1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #include <linux/highmem.h>
10 #include <linux/hrtimer_api.h>
11 #include <linux/ktime_api.h>
12 #include <linux/sched/signal.h>
13 #include <linux/syscalls_api.h>
14 #include <linux/debug_locks.h>
15 #include <linux/prefetch.h>
16 #include <linux/capability.h>
17 #include <linux/pgtable_api.h>
18 #include <linux/wait_bit.h>
19 #include <linux/jiffies.h>
20 #include <linux/spinlock_api.h>
21 #include <linux/cpumask_api.h>
22 #include <linux/lockdep_api.h>
23 #include <linux/hardirq.h>
24 #include <linux/softirq.h>
25 #include <linux/refcount_api.h>
26 #include <linux/topology.h>
27 #include <linux/sched/clock.h>
28 #include <linux/sched/cond_resched.h>
29 #include <linux/sched/cputime.h>
30 #include <linux/sched/debug.h>
31 #include <linux/sched/hotplug.h>
32 #include <linux/sched/init.h>
33 #include <linux/sched/isolation.h>
34 #include <linux/sched/loadavg.h>
35 #include <linux/sched/mm.h>
36 #include <linux/sched/nohz.h>
37 #include <linux/sched/rseq_api.h>
38 #include <linux/sched/rt.h>
40 #include <linux/blkdev.h>
41 #include <linux/context_tracking.h>
42 #include <linux/cpuset.h>
43 #include <linux/delayacct.h>
44 #include <linux/init_task.h>
45 #include <linux/interrupt.h>
46 #include <linux/ioprio.h>
47 #include <linux/kallsyms.h>
48 #include <linux/kcov.h>
49 #include <linux/kprobes.h>
50 #include <linux/llist_api.h>
51 #include <linux/mmu_context.h>
52 #include <linux/mmzone.h>
53 #include <linux/mutex_api.h>
54 #include <linux/nmi.h>
55 #include <linux/nospec.h>
56 #include <linux/perf_event_api.h>
57 #include <linux/profile.h>
58 #include <linux/psi.h>
59 #include <linux/rcuwait_api.h>
60 #include <linux/sched/wake_q.h>
61 #include <linux/scs.h>
62 #include <linux/slab.h>
63 #include <linux/syscalls.h>
64 #include <linux/vtime.h>
65 #include <linux/wait_api.h>
66 #include <linux/workqueue_api.h>
68 #ifdef CONFIG_PREEMPT_DYNAMIC
69 # ifdef CONFIG_GENERIC_ENTRY
70 # include <linux/entry-common.h>
74 #include <uapi/linux/sched/types.h>
76 #include <asm/irq_regs.h>
77 #include <asm/switch_to.h>
80 #define CREATE_TRACE_POINTS
81 #include <linux/sched/rseq_api.h>
82 #include <trace/events/sched.h>
83 #include <trace/events/ipi.h>
84 #undef CREATE_TRACE_POINTS
88 #include "autogroup.h"
90 #include "autogroup.h"
95 #include "../workqueue_internal.h"
96 #include "../../io_uring/io-wq.h"
97 #include "../smpboot.h"
99 EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpu
);
100 EXPORT_TRACEPOINT_SYMBOL_GPL(ipi_send_cpumask
);
103 * Export tracepoints that act as a bare tracehook (ie: have no trace event
104 * associated with them) to allow external modules to probe them.
106 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp
);
107 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp
);
108 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp
);
109 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp
);
110 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp
);
111 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp
);
112 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp
);
113 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp
);
114 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp
);
115 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp
);
116 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp
);
118 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
120 #ifdef CONFIG_SCHED_DEBUG
122 * Debugging: various feature bits
124 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
125 * sysctl_sched_features, defined in sched.h, to allow constants propagation
126 * at compile time and compiler optimization based on features default.
128 #define SCHED_FEAT(name, enabled) \
129 (1UL << __SCHED_FEAT_##name) * enabled |
130 const_debug
unsigned int sysctl_sched_features
=
131 #include "features.h"
136 * Print a warning if need_resched is set for the given duration (if
137 * LATENCY_WARN is enabled).
139 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
142 __read_mostly
int sysctl_resched_latency_warn_ms
= 100;
143 __read_mostly
int sysctl_resched_latency_warn_once
= 1;
144 #endif /* CONFIG_SCHED_DEBUG */
147 * Number of tasks to iterate in a single balance run.
148 * Limited because this is done with IRQs disabled.
150 const_debug
unsigned int sysctl_sched_nr_migrate
= SCHED_NR_MIGRATE_BREAK
;
152 __read_mostly
int scheduler_running
;
154 #ifdef CONFIG_SCHED_CORE
156 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled
);
158 /* kernel prio, less is more */
159 static inline int __task_prio(const struct task_struct
*p
)
161 if (p
->sched_class
== &stop_sched_class
) /* trumps deadline */
164 if (rt_prio(p
->prio
)) /* includes deadline */
165 return p
->prio
; /* [-1, 99] */
167 if (p
->sched_class
== &idle_sched_class
)
168 return MAX_RT_PRIO
+ NICE_WIDTH
; /* 140 */
170 return MAX_RT_PRIO
+ MAX_NICE
; /* 120, squash fair */
180 /* real prio, less is less */
181 static inline bool prio_less(const struct task_struct
*a
,
182 const struct task_struct
*b
, bool in_fi
)
185 int pa
= __task_prio(a
), pb
= __task_prio(b
);
193 if (pa
== -1) /* dl_prio() doesn't work because of stop_class above */
194 return !dl_time_before(a
->dl
.deadline
, b
->dl
.deadline
);
196 if (pa
== MAX_RT_PRIO
+ MAX_NICE
) /* fair */
197 return cfs_prio_less(a
, b
, in_fi
);
202 static inline bool __sched_core_less(const struct task_struct
*a
,
203 const struct task_struct
*b
)
205 if (a
->core_cookie
< b
->core_cookie
)
208 if (a
->core_cookie
> b
->core_cookie
)
211 /* flip prio, so high prio is leftmost */
212 if (prio_less(b
, a
, !!task_rq(a
)->core
->core_forceidle_count
))
218 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
220 static inline bool rb_sched_core_less(struct rb_node
*a
, const struct rb_node
*b
)
222 return __sched_core_less(__node_2_sc(a
), __node_2_sc(b
));
225 static inline int rb_sched_core_cmp(const void *key
, const struct rb_node
*node
)
227 const struct task_struct
*p
= __node_2_sc(node
);
228 unsigned long cookie
= (unsigned long)key
;
230 if (cookie
< p
->core_cookie
)
233 if (cookie
> p
->core_cookie
)
239 void sched_core_enqueue(struct rq
*rq
, struct task_struct
*p
)
241 rq
->core
->core_task_seq
++;
246 rb_add(&p
->core_node
, &rq
->core_tree
, rb_sched_core_less
);
249 void sched_core_dequeue(struct rq
*rq
, struct task_struct
*p
, int flags
)
251 rq
->core
->core_task_seq
++;
253 if (sched_core_enqueued(p
)) {
254 rb_erase(&p
->core_node
, &rq
->core_tree
);
255 RB_CLEAR_NODE(&p
->core_node
);
259 * Migrating the last task off the cpu, with the cpu in forced idle
260 * state. Reschedule to create an accounting edge for forced idle,
261 * and re-examine whether the core is still in forced idle state.
263 if (!(flags
& DEQUEUE_SAVE
) && rq
->nr_running
== 1 &&
264 rq
->core
->core_forceidle_count
&& rq
->curr
== rq
->idle
)
268 static int sched_task_is_throttled(struct task_struct
*p
, int cpu
)
270 if (p
->sched_class
->task_is_throttled
)
271 return p
->sched_class
->task_is_throttled(p
, cpu
);
276 static struct task_struct
*sched_core_next(struct task_struct
*p
, unsigned long cookie
)
278 struct rb_node
*node
= &p
->core_node
;
279 int cpu
= task_cpu(p
);
282 node
= rb_next(node
);
286 p
= __node_2_sc(node
);
287 if (p
->core_cookie
!= cookie
)
290 } while (sched_task_is_throttled(p
, cpu
));
296 * Find left-most (aka, highest priority) and unthrottled task matching @cookie.
297 * If no suitable task is found, NULL will be returned.
299 static struct task_struct
*sched_core_find(struct rq
*rq
, unsigned long cookie
)
301 struct task_struct
*p
;
302 struct rb_node
*node
;
304 node
= rb_find_first((void *)cookie
, &rq
->core_tree
, rb_sched_core_cmp
);
308 p
= __node_2_sc(node
);
309 if (!sched_task_is_throttled(p
, rq
->cpu
))
312 return sched_core_next(p
, cookie
);
316 * Magic required such that:
318 * raw_spin_rq_lock(rq);
320 * raw_spin_rq_unlock(rq);
322 * ends up locking and unlocking the _same_ lock, and all CPUs
323 * always agree on what rq has what lock.
325 * XXX entirely possible to selectively enable cores, don't bother for now.
328 static DEFINE_MUTEX(sched_core_mutex
);
329 static atomic_t sched_core_count
;
330 static struct cpumask sched_core_mask
;
332 static void sched_core_lock(int cpu
, unsigned long *flags
)
334 const struct cpumask
*smt_mask
= cpu_smt_mask(cpu
);
337 local_irq_save(*flags
);
338 for_each_cpu(t
, smt_mask
)
339 raw_spin_lock_nested(&cpu_rq(t
)->__lock
, i
++);
342 static void sched_core_unlock(int cpu
, unsigned long *flags
)
344 const struct cpumask
*smt_mask
= cpu_smt_mask(cpu
);
347 for_each_cpu(t
, smt_mask
)
348 raw_spin_unlock(&cpu_rq(t
)->__lock
);
349 local_irq_restore(*flags
);
352 static void __sched_core_flip(bool enabled
)
360 * Toggle the online cores, one by one.
362 cpumask_copy(&sched_core_mask
, cpu_online_mask
);
363 for_each_cpu(cpu
, &sched_core_mask
) {
364 const struct cpumask
*smt_mask
= cpu_smt_mask(cpu
);
366 sched_core_lock(cpu
, &flags
);
368 for_each_cpu(t
, smt_mask
)
369 cpu_rq(t
)->core_enabled
= enabled
;
371 cpu_rq(cpu
)->core
->core_forceidle_start
= 0;
373 sched_core_unlock(cpu
, &flags
);
375 cpumask_andnot(&sched_core_mask
, &sched_core_mask
, smt_mask
);
379 * Toggle the offline CPUs.
381 for_each_cpu_andnot(cpu
, cpu_possible_mask
, cpu_online_mask
)
382 cpu_rq(cpu
)->core_enabled
= enabled
;
387 static void sched_core_assert_empty(void)
391 for_each_possible_cpu(cpu
)
392 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu
)->core_tree
));
395 static void __sched_core_enable(void)
397 static_branch_enable(&__sched_core_enabled
);
399 * Ensure all previous instances of raw_spin_rq_*lock() have finished
400 * and future ones will observe !sched_core_disabled().
403 __sched_core_flip(true);
404 sched_core_assert_empty();
407 static void __sched_core_disable(void)
409 sched_core_assert_empty();
410 __sched_core_flip(false);
411 static_branch_disable(&__sched_core_enabled
);
414 void sched_core_get(void)
416 if (atomic_inc_not_zero(&sched_core_count
))
419 mutex_lock(&sched_core_mutex
);
420 if (!atomic_read(&sched_core_count
))
421 __sched_core_enable();
423 smp_mb__before_atomic();
424 atomic_inc(&sched_core_count
);
425 mutex_unlock(&sched_core_mutex
);
428 static void __sched_core_put(struct work_struct
*work
)
430 if (atomic_dec_and_mutex_lock(&sched_core_count
, &sched_core_mutex
)) {
431 __sched_core_disable();
432 mutex_unlock(&sched_core_mutex
);
436 void sched_core_put(void)
438 static DECLARE_WORK(_work
, __sched_core_put
);
441 * "There can be only one"
443 * Either this is the last one, or we don't actually need to do any
444 * 'work'. If it is the last *again*, we rely on
445 * WORK_STRUCT_PENDING_BIT.
447 if (!atomic_add_unless(&sched_core_count
, -1, 1))
448 schedule_work(&_work
);
451 #else /* !CONFIG_SCHED_CORE */
453 static inline void sched_core_enqueue(struct rq
*rq
, struct task_struct
*p
) { }
455 sched_core_dequeue(struct rq
*rq
, struct task_struct
*p
, int flags
) { }
457 #endif /* CONFIG_SCHED_CORE */
460 * Serialization rules:
466 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
469 * rq2->lock where: rq1 < rq2
473 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
474 * local CPU's rq->lock, it optionally removes the task from the runqueue and
475 * always looks at the local rq data structures to find the most eligible task
478 * Task enqueue is also under rq->lock, possibly taken from another CPU.
479 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
480 * the local CPU to avoid bouncing the runqueue state around [ see
481 * ttwu_queue_wakelist() ]
483 * Task wakeup, specifically wakeups that involve migration, are horribly
484 * complicated to avoid having to take two rq->locks.
488 * System-calls and anything external will use task_rq_lock() which acquires
489 * both p->pi_lock and rq->lock. As a consequence the state they change is
490 * stable while holding either lock:
492 * - sched_setaffinity()/
493 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
494 * - set_user_nice(): p->se.load, p->*prio
495 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
496 * p->se.load, p->rt_priority,
497 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
498 * - sched_setnuma(): p->numa_preferred_nid
499 * - sched_move_task(): p->sched_task_group
500 * - uclamp_update_active() p->uclamp*
502 * p->state <- TASK_*:
504 * is changed locklessly using set_current_state(), __set_current_state() or
505 * set_special_state(), see their respective comments, or by
506 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
509 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
511 * is set by activate_task() and cleared by deactivate_task(), under
512 * rq->lock. Non-zero indicates the task is runnable, the special
513 * ON_RQ_MIGRATING state is used for migration without holding both
514 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
516 * p->on_cpu <- { 0, 1 }:
518 * is set by prepare_task() and cleared by finish_task() such that it will be
519 * set before p is scheduled-in and cleared after p is scheduled-out, both
520 * under rq->lock. Non-zero indicates the task is running on its CPU.
522 * [ The astute reader will observe that it is possible for two tasks on one
523 * CPU to have ->on_cpu = 1 at the same time. ]
525 * task_cpu(p): is changed by set_task_cpu(), the rules are:
527 * - Don't call set_task_cpu() on a blocked task:
529 * We don't care what CPU we're not running on, this simplifies hotplug,
530 * the CPU assignment of blocked tasks isn't required to be valid.
532 * - for try_to_wake_up(), called under p->pi_lock:
534 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
536 * - for migration called under rq->lock:
537 * [ see task_on_rq_migrating() in task_rq_lock() ]
539 * o move_queued_task()
542 * - for migration called under double_rq_lock():
544 * o __migrate_swap_task()
545 * o push_rt_task() / pull_rt_task()
546 * o push_dl_task() / pull_dl_task()
547 * o dl_task_offline_migration()
551 void raw_spin_rq_lock_nested(struct rq
*rq
, int subclass
)
553 raw_spinlock_t
*lock
;
555 /* Matches synchronize_rcu() in __sched_core_enable() */
557 if (sched_core_disabled()) {
558 raw_spin_lock_nested(&rq
->__lock
, subclass
);
559 /* preempt_count *MUST* be > 1 */
560 preempt_enable_no_resched();
565 lock
= __rq_lockp(rq
);
566 raw_spin_lock_nested(lock
, subclass
);
567 if (likely(lock
== __rq_lockp(rq
))) {
568 /* preempt_count *MUST* be > 1 */
569 preempt_enable_no_resched();
572 raw_spin_unlock(lock
);
576 bool raw_spin_rq_trylock(struct rq
*rq
)
578 raw_spinlock_t
*lock
;
581 /* Matches synchronize_rcu() in __sched_core_enable() */
583 if (sched_core_disabled()) {
584 ret
= raw_spin_trylock(&rq
->__lock
);
590 lock
= __rq_lockp(rq
);
591 ret
= raw_spin_trylock(lock
);
592 if (!ret
|| (likely(lock
== __rq_lockp(rq
)))) {
596 raw_spin_unlock(lock
);
600 void raw_spin_rq_unlock(struct rq
*rq
)
602 raw_spin_unlock(rq_lockp(rq
));
607 * double_rq_lock - safely lock two runqueues
609 void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
611 lockdep_assert_irqs_disabled();
613 if (rq_order_less(rq2
, rq1
))
616 raw_spin_rq_lock(rq1
);
617 if (__rq_lockp(rq1
) != __rq_lockp(rq2
))
618 raw_spin_rq_lock_nested(rq2
, SINGLE_DEPTH_NESTING
);
620 double_rq_clock_clear_update(rq1
, rq2
);
625 * __task_rq_lock - lock the rq @p resides on.
627 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
632 lockdep_assert_held(&p
->pi_lock
);
636 raw_spin_rq_lock(rq
);
637 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
641 raw_spin_rq_unlock(rq
);
643 while (unlikely(task_on_rq_migrating(p
)))
649 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
651 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
652 __acquires(p
->pi_lock
)
658 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
660 raw_spin_rq_lock(rq
);
662 * move_queued_task() task_rq_lock()
665 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
666 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
667 * [S] ->cpu = new_cpu [L] task_rq()
671 * If we observe the old CPU in task_rq_lock(), the acquire of
672 * the old rq->lock will fully serialize against the stores.
674 * If we observe the new CPU in task_rq_lock(), the address
675 * dependency headed by '[L] rq = task_rq()' and the acquire
676 * will pair with the WMB to ensure we then also see migrating.
678 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
682 raw_spin_rq_unlock(rq
);
683 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
685 while (unlikely(task_on_rq_migrating(p
)))
691 * RQ-clock updating methods:
694 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
697 * In theory, the compile should just see 0 here, and optimize out the call
698 * to sched_rt_avg_update. But I don't trust it...
700 s64 __maybe_unused steal
= 0, irq_delta
= 0;
702 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
703 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
706 * Since irq_time is only updated on {soft,}irq_exit, we might run into
707 * this case when a previous update_rq_clock() happened inside a
710 * When this happens, we stop ->clock_task and only update the
711 * prev_irq_time stamp to account for the part that fit, so that a next
712 * update will consume the rest. This ensures ->clock_task is
715 * It does however cause some slight miss-attribution of {soft,}irq
716 * time, a more accurate solution would be to update the irq_time using
717 * the current rq->clock timestamp, except that would require using
720 if (irq_delta
> delta
)
723 rq
->prev_irq_time
+= irq_delta
;
725 psi_account_irqtime(rq
->curr
, irq_delta
);
726 delayacct_irq(rq
->curr
, irq_delta
);
728 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
729 if (static_key_false((¶virt_steal_rq_enabled
))) {
730 steal
= paravirt_steal_clock(cpu_of(rq
));
731 steal
-= rq
->prev_steal_time_rq
;
733 if (unlikely(steal
> delta
))
736 rq
->prev_steal_time_rq
+= steal
;
741 rq
->clock_task
+= delta
;
743 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
744 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
745 update_irq_load_avg(rq
, irq_delta
+ steal
);
747 update_rq_clock_pelt(rq
, delta
);
750 void update_rq_clock(struct rq
*rq
)
754 lockdep_assert_rq_held(rq
);
756 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
759 #ifdef CONFIG_SCHED_DEBUG
760 if (sched_feat(WARN_DOUBLE_CLOCK
))
761 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
762 rq
->clock_update_flags
|= RQCF_UPDATED
;
765 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
769 update_rq_clock_task(rq
, delta
);
772 #ifdef CONFIG_SCHED_HRTICK
774 * Use HR-timers to deliver accurate preemption points.
777 static void hrtick_clear(struct rq
*rq
)
779 if (hrtimer_active(&rq
->hrtick_timer
))
780 hrtimer_cancel(&rq
->hrtick_timer
);
784 * High-resolution timer tick.
785 * Runs from hardirq context with interrupts disabled.
787 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
789 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
792 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
796 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
799 return HRTIMER_NORESTART
;
804 static void __hrtick_restart(struct rq
*rq
)
806 struct hrtimer
*timer
= &rq
->hrtick_timer
;
807 ktime_t time
= rq
->hrtick_time
;
809 hrtimer_start(timer
, time
, HRTIMER_MODE_ABS_PINNED_HARD
);
813 * called from hardirq (IPI) context
815 static void __hrtick_start(void *arg
)
821 __hrtick_restart(rq
);
826 * Called to set the hrtick timer state.
828 * called with rq->lock held and irqs disabled
830 void hrtick_start(struct rq
*rq
, u64 delay
)
832 struct hrtimer
*timer
= &rq
->hrtick_timer
;
836 * Don't schedule slices shorter than 10000ns, that just
837 * doesn't make sense and can cause timer DoS.
839 delta
= max_t(s64
, delay
, 10000LL);
840 rq
->hrtick_time
= ktime_add_ns(timer
->base
->get_time(), delta
);
843 __hrtick_restart(rq
);
845 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
850 * Called to set the hrtick timer state.
852 * called with rq->lock held and irqs disabled
854 void hrtick_start(struct rq
*rq
, u64 delay
)
857 * Don't schedule slices shorter than 10000ns, that just
858 * doesn't make sense. Rely on vruntime for fairness.
860 delay
= max_t(u64
, delay
, 10000LL);
861 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
862 HRTIMER_MODE_REL_PINNED_HARD
);
865 #endif /* CONFIG_SMP */
867 static void hrtick_rq_init(struct rq
*rq
)
870 INIT_CSD(&rq
->hrtick_csd
, __hrtick_start
, rq
);
872 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
873 rq
->hrtick_timer
.function
= hrtick
;
875 #else /* CONFIG_SCHED_HRTICK */
876 static inline void hrtick_clear(struct rq
*rq
)
880 static inline void hrtick_rq_init(struct rq
*rq
)
883 #endif /* CONFIG_SCHED_HRTICK */
886 * cmpxchg based fetch_or, macro so it works for different integer types
888 #define fetch_or(ptr, mask) \
890 typeof(ptr) _ptr = (ptr); \
891 typeof(mask) _mask = (mask); \
892 typeof(*_ptr) _val = *_ptr; \
895 } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
899 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
901 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
902 * this avoids any races wrt polling state changes and thereby avoids
905 static inline bool set_nr_and_not_polling(struct task_struct
*p
)
907 struct thread_info
*ti
= task_thread_info(p
);
908 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
912 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
914 * If this returns true, then the idle task promises to call
915 * sched_ttwu_pending() and reschedule soon.
917 static bool set_nr_if_polling(struct task_struct
*p
)
919 struct thread_info
*ti
= task_thread_info(p
);
920 typeof(ti
->flags
) val
= READ_ONCE(ti
->flags
);
923 if (!(val
& _TIF_POLLING_NRFLAG
))
925 if (val
& _TIF_NEED_RESCHED
)
927 if (try_cmpxchg(&ti
->flags
, &val
, val
| _TIF_NEED_RESCHED
))
934 static inline bool set_nr_and_not_polling(struct task_struct
*p
)
936 set_tsk_need_resched(p
);
941 static inline bool set_nr_if_polling(struct task_struct
*p
)
948 static bool __wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
950 struct wake_q_node
*node
= &task
->wake_q
;
953 * Atomically grab the task, if ->wake_q is !nil already it means
954 * it's already queued (either by us or someone else) and will get the
955 * wakeup due to that.
957 * In order to ensure that a pending wakeup will observe our pending
958 * state, even in the failed case, an explicit smp_mb() must be used.
960 smp_mb__before_atomic();
961 if (unlikely(cmpxchg_relaxed(&node
->next
, NULL
, WAKE_Q_TAIL
)))
965 * The head is context local, there can be no concurrency.
968 head
->lastp
= &node
->next
;
973 * wake_q_add() - queue a wakeup for 'later' waking.
974 * @head: the wake_q_head to add @task to
975 * @task: the task to queue for 'later' wakeup
977 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
978 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
981 * This function must be used as-if it were wake_up_process(); IOW the task
982 * must be ready to be woken at this location.
984 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
986 if (__wake_q_add(head
, task
))
987 get_task_struct(task
);
991 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
992 * @head: the wake_q_head to add @task to
993 * @task: the task to queue for 'later' wakeup
995 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
996 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
999 * This function must be used as-if it were wake_up_process(); IOW the task
1000 * must be ready to be woken at this location.
1002 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
1003 * that already hold reference to @task can call the 'safe' version and trust
1004 * wake_q to do the right thing depending whether or not the @task is already
1005 * queued for wakeup.
1007 void wake_q_add_safe(struct wake_q_head
*head
, struct task_struct
*task
)
1009 if (!__wake_q_add(head
, task
))
1010 put_task_struct(task
);
1013 void wake_up_q(struct wake_q_head
*head
)
1015 struct wake_q_node
*node
= head
->first
;
1017 while (node
!= WAKE_Q_TAIL
) {
1018 struct task_struct
*task
;
1020 task
= container_of(node
, struct task_struct
, wake_q
);
1021 /* Task can safely be re-inserted now: */
1023 task
->wake_q
.next
= NULL
;
1026 * wake_up_process() executes a full barrier, which pairs with
1027 * the queueing in wake_q_add() so as not to miss wakeups.
1029 wake_up_process(task
);
1030 put_task_struct(task
);
1035 * resched_curr - mark rq's current task 'to be rescheduled now'.
1037 * On UP this means the setting of the need_resched flag, on SMP it
1038 * might also involve a cross-CPU call to trigger the scheduler on
1041 void resched_curr(struct rq
*rq
)
1043 struct task_struct
*curr
= rq
->curr
;
1046 lockdep_assert_rq_held(rq
);
1048 if (test_tsk_need_resched(curr
))
1053 if (cpu
== smp_processor_id()) {
1054 set_tsk_need_resched(curr
);
1055 set_preempt_need_resched();
1059 if (set_nr_and_not_polling(curr
))
1060 smp_send_reschedule(cpu
);
1062 trace_sched_wake_idle_without_ipi(cpu
);
1065 void resched_cpu(int cpu
)
1067 struct rq
*rq
= cpu_rq(cpu
);
1068 unsigned long flags
;
1070 raw_spin_rq_lock_irqsave(rq
, flags
);
1071 if (cpu_online(cpu
) || cpu
== smp_processor_id())
1073 raw_spin_rq_unlock_irqrestore(rq
, flags
);
1077 #ifdef CONFIG_NO_HZ_COMMON
1079 * In the semi idle case, use the nearest busy CPU for migrating timers
1080 * from an idle CPU. This is good for power-savings.
1082 * We don't do similar optimization for completely idle system, as
1083 * selecting an idle CPU will add more delays to the timers than intended
1084 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1086 int get_nohz_timer_target(void)
1088 int i
, cpu
= smp_processor_id(), default_cpu
= -1;
1089 struct sched_domain
*sd
;
1090 const struct cpumask
*hk_mask
;
1092 if (housekeeping_cpu(cpu
, HK_TYPE_TIMER
)) {
1098 hk_mask
= housekeeping_cpumask(HK_TYPE_TIMER
);
1101 for_each_domain(cpu
, sd
) {
1102 for_each_cpu_and(i
, sched_domain_span(sd
), hk_mask
) {
1113 if (default_cpu
== -1)
1114 default_cpu
= housekeeping_any_cpu(HK_TYPE_TIMER
);
1122 * When add_timer_on() enqueues a timer into the timer wheel of an
1123 * idle CPU then this timer might expire before the next timer event
1124 * which is scheduled to wake up that CPU. In case of a completely
1125 * idle system the next event might even be infinite time into the
1126 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1127 * leaves the inner idle loop so the newly added timer is taken into
1128 * account when the CPU goes back to idle and evaluates the timer
1129 * wheel for the next timer event.
1131 static void wake_up_idle_cpu(int cpu
)
1133 struct rq
*rq
= cpu_rq(cpu
);
1135 if (cpu
== smp_processor_id())
1138 if (set_nr_and_not_polling(rq
->idle
))
1139 smp_send_reschedule(cpu
);
1141 trace_sched_wake_idle_without_ipi(cpu
);
1144 static bool wake_up_full_nohz_cpu(int cpu
)
1147 * We just need the target to call irq_exit() and re-evaluate
1148 * the next tick. The nohz full kick at least implies that.
1149 * If needed we can still optimize that later with an
1152 if (cpu_is_offline(cpu
))
1153 return true; /* Don't try to wake offline CPUs. */
1154 if (tick_nohz_full_cpu(cpu
)) {
1155 if (cpu
!= smp_processor_id() ||
1156 tick_nohz_tick_stopped())
1157 tick_nohz_full_kick_cpu(cpu
);
1165 * Wake up the specified CPU. If the CPU is going offline, it is the
1166 * caller's responsibility to deal with the lost wakeup, for example,
1167 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1169 void wake_up_nohz_cpu(int cpu
)
1171 if (!wake_up_full_nohz_cpu(cpu
))
1172 wake_up_idle_cpu(cpu
);
1175 static void nohz_csd_func(void *info
)
1177 struct rq
*rq
= info
;
1178 int cpu
= cpu_of(rq
);
1182 * Release the rq::nohz_csd.
1184 flags
= atomic_fetch_andnot(NOHZ_KICK_MASK
| NOHZ_NEWILB_KICK
, nohz_flags(cpu
));
1185 WARN_ON(!(flags
& NOHZ_KICK_MASK
));
1187 rq
->idle_balance
= idle_cpu(cpu
);
1188 if (rq
->idle_balance
&& !need_resched()) {
1189 rq
->nohz_idle_balance
= flags
;
1190 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1194 #endif /* CONFIG_NO_HZ_COMMON */
1196 #ifdef CONFIG_NO_HZ_FULL
1197 bool sched_can_stop_tick(struct rq
*rq
)
1199 int fifo_nr_running
;
1201 /* Deadline tasks, even if single, need the tick */
1202 if (rq
->dl
.dl_nr_running
)
1206 * If there are more than one RR tasks, we need the tick to affect the
1207 * actual RR behaviour.
1209 if (rq
->rt
.rr_nr_running
) {
1210 if (rq
->rt
.rr_nr_running
== 1)
1217 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1218 * forced preemption between FIFO tasks.
1220 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
1221 if (fifo_nr_running
)
1225 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1226 * if there's more than one we need the tick for involuntary
1229 if (rq
->nr_running
> 1)
1234 #endif /* CONFIG_NO_HZ_FULL */
1235 #endif /* CONFIG_SMP */
1237 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1238 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1240 * Iterate task_group tree rooted at *from, calling @down when first entering a
1241 * node and @up when leaving it for the final time.
1243 * Caller must hold rcu_lock or sufficient equivalent.
1245 int walk_tg_tree_from(struct task_group
*from
,
1246 tg_visitor down
, tg_visitor up
, void *data
)
1248 struct task_group
*parent
, *child
;
1254 ret
= (*down
)(parent
, data
);
1257 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1264 ret
= (*up
)(parent
, data
);
1265 if (ret
|| parent
== from
)
1269 parent
= parent
->parent
;
1276 int tg_nop(struct task_group
*tg
, void *data
)
1282 static void set_load_weight(struct task_struct
*p
, bool update_load
)
1284 int prio
= p
->static_prio
- MAX_RT_PRIO
;
1285 struct load_weight
*load
= &p
->se
.load
;
1288 * SCHED_IDLE tasks get minimal weight:
1290 if (task_has_idle_policy(p
)) {
1291 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
1292 load
->inv_weight
= WMULT_IDLEPRIO
;
1297 * SCHED_OTHER tasks have to update their load when changing their
1300 if (update_load
&& p
->sched_class
== &fair_sched_class
) {
1301 reweight_task(p
, prio
);
1303 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
1304 load
->inv_weight
= sched_prio_to_wmult
[prio
];
1308 #ifdef CONFIG_UCLAMP_TASK
1310 * Serializes updates of utilization clamp values
1312 * The (slow-path) user-space triggers utilization clamp value updates which
1313 * can require updates on (fast-path) scheduler's data structures used to
1314 * support enqueue/dequeue operations.
1315 * While the per-CPU rq lock protects fast-path update operations, user-space
1316 * requests are serialized using a mutex to reduce the risk of conflicting
1317 * updates or API abuses.
1319 static DEFINE_MUTEX(uclamp_mutex
);
1321 /* Max allowed minimum utilization */
1322 static unsigned int __maybe_unused sysctl_sched_uclamp_util_min
= SCHED_CAPACITY_SCALE
;
1324 /* Max allowed maximum utilization */
1325 static unsigned int __maybe_unused sysctl_sched_uclamp_util_max
= SCHED_CAPACITY_SCALE
;
1328 * By default RT tasks run at the maximum performance point/capacity of the
1329 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1330 * SCHED_CAPACITY_SCALE.
1332 * This knob allows admins to change the default behavior when uclamp is being
1333 * used. In battery powered devices, particularly, running at the maximum
1334 * capacity and frequency will increase energy consumption and shorten the
1337 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1339 * This knob will not override the system default sched_util_clamp_min defined
1342 static unsigned int sysctl_sched_uclamp_util_min_rt_default
= SCHED_CAPACITY_SCALE
;
1344 /* All clamps are required to be less or equal than these values */
1345 static struct uclamp_se uclamp_default
[UCLAMP_CNT
];
1348 * This static key is used to reduce the uclamp overhead in the fast path. It
1349 * primarily disables the call to uclamp_rq_{inc, dec}() in
1350 * enqueue/dequeue_task().
1352 * This allows users to continue to enable uclamp in their kernel config with
1353 * minimum uclamp overhead in the fast path.
1355 * As soon as userspace modifies any of the uclamp knobs, the static key is
1356 * enabled, since we have an actual users that make use of uclamp
1359 * The knobs that would enable this static key are:
1361 * * A task modifying its uclamp value with sched_setattr().
1362 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1363 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1365 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used
);
1367 /* Integer rounded range for each bucket */
1368 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1370 #define for_each_clamp_id(clamp_id) \
1371 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1373 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value
)
1375 return min_t(unsigned int, clamp_value
/ UCLAMP_BUCKET_DELTA
, UCLAMP_BUCKETS
- 1);
1378 static inline unsigned int uclamp_none(enum uclamp_id clamp_id
)
1380 if (clamp_id
== UCLAMP_MIN
)
1382 return SCHED_CAPACITY_SCALE
;
1385 static inline void uclamp_se_set(struct uclamp_se
*uc_se
,
1386 unsigned int value
, bool user_defined
)
1388 uc_se
->value
= value
;
1389 uc_se
->bucket_id
= uclamp_bucket_id(value
);
1390 uc_se
->user_defined
= user_defined
;
1393 static inline unsigned int
1394 uclamp_idle_value(struct rq
*rq
, enum uclamp_id clamp_id
,
1395 unsigned int clamp_value
)
1398 * Avoid blocked utilization pushing up the frequency when we go
1399 * idle (which drops the max-clamp) by retaining the last known
1402 if (clamp_id
== UCLAMP_MAX
) {
1403 rq
->uclamp_flags
|= UCLAMP_FLAG_IDLE
;
1407 return uclamp_none(UCLAMP_MIN
);
1410 static inline void uclamp_idle_reset(struct rq
*rq
, enum uclamp_id clamp_id
,
1411 unsigned int clamp_value
)
1413 /* Reset max-clamp retention only on idle exit */
1414 if (!(rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
))
1417 uclamp_rq_set(rq
, clamp_id
, clamp_value
);
1421 unsigned int uclamp_rq_max_value(struct rq
*rq
, enum uclamp_id clamp_id
,
1422 unsigned int clamp_value
)
1424 struct uclamp_bucket
*bucket
= rq
->uclamp
[clamp_id
].bucket
;
1425 int bucket_id
= UCLAMP_BUCKETS
- 1;
1428 * Since both min and max clamps are max aggregated, find the
1429 * top most bucket with tasks in.
1431 for ( ; bucket_id
>= 0; bucket_id
--) {
1432 if (!bucket
[bucket_id
].tasks
)
1434 return bucket
[bucket_id
].value
;
1437 /* No tasks -- default clamp values */
1438 return uclamp_idle_value(rq
, clamp_id
, clamp_value
);
1441 static void __uclamp_update_util_min_rt_default(struct task_struct
*p
)
1443 unsigned int default_util_min
;
1444 struct uclamp_se
*uc_se
;
1446 lockdep_assert_held(&p
->pi_lock
);
1448 uc_se
= &p
->uclamp_req
[UCLAMP_MIN
];
1450 /* Only sync if user didn't override the default */
1451 if (uc_se
->user_defined
)
1454 default_util_min
= sysctl_sched_uclamp_util_min_rt_default
;
1455 uclamp_se_set(uc_se
, default_util_min
, false);
1458 static void uclamp_update_util_min_rt_default(struct task_struct
*p
)
1466 /* Protect updates to p->uclamp_* */
1467 rq
= task_rq_lock(p
, &rf
);
1468 __uclamp_update_util_min_rt_default(p
);
1469 task_rq_unlock(rq
, p
, &rf
);
1472 static inline struct uclamp_se
1473 uclamp_tg_restrict(struct task_struct
*p
, enum uclamp_id clamp_id
)
1475 /* Copy by value as we could modify it */
1476 struct uclamp_se uc_req
= p
->uclamp_req
[clamp_id
];
1477 #ifdef CONFIG_UCLAMP_TASK_GROUP
1478 unsigned int tg_min
, tg_max
, value
;
1481 * Tasks in autogroups or root task group will be
1482 * restricted by system defaults.
1484 if (task_group_is_autogroup(task_group(p
)))
1486 if (task_group(p
) == &root_task_group
)
1489 tg_min
= task_group(p
)->uclamp
[UCLAMP_MIN
].value
;
1490 tg_max
= task_group(p
)->uclamp
[UCLAMP_MAX
].value
;
1491 value
= uc_req
.value
;
1492 value
= clamp(value
, tg_min
, tg_max
);
1493 uclamp_se_set(&uc_req
, value
, false);
1500 * The effective clamp bucket index of a task depends on, by increasing
1502 * - the task specific clamp value, when explicitly requested from userspace
1503 * - the task group effective clamp value, for tasks not either in the root
1504 * group or in an autogroup
1505 * - the system default clamp value, defined by the sysadmin
1507 static inline struct uclamp_se
1508 uclamp_eff_get(struct task_struct
*p
, enum uclamp_id clamp_id
)
1510 struct uclamp_se uc_req
= uclamp_tg_restrict(p
, clamp_id
);
1511 struct uclamp_se uc_max
= uclamp_default
[clamp_id
];
1513 /* System default restrictions always apply */
1514 if (unlikely(uc_req
.value
> uc_max
.value
))
1520 unsigned long uclamp_eff_value(struct task_struct
*p
, enum uclamp_id clamp_id
)
1522 struct uclamp_se uc_eff
;
1524 /* Task currently refcounted: use back-annotated (effective) value */
1525 if (p
->uclamp
[clamp_id
].active
)
1526 return (unsigned long)p
->uclamp
[clamp_id
].value
;
1528 uc_eff
= uclamp_eff_get(p
, clamp_id
);
1530 return (unsigned long)uc_eff
.value
;
1534 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1535 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1536 * updates the rq's clamp value if required.
1538 * Tasks can have a task-specific value requested from user-space, track
1539 * within each bucket the maximum value for tasks refcounted in it.
