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
1020 * Pairs with the control dependency and rmb in try_to_wake_up().
1025 #ifdef CONFIG_DEBUG_SPINLOCK
1026 /* this is a valid case when another task releases the spinlock */
1027 rq->lock.owner = current;
1030 * If we are tracking spinlock dependencies then we have to
1031 * fix up the runqueue lock - which gets 'carried over' from
1032 * prev into current:
1034 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1036 raw_spin_unlock_irq(&rq->lock);
1039 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1040 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1044 * We can optimise this out completely for !SMP, because the
1045 * SMP rebalancing from interrupt is the only thing that cares
1050 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1051 raw_spin_unlock_irq(&rq->lock);
1053 raw_spin_unlock(&rq->lock);
1057 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1061 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1062 * We must ensure this doesn't happen until the switch is completely
1068 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1072 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1075 * __task_rq_lock - lock the rq @p resides on.
1077 static inline struct rq *__task_rq_lock(struct task_struct *p)
1078 __acquires(rq->lock)
1082 lockdep_assert_held(&p->pi_lock);
1086 raw_spin_lock(&rq->lock);
1087 if (likely(rq == task_rq(p)))
1089 raw_spin_unlock(&rq->lock);
1094 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1096 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1097 __acquires(p->pi_lock)
1098 __acquires(rq->lock)
1103 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1105 raw_spin_lock(&rq->lock);
1106 if (likely(rq == task_rq(p)))
1108 raw_spin_unlock(&rq->lock);
1109 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1113 static void __task_rq_unlock(struct rq *rq)
1114 __releases(rq->lock)
1116 raw_spin_unlock(&rq->lock);
1120 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1121 __releases(rq->lock)
1122 __releases(p->pi_lock)
1124 raw_spin_unlock(&rq->lock);
1125 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1129 * this_rq_lock - lock this runqueue and disable interrupts.
1131 static struct rq *this_rq_lock(void)
1132 __acquires(rq->lock)
1136 local_irq_disable();
1138 raw_spin_lock(&rq->lock);
1143 #ifdef CONFIG_SCHED_HRTICK
1145 * Use HR-timers to deliver accurate preemption points.
1147 * Its all a bit involved since we cannot program an hrt while holding the
1148 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1151 * When we get rescheduled we reprogram the hrtick_timer outside of the
1157 * - enabled by features
1158 * - hrtimer is actually high res
1160 static inline int hrtick_enabled(struct rq *rq)
1162 if (!sched_feat(HRTICK))
1164 if (!cpu_active(cpu_of(rq)))
1166 return hrtimer_is_hres_active(&rq->hrtick_timer);
1169 static void hrtick_clear(struct rq *rq)
1171 if (hrtimer_active(&rq->hrtick_timer))
1172 hrtimer_cancel(&rq->hrtick_timer);
1176 * High-resolution timer tick.
1177 * Runs from hardirq context with interrupts disabled.
1179 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1181 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1183 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1185 raw_spin_lock(&rq->lock);
1186 update_rq_clock(rq);
1187 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1188 raw_spin_unlock(&rq->lock);
1190 return HRTIMER_NORESTART;
1195 * called from hardirq (IPI) context
1197 static void __hrtick_start(void *arg)
1199 struct rq *rq = arg;
1201 raw_spin_lock(&rq->lock);
1202 hrtimer_restart(&rq->hrtick_timer);
1203 rq->hrtick_csd_pending = 0;
1204 raw_spin_unlock(&rq->lock);
1208 * Called to set the hrtick timer state.
1210 * called with rq->lock held and irqs disabled
1212 static void hrtick_start(struct rq *rq, u64 delay)
1214 struct hrtimer *timer = &rq->hrtick_timer;
1215 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1217 hrtimer_set_expires(timer, time);
1219 if (rq == this_rq()) {
1220 hrtimer_restart(timer);
1221 } else if (!rq->hrtick_csd_pending) {
1222 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1223 rq->hrtick_csd_pending = 1;
1228 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1230 int cpu = (int)(long)hcpu;
1233 case CPU_UP_CANCELED:
1234 case CPU_UP_CANCELED_FROZEN:
1235 case CPU_DOWN_PREPARE:
1236 case CPU_DOWN_PREPARE_FROZEN:
1238 case CPU_DEAD_FROZEN:
1239 hrtick_clear(cpu_rq(cpu));
1246 static __init void init_hrtick(void)
1248 hotcpu_notifier(hotplug_hrtick, 0);
1252 * Called to set the hrtick timer state.
1254 * called with rq->lock held and irqs disabled
1256 static void hrtick_start(struct rq *rq, u64 delay)
1258 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1259 HRTIMER_MODE_REL_PINNED, 0);
1262 static inline void init_hrtick(void)
1265 #endif /* CONFIG_SMP */
1267 static void init_rq_hrtick(struct rq *rq)
1270 rq->hrtick_csd_pending = 0;
1272 rq->hrtick_csd.flags = 0;
1273 rq->hrtick_csd.func = __hrtick_start;
1274 rq->hrtick_csd.info = rq;
1277 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1278 rq->hrtick_timer.function = hrtick;
1280 #else /* CONFIG_SCHED_HRTICK */
1281 static inline void hrtick_clear(struct rq *rq)
1285 static inline void init_rq_hrtick(struct rq *rq)
1289 static inline void init_hrtick(void)
1292 #endif /* CONFIG_SCHED_HRTICK */
1295 * resched_task - mark a task 'to be rescheduled now'.
1297 * On UP this means the setting of the need_resched flag, on SMP it
1298 * might also involve a cross-CPU call to trigger the scheduler on
1303 #ifndef tsk_is_polling
1304 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1307 static void resched_task(struct task_struct *p)
1311 assert_raw_spin_locked(&task_rq(p)->lock);
1313 if (test_tsk_need_resched(p))
1316 set_tsk_need_resched(p);
1319 if (cpu == smp_processor_id())
1322 /* NEED_RESCHED must be visible before we test polling */
1324 if (!tsk_is_polling(p))
1325 smp_send_reschedule(cpu);
1328 static void resched_cpu(int cpu)
1330 struct rq *rq = cpu_rq(cpu);
1331 unsigned long flags;
1333 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1335 resched_task(cpu_curr(cpu));
1336 raw_spin_unlock_irqrestore(&rq->lock, flags);
1341 * In the semi idle case, use the nearest busy cpu for migrating timers
1342 * from an idle cpu. This is good for power-savings.
1344 * We don't do similar optimization for completely idle system, as
1345 * selecting an idle cpu will add more delays to the timers than intended
1346 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1348 int get_nohz_timer_target(void)
1350 int cpu = smp_processor_id();
1352 struct sched_domain *sd;
1355 for_each_domain(cpu, sd) {
1356 for_each_cpu(i, sched_domain_span(sd)) {
1368 * When add_timer_on() enqueues a timer into the timer wheel of an
1369 * idle CPU then this timer might expire before the next timer event
1370 * which is scheduled to wake up that CPU. In case of a completely
1371 * idle system the next event might even be infinite time into the
1372 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1373 * leaves the inner idle loop so the newly added timer is taken into
1374 * account when the CPU goes back to idle and evaluates the timer
1375 * wheel for the next timer event.
1377 void wake_up_idle_cpu(int cpu)
1379 struct rq *rq = cpu_rq(cpu);
1381 if (cpu == smp_processor_id())
1385 * This is safe, as this function is called with the timer
1386 * wheel base lock of (cpu) held. When the CPU is on the way
1387 * to idle and has not yet set rq->curr to idle then it will
1388 * be serialized on the timer wheel base lock and take the new
1389 * timer into account automatically.
1391 if (rq->curr != rq->idle)
1395 * We can set TIF_RESCHED on the idle task of the other CPU
1396 * lockless. The worst case is that the other CPU runs the
1397 * idle task through an additional NOOP schedule()
1399 set_tsk_need_resched(rq->idle);
1401 /* NEED_RESCHED must be visible before we test polling */
1403 if (!tsk_is_polling(rq->idle))
1404 smp_send_reschedule(cpu);
1407 static inline bool got_nohz_idle_kick(void)
1409 return idle_cpu(smp_processor_id()) && this_rq()->nohz_balance_kick;
1412 #else /* CONFIG_NO_HZ */
1414 static inline bool got_nohz_idle_kick(void)
1419 #endif /* CONFIG_NO_HZ */
1421 static u64 sched_avg_period(void)
1423 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1426 static void sched_avg_update(struct rq *rq)
1428 s64 period = sched_avg_period();
1430 while ((s64)(rq->clock - rq->age_stamp) > period) {
1432 * Inline assembly required to prevent the compiler
1433 * optimising this loop into a divmod call.
1434 * See __iter_div_u64_rem() for another example of this.
