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/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
74 #include <linux/init_task.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
79 #ifdef CONFIG_PARAVIRT
80 #include <asm/paravirt.h>
83 #include "sched_cpupri.h"
84 #include "workqueue_sched.h"
85 #include "sched_autogroup.h"
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/sched.h>
91 * Convert user-nice values [ -20 ... 0 ... 19 ]
92 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
95 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
96 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
97 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
100 * 'User priority' is the nice value converted to something we
101 * can work with better when scaling various scheduler parameters,
102 * it's a [ 0 ... 39 ] range.
104 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
105 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
106 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
109 * Helpers for converting nanosecond timing to jiffy resolution
111 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
113 #define NICE_0_LOAD SCHED_LOAD_SCALE
114 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
117 * These are the 'tuning knobs' of the scheduler:
119 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
120 * Timeslices get refilled after they expire.
122 #define DEF_TIMESLICE (100 * HZ / 1000)
125 * single value that denotes runtime == period, ie unlimited time.
127 #define RUNTIME_INF ((u64)~0ULL)
129 static inline int rt_policy(int policy)
131 if (policy == SCHED_FIFO || policy == SCHED_RR)
136 static inline int task_has_rt_policy(struct task_struct *p)
138 return rt_policy(p->policy);
142 * This is the priority-queue data structure of the RT scheduling class:
144 struct rt_prio_array {
145 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
146 struct list_head queue[MAX_RT_PRIO];
149 struct rt_bandwidth {
150 /* nests inside the rq lock: */
151 raw_spinlock_t rt_runtime_lock;
154 struct hrtimer rt_period_timer;
157 static struct rt_bandwidth def_rt_bandwidth;
159 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
161 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
163 struct rt_bandwidth *rt_b =
164 container_of(timer, struct rt_bandwidth, rt_period_timer);
170 now = hrtimer_cb_get_time(timer);
171 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
176 idle = do_sched_rt_period_timer(rt_b, overrun);
179 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
183 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
185 rt_b->rt_period = ns_to_ktime(period);
186 rt_b->rt_runtime = runtime;
188 raw_spin_lock_init(&rt_b->rt_runtime_lock);
190 hrtimer_init(&rt_b->rt_period_timer,
191 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
192 rt_b->rt_period_timer.function = sched_rt_period_timer;
195 static inline int rt_bandwidth_enabled(void)
197 return sysctl_sched_rt_runtime >= 0;
200 static void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
203 ktime_t soft, hard, now;
206 if (hrtimer_active(period_timer))
209 now = hrtimer_cb_get_time(period_timer);
210 hrtimer_forward(period_timer, now, period);
212 soft = hrtimer_get_softexpires(period_timer);
213 hard = hrtimer_get_expires(period_timer);
214 delta = ktime_to_ns(ktime_sub(hard, soft));
215 __hrtimer_start_range_ns(period_timer, soft, delta,
216 HRTIMER_MODE_ABS_PINNED, 0);
220 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 raw_spin_lock(&rt_b->rt_runtime_lock);
229 start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
230 raw_spin_unlock(&rt_b->rt_runtime_lock);
233 #ifdef CONFIG_RT_GROUP_SCHED
234 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
236 hrtimer_cancel(&rt_b->rt_period_timer);
241 * sched_domains_mutex serializes calls to init_sched_domains,
242 * detach_destroy_domains and partition_sched_domains.
244 static DEFINE_MUTEX(sched_domains_mutex);
246 #ifdef CONFIG_CGROUP_SCHED
248 #include <linux/cgroup.h>
252 static LIST_HEAD(task_groups);
254 struct cfs_bandwidth {
255 #ifdef CONFIG_CFS_BANDWIDTH
259 s64 hierarchal_quota;
262 int idle, timer_active;
263 struct hrtimer period_timer, slack_timer;
264 struct list_head throttled_cfs_rq;
267 int nr_periods, nr_throttled;
272 /* task group related information */
274 struct cgroup_subsys_state css;
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity **se;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq **cfs_rq;
281 unsigned long shares;
283 atomic_t load_weight;
286 #ifdef CONFIG_RT_GROUP_SCHED
287 struct sched_rt_entity **rt_se;
288 struct rt_rq **rt_rq;
290 struct rt_bandwidth rt_bandwidth;
294 struct list_head list;
296 struct task_group *parent;
297 struct list_head siblings;
298 struct list_head children;
300 #ifdef CONFIG_SCHED_AUTOGROUP
301 struct autogroup *autogroup;
304 struct cfs_bandwidth cfs_bandwidth;
307 /* task_group_lock serializes the addition/removal of task groups */
308 static DEFINE_SPINLOCK(task_group_lock);
310 #ifdef CONFIG_FAIR_GROUP_SCHED
312 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
315 * A weight of 0 or 1 can cause arithmetics problems.
316 * A weight of a cfs_rq is the sum of weights of which entities
317 * are queued on this cfs_rq, so a weight of a entity should not be
318 * too large, so as the shares value of a task group.
319 * (The default weight is 1024 - so there's no practical
320 * limitation from this.)
322 #define MIN_SHARES (1UL << 1)
323 #define MAX_SHARES (1UL << 18)
325 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
328 /* Default task group.
329 * Every task in system belong to this group at bootup.
331 struct task_group root_task_group;
333 #endif /* CONFIG_CGROUP_SCHED */
335 /* CFS-related fields in a runqueue */
337 struct load_weight load;
338 unsigned long nr_running, h_nr_running;
343 u64 min_vruntime_copy;
346 struct rb_root tasks_timeline;
347 struct rb_node *rb_leftmost;
349 struct list_head tasks;
350 struct list_head *balance_iterator;
353 * 'curr' points to currently running entity on this cfs_rq.
354 * It is set to NULL otherwise (i.e when none are currently running).
356 struct sched_entity *curr, *next, *last, *skip;
358 #ifdef CONFIG_SCHED_DEBUG
359 unsigned int nr_spread_over;
362 #ifdef CONFIG_FAIR_GROUP_SCHED
363 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
366 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
367 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
368 * (like users, containers etc.)
370 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
371 * list is used during load balance.
374 struct list_head leaf_cfs_rq_list;
375 struct task_group *tg; /* group that "owns" this runqueue */
379 * the part of load.weight contributed by tasks
381 unsigned long task_weight;
384 * h_load = weight * f(tg)
386 * Where f(tg) is the recursive weight fraction assigned to
389 unsigned long h_load;
392 * Maintaining per-cpu shares distribution for group scheduling
394 * load_stamp is the last time we updated the load average
395 * load_last is the last time we updated the load average and saw load
396 * load_unacc_exec_time is currently unaccounted execution time
400 u64 load_stamp, load_last, load_unacc_exec_time;
402 unsigned long load_contribution;
404 #ifdef CONFIG_CFS_BANDWIDTH
407 s64 runtime_remaining;
409 u64 throttled_timestamp;
410 int throttled, throttle_count;
411 struct list_head throttled_list;
416 #ifdef CONFIG_FAIR_GROUP_SCHED
417 #ifdef CONFIG_CFS_BANDWIDTH
418 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
420 return &tg->cfs_bandwidth;
423 static inline u64 default_cfs_period(void);
424 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
425 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
427 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
429 struct cfs_bandwidth *cfs_b =
430 container_of(timer, struct cfs_bandwidth, slack_timer);
431 do_sched_cfs_slack_timer(cfs_b);
433 return HRTIMER_NORESTART;
436 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
438 struct cfs_bandwidth *cfs_b =
439 container_of(timer, struct cfs_bandwidth, period_timer);
445 now = hrtimer_cb_get_time(timer);
446 overrun = hrtimer_forward(timer, now, cfs_b->period);
451 idle = do_sched_cfs_period_timer(cfs_b, overrun);
454 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
457 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
459 raw_spin_lock_init(&cfs_b->lock);
461 cfs_b->quota = RUNTIME_INF;
462 cfs_b->period = ns_to_ktime(default_cfs_period());
464 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
465 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
466 cfs_b->period_timer.function = sched_cfs_period_timer;
467 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
468 cfs_b->slack_timer.function = sched_cfs_slack_timer;
471 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
473 cfs_rq->runtime_enabled = 0;
474 INIT_LIST_HEAD(&cfs_rq->throttled_list);
477 /* requires cfs_b->lock, may release to reprogram timer */
478 static void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
481 * The timer may be active because we're trying to set a new bandwidth
482 * period or because we're racing with the tear-down path
483 * (timer_active==0 becomes visible before the hrtimer call-back
484 * terminates). In either case we ensure that it's re-programmed
486 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
487 raw_spin_unlock(&cfs_b->lock);
488 /* ensure cfs_b->lock is available while we wait */
489 hrtimer_cancel(&cfs_b->period_timer);
491 raw_spin_lock(&cfs_b->lock);
492 /* if someone else restarted the timer then we're done */
493 if (cfs_b->timer_active)
497 cfs_b->timer_active = 1;
498 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
501 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
503 hrtimer_cancel(&cfs_b->period_timer);
504 hrtimer_cancel(&cfs_b->slack_timer);
507 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
508 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
509 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
511 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
515 #endif /* CONFIG_CFS_BANDWIDTH */
516 #endif /* CONFIG_FAIR_GROUP_SCHED */
518 /* Real-Time classes' related field in a runqueue: */
520 struct rt_prio_array active;
521 unsigned long rt_nr_running;
522 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
524 int curr; /* highest queued rt task prio */
526 int next; /* next highest */
531 unsigned long rt_nr_migratory;
532 unsigned long rt_nr_total;
534 struct plist_head pushable_tasks;
539 /* Nests inside the rq lock: */
540 raw_spinlock_t rt_runtime_lock;
542 #ifdef CONFIG_RT_GROUP_SCHED
543 unsigned long rt_nr_boosted;
546 struct list_head leaf_rt_rq_list;
547 struct task_group *tg;
554 * We add the notion of a root-domain which will be used to define per-domain
555 * variables. Each exclusive cpuset essentially defines an island domain by
556 * fully partitioning the member cpus from any other cpuset. Whenever a new
557 * exclusive cpuset is created, we also create and attach a new root-domain
566 cpumask_var_t online;
569 * The "RT overload" flag: it gets set if a CPU has more than
570 * one runnable RT task.
