static_prio = NICE_TO_PRIO(0);
4565 p->prio = p->normal_prio = p->static_prio;
4566 set_load_weight(p, false);
4569 * We don't need the reset flag anymore after the fork. It has
4570 * fulfilled its duty:
4572 p->sched_reset_on_fork = 0;
4575 if (dl_prio(p->prio))
4577 else if (rt_prio(p->prio))
4578 p->sched_class = &rt_sched_class;
4580 p->sched_class = &fair_sched_class;
4582 init_entity_runnable_average(&p->se);
4585 #ifdef CONFIG_SCHED_INFO
4586 if (likely(sched_info_on()))
4587 memset(&p->sched_info, 0, sizeof(p->sched_info));
4589 #if defined(CONFIG_SMP)
4592 init_task_preempt_count(p);
4594 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4595 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4600 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4602 unsigned long flags;
4605 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4606 * required yet, but lockdep gets upset if rules are violated.
4608 raw_spin_lock_irqsave(&p->pi_lock, flags);
4609 #ifdef CONFIG_CGROUP_SCHED
4611 struct task_group *tg;
4612 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4613 struct task_group, css);
4614 tg = autogroup_task_group(p, tg);
4615 p->sched_task_group = tg;
4620 * We're setting the CPU for the first time, we don't migrate,
4621 * so use __set_task_cpu().
4623 __set_task_cpu(p, smp_processor_id());
4624 if (p->sched_class->task_fork)
4625 p->sched_class->task_fork(p);
4626 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4629 void sched_post_fork(struct task_struct *p)
4631 uclamp_post_fork(p);
4634 unsigned long to_ratio(u64 period, u64 runtime)
4636 if (runtime == RUNTIME_INF)
4640 * Doing this here saves a lot of checks in all
4641 * the calling paths, and returning zero seems
4642 * safe for them anyway.
4647 return div64_u64(runtime << BW_SHIFT, period);
4651 * wake_up_new_task - wake up a newly created task for the first time.
4653 * This function will do some initial scheduler statistics housekeeping
4654 * that must be done for every newly created context, then puts the task
4655 * on the runqueue and wakes it.
4657 void wake_up_new_task(struct task_struct *p)
4662 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4663 WRITE_ONCE(p->__state, TASK_RUNNING);
4666 * Fork balancing, do it here and not earlier because:
4667 * - cpus_ptr can change in the fork path
4668 * - any previously selected CPU might disappear through hotplug
4670 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4671 * as we're not fully set-up yet.
4673 p->recent_used_cpu = task_cpu(p);
4675 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4677 rq = __task_rq_lock(p, &rf);
4678 update_rq_clock(rq);
4679 post_init_entity_util_avg(p);
4681 activate_task(rq, p, ENQUEUE_NOCLOCK);
4682 trace_sched_wakeup_new(p);
4683 check_preempt_curr(rq, p, WF_FORK);
4685 if (p->sched_class->task_woken) {
4687 * Nothing relies on rq->lock after this, so it's fine to
4690 rq_unpin_lock(rq, &rf);
4691 p->sched_class->task_woken(rq, p);
4692 rq_repin_lock(rq, &rf);
4695 task_rq_unlock(rq, p, &rf);
4698 #ifdef CONFIG_PREEMPT_NOTIFIERS
4700 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4702 void preempt_notifier_inc(void)
4704 static_branch_inc(&preempt_notifier_key);
4706 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4708 void preempt_notifier_dec(void)
4710 static_branch_dec(&preempt_notifier_key);
4712 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4715 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4716 * @notifier: notifier struct to register
4718 void preempt_notifier_register(struct preempt_notifier *notifier)
4720 if (!static_branch_unlikely(&preempt_notifier_key))
4721 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4723 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4725 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4728 * preempt_notifier_unregister - no longer interested in preemption notifications
4729 * @notifier: notifier struct to unregister
4731 * This is *not* safe to call from within a preemption notifier.
4733 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4735 hlist_del(¬ifier->link);
4737 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4739 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4741 struct preempt_notifier *notifier;
4743 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4744 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4747 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4749 if (static_branch_unlikely(&preempt_notifier_key))
4750 __fire_sched_in_preempt_notifiers(curr);
4754 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4755 struct task_struct *next)
4757 struct preempt_notifier *notifier;
4759 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4760 notifier->ops->sched_out(notifier, next);
4763 static __always_inline void
4764 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4765 struct task_struct *next)
4767 if (static_branch_unlikely(&preempt_notifier_key))
4768 __fire_sched_out_preempt_notifiers(curr, next);
4771 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4773 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4778 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4779 struct task_struct *next)
4783 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4785 static inline void prepare_task(struct task_struct *next)
4789 * Claim the task as running, we do this before switching to it
4790 * such that any running task will have this set.
4792 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4793 * its ordering comment.
4795 WRITE_ONCE(next->on_cpu, 1);
4799 static inline void finish_task(struct task_struct *prev)
4803 * This must be the very last reference to @prev from this CPU. After
4804 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4805 * must ensure this doesn't happen until the switch is completely
4808 * In particular, the load of prev->state in finish_task_switch() must
4809 * happen before this.
4811 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4813 smp_store_release(&prev->on_cpu, 0);
4819 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4821 void (*func)(struct rq *rq);
4822 struct callback_head *next;
4824 lockdep_assert_rq_held(rq);
4827 func = (void (*)(struct rq *))head->func;
4836 static void balance_push(struct rq *rq);
4839 * balance_push_callback is a right abuse of the callback interface and plays
4840 * by significantly different rules.
4842 * Where the normal balance_callback's purpose is to be ran in the same context
4843 * that queued it (only later, when it's safe to drop rq->lock again),
4844 * balance_push_callback is specifically targeted at __schedule().
4846 * This abuse is tolerated because it places all the unlikely/odd cases behind
4847 * a single test, namely: rq->balance_callback == NULL.
4849 struct callback_head balance_push_callback = {
4851 .func = (void (*)(struct callback_head *))balance_push,
4854 static inline struct callback_head *
4855 __splice_balance_callbacks(struct rq *rq, bool split)
4857 struct callback_head *head = rq->balance_callback;
4862 lockdep_assert_rq_held(rq);
4864 * Must not take balance_push_callback off the list when
4865 * splice_balance_callbacks() and balance_callbacks() are not
4866 * in the same rq->lock section.
4868 * In that case it would be possible for __schedule() to interleave
4869 * and observe the list empty.
4871 if (split && head == &balance_push_callback)
4874 rq->balance_callback = NULL;
4879 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4881 return __splice_balance_callbacks(rq, true);
4884 static void __balance_callbacks(struct rq *rq)
4886 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
4889 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4891 unsigned long flags;
4893 if (unlikely(head)) {
4894 raw_spin_rq_lock_irqsave(rq, flags);
4895 do_balance_callbacks(rq, head);
4896 raw_spin_rq_unlock_irqrestore(rq, flags);
4902 static inline void __balance_callbacks(struct rq *rq)
4906 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4911 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4918 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4921 * Since the runqueue lock will be released by the next
4922 * task (which is an invalid locking op but in the case
4923 * of the scheduler it's an obvious special-case), so we
4924 * do an early lockdep release here:
4926 rq_unpin_lock(rq, rf);
4927 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4928 #ifdef CONFIG_DEBUG_SPINLOCK
4929 /* this is a valid case when another task releases the spinlock */
4930 rq_lockp(rq)->owner = next;
4934 static inline void finish_lock_switch(struct rq *rq)
4937 * If we are tracking spinlock dependencies then we have to
4938 * fix up the runqueue lock - which gets 'carried over' from
4939 * prev into current:
4941 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4942 __balance_callbacks(rq);
4943 raw_spin_rq_unlock_irq(rq);
4947 * NOP if the arch has not defined these:
4950 #ifndef prepare_arch_switch
4951 # define prepare_arch_switch(next) do { } while (0)
4954 #ifndef finish_arch_post_lock_switch
4955 # define finish_arch_post_lock_switch() do { } while (0)
4958 static inline void kmap_local_sched_out(void)
4960 #ifdef CONFIG_KMAP_LOCAL
4961 if (unlikely(current->kmap_ctrl.idx))
4962 __kmap_local_sched_out();
4966 static inline void kmap_local_sched_in(void)
4968 #ifdef CONFIG_KMAP_LOCAL
4969 if (unlikely(current->kmap_ctrl.idx))
4970 __kmap_local_sched_in();
4975 * prepare_task_switch - prepare to switch tasks
4976 * @rq: the runqueue preparing to switch
4977 * @prev: the current task that is being switched out
4978 * @next: the task we are going to switch to.
4980 * This is called with the rq lock held and interrupts off. It must
4981 * be paired with a subsequent finish_task_switch after the context
4984 * prepare_task_switch sets up locking and calls architecture specific
4988 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4989 struct task_struct *next)
4991 kcov_prepare_switch(prev);
4992 sched_info_switch(rq, prev, next);
4993 perf_event_task_sched_out(prev, next);
4995 fire_sched_out_preempt_notifiers(prev, next);
4996 kmap_local_sched_out();
4998 prepare_arch_switch(next);
5002 * finish_task_switch - clean up after a task-switch
5003 * @prev: the thread we just switched away from.
5005 * finish_task_switch must be called after the context switch, paired
5006 * with a prepare_task_switch call before the context switch.
5007 * finish_task_switch will reconcile locking set up by prepare_task_switch,
5008 * and do any other architecture-specific cleanup actions.
5010 * Note that we may have delayed dropping an mm in context_switch(). If
5011 * so, we finish that here outside of the runqueue lock. (Doing it
5012 * with the lock held can cause deadlocks; see schedule() for
5015 * The context switch have flipped the stack from under us and restored the
5016 * local variables which were saved when this task called schedule() in the
5017 * past. prev == current is still correct but we need to recalculate this_rq
5018 * because prev may have moved to another CPU.
5020 static struct rq *finish_task_switch(struct task_struct *prev)
5021 __releases(rq->lock)
5023 struct rq *rq = this_rq();
5024 struct mm_struct *mm = rq->prev_mm;
5025 unsigned int prev_state;
5028 * The previous task will have left us with a preempt_count of 2
5029 * because it left us after:
5032 * preempt_disable(); // 1
5034 * raw_spin_lock_irq(&rq->lock) // 2
5036 * Also, see FORK_PREEMPT_COUNT.
5038 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5039 "corrupted preempt_count: %s/%d/0x%x\n",
5040 current->comm, current->pid, preempt_count()))
5041 preempt_count_set(FORK_PREEMPT_COUNT);
5046 * A task struct has one reference for the use as "current".
5047 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5048 * schedule one last time. The schedule call will never return, and
5049 * the scheduled task must drop that reference.
5051 * We must observe prev->state before clearing prev->on_cpu (in
5052 * finish_task), otherwise a concurrent wakeup can get prev
5053 * running on another CPU and we could rave with its RUNNING -> DEAD
5054 * transition, resulting in a double drop.
5056 prev_state = READ_ONCE(prev->__state);
5057 vtime_task_switch(prev);
5058 perf_event_task_sched_in(prev, current);
5060 tick_nohz_task_switch();
5061 finish_lock_switch(rq);
5062 finish_arch_post_lock_switch();
5063 kcov_finish_switch(current);
5065 * kmap_local_sched_out() is invoked with rq::lock held and
5066 * interrupts disabled. There is no requirement for that, but the
5067 * sched out code does not have an interrupt enabled section.
5068 * Restoring the maps on sched in does not require interrupts being
5071 kmap_local_sched_in();
5073 fire_sched_in_preempt_notifiers(current);
5075 * When switching through a kernel thread, the loop in
5076 * membarrier_{private,global}_expedited() may have observed that
5077 * kernel thread and not issued an IPI. It is therefore possible to
5078 * schedule between user->kernel->user threads without passing though
5079 * switch_mm(). Membarrier requires a barrier after storing to
5080 * rq->curr, before returning to userspace, so provide them here:
5082 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5083 * provided by mmdrop(),
5084 * - a sync_core for SYNC_CORE.
5087 membarrier_mm_sync_core_before_usermode(mm);
5090 if (unlikely(prev_state == TASK_DEAD)) {
5091 if (prev->sched_class->task_dead)
5092 prev->sched_class->task_dead(prev);
5094 /* Task is done with its stack. */
5095 put_task_stack(prev);
5097 put_task_struct_rcu_user(prev);
5104 * schedule_tail - first thing a freshly forked thread must call.
5105 * @prev: the thread we just switched away from.
5107 asmlinkage __visible void schedule_tail(struct task_struct *prev)
5108 __releases(rq->lock)
5111 * New tasks start with FORK_PREEMPT_COUNT, see there and
5112 * finish_task_switch() for details.
5114 * finish_task_switch() will drop rq->lock() and lower preempt_count
5115 * and the preempt_enable() will end up enabling preemption (on
5116 * PREEMPT_COUNT kernels).
5119 finish_task_switch(prev);
5122 if (current->set_child_tid)
5123 put_user(task_pid_vnr(current), current->set_child_tid);
5125 calculate_sigpending();
5129 * context_switch - switch to the new MM and the new thread's register state.
5131 static __always_inline struct rq *
5132 context_switch(struct rq *rq, struct task_struct *prev,
5133 struct task_struct *next, struct rq_flags *rf)
5135 prepare_task_switch(rq, prev, next);
5138 * For paravirt, this is coupled with an exit in switch_to to
5139 * combine the page table reload and the switch backend into
5142 arch_start_context_switch(prev);
5145 * kernel -> kernel lazy + transfer active
5146 * user -> kernel lazy + mmgrab() active
5148 * kernel -> user switch + mmdrop() active
5149 * user -> user switch
5151 if (!next->mm) { // to kernel
5152 enter_lazy_tlb(prev->active_mm, next);
5154 next->active_mm = prev->active_mm;
5155 if (prev->mm) // from user
5156 mmgrab(prev->active_mm);
5158 prev->active_mm = NULL;
5160 membarrier_switch_mm(rq, prev->active_mm, next->mm);
5162 * sys_membarrier() requires an smp_mb() between setting
5163 * rq->curr / membarrier_switch_mm() and returning to userspace.
5165 * The below provides this either through switch_mm(), or in
5166 * case 'prev->active_mm == next->mm' through
5167 * finish_task_switch()'s mmdrop().
5169 switch_mm_irqs_off(prev->active_mm, next->mm, next);
5171 if (!prev->mm) { // from kernel
5172 /* will mmdrop() in finish_task_switch(). */
5173 rq->prev_mm = prev->active_mm;
5174 prev->active_mm = NULL;
5178 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5180 prepare_lock_switch(rq, next, rf);
5182 /* Here we just switch the register state and the stack. */
5183 switch_to(prev, next, prev);
5186 return finish_task_switch(prev);
5190 * nr_running and nr_context_switches:
5192 * externally visible scheduler statistics: current number of runnable
5193 * threads, total number of context switches performed since bootup.
5195 unsigned int nr_running(void)
5197 unsigned int i, sum = 0;
5199 for_each_online_cpu(i)
5200 sum += cpu_rq(i)->nr_running;
5206 * Check if only the current task is running on the CPU.
5208 * Caution: this function does not check that the caller has disabled
5209 * preemption, thus the result might have a time-of-check-to-time-of-use
5210 * race. The caller is responsible to use it correctly, for example:
5212 * - from a non-preemptible section (of course)
5214 * - from a thread that is bound to a single CPU
5216 * - in a loop with very short iterations (e.g. a polling loop)
5218 bool single_task_running(void)
5220 return raw_rq()->nr_running == 1;
5222 EXPORT_SYMBOL(single_task_running);
5224 unsigned long long nr_context_switches(void)
5227 unsigned long long sum = 0;
5229 for_each_possible_cpu(i)
5230 sum += cpu_rq(i)->nr_switches;
5236 * Consumers of these two interfaces, like for example the cpuidle menu
5237 * governor, are using nonsensical data. Preferring shallow idle state selection
5238 * for a CPU that has IO-wait which might not even end up running the task when
5239 * it does become runnable.
5242 unsigned int nr_iowait_cpu(int cpu)
5244 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5248 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5250 * The idea behind IO-wait account is to account the idle time that we could
5251 * have spend running if it were not for IO. That is, if we were to improve the
5252 * storage performance, we'd have a proportional reduction in IO-wait time.
5254 * This all works nicely on UP, where, when a task blocks on IO, we account
5255 * idle time as IO-wait, because if the storage were faster, it could've been
5256 * running and we'd not be idle.
5258 * This has been extended to SMP, by doing the same for each CPU. This however
5261 * Imagine for instance the case where two tasks block on one CPU, only the one
5262 * CPU will have IO-wait accounted, while the other has regular idle. Even
5263 * though, if the storage were faster, both could've ran at the same time,
5264 * utilising both CPUs.
5266 * This means, that when looking globally, the current IO-wait accounting on
5267 * SMP is a lower bound, by reason of under accounting.
5269 * Worse, since the numbers are provided per CPU, they are sometimes
5270 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5271 * associated with any one particular CPU, it can wake to another CPU than it
5272 * blocked on. This means the per CPU IO-wait number is meaningless.
5274 * Task CPU affinities can make all that even more 'interesting'.
5277 unsigned int nr_iowait(void)
5279 unsigned int i, sum = 0;
5281 for_each_possible_cpu(i)
5282 sum += nr_iowait_cpu(i);
5290 * sched_exec - execve() is a valuable balancing opportunity, because at
5291 * this point the task has the smallest effective memory and cache footprint.
5293 void sched_exec(void)
5295 struct task_struct *p = current;
5296 unsigned long flags;
5299 raw_spin_lock_irqsave(&p->pi_lock, flags);
5300 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5301 if (dest_cpu == smp_processor_id())
5304 if (likely(cpu_active(dest_cpu))) {
5305 struct migration_arg arg = { p, dest_cpu };
5307 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5308 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5312 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5317 DEFINE_PER_CPU(struct kernel_stat, kstat);
5318 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5320 EXPORT_PER_CPU_SYMBOL(kstat);
5321 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5324 * The function fair_sched_class.update_curr accesses the struct curr
5325 * and its field curr->exec_start; when called from task_sched_runtime(),
5326 * we observe a high rate of cache misses in practice.
5327 * Prefetching this data results in improved performance.
5329 static inline void prefetch_curr_exec_start(struct task_struct *p)
5331 #ifdef CONFIG_FAIR_GROUP_SCHED
5332 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5334 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5337 prefetch(&curr->exec_start);
5341 * Return accounted runtime for the task.
5342 * In case the task is currently running, return the runtime plus current's
5343 * pending runtime that have not been accounted yet.
5345 unsigned long long task_sched_runtime(struct task_struct *p)
5351 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5353 * 64-bit doesn't need locks to atomically read a 64-bit value.
