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 use task_subsys_state_check() and extend the RCU verification with
750 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
751 * task it moves into the cgroup. Therefore by holding either of those locks,
752 * we pin the task to the current cgroup.
754 static inline struct task_group *task_group(struct task_struct *p)
756 struct task_group *tg;
757 struct cgroup_subsys_state *css;
759 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
760 lockdep_is_held(&p->pi_lock) ||
761 lockdep_is_held(&task_rq(p)->lock));
762 tg = container_of(css, struct task_group, css);
764 return autogroup_task_group(p, tg);
767 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
768 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
770 #ifdef CONFIG_FAIR_GROUP_SCHED
771 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
772 p->se.parent = task_group(p)->se[cpu];
775 #ifdef CONFIG_RT_GROUP_SCHED
776 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
777 p->rt.parent = task_group(p)->rt_se[cpu];
781 #else /* CONFIG_CGROUP_SCHED */
783 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
784 static inline struct task_group *task_group(struct task_struct *p)
789 #endif /* CONFIG_CGROUP_SCHED */
791 static void update_rq_clock_task(struct rq *rq, s64 delta);
793 static void update_rq_clock(struct rq *rq)
797 if (rq->skip_clock_update > 0)
800 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
802 update_rq_clock_task(rq, delta);
806 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
808 #ifdef CONFIG_SCHED_DEBUG
809 # define const_debug __read_mostly
811 # define const_debug static const
815 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
816 * @cpu: the processor in question.
818 * This interface allows printk to be called with the runqueue lock
819 * held and know whether or not it is OK to wake up the klogd.
821 int runqueue_is_locked(int cpu)
823 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
827 * Debugging: various feature bits
830 #define SCHED_FEAT(name, enabled) \
831 __SCHED_FEAT_##name ,
834 #include "sched_features.h"
839 #define SCHED_FEAT(name, enabled) \
840 (1UL << __SCHED_FEAT_##name) * enabled |
842 const_debug unsigned int sysctl_sched_features =
843 #include "sched_features.h"
848 #ifdef CONFIG_SCHED_DEBUG
849 #define SCHED_FEAT(name, enabled) \
852 static __read_mostly char *sched_feat_names[] = {
853 #include "sched_features.h"
859 static int sched_feat_show(struct seq_file *m, void *v)
863 for (i = 0; sched_feat_names[i]; i++) {
864 if (!(sysctl_sched_features & (1UL << i)))
866 seq_printf(m, "%s ", sched_feat_names[i]);
874 sched_feat_write(struct file *filp, const char __user *ubuf,
875 size_t cnt, loff_t *ppos)
885 if (copy_from_user(&buf, ubuf, cnt))
891 if (strncmp(cmp, "NO_", 3) == 0) {
896 for (i = 0; sched_feat_names[i]; i++) {
897 if (strcmp(cmp, sched_feat_names[i]) == 0) {
899 sysctl_sched_features &= ~(1UL << i);
901 sysctl_sched_features |= (1UL << i);
906 if (!sched_feat_names[i])
914 static int sched_feat_open(struct inode *inode, struct file *filp)
916 return single_open(filp, sched_feat_show, NULL);
919 static const struct file_operations sched_feat_fops = {
920 .open = sched_feat_open,
921 .write = sched_feat_write,
924 .release = single_release,
927 static __init int sched_init_debug(void)
929 debugfs_create_file("sched_features", 0644, NULL, NULL,
934 late_initcall(sched_init_debug);
938 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
941 * Number of tasks to iterate in a single balance run.
942 * Limited because this is done with IRQs disabled.
944 const_debug unsigned int sysctl_sched_nr_migrate = 32;
947 * period over which we average the RT time consumption, measured
952 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
955 * period over which we measure -rt task cpu usage in us.
958 unsigned int sysctl_sched_rt_period = 1000000;
960 static __read_mostly int scheduler_running;
963 * part of the period that we allow rt tasks to run in us.
966 int sysctl_sched_rt_runtime = 950000;
968 static inline u64 global_rt_period(void)
970 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
973 static inline u64 global_rt_runtime(void)
975 if (sysctl_sched_rt_runtime < 0)
978 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
981 #ifndef prepare_arch_switch
982 # define prepare_arch_switch(next) do { } while (0)
984 #ifndef finish_arch_switch
985 # define finish_arch_switch(prev) do { } while (0)
988 static inline int task_current(struct rq *rq, struct task_struct *p)
990 return rq->curr == p;
993 static inline int task_running(struct rq *rq, struct task_struct *p)
998 return task_current(rq, p);
1002 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1003 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1007 * We can optimise this out completely for !SMP, because the
1008 * SMP rebalancing from interrupt is the only thing that cares
1015 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1019 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1020 * We must ensure this doesn't happen until the switch is completely
1026 #ifdef CONFIG_DEBUG_SPINLOCK
1027 /* this is a valid case when another task releases the spinlock */
1028 rq->lock.owner = current;
1031 * If we are tracking spinlock dependencies then we have to
1032 * fix up the runqueue lock - which gets 'carried over' from
1033 * prev into current:
1035 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1037 raw_spin_unlock_irq(&rq->lock);
1040 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1041 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1045 * We can optimise this out completely for !SMP, because the
1046 * SMP rebalancing from interrupt is the only thing that cares
1051 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1052 raw_spin_unlock_irq(&rq->lock);
1054 raw_spin_unlock(&rq->lock);
1058 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1062 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1063 * We must ensure this doesn't happen until the switch is completely
1069 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1073 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1076 * __task_rq_lock - lock the rq @p resides on.
1078 static inline struct rq *__task_rq_lock(struct task_struct *p)
1079 __acquires(rq->lock)
1083 lockdep_assert_held(&p->pi_lock);
1087 raw_spin_lock(&rq->lock);
1088 if (likely(rq == task_rq(p)))
1090 raw_spin_unlock(&rq->lock);
1095 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1097 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1098 __acquires(p->pi_lock)
1099 __acquires(rq->lock)
1104 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1106 raw_spin_lock(&rq->lock);
1107 if (likely(rq == task_rq(p)))
1109 raw_spin_unlock(&rq->lock);
1110 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1114 static void __task_rq_unlock(struct rq *rq)
1115 __releases(rq->lock)
1117 raw_spin_unlock(&rq->lock);
1121 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1122 __releases(rq->lock)
1123 __releases(p->pi_lock)
1125 raw_spin_unlock(&rq->lock);
1126 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1130 * this_rq_lock - lock this runqueue and disable interrupts.
1132 static struct rq *this_rq_lock(void)
1133 __acquires(rq->lock)
1137 local_irq_disable();
1139 raw_spin_lock(&rq->lock);
1144 #ifdef CONFIG_SCHED_HRTICK
1146 * Use HR-timers to deliver accurate preemption points.
1148 * Its all a bit involved since we cannot program an hrt while holding the
1149 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1152 * When we get rescheduled we reprogram the hrtick_timer outside of the
1158 * - enabled by features
1159 * - hrtimer is actually high res
1161 static inline int hrtick_enabled(struct rq *rq)
1163 if (!sched_feat(HRTICK))
1165 if (!cpu_active(cpu_of(rq)))
1167 return hrtimer_is_hres_active(&rq->hrtick_timer);
1170 static void hrtick_clear(struct rq *rq)
1172 if (hrtimer_active(&rq->hrtick_timer))
1173 hrtimer_cancel(&rq->hrtick_timer);
1177 * High-resolution timer tick.
1178 * Runs from hardirq context with interrupts disabled.
1180 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1182 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1184 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1186 raw_spin_lock(&rq->lock);
1187 update_rq_clock(rq);
1188 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1189 raw_spin_unlock(&rq->lock);
1191 return HRTIMER_NORESTART;
1196 * called from hardirq (IPI) context
1198 static void __hrtick_start(void *arg)
1200 struct rq *rq = arg;
1202 raw_spin_lock(&rq->lock);
1203 hrtimer_restart(&rq->hrtick_timer);
1204 rq->hrtick_csd_pending = 0;
1205 raw_spin_unlock(&rq->lock);
1209 * Called to set the hrtick timer state.
1211 * called with rq->lock held and irqs disabled
1213 static void hrtick_start(struct rq *rq, u64 delay)
1215 struct hrtimer *timer = &rq->hrtick_timer;
1216 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1218 hrtimer_set_expires(timer, time);
1220 if (rq == this_rq()) {
1221 hrtimer_restart(timer);
1222 } else if (!rq->hrtick_csd_pending) {
1223 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1224 rq->hrtick_csd_pending = 1;
1229 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1231 int cpu = (int)(long)hcpu;
1234 case CPU_UP_CANCELED:
1235 case CPU_UP_CANCELED_FROZEN:
1236 case CPU_DOWN_PREPARE:
1237 case CPU_DOWN_PREPARE_FROZEN:
1239 case CPU_DEAD_FROZEN:
1240 hrtick_clear(cpu_rq(cpu));
1247 static __init void init_hrtick(void)
1249 hotcpu_notifier(hotplug_hrtick, 0);
1253 * Called to set the hrtick timer state.
1255 * called with rq->lock held and irqs disabled
1257 static void hrtick_start(struct rq *rq, u64 delay)
1259 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1260 HRTIMER_MODE_REL_PINNED, 0);
1263 static inline void init_hrtick(void)
1266 #endif /* CONFIG_SMP */
1268 static void init_rq_hrtick(struct rq *rq)
1271 rq->hrtick_csd_pending = 0;
1273 rq->hrtick_csd.flags = 0;
1274 rq->hrtick_csd.func = __hrtick_start;
1275 rq->hrtick_csd.info = rq;
1278 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1279 rq->hrtick_timer.function = hrtick;
1281 #else /* CONFIG_SCHED_HRTICK */
1282 static inline void hrtick_clear(struct rq *rq)
1286 static inline void init_rq_hrtick(struct rq *rq)
1290 static inline void init_hrtick(void)
1293 #endif /* CONFIG_SCHED_HRTICK */
1296 * resched_task - mark a task 'to be rescheduled now'.
1298 * On UP this means the setting of the need_resched flag, on SMP it
1299 * might also involve a cross-CPU call to trigger the scheduler on
1304 #ifndef tsk_is_polling
1305 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1308 static void resched_task(struct task_struct *p)
1312 assert_raw_spin_locked(&task_rq(p)->lock);
1314 if (test_tsk_need_resched(p))
1317 set_tsk_need_resched(p);
1320 if (cpu == smp_processor_id())
1323 /* NEED_RESCHED must be visible before we test polling */
1325 if (!tsk_is_polling(p))
1326 smp_send_reschedule(cpu);
1329 static void resched_cpu(int cpu)
1331 struct rq *rq = cpu_rq(cpu);
1332 unsigned long flags;
1334 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1336 resched_task(cpu_curr(cpu));
1337 raw_spin_unlock_irqrestore(&rq->lock, flags);
1342 * In the semi idle case, use the nearest busy cpu for migrating timers
1343 * from an idle cpu. This is good for power-savings.
1345 * We don't do similar optimization for completely idle system, as
1346 * selecting an idle cpu will add more delays to the timers than intended
1347 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1349 int get_nohz_timer_target(void)
1351 int cpu = smp_processor_id();
1353 struct sched_domain *sd;
1356 for_each_domain(cpu, sd) {
1357 for_each_cpu(i, sched_domain_span(sd)) {
1369 * When add_timer_on() enqueues a timer into the timer wheel of an
1370 * idle CPU then this timer might expire before the next timer event
1371 * which is scheduled to wake up that CPU. In case of a completely
1372 * idle system the next event might even be infinite time into the
1373 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1374 * leaves the inner idle loop so the newly added timer is taken into
1375 * account when the CPU goes back to idle and evaluates the timer
1376 * wheel for the next timer event.
1378 void wake_up_idle_cpu(int cpu)
1380 struct rq *rq = cpu_rq(cpu);
1382 if (cpu == smp_processor_id())
1386 * This is safe, as this function is called with the timer
1387 * wheel base lock of (cpu) held. When the CPU is on the way
1388 * to idle and has not yet set rq->curr to idle then it will
1389 * be serialized on the timer wheel base lock and take the new
1390 * timer into account automatically.
1392 if (rq->curr != rq->idle)
1396 * We can set TIF_RESCHED on the idle task of the other CPU
1397 * lockless. The worst case is that the other CPU runs the
1398 * idle task through an additional NOOP schedule()
1400 set_tsk_need_resched(rq->idle);
1402 /* NEED_RESCHED must be visible before we test polling */
1404 if (!tsk_is_polling(rq->idle))
1405 smp_send_reschedule(cpu);
1408 static inline bool got_nohz_idle_kick(void)
1410 return idle_cpu(smp_processor_id()) && this_rq()->nohz_balance_kick;
1413 #else /* CONFIG_NO_HZ */
1415 static inline bool got_nohz_idle_kick(void)
1420 #endif /* CONFIG_NO_HZ */
1422 static u64 sched_avg_period(void)
1424 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1427 static void sched_avg_update(struct rq *rq)
1429 s64 period = sched_avg_period();
1431 while ((s64)(rq->clock - rq->age_stamp) > period) {
1433 * Inline assembly required to prevent the compiler
1434 * optimising this loop into a divmod call.
1435 * See __iter_div_u64_rem() for another example of this.
1437 asm("" : "+rm" (rq->age_stamp));
1438 rq->age_stamp += period;
1443 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1445 rq->rt_avg += rt_delta;
1446 sched_avg_update(rq);
1449 #else /* !CONFIG_SMP */
1450 static void resched_task(struct task_struct *p)
1452 assert_raw_spin_locked(&task_rq(p)->lock);
1453 set_tsk_need_resched(p);
1456 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1460 static void sched_avg_update(struct rq *rq)
1463 #endif /* CONFIG_SMP */
1465 #if BITS_PER_LONG == 32
1466 # define WMULT_CONST (~0UL)
1468 # define WMULT_CONST (1UL << 32)
1471 #define WMULT_SHIFT 32
1474 * Shift right and round:
1476 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1479 * delta *= weight / lw
1481 static unsigned long
1482 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1483 struct load_weight *lw)
1488 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1489 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1490 * 2^SCHED_LOAD_RESOLUTION.
1492 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1493 tmp = (u64)delta_exec * scale_load_down(weight);
1495 tmp = (u64)delta_exec;
1497 if (!lw->inv_weight) {
1498 unsigned long w = scale_load_down(lw->weight);
1500 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1502 else if (unlikely(!w))
1503 lw->inv_weight = WMULT_CONST;
1505 lw->inv_weight = WMULT_CONST / w;
1509 * Check whether we'd overflow the 64-bit multiplication:
1511 if (unlikely(tmp > WMULT_CONST))
1512 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1515 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1517 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1520 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1526 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1532 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1539 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1540 * of tasks with abnormal "nice" values across CPUs the contribution that
1541 * each task makes to its run queue's load is weighted according to its
1542 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1543 * scaled version of the new time slice allocation that they receive on time
1547 #define WEIGHT_IDLEPRIO 3
1548 #define WMULT_IDLEPRIO 1431655765
1551 * Nice levels are multiplicative, with a gentle 10% change for every
1552 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1553 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1554 * that remained on nice 0.
1556 * The "10% effect" is relative and cumulative: from _any_ nice level,
1557 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1558 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1559 * If a task goes up by ~10% and another task goes down by ~10% then
1560 * the relative distance between them is ~25%.)
1562 static const int prio_to_weight[40] = {
1563 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1564 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1565 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1566 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1567 /* 0 */ 1024, 820, 655, 526, 423,
1568 /* 5 */ 335, 272, 215, 172, 137,
1569 /* 10 */ 110, 87, 70, 56, 45,
1570 /* 15 */ 36, 29, 23, 18, 15,
1574 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1576 * In cases where the weight does not change often, we can use the
1577 * precalculated inverse to speed up arithmetics by turning divisions
1578 * into multiplications:
1580 static const u32 prio_to_wmult[40] = {
1581 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1582 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1583 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1584 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1585 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1586 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1587 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1588 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1591 /* Time spent by the tasks of the cpu accounting group executing in ... */
1592 enum cpuacct_stat_index {
1593 CPUACCT_STAT_USER, /* ... user mode */
1594 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1596 CPUACCT_STAT_NSTATS,
1599 #ifdef CONFIG_CGROUP_CPUACCT
1600 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1601 static void cpuacct_update_stats(struct task_struct *tsk,
1602 enum cpuacct_stat_index idx, cputime_t val);
1604 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1605 static inline void cpuacct_update_stats(struct task_struct *tsk,
1606 enum cpuacct_stat_index idx, cputime_t val) {}
1609 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1611 update_load_add(&rq->load, load);
1614 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1616 update_load_sub(&rq->load, load);
1619 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1620 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1621 typedef int (*tg_visitor)(struct task_group *, void *);
1624 * Iterate task_group tree rooted at *from, calling @down when first entering a
1625 * node and @up when leaving it for the final time.
1627 * Caller must hold rcu_lock or sufficient equivalent.
1629 static int walk_tg_tree_from(struct task_group *from,
1630 tg_visitor down, tg_visitor up, void *data)
1632 struct task_group *parent, *child;
1638 ret = (*down)(parent, data);
1641 list_for_each_entry_rcu(child, &parent->children, siblings) {
1648 ret = (*up)(parent, data);
1649 if (ret || parent == from)
1653 parent = parent->parent;
1661 * Iterate the full tree, calling @down when first entering a node and @up when
1662 * leaving it for the final time.
1664 * Caller must hold rcu_lock or sufficient equivalent.
1667 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1669 return walk_tg_tree_from(&root_task_group, down, up, data);
1672 static int tg_nop(struct task_group *tg, void *data)
1679 /* Used instead of source_load when we know the type == 0 */
1680 static unsigned long weighted_cpuload(const int cpu)
1682 return cpu_rq(cpu)->load.weight;
1686 * Return a low guess at the load of a migration-source cpu weighted
1687 * according to the scheduling class and "nice" value.
1689 * We want to under-estimate the load of migration sources, to
1690 * balance conservatively.
1692 static unsigned long source_load(int cpu, int type)
1694 struct rq *rq = cpu_rq(cpu);
1695 unsigned long total = weighted_cpuload(cpu);
1697 if (type == 0 || !sched_feat(LB_BIAS))
1700 return min(rq->cpu_load[type-1], total);
1704 * Return a high guess at the load of a migration-target cpu weighted
1705 * according to the scheduling class and "nice" value.
1707 static unsigned long target_load(int cpu, int type)
1709 struct rq *rq = cpu_rq(cpu);
1710 unsigned long total = weighted_cpuload(cpu);
1712 if (type == 0 || !sched_feat(LB_BIAS))
1715 return max(rq->cpu_load[type-1], total);
1718 static unsigned long power_of(int cpu)
1720 return cpu_rq(cpu)->cpu_power;
1723 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1725 static unsigned long cpu_avg_load_per_task(int cpu)
1727 struct rq *rq = cpu_rq(cpu);
1728 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1731 return rq->load.weight / nr_running;
1736 #ifdef CONFIG_PREEMPT
1738 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1741 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1742 * way at the expense of forcing extra atomic operations in all
1743 * invocations. This assures that the double_lock is acquired using the
1744 * same underlying policy as the spinlock_t on this architecture, which
1745 * reduces latency compared to the unfair variant below. However, it
1746 * also adds more overhead and therefore may reduce throughput.
1748 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1749 __releases(this_rq->lock)
1750 __acquires(busiest->lock)
1751 __acquires(this_rq->lock)
1753 raw_spin_unlock(&this_rq->lock);
1754 double_rq_lock(this_rq, busiest);
1761 * Unfair double_lock_balance: Optimizes throughput at the expense of
1762 * latency by eliminating extra atomic operations when the locks are
1763 * already in proper order on entry. This favors lower cpu-ids and will
1764 * grant the double lock to lower cpus over higher ids under contention,
1765 * regardless of entry order into the function.
1767 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1768 __releases(this_rq->lock)
1769 __acquires(busiest->lock)
1770 __acquires(this_rq->lock)
1774 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1775 if (busiest < this_rq) {
1776 raw_spin_unlock(&this_rq->lock);
1777 raw_spin_lock(&busiest->lock);
1778 raw_spin_lock_nested(&this_rq->lock,
1779 SINGLE_DEPTH_NESTING);
1782 raw_spin_lock_nested(&busiest->lock,
1783 SINGLE_DEPTH_NESTING);
1788 #endif /* CONFIG_PREEMPT */
1791 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1793 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1795 if (unlikely(!irqs_disabled())) {
1796 /* printk() doesn't work good under rq->lock */
1797 raw_spin_unlock(&this_rq->lock);
1801 return _double_lock_balance(this_rq, busiest);
1804 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1805 __releases(busiest->lock)
1807 raw_spin_unlock(&busiest->lock);
1808 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1812 * double_rq_lock - safely lock two runqueues
1814 * Note this does not disable interrupts like task_rq_lock,
1815 * you need to do so manually before calling.
1817 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1818 __acquires(rq1->lock)
1819 __acquires(rq2->lock)
1821 BUG_ON(!irqs_disabled());
1823 raw_spin_lock(&rq1->lock);
1824 __acquire(rq2->lock); /* Fake it out ;) */
1827 raw_spin_lock(&rq1->lock);
1828 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1830 raw_spin_lock(&rq2->lock);
1831 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1837 * double_rq_unlock - safely unlock two runqueues
1839 * Note this does not restore interrupts like task_rq_unlock,
1840 * you need to do so manually after calling.
1842 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1843 __releases(rq1->lock)
1844 __releases(rq2->lock)
1846 raw_spin_unlock(&rq1->lock);
1848 raw_spin_unlock(&rq2->lock);
1850 __release(rq2->lock);
1853 #else /* CONFIG_SMP */
1856 * double_rq_lock - safely lock two runqueues
1858 * Note this does not disable interrupts like task_rq_lock,
1859 * you need to do so manually before calling.
1861 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1862 __acquires(rq1->lock)
1863 __acquires(rq2->lock)
1865 BUG_ON(!irqs_disabled());
1867 raw_spin_lock(&rq1->lock);
1868 __acquire(rq2->lock); /* Fake it out ;) */
1872 * double_rq_unlock - safely unlock two runqueues
1874 * Note this does not restore interrupts like task_rq_unlock,
1875 * you need to do so manually after calling.
1877 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1878 __releases(rq1->lock)
1879 __releases(rq2->lock)
1882 raw_spin_unlock(&rq1->lock);
1883 __release(rq2->lock);
1888 static void update_sysctl(void);
1889 static int get_update_sysctl_factor(void);
1890 static void update_cpu_load(struct rq *this_rq);
1892 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1894 set_task_rq(p, cpu);
1897 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1898 * successfully executed on another CPU. We must ensure that updates of
1899 * per-task data have been completed by this moment.
1902 task_thread_info(p)->cpu = cpu;
1906 static const struct sched_class rt_sched_class;
1908 #define sched_class_highest (&stop_sched_class)
1909 #define for_each_class(class) \
1910 for (class = sched_class_highest; class; class = class->next)
1912 #include "sched_stats.h"
1914 static void inc_nr_running(struct rq *rq)
1919 static void dec_nr_running(struct rq *rq)
1924 static void set_load_weight(struct task_struct *p)
1926 int prio = p->static_prio - MAX_RT_PRIO;
1927 struct load_weight *load = &p->se.load;
1930 * SCHED_IDLE tasks get minimal weight:
1932 if (p->policy == SCHED_IDLE) {
1933 load->weight = scale_load(WEIGHT_IDLEPRIO);
1934 load->inv_weight = WMULT_IDLEPRIO;
1938 load->weight = scale_load(prio_to_weight[prio]);
1939 load->inv_weight = prio_to_wmult[prio];
1942 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1944 update_rq_clock(rq);
1945 sched_info_queued(p);
1946 p->sched_class->enqueue_task(rq, p, flags);
1949 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1951 update_rq_clock(rq);
1952 sched_info_dequeued(p);
1953 p->sched_class->dequeue_task(rq, p, flags);
1957 * activate_task - move a task to the runqueue.