572 cpumask_var_t rto_mask;
573 struct cpupri cpupri;
577 * By default the system creates a single root-domain with all cpus as
578 * members (mimicking the global state we have today).
580 static struct root_domain def_root_domain;
582 #endif /* CONFIG_SMP */
585 * This is the main, per-CPU runqueue data structure.
587 * Locking rule: those places that want to lock multiple runqueues
588 * (such as the load balancing or the thread migration code), lock
589 * acquire operations must be ordered by ascending &runqueue.
596 * nr_running and cpu_load should be in the same cacheline because
597 * remote CPUs use both these fields when doing load calculation.
599 unsigned long nr_running;
600 #define CPU_LOAD_IDX_MAX 5
601 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
602 unsigned long last_load_update_tick;
605 unsigned char nohz_balance_kick;
607 int skip_clock_update;
609 /* capture load from *all* tasks on this cpu: */
610 struct load_weight load;
611 unsigned long nr_load_updates;
617 #ifdef CONFIG_FAIR_GROUP_SCHED
618 /* list of leaf cfs_rq on this cpu: */
619 struct list_head leaf_cfs_rq_list;
621 #ifdef CONFIG_RT_GROUP_SCHED
622 struct list_head leaf_rt_rq_list;
626 * This is part of a global counter where only the total sum
627 * over all CPUs matters. A task can increase this counter on
628 * one CPU and if it got migrated afterwards it may decrease
629 * it on another CPU. Always updated under the runqueue lock:
631 unsigned long nr_uninterruptible;
633 struct task_struct *curr, *idle, *stop;
634 unsigned long next_balance;
635 struct mm_struct *prev_mm;
643 struct root_domain *rd;
644 struct sched_domain *sd;
646 unsigned long cpu_power;
648 unsigned char idle_balance;
649 /* For active balancing */
653 struct cpu_stop_work active_balance_work;
654 /* cpu of this runqueue: */
664 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
667 #ifdef CONFIG_PARAVIRT
670 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
671 u64 prev_steal_time_rq;
674 /* calc_load related fields */
675 unsigned long calc_load_update;
676 long calc_load_active;
678 #ifdef CONFIG_SCHED_HRTICK
680 int hrtick_csd_pending;
681 struct call_single_data hrtick_csd;
683 struct hrtimer hrtick_timer;
686 #ifdef CONFIG_SCHEDSTATS
688 struct sched_info rq_sched_info;
689 unsigned long long rq_cpu_time;
690 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
692 /* sys_sched_yield() stats */
693 unsigned int yld_count;
695 /* schedule() stats */
696 unsigned int sched_switch;
697 unsigned int sched_count;
698 unsigned int sched_goidle;
700 /* try_to_wake_up() stats */
701 unsigned int ttwu_count;
702 unsigned int ttwu_local;
706 struct llist_head wake_list;
710 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
713 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
715 static inline int cpu_of(struct rq *rq)
724 #define rcu_dereference_check_sched_domain(p) \
725 rcu_dereference_check((p), \
726 lockdep_is_held(&sched_domains_mutex))
729 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
730 * See detach_destroy_domains: synchronize_sched for details.
732 * The domain tree of any CPU may only be accessed from within
733 * preempt-disabled sections.
735 #define for_each_domain(cpu, __sd) \
736 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
738 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
739 #define this_rq() (&__get_cpu_var(runqueues))
740 #define task_rq(p) cpu_rq(task_cpu(p))
741 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
742 #define raw_rq() (&__raw_get_cpu_var(runqueues))
744 #ifdef CONFIG_CGROUP_SCHED
747 * Return the group to which this tasks belongs.
749 * We cannot use task_subsys_state() and friends because the cgroup
750 * subsystem changes that value before the cgroup_subsys::attach() method
751 * is called, therefore we cannot pin it and might observe the wrong value.
753 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
754 * core changes this before calling sched_move_task().
756 * Instead we use a 'copy' which is updated from sched_move_task() while
757 * holding both task_struct::pi_lock and rq::lock.
759 static inline struct task_group *task_group(struct task_struct *p)
761 return p->sched_task_group;
764 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
765 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
767 #ifdef CONFIG_FAIR_GROUP_SCHED
768 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
769 p->se.parent = task_group(p)->se[cpu];
772 #ifdef CONFIG_RT_GROUP_SCHED
773 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
774 p->rt.parent = task_group(p)->rt_se[cpu];
778 #else /* CONFIG_CGROUP_SCHED */
780 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
781 static inline struct task_group *task_group(struct task_struct *p)
786 #endif /* CONFIG_CGROUP_SCHED */
788 static void update_rq_clock_task(struct rq *rq, s64 delta);
790 static void update_rq_clock(struct rq *rq)
794 if (rq->skip_clock_update > 0)
797 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
799 update_rq_clock_task(rq, delta);
803 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
805 #ifdef CONFIG_SCHED_DEBUG
806 # define const_debug __read_mostly
808 # define const_debug static const
812 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
813 * @cpu: the processor in question.
815 * This interface allows printk to be called with the runqueue lock
816 * held and know whether or not it is OK to wake up the klogd.
818 int runqueue_is_locked(int cpu)
820 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
824 * Debugging: various feature bits
827 #define SCHED_FEAT(name, enabled) \
828 __SCHED_FEAT_##name ,
831 #include "sched_features.h"
836 #define SCHED_FEAT(name, enabled) \
837 (1UL << __SCHED_FEAT_##name) * enabled |
839 const_debug unsigned int sysctl_sched_features =
840 #include "sched_features.h"
845 #ifdef CONFIG_SCHED_DEBUG
846 #define SCHED_FEAT(name, enabled) \
849 static __read_mostly char *sched_feat_names[] = {
850 #include "sched_features.h"
856 static int sched_feat_show(struct seq_file *m, void *v)
860 for (i = 0; sched_feat_names[i]; i++) {
861 if (!(sysctl_sched_features & (1UL << i)))
863 seq_printf(m, "%s ", sched_feat_names[i]);
871 sched_feat_write(struct file *filp, const char __user *ubuf,
872 size_t cnt, loff_t *ppos)
882 if (copy_from_user(&buf, ubuf, cnt))
888 if (strncmp(cmp, "NO_", 3) == 0) {
893 for (i = 0; sched_feat_names[i]; i++) {
894 if (strcmp(cmp, sched_feat_names[i]) == 0) {
896 sysctl_sched_features &= ~(1UL << i);
898 sysctl_sched_features |= (1UL << i);
903 if (!sched_feat_names[i])
911 static int sched_feat_open(struct inode *inode, struct file *filp)
913 return single_open(filp, sched_feat_show, NULL);
916 static const struct file_operations sched_feat_fops = {
917 .open = sched_feat_open,
918 .write = sched_feat_write,
921 .release = single_release,
924 static __init int sched_init_debug(void)
926 debugfs_create_file("sched_features", 0644, NULL, NULL,
931 late_initcall(sched_init_debug);
935 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
938 * Number of tasks to iterate in a single balance run.
939 * Limited because this is done with IRQs disabled.
941 const_debug unsigned int sysctl_sched_nr_migrate = 32;
944 * period over which we average the RT time consumption, measured
949 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
952 * period over which we measure -rt task cpu usage in us.
955 unsigned int sysctl_sched_rt_period = 1000000;
957 static __read_mostly int scheduler_running;
960 * part of the period that we allow rt tasks to run in us.
963 int sysctl_sched_rt_runtime = 950000;
965 static inline u64 global_rt_period(void)
967 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
970 static inline u64 global_rt_runtime(void)
972 if (sysctl_sched_rt_runtime < 0)
975 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
978 #ifndef prepare_arch_switch
979 # define prepare_arch_switch(next) do { } while (0)
981 #ifndef finish_arch_switch
982 # define finish_arch_switch(prev) do { } while (0)
985 static inline int task_current(struct rq *rq, struct task_struct *p)
987 return rq->curr == p;
990 static inline int task_running(struct rq *rq, struct task_struct *p)
995 return task_current(rq, p);
999 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1000 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1004 * We can optimise this out completely for !SMP, because the
1005 * SMP rebalancing from interrupt is the only thing that cares
1012 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1016 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1017 * We must ensure this doesn't happen until the switch is completely
1023 #ifdef CONFIG_DEBUG_SPINLOCK
1024 /* this is a valid case when another task releases the spinlock */
1025 rq->lock.owner = current;
1028 * If we are tracking spinlock dependencies then we have to
1029 * fix up the runqueue lock - which gets 'carried over' from
1030 * prev into current:
1032 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1034 raw_spin_unlock_irq(&rq->lock);
1037 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1038 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1042 * We can optimise this out completely for !SMP, because the
1043 * SMP rebalancing from interrupt is the only thing that cares
1048 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1049 raw_spin_unlock_irq(&rq->lock);
1051 raw_spin_unlock(&rq->lock);
1055 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1059 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1060 * We must ensure this doesn't happen until the switch is completely
1066 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1070 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1073 * __task_rq_lock - lock the rq @p resides on.