5354 * So we have a optimization chance when the task's delta_exec is 0.
5355 * Reading ->on_cpu is racy, but this is ok.
5357 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5358 * If we race with it entering CPU, unaccounted time is 0. This is
5359 * indistinguishable from the read occurring a few cycles earlier.
5360 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5361 * been accounted, so we're correct here as well.
5363 if (!p->on_cpu || !task_on_rq_queued(p))
5364 return p->se.sum_exec_runtime;
5367 rq = task_rq_lock(p, &rf);
5369 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5370 * project cycles that may never be accounted to this
5371 * thread, breaking clock_gettime().
5373 if (task_current(rq, p) && task_on_rq_queued(p)) {
5374 prefetch_curr_exec_start(p);
5375 update_rq_clock(rq);
5376 p->sched_class->update_curr(rq);
5378 ns = p->se.sum_exec_runtime;
5379 task_rq_unlock(rq, p, &rf);
5384 #ifdef CONFIG_SCHED_DEBUG
5385 static u64 cpu_resched_latency(struct rq *rq)
5387 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5388 u64 resched_latency, now = rq_clock(rq);
5389 static bool warned_once;
5391 if (sysctl_resched_latency_warn_once && warned_once)
5394 if (!need_resched() || !latency_warn_ms)
5397 if (system_state == SYSTEM_BOOTING)
5400 if (!rq->last_seen_need_resched_ns) {
5401 rq->last_seen_need_resched_ns = now;
5402 rq->ticks_without_resched = 0;
5406 rq->ticks_without_resched++;
5407 resched_latency = now - rq->last_seen_need_resched_ns;
5408 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5413 return resched_latency;
5416 static int __init setup_resched_latency_warn_ms(char *str)
5420 if ((kstrtol(str, 0, &val))) {
5421 pr_warn("Unable to set resched_latency_warn_ms\n");
5425 sysctl_resched_latency_warn_ms = val;
5428 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5430 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5431 #endif /* CONFIG_SCHED_DEBUG */
5434 * This function gets called by the timer code, with HZ frequency.
5435 * We call it with interrupts disabled.
5437 void scheduler_tick(void)
5439 int cpu = smp_processor_id();
5440 struct rq *rq = cpu_rq(cpu);
5441 struct task_struct *curr = rq->curr;
5443 unsigned long thermal_pressure;
5444 u64 resched_latency;
5446 arch_scale_freq_tick();
5451 update_rq_clock(rq);
5452 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5453 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5454 curr->sched_class->task_tick(rq, curr, 0);
5455 if (sched_feat(LATENCY_WARN))
5456 resched_latency = cpu_resched_latency(rq);
5457 calc_global_load_tick(rq);
5458 sched_core_tick(rq);
5462 if (sched_feat(LATENCY_WARN) && resched_latency)
5463 resched_latency_warn(cpu, resched_latency);
5465 perf_event_task_tick();
5468 rq->idle_balance = idle_cpu(cpu);
5469 trigger_load_balance(rq);
5473 #ifdef CONFIG_NO_HZ_FULL
5478 struct delayed_work work;
5480 /* Values for ->state, see diagram below. */
5481 #define TICK_SCHED_REMOTE_OFFLINE 0
5482 #define TICK_SCHED_REMOTE_OFFLINING 1
5483 #define TICK_SCHED_REMOTE_RUNNING 2
5486 * State diagram for ->state:
5489 * TICK_SCHED_REMOTE_OFFLINE
5492 * | | sched_tick_remote()
5495 * +--TICK_SCHED_REMOTE_OFFLINING
5498 * sched_tick_start() | | sched_tick_stop()
5501 * TICK_SCHED_REMOTE_RUNNING
5504 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5505 * and sched_tick_start() are happy to leave the state in RUNNING.
5508 static struct tick_work __percpu *tick_work_cpu;
5510 static void sched_tick_remote(struct work_struct *work)
5512 struct delayed_work *dwork = to_delayed_work(work);
5513 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5514 int cpu = twork->cpu;
5515 struct rq *rq = cpu_rq(cpu);
5516 struct task_struct *curr;
5522 * Handle the tick only if it appears the remote CPU is running in full
5523 * dynticks mode. The check is racy by nature, but missing a tick or
5524 * having one too much is no big deal because the scheduler tick updates
5525 * statistics and checks timeslices in a time-independent way, regardless
5526 * of when exactly it is running.
5528 if (!tick_nohz_tick_stopped_cpu(cpu))
5531 rq_lock_irq(rq, &rf);
5533 if (cpu_is_offline(cpu))
5536 update_rq_clock(rq);
5538 if (!is_idle_task(curr)) {
5540 * Make sure the next tick runs within a reasonable
5543 delta = rq_clock_task(rq) - curr->se.exec_start;
5544 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5546 curr->sched_class->task_tick(rq, curr, 0);
5548 calc_load_nohz_remote(rq);
5550 rq_unlock_irq(rq, &rf);
5554 * Run the remote tick once per second (1Hz). This arbitrary
5555 * frequency is large enough to avoid overload but short enough
5556 * to keep scheduler internal stats reasonably up to date. But
5557 * first update state to reflect hotplug activity if required.
5559 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5560 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5561 if (os == TICK_SCHED_REMOTE_RUNNING)
5562 queue_delayed_work(system_unbound_wq, dwork, HZ);
5565 static void sched_tick_start(int cpu)
5568 struct tick_work *twork;
5570 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5573 WARN_ON_ONCE(!tick_work_cpu);
5575 twork = per_cpu_ptr(tick_work_cpu, cpu);
5576 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5577 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5578 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5580 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5581 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5585 #ifdef CONFIG_HOTPLUG_CPU
5586 static void sched_tick_stop(int cpu)
5588 struct tick_work *twork;
5591 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5594 WARN_ON_ONCE(!tick_work_cpu);
5596 twork = per_cpu_ptr(tick_work_cpu, cpu);
5597 /* There cannot be competing actions, but don't rely on stop-machine. */
5598 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5599 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5600 /* Don't cancel, as this would mess up the state machine. */
5602 #endif /* CONFIG_HOTPLUG_CPU */
5604 int __init sched_tick_offload_init(void)
5606 tick_work_cpu = alloc_percpu(struct tick_work);
5607 BUG_ON(!tick_work_cpu);
5611 #else /* !CONFIG_NO_HZ_FULL */
5612 static inline void sched_tick_start(int cpu) { }
5613 static inline void sched_tick_stop(int cpu) { }
5616 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5617 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5619 * If the value passed in is equal to the current preempt count
5620 * then we just disabled preemption. Start timing the latency.
5622 static inline void preempt_latency_start(int val)
5624 if (preempt_count() == val) {
5625 unsigned long ip = get_lock_parent_ip();
5626 #ifdef CONFIG_DEBUG_PREEMPT
5627 current->preempt_disable_ip = ip;
5629 trace_preempt_off(CALLER_ADDR0, ip);
5633 void preempt_count_add(int val)
5635 #ifdef CONFIG_DEBUG_PREEMPT
5639 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5642 __preempt_count_add(val);
5643 #ifdef CONFIG_DEBUG_PREEMPT
5645 * Spinlock count overflowing soon?
5647 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5650 preempt_latency_start(val);
5652 EXPORT_SYMBOL(preempt_count_add);
5653 NOKPROBE_SYMBOL(preempt_count_add);
5656 * If the value passed in equals to the current preempt count
5657 * then we just enabled preemption. Stop timing the latency.
5659 static inline void preempt_latency_stop(int val)
5661 if (preempt_count() == val)
5662 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5665 void preempt_count_sub(int val)
5667 #ifdef CONFIG_DEBUG_PREEMPT
5671 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5674 * Is the spinlock portion underflowing?
5676 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5677 !(preempt_count() & PREEMPT_MASK)))
5681 preempt_latency_stop(val);
5682 __preempt_count_sub(val);
5684 EXPORT_SYMBOL(preempt_count_sub);
5685 NOKPROBE_SYMBOL(preempt_count_sub);
5688 static inline void preempt_latency_start(int val) { }
5689 static inline void preempt_latency_stop(int val) { }
5692 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5694 #ifdef CONFIG_DEBUG_PREEMPT
5695 return p->preempt_disable_ip;
5702 * Print scheduling while atomic bug:
5704 static noinline void __schedule_bug(struct task_struct *prev)
5706 /* Save this before calling printk(), since that will clobber it */
5707 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5709 if (oops_in_progress)
5712 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5713 prev->comm, prev->pid, preempt_count());
5715 debug_show_held_locks(prev);
5717 if (irqs_disabled())
5718 print_irqtrace_events(prev);
5719 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5720 && in_atomic_preempt_off()) {
5721 pr_err("Preemption disabled at:");
5722 print_ip_sym(KERN_ERR, preempt_disable_ip);
5725 panic("scheduling while atomic\n");
5728 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5732 * Various schedule()-time debugging checks and statistics:
5734 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5736 #ifdef CONFIG_SCHED_STACK_END_CHECK
5737 if (task_stack_end_corrupted(prev))
5738 panic("corrupted stack end detected inside scheduler\n");
5740 if (task_scs_end_corrupted(prev))
5741 panic("corrupted shadow stack detected inside scheduler\n");
5744 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5745 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5746 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5747 prev->comm, prev->pid, prev->non_block_count);
5749 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5753 if (unlikely(in_atomic_preempt_off())) {
5754 __schedule_bug(prev);
5755 preempt_count_set(PREEMPT_DISABLED);
5758 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5760 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5762 schedstat_inc(this_rq()->sched_count);
5765 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5766 struct rq_flags *rf)
5769 const struct sched_class *class;
5771 * We must do the balancing pass before put_prev_task(), such
5772 * that when we release the rq->lock the task is in the same
5773 * state as before we took rq->lock.
5775 * We can terminate the balance pass as soon as we know there is
5776 * a runnable task of @class priority or higher.
5778 for_class_range(class, prev->sched_class, &idle_sched_class) {
5779 if (class->balance(rq, prev, rf))
5784 put_prev_task(rq, prev);
5788 * Pick up the highest-prio task:
5790 static inline struct task_struct *
5791 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5793 const struct sched_class *class;
5794 struct task_struct *p;
5797 * Optimization: we know that if all tasks are in the fair class we can
5798 * call that function directly, but only if the @prev task wasn't of a
5799 * higher scheduling class, because otherwise those lose the
5800 * opportunity to pull in more work from other CPUs.
5802 if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
5803 rq->nr_running == rq->cfs.h_nr_running)) {
5805 p = pick_next_task_fair(rq, prev, rf);
5806 if (unlikely(p == RETRY_TASK))
5809 /* Assume the next prioritized class is idle_sched_class */
5811 put_prev_task(rq, prev);
5812 p = pick_next_task_idle(rq);
5819 put_prev_task_balance(rq, prev, rf);
5821 for_each_class(class) {
5822 p = class->pick_next_task(rq);
5827 BUG(); /* The idle class should always have a runnable task. */
5830 #ifdef CONFIG_SCHED_CORE
5831 static inline bool is_task_rq_idle(struct task_struct *t)
5833 return (task_rq(t)->idle == t);
5836 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5838 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5841 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5843 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5846 return a->core_cookie == b->core_cookie;
5849 static inline struct task_struct *pick_task(struct rq *rq)
5851 const struct sched_class *class;
5852 struct task_struct *p;
5854 for_each_class(class) {
5855 p = class->pick_task(rq);
5860 BUG(); /* The idle class should always have a runnable task. */
5863 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5865 static void queue_core_balance(struct rq *rq);
5867 static struct task_struct *
5868 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5870 struct task_struct *next, *p, *max = NULL;
5871 const struct cpumask *smt_mask;
5872 bool fi_before = false;
5873 bool core_clock_updated = (rq == rq->core);
5874 unsigned long cookie;
5875 int i, cpu, occ = 0;
5879 if (!sched_core_enabled(rq))
5880 return __pick_next_task(rq, prev, rf);
5884 /* Stopper task is switching into idle, no need core-wide selection. */
5885 if (cpu_is_offline(cpu)) {
5887 * Reset core_pick so that we don't enter the fastpath when
5888 * coming online. core_pick would already be migrated to
5889 * another cpu during offline.
5891 rq->core_pick = NULL;
5892 return __pick_next_task(rq, prev, rf);
5896 * If there were no {en,de}queues since we picked (IOW, the task
5897 * pointers are all still valid), and we haven't scheduled the last
5898 * pick yet, do so now.
5900 * rq->core_pick can be NULL if no selection was made for a CPU because
5901 * it was either offline or went offline during a sibling's core-wide
5902 * selection. In this case, do a core-wide selection.
5904 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5905 rq->core->core_pick_seq != rq->core_sched_seq &&
5907 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5909 next = rq->core_pick;
5911 put_prev_task(rq, prev);
5912 set_next_task(rq, next);
5915 rq->core_pick = NULL;
5919 put_prev_task_balance(rq, prev, rf);
5921 smt_mask = cpu_smt_mask(cpu);
5922 need_sync = !!rq->core->core_cookie;
5925 rq->core->core_cookie = 0UL;
5926 if (rq->core->core_forceidle_count) {
5927 if (!core_clock_updated) {
5928 update_rq_clock(rq->core);
5929 core_clock_updated = true;
5931 sched_core_account_forceidle(rq);
5932 /* reset after accounting force idle */
5933 rq->core->core_forceidle_start = 0;
5934 rq->core->core_forceidle_count = 0;
5935 rq->core->core_forceidle_occupation = 0;
5941 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5943 * @task_seq guards the task state ({en,de}queues)
5944 * @pick_seq is the @task_seq we did a selection on
5945 * @sched_seq is the @pick_seq we scheduled
5947 * However, preemptions can cause multiple picks on the same task set.
5948 * 'Fix' this by also increasing @task_seq for every pick.
5950 rq->core->core_task_seq++;
5953 * Optimize for common case where this CPU has no cookies
5954 * and there are no cookied tasks running on siblings.
5957 next = pick_task(rq);
5958 if (!next->core_cookie) {
5959 rq->core_pick = NULL;
5961 * For robustness, update the min_vruntime_fi for
5962 * unconstrained picks as well.
5964 WARN_ON_ONCE(fi_before);
5965 task_vruntime_update(rq, next, false);
5971 * For each thread: do the regular task pick and find the max prio task
5974 * Tie-break prio towards the current CPU
5976 for_each_cpu_wrap(i, smt_mask, cpu) {
5980 * Current cpu always has its clock updated on entrance to
5981 * pick_next_task(). If the current cpu is not the core,
5982 * the core may also have been updated above.
5984 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
5985 update_rq_clock(rq_i);
5987 p = rq_i->core_pick = pick_task(rq_i);
5988 if (!max || prio_less(max, p, fi_before))
5992 cookie = rq->core->core_cookie = max->core_cookie;
5995 * For each thread: try and find a runnable task that matches @max or
5998 for_each_cpu(i, smt_mask) {
6000 p = rq_i->core_pick;
6002 if (!cookie_equals(p, cookie)) {
6005 p = sched_core_find(rq_i, cookie);
6007 p = idle_sched_class.pick_task(rq_i);
6010 rq_i->core_pick = p;
6012 if (p == rq_i->idle) {
6013 if (rq_i->nr_running) {
6014 rq->core->core_forceidle_count++;
6016 rq->core->core_forceidle_seq++;
6023 if (schedstat_enabled() && rq->core->core_forceidle_count) {
6024 rq->core->core_forceidle_start = rq_clock(rq->core);
6025 rq->core->core_forceidle_occupation = occ;
6028 rq->core->core_pick_seq = rq->core->core_task_seq;
6029 next = rq->core_pick;
6030 rq->core_sched_seq = rq->core->core_pick_seq;
6032 /* Something should have been selected for current CPU */
6033 WARN_ON_ONCE(!next);
6036 * Reschedule siblings
6038 * NOTE: L1TF -- at this point we're no longer running the old task and
6039 * sending an IPI (below) ensures the sibling will no longer be running
6040 * their task. This ensures there is no inter-sibling overlap between
6041 * non-matching user state.
6043 for_each_cpu(i, smt_mask) {
6047 * An online sibling might have gone offline before a task
6048 * could be picked for it, or it might be offline but later
6049 * happen to come online, but its too late and nothing was
6050 * picked for it. That's Ok - it will pick tasks for itself,
6053 if (!rq_i->core_pick)
6057 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6058 * fi_before fi update?
6064 if (!(fi_before && rq->core->core_forceidle_count))
6065 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
6067 rq_i->core_pick->core_occupation = occ;
6070 rq_i->core_pick = NULL;
6074 /* Did we break L1TF mitigation requirements? */
6075 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6077 if (rq_i->curr == rq_i->core_pick) {
6078 rq_i->core_pick = NULL;
6086 set_next_task(rq, next);
6088 if (rq->core->core_forceidle_count && next == rq->idle)
6089 queue_core_balance(rq);
6094 static bool try_steal_cookie(int this, int that)
6096 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6097 struct task_struct *p;
6098 unsigned long cookie;
6099 bool success = false;
6101 local_irq_disable();
6102 double_rq_lock(dst, src);
6104 cookie = dst->core->core_cookie;
6108 if (dst->curr != dst->idle)
6111 p = sched_core_find(src, cookie);
6116 if (p == src->core_pick || p == src->curr)
6119 if (!is_cpu_allowed(p, this))
6122 if (p->core_occupation > dst->idle->core_occupation)
6125 deactivate_task(src, p, 0);
6126 set_task_cpu(p, this);
6127 activate_task(dst, p, 0);
6135 p = sched_core_next(p, cookie);
6139 double_rq_unlock(dst, src);
6145 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6149 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
6156 if (try_steal_cookie(cpu, i))
6163 static void sched_core_balance(struct rq *rq)
6165 struct sched_domain *sd;
6166 int cpu = cpu_of(rq);
6170 raw_spin_rq_unlock_irq(rq);
6171 for_each_domain(cpu, sd) {
6175 if (steal_cookie_task(cpu, sd))
6178 raw_spin_rq_lock_irq(rq);
6183 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
6185 static void queue_core_balance(struct rq *rq)
6187 if (!sched_core_enabled(rq))
6190 if (!rq->core->core_cookie)
6193 if (!rq->nr_running) /* not forced idle */
6196 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6199 static void sched_core_cpu_starting(unsigned int cpu)
6201 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6202 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6203 unsigned long flags;
6206 sched_core_lock(cpu, &flags);
6208 WARN_ON_ONCE(rq->core != rq);
6210 /* if we're the first, we'll be our own leader */
6211 if (cpumask_weight(smt_mask) == 1)
6214 /* find the leader */
6215 for_each_cpu(t, smt_mask) {
6219 if (rq->core == rq) {
6225 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6228 /* install and validate core_rq */
6229 for_each_cpu(t, smt_mask) {
6235 WARN_ON_ONCE(rq->core != core_rq);
6239 sched_core_unlock(cpu, &flags);
6242 static void sched_core_cpu_deactivate(unsigned int cpu)
6244 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6245 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6246 unsigned long flags;
6249 sched_core_lock(cpu, &flags);
6251 /* if we're the last man standing, nothing to do */
6252 if (cpumask_weight(smt_mask) == 1) {
6253 WARN_ON_ONCE(rq->core != rq);
6257 /* if we're not the leader, nothing to do */
6261 /* find a new leader */
6262 for_each_cpu(t, smt_mask) {
6265 core_rq = cpu_rq(t);
6269 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6272 /* copy the shared state to the new leader */
6273 core_rq->core_task_seq = rq->core_task_seq;
6274 core_rq->core_pick_seq = rq->core_pick_seq;
6275 core_rq->core_cookie = rq->core_cookie;
6276 core_rq->core_forceidle_count = rq->core_forceidle_count;
6277 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6278 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6281 * Accounting edge for forced idle is handled in pick_next_task().
