4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
74 #include <linux/init_task.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
79 #ifdef CONFIG_PARAVIRT
80 #include <asm/paravirt.h>
83 #include "sched_cpupri.h"
84 #include "workqueue_sched.h"
85 #include "sched_autogroup.h"
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/sched.h>
91 * Convert user-nice values [ -20 ... 0 ... 19 ]
92 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
95 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
96 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
97 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
100 * 'User priority' is the nice value converted to something we
101 * can work with better when scaling various scheduler parameters,
102 * it's a [ 0 ... 39 ] range.
104 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
105 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
106 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
109 * Helpers for converting nanosecond timing to jiffy resolution
111 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
113 #define NICE_0_LOAD SCHED_LOAD_SCALE
114 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
117 * These are the 'tuning knobs' of the scheduler:
119 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
120 * Timeslices get refilled after they expire.
122 #define DEF_TIMESLICE (100 * HZ / 1000)
125 * single value that denotes runtime == period, ie unlimited time.
127 #define RUNTIME_INF ((u64)~0ULL)
129 static inline int rt_policy(int policy)
131 if (policy == SCHED_FIFO || policy == SCHED_RR)
136 static inline int task_has_rt_policy(struct task_struct *p)
138 return rt_policy(p->policy);
142 * This is the priority-queue data structure of the RT scheduling class:
144 struct rt_prio_array {
145 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
146 struct list_head queue[MAX_RT_PRIO];
149 struct rt_bandwidth {
150 /* nests inside the rq lock: */
151 raw_spinlock_t rt_runtime_lock;
154 struct hrtimer rt_period_timer;
157 static struct rt_bandwidth def_rt_bandwidth;
159 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
161 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
163 struct rt_bandwidth *rt_b =
164 container_of(timer, struct rt_bandwidth, rt_period_timer);
170 now = hrtimer_cb_get_time(timer);
171 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
176 idle = do_sched_rt_period_timer(rt_b, overrun);
179 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
183 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
185 rt_b->rt_period = ns_to_ktime(period);
186 rt_b->rt_runtime = runtime;
188 raw_spin_lock_init(&rt_b->rt_runtime_lock);
190 hrtimer_init(&rt_b->rt_period_timer,
191 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
192 rt_b->rt_period_timer.function = sched_rt_period_timer;
195 static inline int rt_bandwidth_enabled(void)
197 return sysctl_sched_rt_runtime >= 0;
200 static void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
203 ktime_t soft, hard, now;
206 if (hrtimer_active(period_timer))
209 now = hrtimer_cb_get_time(period_timer);
210 hrtimer_forward(period_timer, now, period);
212 soft = hrtimer_get_softexpires(period_timer);
213 hard = hrtimer_get_expires(period_timer);
214 delta = ktime_to_ns(ktime_sub(hard, soft));
215 __hrtimer_start_range_ns(period_timer, soft, delta,
216 HRTIMER_MODE_ABS_PINNED, 0);
220 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 raw_spin_lock(&rt_b->rt_runtime_lock);
229 start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
230 raw_spin_unlock(&rt_b->rt_runtime_lock);
233 #ifdef CONFIG_RT_GROUP_SCHED
234 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
236 hrtimer_cancel(&rt_b->rt_period_timer);
241 * sched_domains_mutex serializes calls to init_sched_domains,
242 * detach_destroy_domains and partition_sched_domains.
244 static DEFINE_MUTEX(sched_domains_mutex);
246 #ifdef CONFIG_CGROUP_SCHED
248 #include <linux/cgroup.h>
252 static LIST_HEAD(task_groups);
254 struct cfs_bandwidth {
255 #ifdef CONFIG_CFS_BANDWIDTH
259 s64 hierarchal_quota;
262 int idle, timer_active;
263 struct hrtimer period_timer, slack_timer;
264 struct list_head throttled_cfs_rq;
267 int nr_periods, nr_throttled;
272 /* task group related information */
274 struct cgroup_subsys_state css;
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity **se;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq **cfs_rq;
281 unsigned long shares;
283 atomic_t load_weight;
286 #ifdef CONFIG_RT_GROUP_SCHED
287 struct sched_rt_entity **rt_se;
288 struct rt_rq **rt_rq;
290 struct rt_bandwidth rt_bandwidth;
294 struct list_head list;
296 struct task_group *parent;
297 struct list_head siblings;
298 struct list_head children;
300 #ifdef CONFIG_SCHED_AUTOGROUP
301 struct autogroup *autogroup;
304 struct cfs_bandwidth cfs_bandwidth;
307 /* task_group_lock serializes the addition/removal of task groups */
308 static DEFINE_SPINLOCK(task_group_lock);
310 #ifdef CONFIG_FAIR_GROUP_SCHED
312 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
315 * A weight of 0 or 1 can cause arithmetics problems.
316 * A weight of a cfs_rq is the sum of weights of which entities
317 * are queued on this cfs_rq, so a weight of a entity should not be
318 * too large, so as the shares value of a task group.
319 * (The default weight is 1024 - so there's no practical
320 * limitation from this.)
322 #define MIN_SHARES (1UL << 1)
323 #define MAX_SHARES (1UL << 18)
325 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
328 /* Default task group.
329 * Every task in system belong to this group at bootup.
331 struct task_group root_task_group;
333 #endif /* CONFIG_CGROUP_SCHED */
335 /* CFS-related fields in a runqueue */
337 struct load_weight load;
338 unsigned long nr_running, h_nr_running;
343 u64 min_vruntime_copy;
346 struct rb_root tasks_timeline;
347 struct rb_node *rb_leftmost;
349 struct list_head tasks;
350 struct list_head *balance_iterator;
353 * 'curr' points to currently running entity on this cfs_rq.
354 * It is set to NULL otherwise (i.e when none are currently running).
356 struct sched_entity *curr, *next, *last, *skip;
358 #ifdef CONFIG_SCHED_DEBUG
359 unsigned int nr_spread_over;
362 #ifdef CONFIG_FAIR_GROUP_SCHED
363 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
366 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
367 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
368 * (like users, containers etc.)
370 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
371 * list is used during load balance.