1075 static inline struct rq *__task_rq_lock(struct task_struct *p)
1076 __acquires(rq->lock)
1080 lockdep_assert_held(&p->pi_lock);
1084 raw_spin_lock(&rq->lock);
1085 if (likely(rq == task_rq(p)))
1087 raw_spin_unlock(&rq->lock);
1092 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1094 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1095 __acquires(p->pi_lock)
1096 __acquires(rq->lock)
1101 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1103 raw_spin_lock(&rq->lock);
1104 if (likely(rq == task_rq(p)))
1106 raw_spin_unlock(&rq->lock);
1107 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1111 static void __task_rq_unlock(struct rq *rq)
1112 __releases(rq->lock)
1114 raw_spin_unlock(&rq->lock);
1118 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1119 __releases(rq->lock)
1120 __releases(p->pi_lock)
1122 raw_spin_unlock(&rq->lock);
1123 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1127 * this_rq_lock - lock this runqueue and disable interrupts.
1129 static struct rq *this_rq_lock(void)
1130 __acquires(rq->lock)
1134 local_irq_disable();
1136 raw_spin_lock(&rq->lock);
1141 #ifdef CONFIG_SCHED_HRTICK
1143 * Use HR-timers to deliver accurate preemption points.
1145 * Its all a bit involved since we cannot program an hrt while holding the
1146 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1149 * When we get rescheduled we reprogram the hrtick_timer outside of the
1155 * - enabled by features
1156 * - hrtimer is actually high res
1158 static inline int hrtick_enabled(struct rq *rq)
1160 if (!sched_feat(HRTICK))
1162 if (!cpu_active(cpu_of(rq)))
1164 return hrtimer_is_hres_active(&rq->hrtick_timer);
1167 static void hrtick_clear(struct rq *rq)
1169 if (hrtimer_active(&rq->hrtick_timer))
1170 hrtimer_cancel(&rq->hrtick_timer);
1174 * High-resolution timer tick.
1175 * Runs from hardirq context with interrupts disabled.
1177 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1179 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1181 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1183 raw_spin_lock(&rq->lock);
1184 update_rq_clock(rq);
1185 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1186 raw_spin_unlock(&rq->lock);
1188 return HRTIMER_NORESTART;
1193 * called from hardirq (IPI) context
1195 static void __hrtick_start(void *arg)
1197 struct rq *rq = arg;
1199 raw_spin_lock(&rq->lock);
1200 hrtimer_restart(&rq->hrtick_timer);
1201 rq->hrtick_csd_pending = 0;
1202 raw_spin_unlock(&rq->lock);
1206 * Called to set the hrtick timer state.
1208 * called with rq->lock held and irqs disabled
1210 static void hrtick_start(struct rq *rq, u64 delay)
1212 struct hrtimer *timer = &rq->hrtick_timer;
1213 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1215 hrtimer_set_expires(timer, time);
1217 if (rq == this_rq()) {
1218 hrtimer_restart(timer);
1219 } else if (!rq->hrtick_csd_pending) {
1220 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1221 rq->hrtick_csd_pending = 1;
1226 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1228 int cpu = (int)(long)hcpu;
1231 case CPU_UP_CANCELED:
1232 case CPU_UP_CANCELED_FROZEN:
1233 case CPU_DOWN_PREPARE:
1234 case CPU_DOWN_PREPARE_FROZEN:
1236 case CPU_DEAD_FROZEN:
1237 hrtick_clear(cpu_rq(cpu));
1244 static __init void init_hrtick(void)
1246 hotcpu_notifier(hotplug_hrtick, 0);
1250 * Called to set the hrtick timer state.
1252 * called with rq->lock held and irqs disabled
1254 static void hrtick_start(struct rq *rq, u64 delay)
1256 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1257 HRTIMER_MODE_REL_PINNED, 0);
1260 static inline void init_hrtick(void)
1263 #endif /* CONFIG_SMP */
1265 static void init_rq_hrtick(struct rq *rq)
1268 rq->hrtick_csd_pending = 0;
1270 rq->hrtick_csd.flags = 0;
1271 rq->hrtick_csd.func = __hrtick_start;
1272 rq->hrtick_csd.info = rq;
1275 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1276 rq->hrtick_timer.function = hrtick;
1278 #else /* CONFIG_SCHED_HRTICK */
1279 static inline void hrtick_clear(struct rq *rq)
1283 static inline void init_rq_hrtick(struct rq *rq)
1287 static inline void init_hrtick(void)
1290 #endif /* CONFIG_SCHED_HRTICK */
1293 * resched_task - mark a task 'to be rescheduled now'.
1295 * On UP this means the setting of the need_resched flag, on SMP it
1296 * might also involve a cross-CPU call to trigger the scheduler on
1301 #ifndef tsk_is_polling
1302 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1305 static void resched_task(struct task_struct *p)
1309 assert_raw_spin_locked(&task_rq(p)->lock);
1311 if (test_tsk_need_resched(p))
1314 set_tsk_need_resched(p);
1317 if (cpu == smp_processor_id())
1320 /* NEED_RESCHED must be visible before we test polling */
1322 if (!tsk_is_polling(p))
1323 smp_send_reschedule(cpu);
1326 static void resched_cpu(int cpu)
1328 struct rq *rq = cpu_rq(cpu);
1329 unsigned long flags;
1331 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1333 resched_task(cpu_curr(cpu));
1334 raw_spin_unlock_irqrestore(&rq->lock, flags);
1339 * In the semi idle case, use the nearest busy cpu for migrating timers
1340 * from an idle cpu. This is good for power-savings.
1342 * We don't do similar optimization for completely idle system, as
1343 * selecting an idle cpu will add more delays to the timers than intended
1344 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1346 int get_nohz_timer_target(void)
1348 int cpu = smp_processor_id();
1350 struct sched_domain *sd;
1353 for_each_domain(cpu, sd) {
1354 for_each_cpu(i, sched_domain_span(sd)) {
1366 * When add_timer_on() enqueues a timer into the timer wheel of an
1367 * idle CPU then this timer might expire before the next timer event
1368 * which is scheduled to wake up that CPU. In case of a completely
1369 * idle system the next event might even be infinite time into the
1370 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1371 * leaves the inner idle loop so the newly added timer is taken into
1372 * account when the CPU goes back to idle and evaluates the timer
1373 * wheel for the next timer event.
1375 void wake_up_idle_cpu(int cpu)
1377 struct rq *rq = cpu_rq(cpu);
1379 if (cpu == smp_processor_id())
1383 * This is safe, as this function is called with the timer
1384 * wheel base lock of (cpu) held. When the CPU is on the way
1385 * to idle and has not yet set rq->curr to idle then it will
1386 * be serialized on the timer wheel base lock and take the new
1387 * timer into account automatically.
1389 if (rq->curr != rq->idle)
1393 * We can set TIF_RESCHED on the idle task of the other CPU
1394 * lockless. The worst case is that the other CPU runs the
1395 * idle task through an additional NOOP schedule()
1397 set_tsk_need_resched(rq->idle);
1399 /* NEED_RESCHED must be visible before we test polling */
1401 if (!tsk_is_polling(rq->idle))
1402 smp_send_reschedule(cpu);
1405 static inline bool got_nohz_idle_kick(void)
1407 return idle_cpu(smp_processor_id()) && this_rq()->nohz_balance_kick;
1410 #else /* CONFIG_NO_HZ */
1412 static inline bool got_nohz_idle_kick(void)
1417 #endif /* CONFIG_NO_HZ */
1419 static u64 sched_avg_period(void)
1421 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1424 static void sched_avg_update(struct rq *rq)
1426 s64 period = sched_avg_period();
1428 while ((s64)(rq->clock - rq->age_stamp) > period) {
1430 * Inline assembly required to prevent the compiler
1431 * optimising this loop into a divmod call.
1432 * See __iter_div_u64_rem() for another example of this.
1434 asm("" : "+rm" (rq->age_stamp));
1435 rq->age_stamp += period;
1440 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1442 rq->rt_avg += rt_delta;
1443 sched_avg_update(rq);
1446 #else /* !CONFIG_SMP */
1447 static void resched_task(struct task_struct *p)
1449 assert_raw_spin_locked(&task_rq(p)->lock);
1450 set_tsk_need_resched(p);
1453 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1457 static void sched_avg_update(struct rq *rq)
1460 #endif /* CONFIG_SMP */
1462 #if BITS_PER_LONG == 32
1463 # define WMULT_CONST (~0UL)
1465 # define WMULT_CONST (1UL << 32)
1468 #define WMULT_SHIFT 32
1471 * Shift right and round:
1473 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1476 * delta *= weight / lw
1478 static unsigned long
1479 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1480 struct load_weight *lw)
1485 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1486 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1487 * 2^SCHED_LOAD_RESOLUTION.
