4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex
);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
96 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
98 void update_rq_clock(struct rq
*rq
)
102 lockdep_assert_held(&rq
->lock
);
104 if (rq
->clock_skip_update
& RQCF_ACT_SKIP
)
107 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
111 update_rq_clock_task(rq
, delta
);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug
unsigned int sysctl_sched_features
=
122 #include "features.h"
127 #ifdef CONFIG_SCHED_DEBUG
128 #define SCHED_FEAT(name, enabled) \
131 static const char * const sched_feat_names
[] = {
132 #include "features.h"
137 static int sched_feat_show(struct seq_file
*m
, void *v
)
141 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
142 if (!(sysctl_sched_features
& (1UL << i
)))
144 seq_printf(m
, "%s ", sched_feat_names
[i
]);
151 #ifdef HAVE_JUMP_LABEL
153 #define jump_label_key__true STATIC_KEY_INIT_TRUE
154 #define jump_label_key__false STATIC_KEY_INIT_FALSE
156 #define SCHED_FEAT(name, enabled) \
157 jump_label_key__##enabled ,
159 struct static_key sched_feat_keys
[__SCHED_FEAT_NR
] = {
160 #include "features.h"
165 static void sched_feat_disable(int i
)
167 if (static_key_enabled(&sched_feat_keys
[i
]))
168 static_key_slow_dec(&sched_feat_keys
[i
]);
171 static void sched_feat_enable(int i
)
173 if (!static_key_enabled(&sched_feat_keys
[i
]))
174 static_key_slow_inc(&sched_feat_keys
[i
]);
177 static void sched_feat_disable(int i
) { };
178 static void sched_feat_enable(int i
) { };
179 #endif /* HAVE_JUMP_LABEL */
181 static int sched_feat_set(char *cmp
)
186 if (strncmp(cmp
, "NO_", 3) == 0) {
191 for (i
= 0; i
< __SCHED_FEAT_NR
; i
++) {
192 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
194 sysctl_sched_features
&= ~(1UL << i
);
195 sched_feat_disable(i
);
197 sysctl_sched_features
|= (1UL << i
);
198 sched_feat_enable(i
);
208 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
209 size_t cnt
, loff_t
*ppos
)
219 if (copy_from_user(&buf
, ubuf
, cnt
))
225 /* Ensure the static_key remains in a consistent state */
226 inode
= file_inode(filp
);
227 mutex_lock(&inode
->i_mutex
);
228 i
= sched_feat_set(cmp
);
229 mutex_unlock(&inode
->i_mutex
);
230 if (i
== __SCHED_FEAT_NR
)
238 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
240 return single_open(filp
, sched_feat_show
, NULL
);
243 static const struct file_operations sched_feat_fops
= {
244 .open
= sched_feat_open
,
245 .write
= sched_feat_write
,
248 .release
= single_release
,
251 static __init
int sched_init_debug(void)
253 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
258 late_initcall(sched_init_debug
);
259 #endif /* CONFIG_SCHED_DEBUG */
262 * Number of tasks to iterate in a single balance run.
263 * Limited because this is done with IRQs disabled.
265 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
268 * period over which we average the RT time consumption, measured
273 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
276 * period over which we measure -rt task cpu usage in us.
279 unsigned int sysctl_sched_rt_period
= 1000000;
281 __read_mostly
int scheduler_running
;
284 * part of the period that we allow rt tasks to run in us.
287 int sysctl_sched_rt_runtime
= 950000;
289 /* cpus with isolated domains */
290 cpumask_var_t cpu_isolated_map
;
293 * this_rq_lock - lock this runqueue and disable interrupts.
295 static struct rq
*this_rq_lock(void)
302 raw_spin_lock(&rq
->lock
);
307 #ifdef CONFIG_SCHED_HRTICK
309 * Use HR-timers to deliver accurate preemption points.
312 static void hrtick_clear(struct rq
*rq
)
314 if (hrtimer_active(&rq
->hrtick_timer
))
315 hrtimer_cancel(&rq
->hrtick_timer
);
319 * High-resolution timer tick.
320 * Runs from hardirq context with interrupts disabled.
322 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
324 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
326 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
328 raw_spin_lock(&rq
->lock
);
330 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
331 raw_spin_unlock(&rq
->lock
);
333 return HRTIMER_NORESTART
;
338 static void __hrtick_restart(struct rq
*rq
)
340 struct hrtimer
*timer
= &rq
->hrtick_timer
;
342 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
346 * called from hardirq (IPI) context
348 static void __hrtick_start(void *arg
)
352 raw_spin_lock(&rq
->lock
);
353 __hrtick_restart(rq
);
354 rq
->hrtick_csd_pending
= 0;
355 raw_spin_unlock(&rq
->lock
);
359 * Called to set the hrtick timer state.
361 * called with rq->lock held and irqs disabled
363 void hrtick_start(struct rq
*rq
, u64 delay
)
365 struct hrtimer
*timer
= &rq
->hrtick_timer
;
370 * Don't schedule slices shorter than 10000ns, that just
371 * doesn't make sense and can cause timer DoS.
373 delta
= max_t(s64
, delay
, 10000LL);
374 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
376 hrtimer_set_expires(timer
, time
);
378 if (rq
== this_rq()) {
379 __hrtick_restart(rq
);
380 } else if (!rq
->hrtick_csd_pending
) {
381 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
382 rq
->hrtick_csd_pending
= 1;
387 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
389 int cpu
= (int)(long)hcpu
;
392 case CPU_UP_CANCELED
:
393 case CPU_UP_CANCELED_FROZEN
:
394 case CPU_DOWN_PREPARE
:
395 case CPU_DOWN_PREPARE_FROZEN
:
397 case CPU_DEAD_FROZEN
:
398 hrtick_clear(cpu_rq(cpu
));
405 static __init
void init_hrtick(void)
407 hotcpu_notifier(hotplug_hrtick
, 0);
411 * Called to set the hrtick timer state.
413 * called with rq->lock held and irqs disabled
415 void hrtick_start(struct rq
*rq
, u64 delay
)
418 * Don't schedule slices shorter than 10000ns, that just
419 * doesn't make sense. Rely on vruntime for fairness.
421 delay
= max_t(u64
, delay
, 10000LL);
422 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
423 HRTIMER_MODE_REL_PINNED
);
426 static inline void init_hrtick(void)
429 #endif /* CONFIG_SMP */
431 static void init_rq_hrtick(struct rq
*rq
)
434 rq
->hrtick_csd_pending
= 0;
436 rq
->hrtick_csd
.flags
= 0;
437 rq
->hrtick_csd
.func
= __hrtick_start
;
438 rq
->hrtick_csd
.info
= rq
;
441 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
442 rq
->hrtick_timer
.function
= hrtick
;
444 #else /* CONFIG_SCHED_HRTICK */
445 static inline void hrtick_clear(struct rq
*rq
)
449 static inline void init_rq_hrtick(struct rq
*rq
)
453 static inline void init_hrtick(void)
456 #endif /* CONFIG_SCHED_HRTICK */
459 * cmpxchg based fetch_or, macro so it works for different integer types
461 #define fetch_or(ptr, val) \
462 ({ typeof(*(ptr)) __old, __val = *(ptr); \
464 __old = cmpxchg((ptr), __val, __val | (val)); \
465 if (__old == __val) \
472 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
474 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
475 * this avoids any races wrt polling state changes and thereby avoids
478 static bool set_nr_and_not_polling(struct task_struct
*p
)
480 struct thread_info
*ti
= task_thread_info(p
);
481 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
485 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
487 * If this returns true, then the idle task promises to call
488 * sched_ttwu_pending() and reschedule soon.
490 static bool set_nr_if_polling(struct task_struct
*p
)
492 struct thread_info
*ti
= task_thread_info(p
);
493 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
496 if (!(val
& _TIF_POLLING_NRFLAG
))
498 if (val
& _TIF_NEED_RESCHED
)
500 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
509 static bool set_nr_and_not_polling(struct task_struct
*p
)
511 set_tsk_need_resched(p
);
516 static bool set_nr_if_polling(struct task_struct
*p
)
523 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
525 struct wake_q_node
*node
= &task
->wake_q
;
528 * Atomically grab the task, if ->wake_q is !nil already it means
529 * its already queued (either by us or someone else) and will get the
530 * wakeup due to that.
532 * This cmpxchg() implies a full barrier, which pairs with the write
533 * barrier implied by the wakeup in wake_up_list().
535 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
538 get_task_struct(task
);
541 * The head is context local, there can be no concurrency.
544 head
->lastp
= &node
->next
;
547 void wake_up_q(struct wake_q_head
*head
)
549 struct wake_q_node
*node
= head
->first
;
551 while (node
!= WAKE_Q_TAIL
) {
552 struct task_struct
*task
;
554 task
= container_of(node
, struct task_struct
, wake_q
);
556 /* task can safely be re-inserted now */
558 task
->wake_q
.next
= NULL
;
561 * wake_up_process() implies a wmb() to pair with the queueing
562 * in wake_q_add() so as not to miss wakeups.
564 wake_up_process(task
);
565 put_task_struct(task
);
570 * resched_curr - mark rq's current task 'to be rescheduled now'.
572 * On UP this means the setting of the need_resched flag, on SMP it
573 * might also involve a cross-CPU call to trigger the scheduler on
576 void resched_curr(struct rq
*rq
)
578 struct task_struct
*curr
= rq
->curr
;
581 lockdep_assert_held(&rq
->lock
);
583 if (test_tsk_need_resched(curr
))
588 if (cpu
== smp_processor_id()) {
589 set_tsk_need_resched(curr
);
590 set_preempt_need_resched();
594 if (set_nr_and_not_polling(curr
))
595 smp_send_reschedule(cpu
);
597 trace_sched_wake_idle_without_ipi(cpu
);
600 void resched_cpu(int cpu
)
602 struct rq
*rq
= cpu_rq(cpu
);
605 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
608 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
612 #ifdef CONFIG_NO_HZ_COMMON
614 * In the semi idle case, use the nearest busy cpu for migrating timers
615 * from an idle cpu. This is good for power-savings.
617 * We don't do similar optimization for completely idle system, as
618 * selecting an idle cpu will add more delays to the timers than intended
619 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
621 int get_nohz_timer_target(int pinned
)
623 int cpu
= smp_processor_id();
625 struct sched_domain
*sd
;
627 if (pinned
|| !get_sysctl_timer_migration() || !idle_cpu(cpu
))
631 for_each_domain(cpu
, sd
) {
632 for_each_cpu(i
, sched_domain_span(sd
)) {
644 * When add_timer_on() enqueues a timer into the timer wheel of an
645 * idle CPU then this timer might expire before the next timer event
646 * which is scheduled to wake up that CPU. In case of a completely
647 * idle system the next event might even be infinite time into the
648 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
649 * leaves the inner idle loop so the newly added timer is taken into
650 * account when the CPU goes back to idle and evaluates the timer
651 * wheel for the next timer event.
653 static void wake_up_idle_cpu(int cpu
)
655 struct rq
*rq
= cpu_rq(cpu
);
657 if (cpu
== smp_processor_id())
660 if (set_nr_and_not_polling(rq
->idle
))
661 smp_send_reschedule(cpu
);
663 trace_sched_wake_idle_without_ipi(cpu
);
666 static bool wake_up_full_nohz_cpu(int cpu
)
669 * We just need the target to call irq_exit() and re-evaluate
670 * the next tick. The nohz full kick at least implies that.
671 * If needed we can still optimize that later with an
674 if (tick_nohz_full_cpu(cpu
)) {
675 if (cpu
!= smp_processor_id() ||
676 tick_nohz_tick_stopped())
677 tick_nohz_full_kick_cpu(cpu
);
684 void wake_up_nohz_cpu(int cpu
)
686 if (!wake_up_full_nohz_cpu(cpu
))
687 wake_up_idle_cpu(cpu
);
690 static inline bool got_nohz_idle_kick(void)
692 int cpu
= smp_processor_id();
694 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
697 if (idle_cpu(cpu
) && !need_resched())
701 * We can't run Idle Load Balance on this CPU for this time so we
702 * cancel it and clear NOHZ_BALANCE_KICK
704 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
708 #else /* CONFIG_NO_HZ_COMMON */
710 static inline bool got_nohz_idle_kick(void)
715 #endif /* CONFIG_NO_HZ_COMMON */
717 #ifdef CONFIG_NO_HZ_FULL
718 bool sched_can_stop_tick(void)
721 * FIFO realtime policy runs the highest priority task. Other runnable
722 * tasks are of a lower priority. The scheduler tick does nothing.
724 if (current
->policy
== SCHED_FIFO
)
728 * Round-robin realtime tasks time slice with other tasks at the same
729 * realtime priority. Is this task the only one at this priority?
731 if (current
->policy
== SCHED_RR
) {
732 struct sched_rt_entity
*rt_se
= ¤t
->rt
;
734 return rt_se
->run_list
.prev
== rt_se
->run_list
.next
;
738 * More than one running task need preemption.
739 * nr_running update is assumed to be visible
740 * after IPI is sent from wakers.
742 if (this_rq()->nr_running
> 1)
747 #endif /* CONFIG_NO_HZ_FULL */
749 void sched_avg_update(struct rq
*rq
)
751 s64 period
= sched_avg_period();
753 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
755 * Inline assembly required to prevent the compiler
756 * optimising this loop into a divmod call.
757 * See __iter_div_u64_rem() for another example of this.
759 asm("" : "+rm" (rq
->age_stamp
));
760 rq
->age_stamp
+= period
;
765 #endif /* CONFIG_SMP */
767 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
768 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
770 * Iterate task_group tree rooted at *from, calling @down when first entering a
771 * node and @up when leaving it for the final time.
773 * Caller must hold rcu_lock or sufficient equivalent.
775 int walk_tg_tree_from(struct task_group
*from
,
776 tg_visitor down
, tg_visitor up
, void *data
)
778 struct task_group
*parent
, *child
;
784 ret
= (*down
)(parent
, data
);
787 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
794 ret
= (*up
)(parent
, data
);
795 if (ret
|| parent
== from
)
799 parent
= parent
->parent
;
806 int tg_nop(struct task_group
*tg
, void *data
)
812 static void set_load_weight(struct task_struct
*p
)
814 int prio
= p
->static_prio
- MAX_RT_PRIO
;
815 struct load_weight
*load
= &p
->se
.load
;
818 * SCHED_IDLE tasks get minimal weight:
820 if (p
->policy
== SCHED_IDLE
) {
821 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
822 load
->inv_weight
= WMULT_IDLEPRIO
;
826 load
->weight
= scale_load(prio_to_weight
[prio
]);
827 load
->inv_weight
= prio_to_wmult
[prio
];
830 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
833 sched_info_queued(rq
, p
);
834 p
->sched_class
->enqueue_task(rq
, p
, flags
);
837 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
840 sched_info_dequeued(rq
, p
);
841 p
->sched_class
->dequeue_task(rq
, p
, flags
);
844 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
846 if (task_contributes_to_load(p
))
847 rq
->nr_uninterruptible
--;
849 enqueue_task(rq
, p
, flags
);
852 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
854 if (task_contributes_to_load(p
))
855 rq
->nr_uninterruptible
++;
857 dequeue_task(rq
, p
, flags
);
860 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
863 * In theory, the compile should just see 0 here, and optimize out the call
864 * to sched_rt_avg_update. But I don't trust it...
866 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
867 s64 steal
= 0, irq_delta
= 0;
869 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
870 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
873 * Since irq_time is only updated on {soft,}irq_exit, we might run into
874 * this case when a previous update_rq_clock() happened inside a
877 * When this happens, we stop ->clock_task and only update the
878 * prev_irq_time stamp to account for the part that fit, so that a next
879 * update will consume the rest. This ensures ->clock_task is
882 * It does however cause some slight miss-attribution of {soft,}irq
883 * time, a more accurate solution would be to update the irq_time using
884 * the current rq->clock timestamp, except that would require using
887 if (irq_delta
> delta
)
890 rq
->prev_irq_time
+= irq_delta
;
893 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
894 if (static_key_false((¶virt_steal_rq_enabled
))) {
895 steal
= paravirt_steal_clock(cpu_of(rq
));
896 steal
-= rq
->prev_steal_time_rq
;
898 if (unlikely(steal
> delta
))
901 rq
->prev_steal_time_rq
+= steal
;
906 rq
->clock_task
+= delta
;
908 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
909 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
910 sched_rt_avg_update(rq
, irq_delta
+ steal
);
914 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
916 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
917 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
921 * Make it appear like a SCHED_FIFO task, its something
922 * userspace knows about and won't get confused about.
924 * Also, it will make PI more or less work without too
925 * much confusion -- but then, stop work should not
926 * rely on PI working anyway.
928 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
930 stop
->sched_class
= &stop_sched_class
;
933 cpu_rq(cpu
)->stop
= stop
;
937 * Reset it back to a normal scheduling class so that
938 * it can die in pieces.
940 old_stop
->sched_class
= &rt_sched_class
;
945 * __normal_prio - return the priority that is based on the static prio
947 static inline int __normal_prio(struct task_struct
*p
)
949 return p
->static_prio
;
953 * Calculate the expected normal priority: i.e. priority
954 * without taking RT-inheritance into account. Might be
955 * boosted by interactivity modifiers. Changes upon fork,
956 * setprio syscalls, and whenever the interactivity
957 * estimator recalculates.
959 static inline int normal_prio(struct task_struct
*p
)
963 if (task_has_dl_policy(p
))
964 prio
= MAX_DL_PRIO
-1;
965 else if (task_has_rt_policy(p
))
966 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
968 prio
= __normal_prio(p
);
973 * Calculate the current priority, i.e. the priority
974 * taken into account by the scheduler. This value might
975 * be boosted by RT tasks, or might be boosted by
976 * interactivity modifiers. Will be RT if the task got
977 * RT-boosted. If not then it returns p->normal_prio.
979 static int effective_prio(struct task_struct
*p
)
981 p
->normal_prio
= normal_prio(p
);
983 * If we are RT tasks or we were boosted to RT priority,
984 * keep the priority unchanged. Otherwise, update priority
985 * to the normal priority:
987 if (!rt_prio(p
->prio
))
988 return p
->normal_prio
;
993 * task_curr - is this task currently executing on a CPU?
994 * @p: the task in question.
996 * Return: 1 if the task is currently executing. 0 otherwise.
998 inline int task_curr(const struct task_struct
*p
)
1000 return cpu_curr(task_cpu(p
)) == p
;
1004 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1005 * use the balance_callback list if you want balancing.
1007 * this means any call to check_class_changed() must be followed by a call to
1008 * balance_callback().
1010 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1011 const struct sched_class
*prev_class
,
1014 if (prev_class
!= p
->sched_class
) {
1015 if (prev_class
->switched_from
)
1016 prev_class
->switched_from(rq
, p
);
1018 p
->sched_class
->switched_to(rq
, p
);
1019 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1020 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1023 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1025 const struct sched_class
*class;
1027 if (p
->sched_class
== rq
->curr
->sched_class
) {
1028 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1030 for_each_class(class) {
1031 if (class == rq
->curr
->sched_class
)
1033 if (class == p
->sched_class
) {
1041 * A queue event has occurred, and we're going to schedule. In
1042 * this case, we can save a useless back to back clock update.
1044 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1045 rq_clock_skip_update(rq
, true);
1049 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1051 #ifdef CONFIG_SCHED_DEBUG
1053 * We should never call set_task_cpu() on a blocked task,
1054 * ttwu() will sort out the placement.
1056 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1059 #ifdef CONFIG_LOCKDEP
1061 * The caller should hold either p->pi_lock or rq->lock, when changing
1062 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1064 * sched_move_task() holds both and thus holding either pins the cgroup,
1067 * Furthermore, all task_rq users should acquire both locks, see
1070 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1071 lockdep_is_held(&task_rq(p
)->lock
)));
1075 trace_sched_migrate_task(p
, new_cpu
);
1077 if (task_cpu(p
) != new_cpu
) {
1078 if (p
->sched_class
->migrate_task_rq
)
1079 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1080 p
->se
.nr_migrations
++;
1081 perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 0);
1084 __set_task_cpu(p
, new_cpu
);
1087 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1089 if (task_on_rq_queued(p
)) {
1090 struct rq
*src_rq
, *dst_rq
;
1092 src_rq
= task_rq(p
);
1093 dst_rq
= cpu_rq(cpu
);
1095 deactivate_task(src_rq
, p
, 0);
1096 set_task_cpu(p
, cpu
);
1097 activate_task(dst_rq
, p
, 0);
1098 check_preempt_curr(dst_rq
, p
, 0);
1101 * Task isn't running anymore; make it appear like we migrated
1102 * it before it went to sleep. This means on wakeup we make the
1103 * previous cpu our targer instead of where it really is.
1109 struct migration_swap_arg
{
1110 struct task_struct
*src_task
, *dst_task
;
1111 int src_cpu
, dst_cpu
;
1114 static int migrate_swap_stop(void *data
)
1116 struct migration_swap_arg
*arg
= data
;
1117 struct rq
*src_rq
, *dst_rq
;
1120 src_rq
= cpu_rq(arg
->src_cpu
);
1121 dst_rq
= cpu_rq(arg
->dst_cpu
);
1123 double_raw_lock(&arg
->src_task
->pi_lock
,
1124 &arg
->dst_task
->pi_lock
);
1125 double_rq_lock(src_rq
, dst_rq
);
1126 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1129 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1132 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1135 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1138 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1139 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1144 double_rq_unlock(src_rq
, dst_rq
);
1145 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1146 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1152 * Cross migrate two tasks
1154 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1156 struct migration_swap_arg arg
;
1159 arg
= (struct migration_swap_arg
){
1161 .src_cpu
= task_cpu(cur
),
1163 .dst_cpu
= task_cpu(p
),
1166 if (arg
.src_cpu
== arg
.dst_cpu
)
1170 * These three tests are all lockless; this is OK since all of them
1171 * will be re-checked with proper locks held further down the line.
1173 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1176 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1179 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1182 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1183 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1189 struct migration_arg
{
1190 struct task_struct
*task
;
1194 static int migration_cpu_stop(void *data
);
1197 * wait_task_inactive - wait for a thread to unschedule.