374 struct list_head leaf_cfs_rq_list;
375 struct task_group *tg; /* group that "owns" this runqueue */
379 * the part of load.weight contributed by tasks
381 unsigned long task_weight;
384 * h_load = weight * f(tg)
386 * Where f(tg) is the recursive weight fraction assigned to
389 unsigned long h_load;
392 * Maintaining per-cpu shares distribution for group scheduling
394 * load_stamp is the last time we updated the load average
395 * load_last is the last time we updated the load average and saw load
396 * load_unacc_exec_time is currently unaccounted execution time
400 u64 load_stamp, load_last, load_unacc_exec_time;
402 unsigned long load_contribution;
404 #ifdef CONFIG_CFS_BANDWIDTH
407 s64 runtime_remaining;
409 u64 throttled_timestamp;
410 int throttled, throttle_count;
411 struct list_head throttled_list;
416 #ifdef CONFIG_FAIR_GROUP_SCHED
417 #ifdef CONFIG_CFS_BANDWIDTH
418 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
420 return &tg->cfs_bandwidth;
423 static inline u64 default_cfs_period(void);
424 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
425 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
427 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
429 struct cfs_bandwidth *cfs_b =
430 container_of(timer, struct cfs_bandwidth, slack_timer);
431 do_sched_cfs_slack_timer(cfs_b);
433 return HRTIMER_NORESTART;
436 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
438 struct cfs_bandwidth *cfs_b =
439 container_of(timer, struct cfs_bandwidth, period_timer);
445 now = hrtimer_cb_get_time(timer);
446 overrun = hrtimer_forward(timer, now, cfs_b->period);
451 idle = do_sched_cfs_period_timer(cfs_b, overrun);
454 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
457 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
459 raw_spin_lock_init(&cfs_b->lock);
461 cfs_b->quota = RUNTIME_INF;
462 cfs_b->period = ns_to_ktime(default_cfs_period());
464 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
465 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
466 cfs_b->period_timer.function = sched_cfs_period_timer;
467 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
468 cfs_b->slack_timer.function = sched_cfs_slack_timer;
471 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
473 cfs_rq->runtime_enabled = 0;
474 INIT_LIST_HEAD(&cfs_rq->throttled_list);
477 /* requires cfs_b->lock, may release to reprogram timer */
478 static void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
481 * The timer may be active because we're trying to set a new bandwidth
482 * period or because we're racing with the tear-down path
483 * (timer_active==0 becomes visible before the hrtimer call-back
484 * terminates). In either case we ensure that it's re-programmed
486 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
487 raw_spin_unlock(&cfs_b->lock);
488 /* ensure cfs_b->lock is available while we wait */
489 hrtimer_cancel(&cfs_b->period_timer);
491 raw_spin_lock(&cfs_b->lock);
492 /* if someone else restarted the timer then we're done */
493 if (cfs_b->timer_active)
497 cfs_b->timer_active = 1;
498 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
501 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
503 hrtimer_cancel(&cfs_b->period_timer);
504 hrtimer_cancel(&cfs_b->slack_timer);
507 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
508 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
509 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
511 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
515 #endif /* CONFIG_CFS_BANDWIDTH */
516 #endif /* CONFIG_FAIR_GROUP_SCHED */
518 /* Real-Time classes' related field in a runqueue: */
520 struct rt_prio_array active;
521 unsigned long rt_nr_running;
522 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
524 int curr; /* highest queued rt task prio */
526 int next; /* next highest */
531 unsigned long rt_nr_migratory;
532 unsigned long rt_nr_total;
534 struct plist_head pushable_tasks;
539 /* Nests inside the rq lock: */
540 raw_spinlock_t rt_runtime_lock;
542 #ifdef CONFIG_RT_GROUP_SCHED
543 unsigned long rt_nr_boosted;
546 struct list_head leaf_rt_rq_list;
547 struct task_group *tg;
554 * We add the notion of a root-domain which will be used to define per-domain
555 * variables. Each exclusive cpuset essentially defines an island domain by
556 * fully partitioning the member cpus from any other cpuset. Whenever a new
557 * exclusive cpuset is created, we also create and attach a new root-domain
566 cpumask_var_t online;
569 * The "RT overload" flag: it gets set if a CPU has more than
570 * one runnable RT task.
572 cpumask_var_t rto_mask;
573 struct cpupri cpupri;
577 * By default the system creates a single root-domain with all cpus as
578 * members (mimicking the global state we have today).
580 static struct root_domain def_root_domain;
582 #endif /* CONFIG_SMP */
585 * This is the main, per-CPU runqueue data structure.
587 * Locking rule: those places that want to lock multiple runqueues
588 * (such as the load balancing or the thread migration code), lock
589 * acquire operations must be ordered by ascending &runqueue.
596 * nr_running and cpu_load should be in the same cacheline because
597 * remote CPUs use both these fields when doing load calculation.
599 unsigned long nr_running;
600 #define CPU_LOAD_IDX_MAX 5
601 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
602 unsigned long last_load_update_tick;
605 unsigned char nohz_balance_kick;
607 int skip_clock_update;
609 /* capture load from *all* tasks on this cpu: */
610 struct load_weight load;
611 unsigned long nr_load_updates;
617 #ifdef CONFIG_FAIR_GROUP_SCHED
618 /* list of leaf cfs_rq on this cpu: */
619 struct list_head leaf_cfs_rq_list;
621 #ifdef CONFIG_RT_GROUP_SCHED
622 struct list_head leaf_rt_rq_list;
626 * This is part of a global counter where only the total sum
627 * over all CPUs matters. A task can increase this counter on
628 * one CPU and if it got migrated afterwards it may decrease
629 * it on another CPU. Always updated under the runqueue lock:
631 unsigned long nr_uninterruptible;
633 struct task_struct *curr, *idle, *stop;
634 unsigned long next_balance;
635 struct mm_struct *prev_mm;
643 struct root_domain *rd;
644 struct sched_domain *sd;
646 unsigned long cpu_power;
648 unsigned char idle_balance;
649 /* For active balancing */
653 struct cpu_stop_work active_balance_work;
654 /* cpu of this runqueue: */
664 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
667 #ifdef CONFIG_PARAVIRT
670 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
671 u64 prev_steal_time_rq;
674 /* calc_load related fields */
675 unsigned long calc_load_update;
676 long calc_load_active;
678 #ifdef CONFIG_SCHED_HRTICK
680 int hrtick_csd_pending;
681 struct call_single_data hrtick_csd;
683 struct hrtimer hrtick_timer;
686 #ifdef CONFIG_SCHEDSTATS
688 struct sched_info rq_sched_info;
689 unsigned long long rq_cpu_time;
690 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
692 /* sys_sched_yield() stats */
693 unsigned int yld_count;
695 /* schedule() stats */
696 unsigned int sched_switch;
697 unsigned int sched_count;
698 unsigned int sched_goidle;
700 /* try_to_wake_up() stats */
701 unsigned int ttwu_count;
702 unsigned int ttwu_local;
706 struct llist_head wake_list;
710 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
713 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
715 static inline int cpu_of(struct rq *rq)
724 #define rcu_dereference_check_sched_domain(p) \
725 rcu_dereference_check((p), \
726 lockdep_is_held(&sched_domains_mutex))
729 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
730 * See detach_destroy_domains: synchronize_sched for details.
732 * The domain tree of any CPU may only be accessed from within
733 * preempt-disabled sections.