1489 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1490 tmp = (u64)delta_exec * scale_load_down(weight);
1492 tmp = (u64)delta_exec;
1494 if (!lw->inv_weight) {
1495 unsigned long w = scale_load_down(lw->weight);
1497 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1499 else if (unlikely(!w))
1500 lw->inv_weight = WMULT_CONST;
1502 lw->inv_weight = WMULT_CONST / w;
1506 * Check whether we'd overflow the 64-bit multiplication:
1508 if (unlikely(tmp > WMULT_CONST))
1509 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1512 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1514 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1517 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1523 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1529 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1536 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1537 * of tasks with abnormal "nice" values across CPUs the contribution that
1538 * each task makes to its run queue's load is weighted according to its
1539 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1540 * scaled version of the new time slice allocation that they receive on time
1544 #define WEIGHT_IDLEPRIO 3
1545 #define WMULT_IDLEPRIO 1431655765
1548 * Nice levels are multiplicative, with a gentle 10% change for every
1549 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1550 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1551 * that remained on nice 0.
1553 * The "10% effect" is relative and cumulative: from _any_ nice level,
1554 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1555 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1556 * If a task goes up by ~10% and another task goes down by ~10% then
1557 * the relative distance between them is ~25%.)
1559 static const int prio_to_weight[40] = {
1560 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1561 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1562 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1563 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1564 /* 0 */ 1024, 820, 655, 526, 423,
1565 /* 5 */ 335, 272, 215, 172, 137,
1566 /* 10 */ 110, 87, 70, 56, 45,
1567 /* 15 */ 36, 29, 23, 18, 15,
1571 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1573 * In cases where the weight does not change often, we can use the
1574 * precalculated inverse to speed up arithmetics by turning divisions
1575 * into multiplications:
1577 static const u32 prio_to_wmult[40] = {
1578 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1579 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1580 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1581 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1582 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1583 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1584 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1585 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1588 /* Time spent by the tasks of the cpu accounting group executing in ... */
1589 enum cpuacct_stat_index {
1590 CPUACCT_STAT_USER, /* ... user mode */
1591 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1593 CPUACCT_STAT_NSTATS,
1596 #ifdef CONFIG_CGROUP_CPUACCT
1597 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1598 static void cpuacct_update_stats(struct task_struct *tsk,
1599 enum cpuacct_stat_index idx, cputime_t val);
1601 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1602 static inline void cpuacct_update_stats(struct task_struct *tsk,
1603 enum cpuacct_stat_index idx, cputime_t val) {}
1606 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1608 update_load_add(&rq->load, load);
1611 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1613 update_load_sub(&rq->load, load);
1616 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1617 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1618 typedef int (*tg_visitor)(struct task_group *, void *);
1621 * Iterate task_group tree rooted at *from, calling @down when first entering a
1622 * node and @up when leaving it for the final time.
1624 * Caller must hold rcu_lock or sufficient equivalent.
1626 static int walk_tg_tree_from(struct task_group *from,
1627 tg_visitor down, tg_visitor up, void *data)
1629 struct task_group *parent, *child;
1635 ret = (*down)(parent, data);
1638 list_for_each_entry_rcu(child, &parent->children, siblings) {
1645 ret = (*up)(parent, data);
1646 if (ret || parent == from)
1650 parent = parent->parent;
1658 * Iterate the full tree, calling @down when first entering a node and @up when
1659 * leaving it for the final time.
1661 * Caller must hold rcu_lock or sufficient equivalent.
1664 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1666 return walk_tg_tree_from(&root_task_group, down, up, data);
1669 static int tg_nop(struct task_group *tg, void *data)
1676 /* Used instead of source_load when we know the type == 0 */
1677 static unsigned long weighted_cpuload(const int cpu)
1679 return cpu_rq(cpu)->load.weight;
1683 * Return a low guess at the load of a migration-source cpu weighted
1684 * according to the scheduling class and "nice" value.
1686 * We want to under-estimate the load of migration sources, to
1687 * balance conservatively.
1689 static unsigned long source_load(int cpu, int type)
1691 struct rq *rq = cpu_rq(cpu);
1692 unsigned long total = weighted_cpuload(cpu);
1694 if (type == 0 || !sched_feat(LB_BIAS))
1697 return min(rq->cpu_load[type-1], total);
1701 * Return a high guess at the load of a migration-target cpu weighted
1702 * according to the scheduling class and "nice" value.
1704 static unsigned long target_load(int cpu, int type)
1706 struct rq *rq = cpu_rq(cpu);
1707 unsigned long total = weighted_cpuload(cpu);
1709 if (type == 0 || !sched_feat(LB_BIAS))
1712 return max(rq->cpu_load[type-1], total);
1715 static unsigned long power_of(int cpu)
1717 return cpu_rq(cpu)->cpu_power;
1720 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1722 static unsigned long cpu_avg_load_per_task(int cpu)
1724 struct rq *rq = cpu_rq(cpu);
1725 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1728 return rq->load.weight / nr_running;
1733 #ifdef CONFIG_PREEMPT
1735 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1738 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1739 * way at the expense of forcing extra atomic operations in all
1740 * invocations. This assures that the double_lock is acquired using the
1741 * same underlying policy as the spinlock_t on this architecture, which
1742 * reduces latency compared to the unfair variant below. However, it
1743 * also adds more overhead and therefore may reduce throughput.
1745 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1746 __releases(this_rq->lock)
1747 __acquires(busiest->lock)
1748 __acquires(this_rq->lock)
1750 raw_spin_unlock(&this_rq->lock);
1751 double_rq_lock(this_rq, busiest);
1758 * Unfair double_lock_balance: Optimizes throughput at the expense of
1759 * latency by eliminating extra atomic operations when the locks are
1760 * already in proper order on entry. This favors lower cpu-ids and will
1761 * grant the double lock to lower cpus over higher ids under contention,
1762 * regardless of entry order into the function.
1764 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1765 __releases(this_rq->lock)
1766 __acquires(busiest->lock)
1767 __acquires(this_rq->lock)
1771 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1772 if (busiest < this_rq) {
1773 raw_spin_unlock(&this_rq->lock);
1774 raw_spin_lock(&busiest->lock);
1775 raw_spin_lock_nested(&this_rq->lock,
1776 SINGLE_DEPTH_NESTING);
1779 raw_spin_lock_nested(&busiest->lock,
1780 SINGLE_DEPTH_NESTING);
1785 #endif /* CONFIG_PREEMPT */
1788 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1790 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1792 if (unlikely(!irqs_disabled())) {
1793 /* printk() doesn't work good under rq->lock */
1794 raw_spin_unlock(&this_rq->lock);
1798 return _double_lock_balance(this_rq, busiest);
1801 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1802 __releases(busiest->lock)
1804 raw_spin_unlock(&busiest->lock);
1805 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1809 * double_rq_lock - safely lock two runqueues
1811 * Note this does not disable interrupts like task_rq_lock,
1812 * you need to do so manually before calling.
1814 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1815 __acquires(rq1->lock)
1816 __acquires(rq2->lock)
1818 BUG_ON(!irqs_disabled());
1820 raw_spin_lock(&rq1->lock);
1821 __acquire(rq2->lock); /* Fake it out ;) */
1824 raw_spin_lock(&rq1->lock);
1825 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1827 raw_spin_lock(&rq2->lock);
1828 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1834 * double_rq_unlock - safely unlock two runqueues
1836 * Note this does not restore interrupts like task_rq_unlock,
1837 * you need to do so manually after calling.
1839 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1840 __releases(rq1->lock)
1841 __releases(rq2->lock)
1843 raw_spin_unlock(&rq1->lock);
1845 raw_spin_unlock(&rq2->lock);
1847 __release(rq2->lock);
1850 #else /* CONFIG_SMP */
1853 * double_rq_lock - safely lock two runqueues
1855 * Note this does not disable interrupts like task_rq_lock,
1856 * you need to do so manually before calling.
1858 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1859 __acquires(rq1->lock)
1860 __acquires(rq2->lock)
1862 BUG_ON(!irqs_disabled());
1864 raw_spin_lock(&rq1->lock);
1865 __acquire(rq2->lock); /* Fake it out ;) */
1869 * double_rq_unlock - safely unlock two runqueues
1871 * Note this does not restore interrupts like task_rq_unlock,
1872 * you need to do so manually after calling.
1874 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1875 __releases(rq1->lock)
1876 __releases(rq2->lock)
1879 raw_spin_unlock(&rq1->lock);
1880 __release(rq2->lock);
1885 static void update_sysctl(void);
1886 static int get_update_sysctl_factor(void);
1887 static void update_idle_cpu_load(struct rq *this_rq);
1889 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1891 set_task_rq(p, cpu);
1894 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1895 * successfully executed on another CPU. We must ensure that updates of
1896 * per-task data have been completed by this moment.