1199 * If @match_state is nonzero, it's the @p->state value just checked and
1200 * not expected to change. If it changes, i.e. @p might have woken up,
1201 * then return zero. When we succeed in waiting for @p to be off its CPU,
1202 * we return a positive number (its total switch count). If a second call
1203 * a short while later returns the same number, the caller can be sure that
1204 * @p has remained unscheduled the whole time.
1206 * The caller must ensure that the task *will* unschedule sometime soon,
1207 * else this function might spin for a *long* time. This function can't
1208 * be called with interrupts off, or it may introduce deadlock with
1209 * smp_call_function() if an IPI is sent by the same process we are
1210 * waiting to become inactive.
1212 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1214 unsigned long flags
;
1215 int running
, queued
;
1221 * We do the initial early heuristics without holding
1222 * any task-queue locks at all. We'll only try to get
1223 * the runqueue lock when things look like they will
1229 * If the task is actively running on another CPU
1230 * still, just relax and busy-wait without holding
1233 * NOTE! Since we don't hold any locks, it's not
1234 * even sure that "rq" stays as the right runqueue!
1235 * But we don't care, since "task_running()" will
1236 * return false if the runqueue has changed and p
1237 * is actually now running somewhere else!
1239 while (task_running(rq
, p
)) {
1240 if (match_state
&& unlikely(p
->state
!= match_state
))
1246 * Ok, time to look more closely! We need the rq
1247 * lock now, to be *sure*. If we're wrong, we'll
1248 * just go back and repeat.
1250 rq
= task_rq_lock(p
, &flags
);
1251 trace_sched_wait_task(p
);
1252 running
= task_running(rq
, p
);
1253 queued
= task_on_rq_queued(p
);
1255 if (!match_state
|| p
->state
== match_state
)
1256 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1257 task_rq_unlock(rq
, p
, &flags
);
1260 * If it changed from the expected state, bail out now.
1262 if (unlikely(!ncsw
))
1266 * Was it really running after all now that we
1267 * checked with the proper locks actually held?
1269 * Oops. Go back and try again..
1271 if (unlikely(running
)) {
1277 * It's not enough that it's not actively running,
1278 * it must be off the runqueue _entirely_, and not
1281 * So if it was still runnable (but just not actively
1282 * running right now), it's preempted, and we should
1283 * yield - it could be a while.
1285 if (unlikely(queued
)) {
1286 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1288 set_current_state(TASK_UNINTERRUPTIBLE
);
1289 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1294 * Ahh, all good. It wasn't running, and it wasn't
1295 * runnable, which means that it will never become
1296 * running in the future either. We're all done!
1305 * kick_process - kick a running thread to enter/exit the kernel
1306 * @p: the to-be-kicked thread
1308 * Cause a process which is running on another CPU to enter
1309 * kernel-mode, without any delay. (to get signals handled.)
1311 * NOTE: this function doesn't have to take the runqueue lock,
1312 * because all it wants to ensure is that the remote task enters
1313 * the kernel. If the IPI races and the task has been migrated
1314 * to another CPU then no harm is done and the purpose has been
1317 void kick_process(struct task_struct
*p
)
1323 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1324 smp_send_reschedule(cpu
);
1327 EXPORT_SYMBOL_GPL(kick_process
);
1328 #endif /* CONFIG_SMP */
1332 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1334 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1336 int nid
= cpu_to_node(cpu
);
1337 const struct cpumask
*nodemask
= NULL
;
1338 enum { cpuset
, possible
, fail
} state
= cpuset
;
1342 * If the node that the cpu is on has been offlined, cpu_to_node()
1343 * will return -1. There is no cpu on the node, and we should
1344 * select the cpu on the other node.
1347 nodemask
= cpumask_of_node(nid
);
1349 /* Look for allowed, online CPU in same node. */
1350 for_each_cpu(dest_cpu
, nodemask
) {
1351 if (!cpu_online(dest_cpu
))
1353 if (!cpu_active(dest_cpu
))
1355 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1361 /* Any allowed, online CPU? */
1362 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1363 if (!cpu_online(dest_cpu
))
1365 if (!cpu_active(dest_cpu
))
1372 /* No more Mr. Nice Guy. */
1373 cpuset_cpus_allowed_fallback(p
);
1378 do_set_cpus_allowed(p
, cpu_possible_mask
);
1389 if (state
!= cpuset
) {
1391 * Don't tell them about moving exiting tasks or
1392 * kernel threads (both mm NULL), since they never
1395 if (p
->mm
&& printk_ratelimit()) {
1396 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1397 task_pid_nr(p
), p
->comm
, cpu
);
1405 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1408 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1410 if (p
->nr_cpus_allowed
> 1)
1411 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1414 * In order not to call set_task_cpu() on a blocking task we need
1415 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1418 * Since this is common to all placement strategies, this lives here.
1420 * [ this allows ->select_task() to simply return task_cpu(p) and
1421 * not worry about this generic constraint ]
1423 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1425 cpu
= select_fallback_rq(task_cpu(p
), p
);
1430 static void update_avg(u64
*avg
, u64 sample
)
1432 s64 diff
= sample
- *avg
;
1438 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1440 #ifdef CONFIG_SCHEDSTATS
1441 struct rq
*rq
= this_rq();
1444 int this_cpu
= smp_processor_id();
1446 if (cpu
== this_cpu
) {
1447 schedstat_inc(rq
, ttwu_local
);
1448 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1450 struct sched_domain
*sd
;
1452 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1454 for_each_domain(this_cpu
, sd
) {
1455 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1456 schedstat_inc(sd
, ttwu_wake_remote
);
1463 if (wake_flags
& WF_MIGRATED
)
1464 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1466 #endif /* CONFIG_SMP */
1468 schedstat_inc(rq
, ttwu_count
);
1469 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1471 if (wake_flags
& WF_SYNC
)
1472 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1474 #endif /* CONFIG_SCHEDSTATS */
1477 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1479 activate_task(rq
, p
, en_flags
);
1480 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1482 /* if a worker is waking up, notify workqueue */
1483 if (p
->flags
& PF_WQ_WORKER
)
1484 wq_worker_waking_up(p
, cpu_of(rq
));
1488 * Mark the task runnable and perform wakeup-preemption.
1491 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1493 check_preempt_curr(rq
, p
, wake_flags
);
1494 trace_sched_wakeup(p
, true);
1496 p
->state
= TASK_RUNNING
;
1498 if (p
->sched_class
->task_woken
) {
1500 * XXX can drop rq->lock; most likely ok.
1502 p
->sched_class
->task_woken(rq
, p
);
1505 if (rq
->idle_stamp
) {
1506 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1507 u64 max
= 2*rq
->max_idle_balance_cost
;
1509 update_avg(&rq
->avg_idle
, delta
);
1511 if (rq
->avg_idle
> max
)
1520 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1523 if (p
->sched_contributes_to_load
)
1524 rq
->nr_uninterruptible
--;
1527 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1528 ttwu_do_wakeup(rq
, p
, wake_flags
);
1532 * Called in case the task @p isn't fully descheduled from its runqueue,
1533 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1534 * since all we need to do is flip p->state to TASK_RUNNING, since
1535 * the task is still ->on_rq.
1537 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1542 rq
= __task_rq_lock(p
);
1543 if (task_on_rq_queued(p
)) {
1544 /* check_preempt_curr() may use rq clock */
1545 update_rq_clock(rq
);
1546 ttwu_do_wakeup(rq
, p
, wake_flags
);
1549 __task_rq_unlock(rq
);
1555 void sched_ttwu_pending(void)
1557 struct rq
*rq
= this_rq();
1558 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1559 struct task_struct
*p
;
1560 unsigned long flags
;
1565 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1568 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1569 llist
= llist_next(llist
);
1570 ttwu_do_activate(rq
, p
, 0);
1573 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1576 void scheduler_ipi(void)
1579 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1580 * TIF_NEED_RESCHED remotely (for the first time) will also send
1583 preempt_fold_need_resched();
1585 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1589 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1590 * traditionally all their work was done from the interrupt return
1591 * path. Now that we actually do some work, we need to make sure
1594 * Some archs already do call them, luckily irq_enter/exit nest
1597 * Arguably we should visit all archs and update all handlers,
1598 * however a fair share of IPIs are still resched only so this would
1599 * somewhat pessimize the simple resched case.
1602 sched_ttwu_pending();
1605 * Check if someone kicked us for doing the nohz idle load balance.
1607 if (unlikely(got_nohz_idle_kick())) {
1608 this_rq()->idle_balance
= 1;
1609 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1614 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1616 struct rq
*rq
= cpu_rq(cpu
);
1618 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1619 if (!set_nr_if_polling(rq
->idle
))
1620 smp_send_reschedule(cpu
);
1622 trace_sched_wake_idle_without_ipi(cpu
);
1626 void wake_up_if_idle(int cpu
)
1628 struct rq
*rq
= cpu_rq(cpu
);
1629 unsigned long flags
;
1633 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1636 if (set_nr_if_polling(rq
->idle
)) {
1637 trace_sched_wake_idle_without_ipi(cpu
);
1639 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1640 if (is_idle_task(rq
->curr
))
1641 smp_send_reschedule(cpu
);
1642 /* Else cpu is not in idle, do nothing here */
1643 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1650 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1652 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1654 #endif /* CONFIG_SMP */
1656 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1658 struct rq
*rq
= cpu_rq(cpu
);
1660 #if defined(CONFIG_SMP)
1661 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1662 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1663 ttwu_queue_remote(p
, cpu
);
1668 raw_spin_lock(&rq
->lock
);
1669 ttwu_do_activate(rq
, p
, 0);
1670 raw_spin_unlock(&rq
->lock
);
1674 * try_to_wake_up - wake up a thread
1675 * @p: the thread to be awakened
1676 * @state: the mask of task states that can be woken
1677 * @wake_flags: wake modifier flags (WF_*)
1679 * Put it on the run-queue if it's not already there. The "current"
1680 * thread is always on the run-queue (except when the actual
1681 * re-schedule is in progress), and as such you're allowed to do
1682 * the simpler "current->state = TASK_RUNNING" to mark yourself
1683 * runnable without the overhead of this.
1685 * Return: %true if @p was woken up, %false if it was already running.
1686 * or @state didn't match @p's state.
1689 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1691 unsigned long flags
;
1692 int cpu
, success
= 0;
1695 * If we are going to wake up a thread waiting for CONDITION we
1696 * need to ensure that CONDITION=1 done by the caller can not be
1697 * reordered with p->state check below. This pairs with mb() in
1698 * set_current_state() the waiting thread does.
1700 smp_mb__before_spinlock();
1701 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1702 if (!(p
->state
& state
))
1705 success
= 1; /* we're going to change ->state */
1708 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1713 * If the owning (remote) cpu is still in the middle of schedule() with
1714 * this task as prev, wait until its done referencing the task.
1719 * Pairs with the smp_wmb() in finish_lock_switch().
1723 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
1724 p
->state
= TASK_WAKING
;
1726 if (p
->sched_class
->task_waking
)
1727 p
->sched_class
->task_waking(p
);
1729 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
1730 if (task_cpu(p
) != cpu
) {
1731 wake_flags
|= WF_MIGRATED
;
1732 set_task_cpu(p
, cpu
);
1734 #endif /* CONFIG_SMP */
1738 ttwu_stat(p
, cpu
, wake_flags
);
1740 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1746 * try_to_wake_up_local - try to wake up a local task with rq lock held
1747 * @p: the thread to be awakened
1749 * Put @p on the run-queue if it's not already there. The caller must
1750 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1753 static void try_to_wake_up_local(struct task_struct
*p
)
1755 struct rq
*rq
= task_rq(p
);
1757 if (WARN_ON_ONCE(rq
!= this_rq()) ||
1758 WARN_ON_ONCE(p
== current
))
1761 lockdep_assert_held(&rq
->lock
);
1763 if (!raw_spin_trylock(&p
->pi_lock
)) {
1764 raw_spin_unlock(&rq
->lock
);
1765 raw_spin_lock(&p
->pi_lock
);
1766 raw_spin_lock(&rq
->lock
);
1769 if (!(p
->state
& TASK_NORMAL
))
1772 if (!task_on_rq_queued(p
))
1773 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
1775 ttwu_do_wakeup(rq
, p
, 0);
1776 ttwu_stat(p
, smp_processor_id(), 0);
1778 raw_spin_unlock(&p
->pi_lock
);
1782 * wake_up_process - Wake up a specific process
1783 * @p: The process to be woken up.
1785 * Attempt to wake up the nominated process and move it to the set of runnable
1788 * Return: 1 if the process was woken up, 0 if it was already running.
1790 * It may be assumed that this function implies a write memory barrier before
1791 * changing the task state if and only if any tasks are woken up.
1793 int wake_up_process(struct task_struct
*p
)
1795 WARN_ON(task_is_stopped_or_traced(p
));
1796 return try_to_wake_up(p
, TASK_NORMAL
, 0);
1798 EXPORT_SYMBOL(wake_up_process
);
1800 int wake_up_state(struct task_struct
*p
, unsigned int state
)
1802 return try_to_wake_up(p
, state
, 0);
1806 * This function clears the sched_dl_entity static params.
1808 void __dl_clear_params(struct task_struct
*p
)
1810 struct sched_dl_entity
*dl_se
= &p
->dl
;
1812 dl_se
->dl_runtime
= 0;
1813 dl_se
->dl_deadline
= 0;
1814 dl_se
->dl_period
= 0;
1818 dl_se
->dl_throttled
= 0;
1820 dl_se
->dl_yielded
= 0;
1824 * Perform scheduler related setup for a newly forked process p.
1825 * p is forked by current.
1827 * __sched_fork() is basic setup used by init_idle() too:
1829 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1834 p
->se
.exec_start
= 0;
1835 p
->se
.sum_exec_runtime
= 0;
1836 p
->se
.prev_sum_exec_runtime
= 0;
1837 p
->se
.nr_migrations
= 0;
1840 p
->se
.avg
.decay_count
= 0;
1842 INIT_LIST_HEAD(&p
->se
.group_node
);
1844 #ifdef CONFIG_SCHEDSTATS
1845 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
1848 RB_CLEAR_NODE(&p
->dl
.rb_node
);
1849 init_dl_task_timer(&p
->dl
);
1850 __dl_clear_params(p
);
1852 INIT_LIST_HEAD(&p
->rt
.run_list
);
1854 #ifdef CONFIG_PREEMPT_NOTIFIERS
1855 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
1858 #ifdef CONFIG_NUMA_BALANCING
1859 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
1860 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1861 p
->mm
->numa_scan_seq
= 0;
1864 if (clone_flags
& CLONE_VM
)
1865 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
1867 p
->numa_preferred_nid
= -1;
1869 p
->node_stamp
= 0ULL;
1870 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
1871 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
1872 p
->numa_work
.next
= &p
->numa_work
;
1873 p
->numa_faults
= NULL
;
1874 p
->last_task_numa_placement
= 0;
1875 p
->last_sum_exec_runtime
= 0;
1877 p
->numa_group
= NULL
;
1878 #endif /* CONFIG_NUMA_BALANCING */
1881 #ifdef CONFIG_NUMA_BALANCING
1882 #ifdef CONFIG_SCHED_DEBUG
1883 void set_numabalancing_state(bool enabled
)
1886 sched_feat_set("NUMA");
1888 sched_feat_set("NO_NUMA");
1891 __read_mostly
bool numabalancing_enabled
;
1893 void set_numabalancing_state(bool enabled
)
1895 numabalancing_enabled
= enabled
;
1897 #endif /* CONFIG_SCHED_DEBUG */
1899 #ifdef CONFIG_PROC_SYSCTL
1900 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
1901 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1905 int state
= numabalancing_enabled
;
1907 if (write
&& !capable(CAP_SYS_ADMIN
))
1912 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
1916 set_numabalancing_state(state
);
1923 * fork()/clone()-time setup:
1925 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
1927 unsigned long flags
;
1928 int cpu
= get_cpu();
1930 __sched_fork(clone_flags
, p
);
1932 * We mark the process as running here. This guarantees that
1933 * nobody will actually run it, and a signal or other external
1934 * event cannot wake it up and insert it on the runqueue either.
1936 p
->state
= TASK_RUNNING
;
1939 * Make sure we do not leak PI boosting priority to the child.
1941 p
->prio
= current
->normal_prio
;
1944 * Revert to default priority/policy on fork if requested.
1946 if (unlikely(p
->sched_reset_on_fork
)) {
1947 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
1948 p
->policy
= SCHED_NORMAL
;
1949 p
->static_prio
= NICE_TO_PRIO(0);
1951 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
1952 p
->static_prio
= NICE_TO_PRIO(0);
1954 p
->prio
= p
->normal_prio
= __normal_prio(p
);
1958 * We don't need the reset flag anymore after the fork. It has
1959 * fulfilled its duty:
1961 p
->sched_reset_on_fork
= 0;
1964 if (dl_prio(p
->prio
)) {
1967 } else if (rt_prio(p
->prio
)) {
1968 p
->sched_class
= &rt_sched_class
;
1970 p
->sched_class
= &fair_sched_class
;
1973 if (p
->sched_class
->task_fork
)
1974 p
->sched_class
->task_fork(p
);
1977 * The child is not yet in the pid-hash so no cgroup attach races,
1978 * and the cgroup is pinned to this child due to cgroup_fork()
1979 * is ran before sched_fork().
1981 * Silence PROVE_RCU.
1983 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1984 set_task_cpu(p
, cpu
);
1985 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
1987 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1988 if (likely(sched_info_on()))
1989 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1991 #if defined(CONFIG_SMP)
1994 init_task_preempt_count(p
);
1996 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
1997 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2004 unsigned long to_ratio(u64 period
, u64 runtime
)
2006 if (runtime
== RUNTIME_INF
)
2010 * Doing this here saves a lot of checks in all
2011 * the calling paths, and returning zero seems
2012 * safe for them anyway.
2017 return div64_u64(runtime
<< 20, period
);
2021 inline struct dl_bw
*dl_bw_of(int i
)
2023 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2024 "sched RCU must be held");
2025 return &cpu_rq(i
)->rd
->dl_bw
;
2028 static inline int dl_bw_cpus(int i
)
2030 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2033 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2034 "sched RCU must be held");
2035 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2041 inline struct dl_bw
*dl_bw_of(int i
)
2043 return &cpu_rq(i
)->dl
.dl_bw
;
2046 static inline int dl_bw_cpus(int i
)
2053 * We must be sure that accepting a new task (or allowing changing the
2054 * parameters of an existing one) is consistent with the bandwidth
2055 * constraints. If yes, this function also accordingly updates the currently
2056 * allocated bandwidth to reflect the new situation.
2058 * This function is called while holding p's rq->lock.
2060 * XXX we should delay bw change until the task's 0-lag point, see
2063 static int dl_overflow(struct task_struct
*p
, int policy
,
2064 const struct sched_attr
*attr
)
2067 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2068 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2069 u64 runtime
= attr
->sched_runtime
;
2070 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2073 if (new_bw
== p
->dl
.dl_bw
)
2077 * Either if a task, enters, leave, or stays -deadline but changes
2078 * its parameters, we may need to update accordingly the total
2079 * allocated bandwidth of the container.
2081 raw_spin_lock(&dl_b
->lock
);
2082 cpus
= dl_bw_cpus(task_cpu(p
));
2083 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2084 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2085 __dl_add(dl_b
, new_bw
);
2087 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2088 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2089 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2090 __dl_add(dl_b
, new_bw
);
2092 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2093 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2096 raw_spin_unlock(&dl_b
->lock
);
2101 extern void init_dl_bw(struct dl_bw
*dl_b
);
2104 * wake_up_new_task - wake up a newly created task for the first time.
2106 * This function will do some initial scheduler statistics housekeeping
2107 * that must be done for every newly created context, then puts the task
2108 * on the runqueue and wakes it.
2110 void wake_up_new_task(struct task_struct
*p
)
2112 unsigned long flags
;
2115 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2118 * Fork balancing, do it here and not earlier because:
2119 * - cpus_allowed can change in the fork path
2120 * - any previously selected cpu might disappear through hotplug
2122 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2125 /* Initialize new task's runnable average */
2126 init_task_runnable_average(p
);
2127 rq
= __task_rq_lock(p
);
2128 activate_task(rq
, p
, 0);
2129 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2130 trace_sched_wakeup_new(p
, true);
2131 check_preempt_curr(rq
, p
, WF_FORK
);
2133 if (p
->sched_class
->task_woken
)
2134 p
->sched_class
->task_woken(rq
, p
);
2136 task_rq_unlock(rq
, p
, &flags
);
2139 #ifdef CONFIG_PREEMPT_NOTIFIERS
2142 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2143 * @notifier: notifier struct to register
2145 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2147 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2149 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2152 * preempt_notifier_unregister - no longer interested in preemption notifications
2153 * @notifier: notifier struct to unregister
2155 * This is safe to call from within a preemption notifier.
2157 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2159 hlist_del(¬ifier
->link
);
2161 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2163 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2165 struct preempt_notifier
*notifier
;
2167 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2168 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2172 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2173 struct task_struct
*next
)
2175 struct preempt_notifier
*notifier
;
2177 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2178 notifier
->ops
->sched_out(notifier
, next
);
2181 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2183 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2188 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2189 struct task_struct
*next
)
2193 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2196 * prepare_task_switch - prepare to switch tasks
2197 * @rq: the runqueue preparing to switch
2198 * @prev: the current task that is being switched out
2199 * @next: the task we are going to switch to.
2201 * This is called with the rq lock held and interrupts off. It must
2202 * be paired with a subsequent finish_task_switch after the context
2205 * prepare_task_switch sets up locking and calls architecture specific
2209 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2210 struct task_struct
*next
)
2212 trace_sched_switch(prev
, next
);
2213 sched_info_switch(rq
, prev
, next
);
2214 perf_event_task_sched_out(prev
, next
);
2215 fire_sched_out_preempt_notifiers(prev
, next
);
2216 prepare_lock_switch(rq
, next
);
2217 prepare_arch_switch(next
);
2221 * finish_task_switch - clean up after a task-switch
2222 * @prev: the thread we just switched away from.