735 #define for_each_domain(cpu, __sd) \
736 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
738 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
739 #define this_rq() (&__get_cpu_var(runqueues))
740 #define task_rq(p) cpu_rq(task_cpu(p))
741 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
742 #define raw_rq() (&__raw_get_cpu_var(runqueues))
744 #ifdef CONFIG_CGROUP_SCHED
747 * Return the group to which this tasks belongs.
749 * We cannot use task_subsys_state() and friends because the cgroup
750 * subsystem changes that value before the cgroup_subsys::attach() method
751 * is called, therefore we cannot pin it and might observe the wrong value.
753 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
754 * core changes this before calling sched_move_task().
756 * Instead we use a 'copy' which is updated from sched_move_task() while
757 * holding both task_struct::pi_lock and rq::lock.
759 static inline struct task_group *task_group(struct task_struct *p)
761 return p->sched_task_group;
764 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
765 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
767 #ifdef CONFIG_FAIR_GROUP_SCHED
768 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
769 p->se.parent = task_group(p)->se[cpu];
772 #ifdef CONFIG_RT_GROUP_SCHED
773 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
774 p->rt.parent = task_group(p)->rt_se[cpu];
778 #else /* CONFIG_CGROUP_SCHED */
780 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
781 static inline struct task_group *task_group(struct task_struct *p)
786 #endif /* CONFIG_CGROUP_SCHED */
788 static void update_rq_clock_task(struct rq *rq, s64 delta);
790 static void update_rq_clock(struct rq *rq)
794 if (rq->skip_clock_update > 0)
797 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
799 update_rq_clock_task(rq, delta);
803 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
805 #ifdef CONFIG_SCHED_DEBUG
806 # define const_debug __read_mostly
808 # define const_debug static const
812 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
813 * @cpu: the processor in question.
815 * This interface allows printk to be called with the runqueue lock
816 * held and know whether or not it is OK to wake up the klogd.
818 int runqueue_is_locked(int cpu)
820 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
824 * Debugging: various feature bits
827 #define SCHED_FEAT(name, enabled) \
828 __SCHED_FEAT_##name ,
831 #include "sched_features.h"
836 #define SCHED_FEAT(name, enabled) \
837 (1UL << __SCHED_FEAT_##name) * enabled |
839 const_debug unsigned int sysctl_sched_features =
840 #include "sched_features.h"
845 #ifdef CONFIG_SCHED_DEBUG
846 #define SCHED_FEAT(name, enabled) \
849 static __read_mostly char *sched_feat_names[] = {
850 #include "sched_features.h"
856 static int sched_feat_show(struct seq_file *m, void *v)
860 for (i = 0; sched_feat_names[i]; i++) {
861 if (!(sysctl_sched_features & (1UL << i)))
863 seq_printf(m, "%s ", sched_feat_names[i]);
871 sched_feat_write(struct file *filp, const char __user *ubuf,
872 size_t cnt, loff_t *ppos)
882 if (copy_from_user(&buf, ubuf, cnt))
888 if (strncmp(cmp, "NO_", 3) == 0) {
893 for (i = 0; sched_feat_names[i]; i++) {
894 if (strcmp(cmp, sched_feat_names[i]) == 0) {
896 sysctl_sched_features &= ~(1UL << i);
898 sysctl_sched_features |= (1UL << i);
903 if (!sched_feat_names[i])
911 static int sched_feat_open(struct inode *inode, struct file *filp)
913 return single_open(filp, sched_feat_show, NULL);
916 static const struct file_operations sched_feat_fops = {
917 .open = sched_feat_open,
918 .write = sched_feat_write,
921 .release = single_release,
924 static __init int sched_init_debug(void)
926 debugfs_create_file("sched_features", 0644, NULL, NULL,
931 late_initcall(sched_init_debug);
935 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
938 * Number of tasks to iterate in a single balance run.
939 * Limited because this is done with IRQs disabled.
941 const_debug unsigned int sysctl_sched_nr_migrate = 32;
944 * period over which we average the RT time consumption, measured
949 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
952 * period over which we measure -rt task cpu usage in us.
955 unsigned int sysctl_sched_rt_period = 1000000;
957 static __read_mostly int scheduler_running;
960 * part of the period that we allow rt tasks to run in us.
963 int sysctl_sched_rt_runtime = 950000;
965 static inline u64 global_rt_period(void)
967 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
970 static inline u64 global_rt_runtime(void)
972 if (sysctl_sched_rt_runtime < 0)
975 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
978 #ifndef prepare_arch_switch
979 # define prepare_arch_switch(next) do { } while (0)
981 #ifndef finish_arch_switch
982 # define finish_arch_switch(prev) do { } while (0)
985 static inline int task_current(struct rq *rq, struct task_struct *p)
987 return rq->curr == p;
990 static inline int task_running(struct rq *rq, struct task_struct *p)
995 return task_current(rq, p);
999 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1000 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1004 * We can optimise this out completely for !SMP, because the
1005 * SMP rebalancing from interrupt is the only thing that cares
1012 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1016 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1017 * We must ensure this doesn't happen until the switch is completely
1020 * Pairs with the control dependency and rmb in try_to_wake_up().
1025 #ifdef CONFIG_DEBUG_SPINLOCK
1026 /* this is a valid case when another task releases the spinlock */
1027 rq->lock.owner = current;
1030 * If we are tracking spinlock dependencies then we have to
1031 * fix up the runqueue lock - which gets 'carried over' from
1032 * prev into current:
1034 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1036 raw_spin_unlock_irq(&rq->lock);
1039 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1040 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1044 * We can optimise this out completely for !SMP, because the
1045 * SMP rebalancing from interrupt is the only thing that cares
1050 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1051 raw_spin_unlock_irq(&rq->lock);
1053 raw_spin_unlock(&rq->lock);
1057 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1061 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1062 * We must ensure this doesn't happen until the switch is completely
1068 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1072 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1075 * __task_rq_lock - lock the rq @p resides on.
1077 static inline struct rq *__task_rq_lock(struct task_struct *p)
1078 __acquires(rq->lock)
1082 lockdep_assert_held(&p->pi_lock);
1086 raw_spin_lock(&rq->lock);
1087 if (likely(rq == task_rq(p)))
1089 raw_spin_unlock(&rq->lock);
1094 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1096 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1097 __acquires(p->pi_lock)
1098 __acquires(rq->lock)
1103 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1105 raw_spin_lock(&rq->lock);
1106 if (likely(rq == task_rq(p)))
1108 raw_spin_unlock(&rq->lock);
1109 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1113 static void __task_rq_unlock(struct rq *rq)
1114 __releases(rq->lock)
1116 raw_spin_unlock(&rq->lock);
1120 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1121 __releases(rq->lock)
1122 __releases(p->pi_lock)
1124 raw_spin_unlock(&rq->lock);
1125 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1129 * this_rq_lock - lock this runqueue and disable interrupts.