1959 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1961 if (task_contributes_to_load(p))
1962 rq->nr_uninterruptible--;
1964 enqueue_task(rq, p, flags);
1968 * deactivate_task - remove a task from the runqueue.
1970 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1972 if (task_contributes_to_load(p))
1973 rq->nr_uninterruptible++;
1975 dequeue_task(rq, p, flags);
1978 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1981 * There are no locks covering percpu hardirq/softirq time.
1982 * They are only modified in account_system_vtime, on corresponding CPU
1983 * with interrupts disabled. So, writes are safe.
1984 * They are read and saved off onto struct rq in update_rq_clock().
1985 * This may result in other CPU reading this CPU's irq time and can
1986 * race with irq/account_system_vtime on this CPU. We would either get old
1987 * or new value with a side effect of accounting a slice of irq time to wrong
1988 * task when irq is in progress while we read rq->clock. That is a worthy
1989 * compromise in place of having locks on each irq in account_system_time.
1991 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1992 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1994 static DEFINE_PER_CPU(u64, irq_start_time);
1995 static int sched_clock_irqtime;
1997 void enable_sched_clock_irqtime(void)
1999 sched_clock_irqtime = 1;
2002 void disable_sched_clock_irqtime(void)
2004 sched_clock_irqtime = 0;
2007 #ifndef CONFIG_64BIT
2008 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
2010 static inline void irq_time_write_begin(void)
2012 __this_cpu_inc(irq_time_seq.sequence);
2016 static inline void irq_time_write_end(void)
2019 __this_cpu_inc(irq_time_seq.sequence);
2022 static inline u64 irq_time_read(int cpu)
2028 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
2029 irq_time = per_cpu(cpu_softirq_time, cpu) +
2030 per_cpu(cpu_hardirq_time, cpu);
2031 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
2035 #else /* CONFIG_64BIT */
2036 static inline void irq_time_write_begin(void)
2040 static inline void irq_time_write_end(void)
2044 static inline u64 irq_time_read(int cpu)
2046 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
2048 #endif /* CONFIG_64BIT */
2051 * Called before incrementing preempt_count on {soft,}irq_enter
2052 * and before decrementing preempt_count on {soft,}irq_exit.
2054 void account_system_vtime(struct task_struct *curr)
2056 unsigned long flags;
2060 if (!sched_clock_irqtime)
2063 local_irq_save(flags);
2065 cpu = smp_processor_id();
2066 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2067 __this_cpu_add(irq_start_time, delta);
2069 irq_time_write_begin();
2071 * We do not account for softirq time from ksoftirqd here.
2072 * We want to continue accounting softirq time to ksoftirqd thread
2073 * in that case, so as not to confuse scheduler with a special task
2074 * that do not consume any time, but still wants to run.
2076 if (hardirq_count())
2077 __this_cpu_add(cpu_hardirq_time, delta);
2078 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
2079 __this_cpu_add(cpu_softirq_time, delta);
2081 irq_time_write_end();
2082 local_irq_restore(flags);
2084 EXPORT_SYMBOL_GPL(account_system_vtime);
2086 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2088 #ifdef CONFIG_PARAVIRT
2089 static inline u64 steal_ticks(u64 steal)
2091 if (unlikely(steal > NSEC_PER_SEC))
2092 return div_u64(steal, TICK_NSEC);
2094 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
2098 static void update_rq_clock_task(struct rq *rq, s64 delta)
2101 * In theory, the compile should just see 0 here, and optimize out the call
2102 * to sched_rt_avg_update. But I don't trust it...
2104 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2105 s64 steal = 0, irq_delta = 0;
2107 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2108 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2111 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2112 * this case when a previous update_rq_clock() happened inside a
2113 * {soft,}irq region.
2115 * When this happens, we stop ->clock_task and only update the
2116 * prev_irq_time stamp to account for the part that fit, so that a next
2117 * update will consume the rest. This ensures ->clock_task is
2120 * It does however cause some slight miss-attribution of {soft,}irq
2121 * time, a more accurate solution would be to update the irq_time using
2122 * the current rq->clock timestamp, except that would require using
2125 if (irq_delta > delta)
2128 rq->prev_irq_time += irq_delta;
2131 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2132 if (static_branch((¶virt_steal_rq_enabled))) {
2135 steal = paravirt_steal_clock(cpu_of(rq));
2136 steal -= rq->prev_steal_time_rq;
2138 if (unlikely(steal > delta))
2141 st = steal_ticks(steal);
2142 steal = st * TICK_NSEC;
2144 rq->prev_steal_time_rq += steal;
2150 rq->clock_task += delta;
2152 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2153 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2154 sched_rt_avg_update(rq, irq_delta + steal);
2158 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2159 static int irqtime_account_hi_update(void)
2161 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2162 unsigned long flags;
2166 local_irq_save(flags);
2167 latest_ns = this_cpu_read(cpu_hardirq_time);
2168 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2170 local_irq_restore(flags);
2174 static int irqtime_account_si_update(void)
2176 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2177 unsigned long flags;
2181 local_irq_save(flags);
2182 latest_ns = this_cpu_read(cpu_softirq_time);
2183 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2185 local_irq_restore(flags);
2189 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2191 #define sched_clock_irqtime (0)
2195 #include "sched_idletask.c"
2196 #include "sched_fair.c"
2197 #include "sched_rt.c"
2198 #include "sched_autogroup.c"
2199 #include "sched_stoptask.c"
2200 #ifdef CONFIG_SCHED_DEBUG
2201 # include "sched_debug.c"
2204 void sched_set_stop_task(int cpu, struct task_struct *stop)
2206 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2207 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2211 * Make it appear like a SCHED_FIFO task, its something
2212 * userspace knows about and won't get confused about.
2214 * Also, it will make PI more or less work without too
2215 * much confusion -- but then, stop work should not
2216 * rely on PI working anyway.
2218 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2220 stop->sched_class = &stop_sched_class;
2223 cpu_rq(cpu)->stop = stop;
2227 * Reset it back to a normal scheduling class so that
2228 * it can die in pieces.
2230 old_stop->sched_class = &rt_sched_class;
2235 * __normal_prio - return the priority that is based on the static prio
2237 static inline int __normal_prio(struct task_struct *p)
2239 return p->static_prio;
2243 * Calculate the expected normal priority: i.e. priority
2244 * without taking RT-inheritance into account. Might be
2245 * boosted by interactivity modifiers. Changes upon fork,
2246 * setprio syscalls, and whenever the interactivity
2247 * estimator recalculates.
2249 static inline int normal_prio(struct task_struct *p)
2253 if (task_has_rt_policy(p))
2254 prio = MAX_RT_PRIO-1 - p->rt_priority;
2256 prio = __normal_prio(p);
2261 * Calculate the current priority, i.e. the priority
2262 * taken into account by the scheduler. This value might
2263 * be boosted by RT tasks, or might be boosted by
2264 * interactivity modifiers. Will be RT if the task got
2265 * RT-boosted. If not then it returns p->normal_prio.
2267 static int effective_prio(struct task_struct *p)
2269 p->normal_prio = normal_prio(p);
2271 * If we are RT tasks or we were boosted to RT priority,
2272 * keep the priority unchanged. Otherwise, update priority
2273 * to the normal priority:
2275 if (!rt_prio(p->prio))
2276 return p->normal_prio;
2281 * task_curr - is this task currently executing on a CPU?
2282 * @p: the task in question.
2284 inline int task_curr(const struct task_struct *p)
2286 return cpu_curr(task_cpu(p)) == p;
2289 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2290 const struct sched_class *prev_class,
2293 if (prev_class != p->sched_class) {
2294 if (prev_class->switched_from)
2295 prev_class->switched_from(rq, p);
2296 p->sched_class->switched_to(rq, p);
2297 } else if (oldprio != p->prio)
2298 p->sched_class->prio_changed(rq, p, oldprio);
2301 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2303 const struct sched_class *class;
2305 if (p->sched_class == rq->curr->sched_class) {
2306 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2308 for_each_class(class) {
2309 if (class == rq->curr->sched_class)
2311 if (class == p->sched_class) {
2312 resched_task(rq->curr);
2319 * A queue event has occurred, and we're going to schedule. In
2320 * this case, we can save a useless back to back clock update.
2322 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2323 rq->skip_clock_update = 1;
2328 * Is this task likely cache-hot:
2331 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2335 if (p->sched_class != &fair_sched_class)
2338 if (unlikely(p->policy == SCHED_IDLE))
2342 * Buddy candidates are cache hot:
2344 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2345 (&p->se == cfs_rq_of(&p->se)->next ||
2346 &p->se == cfs_rq_of(&p->se)->last))
2349 if (sysctl_sched_migration_cost == -1)
2351 if (sysctl_sched_migration_cost == 0)
2354 delta = now - p->se.exec_start;
2356 return delta < (s64)sysctl_sched_migration_cost;
2359 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2361 #ifdef CONFIG_SCHED_DEBUG
2363 * We should never call set_task_cpu() on a blocked task,
2364 * ttwu() will sort out the placement.
2366 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2367 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2369 #ifdef CONFIG_LOCKDEP
2371 * The caller should hold either p->pi_lock or rq->lock, when changing
2372 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2374 * sched_move_task() holds both and thus holding either pins the cgroup,
2375 * see set_task_rq().
2377 * Furthermore, all task_rq users should acquire both locks, see
2380 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2381 lockdep_is_held(&task_rq(p)->lock)));
2385 trace_sched_migrate_task(p, new_cpu);
2387 if (task_cpu(p) != new_cpu) {
2388 p->se.nr_migrations++;
2389 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2392 __set_task_cpu(p, new_cpu);
2395 struct migration_arg {
2396 struct task_struct *task;
2400 static int migration_cpu_stop(void *data);
2403 * wait_task_inactive - wait for a thread to unschedule.
2405 * If @match_state is nonzero, it's the @p->state value just checked and
2406 * not expected to change. If it changes, i.e. @p might have woken up,
2407 * then return zero. When we succeed in waiting for @p to be off its CPU,
2408 * we return a positive number (its total switch count). If a second call
2409 * a short while later returns the same number, the caller can be sure that
2410 * @p has remained unscheduled the whole time.
2412 * The caller must ensure that the task *will* unschedule sometime soon,
2413 * else this function might spin for a *long* time. This function can't
2414 * be called with interrupts off, or it may introduce deadlock with
2415 * smp_call_function() if an IPI is sent by the same process we are
2416 * waiting to become inactive.
2418 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2420 unsigned long flags;
2427 * We do the initial early heuristics without holding
2428 * any task-queue locks at all. We'll only try to get
2429 * the runqueue lock when things look like they will
2435 * If the task is actively running on another CPU
2436 * still, just relax and busy-wait without holding
2439 * NOTE! Since we don't hold any locks, it's not
2440 * even sure that "rq" stays as the right runqueue!
2441 * But we don't care, since "task_running()" will
2442 * return false if the runqueue has changed and p
2443 * is actually now running somewhere else!
2445 while (task_running(rq, p)) {
2446 if (match_state && unlikely(p->state != match_state))
2452 * Ok, time to look more closely! We need the rq
2453 * lock now, to be *sure*. If we're wrong, we'll
2454 * just go back and repeat.
2456 rq = task_rq_lock(p, &flags);
2457 trace_sched_wait_task(p);
2458 running = task_running(rq, p);
2461 if (!match_state || p->state == match_state)
2462 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2463 task_rq_unlock(rq, p, &flags);
2466 * If it changed from the expected state, bail out now.
2468 if (unlikely(!ncsw))
2472 * Was it really running after all now that we
2473 * checked with the proper locks actually held?
2475 * Oops. Go back and try again..
2477 if (unlikely(running)) {
2483 * It's not enough that it's not actively running,
2484 * it must be off the runqueue _entirely_, and not
2487 * So if it was still runnable (but just not actively
2488 * running right now), it's preempted, and we should
2489 * yield - it could be a while.
2491 if (unlikely(on_rq)) {
2492 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2494 set_current_state(TASK_UNINTERRUPTIBLE);
2495 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2500 * Ahh, all good. It wasn't running, and it wasn't
2501 * runnable, which means that it will never become
2502 * running in the future either. We're all done!
2511 * kick_process - kick a running thread to enter/exit the kernel
2512 * @p: the to-be-kicked thread
2514 * Cause a process which is running on another CPU to enter
2515 * kernel-mode, without any delay. (to get signals handled.)
2517 * NOTE: this function doesn't have to take the runqueue lock,
2518 * because all it wants to ensure is that the remote task enters
2519 * the kernel. If the IPI races and the task has been migrated
2520 * to another CPU then no harm is done and the purpose has been
2523 void kick_process(struct task_struct *p)
2529 if ((cpu != smp_processor_id()) && task_curr(p))
2530 smp_send_reschedule(cpu);
2533 EXPORT_SYMBOL_GPL(kick_process);
2534 #endif /* CONFIG_SMP */
2538 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2540 static int select_fallback_rq(int cpu, struct task_struct *p)
2543 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2545 /* Look for allowed, online CPU in same node. */
2546 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2547 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
2550 /* Any allowed, online CPU? */
2551 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
2552 if (dest_cpu < nr_cpu_ids)
2555 /* No more Mr. Nice Guy. */
2556 dest_cpu = cpuset_cpus_allowed_fallback(p);
2558 * Don't tell them about moving exiting tasks or
2559 * kernel threads (both mm NULL), since they never
2562 if (p->mm && printk_ratelimit()) {
2563 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2564 task_pid_nr(p), p->comm, cpu);
2571 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2574 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2576 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2579 * In order not to call set_task_cpu() on a blocking task we need
2580 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2583 * Since this is common to all placement strategies, this lives here.
2585 * [ this allows ->select_task() to simply return task_cpu(p) and
2586 * not worry about this generic constraint ]
2588 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
2590 cpu = select_fallback_rq(task_cpu(p), p);
2595 static void update_avg(u64 *avg, u64 sample)
2597 s64 diff = sample - *avg;
2603 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2605 #ifdef CONFIG_SCHEDSTATS
2606 struct rq *rq = this_rq();
2609 int this_cpu = smp_processor_id();
2611 if (cpu == this_cpu) {
2612 schedstat_inc(rq, ttwu_local);
2613 schedstat_inc(p, se.statistics.nr_wakeups_local);
2615 struct sched_domain *sd;
2617 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2619 for_each_domain(this_cpu, sd) {
2620 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2621 schedstat_inc(sd, ttwu_wake_remote);
2628 if (wake_flags & WF_MIGRATED)
2629 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2631 #endif /* CONFIG_SMP */
2633 schedstat_inc(rq, ttwu_count);
2634 schedstat_inc(p, se.statistics.nr_wakeups);
2636 if (wake_flags & WF_SYNC)
2637 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2639 #endif /* CONFIG_SCHEDSTATS */
2642 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2644 activate_task(rq, p, en_flags);
2647 /* if a worker is waking up, notify workqueue */
2648 if (p->flags & PF_WQ_WORKER)
2649 wq_worker_waking_up(p, cpu_of(rq));
2653 * Mark the task runnable and perform wakeup-preemption.
2656 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2658 trace_sched_wakeup(p, true);
2659 check_preempt_curr(rq, p, wake_flags);
2661 p->state = TASK_RUNNING;
2663 if (p->sched_class->task_woken)
2664 p->sched_class->task_woken(rq, p);
2666 if (rq->idle_stamp) {
2667 u64 delta = rq->clock - rq->idle_stamp;
2668 u64 max = 2*sysctl_sched_migration_cost;
2673 update_avg(&rq->avg_idle, delta);
2680 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2683 if (p->sched_contributes_to_load)
2684 rq->nr_uninterruptible--;
2687 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2688 ttwu_do_wakeup(rq, p, wake_flags);
2692 * Called in case the task @p isn't fully descheduled from its runqueue,
2693 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2694 * since all we need to do is flip p->state to TASK_RUNNING, since
2695 * the task is still ->on_rq.
2697 static int ttwu_remote(struct task_struct *p, int wake_flags)
2702 rq = __task_rq_lock(p);
2704 ttwu_do_wakeup(rq, p, wake_flags);
2707 __task_rq_unlock(rq);
2713 static void sched_ttwu_pending(void)
2715 struct rq *rq = this_rq();
2716 struct llist_node *llist = llist_del_all(&rq->wake_list);
2717 struct task_struct *p;
2719 raw_spin_lock(&rq->lock);
2722 p = llist_entry(llist, struct task_struct, wake_entry);
2723 llist = llist_next(llist);
2724 ttwu_do_activate(rq, p, 0);
2727 raw_spin_unlock(&rq->lock);
2730 void scheduler_ipi(void)
2732 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2736 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2737 * traditionally all their work was done from the interrupt return
2738 * path. Now that we actually do some work, we need to make sure
2741 * Some archs already do call them, luckily irq_enter/exit nest
2744 * Arguably we should visit all archs and update all handlers,
2745 * however a fair share of IPIs are still resched only so this would
2746 * somewhat pessimize the simple resched case.
2749 sched_ttwu_pending();
2752 * Check if someone kicked us for doing the nohz idle load balance.
2754 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
2755 this_rq()->idle_balance = 1;
2756 raise_softirq_irqoff(SCHED_SOFTIRQ);
2761 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2763 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
2764 smp_send_reschedule(cpu);
2767 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2768 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2773 rq = __task_rq_lock(p);
2775 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2776 ttwu_do_wakeup(rq, p, wake_flags);
2779 __task_rq_unlock(rq);
2784 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2785 #endif /* CONFIG_SMP */
2787 static void ttwu_queue(struct task_struct *p, int cpu)
2789 struct rq *rq = cpu_rq(cpu);
2791 #if defined(CONFIG_SMP)
2792 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2793 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2794 ttwu_queue_remote(p, cpu);
2799 raw_spin_lock(&rq->lock);
2800 ttwu_do_activate(rq, p, 0);
2801 raw_spin_unlock(&rq->lock);
2805 * try_to_wake_up - wake up a thread
2806 * @p: the thread to be awakened
2807 * @state: the mask of task states that can be woken
2808 * @wake_flags: wake modifier flags (WF_*)
2810 * Put it on the run-queue if it's not already there. The "current"
2811 * thread is always on the run-queue (except when the actual
2812 * re-schedule is in progress), and as such you're allowed to do
2813 * the simpler "current->state = TASK_RUNNING" to mark yourself
2814 * runnable without the overhead of this.
2816 * Returns %true if @p was woken up, %false if it was already running
2817 * or @state didn't match @p's state.
2820 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2822 unsigned long flags;
2823 int cpu, success = 0;
2826 raw_spin_lock_irqsave(&p->pi_lock, flags);
2827 if (!(p->state & state))
2830 success = 1; /* we're going to change ->state */
2833 if (p->on_rq && ttwu_remote(p, wake_flags))
2838 * If the owning (remote) cpu is still in the middle of schedule() with
2839 * this task as prev, wait until its done referencing the task.
2842 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2844 * In case the architecture enables interrupts in
2845 * context_switch(), we cannot busy wait, since that
2846 * would lead to deadlocks when an interrupt hits and
2847 * tries to wake up @prev. So bail and do a complete
2850 if (ttwu_activate_remote(p, wake_flags))
2857 * Pairs with the smp_wmb() in finish_lock_switch().
2861 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2862 p->state = TASK_WAKING;
2864 if (p->sched_class->task_waking)
2865 p->sched_class->task_waking(p);
2867 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2868 if (task_cpu(p) != cpu) {
2869 wake_flags |= WF_MIGRATED;
2870 set_task_cpu(p, cpu);
2872 #endif /* CONFIG_SMP */
2876 ttwu_stat(p, cpu, wake_flags);
2878 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2884 * try_to_wake_up_local - try to wake up a local task with rq lock held
2885 * @p: the thread to be awakened
2887 * Put @p on the run-queue if it's not already there. The caller must
2888 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2891 static void try_to_wake_up_local(struct task_struct *p)
2893 struct rq *rq = task_rq(p);
2895 BUG_ON(rq != this_rq());
2896 BUG_ON(p == current);
2897 lockdep_assert_held(&rq->lock);
2899 if (!raw_spin_trylock(&p->pi_lock)) {
2900 raw_spin_unlock(&rq->lock);
2901 raw_spin_lock(&p->pi_lock);
2902 raw_spin_lock(&rq->lock);
2905 if (!(p->state & TASK_NORMAL))
2909 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2911 ttwu_do_wakeup(rq, p, 0);
2912 ttwu_stat(p, smp_processor_id(), 0);
2914 raw_spin_unlock(&p->pi_lock);
2918 * wake_up_process - Wake up a specific process
2919 * @p: The process to be woken up.
2921 * Attempt to wake up the nominated process and move it to the set of runnable
2922 * processes. Returns 1 if the process was woken up, 0 if it was already
2925 * It may be assumed that this function implies a write memory barrier before
2926 * changing the task state if and only if any tasks are woken up.
2928 int wake_up_process(struct task_struct *p)
2930 return try_to_wake_up(p, TASK_ALL, 0);
2932 EXPORT_SYMBOL(wake_up_process);
2934 int wake_up_state(struct task_struct *p, unsigned int state)
2936 return try_to_wake_up(p, state, 0);
2940 * Perform scheduler related setup for a newly forked process p.
2941 * p is forked by current.
2943 * __sched_fork() is basic setup used by init_idle() too:
2945 static void __sched_fork(struct task_struct *p)
2950 p->se.exec_start = 0;
2951 p->se.sum_exec_runtime = 0;
2952 p->se.prev_sum_exec_runtime = 0;
2953 p->se.nr_migrations = 0;
2955 INIT_LIST_HEAD(&p->se.group_node);
2957 #ifdef CONFIG_SCHEDSTATS
2958 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2961 INIT_LIST_HEAD(&p->rt.run_list);
2963 #ifdef CONFIG_PREEMPT_NOTIFIERS
2964 INIT_HLIST_HEAD(&p->preempt_notifiers);
2969 * fork()/clone()-time setup:
2971 void sched_fork(struct task_struct *p)
2973 unsigned long flags;
2974 int cpu = get_cpu();
2978 * We mark the process as running here. This guarantees that
2979 * nobody will actually run it, and a signal or other external
2980 * event cannot wake it up and insert it on the runqueue either.
2982 p->state = TASK_RUNNING;
2985 * Make sure we do not leak PI boosting priority to the child.
2987 p->prio = current->normal_prio;
2990 * Revert to default priority/policy on fork if requested.
2992 if (unlikely(p->sched_reset_on_fork)) {
2993 if (task_has_rt_policy(p)) {
2994 p->policy = SCHED_NORMAL;
2995 p->static_prio = NICE_TO_PRIO(0);
2997 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2998 p->static_prio = NICE_TO_PRIO(0);
3000 p->prio = p->normal_prio = __normal_prio(p);
3004 * We don't need the reset flag anymore after the fork. It has
3005 * fulfilled its duty:
3007 p->sched_reset_on_fork = 0;
3010 if (!rt_prio(p->prio))
3011 p->sched_class = &fair_sched_class;
3013 if (p->sched_class->task_fork)
3014 p->sched_class->task_fork(p);
3017 * The child is not yet in the pid-hash so no cgroup attach races,
3018 * and the cgroup is pinned to this child due to cgroup_fork()
3019 * is ran before sched_fork().
3021 * Silence PROVE_RCU.
3023 raw_spin_lock_irqsave(&p->pi_lock, flags);
3024 set_task_cpu(p, cpu);
3025 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3027 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3028 if (likely(sched_info_on()))
3029 memset(&p->sched_info, 0, sizeof(p->sched_info));
3031 #if defined(CONFIG_SMP)
3034 #ifdef CONFIG_PREEMPT_COUNT
3035 /* Want to start with kernel preemption disabled. */
3036 task_thread_info(p)->preempt_count = 1;
3039 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3046 * wake_up_new_task - wake up a newly created task for the first time.