1436 asm("" : "+rm" (rq->age_stamp));
1437 rq->age_stamp += period;
1442 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1444 rq->rt_avg += rt_delta;
1445 sched_avg_update(rq);
1448 #else /* !CONFIG_SMP */
1449 static void resched_task(struct task_struct *p)
1451 assert_raw_spin_locked(&task_rq(p)->lock);
1452 set_tsk_need_resched(p);
1455 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1459 static void sched_avg_update(struct rq *rq)
1462 #endif /* CONFIG_SMP */
1464 #if BITS_PER_LONG == 32
1465 # define WMULT_CONST (~0UL)
1467 # define WMULT_CONST (1UL << 32)
1470 #define WMULT_SHIFT 32
1473 * Shift right and round:
1475 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1478 * delta *= weight / lw
1480 static unsigned long
1481 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1482 struct load_weight *lw)
1487 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1488 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1489 * 2^SCHED_LOAD_RESOLUTION.
1491 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1492 tmp = (u64)delta_exec * scale_load_down(weight);
1494 tmp = (u64)delta_exec;
1496 if (!lw->inv_weight) {
1497 unsigned long w = scale_load_down(lw->weight);
1499 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1501 else if (unlikely(!w))
1502 lw->inv_weight = WMULT_CONST;
1504 lw->inv_weight = WMULT_CONST / w;
1508 * Check whether we'd overflow the 64-bit multiplication:
1510 if (unlikely(tmp > WMULT_CONST))
1511 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1514 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1516 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1519 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1525 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1531 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1538 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1539 * of tasks with abnormal "nice" values across CPUs the contribution that
1540 * each task makes to its run queue's load is weighted according to its
1541 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1542 * scaled version of the new time slice allocation that they receive on time
1546 #define WEIGHT_IDLEPRIO 3
1547 #define WMULT_IDLEPRIO 1431655765
1550 * Nice levels are multiplicative, with a gentle 10% change for every
1551 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1552 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1553 * that remained on nice 0.
1555 * The "10% effect" is relative and cumulative: from _any_ nice level,
1556 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1557 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1558 * If a task goes up by ~10% and another task goes down by ~10% then
1559 * the relative distance between them is ~25%.)
1561 static const int prio_to_weight[40] = {
1562 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1563 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1564 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1565 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1566 /* 0 */ 1024, 820, 655, 526, 423,
1567 /* 5 */ 335, 272, 215, 172, 137,
1568 /* 10 */ 110, 87, 70, 56, 45,
1569 /* 15 */ 36, 29, 23, 18, 15,
1573 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1575 * In cases where the weight does not change often, we can use the
1576 * precalculated inverse to speed up arithmetics by turning divisions
1577 * into multiplications:
1579 static const u32 prio_to_wmult[40] = {
1580 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1581 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1582 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1583 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1584 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1585 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1586 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1587 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1590 /* Time spent by the tasks of the cpu accounting group executing in ... */
1591 enum cpuacct_stat_index {
1592 CPUACCT_STAT_USER, /* ... user mode */
1593 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1595 CPUACCT_STAT_NSTATS,
1598 #ifdef CONFIG_CGROUP_CPUACCT
1599 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1600 static void cpuacct_update_stats(struct task_struct *tsk,
1601 enum cpuacct_stat_index idx, cputime_t val);
1603 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1604 static inline void cpuacct_update_stats(struct task_struct *tsk,
1605 enum cpuacct_stat_index idx, cputime_t val) {}
1608 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1610 update_load_add(&rq->load, load);
1613 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1615 update_load_sub(&rq->load, load);
1618 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1619 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1620 typedef int (*tg_visitor)(struct task_group *, void *);
1623 * Iterate task_group tree rooted at *from, calling @down when first entering a
1624 * node and @up when leaving it for the final time.
1626 * Caller must hold rcu_lock or sufficient equivalent.
1628 static int walk_tg_tree_from(struct task_group *from,
1629 tg_visitor down, tg_visitor up, void *data)
1631 struct task_group *parent, *child;
1637 ret = (*down)(parent, data);
1640 list_for_each_entry_rcu(child, &parent->children, siblings) {
1647 ret = (*up)(parent, data);
1648 if (ret || parent == from)
1652 parent = parent->parent;
1660 * Iterate the full tree, calling @down when first entering a node and @up when
1661 * leaving it for the final time.
1663 * Caller must hold rcu_lock or sufficient equivalent.
1666 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1668 return walk_tg_tree_from(&root_task_group, down, up, data);
1671 static int tg_nop(struct task_group *tg, void *data)
1678 /* Used instead of source_load when we know the type == 0 */
1679 static unsigned long weighted_cpuload(const int cpu)
1681 return cpu_rq(cpu)->load.weight;
1685 * Return a low guess at the load of a migration-source cpu weighted
1686 * according to the scheduling class and "nice" value.
1688 * We want to under-estimate the load of migration sources, to
1689 * balance conservatively.
1691 static unsigned long source_load(int cpu, int type)
1693 struct rq *rq = cpu_rq(cpu);
1694 unsigned long total = weighted_cpuload(cpu);
1696 if (type == 0 || !sched_feat(LB_BIAS))
1699 return min(rq->cpu_load[type-1], total);
1703 * Return a high guess at the load of a migration-target cpu weighted
1704 * according to the scheduling class and "nice" value.
1706 static unsigned long target_load(int cpu, int type)
1708 struct rq *rq = cpu_rq(cpu);
1709 unsigned long total = weighted_cpuload(cpu);
1711 if (type == 0 || !sched_feat(LB_BIAS))
1714 return max(rq->cpu_load[type-1], total);
1717 static unsigned long power_of(int cpu)
1719 return cpu_rq(cpu)->cpu_power;
1722 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1724 static unsigned long cpu_avg_load_per_task(int cpu)
1726 struct rq *rq = cpu_rq(cpu);
1727 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1730 return rq->load.weight / nr_running;
1735 #ifdef CONFIG_PREEMPT
1737 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1740 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1741 * way at the expense of forcing extra atomic operations in all
1742 * invocations. This assures that the double_lock is acquired using the
1743 * same underlying policy as the spinlock_t on this architecture, which
1744 * reduces latency compared to the unfair variant below. However, it
1745 * also adds more overhead and therefore may reduce throughput.
1747 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1748 __releases(this_rq->lock)
1749 __acquires(busiest->lock)
1750 __acquires(this_rq->lock)
1752 raw_spin_unlock(&this_rq->lock);
1753 double_rq_lock(this_rq, busiest);
1760 * Unfair double_lock_balance: Optimizes throughput at the expense of
1761 * latency by eliminating extra atomic operations when the locks are
1762 * already in proper order on entry. This favors lower cpu-ids and will
1763 * grant the double lock to lower cpus over higher ids under contention,
1764 * regardless of entry order into the function.
1766 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1767 __releases(this_rq->lock)
1768 __acquires(busiest->lock)
1769 __acquires(this_rq->lock)
1773 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1774 if (busiest < this_rq) {
1775 raw_spin_unlock(&this_rq->lock);
1776 raw_spin_lock(&busiest->lock);
1777 raw_spin_lock_nested(&this_rq->lock,
1778 SINGLE_DEPTH_NESTING);
1781 raw_spin_lock_nested(&busiest->lock,
1782 SINGLE_DEPTH_NESTING);
1787 #endif /* CONFIG_PREEMPT */
1790 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1792 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1794 if (unlikely(!irqs_disabled())) {
1795 /* printk() doesn't work good under rq->lock */
1796 raw_spin_unlock(&this_rq->lock);
1800 return _double_lock_balance(this_rq, busiest);
1803 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1804 __releases(busiest->lock)
1806 raw_spin_unlock(&busiest->lock);
1807 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1811 * double_rq_lock - safely lock two runqueues
1813 * Note this does not disable interrupts like task_rq_lock,
1814 * you need to do so manually before calling.
1816 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1817 __acquires(rq1->lock)
1818 __acquires(rq2->lock)
1820 BUG_ON(!irqs_disabled());
1822 raw_spin_lock(&rq1->lock);
1823 __acquire(rq2->lock); /* Fake it out ;) */
1826 raw_spin_lock(&rq1->lock);
1827 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1829 raw_spin_lock(&rq2->lock);
1830 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1836 * double_rq_unlock - safely unlock two runqueues
1838 * Note this does not restore interrupts like task_rq_unlock,
1839 * you need to do so manually after calling.
1841 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1842 __releases(rq1->lock)
1843 __releases(rq2->lock)
1845 raw_spin_unlock(&rq1->lock);
1847 raw_spin_unlock(&rq2->lock);
1849 __release(rq2->lock);
1852 #else /* CONFIG_SMP */
1855 * double_rq_lock - safely lock two runqueues
1857 * Note this does not disable interrupts like task_rq_lock,
1858 * you need to do so manually before calling.