1540 * This "local max aggregation" allows to track the exact "requested" value
1541 * for each bucket when all its RUNNABLE tasks require the same clamp.
1543 static inline void uclamp_rq_inc_id(struct rq
*rq
, struct task_struct
*p
,
1544 enum uclamp_id clamp_id
)
1546 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
1547 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
1548 struct uclamp_bucket
*bucket
;
1550 lockdep_assert_rq_held(rq
);
1552 /* Update task effective clamp */
1553 p
->uclamp
[clamp_id
] = uclamp_eff_get(p
, clamp_id
);
1555 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
1557 uc_se
->active
= true;
1559 uclamp_idle_reset(rq
, clamp_id
, uc_se
->value
);
1562 * Local max aggregation: rq buckets always track the max
1563 * "requested" clamp value of its RUNNABLE tasks.
1565 if (bucket
->tasks
== 1 || uc_se
->value
> bucket
->value
)
1566 bucket
->value
= uc_se
->value
;
1568 if (uc_se
->value
> uclamp_rq_get(rq
, clamp_id
))
1569 uclamp_rq_set(rq
, clamp_id
, uc_se
->value
);
1573 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1574 * is released. If this is the last task reference counting the rq's max
1575 * active clamp value, then the rq's clamp value is updated.
1577 * Both refcounted tasks and rq's cached clamp values are expected to be
1578 * always valid. If it's detected they are not, as defensive programming,
1579 * enforce the expected state and warn.
1581 static inline void uclamp_rq_dec_id(struct rq
*rq
, struct task_struct
*p
,
1582 enum uclamp_id clamp_id
)
1584 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
1585 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
1586 struct uclamp_bucket
*bucket
;
1587 unsigned int bkt_clamp
;
1588 unsigned int rq_clamp
;
1590 lockdep_assert_rq_held(rq
);
1593 * If sched_uclamp_used was enabled after task @p was enqueued,
1594 * we could end up with unbalanced call to uclamp_rq_dec_id().
1596 * In this case the uc_se->active flag should be false since no uclamp
1597 * accounting was performed at enqueue time and we can just return
1600 * Need to be careful of the following enqueue/dequeue ordering
1604 * // sched_uclamp_used gets enabled
1607 * // Must not decrement bucket->tasks here
1610 * where we could end up with stale data in uc_se and
1611 * bucket[uc_se->bucket_id].
1613 * The following check here eliminates the possibility of such race.
1615 if (unlikely(!uc_se
->active
))
1618 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
1620 SCHED_WARN_ON(!bucket
->tasks
);
1621 if (likely(bucket
->tasks
))
1624 uc_se
->active
= false;
1627 * Keep "local max aggregation" simple and accept to (possibly)
1628 * overboost some RUNNABLE tasks in the same bucket.
1629 * The rq clamp bucket value is reset to its base value whenever
1630 * there are no more RUNNABLE tasks refcounting it.
1632 if (likely(bucket
->tasks
))
1635 rq_clamp
= uclamp_rq_get(rq
, clamp_id
);
1637 * Defensive programming: this should never happen. If it happens,
1638 * e.g. due to future modification, warn and fixup the expected value.
1640 SCHED_WARN_ON(bucket
->value
> rq_clamp
);
1641 if (bucket
->value
>= rq_clamp
) {
1642 bkt_clamp
= uclamp_rq_max_value(rq
, clamp_id
, uc_se
->value
);
1643 uclamp_rq_set(rq
, clamp_id
, bkt_clamp
);
1647 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
)
1649 enum uclamp_id clamp_id
;
1652 * Avoid any overhead until uclamp is actually used by the userspace.
1654 * The condition is constructed such that a NOP is generated when
1655 * sched_uclamp_used is disabled.
1657 if (!static_branch_unlikely(&sched_uclamp_used
))
1660 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1663 for_each_clamp_id(clamp_id
)
1664 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1666 /* Reset clamp idle holding when there is one RUNNABLE task */
1667 if (rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
)
1668 rq
->uclamp_flags
&= ~UCLAMP_FLAG_IDLE
;
1671 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
)
1673 enum uclamp_id clamp_id
;
1676 * Avoid any overhead until uclamp is actually used by the userspace.
1678 * The condition is constructed such that a NOP is generated when
1679 * sched_uclamp_used is disabled.
1681 if (!static_branch_unlikely(&sched_uclamp_used
))
1684 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1687 for_each_clamp_id(clamp_id
)
1688 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1691 static inline void uclamp_rq_reinc_id(struct rq
*rq
, struct task_struct
*p
,
1692 enum uclamp_id clamp_id
)
1694 if (!p
->uclamp
[clamp_id
].active
)
1697 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1698 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1701 * Make sure to clear the idle flag if we've transiently reached 0
1702 * active tasks on rq.
1704 if (clamp_id
== UCLAMP_MAX
&& (rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
))
1705 rq
->uclamp_flags
&= ~UCLAMP_FLAG_IDLE
;
1709 uclamp_update_active(struct task_struct
*p
)
1711 enum uclamp_id clamp_id
;
1716 * Lock the task and the rq where the task is (or was) queued.
1718 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1719 * price to pay to safely serialize util_{min,max} updates with
1720 * enqueues, dequeues and migration operations.
1721 * This is the same locking schema used by __set_cpus_allowed_ptr().
1723 rq
= task_rq_lock(p
, &rf
);
1726 * Setting the clamp bucket is serialized by task_rq_lock().
1727 * If the task is not yet RUNNABLE and its task_struct is not
1728 * affecting a valid clamp bucket, the next time it's enqueued,
1729 * it will already see the updated clamp bucket value.
1731 for_each_clamp_id(clamp_id
)
1732 uclamp_rq_reinc_id(rq
, p
, clamp_id
);
1734 task_rq_unlock(rq
, p
, &rf
);
1737 #ifdef CONFIG_UCLAMP_TASK_GROUP
1739 uclamp_update_active_tasks(struct cgroup_subsys_state
*css
)
1741 struct css_task_iter it
;
1742 struct task_struct
*p
;
1744 css_task_iter_start(css
, 0, &it
);
1745 while ((p
= css_task_iter_next(&it
)))
1746 uclamp_update_active(p
);
1747 css_task_iter_end(&it
);
1750 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
);
1753 #ifdef CONFIG_SYSCTL
1754 #ifdef CONFIG_UCLAMP_TASK
1755 #ifdef CONFIG_UCLAMP_TASK_GROUP
1756 static void uclamp_update_root_tg(void)
1758 struct task_group
*tg
= &root_task_group
;
1760 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MIN
],
1761 sysctl_sched_uclamp_util_min
, false);
1762 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MAX
],
1763 sysctl_sched_uclamp_util_max
, false);
1766 cpu_util_update_eff(&root_task_group
.css
);
1770 static void uclamp_update_root_tg(void) { }
1773 static void uclamp_sync_util_min_rt_default(void)
1775 struct task_struct
*g
, *p
;
1778 * copy_process() sysctl_uclamp
1779 * uclamp_min_rt = X;
1780 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1781 * // link thread smp_mb__after_spinlock()
1782 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1783 * sched_post_fork() for_each_process_thread()
1784 * __uclamp_sync_rt() __uclamp_sync_rt()
1786 * Ensures that either sched_post_fork() will observe the new
1787 * uclamp_min_rt or for_each_process_thread() will observe the new
1790 read_lock(&tasklist_lock
);
1791 smp_mb__after_spinlock();
1792 read_unlock(&tasklist_lock
);
1795 for_each_process_thread(g
, p
)
1796 uclamp_update_util_min_rt_default(p
);
1800 static int sysctl_sched_uclamp_handler(struct ctl_table
*table
, int write
,
1801 void *buffer
, size_t *lenp
, loff_t
*ppos
)
1803 bool update_root_tg
= false;
1804 int old_min
, old_max
, old_min_rt
;
1807 mutex_lock(&uclamp_mutex
);
1808 old_min
= sysctl_sched_uclamp_util_min
;
1809 old_max
= sysctl_sched_uclamp_util_max
;
1810 old_min_rt
= sysctl_sched_uclamp_util_min_rt_default
;
1812 result
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
1818 if (sysctl_sched_uclamp_util_min
> sysctl_sched_uclamp_util_max
||
1819 sysctl_sched_uclamp_util_max
> SCHED_CAPACITY_SCALE
||
1820 sysctl_sched_uclamp_util_min_rt_default
> SCHED_CAPACITY_SCALE
) {
1826 if (old_min
!= sysctl_sched_uclamp_util_min
) {
1827 uclamp_se_set(&uclamp_default
[UCLAMP_MIN
],
1828 sysctl_sched_uclamp_util_min
, false);
1829 update_root_tg
= true;
1831 if (old_max
!= sysctl_sched_uclamp_util_max
) {
1832 uclamp_se_set(&uclamp_default
[UCLAMP_MAX
],
1833 sysctl_sched_uclamp_util_max
, false);
1834 update_root_tg
= true;
1837 if (update_root_tg
) {
1838 static_branch_enable(&sched_uclamp_used
);
1839 uclamp_update_root_tg();
1842 if (old_min_rt
!= sysctl_sched_uclamp_util_min_rt_default
) {
1843 static_branch_enable(&sched_uclamp_used
);
1844 uclamp_sync_util_min_rt_default();
1848 * We update all RUNNABLE tasks only when task groups are in use.
1849 * Otherwise, keep it simple and do just a lazy update at each next
1850 * task enqueue time.
1856 sysctl_sched_uclamp_util_min
= old_min
;
1857 sysctl_sched_uclamp_util_max
= old_max
;
1858 sysctl_sched_uclamp_util_min_rt_default
= old_min_rt
;
1860 mutex_unlock(&uclamp_mutex
);
1867 static int uclamp_validate(struct task_struct
*p
,
1868 const struct sched_attr
*attr
)
1870 int util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
1871 int util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
1873 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
) {
1874 util_min
= attr
->sched_util_min
;
1876 if (util_min
+ 1 > SCHED_CAPACITY_SCALE
+ 1)
1880 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
) {
1881 util_max
= attr
->sched_util_max
;
1883 if (util_max
+ 1 > SCHED_CAPACITY_SCALE
+ 1)
1887 if (util_min
!= -1 && util_max
!= -1 && util_min
> util_max
)
1891 * We have valid uclamp attributes; make sure uclamp is enabled.
1893 * We need to do that here, because enabling static branches is a
1894 * blocking operation which obviously cannot be done while holding
1897 static_branch_enable(&sched_uclamp_used
);
1902 static bool uclamp_reset(const struct sched_attr
*attr
,
1903 enum uclamp_id clamp_id
,
1904 struct uclamp_se
*uc_se
)
1906 /* Reset on sched class change for a non user-defined clamp value. */
1907 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)) &&
1908 !uc_se
->user_defined
)
1911 /* Reset on sched_util_{min,max} == -1. */
1912 if (clamp_id
== UCLAMP_MIN
&&
1913 attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
&&
1914 attr
->sched_util_min
== -1) {
1918 if (clamp_id
== UCLAMP_MAX
&&
1919 attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
&&
1920 attr
->sched_util_max
== -1) {
1927 static void __setscheduler_uclamp(struct task_struct
*p
,
1928 const struct sched_attr
*attr
)
1930 enum uclamp_id clamp_id
;
1932 for_each_clamp_id(clamp_id
) {
1933 struct uclamp_se
*uc_se
= &p
->uclamp_req
[clamp_id
];
1936 if (!uclamp_reset(attr
, clamp_id
, uc_se
))
1940 * RT by default have a 100% boost value that could be modified
1943 if (unlikely(rt_task(p
) && clamp_id
== UCLAMP_MIN
))
1944 value
= sysctl_sched_uclamp_util_min_rt_default
;
1946 value
= uclamp_none(clamp_id
);
1948 uclamp_se_set(uc_se
, value
, false);
1952 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)))
1955 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
&&
1956 attr
->sched_util_min
!= -1) {
1957 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MIN
],
1958 attr
->sched_util_min
, true);
1961 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
&&
1962 attr
->sched_util_max
!= -1) {
1963 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MAX
],
1964 attr
->sched_util_max
, true);
1968 static void uclamp_fork(struct task_struct
*p
)
1970 enum uclamp_id clamp_id
;
1973 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1974 * as the task is still at its early fork stages.
1976 for_each_clamp_id(clamp_id
)
1977 p
->uclamp
[clamp_id
].active
= false;
1979 if (likely(!p
->sched_reset_on_fork
))
1982 for_each_clamp_id(clamp_id
) {
1983 uclamp_se_set(&p
->uclamp_req
[clamp_id
],
1984 uclamp_none(clamp_id
), false);
1988 static void uclamp_post_fork(struct task_struct
*p
)
1990 uclamp_update_util_min_rt_default(p
);
1993 static void __init
init_uclamp_rq(struct rq
*rq
)
1995 enum uclamp_id clamp_id
;
1996 struct uclamp_rq
*uc_rq
= rq
->uclamp
;
1998 for_each_clamp_id(clamp_id
) {
1999 uc_rq
[clamp_id
] = (struct uclamp_rq
) {
2000 .value
= uclamp_none(clamp_id
)
2004 rq
->uclamp_flags
= UCLAMP_FLAG_IDLE
;
2007 static void __init
init_uclamp(void)
2009 struct uclamp_se uc_max
= {};
2010 enum uclamp_id clamp_id
;
2013 for_each_possible_cpu(cpu
)
2014 init_uclamp_rq(cpu_rq(cpu
));
2016 for_each_clamp_id(clamp_id
) {
2017 uclamp_se_set(&init_task
.uclamp_req
[clamp_id
],
2018 uclamp_none(clamp_id
), false);
2021 /* System defaults allow max clamp values for both indexes */
2022 uclamp_se_set(&uc_max
, uclamp_none(UCLAMP_MAX
), false);
2023 for_each_clamp_id(clamp_id
) {
2024 uclamp_default
[clamp_id
] = uc_max
;
2025 #ifdef CONFIG_UCLAMP_TASK_GROUP
2026 root_task_group
.uclamp_req
[clamp_id
] = uc_max
;
2027 root_task_group
.uclamp
[clamp_id
] = uc_max
;
2032 #else /* CONFIG_UCLAMP_TASK */
2033 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
) { }
2034 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
) { }
2035 static inline int uclamp_validate(struct task_struct
*p
,
2036 const struct sched_attr
*attr
)
2040 static void __setscheduler_uclamp(struct task_struct
*p
,
2041 const struct sched_attr
*attr
) { }
2042 static inline void uclamp_fork(struct task_struct
*p
) { }
2043 static inline void uclamp_post_fork(struct task_struct
*p
) { }
2044 static inline void init_uclamp(void) { }
2045 #endif /* CONFIG_UCLAMP_TASK */
2047 bool sched_task_on_rq(struct task_struct
*p
)
2049 return task_on_rq_queued(p
);
2052 unsigned long get_wchan(struct task_struct
*p
)
2054 unsigned long ip
= 0;
2057 if (!p
|| p
== current
)
2060 /* Only get wchan if task is blocked and we can keep it that way. */
2061 raw_spin_lock_irq(&p
->pi_lock
);
2062 state
= READ_ONCE(p
->__state
);
2063 smp_rmb(); /* see try_to_wake_up() */
2064 if (state
!= TASK_RUNNING
&& state
!= TASK_WAKING
&& !p
->on_rq
)
2065 ip
= __get_wchan(p
);
2066 raw_spin_unlock_irq(&p
->pi_lock
);
2071 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
2073 if (!(flags
& ENQUEUE_NOCLOCK
))
2074 update_rq_clock(rq
);
2076 if (!(flags
& ENQUEUE_RESTORE
)) {
2077 sched_info_enqueue(rq
, p
);
2078 psi_enqueue(p
, (flags
& ENQUEUE_WAKEUP
) && !(flags
& ENQUEUE_MIGRATED
));
2081 uclamp_rq_inc(rq
, p
);
2082 p
->sched_class
->enqueue_task(rq
, p
, flags
);
2084 if (sched_core_enabled(rq
))
2085 sched_core_enqueue(rq
, p
);
2088 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
2090 if (sched_core_enabled(rq
))
2091 sched_core_dequeue(rq
, p
, flags
);
2093 if (!(flags
& DEQUEUE_NOCLOCK
))
2094 update_rq_clock(rq
);
2096 if (!(flags
& DEQUEUE_SAVE
)) {
2097 sched_info_dequeue(rq
, p
);
2098 psi_dequeue(p
, flags
& DEQUEUE_SLEEP
);
2101 uclamp_rq_dec(rq
, p
);
2102 p
->sched_class
->dequeue_task(rq
, p
, flags
);
2105 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
2107 if (task_on_rq_migrating(p
))
2108 flags
|= ENQUEUE_MIGRATED
;
2109 if (flags
& ENQUEUE_MIGRATED
)
2110 sched_mm_cid_migrate_to(rq
, p
);
2112 enqueue_task(rq
, p
, flags
);
2114 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2117 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
2119 p
->on_rq
= (flags
& DEQUEUE_SLEEP
) ? 0 : TASK_ON_RQ_MIGRATING
;
2121 dequeue_task(rq
, p
, flags
);
2124 static inline int __normal_prio(int policy
, int rt_prio
, int nice
)
2128 if (dl_policy(policy
))
2129 prio
= MAX_DL_PRIO
- 1;
2130 else if (rt_policy(policy
))
2131 prio
= MAX_RT_PRIO
- 1 - rt_prio
;
2133 prio
= NICE_TO_PRIO(nice
);
2139 * Calculate the expected normal priority: i.e. priority
2140 * without taking RT-inheritance into account. Might be
2141 * boosted by interactivity modifiers. Changes upon fork,
2142 * setprio syscalls, and whenever the interactivity
2143 * estimator recalculates.
2145 static inline int normal_prio(struct task_struct
*p
)
2147 return __normal_prio(p
->policy
, p
->rt_priority
, PRIO_TO_NICE(p
->static_prio
));
2151 * Calculate the current priority, i.e. the priority
2152 * taken into account by the scheduler. This value might
2153 * be boosted by RT tasks, or might be boosted by
2154 * interactivity modifiers. Will be RT if the task got
2155 * RT-boosted. If not then it returns p->normal_prio.
2157 static int effective_prio(struct task_struct
*p
)
2159 p
->normal_prio
= normal_prio(p
);
2161 * If we are RT tasks or we were boosted to RT priority,
2162 * keep the priority unchanged. Otherwise, update priority
2163 * to the normal priority:
2165 if (!rt_prio(p
->prio
))
2166 return p
->normal_prio
;
2171 * task_curr - is this task currently executing on a CPU?
2172 * @p: the task in question.
2174 * Return: 1 if the task is currently executing. 0 otherwise.
2176 inline int task_curr(const struct task_struct
*p
)
2178 return cpu_curr(task_cpu(p
)) == p
;
2182 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2183 * use the balance_callback list if you want balancing.
2185 * this means any call to check_class_changed() must be followed by a call to
2186 * balance_callback().
2188 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2189 const struct sched_class
*prev_class
,
2192 if (prev_class
!= p
->sched_class
) {
2193 if (prev_class
->switched_from
)
2194 prev_class
->switched_from(rq
, p
);
2196 p
->sched_class
->switched_to(rq
, p
);
2197 } else if (oldprio
!= p
->prio
|| dl_task(p
))
2198 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
2201 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
2203 if (p
->sched_class
== rq
->curr
->sched_class
)
2204 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
2205 else if (sched_class_above(p
->sched_class
, rq
->curr
->sched_class
))
2209 * A queue event has occurred, and we're going to schedule. In
2210 * this case, we can save a useless back to back clock update.
2212 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
2213 rq_clock_skip_update(rq
);
2219 __do_set_cpus_allowed(struct task_struct
*p
, struct affinity_context
*ctx
);
2221 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
2222 struct affinity_context
*ctx
);
2224 static void migrate_disable_switch(struct rq
*rq
, struct task_struct
*p
)
2226 struct affinity_context ac
= {
2227 .new_mask
= cpumask_of(rq
->cpu
),
2228 .flags
= SCA_MIGRATE_DISABLE
,
2231 if (likely(!p
->migration_disabled
))
2234 if (p
->cpus_ptr
!= &p
->cpus_mask
)
2238 * Violates locking rules! see comment in __do_set_cpus_allowed().
2240 __do_set_cpus_allowed(p
, &ac
);
2243 void migrate_disable(void)
2245 struct task_struct
*p
= current
;
2247 if (p
->migration_disabled
) {
2248 p
->migration_disabled
++;
2253 this_rq()->nr_pinned
++;
2254 p
->migration_disabled
= 1;
2257 EXPORT_SYMBOL_GPL(migrate_disable
);
2259 void migrate_enable(void)
2261 struct task_struct
*p
= current
;
2262 struct affinity_context ac
= {
2263 .new_mask
= &p
->cpus_mask
,
2264 .flags
= SCA_MIGRATE_ENABLE
,
2267 if (p
->migration_disabled
> 1) {
2268 p
->migration_disabled
--;
2272 if (WARN_ON_ONCE(!p
->migration_disabled
))
2276 * Ensure stop_task runs either before or after this, and that
2277 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2280 if (p
->cpus_ptr
!= &p
->cpus_mask
)
2281 __set_cpus_allowed_ptr(p
, &ac
);
2283 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2284 * regular cpus_mask, otherwise things that race (eg.
2285 * select_fallback_rq) get confused.
2288 p
->migration_disabled
= 0;
2289 this_rq()->nr_pinned
--;
2292 EXPORT_SYMBOL_GPL(migrate_enable
);
2294 static inline bool rq_has_pinned_tasks(struct rq
*rq
)
2296 return rq
->nr_pinned
;
2300 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2301 * __set_cpus_allowed_ptr() and select_fallback_rq().
2303 static inline bool is_cpu_allowed(struct task_struct
*p
, int cpu
)
2305 /* When not in the task's cpumask, no point in looking further. */
2306 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
2309 /* migrate_disabled() must be allowed to finish. */
2310 if (is_migration_disabled(p
))
2311 return cpu_online(cpu
);
2313 /* Non kernel threads are not allowed during either online or offline. */
2314 if (!(p
->flags
& PF_KTHREAD
))
2315 return cpu_active(cpu
) && task_cpu_possible(cpu
, p
);
2317 /* KTHREAD_IS_PER_CPU is always allowed. */
2318 if (kthread_is_per_cpu(p
))
2319 return cpu_online(cpu
);
2321 /* Regular kernel threads don't get to stay during offline. */
2325 /* But are allowed during online. */
2326 return cpu_online(cpu
);
2330 * This is how migration works:
2332 * 1) we invoke migration_cpu_stop() on the target CPU using
2334 * 2) stopper starts to run (implicitly forcing the migrated thread
2336 * 3) it checks whether the migrated task is still in the wrong runqueue.
2337 * 4) if it's in the wrong runqueue then the migration thread removes
2338 * it and puts it into the right queue.
2339 * 5) stopper completes and stop_one_cpu() returns and the migration
2344 * move_queued_task - move a queued task to new rq.
2346 * Returns (locked) new rq. Old rq's lock is released.
2348 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
2349 struct task_struct
*p
, int new_cpu
)
2351 lockdep_assert_rq_held(rq
);
2353 deactivate_task(rq
, p
, DEQUEUE_NOCLOCK
);
2354 set_task_cpu(p
, new_cpu
);
2357 rq
= cpu_rq(new_cpu
);
2360 WARN_ON_ONCE(task_cpu(p
) != new_cpu
);
2361 activate_task(rq
, p
, 0);
2362 check_preempt_curr(rq
, p
, 0);
2367 struct migration_arg
{
2368 struct task_struct
*task
;
2370 struct set_affinity_pending
*pending
;
2374 * @refs: number of wait_for_completion()
2375 * @stop_pending: is @stop_work in use
2377 struct set_affinity_pending
{
2379 unsigned int stop_pending
;
2380 struct completion done
;
2381 struct cpu_stop_work stop_work
;
2382 struct migration_arg arg
;
2386 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2387 * this because either it can't run here any more (set_cpus_allowed()
2388 * away from this CPU, or CPU going down), or because we're
2389 * attempting to rebalance this task on exec (sched_exec).
2391 * So we race with normal scheduler movements, but that's OK, as long
2392 * as the task is no longer on this CPU.
2394 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
2395 struct task_struct
*p
, int dest_cpu
)
2397 /* Affinity changed (again). */
2398 if (!is_cpu_allowed(p
, dest_cpu
))
2401 update_rq_clock(rq
);
2402 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
2408 * migration_cpu_stop - this will be executed by a highprio stopper thread
2409 * and performs thread migration by bumping thread off CPU then
2410 * 'pushing' onto another runqueue.
2412 static int migration_cpu_stop(void *data
)
2414 struct migration_arg
*arg
= data
;
2415 struct set_affinity_pending
*pending
= arg
->pending
;
2416 struct task_struct
*p
= arg
->task
;
2417 struct rq
*rq
= this_rq();
2418 bool complete
= false;
2422 * The original target CPU might have gone down and we might
2423 * be on another CPU but it doesn't matter.
2425 local_irq_save(rf
.flags
);
2427 * We need to explicitly wake pending tasks before running
2428 * __migrate_task() such that we will not miss enforcing cpus_ptr
2429 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2431 flush_smp_call_function_queue();
2433 raw_spin_lock(&p
->pi_lock
);
2437 * If we were passed a pending, then ->stop_pending was set, thus
2438 * p->migration_pending must have remained stable.
2440 WARN_ON_ONCE(pending
&& pending
!= p
->migration_pending
);
2443 * If task_rq(p) != rq, it cannot be migrated here, because we're
2444 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2445 * we're holding p->pi_lock.
2447 if (task_rq(p
) == rq
) {
2448 if (is_migration_disabled(p
))
2452 p
->migration_pending
= NULL
;
2455 if (cpumask_test_cpu(task_cpu(p
), &p
->cpus_mask
))
2459 if (task_on_rq_queued(p
))
2460 rq
= __migrate_task(rq
, &rf
, p
, arg
->dest_cpu
);
2462 p
->wake_cpu
= arg
->dest_cpu
;
2465 * XXX __migrate_task() can fail, at which point we might end
2466 * up running on a dodgy CPU, AFAICT this can only happen
2467 * during CPU hotplug, at which point we'll get pushed out
2468 * anyway, so it's probably not a big deal.
2471 } else if (pending
) {
2473 * This happens when we get migrated between migrate_enable()'s
2474 * preempt_enable() and scheduling the stopper task. At that
2475 * point we're a regular task again and not current anymore.
2477 * A !PREEMPT kernel has a giant hole here, which makes it far
2482 * The task moved before the stopper got to run. We're holding
2483 * ->pi_lock, so the allowed mask is stable - if it got
2484 * somewhere allowed, we're done.
2486 if (cpumask_test_cpu(task_cpu(p
), p
->cpus_ptr
)) {
2487 p
->migration_pending
= NULL
;
2493 * When migrate_enable() hits a rq mis-match we can't reliably
2494 * determine is_migration_disabled() and so have to chase after
2497 WARN_ON_ONCE(!pending
->stop_pending
);
2498 task_rq_unlock(rq
, p
, &rf
);
2499 stop_one_cpu_nowait(task_cpu(p
), migration_cpu_stop
,
2500 &pending
->arg
, &pending
->stop_work
);
2505 pending
->stop_pending
= false;
2506 task_rq_unlock(rq
, p
, &rf
);
2509 complete_all(&pending
->done
);
2514 int push_cpu_stop(void *arg
)
2516 struct rq
*lowest_rq
= NULL
, *rq
= this_rq();
2517 struct task_struct
*p
= arg
;
2519 raw_spin_lock_irq(&p
->pi_lock
);
2520 raw_spin_rq_lock(rq
);
2522 if (task_rq(p
) != rq
)
2525 if (is_migration_disabled(p
)) {
2526 p
->migration_flags
|= MDF_PUSH
;
2530 p
->migration_flags
&= ~MDF_PUSH
;
2532 if (p
->sched_class
->find_lock_rq
)
2533 lowest_rq
= p
->sched_class
->find_lock_rq(p
, rq
);
2538 // XXX validate p is still the highest prio task
2539 if (task_rq(p
) == rq
) {
2540 deactivate_task(rq
, p
, 0);
2541 set_task_cpu(p
, lowest_rq
->cpu
);
2542 activate_task(lowest_rq
, p
, 0);
2543 resched_curr(lowest_rq
);
2546 double_unlock_balance(rq
, lowest_rq
);
2549 rq
->push_busy
= false;
2550 raw_spin_rq_unlock(rq
);
2551 raw_spin_unlock_irq(&p
->pi_lock
);
2558 * sched_class::set_cpus_allowed must do the below, but is not required to
2559 * actually call this function.
2561 void set_cpus_allowed_common(struct task_struct
*p
, struct affinity_context
*ctx
)
2563 if (ctx
->flags
& (SCA_MIGRATE_ENABLE
| SCA_MIGRATE_DISABLE
)) {
2564 p
->cpus_ptr
= ctx
->new_mask
;
2568 cpumask_copy(&p
->cpus_mask
, ctx
->new_mask
);
2569 p
->nr_cpus_allowed
= cpumask_weight(ctx
->new_mask
);
2572 * Swap in a new user_cpus_ptr if SCA_USER flag set
2574 if (ctx
->flags
& SCA_USER
)
2575 swap(p
->user_cpus_ptr
, ctx
->user_mask
);
2579 __do_set_cpus_allowed(struct task_struct
*p
, struct affinity_context
*ctx
)
2581 struct rq
*rq
= task_rq(p
);
2582 bool queued
, running
;
2585 * This here violates the locking rules for affinity, since we're only
2586 * supposed to change these variables while holding both rq->lock and
2589 * HOWEVER, it magically works, because ttwu() is the only code that
2590 * accesses these variables under p->pi_lock and only does so after
2591 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2592 * before finish_task().
2594 * XXX do further audits, this smells like something putrid.
2596 if (ctx
->flags
& SCA_MIGRATE_DISABLE
)
2597 SCHED_WARN_ON(!p
->on_cpu
);
2599 lockdep_assert_held(&p
->pi_lock
);
2601 queued
= task_on_rq_queued(p
);
2602 running
= task_current(rq
, p
);
2606 * Because __kthread_bind() calls this on blocked tasks without
2609 lockdep_assert_rq_held(rq
);
2610 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
2613 put_prev_task(rq
, p
);
2615 p
->sched_class
->set_cpus_allowed(p
, ctx
);
2618 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
2620 set_next_task(rq
, p
);
2624 * Used for kthread_bind() and select_fallback_rq(), in both cases the user
2625 * affinity (if any) should be destroyed too.
2627 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
2629 struct affinity_context ac
= {
2630 .new_mask
= new_mask
,
2632 .flags
= SCA_USER
, /* clear the user requested mask */
2634 union cpumask_rcuhead
{
2636 struct rcu_head rcu
;
2639 __do_set_cpus_allowed(p
, &ac
);
2642 * Because this is called with p->pi_lock held, it is not possible
2643 * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
2646 kfree_rcu((union cpumask_rcuhead
*)ac
.user_mask
, rcu
);
2649 static cpumask_t
*alloc_user_cpus_ptr(int node
)
2652 * See do_set_cpus_allowed() above for the rcu_head usage.
2654 int size
= max_t(int, cpumask_size(), sizeof(struct rcu_head
));
2656 return kmalloc_node(size
, GFP_KERNEL
, node
);
2659 int dup_user_cpus_ptr(struct task_struct
*dst
, struct task_struct
*src
,
2662 cpumask_t
*user_mask
;
2663 unsigned long flags
;
2666 * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2667 * may differ by now due to racing.
2669 dst
->user_cpus_ptr
= NULL
;
2672 * This check is racy and losing the race is a valid situation.
2673 * It is not worth the extra overhead of taking the pi_lock on
2676 if (data_race(!src
->user_cpus_ptr
))
2679 user_mask
= alloc_user_cpus_ptr(node
);
2684 * Use pi_lock to protect content of user_cpus_ptr
2686 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2687 * do_set_cpus_allowed().
2689 raw_spin_lock_irqsave(&src
->pi_lock
, flags
);
2690 if (src
->user_cpus_ptr
) {
2691 swap(dst
->user_cpus_ptr
, user_mask
);
2692 cpumask_copy(dst
->user_cpus_ptr
, src
->user_cpus_ptr
);
2694 raw_spin_unlock_irqrestore(&src
->pi_lock
, flags
);
2696 if (unlikely(user_mask
))
2702 static inline struct cpumask
*clear_user_cpus_ptr(struct task_struct
*p
)
2704 struct cpumask
*user_mask
= NULL
;
2706 swap(p
->user_cpus_ptr
, user_mask
);
2711 void release_user_cpus_ptr(struct task_struct
*p
)
2713 kfree(clear_user_cpus_ptr(p
));
2717 * This function is wildly self concurrent; here be dragons.
2720 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2721 * designated task is enqueued on an allowed CPU. If that task is currently
2722 * running, we have to kick it out using the CPU stopper.
2724 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2727 * Initial conditions: P0->cpus_mask = [0, 1]
2731 * migrate_disable();
2733 * set_cpus_allowed_ptr(P0, [1]);
2735 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2736 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2737 * This means we need the following scheme:
2741 * migrate_disable();
2743 * set_cpus_allowed_ptr(P0, [1]);
2747 * __set_cpus_allowed_ptr();
2748 * <wakes local stopper>
2749 * `--> <woken on migration completion>
2751 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2752 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2753 * task p are serialized by p->pi_lock, which we can leverage: the one that
2754 * should come into effect at the end of the Migrate-Disable region is the last
2755 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2756 * but we still need to properly signal those waiting tasks at the appropriate
2759 * This is implemented using struct set_affinity_pending. The first
2760 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2761 * setup an instance of that struct and install it on the targeted task_struct.
2762 * Any and all further callers will reuse that instance. Those then wait for
2763 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2764 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2767 * (1) In the cases covered above. There is one more where the completion is
2768 * signaled within affine_move_task() itself: when a subsequent affinity request
2769 * occurs after the stopper bailed out due to the targeted task still being
2770 * Migrate-Disable. Consider:
2772 * Initial conditions: P0->cpus_mask = [0, 1]
2776 * migrate_disable();
2778 * set_cpus_allowed_ptr(P0, [1]);
2781 * migration_cpu_stop()
2782 * is_migration_disabled()
2784 * set_cpus_allowed_ptr(P0, [0, 1]);
2785 * <signal completion>
2788 * Note that the above is safe vs a concurrent migrate_enable(), as any
2789 * pending affinity completion is preceded by an uninstallation of
2790 * p->migration_pending done with p->pi_lock held.
2792 static int affine_move_task(struct rq
*rq
, struct task_struct
*p
, struct rq_flags
*rf
,
2793 int dest_cpu
, unsigned int flags
)
2794 __releases(rq
->lock
)
2795 __releases(p
->pi_lock
)
2797 struct set_affinity_pending my_pending
= { }, *pending
= NULL
;
2798 bool stop_pending
, complete
= false;
2800 /* Can the task run on the task's current CPU? If so, we're done */
2801 if (cpumask_test_cpu(task_cpu(p
), &p
->cpus_mask
)) {
2802 struct task_struct
*push_task
= NULL
;
2804 if ((flags
& SCA_MIGRATE_ENABLE
) &&
2805 (p
->migration_flags
& MDF_PUSH
) && !rq
->push_busy
) {
2806 rq
->push_busy
= true;
2807 push_task
= get_task_struct(p
);
2811 * If there are pending waiters, but no pending stop_work,
2812 * then complete now.
2814 pending
= p
->migration_pending
;
2815 if (pending
&& !pending
->stop_pending
) {
2816 p
->migration_pending
= NULL
;
2820 task_rq_unlock(rq
, p
, rf
);
2823 stop_one_cpu_nowait(rq
->cpu
, push_cpu_stop
,
2828 complete_all(&pending
->done
);
2833 if (!(flags
& SCA_MIGRATE_ENABLE
)) {
2834 /* serialized by p->pi_lock */
2835 if (!p
->migration_pending
) {
2836 /* Install the request */
2837 refcount_set(&my_pending
.refs
, 1);
2838 init_completion(&my_pending
.done
);
2839 my_pending
.arg
= (struct migration_arg
) {
2841 .dest_cpu
= dest_cpu
,
2842 .pending
= &my_pending
,
2845 p
->migration_pending
= &my_pending
;
2847 pending
= p
->migration_pending
;
2848 refcount_inc(&pending
->refs
);
2850 * Affinity has changed, but we've already installed a
2851 * pending. migration_cpu_stop() *must* see this, else
2852 * we risk a completion of the pending despite having a
2853 * task on a disallowed CPU.
2855 * Serialized by p->pi_lock, so this is safe.
2857 pending
->arg
.dest_cpu
= dest_cpu
;
2860 pending
= p
->migration_pending
;
2862 * - !MIGRATE_ENABLE:
2863 * we'll have installed a pending if there wasn't one already.
2866 * we're here because the current CPU isn't matching anymore,
2867 * the only way that can happen is because of a concurrent
2868 * set_cpus_allowed_ptr() call, which should then still be
2869 * pending completion.
2871 * Either way, we really should have a @pending here.
2873 if (WARN_ON_ONCE(!pending
)) {
2874 task_rq_unlock(rq
, p
, rf
);
2878 if (task_on_cpu(rq
, p
) || READ_ONCE(p
->__state
) == TASK_WAKING
) {
2880 * MIGRATE_ENABLE gets here because 'p == current', but for
2881 * anything else we cannot do is_migration_disabled(), punt
2882 * and have the stopper function handle it all race-free.
2884 stop_pending
= pending
->stop_pending
;
2886 pending
->stop_pending
= true;
2888 if (flags
& SCA_MIGRATE_ENABLE
)
2889 p
->migration_flags
&= ~MDF_PUSH
;
2891 task_rq_unlock(rq
, p
, rf
);
2893 if (!stop_pending
) {
2894 stop_one_cpu_nowait(cpu_of(rq
), migration_cpu_stop
,
2895 &pending
->arg
, &pending
->stop_work
);
2898 if (flags
& SCA_MIGRATE_ENABLE
)
2902 if (!is_migration_disabled(p
)) {
2903 if (task_on_rq_queued(p
))
2904 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
2906 if (!pending
->stop_pending
) {
2907 p
->migration_pending
= NULL
;
2911 task_rq_unlock(rq
, p
, rf
);
2914 complete_all(&pending
->done
);
2917 wait_for_completion(&pending
->done
);
2919 if (refcount_dec_and_test(&pending
->refs
))
2920 wake_up_var(&pending
->refs
); /* No UaF, just an address */
2923 * Block the original owner of &pending until all subsequent callers
2924 * have seen the completion and decremented the refcount
2926 wait_var_event(&my_pending
.refs
, !refcount_read(&my_pending
.refs
));
2929 WARN_ON_ONCE(my_pending
.stop_pending
);
2935 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2937 static int __set_cpus_allowed_ptr_locked(struct task_struct
*p
,
2938 struct affinity_context
*ctx
,
2940 struct rq_flags
*rf
)
2941 __releases(rq
->lock
)
2942 __releases(p
->pi_lock
)
2944 const struct cpumask
*cpu_allowed_mask
= task_cpu_possible_mask(p
);
2945 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
2946 bool kthread
= p
->flags
& PF_KTHREAD
;
2947 unsigned int dest_cpu
;
2950 update_rq_clock(rq
);
2952 if (kthread
|| is_migration_disabled(p
)) {
2954 * Kernel threads are allowed on online && !active CPUs,
2955 * however, during cpu-hot-unplug, even these might get pushed
2956 * away if not KTHREAD_IS_PER_CPU.