2224 * finish_task_switch must be called after the context switch, paired
2225 * with a prepare_task_switch call before the context switch.
2226 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2227 * and do any other architecture-specific cleanup actions.
2229 * Note that we may have delayed dropping an mm in context_switch(). If
2230 * so, we finish that here outside of the runqueue lock. (Doing it
2231 * with the lock held can cause deadlocks; see schedule() for
2234 * The context switch have flipped the stack from under us and restored the
2235 * local variables which were saved when this task called schedule() in the
2236 * past. prev == current is still correct but we need to recalculate this_rq
2237 * because prev may have moved to another CPU.
2239 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2240 __releases(rq
->lock
)
2242 struct rq
*rq
= this_rq();
2243 struct mm_struct
*mm
= rq
->prev_mm
;
2249 * A task struct has one reference for the use as "current".
2250 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2251 * schedule one last time. The schedule call will never return, and
2252 * the scheduled task must drop that reference.
2253 * The test for TASK_DEAD must occur while the runqueue locks are
2254 * still held, otherwise prev could be scheduled on another cpu, die
2255 * there before we look at prev->state, and then the reference would
2257 * Manfred Spraul <manfred@colorfullife.com>
2259 prev_state
= prev
->state
;
2260 vtime_task_switch(prev
);
2261 finish_arch_switch(prev
);
2262 perf_event_task_sched_in(prev
, current
);
2263 finish_lock_switch(rq
, prev
);
2264 finish_arch_post_lock_switch();
2266 fire_sched_in_preempt_notifiers(current
);
2269 if (unlikely(prev_state
== TASK_DEAD
)) {
2270 if (prev
->sched_class
->task_dead
)
2271 prev
->sched_class
->task_dead(prev
);
2274 * Remove function-return probe instances associated with this
2275 * task and put them back on the free list.
2277 kprobe_flush_task(prev
);
2278 put_task_struct(prev
);
2281 tick_nohz_task_switch(current
);
2287 /* rq->lock is NOT held, but preemption is disabled */
2288 static void __balance_callback(struct rq
*rq
)
2290 struct callback_head
*head
, *next
;
2291 void (*func
)(struct rq
*rq
);
2292 unsigned long flags
;
2294 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2295 head
= rq
->balance_callback
;
2296 rq
->balance_callback
= NULL
;
2298 func
= (void (*)(struct rq
*))head
->func
;
2305 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2308 static inline void balance_callback(struct rq
*rq
)
2310 if (unlikely(rq
->balance_callback
))
2311 __balance_callback(rq
);
2316 static inline void balance_callback(struct rq
*rq
)
2323 * schedule_tail - first thing a freshly forked thread must call.
2324 * @prev: the thread we just switched away from.
2326 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2327 __releases(rq
->lock
)
2331 /* finish_task_switch() drops rq->lock and enables preemtion */
2333 rq
= finish_task_switch(prev
);
2334 balance_callback(rq
);
2337 if (current
->set_child_tid
)
2338 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2342 * context_switch - switch to the new MM and the new thread's register state.
2344 static inline struct rq
*
2345 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2346 struct task_struct
*next
)
2348 struct mm_struct
*mm
, *oldmm
;
2350 prepare_task_switch(rq
, prev
, next
);
2353 oldmm
= prev
->active_mm
;
2355 * For paravirt, this is coupled with an exit in switch_to to
2356 * combine the page table reload and the switch backend into
2359 arch_start_context_switch(prev
);
2362 next
->active_mm
= oldmm
;
2363 atomic_inc(&oldmm
->mm_count
);
2364 enter_lazy_tlb(oldmm
, next
);
2366 switch_mm(oldmm
, mm
, next
);
2369 prev
->active_mm
= NULL
;
2370 rq
->prev_mm
= oldmm
;
2373 * Since the runqueue lock will be released by the next
2374 * task (which is an invalid locking op but in the case
2375 * of the scheduler it's an obvious special-case), so we
2376 * do an early lockdep release here:
2378 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2380 context_tracking_task_switch(prev
, next
);
2381 /* Here we just switch the register state and the stack. */
2382 switch_to(prev
, next
, prev
);
2385 return finish_task_switch(prev
);
2389 * nr_running and nr_context_switches:
2391 * externally visible scheduler statistics: current number of runnable
2392 * threads, total number of context switches performed since bootup.
2394 unsigned long nr_running(void)
2396 unsigned long i
, sum
= 0;
2398 for_each_online_cpu(i
)
2399 sum
+= cpu_rq(i
)->nr_running
;
2405 * Check if only the current task is running on the cpu.
2407 bool single_task_running(void)
2409 if (cpu_rq(smp_processor_id())->nr_running
== 1)
2414 EXPORT_SYMBOL(single_task_running
);
2416 unsigned long long nr_context_switches(void)
2419 unsigned long long sum
= 0;
2421 for_each_possible_cpu(i
)
2422 sum
+= cpu_rq(i
)->nr_switches
;
2427 unsigned long nr_iowait(void)
2429 unsigned long i
, sum
= 0;
2431 for_each_possible_cpu(i
)
2432 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2437 unsigned long nr_iowait_cpu(int cpu
)
2439 struct rq
*this = cpu_rq(cpu
);
2440 return atomic_read(&this->nr_iowait
);
2443 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2445 struct rq
*rq
= this_rq();
2446 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2447 *load
= rq
->load
.weight
;
2453 * sched_exec - execve() is a valuable balancing opportunity, because at
2454 * this point the task has the smallest effective memory and cache footprint.
2456 void sched_exec(void)
2458 struct task_struct
*p
= current
;
2459 unsigned long flags
;
2462 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2463 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2464 if (dest_cpu
== smp_processor_id())
2467 if (likely(cpu_active(dest_cpu
))) {
2468 struct migration_arg arg
= { p
, dest_cpu
};
2470 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2471 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2475 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2480 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2481 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2483 EXPORT_PER_CPU_SYMBOL(kstat
);
2484 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2487 * Return accounted runtime for the task.
2488 * In case the task is currently running, return the runtime plus current's
2489 * pending runtime that have not been accounted yet.
2491 unsigned long long task_sched_runtime(struct task_struct
*p
)
2493 unsigned long flags
;
2497 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2499 * 64-bit doesn't need locks to atomically read a 64bit value.
2500 * So we have a optimization chance when the task's delta_exec is 0.
2501 * Reading ->on_cpu is racy, but this is ok.
2503 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2504 * If we race with it entering cpu, unaccounted time is 0. This is
2505 * indistinguishable from the read occurring a few cycles earlier.
2506 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2507 * been accounted, so we're correct here as well.
2509 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2510 return p
->se
.sum_exec_runtime
;
2513 rq
= task_rq_lock(p
, &flags
);
2515 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2516 * project cycles that may never be accounted to this
2517 * thread, breaking clock_gettime().
2519 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2520 update_rq_clock(rq
);
2521 p
->sched_class
->update_curr(rq
);
2523 ns
= p
->se
.sum_exec_runtime
;
2524 task_rq_unlock(rq
, p
, &flags
);
2530 * This function gets called by the timer code, with HZ frequency.
2531 * We call it with interrupts disabled.
2533 void scheduler_tick(void)
2535 int cpu
= smp_processor_id();
2536 struct rq
*rq
= cpu_rq(cpu
);
2537 struct task_struct
*curr
= rq
->curr
;
2541 raw_spin_lock(&rq
->lock
);
2542 update_rq_clock(rq
);
2543 curr
->sched_class
->task_tick(rq
, curr
, 0);
2544 update_cpu_load_active(rq
);
2545 calc_global_load_tick(rq
);
2546 raw_spin_unlock(&rq
->lock
);
2548 perf_event_task_tick();
2551 rq
->idle_balance
= idle_cpu(cpu
);
2552 trigger_load_balance(rq
);
2554 rq_last_tick_reset(rq
);
2557 #ifdef CONFIG_NO_HZ_FULL
2559 * scheduler_tick_max_deferment
2561 * Keep at least one tick per second when a single
2562 * active task is running because the scheduler doesn't
2563 * yet completely support full dynticks environment.
2565 * This makes sure that uptime, CFS vruntime, load
2566 * balancing, etc... continue to move forward, even
2567 * with a very low granularity.
2569 * Return: Maximum deferment in nanoseconds.
2571 u64
scheduler_tick_max_deferment(void)
2573 struct rq
*rq
= this_rq();
2574 unsigned long next
, now
= READ_ONCE(jiffies
);
2576 next
= rq
->last_sched_tick
+ HZ
;
2578 if (time_before_eq(next
, now
))
2581 return jiffies_to_nsecs(next
- now
);
2585 notrace
unsigned long get_parent_ip(unsigned long addr
)
2587 if (in_lock_functions(addr
)) {
2588 addr
= CALLER_ADDR2
;
2589 if (in_lock_functions(addr
))
2590 addr
= CALLER_ADDR3
;
2595 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2596 defined(CONFIG_PREEMPT_TRACER))
2598 void preempt_count_add(int val
)
2600 #ifdef CONFIG_DEBUG_PREEMPT
2604 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2607 __preempt_count_add(val
);
2608 #ifdef CONFIG_DEBUG_PREEMPT
2610 * Spinlock count overflowing soon?
2612 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
2615 if (preempt_count() == val
) {
2616 unsigned long ip
= get_parent_ip(CALLER_ADDR1
);
2617 #ifdef CONFIG_DEBUG_PREEMPT
2618 current
->preempt_disable_ip
= ip
;
2620 trace_preempt_off(CALLER_ADDR0
, ip
);
2623 EXPORT_SYMBOL(preempt_count_add
);
2624 NOKPROBE_SYMBOL(preempt_count_add
);
2626 void preempt_count_sub(int val
)
2628 #ifdef CONFIG_DEBUG_PREEMPT
2632 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
2635 * Is the spinlock portion underflowing?
2637 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
2638 !(preempt_count() & PREEMPT_MASK
)))
2642 if (preempt_count() == val
)
2643 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
2644 __preempt_count_sub(val
);
2646 EXPORT_SYMBOL(preempt_count_sub
);
2647 NOKPROBE_SYMBOL(preempt_count_sub
);
2652 * Print scheduling while atomic bug:
2654 static noinline
void __schedule_bug(struct task_struct
*prev
)
2656 if (oops_in_progress
)
2659 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
2660 prev
->comm
, prev
->pid
, preempt_count());
2662 debug_show_held_locks(prev
);
2664 if (irqs_disabled())
2665 print_irqtrace_events(prev
);
2666 #ifdef CONFIG_DEBUG_PREEMPT
2667 if (in_atomic_preempt_off()) {
2668 pr_err("Preemption disabled at:");
2669 print_ip_sym(current
->preempt_disable_ip
);
2674 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
2678 * Various schedule()-time debugging checks and statistics:
2680 static inline void schedule_debug(struct task_struct
*prev
)
2682 #ifdef CONFIG_SCHED_STACK_END_CHECK
2683 BUG_ON(unlikely(task_stack_end_corrupted(prev
)));
2686 * Test if we are atomic. Since do_exit() needs to call into
2687 * schedule() atomically, we ignore that path. Otherwise whine
2688 * if we are scheduling when we should not.
2690 if (unlikely(in_atomic_preempt_off() && prev
->state
!= TASK_DEAD
))
2691 __schedule_bug(prev
);
2694 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
2696 schedstat_inc(this_rq(), sched_count
);
2700 * Pick up the highest-prio task:
2702 static inline struct task_struct
*
2703 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
2705 const struct sched_class
*class = &fair_sched_class
;
2706 struct task_struct
*p
;
2709 * Optimization: we know that if all tasks are in
2710 * the fair class we can call that function directly:
2712 if (likely(prev
->sched_class
== class &&
2713 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
2714 p
= fair_sched_class
.pick_next_task(rq
, prev
);
2715 if (unlikely(p
== RETRY_TASK
))
2718 /* assumes fair_sched_class->next == idle_sched_class */
2720 p
= idle_sched_class
.pick_next_task(rq
, prev
);
2726 for_each_class(class) {
2727 p
= class->pick_next_task(rq
, prev
);
2729 if (unlikely(p
== RETRY_TASK
))
2735 BUG(); /* the idle class will always have a runnable task */
2739 * __schedule() is the main scheduler function.
2741 * The main means of driving the scheduler and thus entering this function are:
2743 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2745 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2746 * paths. For example, see arch/x86/entry_64.S.
2748 * To drive preemption between tasks, the scheduler sets the flag in timer
2749 * interrupt handler scheduler_tick().
2751 * 3. Wakeups don't really cause entry into schedule(). They add a
2752 * task to the run-queue and that's it.
2754 * Now, if the new task added to the run-queue preempts the current
2755 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2756 * called on the nearest possible occasion:
2758 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2760 * - in syscall or exception context, at the next outmost
2761 * preempt_enable(). (this might be as soon as the wake_up()'s
2764 * - in IRQ context, return from interrupt-handler to
2765 * preemptible context
2767 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2770 * - cond_resched() call
2771 * - explicit schedule() call
2772 * - return from syscall or exception to user-space
2773 * - return from interrupt-handler to user-space
2775 * WARNING: must be called with preemption disabled!
2777 static void __sched
__schedule(void)
2779 struct task_struct
*prev
, *next
;
2780 unsigned long *switch_count
;
2784 cpu
= smp_processor_id();
2786 rcu_note_context_switch();
2789 schedule_debug(prev
);
2791 if (sched_feat(HRTICK
))
2795 * Make sure that signal_pending_state()->signal_pending() below
2796 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2797 * done by the caller to avoid the race with signal_wake_up().
2799 smp_mb__before_spinlock();
2800 raw_spin_lock_irq(&rq
->lock
);
2802 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
2804 switch_count
= &prev
->nivcsw
;
2805 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
2806 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
2807 prev
->state
= TASK_RUNNING
;
2809 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
2813 * If a worker went to sleep, notify and ask workqueue
2814 * whether it wants to wake up a task to maintain
2817 if (prev
->flags
& PF_WQ_WORKER
) {
2818 struct task_struct
*to_wakeup
;
2820 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
2822 try_to_wake_up_local(to_wakeup
);
2825 switch_count
= &prev
->nvcsw
;
2828 if (task_on_rq_queued(prev
))
2829 update_rq_clock(rq
);
2831 next
= pick_next_task(rq
, prev
);
2832 clear_tsk_need_resched(prev
);
2833 clear_preempt_need_resched();
2834 rq
->clock_skip_update
= 0;
2836 if (likely(prev
!= next
)) {
2841 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
2844 raw_spin_unlock_irq(&rq
->lock
);
2846 balance_callback(rq
);
2849 static inline void sched_submit_work(struct task_struct
*tsk
)
2851 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
2854 * If we are going to sleep and we have plugged IO queued,
2855 * make sure to submit it to avoid deadlocks.
2857 if (blk_needs_flush_plug(tsk
))
2858 blk_schedule_flush_plug(tsk
);
2861 asmlinkage __visible
void __sched
schedule(void)
2863 struct task_struct
*tsk
= current
;
2865 sched_submit_work(tsk
);
2869 sched_preempt_enable_no_resched();
2870 } while (need_resched());
2872 EXPORT_SYMBOL(schedule
);
2874 #ifdef CONFIG_CONTEXT_TRACKING
2875 asmlinkage __visible
void __sched
schedule_user(void)
2878 * If we come here after a random call to set_need_resched(),
2879 * or we have been woken up remotely but the IPI has not yet arrived,
2880 * we haven't yet exited the RCU idle mode. Do it here manually until
2881 * we find a better solution.
2883 * NB: There are buggy callers of this function. Ideally we
2884 * should warn if prev_state != CONTEXT_USER, but that will trigger
2885 * too frequently to make sense yet.
2887 enum ctx_state prev_state
= exception_enter();
2889 exception_exit(prev_state
);
2894 * schedule_preempt_disabled - called with preemption disabled
2896 * Returns with preemption disabled. Note: preempt_count must be 1
2898 void __sched
schedule_preempt_disabled(void)
2900 sched_preempt_enable_no_resched();
2905 static void __sched notrace
preempt_schedule_common(void)
2908 preempt_active_enter();
2910 preempt_active_exit();
2913 * Check again in case we missed a preemption opportunity
2914 * between schedule and now.
2916 } while (need_resched());
2919 #ifdef CONFIG_PREEMPT
2921 * this is the entry point to schedule() from in-kernel preemption
2922 * off of preempt_enable. Kernel preemptions off return from interrupt
2923 * occur there and call schedule directly.
2925 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
2928 * If there is a non-zero preempt_count or interrupts are disabled,
2929 * we do not want to preempt the current task. Just return..
2931 if (likely(!preemptible()))
2934 preempt_schedule_common();
2936 NOKPROBE_SYMBOL(preempt_schedule
);
2937 EXPORT_SYMBOL(preempt_schedule
);
2940 * preempt_schedule_notrace - preempt_schedule called by tracing
2942 * The tracing infrastructure uses preempt_enable_notrace to prevent
2943 * recursion and tracing preempt enabling caused by the tracing
2944 * infrastructure itself. But as tracing can happen in areas coming
2945 * from userspace or just about to enter userspace, a preempt enable
2946 * can occur before user_exit() is called. This will cause the scheduler
2947 * to be called when the system is still in usermode.
2949 * To prevent this, the preempt_enable_notrace will use this function
2950 * instead of preempt_schedule() to exit user context if needed before
2951 * calling the scheduler.
2953 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
2955 enum ctx_state prev_ctx
;
2957 if (likely(!preemptible()))
2962 * Use raw __prempt_count() ops that don't call function.
2963 * We can't call functions before disabling preemption which
2964 * disarm preemption tracing recursions.
2966 __preempt_count_add(PREEMPT_ACTIVE
+ PREEMPT_DISABLE_OFFSET
);
2969 * Needs preempt disabled in case user_exit() is traced
2970 * and the tracer calls preempt_enable_notrace() causing
2971 * an infinite recursion.
2973 prev_ctx
= exception_enter();
2975 exception_exit(prev_ctx
);
2978 __preempt_count_sub(PREEMPT_ACTIVE
+ PREEMPT_DISABLE_OFFSET
);
2979 } while (need_resched());
2981 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
2983 #endif /* CONFIG_PREEMPT */
2986 * this is the entry point to schedule() from kernel preemption
2987 * off of irq context.
2988 * Note, that this is called and return with irqs disabled. This will
2989 * protect us against recursive calling from irq.
2991 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
2993 enum ctx_state prev_state
;
2995 /* Catch callers which need to be fixed */
2996 BUG_ON(preempt_count() || !irqs_disabled());
2998 prev_state
= exception_enter();
3001 preempt_active_enter();
3004 local_irq_disable();
3005 preempt_active_exit();
3006 } while (need_resched());
3008 exception_exit(prev_state
);
3011 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3014 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3016 EXPORT_SYMBOL(default_wake_function
);
3018 #ifdef CONFIG_RT_MUTEXES
3021 * rt_mutex_setprio - set the current priority of a task
3023 * @prio: prio value (kernel-internal form)
3025 * This function changes the 'effective' priority of a task. It does
3026 * not touch ->normal_prio like __setscheduler().
3028 * Used by the rt_mutex code to implement priority inheritance
3029 * logic. Call site only calls if the priority of the task changed.
3031 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3033 int oldprio
, queued
, running
, enqueue_flag
= 0;
3035 const struct sched_class
*prev_class
;
3037 BUG_ON(prio
> MAX_PRIO
);
3039 rq
= __task_rq_lock(p
);
3042 * Idle task boosting is a nono in general. There is one
3043 * exception, when PREEMPT_RT and NOHZ is active:
3045 * The idle task calls get_next_timer_interrupt() and holds
3046 * the timer wheel base->lock on the CPU and another CPU wants
3047 * to access the timer (probably to cancel it). We can safely
3048 * ignore the boosting request, as the idle CPU runs this code
3049 * with interrupts disabled and will complete the lock
3050 * protected section without being interrupted. So there is no
3051 * real need to boost.
3053 if (unlikely(p
== rq
->idle
)) {
3054 WARN_ON(p
!= rq
->curr
);
3055 WARN_ON(p
->pi_blocked_on
);
3059 trace_sched_pi_setprio(p
, prio
);
3061 prev_class
= p
->sched_class
;
3062 queued
= task_on_rq_queued(p
);
3063 running
= task_current(rq
, p
);
3065 dequeue_task(rq
, p
, 0);
3067 put_prev_task(rq
, p
);
3070 * Boosting condition are:
3071 * 1. -rt task is running and holds mutex A
3072 * --> -dl task blocks on mutex A
3074 * 2. -dl task is running and holds mutex A
3075 * --> -dl task blocks on mutex A and could preempt the
3078 if (dl_prio(prio
)) {
3079 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3080 if (!dl_prio(p
->normal_prio
) ||
3081 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3082 p
->dl
.dl_boosted
= 1;
3083 p
->dl
.dl_throttled
= 0;
3084 enqueue_flag
= ENQUEUE_REPLENISH
;
3086 p
->dl
.dl_boosted
= 0;
3087 p
->sched_class
= &dl_sched_class
;
3088 } else if (rt_prio(prio
)) {
3089 if (dl_prio(oldprio
))
3090 p
->dl
.dl_boosted
= 0;
3092 enqueue_flag
= ENQUEUE_HEAD
;
3093 p
->sched_class
= &rt_sched_class
;
3095 if (dl_prio(oldprio
))
3096 p
->dl
.dl_boosted
= 0;
3097 if (rt_prio(oldprio
))
3099 p
->sched_class
= &fair_sched_class
;
3105 p
->sched_class
->set_curr_task(rq
);
3107 enqueue_task(rq
, p
, enqueue_flag
);
3109 check_class_changed(rq
, p
, prev_class
, oldprio
);
3111 preempt_disable(); /* avoid rq from going away on us */
3112 __task_rq_unlock(rq
);
3114 balance_callback(rq
);
3119 void set_user_nice(struct task_struct
*p
, long nice
)
3121 int old_prio
, delta
, queued
;
3122 unsigned long flags
;
3125 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3128 * We have to be careful, if called from sys_setpriority(),
3129 * the task might be in the middle of scheduling on another CPU.