1899 task_thread_info(p)->cpu = cpu;
1903 static const struct sched_class rt_sched_class;
1905 #define sched_class_highest (&stop_sched_class)
1906 #define for_each_class(class) \
1907 for (class = sched_class_highest; class; class = class->next)
1909 #include "sched_stats.h"
1911 static void inc_nr_running(struct rq *rq)
1916 static void dec_nr_running(struct rq *rq)
1921 static void set_load_weight(struct task_struct *p)
1923 int prio = p->static_prio - MAX_RT_PRIO;
1924 struct load_weight *load = &p->se.load;
1927 * SCHED_IDLE tasks get minimal weight:
1929 if (p->policy == SCHED_IDLE) {
1930 load->weight = scale_load(WEIGHT_IDLEPRIO);
1931 load->inv_weight = WMULT_IDLEPRIO;
1935 load->weight = scale_load(prio_to_weight[prio]);
1936 load->inv_weight = prio_to_wmult[prio];
1939 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1941 update_rq_clock(rq);
1942 sched_info_queued(p);
1943 p->sched_class->enqueue_task(rq, p, flags);
1946 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1948 update_rq_clock(rq);
1949 sched_info_dequeued(p);
1950 p->sched_class->dequeue_task(rq, p, flags);
1954 * activate_task - move a task to the runqueue.
1956 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1958 if (task_contributes_to_load(p))
1959 rq->nr_uninterruptible--;
1961 enqueue_task(rq, p, flags);
1965 * deactivate_task - remove a task from the runqueue.
1967 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1969 if (task_contributes_to_load(p))
1970 rq->nr_uninterruptible++;
1972 dequeue_task(rq, p, flags);
1975 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1978 * There are no locks covering percpu hardirq/softirq time.
1979 * They are only modified in account_system_vtime, on corresponding CPU
1980 * with interrupts disabled. So, writes are safe.
1981 * They are read and saved off onto struct rq in update_rq_clock().
1982 * This may result in other CPU reading this CPU's irq time and can
1983 * race with irq/account_system_vtime on this CPU. We would either get old
1984 * or new value with a side effect of accounting a slice of irq time to wrong
1985 * task when irq is in progress while we read rq->clock. That is a worthy
1986 * compromise in place of having locks on each irq in account_system_time.
1988 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1989 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1991 static DEFINE_PER_CPU(u64, irq_start_time);
1992 static int sched_clock_irqtime;
1994 void enable_sched_clock_irqtime(void)
1996 sched_clock_irqtime = 1;
1999 void disable_sched_clock_irqtime(void)
2001 sched_clock_irqtime = 0;
2004 #ifndef CONFIG_64BIT
2005 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
2007 static inline void irq_time_write_begin(void)
2009 __this_cpu_inc(irq_time_seq.sequence);
2013 static inline void irq_time_write_end(void)
2016 __this_cpu_inc(irq_time_seq.sequence);
2019 static inline u64 irq_time_read(int cpu)
2025 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
2026 irq_time = per_cpu(cpu_softirq_time, cpu) +
2027 per_cpu(cpu_hardirq_time, cpu);
2028 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
2032 #else /* CONFIG_64BIT */
2033 static inline void irq_time_write_begin(void)
2037 static inline void irq_time_write_end(void)
2041 static inline u64 irq_time_read(int cpu)
2043 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
2045 #endif /* CONFIG_64BIT */
2048 * Called before incrementing preempt_count on {soft,}irq_enter
2049 * and before decrementing preempt_count on {soft,}irq_exit.
2051 void account_system_vtime(struct task_struct *curr)
2053 unsigned long flags;
2057 if (!sched_clock_irqtime)
2060 local_irq_save(flags);
2062 cpu = smp_processor_id();
2063 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2064 __this_cpu_add(irq_start_time, delta);
2066 irq_time_write_begin();
2068 * We do not account for softirq time from ksoftirqd here.
2069 * We want to continue accounting softirq time to ksoftirqd thread
2070 * in that case, so as not to confuse scheduler with a special task
2071 * that do not consume any time, but still wants to run.
2073 if (hardirq_count())
2074 __this_cpu_add(cpu_hardirq_time, delta);
2075 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
2076 __this_cpu_add(cpu_softirq_time, delta);
2078 irq_time_write_end();
2079 local_irq_restore(flags);
2081 EXPORT_SYMBOL_GPL(account_system_vtime);
2083 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2085 #ifdef CONFIG_PARAVIRT
2086 static inline u64 steal_ticks(u64 steal)
2088 if (unlikely(steal > NSEC_PER_SEC))
2089 return div_u64(steal, TICK_NSEC);
2091 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
2095 static void update_rq_clock_task(struct rq *rq, s64 delta)
2098 * In theory, the compile should just see 0 here, and optimize out the call
2099 * to sched_rt_avg_update. But I don't trust it...
2101 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2102 s64 steal = 0, irq_delta = 0;
2104 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2105 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2108 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2109 * this case when a previous update_rq_clock() happened inside a
2110 * {soft,}irq region.
2112 * When this happens, we stop ->clock_task and only update the
2113 * prev_irq_time stamp to account for the part that fit, so that a next
2114 * update will consume the rest. This ensures ->clock_task is
2117 * It does however cause some slight miss-attribution of {soft,}irq
2118 * time, a more accurate solution would be to update the irq_time using
2119 * the current rq->clock timestamp, except that would require using
2122 if (irq_delta > delta)
2125 rq->prev_irq_time += irq_delta;
2128 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2129 if (static_branch((¶virt_steal_rq_enabled))) {
2132 steal = paravirt_steal_clock(cpu_of(rq));
2133 steal -= rq->prev_steal_time_rq;
2135 if (unlikely(steal > delta))
2138 st = steal_ticks(steal);
2139 steal = st * TICK_NSEC;
2141 rq->prev_steal_time_rq += steal;
2147 rq->clock_task += delta;
2149 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2150 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2151 sched_rt_avg_update(rq, irq_delta + steal);
2155 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2156 static int irqtime_account_hi_update(void)
2158 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2159 unsigned long flags;
2163 local_irq_save(flags);
2164 latest_ns = this_cpu_read(cpu_hardirq_time);
2165 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2167 local_irq_restore(flags);
2171 static int irqtime_account_si_update(void)
2173 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2174 unsigned long flags;
2178 local_irq_save(flags);
2179 latest_ns = this_cpu_read(cpu_softirq_time);
2180 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2182 local_irq_restore(flags);
2186 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2188 #define sched_clock_irqtime (0)
2193 static void unthrottle_offline_cfs_rqs(struct rq *rq);
2196 #include "sched_idletask.c"
2197 #include "sched_fair.c"
2198 #include "sched_rt.c"
2199 #include "sched_autogroup.c"
2200 #include "sched_stoptask.c"
2201 #ifdef CONFIG_SCHED_DEBUG
2202 # include "sched_debug.c"
2205 void sched_set_stop_task(int cpu, struct task_struct *stop)
2207 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2208 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2212 * Make it appear like a SCHED_FIFO task, its something
2213 * userspace knows about and won't get confused about.
2215 * Also, it will make PI more or less work without too
2216 * much confusion -- but then, stop work should not
2217 * rely on PI working anyway.
2219 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2221 stop->sched_class = &stop_sched_class;
2224 cpu_rq(cpu)->stop = stop;
2228 * Reset it back to a normal scheduling class so that
2229 * it can die in pieces.
2231 old_stop->sched_class = &rt_sched_class;
2236 * __normal_prio - return the priority that is based on the static prio
2238 static inline int __normal_prio(struct task_struct *p)
2240 return p->static_prio;
2244 * Calculate the expected normal priority: i.e. priority
2245 * without taking RT-inheritance into account. Might be
2246 * boosted by interactivity modifiers. Changes upon fork,
2247 * setprio syscalls, and whenever the interactivity
2248 * estimator recalculates.
2250 static inline int normal_prio(struct task_struct *p)
2254 if (task_has_rt_policy(p))
2255 prio = MAX_RT_PRIO-1 - p->rt_priority;
2257 prio = __normal_prio(p);
2262 * Calculate the current priority, i.e. the priority
2263 * taken into account by the scheduler. This value might
2264 * be boosted by RT tasks, or might be boosted by
2265 * interactivity modifiers. Will be RT if the task got
2266 * RT-boosted. If not then it returns p->normal_prio.
2268 static int effective_prio(struct task_struct *p)
2270 p->normal_prio = normal_prio(p);
2272 * If we are RT tasks or we were boosted to RT priority,
2273 * keep the priority unchanged. Otherwise, update priority
2274 * to the normal priority:
2276 if (!rt_prio(p->prio))
2277 return p->normal_prio;
2282 * task_curr - is this task currently executing on a CPU?
2283 * @p: the task in question.
2285 inline int task_curr(const struct task_struct *p)
2287 return cpu_curr(task_cpu(p)) == p;
2290 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2291 const struct sched_class *prev_class,
2294 if (prev_class != p->sched_class) {
2295 if (prev_class->switched_from)
2296 prev_class->switched_from(rq, p);
2297 p->sched_class->switched_to(rq, p);
2298 } else if (oldprio != p->prio)
2299 p->sched_class->prio_changed(rq, p, oldprio);
2302 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2304 const struct sched_class *class;
2306 if (p->sched_class == rq->curr->sched_class) {
2307 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2309 for_each_class(class) {
2310 if (class == rq->curr->sched_class)
2312 if (class == p->sched_class) {
2313 resched_task(rq->curr);
2320 * A queue event has occurred, and we're going to schedule. In
2321 * this case, we can save a useless back to back clock update.