2958 * Specifically, migration_disabled() tasks must not fail the
2959 * cpumask_any_and_distribute() pick below, esp. so on
2960 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2961 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2963 cpu_valid_mask
= cpu_online_mask
;
2966 if (!kthread
&& !cpumask_subset(ctx
->new_mask
, cpu_allowed_mask
)) {
2972 * Must re-check here, to close a race against __kthread_bind(),
2973 * sched_setaffinity() is not guaranteed to observe the flag.
2975 if ((ctx
->flags
& SCA_CHECK
) && (p
->flags
& PF_NO_SETAFFINITY
)) {
2980 if (!(ctx
->flags
& SCA_MIGRATE_ENABLE
)) {
2981 if (cpumask_equal(&p
->cpus_mask
, ctx
->new_mask
)) {
2982 if (ctx
->flags
& SCA_USER
)
2983 swap(p
->user_cpus_ptr
, ctx
->user_mask
);
2987 if (WARN_ON_ONCE(p
== current
&&
2988 is_migration_disabled(p
) &&
2989 !cpumask_test_cpu(task_cpu(p
), ctx
->new_mask
))) {
2996 * Picking a ~random cpu helps in cases where we are changing affinity
2997 * for groups of tasks (ie. cpuset), so that load balancing is not
2998 * immediately required to distribute the tasks within their new mask.
3000 dest_cpu
= cpumask_any_and_distribute(cpu_valid_mask
, ctx
->new_mask
);
3001 if (dest_cpu
>= nr_cpu_ids
) {
3006 __do_set_cpus_allowed(p
, ctx
);
3008 return affine_move_task(rq
, p
, rf
, dest_cpu
, ctx
->flags
);
3011 task_rq_unlock(rq
, p
, rf
);
3017 * Change a given task's CPU affinity. Migrate the thread to a
3018 * proper CPU and schedule it away if the CPU it's executing on
3019 * is removed from the allowed bitmask.
3021 * NOTE: the caller must have a valid reference to the task, the
3022 * task must not exit() & deallocate itself prematurely. The
3023 * call is not atomic; no spinlocks may be held.
3025 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
3026 struct affinity_context
*ctx
)
3031 rq
= task_rq_lock(p
, &rf
);
3033 * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
3036 if (p
->user_cpus_ptr
&&
3037 !(ctx
->flags
& (SCA_USER
| SCA_MIGRATE_ENABLE
| SCA_MIGRATE_DISABLE
)) &&
3038 cpumask_and(rq
->scratch_mask
, ctx
->new_mask
, p
->user_cpus_ptr
))
3039 ctx
->new_mask
= rq
->scratch_mask
;
3041 return __set_cpus_allowed_ptr_locked(p
, ctx
, rq
, &rf
);
3044 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
3046 struct affinity_context ac
= {
3047 .new_mask
= new_mask
,
3051 return __set_cpus_allowed_ptr(p
, &ac
);
3053 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
3056 * Change a given task's CPU affinity to the intersection of its current
3057 * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
3058 * If user_cpus_ptr is defined, use it as the basis for restricting CPU
3059 * affinity or use cpu_online_mask instead.
3061 * If the resulting mask is empty, leave the affinity unchanged and return
3064 static int restrict_cpus_allowed_ptr(struct task_struct
*p
,
3065 struct cpumask
*new_mask
,
3066 const struct cpumask
*subset_mask
)
3068 struct affinity_context ac
= {
3069 .new_mask
= new_mask
,
3076 rq
= task_rq_lock(p
, &rf
);
3079 * Forcefully restricting the affinity of a deadline task is
3080 * likely to cause problems, so fail and noisily override the
3083 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
3088 if (!cpumask_and(new_mask
, task_user_cpus(p
), subset_mask
)) {
3093 return __set_cpus_allowed_ptr_locked(p
, &ac
, rq
, &rf
);
3096 task_rq_unlock(rq
, p
, &rf
);
3101 * Restrict the CPU affinity of task @p so that it is a subset of
3102 * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
3103 * old affinity mask. If the resulting mask is empty, we warn and walk
3104 * up the cpuset hierarchy until we find a suitable mask.
3106 void force_compatible_cpus_allowed_ptr(struct task_struct
*p
)
3108 cpumask_var_t new_mask
;
3109 const struct cpumask
*override_mask
= task_cpu_possible_mask(p
);
3111 alloc_cpumask_var(&new_mask
, GFP_KERNEL
);
3114 * __migrate_task() can fail silently in the face of concurrent
3115 * offlining of the chosen destination CPU, so take the hotplug
3116 * lock to ensure that the migration succeeds.
3119 if (!cpumask_available(new_mask
))
3122 if (!restrict_cpus_allowed_ptr(p
, new_mask
, override_mask
))
3126 * We failed to find a valid subset of the affinity mask for the
3127 * task, so override it based on its cpuset hierarchy.
3129 cpuset_cpus_allowed(p
, new_mask
);
3130 override_mask
= new_mask
;
3133 if (printk_ratelimit()) {
3134 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3135 task_pid_nr(p
), p
->comm
,
3136 cpumask_pr_args(override_mask
));
3139 WARN_ON(set_cpus_allowed_ptr(p
, override_mask
));
3142 free_cpumask_var(new_mask
);
3146 __sched_setaffinity(struct task_struct
*p
, struct affinity_context
*ctx
);
3149 * Restore the affinity of a task @p which was previously restricted by a
3150 * call to force_compatible_cpus_allowed_ptr().
3152 * It is the caller's responsibility to serialise this with any calls to
3153 * force_compatible_cpus_allowed_ptr(@p).
3155 void relax_compatible_cpus_allowed_ptr(struct task_struct
*p
)
3157 struct affinity_context ac
= {
3158 .new_mask
= task_user_cpus(p
),
3164 * Try to restore the old affinity mask with __sched_setaffinity().
3165 * Cpuset masking will be done there too.
3167 ret
= __sched_setaffinity(p
, &ac
);
3171 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
3173 #ifdef CONFIG_SCHED_DEBUG
3174 unsigned int state
= READ_ONCE(p
->__state
);
3177 * We should never call set_task_cpu() on a blocked task,
3178 * ttwu() will sort out the placement.
3180 WARN_ON_ONCE(state
!= TASK_RUNNING
&& state
!= TASK_WAKING
&& !p
->on_rq
);
3183 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3184 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3185 * time relying on p->on_rq.
3187 WARN_ON_ONCE(state
== TASK_RUNNING
&&
3188 p
->sched_class
== &fair_sched_class
&&
3189 (p
->on_rq
&& !task_on_rq_migrating(p
)));
3191 #ifdef CONFIG_LOCKDEP
3193 * The caller should hold either p->pi_lock or rq->lock, when changing
3194 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3196 * sched_move_task() holds both and thus holding either pins the cgroup,
3199 * Furthermore, all task_rq users should acquire both locks, see
3202 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
3203 lockdep_is_held(__rq_lockp(task_rq(p
)))));
3206 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3208 WARN_ON_ONCE(!cpu_online(new_cpu
));
3210 WARN_ON_ONCE(is_migration_disabled(p
));
3213 trace_sched_migrate_task(p
, new_cpu
);
3215 if (task_cpu(p
) != new_cpu
) {
3216 if (p
->sched_class
->migrate_task_rq
)
3217 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
3218 p
->se
.nr_migrations
++;
3220 sched_mm_cid_migrate_from(p
);
3221 perf_event_task_migrate(p
);
3224 __set_task_cpu(p
, new_cpu
);
3227 #ifdef CONFIG_NUMA_BALANCING
3228 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
3230 if (task_on_rq_queued(p
)) {
3231 struct rq
*src_rq
, *dst_rq
;
3232 struct rq_flags srf
, drf
;
3234 src_rq
= task_rq(p
);
3235 dst_rq
= cpu_rq(cpu
);
3237 rq_pin_lock(src_rq
, &srf
);
3238 rq_pin_lock(dst_rq
, &drf
);
3240 deactivate_task(src_rq
, p
, 0);
3241 set_task_cpu(p
, cpu
);
3242 activate_task(dst_rq
, p
, 0);
3243 check_preempt_curr(dst_rq
, p
, 0);
3245 rq_unpin_lock(dst_rq
, &drf
);
3246 rq_unpin_lock(src_rq
, &srf
);
3250 * Task isn't running anymore; make it appear like we migrated
3251 * it before it went to sleep. This means on wakeup we make the
3252 * previous CPU our target instead of where it really is.
3258 struct migration_swap_arg
{
3259 struct task_struct
*src_task
, *dst_task
;
3260 int src_cpu
, dst_cpu
;
3263 static int migrate_swap_stop(void *data
)
3265 struct migration_swap_arg
*arg
= data
;
3266 struct rq
*src_rq
, *dst_rq
;
3269 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
3272 src_rq
= cpu_rq(arg
->src_cpu
);
3273 dst_rq
= cpu_rq(arg
->dst_cpu
);
3275 double_raw_lock(&arg
->src_task
->pi_lock
,
3276 &arg
->dst_task
->pi_lock
);
3277 double_rq_lock(src_rq
, dst_rq
);
3279 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
3282 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
3285 if (!cpumask_test_cpu(arg
->dst_cpu
, arg
->src_task
->cpus_ptr
))
3288 if (!cpumask_test_cpu(arg
->src_cpu
, arg
->dst_task
->cpus_ptr
))
3291 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
3292 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
3297 double_rq_unlock(src_rq
, dst_rq
);
3298 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
3299 raw_spin_unlock(&arg
->src_task
->pi_lock
);
3305 * Cross migrate two tasks
3307 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
,
3308 int target_cpu
, int curr_cpu
)
3310 struct migration_swap_arg arg
;
3313 arg
= (struct migration_swap_arg
){
3315 .src_cpu
= curr_cpu
,
3317 .dst_cpu
= target_cpu
,
3320 if (arg
.src_cpu
== arg
.dst_cpu
)
3324 * These three tests are all lockless; this is OK since all of them
3325 * will be re-checked with proper locks held further down the line.
3327 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
3330 if (!cpumask_test_cpu(arg
.dst_cpu
, arg
.src_task
->cpus_ptr
))
3333 if (!cpumask_test_cpu(arg
.src_cpu
, arg
.dst_task
->cpus_ptr
))
3336 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
3337 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
3342 #endif /* CONFIG_NUMA_BALANCING */
3345 * wait_task_inactive - wait for a thread to unschedule.
3347 * Wait for the thread to block in any of the states set in @match_state.
3348 * If it changes, i.e. @p might have woken up, then return zero. When we
3349 * succeed in waiting for @p to be off its CPU, we return a positive number
3350 * (its total switch count). If a second call a short while later returns the
3351 * same number, the caller can be sure that @p has remained unscheduled the
3354 * The caller must ensure that the task *will* unschedule sometime soon,
3355 * else this function might spin for a *long* time. This function can't
3356 * be called with interrupts off, or it may introduce deadlock with
3357 * smp_call_function() if an IPI is sent by the same process we are
3358 * waiting to become inactive.
3360 unsigned long wait_task_inactive(struct task_struct
*p
, unsigned int match_state
)
3362 int running
, queued
;
3369 * We do the initial early heuristics without holding
3370 * any task-queue locks at all. We'll only try to get
3371 * the runqueue lock when things look like they will
3377 * If the task is actively running on another CPU
3378 * still, just relax and busy-wait without holding
3381 * NOTE! Since we don't hold any locks, it's not
3382 * even sure that "rq" stays as the right runqueue!
3383 * But we don't care, since "task_on_cpu()" will
3384 * return false if the runqueue has changed and p
3385 * is actually now running somewhere else!
3387 while (task_on_cpu(rq
, p
)) {
3388 if (!(READ_ONCE(p
->__state
) & match_state
))
3394 * Ok, time to look more closely! We need the rq
3395 * lock now, to be *sure*. If we're wrong, we'll
3396 * just go back and repeat.
3398 rq
= task_rq_lock(p
, &rf
);
3399 trace_sched_wait_task(p
);
3400 running
= task_on_cpu(rq
, p
);
3401 queued
= task_on_rq_queued(p
);
3403 if (READ_ONCE(p
->__state
) & match_state
)
3404 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
3405 task_rq_unlock(rq
, p
, &rf
);
3408 * If it changed from the expected state, bail out now.
3410 if (unlikely(!ncsw
))
3414 * Was it really running after all now that we
3415 * checked with the proper locks actually held?
3417 * Oops. Go back and try again..
3419 if (unlikely(running
)) {
3425 * It's not enough that it's not actively running,
3426 * it must be off the runqueue _entirely_, and not
3429 * So if it was still runnable (but just not actively
3430 * running right now), it's preempted, and we should
3431 * yield - it could be a while.
3433 if (unlikely(queued
)) {
3434 ktime_t to
= NSEC_PER_SEC
/ HZ
;
3436 set_current_state(TASK_UNINTERRUPTIBLE
);
3437 schedule_hrtimeout(&to
, HRTIMER_MODE_REL_HARD
);
3442 * Ahh, all good. It wasn't running, and it wasn't
3443 * runnable, which means that it will never become
3444 * running in the future either. We're all done!
3453 * kick_process - kick a running thread to enter/exit the kernel
3454 * @p: the to-be-kicked thread
3456 * Cause a process which is running on another CPU to enter
3457 * kernel-mode, without any delay. (to get signals handled.)
3459 * NOTE: this function doesn't have to take the runqueue lock,
3460 * because all it wants to ensure is that the remote task enters
3461 * the kernel. If the IPI races and the task has been migrated
3462 * to another CPU then no harm is done and the purpose has been
3465 void kick_process(struct task_struct
*p
)
3471 if ((cpu
!= smp_processor_id()) && task_curr(p
))
3472 smp_send_reschedule(cpu
);
3475 EXPORT_SYMBOL_GPL(kick_process
);
3478 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3480 * A few notes on cpu_active vs cpu_online:
3482 * - cpu_active must be a subset of cpu_online
3484 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3485 * see __set_cpus_allowed_ptr(). At this point the newly online
3486 * CPU isn't yet part of the sched domains, and balancing will not
3489 * - on CPU-down we clear cpu_active() to mask the sched domains and
3490 * avoid the load balancer to place new tasks on the to be removed
3491 * CPU. Existing tasks will remain running there and will be taken
3494 * This means that fallback selection must not select !active CPUs.
3495 * And can assume that any active CPU must be online. Conversely
3496 * select_task_rq() below may allow selection of !active CPUs in order
3497 * to satisfy the above rules.
3499 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
3501 int nid
= cpu_to_node(cpu
);
3502 const struct cpumask
*nodemask
= NULL
;
3503 enum { cpuset
, possible
, fail
} state
= cpuset
;
3507 * If the node that the CPU is on has been offlined, cpu_to_node()
3508 * will return -1. There is no CPU on the node, and we should
3509 * select the CPU on the other node.
3512 nodemask
= cpumask_of_node(nid
);
3514 /* Look for allowed, online CPU in same node. */
3515 for_each_cpu(dest_cpu
, nodemask
) {
3516 if (is_cpu_allowed(p
, dest_cpu
))
3522 /* Any allowed, online CPU? */
3523 for_each_cpu(dest_cpu
, p
->cpus_ptr
) {
3524 if (!is_cpu_allowed(p
, dest_cpu
))
3530 /* No more Mr. Nice Guy. */
3533 if (cpuset_cpus_allowed_fallback(p
)) {
3540 * XXX When called from select_task_rq() we only
3541 * hold p->pi_lock and again violate locking order.
3543 * More yuck to audit.
3545 do_set_cpus_allowed(p
, task_cpu_possible_mask(p
));
3555 if (state
!= cpuset
) {
3557 * Don't tell them about moving exiting tasks or
3558 * kernel threads (both mm NULL), since they never
3561 if (p
->mm
&& printk_ratelimit()) {
3562 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3563 task_pid_nr(p
), p
->comm
, cpu
);
3571 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3574 int select_task_rq(struct task_struct
*p
, int cpu
, int wake_flags
)
3576 lockdep_assert_held(&p
->pi_lock
);
3578 if (p
->nr_cpus_allowed
> 1 && !is_migration_disabled(p
))
3579 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, wake_flags
);
3581 cpu
= cpumask_any(p
->cpus_ptr
);
3584 * In order not to call set_task_cpu() on a blocking task we need
3585 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3588 * Since this is common to all placement strategies, this lives here.
3590 * [ this allows ->select_task() to simply return task_cpu(p) and
3591 * not worry about this generic constraint ]
3593 if (unlikely(!is_cpu_allowed(p
, cpu
)))
3594 cpu
= select_fallback_rq(task_cpu(p
), p
);
3599 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
3601 static struct lock_class_key stop_pi_lock
;
3602 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
3603 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
3607 * Make it appear like a SCHED_FIFO task, its something
3608 * userspace knows about and won't get confused about.
3610 * Also, it will make PI more or less work without too
3611 * much confusion -- but then, stop work should not
3612 * rely on PI working anyway.
3614 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
3616 stop
->sched_class
= &stop_sched_class
;
3619 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3620 * adjust the effective priority of a task. As a result,
3621 * rt_mutex_setprio() can trigger (RT) balancing operations,
3622 * which can then trigger wakeups of the stop thread to push
3623 * around the current task.
3625 * The stop task itself will never be part of the PI-chain, it
3626 * never blocks, therefore that ->pi_lock recursion is safe.
3627 * Tell lockdep about this by placing the stop->pi_lock in its
3630 lockdep_set_class(&stop
->pi_lock
, &stop_pi_lock
);
3633 cpu_rq(cpu
)->stop
= stop
;
3637 * Reset it back to a normal scheduling class so that
3638 * it can die in pieces.
3640 old_stop
->sched_class
= &rt_sched_class
;
3644 #else /* CONFIG_SMP */
3646 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
3647 struct affinity_context
*ctx
)
3649 return set_cpus_allowed_ptr(p
, ctx
->new_mask
);
3652 static inline void migrate_disable_switch(struct rq
*rq
, struct task_struct
*p
) { }
3654 static inline bool rq_has_pinned_tasks(struct rq
*rq
)
3659 static inline cpumask_t
*alloc_user_cpus_ptr(int node
)
3664 #endif /* !CONFIG_SMP */
3667 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
3671 if (!schedstat_enabled())
3677 if (cpu
== rq
->cpu
) {
3678 __schedstat_inc(rq
->ttwu_local
);
3679 __schedstat_inc(p
->stats
.nr_wakeups_local
);
3681 struct sched_domain
*sd
;
3683 __schedstat_inc(p
->stats
.nr_wakeups_remote
);
3685 for_each_domain(rq
->cpu
, sd
) {
3686 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
3687 __schedstat_inc(sd
->ttwu_wake_remote
);
3694 if (wake_flags
& WF_MIGRATED
)
3695 __schedstat_inc(p
->stats
.nr_wakeups_migrate
);
3696 #endif /* CONFIG_SMP */
3698 __schedstat_inc(rq
->ttwu_count
);
3699 __schedstat_inc(p
->stats
.nr_wakeups
);
3701 if (wake_flags
& WF_SYNC
)
3702 __schedstat_inc(p
->stats
.nr_wakeups_sync
);
3706 * Mark the task runnable.
3708 static inline void ttwu_do_wakeup(struct task_struct
*p
)
3710 WRITE_ONCE(p
->__state
, TASK_RUNNING
);
3711 trace_sched_wakeup(p
);
3715 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
3716 struct rq_flags
*rf
)
3718 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
3720 lockdep_assert_rq_held(rq
);
3722 if (p
->sched_contributes_to_load
)
3723 rq
->nr_uninterruptible
--;
3726 if (wake_flags
& WF_MIGRATED
)
3727 en_flags
|= ENQUEUE_MIGRATED
;
3731 delayacct_blkio_end(p
);
3732 atomic_dec(&task_rq(p
)->nr_iowait
);
3735 activate_task(rq
, p
, en_flags
);
3736 check_preempt_curr(rq
, p
, wake_flags
);
3741 if (p
->sched_class
->task_woken
) {
3743 * Our task @p is fully woken up and running; so it's safe to
3744 * drop the rq->lock, hereafter rq is only used for statistics.
3746 rq_unpin_lock(rq
, rf
);
3747 p
->sched_class
->task_woken(rq
, p
);
3748 rq_repin_lock(rq
, rf
);
3751 if (rq
->idle_stamp
) {
3752 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
3753 u64 max
= 2*rq
->max_idle_balance_cost
;
3755 update_avg(&rq
->avg_idle
, delta
);
3757 if (rq
->avg_idle
> max
)
3760 rq
->wake_stamp
= jiffies
;
3761 rq
->wake_avg_idle
= rq
->avg_idle
/ 2;
3769 * Consider @p being inside a wait loop:
3772 * set_current_state(TASK_UNINTERRUPTIBLE);
3779 * __set_current_state(TASK_RUNNING);
3781 * between set_current_state() and schedule(). In this case @p is still
3782 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3785 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3786 * then schedule() must still happen and p->state can be changed to
3787 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3788 * need to do a full wakeup with enqueue.
3790 * Returns: %true when the wakeup is done,
3793 static int ttwu_runnable(struct task_struct
*p
, int wake_flags
)
3799 rq
= __task_rq_lock(p
, &rf
);
3800 if (task_on_rq_queued(p
)) {
3801 if (!task_on_cpu(rq
, p
)) {
3803 * When on_rq && !on_cpu the task is preempted, see if
3804 * it should preempt the task that is current now.
3806 update_rq_clock(rq
);
3807 check_preempt_curr(rq
, p
, wake_flags
);
3812 __task_rq_unlock(rq
, &rf
);
3818 void sched_ttwu_pending(void *arg
)
3820 struct llist_node
*llist
= arg
;
3821 struct rq
*rq
= this_rq();
3822 struct task_struct
*p
, *t
;
3828 rq_lock_irqsave(rq
, &rf
);
3829 update_rq_clock(rq
);
3831 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
.llist
) {
3832 if (WARN_ON_ONCE(p
->on_cpu
))
3833 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
3835 if (WARN_ON_ONCE(task_cpu(p
) != cpu_of(rq
)))
3836 set_task_cpu(p
, cpu_of(rq
));
3838 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
3842 * Must be after enqueueing at least once task such that
3843 * idle_cpu() does not observe a false-negative -- if it does,
3844 * it is possible for select_idle_siblings() to stack a number
3845 * of tasks on this CPU during that window.
3847 * It is ok to clear ttwu_pending when another task pending.
3848 * We will receive IPI after local irq enabled and then enqueue it.
3849 * Since now nr_running > 0, idle_cpu() will always get correct result.
3851 WRITE_ONCE(rq
->ttwu_pending
, 0);
3852 rq_unlock_irqrestore(rq
, &rf
);
3856 * Prepare the scene for sending an IPI for a remote smp_call
3858 * Returns true if the caller can proceed with sending the IPI.
3859 * Returns false otherwise.
3861 bool call_function_single_prep_ipi(int cpu
)
3863 if (set_nr_if_polling(cpu_rq(cpu
)->idle
)) {
3864 trace_sched_wake_idle_without_ipi(cpu
);
3872 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3873 * necessary. The wakee CPU on receipt of the IPI will queue the task
3874 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3875 * of the wakeup instead of the waker.
3877 static void __ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3879 struct rq
*rq
= cpu_rq(cpu
);
3881 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
3883 WRITE_ONCE(rq
->ttwu_pending
, 1);
3884 __smp_call_single_queue(cpu
, &p
->wake_entry
.llist
);
3887 void wake_up_if_idle(int cpu
)
3889 struct rq
*rq
= cpu_rq(cpu
);
3894 if (!is_idle_task(rcu_dereference(rq
->curr
)))
3897 rq_lock_irqsave(rq
, &rf
);
3898 if (is_idle_task(rq
->curr
))
3900 /* Else CPU is not idle, do nothing here: */
3901 rq_unlock_irqrestore(rq
, &rf
);
3907 bool cpus_share_cache(int this_cpu
, int that_cpu
)
3909 if (this_cpu
== that_cpu
)
3912 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
3915 static inline bool ttwu_queue_cond(struct task_struct
*p
, int cpu
)
3918 * Do not complicate things with the async wake_list while the CPU is
3921 if (!cpu_active(cpu
))
3924 /* Ensure the task will still be allowed to run on the CPU. */
3925 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
3929 * If the CPU does not share cache, then queue the task on the
3930 * remote rqs wakelist to avoid accessing remote data.
3932 if (!cpus_share_cache(smp_processor_id(), cpu
))
3935 if (cpu
== smp_processor_id())
3939 * If the wakee cpu is idle, or the task is descheduling and the
3940 * only running task on the CPU, then use the wakelist to offload
3941 * the task activation to the idle (or soon-to-be-idle) CPU as
3942 * the current CPU is likely busy. nr_running is checked to
3943 * avoid unnecessary task stacking.
3945 * Note that we can only get here with (wakee) p->on_rq=0,
3946 * p->on_cpu can be whatever, we've done the dequeue, so
3947 * the wakee has been accounted out of ->nr_running.
3949 if (!cpu_rq(cpu
)->nr_running
)
3955 static bool ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3957 if (sched_feat(TTWU_QUEUE
) && ttwu_queue_cond(p
, cpu
)) {
3958 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
3959 __ttwu_queue_wakelist(p
, cpu
, wake_flags
);
3966 #else /* !CONFIG_SMP */
3968 static inline bool ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3973 #endif /* CONFIG_SMP */
3975 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
3977 struct rq
*rq
= cpu_rq(cpu
);
3980 if (ttwu_queue_wakelist(p
, cpu
, wake_flags
))
3984 update_rq_clock(rq
);
3985 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
3990 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3992 * The caller holds p::pi_lock if p != current or has preemption
3993 * disabled when p == current.
3995 * The rules of PREEMPT_RT saved_state:
3997 * The related locking code always holds p::pi_lock when updating
3998 * p::saved_state, which means the code is fully serialized in both cases.
4000 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
4001 * bits set. This allows to distinguish all wakeup scenarios.
4003 static __always_inline
4004 bool ttwu_state_match(struct task_struct
*p
, unsigned int state
, int *success
)
4006 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)) {
4007 WARN_ON_ONCE((state
& TASK_RTLOCK_WAIT
) &&
4008 state
!= TASK_RTLOCK_WAIT
);
4011 if (READ_ONCE(p
->__state
) & state
) {
4016 #ifdef CONFIG_PREEMPT_RT
4018 * Saved state preserves the task state across blocking on
4019 * an RT lock. If the state matches, set p::saved_state to
4020 * TASK_RUNNING, but do not wake the task because it waits
4021 * for a lock wakeup. Also indicate success because from
4022 * the regular waker's point of view this has succeeded.
4024 * After acquiring the lock the task will restore p::__state
4025 * from p::saved_state which ensures that the regular
4026 * wakeup is not lost. The restore will also set
4027 * p::saved_state to TASK_RUNNING so any further tests will
4028 * not result in false positives vs. @success
4030 if (p
->saved_state
& state
) {
4031 p
->saved_state
= TASK_RUNNING
;
4039 * Notes on Program-Order guarantees on SMP systems.
4043 * The basic program-order guarantee on SMP systems is that when a task [t]
4044 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
4045 * execution on its new CPU [c1].
4047 * For migration (of runnable tasks) this is provided by the following means:
4049 * A) UNLOCK of the rq(c0)->lock scheduling out task t
4050 * B) migration for t is required to synchronize *both* rq(c0)->lock and
4051 * rq(c1)->lock (if not at the same time, then in that order).
4052 * C) LOCK of the rq(c1)->lock scheduling in task
4054 * Release/acquire chaining guarantees that B happens after A and C after B.
4055 * Note: the CPU doing B need not be c0 or c1
4064 * UNLOCK rq(0)->lock
4066 * LOCK rq(0)->lock // orders against CPU0
4068 * UNLOCK rq(0)->lock
4072 * UNLOCK rq(1)->lock
4074 * LOCK rq(1)->lock // orders against CPU2
4077 * UNLOCK rq(1)->lock
4080 * BLOCKING -- aka. SLEEP + WAKEUP
4082 * For blocking we (obviously) need to provide the same guarantee as for
4083 * migration. However the means are completely different as there is no lock
4084 * chain to provide order. Instead we do:
4086 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
4087 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
4091 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
4093 * LOCK rq(0)->lock LOCK X->pi_lock
4096 * smp_store_release(X->on_cpu, 0);
4098 * smp_cond_load_acquire(&X->on_cpu, !VAL);
4104 * X->state = RUNNING
4105 * UNLOCK rq(2)->lock
4107 * LOCK rq(2)->lock // orders against CPU1
4110 * UNLOCK rq(2)->lock
4113 * UNLOCK rq(0)->lock
4116 * However, for wakeups there is a second guarantee we must provide, namely we
4117 * must ensure that CONDITION=1 done by the caller can not be reordered with
4118 * accesses to the task state; see try_to_wake_up() and set_current_state().
4122 * try_to_wake_up - wake up a thread
4123 * @p: the thread to be awakened
4124 * @state: the mask of task states that can be woken
4125 * @wake_flags: wake modifier flags (WF_*)
4127 * Conceptually does:
4129 * If (@state & @p->state) @p->state = TASK_RUNNING.
4131 * If the task was not queued/runnable, also place it back on a runqueue.
4133 * This function is atomic against schedule() which would dequeue the task.
4135 * It issues a full memory barrier before accessing @p->state, see the comment
4136 * with set_current_state().
4138 * Uses p->pi_lock to serialize against concurrent wake-ups.
4140 * Relies on p->pi_lock stabilizing:
4143 * - p->sched_task_group
4144 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4146 * Tries really hard to only take one task_rq(p)->lock for performance.
4147 * Takes rq->lock in:
4148 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4149 * - ttwu_queue() -- new rq, for enqueue of the task;
4150 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4152 * As a consequence we race really badly with just about everything. See the
4153 * many memory barriers and their comments for details.
4155 * Return: %true if @p->state changes (an actual wakeup was done),
4159 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
4161 unsigned long flags
;
4162 int cpu
, success
= 0;
4167 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4168 * == smp_processor_id()'. Together this means we can special
4169 * case the whole 'p->on_rq && ttwu_runnable()' case below
4170 * without taking any locks.
4173 * - we rely on Program-Order guarantees for all the ordering,
4174 * - we're serialized against set_special_state() by virtue of
4175 * it disabling IRQs (this allows not taking ->pi_lock).
4177 if (!ttwu_state_match(p
, state
, &success
))
4180 trace_sched_waking(p
);
4186 * If we are going to wake up a thread waiting for CONDITION we
4187 * need to ensure that CONDITION=1 done by the caller can not be
4188 * reordered with p->state check below. This pairs with smp_store_mb()
4189 * in set_current_state() that the waiting thread does.
4191 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4192 smp_mb__after_spinlock();
4193 if (!ttwu_state_match(p
, state
, &success
))
4196 trace_sched_waking(p
);
4199 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4200 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4201 * in smp_cond_load_acquire() below.
4203 * sched_ttwu_pending() try_to_wake_up()
4204 * STORE p->on_rq = 1 LOAD p->state
4207 * __schedule() (switch to task 'p')
4208 * LOCK rq->lock smp_rmb();
4209 * smp_mb__after_spinlock();
4213 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4215 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4216 * __schedule(). See the comment for smp_mb__after_spinlock().
4218 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4221 if (READ_ONCE(p
->on_rq
) && ttwu_runnable(p
, wake_flags
))
4226 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4227 * possible to, falsely, observe p->on_cpu == 0.
4229 * One must be running (->on_cpu == 1) in order to remove oneself
4230 * from the runqueue.
4232 * __schedule() (switch to task 'p') try_to_wake_up()
4233 * STORE p->on_cpu = 1 LOAD p->on_rq
4236 * __schedule() (put 'p' to sleep)
4237 * LOCK rq->lock smp_rmb();
4238 * smp_mb__after_spinlock();
4239 * STORE p->on_rq = 0 LOAD p->on_cpu
4241 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4242 * __schedule(). See the comment for smp_mb__after_spinlock().
4244 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4245 * schedule()'s deactivate_task() has 'happened' and p will no longer
4246 * care about it's own p->state. See the comment in __schedule().
4248 smp_acquire__after_ctrl_dep();
4251 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4252 * == 0), which means we need to do an enqueue, change p->state to
4253 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4254 * enqueue, such as ttwu_queue_wakelist().
4256 WRITE_ONCE(p
->__state
, TASK_WAKING
);
4259 * If the owning (remote) CPU is still in the middle of schedule() with
4260 * this task as prev, considering queueing p on the remote CPUs wake_list
4261 * which potentially sends an IPI instead of spinning on p->on_cpu to
4262 * let the waker make forward progress. This is safe because IRQs are
4263 * disabled and the IPI will deliver after on_cpu is cleared.
4265 * Ensure we load task_cpu(p) after p->on_cpu:
4267 * set_task_cpu(p, cpu);
4268 * STORE p->cpu = @cpu
4269 * __schedule() (switch to task 'p')
4271 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4272 * STORE p->on_cpu = 1 LOAD p->cpu
4274 * to ensure we observe the correct CPU on which the task is currently
4277 if (smp_load_acquire(&p
->on_cpu
) &&
4278 ttwu_queue_wakelist(p
, task_cpu(p
), wake_flags
))
4282 * If the owning (remote) CPU is still in the middle of schedule() with
4283 * this task as prev, wait until it's done referencing the task.
4285 * Pairs with the smp_store_release() in finish_task().
4287 * This ensures that tasks getting woken will be fully ordered against
4288 * their previous state and preserve Program Order.
4290 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
4292 cpu
= select_task_rq(p
, p
->wake_cpu
, wake_flags
| WF_TTWU
);
4293 if (task_cpu(p
) != cpu
) {
4295 delayacct_blkio_end(p
);
4296 atomic_dec(&task_rq(p
)->nr_iowait
);
4299 wake_flags
|= WF_MIGRATED
;
4300 psi_ttwu_dequeue(p
);
4301 set_task_cpu(p
, cpu
);
4305 #endif /* CONFIG_SMP */
4307 ttwu_queue(p
, cpu
, wake_flags
);
4309 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4312 ttwu_stat(p
, task_cpu(p
), wake_flags
);
4318 static bool __task_needs_rq_lock(struct task_struct
*p
)
4320 unsigned int state
= READ_ONCE(p
->__state
);
4323 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4324 * the task is blocked. Make sure to check @state since ttwu() can drop
4325 * locks at the end, see ttwu_queue_wakelist().
4327 if (state
== TASK_RUNNING
|| state
== TASK_WAKING
)
4331 * Ensure we load p->on_rq after p->__state, otherwise it would be
4332 * possible to, falsely, observe p->on_rq == 0.
4334 * See try_to_wake_up() for a longer comment.
4342 * Ensure the task has finished __schedule() and will not be referenced
4343 * anymore. Again, see try_to_wake_up() for a longer comment.
4346 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
4353 * task_call_func - Invoke a function on task in fixed state
4354 * @p: Process for which the function is to be invoked, can be @current.
4355 * @func: Function to invoke.
4356 * @arg: Argument to function.
4358 * Fix the task in it's current state by avoiding wakeups and or rq operations
4359 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4360 * to work out what the state is, if required. Given that @func can be invoked
4361 * with a runqueue lock held, it had better be quite lightweight.
4364 * Whatever @func returns
4366 int task_call_func(struct task_struct
*p
, task_call_f func
, void *arg
)
4368 struct rq
*rq
= NULL
;
4372 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
4374 if (__task_needs_rq_lock(p
))
4375 rq
= __task_rq_lock(p
, &rf
);
4378 * At this point the task is pinned; either:
4379 * - blocked and we're holding off wakeups (pi->lock)
4380 * - woken, and we're holding off enqueue (rq->lock)
4381 * - queued, and we're holding off schedule (rq->lock)
4382 * - running, and we're holding off de-schedule (rq->lock)
4384 * The called function (@func) can use: task_curr(), p->on_rq and
4385 * p->__state to differentiate between these states.
4392 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
.flags
);
4397 * cpu_curr_snapshot - Return a snapshot of the currently running task
4398 * @cpu: The CPU on which to snapshot the task.
4400 * Returns the task_struct pointer of the task "currently" running on
4401 * the specified CPU. If the same task is running on that CPU throughout,
4402 * the return value will be a pointer to that task's task_struct structure.
4403 * If the CPU did any context switches even vaguely concurrently with the
4404 * execution of this function, the return value will be a pointer to the
4405 * task_struct structure of a randomly chosen task that was running on
4406 * that CPU somewhere around the time that this function was executing.
4408 * If the specified CPU was offline, the return value is whatever it
4409 * is, perhaps a pointer to the task_struct structure of that CPU's idle
4410 * task, but there is no guarantee. Callers wishing a useful return
4411 * value must take some action to ensure that the specified CPU remains
4412 * online throughout.
4414 * This function executes full memory barriers before and after fetching
4415 * the pointer, which permits the caller to confine this function's fetch
4416 * with respect to the caller's accesses to other shared variables.
4418 struct task_struct
*cpu_curr_snapshot(int cpu
)
4420 struct task_struct
*t
;
4422 smp_mb(); /* Pairing determined by caller's synchronization design. */
4423 t
= rcu_dereference(cpu_curr(cpu
));
4424 smp_mb(); /* Pairing determined by caller's synchronization design. */
4429 * wake_up_process - Wake up a specific process
4430 * @p: The process to be woken up.
4432 * Attempt to wake up the nominated process and move it to the set of runnable
4435 * Return: 1 if the process was woken up, 0 if it was already running.
4437 * This function executes a full memory barrier before accessing the task state.
4439 int wake_up_process(struct task_struct
*p
)
4441 return try_to_wake_up(p
, TASK_NORMAL
, 0);
4443 EXPORT_SYMBOL(wake_up_process
);
4445 int wake_up_state(struct task_struct
*p
, unsigned int state
)
4447 return try_to_wake_up(p
, state
, 0);
4451 * Perform scheduler related setup for a newly forked process p.
4452 * p is forked by current.