3131 rq
= task_rq_lock(p
, &flags
);
3133 * The RT priorities are set via sched_setscheduler(), but we still
3134 * allow the 'normal' nice value to be set - but as expected
3135 * it wont have any effect on scheduling until the task is
3136 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3138 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3139 p
->static_prio
= NICE_TO_PRIO(nice
);
3142 queued
= task_on_rq_queued(p
);
3144 dequeue_task(rq
, p
, 0);
3146 p
->static_prio
= NICE_TO_PRIO(nice
);
3149 p
->prio
= effective_prio(p
);
3150 delta
= p
->prio
- old_prio
;
3153 enqueue_task(rq
, p
, 0);
3155 * If the task increased its priority or is running and
3156 * lowered its priority, then reschedule its CPU:
3158 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3162 task_rq_unlock(rq
, p
, &flags
);
3164 EXPORT_SYMBOL(set_user_nice
);
3167 * can_nice - check if a task can reduce its nice value
3171 int can_nice(const struct task_struct
*p
, const int nice
)
3173 /* convert nice value [19,-20] to rlimit style value [1,40] */
3174 int nice_rlim
= nice_to_rlimit(nice
);
3176 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3177 capable(CAP_SYS_NICE
));
3180 #ifdef __ARCH_WANT_SYS_NICE
3183 * sys_nice - change the priority of the current process.
3184 * @increment: priority increment
3186 * sys_setpriority is a more generic, but much slower function that
3187 * does similar things.
3189 SYSCALL_DEFINE1(nice
, int, increment
)
3194 * Setpriority might change our priority at the same moment.
3195 * We don't have to worry. Conceptually one call occurs first
3196 * and we have a single winner.
3198 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3199 nice
= task_nice(current
) + increment
;
3201 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3202 if (increment
< 0 && !can_nice(current
, nice
))
3205 retval
= security_task_setnice(current
, nice
);
3209 set_user_nice(current
, nice
);
3216 * task_prio - return the priority value of a given task.
3217 * @p: the task in question.
3219 * Return: The priority value as seen by users in /proc.
3220 * RT tasks are offset by -200. Normal tasks are centered
3221 * around 0, value goes from -16 to +15.
3223 int task_prio(const struct task_struct
*p
)
3225 return p
->prio
- MAX_RT_PRIO
;
3229 * idle_cpu - is a given cpu idle currently?
3230 * @cpu: the processor in question.
3232 * Return: 1 if the CPU is currently idle. 0 otherwise.
3234 int idle_cpu(int cpu
)
3236 struct rq
*rq
= cpu_rq(cpu
);
3238 if (rq
->curr
!= rq
->idle
)
3245 if (!llist_empty(&rq
->wake_list
))
3253 * idle_task - return the idle task for a given cpu.
3254 * @cpu: the processor in question.
3256 * Return: The idle task for the cpu @cpu.
3258 struct task_struct
*idle_task(int cpu
)
3260 return cpu_rq(cpu
)->idle
;
3264 * find_process_by_pid - find a process with a matching PID value.
3265 * @pid: the pid in question.
3267 * The task of @pid, if found. %NULL otherwise.
3269 static struct task_struct
*find_process_by_pid(pid_t pid
)
3271 return pid
? find_task_by_vpid(pid
) : current
;
3275 * This function initializes the sched_dl_entity of a newly becoming
3276 * SCHED_DEADLINE task.
3278 * Only the static values are considered here, the actual runtime and the
3279 * absolute deadline will be properly calculated when the task is enqueued
3280 * for the first time with its new policy.
3283 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3285 struct sched_dl_entity
*dl_se
= &p
->dl
;
3287 dl_se
->dl_runtime
= attr
->sched_runtime
;
3288 dl_se
->dl_deadline
= attr
->sched_deadline
;
3289 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3290 dl_se
->flags
= attr
->sched_flags
;
3291 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3294 * Changing the parameters of a task is 'tricky' and we're not doing
3295 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3297 * What we SHOULD do is delay the bandwidth release until the 0-lag
3298 * point. This would include retaining the task_struct until that time
3299 * and change dl_overflow() to not immediately decrement the current
3302 * Instead we retain the current runtime/deadline and let the new
3303 * parameters take effect after the current reservation period lapses.
3304 * This is safe (albeit pessimistic) because the 0-lag point is always
3305 * before the current scheduling deadline.
3307 * We can still have temporary overloads because we do not delay the
3308 * change in bandwidth until that time; so admission control is
3309 * not on the safe side. It does however guarantee tasks will never
3310 * consume more than promised.
3315 * sched_setparam() passes in -1 for its policy, to let the functions
3316 * it calls know not to change it.
3318 #define SETPARAM_POLICY -1
3320 static void __setscheduler_params(struct task_struct
*p
,
3321 const struct sched_attr
*attr
)
3323 int policy
= attr
->sched_policy
;
3325 if (policy
== SETPARAM_POLICY
)
3330 if (dl_policy(policy
))
3331 __setparam_dl(p
, attr
);
3332 else if (fair_policy(policy
))
3333 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3336 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3337 * !rt_policy. Always setting this ensures that things like
3338 * getparam()/getattr() don't report silly values for !rt tasks.
3340 p
->rt_priority
= attr
->sched_priority
;
3341 p
->normal_prio
= normal_prio(p
);
3345 /* Actually do priority change: must hold pi & rq lock. */
3346 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3347 const struct sched_attr
*attr
, bool keep_boost
)
3349 __setscheduler_params(p
, attr
);
3352 * Keep a potential priority boosting if called from
3353 * sched_setscheduler().
3356 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3358 p
->prio
= normal_prio(p
);
3360 if (dl_prio(p
->prio
))
3361 p
->sched_class
= &dl_sched_class
;
3362 else if (rt_prio(p
->prio
))
3363 p
->sched_class
= &rt_sched_class
;
3365 p
->sched_class
= &fair_sched_class
;
3369 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3371 struct sched_dl_entity
*dl_se
= &p
->dl
;
3373 attr
->sched_priority
= p
->rt_priority
;
3374 attr
->sched_runtime
= dl_se
->dl_runtime
;
3375 attr
->sched_deadline
= dl_se
->dl_deadline
;
3376 attr
->sched_period
= dl_se
->dl_period
;
3377 attr
->sched_flags
= dl_se
->flags
;
3381 * This function validates the new parameters of a -deadline task.
3382 * We ask for the deadline not being zero, and greater or equal
3383 * than the runtime, as well as the period of being zero or
3384 * greater than deadline. Furthermore, we have to be sure that
3385 * user parameters are above the internal resolution of 1us (we
3386 * check sched_runtime only since it is always the smaller one) and
3387 * below 2^63 ns (we have to check both sched_deadline and
3388 * sched_period, as the latter can be zero).
3391 __checkparam_dl(const struct sched_attr
*attr
)
3394 if (attr
->sched_deadline
== 0)
3398 * Since we truncate DL_SCALE bits, make sure we're at least
3401 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3405 * Since we use the MSB for wrap-around and sign issues, make
3406 * sure it's not set (mind that period can be equal to zero).
3408 if (attr
->sched_deadline
& (1ULL << 63) ||
3409 attr
->sched_period
& (1ULL << 63))
3412 /* runtime <= deadline <= period (if period != 0) */
3413 if ((attr
->sched_period
!= 0 &&
3414 attr
->sched_period
< attr
->sched_deadline
) ||
3415 attr
->sched_deadline
< attr
->sched_runtime
)
3422 * check the target process has a UID that matches the current process's
3424 static bool check_same_owner(struct task_struct
*p
)
3426 const struct cred
*cred
= current_cred(), *pcred
;
3430 pcred
= __task_cred(p
);
3431 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3432 uid_eq(cred
->euid
, pcred
->uid
));
3437 static bool dl_param_changed(struct task_struct
*p
,
3438 const struct sched_attr
*attr
)
3440 struct sched_dl_entity
*dl_se
= &p
->dl
;
3442 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3443 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3444 dl_se
->dl_period
!= attr
->sched_period
||
3445 dl_se
->flags
!= attr
->sched_flags
)
3451 static int __sched_setscheduler(struct task_struct
*p
,
3452 const struct sched_attr
*attr
,
3455 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3456 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3457 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3458 int new_effective_prio
, policy
= attr
->sched_policy
;
3459 unsigned long flags
;
3460 const struct sched_class
*prev_class
;
3464 /* may grab non-irq protected spin_locks */
3465 BUG_ON(in_interrupt());
3467 /* double check policy once rq lock held */
3469 reset_on_fork
= p
->sched_reset_on_fork
;
3470 policy
= oldpolicy
= p
->policy
;
3472 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3474 if (policy
!= SCHED_DEADLINE
&&
3475 policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
3476 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
3477 policy
!= SCHED_IDLE
)
3481 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3485 * Valid priorities for SCHED_FIFO and SCHED_RR are
3486 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3487 * SCHED_BATCH and SCHED_IDLE is 0.
3489 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3490 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3492 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3493 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3497 * Allow unprivileged RT tasks to decrease priority:
3499 if (user
&& !capable(CAP_SYS_NICE
)) {
3500 if (fair_policy(policy
)) {
3501 if (attr
->sched_nice
< task_nice(p
) &&
3502 !can_nice(p
, attr
->sched_nice
))
3506 if (rt_policy(policy
)) {
3507 unsigned long rlim_rtprio
=
3508 task_rlimit(p
, RLIMIT_RTPRIO
);
3510 /* can't set/change the rt policy */
3511 if (policy
!= p
->policy
&& !rlim_rtprio
)
3514 /* can't increase priority */
3515 if (attr
->sched_priority
> p
->rt_priority
&&
3516 attr
->sched_priority
> rlim_rtprio
)
3521 * Can't set/change SCHED_DEADLINE policy at all for now
3522 * (safest behavior); in the future we would like to allow
3523 * unprivileged DL tasks to increase their relative deadline
3524 * or reduce their runtime (both ways reducing utilization)
3526 if (dl_policy(policy
))
3530 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3531 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3533 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
3534 if (!can_nice(p
, task_nice(p
)))
3538 /* can't change other user's priorities */
3539 if (!check_same_owner(p
))
3542 /* Normal users shall not reset the sched_reset_on_fork flag */
3543 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
3548 retval
= security_task_setscheduler(p
);
3554 * make sure no PI-waiters arrive (or leave) while we are
3555 * changing the priority of the task:
3557 * To be able to change p->policy safely, the appropriate
3558 * runqueue lock must be held.
3560 rq
= task_rq_lock(p
, &flags
);
3563 * Changing the policy of the stop threads its a very bad idea
3565 if (p
== rq
->stop
) {
3566 task_rq_unlock(rq
, p
, &flags
);
3571 * If not changing anything there's no need to proceed further,
3572 * but store a possible modification of reset_on_fork.
3574 if (unlikely(policy
== p
->policy
)) {
3575 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
3577 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
3579 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
3582 p
->sched_reset_on_fork
= reset_on_fork
;
3583 task_rq_unlock(rq
, p
, &flags
);
3589 #ifdef CONFIG_RT_GROUP_SCHED
3591 * Do not allow realtime tasks into groups that have no runtime
3594 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
3595 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
3596 !task_group_is_autogroup(task_group(p
))) {
3597 task_rq_unlock(rq
, p
, &flags
);
3602 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
3603 cpumask_t
*span
= rq
->rd
->span
;
3606 * Don't allow tasks with an affinity mask smaller than
3607 * the entire root_domain to become SCHED_DEADLINE. We
3608 * will also fail if there's no bandwidth available.
3610 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
3611 rq
->rd
->dl_bw
.bw
== 0) {
3612 task_rq_unlock(rq
, p
, &flags
);
3619 /* recheck policy now with rq lock held */
3620 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
3621 policy
= oldpolicy
= -1;
3622 task_rq_unlock(rq
, p
, &flags
);
3627 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3628 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3631 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
3632 task_rq_unlock(rq
, p
, &flags
);
3636 p
->sched_reset_on_fork
= reset_on_fork
;
3641 * Take priority boosted tasks into account. If the new
3642 * effective priority is unchanged, we just store the new
3643 * normal parameters and do not touch the scheduler class and
3644 * the runqueue. This will be done when the task deboost
3647 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
3648 if (new_effective_prio
== oldprio
) {
3649 __setscheduler_params(p
, attr
);
3650 task_rq_unlock(rq
, p
, &flags
);
3655 queued
= task_on_rq_queued(p
);
3656 running
= task_current(rq
, p
);
3658 dequeue_task(rq
, p
, 0);
3660 put_prev_task(rq
, p
);
3662 prev_class
= p
->sched_class
;
3663 __setscheduler(rq
, p
, attr
, pi
);
3666 p
->sched_class
->set_curr_task(rq
);
3669 * We enqueue to tail when the priority of a task is
3670 * increased (user space view).
3672 enqueue_task(rq
, p
, oldprio
<= p
->prio
? ENQUEUE_HEAD
: 0);
3675 check_class_changed(rq
, p
, prev_class
, oldprio
);
3676 preempt_disable(); /* avoid rq from going away on us */
3677 task_rq_unlock(rq
, p
, &flags
);
3680 rt_mutex_adjust_pi(p
);
3683 * Run balance callbacks after we've adjusted the PI chain.
3685 balance_callback(rq
);
3691 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
3692 const struct sched_param
*param
, bool check
)
3694 struct sched_attr attr
= {
3695 .sched_policy
= policy
,
3696 .sched_priority
= param
->sched_priority
,
3697 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
3700 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3701 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
3702 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
3703 policy
&= ~SCHED_RESET_ON_FORK
;
3704 attr
.sched_policy
= policy
;
3707 return __sched_setscheduler(p
, &attr
, check
, true);
3710 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3711 * @p: the task in question.
3712 * @policy: new policy.
3713 * @param: structure containing the new RT priority.
3715 * Return: 0 on success. An error code otherwise.
3717 * NOTE that the task may be already dead.
3719 int sched_setscheduler(struct task_struct
*p
, int policy
,
3720 const struct sched_param
*param
)
3722 return _sched_setscheduler(p
, policy
, param
, true);
3724 EXPORT_SYMBOL_GPL(sched_setscheduler
);
3726 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
3728 return __sched_setscheduler(p
, attr
, true, true);
3730 EXPORT_SYMBOL_GPL(sched_setattr
);
3733 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3734 * @p: the task in question.
3735 * @policy: new policy.
3736 * @param: structure containing the new RT priority.
3738 * Just like sched_setscheduler, only don't bother checking if the
3739 * current context has permission. For example, this is needed in
3740 * stop_machine(): we create temporary high priority worker threads,
3741 * but our caller might not have that capability.
3743 * Return: 0 on success. An error code otherwise.
3745 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
3746 const struct sched_param
*param
)
3748 return _sched_setscheduler(p
, policy
, param
, false);
3752 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
3754 struct sched_param lparam
;
3755 struct task_struct
*p
;
3758 if (!param
|| pid
< 0)
3760 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
3765 p
= find_process_by_pid(pid
);
3767 retval
= sched_setscheduler(p
, policy
, &lparam
);
3774 * Mimics kernel/events/core.c perf_copy_attr().
3776 static int sched_copy_attr(struct sched_attr __user
*uattr
,
3777 struct sched_attr
*attr
)
3782 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
3786 * zero the full structure, so that a short copy will be nice.
3788 memset(attr
, 0, sizeof(*attr
));
3790 ret
= get_user(size
, &uattr
->size
);
3794 if (size
> PAGE_SIZE
) /* silly large */
3797 if (!size
) /* abi compat */
3798 size
= SCHED_ATTR_SIZE_VER0
;
3800 if (size
< SCHED_ATTR_SIZE_VER0
)
3804 * If we're handed a bigger struct than we know of,
3805 * ensure all the unknown bits are 0 - i.e. new
3806 * user-space does not rely on any kernel feature
3807 * extensions we dont know about yet.
3809 if (size
> sizeof(*attr
)) {
3810 unsigned char __user
*addr
;
3811 unsigned char __user
*end
;
3814 addr
= (void __user
*)uattr
+ sizeof(*attr
);
3815 end
= (void __user
*)uattr
+ size
;
3817 for (; addr
< end
; addr
++) {
3818 ret
= get_user(val
, addr
);
3824 size
= sizeof(*attr
);
3827 ret
= copy_from_user(attr
, uattr
, size
);
3832 * XXX: do we want to be lenient like existing syscalls; or do we want
3833 * to be strict and return an error on out-of-bounds values?
3835 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
3840 put_user(sizeof(*attr
), &uattr
->size
);
3845 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3846 * @pid: the pid in question.
3847 * @policy: new policy.
3848 * @param: structure containing the new RT priority.
3850 * Return: 0 on success. An error code otherwise.
3852 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
3853 struct sched_param __user
*, param
)
3855 /* negative values for policy are not valid */
3859 return do_sched_setscheduler(pid
, policy
, param
);
3863 * sys_sched_setparam - set/change the RT priority of a thread
3864 * @pid: the pid in question.
3865 * @param: structure containing the new RT priority.
3867 * Return: 0 on success. An error code otherwise.
3869 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3871 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
3875 * sys_sched_setattr - same as above, but with extended sched_attr
3876 * @pid: the pid in question.
3877 * @uattr: structure containing the extended parameters.
3878 * @flags: for future extension.
3880 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
3881 unsigned int, flags
)
3883 struct sched_attr attr
;
3884 struct task_struct
*p
;
3887 if (!uattr
|| pid
< 0 || flags
)
3890 retval
= sched_copy_attr(uattr
, &attr
);
3894 if ((int)attr
.sched_policy
< 0)
3899 p
= find_process_by_pid(pid
);
3901 retval
= sched_setattr(p
, &attr
);
3908 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3909 * @pid: the pid in question.
3911 * Return: On success, the policy of the thread. Otherwise, a negative error
3914 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
3916 struct task_struct
*p
;
3924 p
= find_process_by_pid(pid
);
3926 retval
= security_task_getscheduler(p
);
3929 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
3936 * sys_sched_getparam - get the RT priority of a thread
3937 * @pid: the pid in question.
3938 * @param: structure containing the RT priority.
3940 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3943 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
3945 struct sched_param lp
= { .sched_priority
= 0 };
3946 struct task_struct
*p
;
3949 if (!param
|| pid
< 0)
3953 p
= find_process_by_pid(pid
);
3958 retval
= security_task_getscheduler(p
);
3962 if (task_has_rt_policy(p
))
3963 lp
.sched_priority
= p
->rt_priority
;
3967 * This one might sleep, we cannot do it with a spinlock held ...
3969 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
3978 static int sched_read_attr(struct sched_attr __user
*uattr
,
3979 struct sched_attr
*attr
,
3984 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
3988 * If we're handed a smaller struct than we know of,
3989 * ensure all the unknown bits are 0 - i.e. old
3990 * user-space does not get uncomplete information.
3992 if (usize
< sizeof(*attr
)) {
3993 unsigned char *addr
;
3996 addr
= (void *)attr
+ usize
;
3997 end
= (void *)attr
+ sizeof(*attr
);
3999 for (; addr
< end
; addr
++) {
4007 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4015 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4016 * @pid: the pid in question.
4017 * @uattr: structure containing the extended parameters.
4018 * @size: sizeof(attr) for fwd/bwd comp.
4019 * @flags: for future extension.
4021 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4022 unsigned int, size
, unsigned int, flags
)
4024 struct sched_attr attr
= {
4025 .size
= sizeof(struct sched_attr
),
4027 struct task_struct
*p
;
4030 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4031 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4035 p
= find_process_by_pid(pid
);
4040 retval
= security_task_getscheduler(p
);
4044 attr
.sched_policy
= p
->policy
;
4045 if (p
->sched_reset_on_fork
)
4046 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4047 if (task_has_dl_policy(p
))
4048 __getparam_dl(p
, &attr
);
4049 else if (task_has_rt_policy(p
))
4050 attr
.sched_priority
= p
->rt_priority
;
4052 attr
.sched_nice
= task_nice(p
);
4056 retval
= sched_read_attr(uattr
, &attr
, size
);
4064 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4066 cpumask_var_t cpus_allowed
, new_mask
;
4067 struct task_struct
*p
;
4072 p
= find_process_by_pid(pid
);
4078 /* Prevent p going away */
4082 if (p
->flags
& PF_NO_SETAFFINITY
) {
4086 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4090 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4092 goto out_free_cpus_allowed
;
4095 if (!check_same_owner(p
)) {
4097 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4099 goto out_free_new_mask
;
4104 retval
= security_task_setscheduler(p
);
4106 goto out_free_new_mask
;
4109 cpuset_cpus_allowed(p
, cpus_allowed
);
4110 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4113 * Since bandwidth control happens on root_domain basis,
4114 * if admission test is enabled, we only admit -deadline
4115 * tasks allowed to run on all the CPUs in the task's
4119 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4121 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4124 goto out_free_new_mask
;
4130 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4133 cpuset_cpus_allowed(p
, cpus_allowed
);
4134 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4136 * We must have raced with a concurrent cpuset
4137 * update. Just reset the cpus_allowed to the
4138 * cpuset's cpus_allowed
4140 cpumask_copy(new_mask
, cpus_allowed
);
4145 free_cpumask_var(new_mask
);
4146 out_free_cpus_allowed
:
4147 free_cpumask_var(cpus_allowed
);
4153 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4154 struct cpumask
*new_mask
)
4156 if (len
< cpumask_size())
4157 cpumask_clear(new_mask
);
4158 else if (len
> cpumask_size())
4159 len
= cpumask_size();
4161 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4165 * sys_sched_setaffinity - set the cpu affinity of a process
4166 * @pid: pid of the process
4167 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4168 * @user_mask_ptr: user-space pointer to the new cpu mask
4170 * Return: 0 on success. An error code otherwise.
4172 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4173 unsigned long __user
*, user_mask_ptr
)
4175 cpumask_var_t new_mask
;
4178 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4181 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4183 retval
= sched_setaffinity(pid
, new_mask
);
4184 free_cpumask_var(new_mask
);
4188 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4190 struct task_struct
*p
;
4191 unsigned long flags
;
4197 p
= find_process_by_pid(pid
);
4201 retval
= security_task_getscheduler(p
);
4205 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4206 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4207 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4216 * sys_sched_getaffinity - get the cpu affinity of a process
4217 * @pid: pid of the process
4218 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4219 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4221 * Return: 0 on success. An error code otherwise.