1860 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1861 __acquires(rq1->lock)
1862 __acquires(rq2->lock)
1864 BUG_ON(!irqs_disabled());
1866 raw_spin_lock(&rq1->lock);
1867 __acquire(rq2->lock); /* Fake it out ;) */
1871 * double_rq_unlock - safely unlock two runqueues
1873 * Note this does not restore interrupts like task_rq_unlock,
1874 * you need to do so manually after calling.
1876 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1877 __releases(rq1->lock)
1878 __releases(rq2->lock)
1881 raw_spin_unlock(&rq1->lock);
1882 __release(rq2->lock);
1887 static void update_sysctl(void);
1888 static int get_update_sysctl_factor(void);
1889 static void update_idle_cpu_load(struct rq *this_rq);
1891 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1893 set_task_rq(p, cpu);
1896 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1897 * successfully executed on another CPU. We must ensure that updates of
1898 * per-task data have been completed by this moment.
1901 task_thread_info(p)->cpu = cpu;
1905 static const struct sched_class rt_sched_class;
1907 #define sched_class_highest (&stop_sched_class)
1908 #define for_each_class(class) \
1909 for (class = sched_class_highest; class; class = class->next)
1911 #include "sched_stats.h"
1913 static void inc_nr_running(struct rq *rq)
1918 static void dec_nr_running(struct rq *rq)
1923 static void set_load_weight(struct task_struct *p)
1925 int prio = p->static_prio - MAX_RT_PRIO;
1926 struct load_weight *load = &p->se.load;
1929 * SCHED_IDLE tasks get minimal weight:
1931 if (p->policy == SCHED_IDLE) {
1932 load->weight = scale_load(WEIGHT_IDLEPRIO);
1933 load->inv_weight = WMULT_IDLEPRIO;
1937 load->weight = scale_load(prio_to_weight[prio]);
1938 load->inv_weight = prio_to_wmult[prio];
1941 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1943 update_rq_clock(rq);
1944 sched_info_queued(p);
1945 p->sched_class->enqueue_task(rq, p, flags);
1948 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1950 update_rq_clock(rq);
1951 sched_info_dequeued(p);
1952 p->sched_class->dequeue_task(rq, p, flags);
1956 * activate_task - move a task to the runqueue.
1958 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1960 if (task_contributes_to_load(p))
1961 rq->nr_uninterruptible--;
1963 enqueue_task(rq, p, flags);
1967 * deactivate_task - remove a task from the runqueue.
1969 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1971 if (task_contributes_to_load(p))
1972 rq->nr_uninterruptible++;
1974 dequeue_task(rq, p, flags);
1977 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1980 * There are no locks covering percpu hardirq/softirq time.
1981 * They are only modified in account_system_vtime, on corresponding CPU
1982 * with interrupts disabled. So, writes are safe.
1983 * They are read and saved off onto struct rq in update_rq_clock().
1984 * This may result in other CPU reading this CPU's irq time and can
1985 * race with irq/account_system_vtime on this CPU. We would either get old
1986 * or new value with a side effect of accounting a slice of irq time to wrong
1987 * task when irq is in progress while we read rq->clock. That is a worthy
1988 * compromise in place of having locks on each irq in account_system_time.
1990 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1991 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1993 static DEFINE_PER_CPU(u64, irq_start_time);
1994 static int sched_clock_irqtime;
1996 void enable_sched_clock_irqtime(void)
1998 sched_clock_irqtime = 1;
2001 void disable_sched_clock_irqtime(void)
2003 sched_clock_irqtime = 0;
2006 #ifndef CONFIG_64BIT
2007 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
2009 static inline void irq_time_write_begin(void)
2011 __this_cpu_inc(irq_time_seq.sequence);
2015 static inline void irq_time_write_end(void)
2018 __this_cpu_inc(irq_time_seq.sequence);
2021 static inline u64 irq_time_read(int cpu)
2027 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
2028 irq_time = per_cpu(cpu_softirq_time, cpu) +
2029 per_cpu(cpu_hardirq_time, cpu);
2030 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
2034 #else /* CONFIG_64BIT */
2035 static inline void irq_time_write_begin(void)
2039 static inline void irq_time_write_end(void)
2043 static inline u64 irq_time_read(int cpu)
2045 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
2047 #endif /* CONFIG_64BIT */
2050 * Called before incrementing preempt_count on {soft,}irq_enter
2051 * and before decrementing preempt_count on {soft,}irq_exit.
2053 void account_system_vtime(struct task_struct *curr)
2055 unsigned long flags;
2059 if (!sched_clock_irqtime)
2062 local_irq_save(flags);
2064 cpu = smp_processor_id();
2065 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2066 __this_cpu_add(irq_start_time, delta);
2068 irq_time_write_begin();
2070 * We do not account for softirq time from ksoftirqd here.
2071 * We want to continue accounting softirq time to ksoftirqd thread
2072 * in that case, so as not to confuse scheduler with a special task
2073 * that do not consume any time, but still wants to run.
2075 if (hardirq_count())
2076 __this_cpu_add(cpu_hardirq_time, delta);
2077 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
2078 __this_cpu_add(cpu_softirq_time, delta);
2080 irq_time_write_end();
2081 local_irq_restore(flags);
2083 EXPORT_SYMBOL_GPL(account_system_vtime);
2085 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2087 #ifdef CONFIG_PARAVIRT
2088 static inline u64 steal_ticks(u64 steal)
2090 if (unlikely(steal > NSEC_PER_SEC))
2091 return div_u64(steal, TICK_NSEC);
2093 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
2097 static void update_rq_clock_task(struct rq *rq, s64 delta)
2100 * In theory, the compile should just see 0 here, and optimize out the call
2101 * to sched_rt_avg_update. But I don't trust it...
2103 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2104 s64 steal = 0, irq_delta = 0;
2106 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2107 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2110 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2111 * this case when a previous update_rq_clock() happened inside a
2112 * {soft,}irq region.
2114 * When this happens, we stop ->clock_task and only update the
2115 * prev_irq_time stamp to account for the part that fit, so that a next
2116 * update will consume the rest. This ensures ->clock_task is
2119 * It does however cause some slight miss-attribution of {soft,}irq
2120 * time, a more accurate solution would be to update the irq_time using
2121 * the current rq->clock timestamp, except that would require using
2124 if (irq_delta > delta)
2127 rq->prev_irq_time += irq_delta;
2130 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2131 if (static_branch((¶virt_steal_rq_enabled))) {
2134 steal = paravirt_steal_clock(cpu_of(rq));
2135 steal -= rq->prev_steal_time_rq;
2137 if (unlikely(steal > delta))
2140 st = steal_ticks(steal);
2141 steal = st * TICK_NSEC;
2143 rq->prev_steal_time_rq += steal;
2149 rq->clock_task += delta;
2151 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2152 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2153 sched_rt_avg_update(rq, irq_delta + steal);
2157 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2158 static int irqtime_account_hi_update(void)
2160 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2161 unsigned long flags;
2165 local_irq_save(flags);
2166 latest_ns = this_cpu_read(cpu_hardirq_time);
2167 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2169 local_irq_restore(flags);
2173 static int irqtime_account_si_update(void)
2175 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2176 unsigned long flags;
2180 local_irq_save(flags);
2181 latest_ns = this_cpu_read(cpu_softirq_time);
2182 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2184 local_irq_restore(flags);
2188 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2190 #define sched_clock_irqtime (0)
2195 static void unthrottle_offline_cfs_rqs(struct rq *rq);
2198 #include "sched_idletask.c"
2199 #include "sched_fair.c"
2200 #include "sched_rt.c"
2201 #include "sched_autogroup.c"
2202 #include "sched_stoptask.c"
2203 #ifdef CONFIG_SCHED_DEBUG
2204 # include "sched_debug.c"
2207 void sched_set_stop_task(int cpu, struct task_struct *stop)
2209 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2210 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2214 * Make it appear like a SCHED_FIFO task, its something
2215 * userspace knows about and won't get confused about.
2217 * Also, it will make PI more or less work without too
2218 * much confusion -- but then, stop work should not
2219 * rely on PI working anyway.
2221 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2223 stop->sched_class = &stop_sched_class;
2226 cpu_rq(cpu)->stop = stop;
2230 * Reset it back to a normal scheduling class so that
2231 * it can die in pieces.
2233 old_stop->sched_class = &rt_sched_class;
2238 * __normal_prio - return the priority that is based on the static prio
2240 static inline int __normal_prio(struct task_struct *p)
2242 return p->static_prio;
2246 * Calculate the expected normal priority: i.e. priority
2247 * without taking RT-inheritance into account. Might be
2248 * boosted by interactivity modifiers. Changes upon fork,
2249 * setprio syscalls, and whenever the interactivity
2250 * estimator recalculates.