1131 static struct rq *this_rq_lock(void)
1132 __acquires(rq->lock)
1136 local_irq_disable();
1138 raw_spin_lock(&rq->lock);
1143 #ifdef CONFIG_SCHED_HRTICK
1145 * Use HR-timers to deliver accurate preemption points.
1147 * Its all a bit involved since we cannot program an hrt while holding the
1148 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1151 * When we get rescheduled we reprogram the hrtick_timer outside of the
1157 * - enabled by features
1158 * - hrtimer is actually high res
1160 static inline int hrtick_enabled(struct rq *rq)
1162 if (!sched_feat(HRTICK))
1164 if (!cpu_active(cpu_of(rq)))
1166 return hrtimer_is_hres_active(&rq->hrtick_timer);
1169 static void hrtick_clear(struct rq *rq)
1171 if (hrtimer_active(&rq->hrtick_timer))
1172 hrtimer_cancel(&rq->hrtick_timer);
1176 * High-resolution timer tick.
1177 * Runs from hardirq context with interrupts disabled.
1179 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1181 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1183 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1185 raw_spin_lock(&rq->lock);
1186 update_rq_clock(rq);
1187 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1188 raw_spin_unlock(&rq->lock);
1190 return HRTIMER_NORESTART;
1195 * called from hardirq (IPI) context
1197 static void __hrtick_start(void *arg)
1199 struct rq *rq = arg;
1201 raw_spin_lock(&rq->lock);
1202 hrtimer_restart(&rq->hrtick_timer);
1203 rq->hrtick_csd_pending = 0;
1204 raw_spin_unlock(&rq->lock);
1208 * Called to set the hrtick timer state.
1210 * called with rq->lock held and irqs disabled
1212 static void hrtick_start(struct rq *rq, u64 delay)
1214 struct hrtimer *timer = &rq->hrtick_timer;
1215 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1217 hrtimer_set_expires(timer, time);
1219 if (rq == this_rq()) {
1220 hrtimer_restart(timer);
1221 } else if (!rq->hrtick_csd_pending) {
1222 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1223 rq->hrtick_csd_pending = 1;
1228 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1230 int cpu = (int)(long)hcpu;
1233 case CPU_UP_CANCELED:
1234 case CPU_UP_CANCELED_FROZEN:
1235 case CPU_DOWN_PREPARE:
1236 case CPU_DOWN_PREPARE_FROZEN:
1238 case CPU_DEAD_FROZEN:
1239 hrtick_clear(cpu_rq(cpu));
1246 static __init void init_hrtick(void)
1248 hotcpu_notifier(hotplug_hrtick, 0);
1252 * Called to set the hrtick timer state.
1254 * called with rq->lock held and irqs disabled
1256 static void hrtick_start(struct rq *rq, u64 delay)
1258 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1259 HRTIMER_MODE_REL_PINNED, 0);
1262 static inline void init_hrtick(void)
1265 #endif /* CONFIG_SMP */
1267 static void init_rq_hrtick(struct rq *rq)
1270 rq->hrtick_csd_pending = 0;
1272 rq->hrtick_csd.flags = 0;
1273 rq->hrtick_csd.func = __hrtick_start;
1274 rq->hrtick_csd.info = rq;
1277 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1278 rq->hrtick_timer.function = hrtick;
1280 #else /* CONFIG_SCHED_HRTICK */
1281 static inline void hrtick_clear(struct rq *rq)
1285 static inline void init_rq_hrtick(struct rq *rq)
1289 static inline void init_hrtick(void)
1292 #endif /* CONFIG_SCHED_HRTICK */
1295 * resched_task - mark a task 'to be rescheduled now'.
1297 * On UP this means the setting of the need_resched flag, on SMP it
1298 * might also involve a cross-CPU call to trigger the scheduler on
1303 #ifndef tsk_is_polling
1304 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1307 static void resched_task(struct task_struct *p)
1311 assert_raw_spin_locked(&task_rq(p)->lock);
1313 if (test_tsk_need_resched(p))
1316 set_tsk_need_resched(p);
1319 if (cpu == smp_processor_id())
1322 /* NEED_RESCHED must be visible before we test polling */
1324 if (!tsk_is_polling(p))
1325 smp_send_reschedule(cpu);
1328 static void resched_cpu(int cpu)
1330 struct rq *rq = cpu_rq(cpu);
1331 unsigned long flags;
1333 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1335 resched_task(cpu_curr(cpu));
1336 raw_spin_unlock_irqrestore(&rq->lock, flags);
1341 * In the semi idle case, use the nearest busy cpu for migrating timers
1342 * from an idle cpu. This is good for power-savings.
1344 * We don't do similar optimization for completely idle system, as
1345 * selecting an idle cpu will add more delays to the timers than intended
1346 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1348 int get_nohz_timer_target(void)
1350 int cpu = smp_processor_id();
1352 struct sched_domain *sd;
1355 for_each_domain(cpu, sd) {
1356 for_each_cpu(i, sched_domain_span(sd)) {
1368 * When add_timer_on() enqueues a timer into the timer wheel of an
1369 * idle CPU then this timer might expire before the next timer event
1370 * which is scheduled to wake up that CPU. In case of a completely
1371 * idle system the next event might even be infinite time into the
1372 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1373 * leaves the inner idle loop so the newly added timer is taken into
1374 * account when the CPU goes back to idle and evaluates the timer
1375 * wheel for the next timer event.
1377 void wake_up_idle_cpu(int cpu)
1379 struct rq *rq = cpu_rq(cpu);
1381 if (cpu == smp_processor_id())
1385 * This is safe, as this function is called with the timer
1386 * wheel base lock of (cpu) held. When the CPU is on the way
1387 * to idle and has not yet set rq->curr to idle then it will
1388 * be serialized on the timer wheel base lock and take the new
1389 * timer into account automatically.
1391 if (rq->curr != rq->idle)
1395 * We can set TIF_RESCHED on the idle task of the other CPU
1396 * lockless. The worst case is that the other CPU runs the
1397 * idle task through an additional NOOP schedule()
1399 set_tsk_need_resched(rq->idle);
1401 /* NEED_RESCHED must be visible before we test polling */
1403 if (!tsk_is_polling(rq->idle))
1404 smp_send_reschedule(cpu);
1407 static inline bool got_nohz_idle_kick(void)
1409 return idle_cpu(smp_processor_id()) && this_rq()->nohz_balance_kick;
1412 #else /* CONFIG_NO_HZ */
1414 static inline bool got_nohz_idle_kick(void)
1419 #endif /* CONFIG_NO_HZ */
1421 static u64 sched_avg_period(void)
1423 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1426 static void sched_avg_update(struct rq *rq)
1428 s64 period = sched_avg_period();
1430 while ((s64)(rq->clock - rq->age_stamp) > period) {
1432 * Inline assembly required to prevent the compiler
1433 * optimising this loop into a divmod call.