3048 * This function will do some initial scheduler statistics housekeeping
3049 * that must be done for every newly created context, then puts the task
3050 * on the runqueue and wakes it.
3052 void wake_up_new_task(struct task_struct *p)
3054 unsigned long flags;
3057 raw_spin_lock_irqsave(&p->pi_lock, flags);
3060 * Fork balancing, do it here and not earlier because:
3061 * - cpus_allowed can change in the fork path
3062 * - any previously selected cpu might disappear through hotplug
3064 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3067 rq = __task_rq_lock(p);
3068 activate_task(rq, p, 0);
3070 trace_sched_wakeup_new(p, true);
3071 check_preempt_curr(rq, p, WF_FORK);
3073 if (p->sched_class->task_woken)
3074 p->sched_class->task_woken(rq, p);
3076 task_rq_unlock(rq, p, &flags);
3079 #ifdef CONFIG_PREEMPT_NOTIFIERS
3082 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3083 * @notifier: notifier struct to register
3085 void preempt_notifier_register(struct preempt_notifier *notifier)
3087 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3089 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3092 * preempt_notifier_unregister - no longer interested in preemption notifications
3093 * @notifier: notifier struct to unregister
3095 * This is safe to call from within a preemption notifier.
3097 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3099 hlist_del(¬ifier->link);
3101 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3103 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3105 struct preempt_notifier *notifier;
3106 struct hlist_node *node;
3108 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3109 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3113 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3114 struct task_struct *next)
3116 struct preempt_notifier *notifier;
3117 struct hlist_node *node;
3119 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3120 notifier->ops->sched_out(notifier, next);
3123 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3125 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3130 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3131 struct task_struct *next)
3135 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3138 * prepare_task_switch - prepare to switch tasks
3139 * @rq: the runqueue preparing to switch
3140 * @prev: the current task that is being switched out
3141 * @next: the task we are going to switch to.
3143 * This is called with the rq lock held and interrupts off. It must
3144 * be paired with a subsequent finish_task_switch after the context
3147 * prepare_task_switch sets up locking and calls architecture specific
3151 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3152 struct task_struct *next)
3154 sched_info_switch(prev, next);
3155 perf_event_task_sched_out(prev, next);
3156 fire_sched_out_preempt_notifiers(prev, next);
3157 prepare_lock_switch(rq, next);
3158 prepare_arch_switch(next);
3159 trace_sched_switch(prev, next);
3163 * finish_task_switch - clean up after a task-switch
3164 * @rq: runqueue associated with task-switch
3165 * @prev: the thread we just switched away from.
3167 * finish_task_switch must be called after the context switch, paired
3168 * with a prepare_task_switch call before the context switch.
3169 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3170 * and do any other architecture-specific cleanup actions.
3172 * Note that we may have delayed dropping an mm in context_switch(). If
3173 * so, we finish that here outside of the runqueue lock. (Doing it
3174 * with the lock held can cause deadlocks; see schedule() for
3177 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3178 __releases(rq->lock)
3180 struct mm_struct *mm = rq->prev_mm;
3186 * A task struct has one reference for the use as "current".
3187 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3188 * schedule one last time. The schedule call will never return, and
3189 * the scheduled task must drop that reference.
3190 * The test for TASK_DEAD must occur while the runqueue locks are
3191 * still held, otherwise prev could be scheduled on another cpu, die
3192 * there before we look at prev->state, and then the reference would
3194 * Manfred Spraul <manfred@colorfullife.com>
3196 prev_state = prev->state;
3197 finish_arch_switch(prev);
3198 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3199 local_irq_disable();
3200 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3201 perf_event_task_sched_in(prev, current);
3202 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3204 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3205 finish_lock_switch(rq, prev);
3207 fire_sched_in_preempt_notifiers(current);
3210 if (unlikely(prev_state == TASK_DEAD)) {
3212 * Remove function-return probe instances associated with this
3213 * task and put them back on the free list.
3215 kprobe_flush_task(prev);
3216 put_task_struct(prev);
3222 /* assumes rq->lock is held */
3223 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3225 if (prev->sched_class->pre_schedule)
3226 prev->sched_class->pre_schedule(rq, prev);
3229 /* rq->lock is NOT held, but preemption is disabled */
3230 static inline void post_schedule(struct rq *rq)
3232 if (rq->post_schedule) {
3233 unsigned long flags;
3235 raw_spin_lock_irqsave(&rq->lock, flags);
3236 if (rq->curr->sched_class->post_schedule)
3237 rq->curr->sched_class->post_schedule(rq);
3238 raw_spin_unlock_irqrestore(&rq->lock, flags);
3240 rq->post_schedule = 0;
3246 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3250 static inline void post_schedule(struct rq *rq)
3257 * schedule_tail - first thing a freshly forked thread must call.
3258 * @prev: the thread we just switched away from.
3260 asmlinkage void schedule_tail(struct task_struct *prev)
3261 __releases(rq->lock)
3263 struct rq *rq = this_rq();
3265 finish_task_switch(rq, prev);
3268 * FIXME: do we need to worry about rq being invalidated by the
3273 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3274 /* In this case, finish_task_switch does not reenable preemption */
3277 if (current->set_child_tid)
3278 put_user(task_pid_vnr(current), current->set_child_tid);
3282 * context_switch - switch to the new MM and the new
3283 * thread's register state.
3286 context_switch(struct rq *rq, struct task_struct *prev,
3287 struct task_struct *next)
3289 struct mm_struct *mm, *oldmm;
3291 prepare_task_switch(rq, prev, next);
3294 oldmm = prev->active_mm;
3296 * For paravirt, this is coupled with an exit in switch_to to
3297 * combine the page table reload and the switch backend into
3300 arch_start_context_switch(prev);
3303 next->active_mm = oldmm;
3304 atomic_inc(&oldmm->mm_count);
3305 enter_lazy_tlb(oldmm, next);
3307 switch_mm(oldmm, mm, next);
3310 prev->active_mm = NULL;
3311 rq->prev_mm = oldmm;
3314 * Since the runqueue lock will be released by the next
3315 * task (which is an invalid locking op but in the case
3316 * of the scheduler it's an obvious special-case), so we
3317 * do an early lockdep release here:
3319 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3320 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3323 /* Here we just switch the register state and the stack. */
3324 switch_to(prev, next, prev);
3328 * this_rq must be evaluated again because prev may have moved
3329 * CPUs since it called schedule(), thus the 'rq' on its stack
3330 * frame will be invalid.
3332 finish_task_switch(this_rq(), prev);
3336 * nr_running, nr_uninterruptible and nr_context_switches:
3338 * externally visible scheduler statistics: current number of runnable
3339 * threads, current number of uninterruptible-sleeping threads, total
3340 * number of context switches performed since bootup.
3342 unsigned long nr_running(void)
3344 unsigned long i, sum = 0;
3346 for_each_online_cpu(i)
3347 sum += cpu_rq(i)->nr_running;
3352 unsigned long nr_uninterruptible(void)
3354 unsigned long i, sum = 0;
3356 for_each_possible_cpu(i)
3357 sum += cpu_rq(i)->nr_uninterruptible;
3360 * Since we read the counters lockless, it might be slightly
3361 * inaccurate. Do not allow it to go below zero though:
3363 if (unlikely((long)sum < 0))
3369 unsigned long long nr_context_switches(void)
3372 unsigned long long sum = 0;
3374 for_each_possible_cpu(i)
3375 sum += cpu_rq(i)->nr_switches;
3380 unsigned long nr_iowait(void)
3382 unsigned long i, sum = 0;
3384 for_each_possible_cpu(i)
3385 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3390 unsigned long nr_iowait_cpu(int cpu)
3392 struct rq *this = cpu_rq(cpu);
3393 return atomic_read(&this->nr_iowait);
3396 unsigned long this_cpu_load(void)
3398 struct rq *this = this_rq();
3399 return this->cpu_load[0];
3404 * Global load-average calculations
3406 * We take a distributed and async approach to calculating the global load-avg
3407 * in order to minimize overhead.
3409 * The global load average is an exponentially decaying average of nr_running +
3410 * nr_uninterruptible.
3412 * Once every LOAD_FREQ:
3415 * for_each_possible_cpu(cpu)
3416 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
3418 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
3420 * Due to a number of reasons the above turns in the mess below:
3422 * - for_each_possible_cpu() is prohibitively expensive on machines with
3423 * serious number of cpus, therefore we need to take a distributed approach
3424 * to calculating nr_active.
3426 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
3427 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
3429 * So assuming nr_active := 0 when we start out -- true per definition, we
3430 * can simply take per-cpu deltas and fold those into a global accumulate
3431 * to obtain the same result. See calc_load_fold_active().
3433 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
3434 * across the machine, we assume 10 ticks is sufficient time for every
3435 * cpu to have completed this task.
3437 * This places an upper-bound on the IRQ-off latency of the machine. Then
3438 * again, being late doesn't loose the delta, just wrecks the sample.
3440 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
3441 * this would add another cross-cpu cacheline miss and atomic operation
3442 * to the wakeup path. Instead we increment on whatever cpu the task ran
3443 * when it went into uninterruptible state and decrement on whatever cpu
3444 * did the wakeup. This means that only the sum of nr_uninterruptible over
3445 * all cpus yields the correct result.
3447 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
3450 /* Variables and functions for calc_load */
3451 static atomic_long_t calc_load_tasks;
3452 static unsigned long calc_load_update;
3453 unsigned long avenrun[3];
3454 EXPORT_SYMBOL(avenrun); /* should be removed */
3457 * get_avenrun - get the load average array
3458 * @loads: pointer to dest load array
3459 * @offset: offset to add
3460 * @shift: shift count to shift the result left
3462 * These values are estimates at best, so no need for locking.
3464 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3466 loads[0] = (avenrun[0] + offset) << shift;
3467 loads[1] = (avenrun[1] + offset) << shift;
3468 loads[2] = (avenrun[2] + offset) << shift;
3471 static long calc_load_fold_active(struct rq *this_rq)
3473 long nr_active, delta = 0;
3475 nr_active = this_rq->nr_running;
3476 nr_active += (long) this_rq->nr_uninterruptible;
3478 if (nr_active != this_rq->calc_load_active) {
3479 delta = nr_active - this_rq->calc_load_active;
3480 this_rq->calc_load_active = nr_active;
3487 * a1 = a0 * e + a * (1 - e)
3489 static unsigned long
3490 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3493 load += active * (FIXED_1 - exp);
3494 load += 1UL << (FSHIFT - 1);
3495 return load >> FSHIFT;
3500 * Handle NO_HZ for the global load-average.
3502 * Since the above described distributed algorithm to compute the global
3503 * load-average relies on per-cpu sampling from the tick, it is affected by
3506 * The basic idea is to fold the nr_active delta into a global idle-delta upon
3507 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
3508 * when we read the global state.
3510 * Obviously reality has to ruin such a delightfully simple scheme:
3512 * - When we go NO_HZ idle during the window, we can negate our sample
3513 * contribution, causing under-accounting.
3515 * We avoid this by keeping two idle-delta counters and flipping them
3516 * when the window starts, thus separating old and new NO_HZ load.
3518 * The only trick is the slight shift in index flip for read vs write.
3522 * |-|-----------|-|-----------|-|-----------|-|
3523 * r:0 0 1 1 0 0 1 1 0
3524 * w:0 1 1 0 0 1 1 0 0
3526 * This ensures we'll fold the old idle contribution in this window while
3527 * accumlating the new one.
3529 * - When we wake up from NO_HZ idle during the window, we push up our
3530 * contribution, since we effectively move our sample point to a known
3533 * This is solved by pushing the window forward, and thus skipping the
3534 * sample, for this cpu (effectively using the idle-delta for this cpu which
3535 * was in effect at the time the window opened). This also solves the issue
3536 * of having to deal with a cpu having been in NOHZ idle for multiple
3537 * LOAD_FREQ intervals.
3539 * When making the ILB scale, we should try to pull this in as well.
3541 static atomic_long_t calc_load_idle[2];
3542 static int calc_load_idx;
3544 static inline int calc_load_write_idx(void)
3546 int idx = calc_load_idx;
3549 * See calc_global_nohz(), if we observe the new index, we also
3550 * need to observe the new update time.
3555 * If the folding window started, make sure we start writing in the
3558 if (!time_before(jiffies, calc_load_update))
3564 static inline int calc_load_read_idx(void)
3566 return calc_load_idx & 1;
3569 void calc_load_enter_idle(void)
3571 struct rq *this_rq = this_rq();
3575 * We're going into NOHZ mode, if there's any pending delta, fold it
3576 * into the pending idle delta.
3578 delta = calc_load_fold_active(this_rq);
3580 int idx = calc_load_write_idx();
3581 atomic_long_add(delta, &calc_load_idle[idx]);
3585 void calc_load_exit_idle(void)
3587 struct rq *this_rq = this_rq();
3590 * If we're still before the sample window, we're done.
3592 if (time_before(jiffies, this_rq->calc_load_update))
3596 * We woke inside or after the sample window, this means we're already
3597 * accounted through the nohz accounting, so skip the entire deal and
3598 * sync up for the next window.
3600 this_rq->calc_load_update = calc_load_update;
3601 if (time_before(jiffies, this_rq->calc_load_update + 10))
3602 this_rq->calc_load_update += LOAD_FREQ;
3605 static long calc_load_fold_idle(void)
3607 int idx = calc_load_read_idx();
3610 if (atomic_long_read(&calc_load_idle[idx]))
3611 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
3617 * fixed_power_int - compute: x^n, in O(log n) time
3619 * @x: base of the power
3620 * @frac_bits: fractional bits of @x
3621 * @n: power to raise @x to.
3623 * By exploiting the relation between the definition of the natural power
3624 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3625 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3626 * (where: n_i \elem {0, 1}, the binary vector representing n),
3627 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3628 * of course trivially computable in O(log_2 n), the length of our binary
3631 static unsigned long
3632 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3634 unsigned long result = 1UL << frac_bits;
3639 result += 1UL << (frac_bits - 1);
3640 result >>= frac_bits;
3646 x += 1UL << (frac_bits - 1);
3654 * a1 = a0 * e + a * (1 - e)
3656 * a2 = a1 * e + a * (1 - e)
3657 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3658 * = a0 * e^2 + a * (1 - e) * (1 + e)
3660 * a3 = a2 * e + a * (1 - e)
3661 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3662 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3666 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3667 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3668 * = a0 * e^n + a * (1 - e^n)
3670 * [1] application of the geometric series:
3673 * S_n := \Sum x^i = -------------
3676 static unsigned long
3677 calc_load_n(unsigned long load, unsigned long exp,
3678 unsigned long active, unsigned int n)
3681 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3685 * NO_HZ can leave us missing all per-cpu ticks calling
3686 * calc_load_account_active(), but since an idle CPU folds its delta into
3687 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3688 * in the pending idle delta if our idle period crossed a load cycle boundary.
3690 * Once we've updated the global active value, we need to apply the exponential
3691 * weights adjusted to the number of cycles missed.
3693 static void calc_global_nohz(void)
3695 long delta, active, n;
3697 if (!time_before(jiffies, calc_load_update + 10)) {
3699 * Catch-up, fold however many we are behind still
3701 delta = jiffies - calc_load_update - 10;
3702 n = 1 + (delta / LOAD_FREQ);
3704 active = atomic_long_read(&calc_load_tasks);
3705 active = active > 0 ? active * FIXED_1 : 0;
3707 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3708 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3709 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3711 calc_load_update += n * LOAD_FREQ;
3715 * Flip the idle index...
3717 * Make sure we first write the new time then flip the index, so that
3718 * calc_load_write_idx() will see the new time when it reads the new
3719 * index, this avoids a double flip messing things up.
3724 #else /* !CONFIG_NO_HZ */
3726 static inline long calc_load_fold_idle(void) { return 0; }
3727 static inline void calc_global_nohz(void) { }
3729 #endif /* CONFIG_NO_HZ */
3732 * calc_load - update the avenrun load estimates 10 ticks after the
3733 * CPUs have updated calc_load_tasks.
3735 void calc_global_load(unsigned long ticks)
3739 if (time_before(jiffies, calc_load_update + 10))
3743 * Fold the 'old' idle-delta to include all NO_HZ cpus.
3745 delta = calc_load_fold_idle();
3747 atomic_long_add(delta, &calc_load_tasks);
3749 active = atomic_long_read(&calc_load_tasks);
3750 active = active > 0 ? active * FIXED_1 : 0;
3752 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3753 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3754 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3756 calc_load_update += LOAD_FREQ;
3759 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
3765 * Called from update_cpu_load() to periodically update this CPU's
3768 static void calc_load_account_active(struct rq *this_rq)
3772 if (time_before(jiffies, this_rq->calc_load_update))
3775 delta = calc_load_fold_active(this_rq);
3777 atomic_long_add(delta, &calc_load_tasks);
3779 this_rq->calc_load_update += LOAD_FREQ;
3783 * End of global load-average stuff
3787 * The exact cpuload at various idx values, calculated at every tick would be
3788 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3790 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3791 * on nth tick when cpu may be busy, then we have:
3792 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3793 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3795 * decay_load_missed() below does efficient calculation of
3796 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3797 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3799 * The calculation is approximated on a 128 point scale.
3800 * degrade_zero_ticks is the number of ticks after which load at any
3801 * particular idx is approximated to be zero.
3802 * degrade_factor is a precomputed table, a row for each load idx.
3803 * Each column corresponds to degradation factor for a power of two ticks,
3804 * based on 128 point scale.
3806 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3807 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3809 * With this power of 2 load factors, we can degrade the load n times
3810 * by looking at 1 bits in n and doing as many mult/shift instead of
3811 * n mult/shifts needed by the exact degradation.
3813 #define DEGRADE_SHIFT 7
3814 static const unsigned char
3815 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3816 static const unsigned char
3817 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3818 {0, 0, 0, 0, 0, 0, 0, 0},
3819 {64, 32, 8, 0, 0, 0, 0, 0},
3820 {96, 72, 40, 12, 1, 0, 0},
3821 {112, 98, 75, 43, 15, 1, 0},
3822 {120, 112, 98, 76, 45, 16, 2} };
3825 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3826 * would be when CPU is idle and so we just decay the old load without
3827 * adding any new load.
3829 static unsigned long
3830 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3834 if (!missed_updates)
3837 if (missed_updates >= degrade_zero_ticks[idx])
3841 return load >> missed_updates;
3843 while (missed_updates) {
3844 if (missed_updates % 2)
3845 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3847 missed_updates >>= 1;
3854 * Update rq->cpu_load[] statistics. This function is usually called every
3855 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3856 * every tick. We fix it up based on jiffies.
3858 static void update_cpu_load(struct rq *this_rq)
3860 unsigned long this_load = this_rq->load.weight;
3861 unsigned long curr_jiffies = jiffies;
3862 unsigned long pending_updates;
3865 this_rq->nr_load_updates++;
3867 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3868 if (curr_jiffies == this_rq->last_load_update_tick)
3871 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3872 this_rq->last_load_update_tick = curr_jiffies;
3874 /* Update our load: */
3875 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3876 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3877 unsigned long old_load, new_load;
3879 /* scale is effectively 1 << i now, and >> i divides by scale */
3881 old_load = this_rq->cpu_load[i];
3882 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3883 new_load = this_load;
3885 * Round up the averaging division if load is increasing. This
3886 * prevents us from getting stuck on 9 if the load is 10, for
3889 if (new_load > old_load)
3890 new_load += scale - 1;
3892 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3895 sched_avg_update(this_rq);
3898 static void update_cpu_load_active(struct rq *this_rq)
3900 update_cpu_load(this_rq);
3902 calc_load_account_active(this_rq);
3908 * sched_exec - execve() is a valuable balancing opportunity, because at
3909 * this point the task has the smallest effective memory and cache footprint.
3911 void sched_exec(void)
3913 struct task_struct *p = current;
3914 unsigned long flags;
3917 raw_spin_lock_irqsave(&p->pi_lock, flags);
3918 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3919 if (dest_cpu == smp_processor_id())
3922 if (likely(cpu_active(dest_cpu))) {
3923 struct migration_arg arg = { p, dest_cpu };
3925 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3926 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3930 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3935 DEFINE_PER_CPU(struct kernel_stat, kstat);
3937 EXPORT_PER_CPU_SYMBOL(kstat);
3940 * Return any ns on the sched_clock that have not yet been accounted in
3941 * @p in case that task is currently running.
3943 * Called with task_rq_lock() held on @rq.
3945 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3949 if (task_current(rq, p)) {
3950 update_rq_clock(rq);
3951 ns = rq->clock_task - p->se.exec_start;
3959 unsigned long long task_delta_exec(struct task_struct *p)
3961 unsigned long flags;
3965 rq = task_rq_lock(p, &flags);
3966 ns = do_task_delta_exec(p, rq);
3967 task_rq_unlock(rq, p, &flags);
3973 * Return accounted runtime for the task.
3974 * In case the task is currently running, return the runtime plus current's
3975 * pending runtime that have not been accounted yet.
3977 unsigned long long task_sched_runtime(struct task_struct *p)
3979 unsigned long flags;
3983 rq = task_rq_lock(p, &flags);
3984 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3985 task_rq_unlock(rq, p, &flags);
3991 * Account user cpu time to a process.
3992 * @p: the process that the cpu time gets accounted to
3993 * @cputime: the cpu time spent in user space since the last update
3994 * @cputime_scaled: cputime scaled by cpu frequency
3996 void account_user_time(struct task_struct *p, cputime_t cputime,
3997 cputime_t cputime_scaled)
3999 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4002 /* Add user time to process. */
4003 p->utime = cputime_add(p->utime, cputime);
4004 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4005 account_group_user_time(p, cputime);
4007 /* Add user time to cpustat. */
4008 tmp = cputime_to_cputime64(cputime);
4009 if (TASK_NICE(p) > 0)
4010 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4012 cpustat->user = cputime64_add(cpustat->user, tmp);
4014 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4015 /* Account for user time used */
4016 acct_update_integrals(p);
4020 * Account guest cpu time to a process.
4021 * @p: the process that the cpu time gets accounted to
4022 * @cputime: the cpu time spent in virtual machine since the last update
4023 * @cputime_scaled: cputime scaled by cpu frequency
4025 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4026 cputime_t cputime_scaled)
4029 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4031 tmp = cputime_to_cputime64(cputime);
4033 /* Add guest time to process. */
4034 p->utime = cputime_add(p->utime, cputime);
4035 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4036 account_group_user_time(p, cputime);
4037 p->gtime = cputime_add(p->gtime, cputime);
4039 /* Add guest time to cpustat. */
4040 if (TASK_NICE(p) > 0) {
4041 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4042 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
4044 cpustat->user = cputime64_add(cpustat->user, tmp);
4045 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4050 * Account system cpu time to a process and desired cpustat field
4051 * @p: the process that the cpu time gets accounted to
4052 * @cputime: the cpu time spent in kernel space since the last update
4053 * @cputime_scaled: cputime scaled by cpu frequency
4054 * @target_cputime64: pointer to cpustat field that has to be updated
4057 void __account_system_time(struct task_struct *p, cputime_t cputime,
4058 cputime_t cputime_scaled, cputime64_t *target_cputime64)
4060 cputime64_t tmp = cputime_to_cputime64(cputime);
4062 /* Add system time to process. */
4063 p->stime = cputime_add(p->stime, cputime);
4064 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4065 account_group_system_time(p, cputime);
4067 /* Add system time to cpustat. */
4068 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
4069 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4071 /* Account for system time used */
4072 acct_update_integrals(p);
4076 * Account system cpu time to a process.