4223 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4224 unsigned long __user
*, user_mask_ptr
)
4229 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4231 if (len
& (sizeof(unsigned long)-1))
4234 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4237 ret
= sched_getaffinity(pid
, mask
);
4239 size_t retlen
= min_t(size_t, len
, cpumask_size());
4241 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4246 free_cpumask_var(mask
);
4252 * sys_sched_yield - yield the current processor to other threads.
4254 * This function yields the current CPU to other tasks. If there are no
4255 * other threads running on this CPU then this function will return.
4259 SYSCALL_DEFINE0(sched_yield
)
4261 struct rq
*rq
= this_rq_lock();
4263 schedstat_inc(rq
, yld_count
);
4264 current
->sched_class
->yield_task(rq
);
4267 * Since we are going to call schedule() anyway, there's
4268 * no need to preempt or enable interrupts:
4270 __release(rq
->lock
);
4271 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4272 do_raw_spin_unlock(&rq
->lock
);
4273 sched_preempt_enable_no_resched();
4280 int __sched
_cond_resched(void)
4282 if (should_resched()) {
4283 preempt_schedule_common();
4288 EXPORT_SYMBOL(_cond_resched
);
4291 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4292 * call schedule, and on return reacquire the lock.
4294 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4295 * operations here to prevent schedule() from being called twice (once via
4296 * spin_unlock(), once by hand).
4298 int __cond_resched_lock(spinlock_t
*lock
)
4300 int resched
= should_resched();
4303 lockdep_assert_held(lock
);
4305 if (spin_needbreak(lock
) || resched
) {
4308 preempt_schedule_common();
4316 EXPORT_SYMBOL(__cond_resched_lock
);
4318 int __sched
__cond_resched_softirq(void)
4320 BUG_ON(!in_softirq());
4322 if (should_resched()) {
4324 preempt_schedule_common();
4330 EXPORT_SYMBOL(__cond_resched_softirq
);
4333 * yield - yield the current processor to other threads.
4335 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4337 * The scheduler is at all times free to pick the calling task as the most
4338 * eligible task to run, if removing the yield() call from your code breaks
4339 * it, its already broken.
4341 * Typical broken usage is:
4346 * where one assumes that yield() will let 'the other' process run that will
4347 * make event true. If the current task is a SCHED_FIFO task that will never
4348 * happen. Never use yield() as a progress guarantee!!
4350 * If you want to use yield() to wait for something, use wait_event().
4351 * If you want to use yield() to be 'nice' for others, use cond_resched().
4352 * If you still want to use yield(), do not!
4354 void __sched
yield(void)
4356 set_current_state(TASK_RUNNING
);
4359 EXPORT_SYMBOL(yield
);
4362 * yield_to - yield the current processor to another thread in
4363 * your thread group, or accelerate that thread toward the
4364 * processor it's on.
4366 * @preempt: whether task preemption is allowed or not
4368 * It's the caller's job to ensure that the target task struct
4369 * can't go away on us before we can do any checks.
4372 * true (>0) if we indeed boosted the target task.
4373 * false (0) if we failed to boost the target.
4374 * -ESRCH if there's no task to yield to.
4376 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4378 struct task_struct
*curr
= current
;
4379 struct rq
*rq
, *p_rq
;
4380 unsigned long flags
;
4383 local_irq_save(flags
);
4389 * If we're the only runnable task on the rq and target rq also
4390 * has only one task, there's absolutely no point in yielding.
4392 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4397 double_rq_lock(rq
, p_rq
);
4398 if (task_rq(p
) != p_rq
) {
4399 double_rq_unlock(rq
, p_rq
);
4403 if (!curr
->sched_class
->yield_to_task
)
4406 if (curr
->sched_class
!= p
->sched_class
)
4409 if (task_running(p_rq
, p
) || p
->state
)
4412 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4414 schedstat_inc(rq
, yld_count
);
4416 * Make p's CPU reschedule; pick_next_entity takes care of
4419 if (preempt
&& rq
!= p_rq
)
4424 double_rq_unlock(rq
, p_rq
);
4426 local_irq_restore(flags
);
4433 EXPORT_SYMBOL_GPL(yield_to
);
4436 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4437 * that process accounting knows that this is a task in IO wait state.
4439 long __sched
io_schedule_timeout(long timeout
)
4441 int old_iowait
= current
->in_iowait
;
4445 current
->in_iowait
= 1;
4446 blk_schedule_flush_plug(current
);
4448 delayacct_blkio_start();
4450 atomic_inc(&rq
->nr_iowait
);
4451 ret
= schedule_timeout(timeout
);
4452 current
->in_iowait
= old_iowait
;
4453 atomic_dec(&rq
->nr_iowait
);
4454 delayacct_blkio_end();
4458 EXPORT_SYMBOL(io_schedule_timeout
);
4461 * sys_sched_get_priority_max - return maximum RT priority.
4462 * @policy: scheduling class.
4464 * Return: On success, this syscall returns the maximum
4465 * rt_priority that can be used by a given scheduling class.
4466 * On failure, a negative error code is returned.
4468 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4475 ret
= MAX_USER_RT_PRIO
-1;
4477 case SCHED_DEADLINE
:
4488 * sys_sched_get_priority_min - return minimum RT priority.
4489 * @policy: scheduling class.
4491 * Return: On success, this syscall returns the minimum
4492 * rt_priority that can be used by a given scheduling class.
4493 * On failure, a negative error code is returned.
4495 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4504 case SCHED_DEADLINE
:
4514 * sys_sched_rr_get_interval - return the default timeslice of a process.
4515 * @pid: pid of the process.
4516 * @interval: userspace pointer to the timeslice value.
4518 * this syscall writes the default timeslice value of a given process
4519 * into the user-space timespec buffer. A value of '0' means infinity.
4521 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4524 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
4525 struct timespec __user
*, interval
)
4527 struct task_struct
*p
;
4528 unsigned int time_slice
;
4529 unsigned long flags
;
4539 p
= find_process_by_pid(pid
);
4543 retval
= security_task_getscheduler(p
);
4547 rq
= task_rq_lock(p
, &flags
);
4549 if (p
->sched_class
->get_rr_interval
)
4550 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
4551 task_rq_unlock(rq
, p
, &flags
);
4554 jiffies_to_timespec(time_slice
, &t
);
4555 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4563 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
4565 void sched_show_task(struct task_struct
*p
)
4567 unsigned long free
= 0;
4569 unsigned long state
= p
->state
;
4572 state
= __ffs(state
) + 1;
4573 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
4574 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4575 #if BITS_PER_LONG == 32
4576 if (state
== TASK_RUNNING
)
4577 printk(KERN_CONT
" running ");
4579 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
4581 if (state
== TASK_RUNNING
)
4582 printk(KERN_CONT
" running task ");
4584 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
4586 #ifdef CONFIG_DEBUG_STACK_USAGE
4587 free
= stack_not_used(p
);
4592 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
4594 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
4595 task_pid_nr(p
), ppid
,
4596 (unsigned long)task_thread_info(p
)->flags
);
4598 print_worker_info(KERN_INFO
, p
);
4599 show_stack(p
, NULL
);
4602 void show_state_filter(unsigned long state_filter
)
4604 struct task_struct
*g
, *p
;
4606 #if BITS_PER_LONG == 32
4608 " task PC stack pid father\n");
4611 " task PC stack pid father\n");
4614 for_each_process_thread(g
, p
) {
4616 * reset the NMI-timeout, listing all files on a slow
4617 * console might take a lot of time:
4619 touch_nmi_watchdog();
4620 if (!state_filter
|| (p
->state
& state_filter
))
4624 touch_all_softlockup_watchdogs();
4626 #ifdef CONFIG_SCHED_DEBUG
4627 sysrq_sched_debug_show();
4631 * Only show locks if all tasks are dumped:
4634 debug_show_all_locks();
4637 void init_idle_bootup_task(struct task_struct
*idle
)
4639 idle
->sched_class
= &idle_sched_class
;
4643 * init_idle - set up an idle thread for a given CPU
4644 * @idle: task in question
4645 * @cpu: cpu the idle task belongs to
4647 * NOTE: this function does not set the idle thread's NEED_RESCHED
4648 * flag, to make booting more robust.
4650 void init_idle(struct task_struct
*idle
, int cpu
)
4652 struct rq
*rq
= cpu_rq(cpu
);
4653 unsigned long flags
;
4655 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4657 __sched_fork(0, idle
);
4658 idle
->state
= TASK_RUNNING
;
4659 idle
->se
.exec_start
= sched_clock();
4661 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
4663 * We're having a chicken and egg problem, even though we are
4664 * holding rq->lock, the cpu isn't yet set to this cpu so the
4665 * lockdep check in task_group() will fail.
4667 * Similar case to sched_fork(). / Alternatively we could
4668 * use task_rq_lock() here and obtain the other rq->lock.
4673 __set_task_cpu(idle
, cpu
);
4676 rq
->curr
= rq
->idle
= idle
;
4677 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
4678 #if defined(CONFIG_SMP)
4681 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4683 /* Set the preempt count _outside_ the spinlocks! */
4684 init_idle_preempt_count(idle
, cpu
);
4687 * The idle tasks have their own, simple scheduling class:
4689 idle
->sched_class
= &idle_sched_class
;
4690 ftrace_graph_init_idle_task(idle
, cpu
);
4691 vtime_init_idle(idle
, cpu
);
4692 #if defined(CONFIG_SMP)
4693 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
4697 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
4698 const struct cpumask
*trial
)
4700 int ret
= 1, trial_cpus
;
4701 struct dl_bw
*cur_dl_b
;
4702 unsigned long flags
;
4704 if (!cpumask_weight(cur
))
4707 rcu_read_lock_sched();
4708 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
4709 trial_cpus
= cpumask_weight(trial
);
4711 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
4712 if (cur_dl_b
->bw
!= -1 &&
4713 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
4715 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
4716 rcu_read_unlock_sched();
4721 int task_can_attach(struct task_struct
*p
,
4722 const struct cpumask
*cs_cpus_allowed
)
4727 * Kthreads which disallow setaffinity shouldn't be moved
4728 * to a new cpuset; we don't want to change their cpu
4729 * affinity and isolating such threads by their set of
4730 * allowed nodes is unnecessary. Thus, cpusets are not
4731 * applicable for such threads. This prevents checking for
4732 * success of set_cpus_allowed_ptr() on all attached tasks
4733 * before cpus_allowed may be changed.
4735 if (p
->flags
& PF_NO_SETAFFINITY
) {
4741 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
4743 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
4748 unsigned long flags
;
4750 rcu_read_lock_sched();
4751 dl_b
= dl_bw_of(dest_cpu
);
4752 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
4753 cpus
= dl_bw_cpus(dest_cpu
);
4754 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
4759 * We reserve space for this task in the destination
4760 * root_domain, as we can't fail after this point.
4761 * We will free resources in the source root_domain
4762 * later on (see set_cpus_allowed_dl()).
4764 __dl_add(dl_b
, p
->dl
.dl_bw
);
4766 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
4767 rcu_read_unlock_sched();
4777 * move_queued_task - move a queued task to new rq.
4779 * Returns (locked) new rq. Old rq's lock is released.
4781 static struct rq
*move_queued_task(struct task_struct
*p
, int new_cpu
)
4783 struct rq
*rq
= task_rq(p
);
4785 lockdep_assert_held(&rq
->lock
);
4787 dequeue_task(rq
, p
, 0);
4788 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
4789 set_task_cpu(p
, new_cpu
);
4790 raw_spin_unlock(&rq
->lock
);
4792 rq
= cpu_rq(new_cpu
);
4794 raw_spin_lock(&rq
->lock
);
4795 BUG_ON(task_cpu(p
) != new_cpu
);
4796 p
->on_rq
= TASK_ON_RQ_QUEUED
;
4797 enqueue_task(rq
, p
, 0);
4798 check_preempt_curr(rq
, p
, 0);
4803 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
4805 if (p
->sched_class
->set_cpus_allowed
)
4806 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
4808 cpumask_copy(&p
->cpus_allowed
, new_mask
);
4809 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
4813 * This is how migration works:
4815 * 1) we invoke migration_cpu_stop() on the target CPU using
4817 * 2) stopper starts to run (implicitly forcing the migrated thread
4819 * 3) it checks whether the migrated task is still in the wrong runqueue.
4820 * 4) if it's in the wrong runqueue then the migration thread removes
4821 * it and puts it into the right queue.
4822 * 5) stopper completes and stop_one_cpu() returns and the migration
4827 * Change a given task's CPU affinity. Migrate the thread to a
4828 * proper CPU and schedule it away if the CPU it's executing on
4829 * is removed from the allowed bitmask.
4831 * NOTE: the caller must have a valid reference to the task, the
4832 * task must not exit() & deallocate itself prematurely. The
4833 * call is not atomic; no spinlocks may be held.
4835 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
4837 unsigned long flags
;
4839 unsigned int dest_cpu
;
4842 rq
= task_rq_lock(p
, &flags
);
4844 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
4847 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
4852 do_set_cpus_allowed(p
, new_mask
);
4854 /* Can the task run on the task's current CPU? If so, we're done */
4855 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
4858 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
4859 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
4860 struct migration_arg arg
= { p
, dest_cpu
};
4861 /* Need help from migration thread: drop lock and wait. */
4862 task_rq_unlock(rq
, p
, &flags
);
4863 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
4864 tlb_migrate_finish(p
->mm
);
4866 } else if (task_on_rq_queued(p
))
4867 rq
= move_queued_task(p
, dest_cpu
);
4869 task_rq_unlock(rq
, p
, &flags
);
4873 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
4876 * Move (not current) task off this cpu, onto dest cpu. We're doing
4877 * this because either it can't run here any more (set_cpus_allowed()
4878 * away from this CPU, or CPU going down), or because we're
4879 * attempting to rebalance this task on exec (sched_exec).
4881 * So we race with normal scheduler movements, but that's OK, as long
4882 * as the task is no longer on this CPU.
4884 * Returns non-zero if task was successfully migrated.
4886 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
4891 if (unlikely(!cpu_active(dest_cpu
)))
4894 rq
= cpu_rq(src_cpu
);
4896 raw_spin_lock(&p
->pi_lock
);
4897 raw_spin_lock(&rq
->lock
);
4898 /* Already moved. */
4899 if (task_cpu(p
) != src_cpu
)
4902 /* Affinity changed (again). */
4903 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
4907 * If we're not on a rq, the next wake-up will ensure we're
4910 if (task_on_rq_queued(p
))
4911 rq
= move_queued_task(p
, dest_cpu
);
4915 raw_spin_unlock(&rq
->lock
);
4916 raw_spin_unlock(&p
->pi_lock
);
4920 #ifdef CONFIG_NUMA_BALANCING
4921 /* Migrate current task p to target_cpu */
4922 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
4924 struct migration_arg arg
= { p
, target_cpu
};
4925 int curr_cpu
= task_cpu(p
);
4927 if (curr_cpu
== target_cpu
)
4930 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
4933 /* TODO: This is not properly updating schedstats */
4935 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
4936 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
4940 * Requeue a task on a given node and accurately track the number of NUMA
4941 * tasks on the runqueues
4943 void sched_setnuma(struct task_struct
*p
, int nid
)
4946 unsigned long flags
;
4947 bool queued
, running
;
4949 rq
= task_rq_lock(p
, &flags
);
4950 queued
= task_on_rq_queued(p
);
4951 running
= task_current(rq
, p
);
4954 dequeue_task(rq
, p
, 0);
4956 put_prev_task(rq
, p
);
4958 p
->numa_preferred_nid
= nid
;
4961 p
->sched_class
->set_curr_task(rq
);
4963 enqueue_task(rq
, p
, 0);
4964 task_rq_unlock(rq
, p
, &flags
);
4969 * migration_cpu_stop - this will be executed by a highprio stopper thread
4970 * and performs thread migration by bumping thread off CPU then
4971 * 'pushing' onto another runqueue.
4973 static int migration_cpu_stop(void *data
)
4975 struct migration_arg
*arg
= data
;
4978 * The original target cpu might have gone down and we might
4979 * be on another cpu but it doesn't matter.
4981 local_irq_disable();
4983 * We need to explicitly wake pending tasks before running
4984 * __migrate_task() such that we will not miss enforcing cpus_allowed
4985 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4987 sched_ttwu_pending();
4988 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
4993 #ifdef CONFIG_HOTPLUG_CPU
4996 * Ensures that the idle task is using init_mm right before its cpu goes
4999 void idle_task_exit(void)
5001 struct mm_struct
*mm
= current
->active_mm
;
5003 BUG_ON(cpu_online(smp_processor_id()));
5005 if (mm
!= &init_mm
) {
5006 switch_mm(mm
, &init_mm
, current
);
5007 finish_arch_post_lock_switch();
5013 * Since this CPU is going 'away' for a while, fold any nr_active delta
5014 * we might have. Assumes we're called after migrate_tasks() so that the
5015 * nr_active count is stable.
5017 * Also see the comment "Global load-average calculations".
5019 static void calc_load_migrate(struct rq
*rq
)
5021 long delta
= calc_load_fold_active(rq
);
5023 atomic_long_add(delta
, &calc_load_tasks
);
5026 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5030 static const struct sched_class fake_sched_class
= {
5031 .put_prev_task
= put_prev_task_fake
,
5034 static struct task_struct fake_task
= {
5036 * Avoid pull_{rt,dl}_task()
5038 .prio
= MAX_PRIO
+ 1,
5039 .sched_class
= &fake_sched_class
,
5043 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5044 * try_to_wake_up()->select_task_rq().
5046 * Called with rq->lock held even though we'er in stop_machine() and
5047 * there's no concurrency possible, we hold the required locks anyway
5048 * because of lock validation efforts.
5050 static void migrate_tasks(unsigned int dead_cpu
)
5052 struct rq
*rq
= cpu_rq(dead_cpu
);
5053 struct task_struct
*next
, *stop
= rq
->stop
;
5057 * Fudge the rq selection such that the below task selection loop
5058 * doesn't get stuck on the currently eligible stop task.
5060 * We're currently inside stop_machine() and the rq is either stuck
5061 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5062 * either way we should never end up calling schedule() until we're
5068 * put_prev_task() and pick_next_task() sched
5069 * class method both need to have an up-to-date
5070 * value of rq->clock[_task]
5072 update_rq_clock(rq
);
5076 * There's this thread running, bail when that's the only
5079 if (rq
->nr_running
== 1)
5082 next
= pick_next_task(rq
, &fake_task
);
5084 next
->sched_class
->put_prev_task(rq
, next
);
5086 /* Find suitable destination for @next, with force if needed. */
5087 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5088 raw_spin_unlock(&rq
->lock
);
5090 __migrate_task(next
, dead_cpu
, dest_cpu
);
5092 raw_spin_lock(&rq
->lock
);
5098 #endif /* CONFIG_HOTPLUG_CPU */
5100 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5102 static struct ctl_table sd_ctl_dir
[] = {
5104 .procname
= "sched_domain",
5110 static struct ctl_table sd_ctl_root
[] = {
5112 .procname
= "kernel",
5114 .child
= sd_ctl_dir
,
5119 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5121 struct ctl_table
*entry
=
5122 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5127 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5129 struct ctl_table
*entry
;
5132 * In the intermediate directories, both the child directory and
5133 * procname are dynamically allocated and could fail but the mode
5134 * will always be set. In the lowest directory the names are
5135 * static strings and all have proc handlers.
5137 for (entry
= *tablep
; entry
->mode
; entry
++) {
5139 sd_free_ctl_entry(&entry
->child
);
5140 if (entry
->proc_handler
== NULL
)
5141 kfree(entry
->procname
);
5148 static int min_load_idx
= 0;
5149 static int max_load_idx
= CPU_LOAD_IDX_MAX
-1;
5152 set_table_entry(struct ctl_table
*entry
,
5153 const char *procname
, void *data
, int maxlen
,
5154 umode_t mode
, proc_handler
*proc_handler
,
5157 entry
->procname
= procname
;
5159 entry
->maxlen
= maxlen
;
5161 entry
->proc_handler
= proc_handler
;
5164 entry
->extra1
= &min_load_idx
;
5165 entry
->extra2
= &max_load_idx
;
5169 static struct ctl_table
*
5170 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5172 struct ctl_table
*table
= sd_alloc_ctl_entry(14);
5177 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5178 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5179 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5180 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5181 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5182 sizeof(int), 0644, proc_dointvec_minmax
, true);
5183 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5184 sizeof(int), 0644, proc_dointvec_minmax
, true);
5185 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5186 sizeof(int), 0644, proc_dointvec_minmax
, true);
5187 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5188 sizeof(int), 0644, proc_dointvec_minmax
, true);
5189 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5190 sizeof(int), 0644, proc_dointvec_minmax
, true);
5191 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5192 sizeof(int), 0644, proc_dointvec_minmax
, false);
5193 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5194 sizeof(int), 0644, proc_dointvec_minmax
, false);
5195 set_table_entry(&table
[9], "cache_nice_tries",
5196 &sd
->cache_nice_tries
,
5197 sizeof(int), 0644, proc_dointvec_minmax
, false);
5198 set_table_entry(&table
[10], "flags", &sd
->flags
,
5199 sizeof(int), 0644, proc_dointvec_minmax
, false);
5200 set_table_entry(&table
[11], "max_newidle_lb_cost",
5201 &sd
->max_newidle_lb_cost
,
5202 sizeof(long), 0644, proc_doulongvec_minmax
, false);
5203 set_table_entry(&table
[12], "name", sd
->name
,
5204 CORENAME_MAX_SIZE
, 0444, proc_dostring
, false);
5205 /* &table[13] is terminator */
5210 static struct ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5212 struct ctl_table
*entry
, *table
;
5213 struct sched_domain
*sd
;
5214 int domain_num
= 0, i
;
5217 for_each_domain(cpu
, sd
)
5219 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5224 for_each_domain(cpu
, sd
) {
5225 snprintf(buf
, 32, "domain%d", i
);
5226 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5228 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5235 static struct ctl_table_header
*sd_sysctl_header
;
5236 static void register_sched_domain_sysctl(void)
5238 int i
, cpu_num
= num_possible_cpus();
5239 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5242 WARN_ON(sd_ctl_dir
[0].child
);
5243 sd_ctl_dir
[0].child
= entry
;
5248 for_each_possible_cpu(i
) {
5249 snprintf(buf
, 32, "cpu%d", i
);
5250 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5252 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5256 WARN_ON(sd_sysctl_header
);
5257 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5260 /* may be called multiple times per register */
5261 static void unregister_sched_domain_sysctl(void)
5263 if (sd_sysctl_header
)
5264 unregister_sysctl_table(sd_sysctl_header
);
5265 sd_sysctl_header
= NULL
;
5266 if (sd_ctl_dir
[0].child
)
5267 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5270 static void register_sched_domain_sysctl(void)
5273 static void unregister_sched_domain_sysctl(void)
5278 static void set_rq_online(struct rq
*rq
)
5281 const struct sched_class
*class;
5283 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5286 for_each_class(class) {
5287 if (class->rq_online
)
5288 class->rq_online(rq
);
5293 static void set_rq_offline(struct rq
*rq
)
5296 const struct sched_class
*class;
5298 for_each_class(class) {
5299 if (class->rq_offline
)
5300 class->rq_offline(rq
);
5303 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5309 * migration_call - callback that gets triggered when a CPU is added.