1434 * See __iter_div_u64_rem() for another example of this.
1436 asm("" : "+rm" (rq->age_stamp));
1437 rq->age_stamp += period;
1442 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1444 rq->rt_avg += rt_delta;
1445 sched_avg_update(rq);
1448 #else /* !CONFIG_SMP */
1449 static void resched_task(struct task_struct *p)
1451 assert_raw_spin_locked(&task_rq(p)->lock);
1452 set_tsk_need_resched(p);
1455 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1459 static void sched_avg_update(struct rq *rq)
1462 #endif /* CONFIG_SMP */
1464 #if BITS_PER_LONG == 32
1465 # define WMULT_CONST (~0UL)
1467 # define WMULT_CONST (1UL << 32)
1470 #define WMULT_SHIFT 32
1473 * Shift right and round:
1475 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1478 * delta *= weight / lw
1480 static unsigned long
1481 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1482 struct load_weight *lw)
1487 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1488 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1489 * 2^SCHED_LOAD_RESOLUTION.
1491 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1492 tmp = (u64)delta_exec * scale_load_down(weight);
1494 tmp = (u64)delta_exec;
1496 if (!lw->inv_weight) {
1497 unsigned long w = scale_load_down(lw->weight);
1499 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1501 else if (unlikely(!w))
1502 lw->inv_weight = WMULT_CONST;
1504 lw->inv_weight = WMULT_CONST / w;
1508 * Check whether we'd overflow the 64-bit multiplication:
1510 if (unlikely(tmp > WMULT_CONST))
1511 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1514 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1516 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1519 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1525 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1531 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1538 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1539 * of tasks with abnormal "nice" values across CPUs the contribution that
1540 * each task makes to its run queue's load is weighted according to its
1541 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1542 * scaled version of the new time slice allocation that they receive on time
1546 #define WEIGHT_IDLEPRIO 3
1547 #define WMULT_IDLEPRIO 1431655765
1550 * Nice levels are multiplicative, with a gentle 10% change for every
1551 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1552 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1553 * that remained on nice 0.
1555 * The "10% effect" is relative and cumulative: from _any_ nice level,
1556 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1557 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1558 * If a task goes up by ~10% and another task goes down by ~10% then
1559 * the relative distance between them is ~25%.)
1561 static const int prio_to_weight[40] = {
1562 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1563 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1564 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1565 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1566 /* 0 */ 1024, 820, 655, 526, 423,
1567 /* 5 */ 335, 272, 215, 172, 137,
1568 /* 10 */ 110, 87, 70, 56, 45,
1569 /* 15 */ 36, 29, 23, 18, 15,
1573 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1575 * In cases where the weight does not change often, we can use the
1576 * precalculated inverse to speed up arithmetics by turning divisions
1577 * into multiplications:
1579 static const u32 prio_to_wmult[40] = {
1580 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1581 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1582 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1583 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1584 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1585 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1586 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1587 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1590 /* Time spent by the tasks of the cpu accounting group executing in ... */
1591 enum cpuacct_stat_index {
1592 CPUACCT_STAT_USER, /* ... user mode */
1593 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1595 CPUACCT_STAT_NSTATS,
1598 #ifdef CONFIG_CGROUP_CPUACCT
1599 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1600 static void cpuacct_update_stats(struct task_struct *tsk,
1601 enum cpuacct_stat_index idx, cputime_t val);
1603 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1604 static inline void cpuacct_update_stats(struct task_struct *tsk,
1605 enum cpuacct_stat_index idx, cputime_t val) {}
1608 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1610 update_load_add(&rq->load, load);
1613 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1615 update_load_sub(&rq->load, load);
1618 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1619 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1620 typedef int (*tg_visitor)(struct task_group *, void *);
1623 * Iterate task_group tree rooted at *from, calling @down when first entering a
1624 * node and @up when leaving it for the final time.
1626 * Caller must hold rcu_lock or sufficient equivalent.
1628 static int walk_tg_tree_from(struct task_group *from,
1629 tg_visitor down, tg_visitor up, void *data)
1631 struct task_group *parent, *child;
1637 ret = (*down)(parent, data);
1640 list_for_each_entry_rcu(child, &parent->children, siblings) {
1647 ret = (*up)(parent, data);
1648 if (ret || parent == from)
1652 parent = parent->parent;
1660 * Iterate the full tree, calling @down when first entering a node and @up when
1661 * leaving it for the final time.
1663 * Caller must hold rcu_lock or sufficient equivalent.
1666 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1668 return walk_tg_tree_from(&root_task_group, down, up, data);
1671 static int tg_nop(struct task_group *tg, void *data)
1678 /* Used instead of source_load when we know the type == 0 */
1679 static unsigned long weighted_cpuload(const int cpu)
1681 return cpu_rq(cpu)->load.weight;
1685 * Return a low guess at the load of a migration-source cpu weighted
1686 * according to the scheduling class and "nice" value.
1688 * We want to under-estimate the load of migration sources, to
1689 * balance conservatively.
1691 static unsigned long source_load(int cpu, int type)
1693 struct rq *rq = cpu_rq(cpu);
1694 unsigned long total = weighted_cpuload(cpu);
1696 if (type == 0 || !sched_feat(LB_BIAS))
1699 return min(rq->cpu_load[type-1], total);
1703 * Return a high guess at the load of a migration-target cpu weighted
1704 * according to the scheduling class and "nice" value.
1706 static unsigned long target_load(int cpu, int type)
1708 struct rq *rq = cpu_rq(cpu);
1709 unsigned long total = weighted_cpuload(cpu);
1711 if (type == 0 || !sched_feat(LB_BIAS))
1714 return max(rq->cpu_load[type-1], total);
1717 static unsigned long power_of(int cpu)
1719 return cpu_rq(cpu)->cpu_power;
1722 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1724 static unsigned long cpu_avg_load_per_task(int cpu)
1726 struct rq *rq = cpu_rq(cpu);
1727 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1730 return rq->load.weight / nr_running;
1735 #ifdef CONFIG_PREEMPT
1737 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1740 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1741 * way at the expense of forcing extra atomic operations in all
1742 * invocations. This assures that the double_lock is acquired using the
1743 * same underlying policy as the spinlock_t on this architecture, which
1744 * reduces latency compared to the unfair variant below. However, it
1745 * also adds more overhead and therefore may reduce throughput.
1747 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1748 __releases(this_rq->lock)
1749 __acquires(busiest->lock)
1750 __acquires(this_rq->lock)
1752 raw_spin_unlock(&this_rq->lock);
1753 double_rq_lock(this_rq, busiest);
1760 * Unfair double_lock_balance: Optimizes throughput at the expense of
1761 * latency by eliminating extra atomic operations when the locks are
1762 * already in proper order on entry. This favors lower cpu-ids and will
1763 * grant the double lock to lower cpus over higher ids under contention,
1764 * regardless of entry order into the function.