2252 static inline int normal_prio(struct task_struct *p)
2256 if (task_has_rt_policy(p))
2257 prio = MAX_RT_PRIO-1 - p->rt_priority;
2259 prio = __normal_prio(p);
2264 * Calculate the current priority, i.e. the priority
2265 * taken into account by the scheduler. This value might
2266 * be boosted by RT tasks, or might be boosted by
2267 * interactivity modifiers. Will be RT if the task got
2268 * RT-boosted. If not then it returns p->normal_prio.
2270 static int effective_prio(struct task_struct *p)
2272 p->normal_prio = normal_prio(p);
2274 * If we are RT tasks or we were boosted to RT priority,
2275 * keep the priority unchanged. Otherwise, update priority
2276 * to the normal priority:
2278 if (!rt_prio(p->prio))
2279 return p->normal_prio;
2284 * task_curr - is this task currently executing on a CPU?
2285 * @p: the task in question.
2287 inline int task_curr(const struct task_struct *p)
2289 return cpu_curr(task_cpu(p)) == p;
2292 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2293 const struct sched_class *prev_class,
2296 if (prev_class != p->sched_class) {
2297 if (prev_class->switched_from)
2298 prev_class->switched_from(rq, p);
2299 p->sched_class->switched_to(rq, p);
2300 } else if (oldprio != p->prio)
2301 p->sched_class->prio_changed(rq, p, oldprio);
2304 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2306 const struct sched_class *class;
2308 if (p->sched_class == rq->curr->sched_class) {
2309 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2311 for_each_class(class) {
2312 if (class == rq->curr->sched_class)
2314 if (class == p->sched_class) {
2315 resched_task(rq->curr);
2322 * A queue event has occurred, and we're going to schedule. In
2323 * this case, we can save a useless back to back clock update.
2325 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2326 rq->skip_clock_update = 1;
2331 * Is this task likely cache-hot:
2334 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2338 if (p->sched_class != &fair_sched_class)
2341 if (unlikely(p->policy == SCHED_IDLE))
2345 * Buddy candidates are cache hot:
2347 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2348 (&p->se == cfs_rq_of(&p->se)->next ||
2349 &p->se == cfs_rq_of(&p->se)->last))
2352 if (sysctl_sched_migration_cost == -1)
2354 if (sysctl_sched_migration_cost == 0)
2357 delta = now - p->se.exec_start;
2359 return delta < (s64)sysctl_sched_migration_cost;
2362 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2364 #ifdef CONFIG_SCHED_DEBUG
2366 * We should never call set_task_cpu() on a blocked task,
2367 * ttwu() will sort out the placement.
2369 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2370 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2372 #ifdef CONFIG_LOCKDEP
2374 * The caller should hold either p->pi_lock or rq->lock, when changing
2375 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2377 * sched_move_task() holds both and thus holding either pins the cgroup,
2380 * Furthermore, all task_rq users should acquire both locks, see
2383 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2384 lockdep_is_held(&task_rq(p)->lock)));
2388 trace_sched_migrate_task(p, new_cpu);
2390 if (task_cpu(p) != new_cpu) {
2391 p->se.nr_migrations++;
2392 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2395 __set_task_cpu(p, new_cpu);
2398 struct migration_arg {
2399 struct task_struct *task;
2403 static int migration_cpu_stop(void *data);
2406 * wait_task_inactive - wait for a thread to unschedule.
2408 * If @match_state is nonzero, it's the @p->state value just checked and
2409 * not expected to change. If it changes, i.e. @p might have woken up,
2410 * then return zero. When we succeed in waiting for @p to be off its CPU,
2411 * we return a positive number (its total switch count). If a second call
2412 * a short while later returns the same number, the caller can be sure that
2413 * @p has remained unscheduled the whole time.
2415 * The caller must ensure that the task *will* unschedule sometime soon,
2416 * else this function might spin for a *long* time. This function can't
2417 * be called with interrupts off, or it may introduce deadlock with
2418 * smp_call_function() if an IPI is sent by the same process we are
2419 * waiting to become inactive.
2421 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2423 unsigned long flags;
2430 * We do the initial early heuristics without holding
2431 * any task-queue locks at all. We'll only try to get
2432 * the runqueue lock when things look like they will
2438 * If the task is actively running on another CPU
2439 * still, just relax and busy-wait without holding
2442 * NOTE! Since we don't hold any locks, it's not
2443 * even sure that "rq" stays as the right runqueue!
2444 * But we don't care, since "task_running()" will
2445 * return false if the runqueue has changed and p
2446 * is actually now running somewhere else!
2448 while (task_running(rq, p)) {
2449 if (match_state && unlikely(p->state != match_state))
2455 * Ok, time to look more closely! We need the rq
2456 * lock now, to be *sure*. If we're wrong, we'll
2457 * just go back and repeat.
2459 rq = task_rq_lock(p, &flags);
2460 trace_sched_wait_task(p);
2461 running = task_running(rq, p);
2464 if (!match_state || p->state == match_state)
2465 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2466 task_rq_unlock(rq, p, &flags);
2469 * If it changed from the expected state, bail out now.
2471 if (unlikely(!ncsw))
2475 * Was it really running after all now that we
2476 * checked with the proper locks actually held?
2478 * Oops. Go back and try again..
2480 if (unlikely(running)) {
2486 * It's not enough that it's not actively running,
2487 * it must be off the runqueue _entirely_, and not
2490 * So if it was still runnable (but just not actively
2491 * running right now), it's preempted, and we should
2492 * yield - it could be a while.
2494 if (unlikely(on_rq)) {
2495 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2497 set_current_state(TASK_UNINTERRUPTIBLE);
2498 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2503 * Ahh, all good. It wasn't running, and it wasn't
2504 * runnable, which means that it will never become
2505 * running in the future either. We're all done!
2514 * kick_process - kick a running thread to enter/exit the kernel
2515 * @p: the to-be-kicked thread
2517 * Cause a process which is running on another CPU to enter
2518 * kernel-mode, without any delay. (to get signals handled.)
2520 * NOTE: this function doesn't have to take the runqueue lock,
2521 * because all it wants to ensure is that the remote task enters
2522 * the kernel. If the IPI races and the task has been migrated
2523 * to another CPU then no harm is done and the purpose has been
2526 void kick_process(struct task_struct *p)
2532 if ((cpu != smp_processor_id()) && task_curr(p))
2533 smp_send_reschedule(cpu);
2536 EXPORT_SYMBOL_GPL(kick_process);
2537 #endif /* CONFIG_SMP */
2541 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2543 static int select_fallback_rq(int cpu, struct task_struct *p)
2546 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2548 /* Look for allowed, online CPU in same node. */
2549 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2550 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
2553 /* Any allowed, online CPU? */
2554 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
2555 if (dest_cpu < nr_cpu_ids)
2558 /* No more Mr. Nice Guy. */
2559 dest_cpu = cpuset_cpus_allowed_fallback(p);
2561 * Don't tell them about moving exiting tasks or
2562 * kernel threads (both mm NULL), since they never
2565 if (p->mm && printk_ratelimit()) {
2566 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2567 task_pid_nr(p), p->comm, cpu);
2574 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2577 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2579 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2582 * In order not to call set_task_cpu() on a blocking task we need
2583 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2586 * Since this is common to all placement strategies, this lives here.
2588 * [ this allows ->select_task() to simply return task_cpu(p) and
2589 * not worry about this generic constraint ]
2591 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
2593 cpu = select_fallback_rq(task_cpu(p), p);
2598 static void update_avg(u64 *avg, u64 sample)
2600 s64 diff = sample - *avg;
2606 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2608 #ifdef CONFIG_SCHEDSTATS
2609 struct rq *rq = this_rq();
2612 int this_cpu = smp_processor_id();
2614 if (cpu == this_cpu) {
2615 schedstat_inc(rq, ttwu_local);
2616 schedstat_inc(p, se.statistics.nr_wakeups_local);
2618 struct sched_domain *sd;
2620 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2622 for_each_domain(this_cpu, sd) {
2623 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2624 schedstat_inc(sd, ttwu_wake_remote);
2631 if (wake_flags & WF_MIGRATED)
2632 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2634 #endif /* CONFIG_SMP */
2636 schedstat_inc(rq, ttwu_count);
2637 schedstat_inc(p, se.statistics.nr_wakeups);
2639 if (wake_flags & WF_SYNC)
2640 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2642 #endif /* CONFIG_SCHEDSTATS */
2645 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2647 activate_task(rq, p, en_flags);
2650 /* if a worker is waking up, notify workqueue */
2651 if (p->flags & PF_WQ_WORKER)
2652 wq_worker_waking_up(p, cpu_of(rq));
2656 * Mark the task runnable and perform wakeup-preemption.