5310 * Here we can start up the necessary migration thread for the new CPU.
5313 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5315 int cpu
= (long)hcpu
;
5316 unsigned long flags
;
5317 struct rq
*rq
= cpu_rq(cpu
);
5319 switch (action
& ~CPU_TASKS_FROZEN
) {
5321 case CPU_UP_PREPARE
:
5322 rq
->calc_load_update
= calc_load_update
;
5326 /* Update our root-domain */
5327 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5329 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5333 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5336 #ifdef CONFIG_HOTPLUG_CPU
5338 sched_ttwu_pending();
5339 /* Update our root-domain */
5340 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5342 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5346 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5347 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5351 calc_load_migrate(rq
);
5356 update_max_interval();
5362 * Register at high priority so that task migration (migrate_all_tasks)
5363 * happens before everything else. This has to be lower priority than
5364 * the notifier in the perf_event subsystem, though.
5366 static struct notifier_block migration_notifier
= {
5367 .notifier_call
= migration_call
,
5368 .priority
= CPU_PRI_MIGRATION
,
5371 static void set_cpu_rq_start_time(void)
5373 int cpu
= smp_processor_id();
5374 struct rq
*rq
= cpu_rq(cpu
);
5375 rq
->age_stamp
= sched_clock_cpu(cpu
);
5378 static int sched_cpu_active(struct notifier_block
*nfb
,
5379 unsigned long action
, void *hcpu
)
5381 switch (action
& ~CPU_TASKS_FROZEN
) {
5383 set_cpu_rq_start_time();
5385 case CPU_DOWN_FAILED
:
5386 set_cpu_active((long)hcpu
, true);
5393 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5394 unsigned long action
, void *hcpu
)
5396 switch (action
& ~CPU_TASKS_FROZEN
) {
5397 case CPU_DOWN_PREPARE
:
5398 set_cpu_active((long)hcpu
, false);
5405 static int __init
migration_init(void)
5407 void *cpu
= (void *)(long)smp_processor_id();
5410 /* Initialize migration for the boot CPU */
5411 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5412 BUG_ON(err
== NOTIFY_BAD
);
5413 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5414 register_cpu_notifier(&migration_notifier
);
5416 /* Register cpu active notifiers */
5417 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5418 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5422 early_initcall(migration_init
);
5427 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5429 #ifdef CONFIG_SCHED_DEBUG
5431 static __read_mostly
int sched_debug_enabled
;
5433 static int __init
sched_debug_setup(char *str
)
5435 sched_debug_enabled
= 1;
5439 early_param("sched_debug", sched_debug_setup
);
5441 static inline bool sched_debug(void)
5443 return sched_debug_enabled
;
5446 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5447 struct cpumask
*groupmask
)
5449 struct sched_group
*group
= sd
->groups
;
5451 cpumask_clear(groupmask
);
5453 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5455 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5456 printk("does not load-balance\n");
5458 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5463 printk(KERN_CONT
"span %*pbl level %s\n",
5464 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5466 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5467 printk(KERN_ERR
"ERROR: domain->span does not contain "
5470 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5471 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5475 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5479 printk(KERN_ERR
"ERROR: group is NULL\n");
5483 if (!cpumask_weight(sched_group_cpus(group
))) {
5484 printk(KERN_CONT
"\n");
5485 printk(KERN_ERR
"ERROR: empty group\n");
5489 if (!(sd
->flags
& SD_OVERLAP
) &&
5490 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5491 printk(KERN_CONT
"\n");
5492 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5496 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5498 printk(KERN_CONT
" %*pbl",
5499 cpumask_pr_args(sched_group_cpus(group
)));
5500 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5501 printk(KERN_CONT
" (cpu_capacity = %d)",
5502 group
->sgc
->capacity
);
5505 group
= group
->next
;
5506 } while (group
!= sd
->groups
);
5507 printk(KERN_CONT
"\n");
5509 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5510 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5513 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5514 printk(KERN_ERR
"ERROR: parent span is not a superset "
5515 "of domain->span\n");
5519 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5523 if (!sched_debug_enabled
)
5527 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5531 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5534 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5542 #else /* !CONFIG_SCHED_DEBUG */
5543 # define sched_domain_debug(sd, cpu) do { } while (0)
5544 static inline bool sched_debug(void)
5548 #endif /* CONFIG_SCHED_DEBUG */
5550 static int sd_degenerate(struct sched_domain
*sd
)
5552 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5555 /* Following flags need at least 2 groups */
5556 if (sd
->flags
& (SD_LOAD_BALANCE
|
5557 SD_BALANCE_NEWIDLE
|
5560 SD_SHARE_CPUCAPACITY
|
5561 SD_SHARE_PKG_RESOURCES
|
5562 SD_SHARE_POWERDOMAIN
)) {
5563 if (sd
->groups
!= sd
->groups
->next
)
5567 /* Following flags don't use groups */
5568 if (sd
->flags
& (SD_WAKE_AFFINE
))
5575 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5577 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5579 if (sd_degenerate(parent
))
5582 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5585 /* Flags needing groups don't count if only 1 group in parent */
5586 if (parent
->groups
== parent
->groups
->next
) {
5587 pflags
&= ~(SD_LOAD_BALANCE
|
5588 SD_BALANCE_NEWIDLE
|
5591 SD_SHARE_CPUCAPACITY
|
5592 SD_SHARE_PKG_RESOURCES
|
5594 SD_SHARE_POWERDOMAIN
);
5595 if (nr_node_ids
== 1)
5596 pflags
&= ~SD_SERIALIZE
;
5598 if (~cflags
& pflags
)
5604 static void free_rootdomain(struct rcu_head
*rcu
)
5606 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5608 cpupri_cleanup(&rd
->cpupri
);
5609 cpudl_cleanup(&rd
->cpudl
);
5610 free_cpumask_var(rd
->dlo_mask
);
5611 free_cpumask_var(rd
->rto_mask
);
5612 free_cpumask_var(rd
->online
);
5613 free_cpumask_var(rd
->span
);
5617 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5619 struct root_domain
*old_rd
= NULL
;
5620 unsigned long flags
;
5622 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5627 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5630 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5633 * If we dont want to free the old_rd yet then
5634 * set old_rd to NULL to skip the freeing later
5637 if (!atomic_dec_and_test(&old_rd
->refcount
))
5641 atomic_inc(&rd
->refcount
);
5644 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5645 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5648 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5651 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5654 static int init_rootdomain(struct root_domain
*rd
)
5656 memset(rd
, 0, sizeof(*rd
));
5658 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5660 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5662 if (!alloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5664 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5667 init_dl_bw(&rd
->dl_bw
);
5668 if (cpudl_init(&rd
->cpudl
) != 0)
5671 if (cpupri_init(&rd
->cpupri
) != 0)
5676 free_cpumask_var(rd
->rto_mask
);
5678 free_cpumask_var(rd
->dlo_mask
);
5680 free_cpumask_var(rd
->online
);
5682 free_cpumask_var(rd
->span
);
5688 * By default the system creates a single root-domain with all cpus as
5689 * members (mimicking the global state we have today).
5691 struct root_domain def_root_domain
;
5693 static void init_defrootdomain(void)
5695 init_rootdomain(&def_root_domain
);
5697 atomic_set(&def_root_domain
.refcount
, 1);
5700 static struct root_domain
*alloc_rootdomain(void)
5702 struct root_domain
*rd
;
5704 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5708 if (init_rootdomain(rd
) != 0) {
5716 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5718 struct sched_group
*tmp
, *first
;
5727 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5732 } while (sg
!= first
);
5735 static void free_sched_domain(struct rcu_head
*rcu
)
5737 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5740 * If its an overlapping domain it has private groups, iterate and
5743 if (sd
->flags
& SD_OVERLAP
) {
5744 free_sched_groups(sd
->groups
, 1);
5745 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5746 kfree(sd
->groups
->sgc
);
5752 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5754 call_rcu(&sd
->rcu
, free_sched_domain
);
5757 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5759 for (; sd
; sd
= sd
->parent
)
5760 destroy_sched_domain(sd
, cpu
);
5764 * Keep a special pointer to the highest sched_domain that has
5765 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5766 * allows us to avoid some pointer chasing select_idle_sibling().
5768 * Also keep a unique ID per domain (we use the first cpu number in
5769 * the cpumask of the domain), this allows us to quickly tell if
5770 * two cpus are in the same cache domain, see cpus_share_cache().
5772 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5773 DEFINE_PER_CPU(int, sd_llc_size
);
5774 DEFINE_PER_CPU(int, sd_llc_id
);
5775 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5776 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5777 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5779 static void update_top_cache_domain(int cpu
)
5781 struct sched_domain
*sd
;
5782 struct sched_domain
*busy_sd
= NULL
;
5786 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5788 id
= cpumask_first(sched_domain_span(sd
));
5789 size
= cpumask_weight(sched_domain_span(sd
));
5790 busy_sd
= sd
->parent
; /* sd_busy */
5792 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5794 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5795 per_cpu(sd_llc_size
, cpu
) = size
;
5796 per_cpu(sd_llc_id
, cpu
) = id
;
5798 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5799 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5801 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5802 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5806 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5807 * hold the hotplug lock.
5810 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5812 struct rq
*rq
= cpu_rq(cpu
);
5813 struct sched_domain
*tmp
;
5815 /* Remove the sched domains which do not contribute to scheduling. */
5816 for (tmp
= sd
; tmp
; ) {
5817 struct sched_domain
*parent
= tmp
->parent
;
5821 if (sd_parent_degenerate(tmp
, parent
)) {
5822 tmp
->parent
= parent
->parent
;
5824 parent
->parent
->child
= tmp
;
5826 * Transfer SD_PREFER_SIBLING down in case of a
5827 * degenerate parent; the spans match for this
5828 * so the property transfers.
5830 if (parent
->flags
& SD_PREFER_SIBLING
)
5831 tmp
->flags
|= SD_PREFER_SIBLING
;
5832 destroy_sched_domain(parent
, cpu
);
5837 if (sd
&& sd_degenerate(sd
)) {
5840 destroy_sched_domain(tmp
, cpu
);
5845 sched_domain_debug(sd
, cpu
);
5847 rq_attach_root(rq
, rd
);
5849 rcu_assign_pointer(rq
->sd
, sd
);
5850 destroy_sched_domains(tmp
, cpu
);
5852 update_top_cache_domain(cpu
);
5855 /* Setup the mask of cpus configured for isolated domains */
5856 static int __init
isolated_cpu_setup(char *str
)
5858 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
5859 cpulist_parse(str
, cpu_isolated_map
);
5863 __setup("isolcpus=", isolated_cpu_setup
);
5866 struct sched_domain
** __percpu sd
;
5867 struct root_domain
*rd
;
5878 * Build an iteration mask that can exclude certain CPUs from the upwards
5881 * Asymmetric node setups can result in situations where the domain tree is of
5882 * unequal depth, make sure to skip domains that already cover the entire
5885 * In that case build_sched_domains() will have terminated the iteration early
5886 * and our sibling sd spans will be empty. Domains should always include the
5887 * cpu they're built on, so check that.
5890 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
5892 const struct cpumask
*span
= sched_domain_span(sd
);
5893 struct sd_data
*sdd
= sd
->private;
5894 struct sched_domain
*sibling
;
5897 for_each_cpu(i
, span
) {
5898 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5899 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5902 cpumask_set_cpu(i
, sched_group_mask(sg
));
5907 * Return the canonical balance cpu for this group, this is the first cpu
5908 * of this group that's also in the iteration mask.
5910 int group_balance_cpu(struct sched_group
*sg
)
5912 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
5916 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
5918 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
5919 const struct cpumask
*span
= sched_domain_span(sd
);
5920 struct cpumask
*covered
= sched_domains_tmpmask
;
5921 struct sd_data
*sdd
= sd
->private;
5922 struct sched_domain
*sibling
;
5925 cpumask_clear(covered
);
5927 for_each_cpu(i
, span
) {
5928 struct cpumask
*sg_span
;
5930 if (cpumask_test_cpu(i
, covered
))
5933 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
5935 /* See the comment near build_group_mask(). */
5936 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
5939 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
5940 GFP_KERNEL
, cpu_to_node(cpu
));
5945 sg_span
= sched_group_cpus(sg
);
5947 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
5949 cpumask_set_cpu(i
, sg_span
);
5951 cpumask_or(covered
, covered
, sg_span
);
5953 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
5954 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
5955 build_group_mask(sd
, sg
);
5958 * Initialize sgc->capacity such that even if we mess up the
5959 * domains and no possible iteration will get us here, we won't
5962 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
5965 * Make sure the first group of this domain contains the
5966 * canonical balance cpu. Otherwise the sched_domain iteration
5967 * breaks. See update_sg_lb_stats().
5969 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
5970 group_balance_cpu(sg
) == cpu
)
5980 sd
->groups
= groups
;
5985 free_sched_groups(first
, 0);
5990 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
5992 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
5993 struct sched_domain
*child
= sd
->child
;
5996 cpu
= cpumask_first(sched_domain_span(child
));
5999 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6000 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6001 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6008 * build_sched_groups will build a circular linked list of the groups
6009 * covered by the given span, and will set each group's ->cpumask correctly,
6010 * and ->cpu_capacity to 0.
6012 * Assumes the sched_domain tree is fully constructed
6015 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6017 struct sched_group
*first
= NULL
, *last
= NULL
;
6018 struct sd_data
*sdd
= sd
->private;
6019 const struct cpumask
*span
= sched_domain_span(sd
);
6020 struct cpumask
*covered
;
6023 get_group(cpu
, sdd
, &sd
->groups
);
6024 atomic_inc(&sd
->groups
->ref
);
6026 if (cpu
!= cpumask_first(span
))
6029 lockdep_assert_held(&sched_domains_mutex
);
6030 covered
= sched_domains_tmpmask
;
6032 cpumask_clear(covered
);
6034 for_each_cpu(i
, span
) {
6035 struct sched_group
*sg
;
6038 if (cpumask_test_cpu(i
, covered
))
6041 group
= get_group(i
, sdd
, &sg
);
6042 cpumask_setall(sched_group_mask(sg
));
6044 for_each_cpu(j
, span
) {
6045 if (get_group(j
, sdd
, NULL
) != group
)
6048 cpumask_set_cpu(j
, covered
);
6049 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6064 * Initialize sched groups cpu_capacity.
6066 * cpu_capacity indicates the capacity of sched group, which is used while
6067 * distributing the load between different sched groups in a sched domain.
6068 * Typically cpu_capacity for all the groups in a sched domain will be same
6069 * unless there are asymmetries in the topology. If there are asymmetries,
6070 * group having more cpu_capacity will pickup more load compared to the
6071 * group having less cpu_capacity.
6073 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6075 struct sched_group
*sg
= sd
->groups
;
6080 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6082 } while (sg
!= sd
->groups
);
6084 if (cpu
!= group_balance_cpu(sg
))
6087 update_group_capacity(sd
, cpu
);
6088 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6092 * Initializers for schedule domains
6093 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6096 static int default_relax_domain_level
= -1;
6097 int sched_domain_level_max
;
6099 static int __init
setup_relax_domain_level(char *str
)
6101 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6102 pr_warn("Unable to set relax_domain_level\n");
6106 __setup("relax_domain_level=", setup_relax_domain_level
);
6108 static void set_domain_attribute(struct sched_domain
*sd
,
6109 struct sched_domain_attr
*attr
)
6113 if (!attr
|| attr
->relax_domain_level
< 0) {
6114 if (default_relax_domain_level
< 0)
6117 request
= default_relax_domain_level
;
6119 request
= attr
->relax_domain_level
;
6120 if (request
< sd
->level
) {
6121 /* turn off idle balance on this domain */
6122 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6124 /* turn on idle balance on this domain */
6125 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6129 static void __sdt_free(const struct cpumask
*cpu_map
);
6130 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6132 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6133 const struct cpumask
*cpu_map
)
6137 if (!atomic_read(&d
->rd
->refcount
))
6138 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6140 free_percpu(d
->sd
); /* fall through */
6142 __sdt_free(cpu_map
); /* fall through */
6148 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6149 const struct cpumask
*cpu_map
)
6151 memset(d
, 0, sizeof(*d
));
6153 if (__sdt_alloc(cpu_map
))
6154 return sa_sd_storage
;
6155 d
->sd
= alloc_percpu(struct sched_domain
*);
6157 return sa_sd_storage
;
6158 d
->rd
= alloc_rootdomain();
6161 return sa_rootdomain
;
6165 * NULL the sd_data elements we've used to build the sched_domain and
6166 * sched_group structure so that the subsequent __free_domain_allocs()
6167 * will not free the data we're using.
6169 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6171 struct sd_data
*sdd
= sd
->private;
6173 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6174 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6176 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6177 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6179 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6180 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6184 static int sched_domains_numa_levels
;
6185 enum numa_topology_type sched_numa_topology_type
;
6186 static int *sched_domains_numa_distance
;
6187 int sched_max_numa_distance
;
6188 static struct cpumask
***sched_domains_numa_masks
;
6189 static int sched_domains_curr_level
;
6193 * SD_flags allowed in topology descriptions.
6195 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6196 * SD_SHARE_PKG_RESOURCES - describes shared caches
6197 * SD_NUMA - describes NUMA topologies
6198 * SD_SHARE_POWERDOMAIN - describes shared power domain
6201 * SD_ASYM_PACKING - describes SMT quirks
6203 #define TOPOLOGY_SD_FLAGS \
6204 (SD_SHARE_CPUCAPACITY | \
6205 SD_SHARE_PKG_RESOURCES | \
6208 SD_SHARE_POWERDOMAIN)
6210 static struct sched_domain
*
6211 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6213 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6214 int sd_weight
, sd_flags
= 0;
6218 * Ugly hack to pass state to sd_numa_mask()...
6220 sched_domains_curr_level
= tl
->numa_level
;
6223 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6226 sd_flags
= (*tl
->sd_flags
)();
6227 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6228 "wrong sd_flags in topology description\n"))
6229 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6231 *sd
= (struct sched_domain
){
6232 .min_interval
= sd_weight
,
6233 .max_interval
= 2*sd_weight
,
6235 .imbalance_pct
= 125,
6237 .cache_nice_tries
= 0,
6244 .flags
= 1*SD_LOAD_BALANCE
6245 | 1*SD_BALANCE_NEWIDLE
6250 | 0*SD_SHARE_CPUCAPACITY
6251 | 0*SD_SHARE_PKG_RESOURCES
6253 | 0*SD_PREFER_SIBLING
6258 .last_balance
= jiffies
,
6259 .balance_interval
= sd_weight
,
6261 .max_newidle_lb_cost
= 0,
6262 .next_decay_max_lb_cost
= jiffies
,
6263 #ifdef CONFIG_SCHED_DEBUG
6269 * Convert topological properties into behaviour.
6272 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6273 sd
->flags
|= SD_PREFER_SIBLING
;
6274 sd
->imbalance_pct
= 110;
6275 sd
->smt_gain
= 1178; /* ~15% */
6277 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6278 sd
->imbalance_pct
= 117;
6279 sd
->cache_nice_tries
= 1;
6283 } else if (sd
->flags
& SD_NUMA
) {
6284 sd
->cache_nice_tries
= 2;
6288 sd
->flags
|= SD_SERIALIZE
;
6289 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6290 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6297 sd
->flags
|= SD_PREFER_SIBLING
;
6298 sd
->cache_nice_tries
= 1;
6303 sd
->private = &tl
->data
;
6309 * Topology list, bottom-up.
6311 static struct sched_domain_topology_level default_topology
[] = {
6312 #ifdef CONFIG_SCHED_SMT
6313 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6315 #ifdef CONFIG_SCHED_MC
6316 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6318 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6322 struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
6324 #define for_each_sd_topology(tl) \
6325 for (tl = sched_domain_topology; tl->mask; tl++)
6327 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6329 sched_domain_topology
= tl
;
6334 static const struct cpumask
*sd_numa_mask(int cpu
)
6336 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6339 static void sched_numa_warn(const char *str
)
6341 static int done
= false;
6349 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6351 for (i
= 0; i
< nr_node_ids
; i
++) {
6352 printk(KERN_WARNING
" ");
6353 for (j
= 0; j
< nr_node_ids
; j
++)
6354 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6355 printk(KERN_CONT
"\n");
6357 printk(KERN_WARNING
"\n");
6360 bool find_numa_distance(int distance
)
6364 if (distance
== node_distance(0, 0))
6367 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6368 if (sched_domains_numa_distance
[i
] == distance
)
6376 * A system can have three types of NUMA topology:
6377 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6378 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6379 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6381 * The difference between a glueless mesh topology and a backplane
6382 * topology lies in whether communication between not directly
6383 * connected nodes goes through intermediary nodes (where programs
6384 * could run), or through backplane controllers. This affects
6385 * placement of programs.
6387 * The type of topology can be discerned with the following tests:
6388 * - If the maximum distance between any nodes is 1 hop, the system
6389 * is directly connected.