1766 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1767 __releases(this_rq->lock)
1768 __acquires(busiest->lock)
1769 __acquires(this_rq->lock)
1773 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1774 if (busiest < this_rq) {
1775 raw_spin_unlock(&this_rq->lock);
1776 raw_spin_lock(&busiest->lock);
1777 raw_spin_lock_nested(&this_rq->lock,
1778 SINGLE_DEPTH_NESTING);
1781 raw_spin_lock_nested(&busiest->lock,
1782 SINGLE_DEPTH_NESTING);
1787 #endif /* CONFIG_PREEMPT */
1790 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1792 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1794 if (unlikely(!irqs_disabled())) {
1795 /* printk() doesn't work good under rq->lock */
1796 raw_spin_unlock(&this_rq->lock);
1800 return _double_lock_balance(this_rq, busiest);
1803 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1804 __releases(busiest->lock)
1806 raw_spin_unlock(&busiest->lock);
1807 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1811 * double_rq_lock - safely lock two runqueues
1813 * Note this does not disable interrupts like task_rq_lock,
1814 * you need to do so manually before calling.
1816 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1817 __acquires(rq1->lock)
1818 __acquires(rq2->lock)
1820 BUG_ON(!irqs_disabled());
1822 raw_spin_lock(&rq1->lock);
1823 __acquire(rq2->lock); /* Fake it out ;) */
1826 raw_spin_lock(&rq1->lock);
1827 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1829 raw_spin_lock(&rq2->lock);
1830 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1836 * double_rq_unlock - safely unlock two runqueues
1838 * Note this does not restore interrupts like task_rq_unlock,
1839 * you need to do so manually after calling.
1841 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1842 __releases(rq1->lock)
1843 __releases(rq2->lock)
1845 raw_spin_unlock(&rq1->lock);
1847 raw_spin_unlock(&rq2->lock);
1849 __release(rq2->lock);
1852 #else /* CONFIG_SMP */
1855 * double_rq_lock - safely lock two runqueues
1857 * Note this does not disable interrupts like task_rq_lock,
1858 * you need to do so manually before calling.
1860 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1861 __acquires(rq1->lock)
1862 __acquires(rq2->lock)
1864 BUG_ON(!irqs_disabled());
1866 raw_spin_lock(&rq1->lock);
1867 __acquire(rq2->lock); /* Fake it out ;) */
1871 * double_rq_unlock - safely unlock two runqueues
1873 * Note this does not restore interrupts like task_rq_unlock,
1874 * you need to do so manually after calling.
1876 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1877 __releases(rq1->lock)
1878 __releases(rq2->lock)
1881 raw_spin_unlock(&rq1->lock);
1882 __release(rq2->lock);
1887 static void update_sysctl(void);
1888 static int get_update_sysctl_factor(void);
1889 static void update_idle_cpu_load(struct rq *this_rq);
1891 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1893 set_task_rq(p, cpu);
1896 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1897 * successfully executed on another CPU. We must ensure that updates of
1898 * per-task data have been completed by this moment.
1901 task_thread_info(p)->cpu = cpu;
1905 static const struct sched_class rt_sched_class;
1907 #define sched_class_highest (&stop_sched_class)
1908 #define for_each_class(class) \
1909 for (class = sched_class_highest; class; class = class->next)
1911 #include "sched_stats.h"
1913 static void inc_nr_running(struct rq *rq)
1918 static void dec_nr_running(struct rq *rq)
1923 static void set_load_weight(struct task_struct *p)
1925 int prio = p->static_prio - MAX_RT_PRIO;
1926 struct load_weight *load = &p->se.load;
1929 * SCHED_IDLE tasks get minimal weight:
1931 if (p->policy == SCHED_IDLE) {
1932 load->weight = scale_load(WEIGHT_IDLEPRIO);
1933 load->inv_weight = WMULT_IDLEPRIO;
1937 load->weight = scale_load(prio_to_weight[prio]);
1938 load->inv_weight = prio_to_wmult[prio];
1941 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1943 update_rq_clock(rq);
1944 sched_info_queued(p);
1945 p->sched_class->enqueue_task(rq, p, flags);
1948 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1950 update_rq_clock(rq);
1951 sched_info_dequeued(p);
1952 p->sched_class->dequeue_task(rq, p, flags);
1956 * activate_task - move a task to the runqueue.
1958 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1960 if (task_contributes_to_load(p))
1961 rq->nr_uninterruptible--;
1963 enqueue_task(rq, p, flags);
1967 * deactivate_task - remove a task from the runqueue.
1969 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1971 if (task_contributes_to_load(p))
1972 rq->nr_uninterruptible++;
1974 dequeue_task(rq, p, flags);
1977 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1980 * There are no locks covering percpu hardirq/softirq time.
1981 * They are only modified in account_system_vtime, on corresponding CPU
1982 * with interrupts disabled. So, writes are safe.
1983 * They are read and saved off onto struct rq in update_rq_clock().
1984 * This may result in other CPU reading this CPU's irq time and can
1985 * race with irq/account_system_vtime on this CPU. We would either get old
1986 * or new value with a side effect of accounting a slice of irq time to wrong
1987 * task when irq is in progress while we read rq->clock. That is a worthy
1988 * compromise in place of having locks on each irq in account_system_time.
1990 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1991 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1993 static DEFINE_PER_CPU(u64, irq_start_time);
1994 static int sched_clock_irqtime;
1996 void enable_sched_clock_irqtime(void)
1998 sched_clock_irqtime = 1;
2001 void disable_sched_clock_irqtime(void)
2003 sched_clock_irqtime = 0;
2006 #ifndef CONFIG_64BIT
2007 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
2009 static inline void irq_time_write_begin(void)
2011 __this_cpu_inc(irq_time_seq.sequence);
2015 static inline void irq_time_write_end(void)
2018 __this_cpu_inc(irq_time_seq.sequence);
2021 static inline u64 irq_time_read(int cpu)
2027 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
2028 irq_time = per_cpu(cpu_softirq_time, cpu) +
2029 per_cpu(cpu_hardirq_time, cpu);
2030 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
2034 #else /* CONFIG_64BIT */
2035 static inline void irq_time_write_begin(void)
2039 static inline void irq_time_write_end(void)
2043 static inline u64 irq_time_read(int cpu)
2045 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
2047 #endif /* CONFIG_64BIT */
2050 * Called before incrementing preempt_count on {soft,}irq_enter
2051 * and before decrementing preempt_count on {soft,}irq_exit.