2659 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2661 trace_sched_wakeup(p, true);
2662 check_preempt_curr(rq, p, wake_flags);
2664 p->state = TASK_RUNNING;
2666 if (p->sched_class->task_woken)
2667 p->sched_class->task_woken(rq, p);
2669 if (rq->idle_stamp) {
2670 u64 delta = rq->clock - rq->idle_stamp;
2671 u64 max = 2*sysctl_sched_migration_cost;
2676 update_avg(&rq->avg_idle, delta);
2683 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2686 if (p->sched_contributes_to_load)
2687 rq->nr_uninterruptible--;
2690 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2691 ttwu_do_wakeup(rq, p, wake_flags);
2695 * Called in case the task @p isn't fully descheduled from its runqueue,
2696 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2697 * since all we need to do is flip p->state to TASK_RUNNING, since
2698 * the task is still ->on_rq.
2700 static int ttwu_remote(struct task_struct *p, int wake_flags)
2705 rq = __task_rq_lock(p);
2707 ttwu_do_wakeup(rq, p, wake_flags);
2710 __task_rq_unlock(rq);
2716 static void sched_ttwu_pending(void)
2718 struct rq *rq = this_rq();
2719 struct llist_node *llist = llist_del_all(&rq->wake_list);
2720 struct task_struct *p;
2722 raw_spin_lock(&rq->lock);
2725 p = llist_entry(llist, struct task_struct, wake_entry);
2726 llist = llist_next(llist);
2727 ttwu_do_activate(rq, p, 0);
2730 raw_spin_unlock(&rq->lock);
2733 void scheduler_ipi(void)
2735 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2739 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2740 * traditionally all their work was done from the interrupt return
2741 * path. Now that we actually do some work, we need to make sure
2744 * Some archs already do call them, luckily irq_enter/exit nest
2747 * Arguably we should visit all archs and update all handlers,
2748 * however a fair share of IPIs are still resched only so this would
2749 * somewhat pessimize the simple resched case.
2752 sched_ttwu_pending();
2755 * Check if someone kicked us for doing the nohz idle load balance.
2757 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
2758 this_rq()->idle_balance = 1;
2759 raise_softirq_irqoff(SCHED_SOFTIRQ);
2764 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2766 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
2767 smp_send_reschedule(cpu);
2770 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2771 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2776 rq = __task_rq_lock(p);
2778 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2779 ttwu_do_wakeup(rq, p, wake_flags);
2782 __task_rq_unlock(rq);
2787 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2788 #endif /* CONFIG_SMP */
2790 static void ttwu_queue(struct task_struct *p, int cpu)
2792 struct rq *rq = cpu_rq(cpu);
2794 #if defined(CONFIG_SMP)
2795 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2796 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2797 ttwu_queue_remote(p, cpu);
2802 raw_spin_lock(&rq->lock);
2803 ttwu_do_activate(rq, p, 0);
2804 raw_spin_unlock(&rq->lock);
2808 * try_to_wake_up - wake up a thread
2809 * @p: the thread to be awakened
2810 * @state: the mask of task states that can be woken
2811 * @wake_flags: wake modifier flags (WF_*)
2813 * Put it on the run-queue if it's not already there. The "current"
2814 * thread is always on the run-queue (except when the actual
2815 * re-schedule is in progress), and as such you're allowed to do
2816 * the simpler "current->state = TASK_RUNNING" to mark yourself
2817 * runnable without the overhead of this.
2819 * Returns %true if @p was woken up, %false if it was already running
2820 * or @state didn't match @p's state.
2823 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2825 unsigned long flags;
2826 int cpu, success = 0;
2829 raw_spin_lock_irqsave(&p->pi_lock, flags);
2830 if (!(p->state & state))
2833 success = 1; /* we're going to change ->state */
2836 if (p->on_rq && ttwu_remote(p, wake_flags))
2841 * If the owning (remote) cpu is still in the middle of schedule() with
2842 * this task as prev, wait until its done referencing the task.
2845 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2847 * In case the architecture enables interrupts in
2848 * context_switch(), we cannot busy wait, since that
2849 * would lead to deadlocks when an interrupt hits and
2850 * tries to wake up @prev. So bail and do a complete
2853 if (ttwu_activate_remote(p, wake_flags))
2860 * Pairs with the smp_wmb() in finish_lock_switch().
2864 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2865 p->state = TASK_WAKING;
2867 if (p->sched_class->task_waking)
2868 p->sched_class->task_waking(p);
2870 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2871 if (task_cpu(p) != cpu) {
2872 wake_flags |= WF_MIGRATED;
2873 set_task_cpu(p, cpu);
2875 #endif /* CONFIG_SMP */
2879 ttwu_stat(p, cpu, wake_flags);
2881 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2887 * try_to_wake_up_local - try to wake up a local task with rq lock held
2888 * @p: the thread to be awakened
2890 * Put @p on the run-queue if it's not already there. The caller must
2891 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2894 static void try_to_wake_up_local(struct task_struct *p)
2896 struct rq *rq = task_rq(p);
2898 if (WARN_ON_ONCE(rq != this_rq()) ||
2899 WARN_ON_ONCE(p == current))
2902 lockdep_assert_held(&rq->lock);
2904 if (!raw_spin_trylock(&p->pi_lock)) {
2905 raw_spin_unlock(&rq->lock);
2906 raw_spin_lock(&p->pi_lock);
2907 raw_spin_lock(&rq->lock);
2910 if (!(p->state & TASK_NORMAL))
2914 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2916 ttwu_do_wakeup(rq, p, 0);
2917 ttwu_stat(p, smp_processor_id(), 0);
2919 raw_spin_unlock(&p->pi_lock);
2923 * wake_up_process - Wake up a specific process
2924 * @p: The process to be woken up.
2926 * Attempt to wake up the nominated process and move it to the set of runnable
2927 * processes. Returns 1 if the process was woken up, 0 if it was already
2930 * It may be assumed that this function implies a write memory barrier before
2931 * changing the task state if and only if any tasks are woken up.
2933 int wake_up_process(struct task_struct *p)
2935 return try_to_wake_up(p, TASK_NORMAL, 0);
2937 EXPORT_SYMBOL(wake_up_process);
2939 int wake_up_state(struct task_struct *p, unsigned int state)
2941 return try_to_wake_up(p, state, 0);
2945 * Perform scheduler related setup for a newly forked process p.
2946 * p is forked by current.
2948 * __sched_fork() is basic setup used by init_idle() too:
2950 static void __sched_fork(struct task_struct *p)
2955 p->se.exec_start = 0;
2956 p->se.sum_exec_runtime = 0;
2957 p->se.prev_sum_exec_runtime = 0;
2958 p->se.nr_migrations = 0;
2960 INIT_LIST_HEAD(&p->se.group_node);
2962 #ifdef CONFIG_SCHEDSTATS
2963 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2966 INIT_LIST_HEAD(&p->rt.run_list);
2968 #ifdef CONFIG_PREEMPT_NOTIFIERS
2969 INIT_HLIST_HEAD(&p->preempt_notifiers);
2974 * fork()/clone()-time setup:
2976 void sched_fork(struct task_struct *p)
2978 unsigned long flags;
2979 int cpu = get_cpu();
2983 * We mark the process as running here. This guarantees that
2984 * nobody will actually run it, and a signal or other external
2985 * event cannot wake it up and insert it on the runqueue either.
2987 p->state = TASK_RUNNING;
2990 * Make sure we do not leak PI boosting priority to the child.
2992 p->prio = current->normal_prio;
2995 * Revert to default priority/policy on fork if requested.
2997 if (unlikely(p->sched_reset_on_fork)) {
2998 if (task_has_rt_policy(p)) {
2999 p->policy = SCHED_NORMAL;
3000 p->static_prio = NICE_TO_PRIO(0);
3002 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3003 p->static_prio = NICE_TO_PRIO(0);
3005 p->prio = p->normal_prio = __normal_prio(p);
3009 * We don't need the reset flag anymore after the fork. It has
3010 * fulfilled its duty:
3012 p->sched_reset_on_fork = 0;
3015 if (!rt_prio(p->prio))
3016 p->sched_class = &fair_sched_class;
3018 if (p->sched_class->task_fork)
3019 p->sched_class->task_fork(p);
3022 * The child is not yet in the pid-hash so no cgroup attach races,
3023 * and the cgroup is pinned to this child due to cgroup_fork()
3024 * is ran before sched_fork().
3026 * Silence PROVE_RCU.
3028 raw_spin_lock_irqsave(&p->pi_lock, flags);
3029 set_task_cpu(p, cpu);
3030 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3032 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3033 if (likely(sched_info_on()))
3034 memset(&p->sched_info, 0, sizeof(p->sched_info));
3036 #if defined(CONFIG_SMP)
3039 #ifdef CONFIG_PREEMPT_COUNT
3040 /* Want to start with kernel preemption disabled. */
3041 task_thread_info(p)->preempt_count = 1;
3044 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3051 * wake_up_new_task - wake up a newly created task for the first time.