6282 * Don't need another one here, since the hotplug thread shouldn't
6285 core_rq->core_forceidle_start = 0;
6287 /* install new leader */
6288 for_each_cpu(t, smt_mask) {
6294 sched_core_unlock(cpu, &flags);
6297 static inline void sched_core_cpu_dying(unsigned int cpu)
6299 struct rq *rq = cpu_rq(cpu);
6305 #else /* !CONFIG_SCHED_CORE */
6307 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6308 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6309 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6311 static struct task_struct *
6312 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6314 return __pick_next_task(rq, prev, rf);
6317 #endif /* CONFIG_SCHED_CORE */
6320 * Constants for the sched_mode argument of __schedule().
6322 * The mode argument allows RT enabled kernels to differentiate a
6323 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6324 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6325 * optimize the AND operation out and just check for zero.
6328 #define SM_PREEMPT 0x1
6329 #define SM_RTLOCK_WAIT 0x2
6331 #ifndef CONFIG_PREEMPT_RT
6332 # define SM_MASK_PREEMPT (~0U)
6334 # define SM_MASK_PREEMPT SM_PREEMPT
6338 * __schedule() is the main scheduler function.
6340 * The main means of driving the scheduler and thus entering this function are:
6342 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6344 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6345 * paths. For example, see arch/x86/entry_64.S.
6347 * To drive preemption between tasks, the scheduler sets the flag in timer
6348 * interrupt handler scheduler_tick().
6350 * 3. Wakeups don't really cause entry into schedule(). They add a
6351 * task to the run-queue and that's it.
6353 * Now, if the new task added to the run-queue preempts the current
6354 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6355 * called on the nearest possible occasion:
6357 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6359 * - in syscall or exception context, at the next outmost
6360 * preempt_enable(). (this might be as soon as the wake_up()'s
6363 * - in IRQ context, return from interrupt-handler to
6364 * preemptible context
6366 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6369 * - cond_resched() call
6370 * - explicit schedule() call
6371 * - return from syscall or exception to user-space
6372 * - return from interrupt-handler to user-space
6374 * WARNING: must be called with preemption disabled!
6376 static void __sched notrace __schedule(unsigned int sched_mode)
6378 struct task_struct *prev, *next;
6379 unsigned long *switch_count;
6380 unsigned long prev_state;
6385 cpu = smp_processor_id();
6389 schedule_debug(prev, !!sched_mode);
6391 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6394 local_irq_disable();
6395 rcu_note_context_switch(!!sched_mode);
6398 * Make sure that signal_pending_state()->signal_pending() below
6399 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6400 * done by the caller to avoid the race with signal_wake_up():
6402 * __set_current_state(@state) signal_wake_up()
6403 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6404 * wake_up_state(p, state)
6405 * LOCK rq->lock LOCK p->pi_state
6406 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6407 * if (signal_pending_state()) if (p->state & @state)
6409 * Also, the membarrier system call requires a full memory barrier
6410 * after coming from user-space, before storing to rq->curr.
6413 smp_mb__after_spinlock();
6415 /* Promote REQ to ACT */
6416 rq->clock_update_flags <<= 1;
6417 update_rq_clock(rq);
6419 switch_count = &prev->nivcsw;
6422 * We must load prev->state once (task_struct::state is volatile), such
6423 * that we form a control dependency vs deactivate_task() below.
6425 prev_state = READ_ONCE(prev->__state);
6426 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6427 if (signal_pending_state(prev_state, prev)) {
6428 WRITE_ONCE(prev->__state, TASK_RUNNING);
6430 prev->sched_contributes_to_load =
6431 (prev_state & TASK_UNINTERRUPTIBLE) &&
6432 !(prev_state & TASK_NOLOAD) &&
6433 !(prev->flags & PF_FROZEN);
6435 if (prev->sched_contributes_to_load)
6436 rq->nr_uninterruptible++;
6439 * __schedule() ttwu()
6440 * prev_state = prev->state; if (p->on_rq && ...)
6441 * if (prev_state) goto out;
6442 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6443 * p->state = TASK_WAKING
6445 * Where __schedule() and ttwu() have matching control dependencies.
6447 * After this, schedule() must not care about p->state any more.
6449 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6451 if (prev->in_iowait) {
6452 atomic_inc(&rq->nr_iowait);
6453 delayacct_blkio_start();
6456 switch_count = &prev->nvcsw;
6459 next = pick_next_task(rq, prev, &rf);
6460 clear_tsk_need_resched(prev);
6461 clear_preempt_need_resched();
6462 #ifdef CONFIG_SCHED_DEBUG
6463 rq->last_seen_need_resched_ns = 0;
6466 if (likely(prev != next)) {
6469 * RCU users of rcu_dereference(rq->curr) may not see
6470 * changes to task_struct made by pick_next_task().
6472 RCU_INIT_POINTER(rq->curr, next);
6474 * The membarrier system call requires each architecture
6475 * to have a full memory barrier after updating
6476 * rq->curr, before returning to user-space.
6478 * Here are the schemes providing that barrier on the
6479 * various architectures:
6480 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6481 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6482 * - finish_lock_switch() for weakly-ordered
6483 * architectures where spin_unlock is a full barrier,
6484 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6485 * is a RELEASE barrier),
6489 migrate_disable_switch(rq, prev);
6490 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6492 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
6494 /* Also unlocks the rq: */
6495 rq = context_switch(rq, prev, next, &rf);
6497 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6499 rq_unpin_lock(rq, &rf);
6500 __balance_callbacks(rq);
6501 raw_spin_rq_unlock_irq(rq);
6505 void __noreturn do_task_dead(void)
6507 /* Causes final put_task_struct in finish_task_switch(): */
6508 set_special_state(TASK_DEAD);
6510 /* Tell freezer to ignore us: */
6511 current->flags |= PF_NOFREEZE;
6513 __schedule(SM_NONE);
6516 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6521 static inline void sched_submit_work(struct task_struct *tsk)
6523 unsigned int task_flags;
6525 if (task_is_running(tsk))
6528 task_flags = tsk->flags;
6530 * If a worker goes to sleep, notify and ask workqueue whether it
6531 * wants to wake up a task to maintain concurrency.
6533 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6534 if (task_flags & PF_WQ_WORKER)
6535 wq_worker_sleeping(tsk);
6537 io_wq_worker_sleeping(tsk);
6541 * spinlock and rwlock must not flush block requests. This will
6542 * deadlock if the callback attempts to acquire a lock which is
6545 SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6548 * If we are going to sleep and we have plugged IO queued,
6549 * make sure to submit it to avoid deadlocks.
6551 blk_flush_plug(tsk->plug, true);
6554 static void sched_update_worker(struct task_struct *tsk)
6556 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6557 if (tsk->flags & PF_WQ_WORKER)
6558 wq_worker_running(tsk);
6560 io_wq_worker_running(tsk);
6564 asmlinkage __visible void __sched schedule(void)
6566 struct task_struct *tsk = current;
6568 sched_submit_work(tsk);
6571 __schedule(SM_NONE);
6572 sched_preempt_enable_no_resched();
6573 } while (need_resched());
6574 sched_update_worker(tsk);
6576 EXPORT_SYMBOL(schedule);
6579 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6580 * state (have scheduled out non-voluntarily) by making sure that all
6581 * tasks have either left the run queue or have gone into user space.
6582 * As idle tasks do not do either, they must not ever be preempted
6583 * (schedule out non-voluntarily).
6585 * schedule_idle() is similar to schedule_preempt_disable() except that it
6586 * never enables preemption because it does not call sched_submit_work().
6588 void __sched schedule_idle(void)
6591 * As this skips calling sched_submit_work(), which the idle task does
6592 * regardless because that function is a nop when the task is in a
6593 * TASK_RUNNING state, make sure this isn't used someplace that the
6594 * current task can be in any other state. Note, idle is always in the
6595 * TASK_RUNNING state.
6597 WARN_ON_ONCE(current->__state);
6599 __schedule(SM_NONE);
6600 } while (need_resched());
6603 #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6604 asmlinkage __visible void __sched schedule_user(void)
6607 * If we come here after a random call to set_need_resched(),
6608 * or we have been woken up remotely but the IPI has not yet arrived,
6609 * we haven't yet exited the RCU idle mode. Do it here manually until
6610 * we find a better solution.
6612 * NB: There are buggy callers of this function. Ideally we
6613 * should warn if prev_state != CONTEXT_USER, but that will trigger
6614 * too frequently to make sense yet.
6616 enum ctx_state prev_state = exception_enter();
6618 exception_exit(prev_state);
6623 * schedule_preempt_disabled - called with preemption disabled
6625 * Returns with preemption disabled. Note: preempt_count must be 1
6627 void __sched schedule_preempt_disabled(void)
6629 sched_preempt_enable_no_resched();
6634 #ifdef CONFIG_PREEMPT_RT
6635 void __sched notrace schedule_rtlock(void)
6639 __schedule(SM_RTLOCK_WAIT);
6640 sched_preempt_enable_no_resched();
6641 } while (need_resched());
6643 NOKPROBE_SYMBOL(schedule_rtlock);
6646 static void __sched notrace preempt_schedule_common(void)
6650 * Because the function tracer can trace preempt_count_sub()
6651 * and it also uses preempt_enable/disable_notrace(), if
6652 * NEED_RESCHED is set, the preempt_enable_notrace() called
6653 * by the function tracer will call this function again and
6654 * cause infinite recursion.
6656 * Preemption must be disabled here before the function
6657 * tracer can trace. Break up preempt_disable() into two
6658 * calls. One to disable preemption without fear of being
6659 * traced. The other to still record the preemption latency,
6660 * which can also be traced by the function tracer.
6662 preempt_disable_notrace();
6663 preempt_latency_start(1);
6664 __schedule(SM_PREEMPT);
6665 preempt_latency_stop(1);
6666 preempt_enable_no_resched_notrace();
6669 * Check again in case we missed a preemption opportunity
6670 * between schedule and now.
6672 } while (need_resched());
6675 #ifdef CONFIG_PREEMPTION
6677 * This is the entry point to schedule() from in-kernel preemption
6678 * off of preempt_enable.
6680 asmlinkage __visible void __sched notrace preempt_schedule(void)
6683 * If there is a non-zero preempt_count or interrupts are disabled,
6684 * we do not want to preempt the current task. Just return..
6686 if (likely(!preemptible()))
6688 preempt_schedule_common();
6690 NOKPROBE_SYMBOL(preempt_schedule);
6691 EXPORT_SYMBOL(preempt_schedule);
6693 #ifdef CONFIG_PREEMPT_DYNAMIC
6694 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6695 #ifndef preempt_schedule_dynamic_enabled
6696 #define preempt_schedule_dynamic_enabled preempt_schedule
6697 #define preempt_schedule_dynamic_disabled NULL
6699 DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
6700 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6701 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6702 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6703 void __sched notrace dynamic_preempt_schedule(void)
6705 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6709 NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6710 EXPORT_SYMBOL(dynamic_preempt_schedule);
6715 * preempt_schedule_notrace - preempt_schedule called by tracing
6717 * The tracing infrastructure uses preempt_enable_notrace to prevent
6718 * recursion and tracing preempt enabling caused by the tracing
6719 * infrastructure itself. But as tracing can happen in areas coming
6720 * from userspace or just about to enter userspace, a preempt enable
6721 * can occur before user_exit() is called. This will cause the scheduler
6722 * to be called when the system is still in usermode.
6724 * To prevent this, the preempt_enable_notrace will use this function
6725 * instead of preempt_schedule() to exit user context if needed before
6726 * calling the scheduler.
6728 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6730 enum ctx_state prev_ctx;
6732 if (likely(!preemptible()))
6737 * Because the function tracer can trace preempt_count_sub()
6738 * and it also uses preempt_enable/disable_notrace(), if
6739 * NEED_RESCHED is set, the preempt_enable_notrace() called
6740 * by the function tracer will call this function again and
6741 * cause infinite recursion.
6743 * Preemption must be disabled here before the function
6744 * tracer can trace. Break up preempt_disable() into two
6745 * calls. One to disable preemption without fear of being
6746 * traced. The other to still record the preemption latency,
6747 * which can also be traced by the function tracer.
6749 preempt_disable_notrace();
6750 preempt_latency_start(1);
6752 * Needs preempt disabled in case user_exit() is traced
6753 * and the tracer calls preempt_enable_notrace() causing
6754 * an infinite recursion.
6756 prev_ctx = exception_enter();
6757 __schedule(SM_PREEMPT);
6758 exception_exit(prev_ctx);
6760 preempt_latency_stop(1);
6761 preempt_enable_no_resched_notrace();
6762 } while (need_resched());
6764 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6766 #ifdef CONFIG_PREEMPT_DYNAMIC
6767 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6768 #ifndef preempt_schedule_notrace_dynamic_enabled
6769 #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6770 #define preempt_schedule_notrace_dynamic_disabled NULL
6772 DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
6773 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6774 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6775 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6776 void __sched notrace dynamic_preempt_schedule_notrace(void)
6778 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6780 preempt_schedule_notrace();
6782 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6783 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6787 #endif /* CONFIG_PREEMPTION */
6790 * This is the entry point to schedule() from kernel preemption
6791 * off of irq context.
6792 * Note, that this is called and return with irqs disabled. This will
6793 * protect us against recursive calling from irq.
6795 asmlinkage __visible void __sched preempt_schedule_irq(void)
6797 enum ctx_state prev_state;
6799 /* Catch callers which need to be fixed */
6800 BUG_ON(preempt_count() || !irqs_disabled());
6802 prev_state = exception_enter();
6807 __schedule(SM_PREEMPT);
6808 local_irq_disable();
6809 sched_preempt_enable_no_resched();
6810 } while (need_resched());
6812 exception_exit(prev_state);
6815 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6818 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6819 return try_to_wake_up(curr->private, mode, wake_flags);
6821 EXPORT_SYMBOL(default_wake_function);
6823 static void __setscheduler_prio(struct task_struct *p, int prio)
6826 p->sched_class = &dl_sched_class;
6827 else if (rt_prio(prio))
6828 p->sched_class = &rt_sched_class;
6830 p->sched_class = &fair_sched_class;
6835 #ifdef CONFIG_RT_MUTEXES
6837 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6840 prio = min(prio, pi_task->prio);
6845 static inline int rt_effective_prio(struct task_struct *p, int prio)
6847 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6849 return __rt_effective_prio(pi_task, prio);
6853 * rt_mutex_setprio - set the current priority of a task
6855 * @pi_task: donor task
6857 * This function changes the 'effective' priority of a task. It does
6858 * not touch ->normal_prio like __setscheduler().
6860 * Used by the rt_mutex code to implement priority inheritance
6861 * logic. Call site only calls if the priority of the task changed.
6863 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6865 int prio, oldprio, queued, running, queue_flag =
6866 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6867 const struct sched_class *prev_class;
6871 /* XXX used to be waiter->prio, not waiter->task->prio */
6872 prio = __rt_effective_prio(pi_task, p->normal_prio);
6875 * If nothing changed; bail early.
6877 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6880 rq = __task_rq_lock(p, &rf);
6881 update_rq_clock(rq);
6883 * Set under pi_lock && rq->lock, such that the value can be used under
6886 * Note that there is loads of tricky to make this pointer cache work
6887 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6888 * ensure a task is de-boosted (pi_task is set to NULL) before the
6889 * task is allowed to run again (and can exit). This ensures the pointer
6890 * points to a blocked task -- which guarantees the task is present.
6892 p->pi_top_task = pi_task;
6895 * For FIFO/RR we only need to set prio, if that matches we're done.
6897 if (prio == p->prio && !dl_prio(prio))
6901 * Idle task boosting is a nono in general. There is one
6902 * exception, when PREEMPT_RT and NOHZ is active:
6904 * The idle task calls get_next_timer_interrupt() and holds
6905 * the timer wheel base->lock on the CPU and another CPU wants
6906 * to access the timer (probably to cancel it). We can safely
6907 * ignore the boosting request, as the idle CPU runs this code
6908 * with interrupts disabled and will complete the lock
6909 * protected section without being interrupted. So there is no
6910 * real need to boost.
6912 if (unlikely(p == rq->idle)) {
6913 WARN_ON(p != rq->curr);
6914 WARN_ON(p->pi_blocked_on);
6918 trace_sched_pi_setprio(p, pi_task);
6921 if (oldprio == prio)
6922 queue_flag &= ~DEQUEUE_MOVE;
6924 prev_class = p->sched_class;
6925 queued = task_on_rq_queued(p);
6926 running = task_current(rq, p);
6928 dequeue_task(rq, p, queue_flag);
6930 put_prev_task(rq, p);
6933 * Boosting condition are:
6934 * 1. -rt task is running and holds mutex A
6935 * --> -dl task blocks on mutex A
6937 * 2. -dl task is running and holds mutex A
6938 * --> -dl task blocks on mutex A and could preempt the
6941 if (dl_prio(prio)) {
6942 if (!dl_prio(p->normal_prio) ||
6943 (pi_task && dl_prio(pi_task->prio) &&
6944 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6945 p->dl.pi_se = pi_task->dl.pi_se;
6946 queue_flag |= ENQUEUE_REPLENISH;
6948 p->dl.pi_se = &p->dl;
6950 } else if (rt_prio(prio)) {
6951 if (dl_prio(oldprio))
6952 p->dl.pi_se = &p->dl;
6954 queue_flag |= ENQUEUE_HEAD;
6956 if (dl_prio(oldprio))
6957 p->dl.pi_se = &p->dl;
6958 if (rt_prio(oldprio))
6962 __setscheduler_prio(p, prio);
6965 enqueue_task(rq, p, queue_flag);
6967 set_next_task(rq, p);
6969 check_class_changed(rq, p, prev_class, oldprio);
6971 /* Avoid rq from going away on us: */
6974 rq_unpin_lock(rq, &rf);
6975 __balance_callbacks(rq);
6976 raw_spin_rq_unlock(rq);
6981 static inline int rt_effective_prio(struct task_struct *p, int prio)
6987 void set_user_nice(struct task_struct *p, long nice)
6989 bool queued, running;
6994 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6997 * We have to be careful, if called from sys_setpriority(),
6998 * the task might be in the middle of scheduling on another CPU.