4454 * __sched_fork() is basic setup used by init_idle() too:
4456 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
4461 p
->se
.exec_start
= 0;
4462 p
->se
.sum_exec_runtime
= 0;
4463 p
->se
.prev_sum_exec_runtime
= 0;
4464 p
->se
.nr_migrations
= 0;
4466 INIT_LIST_HEAD(&p
->se
.group_node
);
4468 #ifdef CONFIG_FAIR_GROUP_SCHED
4469 p
->se
.cfs_rq
= NULL
;
4472 #ifdef CONFIG_SCHEDSTATS
4473 /* Even if schedstat is disabled, there should not be garbage */
4474 memset(&p
->stats
, 0, sizeof(p
->stats
));
4477 RB_CLEAR_NODE(&p
->dl
.rb_node
);
4478 init_dl_task_timer(&p
->dl
);
4479 init_dl_inactive_task_timer(&p
->dl
);
4480 __dl_clear_params(p
);
4482 INIT_LIST_HEAD(&p
->rt
.run_list
);
4484 p
->rt
.time_slice
= sched_rr_timeslice
;
4488 #ifdef CONFIG_PREEMPT_NOTIFIERS
4489 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
4492 #ifdef CONFIG_COMPACTION
4493 p
->capture_control
= NULL
;
4495 init_numa_balancing(clone_flags
, p
);
4497 p
->wake_entry
.u_flags
= CSD_TYPE_TTWU
;
4498 p
->migration_pending
= NULL
;
4500 init_sched_mm_cid(p
);
4503 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
4505 #ifdef CONFIG_NUMA_BALANCING
4507 int sysctl_numa_balancing_mode
;
4509 static void __set_numabalancing_state(bool enabled
)
4512 static_branch_enable(&sched_numa_balancing
);
4514 static_branch_disable(&sched_numa_balancing
);
4517 void set_numabalancing_state(bool enabled
)
4520 sysctl_numa_balancing_mode
= NUMA_BALANCING_NORMAL
;
4522 sysctl_numa_balancing_mode
= NUMA_BALANCING_DISABLED
;
4523 __set_numabalancing_state(enabled
);
4526 #ifdef CONFIG_PROC_SYSCTL
4527 static void reset_memory_tiering(void)
4529 struct pglist_data
*pgdat
;
4531 for_each_online_pgdat(pgdat
) {
4532 pgdat
->nbp_threshold
= 0;
4533 pgdat
->nbp_th_nr_cand
= node_page_state(pgdat
, PGPROMOTE_CANDIDATE
);
4534 pgdat
->nbp_th_start
= jiffies_to_msecs(jiffies
);
4538 static int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
4539 void *buffer
, size_t *lenp
, loff_t
*ppos
)
4543 int state
= sysctl_numa_balancing_mode
;
4545 if (write
&& !capable(CAP_SYS_ADMIN
))
4550 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
4554 if (!(sysctl_numa_balancing_mode
& NUMA_BALANCING_MEMORY_TIERING
) &&
4555 (state
& NUMA_BALANCING_MEMORY_TIERING
))
4556 reset_memory_tiering();
4557 sysctl_numa_balancing_mode
= state
;
4558 __set_numabalancing_state(state
);
4565 #ifdef CONFIG_SCHEDSTATS
4567 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
4569 static void set_schedstats(bool enabled
)
4572 static_branch_enable(&sched_schedstats
);
4574 static_branch_disable(&sched_schedstats
);
4577 void force_schedstat_enabled(void)
4579 if (!schedstat_enabled()) {
4580 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4581 static_branch_enable(&sched_schedstats
);
4585 static int __init
setup_schedstats(char *str
)
4591 if (!strcmp(str
, "enable")) {
4592 set_schedstats(true);
4594 } else if (!strcmp(str
, "disable")) {
4595 set_schedstats(false);
4600 pr_warn("Unable to parse schedstats=\n");
4604 __setup("schedstats=", setup_schedstats
);
4606 #ifdef CONFIG_PROC_SYSCTL
4607 static int sysctl_schedstats(struct ctl_table
*table
, int write
, void *buffer
,
4608 size_t *lenp
, loff_t
*ppos
)
4612 int state
= static_branch_likely(&sched_schedstats
);
4614 if (write
&& !capable(CAP_SYS_ADMIN
))
4619 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
4623 set_schedstats(state
);
4626 #endif /* CONFIG_PROC_SYSCTL */
4627 #endif /* CONFIG_SCHEDSTATS */
4629 #ifdef CONFIG_SYSCTL
4630 static struct ctl_table sched_core_sysctls
[] = {
4631 #ifdef CONFIG_SCHEDSTATS
4633 .procname
= "sched_schedstats",
4635 .maxlen
= sizeof(unsigned int),
4637 .proc_handler
= sysctl_schedstats
,
4638 .extra1
= SYSCTL_ZERO
,
4639 .extra2
= SYSCTL_ONE
,
4641 #endif /* CONFIG_SCHEDSTATS */
4642 #ifdef CONFIG_UCLAMP_TASK
4644 .procname
= "sched_util_clamp_min",
4645 .data
= &sysctl_sched_uclamp_util_min
,
4646 .maxlen
= sizeof(unsigned int),
4648 .proc_handler
= sysctl_sched_uclamp_handler
,
4651 .procname
= "sched_util_clamp_max",
4652 .data
= &sysctl_sched_uclamp_util_max
,
4653 .maxlen
= sizeof(unsigned int),
4655 .proc_handler
= sysctl_sched_uclamp_handler
,
4658 .procname
= "sched_util_clamp_min_rt_default",
4659 .data
= &sysctl_sched_uclamp_util_min_rt_default
,
4660 .maxlen
= sizeof(unsigned int),
4662 .proc_handler
= sysctl_sched_uclamp_handler
,
4664 #endif /* CONFIG_UCLAMP_TASK */
4665 #ifdef CONFIG_NUMA_BALANCING
4667 .procname
= "numa_balancing",
4668 .data
= NULL
, /* filled in by handler */
4669 .maxlen
= sizeof(unsigned int),
4671 .proc_handler
= sysctl_numa_balancing
,
4672 .extra1
= SYSCTL_ZERO
,
4673 .extra2
= SYSCTL_FOUR
,
4675 #endif /* CONFIG_NUMA_BALANCING */
4678 static int __init
sched_core_sysctl_init(void)
4680 register_sysctl_init("kernel", sched_core_sysctls
);
4683 late_initcall(sched_core_sysctl_init
);
4684 #endif /* CONFIG_SYSCTL */
4687 * fork()/clone()-time setup:
4689 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
4691 __sched_fork(clone_flags
, p
);
4693 * We mark the process as NEW here. This guarantees that
4694 * nobody will actually run it, and a signal or other external
4695 * event cannot wake it up and insert it on the runqueue either.
4697 p
->__state
= TASK_NEW
;
4700 * Make sure we do not leak PI boosting priority to the child.
4702 p
->prio
= current
->normal_prio
;
4707 * Revert to default priority/policy on fork if requested.
4709 if (unlikely(p
->sched_reset_on_fork
)) {
4710 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
4711 p
->policy
= SCHED_NORMAL
;
4712 p
->static_prio
= NICE_TO_PRIO(0);
4714 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
4715 p
->static_prio
= NICE_TO_PRIO(0);
4717 p
->prio
= p
->normal_prio
= p
->static_prio
;
4718 set_load_weight(p
, false);
4721 * We don't need the reset flag anymore after the fork. It has
4722 * fulfilled its duty:
4724 p
->sched_reset_on_fork
= 0;
4727 if (dl_prio(p
->prio
))
4729 else if (rt_prio(p
->prio
))
4730 p
->sched_class
= &rt_sched_class
;
4732 p
->sched_class
= &fair_sched_class
;
4734 init_entity_runnable_average(&p
->se
);
4737 #ifdef CONFIG_SCHED_INFO
4738 if (likely(sched_info_on()))
4739 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
4741 #if defined(CONFIG_SMP)
4744 init_task_preempt_count(p
);
4746 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
4747 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
4752 void sched_cgroup_fork(struct task_struct
*p
, struct kernel_clone_args
*kargs
)
4754 unsigned long flags
;
4757 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4758 * required yet, but lockdep gets upset if rules are violated.
4760 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4761 #ifdef CONFIG_CGROUP_SCHED
4763 struct task_group
*tg
;
4764 tg
= container_of(kargs
->cset
->subsys
[cpu_cgrp_id
],
4765 struct task_group
, css
);
4766 tg
= autogroup_task_group(p
, tg
);
4767 p
->sched_task_group
= tg
;
4772 * We're setting the CPU for the first time, we don't migrate,
4773 * so use __set_task_cpu().
4775 __set_task_cpu(p
, smp_processor_id());
4776 if (p
->sched_class
->task_fork
)
4777 p
->sched_class
->task_fork(p
);
4778 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4781 void sched_post_fork(struct task_struct
*p
)
4783 uclamp_post_fork(p
);
4786 unsigned long to_ratio(u64 period
, u64 runtime
)
4788 if (runtime
== RUNTIME_INF
)
4792 * Doing this here saves a lot of checks in all
4793 * the calling paths, and returning zero seems
4794 * safe for them anyway.
4799 return div64_u64(runtime
<< BW_SHIFT
, period
);
4803 * wake_up_new_task - wake up a newly created task for the first time.
4805 * This function will do some initial scheduler statistics housekeeping
4806 * that must be done for every newly created context, then puts the task
4807 * on the runqueue and wakes it.
4809 void wake_up_new_task(struct task_struct
*p
)
4814 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
4815 WRITE_ONCE(p
->__state
, TASK_RUNNING
);
4818 * Fork balancing, do it here and not earlier because:
4819 * - cpus_ptr can change in the fork path
4820 * - any previously selected CPU might disappear through hotplug
4822 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4823 * as we're not fully set-up yet.
4825 p
->recent_used_cpu
= task_cpu(p
);
4827 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), WF_FORK
));
4829 rq
= __task_rq_lock(p
, &rf
);
4830 update_rq_clock(rq
);
4831 post_init_entity_util_avg(p
);
4833 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
4834 trace_sched_wakeup_new(p
);
4835 check_preempt_curr(rq
, p
, WF_FORK
);
4837 if (p
->sched_class
->task_woken
) {
4839 * Nothing relies on rq->lock after this, so it's fine to
4842 rq_unpin_lock(rq
, &rf
);
4843 p
->sched_class
->task_woken(rq
, p
);
4844 rq_repin_lock(rq
, &rf
);
4847 task_rq_unlock(rq
, p
, &rf
);
4850 #ifdef CONFIG_PREEMPT_NOTIFIERS
4852 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key
);
4854 void preempt_notifier_inc(void)
4856 static_branch_inc(&preempt_notifier_key
);
4858 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
4860 void preempt_notifier_dec(void)
4862 static_branch_dec(&preempt_notifier_key
);
4864 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
4867 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4868 * @notifier: notifier struct to register
4870 void preempt_notifier_register(struct preempt_notifier
*notifier
)
4872 if (!static_branch_unlikely(&preempt_notifier_key
))
4873 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4875 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
4877 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
4880 * preempt_notifier_unregister - no longer interested in preemption notifications
4881 * @notifier: notifier struct to unregister
4883 * This is *not* safe to call from within a preemption notifier.
4885 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
4887 hlist_del(¬ifier
->link
);
4889 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
4891 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
4893 struct preempt_notifier
*notifier
;
4895 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
4896 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
4899 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
4901 if (static_branch_unlikely(&preempt_notifier_key
))
4902 __fire_sched_in_preempt_notifiers(curr
);
4906 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
4907 struct task_struct
*next
)
4909 struct preempt_notifier
*notifier
;
4911 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
4912 notifier
->ops
->sched_out(notifier
, next
);
4915 static __always_inline
void
4916 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
4917 struct task_struct
*next
)
4919 if (static_branch_unlikely(&preempt_notifier_key
))
4920 __fire_sched_out_preempt_notifiers(curr
, next
);
4923 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4925 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
4930 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
4931 struct task_struct
*next
)
4935 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4937 static inline void prepare_task(struct task_struct
*next
)
4941 * Claim the task as running, we do this before switching to it
4942 * such that any running task will have this set.
4944 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4945 * its ordering comment.
4947 WRITE_ONCE(next
->on_cpu
, 1);
4951 static inline void finish_task(struct task_struct
*prev
)
4955 * This must be the very last reference to @prev from this CPU. After
4956 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4957 * must ensure this doesn't happen until the switch is completely
4960 * In particular, the load of prev->state in finish_task_switch() must
4961 * happen before this.
4963 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4965 smp_store_release(&prev
->on_cpu
, 0);
4971 static void do_balance_callbacks(struct rq
*rq
, struct balance_callback
*head
)
4973 void (*func
)(struct rq
*rq
);
4974 struct balance_callback
*next
;
4976 lockdep_assert_rq_held(rq
);
4979 func
= (void (*)(struct rq
*))head
->func
;
4988 static void balance_push(struct rq
*rq
);
4991 * balance_push_callback is a right abuse of the callback interface and plays
4992 * by significantly different rules.
4994 * Where the normal balance_callback's purpose is to be ran in the same context
4995 * that queued it (only later, when it's safe to drop rq->lock again),
4996 * balance_push_callback is specifically targeted at __schedule().
4998 * This abuse is tolerated because it places all the unlikely/odd cases behind
4999 * a single test, namely: rq->balance_callback == NULL.
5001 struct balance_callback balance_push_callback
= {
5003 .func
= balance_push
,
5006 static inline struct balance_callback
*
5007 __splice_balance_callbacks(struct rq
*rq
, bool split
)
5009 struct balance_callback
*head
= rq
->balance_callback
;
5014 lockdep_assert_rq_held(rq
);
5016 * Must not take balance_push_callback off the list when
5017 * splice_balance_callbacks() and balance_callbacks() are not
5018 * in the same rq->lock section.
5020 * In that case it would be possible for __schedule() to interleave
5021 * and observe the list empty.
5023 if (split
&& head
== &balance_push_callback
)
5026 rq
->balance_callback
= NULL
;
5031 static inline struct balance_callback
*splice_balance_callbacks(struct rq
*rq
)
5033 return __splice_balance_callbacks(rq
, true);
5036 static void __balance_callbacks(struct rq
*rq
)
5038 do_balance_callbacks(rq
, __splice_balance_callbacks(rq
, false));
5041 static inline void balance_callbacks(struct rq
*rq
, struct balance_callback
*head
)
5043 unsigned long flags
;
5045 if (unlikely(head
)) {
5046 raw_spin_rq_lock_irqsave(rq
, flags
);
5047 do_balance_callbacks(rq
, head
);
5048 raw_spin_rq_unlock_irqrestore(rq
, flags
);
5054 static inline void __balance_callbacks(struct rq
*rq
)
5058 static inline struct balance_callback
*splice_balance_callbacks(struct rq
*rq
)
5063 static inline void balance_callbacks(struct rq
*rq
, struct balance_callback
*head
)
5070 prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
, struct rq_flags
*rf
)
5073 * Since the runqueue lock will be released by the next
5074 * task (which is an invalid locking op but in the case
5075 * of the scheduler it's an obvious special-case), so we
5076 * do an early lockdep release here:
5078 rq_unpin_lock(rq
, rf
);
5079 spin_release(&__rq_lockp(rq
)->dep_map
, _THIS_IP_
);
5080 #ifdef CONFIG_DEBUG_SPINLOCK
5081 /* this is a valid case when another task releases the spinlock */
5082 rq_lockp(rq
)->owner
= next
;
5086 static inline void finish_lock_switch(struct rq
*rq
)
5089 * If we are tracking spinlock dependencies then we have to
5090 * fix up the runqueue lock - which gets 'carried over' from
5091 * prev into current:
5093 spin_acquire(&__rq_lockp(rq
)->dep_map
, 0, 0, _THIS_IP_
);
5094 __balance_callbacks(rq
);
5095 raw_spin_rq_unlock_irq(rq
);
5099 * NOP if the arch has not defined these:
5102 #ifndef prepare_arch_switch
5103 # define prepare_arch_switch(next) do { } while (0)
5106 #ifndef finish_arch_post_lock_switch
5107 # define finish_arch_post_lock_switch() do { } while (0)
5110 static inline void kmap_local_sched_out(void)
5112 #ifdef CONFIG_KMAP_LOCAL
5113 if (unlikely(current
->kmap_ctrl
.idx
))
5114 __kmap_local_sched_out();
5118 static inline void kmap_local_sched_in(void)
5120 #ifdef CONFIG_KMAP_LOCAL
5121 if (unlikely(current
->kmap_ctrl
.idx
))
5122 __kmap_local_sched_in();
5127 * prepare_task_switch - prepare to switch tasks
5128 * @rq: the runqueue preparing to switch
5129 * @prev: the current task that is being switched out
5130 * @next: the task we are going to switch to.
5132 * This is called with the rq lock held and interrupts off. It must
5133 * be paired with a subsequent finish_task_switch after the context
5136 * prepare_task_switch sets up locking and calls architecture specific
5140 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
5141 struct task_struct
*next
)
5143 kcov_prepare_switch(prev
);
5144 sched_info_switch(rq
, prev
, next
);
5145 perf_event_task_sched_out(prev
, next
);
5147 fire_sched_out_preempt_notifiers(prev
, next
);
5148 kmap_local_sched_out();
5150 prepare_arch_switch(next
);
5154 * finish_task_switch - clean up after a task-switch
5155 * @prev: the thread we just switched away from.
5157 * finish_task_switch must be called after the context switch, paired
5158 * with a prepare_task_switch call before the context switch.
5159 * finish_task_switch will reconcile locking set up by prepare_task_switch,
5160 * and do any other architecture-specific cleanup actions.
5162 * Note that we may have delayed dropping an mm in context_switch(). If
5163 * so, we finish that here outside of the runqueue lock. (Doing it
5164 * with the lock held can cause deadlocks; see schedule() for
5167 * The context switch have flipped the stack from under us and restored the
5168 * local variables which were saved when this task called schedule() in the
5169 * past. prev == current is still correct but we need to recalculate this_rq
5170 * because prev may have moved to another CPU.
5172 static struct rq
*finish_task_switch(struct task_struct
*prev
)
5173 __releases(rq
->lock
)
5175 struct rq
*rq
= this_rq();
5176 struct mm_struct
*mm
= rq
->prev_mm
;
5177 unsigned int prev_state
;
5180 * The previous task will have left us with a preempt_count of 2
5181 * because it left us after:
5184 * preempt_disable(); // 1
5186 * raw_spin_lock_irq(&rq->lock) // 2
5188 * Also, see FORK_PREEMPT_COUNT.
5190 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
5191 "corrupted preempt_count: %s/%d/0x%x\n",
5192 current
->comm
, current
->pid
, preempt_count()))
5193 preempt_count_set(FORK_PREEMPT_COUNT
);
5198 * A task struct has one reference for the use as "current".
5199 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
5200 * schedule one last time. The schedule call will never return, and
5201 * the scheduled task must drop that reference.
5203 * We must observe prev->state before clearing prev->on_cpu (in
5204 * finish_task), otherwise a concurrent wakeup can get prev
5205 * running on another CPU and we could rave with its RUNNING -> DEAD
5206 * transition, resulting in a double drop.
5208 prev_state
= READ_ONCE(prev
->__state
);
5209 vtime_task_switch(prev
);
5210 perf_event_task_sched_in(prev
, current
);
5212 tick_nohz_task_switch();
5213 finish_lock_switch(rq
);
5214 finish_arch_post_lock_switch();
5215 kcov_finish_switch(current
);
5217 * kmap_local_sched_out() is invoked with rq::lock held and
5218 * interrupts disabled. There is no requirement for that, but the
5219 * sched out code does not have an interrupt enabled section.
5220 * Restoring the maps on sched in does not require interrupts being
5223 kmap_local_sched_in();
5225 fire_sched_in_preempt_notifiers(current
);
5227 * When switching through a kernel thread, the loop in
5228 * membarrier_{private,global}_expedited() may have observed that
5229 * kernel thread and not issued an IPI. It is therefore possible to
5230 * schedule between user->kernel->user threads without passing though
5231 * switch_mm(). Membarrier requires a barrier after storing to
5232 * rq->curr, before returning to userspace, so provide them here:
5234 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5235 * provided by mmdrop_lazy_tlb(),
5236 * - a sync_core for SYNC_CORE.
5239 membarrier_mm_sync_core_before_usermode(mm
);
5240 mmdrop_lazy_tlb_sched(mm
);
5243 if (unlikely(prev_state
== TASK_DEAD
)) {
5244 if (prev
->sched_class
->task_dead
)
5245 prev
->sched_class
->task_dead(prev
);
5247 /* Task is done with its stack. */
5248 put_task_stack(prev
);
5250 put_task_struct_rcu_user(prev
);
5257 * schedule_tail - first thing a freshly forked thread must call.
5258 * @prev: the thread we just switched away from.
5260 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
5261 __releases(rq
->lock
)
5264 * New tasks start with FORK_PREEMPT_COUNT, see there and
5265 * finish_task_switch() for details.
5267 * finish_task_switch() will drop rq->lock() and lower preempt_count
5268 * and the preempt_enable() will end up enabling preemption (on
5269 * PREEMPT_COUNT kernels).
5272 finish_task_switch(prev
);
5275 if (current
->set_child_tid
)
5276 put_user(task_pid_vnr(current
), current
->set_child_tid
);
5278 calculate_sigpending();
5282 * context_switch - switch to the new MM and the new thread's register state.
5284 static __always_inline
struct rq
*
5285 context_switch(struct rq
*rq
, struct task_struct
*prev
,
5286 struct task_struct
*next
, struct rq_flags
*rf
)
5288 prepare_task_switch(rq
, prev
, next
);
5291 * For paravirt, this is coupled with an exit in switch_to to
5292 * combine the page table reload and the switch backend into
5295 arch_start_context_switch(prev
);
5298 * kernel -> kernel lazy + transfer active
5299 * user -> kernel lazy + mmgrab_lazy_tlb() active
5301 * kernel -> user switch + mmdrop_lazy_tlb() active
5302 * user -> user switch
5304 * switch_mm_cid() needs to be updated if the barriers provided
5305 * by context_switch() are modified.
5307 if (!next
->mm
) { // to kernel
5308 enter_lazy_tlb(prev
->active_mm
, next
);
5310 next
->active_mm
= prev
->active_mm
;
5311 if (prev
->mm
) // from user
5312 mmgrab_lazy_tlb(prev
->active_mm
);
5314 prev
->active_mm
= NULL
;
5316 membarrier_switch_mm(rq
, prev
->active_mm
, next
->mm
);
5318 * sys_membarrier() requires an smp_mb() between setting
5319 * rq->curr / membarrier_switch_mm() and returning to userspace.
5321 * The below provides this either through switch_mm(), or in
5322 * case 'prev->active_mm == next->mm' through
5323 * finish_task_switch()'s mmdrop().
5325 switch_mm_irqs_off(prev
->active_mm
, next
->mm
, next
);
5326 lru_gen_use_mm(next
->mm
);
5328 if (!prev
->mm
) { // from kernel
5329 /* will mmdrop_lazy_tlb() in finish_task_switch(). */
5330 rq
->prev_mm
= prev
->active_mm
;
5331 prev
->active_mm
= NULL
;
5335 /* switch_mm_cid() requires the memory barriers above. */
5336 switch_mm_cid(rq
, prev
, next
);
5338 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
5340 prepare_lock_switch(rq
, next
, rf
);
5342 /* Here we just switch the register state and the stack. */
5343 switch_to(prev
, next
, prev
);
5346 return finish_task_switch(prev
);
5350 * nr_running and nr_context_switches:
5352 * externally visible scheduler statistics: current number of runnable
5353 * threads, total number of context switches performed since bootup.
5355 unsigned int nr_running(void)
5357 unsigned int i
, sum
= 0;
5359 for_each_online_cpu(i
)
5360 sum
+= cpu_rq(i
)->nr_running
;
5366 * Check if only the current task is running on the CPU.
5368 * Caution: this function does not check that the caller has disabled
5369 * preemption, thus the result might have a time-of-check-to-time-of-use
5370 * race. The caller is responsible to use it correctly, for example:
5372 * - from a non-preemptible section (of course)
5374 * - from a thread that is bound to a single CPU
5376 * - in a loop with very short iterations (e.g. a polling loop)
5378 bool single_task_running(void)
5380 return raw_rq()->nr_running
== 1;
5382 EXPORT_SYMBOL(single_task_running
);
5384 unsigned long long nr_context_switches_cpu(int cpu
)
5386 return cpu_rq(cpu
)->nr_switches
;
5389 unsigned long long nr_context_switches(void)
5392 unsigned long long sum
= 0;
5394 for_each_possible_cpu(i
)
5395 sum
+= cpu_rq(i
)->nr_switches
;
5401 * Consumers of these two interfaces, like for example the cpuidle menu
5402 * governor, are using nonsensical data. Preferring shallow idle state selection
5403 * for a CPU that has IO-wait which might not even end up running the task when
5404 * it does become runnable.
5407 unsigned int nr_iowait_cpu(int cpu
)
5409 return atomic_read(&cpu_rq(cpu
)->nr_iowait
);
5413 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5415 * The idea behind IO-wait account is to account the idle time that we could
5416 * have spend running if it were not for IO. That is, if we were to improve the
5417 * storage performance, we'd have a proportional reduction in IO-wait time.
5419 * This all works nicely on UP, where, when a task blocks on IO, we account
5420 * idle time as IO-wait, because if the storage were faster, it could've been
5421 * running and we'd not be idle.
5423 * This has been extended to SMP, by doing the same for each CPU. This however
5426 * Imagine for instance the case where two tasks block on one CPU, only the one
5427 * CPU will have IO-wait accounted, while the other has regular idle. Even
5428 * though, if the storage were faster, both could've ran at the same time,
5429 * utilising both CPUs.
5431 * This means, that when looking globally, the current IO-wait accounting on
5432 * SMP is a lower bound, by reason of under accounting.
5434 * Worse, since the numbers are provided per CPU, they are sometimes
5435 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5436 * associated with any one particular CPU, it can wake to another CPU than it
5437 * blocked on. This means the per CPU IO-wait number is meaningless.
5439 * Task CPU affinities can make all that even more 'interesting'.
5442 unsigned int nr_iowait(void)
5444 unsigned int i
, sum
= 0;
5446 for_each_possible_cpu(i
)
5447 sum
+= nr_iowait_cpu(i
);
5455 * sched_exec - execve() is a valuable balancing opportunity, because at
5456 * this point the task has the smallest effective memory and cache footprint.
5458 void sched_exec(void)
5460 struct task_struct
*p
= current
;
5461 unsigned long flags
;
5464 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
5465 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), WF_EXEC
);
5466 if (dest_cpu
== smp_processor_id())
5469 if (likely(cpu_active(dest_cpu
))) {
5470 struct migration_arg arg
= { p
, dest_cpu
};
5472 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5473 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
5477 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5482 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
5483 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
5485 EXPORT_PER_CPU_SYMBOL(kstat
);
5486 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
5489 * The function fair_sched_class.update_curr accesses the struct curr
5490 * and its field curr->exec_start; when called from task_sched_runtime(),
5491 * we observe a high rate of cache misses in practice.
5492 * Prefetching this data results in improved performance.
5494 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
5496 #ifdef CONFIG_FAIR_GROUP_SCHED
5497 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
5499 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
5502 prefetch(&curr
->exec_start
);
5506 * Return accounted runtime for the task.
5507 * In case the task is currently running, return the runtime plus current's
5508 * pending runtime that have not been accounted yet.
5510 unsigned long long task_sched_runtime(struct task_struct
*p
)
5516 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5518 * 64-bit doesn't need locks to atomically read a 64-bit value.
5519 * So we have a optimization chance when the task's delta_exec is 0.
5520 * Reading ->on_cpu is racy, but this is ok.
5522 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5523 * If we race with it entering CPU, unaccounted time is 0. This is
5524 * indistinguishable from the read occurring a few cycles earlier.
5525 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5526 * been accounted, so we're correct here as well.
5528 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
5529 return p
->se
.sum_exec_runtime
;
5532 rq
= task_rq_lock(p
, &rf
);
5534 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5535 * project cycles that may never be accounted to this
5536 * thread, breaking clock_gettime().
5538 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
5539 prefetch_curr_exec_start(p
);
5540 update_rq_clock(rq
);
5541 p
->sched_class
->update_curr(rq
);
5543 ns
= p
->se
.sum_exec_runtime
;
5544 task_rq_unlock(rq
, p
, &rf
);
5549 #ifdef CONFIG_SCHED_DEBUG
5550 static u64
cpu_resched_latency(struct rq
*rq
)
5552 int latency_warn_ms
= READ_ONCE(sysctl_resched_latency_warn_ms
);
5553 u64 resched_latency
, now
= rq_clock(rq
);
5554 static bool warned_once
;
5556 if (sysctl_resched_latency_warn_once
&& warned_once
)
5559 if (!need_resched() || !latency_warn_ms
)
5562 if (system_state
== SYSTEM_BOOTING
)
5565 if (!rq
->last_seen_need_resched_ns
) {
5566 rq
->last_seen_need_resched_ns
= now
;
5567 rq
->ticks_without_resched
= 0;
5571 rq
->ticks_without_resched
++;
5572 resched_latency
= now
- rq
->last_seen_need_resched_ns
;
5573 if (resched_latency
<= latency_warn_ms
* NSEC_PER_MSEC
)
5578 return resched_latency
;
5581 static int __init
setup_resched_latency_warn_ms(char *str
)
5585 if ((kstrtol(str
, 0, &val
))) {
5586 pr_warn("Unable to set resched_latency_warn_ms\n");
5590 sysctl_resched_latency_warn_ms
= val
;
5593 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms
);
5595 static inline u64
cpu_resched_latency(struct rq
*rq
) { return 0; }
5596 #endif /* CONFIG_SCHED_DEBUG */
5599 * This function gets called by the timer code, with HZ frequency.
5600 * We call it with interrupts disabled.
5602 void scheduler_tick(void)
5604 int cpu
= smp_processor_id();
5605 struct rq
*rq
= cpu_rq(cpu
);
5606 struct task_struct
*curr
= rq
->curr
;
5608 unsigned long thermal_pressure
;
5609 u64 resched_latency
;
5611 if (housekeeping_cpu(cpu
, HK_TYPE_TICK
))
5612 arch_scale_freq_tick();
5618 update_rq_clock(rq
);
5619 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
5620 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
);
5621 curr
->sched_class
->task_tick(rq
, curr
, 0);
5622 if (sched_feat(LATENCY_WARN
))
5623 resched_latency
= cpu_resched_latency(rq
);
5624 calc_global_load_tick(rq
);
5625 sched_core_tick(rq
);
5626 task_tick_mm_cid(rq
, curr
);
5630 if (sched_feat(LATENCY_WARN
) && resched_latency
)
5631 resched_latency_warn(cpu
, resched_latency
);
5633 perf_event_task_tick();
5636 rq
->idle_balance
= idle_cpu(cpu
);
5637 trigger_load_balance(rq
);
5641 #ifdef CONFIG_NO_HZ_FULL
5646 struct delayed_work work
;
5648 /* Values for ->state, see diagram below. */
5649 #define TICK_SCHED_REMOTE_OFFLINE 0
5650 #define TICK_SCHED_REMOTE_OFFLINING 1
5651 #define TICK_SCHED_REMOTE_RUNNING 2
5654 * State diagram for ->state:
5657 * TICK_SCHED_REMOTE_OFFLINE
5660 * | | sched_tick_remote()
5663 * +--TICK_SCHED_REMOTE_OFFLINING
5666 * sched_tick_start() | | sched_tick_stop()
5669 * TICK_SCHED_REMOTE_RUNNING
5672 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5673 * and sched_tick_start() are happy to leave the state in RUNNING.
5676 static struct tick_work __percpu
*tick_work_cpu
;
5678 static void sched_tick_remote(struct work_struct
*work
)
5680 struct delayed_work
*dwork
= to_delayed_work(work
);
5681 struct tick_work
*twork
= container_of(dwork
, struct tick_work
, work
);
5682 int cpu
= twork
->cpu
;
5683 struct rq
*rq
= cpu_rq(cpu
);
5684 struct task_struct
*curr
;
5690 * Handle the tick only if it appears the remote CPU is running in full
5691 * dynticks mode. The check is racy by nature, but missing a tick or
5692 * having one too much is no big deal because the scheduler tick updates
5693 * statistics and checks timeslices in a time-independent way, regardless
5694 * of when exactly it is running.
5696 if (!tick_nohz_tick_stopped_cpu(cpu
))
5699 rq_lock_irq(rq
, &rf
);
5701 if (cpu_is_offline(cpu
))
5704 update_rq_clock(rq
);
5706 if (!is_idle_task(curr
)) {
5708 * Make sure the next tick runs within a reasonable
5711 delta
= rq_clock_task(rq
) - curr
->se
.exec_start
;
5712 WARN_ON_ONCE(delta
> (u64
)NSEC_PER_SEC
* 3);
5714 curr
->sched_class
->task_tick(rq
, curr
, 0);
5716 calc_load_nohz_remote(rq
);
5718 rq_unlock_irq(rq
, &rf
);
5722 * Run the remote tick once per second (1Hz). This arbitrary
5723 * frequency is large enough to avoid overload but short enough
5724 * to keep scheduler internal stats reasonably up to date. But
5725 * first update state to reflect hotplug activity if required.
5727 os
= atomic_fetch_add_unless(&twork
->state
, -1, TICK_SCHED_REMOTE_RUNNING
);
5728 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_OFFLINE
);
5729 if (os
== TICK_SCHED_REMOTE_RUNNING
)
5730 queue_delayed_work(system_unbound_wq
, dwork
, HZ
);
5733 static void sched_tick_start(int cpu
)
5736 struct tick_work
*twork
;
5738 if (housekeeping_cpu(cpu
, HK_TYPE_TICK
))
5741 WARN_ON_ONCE(!tick_work_cpu
);
5743 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
5744 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_RUNNING
);
5745 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_RUNNING
);
5746 if (os
== TICK_SCHED_REMOTE_OFFLINE
) {
5748 INIT_DELAYED_WORK(&twork
->work
, sched_tick_remote
);
5749 queue_delayed_work(system_unbound_wq
, &twork
->work
, HZ
);
5753 #ifdef CONFIG_HOTPLUG_CPU
5754 static void sched_tick_stop(int cpu
)
5756 struct tick_work
*twork
;
5759 if (housekeeping_cpu(cpu
, HK_TYPE_TICK
))
5762 WARN_ON_ONCE(!tick_work_cpu
);
5764 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
5765 /* There cannot be competing actions, but don't rely on stop-machine. */
5766 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_OFFLINING
);
5767 WARN_ON_ONCE(os
!= TICK_SCHED_REMOTE_RUNNING
);
5768 /* Don't cancel, as this would mess up the state machine. */
5770 #endif /* CONFIG_HOTPLUG_CPU */
5772 int __init
sched_tick_offload_init(void)
5774 tick_work_cpu
= alloc_percpu(struct tick_work
);
5775 BUG_ON(!tick_work_cpu
);
5779 #else /* !CONFIG_NO_HZ_FULL */
5780 static inline void sched_tick_start(int cpu
) { }
5781 static inline void sched_tick_stop(int cpu
) { }
5784 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5785 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5787 * If the value passed in is equal to the current preempt count
5788 * then we just disabled preemption. Start timing the latency.
5790 static inline void preempt_latency_start(int val
)
5792 if (preempt_count() == val
) {
5793 unsigned long ip
= get_lock_parent_ip();
5794 #ifdef CONFIG_DEBUG_PREEMPT
5795 current
->preempt_disable_ip
= ip
;
5797 trace_preempt_off(CALLER_ADDR0
, ip
);
5801 void preempt_count_add(int val
)
5803 #ifdef CONFIG_DEBUG_PREEMPT
5807 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5810 __preempt_count_add(val
);
5811 #ifdef CONFIG_DEBUG_PREEMPT
5813 * Spinlock count overflowing soon?
5815 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5818 preempt_latency_start(val
);
5820 EXPORT_SYMBOL(preempt_count_add
);
5821 NOKPROBE_SYMBOL(preempt_count_add
);
5824 * If the value passed in equals to the current preempt count
5825 * then we just enabled preemption. Stop timing the latency.
5827 static inline void preempt_latency_stop(int val
)
5829 if (preempt_count() == val
)
5830 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
5833 void preempt_count_sub(int val
)
5835 #ifdef CONFIG_DEBUG_PREEMPT
5839 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5842 * Is the spinlock portion underflowing?
5844 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5845 !(preempt_count() & PREEMPT_MASK
)))
5849 preempt_latency_stop(val
);
5850 __preempt_count_sub(val
);
5852 EXPORT_SYMBOL(preempt_count_sub
);
5853 NOKPROBE_SYMBOL(preempt_count_sub
);
5856 static inline void preempt_latency_start(int val
) { }
5857 static inline void preempt_latency_stop(int val
) { }
5860 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
5862 #ifdef CONFIG_DEBUG_PREEMPT
5863 return p
->preempt_disable_ip
;
5870 * Print scheduling while atomic bug:
5872 static noinline
void __schedule_bug(struct task_struct
*prev
)
5874 /* Save this before calling printk(), since that will clobber it */
5875 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
5877 if (oops_in_progress
)
5880 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5881 prev
->comm
, prev
->pid
, preempt_count());
5883 debug_show_held_locks(prev
);
5885 if (irqs_disabled())
5886 print_irqtrace_events(prev
);
5887 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
5888 && in_atomic_preempt_off()) {
5889 pr_err("Preemption disabled at:");
5890 print_ip_sym(KERN_ERR
, preempt_disable_ip
);
5892 check_panic_on_warn("scheduling while atomic");
5895 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
5899 * Various schedule()-time debugging checks and statistics:
5901 static inline void schedule_debug(struct task_struct
*prev
, bool preempt
)
5903 #ifdef CONFIG_SCHED_STACK_END_CHECK
5904 if (task_stack_end_corrupted(prev
))
5905 panic("corrupted stack end detected inside scheduler\n");
5907 if (task_scs_end_corrupted(prev
))
5908 panic("corrupted shadow stack detected inside scheduler\n");
5911 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5912 if (!preempt
&& READ_ONCE(prev
->__state
) && prev
->non_block_count
) {
5913 printk(KERN_ERR
"BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5914 prev
->comm
, prev
->pid
, prev
->non_block_count
);
5916 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
5920 if (unlikely(in_atomic_preempt_off())) {
5921 __schedule_bug(prev
);
5922 preempt_count_set(PREEMPT_DISABLED
);
5925 SCHED_WARN_ON(ct_state() == CONTEXT_USER
);
5927 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5929 schedstat_inc(this_rq()->sched_count
);
5932 static void put_prev_task_balance(struct rq
*rq
, struct task_struct
*prev
,
5933 struct rq_flags
*rf
)
5936 const struct sched_class
*class;
5938 * We must do the balancing pass before put_prev_task(), such
5939 * that when we release the rq->lock the task is in the same
5940 * state as before we took rq->lock.
5942 * We can terminate the balance pass as soon as we know there is
5943 * a runnable task of @class priority or higher.
5945 for_class_range(class, prev
->sched_class
, &idle_sched_class
) {
5946 if (class->balance(rq
, prev
, rf
))
5951 put_prev_task(rq
, prev
);
5955 * Pick up the highest-prio task:
5957 static inline struct task_struct
*
5958 __pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
5960 const struct sched_class
*class;
5961 struct task_struct
*p
;
5964 * Optimization: we know that if all tasks are in the fair class we can
5965 * call that function directly, but only if the @prev task wasn't of a
5966 * higher scheduling class, because otherwise those lose the
5967 * opportunity to pull in more work from other CPUs.