2053 void account_system_vtime(struct task_struct *curr)
2055 unsigned long flags;
2059 if (!sched_clock_irqtime)
2062 local_irq_save(flags);
2064 cpu = smp_processor_id();
2065 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2066 __this_cpu_add(irq_start_time, delta);
2068 irq_time_write_begin();
2070 * We do not account for softirq time from ksoftirqd here.
2071 * We want to continue accounting softirq time to ksoftirqd thread
2072 * in that case, so as not to confuse scheduler with a special task
2073 * that do not consume any time, but still wants to run.
2075 if (hardirq_count())
2076 __this_cpu_add(cpu_hardirq_time, delta);
2077 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
2078 __this_cpu_add(cpu_softirq_time, delta);
2080 irq_time_write_end();
2081 local_irq_restore(flags);
2083 EXPORT_SYMBOL_GPL(account_system_vtime);
2085 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2087 #ifdef CONFIG_PARAVIRT
2088 static inline u64 steal_ticks(u64 steal)
2090 if (unlikely(steal > NSEC_PER_SEC))
2091 return div_u64(steal, TICK_NSEC);
2093 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
2097 static void update_rq_clock_task(struct rq *rq, s64 delta)
2100 * In theory, the compile should just see 0 here, and optimize out the call
2101 * to sched_rt_avg_update. But I don't trust it...
2103 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2104 s64 steal = 0, irq_delta = 0;
2106 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2107 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2110 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2111 * this case when a previous update_rq_clock() happened inside a
2112 * {soft,}irq region.
2114 * When this happens, we stop ->clock_task and only update the
2115 * prev_irq_time stamp to account for the part that fit, so that a next
2116 * update will consume the rest. This ensures ->clock_task is
2119 * It does however cause some slight miss-attribution of {soft,}irq
2120 * time, a more accurate solution would be to update the irq_time using
2121 * the current rq->clock timestamp, except that would require using
2124 if (irq_delta > delta)
2127 rq->prev_irq_time += irq_delta;
2130 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2131 if (static_branch((¶virt_steal_rq_enabled))) {
2134 steal = paravirt_steal_clock(cpu_of(rq));
2135 steal -= rq->prev_steal_time_rq;
2137 if (unlikely(steal > delta))
2140 st = steal_ticks(steal);
2141 steal = st * TICK_NSEC;
2143 rq->prev_steal_time_rq += steal;
2149 rq->clock_task += delta;
2151 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2152 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2153 sched_rt_avg_update(rq, irq_delta + steal);
2157 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2158 static int irqtime_account_hi_update(void)
2160 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2161 unsigned long flags;
2165 local_irq_save(flags);
2166 latest_ns = this_cpu_read(cpu_hardirq_time);
2167 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2169 local_irq_restore(flags);
2173 static int irqtime_account_si_update(void)
2175 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2176 unsigned long flags;
2180 local_irq_save(flags);
2181 latest_ns = this_cpu_read(cpu_softirq_time);
2182 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2184 local_irq_restore(flags);
2188 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2190 #define sched_clock_irqtime (0)
2195 static void unthrottle_offline_cfs_rqs(struct rq *rq);
2198 #include "sched_idletask.c"
2199 #include "sched_fair.c"
2200 #include "sched_rt.c"
2201 #include "sched_autogroup.c"
2202 #include "sched_stoptask.c"
2203 #ifdef CONFIG_SCHED_DEBUG
2204 # include "sched_debug.c"
2207 void sched_set_stop_task(int cpu, struct task_struct *stop)
2209 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2210 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2214 * Make it appear like a SCHED_FIFO task, its something
2215 * userspace knows about and won't get confused about.
2217 * Also, it will make PI more or less work without too
2218 * much confusion -- but then, stop work should not
2219 * rely on PI working anyway.
2221 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2223 stop->sched_class = &stop_sched_class;
2226 cpu_rq(cpu)->stop = stop;
2230 * Reset it back to a normal scheduling class so that
2231 * it can die in pieces.
2233 old_stop->sched_class = &rt_sched_class;
2238 * __normal_prio - return the priority that is based on the static prio
2240 static inline int __normal_prio(struct task_struct *p)
2242 return p->static_prio;
2246 * Calculate the expected normal priority: i.e. priority
2247 * without taking RT-inheritance into account. Might be
2248 * boosted by interactivity modifiers. Changes upon fork,
2249 * setprio syscalls, and whenever the interactivity
2250 * estimator recalculates.
2252 static inline int normal_prio(struct task_struct *p)
2256 if (task_has_rt_policy(p))
2257 prio = MAX_RT_PRIO-1 - p->rt_priority;
2259 prio = __normal_prio(p);
2264 * Calculate the current priority, i.e. the priority
2265 * taken into account by the scheduler. This value might
2266 * be boosted by RT tasks, or might be boosted by
2267 * interactivity modifiers. Will be RT if the task got
2268 * RT-boosted. If not then it returns p->normal_prio.
2270 static int effective_prio(struct task_struct *p)
2272 p->normal_prio = normal_prio(p);
2274 * If we are RT tasks or we were boosted to RT priority,
2275 * keep the priority unchanged. Otherwise, update priority
2276 * to the normal priority:
2278 if (!rt_prio(p->prio))
2279 return p->normal_prio;
2284 * task_curr - is this task currently executing on a CPU?
2285 * @p: the task in question.
2287 inline int task_curr(const struct task_struct *p)
2289 return cpu_curr(task_cpu(p)) == p;
2292 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2293 const struct sched_class *prev_class,
2296 if (prev_class != p->sched_class) {
2297 if (prev_class->switched_from)
2298 prev_class->switched_from(rq, p);
2299 p->sched_class->switched_to(rq, p);
2300 } else if (oldprio != p->prio)
2301 p->sched_class->prio_changed(rq, p, oldprio);
2304 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2306 const struct sched_class *class;
2308 if (p->sched_class == rq->curr->sched_class) {
2309 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2311 for_each_class(class) {
2312 if (class == rq->curr->sched_class)
2314 if (class == p->sched_class) {
2315 resched_task(rq->curr);
2322 * A queue event has occurred, and we're going to schedule. In
2323 * this case, we can save a useless back to back clock update.
2325 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2326 rq->skip_clock_update = 1;
2331 * Is this task likely cache-hot:
2334 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2338 if (p->sched_class != &fair_sched_class)
2341 if (unlikely(p->policy == SCHED_IDLE))
2345 * Buddy candidates are cache hot:
2347 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2348 (&p->se == cfs_rq_of(&p->se)->next ||
2349 &p->se == cfs_rq_of(&p->se)->last))
2352 if (sysctl_sched_migration_cost == -1)
2354 if (sysctl_sched_migration_cost == 0)
2357 delta = now - p->se.exec_start;
2359 return delta < (s64)sysctl_sched_migration_cost;
2362 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2364 #ifdef CONFIG_SCHED_DEBUG
2366 * We should never call set_task_cpu() on a blocked task,
2367 * ttwu() will sort out the placement.