7000 rq = task_rq_lock(p, &rf);
7001 update_rq_clock(rq);
7004 * The RT priorities are set via sched_setscheduler(), but we still
7005 * allow the 'normal' nice value to be set - but as expected
7006 * it won't have any effect on scheduling until the task is
7007 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7009 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7010 p->static_prio = NICE_TO_PRIO(nice);
7013 queued = task_on_rq_queued(p);
7014 running = task_current(rq, p);
7016 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7018 put_prev_task(rq, p);
7020 p->static_prio = NICE_TO_PRIO(nice);
7021 set_load_weight(p, true);
7023 p->prio = effective_prio(p);
7026 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7028 set_next_task(rq, p);
7031 * If the task increased its priority or is running and
7032 * lowered its priority, then reschedule its CPU:
7034 p->sched_class->prio_changed(rq, p, old_prio);
7037 task_rq_unlock(rq, p, &rf);
7039 EXPORT_SYMBOL(set_user_nice);
7042 * is_nice_reduction - check if nice value is an actual reduction
7044 * Similar to can_nice() but does not perform a capability check.
7049 static bool is_nice_reduction(const struct task_struct *p, const int nice)
7051 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7052 int nice_rlim = nice_to_rlimit(nice);
7054 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7058 * can_nice - check if a task can reduce its nice value
7062 int can_nice(const struct task_struct *p, const int nice)
7064 return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
7067 #ifdef __ARCH_WANT_SYS_NICE
7070 * sys_nice - change the priority of the current process.
7071 * @increment: priority increment
7073 * sys_setpriority is a more generic, but much slower function that
7074 * does similar things.
7076 SYSCALL_DEFINE1(nice, int, increment)
7081 * Setpriority might change our priority at the same moment.
7082 * We don't have to worry. Conceptually one call occurs first
7083 * and we have a single winner.
7085 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7086 nice = task_nice(current) + increment;
7088 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7089 if (increment < 0 && !can_nice(current, nice))
7092 retval = security_task_setnice(current, nice);
7096 set_user_nice(current, nice);
7103 * task_prio - return the priority value of a given task.
7104 * @p: the task in question.
7106 * Return: The priority value as seen by users in /proc.
7108 * sched policy return value kernel prio user prio/nice
7110 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7111 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7112 * deadline -101 -1 0
7114 int task_prio(const struct task_struct *p)
7116 return p->prio - MAX_RT_PRIO;
7120 * idle_cpu - is a given CPU idle currently?
7121 * @cpu: the processor in question.
7123 * Return: 1 if the CPU is currently idle. 0 otherwise.
7125 int idle_cpu(int cpu)
7127 struct rq *rq = cpu_rq(cpu);
7129 if (rq->curr != rq->idle)
7136 if (rq->ttwu_pending)
7144 * available_idle_cpu - is a given CPU idle for enqueuing work.
7145 * @cpu: the CPU in question.
7147 * Return: 1 if the CPU is currently idle. 0 otherwise.
7149 int available_idle_cpu(int cpu)
7154 if (vcpu_is_preempted(cpu))
7161 * idle_task - return the idle task for a given CPU.
7162 * @cpu: the processor in question.
7164 * Return: The idle task for the CPU @cpu.
7166 struct task_struct *idle_task(int cpu)
7168 return cpu_rq(cpu)->idle;
7173 * This function computes an effective utilization for the given CPU, to be
7174 * used for frequency selection given the linear relation: f = u * f_max.
7176 * The scheduler tracks the following metrics:
7178 * cpu_util_{cfs,rt,dl,irq}()
7181 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7182 * synchronized windows and are thus directly comparable.
7184 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7185 * which excludes things like IRQ and steal-time. These latter are then accrued
7186 * in the irq utilization.
7188 * The DL bandwidth number otoh is not a measured metric but a value computed
7189 * based on the task model parameters and gives the minimal utilization
7190 * required to meet deadlines.
7192 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7193 enum cpu_util_type type,
7194 struct task_struct *p)
7196 unsigned long dl_util, util, irq, max;
7197 struct rq *rq = cpu_rq(cpu);
7199 max = arch_scale_cpu_capacity(cpu);
7201 if (!uclamp_is_used() &&
7202 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7207 * Early check to see if IRQ/steal time saturates the CPU, can be
7208 * because of inaccuracies in how we track these -- see
7209 * update_irq_load_avg().
7211 irq = cpu_util_irq(rq);
7212 if (unlikely(irq >= max))
7216 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7217 * CFS tasks and we use the same metric to track the effective
7218 * utilization (PELT windows are synchronized) we can directly add them
7219 * to obtain the CPU's actual utilization.
7221 * CFS and RT utilization can be boosted or capped, depending on
7222 * utilization clamp constraints requested by currently RUNNABLE
7224 * When there are no CFS RUNNABLE tasks, clamps are released and
7225 * frequency will be gracefully reduced with the utilization decay.
7227 util = util_cfs + cpu_util_rt(rq);
7228 if (type == FREQUENCY_UTIL)
7229 util = uclamp_rq_util_with(rq, util, p);
7231 dl_util = cpu_util_dl(rq);
7234 * For frequency selection we do not make cpu_util_dl() a permanent part
7235 * of this sum because we want to use cpu_bw_dl() later on, but we need
7236 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7237 * that we select f_max when there is no idle time.
7239 * NOTE: numerical errors or stop class might cause us to not quite hit
7240 * saturation when we should -- something for later.
7242 if (util + dl_util >= max)
7246 * OTOH, for energy computation we need the estimated running time, so
7247 * include util_dl and ignore dl_bw.
7249 if (type == ENERGY_UTIL)
7253 * There is still idle time; further improve the number by using the
7254 * irq metric. Because IRQ/steal time is hidden from the task clock we
7255 * need to scale the task numbers:
7258 * U' = irq + --------- * U
7261 util = scale_irq_capacity(util, irq, max);
7265 * Bandwidth required by DEADLINE must always be granted while, for
7266 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7267 * to gracefully reduce the frequency when no tasks show up for longer
7270 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7271 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7272 * an interface. So, we only do the latter for now.
7274 if (type == FREQUENCY_UTIL)
7275 util += cpu_bw_dl(rq);
7277 return min(max, util);
7280 unsigned long sched_cpu_util(int cpu)
7282 return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
7284 #endif /* CONFIG_SMP */
7287 * find_process_by_pid - find a process with a matching PID value.
7288 * @pid: the pid in question.
7290 * The task of @pid, if found. %NULL otherwise.
7292 static struct task_struct *find_process_by_pid(pid_t pid)
7294 return pid ? find_task_by_vpid(pid) : current;
7298 * sched_setparam() passes in -1 for its policy, to let the functions
7299 * it calls know not to change it.
7301 #define SETPARAM_POLICY -1
7303 static void __setscheduler_params(struct task_struct *p,
7304 const struct sched_attr *attr)
7306 int policy = attr->sched_policy;
7308 if (policy == SETPARAM_POLICY)
7313 if (dl_policy(policy))
7314 __setparam_dl(p, attr);
7315 else if (fair_policy(policy))
7316 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7319 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7320 * !rt_policy. Always setting this ensures that things like
7321 * getparam()/getattr() don't report silly values for !rt tasks.
7323 p->rt_priority = attr->sched_priority;
7324 p->normal_prio = normal_prio(p);
7325 set_load_weight(p, true);
7329 * Check the target process has a UID that matches the current process's:
7331 static bool check_same_owner(struct task_struct *p)
7333 const struct cred *cred = current_cred(), *pcred;
7337 pcred = __task_cred(p);
7338 match = (uid_eq(cred->euid, pcred->euid) ||
7339 uid_eq(cred->euid, pcred->uid));
7345 * Allow unprivileged RT tasks to decrease priority.
7346 * Only issue a capable test if needed and only once to avoid an audit
7347 * event on permitted non-privileged operations:
7349 static int user_check_sched_setscheduler(struct task_struct *p,
7350 const struct sched_attr *attr,
7351 int policy, int reset_on_fork)
7353 if (fair_policy(policy)) {
7354 if (attr->sched_nice < task_nice(p) &&
7355 !is_nice_reduction(p, attr->sched_nice))
7359 if (rt_policy(policy)) {
7360 unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7362 /* Can't set/change the rt policy: */
7363 if (policy != p->policy && !rlim_rtprio)
7366 /* Can't increase priority: */
7367 if (attr->sched_priority > p->rt_priority &&
7368 attr->sched_priority > rlim_rtprio)
7373 * Can't set/change SCHED_DEADLINE policy at all for now
7374 * (safest behavior); in the future we would like to allow
7375 * unprivileged DL tasks to increase their relative deadline
7376 * or reduce their runtime (both ways reducing utilization)
7378 if (dl_policy(policy))
7382 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7383 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7385 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7386 if (!is_nice_reduction(p, task_nice(p)))
7390 /* Can't change other user's priorities: */
7391 if (!check_same_owner(p))
7394 /* Normal users shall not reset the sched_reset_on_fork flag: */
7395 if (p->sched_reset_on_fork && !reset_on_fork)
7401 if (!capable(CAP_SYS_NICE))
7407 static int __sched_setscheduler(struct task_struct *p,
7408 const struct sched_attr *attr,
7411 int oldpolicy = -1, policy = attr->sched_policy;
7412 int retval, oldprio, newprio, queued, running;
7413 const struct sched_class *prev_class;
7414 struct callback_head *head;
7417 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7420 /* The pi code expects interrupts enabled */
7421 BUG_ON(pi && in_interrupt());
7423 /* Double check policy once rq lock held: */
7425 reset_on_fork = p->sched_reset_on_fork;
7426 policy = oldpolicy = p->policy;
7428 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7430 if (!valid_policy(policy))
7434 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7438 * Valid priorities for SCHED_FIFO and SCHED_RR are
7439 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7440 * SCHED_BATCH and SCHED_IDLE is 0.
7442 if (attr->sched_priority > MAX_RT_PRIO-1)
7444 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7445 (rt_policy(policy) != (attr->sched_priority != 0)))
7449 retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7453 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7456 retval = security_task_setscheduler(p);
7461 /* Update task specific "requested" clamps */
7462 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7463 retval = uclamp_validate(p, attr);
7472 * Make sure no PI-waiters arrive (or leave) while we are
7473 * changing the priority of the task:
7475 * To be able to change p->policy safely, the appropriate
7476 * runqueue lock must be held.
7478 rq = task_rq_lock(p, &rf);
7479 update_rq_clock(rq);
7482 * Changing the policy of the stop threads its a very bad idea:
7484 if (p == rq->stop) {
7490 * If not changing anything there's no need to proceed further,
7491 * but store a possible modification of reset_on_fork.
7493 if (unlikely(policy == p->policy)) {
7494 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7496 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7498 if (dl_policy(policy) && dl_param_changed(p, attr))
7500 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7503 p->sched_reset_on_fork = reset_on_fork;
7510 #ifdef CONFIG_RT_GROUP_SCHED
7512 * Do not allow realtime tasks into groups that have no runtime
7515 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7516 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7517 !task_group_is_autogroup(task_group(p))) {
7523 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7524 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7525 cpumask_t *span = rq->rd->span;
7528 * Don't allow tasks with an affinity mask smaller than
7529 * the entire root_domain to become SCHED_DEADLINE. We
7530 * will also fail if there's no bandwidth available.
7532 if (!cpumask_subset(span, p->cpus_ptr) ||
7533 rq->rd->dl_bw.bw == 0) {
7541 /* Re-check policy now with rq lock held: */
7542 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7543 policy = oldpolicy = -1;
7544 task_rq_unlock(rq, p, &rf);
7546 cpuset_read_unlock();
7551 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7552 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7555 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7560 p->sched_reset_on_fork = reset_on_fork;
7563 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7566 * Take priority boosted tasks into account. If the new
7567 * effective priority is unchanged, we just store the new
7568 * normal parameters and do not touch the scheduler class and
7569 * the runqueue. This will be done when the task deboost
7572 newprio = rt_effective_prio(p, newprio);
7573 if (newprio == oldprio)
7574 queue_flags &= ~DEQUEUE_MOVE;
7577 queued = task_on_rq_queued(p);
7578 running = task_current(rq, p);
7580 dequeue_task(rq, p, queue_flags);
7582 put_prev_task(rq, p);
7584 prev_class = p->sched_class;
7586 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7587 __setscheduler_params(p, attr);
7588 __setscheduler_prio(p, newprio);
7590 __setscheduler_uclamp(p, attr);
7594 * We enqueue to tail when the priority of a task is
7595 * increased (user space view).
7597 if (oldprio < p->prio)
7598 queue_flags |= ENQUEUE_HEAD;
7600 enqueue_task(rq, p, queue_flags);
7603 set_next_task(rq, p);
7605 check_class_changed(rq, p, prev_class, oldprio);
7607 /* Avoid rq from going away on us: */
7609 head = splice_balance_callbacks(rq);
7610 task_rq_unlock(rq, p, &rf);
7613 cpuset_read_unlock();
7614 rt_mutex_adjust_pi(p);
7617 /* Run balance callbacks after we've adjusted the PI chain: */
7618 balance_callbacks(rq, head);
7624 task_rq_unlock(rq, p, &rf);
7626 cpuset_read_unlock();
7630 static int _sched_setscheduler(struct task_struct *p, int policy,
7631 const struct sched_param *param, bool check)
7633 struct sched_attr attr = {
7634 .sched_policy = policy,
7635 .sched_priority = param->sched_priority,
7636 .sched_nice = PRIO_TO_NICE(p->static_prio),
7639 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7640 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7641 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7642 policy &= ~SCHED_RESET_ON_FORK;
7643 attr.sched_policy = policy;
7646 return __sched_setscheduler(p, &attr, check, true);
7649 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7650 * @p: the task in question.
7651 * @policy: new policy.
7652 * @param: structure containing the new RT priority.
7654 * Use sched_set_fifo(), read its comment.
7656 * Return: 0 on success. An error code otherwise.
7658 * NOTE that the task may be already dead.
7660 int sched_setscheduler(struct task_struct *p, int policy,
7661 const struct sched_param *param)
7663 return _sched_setscheduler(p, policy, param, true);
7666 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7668 return __sched_setscheduler(p, attr, true, true);
7671 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7673 return __sched_setscheduler(p, attr, false, true);
7675 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7678 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7679 * @p: the task in question.
7680 * @policy: new policy.
7681 * @param: structure containing the new RT priority.
7683 * Just like sched_setscheduler, only don't bother checking if the
7684 * current context has permission. For example, this is needed in
7685 * stop_machine(): we create temporary high priority worker threads,
7686 * but our caller might not have that capability.
7688 * Return: 0 on success. An error code otherwise.
7690 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7691 const struct sched_param *param)
7693 return _sched_setscheduler(p, policy, param, false);
7697 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7698 * incapable of resource management, which is the one thing an OS really should
7701 * This is of course the reason it is limited to privileged users only.
7703 * Worse still; it is fundamentally impossible to compose static priority
7704 * workloads. You cannot take two correctly working static prio workloads
7705 * and smash them together and still expect them to work.
7707 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7711 * The administrator _MUST_ configure the system, the kernel simply doesn't
7712 * know enough information to make a sensible choice.
7714 void sched_set_fifo(struct task_struct *p)
7716 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7717 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7719 EXPORT_SYMBOL_GPL(sched_set_fifo);
7722 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7724 void sched_set_fifo_low(struct task_struct *p)
7726 struct sched_param sp = { .sched_priority = 1 };
7727 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7729 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7731 void sched_set_normal(struct task_struct *p, int nice)
7733 struct sched_attr attr = {
7734 .sched_policy = SCHED_NORMAL,
7737 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7739 EXPORT_SYMBOL_GPL(sched_set_normal);
7742 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7744 struct sched_param lparam;
7745 struct task_struct *p;
7748 if (!param || pid < 0)
7750 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7755 p = find_process_by_pid(pid);
7761 retval = sched_setscheduler(p, policy, &lparam);
7769 * Mimics kernel/events/core.c perf_copy_attr().
7771 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7776 /* Zero the full structure, so that a short copy will be nice: */
7777 memset(attr, 0, sizeof(*attr));
7779 ret = get_user(size, &uattr->size);
7783 /* ABI compatibility quirk: */
7785 size = SCHED_ATTR_SIZE_VER0;
7786 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7789 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7796 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7797 size < SCHED_ATTR_SIZE_VER1)
7801 * XXX: Do we want to be lenient like existing syscalls; or do we want
7802 * to be strict and return an error on out-of-bounds values?
7804 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7809 put_user(sizeof(*attr), &uattr->size);
7813 static void get_params(struct task_struct *p, struct sched_attr *attr)
7815 if (task_has_dl_policy(p))
7816 __getparam_dl(p, attr);
7817 else if (task_has_rt_policy(p))
7818 attr->sched_priority = p->rt_priority;
7820 attr->sched_nice = task_nice(p);
7824 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7825 * @pid: the pid in question.
7826 * @policy: new policy.
7827 * @param: structure containing the new RT priority.
7829 * Return: 0 on success. An error code otherwise.
7831 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7836 return do_sched_setscheduler(pid, policy, param);
7840 * sys_sched_setparam - set/change the RT priority of a thread
7841 * @pid: the pid in question.
7842 * @param: structure containing the new RT priority.
7844 * Return: 0 on success. An error code otherwise.
7846 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7848 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7852 * sys_sched_setattr - same as above, but with extended sched_attr
7853 * @pid: the pid in question.
7854 * @uattr: structure containing the extended parameters.
7855 * @flags: for future extension.
7857 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7858 unsigned int, flags)
7860 struct sched_attr attr;
7861 struct task_struct *p;
7864 if (!uattr || pid < 0 || flags)
7867 retval = sched_copy_attr(uattr, &attr);
7871 if ((int)attr.sched_policy < 0)
7873 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7874 attr.sched_policy = SETPARAM_POLICY;
7878 p = find_process_by_pid(pid);
7884 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7885 get_params(p, &attr);
7886 retval = sched_setattr(p, &attr);
7894 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7895 * @pid: the pid in question.