2323 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2324 rq->skip_clock_update = 1;
2329 * Is this task likely cache-hot:
2332 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2336 if (p->sched_class != &fair_sched_class)
2339 if (unlikely(p->policy == SCHED_IDLE))
2343 * Buddy candidates are cache hot:
2345 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2346 (&p->se == cfs_rq_of(&p->se)->next ||
2347 &p->se == cfs_rq_of(&p->se)->last))
2350 if (sysctl_sched_migration_cost == -1)
2352 if (sysctl_sched_migration_cost == 0)
2355 delta = now - p->se.exec_start;
2357 return delta < (s64)sysctl_sched_migration_cost;
2360 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2362 #ifdef CONFIG_SCHED_DEBUG
2364 * We should never call set_task_cpu() on a blocked task,
2365 * ttwu() will sort out the placement.
2367 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2368 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2370 #ifdef CONFIG_LOCKDEP
2372 * The caller should hold either p->pi_lock or rq->lock, when changing
2373 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2375 * sched_move_task() holds both and thus holding either pins the cgroup,
2378 * Furthermore, all task_rq users should acquire both locks, see
2381 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2382 lockdep_is_held(&task_rq(p)->lock)));
2386 trace_sched_migrate_task(p, new_cpu);
2388 if (task_cpu(p) != new_cpu) {
2389 p->se.nr_migrations++;
2390 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2393 __set_task_cpu(p, new_cpu);
2396 struct migration_arg {
2397 struct task_struct *task;
2401 static int migration_cpu_stop(void *data);
2404 * wait_task_inactive - wait for a thread to unschedule.
2406 * If @match_state is nonzero, it's the @p->state value just checked and
2407 * not expected to change. If it changes, i.e. @p might have woken up,
2408 * then return zero. When we succeed in waiting for @p to be off its CPU,
2409 * we return a positive number (its total switch count). If a second call
2410 * a short while later returns the same number, the caller can be sure that
2411 * @p has remained unscheduled the whole time.
2413 * The caller must ensure that the task *will* unschedule sometime soon,
2414 * else this function might spin for a *long* time. This function can't
2415 * be called with interrupts off, or it may introduce deadlock with
2416 * smp_call_function() if an IPI is sent by the same process we are
2417 * waiting to become inactive.
2419 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2421 unsigned long flags;
2428 * We do the initial early heuristics without holding
2429 * any task-queue locks at all. We'll only try to get
2430 * the runqueue lock when things look like they will
2436 * If the task is actively running on another CPU
2437 * still, just relax and busy-wait without holding
2440 * NOTE! Since we don't hold any locks, it's not
2441 * even sure that "rq" stays as the right runqueue!
2442 * But we don't care, since "task_running()" will
2443 * return false if the runqueue has changed and p
2444 * is actually now running somewhere else!
2446 while (task_running(rq, p)) {
2447 if (match_state && unlikely(p->state != match_state))
2453 * Ok, time to look more closely! We need the rq
2454 * lock now, to be *sure*. If we're wrong, we'll
2455 * just go back and repeat.
2457 rq = task_rq_lock(p, &flags);
2458 trace_sched_wait_task(p);
2459 running = task_running(rq, p);
2462 if (!match_state || p->state == match_state)
2463 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2464 task_rq_unlock(rq, p, &flags);
2467 * If it changed from the expected state, bail out now.
2469 if (unlikely(!ncsw))
2473 * Was it really running after all now that we
2474 * checked with the proper locks actually held?
2476 * Oops. Go back and try again..
2478 if (unlikely(running)) {
2484 * It's not enough that it's not actively running,
2485 * it must be off the runqueue _entirely_, and not
2488 * So if it was still runnable (but just not actively
2489 * running right now), it's preempted, and we should
2490 * yield - it could be a while.
2492 if (unlikely(on_rq)) {
2493 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2495 set_current_state(TASK_UNINTERRUPTIBLE);
2496 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2501 * Ahh, all good. It wasn't running, and it wasn't
2502 * runnable, which means that it will never become
2503 * running in the future either. We're all done!
2512 * kick_process - kick a running thread to enter/exit the kernel
2513 * @p: the to-be-kicked thread
2515 * Cause a process which is running on another CPU to enter
2516 * kernel-mode, without any delay. (to get signals handled.)
2518 * NOTE: this function doesn't have to take the runqueue lock,
2519 * because all it wants to ensure is that the remote task enters
2520 * the kernel. If the IPI races and the task has been migrated
2521 * to another CPU then no harm is done and the purpose has been
2524 void kick_process(struct task_struct *p)
2530 if ((cpu != smp_processor_id()) && task_curr(p))
2531 smp_send_reschedule(cpu);
2534 EXPORT_SYMBOL_GPL(kick_process);
2535 #endif /* CONFIG_SMP */
2539 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2541 static int select_fallback_rq(int cpu, struct task_struct *p)
2544 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2546 /* Look for allowed, online CPU in same node. */
2547 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2548 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
2551 /* Any allowed, online CPU? */
2552 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
2553 if (dest_cpu < nr_cpu_ids)
2556 /* No more Mr. Nice Guy. */
2557 dest_cpu = cpuset_cpus_allowed_fallback(p);
2559 * Don't tell them about moving exiting tasks or
2560 * kernel threads (both mm NULL), since they never
2563 if (p->mm && printk_ratelimit()) {
2564 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2565 task_pid_nr(p), p->comm, cpu);
2572 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2575 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2577 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2580 * In order not to call set_task_cpu() on a blocking task we need
2581 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2584 * Since this is common to all placement strategies, this lives here.
2586 * [ this allows ->select_task() to simply return task_cpu(p) and
2587 * not worry about this generic constraint ]
2589 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
2591 cpu = select_fallback_rq(task_cpu(p), p);
2596 static void update_avg(u64 *avg, u64 sample)
2598 s64 diff = sample - *avg;
2604 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2606 #ifdef CONFIG_SCHEDSTATS
2607 struct rq *rq = this_rq();
2610 int this_cpu = smp_processor_id();
2612 if (cpu == this_cpu) {
2613 schedstat_inc(rq, ttwu_local);
2614 schedstat_inc(p, se.statistics.nr_wakeups_local);
2616 struct sched_domain *sd;
2618 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2620 for_each_domain(this_cpu, sd) {
2621 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2622 schedstat_inc(sd, ttwu_wake_remote);
2629 if (wake_flags & WF_MIGRATED)
2630 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2632 #endif /* CONFIG_SMP */
2634 schedstat_inc(rq, ttwu_count);
2635 schedstat_inc(p, se.statistics.nr_wakeups);
2637 if (wake_flags & WF_SYNC)
2638 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2640 #endif /* CONFIG_SCHEDSTATS */
2643 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2645 activate_task(rq, p, en_flags);
2648 /* if a worker is waking up, notify workqueue */
2649 if (p->flags & PF_WQ_WORKER)
2650 wq_worker_waking_up(p, cpu_of(rq));
2654 * Mark the task runnable and perform wakeup-preemption.
2657 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2659 trace_sched_wakeup(p, true);
2660 check_preempt_curr(rq, p, wake_flags);
2662 p->state = TASK_RUNNING;
2664 if (p->sched_class->task_woken)
2665 p->sched_class->task_woken(rq, p);
2667 if (rq->idle_stamp) {
2668 u64 delta = rq->clock - rq->idle_stamp;
2669 u64 max = 2*sysctl_sched_migration_cost;
2674 update_avg(&rq->avg_idle, delta);
2681 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2684 if (p->sched_contributes_to_load)
2685 rq->nr_uninterruptible--;
2688 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2689 ttwu_do_wakeup(rq, p, wake_flags);
2693 * Called in case the task @p isn't fully descheduled from its runqueue,
2694 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2695 * since all we need to do is flip p->state to TASK_RUNNING, since
2696 * the task is still ->on_rq.
2698 static int ttwu_remote(struct task_struct *p, int wake_flags)
2703 rq = __task_rq_lock(p);
2705 ttwu_do_wakeup(rq, p, wake_flags);
2708 __task_rq_unlock(rq);
2714 static void sched_ttwu_pending(void)
2716 struct rq *rq = this_rq();
2717 struct llist_node *llist = llist_del_all(&rq->wake_list);
2718 struct task_struct *p;
2720 raw_spin_lock(&rq->lock);
2723 p = llist_entry(llist, struct task_struct, wake_entry);
2724 llist = llist_next(llist);
2725 ttwu_do_activate(rq, p, 0);
2728 raw_spin_unlock(&rq->lock);
2731 void scheduler_ipi(void)
2733 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2737 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2738 * traditionally all their work was done from the interrupt return
2739 * path. Now that we actually do some work, we need to make sure
2742 * Some archs already do call them, luckily irq_enter/exit nest
2745 * Arguably we should visit all archs and update all handlers,
2746 * however a fair share of IPIs are still resched only so this would
2747 * somewhat pessimize the simple resched case.
2750 sched_ttwu_pending();
2753 * Check if someone kicked us for doing the nohz idle load balance.
2755 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
2756 this_rq()->idle_balance = 1;
2757 raise_softirq_irqoff(SCHED_SOFTIRQ);
2762 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2764 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
2765 smp_send_reschedule(cpu);
2768 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2769 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2774 rq = __task_rq_lock(p);
2776 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2777 ttwu_do_wakeup(rq, p, wake_flags);
2780 __task_rq_unlock(rq);
2785 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2786 #endif /* CONFIG_SMP */
2788 static void ttwu_queue(struct task_struct *p, int cpu)
2790 struct rq *rq = cpu_rq(cpu);
2792 #if defined(CONFIG_SMP)
2793 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2794 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2795 ttwu_queue_remote(p, cpu);
2800 raw_spin_lock(&rq->lock);
2801 ttwu_do_activate(rq, p, 0);
2802 raw_spin_unlock(&rq->lock);
2806 * try_to_wake_up - wake up a thread
2807 * @p: the thread to be awakened
2808 * @state: the mask of task states that can be woken
2809 * @wake_flags: wake modifier flags (WF_*)
2811 * Put it on the run-queue if it's not already there. The "current"
2812 * thread is always on the run-queue (except when the actual
2813 * re-schedule is in progress), and as such you're allowed to do
2814 * the simpler "current->state = TASK_RUNNING" to mark yourself
2815 * runnable without the overhead of this.