5969 if (likely(!sched_class_above(prev
->sched_class
, &fair_sched_class
) &&
5970 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
5972 p
= pick_next_task_fair(rq
, prev
, rf
);
5973 if (unlikely(p
== RETRY_TASK
))
5976 /* Assume the next prioritized class is idle_sched_class */
5978 put_prev_task(rq
, prev
);
5979 p
= pick_next_task_idle(rq
);
5986 put_prev_task_balance(rq
, prev
, rf
);
5988 for_each_class(class) {
5989 p
= class->pick_next_task(rq
);
5994 BUG(); /* The idle class should always have a runnable task. */
5997 #ifdef CONFIG_SCHED_CORE
5998 static inline bool is_task_rq_idle(struct task_struct
*t
)
6000 return (task_rq(t
)->idle
== t
);
6003 static inline bool cookie_equals(struct task_struct
*a
, unsigned long cookie
)
6005 return is_task_rq_idle(a
) || (a
->core_cookie
== cookie
);
6008 static inline bool cookie_match(struct task_struct
*a
, struct task_struct
*b
)
6010 if (is_task_rq_idle(a
) || is_task_rq_idle(b
))
6013 return a
->core_cookie
== b
->core_cookie
;
6016 static inline struct task_struct
*pick_task(struct rq
*rq
)
6018 const struct sched_class
*class;
6019 struct task_struct
*p
;
6021 for_each_class(class) {
6022 p
= class->pick_task(rq
);
6027 BUG(); /* The idle class should always have a runnable task. */
6030 extern void task_vruntime_update(struct rq
*rq
, struct task_struct
*p
, bool in_fi
);
6032 static void queue_core_balance(struct rq
*rq
);
6034 static struct task_struct
*
6035 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6037 struct task_struct
*next
, *p
, *max
= NULL
;
6038 const struct cpumask
*smt_mask
;
6039 bool fi_before
= false;
6040 bool core_clock_updated
= (rq
== rq
->core
);
6041 unsigned long cookie
;
6042 int i
, cpu
, occ
= 0;
6046 if (!sched_core_enabled(rq
))
6047 return __pick_next_task(rq
, prev
, rf
);
6051 /* Stopper task is switching into idle, no need core-wide selection. */
6052 if (cpu_is_offline(cpu
)) {
6054 * Reset core_pick so that we don't enter the fastpath when
6055 * coming online. core_pick would already be migrated to
6056 * another cpu during offline.
6058 rq
->core_pick
= NULL
;
6059 return __pick_next_task(rq
, prev
, rf
);
6063 * If there were no {en,de}queues since we picked (IOW, the task
6064 * pointers are all still valid), and we haven't scheduled the last
6065 * pick yet, do so now.
6067 * rq->core_pick can be NULL if no selection was made for a CPU because
6068 * it was either offline or went offline during a sibling's core-wide
6069 * selection. In this case, do a core-wide selection.
6071 if (rq
->core
->core_pick_seq
== rq
->core
->core_task_seq
&&
6072 rq
->core
->core_pick_seq
!= rq
->core_sched_seq
&&
6074 WRITE_ONCE(rq
->core_sched_seq
, rq
->core
->core_pick_seq
);
6076 next
= rq
->core_pick
;
6078 put_prev_task(rq
, prev
);
6079 set_next_task(rq
, next
);
6082 rq
->core_pick
= NULL
;
6086 put_prev_task_balance(rq
, prev
, rf
);
6088 smt_mask
= cpu_smt_mask(cpu
);
6089 need_sync
= !!rq
->core
->core_cookie
;
6092 rq
->core
->core_cookie
= 0UL;
6093 if (rq
->core
->core_forceidle_count
) {
6094 if (!core_clock_updated
) {
6095 update_rq_clock(rq
->core
);
6096 core_clock_updated
= true;
6098 sched_core_account_forceidle(rq
);
6099 /* reset after accounting force idle */
6100 rq
->core
->core_forceidle_start
= 0;
6101 rq
->core
->core_forceidle_count
= 0;
6102 rq
->core
->core_forceidle_occupation
= 0;
6108 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
6110 * @task_seq guards the task state ({en,de}queues)
6111 * @pick_seq is the @task_seq we did a selection on
6112 * @sched_seq is the @pick_seq we scheduled
6114 * However, preemptions can cause multiple picks on the same task set.
6115 * 'Fix' this by also increasing @task_seq for every pick.
6117 rq
->core
->core_task_seq
++;
6120 * Optimize for common case where this CPU has no cookies
6121 * and there are no cookied tasks running on siblings.
6124 next
= pick_task(rq
);
6125 if (!next
->core_cookie
) {
6126 rq
->core_pick
= NULL
;
6128 * For robustness, update the min_vruntime_fi for
6129 * unconstrained picks as well.
6131 WARN_ON_ONCE(fi_before
);
6132 task_vruntime_update(rq
, next
, false);
6138 * For each thread: do the regular task pick and find the max prio task
6141 * Tie-break prio towards the current CPU
6143 for_each_cpu_wrap(i
, smt_mask
, cpu
) {
6147 * Current cpu always has its clock updated on entrance to
6148 * pick_next_task(). If the current cpu is not the core,
6149 * the core may also have been updated above.
6151 if (i
!= cpu
&& (rq_i
!= rq
->core
|| !core_clock_updated
))
6152 update_rq_clock(rq_i
);
6154 p
= rq_i
->core_pick
= pick_task(rq_i
);
6155 if (!max
|| prio_less(max
, p
, fi_before
))
6159 cookie
= rq
->core
->core_cookie
= max
->core_cookie
;
6162 * For each thread: try and find a runnable task that matches @max or
6165 for_each_cpu(i
, smt_mask
) {
6167 p
= rq_i
->core_pick
;
6169 if (!cookie_equals(p
, cookie
)) {
6172 p
= sched_core_find(rq_i
, cookie
);
6174 p
= idle_sched_class
.pick_task(rq_i
);
6177 rq_i
->core_pick
= p
;
6179 if (p
== rq_i
->idle
) {
6180 if (rq_i
->nr_running
) {
6181 rq
->core
->core_forceidle_count
++;
6183 rq
->core
->core_forceidle_seq
++;
6190 if (schedstat_enabled() && rq
->core
->core_forceidle_count
) {
6191 rq
->core
->core_forceidle_start
= rq_clock(rq
->core
);
6192 rq
->core
->core_forceidle_occupation
= occ
;
6195 rq
->core
->core_pick_seq
= rq
->core
->core_task_seq
;
6196 next
= rq
->core_pick
;
6197 rq
->core_sched_seq
= rq
->core
->core_pick_seq
;
6199 /* Something should have been selected for current CPU */
6200 WARN_ON_ONCE(!next
);
6203 * Reschedule siblings
6205 * NOTE: L1TF -- at this point we're no longer running the old task and
6206 * sending an IPI (below) ensures the sibling will no longer be running
6207 * their task. This ensures there is no inter-sibling overlap between
6208 * non-matching user state.
6210 for_each_cpu(i
, smt_mask
) {
6214 * An online sibling might have gone offline before a task
6215 * could be picked for it, or it might be offline but later
6216 * happen to come online, but its too late and nothing was
6217 * picked for it. That's Ok - it will pick tasks for itself,
6220 if (!rq_i
->core_pick
)
6224 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6225 * fi_before fi update?
6231 if (!(fi_before
&& rq
->core
->core_forceidle_count
))
6232 task_vruntime_update(rq_i
, rq_i
->core_pick
, !!rq
->core
->core_forceidle_count
);
6234 rq_i
->core_pick
->core_occupation
= occ
;
6237 rq_i
->core_pick
= NULL
;
6241 /* Did we break L1TF mitigation requirements? */
6242 WARN_ON_ONCE(!cookie_match(next
, rq_i
->core_pick
));
6244 if (rq_i
->curr
== rq_i
->core_pick
) {
6245 rq_i
->core_pick
= NULL
;
6253 set_next_task(rq
, next
);
6255 if (rq
->core
->core_forceidle_count
&& next
== rq
->idle
)
6256 queue_core_balance(rq
);
6261 static bool try_steal_cookie(int this, int that
)
6263 struct rq
*dst
= cpu_rq(this), *src
= cpu_rq(that
);
6264 struct task_struct
*p
;
6265 unsigned long cookie
;
6266 bool success
= false;
6268 local_irq_disable();
6269 double_rq_lock(dst
, src
);
6271 cookie
= dst
->core
->core_cookie
;
6275 if (dst
->curr
!= dst
->idle
)
6278 p
= sched_core_find(src
, cookie
);
6283 if (p
== src
->core_pick
|| p
== src
->curr
)
6286 if (!is_cpu_allowed(p
, this))
6289 if (p
->core_occupation
> dst
->idle
->core_occupation
)
6292 * sched_core_find() and sched_core_next() will ensure that task @p
6293 * is not throttled now, we also need to check whether the runqueue
6294 * of the destination CPU is being throttled.
6296 if (sched_task_is_throttled(p
, this))
6299 deactivate_task(src
, p
, 0);
6300 set_task_cpu(p
, this);
6301 activate_task(dst
, p
, 0);
6309 p
= sched_core_next(p
, cookie
);
6313 double_rq_unlock(dst
, src
);
6319 static bool steal_cookie_task(int cpu
, struct sched_domain
*sd
)
6323 for_each_cpu_wrap(i
, sched_domain_span(sd
), cpu
+ 1) {
6330 if (try_steal_cookie(cpu
, i
))
6337 static void sched_core_balance(struct rq
*rq
)
6339 struct sched_domain
*sd
;
6340 int cpu
= cpu_of(rq
);
6344 raw_spin_rq_unlock_irq(rq
);
6345 for_each_domain(cpu
, sd
) {
6349 if (steal_cookie_task(cpu
, sd
))
6352 raw_spin_rq_lock_irq(rq
);
6357 static DEFINE_PER_CPU(struct balance_callback
, core_balance_head
);
6359 static void queue_core_balance(struct rq
*rq
)
6361 if (!sched_core_enabled(rq
))
6364 if (!rq
->core
->core_cookie
)
6367 if (!rq
->nr_running
) /* not forced idle */
6370 queue_balance_callback(rq
, &per_cpu(core_balance_head
, rq
->cpu
), sched_core_balance
);
6373 static void sched_core_cpu_starting(unsigned int cpu
)
6375 const struct cpumask
*smt_mask
= cpu_smt_mask(cpu
);
6376 struct rq
*rq
= cpu_rq(cpu
), *core_rq
= NULL
;
6377 unsigned long flags
;
6380 sched_core_lock(cpu
, &flags
);
6382 WARN_ON_ONCE(rq
->core
!= rq
);
6384 /* if we're the first, we'll be our own leader */
6385 if (cpumask_weight(smt_mask
) == 1)
6388 /* find the leader */
6389 for_each_cpu(t
, smt_mask
) {
6393 if (rq
->core
== rq
) {
6399 if (WARN_ON_ONCE(!core_rq
)) /* whoopsie */
6402 /* install and validate core_rq */
6403 for_each_cpu(t
, smt_mask
) {
6409 WARN_ON_ONCE(rq
->core
!= core_rq
);
6413 sched_core_unlock(cpu
, &flags
);
6416 static void sched_core_cpu_deactivate(unsigned int cpu
)
6418 const struct cpumask
*smt_mask
= cpu_smt_mask(cpu
);
6419 struct rq
*rq
= cpu_rq(cpu
), *core_rq
= NULL
;
6420 unsigned long flags
;
6423 sched_core_lock(cpu
, &flags
);
6425 /* if we're the last man standing, nothing to do */
6426 if (cpumask_weight(smt_mask
) == 1) {
6427 WARN_ON_ONCE(rq
->core
!= rq
);
6431 /* if we're not the leader, nothing to do */
6435 /* find a new leader */
6436 for_each_cpu(t
, smt_mask
) {
6439 core_rq
= cpu_rq(t
);
6443 if (WARN_ON_ONCE(!core_rq
)) /* impossible */
6446 /* copy the shared state to the new leader */
6447 core_rq
->core_task_seq
= rq
->core_task_seq
;
6448 core_rq
->core_pick_seq
= rq
->core_pick_seq
;
6449 core_rq
->core_cookie
= rq
->core_cookie
;
6450 core_rq
->core_forceidle_count
= rq
->core_forceidle_count
;
6451 core_rq
->core_forceidle_seq
= rq
->core_forceidle_seq
;
6452 core_rq
->core_forceidle_occupation
= rq
->core_forceidle_occupation
;
6455 * Accounting edge for forced idle is handled in pick_next_task().
6456 * Don't need another one here, since the hotplug thread shouldn't
6459 core_rq
->core_forceidle_start
= 0;
6461 /* install new leader */
6462 for_each_cpu(t
, smt_mask
) {
6468 sched_core_unlock(cpu
, &flags
);
6471 static inline void sched_core_cpu_dying(unsigned int cpu
)
6473 struct rq
*rq
= cpu_rq(cpu
);
6479 #else /* !CONFIG_SCHED_CORE */
6481 static inline void sched_core_cpu_starting(unsigned int cpu
) {}
6482 static inline void sched_core_cpu_deactivate(unsigned int cpu
) {}
6483 static inline void sched_core_cpu_dying(unsigned int cpu
) {}
6485 static struct task_struct
*
6486 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
6488 return __pick_next_task(rq
, prev
, rf
);
6491 #endif /* CONFIG_SCHED_CORE */
6494 * Constants for the sched_mode argument of __schedule().
6496 * The mode argument allows RT enabled kernels to differentiate a
6497 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6498 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6499 * optimize the AND operation out and just check for zero.
6502 #define SM_PREEMPT 0x1
6503 #define SM_RTLOCK_WAIT 0x2
6505 #ifndef CONFIG_PREEMPT_RT
6506 # define SM_MASK_PREEMPT (~0U)
6508 # define SM_MASK_PREEMPT SM_PREEMPT
6512 * __schedule() is the main scheduler function.
6514 * The main means of driving the scheduler and thus entering this function are:
6516 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6518 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6519 * paths. For example, see arch/x86/entry_64.S.
6521 * To drive preemption between tasks, the scheduler sets the flag in timer
6522 * interrupt handler scheduler_tick().
6524 * 3. Wakeups don't really cause entry into schedule(). They add a
6525 * task to the run-queue and that's it.
6527 * Now, if the new task added to the run-queue preempts the current
6528 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6529 * called on the nearest possible occasion:
6531 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6533 * - in syscall or exception context, at the next outmost
6534 * preempt_enable(). (this might be as soon as the wake_up()'s
6537 * - in IRQ context, return from interrupt-handler to
6538 * preemptible context
6540 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6543 * - cond_resched() call
6544 * - explicit schedule() call
6545 * - return from syscall or exception to user-space
6546 * - return from interrupt-handler to user-space
6548 * WARNING: must be called with preemption disabled!
6550 static void __sched notrace
__schedule(unsigned int sched_mode
)
6552 struct task_struct
*prev
, *next
;
6553 unsigned long *switch_count
;
6554 unsigned long prev_state
;
6559 cpu
= smp_processor_id();
6563 schedule_debug(prev
, !!sched_mode
);
6565 if (sched_feat(HRTICK
) || sched_feat(HRTICK_DL
))
6568 local_irq_disable();
6569 rcu_note_context_switch(!!sched_mode
);
6572 * Make sure that signal_pending_state()->signal_pending() below
6573 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6574 * done by the caller to avoid the race with signal_wake_up():
6576 * __set_current_state(@state) signal_wake_up()
6577 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6578 * wake_up_state(p, state)
6579 * LOCK rq->lock LOCK p->pi_state
6580 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6581 * if (signal_pending_state()) if (p->state & @state)
6583 * Also, the membarrier system call requires a full memory barrier
6584 * after coming from user-space, before storing to rq->curr.
6587 smp_mb__after_spinlock();
6589 /* Promote REQ to ACT */
6590 rq
->clock_update_flags
<<= 1;
6591 update_rq_clock(rq
);
6593 switch_count
= &prev
->nivcsw
;
6596 * We must load prev->state once (task_struct::state is volatile), such
6597 * that we form a control dependency vs deactivate_task() below.
6599 prev_state
= READ_ONCE(prev
->__state
);
6600 if (!(sched_mode
& SM_MASK_PREEMPT
) && prev_state
) {
6601 if (signal_pending_state(prev_state
, prev
)) {
6602 WRITE_ONCE(prev
->__state
, TASK_RUNNING
);
6604 prev
->sched_contributes_to_load
=
6605 (prev_state
& TASK_UNINTERRUPTIBLE
) &&
6606 !(prev_state
& TASK_NOLOAD
) &&
6607 !(prev_state
& TASK_FROZEN
);
6609 if (prev
->sched_contributes_to_load
)
6610 rq
->nr_uninterruptible
++;
6613 * __schedule() ttwu()
6614 * prev_state = prev->state; if (p->on_rq && ...)
6615 * if (prev_state) goto out;
6616 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6617 * p->state = TASK_WAKING
6619 * Where __schedule() and ttwu() have matching control dependencies.
6621 * After this, schedule() must not care about p->state any more.
6623 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
6625 if (prev
->in_iowait
) {
6626 atomic_inc(&rq
->nr_iowait
);
6627 delayacct_blkio_start();
6630 switch_count
= &prev
->nvcsw
;
6633 next
= pick_next_task(rq
, prev
, &rf
);
6634 clear_tsk_need_resched(prev
);
6635 clear_preempt_need_resched();
6636 #ifdef CONFIG_SCHED_DEBUG
6637 rq
->last_seen_need_resched_ns
= 0;
6640 if (likely(prev
!= next
)) {
6643 * RCU users of rcu_dereference(rq->curr) may not see
6644 * changes to task_struct made by pick_next_task().
6646 RCU_INIT_POINTER(rq
->curr
, next
);
6648 * The membarrier system call requires each architecture
6649 * to have a full memory barrier after updating
6650 * rq->curr, before returning to user-space.
6652 * Here are the schemes providing that barrier on the
6653 * various architectures:
6654 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6655 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6656 * - finish_lock_switch() for weakly-ordered
6657 * architectures where spin_unlock is a full barrier,
6658 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6659 * is a RELEASE barrier),
6663 migrate_disable_switch(rq
, prev
);
6664 psi_sched_switch(prev
, next
, !task_on_rq_queued(prev
));
6666 trace_sched_switch(sched_mode
& SM_MASK_PREEMPT
, prev
, next
, prev_state
);
6668 /* Also unlocks the rq: */
6669 rq
= context_switch(rq
, prev
, next
, &rf
);
6671 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
6673 rq_unpin_lock(rq
, &rf
);
6674 __balance_callbacks(rq
);
6675 raw_spin_rq_unlock_irq(rq
);
6679 void __noreturn
do_task_dead(void)
6681 /* Causes final put_task_struct in finish_task_switch(): */
6682 set_special_state(TASK_DEAD
);
6684 /* Tell freezer to ignore us: */
6685 current
->flags
|= PF_NOFREEZE
;
6687 __schedule(SM_NONE
);
6690 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6695 static inline void sched_submit_work(struct task_struct
*tsk
)
6697 unsigned int task_flags
;
6699 if (task_is_running(tsk
))
6702 task_flags
= tsk
->flags
;
6704 * If a worker goes to sleep, notify and ask workqueue whether it
6705 * wants to wake up a task to maintain concurrency.
6707 if (task_flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
6708 if (task_flags
& PF_WQ_WORKER
)
6709 wq_worker_sleeping(tsk
);
6711 io_wq_worker_sleeping(tsk
);
6715 * spinlock and rwlock must not flush block requests. This will
6716 * deadlock if the callback attempts to acquire a lock which is
6719 SCHED_WARN_ON(current
->__state
& TASK_RTLOCK_WAIT
);
6722 * If we are going to sleep and we have plugged IO queued,
6723 * make sure to submit it to avoid deadlocks.
6725 blk_flush_plug(tsk
->plug
, true);
6728 static void sched_update_worker(struct task_struct
*tsk
)
6730 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
6731 if (tsk
->flags
& PF_WQ_WORKER
)
6732 wq_worker_running(tsk
);
6734 io_wq_worker_running(tsk
);
6738 asmlinkage __visible
void __sched
schedule(void)
6740 struct task_struct
*tsk
= current
;
6742 sched_submit_work(tsk
);
6745 __schedule(SM_NONE
);
6746 sched_preempt_enable_no_resched();
6747 } while (need_resched());
6748 sched_update_worker(tsk
);
6750 EXPORT_SYMBOL(schedule
);
6753 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6754 * state (have scheduled out non-voluntarily) by making sure that all
6755 * tasks have either left the run queue or have gone into user space.
6756 * As idle tasks do not do either, they must not ever be preempted
6757 * (schedule out non-voluntarily).
6759 * schedule_idle() is similar to schedule_preempt_disable() except that it
6760 * never enables preemption because it does not call sched_submit_work().
6762 void __sched
schedule_idle(void)
6765 * As this skips calling sched_submit_work(), which the idle task does
6766 * regardless because that function is a nop when the task is in a
6767 * TASK_RUNNING state, make sure this isn't used someplace that the
6768 * current task can be in any other state. Note, idle is always in the
6769 * TASK_RUNNING state.
6771 WARN_ON_ONCE(current
->__state
);
6773 __schedule(SM_NONE
);
6774 } while (need_resched());
6777 #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
6778 asmlinkage __visible
void __sched
schedule_user(void)
6781 * If we come here after a random call to set_need_resched(),
6782 * or we have been woken up remotely but the IPI has not yet arrived,
6783 * we haven't yet exited the RCU idle mode. Do it here manually until
6784 * we find a better solution.
6786 * NB: There are buggy callers of this function. Ideally we
6787 * should warn if prev_state != CONTEXT_USER, but that will trigger
6788 * too frequently to make sense yet.
6790 enum ctx_state prev_state
= exception_enter();
6792 exception_exit(prev_state
);
6797 * schedule_preempt_disabled - called with preemption disabled
6799 * Returns with preemption disabled. Note: preempt_count must be 1
6801 void __sched
schedule_preempt_disabled(void)
6803 sched_preempt_enable_no_resched();
6808 #ifdef CONFIG_PREEMPT_RT
6809 void __sched notrace
schedule_rtlock(void)
6813 __schedule(SM_RTLOCK_WAIT
);
6814 sched_preempt_enable_no_resched();
6815 } while (need_resched());
6817 NOKPROBE_SYMBOL(schedule_rtlock
);
6820 static void __sched notrace
preempt_schedule_common(void)
6824 * Because the function tracer can trace preempt_count_sub()
6825 * and it also uses preempt_enable/disable_notrace(), if
6826 * NEED_RESCHED is set, the preempt_enable_notrace() called
6827 * by the function tracer will call this function again and
6828 * cause infinite recursion.
6830 * Preemption must be disabled here before the function
6831 * tracer can trace. Break up preempt_disable() into two
6832 * calls. One to disable preemption without fear of being
6833 * traced. The other to still record the preemption latency,
6834 * which can also be traced by the function tracer.
6836 preempt_disable_notrace();
6837 preempt_latency_start(1);
6838 __schedule(SM_PREEMPT
);
6839 preempt_latency_stop(1);
6840 preempt_enable_no_resched_notrace();
6843 * Check again in case we missed a preemption opportunity
6844 * between schedule and now.
6846 } while (need_resched());
6849 #ifdef CONFIG_PREEMPTION
6851 * This is the entry point to schedule() from in-kernel preemption
6852 * off of preempt_enable.
6854 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
6857 * If there is a non-zero preempt_count or interrupts are disabled,
6858 * we do not want to preempt the current task. Just return..
6860 if (likely(!preemptible()))
6862 preempt_schedule_common();
6864 NOKPROBE_SYMBOL(preempt_schedule
);
6865 EXPORT_SYMBOL(preempt_schedule
);
6867 #ifdef CONFIG_PREEMPT_DYNAMIC
6868 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6869 #ifndef preempt_schedule_dynamic_enabled
6870 #define preempt_schedule_dynamic_enabled preempt_schedule
6871 #define preempt_schedule_dynamic_disabled NULL
6873 DEFINE_STATIC_CALL(preempt_schedule
, preempt_schedule_dynamic_enabled
);
6874 EXPORT_STATIC_CALL_TRAMP(preempt_schedule
);
6875 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6876 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule
);
6877 void __sched notrace
dynamic_preempt_schedule(void)
6879 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule
))
6883 NOKPROBE_SYMBOL(dynamic_preempt_schedule
);
6884 EXPORT_SYMBOL(dynamic_preempt_schedule
);
6889 * preempt_schedule_notrace - preempt_schedule called by tracing
6891 * The tracing infrastructure uses preempt_enable_notrace to prevent
6892 * recursion and tracing preempt enabling caused by the tracing
6893 * infrastructure itself. But as tracing can happen in areas coming
6894 * from userspace or just about to enter userspace, a preempt enable
6895 * can occur before user_exit() is called. This will cause the scheduler
6896 * to be called when the system is still in usermode.
6898 * To prevent this, the preempt_enable_notrace will use this function
6899 * instead of preempt_schedule() to exit user context if needed before
6900 * calling the scheduler.
6902 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
6904 enum ctx_state prev_ctx
;
6906 if (likely(!preemptible()))
6911 * Because the function tracer can trace preempt_count_sub()
6912 * and it also uses preempt_enable/disable_notrace(), if
6913 * NEED_RESCHED is set, the preempt_enable_notrace() called
6914 * by the function tracer will call this function again and
6915 * cause infinite recursion.
6917 * Preemption must be disabled here before the function
6918 * tracer can trace. Break up preempt_disable() into two
6919 * calls. One to disable preemption without fear of being
6920 * traced. The other to still record the preemption latency,
6921 * which can also be traced by the function tracer.
6923 preempt_disable_notrace();
6924 preempt_latency_start(1);
6926 * Needs preempt disabled in case user_exit() is traced
6927 * and the tracer calls preempt_enable_notrace() causing
6928 * an infinite recursion.
6930 prev_ctx
= exception_enter();
6931 __schedule(SM_PREEMPT
);
6932 exception_exit(prev_ctx
);
6934 preempt_latency_stop(1);
6935 preempt_enable_no_resched_notrace();
6936 } while (need_resched());
6938 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
6940 #ifdef CONFIG_PREEMPT_DYNAMIC
6941 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
6942 #ifndef preempt_schedule_notrace_dynamic_enabled
6943 #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6944 #define preempt_schedule_notrace_dynamic_disabled NULL
6946 DEFINE_STATIC_CALL(preempt_schedule_notrace
, preempt_schedule_notrace_dynamic_enabled
);
6947 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace
);
6948 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6949 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace
);
6950 void __sched notrace
dynamic_preempt_schedule_notrace(void)
6952 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace
))
6954 preempt_schedule_notrace();
6956 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace
);
6957 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace
);
6961 #endif /* CONFIG_PREEMPTION */
6964 * This is the entry point to schedule() from kernel preemption
6965 * off of irq context.
6966 * Note, that this is called and return with irqs disabled. This will
6967 * protect us against recursive calling from irq.
6969 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
6971 enum ctx_state prev_state
;
6973 /* Catch callers which need to be fixed */
6974 BUG_ON(preempt_count() || !irqs_disabled());
6976 prev_state
= exception_enter();
6981 __schedule(SM_PREEMPT
);
6982 local_irq_disable();
6983 sched_preempt_enable_no_resched();
6984 } while (need_resched());
6986 exception_exit(prev_state
);
6989 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
6992 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG
) && wake_flags
& ~WF_SYNC
);
6993 return try_to_wake_up(curr
->private, mode
, wake_flags
);
6995 EXPORT_SYMBOL(default_wake_function
);
6997 static void __setscheduler_prio(struct task_struct
*p
, int prio
)
7000 p
->sched_class
= &dl_sched_class
;
7001 else if (rt_prio(prio
))
7002 p
->sched_class
= &rt_sched_class
;
7004 p
->sched_class
= &fair_sched_class
;
7009 #ifdef CONFIG_RT_MUTEXES
7011 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
7014 prio
= min(prio
, pi_task
->prio
);
7019 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
7021 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
7023 return __rt_effective_prio(pi_task
, prio
);
7027 * rt_mutex_setprio - set the current priority of a task
7029 * @pi_task: donor task
7031 * This function changes the 'effective' priority of a task. It does
7032 * not touch ->normal_prio like __setscheduler().
7034 * Used by the rt_mutex code to implement priority inheritance
7035 * logic. Call site only calls if the priority of the task changed.
7037 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
7039 int prio
, oldprio
, queued
, running
, queue_flag
=
7040 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
7041 const struct sched_class
*prev_class
;
7045 /* XXX used to be waiter->prio, not waiter->task->prio */
7046 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
7049 * If nothing changed; bail early.
7051 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
7054 rq
= __task_rq_lock(p
, &rf
);
7055 update_rq_clock(rq
);
7057 * Set under pi_lock && rq->lock, such that the value can be used under
7060 * Note that there is loads of tricky to make this pointer cache work
7061 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
7062 * ensure a task is de-boosted (pi_task is set to NULL) before the
7063 * task is allowed to run again (and can exit). This ensures the pointer
7064 * points to a blocked task -- which guarantees the task is present.
7066 p
->pi_top_task
= pi_task
;
7069 * For FIFO/RR we only need to set prio, if that matches we're done.
7071 if (prio
== p
->prio
&& !dl_prio(prio
))
7075 * Idle task boosting is a nono in general. There is one
7076 * exception, when PREEMPT_RT and NOHZ is active:
7078 * The idle task calls get_next_timer_interrupt() and holds
7079 * the timer wheel base->lock on the CPU and another CPU wants
7080 * to access the timer (probably to cancel it). We can safely
7081 * ignore the boosting request, as the idle CPU runs this code
7082 * with interrupts disabled and will complete the lock
7083 * protected section without being interrupted. So there is no
7084 * real need to boost.
7086 if (unlikely(p
== rq
->idle
)) {
7087 WARN_ON(p
!= rq
->curr
);
7088 WARN_ON(p
->pi_blocked_on
);
7092 trace_sched_pi_setprio(p
, pi_task
);
7095 if (oldprio
== prio
)
7096 queue_flag
&= ~DEQUEUE_MOVE
;
7098 prev_class
= p
->sched_class
;
7099 queued
= task_on_rq_queued(p
);
7100 running
= task_current(rq
, p
);
7102 dequeue_task(rq
, p
, queue_flag
);
7104 put_prev_task(rq
, p
);
7107 * Boosting condition are:
7108 * 1. -rt task is running and holds mutex A
7109 * --> -dl task blocks on mutex A
7111 * 2. -dl task is running and holds mutex A
7112 * --> -dl task blocks on mutex A and could preempt the
7115 if (dl_prio(prio
)) {
7116 if (!dl_prio(p
->normal_prio
) ||
7117 (pi_task
&& dl_prio(pi_task
->prio
) &&
7118 dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
7119 p
->dl
.pi_se
= pi_task
->dl
.pi_se
;
7120 queue_flag
|= ENQUEUE_REPLENISH
;
7122 p
->dl
.pi_se
= &p
->dl
;
7124 } else if (rt_prio(prio
)) {
7125 if (dl_prio(oldprio
))
7126 p
->dl
.pi_se
= &p
->dl
;
7128 queue_flag
|= ENQUEUE_HEAD
;
7130 if (dl_prio(oldprio
))
7131 p
->dl
.pi_se
= &p
->dl
;
7132 if (rt_prio(oldprio
))
7136 __setscheduler_prio(p
, prio
);
7139 enqueue_task(rq
, p
, queue_flag
);
7141 set_next_task(rq
, p
);
7143 check_class_changed(rq
, p
, prev_class
, oldprio
);
7145 /* Avoid rq from going away on us: */
7148 rq_unpin_lock(rq
, &rf
);
7149 __balance_callbacks(rq
);
7150 raw_spin_rq_unlock(rq
);
7155 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
7161 void set_user_nice(struct task_struct
*p
, long nice
)
7163 bool queued
, running
;
7168 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
7171 * We have to be careful, if called from sys_setpriority(),
7172 * the task might be in the middle of scheduling on another CPU.
7174 rq
= task_rq_lock(p
, &rf
);
7175 update_rq_clock(rq
);
7178 * The RT priorities are set via sched_setscheduler(), but we still
7179 * allow the 'normal' nice value to be set - but as expected
7180 * it won't have any effect on scheduling until the task is
7181 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7183 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
7184 p
->static_prio
= NICE_TO_PRIO(nice
);
7187 queued
= task_on_rq_queued(p
);
7188 running
= task_current(rq
, p
);
7190 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
7192 put_prev_task(rq
, p
);
7194 p
->static_prio
= NICE_TO_PRIO(nice
);
7195 set_load_weight(p
, true);
7197 p
->prio
= effective_prio(p
);
7200 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
7202 set_next_task(rq
, p
);
7205 * If the task increased its priority or is running and
7206 * lowered its priority, then reschedule its CPU:
7208 p
->sched_class
->prio_changed(rq
, p
, old_prio
);
7211 task_rq_unlock(rq
, p
, &rf
);
7213 EXPORT_SYMBOL(set_user_nice
);
7216 * is_nice_reduction - check if nice value is an actual reduction
7218 * Similar to can_nice() but does not perform a capability check.
7223 static bool is_nice_reduction(const struct task_struct
*p
, const int nice
)
7225 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7226 int nice_rlim
= nice_to_rlimit(nice
);
7228 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
));
7232 * can_nice - check if a task can reduce its nice value
7236 int can_nice(const struct task_struct
*p
, const int nice
)
7238 return is_nice_reduction(p
, nice
) || capable(CAP_SYS_NICE
);
7241 #ifdef __ARCH_WANT_SYS_NICE
7244 * sys_nice - change the priority of the current process.
7245 * @increment: priority increment
7247 * sys_setpriority is a more generic, but much slower function that
7248 * does similar things.
7250 SYSCALL_DEFINE1(nice
, int, increment
)
7255 * Setpriority might change our priority at the same moment.
7256 * We don't have to worry. Conceptually one call occurs first
7257 * and we have a single winner.
7259 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
7260 nice
= task_nice(current
) + increment
;
7262 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
7263 if (increment
< 0 && !can_nice(current
, nice
))
7266 retval
= security_task_setnice(current
, nice
);
7270 set_user_nice(current
, nice
);
7277 * task_prio - return the priority value of a given task.
7278 * @p: the task in question.
7280 * Return: The priority value as seen by users in /proc.
7282 * sched policy return value kernel prio user prio/nice
7284 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7285 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7286 * deadline -101 -1 0
7288 int task_prio(const struct task_struct
*p
)
7290 return p
->prio
- MAX_RT_PRIO
;
7294 * idle_cpu - is a given CPU idle currently?
7295 * @cpu: the processor in question.
7297 * Return: 1 if the CPU is currently idle. 0 otherwise.
7299 int idle_cpu(int cpu
)
7301 struct rq
*rq
= cpu_rq(cpu
);
7303 if (rq
->curr
!= rq
->idle
)
7310 if (rq
->ttwu_pending
)
7318 * available_idle_cpu - is a given CPU idle for enqueuing work.
7319 * @cpu: the CPU in question.
7321 * Return: 1 if the CPU is currently idle. 0 otherwise.
7323 int available_idle_cpu(int cpu
)
7328 if (vcpu_is_preempted(cpu
))
7335 * idle_task - return the idle task for a given CPU.
7336 * @cpu: the processor in question.
7338 * Return: The idle task for the CPU @cpu.
7340 struct task_struct
*idle_task(int cpu
)
7342 return cpu_rq(cpu
)->idle
;
7347 * This function computes an effective utilization for the given CPU, to be
7348 * used for frequency selection given the linear relation: f = u * f_max.
7350 * The scheduler tracks the following metrics:
7352 * cpu_util_{cfs,rt,dl,irq}()
7355 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7356 * synchronized windows and are thus directly comparable.
7358 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7359 * which excludes things like IRQ and steal-time. These latter are then accrued
7360 * in the irq utilization.
7362 * The DL bandwidth number otoh is not a measured metric but a value computed
7363 * based on the task model parameters and gives the minimal utilization
7364 * required to meet deadlines.
7366 unsigned long effective_cpu_util(int cpu
, unsigned long util_cfs
,
7367 enum cpu_util_type type
,
7368 struct task_struct
*p
)
7370 unsigned long dl_util
, util
, irq
, max
;
7371 struct rq
*rq
= cpu_rq(cpu
);
7373 max
= arch_scale_cpu_capacity(cpu
);
7375 if (!uclamp_is_used() &&
7376 type
== FREQUENCY_UTIL
&& rt_rq_is_runnable(&rq
->rt
)) {
7381 * Early check to see if IRQ/steal time saturates the CPU, can be
7382 * because of inaccuracies in how we track these -- see
7383 * update_irq_load_avg().
7385 irq
= cpu_util_irq(rq
);
7386 if (unlikely(irq
>= max
))
7390 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7391 * CFS tasks and we use the same metric to track the effective
7392 * utilization (PELT windows are synchronized) we can directly add them
7393 * to obtain the CPU's actual utilization.
7395 * CFS and RT utilization can be boosted or capped, depending on
7396 * utilization clamp constraints requested by currently RUNNABLE
7398 * When there are no CFS RUNNABLE tasks, clamps are released and
7399 * frequency will be gracefully reduced with the utilization decay.
7401 util
= util_cfs
+ cpu_util_rt(rq
);
7402 if (type
== FREQUENCY_UTIL
)
7403 util
= uclamp_rq_util_with(rq
, util
, p
);
7405 dl_util
= cpu_util_dl(rq
);
7408 * For frequency selection we do not make cpu_util_dl() a permanent part
7409 * of this sum because we want to use cpu_bw_dl() later on, but we need
7410 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7411 * that we select f_max when there is no idle time.
7413 * NOTE: numerical errors or stop class might cause us to not quite hit
7414 * saturation when we should -- something for later.
7416 if (util
+ dl_util
>= max
)
7420 * OTOH, for energy computation we need the estimated running time, so
7421 * include util_dl and ignore dl_bw.
7423 if (type
== ENERGY_UTIL
)
7427 * There is still idle time; further improve the number by using the
7428 * irq metric. Because IRQ/steal time is hidden from the task clock we
7429 * need to scale the task numbers:
7432 * U' = irq + --------- * U
7435 util
= scale_irq_capacity(util
, irq
, max
);
7439 * Bandwidth required by DEADLINE must always be granted while, for
7440 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7441 * to gracefully reduce the frequency when no tasks show up for longer
7444 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7445 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7446 * an interface. So, we only do the latter for now.
7448 if (type
== FREQUENCY_UTIL
)
7449 util
+= cpu_bw_dl(rq
);
7451 return min(max
, util
);
7454 unsigned long sched_cpu_util(int cpu
)
7456 return effective_cpu_util(cpu
, cpu_util_cfs(cpu
), ENERGY_UTIL
, NULL
);
7458 #endif /* CONFIG_SMP */
7461 * find_process_by_pid - find a process with a matching PID value.
7462 * @pid: the pid in question.
7464 * The task of @pid, if found. %NULL otherwise.
7466 static struct task_struct
*find_process_by_pid(pid_t pid
)
7468 return pid
? find_task_by_vpid(pid
) : current
;
7472 * sched_setparam() passes in -1 for its policy, to let the functions
7473 * it calls know not to change it.
7475 #define SETPARAM_POLICY -1
7477 static void __setscheduler_params(struct task_struct
*p
,
7478 const struct sched_attr
*attr
)
7480 int policy
= attr
->sched_policy
;
7482 if (policy
== SETPARAM_POLICY
)
7487 if (dl_policy(policy
))
7488 __setparam_dl(p
, attr
);
7489 else if (fair_policy(policy
))
7490 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
7493 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7494 * !rt_policy. Always setting this ensures that things like
7495 * getparam()/getattr() don't report silly values for !rt tasks.