7897 * Return: On success, the policy of the thread. Otherwise, a negative error
7900 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7902 struct task_struct *p;
7910 p = find_process_by_pid(pid);
7912 retval = security_task_getscheduler(p);
7915 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7922 * sys_sched_getparam - get the RT priority of a thread
7923 * @pid: the pid in question.
7924 * @param: structure containing the RT priority.
7926 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7929 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7931 struct sched_param lp = { .sched_priority = 0 };
7932 struct task_struct *p;
7935 if (!param || pid < 0)
7939 p = find_process_by_pid(pid);
7944 retval = security_task_getscheduler(p);
7948 if (task_has_rt_policy(p))
7949 lp.sched_priority = p->rt_priority;
7953 * This one might sleep, we cannot do it with a spinlock held ...
7955 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7965 * Copy the kernel size attribute structure (which might be larger
7966 * than what user-space knows about) to user-space.
7968 * Note that all cases are valid: user-space buffer can be larger or
7969 * smaller than the kernel-space buffer. The usual case is that both
7970 * have the same size.
7973 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7974 struct sched_attr *kattr,
7977 unsigned int ksize = sizeof(*kattr);
7979 if (!access_ok(uattr, usize))
7983 * sched_getattr() ABI forwards and backwards compatibility:
7985 * If usize == ksize then we just copy everything to user-space and all is good.
7987 * If usize < ksize then we only copy as much as user-space has space for,
7988 * this keeps ABI compatibility as well. We skip the rest.
7990 * If usize > ksize then user-space is using a newer version of the ABI,
7991 * which part the kernel doesn't know about. Just ignore it - tooling can
7992 * detect the kernel's knowledge of attributes from the attr->size value
7993 * which is set to ksize in this case.
7995 kattr->size = min(usize, ksize);
7997 if (copy_to_user(uattr, kattr, kattr->size))
8004 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8005 * @pid: the pid in question.
8006 * @uattr: structure containing the extended parameters.
8007 * @usize: sizeof(attr) for fwd/bwd comp.
8008 * @flags: for future extension.
8010 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8011 unsigned int, usize, unsigned int, flags)
8013 struct sched_attr kattr = { };
8014 struct task_struct *p;
8017 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8018 usize < SCHED_ATTR_SIZE_VER0 || flags)
8022 p = find_process_by_pid(pid);
8027 retval = security_task_getscheduler(p);
8031 kattr.sched_policy = p->policy;
8032 if (p->sched_reset_on_fork)
8033 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8034 get_params(p, &kattr);
8035 kattr.sched_flags &= SCHED_FLAG_ALL;
8037 #ifdef CONFIG_UCLAMP_TASK
8039 * This could race with another potential updater, but this is fine
8040 * because it'll correctly read the old or the new value. We don't need
8041 * to guarantee who wins the race as long as it doesn't return garbage.
8043 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8044 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8049 return sched_attr_copy_to_user(uattr, &kattr, usize);
8057 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8062 * If the task isn't a deadline task or admission control is
8063 * disabled then we don't care about affinity changes.
8065 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8069 * Since bandwidth control happens on root_domain basis,
8070 * if admission test is enabled, we only admit -deadline
8071 * tasks allowed to run on all the CPUs in the task's
8075 if (!cpumask_subset(task_rq(p)->rd->span, mask))
8083 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
8086 cpumask_var_t cpus_allowed, new_mask;
8088 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8091 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8093 goto out_free_cpus_allowed;
8096 cpuset_cpus_allowed(p, cpus_allowed);
8097 cpumask_and(new_mask, mask, cpus_allowed);
8099 retval = dl_task_check_affinity(p, new_mask);
8101 goto out_free_new_mask;
8103 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
8105 goto out_free_new_mask;
8107 cpuset_cpus_allowed(p, cpus_allowed);
8108 if (!cpumask_subset(new_mask, cpus_allowed)) {
8110 * We must have raced with a concurrent cpuset update.
8111 * Just reset the cpumask to the cpuset's cpus_allowed.
8113 cpumask_copy(new_mask, cpus_allowed);
8118 free_cpumask_var(new_mask);
8119 out_free_cpus_allowed:
8120 free_cpumask_var(cpus_allowed);
8124 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8126 struct task_struct *p;
8131 p = find_process_by_pid(pid);
8137 /* Prevent p going away */
8141 if (p->flags & PF_NO_SETAFFINITY) {
8146 if (!check_same_owner(p)) {
8148 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8156 retval = security_task_setscheduler(p);
8160 retval = __sched_setaffinity(p, in_mask);
8166 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8167 struct cpumask *new_mask)
8169 if (len < cpumask_size())
8170 cpumask_clear(new_mask);
8171 else if (len > cpumask_size())
8172 len = cpumask_size();
8174 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8178 * sys_sched_setaffinity - set the CPU affinity of a process
8179 * @pid: pid of the process
8180 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8181 * @user_mask_ptr: user-space pointer to the new CPU mask
8183 * Return: 0 on success. An error code otherwise.
8185 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8186 unsigned long __user *, user_mask_ptr)
8188 cpumask_var_t new_mask;
8191 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8194 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8196 retval = sched_setaffinity(pid, new_mask);
8197 free_cpumask_var(new_mask);
8201 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8203 struct task_struct *p;
8204 unsigned long flags;
8210 p = find_process_by_pid(pid);
8214 retval = security_task_getscheduler(p);
8218 raw_spin_lock_irqsave(&p->pi_lock, flags);
8219 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8220 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8229 * sys_sched_getaffinity - get the CPU affinity of a process
8230 * @pid: pid of the process
8231 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8232 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8234 * Return: size of CPU mask copied to user_mask_ptr on success. An
8235 * error code otherwise.
8237 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8238 unsigned long __user *, user_mask_ptr)
8243 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8245 if (len & (sizeof(unsigned long)-1))
8248 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8251 ret = sched_getaffinity(pid, mask);
8253 unsigned int retlen = min(len, cpumask_size());
8255 if (copy_to_user(user_mask_ptr, mask, retlen))
8260 free_cpumask_var(mask);
8265 static void do_sched_yield(void)
8270 rq = this_rq_lock_irq(&rf);
8272 schedstat_inc(rq->yld_count);
8273 current->sched_class->yield_task(rq);
8276 rq_unlock_irq(rq, &rf);
8277 sched_preempt_enable_no_resched();
8283 * sys_sched_yield - yield the current processor to other threads.
8285 * This function yields the current CPU to other tasks. If there are no
8286 * other threads running on this CPU then this function will return.
8290 SYSCALL_DEFINE0(sched_yield)
8296 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8297 int __sched __cond_resched(void)
8299 if (should_resched(0)) {
8300 preempt_schedule_common();
8304 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8305 * whether the current CPU is in an RCU read-side critical section,
8306 * so the tick can report quiescent states even for CPUs looping
8307 * in kernel context. In contrast, in non-preemptible kernels,
8308 * RCU readers leave no in-memory hints, which means that CPU-bound
8309 * processes executing in kernel context might never report an
8310 * RCU quiescent state. Therefore, the following code causes
8311 * cond_resched() to report a quiescent state, but only when RCU
8312 * is in urgent need of one.
8314 #ifndef CONFIG_PREEMPT_RCU
8319 EXPORT_SYMBOL(__cond_resched);
8322 #ifdef CONFIG_PREEMPT_DYNAMIC
8323 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8324 #define cond_resched_dynamic_enabled __cond_resched
8325 #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8326 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8327 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8329 #define might_resched_dynamic_enabled __cond_resched
8330 #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8331 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8332 EXPORT_STATIC_CALL_TRAMP(might_resched);
8333 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8334 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8335 int __sched dynamic_cond_resched(void)
8337 if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8339 return __cond_resched();
8341 EXPORT_SYMBOL(dynamic_cond_resched);
8343 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8344 int __sched dynamic_might_resched(void)
8346 if (!static_branch_unlikely(&sk_dynamic_might_resched))
8348 return __cond_resched();
8350 EXPORT_SYMBOL(dynamic_might_resched);
8355 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8356 * call schedule, and on return reacquire the lock.
8358 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8359 * operations here to prevent schedule() from being called twice (once via
8360 * spin_unlock(), once by hand).
8362 int __cond_resched_lock(spinlock_t *lock)
8364 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8367 lockdep_assert_held(lock);
8369 if (spin_needbreak(lock) || resched) {
8371 if (!_cond_resched())
8378 EXPORT_SYMBOL(__cond_resched_lock);
8380 int __cond_resched_rwlock_read(rwlock_t *lock)
8382 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8385 lockdep_assert_held_read(lock);
8387 if (rwlock_needbreak(lock) || resched) {
8389 if (!_cond_resched())
8396 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8398 int __cond_resched_rwlock_write(rwlock_t *lock)
8400 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8403 lockdep_assert_held_write(lock);
8405 if (rwlock_needbreak(lock) || resched) {
8407 if (!_cond_resched())
8414 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8416 #ifdef CONFIG_PREEMPT_DYNAMIC
8418 #ifdef CONFIG_GENERIC_ENTRY
8419 #include <linux/entry-common.h>
8425 * SC:preempt_schedule
8426 * SC:preempt_schedule_notrace
8427 * SC:irqentry_exit_cond_resched
8431 * cond_resched <- __cond_resched
8432 * might_resched <- RET0
8433 * preempt_schedule <- NOP
8434 * preempt_schedule_notrace <- NOP
8435 * irqentry_exit_cond_resched <- NOP
8438 * cond_resched <- __cond_resched
8439 * might_resched <- __cond_resched
8440 * preempt_schedule <- NOP
8441 * preempt_schedule_notrace <- NOP
8442 * irqentry_exit_cond_resched <- NOP
8445 * cond_resched <- RET0
8446 * might_resched <- RET0
8447 * preempt_schedule <- preempt_schedule
8448 * preempt_schedule_notrace <- preempt_schedule_notrace
8449 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8453 preempt_dynamic_undefined = -1,
8454 preempt_dynamic_none,
8455 preempt_dynamic_voluntary,
8456 preempt_dynamic_full,
8459 int preempt_dynamic_mode = preempt_dynamic_undefined;
8461 int sched_dynamic_mode(const char *str)
8463 if (!strcmp(str, "none"))
8464 return preempt_dynamic_none;
8466 if (!strcmp(str, "voluntary"))
8467 return preempt_dynamic_voluntary;
8469 if (!strcmp(str, "full"))
8470 return preempt_dynamic_full;
8475 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8476 #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8477 #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8478 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8479 #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8480 #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8482 #error "Unsupported PREEMPT_DYNAMIC mechanism"
8485 void sched_dynamic_update(int mode)
8488 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8489 * the ZERO state, which is invalid.
8491 preempt_dynamic_enable(cond_resched);
8492 preempt_dynamic_enable(might_resched);
8493 preempt_dynamic_enable(preempt_schedule);
8494 preempt_dynamic_enable(preempt_schedule_notrace);
8495 preempt_dynamic_enable(irqentry_exit_cond_resched);
8498 case preempt_dynamic_none:
8499 preempt_dynamic_enable(cond_resched);
8500 preempt_dynamic_disable(might_resched);
8501 preempt_dynamic_disable(preempt_schedule);
8502 preempt_dynamic_disable(preempt_schedule_notrace);
8503 preempt_dynamic_disable(irqentry_exit_cond_resched);
8504 pr_info("Dynamic Preempt: none\n");
8507 case preempt_dynamic_voluntary:
8508 preempt_dynamic_enable(cond_resched);
8509 preempt_dynamic_enable(might_resched);
8510 preempt_dynamic_disable(preempt_schedule);
8511 preempt_dynamic_disable(preempt_schedule_notrace);
8512 preempt_dynamic_disable(irqentry_exit_cond_resched);
8513 pr_info("Dynamic Preempt: voluntary\n");
8516 case preempt_dynamic_full:
8517 preempt_dynamic_disable(cond_resched);
8518 preempt_dynamic_disable(might_resched);
8519 preempt_dynamic_enable(preempt_schedule);
8520 preempt_dynamic_enable(preempt_schedule_notrace);
8521 preempt_dynamic_enable(irqentry_exit_cond_resched);
8522 pr_info("Dynamic Preempt: full\n");
8526 preempt_dynamic_mode = mode;
8529 static int __init setup_preempt_mode(char *str)
8531 int mode = sched_dynamic_mode(str);
8533 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8537 sched_dynamic_update(mode);
8540 __setup("preempt=", setup_preempt_mode);
8542 static void __init preempt_dynamic_init(void)
8544 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8545 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8546 sched_dynamic_update(preempt_dynamic_none);
8547 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8548 sched_dynamic_update(preempt_dynamic_voluntary);
8550 /* Default static call setting, nothing to do */
8551 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8552 preempt_dynamic_mode = preempt_dynamic_full;
8553 pr_info("Dynamic Preempt: full\n");
8558 #define PREEMPT_MODEL_ACCESSOR(mode) \
8559 bool preempt_model_##mode(void) \
8561 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8562 return preempt_dynamic_mode == preempt_dynamic_##mode; \
8564 EXPORT_SYMBOL_GPL(preempt_model_##mode)
8566 PREEMPT_MODEL_ACCESSOR(none);
8567 PREEMPT_MODEL_ACCESSOR(voluntary);
8568 PREEMPT_MODEL_ACCESSOR(full);
8570 #else /* !CONFIG_PREEMPT_DYNAMIC */
8572 static inline void preempt_dynamic_init(void) { }
8574 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8577 * yield - yield the current processor to other threads.
8579 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8581 * The scheduler is at all times free to pick the calling task as the most
8582 * eligible task to run, if removing the yield() call from your code breaks
8583 * it, it's already broken.
8585 * Typical broken usage is:
8590 * where one assumes that yield() will let 'the other' process run that will
8591 * make event true. If the current task is a SCHED_FIFO task that will never
8592 * happen. Never use yield() as a progress guarantee!!
8594 * If you want to use yield() to wait for something, use wait_event().
8595 * If you want to use yield() to be 'nice' for others, use cond_resched().
8596 * If you still want to use yield(), do not!
8598 void __sched yield(void)
8600 set_current_state(TASK_RUNNING);
8603 EXPORT_SYMBOL(yield);
8606 * yield_to - yield the current processor to another thread in
8607 * your thread group, or accelerate that thread toward the
8608 * processor it's on.
8610 * @preempt: whether task preemption is allowed or not
8612 * It's the caller's job to ensure that the target task struct
8613 * can't go away on us before we can do any checks.
8616 * true (>0) if we indeed boosted the target task.
8617 * false (0) if we failed to boost the target.
8618 * -ESRCH if there's no task to yield to.
8620 int __sched yield_to(struct task_struct *p, bool preempt)
8622 struct task_struct *curr = current;
8623 struct rq *rq, *p_rq;
8624 unsigned long flags;
8627 local_irq_save(flags);
8633 * If we're the only runnable task on the rq and target rq also
8634 * has only one task, there's absolutely no point in yielding.
8636 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8641 double_rq_lock(rq, p_rq);
8642 if (task_rq(p) != p_rq) {
8643 double_rq_unlock(rq, p_rq);
8647 if (!curr->sched_class->yield_to_task)
8650 if (curr->sched_class != p->sched_class)
8653 if (task_running(p_rq, p) || !task_is_running(p))
8656 yielded = curr->sched_class->yield_to_task(rq, p);
8658 schedstat_inc(rq->yld_count);
8660 * Make p's CPU reschedule; pick_next_entity takes care of
8663 if (preempt && rq != p_rq)
8668 double_rq_unlock(rq, p_rq);
8670 local_irq_restore(flags);
8677 EXPORT_SYMBOL_GPL(yield_to);
8679 int io_schedule_prepare(void)
8681 int old_iowait = current->in_iowait;
8683 current->in_iowait = 1;
8684 blk_flush_plug(current->plug, true);
8688 void io_schedule_finish(int token)
8690 current->in_iowait = token;
8694 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8695 * that process accounting knows that this is a task in IO wait state.
8697 long __sched io_schedule_timeout(long timeout)
8702 token = io_schedule_prepare();
8703 ret = schedule_timeout(timeout);
8704 io_schedule_finish(token);
8708 EXPORT_SYMBOL(io_schedule_timeout);
8710 void __sched io_schedule(void)
8714 token = io_schedule_prepare();
8716 io_schedule_finish(token);
8718 EXPORT_SYMBOL(io_schedule);
8721 * sys_sched_get_priority_max - return maximum RT priority.
8722 * @policy: scheduling class.
8724 * Return: On success, this syscall returns the maximum
8725 * rt_priority that can be used by a given scheduling class.
8726 * On failure, a negative error code is returned.
8728 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8735 ret = MAX_RT_PRIO-1;
8737 case SCHED_DEADLINE:
8748 * sys_sched_get_priority_min - return minimum RT priority.
8749 * @policy: scheduling class.
8751 * Return: On success, this syscall returns the minimum
8752 * rt_priority that can be used by a given scheduling class.
8753 * On failure, a negative error code is returned.
8755 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8764 case SCHED_DEADLINE:
8773 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8775 struct task_struct *p;
8776 unsigned int time_slice;
8786 p = find_process_by_pid(pid);
8790 retval = security_task_getscheduler(p);
8794 rq = task_rq_lock(p, &rf);
8796 if (p->sched_class->get_rr_interval)
8797 time_slice = p->sched_class->get_rr_interval(rq, p);
8798 task_rq_unlock(rq, p, &rf);
8801 jiffies_to_timespec64(time_slice, t);
8810 * sys_sched_rr_get_interval - return the default timeslice of a process.
8811 * @pid: pid of the process.
8812 * @interval: userspace pointer to the timeslice value.
8814 * this syscall writes the default timeslice value of a given process
8815 * into the user-space timespec buffer. A value of '0' means infinity.