2817 * Returns %true if @p was woken up, %false if it was already running
2818 * or @state didn't match @p's state.
2821 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2823 unsigned long flags;
2824 int cpu, success = 0;
2827 raw_spin_lock_irqsave(&p->pi_lock, flags);
2828 if (!(p->state & state))
2831 success = 1; /* we're going to change ->state */
2834 if (p->on_rq && ttwu_remote(p, wake_flags))
2839 * If the owning (remote) cpu is still in the middle of schedule() with
2840 * this task as prev, wait until its done referencing the task.
2843 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2845 * In case the architecture enables interrupts in
2846 * context_switch(), we cannot busy wait, since that
2847 * would lead to deadlocks when an interrupt hits and
2848 * tries to wake up @prev. So bail and do a complete
2851 if (ttwu_activate_remote(p, wake_flags))
2858 * Pairs with the smp_wmb() in finish_lock_switch().
2862 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2863 p->state = TASK_WAKING;
2865 if (p->sched_class->task_waking)
2866 p->sched_class->task_waking(p);
2868 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2869 if (task_cpu(p) != cpu) {
2870 wake_flags |= WF_MIGRATED;
2871 set_task_cpu(p, cpu);
2873 #endif /* CONFIG_SMP */
2877 ttwu_stat(p, cpu, wake_flags);
2879 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2885 * try_to_wake_up_local - try to wake up a local task with rq lock held
2886 * @p: the thread to be awakened
2888 * Put @p on the run-queue if it's not already there. The caller must
2889 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2892 static void try_to_wake_up_local(struct task_struct *p)
2894 struct rq *rq = task_rq(p);
2896 if (WARN_ON_ONCE(rq != this_rq()) ||
2897 WARN_ON_ONCE(p == current))
2900 lockdep_assert_held(&rq->lock);
2902 if (!raw_spin_trylock(&p->pi_lock)) {
2903 raw_spin_unlock(&rq->lock);
2904 raw_spin_lock(&p->pi_lock);
2905 raw_spin_lock(&rq->lock);
2908 if (!(p->state & TASK_NORMAL))
2912 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2914 ttwu_do_wakeup(rq, p, 0);
2915 ttwu_stat(p, smp_processor_id(), 0);
2917 raw_spin_unlock(&p->pi_lock);
2921 * wake_up_process - Wake up a specific process
2922 * @p: The process to be woken up.
2924 * Attempt to wake up the nominated process and move it to the set of runnable
2925 * processes. Returns 1 if the process was woken up, 0 if it was already
2928 * It may be assumed that this function implies a write memory barrier before
2929 * changing the task state if and only if any tasks are woken up.
2931 int wake_up_process(struct task_struct *p)
2933 WARN_ON(task_is_stopped_or_traced(p));
2934 return try_to_wake_up(p, TASK_NORMAL, 0);
2936 EXPORT_SYMBOL(wake_up_process);
2938 int wake_up_state(struct task_struct *p, unsigned int state)
2940 return try_to_wake_up(p, state, 0);
2944 * Perform scheduler related setup for a newly forked process p.
2945 * p is forked by current.
2947 * __sched_fork() is basic setup used by init_idle() too:
2949 static void __sched_fork(struct task_struct *p)
2954 p->se.exec_start = 0;
2955 p->se.sum_exec_runtime = 0;
2956 p->se.prev_sum_exec_runtime = 0;
2957 p->se.nr_migrations = 0;
2959 INIT_LIST_HEAD(&p->se.group_node);
2961 #ifdef CONFIG_SCHEDSTATS
2962 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2965 INIT_LIST_HEAD(&p->rt.run_list);
2967 #ifdef CONFIG_PREEMPT_NOTIFIERS
2968 INIT_HLIST_HEAD(&p->preempt_notifiers);
2973 * fork()/clone()-time setup:
2975 void sched_fork(struct task_struct *p)
2977 unsigned long flags;
2978 int cpu = get_cpu();
2982 * We mark the process as running here. This guarantees that
2983 * nobody will actually run it, and a signal or other external
2984 * event cannot wake it up and insert it on the runqueue either.
2986 p->state = TASK_RUNNING;
2989 * Make sure we do not leak PI boosting priority to the child.
2991 p->prio = current->normal_prio;
2994 * Revert to default priority/policy on fork if requested.
2996 if (unlikely(p->sched_reset_on_fork)) {
2997 if (task_has_rt_policy(p)) {
2998 p->policy = SCHED_NORMAL;
2999 p->static_prio = NICE_TO_PRIO(0);
3001 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3002 p->static_prio = NICE_TO_PRIO(0);
3004 p->prio = p->normal_prio = __normal_prio(p);
3008 * We don't need the reset flag anymore after the fork. It has
3009 * fulfilled its duty:
3011 p->sched_reset_on_fork = 0;
3014 if (!rt_prio(p->prio))
3015 p->sched_class = &fair_sched_class;
3017 if (p->sched_class->task_fork)
3018 p->sched_class->task_fork(p);
3021 * The child is not yet in the pid-hash so no cgroup attach races,
3022 * and the cgroup is pinned to this child due to cgroup_fork()
3023 * is ran before sched_fork().
3025 * Silence PROVE_RCU.
3027 raw_spin_lock_irqsave(&p->pi_lock, flags);
3028 set_task_cpu(p, cpu);
3029 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3031 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3032 if (likely(sched_info_on()))
3033 memset(&p->sched_info, 0, sizeof(p->sched_info));
3035 #if defined(CONFIG_SMP)
3038 #ifdef CONFIG_PREEMPT_COUNT
3039 /* Want to start with kernel preemption disabled. */
3040 task_thread_info(p)->preempt_count = 1;
3043 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3050 * wake_up_new_task - wake up a newly created task for the first time.
3052 * This function will do some initial scheduler statistics housekeeping
3053 * that must be done for every newly created context, then puts the task
3054 * on the runqueue and wakes it.
3056 void wake_up_new_task(struct task_struct *p)
3058 unsigned long flags;
3061 raw_spin_lock_irqsave(&p->pi_lock, flags);
3064 * Fork balancing, do it here and not earlier because:
3065 * - cpus_allowed can change in the fork path
3066 * - any previously selected cpu might disappear through hotplug
3068 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3071 rq = __task_rq_lock(p);
3072 activate_task(rq, p, 0);
3074 trace_sched_wakeup_new(p, true);
3075 check_preempt_curr(rq, p, WF_FORK);
3077 if (p->sched_class->task_woken)
3078 p->sched_class->task_woken(rq, p);
3080 task_rq_unlock(rq, p, &flags);
3083 #ifdef CONFIG_PREEMPT_NOTIFIERS
3086 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3087 * @notifier: notifier struct to register
3089 void preempt_notifier_register(struct preempt_notifier *notifier)
3091 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3093 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3096 * preempt_notifier_unregister - no longer interested in preemption notifications
3097 * @notifier: notifier struct to unregister
3099 * This is safe to call from within a preemption notifier.
3101 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3103 hlist_del(¬ifier->link);
3105 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3107 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3109 struct preempt_notifier *notifier;
3110 struct hlist_node *node;
3112 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3113 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3117 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3118 struct task_struct *next)
3120 struct preempt_notifier *notifier;
3121 struct hlist_node *node;
3123 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3124 notifier->ops->sched_out(notifier, next);
3127 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3129 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3134 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3135 struct task_struct *next)
3139 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3142 * prepare_task_switch - prepare to switch tasks
3143 * @rq: the runqueue preparing to switch
3144 * @prev: the current task that is being switched out
3145 * @next: the task we are going to switch to.
3147 * This is called with the rq lock held and interrupts off. It must
3148 * be paired with a subsequent finish_task_switch after the context
3151 * prepare_task_switch sets up locking and calls architecture specific
3155 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3156 struct task_struct *next)
3158 sched_info_switch(prev, next);
3159 perf_event_task_sched_out(prev, next);
3160 fire_sched_out_preempt_notifiers(prev, next);
3161 prepare_lock_switch(rq, next);
3162 prepare_arch_switch(next);
3163 trace_sched_switch(prev, next);
3167 * finish_task_switch - clean up after a task-switch
3168 * @rq: runqueue associated with task-switch
3169 * @prev: the thread we just switched away from.
3171 * finish_task_switch must be called after the context switch, paired
3172 * with a prepare_task_switch call before the context switch.
3173 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3174 * and do any other architecture-specific cleanup actions.
3176 * Note that we may have delayed dropping an mm in context_switch(). If
3177 * so, we finish that here outside of the runqueue lock. (Doing it
3178 * with the lock held can cause deadlocks; see schedule() for
3181 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3182 __releases(rq->lock)
3184 struct mm_struct *mm = rq->prev_mm;
3190 * A task struct has one reference for the use as "current".