3053 * This function will do some initial scheduler statistics housekeeping
3054 * that must be done for every newly created context, then puts the task
3055 * on the runqueue and wakes it.
3057 void wake_up_new_task(struct task_struct *p)
3059 unsigned long flags;
3062 raw_spin_lock_irqsave(&p->pi_lock, flags);
3065 * Fork balancing, do it here and not earlier because:
3066 * - cpus_allowed can change in the fork path
3067 * - any previously selected cpu might disappear through hotplug
3069 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3072 rq = __task_rq_lock(p);
3073 activate_task(rq, p, 0);
3075 trace_sched_wakeup_new(p, true);
3076 check_preempt_curr(rq, p, WF_FORK);
3078 if (p->sched_class->task_woken)
3079 p->sched_class->task_woken(rq, p);
3081 task_rq_unlock(rq, p, &flags);
3084 #ifdef CONFIG_PREEMPT_NOTIFIERS
3087 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3088 * @notifier: notifier struct to register
3090 void preempt_notifier_register(struct preempt_notifier *notifier)
3092 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3094 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3097 * preempt_notifier_unregister - no longer interested in preemption notifications
3098 * @notifier: notifier struct to unregister
3100 * This is safe to call from within a preemption notifier.
3102 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3104 hlist_del(¬ifier->link);
3106 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3108 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3110 struct preempt_notifier *notifier;
3111 struct hlist_node *node;
3113 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3114 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3118 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3119 struct task_struct *next)
3121 struct preempt_notifier *notifier;
3122 struct hlist_node *node;
3124 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3125 notifier->ops->sched_out(notifier, next);
3128 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3130 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3135 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3136 struct task_struct *next)
3140 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3143 * prepare_task_switch - prepare to switch tasks
3144 * @rq: the runqueue preparing to switch
3145 * @prev: the current task that is being switched out
3146 * @next: the task we are going to switch to.
3148 * This is called with the rq lock held and interrupts off. It must
3149 * be paired with a subsequent finish_task_switch after the context
3152 * prepare_task_switch sets up locking and calls architecture specific
3156 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3157 struct task_struct *next)
3159 sched_info_switch(prev, next);
3160 perf_event_task_sched_out(prev, next);
3161 fire_sched_out_preempt_notifiers(prev, next);
3162 prepare_lock_switch(rq, next);
3163 prepare_arch_switch(next);
3164 trace_sched_switch(prev, next);
3168 * finish_task_switch - clean up after a task-switch
3169 * @rq: runqueue associated with task-switch
3170 * @prev: the thread we just switched away from.
3172 * finish_task_switch must be called after the context switch, paired
3173 * with a prepare_task_switch call before the context switch.
3174 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3175 * and do any other architecture-specific cleanup actions.
3177 * Note that we may have delayed dropping an mm in context_switch(). If
3178 * so, we finish that here outside of the runqueue lock. (Doing it
3179 * with the lock held can cause deadlocks; see schedule() for
3182 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3183 __releases(rq->lock)
3185 struct mm_struct *mm = rq->prev_mm;
3191 * A task struct has one reference for the use as "current".
3192 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3193 * schedule one last time. The schedule call will never return, and
3194 * the scheduled task must drop that reference.
3196 * We must observe prev->state before clearing prev->on_cpu (in
3197 * finish_lock_switch), otherwise a concurrent wakeup can get prev
3198 * running on another CPU and we could rave with its RUNNING -> DEAD
3199 * transition, resulting in a double drop.
3201 prev_state = prev->state;
3202 finish_arch_switch(prev);
3203 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3204 local_irq_disable();
3205 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3206 perf_event_task_sched_in(prev, current);
3207 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3209 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3210 finish_lock_switch(rq, prev);
3212 fire_sched_in_preempt_notifiers(current);
3215 if (unlikely(prev_state == TASK_DEAD)) {
3217 * Remove function-return probe instances associated with this
3218 * task and put them back on the free list.
3220 kprobe_flush_task(prev);
3221 put_task_struct(prev);
3227 /* assumes rq->lock is held */
3228 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3230 if (prev->sched_class->pre_schedule)
3231 prev->sched_class->pre_schedule(rq, prev);
3234 /* rq->lock is NOT held, but preemption is disabled */
3235 static inline void post_schedule(struct rq *rq)
3237 if (rq->post_schedule) {
3238 unsigned long flags;
3240 raw_spin_lock_irqsave(&rq->lock, flags);
3241 if (rq->curr->sched_class->post_schedule)
3242 rq->curr->sched_class->post_schedule(rq);
3243 raw_spin_unlock_irqrestore(&rq->lock, flags);
3245 rq->post_schedule = 0;
3251 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3255 static inline void post_schedule(struct rq *rq)
3262 * schedule_tail - first thing a freshly forked thread must call.
3263 * @prev: the thread we just switched away from.
3265 asmlinkage void schedule_tail(struct task_struct *prev)
3266 __releases(rq->lock)
3268 struct rq *rq = this_rq();
3270 finish_task_switch(rq, prev);
3273 * FIXME: do we need to worry about rq being invalidated by the
3278 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3279 /* In this case, finish_task_switch does not reenable preemption */
3282 if (current->set_child_tid)
3283 put_user(task_pid_vnr(current), current->set_child_tid);
3287 * context_switch - switch to the new MM and the new
3288 * thread's register state.
3291 context_switch(struct rq *rq, struct task_struct *prev,
3292 struct task_struct *next)
3294 struct mm_struct *mm, *oldmm;
3296 prepare_task_switch(rq, prev, next);
3299 oldmm = prev->active_mm;
3301 * For paravirt, this is coupled with an exit in switch_to to
3302 * combine the page table reload and the switch backend into
3305 arch_start_context_switch(prev);
3308 next->active_mm = oldmm;
3309 atomic_inc(&oldmm->mm_count);
3310 enter_lazy_tlb(oldmm, next);
3312 switch_mm(oldmm, mm, next);
3315 prev->active_mm = NULL;
3316 rq->prev_mm = oldmm;
3319 * Since the runqueue lock will be released by the next
3320 * task (which is an invalid locking op but in the case
3321 * of the scheduler it's an obvious special-case), so we
3322 * do an early lockdep release here:
3324 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3325 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3328 /* Here we just switch the register state and the stack. */
3329 switch_to(prev, next, prev);
3333 * this_rq must be evaluated again because prev may have moved
3334 * CPUs since it called schedule(), thus the 'rq' on its stack
3335 * frame will be invalid.
3337 finish_task_switch(this_rq(), prev);
3341 * nr_running, nr_uninterruptible and nr_context_switches:
3343 * externally visible scheduler statistics: current number of runnable
3344 * threads, current number of uninterruptible-sleeping threads, total
3345 * number of context switches performed since bootup.
3347 unsigned long nr_running(void)
3349 unsigned long i, sum = 0;
3351 for_each_online_cpu(i)
3352 sum += cpu_rq(i)->nr_running;
3357 unsigned long nr_uninterruptible(void)
3359 unsigned long i, sum = 0;
3361 for_each_possible_cpu(i)
3362 sum += cpu_rq(i)->nr_uninterruptible;
3365 * Since we read the counters lockless, it might be slightly
3366 * inaccurate. Do not allow it to go below zero though:
3368 if (unlikely((long)sum < 0))
3374 unsigned long long nr_context_switches(void)
3377 unsigned long long sum = 0;
3379 for_each_possible_cpu(i)
3380 sum += cpu_rq(i)->nr_switches;
3385 unsigned long nr_iowait(void)
3387 unsigned long i, sum = 0;
3389 for_each_possible_cpu(i)
3390 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3395 unsigned long nr_iowait_cpu(int cpu)
3397 struct rq *this = cpu_rq(cpu);
3398 return atomic_read(&this->nr_iowait);
3401 unsigned long this_cpu_load(void)
3403 struct rq *this = this_rq();
3404 return this->cpu_load[0];
3409 * Global load-average calculations
3411 * We take a distributed and async approach to calculating the global load-avg
3412 * in order to minimize overhead.
3414 * The global load average is an exponentially decaying average of nr_running +
3415 * nr_uninterruptible.
3417 * Once every LOAD_FREQ:
3420 * for_each_possible_cpu(cpu)
3421 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
3423 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
3425 * Due to a number of reasons the above turns in the mess below:
3427 * - for_each_possible_cpu() is prohibitively expensive on machines with
3428 * serious number of cpus, therefore we need to take a distributed approach
3429 * to calculating nr_active.
3431 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
3432 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
3434 * So assuming nr_active := 0 when we start out -- true per definition, we
3435 * can simply take per-cpu deltas and fold those into a global accumulate
3436 * to obtain the same result. See calc_load_fold_active().
3438 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
3439 * across the machine, we assume 10 ticks is sufficient time for every
3440 * cpu to have completed this task.