8817 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8820 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8821 struct __kernel_timespec __user *, interval)
8823 struct timespec64 t;
8824 int retval = sched_rr_get_interval(pid, &t);
8827 retval = put_timespec64(&t, interval);
8832 #ifdef CONFIG_COMPAT_32BIT_TIME
8833 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8834 struct old_timespec32 __user *, interval)
8836 struct timespec64 t;
8837 int retval = sched_rr_get_interval(pid, &t);
8840 retval = put_old_timespec32(&t, interval);
8845 void sched_show_task(struct task_struct *p)
8847 unsigned long free = 0;
8850 if (!try_get_task_stack(p))
8853 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8855 if (task_is_running(p))
8856 pr_cont(" running task ");
8857 #ifdef CONFIG_DEBUG_STACK_USAGE
8858 free = stack_not_used(p);
8863 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8865 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8866 free, task_pid_nr(p), ppid,
8867 read_task_thread_flags(p));
8869 print_worker_info(KERN_INFO, p);
8870 print_stop_info(KERN_INFO, p);
8871 show_stack(p, NULL, KERN_INFO);
8874 EXPORT_SYMBOL_GPL(sched_show_task);
8877 state_filter_match(unsigned long state_filter, struct task_struct *p)
8879 unsigned int state = READ_ONCE(p->__state);
8881 /* no filter, everything matches */
8885 /* filter, but doesn't match */
8886 if (!(state & state_filter))
8890 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8893 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8900 void show_state_filter(unsigned int state_filter)
8902 struct task_struct *g, *p;
8905 for_each_process_thread(g, p) {
8907 * reset the NMI-timeout, listing all files on a slow
8908 * console might take a lot of time:
8909 * Also, reset softlockup watchdogs on all CPUs, because
8910 * another CPU might be blocked waiting for us to process
8913 touch_nmi_watchdog();
8914 touch_all_softlockup_watchdogs();
8915 if (state_filter_match(state_filter, p))
8919 #ifdef CONFIG_SCHED_DEBUG
8921 sysrq_sched_debug_show();
8925 * Only show locks if all tasks are dumped:
8928 debug_show_all_locks();
8932 * init_idle - set up an idle thread for a given CPU
8933 * @idle: task in question
8934 * @cpu: CPU the idle task belongs to
8936 * NOTE: this function does not set the idle thread's NEED_RESCHED
8937 * flag, to make booting more robust.
8939 void __init init_idle(struct task_struct *idle, int cpu)
8941 struct rq *rq = cpu_rq(cpu);
8942 unsigned long flags;
8944 __sched_fork(0, idle);
8946 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8947 raw_spin_rq_lock(rq);
8949 idle->__state = TASK_RUNNING;
8950 idle->se.exec_start = sched_clock();
8952 * PF_KTHREAD should already be set at this point; regardless, make it
8953 * look like a proper per-CPU kthread.
8955 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8956 kthread_set_per_cpu(idle, cpu);
8960 * It's possible that init_idle() gets called multiple times on a task,
8961 * in that case do_set_cpus_allowed() will not do the right thing.
8963 * And since this is boot we can forgo the serialization.
8965 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8968 * We're having a chicken and egg problem, even though we are
8969 * holding rq->lock, the CPU isn't yet set to this CPU so the
8970 * lockdep check in task_group() will fail.
8972 * Similar case to sched_fork(). / Alternatively we could
8973 * use task_rq_lock() here and obtain the other rq->lock.
8978 __set_task_cpu(idle, cpu);
8982 rcu_assign_pointer(rq->curr, idle);
8983 idle->on_rq = TASK_ON_RQ_QUEUED;
8987 raw_spin_rq_unlock(rq);
8988 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8990 /* Set the preempt count _outside_ the spinlocks! */
8991 init_idle_preempt_count(idle, cpu);
8994 * The idle tasks have their own, simple scheduling class:
8996 idle->sched_class = &idle_sched_class;
8997 ftrace_graph_init_idle_task(idle, cpu);
8998 vtime_init_idle(idle, cpu);
9000 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9006 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9007 const struct cpumask *trial)
9011 if (cpumask_empty(cur))
9014 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
9019 int task_can_attach(struct task_struct *p,
9020 const struct cpumask *cs_effective_cpus)
9025 * Kthreads which disallow setaffinity shouldn't be moved
9026 * to a new cpuset; we don't want to change their CPU
9027 * affinity and isolating such threads by their set of
9028 * allowed nodes is unnecessary. Thus, cpusets are not
9029 * applicable for such threads. This prevents checking for
9030 * success of set_cpus_allowed_ptr() on all attached tasks
9031 * before cpus_mask may be changed.
9033 if (p->flags & PF_NO_SETAFFINITY) {
9038 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
9039 cs_effective_cpus)) {
9040 int cpu = cpumask_any_and(cpu_active_mask, cs_effective_cpus);
9042 if (unlikely(cpu >= nr_cpu_ids))
9044 ret = dl_cpu_busy(cpu, p);
9051 bool sched_smp_initialized __read_mostly;
9053 #ifdef CONFIG_NUMA_BALANCING
9054 /* Migrate current task p to target_cpu */
9055 int migrate_task_to(struct task_struct *p, int target_cpu)
9057 struct migration_arg arg = { p, target_cpu };
9058 int curr_cpu = task_cpu(p);
9060 if (curr_cpu == target_cpu)
9063 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
9066 /* TODO: This is not properly updating schedstats */
9068 trace_sched_move_numa(p, curr_cpu, target_cpu);
9069 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9073 * Requeue a task on a given node and accurately track the number of NUMA
9074 * tasks on the runqueues
9076 void sched_setnuma(struct task_struct *p, int nid)
9078 bool queued, running;
9082 rq = task_rq_lock(p, &rf);
9083 queued = task_on_rq_queued(p);
9084 running = task_current(rq, p);
9087 dequeue_task(rq, p, DEQUEUE_SAVE);
9089 put_prev_task(rq, p);
9091 p->numa_preferred_nid = nid;
9094 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
9096 set_next_task(rq, p);
9097 task_rq_unlock(rq, p, &rf);
9099 #endif /* CONFIG_NUMA_BALANCING */
9101 #ifdef CONFIG_HOTPLUG_CPU
9103 * Ensure that the idle task is using init_mm right before its CPU goes
9106 void idle_task_exit(void)
9108 struct mm_struct *mm = current->active_mm;
9110 BUG_ON(cpu_online(smp_processor_id()));
9111 BUG_ON(current != this_rq()->idle);
9113 if (mm != &init_mm) {
9114 switch_mm(mm, &init_mm, current);
9115 finish_arch_post_lock_switch();
9118 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9121 static int __balance_push_cpu_stop(void *arg)
9123 struct task_struct *p = arg;
9124 struct rq *rq = this_rq();
9128 raw_spin_lock_irq(&p->pi_lock);
9131 update_rq_clock(rq);
9133 if (task_rq(p) == rq && task_on_rq_queued(p)) {
9134 cpu = select_fallback_rq(rq->cpu, p);
9135 rq = __migrate_task(rq, &rf, p, cpu);
9139 raw_spin_unlock_irq(&p->pi_lock);
9146 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9149 * Ensure we only run per-cpu kthreads once the CPU goes !active.
9151 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9152 * effective when the hotplug motion is down.
9154 static void balance_push(struct rq *rq)
9156 struct task_struct *push_task = rq->curr;
9158 lockdep_assert_rq_held(rq);
9161 * Ensure the thing is persistent until balance_push_set(.on = false);
9163 rq->balance_callback = &balance_push_callback;
9166 * Only active while going offline and when invoked on the outgoing
9169 if (!cpu_dying(rq->cpu) || rq != this_rq())
9173 * Both the cpu-hotplug and stop task are in this case and are
9174 * required to complete the hotplug process.
9176 if (kthread_is_per_cpu(push_task) ||
9177 is_migration_disabled(push_task)) {
9180 * If this is the idle task on the outgoing CPU try to wake
9181 * up the hotplug control thread which might wait for the
9182 * last task to vanish. The rcuwait_active() check is
9183 * accurate here because the waiter is pinned on this CPU
9184 * and can't obviously be running in parallel.
9186 * On RT kernels this also has to check whether there are
9187 * pinned and scheduled out tasks on the runqueue. They
9188 * need to leave the migrate disabled section first.
9190 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9191 rcuwait_active(&rq->hotplug_wait)) {
9192 raw_spin_rq_unlock(rq);
9193 rcuwait_wake_up(&rq->hotplug_wait);
9194 raw_spin_rq_lock(rq);
9199 get_task_struct(push_task);
9201 * Temporarily drop rq->lock such that we can wake-up the stop task.
9202 * Both preemption and IRQs are still disabled.
9204 raw_spin_rq_unlock(rq);
9205 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9206 this_cpu_ptr(&push_work));
9208 * At this point need_resched() is true and we'll take the loop in
9209 * schedule(). The next pick is obviously going to be the stop task
9210 * which kthread_is_per_cpu() and will push this task away.
9212 raw_spin_rq_lock(rq);
9215 static void balance_push_set(int cpu, bool on)
9217 struct rq *rq = cpu_rq(cpu);
9220 rq_lock_irqsave(rq, &rf);
9222 WARN_ON_ONCE(rq->balance_callback);
9223 rq->balance_callback = &balance_push_callback;
9224 } else if (rq->balance_callback == &balance_push_callback) {
9225 rq->balance_callback = NULL;
9227 rq_unlock_irqrestore(rq, &rf);
9231 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9232 * inactive. All tasks which are not per CPU kernel threads are either
9233 * pushed off this CPU now via balance_push() or placed on a different CPU
9234 * during wakeup. Wait until the CPU is quiescent.
9236 static void balance_hotplug_wait(void)
9238 struct rq *rq = this_rq();
9240 rcuwait_wait_event(&rq->hotplug_wait,
9241 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9242 TASK_UNINTERRUPTIBLE);
9247 static inline void balance_push(struct rq *rq)
9251 static inline void balance_push_set(int cpu, bool on)
9255 static inline void balance_hotplug_wait(void)
9259 #endif /* CONFIG_HOTPLUG_CPU */
9261 void set_rq_online(struct rq *rq)
9264 const struct sched_class *class;
9266 cpumask_set_cpu(rq->cpu, rq->rd->online);
9269 for_each_class(class) {
9270 if (class->rq_online)
9271 class->rq_online(rq);
9276 void set_rq_offline(struct rq *rq)
9279 const struct sched_class *class;
9281 for_each_class(class) {
9282 if (class->rq_offline)
9283 class->rq_offline(rq);
9286 cpumask_clear_cpu(rq->cpu, rq->rd->online);
9292 * used to mark begin/end of suspend/resume:
9294 static int num_cpus_frozen;
9297 * Update cpusets according to cpu_active mask. If cpusets are
9298 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9299 * around partition_sched_domains().
9301 * If we come here as part of a suspend/resume, don't touch cpusets because we
9302 * want to restore it back to its original state upon resume anyway.
9304 static void cpuset_cpu_active(void)
9306 if (cpuhp_tasks_frozen) {
9308 * num_cpus_frozen tracks how many CPUs are involved in suspend
9309 * resume sequence. As long as this is not the last online
9310 * operation in the resume sequence, just build a single sched
9311 * domain, ignoring cpusets.
9313 partition_sched_domains(1, NULL, NULL);
9314 if (--num_cpus_frozen)
9317 * This is the last CPU online operation. So fall through and
9318 * restore the original sched domains by considering the
9319 * cpuset configurations.
9321 cpuset_force_rebuild();
9323 cpuset_update_active_cpus();
9326 static int cpuset_cpu_inactive(unsigned int cpu)
9328 if (!cpuhp_tasks_frozen) {
9329 int ret = dl_cpu_busy(cpu, NULL);
9333 cpuset_update_active_cpus();
9336 partition_sched_domains(1, NULL, NULL);
9341 int sched_cpu_activate(unsigned int cpu)
9343 struct rq *rq = cpu_rq(cpu);
9347 * Clear the balance_push callback and prepare to schedule
9350 balance_push_set(cpu, false);
9352 #ifdef CONFIG_SCHED_SMT
9354 * When going up, increment the number of cores with SMT present.
9356 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9357 static_branch_inc_cpuslocked(&sched_smt_present);
9359 set_cpu_active(cpu, true);
9361 if (sched_smp_initialized) {
9362 sched_update_numa(cpu, true);
9363 sched_domains_numa_masks_set(cpu);
9364 cpuset_cpu_active();
9368 * Put the rq online, if not already. This happens:
9370 * 1) In the early boot process, because we build the real domains
9371 * after all CPUs have been brought up.
9373 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9376 rq_lock_irqsave(rq, &rf);
9378 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9381 rq_unlock_irqrestore(rq, &rf);
9386 int sched_cpu_deactivate(unsigned int cpu)
9388 struct rq *rq = cpu_rq(cpu);
9393 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9394 * load balancing when not active
9396 nohz_balance_exit_idle(rq);
9398 set_cpu_active(cpu, false);
9401 * From this point forward, this CPU will refuse to run any task that
9402 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9403 * push those tasks away until this gets cleared, see
9404 * sched_cpu_dying().
9406 balance_push_set(cpu, true);
9409 * We've cleared cpu_active_mask / set balance_push, wait for all
9410 * preempt-disabled and RCU users of this state to go away such that
9411 * all new such users will observe it.
9413 * Specifically, we rely on ttwu to no longer target this CPU, see
9414 * ttwu_queue_cond() and is_cpu_allowed().
9416 * Do sync before park smpboot threads to take care the rcu boost case.
9420 rq_lock_irqsave(rq, &rf);
9422 update_rq_clock(rq);
9423 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9426 rq_unlock_irqrestore(rq, &rf);
9428 #ifdef CONFIG_SCHED_SMT
9430 * When going down, decrement the number of cores with SMT present.
9432 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9433 static_branch_dec_cpuslocked(&sched_smt_present);
9435 sched_core_cpu_deactivate(cpu);
9438 if (!sched_smp_initialized)
9441 sched_update_numa(cpu, false);
9442 ret = cpuset_cpu_inactive(cpu);
9444 balance_push_set(cpu, false);
9445 set_cpu_active(cpu, true);
9446 sched_update_numa(cpu, true);
9449 sched_domains_numa_masks_clear(cpu);
9453 static void sched_rq_cpu_starting(unsigned int cpu)
9455 struct rq *rq = cpu_rq(cpu);
9457 rq->calc_load_update = calc_load_update;
9458 update_max_interval();
9461 int sched_cpu_starting(unsigned int cpu)
9463 sched_core_cpu_starting(cpu);
9464 sched_rq_cpu_starting(cpu);
9465 sched_tick_start(cpu);
9469 #ifdef CONFIG_HOTPLUG_CPU
9472 * Invoked immediately before the stopper thread is invoked to bring the
9473 * CPU down completely. At this point all per CPU kthreads except the
9474 * hotplug thread (current) and the stopper thread (inactive) have been
9475 * either parked or have been unbound from the outgoing CPU. Ensure that
9476 * any of those which might be on the way out are gone.
9478 * If after this point a bound task is being woken on this CPU then the
9479 * responsible hotplug callback has failed to do it's job.
9480 * sched_cpu_dying() will catch it with the appropriate fireworks.
9482 int sched_cpu_wait_empty(unsigned int cpu)
9484 balance_hotplug_wait();
9489 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9490 * might have. Called from the CPU stopper task after ensuring that the
9491 * stopper is the last running task on the CPU, so nr_active count is
9492 * stable. We need to take the teardown thread which is calling this into
9493 * account, so we hand in adjust = 1 to the load calculation.
9495 * Also see the comment "Global load-average calculations".
9497 static void calc_load_migrate(struct rq *rq)
9499 long delta = calc_load_fold_active(rq, 1);
9502 atomic_long_add(delta, &calc_load_tasks);
9505 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9507 struct task_struct *g, *p;
9508 int cpu = cpu_of(rq);
9510 lockdep_assert_rq_held(rq);
9512 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9513 for_each_process_thread(g, p) {
9514 if (task_cpu(p) != cpu)
9517 if (!task_on_rq_queued(p))
9520 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9524 int sched_cpu_dying(unsigned int cpu)
9526 struct rq *rq = cpu_rq(cpu);
9529 /* Handle pending wakeups and then migrate everything off */
9530 sched_tick_stop(cpu);
9532 rq_lock_irqsave(rq, &rf);
9533 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9534 WARN(true, "Dying CPU not properly vacated!");
9535 dump_rq_tasks(rq, KERN_WARNING);
9537 rq_unlock_irqrestore(rq, &rf);
9539 calc_load_migrate(rq);
9540 update_max_interval();
9542 sched_core_cpu_dying(cpu);
9547 void __init sched_init_smp(void)
9549 sched_init_numa(NUMA_NO_NODE);
9552 * There's no userspace yet to cause hotplug operations; hence all the
9553 * CPU masks are stable and all blatant races in the below code cannot
9556 mutex_lock(&sched_domains_mutex);
9557 sched_init_domains(cpu_active_mask);
9558 mutex_unlock(&sched_domains_mutex);
9560 /* Move init over to a non-isolated CPU */
9561 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
9563 current->flags &= ~PF_NO_SETAFFINITY;
9564 sched_init_granularity();
9566 init_sched_rt_class();
9567 init_sched_dl_class();
9569 sched_smp_initialized = true;
9572 static int __init migration_init(void)
9574 sched_cpu_starting(smp_processor_id());
9577 early_initcall(migration_init);
9580 void __init sched_init_smp(void)
9582 sched_init_granularity();
9584 #endif /* CONFIG_SMP */
9586 int in_sched_functions(unsigned long addr)
9588 return in_lock_functions(addr) ||
9589 (addr >= (unsigned long)__sched_text_start
9590 && addr < (unsigned long)__sched_text_end);
9593 #ifdef CONFIG_CGROUP_SCHED
9595 * Default task group.
9596 * Every task in system belongs to this group at bootup.
9598 struct task_group root_task_group;
9599 LIST_HEAD(task_groups);
9601 /* Cacheline aligned slab cache for task_group */
9602 static struct kmem_cache *task_group_cache __read_mostly;
9605 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9606 DECLARE_PER_CPU(cpumask_var_t, select_rq_mask);
9608 void __init sched_init(void)
9610 unsigned long ptr = 0;
9613 /* Make sure the linker didn't screw up */
9614 BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9615 &fair_sched_class != &rt_sched_class + 1 ||
9616 &rt_sched_class != &dl_sched_class + 1);
9618 BUG_ON(&dl_sched_class != &stop_sched_class + 1);
9623 #ifdef CONFIG_FAIR_GROUP_SCHED
9624 ptr += 2 * nr_cpu_ids * sizeof(void **);
9626 #ifdef CONFIG_RT_GROUP_SCHED
9627 ptr += 2 * nr_cpu_ids * sizeof(void **);
9630 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9632 #ifdef CONFIG_FAIR_GROUP_SCHED
9633 root_task_group.se = (struct sched_entity **)ptr;
9634 ptr += nr_cpu_ids * sizeof(void **);
9636 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9637 ptr += nr_cpu_ids * sizeof(void **);
9639 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9640 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9641 #endif /* CONFIG_FAIR_GROUP_SCHED */
9642 #ifdef CONFIG_RT_GROUP_SCHED
9643 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9644 ptr += nr_cpu_ids * sizeof(void **);
9646 root_task_group.rt_rq = (struct rt_rq **)ptr;
9647 ptr += nr_cpu_ids * sizeof(void **);
9649 #endif /* CONFIG_RT_GROUP_SCHED */
9651 #ifdef CONFIG_CPUMASK_OFFSTACK
9652 for_each_possible_cpu(i) {
9653 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9654 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9655 per_cpu(select_rq_mask, i) = (cpumask_var_t)kzalloc_node(
9656 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9658 #endif /* CONFIG_CPUMASK_OFFSTACK */
9660 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9663 init_defrootdomain();
9666 #ifdef CONFIG_RT_GROUP_SCHED
9667 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9668 global_rt_period(), global_rt_runtime());
9669 #endif /* CONFIG_RT_GROUP_SCHED */
9671 #ifdef CONFIG_CGROUP_SCHED
9672 task_group_cache = KMEM_CACHE(task_group, 0);
9674 list_add(&root_task_group.list, &task_groups);
9675 INIT_LIST_HEAD(&root_task_group.children);
9676 INIT_LIST_HEAD(&root_task_group.siblings);
9677 autogroup_init(&init_task);
9678 #endif /* CONFIG_CGROUP_SCHED */
9680 for_each_possible_cpu(i) {
9684 raw_spin_lock_init(&rq->__lock);
9686 rq->calc_load_active = 0;
9687 rq->calc_load_update = jiffies + LOAD_FREQ;
9688 init_cfs_rq(&rq->cfs);
9689 init_rt_rq(&rq->rt);
9690 init_dl_rq(&rq->dl);
9691 #ifdef CONFIG_FAIR_GROUP_SCHED
9692 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9693 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9695 * How much CPU bandwidth does root_task_group get?