6390 * - If for two nodes A and B, located N > 1 hops away from each other,
6391 * there is an intermediary node C, which is < N hops away from both
6392 * nodes A and B, the system is a glueless mesh.
6394 static void init_numa_topology_type(void)
6398 n
= sched_max_numa_distance
;
6401 sched_numa_topology_type
= NUMA_DIRECT
;
6403 for_each_online_node(a
) {
6404 for_each_online_node(b
) {
6405 /* Find two nodes furthest removed from each other. */
6406 if (node_distance(a
, b
) < n
)
6409 /* Is there an intermediary node between a and b? */
6410 for_each_online_node(c
) {
6411 if (node_distance(a
, c
) < n
&&
6412 node_distance(b
, c
) < n
) {
6413 sched_numa_topology_type
=
6419 sched_numa_topology_type
= NUMA_BACKPLANE
;
6425 static void sched_init_numa(void)
6427 int next_distance
, curr_distance
= node_distance(0, 0);
6428 struct sched_domain_topology_level
*tl
;
6432 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6433 if (!sched_domains_numa_distance
)
6437 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6438 * unique distances in the node_distance() table.
6440 * Assumes node_distance(0,j) includes all distances in
6441 * node_distance(i,j) in order to avoid cubic time.
6443 next_distance
= curr_distance
;
6444 for (i
= 0; i
< nr_node_ids
; i
++) {
6445 for (j
= 0; j
< nr_node_ids
; j
++) {
6446 for (k
= 0; k
< nr_node_ids
; k
++) {
6447 int distance
= node_distance(i
, k
);
6449 if (distance
> curr_distance
&&
6450 (distance
< next_distance
||
6451 next_distance
== curr_distance
))
6452 next_distance
= distance
;
6455 * While not a strong assumption it would be nice to know
6456 * about cases where if node A is connected to B, B is not
6457 * equally connected to A.
6459 if (sched_debug() && node_distance(k
, i
) != distance
)
6460 sched_numa_warn("Node-distance not symmetric");
6462 if (sched_debug() && i
&& !find_numa_distance(distance
))
6463 sched_numa_warn("Node-0 not representative");
6465 if (next_distance
!= curr_distance
) {
6466 sched_domains_numa_distance
[level
++] = next_distance
;
6467 sched_domains_numa_levels
= level
;
6468 curr_distance
= next_distance
;
6473 * In case of sched_debug() we verify the above assumption.
6483 * 'level' contains the number of unique distances, excluding the
6484 * identity distance node_distance(i,i).
6486 * The sched_domains_numa_distance[] array includes the actual distance
6491 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6492 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6493 * the array will contain less then 'level' members. This could be
6494 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6495 * in other functions.
6497 * We reset it to 'level' at the end of this function.
6499 sched_domains_numa_levels
= 0;
6501 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6502 if (!sched_domains_numa_masks
)
6506 * Now for each level, construct a mask per node which contains all
6507 * cpus of nodes that are that many hops away from us.
6509 for (i
= 0; i
< level
; i
++) {
6510 sched_domains_numa_masks
[i
] =
6511 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6512 if (!sched_domains_numa_masks
[i
])
6515 for (j
= 0; j
< nr_node_ids
; j
++) {
6516 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6520 sched_domains_numa_masks
[i
][j
] = mask
;
6522 for (k
= 0; k
< nr_node_ids
; k
++) {
6523 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6526 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6531 /* Compute default topology size */
6532 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6534 tl
= kzalloc((i
+ level
+ 1) *
6535 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6540 * Copy the default topology bits..
6542 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6543 tl
[i
] = sched_domain_topology
[i
];
6546 * .. and append 'j' levels of NUMA goodness.
6548 for (j
= 0; j
< level
; i
++, j
++) {
6549 tl
[i
] = (struct sched_domain_topology_level
){
6550 .mask
= sd_numa_mask
,
6551 .sd_flags
= cpu_numa_flags
,
6552 .flags
= SDTL_OVERLAP
,
6558 sched_domain_topology
= tl
;
6560 sched_domains_numa_levels
= level
;
6561 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6563 init_numa_topology_type();
6566 static void sched_domains_numa_masks_set(int cpu
)
6569 int node
= cpu_to_node(cpu
);
6571 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6572 for (j
= 0; j
< nr_node_ids
; j
++) {
6573 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6574 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6579 static void sched_domains_numa_masks_clear(int cpu
)
6582 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6583 for (j
= 0; j
< nr_node_ids
; j
++)
6584 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6589 * Update sched_domains_numa_masks[level][node] array when new cpus
6592 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6593 unsigned long action
,
6596 int cpu
= (long)hcpu
;
6598 switch (action
& ~CPU_TASKS_FROZEN
) {
6600 sched_domains_numa_masks_set(cpu
);
6604 sched_domains_numa_masks_clear(cpu
);
6614 static inline void sched_init_numa(void)
6618 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6619 unsigned long action
,
6624 #endif /* CONFIG_NUMA */
6626 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6628 struct sched_domain_topology_level
*tl
;
6631 for_each_sd_topology(tl
) {
6632 struct sd_data
*sdd
= &tl
->data
;
6634 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6638 sdd
->sg
= alloc_percpu(struct sched_group
*);
6642 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6646 for_each_cpu(j
, cpu_map
) {
6647 struct sched_domain
*sd
;
6648 struct sched_group
*sg
;
6649 struct sched_group_capacity
*sgc
;
6651 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6652 GFP_KERNEL
, cpu_to_node(j
));
6656 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6658 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6659 GFP_KERNEL
, cpu_to_node(j
));
6665 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6667 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6668 GFP_KERNEL
, cpu_to_node(j
));
6672 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6679 static void __sdt_free(const struct cpumask
*cpu_map
)
6681 struct sched_domain_topology_level
*tl
;
6684 for_each_sd_topology(tl
) {
6685 struct sd_data
*sdd
= &tl
->data
;
6687 for_each_cpu(j
, cpu_map
) {
6688 struct sched_domain
*sd
;
6691 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6692 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6693 free_sched_groups(sd
->groups
, 0);
6694 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6698 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6700 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6702 free_percpu(sdd
->sd
);
6704 free_percpu(sdd
->sg
);
6706 free_percpu(sdd
->sgc
);
6711 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6712 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6713 struct sched_domain
*child
, int cpu
)
6715 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6719 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6721 sd
->level
= child
->level
+ 1;
6722 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6726 if (!cpumask_subset(sched_domain_span(child
),
6727 sched_domain_span(sd
))) {
6728 pr_err("BUG: arch topology borken\n");
6729 #ifdef CONFIG_SCHED_DEBUG
6730 pr_err(" the %s domain not a subset of the %s domain\n",
6731 child
->name
, sd
->name
);
6733 /* Fixup, ensure @sd has at least @child cpus. */
6734 cpumask_or(sched_domain_span(sd
),
6735 sched_domain_span(sd
),
6736 sched_domain_span(child
));
6740 set_domain_attribute(sd
, attr
);
6746 * Build sched domains for a given set of cpus and attach the sched domains
6747 * to the individual cpus
6749 static int build_sched_domains(const struct cpumask
*cpu_map
,
6750 struct sched_domain_attr
*attr
)
6752 enum s_alloc alloc_state
;
6753 struct sched_domain
*sd
;
6755 int i
, ret
= -ENOMEM
;
6757 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6758 if (alloc_state
!= sa_rootdomain
)
6761 /* Set up domains for cpus specified by the cpu_map. */
6762 for_each_cpu(i
, cpu_map
) {
6763 struct sched_domain_topology_level
*tl
;
6766 for_each_sd_topology(tl
) {
6767 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6768 if (tl
== sched_domain_topology
)
6769 *per_cpu_ptr(d
.sd
, i
) = sd
;
6770 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6771 sd
->flags
|= SD_OVERLAP
;
6772 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6777 /* Build the groups for the domains */
6778 for_each_cpu(i
, cpu_map
) {
6779 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6780 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6781 if (sd
->flags
& SD_OVERLAP
) {
6782 if (build_overlap_sched_groups(sd
, i
))
6785 if (build_sched_groups(sd
, i
))
6791 /* Calculate CPU capacity for physical packages and nodes */
6792 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6793 if (!cpumask_test_cpu(i
, cpu_map
))
6796 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6797 claim_allocations(i
, sd
);
6798 init_sched_groups_capacity(i
, sd
);
6802 /* Attach the domains */
6804 for_each_cpu(i
, cpu_map
) {
6805 sd
= *per_cpu_ptr(d
.sd
, i
);
6806 cpu_attach_domain(sd
, d
.rd
, i
);
6812 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6816 static cpumask_var_t
*doms_cur
; /* current sched domains */
6817 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6818 static struct sched_domain_attr
*dattr_cur
;
6819 /* attribues of custom domains in 'doms_cur' */
6822 * Special case: If a kmalloc of a doms_cur partition (array of
6823 * cpumask) fails, then fallback to a single sched domain,
6824 * as determined by the single cpumask fallback_doms.
6826 static cpumask_var_t fallback_doms
;
6829 * arch_update_cpu_topology lets virtualized architectures update the
6830 * cpu core maps. It is supposed to return 1 if the topology changed
6831 * or 0 if it stayed the same.
6833 int __weak
arch_update_cpu_topology(void)
6838 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6841 cpumask_var_t
*doms
;
6843 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
6846 for (i
= 0; i
< ndoms
; i
++) {
6847 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
6848 free_sched_domains(doms
, i
);
6855 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
6858 for (i
= 0; i
< ndoms
; i
++)
6859 free_cpumask_var(doms
[i
]);
6864 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6865 * For now this just excludes isolated cpus, but could be used to
6866 * exclude other special cases in the future.
6868 static int init_sched_domains(const struct cpumask
*cpu_map
)
6872 arch_update_cpu_topology();
6874 doms_cur
= alloc_sched_domains(ndoms_cur
);
6876 doms_cur
= &fallback_doms
;
6877 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
6878 err
= build_sched_domains(doms_cur
[0], NULL
);
6879 register_sched_domain_sysctl();
6885 * Detach sched domains from a group of cpus specified in cpu_map
6886 * These cpus will now be attached to the NULL domain
6888 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
6893 for_each_cpu(i
, cpu_map
)
6894 cpu_attach_domain(NULL
, &def_root_domain
, i
);
6898 /* handle null as "default" */
6899 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
6900 struct sched_domain_attr
*new, int idx_new
)
6902 struct sched_domain_attr tmp
;
6909 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
6910 new ? (new + idx_new
) : &tmp
,
6911 sizeof(struct sched_domain_attr
));
6915 * Partition sched domains as specified by the 'ndoms_new'
6916 * cpumasks in the array doms_new[] of cpumasks. This compares
6917 * doms_new[] to the current sched domain partitioning, doms_cur[].
6918 * It destroys each deleted domain and builds each new domain.
6920 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6921 * The masks don't intersect (don't overlap.) We should setup one
6922 * sched domain for each mask. CPUs not in any of the cpumasks will
6923 * not be load balanced. If the same cpumask appears both in the
6924 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6927 * The passed in 'doms_new' should be allocated using
6928 * alloc_sched_domains. This routine takes ownership of it and will
6929 * free_sched_domains it when done with it. If the caller failed the
6930 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6931 * and partition_sched_domains() will fallback to the single partition
6932 * 'fallback_doms', it also forces the domains to be rebuilt.
6934 * If doms_new == NULL it will be replaced with cpu_online_mask.
6935 * ndoms_new == 0 is a special case for destroying existing domains,
6936 * and it will not create the default domain.
6938 * Call with hotplug lock held
6940 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
6941 struct sched_domain_attr
*dattr_new
)
6946 mutex_lock(&sched_domains_mutex
);
6948 /* always unregister in case we don't destroy any domains */
6949 unregister_sched_domain_sysctl();
6951 /* Let architecture update cpu core mappings. */
6952 new_topology
= arch_update_cpu_topology();
6954 n
= doms_new
? ndoms_new
: 0;
6956 /* Destroy deleted domains */
6957 for (i
= 0; i
< ndoms_cur
; i
++) {
6958 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6959 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
6960 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
6963 /* no match - a current sched domain not in new doms_new[] */
6964 detach_destroy_domains(doms_cur
[i
]);
6970 if (doms_new
== NULL
) {
6972 doms_new
= &fallback_doms
;
6973 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
6974 WARN_ON_ONCE(dattr_new
);
6977 /* Build new domains */
6978 for (i
= 0; i
< ndoms_new
; i
++) {
6979 for (j
= 0; j
< n
&& !new_topology
; j
++) {
6980 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
6981 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
6984 /* no match - add a new doms_new */
6985 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
6990 /* Remember the new sched domains */
6991 if (doms_cur
!= &fallback_doms
)
6992 free_sched_domains(doms_cur
, ndoms_cur
);
6993 kfree(dattr_cur
); /* kfree(NULL) is safe */
6994 doms_cur
= doms_new
;
6995 dattr_cur
= dattr_new
;
6996 ndoms_cur
= ndoms_new
;
6998 register_sched_domain_sysctl();
7000 mutex_unlock(&sched_domains_mutex
);
7003 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7006 * Update cpusets according to cpu_active mask. If cpusets are
7007 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7008 * around partition_sched_domains().
7010 * If we come here as part of a suspend/resume, don't touch cpusets because we
7011 * want to restore it back to its original state upon resume anyway.
7013 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7017 case CPU_ONLINE_FROZEN
:
7018 case CPU_DOWN_FAILED_FROZEN
:
7021 * num_cpus_frozen tracks how many CPUs are involved in suspend
7022 * resume sequence. As long as this is not the last online
7023 * operation in the resume sequence, just build a single sched
7024 * domain, ignoring cpusets.
7027 if (likely(num_cpus_frozen
)) {
7028 partition_sched_domains(1, NULL
, NULL
);
7033 * This is the last CPU online operation. So fall through and
7034 * restore the original sched domains by considering the
7035 * cpuset configurations.
7039 cpuset_update_active_cpus(true);
7047 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7050 unsigned long flags
;
7051 long cpu
= (long)hcpu
;
7057 case CPU_DOWN_PREPARE
:
7058 rcu_read_lock_sched();
7059 dl_b
= dl_bw_of(cpu
);
7061 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7062 cpus
= dl_bw_cpus(cpu
);
7063 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7064 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7066 rcu_read_unlock_sched();
7069 return notifier_from_errno(-EBUSY
);
7070 cpuset_update_active_cpus(false);
7072 case CPU_DOWN_PREPARE_FROZEN
:
7074 partition_sched_domains(1, NULL
, NULL
);
7082 void __init
sched_init_smp(void)
7084 cpumask_var_t non_isolated_cpus
;
7086 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7087 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7092 * There's no userspace yet to cause hotplug operations; hence all the
7093 * cpu masks are stable and all blatant races in the below code cannot
7096 mutex_lock(&sched_domains_mutex
);
7097 init_sched_domains(cpu_active_mask
);
7098 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7099 if (cpumask_empty(non_isolated_cpus
))
7100 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7101 mutex_unlock(&sched_domains_mutex
);
7103 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7104 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7105 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7109 /* Move init over to a non-isolated CPU */
7110 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7112 sched_init_granularity();
7113 free_cpumask_var(non_isolated_cpus
);
7115 init_sched_rt_class();
7116 init_sched_dl_class();
7119 void __init
sched_init_smp(void)
7121 sched_init_granularity();
7123 #endif /* CONFIG_SMP */
7125 const_debug
unsigned int sysctl_timer_migration
= 1;
7127 int in_sched_functions(unsigned long addr
)
7129 return in_lock_functions(addr
) ||
7130 (addr
>= (unsigned long)__sched_text_start
7131 && addr
< (unsigned long)__sched_text_end
);
7134 #ifdef CONFIG_CGROUP_SCHED
7136 * Default task group.
7137 * Every task in system belongs to this group at bootup.
7139 struct task_group root_task_group
;
7140 LIST_HEAD(task_groups
);
7143 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7145 void __init
sched_init(void)
7148 unsigned long alloc_size
= 0, ptr
;
7150 #ifdef CONFIG_FAIR_GROUP_SCHED
7151 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7153 #ifdef CONFIG_RT_GROUP_SCHED
7154 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7157 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7159 #ifdef CONFIG_FAIR_GROUP_SCHED
7160 root_task_group
.se
= (struct sched_entity
**)ptr
;
7161 ptr
+= nr_cpu_ids
* sizeof(void **);
7163 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7164 ptr
+= nr_cpu_ids
* sizeof(void **);
7166 #endif /* CONFIG_FAIR_GROUP_SCHED */
7167 #ifdef CONFIG_RT_GROUP_SCHED
7168 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7169 ptr
+= nr_cpu_ids
* sizeof(void **);
7171 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7172 ptr
+= nr_cpu_ids
* sizeof(void **);
7174 #endif /* CONFIG_RT_GROUP_SCHED */
7176 #ifdef CONFIG_CPUMASK_OFFSTACK
7177 for_each_possible_cpu(i
) {
7178 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7179 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7181 #endif /* CONFIG_CPUMASK_OFFSTACK */
7183 init_rt_bandwidth(&def_rt_bandwidth
,
7184 global_rt_period(), global_rt_runtime());
7185 init_dl_bandwidth(&def_dl_bandwidth
,
7186 global_rt_period(), global_rt_runtime());
7189 init_defrootdomain();
7192 #ifdef CONFIG_RT_GROUP_SCHED
7193 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7194 global_rt_period(), global_rt_runtime());
7195 #endif /* CONFIG_RT_GROUP_SCHED */
7197 #ifdef CONFIG_CGROUP_SCHED
7198 list_add(&root_task_group
.list
, &task_groups
);
7199 INIT_LIST_HEAD(&root_task_group
.children
);
7200 INIT_LIST_HEAD(&root_task_group
.siblings
);
7201 autogroup_init(&init_task
);
7203 #endif /* CONFIG_CGROUP_SCHED */
7205 for_each_possible_cpu(i
) {
7209 raw_spin_lock_init(&rq
->lock
);
7211 rq
->calc_load_active
= 0;
7212 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7213 init_cfs_rq(&rq
->cfs
);
7214 init_rt_rq(&rq
->rt
);
7215 init_dl_rq(&rq
->dl
);
7216 #ifdef CONFIG_FAIR_GROUP_SCHED
7217 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7218 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7220 * How much cpu bandwidth does root_task_group get?
7222 * In case of task-groups formed thr' the cgroup filesystem, it
7223 * gets 100% of the cpu resources in the system. This overall
7224 * system cpu resource is divided among the tasks of
7225 * root_task_group and its child task-groups in a fair manner,
7226 * based on each entity's (task or task-group's) weight
7227 * (se->load.weight).
7229 * In other words, if root_task_group has 10 tasks of weight
7230 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7231 * then A0's share of the cpu resource is:
7233 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7235 * We achieve this by letting root_task_group's tasks sit
7236 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7238 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7239 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7240 #endif /* CONFIG_FAIR_GROUP_SCHED */
7242 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7243 #ifdef CONFIG_RT_GROUP_SCHED
7244 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7247 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7248 rq
->cpu_load
[j
] = 0;
7250 rq
->last_load_update_tick
= jiffies
;
7255 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7256 rq
->balance_callback
= NULL
;
7257 rq
->active_balance
= 0;
7258 rq
->next_balance
= jiffies
;
7263 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7264 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7266 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7268 rq_attach_root(rq
, &def_root_domain
);
7269 #ifdef CONFIG_NO_HZ_COMMON
7272 #ifdef CONFIG_NO_HZ_FULL
7273 rq
->last_sched_tick
= 0;
7277 atomic_set(&rq
->nr_iowait
, 0);
7280 set_load_weight(&init_task
);
7282 #ifdef CONFIG_PREEMPT_NOTIFIERS
7283 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7287 * The boot idle thread does lazy MMU switching as well:
7289 atomic_inc(&init_mm
.mm_count
);
7290 enter_lazy_tlb(&init_mm
, current
);
7293 * During early bootup we pretend to be a normal task:
7295 current
->sched_class
= &fair_sched_class
;
7298 * Make us the idle thread. Technically, schedule() should not be
7299 * called from this thread, however somewhere below it might be,
7300 * but because we are the idle thread, we just pick up running again
7301 * when this runqueue becomes "idle".
7303 init_idle(current
, smp_processor_id());
7305 calc_load_update
= jiffies
+ LOAD_FREQ
;
7308 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7309 /* May be allocated at isolcpus cmdline parse time */
7310 if (cpu_isolated_map
== NULL
)
7311 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7312 idle_thread_set_boot_cpu();
7313 set_cpu_rq_start_time();
7315 init_sched_fair_class();
7317 scheduler_running
= 1;
7320 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7321 static inline int preempt_count_equals(int preempt_offset
)
7323 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7325 return (nested
== preempt_offset
);
7328 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7331 * Blocking primitives will set (and therefore destroy) current->state,
7332 * since we will exit with TASK_RUNNING make sure we enter with it,
7333 * otherwise we will destroy state.