3442 * This places an upper-bound on the IRQ-off latency of the machine. Then
3443 * again, being late doesn't loose the delta, just wrecks the sample.
3445 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
3446 * this would add another cross-cpu cacheline miss and atomic operation
3447 * to the wakeup path. Instead we increment on whatever cpu the task ran
3448 * when it went into uninterruptible state and decrement on whatever cpu
3449 * did the wakeup. This means that only the sum of nr_uninterruptible over
3450 * all cpus yields the correct result.
3452 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
3455 /* Variables and functions for calc_load */
3456 static atomic_long_t calc_load_tasks;
3457 static unsigned long calc_load_update;
3458 unsigned long avenrun[3];
3459 EXPORT_SYMBOL(avenrun); /* should be removed */
3462 * get_avenrun - get the load average array
3463 * @loads: pointer to dest load array
3464 * @offset: offset to add
3465 * @shift: shift count to shift the result left
3467 * These values are estimates at best, so no need for locking.
3469 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3471 loads[0] = (avenrun[0] + offset) << shift;
3472 loads[1] = (avenrun[1] + offset) << shift;
3473 loads[2] = (avenrun[2] + offset) << shift;
3476 static long calc_load_fold_active(struct rq *this_rq)
3478 long nr_active, delta = 0;
3480 nr_active = this_rq->nr_running;
3481 nr_active += (long) this_rq->nr_uninterruptible;
3483 if (nr_active != this_rq->calc_load_active) {
3484 delta = nr_active - this_rq->calc_load_active;
3485 this_rq->calc_load_active = nr_active;
3492 * a1 = a0 * e + a * (1 - e)
3494 static unsigned long
3495 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3498 load += active * (FIXED_1 - exp);
3499 load += 1UL << (FSHIFT - 1);
3500 return load >> FSHIFT;
3505 * Handle NO_HZ for the global load-average.
3507 * Since the above described distributed algorithm to compute the global
3508 * load-average relies on per-cpu sampling from the tick, it is affected by
3511 * The basic idea is to fold the nr_active delta into a global idle-delta upon
3512 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
3513 * when we read the global state.
3515 * Obviously reality has to ruin such a delightfully simple scheme:
3517 * - When we go NO_HZ idle during the window, we can negate our sample
3518 * contribution, causing under-accounting.
3520 * We avoid this by keeping two idle-delta counters and flipping them
3521 * when the window starts, thus separating old and new NO_HZ load.
3523 * The only trick is the slight shift in index flip for read vs write.
3527 * |-|-----------|-|-----------|-|-----------|-|
3528 * r:0 0 1 1 0 0 1 1 0
3529 * w:0 1 1 0 0 1 1 0 0
3531 * This ensures we'll fold the old idle contribution in this window while
3532 * accumlating the new one.
3534 * - When we wake up from NO_HZ idle during the window, we push up our
3535 * contribution, since we effectively move our sample point to a known
3538 * This is solved by pushing the window forward, and thus skipping the
3539 * sample, for this cpu (effectively using the idle-delta for this cpu which
3540 * was in effect at the time the window opened). This also solves the issue
3541 * of having to deal with a cpu having been in NOHZ idle for multiple
3542 * LOAD_FREQ intervals.
3544 * When making the ILB scale, we should try to pull this in as well.
3546 static atomic_long_t calc_load_idle[2];
3547 static int calc_load_idx;
3549 static inline int calc_load_write_idx(void)
3551 int idx = calc_load_idx;
3554 * See calc_global_nohz(), if we observe the new index, we also
3555 * need to observe the new update time.
3560 * If the folding window started, make sure we start writing in the
3563 if (!time_before(jiffies, calc_load_update))
3569 static inline int calc_load_read_idx(void)
3571 return calc_load_idx & 1;
3574 void calc_load_enter_idle(void)
3576 struct rq *this_rq = this_rq();
3580 * We're going into NOHZ mode, if there's any pending delta, fold it
3581 * into the pending idle delta.
3583 delta = calc_load_fold_active(this_rq);
3585 int idx = calc_load_write_idx();
3586 atomic_long_add(delta, &calc_load_idle[idx]);
3590 void calc_load_exit_idle(void)
3592 struct rq *this_rq = this_rq();
3595 * If we're still before the sample window, we're done.
3597 if (time_before(jiffies, this_rq->calc_load_update))
3601 * We woke inside or after the sample window, this means we're already
3602 * accounted through the nohz accounting, so skip the entire deal and
3603 * sync up for the next window.
3605 this_rq->calc_load_update = calc_load_update;
3606 if (time_before(jiffies, this_rq->calc_load_update + 10))
3607 this_rq->calc_load_update += LOAD_FREQ;
3610 static long calc_load_fold_idle(void)
3612 int idx = calc_load_read_idx();
3615 if (atomic_long_read(&calc_load_idle[idx]))
3616 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
3622 * fixed_power_int - compute: x^n, in O(log n) time
3624 * @x: base of the power
3625 * @frac_bits: fractional bits of @x
3626 * @n: power to raise @x to.
3628 * By exploiting the relation between the definition of the natural power
3629 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3630 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3631 * (where: n_i \elem {0, 1}, the binary vector representing n),
3632 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3633 * of course trivially computable in O(log_2 n), the length of our binary
3636 static unsigned long
3637 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3639 unsigned long result = 1UL << frac_bits;
3644 result += 1UL << (frac_bits - 1);
3645 result >>= frac_bits;
3651 x += 1UL << (frac_bits - 1);
3659 * a1 = a0 * e + a * (1 - e)
3661 * a2 = a1 * e + a * (1 - e)
3662 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3663 * = a0 * e^2 + a * (1 - e) * (1 + e)
3665 * a3 = a2 * e + a * (1 - e)
3666 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3667 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3671 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3672 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3673 * = a0 * e^n + a * (1 - e^n)
3675 * [1] application of the geometric series:
3678 * S_n := \Sum x^i = -------------
3681 static unsigned long
3682 calc_load_n(unsigned long load, unsigned long exp,
3683 unsigned long active, unsigned int n)
3686 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3690 * NO_HZ can leave us missing all per-cpu ticks calling
3691 * calc_load_account_active(), but since an idle CPU folds its delta into
3692 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3693 * in the pending idle delta if our idle period crossed a load cycle boundary.
3695 * Once we've updated the global active value, we need to apply the exponential
3696 * weights adjusted to the number of cycles missed.
3698 static void calc_global_nohz(void)
3700 long delta, active, n;
3702 if (!time_before(jiffies, calc_load_update + 10)) {
3704 * Catch-up, fold however many we are behind still
3706 delta = jiffies - calc_load_update - 10;
3707 n = 1 + (delta / LOAD_FREQ);
3709 active = atomic_long_read(&calc_load_tasks);
3710 active = active > 0 ? active * FIXED_1 : 0;
3712 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3713 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3714 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3716 calc_load_update += n * LOAD_FREQ;
3720 * Flip the idle index...
3722 * Make sure we first write the new time then flip the index, so that
3723 * calc_load_write_idx() will see the new time when it reads the new
3724 * index, this avoids a double flip messing things up.
3729 #else /* !CONFIG_NO_HZ */
3731 static inline long calc_load_fold_idle(void) { return 0; }
3732 static inline void calc_global_nohz(void) { }
3734 #endif /* CONFIG_NO_HZ */
3737 * calc_load - update the avenrun load estimates 10 ticks after the
3738 * CPUs have updated calc_load_tasks.
3740 void calc_global_load(unsigned long ticks)
3744 if (time_before(jiffies, calc_load_update + 10))
3748 * Fold the 'old' idle-delta to include all NO_HZ cpus.
3750 delta = calc_load_fold_idle();
3752 atomic_long_add(delta, &calc_load_tasks);
3754 active = atomic_long_read(&calc_load_tasks);
3755 active = active > 0 ? active * FIXED_1 : 0;
3757 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3758 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3759 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3761 calc_load_update += LOAD_FREQ;
3764 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
3770 * Called from update_cpu_load() to periodically update this CPU's
3773 static void calc_load_account_active(struct rq *this_rq)
3777 if (time_before(jiffies, this_rq->calc_load_update))
3780 delta = calc_load_fold_active(this_rq);
3782 atomic_long_add(delta, &calc_load_tasks);
3784 this_rq->calc_load_update += LOAD_FREQ;
3788 * End of global load-average stuff
3792 * The exact cpuload at various idx values, calculated at every tick would be
3793 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3795 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3796 * on nth tick when cpu may be busy, then we have:
3797 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3798 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3800 * decay_load_missed() below does efficient calculation of
3801 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3802 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3804 * The calculation is approximated on a 128 point scale.
3805 * degrade_zero_ticks is the number of ticks after which load at any
3806 * particular idx is approximated to be zero.
3807 * degrade_factor is a precomputed table, a row for each load idx.