9697 * In case of task-groups formed thr' the cgroup filesystem, it
9698 * gets 100% of the CPU resources in the system. This overall
9699 * system CPU resource is divided among the tasks of
9700 * root_task_group and its child task-groups in a fair manner,
9701 * based on each entity's (task or task-group's) weight
9702 * (se->load.weight).
9704 * In other words, if root_task_group has 10 tasks of weight
9705 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9706 * then A0's share of the CPU resource is:
9708 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9710 * We achieve this by letting root_task_group's tasks sit
9711 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9713 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9714 #endif /* CONFIG_FAIR_GROUP_SCHED */
9716 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9717 #ifdef CONFIG_RT_GROUP_SCHED
9718 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9723 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9724 rq->balance_callback = &balance_push_callback;
9725 rq->active_balance = 0;
9726 rq->next_balance = jiffies;
9731 rq->avg_idle = 2*sysctl_sched_migration_cost;
9732 rq->wake_stamp = jiffies;
9733 rq->wake_avg_idle = rq->avg_idle;
9734 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9736 INIT_LIST_HEAD(&rq->cfs_tasks);
9738 rq_attach_root(rq, &def_root_domain);
9739 #ifdef CONFIG_NO_HZ_COMMON
9740 rq->last_blocked_load_update_tick = jiffies;
9741 atomic_set(&rq->nohz_flags, 0);
9743 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9745 #ifdef CONFIG_HOTPLUG_CPU
9746 rcuwait_init(&rq->hotplug_wait);
9748 #endif /* CONFIG_SMP */
9750 atomic_set(&rq->nr_iowait, 0);
9752 #ifdef CONFIG_SCHED_CORE
9754 rq->core_pick = NULL;
9755 rq->core_enabled = 0;
9756 rq->core_tree = RB_ROOT;
9757 rq->core_forceidle_count = 0;
9758 rq->core_forceidle_occupation = 0;
9759 rq->core_forceidle_start = 0;
9761 rq->core_cookie = 0UL;
9765 set_load_weight(&init_task, false);
9768 * The boot idle thread does lazy MMU switching as well:
9771 enter_lazy_tlb(&init_mm, current);
9774 * The idle task doesn't need the kthread struct to function, but it
9775 * is dressed up as a per-CPU kthread and thus needs to play the part
9776 * if we want to avoid special-casing it in code that deals with per-CPU
9779 WARN_ON(!set_kthread_struct(current));
9782 * Make us the idle thread. Technically, schedule() should not be
9783 * called from this thread, however somewhere below it might be,
9784 * but because we are the idle thread, we just pick up running again
9785 * when this runqueue becomes "idle".
9787 init_idle(current, smp_processor_id());
9789 calc_load_update = jiffies + LOAD_FREQ;
9792 idle_thread_set_boot_cpu();
9793 balance_push_set(smp_processor_id(), false);
9795 init_sched_fair_class();
9801 preempt_dynamic_init();
9803 scheduler_running = 1;
9806 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9808 void __might_sleep(const char *file, int line)
9810 unsigned int state = get_current_state();
9812 * Blocking primitives will set (and therefore destroy) current->state,
9813 * since we will exit with TASK_RUNNING make sure we enter with it,
9814 * otherwise we will destroy state.
9816 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9817 "do not call blocking ops when !TASK_RUNNING; "
9818 "state=%x set at [<%p>] %pS\n", state,
9819 (void *)current->task_state_change,
9820 (void *)current->task_state_change);
9822 __might_resched(file, line, 0);
9824 EXPORT_SYMBOL(__might_sleep);
9826 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9828 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9831 if (preempt_count() == preempt_offset)
9834 pr_err("Preemption disabled at:");
9835 print_ip_sym(KERN_ERR, ip);
9838 static inline bool resched_offsets_ok(unsigned int offsets)
9840 unsigned int nested = preempt_count();
9842 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9844 return nested == offsets;
9847 void __might_resched(const char *file, int line, unsigned int offsets)
9849 /* Ratelimiting timestamp: */
9850 static unsigned long prev_jiffy;
9852 unsigned long preempt_disable_ip;
9854 /* WARN_ON_ONCE() by default, no rate limit required: */
9857 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9858 !is_idle_task(current) && !current->non_block_count) ||
9859 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9863 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9865 prev_jiffy = jiffies;
9867 /* Save this before calling printk(), since that will clobber it: */
9868 preempt_disable_ip = get_preempt_disable_ip(current);
9870 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9872 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9873 in_atomic(), irqs_disabled(), current->non_block_count,
9874 current->pid, current->comm);
9875 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9876 offsets & MIGHT_RESCHED_PREEMPT_MASK);
9878 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9879 pr_err("RCU nest depth: %d, expected: %u\n",
9880 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9883 if (task_stack_end_corrupted(current))
9884 pr_emerg("Thread overran stack, or stack corrupted\n");
9886 debug_show_held_locks(current);
9887 if (irqs_disabled())
9888 print_irqtrace_events(current);
9890 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9891 preempt_disable_ip);
9894 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9896 EXPORT_SYMBOL(__might_resched);
9898 void __cant_sleep(const char *file, int line, int preempt_offset)
9900 static unsigned long prev_jiffy;
9902 if (irqs_disabled())
9905 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9908 if (preempt_count() > preempt_offset)
9911 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9913 prev_jiffy = jiffies;
9915 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9916 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9917 in_atomic(), irqs_disabled(),
9918 current->pid, current->comm);
9920 debug_show_held_locks(current);
9922 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9924 EXPORT_SYMBOL_GPL(__cant_sleep);
9927 void __cant_migrate(const char *file, int line)
9929 static unsigned long prev_jiffy;
9931 if (irqs_disabled())
9934 if (is_migration_disabled(current))
9937 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9940 if (preempt_count() > 0)
9943 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9945 prev_jiffy = jiffies;
9947 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9948 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9949 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9950 current->pid, current->comm);
9952 debug_show_held_locks(current);
9954 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9956 EXPORT_SYMBOL_GPL(__cant_migrate);
9960 #ifdef CONFIG_MAGIC_SYSRQ
9961 void normalize_rt_tasks(void)
9963 struct task_struct *g, *p;
9964 struct sched_attr attr = {
9965 .sched_policy = SCHED_NORMAL,
9968 read_lock(&tasklist_lock);
9969 for_each_process_thread(g, p) {
9971 * Only normalize user tasks:
9973 if (p->flags & PF_KTHREAD)
9976 p->se.exec_start = 0;
9977 schedstat_set(p->stats.wait_start, 0);
9978 schedstat_set(p->stats.sleep_start, 0);
9979 schedstat_set(p->stats.block_start, 0);
9981 if (!dl_task(p) && !rt_task(p)) {
9983 * Renice negative nice level userspace
9986 if (task_nice(p) < 0)
9987 set_user_nice(p, 0);
9991 __sched_setscheduler(p, &attr, false, false);
9993 read_unlock(&tasklist_lock);
9996 #endif /* CONFIG_MAGIC_SYSRQ */
9998 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
10000 * These functions are only useful for the IA64 MCA handling, or kdb.
10002 * They can only be called when the whole system has been
10003 * stopped - every CPU needs to be quiescent, and no scheduling
10004 * activity can take place. Using them for anything else would
10005 * be a serious bug, and as a result, they aren't even visible
10006 * under any other configuration.
10010 * curr_task - return the current task for a given CPU.
10011 * @cpu: the processor in question.
10013 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10015 * Return: The current task for @cpu.
10017 struct task_struct *curr_task(int cpu)
10019 return cpu_curr(cpu);
10022 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
10026 * ia64_set_curr_task - set the current task for a given CPU.
10027 * @cpu: the processor in question.
10028 * @p: the task pointer to set.
10030 * Description: This function must only be used when non-maskable interrupts
10031 * are serviced on a separate stack. It allows the architecture to switch the
10032 * notion of the current task on a CPU in a non-blocking manner. This function
10033 * must be called with all CPU's synchronized, and interrupts disabled, the
10034 * and caller must save the original value of the current task (see
10035 * curr_task() above) and restore that value before reenabling interrupts and
10036 * re-starting the system.
10038 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10040 void ia64_set_curr_task(int cpu, struct task_struct *p)
10047 #ifdef CONFIG_CGROUP_SCHED
10048 /* task_group_lock serializes the addition/removal of task groups */
10049 static DEFINE_SPINLOCK(task_group_lock);
10051 static inline void alloc_uclamp_sched_group(struct task_group *tg,
10052 struct task_group *parent)
10054 #ifdef CONFIG_UCLAMP_TASK_GROUP
10055 enum uclamp_id clamp_id;
10057 for_each_clamp_id(clamp_id) {
10058 uclamp_se_set(&tg->uclamp_req[clamp_id],
10059 uclamp_none(clamp_id), false);
10060 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
10065 static void sched_free_group(struct task_group *tg)
10067 free_fair_sched_group(tg);
10068 free_rt_sched_group(tg);
10069 autogroup_free(tg);
10070 kmem_cache_free(task_group_cache, tg);
10073 static void sched_free_group_rcu(struct rcu_head *rcu)
10075 sched_free_group(container_of(rcu, struct task_group, rcu));
10078 static void sched_unregister_group(struct task_group *tg)
10080 unregister_fair_sched_group(tg);
10081 unregister_rt_sched_group(tg);
10083 * We have to wait for yet another RCU grace period to expire, as
10084 * print_cfs_stats() might run concurrently.
10086 call_rcu(&tg->rcu, sched_free_group_rcu);
10089 /* allocate runqueue etc for a new task group */
10090 struct task_group *sched_create_group(struct task_group *parent)
10092 struct task_group *tg;
10094 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
10096 return ERR_PTR(-ENOMEM);
10098 if (!alloc_fair_sched_group(tg, parent))
10101 if (!alloc_rt_sched_group(tg, parent))
10104 alloc_uclamp_sched_group(tg, parent);
10109 sched_free_group(tg);
10110 return ERR_PTR(-ENOMEM);
10113 void sched_online_group(struct task_group *tg, struct task_group *parent)
10115 unsigned long flags;
10117 spin_lock_irqsave(&task_group_lock, flags);
10118 list_add_rcu(&tg->list, &task_groups);
10120 /* Root should already exist: */
10123 tg->parent = parent;
10124 INIT_LIST_HEAD(&tg->children);
10125 list_add_rcu(&tg->siblings, &parent->children);
10126 spin_unlock_irqrestore(&task_group_lock, flags);
10128 online_fair_sched_group(tg);
10131 /* rcu callback to free various structures associated with a task group */
10132 static void sched_unregister_group_rcu(struct rcu_head *rhp)
10134 /* Now it should be safe to free those cfs_rqs: */
10135 sched_unregister_group(container_of(rhp, struct task_group, rcu));
10138 void sched_destroy_group(struct task_group *tg)
10140 /* Wait for possible concurrent references to cfs_rqs complete: */
10141 call_rcu(&tg->rcu, sched_unregister_group_rcu);
10144 void sched_release_group(struct task_group *tg)
10146 unsigned long flags;
10149 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10150 * sched_cfs_period_timer()).
10152 * For this to be effective, we have to wait for all pending users of
10153 * this task group to leave their RCU critical section to ensure no new
10154 * user will see our dying task group any more. Specifically ensure
10155 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10157 * We therefore defer calling unregister_fair_sched_group() to
10158 * sched_unregister_group() which is guarantied to get called only after the
10159 * current RCU grace period has expired.
10161 spin_lock_irqsave(&task_group_lock, flags);
10162 list_del_rcu(&tg->list);
10163 list_del_rcu(&tg->siblings);
10164 spin_unlock_irqrestore(&task_group_lock, flags);
10167 static void sched_change_group(struct task_struct *tsk, int type)
10169 struct task_group *tg;
10172 * All callers are synchronized by task_rq_lock(); we do not use RCU
10173 * which is pointless here. Thus, we pass "true" to task_css_check()
10174 * to prevent lockdep warnings.
10176 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10177 struct task_group, css);
10178 tg = autogroup_task_group(tsk, tg);
10179 tsk->sched_task_group = tg;
10181 #ifdef CONFIG_FAIR_GROUP_SCHED
10182 if (tsk->sched_class->task_change_group)
10183 tsk->sched_class->task_change_group(tsk, type);
10186 set_task_rq(tsk, task_cpu(tsk));
10190 * Change task's runqueue when it moves between groups.
10192 * The caller of this function should have put the task in its new group by
10193 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10196 void sched_move_task(struct task_struct *tsk)
10198 int queued, running, queue_flags =
10199 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10200 struct rq_flags rf;
10203 rq = task_rq_lock(tsk, &rf);
10204 update_rq_clock(rq);
10206 running = task_current(rq, tsk);
10207 queued = task_on_rq_queued(tsk);
10210 dequeue_task(rq, tsk, queue_flags);
10212 put_prev_task(rq, tsk);
10214 sched_change_group(tsk, TASK_MOVE_GROUP);
10217 enqueue_task(rq, tsk, queue_flags);
10219 set_next_task(rq, tsk);
10221 * After changing group, the running task may have joined a
10222 * throttled one but it's still the running task. Trigger a
10223 * resched to make sure that task can still run.
10228 task_rq_unlock(rq, tsk, &rf);
10231 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10233 return css ? container_of(css, struct task_group, css) : NULL;
10236 static struct cgroup_subsys_state *
10237 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10239 struct task_group *parent = css_tg(parent_css);
10240 struct task_group *tg;
10243 /* This is early initialization for the top cgroup */
10244 return &root_task_group.css;
10247 tg = sched_create_group(parent);
10249 return ERR_PTR(-ENOMEM);
10254 /* Expose task group only after completing cgroup initialization */
10255 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10257 struct task_group *tg = css_tg(css);
10258 struct task_group *parent = css_tg(css->parent);
10261 sched_online_group(tg, parent);
10263 #ifdef CONFIG_UCLAMP_TASK_GROUP
10264 /* Propagate the effective uclamp value for the new group */
10265 mutex_lock(&uclamp_mutex);
10267 cpu_util_update_eff(css);
10269 mutex_unlock(&uclamp_mutex);
10275 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10277 struct task_group *tg = css_tg(css);
10279 sched_release_group(tg);
10282 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10284 struct task_group *tg = css_tg(css);
10287 * Relies on the RCU grace period between css_released() and this.
10289 sched_unregister_group(tg);
10293 * This is called before wake_up_new_task(), therefore we really only
10294 * have to set its group bits, all the other stuff does not apply.
10296 static void cpu_cgroup_fork(struct task_struct *task)
10298 struct rq_flags rf;
10301 rq = task_rq_lock(task, &rf);
10303 update_rq_clock(rq);
10304 sched_change_group(task, TASK_SET_GROUP);
10306 task_rq_unlock(rq, task, &rf);
10309 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10311 struct task_struct *task;
10312 struct cgroup_subsys_state *css;
10315 cgroup_taskset_for_each(task, css, tset) {
10316 #ifdef CONFIG_RT_GROUP_SCHED
10317 if (!sched_rt_can_attach(css_tg(css), task))
10321 * Serialize against wake_up_new_task() such that if it's
10322 * running, we're sure to observe its full state.
10324 raw_spin_lock_irq(&task->pi_lock);
10326 * Avoid calling sched_move_task() before wake_up_new_task()
10327 * has happened. This would lead to problems with PELT, due to
10328 * move wanting to detach+attach while we're not attached yet.
10330 if (READ_ONCE(task->__state) == TASK_NEW)
10332 raw_spin_unlock_irq(&task->pi_lock);
10340 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10342 struct task_struct *task;
10343 struct cgroup_subsys_state *css;
10345 cgroup_taskset_for_each(task, css, tset)
10346 sched_move_task(task);
10349 #ifdef CONFIG_UCLAMP_TASK_GROUP
10350 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10352 struct cgroup_subsys_state *top_css = css;
10353 struct uclamp_se *uc_parent = NULL;
10354 struct uclamp_se *uc_se = NULL;
10355 unsigned int eff[UCLAMP_CNT];
10356 enum uclamp_id clamp_id;
10357 unsigned int clamps;
10359 lockdep_assert_held(&uclamp_mutex);
10360 SCHED_WARN_ON(!rcu_read_lock_held());
10362 css_for_each_descendant_pre(css, top_css) {
10363 uc_parent = css_tg(css)->parent
10364 ? css_tg(css)->parent->uclamp : NULL;
10366 for_each_clamp_id(clamp_id) {
10367 /* Assume effective clamps matches requested clamps */
10368 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10369 /* Cap effective clamps with parent's effective clamps */
10371 eff[clamp_id] > uc_parent[clamp_id].value) {
10372 eff[clamp_id] = uc_parent[clamp_id].value;
10375 /* Ensure protection is always capped by limit */
10376 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10378 /* Propagate most restrictive effective clamps */
10380 uc_se = css_tg(css)->uclamp;
10381 for_each_clamp_id(clamp_id) {
10382 if (eff[clamp_id] == uc_se[clamp_id].value)
10384 uc_se[clamp_id].value = eff[clamp_id];
10385 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10386 clamps |= (0x1 << clamp_id);
10389 css = css_rightmost_descendant(css);
10393 /* Immediately update descendants RUNNABLE tasks */
10394 uclamp_update_active_tasks(css);
10399 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10400 * C expression. Since there is no way to convert a macro argument (N) into a
10401 * character constant, use two levels of macros.