4077 * @p: the process that the cpu time gets accounted to
4078 * @hardirq_offset: the offset to subtract from hardirq_count()
4079 * @cputime: the cpu time spent in kernel space since the last update
4080 * @cputime_scaled: cputime scaled by cpu frequency
4082 void account_system_time(struct task_struct *p, int hardirq_offset,
4083 cputime_t cputime, cputime_t cputime_scaled)
4085 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4086 cputime64_t *target_cputime64;
4088 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4089 account_guest_time(p, cputime, cputime_scaled);
4093 if (hardirq_count() - hardirq_offset)
4094 target_cputime64 = &cpustat->irq;
4095 else if (in_serving_softirq())
4096 target_cputime64 = &cpustat->softirq;
4098 target_cputime64 = &cpustat->system;
4100 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
4104 * Account for involuntary wait time.
4105 * @cputime: the cpu time spent in involuntary wait
4107 void account_steal_time(cputime_t cputime)
4109 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4110 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4112 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4116 * Account for idle time.
4117 * @cputime: the cpu time spent in idle wait
4119 void account_idle_time(cputime_t cputime)
4121 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4122 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4123 struct rq *rq = this_rq();
4125 if (atomic_read(&rq->nr_iowait) > 0)
4126 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4128 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4131 static __always_inline bool steal_account_process_tick(void)
4133 #ifdef CONFIG_PARAVIRT
4134 if (static_branch(¶virt_steal_enabled)) {
4137 steal = paravirt_steal_clock(smp_processor_id());
4138 steal -= this_rq()->prev_steal_time;
4140 st = steal_ticks(steal);
4141 this_rq()->prev_steal_time += st * TICK_NSEC;
4143 account_steal_time(st);
4150 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4152 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4154 * Account a tick to a process and cpustat
4155 * @p: the process that the cpu time gets accounted to
4156 * @user_tick: is the tick from userspace
4157 * @rq: the pointer to rq
4159 * Tick demultiplexing follows the order
4160 * - pending hardirq update
4161 * - pending softirq update
4165 * - check for guest_time
4166 * - else account as system_time
4168 * Check for hardirq is done both for system and user time as there is
4169 * no timer going off while we are on hardirq and hence we may never get an
4170 * opportunity to update it solely in system time.
4171 * p->stime and friends are only updated on system time and not on irq
4172 * softirq as those do not count in task exec_runtime any more.
4174 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4177 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4178 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
4179 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4181 if (steal_account_process_tick())
4184 if (irqtime_account_hi_update()) {
4185 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4186 } else if (irqtime_account_si_update()) {
4187 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4188 } else if (this_cpu_ksoftirqd() == p) {
4190 * ksoftirqd time do not get accounted in cpu_softirq_time.
4191 * So, we have to handle it separately here.
4192 * Also, p->stime needs to be updated for ksoftirqd.
4194 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4196 } else if (user_tick) {
4197 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4198 } else if (p == rq->idle) {
4199 account_idle_time(cputime_one_jiffy);
4200 } else if (p->flags & PF_VCPU) { /* System time or guest time */
4201 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
4203 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4208 static void irqtime_account_idle_ticks(int ticks)
4211 struct rq *rq = this_rq();
4213 for (i = 0; i < ticks; i++)
4214 irqtime_account_process_tick(current, 0, rq);
4216 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4217 static void irqtime_account_idle_ticks(int ticks) {}
4218 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4220 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4223 * Account a single tick of cpu time.
4224 * @p: the process that the cpu time gets accounted to
4225 * @user_tick: indicates if the tick is a user or a system tick
4227 void account_process_tick(struct task_struct *p, int user_tick)
4229 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4230 struct rq *rq = this_rq();
4232 if (sched_clock_irqtime) {
4233 irqtime_account_process_tick(p, user_tick, rq);
4237 if (steal_account_process_tick())
4241 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4242 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4243 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4246 account_idle_time(cputime_one_jiffy);
4250 * Account multiple ticks of steal time.
4251 * @p: the process from which the cpu time has been stolen
4252 * @ticks: number of stolen ticks
4254 void account_steal_ticks(unsigned long ticks)
4256 account_steal_time(jiffies_to_cputime(ticks));
4260 * Account multiple ticks of idle time.
4261 * @ticks: number of stolen ticks
4263 void account_idle_ticks(unsigned long ticks)
4266 if (sched_clock_irqtime) {
4267 irqtime_account_idle_ticks(ticks);
4271 account_idle_time(jiffies_to_cputime(ticks));
4277 * Use precise platform statistics if available:
4279 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4280 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4286 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4288 struct task_cputime cputime;
4290 thread_group_cputime(p, &cputime);
4292 *ut = cputime.utime;
4293 *st = cputime.stime;
4297 #ifndef nsecs_to_cputime
4298 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4301 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4303 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4306 * Use CFS's precise accounting:
4308 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4314 do_div(temp, total);
4315 utime = (cputime_t)temp;
4320 * Compare with previous values, to keep monotonicity:
4322 p->prev_utime = max(p->prev_utime, utime);
4323 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4325 *ut = p->prev_utime;
4326 *st = p->prev_stime;
4330 * Must be called with siglock held.
4332 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4334 struct signal_struct *sig = p->signal;
4335 struct task_cputime cputime;
4336 cputime_t rtime, utime, total;
4338 thread_group_cputime(p, &cputime);
4340 total = cputime_add(cputime.utime, cputime.stime);
4341 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4346 temp *= cputime.utime;
4347 do_div(temp, total);
4348 utime = (cputime_t)temp;
4352 sig->prev_utime = max(sig->prev_utime, utime);
4353 sig->prev_stime = max(sig->prev_stime,
4354 cputime_sub(rtime, sig->prev_utime));
4356 *ut = sig->prev_utime;
4357 *st = sig->prev_stime;
4362 * This function gets called by the timer code, with HZ frequency.
4363 * We call it with interrupts disabled.
4365 void scheduler_tick(void)
4367 int cpu = smp_processor_id();
4368 struct rq *rq = cpu_rq(cpu);
4369 struct task_struct *curr = rq->curr;
4373 raw_spin_lock(&rq->lock);
4374 update_rq_clock(rq);
4375 update_cpu_load_active(rq);
4376 curr->sched_class->task_tick(rq, curr, 0);
4377 raw_spin_unlock(&rq->lock);
4379 perf_event_task_tick();
4382 rq->idle_balance = idle_cpu(cpu);
4383 trigger_load_balance(rq, cpu);
4387 notrace unsigned long get_parent_ip(unsigned long addr)
4389 if (in_lock_functions(addr)) {
4390 addr = CALLER_ADDR2;
4391 if (in_lock_functions(addr))
4392 addr = CALLER_ADDR3;
4397 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4398 defined(CONFIG_PREEMPT_TRACER))
4400 void __kprobes add_preempt_count(int val)
4402 #ifdef CONFIG_DEBUG_PREEMPT
4406 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4409 preempt_count() += val;
4410 #ifdef CONFIG_DEBUG_PREEMPT
4412 * Spinlock count overflowing soon?
4414 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4417 if (preempt_count() == val)
4418 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4420 EXPORT_SYMBOL(add_preempt_count);
4422 void __kprobes sub_preempt_count(int val)
4424 #ifdef CONFIG_DEBUG_PREEMPT
4428 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4431 * Is the spinlock portion underflowing?
4433 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4434 !(preempt_count() & PREEMPT_MASK)))
4438 if (preempt_count() == val)
4439 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4440 preempt_count() -= val;
4442 EXPORT_SYMBOL(sub_preempt_count);
4447 * Print scheduling while atomic bug:
4449 static noinline void __schedule_bug(struct task_struct *prev)
4451 struct pt_regs *regs = get_irq_regs();
4453 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4454 prev->comm, prev->pid, preempt_count());
4456 debug_show_held_locks(prev);
4458 if (irqs_disabled())
4459 print_irqtrace_events(prev);
4468 * Various schedule()-time debugging checks and statistics:
4470 static inline void schedule_debug(struct task_struct *prev)
4473 * Test if we are atomic. Since do_exit() needs to call into
4474 * schedule() atomically, we ignore that path for now.
4475 * Otherwise, whine if we are scheduling when we should not be.
4477 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4478 __schedule_bug(prev);
4481 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4483 schedstat_inc(this_rq(), sched_count);
4486 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4488 if (prev->on_rq || rq->skip_clock_update < 0)
4489 update_rq_clock(rq);
4490 prev->sched_class->put_prev_task(rq, prev);
4494 * Pick up the highest-prio task:
4496 static inline struct task_struct *
4497 pick_next_task(struct rq *rq)
4499 const struct sched_class *class;
4500 struct task_struct *p;
4503 * Optimization: we know that if all tasks are in
4504 * the fair class we can call that function directly:
4506 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4507 p = fair_sched_class.pick_next_task(rq);
4512 for_each_class(class) {
4513 p = class->pick_next_task(rq);
4518 BUG(); /* the idle class will always have a runnable task */
4522 * __schedule() is the main scheduler function.
4524 static void __sched __schedule(void)
4526 struct task_struct *prev, *next;
4527 unsigned long *switch_count;
4533 cpu = smp_processor_id();
4535 rcu_note_context_switch(cpu);
4538 schedule_debug(prev);
4540 if (sched_feat(HRTICK))
4543 raw_spin_lock_irq(&rq->lock);
4545 switch_count = &prev->nivcsw;
4546 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4547 if (unlikely(signal_pending_state(prev->state, prev))) {
4548 prev->state = TASK_RUNNING;
4550 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4554 * If a worker went to sleep, notify and ask workqueue
4555 * whether it wants to wake up a task to maintain
4558 if (prev->flags & PF_WQ_WORKER) {
4559 struct task_struct *to_wakeup;
4561 to_wakeup = wq_worker_sleeping(prev, cpu);
4563 try_to_wake_up_local(to_wakeup);
4566 switch_count = &prev->nvcsw;
4569 pre_schedule(rq, prev);
4571 if (unlikely(!rq->nr_running))
4572 idle_balance(cpu, rq);
4574 put_prev_task(rq, prev);
4575 next = pick_next_task(rq);
4576 clear_tsk_need_resched(prev);
4577 rq->skip_clock_update = 0;
4579 if (likely(prev != next)) {
4584 context_switch(rq, prev, next); /* unlocks the rq */
4586 * The context switch have flipped the stack from under us
4587 * and restored the local variables which were saved when
4588 * this task called schedule() in the past. prev == current
4589 * is still correct, but it can be moved to another cpu/rq.
4591 cpu = smp_processor_id();
4594 raw_spin_unlock_irq(&rq->lock);
4598 preempt_enable_no_resched();
4603 static inline void sched_submit_work(struct task_struct *tsk)
4608 * If we are going to sleep and we have plugged IO queued,
4609 * make sure to submit it to avoid deadlocks.
4611 if (blk_needs_flush_plug(tsk))
4612 blk_schedule_flush_plug(tsk);
4615 asmlinkage void __sched schedule(void)
4617 struct task_struct *tsk = current;
4619 sched_submit_work(tsk);
4622 EXPORT_SYMBOL(schedule);
4624 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4626 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4628 if (lock->owner != owner)
4632 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4633 * lock->owner still matches owner, if that fails, owner might
4634 * point to free()d memory, if it still matches, the rcu_read_lock()
4635 * ensures the memory stays valid.
4639 return owner->on_cpu;
4643 * Look out! "owner" is an entirely speculative pointer
4644 * access and not reliable.
4646 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4648 if (!sched_feat(OWNER_SPIN))
4652 while (owner_running(lock, owner)) {
4656 arch_mutex_cpu_relax();
4661 * We break out the loop above on need_resched() and when the
4662 * owner changed, which is a sign for heavy contention. Return
4663 * success only when lock->owner is NULL.
4665 return lock->owner == NULL;
4669 #ifdef CONFIG_PREEMPT
4671 * this is the entry point to schedule() from in-kernel preemption
4672 * off of preempt_enable. Kernel preemptions off return from interrupt
4673 * occur there and call schedule directly.
4675 asmlinkage void __sched notrace preempt_schedule(void)
4677 struct thread_info *ti = current_thread_info();
4680 * If there is a non-zero preempt_count or interrupts are disabled,
4681 * we do not want to preempt the current task. Just return..
4683 if (likely(ti->preempt_count || irqs_disabled()))
4687 add_preempt_count_notrace(PREEMPT_ACTIVE);
4689 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4692 * Check again in case we missed a preemption opportunity
4693 * between schedule and now.
4696 } while (need_resched());
4698 EXPORT_SYMBOL(preempt_schedule);
4701 * this is the entry point to schedule() from kernel preemption
4702 * off of irq context.
4703 * Note, that this is called and return with irqs disabled. This will
4704 * protect us against recursive calling from irq.
4706 asmlinkage void __sched preempt_schedule_irq(void)
4708 struct thread_info *ti = current_thread_info();
4710 /* Catch callers which need to be fixed */
4711 BUG_ON(ti->preempt_count || !irqs_disabled());
4714 add_preempt_count(PREEMPT_ACTIVE);
4717 local_irq_disable();
4718 sub_preempt_count(PREEMPT_ACTIVE);
4721 * Check again in case we missed a preemption opportunity
4722 * between schedule and now.
4725 } while (need_resched());
4728 #endif /* CONFIG_PREEMPT */
4730 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4733 return try_to_wake_up(curr->private, mode, wake_flags);
4735 EXPORT_SYMBOL(default_wake_function);
4738 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4739 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4740 * number) then we wake all the non-exclusive tasks and one exclusive task.
4742 * There are circumstances in which we can try to wake a task which has already
4743 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4744 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4746 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4747 int nr_exclusive, int wake_flags, void *key)
4749 wait_queue_t *curr, *next;
4751 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4752 unsigned flags = curr->flags;
4754 if (curr->func(curr, mode, wake_flags, key) &&
4755 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4761 * __wake_up - wake up threads blocked on a waitqueue.
4763 * @mode: which threads
4764 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4765 * @key: is directly passed to the wakeup function
4767 * It may be assumed that this function implies a write memory barrier before
4768 * changing the task state if and only if any tasks are woken up.
4770 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4771 int nr_exclusive, void *key)
4773 unsigned long flags;
4775 spin_lock_irqsave(&q->lock, flags);
4776 __wake_up_common(q, mode, nr_exclusive, 0, key);
4777 spin_unlock_irqrestore(&q->lock, flags);
4779 EXPORT_SYMBOL(__wake_up);
4782 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4784 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4786 __wake_up_common(q, mode, 1, 0, NULL);
4788 EXPORT_SYMBOL_GPL(__wake_up_locked);
4790 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4792 __wake_up_common(q, mode, 1, 0, key);
4794 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4797 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4799 * @mode: which threads
4800 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4801 * @key: opaque value to be passed to wakeup targets
4803 * The sync wakeup differs that the waker knows that it will schedule
4804 * away soon, so while the target thread will be woken up, it will not
4805 * be migrated to another CPU - ie. the two threads are 'synchronized'
4806 * with each other. This can prevent needless bouncing between CPUs.
4808 * On UP it can prevent extra preemption.
4810 * It may be assumed that this function implies a write memory barrier before
4811 * changing the task state if and only if any tasks are woken up.
4813 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4814 int nr_exclusive, void *key)
4816 unsigned long flags;
4817 int wake_flags = WF_SYNC;
4822 if (unlikely(!nr_exclusive))
4825 spin_lock_irqsave(&q->lock, flags);
4826 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4827 spin_unlock_irqrestore(&q->lock, flags);
4829 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4832 * __wake_up_sync - see __wake_up_sync_key()
4834 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4836 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4838 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4841 * complete: - signals a single thread waiting on this completion
4842 * @x: holds the state of this particular completion
4844 * This will wake up a single thread waiting on this completion. Threads will be
4845 * awakened in the same order in which they were queued.
4847 * See also complete_all(), wait_for_completion() and related routines.
4849 * It may be assumed that this function implies a write memory barrier before
4850 * changing the task state if and only if any tasks are woken up.
4852 void complete(struct completion *x)
4854 unsigned long flags;
4856 spin_lock_irqsave(&x->wait.lock, flags);
4858 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4859 spin_unlock_irqrestore(&x->wait.lock, flags);
4861 EXPORT_SYMBOL(complete);
4864 * complete_all: - signals all threads waiting on this completion
4865 * @x: holds the state of this particular completion
4867 * This will wake up all threads waiting on this particular completion event.
4869 * It may be assumed that this function implies a write memory barrier before
4870 * changing the task state if and only if any tasks are woken up.
4872 void complete_all(struct completion *x)
4874 unsigned long flags;
4876 spin_lock_irqsave(&x->wait.lock, flags);
4877 x->done += UINT_MAX/2;
4878 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4879 spin_unlock_irqrestore(&x->wait.lock, flags);
4881 EXPORT_SYMBOL(complete_all);
4883 static inline long __sched
4884 do_wait_for_common(struct completion *x, long timeout, int state)
4887 DECLARE_WAITQUEUE(wait, current);
4889 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4891 if (signal_pending_state(state, current)) {
4892 timeout = -ERESTARTSYS;
4895 __set_current_state(state);
4896 spin_unlock_irq(&x->wait.lock);
4897 timeout = schedule_timeout(timeout);
4898 spin_lock_irq(&x->wait.lock);
4899 } while (!x->done && timeout);
4900 __remove_wait_queue(&x->wait, &wait);
4905 return timeout ?: 1;
4909 wait_for_common(struct completion *x, long timeout, int state)
4913 spin_lock_irq(&x->wait.lock);
4914 timeout = do_wait_for_common(x, timeout, state);
4915 spin_unlock_irq(&x->wait.lock);
4920 * wait_for_completion: - waits for completion of a task
4921 * @x: holds the state of this particular completion
4923 * This waits to be signaled for completion of a specific task. It is NOT
4924 * interruptible and there is no timeout.
4926 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4927 * and interrupt capability. Also see complete().
4929 void __sched wait_for_completion(struct completion *x)
4931 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4933 EXPORT_SYMBOL(wait_for_completion);
4936 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4937 * @x: holds the state of this particular completion
4938 * @timeout: timeout value in jiffies
4940 * This waits for either a completion of a specific task to be signaled or for a
4941 * specified timeout to expire. The timeout is in jiffies. It is not
4944 * The return value is 0 if timed out, and positive (at least 1, or number of
4945 * jiffies left till timeout) if completed.
4947 unsigned long __sched
4948 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4950 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4952 EXPORT_SYMBOL(wait_for_completion_timeout);
4955 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4956 * @x: holds the state of this particular completion
4958 * This waits for completion of a specific task to be signaled. It is
4961 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
4963 int __sched wait_for_completion_interruptible(struct completion *x)
4965 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4966 if (t == -ERESTARTSYS)
4970 EXPORT_SYMBOL(wait_for_completion_interruptible);
4973 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4974 * @x: holds the state of this particular completion
4975 * @timeout: timeout value in jiffies
4977 * This waits for either a completion of a specific task to be signaled or for a
4978 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4980 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
4981 * positive (at least 1, or number of jiffies left till timeout) if completed.
4984 wait_for_completion_interruptible_timeout(struct completion *x,
4985 unsigned long timeout)
4987 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4989 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4992 * wait_for_completion_killable: - waits for completion of a task (killable)
4993 * @x: holds the state of this particular completion
4995 * This waits to be signaled for completion of a specific task. It can be
4996 * interrupted by a kill signal.
4998 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
5000 int __sched wait_for_completion_killable(struct completion *x)
5002 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5003 if (t == -ERESTARTSYS)
5007 EXPORT_SYMBOL(wait_for_completion_killable);
5010 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
5011 * @x: holds the state of this particular completion
5012 * @timeout: timeout value in jiffies
5014 * This waits for either a completion of a specific task to be
5015 * signaled or for a specified timeout to expire. It can be
5016 * interrupted by a kill signal. The timeout is in jiffies.
5018 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
5019 * positive (at least 1, or number of jiffies left till timeout) if completed.
5022 wait_for_completion_killable_timeout(struct completion *x,
5023 unsigned long timeout)
5025 return wait_for_common(x, timeout, TASK_KILLABLE);
5027 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
5030 * try_wait_for_completion - try to decrement a completion without blocking
5031 * @x: completion structure
5033 * Returns: 0 if a decrement cannot be done without blocking
5034 * 1 if a decrement succeeded.
5036 * If a completion is being used as a counting completion,
5037 * attempt to decrement the counter without blocking. This
5038 * enables us to avoid waiting if the resource the completion
5039 * is protecting is not available.
5041 bool try_wait_for_completion(struct completion *x)
5043 unsigned long flags;
5046 spin_lock_irqsave(&x->wait.lock, flags);
5051 spin_unlock_irqrestore(&x->wait.lock, flags);
5054 EXPORT_SYMBOL(try_wait_for_completion);
5057 * completion_done - Test to see if a completion has any waiters
5058 * @x: completion structure
5060 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5061 * 1 if there are no waiters.
5064 bool completion_done(struct completion *x)
5066 unsigned long flags;
5069 spin_lock_irqsave(&x->wait.lock, flags);
5072 spin_unlock_irqrestore(&x->wait.lock, flags);
5075 EXPORT_SYMBOL(completion_done);
5078 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5080 unsigned long flags;
5083 init_waitqueue_entry(&wait, current);
5085 __set_current_state(state);
5087 spin_lock_irqsave(&q->lock, flags);
5088 __add_wait_queue(q, &wait);
5089 spin_unlock(&q->lock);
5090 timeout = schedule_timeout(timeout);
5091 spin_lock_irq(&q->lock);
5092 __remove_wait_queue(q, &wait);
5093 spin_unlock_irqrestore(&q->lock, flags);
5098 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5100 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5102 EXPORT_SYMBOL(interruptible_sleep_on);
5105 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5107 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5109 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5111 void __sched sleep_on(wait_queue_head_t *q)
5113 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5115 EXPORT_SYMBOL(sleep_on);
5117 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5119 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5121 EXPORT_SYMBOL(sleep_on_timeout);
5123 #ifdef CONFIG_RT_MUTEXES
5126 * rt_mutex_setprio - set the current priority of a task
5128 * @prio: prio value (kernel-internal form)
5130 * This function changes the 'effective' priority of a task. It does
5131 * not touch ->normal_prio like __setscheduler().
5133 * Used by the rt_mutex code to implement priority inheritance logic.
5135 void rt_mutex_setprio(struct task_struct *p, int prio)
5137 int oldprio, on_rq, running;
5139 const struct sched_class *prev_class;
5141 BUG_ON(prio < 0 || prio > MAX_PRIO);
5143 rq = __task_rq_lock(p);
5145 trace_sched_pi_setprio(p, prio);
5147 prev_class = p->sched_class;
5149 running = task_current(rq, p);
5151 dequeue_task(rq, p, 0);
5153 p->sched_class->put_prev_task(rq, p);
5156 p->sched_class = &rt_sched_class;
5158 p->sched_class = &fair_sched_class;
5163 p->sched_class->set_curr_task(rq);
5165 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
5167 check_class_changed(rq, p, prev_class, oldprio);
5168 __task_rq_unlock(rq);
5173 void set_user_nice(struct task_struct *p, long nice)
5175 int old_prio, delta, on_rq;
5176 unsigned long flags;
5179 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5182 * We have to be careful, if called from sys_setpriority(),
5183 * the task might be in the middle of scheduling on another CPU.
5185 rq = task_rq_lock(p, &flags);
5187 * The RT priorities are set via sched_setscheduler(), but we still
5188 * allow the 'normal' nice value to be set - but as expected
5189 * it wont have any effect on scheduling until the task is
5190 * SCHED_FIFO/SCHED_RR:
5192 if (task_has_rt_policy(p)) {
5193 p->static_prio = NICE_TO_PRIO(nice);
5198 dequeue_task(rq, p, 0);
5200 p->static_prio = NICE_TO_PRIO(nice);
5203 p->prio = effective_prio(p);
5204 delta = p->prio - old_prio;
5207 enqueue_task(rq, p, 0);
5209 * If the task increased its priority or is running and
5210 * lowered its priority, then reschedule its CPU:
5212 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5213 resched_task(rq->curr);
5216 task_rq_unlock(rq, p, &flags);
5218 EXPORT_SYMBOL(set_user_nice);
5221 * can_nice - check if a task can reduce its nice value
5225 int can_nice(const struct task_struct *p, const int nice)
5227 /* convert nice value [19,-20] to rlimit style value [1,40] */
5228 int nice_rlim = 20 - nice;
5230 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5231 capable(CAP_SYS_NICE));
5234 #ifdef __ARCH_WANT_SYS_NICE
5237 * sys_nice - change the priority of the current process.