7335 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7336 "do not call blocking ops when !TASK_RUNNING; "
7337 "state=%lx set at [<%p>] %pS\n",
7339 (void *)current
->task_state_change
,
7340 (void *)current
->task_state_change
);
7342 ___might_sleep(file
, line
, preempt_offset
);
7344 EXPORT_SYMBOL(__might_sleep
);
7346 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7348 static unsigned long prev_jiffy
; /* ratelimiting */
7350 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7351 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7352 !is_idle_task(current
)) ||
7353 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7355 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7357 prev_jiffy
= jiffies
;
7360 "BUG: sleeping function called from invalid context at %s:%d\n",
7363 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7364 in_atomic(), irqs_disabled(),
7365 current
->pid
, current
->comm
);
7367 if (task_stack_end_corrupted(current
))
7368 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7370 debug_show_held_locks(current
);
7371 if (irqs_disabled())
7372 print_irqtrace_events(current
);
7373 #ifdef CONFIG_DEBUG_PREEMPT
7374 if (!preempt_count_equals(preempt_offset
)) {
7375 pr_err("Preemption disabled at:");
7376 print_ip_sym(current
->preempt_disable_ip
);
7382 EXPORT_SYMBOL(___might_sleep
);
7385 #ifdef CONFIG_MAGIC_SYSRQ
7386 void normalize_rt_tasks(void)
7388 struct task_struct
*g
, *p
;
7389 struct sched_attr attr
= {
7390 .sched_policy
= SCHED_NORMAL
,
7393 read_lock(&tasklist_lock
);
7394 for_each_process_thread(g
, p
) {
7396 * Only normalize user tasks:
7398 if (p
->flags
& PF_KTHREAD
)
7401 p
->se
.exec_start
= 0;
7402 #ifdef CONFIG_SCHEDSTATS
7403 p
->se
.statistics
.wait_start
= 0;
7404 p
->se
.statistics
.sleep_start
= 0;
7405 p
->se
.statistics
.block_start
= 0;
7408 if (!dl_task(p
) && !rt_task(p
)) {
7410 * Renice negative nice level userspace
7413 if (task_nice(p
) < 0)
7414 set_user_nice(p
, 0);
7418 __sched_setscheduler(p
, &attr
, false, false);
7420 read_unlock(&tasklist_lock
);
7423 #endif /* CONFIG_MAGIC_SYSRQ */
7425 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7427 * These functions are only useful for the IA64 MCA handling, or kdb.
7429 * They can only be called when the whole system has been
7430 * stopped - every CPU needs to be quiescent, and no scheduling
7431 * activity can take place. Using them for anything else would
7432 * be a serious bug, and as a result, they aren't even visible
7433 * under any other configuration.
7437 * curr_task - return the current task for a given cpu.
7438 * @cpu: the processor in question.
7440 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7442 * Return: The current task for @cpu.
7444 struct task_struct
*curr_task(int cpu
)
7446 return cpu_curr(cpu
);
7449 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7453 * set_curr_task - set the current task for a given cpu.
7454 * @cpu: the processor in question.
7455 * @p: the task pointer to set.
7457 * Description: This function must only be used when non-maskable interrupts
7458 * are serviced on a separate stack. It allows the architecture to switch the
7459 * notion of the current task on a cpu in a non-blocking manner. This function
7460 * must be called with all CPU's synchronized, and interrupts disabled, the
7461 * and caller must save the original value of the current task (see
7462 * curr_task() above) and restore that value before reenabling interrupts and
7463 * re-starting the system.
7465 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7467 void set_curr_task(int cpu
, struct task_struct
*p
)
7474 #ifdef CONFIG_CGROUP_SCHED
7475 /* task_group_lock serializes the addition/removal of task groups */
7476 static DEFINE_SPINLOCK(task_group_lock
);
7478 static void free_sched_group(struct task_group
*tg
)
7480 free_fair_sched_group(tg
);
7481 free_rt_sched_group(tg
);
7486 /* allocate runqueue etc for a new task group */
7487 struct task_group
*sched_create_group(struct task_group
*parent
)
7489 struct task_group
*tg
;
7491 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
7493 return ERR_PTR(-ENOMEM
);
7495 if (!alloc_fair_sched_group(tg
, parent
))
7498 if (!alloc_rt_sched_group(tg
, parent
))
7504 free_sched_group(tg
);
7505 return ERR_PTR(-ENOMEM
);
7508 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7510 unsigned long flags
;
7512 spin_lock_irqsave(&task_group_lock
, flags
);
7513 list_add_rcu(&tg
->list
, &task_groups
);
7515 WARN_ON(!parent
); /* root should already exist */
7517 tg
->parent
= parent
;
7518 INIT_LIST_HEAD(&tg
->children
);
7519 list_add_rcu(&tg
->siblings
, &parent
->children
);
7520 spin_unlock_irqrestore(&task_group_lock
, flags
);
7523 /* rcu callback to free various structures associated with a task group */
7524 static void free_sched_group_rcu(struct rcu_head
*rhp
)
7526 /* now it should be safe to free those cfs_rqs */
7527 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
7530 /* Destroy runqueue etc associated with a task group */
7531 void sched_destroy_group(struct task_group
*tg
)
7533 /* wait for possible concurrent references to cfs_rqs complete */
7534 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
7537 void sched_offline_group(struct task_group
*tg
)
7539 unsigned long flags
;
7542 /* end participation in shares distribution */
7543 for_each_possible_cpu(i
)
7544 unregister_fair_sched_group(tg
, i
);
7546 spin_lock_irqsave(&task_group_lock
, flags
);
7547 list_del_rcu(&tg
->list
);
7548 list_del_rcu(&tg
->siblings
);
7549 spin_unlock_irqrestore(&task_group_lock
, flags
);
7552 /* change task's runqueue when it moves between groups.
7553 * The caller of this function should have put the task in its new group
7554 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7555 * reflect its new group.
7557 void sched_move_task(struct task_struct
*tsk
)
7559 struct task_group
*tg
;
7560 int queued
, running
;
7561 unsigned long flags
;
7564 rq
= task_rq_lock(tsk
, &flags
);
7566 running
= task_current(rq
, tsk
);
7567 queued
= task_on_rq_queued(tsk
);
7570 dequeue_task(rq
, tsk
, 0);
7571 if (unlikely(running
))
7572 put_prev_task(rq
, tsk
);
7575 * All callers are synchronized by task_rq_lock(); we do not use RCU
7576 * which is pointless here. Thus, we pass "true" to task_css_check()
7577 * to prevent lockdep warnings.
7579 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7580 struct task_group
, css
);
7581 tg
= autogroup_task_group(tsk
, tg
);
7582 tsk
->sched_task_group
= tg
;
7584 #ifdef CONFIG_FAIR_GROUP_SCHED
7585 if (tsk
->sched_class
->task_move_group
)
7586 tsk
->sched_class
->task_move_group(tsk
, queued
);
7589 set_task_rq(tsk
, task_cpu(tsk
));
7591 if (unlikely(running
))
7592 tsk
->sched_class
->set_curr_task(rq
);
7594 enqueue_task(rq
, tsk
, 0);
7596 task_rq_unlock(rq
, tsk
, &flags
);
7598 #endif /* CONFIG_CGROUP_SCHED */
7600 #ifdef CONFIG_RT_GROUP_SCHED
7602 * Ensure that the real time constraints are schedulable.
7604 static DEFINE_MUTEX(rt_constraints_mutex
);
7606 /* Must be called with tasklist_lock held */
7607 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7609 struct task_struct
*g
, *p
;
7612 * Autogroups do not have RT tasks; see autogroup_create().
7614 if (task_group_is_autogroup(tg
))
7617 for_each_process_thread(g
, p
) {
7618 if (rt_task(p
) && task_group(p
) == tg
)
7625 struct rt_schedulable_data
{
7626 struct task_group
*tg
;
7631 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7633 struct rt_schedulable_data
*d
= data
;
7634 struct task_group
*child
;
7635 unsigned long total
, sum
= 0;
7636 u64 period
, runtime
;
7638 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7639 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7642 period
= d
->rt_period
;
7643 runtime
= d
->rt_runtime
;
7647 * Cannot have more runtime than the period.
7649 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7653 * Ensure we don't starve existing RT tasks.
7655 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7658 total
= to_ratio(period
, runtime
);
7661 * Nobody can have more than the global setting allows.
7663 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7667 * The sum of our children's runtime should not exceed our own.
7669 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7670 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7671 runtime
= child
->rt_bandwidth
.rt_runtime
;
7673 if (child
== d
->tg
) {
7674 period
= d
->rt_period
;
7675 runtime
= d
->rt_runtime
;
7678 sum
+= to_ratio(period
, runtime
);
7687 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7691 struct rt_schedulable_data data
= {
7693 .rt_period
= period
,
7694 .rt_runtime
= runtime
,
7698 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7704 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7705 u64 rt_period
, u64 rt_runtime
)
7710 * Disallowing the root group RT runtime is BAD, it would disallow the
7711 * kernel creating (and or operating) RT threads.
7713 if (tg
== &root_task_group
&& rt_runtime
== 0)
7716 /* No period doesn't make any sense. */
7720 mutex_lock(&rt_constraints_mutex
);
7721 read_lock(&tasklist_lock
);
7722 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7726 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7727 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7728 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7730 for_each_possible_cpu(i
) {
7731 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7733 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7734 rt_rq
->rt_runtime
= rt_runtime
;
7735 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7737 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7739 read_unlock(&tasklist_lock
);
7740 mutex_unlock(&rt_constraints_mutex
);
7745 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7747 u64 rt_runtime
, rt_period
;
7749 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7750 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7751 if (rt_runtime_us
< 0)
7752 rt_runtime
= RUNTIME_INF
;
7754 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7757 static long sched_group_rt_runtime(struct task_group
*tg
)
7761 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7764 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7765 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7766 return rt_runtime_us
;
7769 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7771 u64 rt_runtime
, rt_period
;
7773 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7774 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7776 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7779 static long sched_group_rt_period(struct task_group
*tg
)
7783 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7784 do_div(rt_period_us
, NSEC_PER_USEC
);
7785 return rt_period_us
;
7787 #endif /* CONFIG_RT_GROUP_SCHED */
7789 #ifdef CONFIG_RT_GROUP_SCHED
7790 static int sched_rt_global_constraints(void)
7794 mutex_lock(&rt_constraints_mutex
);
7795 read_lock(&tasklist_lock
);
7796 ret
= __rt_schedulable(NULL
, 0, 0);
7797 read_unlock(&tasklist_lock
);
7798 mutex_unlock(&rt_constraints_mutex
);
7803 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
7805 /* Don't accept realtime tasks when there is no way for them to run */
7806 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
7812 #else /* !CONFIG_RT_GROUP_SCHED */
7813 static int sched_rt_global_constraints(void)
7815 unsigned long flags
;
7818 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7819 for_each_possible_cpu(i
) {
7820 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
7822 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7823 rt_rq
->rt_runtime
= global_rt_runtime();
7824 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7826 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
7830 #endif /* CONFIG_RT_GROUP_SCHED */
7832 static int sched_dl_global_validate(void)
7834 u64 runtime
= global_rt_runtime();
7835 u64 period
= global_rt_period();
7836 u64 new_bw
= to_ratio(period
, runtime
);
7839 unsigned long flags
;
7842 * Here we want to check the bandwidth not being set to some
7843 * value smaller than the currently allocated bandwidth in
7844 * any of the root_domains.
7846 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7847 * cycling on root_domains... Discussion on different/better
7848 * solutions is welcome!
7850 for_each_possible_cpu(cpu
) {
7851 rcu_read_lock_sched();
7852 dl_b
= dl_bw_of(cpu
);
7854 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7855 if (new_bw
< dl_b
->total_bw
)
7857 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7859 rcu_read_unlock_sched();
7868 static void sched_dl_do_global(void)
7873 unsigned long flags
;
7875 def_dl_bandwidth
.dl_period
= global_rt_period();
7876 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
7878 if (global_rt_runtime() != RUNTIME_INF
)
7879 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
7882 * FIXME: As above...
7884 for_each_possible_cpu(cpu
) {
7885 rcu_read_lock_sched();
7886 dl_b
= dl_bw_of(cpu
);
7888 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7890 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7892 rcu_read_unlock_sched();
7896 static int sched_rt_global_validate(void)
7898 if (sysctl_sched_rt_period
<= 0)
7901 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
7902 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
7908 static void sched_rt_do_global(void)
7910 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
7911 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
7914 int sched_rt_handler(struct ctl_table
*table
, int write
,
7915 void __user
*buffer
, size_t *lenp
,
7918 int old_period
, old_runtime
;
7919 static DEFINE_MUTEX(mutex
);
7923 old_period
= sysctl_sched_rt_period
;
7924 old_runtime
= sysctl_sched_rt_runtime
;
7926 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7928 if (!ret
&& write
) {
7929 ret
= sched_rt_global_validate();
7933 ret
= sched_dl_global_validate();
7937 ret
= sched_rt_global_constraints();
7941 sched_rt_do_global();
7942 sched_dl_do_global();
7946 sysctl_sched_rt_period
= old_period
;
7947 sysctl_sched_rt_runtime
= old_runtime
;
7949 mutex_unlock(&mutex
);
7954 int sched_rr_handler(struct ctl_table
*table
, int write
,
7955 void __user
*buffer
, size_t *lenp
,
7959 static DEFINE_MUTEX(mutex
);
7962 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
7963 /* make sure that internally we keep jiffies */
7964 /* also, writing zero resets timeslice to default */
7965 if (!ret
&& write
) {
7966 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
7967 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
7969 mutex_unlock(&mutex
);
7973 #ifdef CONFIG_CGROUP_SCHED
7975 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7977 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7980 static struct cgroup_subsys_state
*
7981 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7983 struct task_group
*parent
= css_tg(parent_css
);
7984 struct task_group
*tg
;
7987 /* This is early initialization for the top cgroup */
7988 return &root_task_group
.css
;
7991 tg
= sched_create_group(parent
);
7993 return ERR_PTR(-ENOMEM
);
7998 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
8000 struct task_group
*tg
= css_tg(css
);
8001 struct task_group
*parent
= css_tg(css
->parent
);
8004 sched_online_group(tg
, parent
);
8008 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8010 struct task_group
*tg
= css_tg(css
);
8012 sched_destroy_group(tg
);
8015 static void cpu_cgroup_css_offline(struct cgroup_subsys_state
*css
)
8017 struct task_group
*tg
= css_tg(css
);
8019 sched_offline_group(tg
);
8022 static void cpu_cgroup_fork(struct task_struct
*task
)
8024 sched_move_task(task
);
8027 static int cpu_cgroup_can_attach(struct cgroup_subsys_state
*css
,
8028 struct cgroup_taskset
*tset
)
8030 struct task_struct
*task
;
8032 cgroup_taskset_for_each(task
, tset
) {
8033 #ifdef CONFIG_RT_GROUP_SCHED
8034 if (!sched_rt_can_attach(css_tg(css
), task
))
8037 /* We don't support RT-tasks being in separate groups */
8038 if (task
->sched_class
!= &fair_sched_class
)
8045 static void cpu_cgroup_attach(struct cgroup_subsys_state
*css
,
8046 struct cgroup_taskset
*tset
)
8048 struct task_struct
*task
;
8050 cgroup_taskset_for_each(task
, tset
)
8051 sched_move_task(task
);
8054 static void cpu_cgroup_exit(struct cgroup_subsys_state
*css
,
8055 struct cgroup_subsys_state
*old_css
,
8056 struct task_struct
*task
)
8059 * cgroup_exit() is called in the copy_process() failure path.
8060 * Ignore this case since the task hasn't ran yet, this avoids
8061 * trying to poke a half freed task state from generic code.
8063 if (!(task
->flags
& PF_EXITING
))
8066 sched_move_task(task
);
8069 #ifdef CONFIG_FAIR_GROUP_SCHED
8070 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8071 struct cftype
*cftype
, u64 shareval
)
8073 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8076 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8079 struct task_group
*tg
= css_tg(css
);
8081 return (u64
) scale_load_down(tg
->shares
);
8084 #ifdef CONFIG_CFS_BANDWIDTH
8085 static DEFINE_MUTEX(cfs_constraints_mutex
);
8087 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8088 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8090 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8092 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8094 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8095 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8097 if (tg
== &root_task_group
)
8101 * Ensure we have at some amount of bandwidth every period. This is
8102 * to prevent reaching a state of large arrears when throttled via
8103 * entity_tick() resulting in prolonged exit starvation.
8105 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8109 * Likewise, bound things on the otherside by preventing insane quota
8110 * periods. This also allows us to normalize in computing quota
8113 if (period
> max_cfs_quota_period
)
8117 * Prevent race between setting of cfs_rq->runtime_enabled and
8118 * unthrottle_offline_cfs_rqs().
8121 mutex_lock(&cfs_constraints_mutex
);
8122 ret
= __cfs_schedulable(tg
, period
, quota
);
8126 runtime_enabled
= quota
!= RUNTIME_INF
;
8127 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8129 * If we need to toggle cfs_bandwidth_used, off->on must occur
8130 * before making related changes, and on->off must occur afterwards
8132 if (runtime_enabled
&& !runtime_was_enabled
)
8133 cfs_bandwidth_usage_inc();
8134 raw_spin_lock_irq(&cfs_b
->lock
);
8135 cfs_b
->period
= ns_to_ktime(period
);
8136 cfs_b
->quota
= quota
;
8138 __refill_cfs_bandwidth_runtime(cfs_b
);
8139 /* restart the period timer (if active) to handle new period expiry */
8140 if (runtime_enabled
)
8141 start_cfs_bandwidth(cfs_b
);
8142 raw_spin_unlock_irq(&cfs_b
->lock
);
8144 for_each_online_cpu(i
) {
8145 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8146 struct rq
*rq
= cfs_rq
->rq
;
8148 raw_spin_lock_irq(&rq
->lock
);
8149 cfs_rq
->runtime_enabled
= runtime_enabled
;
8150 cfs_rq
->runtime_remaining
= 0;
8152 if (cfs_rq
->throttled
)
8153 unthrottle_cfs_rq(cfs_rq
);
8154 raw_spin_unlock_irq(&rq
->lock
);
8156 if (runtime_was_enabled
&& !runtime_enabled
)
8157 cfs_bandwidth_usage_dec();
8159 mutex_unlock(&cfs_constraints_mutex
);
8165 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8169 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8170 if (cfs_quota_us
< 0)
8171 quota
= RUNTIME_INF
;
8173 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8175 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8178 long tg_get_cfs_quota(struct task_group
*tg
)
8182 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8185 quota_us
= tg
->cfs_bandwidth
.quota
;
8186 do_div(quota_us
, NSEC_PER_USEC
);
8191 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8195 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8196 quota
= tg
->cfs_bandwidth
.quota
;
8198 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8201 long tg_get_cfs_period(struct task_group
*tg
)
8205 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8206 do_div(cfs_period_us
, NSEC_PER_USEC
);
8208 return cfs_period_us
;
8211 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8214 return tg_get_cfs_quota(css_tg(css
));
8217 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8218 struct cftype
*cftype
, s64 cfs_quota_us
)
8220 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8223 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8226 return tg_get_cfs_period(css_tg(css
));
8229 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8230 struct cftype
*cftype
, u64 cfs_period_us
)
8232 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8235 struct cfs_schedulable_data
{
8236 struct task_group
*tg
;
8241 * normalize group quota/period to be quota/max_period
8242 * note: units are usecs
8244 static u64
normalize_cfs_quota(struct task_group
*tg
,
8245 struct cfs_schedulable_data
*d
)
8253 period
= tg_get_cfs_period(tg
);
8254 quota
= tg_get_cfs_quota(tg
);
8257 /* note: these should typically be equivalent */
8258 if (quota
== RUNTIME_INF
|| quota
== -1)
8261 return to_ratio(period
, quota
);
8264 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8266 struct cfs_schedulable_data
*d
= data
;
8267 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8268 s64 quota
= 0, parent_quota
= -1;
8271 quota
= RUNTIME_INF
;
8273 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8275 quota
= normalize_cfs_quota(tg
, d
);
8276 parent_quota
= parent_b
->hierarchical_quota
;
8279 * ensure max(child_quota) <= parent_quota, inherit when no
8282 if (quota
== RUNTIME_INF
)
8283 quota
= parent_quota
;
8284 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8287 cfs_b
->hierarchical_quota
= quota
;
8292 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8295 struct cfs_schedulable_data data
= {
8301 if (quota
!= RUNTIME_INF
) {
8302 do_div(data
.period
, NSEC_PER_USEC
);
8303 do_div(data
.quota
, NSEC_PER_USEC
);
8307 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8313 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8315 struct task_group
*tg
= css_tg(seq_css(sf
));
8316 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8318 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8319 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8320 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8324 #endif /* CONFIG_CFS_BANDWIDTH */
8325 #endif /* CONFIG_FAIR_GROUP_SCHED */
8327 #ifdef CONFIG_RT_GROUP_SCHED
8328 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8329 struct cftype
*cft
, s64 val
)
8331 return sched_group_set_rt_runtime(css_tg(css
), val
);
8334 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8337 return sched_group_rt_runtime(css_tg(css
));
8340 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8341 struct cftype
*cftype
, u64 rt_period_us
)
8343 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8346 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8349 return sched_group_rt_period(css_tg(css
));
8351 #endif /* CONFIG_RT_GROUP_SCHED */
8353 static struct cftype cpu_files
[] = {
8354 #ifdef CONFIG_FAIR_GROUP_SCHED
8357 .read_u64
= cpu_shares_read_u64
,
8358 .write_u64
= cpu_shares_write_u64
,
8361 #ifdef CONFIG_CFS_BANDWIDTH
8363 .name
= "cfs_quota_us",
8364 .read_s64
= cpu_cfs_quota_read_s64
,
8365 .write_s64
= cpu_cfs_quota_write_s64
,
8368 .name
= "cfs_period_us",
8369 .read_u64
= cpu_cfs_period_read_u64
,
8370 .write_u64
= cpu_cfs_period_write_u64
,
8374 .seq_show
= cpu_stats_show
,
8377 #ifdef CONFIG_RT_GROUP_SCHED
8379 .name
= "rt_runtime_us",
8380 .read_s64
= cpu_rt_runtime_read
,
8381 .write_s64
= cpu_rt_runtime_write
,
8384 .name
= "rt_period_us",
8385 .read_u64
= cpu_rt_period_read_uint
,
8386 .write_u64
= cpu_rt_period_write_uint
,
8392 struct cgroup_subsys cpu_cgrp_subsys
= {
8393 .css_alloc
= cpu_cgroup_css_alloc
,
8394 .css_free
= cpu_cgroup_css_free
,
8395 .css_online
= cpu_cgroup_css_online
,
8396 .css_offline
= cpu_cgroup_css_offline
,
8397 .fork
= cpu_cgroup_fork
,
8398 .can_attach
= cpu_cgroup_can_attach
,
8399 .attach
= cpu_cgroup_attach
,
8400 .exit
= cpu_cgroup_exit
,
8401 .legacy_cftypes
= cpu_files
,
8405 #endif /* CONFIG_CGROUP_SCHED */
8407 void dump_cpu_task(int cpu
)
8409 pr_info("Task dump for CPU %d:\n", cpu
);
8410 sched_show_task(cpu_curr(cpu
));