10403 #define _POW10(exp) ((unsigned int)1e##exp)
10404 #define POW10(exp) _POW10(exp)
10406 struct uclamp_request {
10407 #define UCLAMP_PERCENT_SHIFT 2
10408 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10414 static inline struct uclamp_request
10415 capacity_from_percent(char *buf)
10417 struct uclamp_request req = {
10418 .percent = UCLAMP_PERCENT_SCALE,
10419 .util = SCHED_CAPACITY_SCALE,
10424 if (strcmp(buf, "max")) {
10425 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10429 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10434 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10435 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10441 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10442 size_t nbytes, loff_t off,
10443 enum uclamp_id clamp_id)
10445 struct uclamp_request req;
10446 struct task_group *tg;
10448 req = capacity_from_percent(buf);
10452 static_branch_enable(&sched_uclamp_used);
10454 mutex_lock(&uclamp_mutex);
10457 tg = css_tg(of_css(of));
10458 if (tg->uclamp_req[clamp_id].value != req.util)
10459 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10462 * Because of not recoverable conversion rounding we keep track of the
10463 * exact requested value
10465 tg->uclamp_pct[clamp_id] = req.percent;
10467 /* Update effective clamps to track the most restrictive value */
10468 cpu_util_update_eff(of_css(of));
10471 mutex_unlock(&uclamp_mutex);
10476 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10477 char *buf, size_t nbytes,
10480 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10483 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10484 char *buf, size_t nbytes,
10487 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10490 static inline void cpu_uclamp_print(struct seq_file *sf,
10491 enum uclamp_id clamp_id)
10493 struct task_group *tg;
10499 tg = css_tg(seq_css(sf));
10500 util_clamp = tg->uclamp_req[clamp_id].value;
10503 if (util_clamp == SCHED_CAPACITY_SCALE) {
10504 seq_puts(sf, "max\n");
10508 percent = tg->uclamp_pct[clamp_id];
10509 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10510 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10513 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10515 cpu_uclamp_print(sf, UCLAMP_MIN);
10519 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10521 cpu_uclamp_print(sf, UCLAMP_MAX);
10524 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10526 #ifdef CONFIG_FAIR_GROUP_SCHED
10527 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10528 struct cftype *cftype, u64 shareval)
10530 if (shareval > scale_load_down(ULONG_MAX))
10531 shareval = MAX_SHARES;
10532 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10535 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10536 struct cftype *cft)
10538 struct task_group *tg = css_tg(css);
10540 return (u64) scale_load_down(tg->shares);
10543 #ifdef CONFIG_CFS_BANDWIDTH
10544 static DEFINE_MUTEX(cfs_constraints_mutex);
10546 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10547 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10548 /* More than 203 days if BW_SHIFT equals 20. */
10549 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10551 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10553 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10556 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10557 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10559 if (tg == &root_task_group)
10563 * Ensure we have at some amount of bandwidth every period. This is
10564 * to prevent reaching a state of large arrears when throttled via
10565 * entity_tick() resulting in prolonged exit starvation.
10567 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10571 * Likewise, bound things on the other side by preventing insane quota
10572 * periods. This also allows us to normalize in computing quota
10575 if (period > max_cfs_quota_period)
10579 * Bound quota to defend quota against overflow during bandwidth shift.
10581 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10584 if (quota != RUNTIME_INF && (burst > quota ||
10585 burst + quota > max_cfs_runtime))
10589 * Prevent race between setting of cfs_rq->runtime_enabled and
10590 * unthrottle_offline_cfs_rqs().
10593 mutex_lock(&cfs_constraints_mutex);
10594 ret = __cfs_schedulable(tg, period, quota);
10598 runtime_enabled = quota != RUNTIME_INF;
10599 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10601 * If we need to toggle cfs_bandwidth_used, off->on must occur
10602 * before making related changes, and on->off must occur afterwards
10604 if (runtime_enabled && !runtime_was_enabled)
10605 cfs_bandwidth_usage_inc();
10606 raw_spin_lock_irq(&cfs_b->lock);
10607 cfs_b->period = ns_to_ktime(period);
10608 cfs_b->quota = quota;
10609 cfs_b->burst = burst;
10611 __refill_cfs_bandwidth_runtime(cfs_b);
10613 /* Restart the period timer (if active) to handle new period expiry: */
10614 if (runtime_enabled)
10615 start_cfs_bandwidth(cfs_b);
10617 raw_spin_unlock_irq(&cfs_b->lock);
10619 for_each_online_cpu(i) {
10620 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10621 struct rq *rq = cfs_rq->rq;
10622 struct rq_flags rf;
10624 rq_lock_irq(rq, &rf);
10625 cfs_rq->runtime_enabled = runtime_enabled;
10626 cfs_rq->runtime_remaining = 0;
10628 if (cfs_rq->throttled)
10629 unthrottle_cfs_rq(cfs_rq);
10630 rq_unlock_irq(rq, &rf);
10632 if (runtime_was_enabled && !runtime_enabled)
10633 cfs_bandwidth_usage_dec();
10635 mutex_unlock(&cfs_constraints_mutex);
10636 cpus_read_unlock();
10641 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10643 u64 quota, period, burst;
10645 period = ktime_to_ns(tg->cfs_bandwidth.period);
10646 burst = tg->cfs_bandwidth.burst;
10647 if (cfs_quota_us < 0)
10648 quota = RUNTIME_INF;
10649 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10650 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10654 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10657 static long tg_get_cfs_quota(struct task_group *tg)
10661 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10664 quota_us = tg->cfs_bandwidth.quota;
10665 do_div(quota_us, NSEC_PER_USEC);
10670 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10672 u64 quota, period, burst;
10674 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10677 period = (u64)cfs_period_us * NSEC_PER_USEC;
10678 quota = tg->cfs_bandwidth.quota;
10679 burst = tg->cfs_bandwidth.burst;
10681 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10684 static long tg_get_cfs_period(struct task_group *tg)
10688 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10689 do_div(cfs_period_us, NSEC_PER_USEC);
10691 return cfs_period_us;
10694 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10696 u64 quota, period, burst;
10698 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10701 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10702 period = ktime_to_ns(tg->cfs_bandwidth.period);
10703 quota = tg->cfs_bandwidth.quota;
10705 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10708 static long tg_get_cfs_burst(struct task_group *tg)
10712 burst_us = tg->cfs_bandwidth.burst;
10713 do_div(burst_us, NSEC_PER_USEC);
10718 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10719 struct cftype *cft)
10721 return tg_get_cfs_quota(css_tg(css));
10724 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10725 struct cftype *cftype, s64 cfs_quota_us)
10727 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10730 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10731 struct cftype *cft)
10733 return tg_get_cfs_period(css_tg(css));
10736 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10737 struct cftype *cftype, u64 cfs_period_us)
10739 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10742 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10743 struct cftype *cft)
10745 return tg_get_cfs_burst(css_tg(css));
10748 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10749 struct cftype *cftype, u64 cfs_burst_us)
10751 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10754 struct cfs_schedulable_data {
10755 struct task_group *tg;
10760 * normalize group quota/period to be quota/max_period
10761 * note: units are usecs
10763 static u64 normalize_cfs_quota(struct task_group *tg,
10764 struct cfs_schedulable_data *d)
10769 period = d->period;
10772 period = tg_get_cfs_period(tg);
10773 quota = tg_get_cfs_quota(tg);
10776 /* note: these should typically be equivalent */
10777 if (quota == RUNTIME_INF || quota == -1)
10778 return RUNTIME_INF;
10780 return to_ratio(period, quota);
10783 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10785 struct cfs_schedulable_data *d = data;
10786 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10787 s64 quota = 0, parent_quota = -1;
10790 quota = RUNTIME_INF;
10792 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10794 quota = normalize_cfs_quota(tg, d);
10795 parent_quota = parent_b->hierarchical_quota;
10798 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10799 * always take the min. On cgroup1, only inherit when no
10802 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10803 quota = min(quota, parent_quota);
10805 if (quota == RUNTIME_INF)
10806 quota = parent_quota;
10807 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10811 cfs_b->hierarchical_quota = quota;
10816 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10819 struct cfs_schedulable_data data = {
10825 if (quota != RUNTIME_INF) {
10826 do_div(data.period, NSEC_PER_USEC);
10827 do_div(data.quota, NSEC_PER_USEC);
10831 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10837 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10839 struct task_group *tg = css_tg(seq_css(sf));
10840 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10842 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10843 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10844 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10846 if (schedstat_enabled() && tg != &root_task_group) {
10847 struct sched_statistics *stats;
10851 for_each_possible_cpu(i) {
10852 stats = __schedstats_from_se(tg->se[i]);
10853 ws += schedstat_val(stats->wait_sum);
10856 seq_printf(sf, "wait_sum %llu\n", ws);
10859 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10860 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10864 #endif /* CONFIG_CFS_BANDWIDTH */
10865 #endif /* CONFIG_FAIR_GROUP_SCHED */
10867 #ifdef CONFIG_RT_GROUP_SCHED
10868 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10869 struct cftype *cft, s64 val)
10871 return sched_group_set_rt_runtime(css_tg(css), val);
10874 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10875 struct cftype *cft)
10877 return sched_group_rt_runtime(css_tg(css));
10880 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10881 struct cftype *cftype, u64 rt_period_us)
10883 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10886 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10887 struct cftype *cft)
10889 return sched_group_rt_period(css_tg(css));
10891 #endif /* CONFIG_RT_GROUP_SCHED */
10893 #ifdef CONFIG_FAIR_GROUP_SCHED
10894 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10895 struct cftype *cft)
10897 return css_tg(css)->idle;
10900 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10901 struct cftype *cft, s64 idle)
10903 return sched_group_set_idle(css_tg(css), idle);
10907 static struct cftype cpu_legacy_files[] = {
10908 #ifdef CONFIG_FAIR_GROUP_SCHED
10911 .read_u64 = cpu_shares_read_u64,
10912 .write_u64 = cpu_shares_write_u64,
10916 .read_s64 = cpu_idle_read_s64,
10917 .write_s64 = cpu_idle_write_s64,
10920 #ifdef CONFIG_CFS_BANDWIDTH
10922 .name = "cfs_quota_us",
10923 .read_s64 = cpu_cfs_quota_read_s64,
10924 .write_s64 = cpu_cfs_quota_write_s64,
10927 .name = "cfs_period_us",
10928 .read_u64 = cpu_cfs_period_read_u64,
10929 .write_u64 = cpu_cfs_period_write_u64,
10932 .name = "cfs_burst_us",
10933 .read_u64 = cpu_cfs_burst_read_u64,
10934 .write_u64 = cpu_cfs_burst_write_u64,
10938 .seq_show = cpu_cfs_stat_show,
10941 #ifdef CONFIG_RT_GROUP_SCHED
10943 .name = "rt_runtime_us",
10944 .read_s64 = cpu_rt_runtime_read,
10945 .write_s64 = cpu_rt_runtime_write,
10948 .name = "rt_period_us",
10949 .read_u64 = cpu_rt_period_read_uint,
10950 .write_u64 = cpu_rt_period_write_uint,
10953 #ifdef CONFIG_UCLAMP_TASK_GROUP
10955 .name = "uclamp.min",
10956 .flags = CFTYPE_NOT_ON_ROOT,
10957 .seq_show = cpu_uclamp_min_show,
10958 .write = cpu_uclamp_min_write,
10961 .name = "uclamp.max",
10962 .flags = CFTYPE_NOT_ON_ROOT,
10963 .seq_show = cpu_uclamp_max_show,
10964 .write = cpu_uclamp_max_write,
10967 { } /* Terminate */
10970 static int cpu_extra_stat_show(struct seq_file *sf,
10971 struct cgroup_subsys_state *css)
10973 #ifdef CONFIG_CFS_BANDWIDTH
10975 struct task_group *tg = css_tg(css);
10976 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10977 u64 throttled_usec, burst_usec;
10979 throttled_usec = cfs_b->throttled_time;
10980 do_div(throttled_usec, NSEC_PER_USEC);
10981 burst_usec = cfs_b->burst_time;
10982 do_div(burst_usec, NSEC_PER_USEC);
10984 seq_printf(sf, "nr_periods %d\n"
10985 "nr_throttled %d\n"
10986 "throttled_usec %llu\n"
10988 "burst_usec %llu\n",
10989 cfs_b->nr_periods, cfs_b->nr_throttled,
10990 throttled_usec, cfs_b->nr_burst, burst_usec);
10996 #ifdef CONFIG_FAIR_GROUP_SCHED
10997 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10998 struct cftype *cft)
11000 struct task_group *tg = css_tg(css);
11001 u64 weight = scale_load_down(tg->shares);
11003 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11006 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11007 struct cftype *cft, u64 weight)
11010 * cgroup weight knobs should use the common MIN, DFL and MAX
11011 * values which are 1, 100 and 10000 respectively. While it loses
11012 * a bit of range on both ends, it maps pretty well onto the shares
11013 * value used by scheduler and the round-trip conversions preserve
11014 * the original value over the entire range.
11016 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11019 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11021 return sched_group_set_shares(css_tg(css), scale_load(weight));
11024 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11025 struct cftype *cft)
11027 unsigned long weight = scale_load_down(css_tg(css)->shares);
11028 int last_delta = INT_MAX;
11031 /* find the closest nice value to the current weight */
11032 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11033 delta = abs(sched_prio_to_weight[prio] - weight);
11034 if (delta >= last_delta)
11036 last_delta = delta;
11039 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11042 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11043 struct cftype *cft, s64 nice)
11045 unsigned long weight;
11048 if (nice < MIN_NICE || nice > MAX_NICE)
11051 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11052 idx = array_index_nospec(idx, 40);
11053 weight = sched_prio_to_weight[idx];
11055 return sched_group_set_shares(css_tg(css), scale_load(weight));
11059 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11060 long period, long quota)
11063 seq_puts(sf, "max");
11065 seq_printf(sf, "%ld", quota);
11067 seq_printf(sf, " %ld\n", period);
11070 /* caller should put the current value in *@periodp before calling */
11071 static int __maybe_unused cpu_period_quota_parse(char *buf,
11072 u64 *periodp, u64 *quotap)
11074 char tok[21]; /* U64_MAX */
11076 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
11079 *periodp *= NSEC_PER_USEC;
11081 if (sscanf(tok, "%llu", quotap))
11082 *quotap *= NSEC_PER_USEC;
11083 else if (!strcmp(tok, "max"))
11084 *quotap = RUNTIME_INF;
11091 #ifdef CONFIG_CFS_BANDWIDTH
11092 static int cpu_max_show(struct seq_file *sf, void *v)
11094 struct task_group *tg = css_tg(seq_css(sf));
11096 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11100 static ssize_t cpu_max_write(struct kernfs_open_file *of,
11101 char *buf, size_t nbytes, loff_t off)
11103 struct task_group *tg = css_tg(of_css(of));
11104 u64 period = tg_get_cfs_period(tg);
11105 u64 burst = tg_get_cfs_burst(tg);
11109 ret = cpu_period_quota_parse(buf, &period, "a);
11111 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
11112 return ret ?: nbytes;
11116 static struct cftype cpu_files[] = {
11117 #ifdef CONFIG_FAIR_GROUP_SCHED
11120 .flags = CFTYPE_NOT_ON_ROOT,
11121 .read_u64 = cpu_weight_read_u64,
11122 .write_u64 = cpu_weight_write_u64,
11125 .name = "weight.nice",
11126 .flags = CFTYPE_NOT_ON_ROOT,
11127 .read_s64 = cpu_weight_nice_read_s64,
11128 .write_s64 = cpu_weight_nice_write_s64,
11132 .flags = CFTYPE_NOT_ON_ROOT,
11133 .read_s64 = cpu_idle_read_s64,
11134 .write_s64 = cpu_idle_write_s64,
11137 #ifdef CONFIG_CFS_BANDWIDTH
11140 .flags = CFTYPE_NOT_ON_ROOT,
11141 .seq_show = cpu_max_show,
11142 .write = cpu_max_write,
11145 .name = "max.burst",
11146 .flags = CFTYPE_NOT_ON_ROOT,
11147 .read_u64 = cpu_cfs_burst_read_u64,
11148 .write_u64 = cpu_cfs_burst_write_u64,
11151 #ifdef CONFIG_UCLAMP_TASK_GROUP
11153 .name = "uclamp.min",
11154 .flags = CFTYPE_NOT_ON_ROOT,
11155 .seq_show = cpu_uclamp_min_show,
11156 .write = cpu_uclamp_min_write,
11159 .name = "uclamp.max",
11160 .flags = CFTYPE_NOT_ON_ROOT,
11161 .seq_show = cpu_uclamp_max_show,
11162 .write = cpu_uclamp_max_write,
11165 { } /* terminate */
11168 struct cgroup_subsys cpu_cgrp_subsys = {
11169 .css_alloc = cpu_cgroup_css_alloc,
11170 .css_online = cpu_cgroup_css_online,
11171 .css_released = cpu_cgroup_css_released,
11172 .css_free = cpu_cgroup_css_free,
11173 .css_extra_stat_show = cpu_extra_stat_show,
11174 .fork = cpu_cgroup_fork,
11175 .can_attach = cpu_cgroup_can_attach,
11176 .attach = cpu_cgroup_attach,
11177 .legacy_cftypes = cpu_legacy_files,
11178 .dfl_cftypes = cpu_files,
11179 .early_init = true,
11183 #endif /* CONFIG_CGROUP_SCHED */
11185 void dump_cpu_task(int cpu)
11187 if (cpu == smp_processor_id() && in_hardirq()) {
11188 struct pt_regs *regs;
11190 regs = get_irq_regs();
11197 if (trigger_single_cpu_backtrace(cpu))
11200 pr_info("Task dump for CPU %d:\n", cpu);
11201 sched_show_task(cpu_curr(cpu));
11205 * Nice levels are multiplicative, with a gentle 10% change for every
11206 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11207 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11208 * that remained on nice 0.
11210 * The "10% effect" is relative and cumulative: from _any_ nice level,
11211 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11212 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11213 * If a task goes up by ~10% and another task goes down by ~10% then
11214 * the relative distance between them is ~25%.)
11216 const int sched_prio_to_weight[40] = {
11217 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11218 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11219 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11220 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11221 /* 0 */ 1024, 820, 655, 526, 423,
11222 /* 5 */ 335, 272, 215, 172, 137,
11223 /* 10 */ 110, 87, 70, 56, 45,
11224 /* 15 */ 36, 29, 23, 18, 15,
11228 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11230 * In cases where the weight does not change often, we can use the
11231 * precalculated inverse to speed up arithmetics by turning divisions
11232 * into multiplications:
11234 const u32 sched_prio_to_wmult[40] = {
11235 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11236 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11237 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11238 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11239 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11240 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11241 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11242 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11245 void call_trace_sched_update_nr_running(struct rq *rq, int count)
11247 trace_sched_update_nr_running_tp(rq, count);