7497 p
->rt_priority
= attr
->sched_priority
;
7498 p
->normal_prio
= normal_prio(p
);
7499 set_load_weight(p
, true);
7503 * Check the target process has a UID that matches the current process's:
7505 static bool check_same_owner(struct task_struct
*p
)
7507 const struct cred
*cred
= current_cred(), *pcred
;
7511 pcred
= __task_cred(p
);
7512 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
7513 uid_eq(cred
->euid
, pcred
->uid
));
7519 * Allow unprivileged RT tasks to decrease priority.
7520 * Only issue a capable test if needed and only once to avoid an audit
7521 * event on permitted non-privileged operations:
7523 static int user_check_sched_setscheduler(struct task_struct
*p
,
7524 const struct sched_attr
*attr
,
7525 int policy
, int reset_on_fork
)
7527 if (fair_policy(policy
)) {
7528 if (attr
->sched_nice
< task_nice(p
) &&
7529 !is_nice_reduction(p
, attr
->sched_nice
))
7533 if (rt_policy(policy
)) {
7534 unsigned long rlim_rtprio
= task_rlimit(p
, RLIMIT_RTPRIO
);
7536 /* Can't set/change the rt policy: */
7537 if (policy
!= p
->policy
&& !rlim_rtprio
)
7540 /* Can't increase priority: */
7541 if (attr
->sched_priority
> p
->rt_priority
&&
7542 attr
->sched_priority
> rlim_rtprio
)
7547 * Can't set/change SCHED_DEADLINE policy at all for now
7548 * (safest behavior); in the future we would like to allow
7549 * unprivileged DL tasks to increase their relative deadline
7550 * or reduce their runtime (both ways reducing utilization)
7552 if (dl_policy(policy
))
7556 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7557 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7559 if (task_has_idle_policy(p
) && !idle_policy(policy
)) {
7560 if (!is_nice_reduction(p
, task_nice(p
)))
7564 /* Can't change other user's priorities: */
7565 if (!check_same_owner(p
))
7568 /* Normal users shall not reset the sched_reset_on_fork flag: */
7569 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
7575 if (!capable(CAP_SYS_NICE
))
7581 static int __sched_setscheduler(struct task_struct
*p
,
7582 const struct sched_attr
*attr
,
7585 int oldpolicy
= -1, policy
= attr
->sched_policy
;
7586 int retval
, oldprio
, newprio
, queued
, running
;
7587 const struct sched_class
*prev_class
;
7588 struct balance_callback
*head
;
7591 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
7593 bool cpuset_locked
= false;
7595 /* The pi code expects interrupts enabled */
7596 BUG_ON(pi
&& in_interrupt());
7598 /* Double check policy once rq lock held: */
7600 reset_on_fork
= p
->sched_reset_on_fork
;
7601 policy
= oldpolicy
= p
->policy
;
7603 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
7605 if (!valid_policy(policy
))
7609 if (attr
->sched_flags
& ~(SCHED_FLAG_ALL
| SCHED_FLAG_SUGOV
))
7613 * Valid priorities for SCHED_FIFO and SCHED_RR are
7614 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7615 * SCHED_BATCH and SCHED_IDLE is 0.
7617 if (attr
->sched_priority
> MAX_RT_PRIO
-1)
7619 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
7620 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
7624 retval
= user_check_sched_setscheduler(p
, attr
, policy
, reset_on_fork
);
7628 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
7631 retval
= security_task_setscheduler(p
);
7636 /* Update task specific "requested" clamps */
7637 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) {
7638 retval
= uclamp_validate(p
, attr
);
7644 * SCHED_DEADLINE bandwidth accounting relies on stable cpusets
7647 if (dl_policy(policy
) || dl_policy(p
->policy
)) {
7648 cpuset_locked
= true;
7653 * Make sure no PI-waiters arrive (or leave) while we are
7654 * changing the priority of the task:
7656 * To be able to change p->policy safely, the appropriate
7657 * runqueue lock must be held.
7659 rq
= task_rq_lock(p
, &rf
);
7660 update_rq_clock(rq
);
7663 * Changing the policy of the stop threads its a very bad idea:
7665 if (p
== rq
->stop
) {
7671 * If not changing anything there's no need to proceed further,
7672 * but store a possible modification of reset_on_fork.
7674 if (unlikely(policy
== p
->policy
)) {
7675 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
7677 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
7679 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
7681 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)
7684 p
->sched_reset_on_fork
= reset_on_fork
;
7691 #ifdef CONFIG_RT_GROUP_SCHED
7693 * Do not allow realtime tasks into groups that have no runtime
7696 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
7697 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
7698 !task_group_is_autogroup(task_group(p
))) {
7704 if (dl_bandwidth_enabled() && dl_policy(policy
) &&
7705 !(attr
->sched_flags
& SCHED_FLAG_SUGOV
)) {
7706 cpumask_t
*span
= rq
->rd
->span
;
7709 * Don't allow tasks with an affinity mask smaller than
7710 * the entire root_domain to become SCHED_DEADLINE. We
7711 * will also fail if there's no bandwidth available.
7713 if (!cpumask_subset(span
, p
->cpus_ptr
) ||
7714 rq
->rd
->dl_bw
.bw
== 0) {
7722 /* Re-check policy now with rq lock held: */
7723 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
7724 policy
= oldpolicy
= -1;
7725 task_rq_unlock(rq
, p
, &rf
);
7732 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7733 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7736 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
7741 p
->sched_reset_on_fork
= reset_on_fork
;
7744 newprio
= __normal_prio(policy
, attr
->sched_priority
, attr
->sched_nice
);
7747 * Take priority boosted tasks into account. If the new
7748 * effective priority is unchanged, we just store the new
7749 * normal parameters and do not touch the scheduler class and
7750 * the runqueue. This will be done when the task deboost
7753 newprio
= rt_effective_prio(p
, newprio
);
7754 if (newprio
== oldprio
)
7755 queue_flags
&= ~DEQUEUE_MOVE
;
7758 queued
= task_on_rq_queued(p
);
7759 running
= task_current(rq
, p
);
7761 dequeue_task(rq
, p
, queue_flags
);
7763 put_prev_task(rq
, p
);
7765 prev_class
= p
->sched_class
;
7767 if (!(attr
->sched_flags
& SCHED_FLAG_KEEP_PARAMS
)) {
7768 __setscheduler_params(p
, attr
);
7769 __setscheduler_prio(p
, newprio
);
7771 __setscheduler_uclamp(p
, attr
);
7775 * We enqueue to tail when the priority of a task is
7776 * increased (user space view).
7778 if (oldprio
< p
->prio
)
7779 queue_flags
|= ENQUEUE_HEAD
;
7781 enqueue_task(rq
, p
, queue_flags
);
7784 set_next_task(rq
, p
);
7786 check_class_changed(rq
, p
, prev_class
, oldprio
);
7788 /* Avoid rq from going away on us: */
7790 head
= splice_balance_callbacks(rq
);
7791 task_rq_unlock(rq
, p
, &rf
);
7796 rt_mutex_adjust_pi(p
);
7799 /* Run balance callbacks after we've adjusted the PI chain: */
7800 balance_callbacks(rq
, head
);
7806 task_rq_unlock(rq
, p
, &rf
);
7812 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
7813 const struct sched_param
*param
, bool check
)
7815 struct sched_attr attr
= {
7816 .sched_policy
= policy
,
7817 .sched_priority
= param
->sched_priority
,
7818 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
7821 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7822 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
7823 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
7824 policy
&= ~SCHED_RESET_ON_FORK
;
7825 attr
.sched_policy
= policy
;
7828 return __sched_setscheduler(p
, &attr
, check
, true);
7831 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7832 * @p: the task in question.
7833 * @policy: new policy.
7834 * @param: structure containing the new RT priority.
7836 * Use sched_set_fifo(), read its comment.
7838 * Return: 0 on success. An error code otherwise.
7840 * NOTE that the task may be already dead.
7842 int sched_setscheduler(struct task_struct
*p
, int policy
,
7843 const struct sched_param
*param
)
7845 return _sched_setscheduler(p
, policy
, param
, true);
7848 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
7850 return __sched_setscheduler(p
, attr
, true, true);
7853 int sched_setattr_nocheck(struct task_struct
*p
, const struct sched_attr
*attr
)
7855 return __sched_setscheduler(p
, attr
, false, true);
7857 EXPORT_SYMBOL_GPL(sched_setattr_nocheck
);
7860 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7861 * @p: the task in question.
7862 * @policy: new policy.
7863 * @param: structure containing the new RT priority.
7865 * Just like sched_setscheduler, only don't bother checking if the
7866 * current context has permission. For example, this is needed in
7867 * stop_machine(): we create temporary high priority worker threads,
7868 * but our caller might not have that capability.
7870 * Return: 0 on success. An error code otherwise.
7872 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
7873 const struct sched_param
*param
)
7875 return _sched_setscheduler(p
, policy
, param
, false);
7879 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7880 * incapable of resource management, which is the one thing an OS really should
7883 * This is of course the reason it is limited to privileged users only.
7885 * Worse still; it is fundamentally impossible to compose static priority
7886 * workloads. You cannot take two correctly working static prio workloads
7887 * and smash them together and still expect them to work.
7889 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7893 * The administrator _MUST_ configure the system, the kernel simply doesn't
7894 * know enough information to make a sensible choice.
7896 void sched_set_fifo(struct task_struct
*p
)
7898 struct sched_param sp
= { .sched_priority
= MAX_RT_PRIO
/ 2 };
7899 WARN_ON_ONCE(sched_setscheduler_nocheck(p
, SCHED_FIFO
, &sp
) != 0);
7901 EXPORT_SYMBOL_GPL(sched_set_fifo
);
7904 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7906 void sched_set_fifo_low(struct task_struct
*p
)
7908 struct sched_param sp
= { .sched_priority
= 1 };
7909 WARN_ON_ONCE(sched_setscheduler_nocheck(p
, SCHED_FIFO
, &sp
) != 0);
7911 EXPORT_SYMBOL_GPL(sched_set_fifo_low
);
7913 void sched_set_normal(struct task_struct
*p
, int nice
)
7915 struct sched_attr attr
= {
7916 .sched_policy
= SCHED_NORMAL
,
7919 WARN_ON_ONCE(sched_setattr_nocheck(p
, &attr
) != 0);
7921 EXPORT_SYMBOL_GPL(sched_set_normal
);
7924 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
7926 struct sched_param lparam
;
7927 struct task_struct
*p
;
7930 if (!param
|| pid
< 0)
7932 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
7937 p
= find_process_by_pid(pid
);
7943 retval
= sched_setscheduler(p
, policy
, &lparam
);
7951 * Mimics kernel/events/core.c perf_copy_attr().
7953 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
7958 /* Zero the full structure, so that a short copy will be nice: */
7959 memset(attr
, 0, sizeof(*attr
));
7961 ret
= get_user(size
, &uattr
->size
);
7965 /* ABI compatibility quirk: */
7967 size
= SCHED_ATTR_SIZE_VER0
;
7968 if (size
< SCHED_ATTR_SIZE_VER0
|| size
> PAGE_SIZE
)
7971 ret
= copy_struct_from_user(attr
, sizeof(*attr
), uattr
, size
);
7978 if ((attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) &&
7979 size
< SCHED_ATTR_SIZE_VER1
)
7983 * XXX: Do we want to be lenient like existing syscalls; or do we want
7984 * to be strict and return an error on out-of-bounds values?
7986 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
7991 put_user(sizeof(*attr
), &uattr
->size
);
7995 static void get_params(struct task_struct
*p
, struct sched_attr
*attr
)
7997 if (task_has_dl_policy(p
))
7998 __getparam_dl(p
, attr
);
7999 else if (task_has_rt_policy(p
))
8000 attr
->sched_priority
= p
->rt_priority
;
8002 attr
->sched_nice
= task_nice(p
);
8006 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
8007 * @pid: the pid in question.
8008 * @policy: new policy.
8009 * @param: structure containing the new RT priority.
8011 * Return: 0 on success. An error code otherwise.
8013 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
8018 return do_sched_setscheduler(pid
, policy
, param
);
8022 * sys_sched_setparam - set/change the RT priority of a thread
8023 * @pid: the pid in question.
8024 * @param: structure containing the new RT priority.
8026 * Return: 0 on success. An error code otherwise.
8028 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
8030 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
8034 * sys_sched_setattr - same as above, but with extended sched_attr
8035 * @pid: the pid in question.
8036 * @uattr: structure containing the extended parameters.
8037 * @flags: for future extension.
8039 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
8040 unsigned int, flags
)
8042 struct sched_attr attr
;
8043 struct task_struct
*p
;
8046 if (!uattr
|| pid
< 0 || flags
)
8049 retval
= sched_copy_attr(uattr
, &attr
);
8053 if ((int)attr
.sched_policy
< 0)
8055 if (attr
.sched_flags
& SCHED_FLAG_KEEP_POLICY
)
8056 attr
.sched_policy
= SETPARAM_POLICY
;
8060 p
= find_process_by_pid(pid
);
8066 if (attr
.sched_flags
& SCHED_FLAG_KEEP_PARAMS
)
8067 get_params(p
, &attr
);
8068 retval
= sched_setattr(p
, &attr
);
8076 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
8077 * @pid: the pid in question.
8079 * Return: On success, the policy of the thread. Otherwise, a negative error
8082 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
8084 struct task_struct
*p
;
8092 p
= find_process_by_pid(pid
);
8094 retval
= security_task_getscheduler(p
);
8097 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
8104 * sys_sched_getparam - get the RT priority of a thread
8105 * @pid: the pid in question.
8106 * @param: structure containing the RT priority.
8108 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
8111 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
8113 struct sched_param lp
= { .sched_priority
= 0 };
8114 struct task_struct
*p
;
8117 if (!param
|| pid
< 0)
8121 p
= find_process_by_pid(pid
);
8126 retval
= security_task_getscheduler(p
);
8130 if (task_has_rt_policy(p
))
8131 lp
.sched_priority
= p
->rt_priority
;
8135 * This one might sleep, we cannot do it with a spinlock held ...
8137 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
8147 * Copy the kernel size attribute structure (which might be larger
8148 * than what user-space knows about) to user-space.
8150 * Note that all cases are valid: user-space buffer can be larger or
8151 * smaller than the kernel-space buffer. The usual case is that both
8152 * have the same size.
8155 sched_attr_copy_to_user(struct sched_attr __user
*uattr
,
8156 struct sched_attr
*kattr
,
8159 unsigned int ksize
= sizeof(*kattr
);
8161 if (!access_ok(uattr
, usize
))
8165 * sched_getattr() ABI forwards and backwards compatibility:
8167 * If usize == ksize then we just copy everything to user-space and all is good.
8169 * If usize < ksize then we only copy as much as user-space has space for,
8170 * this keeps ABI compatibility as well. We skip the rest.
8172 * If usize > ksize then user-space is using a newer version of the ABI,
8173 * which part the kernel doesn't know about. Just ignore it - tooling can
8174 * detect the kernel's knowledge of attributes from the attr->size value
8175 * which is set to ksize in this case.
8177 kattr
->size
= min(usize
, ksize
);
8179 if (copy_to_user(uattr
, kattr
, kattr
->size
))
8186 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8187 * @pid: the pid in question.
8188 * @uattr: structure containing the extended parameters.
8189 * @usize: sizeof(attr) for fwd/bwd comp.
8190 * @flags: for future extension.
8192 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
8193 unsigned int, usize
, unsigned int, flags
)
8195 struct sched_attr kattr
= { };
8196 struct task_struct
*p
;
8199 if (!uattr
|| pid
< 0 || usize
> PAGE_SIZE
||
8200 usize
< SCHED_ATTR_SIZE_VER0
|| flags
)
8204 p
= find_process_by_pid(pid
);
8209 retval
= security_task_getscheduler(p
);
8213 kattr
.sched_policy
= p
->policy
;
8214 if (p
->sched_reset_on_fork
)
8215 kattr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
8216 get_params(p
, &kattr
);
8217 kattr
.sched_flags
&= SCHED_FLAG_ALL
;
8219 #ifdef CONFIG_UCLAMP_TASK
8221 * This could race with another potential updater, but this is fine
8222 * because it'll correctly read the old or the new value. We don't need
8223 * to guarantee who wins the race as long as it doesn't return garbage.
8225 kattr
.sched_util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
8226 kattr
.sched_util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
8231 return sched_attr_copy_to_user(uattr
, &kattr
, usize
);
8239 int dl_task_check_affinity(struct task_struct
*p
, const struct cpumask
*mask
)
8244 * If the task isn't a deadline task or admission control is
8245 * disabled then we don't care about affinity changes.
8247 if (!task_has_dl_policy(p
) || !dl_bandwidth_enabled())
8251 * Since bandwidth control happens on root_domain basis,
8252 * if admission test is enabled, we only admit -deadline
8253 * tasks allowed to run on all the CPUs in the task's
8257 if (!cpumask_subset(task_rq(p
)->rd
->span
, mask
))
8265 __sched_setaffinity(struct task_struct
*p
, struct affinity_context
*ctx
)
8268 cpumask_var_t cpus_allowed
, new_mask
;
8270 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
))
8273 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
8275 goto out_free_cpus_allowed
;
8278 cpuset_cpus_allowed(p
, cpus_allowed
);
8279 cpumask_and(new_mask
, ctx
->new_mask
, cpus_allowed
);
8281 ctx
->new_mask
= new_mask
;
8282 ctx
->flags
|= SCA_CHECK
;
8284 retval
= dl_task_check_affinity(p
, new_mask
);
8286 goto out_free_new_mask
;
8288 retval
= __set_cpus_allowed_ptr(p
, ctx
);
8290 goto out_free_new_mask
;
8292 cpuset_cpus_allowed(p
, cpus_allowed
);
8293 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
8295 * We must have raced with a concurrent cpuset update.
8296 * Just reset the cpumask to the cpuset's cpus_allowed.
8298 cpumask_copy(new_mask
, cpus_allowed
);
8301 * If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
8302 * will restore the previous user_cpus_ptr value.
8304 * In the unlikely event a previous user_cpus_ptr exists,
8305 * we need to further restrict the mask to what is allowed
8306 * by that old user_cpus_ptr.
8308 if (unlikely((ctx
->flags
& SCA_USER
) && ctx
->user_mask
)) {
8309 bool empty
= !cpumask_and(new_mask
, new_mask
,
8312 if (WARN_ON_ONCE(empty
))
8313 cpumask_copy(new_mask
, cpus_allowed
);
8315 __set_cpus_allowed_ptr(p
, ctx
);
8320 free_cpumask_var(new_mask
);
8321 out_free_cpus_allowed
:
8322 free_cpumask_var(cpus_allowed
);
8326 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
8328 struct affinity_context ac
;
8329 struct cpumask
*user_mask
;
8330 struct task_struct
*p
;
8335 p
= find_process_by_pid(pid
);
8341 /* Prevent p going away */
8345 if (p
->flags
& PF_NO_SETAFFINITY
) {
8350 if (!check_same_owner(p
)) {
8352 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
8360 retval
= security_task_setscheduler(p
);
8365 * With non-SMP configs, user_cpus_ptr/user_mask isn't used and
8366 * alloc_user_cpus_ptr() returns NULL.
8368 user_mask
= alloc_user_cpus_ptr(NUMA_NO_NODE
);
8370 cpumask_copy(user_mask
, in_mask
);
8371 } else if (IS_ENABLED(CONFIG_SMP
)) {
8376 ac
= (struct affinity_context
){
8377 .new_mask
= in_mask
,
8378 .user_mask
= user_mask
,
8382 retval
= __sched_setaffinity(p
, &ac
);
8383 kfree(ac
.user_mask
);
8390 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
8391 struct cpumask
*new_mask
)
8393 if (len
< cpumask_size())
8394 cpumask_clear(new_mask
);
8395 else if (len
> cpumask_size())
8396 len
= cpumask_size();
8398 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
8402 * sys_sched_setaffinity - set the CPU affinity of a process
8403 * @pid: pid of the process
8404 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8405 * @user_mask_ptr: user-space pointer to the new CPU mask
8407 * Return: 0 on success. An error code otherwise.
8409 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
8410 unsigned long __user
*, user_mask_ptr
)
8412 cpumask_var_t new_mask
;
8415 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
8418 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
8420 retval
= sched_setaffinity(pid
, new_mask
);
8421 free_cpumask_var(new_mask
);
8425 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
8427 struct task_struct
*p
;
8428 unsigned long flags
;
8434 p
= find_process_by_pid(pid
);
8438 retval
= security_task_getscheduler(p
);
8442 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
8443 cpumask_and(mask
, &p
->cpus_mask
, cpu_active_mask
);
8444 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
8453 * sys_sched_getaffinity - get the CPU affinity of a process
8454 * @pid: pid of the process
8455 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8456 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8458 * Return: size of CPU mask copied to user_mask_ptr on success. An
8459 * error code otherwise.
8461 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
8462 unsigned long __user
*, user_mask_ptr
)
8467 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
8469 if (len
& (sizeof(unsigned long)-1))
8472 if (!zalloc_cpumask_var(&mask
, GFP_KERNEL
))
8475 ret
= sched_getaffinity(pid
, mask
);
8477 unsigned int retlen
= min(len
, cpumask_size());
8479 if (copy_to_user(user_mask_ptr
, cpumask_bits(mask
), retlen
))
8484 free_cpumask_var(mask
);
8489 static void do_sched_yield(void)
8494 rq
= this_rq_lock_irq(&rf
);
8496 schedstat_inc(rq
->yld_count
);
8497 current
->sched_class
->yield_task(rq
);
8500 rq_unlock_irq(rq
, &rf
);
8501 sched_preempt_enable_no_resched();
8507 * sys_sched_yield - yield the current processor to other threads.
8509 * This function yields the current CPU to other tasks. If there are no
8510 * other threads running on this CPU then this function will return.
8514 SYSCALL_DEFINE0(sched_yield
)
8520 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8521 int __sched
__cond_resched(void)
8523 if (should_resched(0)) {
8524 preempt_schedule_common();
8528 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8529 * whether the current CPU is in an RCU read-side critical section,
8530 * so the tick can report quiescent states even for CPUs looping
8531 * in kernel context. In contrast, in non-preemptible kernels,
8532 * RCU readers leave no in-memory hints, which means that CPU-bound
8533 * processes executing in kernel context might never report an
8534 * RCU quiescent state. Therefore, the following code causes
8535 * cond_resched() to report a quiescent state, but only when RCU
8536 * is in urgent need of one.
8538 #ifndef CONFIG_PREEMPT_RCU
8543 EXPORT_SYMBOL(__cond_resched
);
8546 #ifdef CONFIG_PREEMPT_DYNAMIC
8547 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8548 #define cond_resched_dynamic_enabled __cond_resched
8549 #define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
8550 DEFINE_STATIC_CALL_RET0(cond_resched
, __cond_resched
);
8551 EXPORT_STATIC_CALL_TRAMP(cond_resched
);
8553 #define might_resched_dynamic_enabled __cond_resched
8554 #define might_resched_dynamic_disabled ((void *)&__static_call_return0)
8555 DEFINE_STATIC_CALL_RET0(might_resched
, __cond_resched
);
8556 EXPORT_STATIC_CALL_TRAMP(might_resched
);
8557 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8558 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched
);
8559 int __sched
dynamic_cond_resched(void)
8561 klp_sched_try_switch();
8562 if (!static_branch_unlikely(&sk_dynamic_cond_resched
))
8564 return __cond_resched();
8566 EXPORT_SYMBOL(dynamic_cond_resched
);
8568 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched
);
8569 int __sched
dynamic_might_resched(void)
8571 if (!static_branch_unlikely(&sk_dynamic_might_resched
))
8573 return __cond_resched();
8575 EXPORT_SYMBOL(dynamic_might_resched
);
8580 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8581 * call schedule, and on return reacquire the lock.
8583 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8584 * operations here to prevent schedule() from being called twice (once via
8585 * spin_unlock(), once by hand).
8587 int __cond_resched_lock(spinlock_t
*lock
)
8589 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
8592 lockdep_assert_held(lock
);
8594 if (spin_needbreak(lock
) || resched
) {
8596 if (!_cond_resched())
8603 EXPORT_SYMBOL(__cond_resched_lock
);
8605 int __cond_resched_rwlock_read(rwlock_t
*lock
)
8607 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
8610 lockdep_assert_held_read(lock
);
8612 if (rwlock_needbreak(lock
) || resched
) {
8614 if (!_cond_resched())
8621 EXPORT_SYMBOL(__cond_resched_rwlock_read
);
8623 int __cond_resched_rwlock_write(rwlock_t
*lock
)
8625 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
8628 lockdep_assert_held_write(lock
);
8630 if (rwlock_needbreak(lock
) || resched
) {
8632 if (!_cond_resched())
8639 EXPORT_SYMBOL(__cond_resched_rwlock_write
);
8641 #ifdef CONFIG_PREEMPT_DYNAMIC
8643 #ifdef CONFIG_GENERIC_ENTRY
8644 #include <linux/entry-common.h>
8650 * SC:preempt_schedule
8651 * SC:preempt_schedule_notrace
8652 * SC:irqentry_exit_cond_resched
8656 * cond_resched <- __cond_resched
8657 * might_resched <- RET0
8658 * preempt_schedule <- NOP
8659 * preempt_schedule_notrace <- NOP
8660 * irqentry_exit_cond_resched <- NOP
8663 * cond_resched <- __cond_resched
8664 * might_resched <- __cond_resched
8665 * preempt_schedule <- NOP
8666 * preempt_schedule_notrace <- NOP
8667 * irqentry_exit_cond_resched <- NOP
8670 * cond_resched <- RET0
8671 * might_resched <- RET0
8672 * preempt_schedule <- preempt_schedule
8673 * preempt_schedule_notrace <- preempt_schedule_notrace
8674 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8678 preempt_dynamic_undefined
= -1,
8679 preempt_dynamic_none
,
8680 preempt_dynamic_voluntary
,
8681 preempt_dynamic_full
,
8684 int preempt_dynamic_mode
= preempt_dynamic_undefined
;
8686 int sched_dynamic_mode(const char *str
)
8688 if (!strcmp(str
, "none"))
8689 return preempt_dynamic_none
;
8691 if (!strcmp(str
, "voluntary"))
8692 return preempt_dynamic_voluntary
;
8694 if (!strcmp(str
, "full"))
8695 return preempt_dynamic_full
;
8700 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
8701 #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8702 #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
8703 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8704 #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8705 #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8707 #error "Unsupported PREEMPT_DYNAMIC mechanism"
8710 static DEFINE_MUTEX(sched_dynamic_mutex
);
8711 static bool klp_override
;
8713 static void __sched_dynamic_update(int mode
)
8716 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8717 * the ZERO state, which is invalid.
8720 preempt_dynamic_enable(cond_resched
);
8721 preempt_dynamic_enable(might_resched
);
8722 preempt_dynamic_enable(preempt_schedule
);
8723 preempt_dynamic_enable(preempt_schedule_notrace
);
8724 preempt_dynamic_enable(irqentry_exit_cond_resched
);
8727 case preempt_dynamic_none
:
8729 preempt_dynamic_enable(cond_resched
);
8730 preempt_dynamic_disable(might_resched
);
8731 preempt_dynamic_disable(preempt_schedule
);
8732 preempt_dynamic_disable(preempt_schedule_notrace
);
8733 preempt_dynamic_disable(irqentry_exit_cond_resched
);
8734 if (mode
!= preempt_dynamic_mode
)
8735 pr_info("Dynamic Preempt: none\n");
8738 case preempt_dynamic_voluntary
:
8740 preempt_dynamic_enable(cond_resched
);
8741 preempt_dynamic_enable(might_resched
);
8742 preempt_dynamic_disable(preempt_schedule
);
8743 preempt_dynamic_disable(preempt_schedule_notrace
);
8744 preempt_dynamic_disable(irqentry_exit_cond_resched
);
8745 if (mode
!= preempt_dynamic_mode
)
8746 pr_info("Dynamic Preempt: voluntary\n");
8749 case preempt_dynamic_full
:
8751 preempt_dynamic_disable(cond_resched
);
8752 preempt_dynamic_disable(might_resched
);
8753 preempt_dynamic_enable(preempt_schedule
);
8754 preempt_dynamic_enable(preempt_schedule_notrace
);
8755 preempt_dynamic_enable(irqentry_exit_cond_resched
);
8756 if (mode
!= preempt_dynamic_mode
)
8757 pr_info("Dynamic Preempt: full\n");
8761 preempt_dynamic_mode
= mode
;
8764 void sched_dynamic_update(int mode
)
8766 mutex_lock(&sched_dynamic_mutex
);
8767 __sched_dynamic_update(mode
);
8768 mutex_unlock(&sched_dynamic_mutex
);
8771 #ifdef CONFIG_HAVE_PREEMPT_DYNAMIC_CALL
8773 static int klp_cond_resched(void)
8775 __klp_sched_try_switch();
8776 return __cond_resched();
8779 void sched_dynamic_klp_enable(void)
8781 mutex_lock(&sched_dynamic_mutex
);
8783 klp_override
= true;
8784 static_call_update(cond_resched
, klp_cond_resched
);
8786 mutex_unlock(&sched_dynamic_mutex
);
8789 void sched_dynamic_klp_disable(void)
8791 mutex_lock(&sched_dynamic_mutex
);
8793 klp_override
= false;
8794 __sched_dynamic_update(preempt_dynamic_mode
);
8796 mutex_unlock(&sched_dynamic_mutex
);
8799 #endif /* CONFIG_HAVE_PREEMPT_DYNAMIC_CALL */
8801 static int __init
setup_preempt_mode(char *str
)
8803 int mode
= sched_dynamic_mode(str
);
8805 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str
);
8809 sched_dynamic_update(mode
);
8812 __setup("preempt=", setup_preempt_mode
);
8814 static void __init
preempt_dynamic_init(void)
8816 if (preempt_dynamic_mode
== preempt_dynamic_undefined
) {
8817 if (IS_ENABLED(CONFIG_PREEMPT_NONE
)) {
8818 sched_dynamic_update(preempt_dynamic_none
);
8819 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY
)) {
8820 sched_dynamic_update(preempt_dynamic_voluntary
);
8822 /* Default static call setting, nothing to do */
8823 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT
));
8824 preempt_dynamic_mode
= preempt_dynamic_full
;
8825 pr_info("Dynamic Preempt: full\n");
8830 #define PREEMPT_MODEL_ACCESSOR(mode) \
8831 bool preempt_model_##mode(void) \
8833 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8834 return preempt_dynamic_mode == preempt_dynamic_##mode; \
8836 EXPORT_SYMBOL_GPL(preempt_model_##mode)
8838 PREEMPT_MODEL_ACCESSOR(none
);
8839 PREEMPT_MODEL_ACCESSOR(voluntary
);
8840 PREEMPT_MODEL_ACCESSOR(full
);
8842 #else /* !CONFIG_PREEMPT_DYNAMIC */
8844 static inline void preempt_dynamic_init(void) { }
8846 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8849 * yield - yield the current processor to other threads.
8851 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8853 * The scheduler is at all times free to pick the calling task as the most
8854 * eligible task to run, if removing the yield() call from your code breaks
8855 * it, it's already broken.
8857 * Typical broken usage is:
8862 * where one assumes that yield() will let 'the other' process run that will
8863 * make event true. If the current task is a SCHED_FIFO task that will never
8864 * happen. Never use yield() as a progress guarantee!!
8866 * If you want to use yield() to wait for something, use wait_event().
8867 * If you want to use yield() to be 'nice' for others, use cond_resched().
8868 * If you still want to use yield(), do not!
8870 void __sched
yield(void)
8872 set_current_state(TASK_RUNNING
);
8875 EXPORT_SYMBOL(yield
);
8878 * yield_to - yield the current processor to another thread in
8879 * your thread group, or accelerate that thread toward the
8880 * processor it's on.
8882 * @preempt: whether task preemption is allowed or not
8884 * It's the caller's job to ensure that the target task struct
8885 * can't go away on us before we can do any checks.
8888 * true (>0) if we indeed boosted the target task.
8889 * false (0) if we failed to boost the target.
8890 * -ESRCH if there's no task to yield to.
8892 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
8894 struct task_struct
*curr
= current
;
8895 struct rq
*rq
, *p_rq
;
8896 unsigned long flags
;
8899 local_irq_save(flags
);
8905 * If we're the only runnable task on the rq and target rq also
8906 * has only one task, there's absolutely no point in yielding.
8908 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
8913 double_rq_lock(rq
, p_rq
);
8914 if (task_rq(p
) != p_rq
) {
8915 double_rq_unlock(rq
, p_rq
);
8919 if (!curr
->sched_class
->yield_to_task
)
8922 if (curr
->sched_class
!= p
->sched_class
)
8925 if (task_on_cpu(p_rq
, p
) || !task_is_running(p
))
8928 yielded
= curr
->sched_class
->yield_to_task(rq
, p
);
8930 schedstat_inc(rq
->yld_count
);
8932 * Make p's CPU reschedule; pick_next_entity takes care of
8935 if (preempt
&& rq
!= p_rq
)
8940 double_rq_unlock(rq
, p_rq
);
8942 local_irq_restore(flags
);
8949 EXPORT_SYMBOL_GPL(yield_to
);
8951 int io_schedule_prepare(void)
8953 int old_iowait
= current
->in_iowait
;
8955 current
->in_iowait
= 1;
8956 blk_flush_plug(current
->plug
, true);
8960 void io_schedule_finish(int token
)
8962 current
->in_iowait
= token
;
8966 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8967 * that process accounting knows that this is a task in IO wait state.
8969 long __sched
io_schedule_timeout(long timeout
)
8974 token
= io_schedule_prepare();
8975 ret
= schedule_timeout(timeout
);
8976 io_schedule_finish(token
);
8980 EXPORT_SYMBOL(io_schedule_timeout
);
8982 void __sched
io_schedule(void)
8986 token
= io_schedule_prepare();
8988 io_schedule_finish(token
);
8990 EXPORT_SYMBOL(io_schedule
);
8993 * sys_sched_get_priority_max - return maximum RT priority.
8994 * @policy: scheduling class.
8996 * Return: On success, this syscall returns the maximum
8997 * rt_priority that can be used by a given scheduling class.
8998 * On failure, a negative error code is returned.
9000 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
9007 ret
= MAX_RT_PRIO
-1;
9009 case SCHED_DEADLINE
:
9020 * sys_sched_get_priority_min - return minimum RT priority.
9021 * @policy: scheduling class.
9023 * Return: On success, this syscall returns the minimum
9024 * rt_priority that can be used by a given scheduling class.
9025 * On failure, a negative error code is returned.
9027 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
9036 case SCHED_DEADLINE
:
9045 static int sched_rr_get_interval(pid_t pid
, struct timespec64
*t
)
9047 struct task_struct
*p
;
9048 unsigned int time_slice
;
9058 p
= find_process_by_pid(pid
);
9062 retval
= security_task_getscheduler(p
);
9066 rq
= task_rq_lock(p
, &rf
);
9068 if (p
->sched_class
->get_rr_interval
)
9069 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
9070 task_rq_unlock(rq
, p
, &rf
);
9073 jiffies_to_timespec64(time_slice
, t
);
9082 * sys_sched_rr_get_interval - return the default timeslice of a process.
9083 * @pid: pid of the process.
9084 * @interval: userspace pointer to the timeslice value.
9086 * this syscall writes the default timeslice value of a given process
9087 * into the user-space timespec buffer. A value of '0' means infinity.
9089 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
9092 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
9093 struct __kernel_timespec __user
*, interval
)
9095 struct timespec64 t
;
9096 int retval
= sched_rr_get_interval(pid
, &t
);
9099 retval
= put_timespec64(&t
, interval
);
9104 #ifdef CONFIG_COMPAT_32BIT_TIME
9105 SYSCALL_DEFINE2(sched_rr_get_interval_time32
, pid_t
, pid
,
9106 struct old_timespec32 __user
*, interval
)
9108 struct timespec64 t
;
9109 int retval
= sched_rr_get_interval(pid
, &t
);
9112 retval
= put_old_timespec32(&t
, interval
);
9117 void sched_show_task(struct task_struct
*p
)
9119 unsigned long free
= 0;
9122 if (!try_get_task_stack(p
))
9125 pr_info("task:%-15.15s state:%c", p
->comm
, task_state_to_char(p
));
9127 if (task_is_running(p
))
9128 pr_cont(" running task ");
9129 #ifdef CONFIG_DEBUG_STACK_USAGE
9130 free
= stack_not_used(p
);
9135 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
9137 pr_cont(" stack:%-5lu pid:%-5d ppid:%-6d flags:0x%08lx\n",
9138 free
, task_pid_nr(p
), ppid
,
9139 read_task_thread_flags(p
));
9141 print_worker_info(KERN_INFO
, p
);
9142 print_stop_info(KERN_INFO
, p
);
9143 show_stack(p
, NULL
, KERN_INFO
);
9146 EXPORT_SYMBOL_GPL(sched_show_task
);
9149 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
9151 unsigned int state
= READ_ONCE(p
->__state
);
9153 /* no filter, everything matches */
9157 /* filter, but doesn't match */
9158 if (!(state
& state_filter
))
9162 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
9165 if (state_filter
== TASK_UNINTERRUPTIBLE
&& (state
& TASK_NOLOAD
))
9172 void show_state_filter(unsigned int state_filter
)
9174 struct task_struct
*g
, *p
;
9177 for_each_process_thread(g
, p
) {
9179 * reset the NMI-timeout, listing all files on a slow
9180 * console might take a lot of time:
9181 * Also, reset softlockup watchdogs on all CPUs, because
9182 * another CPU might be blocked waiting for us to process
9185 touch_nmi_watchdog();
9186 touch_all_softlockup_watchdogs();
9187 if (state_filter_match(state_filter
, p
))
9191 #ifdef CONFIG_SCHED_DEBUG
9193 sysrq_sched_debug_show();
9197 * Only show locks if all tasks are dumped:
9200 debug_show_all_locks();
9204 * init_idle - set up an idle thread for a given CPU
9205 * @idle: task in question
9206 * @cpu: CPU the idle task belongs to
9208 * NOTE: this function does not set the idle thread's NEED_RESCHED
9209 * flag, to make booting more robust.
9211 void __init
init_idle(struct task_struct
*idle
, int cpu
)
9214 struct affinity_context ac
= (struct affinity_context
) {
9215 .new_mask
= cpumask_of(cpu
),
9219 struct rq
*rq
= cpu_rq(cpu
);
9220 unsigned long flags
;
9222 __sched_fork(0, idle
);
9224 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
9225 raw_spin_rq_lock(rq
);
9227 idle
->__state
= TASK_RUNNING
;
9228 idle
->se
.exec_start
= sched_clock();
9230 * PF_KTHREAD should already be set at this point; regardless, make it
9231 * look like a proper per-CPU kthread.
9233 idle
->flags
|= PF_IDLE
| PF_KTHREAD
| PF_NO_SETAFFINITY
;
9234 kthread_set_per_cpu(idle
, cpu
);
9238 * It's possible that init_idle() gets called multiple times on a task,
9239 * in that case do_set_cpus_allowed() will not do the right thing.
9241 * And since this is boot we can forgo the serialization.