5238 * @increment: priority increment
5240 * sys_setpriority is a more generic, but much slower function that
5241 * does similar things.
5243 SYSCALL_DEFINE1(nice, int, increment)
5248 * Setpriority might change our priority at the same moment.
5249 * We don't have to worry. Conceptually one call occurs first
5250 * and we have a single winner.
5252 if (increment < -40)
5257 nice = TASK_NICE(current) + increment;
5263 if (increment < 0 && !can_nice(current, nice))
5266 retval = security_task_setnice(current, nice);
5270 set_user_nice(current, nice);
5277 * task_prio - return the priority value of a given task.
5278 * @p: the task in question.
5280 * This is the priority value as seen by users in /proc.
5281 * RT tasks are offset by -200. Normal tasks are centered
5282 * around 0, value goes from -16 to +15.
5284 int task_prio(const struct task_struct *p)
5286 return p->prio - MAX_RT_PRIO;
5290 * task_nice - return the nice value of a given task.
5291 * @p: the task in question.
5293 int task_nice(const struct task_struct *p)
5295 return TASK_NICE(p);
5297 EXPORT_SYMBOL(task_nice);
5300 * idle_cpu - is a given cpu idle currently?
5301 * @cpu: the processor in question.
5303 int idle_cpu(int cpu)
5305 struct rq *rq = cpu_rq(cpu);
5307 if (rq->curr != rq->idle)
5314 if (!llist_empty(&rq->wake_list))
5322 * idle_task - return the idle task for a given cpu.
5323 * @cpu: the processor in question.
5325 struct task_struct *idle_task(int cpu)
5327 return cpu_rq(cpu)->idle;
5331 * find_process_by_pid - find a process with a matching PID value.
5332 * @pid: the pid in question.
5334 static struct task_struct *find_process_by_pid(pid_t pid)
5336 return pid ? find_task_by_vpid(pid) : current;
5339 /* Actually do priority change: must hold rq lock. */
5341 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5344 p->rt_priority = prio;
5345 p->normal_prio = normal_prio(p);
5346 /* we are holding p->pi_lock already */
5347 p->prio = rt_mutex_getprio(p);
5348 if (rt_prio(p->prio))
5349 p->sched_class = &rt_sched_class;
5351 p->sched_class = &fair_sched_class;
5356 * check the target process has a UID that matches the current process's
5358 static bool check_same_owner(struct task_struct *p)
5360 const struct cred *cred = current_cred(), *pcred;
5364 pcred = __task_cred(p);
5365 if (cred->user->user_ns == pcred->user->user_ns)
5366 match = (cred->euid == pcred->euid ||
5367 cred->euid == pcred->uid);
5374 static int __sched_setscheduler(struct task_struct *p, int policy,
5375 const struct sched_param *param, bool user)
5377 int retval, oldprio, oldpolicy = -1, on_rq, running;
5378 unsigned long flags;
5379 const struct sched_class *prev_class;
5383 /* may grab non-irq protected spin_locks */
5384 BUG_ON(in_interrupt());
5386 /* double check policy once rq lock held */
5388 reset_on_fork = p->sched_reset_on_fork;
5389 policy = oldpolicy = p->policy;
5391 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5392 policy &= ~SCHED_RESET_ON_FORK;
5394 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5395 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5396 policy != SCHED_IDLE)
5401 * Valid priorities for SCHED_FIFO and SCHED_RR are
5402 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5403 * SCHED_BATCH and SCHED_IDLE is 0.
5405 if (param->sched_priority < 0 ||
5406 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5407 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5409 if (rt_policy(policy) != (param->sched_priority != 0))
5413 * Allow unprivileged RT tasks to decrease priority:
5415 if (user && !capable(CAP_SYS_NICE)) {
5416 if (rt_policy(policy)) {
5417 unsigned long rlim_rtprio =
5418 task_rlimit(p, RLIMIT_RTPRIO);
5420 /* can't set/change the rt policy */
5421 if (policy != p->policy && !rlim_rtprio)
5424 /* can't increase priority */
5425 if (param->sched_priority > p->rt_priority &&
5426 param->sched_priority > rlim_rtprio)
5431 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5432 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5434 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5435 if (!can_nice(p, TASK_NICE(p)))
5439 /* can't change other user's priorities */
5440 if (!check_same_owner(p))
5443 /* Normal users shall not reset the sched_reset_on_fork flag */
5444 if (p->sched_reset_on_fork && !reset_on_fork)
5449 retval = security_task_setscheduler(p);
5455 * make sure no PI-waiters arrive (or leave) while we are
5456 * changing the priority of the task:
5458 * To be able to change p->policy safely, the appropriate
5459 * runqueue lock must be held.
5461 rq = task_rq_lock(p, &flags);
5464 * Changing the policy of the stop threads its a very bad idea
5466 if (p == rq->stop) {
5467 task_rq_unlock(rq, p, &flags);
5472 * If not changing anything there's no need to proceed further:
5474 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5475 param->sched_priority == p->rt_priority))) {
5477 __task_rq_unlock(rq);
5478 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5482 #ifdef CONFIG_RT_GROUP_SCHED
5485 * Do not allow realtime tasks into groups that have no runtime
5488 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5489 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5490 !task_group_is_autogroup(task_group(p))) {
5491 task_rq_unlock(rq, p, &flags);
5497 /* recheck policy now with rq lock held */
5498 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5499 policy = oldpolicy = -1;
5500 task_rq_unlock(rq, p, &flags);
5504 running = task_current(rq, p);
5506 deactivate_task(rq, p, 0);
5508 p->sched_class->put_prev_task(rq, p);
5510 p->sched_reset_on_fork = reset_on_fork;
5513 prev_class = p->sched_class;
5514 __setscheduler(rq, p, policy, param->sched_priority);
5517 p->sched_class->set_curr_task(rq);
5519 activate_task(rq, p, 0);
5521 check_class_changed(rq, p, prev_class, oldprio);
5522 task_rq_unlock(rq, p, &flags);
5524 rt_mutex_adjust_pi(p);
5530 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5531 * @p: the task in question.
5532 * @policy: new policy.
5533 * @param: structure containing the new RT priority.
5535 * NOTE that the task may be already dead.
5537 int sched_setscheduler(struct task_struct *p, int policy,
5538 const struct sched_param *param)
5540 return __sched_setscheduler(p, policy, param, true);
5542 EXPORT_SYMBOL_GPL(sched_setscheduler);
5545 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5546 * @p: the task in question.
5547 * @policy: new policy.
5548 * @param: structure containing the new RT priority.
5550 * Just like sched_setscheduler, only don't bother checking if the
5551 * current context has permission. For example, this is needed in
5552 * stop_machine(): we create temporary high priority worker threads,
5553 * but our caller might not have that capability.
5555 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5556 const struct sched_param *param)
5558 return __sched_setscheduler(p, policy, param, false);
5562 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5564 struct sched_param lparam;
5565 struct task_struct *p;
5568 if (!param || pid < 0)
5570 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5575 p = find_process_by_pid(pid);
5577 retval = sched_setscheduler(p, policy, &lparam);
5584 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5585 * @pid: the pid in question.
5586 * @policy: new policy.
5587 * @param: structure containing the new RT priority.
5589 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5590 struct sched_param __user *, param)
5592 /* negative values for policy are not valid */
5596 return do_sched_setscheduler(pid, policy, param);
5600 * sys_sched_setparam - set/change the RT priority of a thread
5601 * @pid: the pid in question.
5602 * @param: structure containing the new RT priority.
5604 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5606 return do_sched_setscheduler(pid, -1, param);
5610 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5611 * @pid: the pid in question.
5613 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5615 struct task_struct *p;
5623 p = find_process_by_pid(pid);
5625 retval = security_task_getscheduler(p);
5628 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5635 * sys_sched_getparam - get the RT priority of a thread
5636 * @pid: the pid in question.
5637 * @param: structure containing the RT priority.
5639 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5641 struct sched_param lp;
5642 struct task_struct *p;
5645 if (!param || pid < 0)
5649 p = find_process_by_pid(pid);
5654 retval = security_task_getscheduler(p);
5658 lp.sched_priority = p->rt_priority;
5662 * This one might sleep, we cannot do it with a spinlock held ...
5664 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5673 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5675 cpumask_var_t cpus_allowed, new_mask;
5676 struct task_struct *p;
5682 p = find_process_by_pid(pid);
5689 /* Prevent p going away */
5693 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5697 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5699 goto out_free_cpus_allowed;
5702 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5705 retval = security_task_setscheduler(p);
5709 cpuset_cpus_allowed(p, cpus_allowed);
5710 cpumask_and(new_mask, in_mask, cpus_allowed);
5712 retval = set_cpus_allowed_ptr(p, new_mask);
5715 cpuset_cpus_allowed(p, cpus_allowed);
5716 if (!cpumask_subset(new_mask, cpus_allowed)) {
5718 * We must have raced with a concurrent cpuset
5719 * update. Just reset the cpus_allowed to the
5720 * cpuset's cpus_allowed
5722 cpumask_copy(new_mask, cpus_allowed);
5727 free_cpumask_var(new_mask);
5728 out_free_cpus_allowed:
5729 free_cpumask_var(cpus_allowed);
5736 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5737 struct cpumask *new_mask)
5739 if (len < cpumask_size())
5740 cpumask_clear(new_mask);
5741 else if (len > cpumask_size())
5742 len = cpumask_size();
5744 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5748 * sys_sched_setaffinity - set the cpu affinity of a process
5749 * @pid: pid of the process
5750 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5751 * @user_mask_ptr: user-space pointer to the new cpu mask
5753 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5754 unsigned long __user *, user_mask_ptr)
5756 cpumask_var_t new_mask;
5759 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5762 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5764 retval = sched_setaffinity(pid, new_mask);
5765 free_cpumask_var(new_mask);
5769 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5771 struct task_struct *p;
5772 unsigned long flags;
5779 p = find_process_by_pid(pid);
5783 retval = security_task_getscheduler(p);
5787 raw_spin_lock_irqsave(&p->pi_lock, flags);
5788 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5789 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5799 * sys_sched_getaffinity - get the cpu affinity of a process
5800 * @pid: pid of the process
5801 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5802 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5804 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5805 unsigned long __user *, user_mask_ptr)
5810 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5812 if (len & (sizeof(unsigned long)-1))
5815 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5818 ret = sched_getaffinity(pid, mask);
5820 size_t retlen = min_t(size_t, len, cpumask_size());
5822 if (copy_to_user(user_mask_ptr, mask, retlen))
5827 free_cpumask_var(mask);
5833 * sys_sched_yield - yield the current processor to other threads.
5835 * This function yields the current CPU to other tasks. If there are no
5836 * other threads running on this CPU then this function will return.
5838 SYSCALL_DEFINE0(sched_yield)
5840 struct rq *rq = this_rq_lock();
5842 schedstat_inc(rq, yld_count);
5843 current->sched_class->yield_task(rq);
5846 * Since we are going to call schedule() anyway, there's
5847 * no need to preempt or enable interrupts:
5849 __release(rq->lock);
5850 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5851 do_raw_spin_unlock(&rq->lock);
5852 preempt_enable_no_resched();
5859 static inline int should_resched(void)
5861 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5864 static void __cond_resched(void)
5866 add_preempt_count(PREEMPT_ACTIVE);
5868 sub_preempt_count(PREEMPT_ACTIVE);
5871 int __sched _cond_resched(void)
5873 if (should_resched()) {
5879 EXPORT_SYMBOL(_cond_resched);
5882 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5883 * call schedule, and on return reacquire the lock.
5885 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5886 * operations here to prevent schedule() from being called twice (once via
5887 * spin_unlock(), once by hand).
5889 int __cond_resched_lock(spinlock_t *lock)
5891 int resched = should_resched();
5894 lockdep_assert_held(lock);
5896 if (spin_needbreak(lock) || resched) {
5907 EXPORT_SYMBOL(__cond_resched_lock);
5909 int __sched __cond_resched_softirq(void)
5911 BUG_ON(!in_softirq());
5913 if (should_resched()) {
5921 EXPORT_SYMBOL(__cond_resched_softirq);
5924 * yield - yield the current processor to other threads.
5926 * This is a shortcut for kernel-space yielding - it marks the
5927 * thread runnable and calls sys_sched_yield().
5929 void __sched yield(void)
5931 set_current_state(TASK_RUNNING);
5934 EXPORT_SYMBOL(yield);
5937 * yield_to - yield the current processor to another thread in
5938 * your thread group, or accelerate that thread toward the
5939 * processor it's on.
5941 * @preempt: whether task preemption is allowed or not
5943 * It's the caller's job to ensure that the target task struct
5944 * can't go away on us before we can do any checks.
5946 * Returns true if we indeed boosted the target task.
5948 bool __sched yield_to(struct task_struct *p, bool preempt)
5950 struct task_struct *curr = current;
5951 struct rq *rq, *p_rq;
5952 unsigned long flags;
5955 local_irq_save(flags);
5960 double_rq_lock(rq, p_rq);
5961 while (task_rq(p) != p_rq) {
5962 double_rq_unlock(rq, p_rq);
5966 if (!curr->sched_class->yield_to_task)
5969 if (curr->sched_class != p->sched_class)
5972 if (task_running(p_rq, p) || p->state)
5975 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5977 schedstat_inc(rq, yld_count);
5979 * Make p's CPU reschedule; pick_next_entity takes care of
5982 if (preempt && rq != p_rq)
5983 resched_task(p_rq->curr);
5987 double_rq_unlock(rq, p_rq);
5988 local_irq_restore(flags);
5995 EXPORT_SYMBOL_GPL(yield_to);
5998 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5999 * that process accounting knows that this is a task in IO wait state.
6001 void __sched io_schedule(void)
6003 struct rq *rq = raw_rq();
6005 delayacct_blkio_start();
6006 atomic_inc(&rq->nr_iowait);
6007 blk_flush_plug(current);
6008 current->in_iowait = 1;
6010 current->in_iowait = 0;
6011 atomic_dec(&rq->nr_iowait);
6012 delayacct_blkio_end();
6014 EXPORT_SYMBOL(io_schedule);
6016 long __sched io_schedule_timeout(long timeout)
6018 struct rq *rq = raw_rq();
6021 delayacct_blkio_start();
6022 atomic_inc(&rq->nr_iowait);
6023 blk_flush_plug(current);
6024 current->in_iowait = 1;
6025 ret = schedule_timeout(timeout);
6026 current->in_iowait = 0;
6027 atomic_dec(&rq->nr_iowait);
6028 delayacct_blkio_end();
6033 * sys_sched_get_priority_max - return maximum RT priority.
6034 * @policy: scheduling class.
6036 * this syscall returns the maximum rt_priority that can be used
6037 * by a given scheduling class.
6039 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6046 ret = MAX_USER_RT_PRIO-1;
6058 * sys_sched_get_priority_min - return minimum RT priority.
6059 * @policy: scheduling class.
6061 * this syscall returns the minimum rt_priority that can be used
6062 * by a given scheduling class.
6064 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6082 * sys_sched_rr_get_interval - return the default timeslice of a process.
6083 * @pid: pid of the process.
6084 * @interval: userspace pointer to the timeslice value.
6086 * this syscall writes the default timeslice value of a given process
6087 * into the user-space timespec buffer. A value of '0' means infinity.
6089 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6090 struct timespec __user *, interval)
6092 struct task_struct *p;
6093 unsigned int time_slice;
6094 unsigned long flags;
6104 p = find_process_by_pid(pid);
6108 retval = security_task_getscheduler(p);
6112 rq = task_rq_lock(p, &flags);
6113 time_slice = p->sched_class->get_rr_interval(rq, p);
6114 task_rq_unlock(rq, p, &flags);
6117 jiffies_to_timespec(time_slice, &t);
6118 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6126 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6128 void sched_show_task(struct task_struct *p)
6130 unsigned long free = 0;
6133 state = p->state ? __ffs(p->state) + 1 : 0;
6134 printk(KERN_INFO "%-15.15s %c", p->comm,
6135 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6136 #if BITS_PER_LONG == 32
6137 if (state == TASK_RUNNING)
6138 printk(KERN_CONT " running ");
6140 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6142 if (state == TASK_RUNNING)
6143 printk(KERN_CONT " running task ");
6145 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6147 #ifdef CONFIG_DEBUG_STACK_USAGE
6148 free = stack_not_used(p);
6150 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6151 task_pid_nr(p), task_pid_nr(p->real_parent),
6152 (unsigned long)task_thread_info(p)->flags);
6154 show_stack(p, NULL);
6157 void show_state_filter(unsigned long state_filter)
6159 struct task_struct *g, *p;
6161 #if BITS_PER_LONG == 32
6163 " task PC stack pid father\n");
6166 " task PC stack pid father\n");
6169 do_each_thread(g, p) {
6171 * reset the NMI-timeout, listing all files on a slow
6172 * console might take a lot of time:
6174 touch_nmi_watchdog();
6175 if (!state_filter || (p->state & state_filter))
6177 } while_each_thread(g, p);
6179 touch_all_softlockup_watchdogs();
6181 #ifdef CONFIG_SCHED_DEBUG
6182 sysrq_sched_debug_show();
6186 * Only show locks if all tasks are dumped:
6189 debug_show_all_locks();
6192 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6194 idle->sched_class = &idle_sched_class;
6198 * init_idle - set up an idle thread for a given CPU
6199 * @idle: task in question
6200 * @cpu: cpu the idle task belongs to
6202 * NOTE: this function does not set the idle thread's NEED_RESCHED
6203 * flag, to make booting more robust.
6205 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6207 struct rq *rq = cpu_rq(cpu);
6208 unsigned long flags;
6210 raw_spin_lock_irqsave(&rq->lock, flags);
6213 idle->state = TASK_RUNNING;
6214 idle->se.exec_start = sched_clock();
6216 do_set_cpus_allowed(idle, cpumask_of(cpu));
6218 * We're having a chicken and egg problem, even though we are
6219 * holding rq->lock, the cpu isn't yet set to this cpu so the
6220 * lockdep check in task_group() will fail.
6222 * Similar case to sched_fork(). / Alternatively we could
6223 * use task_rq_lock() here and obtain the other rq->lock.
6228 __set_task_cpu(idle, cpu);
6231 rq->curr = rq->idle = idle;
6232 #if defined(CONFIG_SMP)
6235 raw_spin_unlock_irqrestore(&rq->lock, flags);
6237 /* Set the preempt count _outside_ the spinlocks! */
6238 task_thread_info(idle)->preempt_count = 0;
6241 * The idle tasks have their own, simple scheduling class:
6243 idle->sched_class = &idle_sched_class;
6244 ftrace_graph_init_idle_task(idle, cpu);
6245 #if defined(CONFIG_SMP)
6246 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6251 * Increase the granularity value when there are more CPUs,
6252 * because with more CPUs the 'effective latency' as visible
6253 * to users decreases. But the relationship is not linear,
6254 * so pick a second-best guess by going with the log2 of the
6257 * This idea comes from the SD scheduler of Con Kolivas:
6259 static int get_update_sysctl_factor(void)
6261 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6262 unsigned int factor;
6264 switch (sysctl_sched_tunable_scaling) {
6265 case SCHED_TUNABLESCALING_NONE:
6268 case SCHED_TUNABLESCALING_LINEAR:
6271 case SCHED_TUNABLESCALING_LOG:
6273 factor = 1 + ilog2(cpus);
6280 static void update_sysctl(void)
6282 unsigned int factor = get_update_sysctl_factor();
6284 #define SET_SYSCTL(name) \
6285 (sysctl_##name = (factor) * normalized_sysctl_##name)
6286 SET_SYSCTL(sched_min_granularity);
6287 SET_SYSCTL(sched_latency);
6288 SET_SYSCTL(sched_wakeup_granularity);
6292 static inline void sched_init_granularity(void)
6298 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6300 if (p->sched_class && p->sched_class->set_cpus_allowed)
6301 p->sched_class->set_cpus_allowed(p, new_mask);
6303 cpumask_copy(&p->cpus_allowed, new_mask);
6304 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6308 * This is how migration works:
6310 * 1) we invoke migration_cpu_stop() on the target CPU using
6312 * 2) stopper starts to run (implicitly forcing the migrated thread
6314 * 3) it checks whether the migrated task is still in the wrong runqueue.
6315 * 4) if it's in the wrong runqueue then the migration thread removes
6316 * it and puts it into the right queue.
6317 * 5) stopper completes and stop_one_cpu() returns and the migration
6322 * Change a given task's CPU affinity. Migrate the thread to a
6323 * proper CPU and schedule it away if the CPU it's executing on
6324 * is removed from the allowed bitmask.
6326 * NOTE: the caller must have a valid reference to the task, the
6327 * task must not exit() & deallocate itself prematurely. The
6328 * call is not atomic; no spinlocks may be held.
6330 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6332 unsigned long flags;
6334 unsigned int dest_cpu;
6337 rq = task_rq_lock(p, &flags);
6339 if (cpumask_equal(&p->cpus_allowed, new_mask))
6342 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6347 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6352 do_set_cpus_allowed(p, new_mask);
6354 /* Can the task run on the task's current CPU? If so, we're done */
6355 if (cpumask_test_cpu(task_cpu(p), new_mask))
6358 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6360 struct migration_arg arg = { p, dest_cpu };
6361 /* Need help from migration thread: drop lock and wait. */
6362 task_rq_unlock(rq, p, &flags);
6363 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6364 tlb_migrate_finish(p->mm);
6368 task_rq_unlock(rq, p, &flags);
6372 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6375 * Move (not current) task off this cpu, onto dest cpu. We're doing
6376 * this because either it can't run here any more (set_cpus_allowed()
6377 * away from this CPU, or CPU going down), or because we're
6378 * attempting to rebalance this task on exec (sched_exec).
6380 * So we race with normal scheduler movements, but that's OK, as long
6381 * as the task is no longer on this CPU.
6383 * Returns non-zero if task was successfully migrated.
6385 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6387 struct rq *rq_dest, *rq_src;
6390 if (unlikely(!cpu_active(dest_cpu)))
6393 rq_src = cpu_rq(src_cpu);
6394 rq_dest = cpu_rq(dest_cpu);
6396 raw_spin_lock(&p->pi_lock);
6397 double_rq_lock(rq_src, rq_dest);
6398 /* Already moved. */
6399 if (task_cpu(p) != src_cpu)
6401 /* Affinity changed (again). */
6402 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
6406 * If we're not on a rq, the next wake-up will ensure we're
6410 deactivate_task(rq_src, p, 0);
6411 set_task_cpu(p, dest_cpu);
6412 activate_task(rq_dest, p, 0);
6413 check_preempt_curr(rq_dest, p, 0);
6418 double_rq_unlock(rq_src, rq_dest);
6419 raw_spin_unlock(&p->pi_lock);
6424 * migration_cpu_stop - this will be executed by a highprio stopper thread
6425 * and performs thread migration by bumping thread off CPU then
6426 * 'pushing' onto another runqueue.
6428 static int migration_cpu_stop(void *data)
6430 struct migration_arg *arg = data;
6433 * The original target cpu might have gone down and we might
6434 * be on another cpu but it doesn't matter.