3191 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3192 * schedule one last time. The schedule call will never return, and
3193 * the scheduled task must drop that reference.
3194 * The test for TASK_DEAD must occur while the runqueue locks are
3195 * still held, otherwise prev could be scheduled on another cpu, die
3196 * there before we look at prev->state, and then the reference would
3198 * Manfred Spraul <manfred@colorfullife.com>
3200 prev_state = prev->state;
3201 finish_arch_switch(prev);
3202 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3203 local_irq_disable();
3204 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3205 perf_event_task_sched_in(prev, current);
3206 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3208 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3209 finish_lock_switch(rq, prev);
3211 fire_sched_in_preempt_notifiers(current);
3214 if (unlikely(prev_state == TASK_DEAD)) {
3216 * Remove function-return probe instances associated with this
3217 * task and put them back on the free list.
3219 kprobe_flush_task(prev);
3220 put_task_struct(prev);
3226 /* assumes rq->lock is held */
3227 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3229 if (prev->sched_class->pre_schedule)
3230 prev->sched_class->pre_schedule(rq, prev);
3233 /* rq->lock is NOT held, but preemption is disabled */
3234 static inline void post_schedule(struct rq *rq)
3236 if (rq->post_schedule) {
3237 unsigned long flags;
3239 raw_spin_lock_irqsave(&rq->lock, flags);
3240 if (rq->curr->sched_class->post_schedule)
3241 rq->curr->sched_class->post_schedule(rq);
3242 raw_spin_unlock_irqrestore(&rq->lock, flags);
3244 rq->post_schedule = 0;
3250 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3254 static inline void post_schedule(struct rq *rq)
3261 * schedule_tail - first thing a freshly forked thread must call.
3262 * @prev: the thread we just switched away from.
3264 asmlinkage void schedule_tail(struct task_struct *prev)
3265 __releases(rq->lock)
3267 struct rq *rq = this_rq();
3269 finish_task_switch(rq, prev);
3272 * FIXME: do we need to worry about rq being invalidated by the
3277 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3278 /* In this case, finish_task_switch does not reenable preemption */
3281 if (current->set_child_tid)
3282 put_user(task_pid_vnr(current), current->set_child_tid);
3286 * context_switch - switch to the new MM and the new
3287 * thread's register state.
3290 context_switch(struct rq *rq, struct task_struct *prev,
3291 struct task_struct *next)
3293 struct mm_struct *mm, *oldmm;
3295 prepare_task_switch(rq, prev, next);
3298 oldmm = prev->active_mm;
3300 * For paravirt, this is coupled with an exit in switch_to to
3301 * combine the page table reload and the switch backend into
3304 arch_start_context_switch(prev);
3307 next->active_mm = oldmm;
3308 atomic_inc(&oldmm->mm_count);
3309 enter_lazy_tlb(oldmm, next);
3311 switch_mm(oldmm, mm, next);
3314 prev->active_mm = NULL;
3315 rq->prev_mm = oldmm;
3318 * Since the runqueue lock will be released by the next
3319 * task (which is an invalid locking op but in the case
3320 * of the scheduler it's an obvious special-case), so we
3321 * do an early lockdep release here:
3323 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3324 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3327 /* Here we just switch the register state and the stack. */
3328 switch_to(prev, next, prev);
3332 * this_rq must be evaluated again because prev may have moved
3333 * CPUs since it called schedule(), thus the 'rq' on its stack
3334 * frame will be invalid.
3336 finish_task_switch(this_rq(), prev);
3340 * nr_running, nr_uninterruptible and nr_context_switches:
3342 * externally visible scheduler statistics: current number of runnable
3343 * threads, current number of uninterruptible-sleeping threads, total
3344 * number of context switches performed since bootup.
3346 unsigned long nr_running(void)
3348 unsigned long i, sum = 0;
3350 for_each_online_cpu(i)
3351 sum += cpu_rq(i)->nr_running;
3356 unsigned long nr_uninterruptible(void)
3358 unsigned long i, sum = 0;
3360 for_each_possible_cpu(i)
3361 sum += cpu_rq(i)->nr_uninterruptible;
3364 * Since we read the counters lockless, it might be slightly
3365 * inaccurate. Do not allow it to go below zero though:
3367 if (unlikely((long)sum < 0))
3373 unsigned long long nr_context_switches(void)
3376 unsigned long long sum = 0;
3378 for_each_possible_cpu(i)
3379 sum += cpu_rq(i)->nr_switches;
3384 unsigned long nr_iowait(void)
3386 unsigned long i, sum = 0;
3388 for_each_possible_cpu(i)
3389 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3394 unsigned long nr_iowait_cpu(int cpu)
3396 struct rq *this = cpu_rq(cpu);
3397 return atomic_read(&this->nr_iowait);
3400 unsigned long this_cpu_load(void)
3402 struct rq *this = this_rq();
3403 return this->cpu_load[0];
3408 * Global load-average calculations
3410 * We take a distributed and async approach to calculating the global load-avg
3411 * in order to minimize overhead.
3413 * The global load average is an exponentially decaying average of nr_running +
3414 * nr_uninterruptible.
3416 * Once every LOAD_FREQ:
3419 * for_each_possible_cpu(cpu)
3420 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
3422 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
3424 * Due to a number of reasons the above turns in the mess below:
3426 * - for_each_possible_cpu() is prohibitively expensive on machines with
3427 * serious number of cpus, therefore we need to take a distributed approach
3428 * to calculating nr_active.
3430 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
3431 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
3433 * So assuming nr_active := 0 when we start out -- true per definition, we
3434 * can simply take per-cpu deltas and fold those into a global accumulate
3435 * to obtain the same result. See calc_load_fold_active().
3437 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
3438 * across the machine, we assume 10 ticks is sufficient time for every
3439 * cpu to have completed this task.
3441 * This places an upper-bound on the IRQ-off latency of the machine. Then
3442 * again, being late doesn't loose the delta, just wrecks the sample.
3444 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
3445 * this would add another cross-cpu cacheline miss and atomic operation
3446 * to the wakeup path. Instead we increment on whatever cpu the task ran
3447 * when it went into uninterruptible state and decrement on whatever cpu
3448 * did the wakeup. This means that only the sum of nr_uninterruptible over
3449 * all cpus yields the correct result.
3451 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
3454 /* Variables and functions for calc_load */
3455 static atomic_long_t calc_load_tasks;
3456 static unsigned long calc_load_update;
3457 unsigned long avenrun[3];
3458 EXPORT_SYMBOL(avenrun); /* should be removed */
3461 * get_avenrun - get the load average array
3462 * @loads: pointer to dest load array
3463 * @offset: offset to add
3464 * @shift: shift count to shift the result left
3466 * These values are estimates at best, so no need for locking.
3468 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3470 loads[0] = (avenrun[0] + offset) << shift;
3471 loads[1] = (avenrun[1] + offset) << shift;
3472 loads[2] = (avenrun[2] + offset) << shift;
3475 static long calc_load_fold_active(struct rq *this_rq)
3477 long nr_active, delta = 0;
3479 nr_active = this_rq->nr_running;
3480 nr_active += (long) this_rq->nr_uninterruptible;
3482 if (nr_active != this_rq->calc_load_active) {
3483 delta = nr_active - this_rq->calc_load_active;
3484 this_rq->calc_load_active = nr_active;
3491 * a1 = a0 * e + a * (1 - e)
3493 static unsigned long
3494 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3497 load += active * (FIXED_1 - exp);
3498 load += 1UL << (FSHIFT - 1);
3499 return load >> FSHIFT;
3504 * Handle NO_HZ for the global load-average.
3506 * Since the above described distributed algorithm to compute the global
3507 * load-average relies on per-cpu sampling from the tick, it is affected by
3510 * The basic idea is to fold the nr_active delta into a global idle-delta upon
3511 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
3512 * when we read the global state.
3514 * Obviously reality has to ruin such a delightfully simple scheme:
3516 * - When we go NO_HZ idle during the window, we can negate our sample
3517 * contribution, causing under-accounting.
3519 * We avoid this by keeping two idle-delta counters and flipping them
3520 * when the window starts, thus separating old and new NO_HZ load.
3522 * The only trick is the slight shift in index flip for read vs write.
3526 * |-|-----------|-|-----------|-|-----------|-|
3527 * r:0 0 1 1 0 0 1 1 0
3528 * w:0 1 1 0 0 1 1 0 0
3530 * This ensures we'll fold the old idle contribution in this window while
3531 * accumlating the new one.
3533 * - When we wake up from NO_HZ idle during the window, we push up our
3534 * contribution, since we effectively move our sample point to a known
3537 * This is solved by pushing the window forward, and thus skipping the
3538 * sample, for this cpu (effectively using the idle-delta for this cpu which
3539 * was in effect at the time the window opened). This also solves the issue
3540 * of having to deal with a cpu having been in NOHZ idle for multiple
3541 * LOAD_FREQ intervals.
3543 * When making the ILB scale, we should try to pull this in as well.
3545 static atomic_long_t calc_load_idle[2];
3546 static int calc_load_idx;
3548 static inline int calc_load_write_idx(void)
3550 int idx = calc_load_idx;
3553 * See calc_global_nohz(), if we observe the new index, we also
3554 * need to observe the new update time.
3559 * If the folding window started, make sure we start writing in the
3562 if (!time_before(jiffies, calc_load_update))