9243 set_cpus_allowed_common(idle
, &ac
);
9246 * We're having a chicken and egg problem, even though we are
9247 * holding rq->lock, the CPU isn't yet set to this CPU so the
9248 * lockdep check in task_group() will fail.
9250 * Similar case to sched_fork(). / Alternatively we could
9251 * use task_rq_lock() here and obtain the other rq->lock.
9256 __set_task_cpu(idle
, cpu
);
9260 rcu_assign_pointer(rq
->curr
, idle
);
9261 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
9265 raw_spin_rq_unlock(rq
);
9266 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
9268 /* Set the preempt count _outside_ the spinlocks! */
9269 init_idle_preempt_count(idle
, cpu
);
9272 * The idle tasks have their own, simple scheduling class:
9274 idle
->sched_class
= &idle_sched_class
;
9275 ftrace_graph_init_idle_task(idle
, cpu
);
9276 vtime_init_idle(idle
, cpu
);
9278 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
9284 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
9285 const struct cpumask
*trial
)
9289 if (cpumask_empty(cur
))
9292 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
9297 int task_can_attach(struct task_struct
*p
)
9302 * Kthreads which disallow setaffinity shouldn't be moved
9303 * to a new cpuset; we don't want to change their CPU
9304 * affinity and isolating such threads by their set of
9305 * allowed nodes is unnecessary. Thus, cpusets are not
9306 * applicable for such threads. This prevents checking for
9307 * success of set_cpus_allowed_ptr() on all attached tasks
9308 * before cpus_mask may be changed.
9310 if (p
->flags
& PF_NO_SETAFFINITY
)
9316 bool sched_smp_initialized __read_mostly
;
9318 #ifdef CONFIG_NUMA_BALANCING
9319 /* Migrate current task p to target_cpu */
9320 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
9322 struct migration_arg arg
= { p
, target_cpu
};
9323 int curr_cpu
= task_cpu(p
);
9325 if (curr_cpu
== target_cpu
)
9328 if (!cpumask_test_cpu(target_cpu
, p
->cpus_ptr
))
9331 /* TODO: This is not properly updating schedstats */
9333 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
9334 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
9338 * Requeue a task on a given node and accurately track the number of NUMA
9339 * tasks on the runqueues
9341 void sched_setnuma(struct task_struct
*p
, int nid
)
9343 bool queued
, running
;
9347 rq
= task_rq_lock(p
, &rf
);
9348 queued
= task_on_rq_queued(p
);
9349 running
= task_current(rq
, p
);
9352 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
9354 put_prev_task(rq
, p
);
9356 p
->numa_preferred_nid
= nid
;
9359 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
9361 set_next_task(rq
, p
);
9362 task_rq_unlock(rq
, p
, &rf
);
9364 #endif /* CONFIG_NUMA_BALANCING */
9366 #ifdef CONFIG_HOTPLUG_CPU
9368 * Ensure that the idle task is using init_mm right before its CPU goes
9371 void idle_task_exit(void)
9373 struct mm_struct
*mm
= current
->active_mm
;
9375 BUG_ON(cpu_online(smp_processor_id()));
9376 BUG_ON(current
!= this_rq()->idle
);
9378 if (mm
!= &init_mm
) {
9379 switch_mm(mm
, &init_mm
, current
);
9380 finish_arch_post_lock_switch();
9383 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
9386 static int __balance_push_cpu_stop(void *arg
)
9388 struct task_struct
*p
= arg
;
9389 struct rq
*rq
= this_rq();
9393 raw_spin_lock_irq(&p
->pi_lock
);
9396 update_rq_clock(rq
);
9398 if (task_rq(p
) == rq
&& task_on_rq_queued(p
)) {
9399 cpu
= select_fallback_rq(rq
->cpu
, p
);
9400 rq
= __migrate_task(rq
, &rf
, p
, cpu
);
9404 raw_spin_unlock_irq(&p
->pi_lock
);
9411 static DEFINE_PER_CPU(struct cpu_stop_work
, push_work
);
9414 * Ensure we only run per-cpu kthreads once the CPU goes !active.
9416 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9417 * effective when the hotplug motion is down.
9419 static void balance_push(struct rq
*rq
)
9421 struct task_struct
*push_task
= rq
->curr
;
9423 lockdep_assert_rq_held(rq
);
9426 * Ensure the thing is persistent until balance_push_set(.on = false);
9428 rq
->balance_callback
= &balance_push_callback
;
9431 * Only active while going offline and when invoked on the outgoing
9434 if (!cpu_dying(rq
->cpu
) || rq
!= this_rq())
9438 * Both the cpu-hotplug and stop task are in this case and are
9439 * required to complete the hotplug process.
9441 if (kthread_is_per_cpu(push_task
) ||
9442 is_migration_disabled(push_task
)) {
9445 * If this is the idle task on the outgoing CPU try to wake
9446 * up the hotplug control thread which might wait for the
9447 * last task to vanish. The rcuwait_active() check is
9448 * accurate here because the waiter is pinned on this CPU
9449 * and can't obviously be running in parallel.
9451 * On RT kernels this also has to check whether there are
9452 * pinned and scheduled out tasks on the runqueue. They
9453 * need to leave the migrate disabled section first.
9455 if (!rq
->nr_running
&& !rq_has_pinned_tasks(rq
) &&
9456 rcuwait_active(&rq
->hotplug_wait
)) {
9457 raw_spin_rq_unlock(rq
);
9458 rcuwait_wake_up(&rq
->hotplug_wait
);
9459 raw_spin_rq_lock(rq
);
9464 get_task_struct(push_task
);
9466 * Temporarily drop rq->lock such that we can wake-up the stop task.
9467 * Both preemption and IRQs are still disabled.
9469 raw_spin_rq_unlock(rq
);
9470 stop_one_cpu_nowait(rq
->cpu
, __balance_push_cpu_stop
, push_task
,
9471 this_cpu_ptr(&push_work
));
9473 * At this point need_resched() is true and we'll take the loop in
9474 * schedule(). The next pick is obviously going to be the stop task
9475 * which kthread_is_per_cpu() and will push this task away.
9477 raw_spin_rq_lock(rq
);
9480 static void balance_push_set(int cpu
, bool on
)
9482 struct rq
*rq
= cpu_rq(cpu
);
9485 rq_lock_irqsave(rq
, &rf
);
9487 WARN_ON_ONCE(rq
->balance_callback
);
9488 rq
->balance_callback
= &balance_push_callback
;
9489 } else if (rq
->balance_callback
== &balance_push_callback
) {
9490 rq
->balance_callback
= NULL
;
9492 rq_unlock_irqrestore(rq
, &rf
);
9496 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9497 * inactive. All tasks which are not per CPU kernel threads are either
9498 * pushed off this CPU now via balance_push() or placed on a different CPU
9499 * during wakeup. Wait until the CPU is quiescent.
9501 static void balance_hotplug_wait(void)
9503 struct rq
*rq
= this_rq();
9505 rcuwait_wait_event(&rq
->hotplug_wait
,
9506 rq
->nr_running
== 1 && !rq_has_pinned_tasks(rq
),
9507 TASK_UNINTERRUPTIBLE
);
9512 static inline void balance_push(struct rq
*rq
)
9516 static inline void balance_push_set(int cpu
, bool on
)
9520 static inline void balance_hotplug_wait(void)
9524 #endif /* CONFIG_HOTPLUG_CPU */
9526 void set_rq_online(struct rq
*rq
)
9529 const struct sched_class
*class;
9531 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
9534 for_each_class(class) {
9535 if (class->rq_online
)
9536 class->rq_online(rq
);
9541 void set_rq_offline(struct rq
*rq
)
9544 const struct sched_class
*class;
9546 for_each_class(class) {
9547 if (class->rq_offline
)
9548 class->rq_offline(rq
);
9551 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
9557 * used to mark begin/end of suspend/resume:
9559 static int num_cpus_frozen
;
9562 * Update cpusets according to cpu_active mask. If cpusets are
9563 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9564 * around partition_sched_domains().
9566 * If we come here as part of a suspend/resume, don't touch cpusets because we
9567 * want to restore it back to its original state upon resume anyway.
9569 static void cpuset_cpu_active(void)
9571 if (cpuhp_tasks_frozen
) {
9573 * num_cpus_frozen tracks how many CPUs are involved in suspend
9574 * resume sequence. As long as this is not the last online
9575 * operation in the resume sequence, just build a single sched
9576 * domain, ignoring cpusets.
9578 partition_sched_domains(1, NULL
, NULL
);
9579 if (--num_cpus_frozen
)
9582 * This is the last CPU online operation. So fall through and
9583 * restore the original sched domains by considering the
9584 * cpuset configurations.
9586 cpuset_force_rebuild();
9588 cpuset_update_active_cpus();
9591 static int cpuset_cpu_inactive(unsigned int cpu
)
9593 if (!cpuhp_tasks_frozen
) {
9594 int ret
= dl_bw_check_overflow(cpu
);
9598 cpuset_update_active_cpus();
9601 partition_sched_domains(1, NULL
, NULL
);
9606 int sched_cpu_activate(unsigned int cpu
)
9608 struct rq
*rq
= cpu_rq(cpu
);
9612 * Clear the balance_push callback and prepare to schedule
9615 balance_push_set(cpu
, false);
9617 #ifdef CONFIG_SCHED_SMT
9619 * When going up, increment the number of cores with SMT present.
9621 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
9622 static_branch_inc_cpuslocked(&sched_smt_present
);
9624 set_cpu_active(cpu
, true);
9626 if (sched_smp_initialized
) {
9627 sched_update_numa(cpu
, true);
9628 sched_domains_numa_masks_set(cpu
);
9629 cpuset_cpu_active();
9633 * Put the rq online, if not already. This happens:
9635 * 1) In the early boot process, because we build the real domains
9636 * after all CPUs have been brought up.
9638 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9641 rq_lock_irqsave(rq
, &rf
);
9643 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
9646 rq_unlock_irqrestore(rq
, &rf
);
9651 int sched_cpu_deactivate(unsigned int cpu
)
9653 struct rq
*rq
= cpu_rq(cpu
);
9658 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9659 * load balancing when not active
9661 nohz_balance_exit_idle(rq
);
9663 set_cpu_active(cpu
, false);
9666 * From this point forward, this CPU will refuse to run any task that
9667 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9668 * push those tasks away until this gets cleared, see
9669 * sched_cpu_dying().
9671 balance_push_set(cpu
, true);
9674 * We've cleared cpu_active_mask / set balance_push, wait for all
9675 * preempt-disabled and RCU users of this state to go away such that
9676 * all new such users will observe it.
9678 * Specifically, we rely on ttwu to no longer target this CPU, see
9679 * ttwu_queue_cond() and is_cpu_allowed().
9681 * Do sync before park smpboot threads to take care the rcu boost case.
9685 rq_lock_irqsave(rq
, &rf
);
9687 update_rq_clock(rq
);
9688 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
9691 rq_unlock_irqrestore(rq
, &rf
);
9693 #ifdef CONFIG_SCHED_SMT
9695 * When going down, decrement the number of cores with SMT present.
9697 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
9698 static_branch_dec_cpuslocked(&sched_smt_present
);
9700 sched_core_cpu_deactivate(cpu
);
9703 if (!sched_smp_initialized
)
9706 sched_update_numa(cpu
, false);
9707 ret
= cpuset_cpu_inactive(cpu
);
9709 balance_push_set(cpu
, false);
9710 set_cpu_active(cpu
, true);
9711 sched_update_numa(cpu
, true);
9714 sched_domains_numa_masks_clear(cpu
);
9718 static void sched_rq_cpu_starting(unsigned int cpu
)
9720 struct rq
*rq
= cpu_rq(cpu
);
9722 rq
->calc_load_update
= calc_load_update
;
9723 update_max_interval();
9726 int sched_cpu_starting(unsigned int cpu
)
9728 sched_core_cpu_starting(cpu
);
9729 sched_rq_cpu_starting(cpu
);
9730 sched_tick_start(cpu
);
9734 #ifdef CONFIG_HOTPLUG_CPU
9737 * Invoked immediately before the stopper thread is invoked to bring the
9738 * CPU down completely. At this point all per CPU kthreads except the
9739 * hotplug thread (current) and the stopper thread (inactive) have been
9740 * either parked or have been unbound from the outgoing CPU. Ensure that
9741 * any of those which might be on the way out are gone.
9743 * If after this point a bound task is being woken on this CPU then the
9744 * responsible hotplug callback has failed to do it's job.
9745 * sched_cpu_dying() will catch it with the appropriate fireworks.
9747 int sched_cpu_wait_empty(unsigned int cpu
)
9749 balance_hotplug_wait();
9754 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9755 * might have. Called from the CPU stopper task after ensuring that the
9756 * stopper is the last running task on the CPU, so nr_active count is
9757 * stable. We need to take the teardown thread which is calling this into
9758 * account, so we hand in adjust = 1 to the load calculation.
9760 * Also see the comment "Global load-average calculations".
9762 static void calc_load_migrate(struct rq
*rq
)
9764 long delta
= calc_load_fold_active(rq
, 1);
9767 atomic_long_add(delta
, &calc_load_tasks
);
9770 static void dump_rq_tasks(struct rq
*rq
, const char *loglvl
)
9772 struct task_struct
*g
, *p
;
9773 int cpu
= cpu_of(rq
);
9775 lockdep_assert_rq_held(rq
);
9777 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl
, cpu
, rq
->nr_running
);
9778 for_each_process_thread(g
, p
) {
9779 if (task_cpu(p
) != cpu
)
9782 if (!task_on_rq_queued(p
))
9785 printk("%s\tpid: %d, name: %s\n", loglvl
, p
->pid
, p
->comm
);
9789 int sched_cpu_dying(unsigned int cpu
)
9791 struct rq
*rq
= cpu_rq(cpu
);
9794 /* Handle pending wakeups and then migrate everything off */
9795 sched_tick_stop(cpu
);
9797 rq_lock_irqsave(rq
, &rf
);
9798 if (rq
->nr_running
!= 1 || rq_has_pinned_tasks(rq
)) {
9799 WARN(true, "Dying CPU not properly vacated!");
9800 dump_rq_tasks(rq
, KERN_WARNING
);
9802 rq_unlock_irqrestore(rq
, &rf
);
9804 calc_load_migrate(rq
);
9805 update_max_interval();
9807 sched_core_cpu_dying(cpu
);
9812 void __init
sched_init_smp(void)
9814 sched_init_numa(NUMA_NO_NODE
);
9817 * There's no userspace yet to cause hotplug operations; hence all the
9818 * CPU masks are stable and all blatant races in the below code cannot
9821 mutex_lock(&sched_domains_mutex
);
9822 sched_init_domains(cpu_active_mask
);
9823 mutex_unlock(&sched_domains_mutex
);
9825 /* Move init over to a non-isolated CPU */
9826 if (set_cpus_allowed_ptr(current
, housekeeping_cpumask(HK_TYPE_DOMAIN
)) < 0)
9828 current
->flags
&= ~PF_NO_SETAFFINITY
;
9829 sched_init_granularity();
9831 init_sched_rt_class();
9832 init_sched_dl_class();
9834 sched_smp_initialized
= true;
9837 static int __init
migration_init(void)
9839 sched_cpu_starting(smp_processor_id());
9842 early_initcall(migration_init
);
9845 void __init
sched_init_smp(void)
9847 sched_init_granularity();
9849 #endif /* CONFIG_SMP */
9851 int in_sched_functions(unsigned long addr
)
9853 return in_lock_functions(addr
) ||
9854 (addr
>= (unsigned long)__sched_text_start
9855 && addr
< (unsigned long)__sched_text_end
);
9858 #ifdef CONFIG_CGROUP_SCHED
9860 * Default task group.
9861 * Every task in system belongs to this group at bootup.
9863 struct task_group root_task_group
;
9864 LIST_HEAD(task_groups
);
9866 /* Cacheline aligned slab cache for task_group */
9867 static struct kmem_cache
*task_group_cache __read_mostly
;
9870 void __init
sched_init(void)
9872 unsigned long ptr
= 0;
9875 /* Make sure the linker didn't screw up */
9876 BUG_ON(&idle_sched_class
!= &fair_sched_class
+ 1 ||
9877 &fair_sched_class
!= &rt_sched_class
+ 1 ||
9878 &rt_sched_class
!= &dl_sched_class
+ 1);
9880 BUG_ON(&dl_sched_class
!= &stop_sched_class
+ 1);
9885 #ifdef CONFIG_FAIR_GROUP_SCHED
9886 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
9888 #ifdef CONFIG_RT_GROUP_SCHED
9889 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
9892 ptr
= (unsigned long)kzalloc(ptr
, GFP_NOWAIT
);
9894 #ifdef CONFIG_FAIR_GROUP_SCHED
9895 root_task_group
.se
= (struct sched_entity
**)ptr
;
9896 ptr
+= nr_cpu_ids
* sizeof(void **);
9898 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9899 ptr
+= nr_cpu_ids
* sizeof(void **);
9901 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
9902 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
9903 #endif /* CONFIG_FAIR_GROUP_SCHED */
9904 #ifdef CONFIG_RT_GROUP_SCHED
9905 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9906 ptr
+= nr_cpu_ids
* sizeof(void **);
9908 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9909 ptr
+= nr_cpu_ids
* sizeof(void **);
9911 #endif /* CONFIG_RT_GROUP_SCHED */
9914 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
9917 init_defrootdomain();
9920 #ifdef CONFIG_RT_GROUP_SCHED
9921 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9922 global_rt_period(), global_rt_runtime());
9923 #endif /* CONFIG_RT_GROUP_SCHED */
9925 #ifdef CONFIG_CGROUP_SCHED
9926 task_group_cache
= KMEM_CACHE(task_group
, 0);
9928 list_add(&root_task_group
.list
, &task_groups
);
9929 INIT_LIST_HEAD(&root_task_group
.children
);
9930 INIT_LIST_HEAD(&root_task_group
.siblings
);
9931 autogroup_init(&init_task
);
9932 #endif /* CONFIG_CGROUP_SCHED */
9934 for_each_possible_cpu(i
) {
9938 raw_spin_lock_init(&rq
->__lock
);
9940 rq
->calc_load_active
= 0;
9941 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9942 init_cfs_rq(&rq
->cfs
);
9943 init_rt_rq(&rq
->rt
);
9944 init_dl_rq(&rq
->dl
);
9945 #ifdef CONFIG_FAIR_GROUP_SCHED
9946 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9947 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
9949 * How much CPU bandwidth does root_task_group get?
9951 * In case of task-groups formed thr' the cgroup filesystem, it
9952 * gets 100% of the CPU resources in the system. This overall
9953 * system CPU resource is divided among the tasks of
9954 * root_task_group and its child task-groups in a fair manner,
9955 * based on each entity's (task or task-group's) weight
9956 * (se->load.weight).
9958 * In other words, if root_task_group has 10 tasks of weight
9959 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9960 * then A0's share of the CPU resource is:
9962 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9964 * We achieve this by letting root_task_group's tasks sit
9965 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9967 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
9968 #endif /* CONFIG_FAIR_GROUP_SCHED */
9970 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9971 #ifdef CONFIG_RT_GROUP_SCHED
9972 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
9977 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
9978 rq
->balance_callback
= &balance_push_callback
;
9979 rq
->active_balance
= 0;
9980 rq
->next_balance
= jiffies
;
9985 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
9986 rq
->wake_stamp
= jiffies
;
9987 rq
->wake_avg_idle
= rq
->avg_idle
;
9988 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
9990 INIT_LIST_HEAD(&rq
->cfs_tasks
);
9992 rq_attach_root(rq
, &def_root_domain
);
9993 #ifdef CONFIG_NO_HZ_COMMON
9994 rq
->last_blocked_load_update_tick
= jiffies
;
9995 atomic_set(&rq
->nohz_flags
, 0);
9997 INIT_CSD(&rq
->nohz_csd
, nohz_csd_func
, rq
);
9999 #ifdef CONFIG_HOTPLUG_CPU
10000 rcuwait_init(&rq
->hotplug_wait
);
10002 #endif /* CONFIG_SMP */
10003 hrtick_rq_init(rq
);
10004 atomic_set(&rq
->nr_iowait
, 0);
10006 #ifdef CONFIG_SCHED_CORE
10008 rq
->core_pick
= NULL
;
10009 rq
->core_enabled
= 0;
10010 rq
->core_tree
= RB_ROOT
;
10011 rq
->core_forceidle_count
= 0;
10012 rq
->core_forceidle_occupation
= 0;
10013 rq
->core_forceidle_start
= 0;
10015 rq
->core_cookie
= 0UL;
10017 zalloc_cpumask_var_node(&rq
->scratch_mask
, GFP_KERNEL
, cpu_to_node(i
));
10020 set_load_weight(&init_task
, false);
10023 * The boot idle thread does lazy MMU switching as well:
10025 mmgrab_lazy_tlb(&init_mm
);
10026 enter_lazy_tlb(&init_mm
, current
);
10029 * The idle task doesn't need the kthread struct to function, but it
10030 * is dressed up as a per-CPU kthread and thus needs to play the part
10031 * if we want to avoid special-casing it in code that deals with per-CPU
10034 WARN_ON(!set_kthread_struct(current
));
10037 * Make us the idle thread. Technically, schedule() should not be
10038 * called from this thread, however somewhere below it might be,
10039 * but because we are the idle thread, we just pick up running again
10040 * when this runqueue becomes "idle".
10042 init_idle(current
, smp_processor_id());
10044 calc_load_update
= jiffies
+ LOAD_FREQ
;
10047 idle_thread_set_boot_cpu();
10048 balance_push_set(smp_processor_id(), false);
10050 init_sched_fair_class();
10056 preempt_dynamic_init();
10058 scheduler_running
= 1;
10061 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
10063 void __might_sleep(const char *file
, int line
)
10065 unsigned int state
= get_current_state();
10067 * Blocking primitives will set (and therefore destroy) current->state,
10068 * since we will exit with TASK_RUNNING make sure we enter with it,
10069 * otherwise we will destroy state.
10071 WARN_ONCE(state
!= TASK_RUNNING
&& current
->task_state_change
,
10072 "do not call blocking ops when !TASK_RUNNING; "
10073 "state=%x set at [<%p>] %pS\n", state
,
10074 (void *)current
->task_state_change
,
10075 (void *)current
->task_state_change
);
10077 __might_resched(file
, line
, 0);
10079 EXPORT_SYMBOL(__might_sleep
);
10081 static void print_preempt_disable_ip(int preempt_offset
, unsigned long ip
)
10083 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT
))
10086 if (preempt_count() == preempt_offset
)
10089 pr_err("Preemption disabled at:");
10090 print_ip_sym(KERN_ERR
, ip
);
10093 static inline bool resched_offsets_ok(unsigned int offsets
)
10095 unsigned int nested
= preempt_count();
10097 nested
+= rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT
;
10099 return nested
== offsets
;
10102 void __might_resched(const char *file
, int line
, unsigned int offsets
)
10104 /* Ratelimiting timestamp: */
10105 static unsigned long prev_jiffy
;
10107 unsigned long preempt_disable_ip
;
10109 /* WARN_ON_ONCE() by default, no rate limit required: */
10112 if ((resched_offsets_ok(offsets
) && !irqs_disabled() &&
10113 !is_idle_task(current
) && !current
->non_block_count
) ||
10114 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
10118 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
10120 prev_jiffy
= jiffies
;
10122 /* Save this before calling printk(), since that will clobber it: */
10123 preempt_disable_ip
= get_preempt_disable_ip(current
);
10125 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10127 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10128 in_atomic(), irqs_disabled(), current
->non_block_count
,
10129 current
->pid
, current
->comm
);
10130 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
10131 offsets
& MIGHT_RESCHED_PREEMPT_MASK
);
10133 if (IS_ENABLED(CONFIG_PREEMPT_RCU
)) {
10134 pr_err("RCU nest depth: %d, expected: %u\n",
10135 rcu_preempt_depth(), offsets
>> MIGHT_RESCHED_RCU_SHIFT
);
10138 if (task_stack_end_corrupted(current
))
10139 pr_emerg("Thread overran stack, or stack corrupted\n");
10141 debug_show_held_locks(current
);
10142 if (irqs_disabled())
10143 print_irqtrace_events(current
);
10145 print_preempt_disable_ip(offsets
& MIGHT_RESCHED_PREEMPT_MASK
,
10146 preempt_disable_ip
);
10149 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
10151 EXPORT_SYMBOL(__might_resched
);
10153 void __cant_sleep(const char *file
, int line
, int preempt_offset
)
10155 static unsigned long prev_jiffy
;
10157 if (irqs_disabled())
10160 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
10163 if (preempt_count() > preempt_offset
)
10166 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
10168 prev_jiffy
= jiffies
;
10170 printk(KERN_ERR
"BUG: assuming atomic context at %s:%d\n", file
, line
);
10171 printk(KERN_ERR
"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10172 in_atomic(), irqs_disabled(),
10173 current
->pid
, current
->comm
);
10175 debug_show_held_locks(current
);
10177 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
10179 EXPORT_SYMBOL_GPL(__cant_sleep
);
10182 void __cant_migrate(const char *file
, int line
)
10184 static unsigned long prev_jiffy
;
10186 if (irqs_disabled())
10189 if (is_migration_disabled(current
))
10192 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
10195 if (preempt_count() > 0)
10198 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
10200 prev_jiffy
= jiffies
;
10202 pr_err("BUG: assuming non migratable context at %s:%d\n", file
, line
);
10203 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10204 in_atomic(), irqs_disabled(), is_migration_disabled(current
),
10205 current
->pid
, current
->comm
);
10207 debug_show_held_locks(current
);
10209 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
10211 EXPORT_SYMBOL_GPL(__cant_migrate
);
10215 #ifdef CONFIG_MAGIC_SYSRQ
10216 void normalize_rt_tasks(void)
10218 struct task_struct
*g
, *p
;
10219 struct sched_attr attr
= {
10220 .sched_policy
= SCHED_NORMAL
,
10223 read_lock(&tasklist_lock
);
10224 for_each_process_thread(g
, p
) {
10226 * Only normalize user tasks:
10228 if (p
->flags
& PF_KTHREAD
)
10231 p
->se
.exec_start
= 0;
10232 schedstat_set(p
->stats
.wait_start
, 0);
10233 schedstat_set(p
->stats
.sleep_start
, 0);
10234 schedstat_set(p
->stats
.block_start
, 0);
10236 if (!dl_task(p
) && !rt_task(p
)) {
10238 * Renice negative nice level userspace
10241 if (task_nice(p
) < 0)
10242 set_user_nice(p
, 0);
10246 __sched_setscheduler(p
, &attr
, false, false);
10248 read_unlock(&tasklist_lock
);
10251 #endif /* CONFIG_MAGIC_SYSRQ */
10253 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
10255 * These functions are only useful for the IA64 MCA handling, or kdb.
10257 * They can only be called when the whole system has been
10258 * stopped - every CPU needs to be quiescent, and no scheduling
10259 * activity can take place. Using them for anything else would
10260 * be a serious bug, and as a result, they aren't even visible
10261 * under any other configuration.
10265 * curr_task - return the current task for a given CPU.
10266 * @cpu: the processor in question.
10268 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10270 * Return: The current task for @cpu.
10272 struct task_struct
*curr_task(int cpu
)
10274 return cpu_curr(cpu
);
10277 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
10281 * ia64_set_curr_task - set the current task for a given CPU.
10282 * @cpu: the processor in question.
10283 * @p: the task pointer to set.
10285 * Description: This function must only be used when non-maskable interrupts
10286 * are serviced on a separate stack. It allows the architecture to switch the
10287 * notion of the current task on a CPU in a non-blocking manner. This function
10288 * must be called with all CPU's synchronized, and interrupts disabled, the
10289 * and caller must save the original value of the current task (see
10290 * curr_task() above) and restore that value before reenabling interrupts and
10291 * re-starting the system.
10293 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10295 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
10302 #ifdef CONFIG_CGROUP_SCHED
10303 /* task_group_lock serializes the addition/removal of task groups */
10304 static DEFINE_SPINLOCK(task_group_lock
);
10306 static inline void alloc_uclamp_sched_group(struct task_group
*tg
,
10307 struct task_group
*parent
)
10309 #ifdef CONFIG_UCLAMP_TASK_GROUP
10310 enum uclamp_id clamp_id
;
10312 for_each_clamp_id(clamp_id
) {
10313 uclamp_se_set(&tg
->uclamp_req
[clamp_id
],
10314 uclamp_none(clamp_id
), false);
10315 tg
->uclamp
[clamp_id
] = parent
->uclamp
[clamp_id
];
10320 static void sched_free_group(struct task_group
*tg
)
10322 free_fair_sched_group(tg
);
10323 free_rt_sched_group(tg
);
10324 autogroup_free(tg
);
10325 kmem_cache_free(task_group_cache
, tg
);
10328 static void sched_free_group_rcu(struct rcu_head
*rcu
)
10330 sched_free_group(container_of(rcu
, struct task_group
, rcu
));
10333 static void sched_unregister_group(struct task_group
*tg
)
10335 unregister_fair_sched_group(tg
);
10336 unregister_rt_sched_group(tg
);
10338 * We have to wait for yet another RCU grace period to expire, as
10339 * print_cfs_stats() might run concurrently.
10341 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
10344 /* allocate runqueue etc for a new task group */
10345 struct task_group
*sched_create_group(struct task_group
*parent
)
10347 struct task_group
*tg
;
10349 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
10351 return ERR_PTR(-ENOMEM
);
10353 if (!alloc_fair_sched_group(tg
, parent
))
10356 if (!alloc_rt_sched_group(tg
, parent
))
10359 alloc_uclamp_sched_group(tg
, parent
);
10364 sched_free_group(tg
);
10365 return ERR_PTR(-ENOMEM
);
10368 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
10370 unsigned long flags
;
10372 spin_lock_irqsave(&task_group_lock
, flags
);
10373 list_add_rcu(&tg
->list
, &task_groups
);
10375 /* Root should already exist: */
10378 tg
->parent
= parent
;
10379 INIT_LIST_HEAD(&tg
->children
);
10380 list_add_rcu(&tg
->siblings
, &parent
->children
);
10381 spin_unlock_irqrestore(&task_group_lock
, flags
);
10383 online_fair_sched_group(tg
);
10386 /* rcu callback to free various structures associated with a task group */
10387 static void sched_unregister_group_rcu(struct rcu_head
*rhp
)
10389 /* Now it should be safe to free those cfs_rqs: */
10390 sched_unregister_group(container_of(rhp
, struct task_group
, rcu
));
10393 void sched_destroy_group(struct task_group
*tg
)
10395 /* Wait for possible concurrent references to cfs_rqs complete: */
10396 call_rcu(&tg
->rcu
, sched_unregister_group_rcu
);
10399 void sched_release_group(struct task_group
*tg
)
10401 unsigned long flags
;
10404 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10405 * sched_cfs_period_timer()).
10407 * For this to be effective, we have to wait for all pending users of
10408 * this task group to leave their RCU critical section to ensure no new
10409 * user will see our dying task group any more. Specifically ensure
10410 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10412 * We therefore defer calling unregister_fair_sched_group() to
10413 * sched_unregister_group() which is guarantied to get called only after the
10414 * current RCU grace period has expired.
10416 spin_lock_irqsave(&task_group_lock
, flags
);
10417 list_del_rcu(&tg
->list
);
10418 list_del_rcu(&tg
->siblings
);
10419 spin_unlock_irqrestore(&task_group_lock
, flags
);
10422 static struct task_group
*sched_get_task_group(struct task_struct
*tsk
)
10424 struct task_group
*tg
;
10427 * All callers are synchronized by task_rq_lock(); we do not use RCU
10428 * which is pointless here. Thus, we pass "true" to task_css_check()
10429 * to prevent lockdep warnings.
10431 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
10432 struct task_group
, css
);
10433 tg
= autogroup_task_group(tsk
, tg
);
10438 static void sched_change_group(struct task_struct
*tsk
, struct task_group
*group
)
10440 tsk
->sched_task_group
= group
;
10442 #ifdef CONFIG_FAIR_GROUP_SCHED
10443 if (tsk
->sched_class
->task_change_group
)
10444 tsk
->sched_class
->task_change_group(tsk
);
10447 set_task_rq(tsk
, task_cpu(tsk
));
10451 * Change task's runqueue when it moves between groups.
10453 * The caller of this function should have put the task in its new group by
10454 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10457 void sched_move_task(struct task_struct
*tsk
)
10459 int queued
, running
, queue_flags
=
10460 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
10461 struct task_group
*group
;
10462 struct rq_flags rf
;
10465 rq
= task_rq_lock(tsk
, &rf
);
10467 * Esp. with SCHED_AUTOGROUP enabled it is possible to get superfluous
10470 group
= sched_get_task_group(tsk
);
10471 if (group
== tsk
->sched_task_group
)
10474 update_rq_clock(rq
);
10476 running
= task_current(rq
, tsk
);
10477 queued
= task_on_rq_queued(tsk
);
10480 dequeue_task(rq
, tsk
, queue_flags
);
10482 put_prev_task(rq
, tsk
);
10484 sched_change_group(tsk
, group
);
10487 enqueue_task(rq
, tsk
, queue_flags
);
10489 set_next_task(rq
, tsk
);
10491 * After changing group, the running task may have joined a
10492 * throttled one but it's still the running task. Trigger a
10493 * resched to make sure that task can still run.
10499 task_rq_unlock(rq
, tsk
, &rf
);
10502 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
10504 return css
? container_of(css
, struct task_group
, css
) : NULL
;
10507 static struct cgroup_subsys_state
*
10508 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
10510 struct task_group
*parent
= css_tg(parent_css
);
10511 struct task_group
*tg
;
10514 /* This is early initialization for the top cgroup */
10515 return &root_task_group
.css
;
10518 tg
= sched_create_group(parent
);
10520 return ERR_PTR(-ENOMEM
);
10525 /* Expose task group only after completing cgroup initialization */
10526 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
10528 struct task_group
*tg
= css_tg(css
);
10529 struct task_group
*parent
= css_tg(css
->parent
);
10532 sched_online_group(tg
, parent
);
10534 #ifdef CONFIG_UCLAMP_TASK_GROUP
10535 /* Propagate the effective uclamp value for the new group */
10536 mutex_lock(&uclamp_mutex
);
10538 cpu_util_update_eff(css
);
10540 mutex_unlock(&uclamp_mutex
);
10546 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
10548 struct task_group
*tg
= css_tg(css
);
10550 sched_release_group(tg
);
10553 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
10555 struct task_group
*tg
= css_tg(css
);
10558 * Relies on the RCU grace period between css_released() and this.
10560 sched_unregister_group(tg
);
10563 #ifdef CONFIG_RT_GROUP_SCHED
10564 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
10566 struct task_struct
*task
;
10567 struct cgroup_subsys_state
*css
;
10569 cgroup_taskset_for_each(task
, css
, tset
) {
10570 if (!sched_rt_can_attach(css_tg(css
), task
))
10577 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
10579 struct task_struct
*task
;
10580 struct cgroup_subsys_state
*css
;
10582 cgroup_taskset_for_each(task
, css
, tset
)
10583 sched_move_task(task
);
10586 #ifdef CONFIG_UCLAMP_TASK_GROUP
10587 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
)
10589 struct cgroup_subsys_state
*top_css
= css
;
10590 struct uclamp_se
*uc_parent
= NULL
;
10591 struct uclamp_se
*uc_se
= NULL
;
10592 unsigned int eff
[UCLAMP_CNT
];
10593 enum uclamp_id clamp_id
;
10594 unsigned int clamps
;
10596 lockdep_assert_held(&uclamp_mutex
);
10597 SCHED_WARN_ON(!rcu_read_lock_held());
10599 css_for_each_descendant_pre(css
, top_css
) {
10600 uc_parent
= css_tg(css
)->parent
10601 ? css_tg(css
)->parent
->uclamp
: NULL
;
10603 for_each_clamp_id(clamp_id
) {
10604 /* Assume effective clamps matches requested clamps */
10605 eff
[clamp_id
] = css_tg(css
)->uclamp_req
[clamp_id
].value
;
10606 /* Cap effective clamps with parent's effective clamps */
10608 eff
[clamp_id
] > uc_parent
[clamp_id
].value
) {
10609 eff
[clamp_id
] = uc_parent
[clamp_id
].value
;
10612 /* Ensure protection is always capped by limit */
10613 eff
[UCLAMP_MIN
] = min(eff
[UCLAMP_MIN
], eff
[UCLAMP_MAX
]);
10615 /* Propagate most restrictive effective clamps */
10617 uc_se
= css_tg(css
)->uclamp
;
10618 for_each_clamp_id(clamp_id
) {
10619 if (eff
[clamp_id
] == uc_se
[clamp_id
].value
)
10621 uc_se
[clamp_id
].value
= eff
[clamp_id
];
10622 uc_se
[clamp_id
].bucket_id
= uclamp_bucket_id(eff
[clamp_id
]);
10623 clamps
|= (0x1 << clamp_id
);
10626 css
= css_rightmost_descendant(css
);
10630 /* Immediately update descendants RUNNABLE tasks */
10631 uclamp_update_active_tasks(css
);
10636 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10637 * C expression. Since there is no way to convert a macro argument (N) into a
10638 * character constant, use two levels of macros.