6436 local_irq_disable();
6437 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6442 #ifdef CONFIG_HOTPLUG_CPU
6445 * Ensures that the idle task is using init_mm right before its cpu goes
6448 void idle_task_exit(void)
6450 struct mm_struct *mm = current->active_mm;
6452 BUG_ON(cpu_online(smp_processor_id()));
6455 switch_mm(mm, &init_mm, current);
6460 * While a dead CPU has no uninterruptible tasks queued at this point,
6461 * it might still have a nonzero ->nr_uninterruptible counter, because
6462 * for performance reasons the counter is not stricly tracking tasks to
6463 * their home CPUs. So we just add the counter to another CPU's counter,
6464 * to keep the global sum constant after CPU-down:
6466 static void migrate_nr_uninterruptible(struct rq *rq_src)
6468 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6470 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6471 rq_src->nr_uninterruptible = 0;
6475 * remove the tasks which were accounted by rq from calc_load_tasks.
6477 static void calc_global_load_remove(struct rq *rq)
6479 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6480 rq->calc_load_active = 0;
6483 #ifdef CONFIG_CFS_BANDWIDTH
6484 static void unthrottle_offline_cfs_rqs(struct rq *rq)
6486 struct cfs_rq *cfs_rq;
6488 for_each_leaf_cfs_rq(rq, cfs_rq) {
6489 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6491 if (!cfs_rq->runtime_enabled)
6495 * clock_task is not advancing so we just need to make sure
6496 * there's some valid quota amount
6498 cfs_rq->runtime_remaining = cfs_b->quota;
6499 if (cfs_rq_throttled(cfs_rq))
6500 unthrottle_cfs_rq(cfs_rq);
6504 static void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6508 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6509 * try_to_wake_up()->select_task_rq().
6511 * Called with rq->lock held even though we'er in stop_machine() and
6512 * there's no concurrency possible, we hold the required locks anyway
6513 * because of lock validation efforts.
6515 static void migrate_tasks(unsigned int dead_cpu)
6517 struct rq *rq = cpu_rq(dead_cpu);
6518 struct task_struct *next, *stop = rq->stop;
6522 * Fudge the rq selection such that the below task selection loop
6523 * doesn't get stuck on the currently eligible stop task.
6525 * We're currently inside stop_machine() and the rq is either stuck
6526 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6527 * either way we should never end up calling schedule() until we're
6532 /* Ensure any throttled groups are reachable by pick_next_task */
6533 unthrottle_offline_cfs_rqs(rq);
6537 * There's this thread running, bail when that's the only
6540 if (rq->nr_running == 1)
6543 next = pick_next_task(rq);
6545 next->sched_class->put_prev_task(rq, next);
6547 /* Find suitable destination for @next, with force if needed. */
6548 dest_cpu = select_fallback_rq(dead_cpu, next);
6549 raw_spin_unlock(&rq->lock);
6551 __migrate_task(next, dead_cpu, dest_cpu);
6553 raw_spin_lock(&rq->lock);
6559 #endif /* CONFIG_HOTPLUG_CPU */
6561 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6563 static struct ctl_table sd_ctl_dir[] = {
6565 .procname = "sched_domain",
6571 static struct ctl_table sd_ctl_root[] = {
6573 .procname = "kernel",
6575 .child = sd_ctl_dir,
6580 static struct ctl_table *sd_alloc_ctl_entry(int n)
6582 struct ctl_table *entry =
6583 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6588 static void sd_free_ctl_entry(struct ctl_table **tablep)
6590 struct ctl_table *entry;
6593 * In the intermediate directories, both the child directory and
6594 * procname are dynamically allocated and could fail but the mode
6595 * will always be set. In the lowest directory the names are
6596 * static strings and all have proc handlers.
6598 for (entry = *tablep; entry->mode; entry++) {
6600 sd_free_ctl_entry(&entry->child);
6601 if (entry->proc_handler == NULL)
6602 kfree(entry->procname);
6610 set_table_entry(struct ctl_table *entry,
6611 const char *procname, void *data, int maxlen,
6612 mode_t mode, proc_handler *proc_handler)
6614 entry->procname = procname;
6616 entry->maxlen = maxlen;
6618 entry->proc_handler = proc_handler;
6621 static struct ctl_table *
6622 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6624 struct ctl_table *table = sd_alloc_ctl_entry(13);
6629 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6630 sizeof(long), 0644, proc_doulongvec_minmax);
6631 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6632 sizeof(long), 0644, proc_doulongvec_minmax);
6633 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6634 sizeof(int), 0644, proc_dointvec_minmax);
6635 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6636 sizeof(int), 0644, proc_dointvec_minmax);
6637 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6638 sizeof(int), 0644, proc_dointvec_minmax);
6639 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6640 sizeof(int), 0644, proc_dointvec_minmax);
6641 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6642 sizeof(int), 0644, proc_dointvec_minmax);
6643 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6644 sizeof(int), 0644, proc_dointvec_minmax);
6645 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6646 sizeof(int), 0644, proc_dointvec_minmax);
6647 set_table_entry(&table[9], "cache_nice_tries",
6648 &sd->cache_nice_tries,
6649 sizeof(int), 0644, proc_dointvec_minmax);
6650 set_table_entry(&table[10], "flags", &sd->flags,
6651 sizeof(int), 0644, proc_dointvec_minmax);
6652 set_table_entry(&table[11], "name", sd->name,
6653 CORENAME_MAX_SIZE, 0444, proc_dostring);
6654 /* &table[12] is terminator */
6659 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6661 struct ctl_table *entry, *table;
6662 struct sched_domain *sd;
6663 int domain_num = 0, i;
6666 for_each_domain(cpu, sd)
6668 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6673 for_each_domain(cpu, sd) {
6674 snprintf(buf, 32, "domain%d", i);
6675 entry->procname = kstrdup(buf, GFP_KERNEL);
6677 entry->child = sd_alloc_ctl_domain_table(sd);
6684 static struct ctl_table_header *sd_sysctl_header;
6685 static void register_sched_domain_sysctl(void)
6687 int i, cpu_num = num_possible_cpus();
6688 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6691 WARN_ON(sd_ctl_dir[0].child);
6692 sd_ctl_dir[0].child = entry;
6697 for_each_possible_cpu(i) {
6698 snprintf(buf, 32, "cpu%d", i);
6699 entry->procname = kstrdup(buf, GFP_KERNEL);
6701 entry->child = sd_alloc_ctl_cpu_table(i);
6705 WARN_ON(sd_sysctl_header);
6706 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6709 /* may be called multiple times per register */
6710 static void unregister_sched_domain_sysctl(void)
6712 if (sd_sysctl_header)
6713 unregister_sysctl_table(sd_sysctl_header);
6714 sd_sysctl_header = NULL;
6715 if (sd_ctl_dir[0].child)
6716 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6719 static void register_sched_domain_sysctl(void)
6722 static void unregister_sched_domain_sysctl(void)
6727 static void set_rq_online(struct rq *rq)
6730 const struct sched_class *class;
6732 cpumask_set_cpu(rq->cpu, rq->rd->online);
6735 for_each_class(class) {
6736 if (class->rq_online)
6737 class->rq_online(rq);
6742 static void set_rq_offline(struct rq *rq)
6745 const struct sched_class *class;
6747 for_each_class(class) {
6748 if (class->rq_offline)
6749 class->rq_offline(rq);
6752 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6758 * migration_call - callback that gets triggered when a CPU is added.
6759 * Here we can start up the necessary migration thread for the new CPU.
6761 static int __cpuinit
6762 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6764 int cpu = (long)hcpu;
6765 unsigned long flags;
6766 struct rq *rq = cpu_rq(cpu);
6768 switch (action & ~CPU_TASKS_FROZEN) {
6770 case CPU_UP_PREPARE:
6771 rq->calc_load_update = calc_load_update;
6775 /* Update our root-domain */
6776 raw_spin_lock_irqsave(&rq->lock, flags);
6778 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6782 raw_spin_unlock_irqrestore(&rq->lock, flags);
6785 #ifdef CONFIG_HOTPLUG_CPU
6787 sched_ttwu_pending();
6788 /* Update our root-domain */
6789 raw_spin_lock_irqsave(&rq->lock, flags);
6791 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6795 BUG_ON(rq->nr_running != 1); /* the migration thread */
6796 raw_spin_unlock_irqrestore(&rq->lock, flags);
6798 migrate_nr_uninterruptible(rq);
6799 calc_global_load_remove(rq);
6804 update_max_interval();
6810 * Register at high priority so that task migration (migrate_all_tasks)
6811 * happens before everything else. This has to be lower priority than
6812 * the notifier in the perf_event subsystem, though.
6814 static struct notifier_block __cpuinitdata migration_notifier = {
6815 .notifier_call = migration_call,
6816 .priority = CPU_PRI_MIGRATION,
6819 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6820 unsigned long action, void *hcpu)
6822 switch (action & ~CPU_TASKS_FROZEN) {
6824 case CPU_DOWN_FAILED:
6825 set_cpu_active((long)hcpu, true);
6832 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6833 unsigned long action, void *hcpu)
6835 switch (action & ~CPU_TASKS_FROZEN) {
6836 case CPU_DOWN_PREPARE:
6837 set_cpu_active((long)hcpu, false);
6844 static int __init migration_init(void)
6846 void *cpu = (void *)(long)smp_processor_id();
6849 /* Initialize migration for the boot CPU */
6850 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6851 BUG_ON(err == NOTIFY_BAD);
6852 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6853 register_cpu_notifier(&migration_notifier);
6855 /* Register cpu active notifiers */
6856 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6857 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6861 early_initcall(migration_init);
6866 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6868 #ifdef CONFIG_SCHED_DEBUG
6870 static __read_mostly int sched_domain_debug_enabled;
6872 static int __init sched_domain_debug_setup(char *str)
6874 sched_domain_debug_enabled = 1;
6878 early_param("sched_debug", sched_domain_debug_setup);
6880 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6881 struct cpumask *groupmask)
6883 struct sched_group *group = sd->groups;
6886 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6887 cpumask_clear(groupmask);
6889 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6891 if (!(sd->flags & SD_LOAD_BALANCE)) {
6892 printk("does not load-balance\n");
6894 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6899 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6901 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6902 printk(KERN_ERR "ERROR: domain->span does not contain "
6905 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6906 printk(KERN_ERR "ERROR: domain->groups does not contain"
6910 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6914 printk(KERN_ERR "ERROR: group is NULL\n");
6918 if (!group->sgp->power) {
6919 printk(KERN_CONT "\n");
6920 printk(KERN_ERR "ERROR: domain->cpu_power not "
6925 if (!cpumask_weight(sched_group_cpus(group))) {
6926 printk(KERN_CONT "\n");
6927 printk(KERN_ERR "ERROR: empty group\n");
6931 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6932 printk(KERN_CONT "\n");
6933 printk(KERN_ERR "ERROR: repeated CPUs\n");
6937 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6939 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6941 printk(KERN_CONT " %s", str);
6942 if (group->sgp->power != SCHED_POWER_SCALE) {
6943 printk(KERN_CONT " (cpu_power = %d)",
6947 group = group->next;
6948 } while (group != sd->groups);
6949 printk(KERN_CONT "\n");
6951 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6952 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6955 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6956 printk(KERN_ERR "ERROR: parent span is not a superset "
6957 "of domain->span\n");
6961 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6965 if (!sched_domain_debug_enabled)
6969 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6973 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6976 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6984 #else /* !CONFIG_SCHED_DEBUG */
6985 # define sched_domain_debug(sd, cpu) do { } while (0)
6986 #endif /* CONFIG_SCHED_DEBUG */
6988 static int sd_degenerate(struct sched_domain *sd)
6990 if (cpumask_weight(sched_domain_span(sd)) == 1)
6993 /* Following flags need at least 2 groups */
6994 if (sd->flags & (SD_LOAD_BALANCE |
6995 SD_BALANCE_NEWIDLE |
6999 SD_SHARE_PKG_RESOURCES)) {
7000 if (sd->groups != sd->groups->next)
7004 /* Following flags don't use groups */
7005 if (sd->flags & (SD_WAKE_AFFINE))
7012 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7014 unsigned long cflags = sd->flags, pflags = parent->flags;
7016 if (sd_degenerate(parent))
7019 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7022 /* Flags needing groups don't count if only 1 group in parent */
7023 if (parent->groups == parent->groups->next) {
7024 pflags &= ~(SD_LOAD_BALANCE |
7025 SD_BALANCE_NEWIDLE |
7029 SD_SHARE_PKG_RESOURCES);
7030 if (nr_node_ids == 1)
7031 pflags &= ~SD_SERIALIZE;
7033 if (~cflags & pflags)
7039 static void free_rootdomain(struct rcu_head *rcu)
7041 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
7043 cpupri_cleanup(&rd->cpupri);
7044 free_cpumask_var(rd->rto_mask);
7045 free_cpumask_var(rd->online);
7046 free_cpumask_var(rd->span);
7050 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7052 struct root_domain *old_rd = NULL;
7053 unsigned long flags;
7055 raw_spin_lock_irqsave(&rq->lock, flags);
7060 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7063 cpumask_clear_cpu(rq->cpu, old_rd->span);
7066 * If we dont want to free the old_rt yet then
7067 * set old_rd to NULL to skip the freeing later
7070 if (!atomic_dec_and_test(&old_rd->refcount))
7074 atomic_inc(&rd->refcount);
7077 cpumask_set_cpu(rq->cpu, rd->span);
7078 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7081 raw_spin_unlock_irqrestore(&rq->lock, flags);
7084 call_rcu_sched(&old_rd->rcu, free_rootdomain);
7087 static int init_rootdomain(struct root_domain *rd)
7089 memset(rd, 0, sizeof(*rd));
7091 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7093 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7095 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7098 if (cpupri_init(&rd->cpupri) != 0)
7103 free_cpumask_var(rd->rto_mask);
7105 free_cpumask_var(rd->online);
7107 free_cpumask_var(rd->span);
7112 static void init_defrootdomain(void)
7114 init_rootdomain(&def_root_domain);
7116 atomic_set(&def_root_domain.refcount, 1);
7119 static struct root_domain *alloc_rootdomain(void)
7121 struct root_domain *rd;
7123 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7127 if (init_rootdomain(rd) != 0) {
7135 static void free_sched_groups(struct sched_group *sg, int free_sgp)
7137 struct sched_group *tmp, *first;
7146 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
7151 } while (sg != first);
7154 static void free_sched_domain(struct rcu_head *rcu)
7156 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
7159 * If its an overlapping domain it has private groups, iterate and
7162 if (sd->flags & SD_OVERLAP) {
7163 free_sched_groups(sd->groups, 1);
7164 } else if (atomic_dec_and_test(&sd->groups->ref)) {
7165 kfree(sd->groups->sgp);
7171 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
7173 call_rcu(&sd->rcu, free_sched_domain);
7176 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
7178 for (; sd; sd = sd->parent)
7179 destroy_sched_domain(sd, cpu);
7183 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7184 * hold the hotplug lock.
7187 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7189 struct rq *rq = cpu_rq(cpu);
7190 struct sched_domain *tmp;
7192 /* Remove the sched domains which do not contribute to scheduling. */
7193 for (tmp = sd; tmp; ) {
7194 struct sched_domain *parent = tmp->parent;
7198 if (sd_parent_degenerate(tmp, parent)) {
7199 tmp->parent = parent->parent;
7201 parent->parent->child = tmp;
7202 destroy_sched_domain(parent, cpu);
7207 if (sd && sd_degenerate(sd)) {
7210 destroy_sched_domain(tmp, cpu);
7215 sched_domain_debug(sd, cpu);
7217 rq_attach_root(rq, rd);
7219 rcu_assign_pointer(rq->sd, sd);
7220 destroy_sched_domains(tmp, cpu);
7223 /* cpus with isolated domains */
7224 static cpumask_var_t cpu_isolated_map;
7226 /* Setup the mask of cpus configured for isolated domains */
7227 static int __init isolated_cpu_setup(char *str)
7229 alloc_bootmem_cpumask_var(&cpu_isolated_map);
7230 cpulist_parse(str, cpu_isolated_map);
7234 __setup("isolcpus=", isolated_cpu_setup);
7239 * find_next_best_node - find the next node to include in a sched_domain
7240 * @node: node whose sched_domain we're building
7241 * @used_nodes: nodes already in the sched_domain
7243 * Find the next node to include in a given scheduling domain. Simply
7244 * finds the closest node not already in the @used_nodes map.
7246 * Should use nodemask_t.
7248 static int find_next_best_node(int node, nodemask_t *used_nodes)
7250 int i, n, val, min_val, best_node = -1;
7254 for (i = 0; i < nr_node_ids; i++) {
7255 /* Start at @node */
7256 n = (node + i) % nr_node_ids;
7258 if (!nr_cpus_node(n))
7261 /* Skip already used nodes */
7262 if (node_isset(n, *used_nodes))
7265 /* Simple min distance search */
7266 val = node_distance(node, n);
7268 if (val < min_val) {
7274 if (best_node != -1)
7275 node_set(best_node, *used_nodes);
7280 * sched_domain_node_span - get a cpumask for a node's sched_domain
7281 * @node: node whose cpumask we're constructing
7282 * @span: resulting cpumask
7284 * Given a node, construct a good cpumask for its sched_domain to span. It
7285 * should be one that prevents unnecessary balancing, but also spreads tasks
7288 static void sched_domain_node_span(int node, struct cpumask *span)
7290 nodemask_t used_nodes;
7293 cpumask_clear(span);
7294 nodes_clear(used_nodes);
7296 cpumask_or(span, span, cpumask_of_node(node));
7297 node_set(node, used_nodes);
7299 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7300 int next_node = find_next_best_node(node, &used_nodes);
7303 cpumask_or(span, span, cpumask_of_node(next_node));
7307 static const struct cpumask *cpu_node_mask(int cpu)
7309 lockdep_assert_held(&sched_domains_mutex);
7311 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7313 return sched_domains_tmpmask;
7316 static const struct cpumask *cpu_allnodes_mask(int cpu)
7318 return cpu_possible_mask;
7320 #endif /* CONFIG_NUMA */
7322 static const struct cpumask *cpu_cpu_mask(int cpu)
7324 return cpumask_of_node(cpu_to_node(cpu));
7327 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7330 struct sched_domain **__percpu sd;
7331 struct sched_group **__percpu sg;
7332 struct sched_group_power **__percpu sgp;
7336 struct sched_domain ** __percpu sd;
7337 struct root_domain *rd;
7347 struct sched_domain_topology_level;
7349 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7350 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7352 #define SDTL_OVERLAP 0x01
7354 struct sched_domain_topology_level {
7355 sched_domain_init_f init;
7356 sched_domain_mask_f mask;
7358 struct sd_data data;
7362 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7364 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7365 const struct cpumask *span = sched_domain_span(sd);
7366 struct cpumask *covered = sched_domains_tmpmask;
7367 struct sd_data *sdd = sd->private;
7368 struct sched_domain *child;
7371 cpumask_clear(covered);
7373 for_each_cpu(i, span) {
7374 struct cpumask *sg_span;
7376 if (cpumask_test_cpu(i, covered))
7379 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7380 GFP_KERNEL, cpu_to_node(i));
7385 sg_span = sched_group_cpus(sg);
7387 child = *per_cpu_ptr(sdd->sd, i);
7389 child = child->child;
7390 cpumask_copy(sg_span, sched_domain_span(child));
7392 cpumask_set_cpu(i, sg_span);
7394 cpumask_or(covered, covered, sg_span);
7396 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7397 atomic_inc(&sg->sgp->ref);
7399 if (cpumask_test_cpu(cpu, sg_span))
7409 sd->groups = groups;
7414 free_sched_groups(first, 0);
7419 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7421 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7422 struct sched_domain *child = sd->child;
7425 cpu = cpumask_first(sched_domain_span(child));
7428 *sg = *per_cpu_ptr(sdd->sg, cpu);
7429 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7430 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7437 * build_sched_groups will build a circular linked list of the groups
7438 * covered by the given span, and will set each group's ->cpumask correctly,
7439 * and ->cpu_power to 0.
7441 * Assumes the sched_domain tree is fully constructed
7444 build_sched_groups(struct sched_domain *sd, int cpu)
7446 struct sched_group *first = NULL, *last = NULL;
7447 struct sd_data *sdd = sd->private;
7448 const struct cpumask *span = sched_domain_span(sd);
7449 struct cpumask *covered;
7452 get_group(cpu, sdd, &sd->groups);
7453 atomic_inc(&sd->groups->ref);
7455 if (cpu != cpumask_first(sched_domain_span(sd)))
7458 lockdep_assert_held(&sched_domains_mutex);
7459 covered = sched_domains_tmpmask;
7461 cpumask_clear(covered);
7463 for_each_cpu(i, span) {
7464 struct sched_group *sg;
7465 int group = get_group(i, sdd, &sg);
7468 if (cpumask_test_cpu(i, covered))
7471 cpumask_clear(sched_group_cpus(sg));
7474 for_each_cpu(j, span) {
7475 if (get_group(j, sdd, NULL) != group)
7478 cpumask_set_cpu(j, covered);
7479 cpumask_set_cpu(j, sched_group_cpus(sg));
7494 * Initialize sched groups cpu_power.
7496 * cpu_power indicates the capacity of sched group, which is used while
7497 * distributing the load between different sched groups in a sched domain.
7498 * Typically cpu_power for all the groups in a sched domain will be same unless
7499 * there are asymmetries in the topology. If there are asymmetries, group
7500 * having more cpu_power will pickup more load compared to the group having
7503 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7505 struct sched_group *sg = sd->groups;
7507 WARN_ON(!sd || !sg);
7510 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7512 } while (sg != sd->groups);
7514 if (cpu != group_first_cpu(sg))
7517 update_group_power(sd, cpu);
7521 * Initializers for schedule domains
7522 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7525 #ifdef CONFIG_SCHED_DEBUG
7526 # define SD_INIT_NAME(sd, type) sd->name = #type
7528 # define SD_INIT_NAME(sd, type) do { } while (0)
7531 #define SD_INIT_FUNC(type) \
7532 static noinline struct sched_domain * \
7533 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7535 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7536 *sd = SD_##type##_INIT; \
7537 SD_INIT_NAME(sd, type); \
7538 sd->private = &tl->data; \
7544 SD_INIT_FUNC(ALLNODES)
7547 #ifdef CONFIG_SCHED_SMT
7548 SD_INIT_FUNC(SIBLING)
7550 #ifdef CONFIG_SCHED_MC
7553 #ifdef CONFIG_SCHED_BOOK
7557 static int default_relax_domain_level = -1;
7558 int sched_domain_level_max;
7560 static int __init setup_relax_domain_level(char *str)
7562 if (kstrtoint(str, 0, &default_relax_domain_level))
7563 pr_warn("Unable to set relax_domain_level\n");
7567 __setup("relax_domain_level=", setup_relax_domain_level);
7569 static void set_domain_attribute(struct sched_domain *sd,
7570 struct sched_domain_attr *attr)
7574 if (!attr || attr->relax_domain_level < 0) {
7575 if (default_relax_domain_level < 0)
7578 request = default_relax_domain_level;
7580 request = attr->relax_domain_level;
7581 if (request < sd->level) {
7582 /* turn off idle balance on this domain */
7583 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7585 /* turn on idle balance on this domain */
7586 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7590 static void __sdt_free(const struct cpumask *cpu_map);
7591 static int __sdt_alloc(const struct cpumask *cpu_map);
7593 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7594 const struct cpumask *cpu_map)
7598 if (!atomic_read(&d->rd->refcount))
7599 free_rootdomain(&d->rd->rcu); /* fall through */
7601 free_percpu(d->sd); /* fall through */
7603 __sdt_free(cpu_map); /* fall through */
7609 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7610 const struct cpumask *cpu_map)
7612 memset(d, 0, sizeof(*d));
7614 if (__sdt_alloc(cpu_map))
7615 return sa_sd_storage;
7616 d->sd = alloc_percpu(struct sched_domain *);
7618 return sa_sd_storage;
7619 d->rd = alloc_rootdomain();
7622 return sa_rootdomain;
7626 * NULL the sd_data elements we've used to build the sched_domain and
7627 * sched_group structure so that the subsequent __free_domain_allocs()
7628 * will not free the data we're using.