3808 * Each column corresponds to degradation factor for a power of two ticks,
3809 * based on 128 point scale.
3811 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3812 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3814 * With this power of 2 load factors, we can degrade the load n times
3815 * by looking at 1 bits in n and doing as many mult/shift instead of
3816 * n mult/shifts needed by the exact degradation.
3818 #define DEGRADE_SHIFT 7
3819 static const unsigned char
3820 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3821 static const unsigned char
3822 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3823 {0, 0, 0, 0, 0, 0, 0, 0},
3824 {64, 32, 8, 0, 0, 0, 0, 0},
3825 {96, 72, 40, 12, 1, 0, 0},
3826 {112, 98, 75, 43, 15, 1, 0},
3827 {120, 112, 98, 76, 45, 16, 2} };
3830 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3831 * would be when CPU is idle and so we just decay the old load without
3832 * adding any new load.
3834 static unsigned long
3835 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3839 if (!missed_updates)
3842 if (missed_updates >= degrade_zero_ticks[idx])
3846 return load >> missed_updates;
3848 while (missed_updates) {
3849 if (missed_updates % 2)
3850 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3852 missed_updates >>= 1;
3859 * Update rq->cpu_load[] statistics. This function is usually called every
3860 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3861 * every tick. We fix it up based on jiffies.
3863 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
3864 unsigned long pending_updates)
3868 this_rq->nr_load_updates++;
3870 /* Update our load: */
3871 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3872 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3873 unsigned long old_load, new_load;
3875 /* scale is effectively 1 << i now, and >> i divides by scale */
3877 old_load = this_rq->cpu_load[i];
3878 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3879 new_load = this_load;
3881 * Round up the averaging division if load is increasing. This
3882 * prevents us from getting stuck on 9 if the load is 10, for
3885 if (new_load > old_load)
3886 new_load += scale - 1;
3888 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3891 sched_avg_update(this_rq);
3896 * There is no sane way to deal with nohz on smp when using jiffies because the
3897 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
3898 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
3900 * Therefore we cannot use the delta approach from the regular tick since that
3901 * would seriously skew the load calculation. However we'll make do for those
3902 * updates happening while idle (nohz_idle_balance) or coming out of idle
3903 * (tick_nohz_idle_exit).
3905 * This means we might still be one tick off for nohz periods.
3909 * Called from nohz_idle_balance() to update the load ratings before doing the
3912 static void update_idle_cpu_load(struct rq *this_rq)
3914 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3915 unsigned long load = this_rq->load.weight;
3916 unsigned long pending_updates;
3919 * bail if there's load or we're actually up-to-date.
3921 if (load || curr_jiffies == this_rq->last_load_update_tick)
3924 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3925 this_rq->last_load_update_tick = curr_jiffies;
3927 __update_cpu_load(this_rq, load, pending_updates);
3931 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
3933 void update_cpu_load_nohz(void)
3935 struct rq *this_rq = this_rq();
3936 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3937 unsigned long pending_updates;
3939 if (curr_jiffies == this_rq->last_load_update_tick)
3942 raw_spin_lock(&this_rq->lock);
3943 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3944 if (pending_updates) {
3945 this_rq->last_load_update_tick = curr_jiffies;
3947 * We were idle, this means load 0, the current load might be
3948 * !0 due to remote wakeups and the sort.
3950 __update_cpu_load(this_rq, 0, pending_updates);
3952 raw_spin_unlock(&this_rq->lock);
3954 #endif /* CONFIG_NO_HZ */
3957 * Called from scheduler_tick()
3959 static void update_cpu_load_active(struct rq *this_rq)
3962 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
3964 this_rq->last_load_update_tick = jiffies;
3965 __update_cpu_load(this_rq, this_rq->load.weight, 1);
3967 calc_load_account_active(this_rq);
3973 * sched_exec - execve() is a valuable balancing opportunity, because at
3974 * this point the task has the smallest effective memory and cache footprint.
3976 void sched_exec(void)
3978 struct task_struct *p = current;
3979 unsigned long flags;
3982 raw_spin_lock_irqsave(&p->pi_lock, flags);
3983 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3984 if (dest_cpu == smp_processor_id())
3987 if (likely(cpu_active(dest_cpu))) {
3988 struct migration_arg arg = { p, dest_cpu };
3990 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3991 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3995 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4000 DEFINE_PER_CPU(struct kernel_stat, kstat);
4002 EXPORT_PER_CPU_SYMBOL(kstat);
4005 * Return any ns on the sched_clock that have not yet been accounted in
4006 * @p in case that task is currently running.
4008 * Called with task_rq_lock() held on @rq.
4010 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4014 if (task_current(rq, p)) {
4015 update_rq_clock(rq);
4016 ns = rq->clock_task - p->se.exec_start;
4024 unsigned long long task_delta_exec(struct task_struct *p)
4026 unsigned long flags;
4030 rq = task_rq_lock(p, &flags);
4031 ns = do_task_delta_exec(p, rq);
4032 task_rq_unlock(rq, p, &flags);
4038 * Return accounted runtime for the task.
4039 * In case the task is currently running, return the runtime plus current's
4040 * pending runtime that have not been accounted yet.
4042 unsigned long long task_sched_runtime(struct task_struct *p)
4044 unsigned long flags;
4048 rq = task_rq_lock(p, &flags);
4049 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4050 task_rq_unlock(rq, p, &flags);
4056 * Account user cpu time to a process.
4057 * @p: the process that the cpu time gets accounted to
4058 * @cputime: the cpu time spent in user space since the last update
4059 * @cputime_scaled: cputime scaled by cpu frequency
4061 void account_user_time(struct task_struct *p, cputime_t cputime,
4062 cputime_t cputime_scaled)
4064 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4067 /* Add user time to process. */
4068 p->utime = cputime_add(p->utime, cputime);
4069 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4070 account_group_user_time(p, cputime);
4072 /* Add user time to cpustat. */
4073 tmp = cputime_to_cputime64(cputime);
4074 if (TASK_NICE(p) > 0)
4075 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4077 cpustat->user = cputime64_add(cpustat->user, tmp);
4079 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4080 /* Account for user time used */
4081 acct_update_integrals(p);
4085 * Account guest cpu time to a process.
4086 * @p: the process that the cpu time gets accounted to
4087 * @cputime: the cpu time spent in virtual machine since the last update
4088 * @cputime_scaled: cputime scaled by cpu frequency
4090 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4091 cputime_t cputime_scaled)
4094 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4096 tmp = cputime_to_cputime64(cputime);
4098 /* Add guest time to process. */
4099 p->utime = cputime_add(p->utime, cputime);
4100 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4101 account_group_user_time(p, cputime);
4102 p->gtime = cputime_add(p->gtime, cputime);
4104 /* Add guest time to cpustat. */
4105 if (TASK_NICE(p) > 0) {
4106 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4107 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
4109 cpustat->user = cputime64_add(cpustat->user, tmp);
4110 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4115 * Account system cpu time to a process and desired cpustat field
4116 * @p: the process that the cpu time gets accounted to
4117 * @cputime: the cpu time spent in kernel space since the last update
4118 * @cputime_scaled: cputime scaled by cpu frequency
4119 * @target_cputime64: pointer to cpustat field that has to be updated
4122 void __account_system_time(struct task_struct *p, cputime_t cputime,
4123 cputime_t cputime_scaled, cputime64_t *target_cputime64)
4125 cputime64_t tmp = cputime_to_cputime64(cputime);
4127 /* Add system time to process. */
4128 p->stime = cputime_add(p->stime, cputime);
4129 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4130 account_group_system_time(p, cputime);
4132 /* Add system time to cpustat. */
4133 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
4134 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4136 /* Account for system time used */
4137 acct_update_integrals(p);
4141 * Account system cpu time to a process.
4142 * @p: the process that the cpu time gets accounted to
4143 * @hardirq_offset: the offset to subtract from hardirq_count()
4144 * @cputime: the cpu time spent in kernel space since the last update
4145 * @cputime_scaled: cputime scaled by cpu frequency
4147 void account_system_time(struct task_struct *p, int hardirq_offset,
4148 cputime_t cputime, cputime_t cputime_scaled)
4150 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4151 cputime64_t *target_cputime64;
4153 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4154 account_guest_time(p, cputime, cputime_scaled);
4158 if (hardirq_count() - hardirq_offset)
4159 target_cputime64 = &cpustat->irq;
4160 else if (in_serving_softirq())
4161 target_cputime64 = &cpustat->softirq;
4163 target_cputime64 = &cpustat->system;
4165 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
4169 * Account for involuntary wait time.
4170 * @cputime: the cpu time spent in involuntary wait
4172 void account_steal_time(cputime_t cputime)
4174 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4175 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4177 cpustat->steal = cputime64_add(cpustat->steal, cputime64);