10640 #define _POW10(exp) ((unsigned int)1e##exp)
10641 #define POW10(exp) _POW10(exp)
10643 struct uclamp_request
{
10644 #define UCLAMP_PERCENT_SHIFT 2
10645 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10651 static inline struct uclamp_request
10652 capacity_from_percent(char *buf
)
10654 struct uclamp_request req
= {
10655 .percent
= UCLAMP_PERCENT_SCALE
,
10656 .util
= SCHED_CAPACITY_SCALE
,
10661 if (strcmp(buf
, "max")) {
10662 req
.ret
= cgroup_parse_float(buf
, UCLAMP_PERCENT_SHIFT
,
10666 if ((u64
)req
.percent
> UCLAMP_PERCENT_SCALE
) {
10671 req
.util
= req
.percent
<< SCHED_CAPACITY_SHIFT
;
10672 req
.util
= DIV_ROUND_CLOSEST_ULL(req
.util
, UCLAMP_PERCENT_SCALE
);
10678 static ssize_t
cpu_uclamp_write(struct kernfs_open_file
*of
, char *buf
,
10679 size_t nbytes
, loff_t off
,
10680 enum uclamp_id clamp_id
)
10682 struct uclamp_request req
;
10683 struct task_group
*tg
;
10685 req
= capacity_from_percent(buf
);
10689 static_branch_enable(&sched_uclamp_used
);
10691 mutex_lock(&uclamp_mutex
);
10694 tg
= css_tg(of_css(of
));
10695 if (tg
->uclamp_req
[clamp_id
].value
!= req
.util
)
10696 uclamp_se_set(&tg
->uclamp_req
[clamp_id
], req
.util
, false);
10699 * Because of not recoverable conversion rounding we keep track of the
10700 * exact requested value
10702 tg
->uclamp_pct
[clamp_id
] = req
.percent
;
10704 /* Update effective clamps to track the most restrictive value */
10705 cpu_util_update_eff(of_css(of
));
10708 mutex_unlock(&uclamp_mutex
);
10713 static ssize_t
cpu_uclamp_min_write(struct kernfs_open_file
*of
,
10714 char *buf
, size_t nbytes
,
10717 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MIN
);
10720 static ssize_t
cpu_uclamp_max_write(struct kernfs_open_file
*of
,
10721 char *buf
, size_t nbytes
,
10724 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MAX
);
10727 static inline void cpu_uclamp_print(struct seq_file
*sf
,
10728 enum uclamp_id clamp_id
)
10730 struct task_group
*tg
;
10736 tg
= css_tg(seq_css(sf
));
10737 util_clamp
= tg
->uclamp_req
[clamp_id
].value
;
10740 if (util_clamp
== SCHED_CAPACITY_SCALE
) {
10741 seq_puts(sf
, "max\n");
10745 percent
= tg
->uclamp_pct
[clamp_id
];
10746 percent
= div_u64_rem(percent
, POW10(UCLAMP_PERCENT_SHIFT
), &rem
);
10747 seq_printf(sf
, "%llu.%0*u\n", percent
, UCLAMP_PERCENT_SHIFT
, rem
);
10750 static int cpu_uclamp_min_show(struct seq_file
*sf
, void *v
)
10752 cpu_uclamp_print(sf
, UCLAMP_MIN
);
10756 static int cpu_uclamp_max_show(struct seq_file
*sf
, void *v
)
10758 cpu_uclamp_print(sf
, UCLAMP_MAX
);
10761 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10763 #ifdef CONFIG_FAIR_GROUP_SCHED
10764 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
10765 struct cftype
*cftype
, u64 shareval
)
10767 if (shareval
> scale_load_down(ULONG_MAX
))
10768 shareval
= MAX_SHARES
;
10769 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
10772 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
10773 struct cftype
*cft
)
10775 struct task_group
*tg
= css_tg(css
);
10777 return (u64
) scale_load_down(tg
->shares
);
10780 #ifdef CONFIG_CFS_BANDWIDTH
10781 static DEFINE_MUTEX(cfs_constraints_mutex
);
10783 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
10784 static const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
10785 /* More than 203 days if BW_SHIFT equals 20. */
10786 static const u64 max_cfs_runtime
= MAX_BW
* NSEC_PER_USEC
;
10788 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
10790 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
,
10793 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
10794 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
10796 if (tg
== &root_task_group
)
10800 * Ensure we have at some amount of bandwidth every period. This is
10801 * to prevent reaching a state of large arrears when throttled via
10802 * entity_tick() resulting in prolonged exit starvation.
10804 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
10808 * Likewise, bound things on the other side by preventing insane quota
10809 * periods. This also allows us to normalize in computing quota
10812 if (period
> max_cfs_quota_period
)
10816 * Bound quota to defend quota against overflow during bandwidth shift.
10818 if (quota
!= RUNTIME_INF
&& quota
> max_cfs_runtime
)
10821 if (quota
!= RUNTIME_INF
&& (burst
> quota
||
10822 burst
+ quota
> max_cfs_runtime
))
10826 * Prevent race between setting of cfs_rq->runtime_enabled and
10827 * unthrottle_offline_cfs_rqs().
10830 mutex_lock(&cfs_constraints_mutex
);
10831 ret
= __cfs_schedulable(tg
, period
, quota
);
10835 runtime_enabled
= quota
!= RUNTIME_INF
;
10836 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
10838 * If we need to toggle cfs_bandwidth_used, off->on must occur
10839 * before making related changes, and on->off must occur afterwards
10841 if (runtime_enabled
&& !runtime_was_enabled
)
10842 cfs_bandwidth_usage_inc();
10843 raw_spin_lock_irq(&cfs_b
->lock
);
10844 cfs_b
->period
= ns_to_ktime(period
);
10845 cfs_b
->quota
= quota
;
10846 cfs_b
->burst
= burst
;
10848 __refill_cfs_bandwidth_runtime(cfs_b
);
10850 /* Restart the period timer (if active) to handle new period expiry: */
10851 if (runtime_enabled
)
10852 start_cfs_bandwidth(cfs_b
);
10854 raw_spin_unlock_irq(&cfs_b
->lock
);
10856 for_each_online_cpu(i
) {
10857 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
10858 struct rq
*rq
= cfs_rq
->rq
;
10859 struct rq_flags rf
;
10861 rq_lock_irq(rq
, &rf
);
10862 cfs_rq
->runtime_enabled
= runtime_enabled
;
10863 cfs_rq
->runtime_remaining
= 0;
10865 if (cfs_rq
->throttled
)
10866 unthrottle_cfs_rq(cfs_rq
);
10867 rq_unlock_irq(rq
, &rf
);
10869 if (runtime_was_enabled
&& !runtime_enabled
)
10870 cfs_bandwidth_usage_dec();
10872 mutex_unlock(&cfs_constraints_mutex
);
10873 cpus_read_unlock();
10878 static int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
10880 u64 quota
, period
, burst
;
10882 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
10883 burst
= tg
->cfs_bandwidth
.burst
;
10884 if (cfs_quota_us
< 0)
10885 quota
= RUNTIME_INF
;
10886 else if ((u64
)cfs_quota_us
<= U64_MAX
/ NSEC_PER_USEC
)
10887 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
10891 return tg_set_cfs_bandwidth(tg
, period
, quota
, burst
);
10894 static long tg_get_cfs_quota(struct task_group
*tg
)
10898 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
10901 quota_us
= tg
->cfs_bandwidth
.quota
;
10902 do_div(quota_us
, NSEC_PER_USEC
);
10907 static int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
10909 u64 quota
, period
, burst
;
10911 if ((u64
)cfs_period_us
> U64_MAX
/ NSEC_PER_USEC
)
10914 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
10915 quota
= tg
->cfs_bandwidth
.quota
;
10916 burst
= tg
->cfs_bandwidth
.burst
;
10918 return tg_set_cfs_bandwidth(tg
, period
, quota
, burst
);
10921 static long tg_get_cfs_period(struct task_group
*tg
)
10925 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
10926 do_div(cfs_period_us
, NSEC_PER_USEC
);
10928 return cfs_period_us
;
10931 static int tg_set_cfs_burst(struct task_group
*tg
, long cfs_burst_us
)
10933 u64 quota
, period
, burst
;
10935 if ((u64
)cfs_burst_us
> U64_MAX
/ NSEC_PER_USEC
)
10938 burst
= (u64
)cfs_burst_us
* NSEC_PER_USEC
;
10939 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
10940 quota
= tg
->cfs_bandwidth
.quota
;
10942 return tg_set_cfs_bandwidth(tg
, period
, quota
, burst
);
10945 static long tg_get_cfs_burst(struct task_group
*tg
)
10949 burst_us
= tg
->cfs_bandwidth
.burst
;
10950 do_div(burst_us
, NSEC_PER_USEC
);
10955 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
10956 struct cftype
*cft
)
10958 return tg_get_cfs_quota(css_tg(css
));
10961 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
10962 struct cftype
*cftype
, s64 cfs_quota_us
)
10964 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
10967 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
10968 struct cftype
*cft
)
10970 return tg_get_cfs_period(css_tg(css
));
10973 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
10974 struct cftype
*cftype
, u64 cfs_period_us
)
10976 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
10979 static u64
cpu_cfs_burst_read_u64(struct cgroup_subsys_state
*css
,
10980 struct cftype
*cft
)
10982 return tg_get_cfs_burst(css_tg(css
));
10985 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state
*css
,
10986 struct cftype
*cftype
, u64 cfs_burst_us
)
10988 return tg_set_cfs_burst(css_tg(css
), cfs_burst_us
);
10991 struct cfs_schedulable_data
{
10992 struct task_group
*tg
;
10997 * normalize group quota/period to be quota/max_period
10998 * note: units are usecs
11000 static u64
normalize_cfs_quota(struct task_group
*tg
,
11001 struct cfs_schedulable_data
*d
)
11006 period
= d
->period
;
11009 period
= tg_get_cfs_period(tg
);
11010 quota
= tg_get_cfs_quota(tg
);
11013 /* note: these should typically be equivalent */
11014 if (quota
== RUNTIME_INF
|| quota
== -1)
11015 return RUNTIME_INF
;
11017 return to_ratio(period
, quota
);
11020 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
11022 struct cfs_schedulable_data
*d
= data
;
11023 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
11024 s64 quota
= 0, parent_quota
= -1;
11027 quota
= RUNTIME_INF
;
11029 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
11031 quota
= normalize_cfs_quota(tg
, d
);
11032 parent_quota
= parent_b
->hierarchical_quota
;
11035 * Ensure max(child_quota) <= parent_quota. On cgroup2,
11036 * always take the min. On cgroup1, only inherit when no
11039 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys
)) {
11040 quota
= min(quota
, parent_quota
);
11042 if (quota
== RUNTIME_INF
)
11043 quota
= parent_quota
;
11044 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
11048 cfs_b
->hierarchical_quota
= quota
;
11053 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
11056 struct cfs_schedulable_data data
= {
11062 if (quota
!= RUNTIME_INF
) {
11063 do_div(data
.period
, NSEC_PER_USEC
);
11064 do_div(data
.quota
, NSEC_PER_USEC
);
11068 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
11074 static int cpu_cfs_stat_show(struct seq_file
*sf
, void *v
)
11076 struct task_group
*tg
= css_tg(seq_css(sf
));
11077 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
11079 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
11080 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
11081 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
11083 if (schedstat_enabled() && tg
!= &root_task_group
) {
11084 struct sched_statistics
*stats
;
11088 for_each_possible_cpu(i
) {
11089 stats
= __schedstats_from_se(tg
->se
[i
]);
11090 ws
+= schedstat_val(stats
->wait_sum
);
11093 seq_printf(sf
, "wait_sum %llu\n", ws
);
11096 seq_printf(sf
, "nr_bursts %d\n", cfs_b
->nr_burst
);
11097 seq_printf(sf
, "burst_time %llu\n", cfs_b
->burst_time
);
11101 #endif /* CONFIG_CFS_BANDWIDTH */
11102 #endif /* CONFIG_FAIR_GROUP_SCHED */
11104 #ifdef CONFIG_RT_GROUP_SCHED
11105 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
11106 struct cftype
*cft
, s64 val
)
11108 return sched_group_set_rt_runtime(css_tg(css
), val
);
11111 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
11112 struct cftype
*cft
)
11114 return sched_group_rt_runtime(css_tg(css
));
11117 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
11118 struct cftype
*cftype
, u64 rt_period_us
)
11120 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
11123 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
11124 struct cftype
*cft
)
11126 return sched_group_rt_period(css_tg(css
));
11128 #endif /* CONFIG_RT_GROUP_SCHED */
11130 #ifdef CONFIG_FAIR_GROUP_SCHED
11131 static s64
cpu_idle_read_s64(struct cgroup_subsys_state
*css
,
11132 struct cftype
*cft
)
11134 return css_tg(css
)->idle
;
11137 static int cpu_idle_write_s64(struct cgroup_subsys_state
*css
,
11138 struct cftype
*cft
, s64 idle
)
11140 return sched_group_set_idle(css_tg(css
), idle
);
11144 static struct cftype cpu_legacy_files
[] = {
11145 #ifdef CONFIG_FAIR_GROUP_SCHED
11148 .read_u64
= cpu_shares_read_u64
,
11149 .write_u64
= cpu_shares_write_u64
,
11153 .read_s64
= cpu_idle_read_s64
,
11154 .write_s64
= cpu_idle_write_s64
,
11157 #ifdef CONFIG_CFS_BANDWIDTH
11159 .name
= "cfs_quota_us",
11160 .read_s64
= cpu_cfs_quota_read_s64
,
11161 .write_s64
= cpu_cfs_quota_write_s64
,
11164 .name
= "cfs_period_us",
11165 .read_u64
= cpu_cfs_period_read_u64
,
11166 .write_u64
= cpu_cfs_period_write_u64
,
11169 .name
= "cfs_burst_us",
11170 .read_u64
= cpu_cfs_burst_read_u64
,
11171 .write_u64
= cpu_cfs_burst_write_u64
,
11175 .seq_show
= cpu_cfs_stat_show
,
11178 #ifdef CONFIG_RT_GROUP_SCHED
11180 .name
= "rt_runtime_us",
11181 .read_s64
= cpu_rt_runtime_read
,
11182 .write_s64
= cpu_rt_runtime_write
,
11185 .name
= "rt_period_us",
11186 .read_u64
= cpu_rt_period_read_uint
,
11187 .write_u64
= cpu_rt_period_write_uint
,
11190 #ifdef CONFIG_UCLAMP_TASK_GROUP
11192 .name
= "uclamp.min",
11193 .flags
= CFTYPE_NOT_ON_ROOT
,
11194 .seq_show
= cpu_uclamp_min_show
,
11195 .write
= cpu_uclamp_min_write
,
11198 .name
= "uclamp.max",
11199 .flags
= CFTYPE_NOT_ON_ROOT
,
11200 .seq_show
= cpu_uclamp_max_show
,
11201 .write
= cpu_uclamp_max_write
,
11204 { } /* Terminate */
11207 static int cpu_extra_stat_show(struct seq_file
*sf
,
11208 struct cgroup_subsys_state
*css
)
11210 #ifdef CONFIG_CFS_BANDWIDTH
11212 struct task_group
*tg
= css_tg(css
);
11213 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
11214 u64 throttled_usec
, burst_usec
;
11216 throttled_usec
= cfs_b
->throttled_time
;
11217 do_div(throttled_usec
, NSEC_PER_USEC
);
11218 burst_usec
= cfs_b
->burst_time
;
11219 do_div(burst_usec
, NSEC_PER_USEC
);
11221 seq_printf(sf
, "nr_periods %d\n"
11222 "nr_throttled %d\n"
11223 "throttled_usec %llu\n"
11225 "burst_usec %llu\n",
11226 cfs_b
->nr_periods
, cfs_b
->nr_throttled
,
11227 throttled_usec
, cfs_b
->nr_burst
, burst_usec
);
11233 #ifdef CONFIG_FAIR_GROUP_SCHED
11234 static u64
cpu_weight_read_u64(struct cgroup_subsys_state
*css
,
11235 struct cftype
*cft
)
11237 struct task_group
*tg
= css_tg(css
);
11238 u64 weight
= scale_load_down(tg
->shares
);
11240 return DIV_ROUND_CLOSEST_ULL(weight
* CGROUP_WEIGHT_DFL
, 1024);
11243 static int cpu_weight_write_u64(struct cgroup_subsys_state
*css
,
11244 struct cftype
*cft
, u64 weight
)
11247 * cgroup weight knobs should use the common MIN, DFL and MAX
11248 * values which are 1, 100 and 10000 respectively. While it loses
11249 * a bit of range on both ends, it maps pretty well onto the shares
11250 * value used by scheduler and the round-trip conversions preserve
11251 * the original value over the entire range.
11253 if (weight
< CGROUP_WEIGHT_MIN
|| weight
> CGROUP_WEIGHT_MAX
)
11256 weight
= DIV_ROUND_CLOSEST_ULL(weight
* 1024, CGROUP_WEIGHT_DFL
);
11258 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
11261 static s64
cpu_weight_nice_read_s64(struct cgroup_subsys_state
*css
,
11262 struct cftype
*cft
)
11264 unsigned long weight
= scale_load_down(css_tg(css
)->shares
);
11265 int last_delta
= INT_MAX
;
11268 /* find the closest nice value to the current weight */
11269 for (prio
= 0; prio
< ARRAY_SIZE(sched_prio_to_weight
); prio
++) {
11270 delta
= abs(sched_prio_to_weight
[prio
] - weight
);
11271 if (delta
>= last_delta
)
11273 last_delta
= delta
;
11276 return PRIO_TO_NICE(prio
- 1 + MAX_RT_PRIO
);
11279 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state
*css
,
11280 struct cftype
*cft
, s64 nice
)
11282 unsigned long weight
;
11285 if (nice
< MIN_NICE
|| nice
> MAX_NICE
)
11288 idx
= NICE_TO_PRIO(nice
) - MAX_RT_PRIO
;
11289 idx
= array_index_nospec(idx
, 40);
11290 weight
= sched_prio_to_weight
[idx
];
11292 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
11296 static void __maybe_unused
cpu_period_quota_print(struct seq_file
*sf
,
11297 long period
, long quota
)
11300 seq_puts(sf
, "max");
11302 seq_printf(sf
, "%ld", quota
);
11304 seq_printf(sf
, " %ld\n", period
);
11307 /* caller should put the current value in *@periodp before calling */
11308 static int __maybe_unused
cpu_period_quota_parse(char *buf
,
11309 u64
*periodp
, u64
*quotap
)
11311 char tok
[21]; /* U64_MAX */
11313 if (sscanf(buf
, "%20s %llu", tok
, periodp
) < 1)
11316 *periodp
*= NSEC_PER_USEC
;
11318 if (sscanf(tok
, "%llu", quotap
))
11319 *quotap
*= NSEC_PER_USEC
;
11320 else if (!strcmp(tok
, "max"))
11321 *quotap
= RUNTIME_INF
;
11328 #ifdef CONFIG_CFS_BANDWIDTH
11329 static int cpu_max_show(struct seq_file
*sf
, void *v
)
11331 struct task_group
*tg
= css_tg(seq_css(sf
));
11333 cpu_period_quota_print(sf
, tg_get_cfs_period(tg
), tg_get_cfs_quota(tg
));
11337 static ssize_t
cpu_max_write(struct kernfs_open_file
*of
,
11338 char *buf
, size_t nbytes
, loff_t off
)
11340 struct task_group
*tg
= css_tg(of_css(of
));
11341 u64 period
= tg_get_cfs_period(tg
);
11342 u64 burst
= tg_get_cfs_burst(tg
);
11346 ret
= cpu_period_quota_parse(buf
, &period
, "a
);
11348 ret
= tg_set_cfs_bandwidth(tg
, period
, quota
, burst
);
11349 return ret
?: nbytes
;
11353 static struct cftype cpu_files
[] = {
11354 #ifdef CONFIG_FAIR_GROUP_SCHED
11357 .flags
= CFTYPE_NOT_ON_ROOT
,
11358 .read_u64
= cpu_weight_read_u64
,
11359 .write_u64
= cpu_weight_write_u64
,
11362 .name
= "weight.nice",
11363 .flags
= CFTYPE_NOT_ON_ROOT
,
11364 .read_s64
= cpu_weight_nice_read_s64
,
11365 .write_s64
= cpu_weight_nice_write_s64
,
11369 .flags
= CFTYPE_NOT_ON_ROOT
,
11370 .read_s64
= cpu_idle_read_s64
,
11371 .write_s64
= cpu_idle_write_s64
,
11374 #ifdef CONFIG_CFS_BANDWIDTH
11377 .flags
= CFTYPE_NOT_ON_ROOT
,
11378 .seq_show
= cpu_max_show
,
11379 .write
= cpu_max_write
,
11382 .name
= "max.burst",
11383 .flags
= CFTYPE_NOT_ON_ROOT
,
11384 .read_u64
= cpu_cfs_burst_read_u64
,
11385 .write_u64
= cpu_cfs_burst_write_u64
,
11388 #ifdef CONFIG_UCLAMP_TASK_GROUP
11390 .name
= "uclamp.min",
11391 .flags
= CFTYPE_NOT_ON_ROOT
,
11392 .seq_show
= cpu_uclamp_min_show
,
11393 .write
= cpu_uclamp_min_write
,
11396 .name
= "uclamp.max",
11397 .flags
= CFTYPE_NOT_ON_ROOT
,
11398 .seq_show
= cpu_uclamp_max_show
,
11399 .write
= cpu_uclamp_max_write
,
11402 { } /* terminate */
11405 struct cgroup_subsys cpu_cgrp_subsys
= {
11406 .css_alloc
= cpu_cgroup_css_alloc
,
11407 .css_online
= cpu_cgroup_css_online
,
11408 .css_released
= cpu_cgroup_css_released
,
11409 .css_free
= cpu_cgroup_css_free
,
11410 .css_extra_stat_show
= cpu_extra_stat_show
,
11411 #ifdef CONFIG_RT_GROUP_SCHED
11412 .can_attach
= cpu_cgroup_can_attach
,
11414 .attach
= cpu_cgroup_attach
,
11415 .legacy_cftypes
= cpu_legacy_files
,
11416 .dfl_cftypes
= cpu_files
,
11417 .early_init
= true,
11421 #endif /* CONFIG_CGROUP_SCHED */
11423 void dump_cpu_task(int cpu
)
11425 if (cpu
== smp_processor_id() && in_hardirq()) {
11426 struct pt_regs
*regs
;
11428 regs
= get_irq_regs();
11435 if (trigger_single_cpu_backtrace(cpu
))
11438 pr_info("Task dump for CPU %d:\n", cpu
);
11439 sched_show_task(cpu_curr(cpu
));
11443 * Nice levels are multiplicative, with a gentle 10% change for every
11444 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11445 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11446 * that remained on nice 0.
11448 * The "10% effect" is relative and cumulative: from _any_ nice level,
11449 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11450 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11451 * If a task goes up by ~10% and another task goes down by ~10% then
11452 * the relative distance between them is ~25%.)
11454 const int sched_prio_to_weight
[40] = {
11455 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11456 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11457 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11458 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11459 /* 0 */ 1024, 820, 655, 526, 423,
11460 /* 5 */ 335, 272, 215, 172, 137,
11461 /* 10 */ 110, 87, 70, 56, 45,
11462 /* 15 */ 36, 29, 23, 18, 15,
11466 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11468 * In cases where the weight does not change often, we can use the
11469 * precalculated inverse to speed up arithmetics by turning divisions
11470 * into multiplications:
11472 const u32 sched_prio_to_wmult
[40] = {
11473 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11474 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11475 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11476 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11477 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11478 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11479 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11480 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11483 void call_trace_sched_update_nr_running(struct rq
*rq
, int count
)
11485 trace_sched_update_nr_running_tp(rq
, count
);
11488 #ifdef CONFIG_SCHED_MM_CID
11491 * @cid_lock: Guarantee forward-progress of cid allocation.
11493 * Concurrency ID allocation within a bitmap is mostly lock-free. The cid_lock
11494 * is only used when contention is detected by the lock-free allocation so
11495 * forward progress can be guaranteed.
11497 DEFINE_RAW_SPINLOCK(cid_lock
);
11500 * @use_cid_lock: Select cid allocation behavior: lock-free vs spinlock.
11502 * When @use_cid_lock is 0, the cid allocation is lock-free. When contention is
11503 * detected, it is set to 1 to ensure that all newly coming allocations are
11504 * serialized by @cid_lock until the allocation which detected contention
11505 * completes and sets @use_cid_lock back to 0. This guarantees forward progress
11506 * of a cid allocation.
11511 * mm_cid remote-clear implements a lock-free algorithm to clear per-mm/cpu cid
11512 * concurrently with respect to the execution of the source runqueue context
11515 * There is one basic properties we want to guarantee here:
11517 * (1) Remote-clear should _never_ mark a per-cpu cid UNSET when it is actively
11518 * used by a task. That would lead to concurrent allocation of the cid and
11519 * userspace corruption.
11521 * Provide this guarantee by introducing a Dekker memory ordering to guarantee
11522 * that a pair of loads observe at least one of a pair of stores, which can be
11531 * Which guarantees that x==0 && y==0 is impossible. But rather than using
11532 * values 0 and 1, this algorithm cares about specific state transitions of the
11533 * runqueue current task (as updated by the scheduler context switch), and the
11534 * per-mm/cpu cid value.
11536 * Let's introduce task (Y) which has task->mm == mm and task (N) which has
11537 * task->mm != mm for the rest of the discussion. There are two scheduler state
11538 * transitions on context switch we care about:
11540 * (TSA) Store to rq->curr with transition from (N) to (Y)
11542 * (TSB) Store to rq->curr with transition from (Y) to (N)
11544 * On the remote-clear side, there is one transition we care about:
11546 * (TMA) cmpxchg to *pcpu_cid to set the LAZY flag
11548 * There is also a transition to UNSET state which can be performed from all
11549 * sides (scheduler, remote-clear). It is always performed with a cmpxchg which
11550 * guarantees that only a single thread will succeed:
11552 * (TMB) cmpxchg to *pcpu_cid to mark UNSET
11554 * Just to be clear, what we do _not_ want to happen is a transition to UNSET
11555 * when a thread is actively using the cid (property (1)).
11557 * Let's looks at the relevant combinations of TSA/TSB, and TMA transitions.
11559 * Scenario A) (TSA)+(TMA) (from next task perspective)
11563 * Context switch CS-1 Remote-clear
11564 * - store to rq->curr: (N)->(Y) (TSA) - cmpxchg to *pcpu_id to LAZY (TMA)
11565 * (implied barrier after cmpxchg)
11566 * - switch_mm_cid()
11567 * - memory barrier (see switch_mm_cid()
11568 * comment explaining how this barrier
11569 * is combined with other scheduler
11571 * - mm_cid_get (next)
11572 * - READ_ONCE(*pcpu_cid) - rcu_dereference(src_rq->curr)
11574 * This Dekker ensures that either task (Y) is observed by the
11575 * rcu_dereference() or the LAZY flag is observed by READ_ONCE(), or both are
11578 * If task (Y) store is observed by rcu_dereference(), it means that there is
11579 * still an active task on the cpu. Remote-clear will therefore not transition
11580 * to UNSET, which fulfills property (1).
11582 * If task (Y) is not observed, but the lazy flag is observed by READ_ONCE(),
11583 * it will move its state to UNSET, which clears the percpu cid perhaps
11584 * uselessly (which is not an issue for correctness). Because task (Y) is not
11585 * observed, CPU1 can move ahead to set the state to UNSET. Because moving
11586 * state to UNSET is done with a cmpxchg expecting that the old state has the
11587 * LAZY flag set, only one thread will successfully UNSET.
11589 * If both states (LAZY flag and task (Y)) are observed, the thread on CPU0
11590 * will observe the LAZY flag and transition to UNSET (perhaps uselessly), and
11591 * CPU1 will observe task (Y) and do nothing more, which is fine.
11593 * What we are effectively preventing with this Dekker is a scenario where
11594 * neither LAZY flag nor store (Y) are observed, which would fail property (1)
11595 * because this would UNSET a cid which is actively used.
11598 void sched_mm_cid_migrate_from(struct task_struct
*t
)
11600 t
->migrate_from_cpu
= task_cpu(t
);
11604 int __sched_mm_cid_migrate_from_fetch_cid(struct rq
*src_rq
,
11605 struct task_struct
*t
,
11606 struct mm_cid
*src_pcpu_cid
)
11608 struct mm_struct
*mm
= t
->mm
;
11609 struct task_struct
*src_task
;
11610 int src_cid
, last_mm_cid
;
11615 last_mm_cid
= t
->last_mm_cid
;
11617 * If the migrated task has no last cid, or if the current
11618 * task on src rq uses the cid, it means the source cid does not need
11619 * to be moved to the destination cpu.
11621 if (last_mm_cid
== -1)
11623 src_cid
= READ_ONCE(src_pcpu_cid
->cid
);
11624 if (!mm_cid_is_valid(src_cid
) || last_mm_cid
!= src_cid
)
11628 * If we observe an active task using the mm on this rq, it means we
11629 * are not the last task to be migrated from this cpu for this mm, so
11630 * there is no need to move src_cid to the destination cpu.
11633 src_task
= rcu_dereference(src_rq
->curr
);
11634 if (READ_ONCE(src_task
->mm_cid_active
) && src_task
->mm
== mm
) {
11636 t
->last_mm_cid
= -1;
11645 int __sched_mm_cid_migrate_from_try_steal_cid(struct rq
*src_rq
,
11646 struct task_struct
*t
,
11647 struct mm_cid
*src_pcpu_cid
,
11650 struct task_struct
*src_task
;
11651 struct mm_struct
*mm
= t
->mm
;
11658 * Attempt to clear the source cpu cid to move it to the destination
11661 lazy_cid
= mm_cid_set_lazy_put(src_cid
);
11662 if (!try_cmpxchg(&src_pcpu_cid
->cid
, &src_cid
, lazy_cid
))
11666 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11667 * rq->curr->mm matches the scheduler barrier in context_switch()
11668 * between store to rq->curr and load of prev and next task's
11671 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11672 * rq->curr->mm_cid_active matches the barrier in
11673 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11674 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11675 * load of per-mm/cpu cid.
11679 * If we observe an active task using the mm on this rq after setting
11680 * the lazy-put flag, this task will be responsible for transitioning
11681 * from lazy-put flag set to MM_CID_UNSET.
11684 src_task
= rcu_dereference(src_rq
->curr
);
11685 if (READ_ONCE(src_task
->mm_cid_active
) && src_task
->mm
== mm
) {
11688 * We observed an active task for this mm, there is therefore
11689 * no point in moving this cid to the destination cpu.
11691 t
->last_mm_cid
= -1;
11697 * The src_cid is unused, so it can be unset.
11699 if (!try_cmpxchg(&src_pcpu_cid
->cid
, &lazy_cid
, MM_CID_UNSET
))
11705 * Migration to dst cpu. Called with dst_rq lock held.
11706 * Interrupts are disabled, which keeps the window of cid ownership without the
11707 * source rq lock held small.
11709 void sched_mm_cid_migrate_to(struct rq
*dst_rq
, struct task_struct
*t
)
11711 struct mm_cid
*src_pcpu_cid
, *dst_pcpu_cid
;
11712 struct mm_struct
*mm
= t
->mm
;
11713 int src_cid
, dst_cid
, src_cpu
;
11716 lockdep_assert_rq_held(dst_rq
);
11720 src_cpu
= t
->migrate_from_cpu
;
11721 if (src_cpu
== -1) {
11722 t
->last_mm_cid
= -1;
11726 * Move the src cid if the dst cid is unset. This keeps id
11727 * allocation closest to 0 in cases where few threads migrate around
11730 * If destination cid is already set, we may have to just clear
11731 * the src cid to ensure compactness in frequent migrations
11734 * It is not useful to clear the src cid when the number of threads is
11735 * greater or equal to the number of allowed cpus, because user-space
11736 * can expect that the number of allowed cids can reach the number of
11739 dst_pcpu_cid
= per_cpu_ptr(mm
->pcpu_cid
, cpu_of(dst_rq
));
11740 dst_cid
= READ_ONCE(dst_pcpu_cid
->cid
);
11741 if (!mm_cid_is_unset(dst_cid
) &&
11742 atomic_read(&mm
->mm_users
) >= t
->nr_cpus_allowed
)
11744 src_pcpu_cid
= per_cpu_ptr(mm
->pcpu_cid
, src_cpu
);
11745 src_rq
= cpu_rq(src_cpu
);
11746 src_cid
= __sched_mm_cid_migrate_from_fetch_cid(src_rq
, t
, src_pcpu_cid
);
11749 src_cid
= __sched_mm_cid_migrate_from_try_steal_cid(src_rq
, t
, src_pcpu_cid
,
11753 if (!mm_cid_is_unset(dst_cid
)) {
11754 __mm_cid_put(mm
, src_cid
);
11757 /* Move src_cid to dst cpu. */
11758 mm_cid_snapshot_time(dst_rq
, mm
);
11759 WRITE_ONCE(dst_pcpu_cid
->cid
, src_cid
);
11762 static void sched_mm_cid_remote_clear(struct mm_struct
*mm
, struct mm_cid
*pcpu_cid
,
11765 struct rq
*rq
= cpu_rq(cpu
);
11766 struct task_struct
*t
;
11767 unsigned long flags
;
11770 cid
= READ_ONCE(pcpu_cid
->cid
);
11771 if (!mm_cid_is_valid(cid
))
11775 * Clear the cpu cid if it is set to keep cid allocation compact. If
11776 * there happens to be other tasks left on the source cpu using this
11777 * mm, the next task using this mm will reallocate its cid on context
11780 lazy_cid
= mm_cid_set_lazy_put(cid
);
11781 if (!try_cmpxchg(&pcpu_cid
->cid
, &cid
, lazy_cid
))
11785 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11786 * rq->curr->mm matches the scheduler barrier in context_switch()
11787 * between store to rq->curr and load of prev and next task's
11790 * The implicit barrier after cmpxchg per-mm/cpu cid before loading
11791 * rq->curr->mm_cid_active matches the barrier in
11792 * sched_mm_cid_exit_signals(), sched_mm_cid_before_execve(), and
11793 * sched_mm_cid_after_execve() between store to t->mm_cid_active and
11794 * load of per-mm/cpu cid.
11798 * If we observe an active task using the mm on this rq after setting
11799 * the lazy-put flag, that task will be responsible for transitioning
11800 * from lazy-put flag set to MM_CID_UNSET.
11803 t
= rcu_dereference(rq
->curr
);
11804 if (READ_ONCE(t
->mm_cid_active
) && t
->mm
== mm
) {
11811 * The cid is unused, so it can be unset.
11812 * Disable interrupts to keep the window of cid ownership without rq
11815 local_irq_save(flags
);
11816 if (try_cmpxchg(&pcpu_cid
->cid
, &lazy_cid
, MM_CID_UNSET
))
11817 __mm_cid_put(mm
, cid
);
11818 local_irq_restore(flags
);
11821 static void sched_mm_cid_remote_clear_old(struct mm_struct
*mm
, int cpu
)
11823 struct rq
*rq
= cpu_rq(cpu
);
11824 struct mm_cid
*pcpu_cid
;
11825 struct task_struct
*curr
;
11829 * rq->clock load is racy on 32-bit but one spurious clear once in a
11830 * while is irrelevant.
11832 rq_clock
= READ_ONCE(rq
->clock
);
11833 pcpu_cid
= per_cpu_ptr(mm
->pcpu_cid
, cpu
);
11836 * In order to take care of infrequently scheduled tasks, bump the time
11837 * snapshot associated with this cid if an active task using the mm is
11838 * observed on this rq.
11841 curr
= rcu_dereference(rq
->curr
);
11842 if (READ_ONCE(curr
->mm_cid_active
) && curr
->mm
== mm
) {
11843 WRITE_ONCE(pcpu_cid
->time
, rq_clock
);
11849 if (rq_clock
< pcpu_cid
->time
+ SCHED_MM_CID_PERIOD_NS
)
11851 sched_mm_cid_remote_clear(mm
, pcpu_cid
, cpu
);
11854 static void sched_mm_cid_remote_clear_weight(struct mm_struct
*mm
, int cpu
,
11857 struct mm_cid
*pcpu_cid
;
11860 pcpu_cid
= per_cpu_ptr(mm
->pcpu_cid
, cpu
);
11861 cid
= READ_ONCE(pcpu_cid
->cid
);
11862 if (!mm_cid_is_valid(cid
) || cid
< weight
)
11864 sched_mm_cid_remote_clear(mm
, pcpu_cid
, cpu
);
11867 static void task_mm_cid_work(struct callback_head
*work
)
11869 unsigned long now
= jiffies
, old_scan
, next_scan
;
11870 struct task_struct
*t
= current
;
11871 struct cpumask
*cidmask
;
11872 struct mm_struct
*mm
;
11875 SCHED_WARN_ON(t
!= container_of(work
, struct task_struct
, cid_work
));
11877 work
->next
= work
; /* Prevent double-add */
11878 if (t
->flags
& PF_EXITING
)
11883 old_scan
= READ_ONCE(mm
->mm_cid_next_scan
);
11884 next_scan
= now
+ msecs_to_jiffies(MM_CID_SCAN_DELAY
);
11888 res
= cmpxchg(&mm
->mm_cid_next_scan
, old_scan
, next_scan
);
11889 if (res
!= old_scan
)
11892 old_scan
= next_scan
;
11894 if (time_before(now
, old_scan
))
11896 if (!try_cmpxchg(&mm
->mm_cid_next_scan
, &old_scan
, next_scan
))
11898 cidmask
= mm_cidmask(mm
);
11899 /* Clear cids that were not recently used. */
11900 for_each_possible_cpu(cpu
)
11901 sched_mm_cid_remote_clear_old(mm
, cpu
);
11902 weight
= cpumask_weight(cidmask
);
11904 * Clear cids that are greater or equal to the cidmask weight to
11907 for_each_possible_cpu(cpu
)
11908 sched_mm_cid_remote_clear_weight(mm
, cpu
, weight
);
11911 void init_sched_mm_cid(struct task_struct
*t
)
11913 struct mm_struct
*mm
= t
->mm
;
11917 mm_users
= atomic_read(&mm
->mm_users
);
11919 mm
->mm_cid_next_scan
= jiffies
+ msecs_to_jiffies(MM_CID_SCAN_DELAY
);
11921 t
->cid_work
.next
= &t
->cid_work
; /* Protect against double add */
11922 init_task_work(&t
->cid_work
, task_mm_cid_work
);
11925 void task_tick_mm_cid(struct rq
*rq
, struct task_struct
*curr
)
11927 struct callback_head
*work
= &curr
->cid_work
;
11928 unsigned long now
= jiffies
;
11930 if (!curr
->mm
|| (curr
->flags
& (PF_EXITING
| PF_KTHREAD
)) ||
11931 work
->next
!= work
)
11933 if (time_before(now
, READ_ONCE(curr
->mm
->mm_cid_next_scan
)))
11935 task_work_add(curr
, work
, TWA_RESUME
);
11938 void sched_mm_cid_exit_signals(struct task_struct
*t
)
11940 struct mm_struct
*mm
= t
->mm
;
11941 struct rq_flags rf
;
11949 rq_lock_irqsave(rq
, &rf
);
11950 preempt_enable_no_resched(); /* holding spinlock */
11951 WRITE_ONCE(t
->mm_cid_active
, 0);
11953 * Store t->mm_cid_active before loading per-mm/cpu cid.
11954 * Matches barrier in sched_mm_cid_remote_clear_old().
11958 t
->last_mm_cid
= t
->mm_cid
= -1;
11959 rq_unlock_irqrestore(rq
, &rf
);
11962 void sched_mm_cid_before_execve(struct task_struct
*t
)
11964 struct mm_struct
*mm
= t
->mm
;
11965 struct rq_flags rf
;
11973 rq_lock_irqsave(rq
, &rf
);
11974 preempt_enable_no_resched(); /* holding spinlock */
11975 WRITE_ONCE(t
->mm_cid_active
, 0);
11977 * Store t->mm_cid_active before loading per-mm/cpu cid.
11978 * Matches barrier in sched_mm_cid_remote_clear_old().
11982 t
->last_mm_cid
= t
->mm_cid
= -1;
11983 rq_unlock_irqrestore(rq
, &rf
);
11986 void sched_mm_cid_after_execve(struct task_struct
*t
)
11988 struct mm_struct
*mm
= t
->mm
;
11989 struct rq_flags rf
;
11997 rq_lock_irqsave(rq
, &rf
);
11998 preempt_enable_no_resched(); /* holding spinlock */
11999 WRITE_ONCE(t
->mm_cid_active
, 1);
12001 * Store t->mm_cid_active before loading per-mm/cpu cid.
12002 * Matches barrier in sched_mm_cid_remote_clear_old().
12005 t
->last_mm_cid
= t
->mm_cid
= mm_cid_get(rq
, mm
);
12006 rq_unlock_irqrestore(rq
, &rf
);
12007 rseq_set_notify_resume(t
);
12010 void sched_mm_cid_fork(struct task_struct
*t
)
12012 WARN_ON_ONCE(!t
->mm
|| t
->mm_cid
!= -1);
12013 t
->mm_cid_active
= 1;