7630 static void claim_allocations(int cpu, struct sched_domain *sd)
7632 struct sd_data *sdd = sd->private;
7634 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7635 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7637 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7638 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7640 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7641 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7644 #ifdef CONFIG_SCHED_SMT
7645 static const struct cpumask *cpu_smt_mask(int cpu)
7647 return topology_thread_cpumask(cpu);
7652 * Topology list, bottom-up.
7654 static struct sched_domain_topology_level default_topology[] = {
7655 #ifdef CONFIG_SCHED_SMT
7656 { sd_init_SIBLING, cpu_smt_mask, },
7658 #ifdef CONFIG_SCHED_MC
7659 { sd_init_MC, cpu_coregroup_mask, },
7661 #ifdef CONFIG_SCHED_BOOK
7662 { sd_init_BOOK, cpu_book_mask, },
7664 { sd_init_CPU, cpu_cpu_mask, },
7666 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7667 { sd_init_ALLNODES, cpu_allnodes_mask, },
7672 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7674 static int __sdt_alloc(const struct cpumask *cpu_map)
7676 struct sched_domain_topology_level *tl;
7679 for (tl = sched_domain_topology; tl->init; tl++) {
7680 struct sd_data *sdd = &tl->data;
7682 sdd->sd = alloc_percpu(struct sched_domain *);
7686 sdd->sg = alloc_percpu(struct sched_group *);
7690 sdd->sgp = alloc_percpu(struct sched_group_power *);
7694 for_each_cpu(j, cpu_map) {
7695 struct sched_domain *sd;
7696 struct sched_group *sg;
7697 struct sched_group_power *sgp;
7699 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7700 GFP_KERNEL, cpu_to_node(j));
7704 *per_cpu_ptr(sdd->sd, j) = sd;
7706 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7707 GFP_KERNEL, cpu_to_node(j));
7711 *per_cpu_ptr(sdd->sg, j) = sg;
7713 sgp = kzalloc_node(sizeof(struct sched_group_power),
7714 GFP_KERNEL, cpu_to_node(j));
7718 *per_cpu_ptr(sdd->sgp, j) = sgp;
7725 static void __sdt_free(const struct cpumask *cpu_map)
7727 struct sched_domain_topology_level *tl;
7730 for (tl = sched_domain_topology; tl->init; tl++) {
7731 struct sd_data *sdd = &tl->data;
7733 for_each_cpu(j, cpu_map) {
7734 struct sched_domain *sd;
7737 sd = *per_cpu_ptr(sdd->sd, j);
7738 if (sd && (sd->flags & SD_OVERLAP))
7739 free_sched_groups(sd->groups, 0);
7740 kfree(*per_cpu_ptr(sdd->sd, j));
7744 kfree(*per_cpu_ptr(sdd->sg, j));
7746 kfree(*per_cpu_ptr(sdd->sgp, j));
7748 free_percpu(sdd->sd);
7750 free_percpu(sdd->sg);
7752 free_percpu(sdd->sgp);
7757 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7758 struct s_data *d, const struct cpumask *cpu_map,
7759 struct sched_domain_attr *attr, struct sched_domain *child,
7762 struct sched_domain *sd = tl->init(tl, cpu);
7766 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7768 sd->level = child->level + 1;
7769 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7773 set_domain_attribute(sd, attr);
7779 * Build sched domains for a given set of cpus and attach the sched domains
7780 * to the individual cpus
7782 static int build_sched_domains(const struct cpumask *cpu_map,
7783 struct sched_domain_attr *attr)
7785 enum s_alloc alloc_state = sa_none;
7786 struct sched_domain *sd;
7788 int i, ret = -ENOMEM;
7790 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7791 if (alloc_state != sa_rootdomain)
7794 /* Set up domains for cpus specified by the cpu_map. */
7795 for_each_cpu(i, cpu_map) {
7796 struct sched_domain_topology_level *tl;
7799 for (tl = sched_domain_topology; tl->init; tl++) {
7800 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7801 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7802 sd->flags |= SD_OVERLAP;
7803 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7810 *per_cpu_ptr(d.sd, i) = sd;
7813 /* Build the groups for the domains */
7814 for_each_cpu(i, cpu_map) {
7815 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7816 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7817 if (sd->flags & SD_OVERLAP) {
7818 if (build_overlap_sched_groups(sd, i))
7821 if (build_sched_groups(sd, i))
7827 /* Calculate CPU power for physical packages and nodes */
7828 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7829 if (!cpumask_test_cpu(i, cpu_map))
7832 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7833 claim_allocations(i, sd);
7834 init_sched_groups_power(i, sd);
7838 /* Attach the domains */
7840 for_each_cpu(i, cpu_map) {
7841 sd = *per_cpu_ptr(d.sd, i);
7842 cpu_attach_domain(sd, d.rd, i);
7848 __free_domain_allocs(&d, alloc_state, cpu_map);
7852 static cpumask_var_t *doms_cur; /* current sched domains */
7853 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7854 static struct sched_domain_attr *dattr_cur;
7855 /* attribues of custom domains in 'doms_cur' */
7858 * Special case: If a kmalloc of a doms_cur partition (array of
7859 * cpumask) fails, then fallback to a single sched domain,
7860 * as determined by the single cpumask fallback_doms.
7862 static cpumask_var_t fallback_doms;
7865 * arch_update_cpu_topology lets virtualized architectures update the
7866 * cpu core maps. It is supposed to return 1 if the topology changed
7867 * or 0 if it stayed the same.
7869 int __attribute__((weak)) arch_update_cpu_topology(void)
7874 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7877 cpumask_var_t *doms;
7879 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7882 for (i = 0; i < ndoms; i++) {
7883 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7884 free_sched_domains(doms, i);
7891 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7894 for (i = 0; i < ndoms; i++)
7895 free_cpumask_var(doms[i]);
7900 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7901 * For now this just excludes isolated cpus, but could be used to
7902 * exclude other special cases in the future.
7904 static int init_sched_domains(const struct cpumask *cpu_map)
7908 arch_update_cpu_topology();
7910 doms_cur = alloc_sched_domains(ndoms_cur);
7912 doms_cur = &fallback_doms;
7913 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7915 err = build_sched_domains(doms_cur[0], NULL);
7916 register_sched_domain_sysctl();
7922 * Detach sched domains from a group of cpus specified in cpu_map
7923 * These cpus will now be attached to the NULL domain
7925 static void detach_destroy_domains(const struct cpumask *cpu_map)
7930 for_each_cpu(i, cpu_map)
7931 cpu_attach_domain(NULL, &def_root_domain, i);
7935 /* handle null as "default" */
7936 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7937 struct sched_domain_attr *new, int idx_new)
7939 struct sched_domain_attr tmp;
7946 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7947 new ? (new + idx_new) : &tmp,
7948 sizeof(struct sched_domain_attr));
7952 * Partition sched domains as specified by the 'ndoms_new'
7953 * cpumasks in the array doms_new[] of cpumasks. This compares
7954 * doms_new[] to the current sched domain partitioning, doms_cur[].
7955 * It destroys each deleted domain and builds each new domain.
7957 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7958 * The masks don't intersect (don't overlap.) We should setup one
7959 * sched domain for each mask. CPUs not in any of the cpumasks will
7960 * not be load balanced. If the same cpumask appears both in the
7961 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7964 * The passed in 'doms_new' should be allocated using
7965 * alloc_sched_domains. This routine takes ownership of it and will
7966 * free_sched_domains it when done with it. If the caller failed the
7967 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7968 * and partition_sched_domains() will fallback to the single partition
7969 * 'fallback_doms', it also forces the domains to be rebuilt.
7971 * If doms_new == NULL it will be replaced with cpu_online_mask.
7972 * ndoms_new == 0 is a special case for destroying existing domains,
7973 * and it will not create the default domain.
7975 * Call with hotplug lock held
7977 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7978 struct sched_domain_attr *dattr_new)
7983 mutex_lock(&sched_domains_mutex);
7985 /* always unregister in case we don't destroy any domains */
7986 unregister_sched_domain_sysctl();
7988 /* Let architecture update cpu core mappings. */
7989 new_topology = arch_update_cpu_topology();
7991 n = doms_new ? ndoms_new : 0;
7993 /* Destroy deleted domains */
7994 for (i = 0; i < ndoms_cur; i++) {
7995 for (j = 0; j < n && !new_topology; j++) {
7996 if (cpumask_equal(doms_cur[i], doms_new[j])
7997 && dattrs_equal(dattr_cur, i, dattr_new, j))
8000 /* no match - a current sched domain not in new doms_new[] */
8001 detach_destroy_domains(doms_cur[i]);
8006 if (doms_new == NULL) {
8008 doms_new = &fallback_doms;
8009 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
8010 WARN_ON_ONCE(dattr_new);
8013 /* Build new domains */
8014 for (i = 0; i < ndoms_new; i++) {
8015 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8016 if (cpumask_equal(doms_new[i], doms_cur[j])
8017 && dattrs_equal(dattr_new, i, dattr_cur, j))
8020 /* no match - add a new doms_new */
8021 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
8026 /* Remember the new sched domains */
8027 if (doms_cur != &fallback_doms)
8028 free_sched_domains(doms_cur, ndoms_cur);
8029 kfree(dattr_cur); /* kfree(NULL) is safe */
8030 doms_cur = doms_new;
8031 dattr_cur = dattr_new;
8032 ndoms_cur = ndoms_new;
8034 register_sched_domain_sysctl();
8036 mutex_unlock(&sched_domains_mutex);
8039 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8040 static void reinit_sched_domains(void)
8044 /* Destroy domains first to force the rebuild */
8045 partition_sched_domains(0, NULL, NULL);
8047 rebuild_sched_domains();
8051 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8053 unsigned int level = 0;
8055 if (sscanf(buf, "%u", &level) != 1)
8059 * level is always be positive so don't check for
8060 * level < POWERSAVINGS_BALANCE_NONE which is 0
8061 * What happens on 0 or 1 byte write,
8062 * need to check for count as well?
8065 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8069 sched_smt_power_savings = level;
8071 sched_mc_power_savings = level;
8073 reinit_sched_domains();
8078 #ifdef CONFIG_SCHED_MC
8079 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8080 struct sysdev_class_attribute *attr,
8083 return sprintf(page, "%u\n", sched_mc_power_savings);
8085 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8086 struct sysdev_class_attribute *attr,
8087 const char *buf, size_t count)
8089 return sched_power_savings_store(buf, count, 0);
8091 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8092 sched_mc_power_savings_show,
8093 sched_mc_power_savings_store);
8096 #ifdef CONFIG_SCHED_SMT
8097 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8098 struct sysdev_class_attribute *attr,
8101 return sprintf(page, "%u\n", sched_smt_power_savings);
8103 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8104 struct sysdev_class_attribute *attr,
8105 const char *buf, size_t count)
8107 return sched_power_savings_store(buf, count, 1);
8109 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8110 sched_smt_power_savings_show,
8111 sched_smt_power_savings_store);
8114 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8118 #ifdef CONFIG_SCHED_SMT
8120 err = sysfs_create_file(&cls->kset.kobj,
8121 &attr_sched_smt_power_savings.attr);
8123 #ifdef CONFIG_SCHED_MC
8124 if (!err && mc_capable())
8125 err = sysfs_create_file(&cls->kset.kobj,
8126 &attr_sched_mc_power_savings.attr);
8130 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8133 * Update cpusets according to cpu_active mask. If cpusets are
8134 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8135 * around partition_sched_domains().
8137 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
8140 switch (action & ~CPU_TASKS_FROZEN) {
8142 case CPU_DOWN_FAILED:
8143 cpuset_update_active_cpus();
8150 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
8153 switch (action & ~CPU_TASKS_FROZEN) {
8154 case CPU_DOWN_PREPARE:
8155 cpuset_update_active_cpus();
8162 static int update_runtime(struct notifier_block *nfb,
8163 unsigned long action, void *hcpu)
8165 int cpu = (int)(long)hcpu;
8168 case CPU_DOWN_PREPARE:
8169 case CPU_DOWN_PREPARE_FROZEN:
8170 disable_runtime(cpu_rq(cpu));
8173 case CPU_DOWN_FAILED:
8174 case CPU_DOWN_FAILED_FROZEN:
8176 case CPU_ONLINE_FROZEN:
8177 enable_runtime(cpu_rq(cpu));
8185 void __init sched_init_smp(void)
8187 cpumask_var_t non_isolated_cpus;
8189 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8190 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8193 mutex_lock(&sched_domains_mutex);
8194 init_sched_domains(cpu_active_mask);
8195 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8196 if (cpumask_empty(non_isolated_cpus))
8197 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8198 mutex_unlock(&sched_domains_mutex);
8201 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
8202 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
8204 /* RT runtime code needs to handle some hotplug events */
8205 hotcpu_notifier(update_runtime, 0);
8209 /* Move init over to a non-isolated CPU */
8210 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8212 sched_init_granularity();
8213 free_cpumask_var(non_isolated_cpus);
8215 init_sched_rt_class();
8218 void __init sched_init_smp(void)
8220 sched_init_granularity();
8222 #endif /* CONFIG_SMP */
8224 const_debug unsigned int sysctl_timer_migration = 1;
8226 int in_sched_functions(unsigned long addr)
8228 return in_lock_functions(addr) ||
8229 (addr >= (unsigned long)__sched_text_start
8230 && addr < (unsigned long)__sched_text_end);
8233 static void init_cfs_rq(struct cfs_rq *cfs_rq)
8235 cfs_rq->tasks_timeline = RB_ROOT;
8236 INIT_LIST_HEAD(&cfs_rq->tasks);
8237 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8238 #ifndef CONFIG_64BIT
8239 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8243 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8245 struct rt_prio_array *array;
8248 array = &rt_rq->active;
8249 for (i = 0; i < MAX_RT_PRIO; i++) {
8250 INIT_LIST_HEAD(array->queue + i);
8251 __clear_bit(i, array->bitmap);
8253 /* delimiter for bitsearch: */
8254 __set_bit(MAX_RT_PRIO, array->bitmap);
8256 #if defined CONFIG_SMP
8257 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8258 rt_rq->highest_prio.next = MAX_RT_PRIO;
8259 rt_rq->rt_nr_migratory = 0;
8260 rt_rq->overloaded = 0;
8261 plist_head_init(&rt_rq->pushable_tasks);
8265 rt_rq->rt_throttled = 0;
8266 rt_rq->rt_runtime = 0;
8267 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8270 #ifdef CONFIG_FAIR_GROUP_SCHED
8271 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8272 struct sched_entity *se, int cpu,
8273 struct sched_entity *parent)
8275 struct rq *rq = cpu_rq(cpu);
8280 /* allow initial update_cfs_load() to truncate */
8281 cfs_rq->load_stamp = 1;
8283 init_cfs_rq_runtime(cfs_rq);
8285 tg->cfs_rq[cpu] = cfs_rq;
8288 /* se could be NULL for root_task_group */
8293 se->cfs_rq = &rq->cfs;
8295 se->cfs_rq = parent->my_q;
8298 update_load_set(&se->load, 0);
8299 se->parent = parent;
8303 #ifdef CONFIG_RT_GROUP_SCHED
8304 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8305 struct sched_rt_entity *rt_se, int cpu,
8306 struct sched_rt_entity *parent)
8308 struct rq *rq = cpu_rq(cpu);
8310 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8311 rt_rq->rt_nr_boosted = 0;
8315 tg->rt_rq[cpu] = rt_rq;
8316 tg->rt_se[cpu] = rt_se;
8322 rt_se->rt_rq = &rq->rt;
8324 rt_se->rt_rq = parent->my_q;
8326 rt_se->my_q = rt_rq;
8327 rt_se->parent = parent;
8328 INIT_LIST_HEAD(&rt_se->run_list);
8332 void __init sched_init(void)
8335 unsigned long alloc_size = 0, ptr;
8337 #ifdef CONFIG_FAIR_GROUP_SCHED
8338 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8340 #ifdef CONFIG_RT_GROUP_SCHED
8341 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8343 #ifdef CONFIG_CPUMASK_OFFSTACK
8344 alloc_size += num_possible_cpus() * cpumask_size();
8347 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8349 #ifdef CONFIG_FAIR_GROUP_SCHED
8350 root_task_group.se = (struct sched_entity **)ptr;
8351 ptr += nr_cpu_ids * sizeof(void **);
8353 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8354 ptr += nr_cpu_ids * sizeof(void **);
8356 #endif /* CONFIG_FAIR_GROUP_SCHED */
8357 #ifdef CONFIG_RT_GROUP_SCHED
8358 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8359 ptr += nr_cpu_ids * sizeof(void **);
8361 root_task_group.rt_rq = (struct rt_rq **)ptr;
8362 ptr += nr_cpu_ids * sizeof(void **);
8364 #endif /* CONFIG_RT_GROUP_SCHED */
8365 #ifdef CONFIG_CPUMASK_OFFSTACK
8366 for_each_possible_cpu(i) {
8367 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8368 ptr += cpumask_size();
8370 #endif /* CONFIG_CPUMASK_OFFSTACK */
8374 init_defrootdomain();
8377 init_rt_bandwidth(&def_rt_bandwidth,
8378 global_rt_period(), global_rt_runtime());
8380 #ifdef CONFIG_RT_GROUP_SCHED
8381 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8382 global_rt_period(), global_rt_runtime());
8383 #endif /* CONFIG_RT_GROUP_SCHED */
8385 #ifdef CONFIG_CGROUP_SCHED
8386 list_add(&root_task_group.list, &task_groups);
8387 INIT_LIST_HEAD(&root_task_group.children);
8388 autogroup_init(&init_task);
8389 #endif /* CONFIG_CGROUP_SCHED */
8391 for_each_possible_cpu(i) {
8395 raw_spin_lock_init(&rq->lock);
8397 rq->calc_load_active = 0;
8398 rq->calc_load_update = jiffies + LOAD_FREQ;
8399 init_cfs_rq(&rq->cfs);
8400 init_rt_rq(&rq->rt, rq);
8401 #ifdef CONFIG_FAIR_GROUP_SCHED
8402 root_task_group.shares = root_task_group_load;
8403 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8405 * How much cpu bandwidth does root_task_group get?
8407 * In case of task-groups formed thr' the cgroup filesystem, it
8408 * gets 100% of the cpu resources in the system. This overall
8409 * system cpu resource is divided among the tasks of
8410 * root_task_group and its child task-groups in a fair manner,
8411 * based on each entity's (task or task-group's) weight
8412 * (se->load.weight).
8414 * In other words, if root_task_group has 10 tasks of weight
8415 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8416 * then A0's share of the cpu resource is:
8418 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8420 * We achieve this by letting root_task_group's tasks sit
8421 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8423 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8424 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8425 #endif /* CONFIG_FAIR_GROUP_SCHED */
8427 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8428 #ifdef CONFIG_RT_GROUP_SCHED
8429 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8430 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8433 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8434 rq->cpu_load[j] = 0;
8436 rq->last_load_update_tick = jiffies;
8441 rq->cpu_power = SCHED_POWER_SCALE;
8442 rq->post_schedule = 0;
8443 rq->active_balance = 0;
8444 rq->next_balance = jiffies;
8449 rq->avg_idle = 2*sysctl_sched_migration_cost;
8450 rq_attach_root(rq, &def_root_domain);
8452 rq->nohz_balance_kick = 0;
8456 atomic_set(&rq->nr_iowait, 0);
8459 set_load_weight(&init_task);
8461 #ifdef CONFIG_PREEMPT_NOTIFIERS
8462 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8466 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8469 #ifdef CONFIG_RT_MUTEXES
8470 plist_head_init(&init_task.pi_waiters);
8474 * The boot idle thread does lazy MMU switching as well:
8476 atomic_inc(&init_mm.mm_count);
8477 enter_lazy_tlb(&init_mm, current);
8480 * Make us the idle thread. Technically, schedule() should not be
8481 * called from this thread, however somewhere below it might be,
8482 * but because we are the idle thread, we just pick up running again
8483 * when this runqueue becomes "idle".
8485 init_idle(current, smp_processor_id());
8487 calc_load_update = jiffies + LOAD_FREQ;
8490 * During early bootup we pretend to be a normal task:
8492 current->sched_class = &fair_sched_class;
8495 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8497 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8498 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8499 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8500 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8501 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8503 /* May be allocated at isolcpus cmdline parse time */
8504 if (cpu_isolated_map == NULL)
8505 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8508 scheduler_running = 1;
8511 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8512 static inline int preempt_count_equals(int preempt_offset)
8514 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8516 return (nested == preempt_offset);
8519 void __might_sleep(const char *file, int line, int preempt_offset)
8521 static unsigned long prev_jiffy; /* ratelimiting */
8523 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
8524 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8525 system_state != SYSTEM_RUNNING || oops_in_progress)
8527 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8529 prev_jiffy = jiffies;
8532 "BUG: sleeping function called from invalid context at %s:%d\n",
8535 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8536 in_atomic(), irqs_disabled(),
8537 current->pid, current->comm);
8539 debug_show_held_locks(current);
8540 if (irqs_disabled())
8541 print_irqtrace_events(current);
8544 EXPORT_SYMBOL(__might_sleep);
8547 #ifdef CONFIG_MAGIC_SYSRQ
8548 static void normalize_task(struct rq *rq, struct task_struct *p)
8550 const struct sched_class *prev_class = p->sched_class;
8551 int old_prio = p->prio;
8556 deactivate_task(rq, p, 0);
8557 __setscheduler(rq, p, SCHED_NORMAL, 0);
8559 activate_task(rq, p, 0);
8560 resched_task(rq->curr);
8563 check_class_changed(rq, p, prev_class, old_prio);
8566 void normalize_rt_tasks(void)
8568 struct task_struct *g, *p;
8569 unsigned long flags;
8572 read_lock_irqsave(&tasklist_lock, flags);
8573 do_each_thread(g, p) {
8575 * Only normalize user tasks:
8580 p->se.exec_start = 0;
8581 #ifdef CONFIG_SCHEDSTATS
8582 p->se.statistics.wait_start = 0;
8583 p->se.statistics.sleep_start = 0;
8584 p->se.statistics.block_start = 0;
8589 * Renice negative nice level userspace
8592 if (TASK_NICE(p) < 0 && p->mm)
8593 set_user_nice(p, 0);
8597 raw_spin_lock(&p->pi_lock);
8598 rq = __task_rq_lock(p);
8600 normalize_task(rq, p);
8602 __task_rq_unlock(rq);
8603 raw_spin_unlock(&p->pi_lock);
8604 } while_each_thread(g, p);
8606 read_unlock_irqrestore(&tasklist_lock, flags);
8609 #endif /* CONFIG_MAGIC_SYSRQ */
8611 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8613 * These functions are only useful for the IA64 MCA handling, or kdb.
8615 * They can only be called when the whole system has been
8616 * stopped - every CPU needs to be quiescent, and no scheduling
8617 * activity can take place. Using them for anything else would
8618 * be a serious bug, and as a result, they aren't even visible
8619 * under any other configuration.
8623 * curr_task - return the current task for a given cpu.
8624 * @cpu: the processor in question.
8626 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8628 struct task_struct *curr_task(int cpu)
8630 return cpu_curr(cpu);
8633 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8637 * set_curr_task - set the current task for a given cpu.
8638 * @cpu: the processor in question.
8639 * @p: the task pointer to set.
8641 * Description: This function must only be used when non-maskable interrupts
8642 * are serviced on a separate stack. It allows the architecture to switch the
8643 * notion of the current task on a cpu in a non-blocking manner. This function
8644 * must be called with all CPU's synchronized, and interrupts disabled, the
8645 * and caller must save the original value of the current task (see
8646 * curr_task() above) and restore that value before reenabling interrupts and
8647 * re-starting the system.
8649 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8651 void set_curr_task(int cpu, struct task_struct *p)
8658 #ifdef CONFIG_FAIR_GROUP_SCHED
8659 static void free_fair_sched_group(struct task_group *tg)
8663 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8665 for_each_possible_cpu(i) {
8667 kfree(tg->cfs_rq[i]);
8677 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8679 struct cfs_rq *cfs_rq;
8680 struct sched_entity *se;
8683 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8686 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8690 tg->shares = NICE_0_LOAD;
8692 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8694 for_each_possible_cpu(i) {
8695 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8696 GFP_KERNEL, cpu_to_node(i));
8700 se = kzalloc_node(sizeof(struct sched_entity),
8701 GFP_KERNEL, cpu_to_node(i));
8705 init_cfs_rq(cfs_rq);
8706 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8717 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8719 struct rq *rq = cpu_rq(cpu);
8720 unsigned long flags;
8723 * Only empty task groups can be destroyed; so we can speculatively
8724 * check on_list without danger of it being re-added.
8726 if (!tg->cfs_rq[cpu]->on_list)
8729 raw_spin_lock_irqsave(&rq->lock, flags);
8730 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8731 raw_spin_unlock_irqrestore(&rq->lock, flags);
8733 #else /* !CONFIG_FAIR_GROUP_SCHED */
8734 static inline void free_fair_sched_group(struct task_group *tg)
8739 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8744 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8747 #endif /* CONFIG_FAIR_GROUP_SCHED */
8749 #ifdef CONFIG_RT_GROUP_SCHED
8750 static void free_rt_sched_group(struct task_group *tg)
8755 destroy_rt_bandwidth(&tg->rt_bandwidth);
8757 for_each_possible_cpu(i) {
8759 kfree(tg->rt_rq[i]);
8761 kfree(tg->rt_se[i]);
8769 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8771 struct rt_rq *rt_rq;
8772 struct sched_rt_entity *rt_se;
8775 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8778 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8782 init_rt_bandwidth(&tg->rt_bandwidth,
8783 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8785 for_each_possible_cpu(i) {
8786 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8787 GFP_KERNEL, cpu_to_node(i));
8791 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8792 GFP_KERNEL, cpu_to_node(i));
8796 init_rt_rq(rt_rq, cpu_rq(i));
8797 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8798 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8808 #else /* !CONFIG_RT_GROUP_SCHED */
8809 static inline void free_rt_sched_group(struct task_group *tg)
8814 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8818 #endif /* CONFIG_RT_GROUP_SCHED */
8820 #ifdef CONFIG_CGROUP_SCHED
8821 static void free_sched_group(struct task_group *tg)
8823 free_fair_sched_group(tg);
8824 free_rt_sched_group(tg);
8829 /* allocate runqueue etc for a new task group */
8830 struct task_group *sched_create_group(struct task_group *parent)
8832 struct task_group *tg;
8833 unsigned long flags;
8835 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8837 return ERR_PTR(-ENOMEM);
8839 if (!alloc_fair_sched_group(tg, parent))
8842 if (!alloc_rt_sched_group(tg, parent))
8845 spin_lock_irqsave(&task_group_lock, flags);
8846 list_add_rcu(&tg->list, &task_groups);
8848 WARN_ON(!parent); /* root should already exist */
8850 tg->parent = parent;
8851 INIT_LIST_HEAD(&tg->children);
8852 list_add_rcu(&tg->siblings, &parent->children);
8853 spin_unlock_irqrestore(&task_group_lock, flags);
8858 free_sched_group(tg);
8859 return ERR_PTR(-ENOMEM);
8862 /* rcu callback to free various structures associated with a task group */
8863 static void free_sched_group_rcu(struct rcu_head *rhp)
8865 /* now it should be safe to free those cfs_rqs */
8866 free_sched_group(container_of(rhp, struct task_group, rcu));
8869 /* Destroy runqueue etc associated with a task group */
8870 void sched_destroy_group(struct task_group *tg)
8872 unsigned long flags;
8875 /* end participation in shares distribution */
8876 for_each_possible_cpu(i)
8877 unregister_fair_sched_group(tg, i);
8879 spin_lock_irqsave(&task_group_lock, flags);
8880 list_del_rcu(&tg->list);
8881 list_del_rcu(&tg->siblings);
8882 spin_unlock_irqrestore(&task_group_lock, flags);
8884 /* wait for possible concurrent references to cfs_rqs complete */
8885 call_rcu(&tg->rcu, free_sched_group_rcu);
8888 /* change task's runqueue when it moves between groups.
8889 * The caller of this function should have put the task in its new group
8890 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8891 * reflect its new group.
8893 void sched_move_task(struct task_struct *tsk)
8896 unsigned long flags;
8899 rq = task_rq_lock(tsk, &flags);
8901 running = task_current(rq, tsk);
8905 dequeue_task(rq, tsk, 0);
8906 if (unlikely(running))
8907 tsk->sched_class->put_prev_task(rq, tsk);
8909 #ifdef CONFIG_FAIR_GROUP_SCHED
8910 if (tsk->sched_class->task_move_group)
8911 tsk->sched_class->task_move_group(tsk, on_rq);
8914 set_task_rq(tsk, task_cpu(tsk));
8916 if (unlikely(running))
8917 tsk->sched_class->set_curr_task(rq);
8919 enqueue_task(rq, tsk, 0);
8921 task_rq_unlock(rq, tsk, &flags);
8923 #endif /* CONFIG_CGROUP_SCHED */
8925 #ifdef CONFIG_FAIR_GROUP_SCHED
8926 static DEFINE_MUTEX(shares_mutex);
8928 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8931 unsigned long flags;
8934 * We can't change the weight of the root cgroup.
8939 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8941 mutex_lock(&shares_mutex);
8942 if (tg->shares == shares)
8945 tg->shares = shares;
8946 for_each_possible_cpu(i) {
8947 struct rq *rq = cpu_rq(i);
8948 struct sched_entity *se;
8951 /* Propagate contribution to hierarchy */
8952 raw_spin_lock_irqsave(&rq->lock, flags);
8953 for_each_sched_entity(se)
8954 update_cfs_shares(group_cfs_rq(se));
8955 raw_spin_unlock_irqrestore(&rq->lock, flags);
8959 mutex_unlock(&shares_mutex);
8963 unsigned long sched_group_shares(struct task_group *tg)
8969 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
8970 static unsigned long to_ratio(u64 period, u64 runtime)
8972 if (runtime == RUNTIME_INF)
8975 return div64_u64(runtime << 20, period);
8979 #ifdef CONFIG_RT_GROUP_SCHED
8981 * Ensure that the real time constraints are schedulable.
8983 static DEFINE_MUTEX(rt_constraints_mutex);
8985 /* Must be called with tasklist_lock held */
8986 static inline int tg_has_rt_tasks(struct task_group *tg)
8988 struct task_struct *g, *p;
8990 do_each_thread(g, p) {
8991 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8993 } while_each_thread(g, p);
8998 struct rt_schedulable_data {
8999 struct task_group *tg;
9004 static int tg_rt_schedulable(struct task_group *tg, void *data)
9006 struct rt_schedulable_data *d = data;
9007 struct task_group *child;
9008 unsigned long total, sum = 0;
9009 u64 period, runtime;
9011 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9012 runtime = tg->rt_bandwidth.rt_runtime;
9015 period = d->rt_period;
9016 runtime = d->rt_runtime;
9020 * Cannot have more runtime than the period.
9022 if (runtime > period && runtime != RUNTIME_INF)
9026 * Ensure we don't starve existing RT tasks.
9028 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9031 total = to_ratio(period, runtime);
9034 * Nobody can have more than the global setting allows.
9036 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9040 * The sum of our children's runtime should not exceed our own.
9042 list_for_each_entry_rcu(child, &tg->children, siblings) {
9043 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9044 runtime = child->rt_bandwidth.rt_runtime;
9046 if (child == d->tg) {
9047 period = d->rt_period;
9048 runtime = d->rt_runtime;
9051 sum += to_ratio(period, runtime);
9060 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9064 struct rt_schedulable_data data = {
9066 .rt_period = period,
9067 .rt_runtime = runtime,
9071 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
9077 static int tg_set_rt_bandwidth(struct task_group *tg,
9078 u64 rt_period, u64 rt_runtime)
9082 mutex_lock(&rt_constraints_mutex);
9083 read_lock(&tasklist_lock);
9084 err = __rt_schedulable(tg, rt_period, rt_runtime);
9088 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9089 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9090 tg->rt_bandwidth.rt_runtime = rt_runtime;
9092 for_each_possible_cpu(i) {
9093 struct rt_rq *rt_rq = tg->rt_rq[i];
9095 raw_spin_lock(&rt_rq->rt_runtime_lock);
9096 rt_rq->rt_runtime = rt_runtime;
9097 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9099 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9101 read_unlock(&tasklist_lock);
9102 mutex_unlock(&rt_constraints_mutex);
9107 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9109 u64 rt_runtime, rt_period;
9111 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9112 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9113 if (rt_runtime_us < 0)
9114 rt_runtime = RUNTIME_INF;
9116 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9119 long sched_group_rt_runtime(struct task_group *tg)
9123 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9126 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9127 do_div(rt_runtime_us, NSEC_PER_USEC);
9128 return rt_runtime_us;
9131 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9133 u64 rt_runtime, rt_period;
9135 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9136 rt_runtime = tg->rt_bandwidth.rt_runtime;
9141 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9144 long sched_group_rt_period(struct task_group *tg)
9148 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9149 do_div(rt_period_us, NSEC_PER_USEC);
9150 return rt_period_us;
9153 static int sched_rt_global_constraints(void)
9155 u64 runtime, period;
9158 if (sysctl_sched_rt_period <= 0)
9161 runtime = global_rt_runtime();
9162 period = global_rt_period();
9165 * Sanity check on the sysctl variables.
9167 if (runtime > period && runtime != RUNTIME_INF)
9170 mutex_lock(&rt_constraints_mutex);
9171 read_lock(&tasklist_lock);
9172 ret = __rt_schedulable(NULL, 0, 0);
9173 read_unlock(&tasklist_lock);
9174 mutex_unlock(&rt_constraints_mutex);
9179 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9181 /* Don't accept realtime tasks when there is no way for them to run */
9182 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9188 #else /* !CONFIG_RT_GROUP_SCHED */
9189 static int sched_rt_global_constraints(void)
9191 unsigned long flags;
9194 if (sysctl_sched_rt_period <= 0)
9198 * There's always some RT tasks in the root group
9199 * -- migration, kstopmachine etc..
9201 if (sysctl_sched_rt_runtime == 0)
9204 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9205 for_each_possible_cpu(i) {
9206 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9208 raw_spin_lock(&rt_rq->rt_runtime_lock);
9209 rt_rq->rt_runtime = global_rt_runtime();
9210 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9212 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9216 #endif /* CONFIG_RT_GROUP_SCHED */
9218 int sched_rt_handler(struct ctl_table *table, int write,
9219 void __user *buffer, size_t *lenp,
9223 int old_period, old_runtime;
9224 static DEFINE_MUTEX(mutex);
9227 old_period = sysctl_sched_rt_period;
9228 old_runtime = sysctl_sched_rt_runtime;
9230 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9232 if (!ret && write) {
9233 ret = sched_rt_global_constraints();
9235 sysctl_sched_rt_period = old_period;
9236 sysctl_sched_rt_runtime = old_runtime;
9238 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9239 def_rt_bandwidth.rt_period =
9240 ns_to_ktime(global_rt_period());
9243 mutex_unlock(&mutex);
9248 #ifdef CONFIG_CGROUP_SCHED
9250 /* return corresponding task_group object of a cgroup */
9251 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9253 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9254 struct task_group, css);
9257 static struct cgroup_subsys_state *
9258 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9260 struct task_group *tg, *parent;
9262 if (!cgrp->parent) {
9263 /* This is early initialization for the top cgroup */
9264 return &root_task_group.css;
9267 parent = cgroup_tg(cgrp->parent);
9268 tg = sched_create_group(parent);
9270 return ERR_PTR(-ENOMEM);
9276 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9278 struct task_group *tg = cgroup_tg(cgrp);
9280 sched_destroy_group(tg);
9284 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9286 #ifdef CONFIG_RT_GROUP_SCHED
9287 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9290 /* We don't support RT-tasks being in separate groups */
9291 if (tsk->sched_class != &fair_sched_class)
9298 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9300 sched_move_task(tsk);
9304 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9305 struct cgroup *old_cgrp, struct task_struct *task)
9308 * cgroup_exit() is called in the copy_process() failure path.
9309 * Ignore this case since the task hasn't ran yet, this avoids
9310 * trying to poke a half freed task state from generic code.
9312 if (!(task->flags & PF_EXITING))
9315 sched_move_task(task);
9318 #ifdef CONFIG_FAIR_GROUP_SCHED
9319 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9322 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9325 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9327 struct task_group *tg = cgroup_tg(cgrp);
9329 return (u64) scale_load_down(tg->shares);
9332 #ifdef CONFIG_CFS_BANDWIDTH
9333 static DEFINE_MUTEX(cfs_constraints_mutex);
9335 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9336 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9338 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9340 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9342 int i, ret = 0, runtime_enabled;
9343 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9345 if (tg == &root_task_group)
9349 * Ensure we have at some amount of bandwidth every period. This is
9350 * to prevent reaching a state of large arrears when throttled via
9351 * entity_tick() resulting in prolonged exit starvation.
9353 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9357 * Likewise, bound things on the otherside by preventing insane quota
9358 * periods. This also allows us to normalize in computing quota
9361 if (period > max_cfs_quota_period)
9364 mutex_lock(&cfs_constraints_mutex);
9365 ret = __cfs_schedulable(tg, period, quota);
9369 runtime_enabled = quota != RUNTIME_INF;
9370 raw_spin_lock_irq(&cfs_b->lock);
9371 cfs_b->period = ns_to_ktime(period);
9372 cfs_b->quota = quota;
9374 __refill_cfs_bandwidth_runtime(cfs_b);
9375 /* restart the period timer (if active) to handle new period expiry */
9376 if (runtime_enabled && cfs_b->timer_active) {
9377 /* force a reprogram */
9378 cfs_b->timer_active = 0;
9379 __start_cfs_bandwidth(cfs_b);
9381 raw_spin_unlock_irq(&cfs_b->lock);
9383 for_each_possible_cpu(i) {
9384 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9385 struct rq *rq = rq_of(cfs_rq);
9387 raw_spin_lock_irq(&rq->lock);
9388 cfs_rq->runtime_enabled = runtime_enabled;
9389 cfs_rq->runtime_remaining = 0;
9391 if (cfs_rq_throttled(cfs_rq))
9392 unthrottle_cfs_rq(cfs_rq);
9393 raw_spin_unlock_irq(&rq->lock);
9396 mutex_unlock(&cfs_constraints_mutex);
9401 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9405 period = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9406 if (cfs_quota_us < 0)
9407 quota = RUNTIME_INF;
9409 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9411 return tg_set_cfs_bandwidth(tg, period, quota);
9414 long tg_get_cfs_quota(struct task_group *tg)
9418 if (tg_cfs_bandwidth(tg)->quota == RUNTIME_INF)
9421 quota_us = tg_cfs_bandwidth(tg)->quota;
9422 do_div(quota_us, NSEC_PER_USEC);
9427 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9431 period = (u64)cfs_period_us * NSEC_PER_USEC;
9432 quota = tg_cfs_bandwidth(tg)->quota;
9437 return tg_set_cfs_bandwidth(tg, period, quota);
9440 long tg_get_cfs_period(struct task_group *tg)
9444 cfs_period_us = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9445 do_div(cfs_period_us, NSEC_PER_USEC);
9447 return cfs_period_us;
9450 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
9452 return tg_get_cfs_quota(cgroup_tg(cgrp));
9455 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
9458 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
9461 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
9463 return tg_get_cfs_period(cgroup_tg(cgrp));
9466 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9469 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
9472 struct cfs_schedulable_data {
9473 struct task_group *tg;
9478 * normalize group quota/period to be quota/max_period
9479 * note: units are usecs
9481 static u64 normalize_cfs_quota(struct task_group *tg,
9482 struct cfs_schedulable_data *d)
9490 period = tg_get_cfs_period(tg);
9491 quota = tg_get_cfs_quota(tg);
9494 /* note: these should typically be equivalent */
9495 if (quota == RUNTIME_INF || quota == -1)
9498 return to_ratio(period, quota);
9501 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9503 struct cfs_schedulable_data *d = data;
9504 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9505 s64 quota = 0, parent_quota = -1;
9508 quota = RUNTIME_INF;
9510 struct cfs_bandwidth *parent_b = tg_cfs_bandwidth(tg->parent);
9512 quota = normalize_cfs_quota(tg, d);
9513 parent_quota = parent_b->hierarchal_quota;
9516 * ensure max(child_quota) <= parent_quota, inherit when no
9519 if (quota == RUNTIME_INF)
9520 quota = parent_quota;
9521 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9524 cfs_b->hierarchal_quota = quota;
9529 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9532 struct cfs_schedulable_data data = {
9538 if (quota != RUNTIME_INF) {
9539 do_div(data.period, NSEC_PER_USEC);
9540 do_div(data.quota, NSEC_PER_USEC);
9544 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9550 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
9551 struct cgroup_map_cb *cb)
9553 struct task_group *tg = cgroup_tg(cgrp);
9554 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9556 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
9557 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
9558 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
9562 #endif /* CONFIG_CFS_BANDWIDTH */
9563 #endif /* CONFIG_FAIR_GROUP_SCHED */
9565 #ifdef CONFIG_RT_GROUP_SCHED
9566 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9569 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9572 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9574 return sched_group_rt_runtime(cgroup_tg(cgrp));
9577 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9580 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9583 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9585 return sched_group_rt_period(cgroup_tg(cgrp));
9587 #endif /* CONFIG_RT_GROUP_SCHED */
9589 static struct cftype cpu_files[] = {
9590 #ifdef CONFIG_FAIR_GROUP_SCHED
9593 .read_u64 = cpu_shares_read_u64,
9594 .write_u64 = cpu_shares_write_u64,
9597 #ifdef CONFIG_CFS_BANDWIDTH
9599 .name = "cfs_quota_us",
9600 .read_s64 = cpu_cfs_quota_read_s64,
9601 .write_s64 = cpu_cfs_quota_write_s64,
9604 .name = "cfs_period_us",
9605 .read_u64 = cpu_cfs_period_read_u64,
9606 .write_u64 = cpu_cfs_period_write_u64,
9610 .read_map = cpu_stats_show,
9613 #ifdef CONFIG_RT_GROUP_SCHED
9615 .name = "rt_runtime_us",
9616 .read_s64 = cpu_rt_runtime_read,
9617 .write_s64 = cpu_rt_runtime_write,
9620 .name = "rt_period_us",
9621 .read_u64 = cpu_rt_period_read_uint,
9622 .write_u64 = cpu_rt_period_write_uint,
9627 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9629 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9632 struct cgroup_subsys cpu_cgroup_subsys = {
9634 .create = cpu_cgroup_create,
9635 .destroy = cpu_cgroup_destroy,
9636 .can_attach_task = cpu_cgroup_can_attach_task,
9637 .attach_task = cpu_cgroup_attach_task,
9638 .exit = cpu_cgroup_exit,
9639 .populate = cpu_cgroup_populate,
9640 .subsys_id = cpu_cgroup_subsys_id,
9644 #endif /* CONFIG_CGROUP_SCHED */
9646 #ifdef CONFIG_CGROUP_CPUACCT
9649 * CPU accounting code for task groups.
9651 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9652 * (balbir@in.ibm.com).
9655 /* track cpu usage of a group of tasks and its child groups */
9657 struct cgroup_subsys_state css;
9658 /* cpuusage holds pointer to a u64-type object on every cpu */
9659 u64 __percpu *cpuusage;
9660 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9661 struct cpuacct *parent;
9664 struct cgroup_subsys cpuacct_subsys;
9666 /* return cpu accounting group corresponding to this container */
9667 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9669 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9670 struct cpuacct, css);
9673 /* return cpu accounting group to which this task belongs */
9674 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9676 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9677 struct cpuacct, css);
9680 /* create a new cpu accounting group */
9681 static struct cgroup_subsys_state *cpuacct_create(
9682 struct cgroup_subsys *ss, struct cgroup *cgrp)
9684 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9690 ca->cpuusage = alloc_percpu(u64);
9694 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9695 if (percpu_counter_init(&ca->cpustat[i], 0))
9696 goto out_free_counters;
9699 ca->parent = cgroup_ca(cgrp->parent);
9705 percpu_counter_destroy(&ca->cpustat[i]);
9706 free_percpu(ca->cpuusage);
9710 return ERR_PTR(-ENOMEM);
9713 /* destroy an existing cpu accounting group */
9715 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9717 struct cpuacct *ca = cgroup_ca(cgrp);
9720 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9721 percpu_counter_destroy(&ca->cpustat[i]);
9722 free_percpu(ca->cpuusage);
9726 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9728 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9731 #ifndef CONFIG_64BIT
9733 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9735 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9737 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9745 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9747 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9749 #ifndef CONFIG_64BIT
9751 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9753 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9755 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9761 /* return total cpu usage (in nanoseconds) of a group */
9762 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9764 struct cpuacct *ca = cgroup_ca(cgrp);
9765 u64 totalcpuusage = 0;
9768 for_each_present_cpu(i)
9769 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9771 return totalcpuusage;
9774 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9777 struct cpuacct *ca = cgroup_ca(cgrp);
9786 for_each_present_cpu(i)
9787 cpuacct_cpuusage_write(ca, i, 0);
9793 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9796 struct cpuacct *ca = cgroup_ca(cgroup);
9800 for_each_present_cpu(i) {
9801 percpu = cpuacct_cpuusage_read(ca, i);
9802 seq_printf(m, "%llu ", (unsigned long long) percpu);
9804 seq_printf(m, "\n");
9808 static const char *cpuacct_stat_desc[] = {
9809 [CPUACCT_STAT_USER] = "user",
9810 [CPUACCT_STAT_SYSTEM] = "system",
9813 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9814 struct cgroup_map_cb *cb)
9816 struct cpuacct *ca = cgroup_ca(cgrp);
9819 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9820 s64 val = percpu_counter_read(&ca->cpustat[i]);
9821 val = cputime64_to_clock_t(val);
9822 cb->fill(cb, cpuacct_stat_desc[i], val);
9827 static struct cftype files[] = {
9830 .read_u64 = cpuusage_read,
9831 .write_u64 = cpuusage_write,
9834 .name = "usage_percpu",
9835 .read_seq_string = cpuacct_percpu_seq_read,
9839 .read_map = cpuacct_stats_show,
9843 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9845 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9849 * charge this task's execution time to its accounting group.
9851 * called with rq->lock held.
9853 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9858 if (unlikely(!cpuacct_subsys.active))
9861 cpu = task_cpu(tsk);
9867 for (; ca; ca = ca->parent) {
9868 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9869 *cpuusage += cputime;
9876 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9877 * in cputime_t units. As a result, cpuacct_update_stats calls
9878 * percpu_counter_add with values large enough to always overflow the
9879 * per cpu batch limit causing bad SMP scalability.
9881 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9882 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9883 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9886 #define CPUACCT_BATCH \
9887 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9889 #define CPUACCT_BATCH 0
9893 * Charge the system/user time to the task's accounting group.
9895 static void cpuacct_update_stats(struct task_struct *tsk,
9896 enum cpuacct_stat_index idx, cputime_t val)
9899 int batch = CPUACCT_BATCH;
9901 if (unlikely(!cpuacct_subsys.active))
9908 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9914 struct cgroup_subsys cpuacct_subsys = {
9916 .create = cpuacct_create,
9917 .destroy = cpuacct_destroy,
9918 .populate = cpuacct_populate,
9919 .subsys_id = cpuacct_subsys_id,
9921 #endif /* CONFIG_CGROUP_CPUACCT */