2369 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2370 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2372 #ifdef CONFIG_LOCKDEP
2374 * The caller should hold either p->pi_lock or rq->lock, when changing
2375 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2377 * sched_move_task() holds both and thus holding either pins the cgroup,
2380 * Furthermore, all task_rq users should acquire both locks, see
2383 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2384 lockdep_is_held(&task_rq(p)->lock)));
2388 trace_sched_migrate_task(p, new_cpu);
2390 if (task_cpu(p) != new_cpu) {
2391 p->se.nr_migrations++;
2392 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2395 __set_task_cpu(p, new_cpu);
2398 struct migration_arg {
2399 struct task_struct *task;
2403 static int migration_cpu_stop(void *data);
2406 * wait_task_inactive - wait for a thread to unschedule.
2408 * If @match_state is nonzero, it's the @p->state value just checked and
2409 * not expected to change. If it changes, i.e. @p might have woken up,
2410 * then return zero. When we succeed in waiting for @p to be off its CPU,
2411 * we return a positive number (its total switch count). If a second call
2412 * a short while later returns the same number, the caller can be sure that
2413 * @p has remained unscheduled the whole time.
2415 * The caller must ensure that the task *will* unschedule sometime soon,
2416 * else this function might spin for a *long* time. This function can't
2417 * be called with interrupts off, or it may introduce deadlock with
2418 * smp_call_function() if an IPI is sent by the same process we are
2419 * waiting to become inactive.
2421 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2423 unsigned long flags;
2430 * We do the initial early heuristics without holding
2431 * any task-queue locks at all. We'll only try to get
2432 * the runqueue lock when things look like they will
2438 * If the task is actively running on another CPU
2439 * still, just relax and busy-wait without holding
2442 * NOTE! Since we don't hold any locks, it's not
2443 * even sure that "rq" stays as the right runqueue!
2444 * But we don't care, since "task_running()" will
2445 * return false if the runqueue has changed and p
2446 * is actually now running somewhere else!
2448 while (task_running(rq, p)) {
2449 if (match_state && unlikely(p->state != match_state))
2455 * Ok, time to look more closely! We need the rq
2456 * lock now, to be *sure*. If we're wrong, we'll
2457 * just go back and repeat.
2459 rq = task_rq_lock(p, &flags);
2460 trace_sched_wait_task(p);
2461 running = task_running(rq, p);
2464 if (!match_state || p->state == match_state)
2465 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2466 task_rq_unlock(rq, p, &flags);
2469 * If it changed from the expected state, bail out now.
2471 if (unlikely(!ncsw))
2475 * Was it really running after all now that we
2476 * checked with the proper locks actually held?
2478 * Oops. Go back and try again..
2480 if (unlikely(running)) {
2486 * It's not enough that it's not actively running,
2487 * it must be off the runqueue _entirely_, and not
2490 * So if it was still runnable (but just not actively
2491 * running right now), it's preempted, and we should
2492 * yield - it could be a while.
2494 if (unlikely(on_rq)) {
2495 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2497 set_current_state(TASK_UNINTERRUPTIBLE);
2498 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2503 * Ahh, all good. It wasn't running, and it wasn't
2504 * runnable, which means that it will never become
2505 * running in the future either. We're all done!
2514 * kick_process - kick a running thread to enter/exit the kernel
2515 * @p: the to-be-kicked thread
2517 * Cause a process which is running on another CPU to enter
2518 * kernel-mode, without any delay. (to get signals handled.)
2520 * NOTE: this function doesn't have to take the runqueue lock,
2521 * because all it wants to ensure is that the remote task enters
2522 * the kernel. If the IPI races and the task has been migrated
2523 * to another CPU then no harm is done and the purpose has been
2526 void kick_process(struct task_struct *p)
2532 if ((cpu != smp_processor_id()) && task_curr(p))
2533 smp_send_reschedule(cpu);
2536 EXPORT_SYMBOL_GPL(kick_process);
2537 #endif /* CONFIG_SMP */
2541 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2543 static int select_fallback_rq(int cpu, struct task_struct *p)
2546 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2548 /* Look for allowed, online CPU in same node. */
2549 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2550 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
2553 /* Any allowed, online CPU? */
2554 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
2555 if (dest_cpu < nr_cpu_ids)
2558 /* No more Mr. Nice Guy. */
2559 dest_cpu = cpuset_cpus_allowed_fallback(p);
2561 * Don't tell them about moving exiting tasks or
2562 * kernel threads (both mm NULL), since they never
2565 if (p->mm && printk_ratelimit()) {
2566 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2567 task_pid_nr(p), p->comm, cpu);
2574 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2577 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2579 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2582 * In order not to call set_task_cpu() on a blocking task we need
2583 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2586 * Since this is common to all placement strategies, this lives here.
2588 * [ this allows ->select_task() to simply return task_cpu(p) and
2589 * not worry about this generic constraint ]
2591 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
2593 cpu = select_fallback_rq(task_cpu(p), p);
2598 static void update_avg(u64 *avg, u64 sample)
2600 s64 diff = sample - *avg;
2606 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2608 #ifdef CONFIG_SCHEDSTATS
2609 struct rq *rq = this_rq();
2612 int this_cpu = smp_processor_id();
2614 if (cpu == this_cpu) {
2615 schedstat_inc(rq, ttwu_local);
2616 schedstat_inc(p, se.statistics.nr_wakeups_local);
2618 struct sched_domain *sd;
2620 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2622 for_each_domain(this_cpu, sd) {
2623 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2624 schedstat_inc(sd, ttwu_wake_remote);
2631 if (wake_flags & WF_MIGRATED)
2632 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2634 #endif /* CONFIG_SMP */
2636 schedstat_inc(rq, ttwu_count);
2637 schedstat_inc(p, se.statistics.nr_wakeups);
2639 if (wake_flags & WF_SYNC)
2640 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2642 #endif /* CONFIG_SCHEDSTATS */
2645 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2647 activate_task(rq, p, en_flags);
2650 /* if a worker is waking up, notify workqueue */
2651 if (p->flags & PF_WQ_WORKER)
2652 wq_worker_waking_up(p, cpu_of(rq));
2656 * Mark the task runnable and perform wakeup-preemption.
2659 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2661 trace_sched_wakeup(p, true);
2662 check_preempt_curr(rq, p, wake_flags);
2664 p->state = TASK_RUNNING;
2666 if (p->sched_class->task_woken)
2667 p->sched_class->task_woken(rq, p);
2669 if (rq->idle_stamp) {
2670 u64 delta = rq->clock - rq->idle_stamp;
2671 u64 max = 2*sysctl_sched_migration_cost;
2676 update_avg(&rq->avg_idle, delta);
2683 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2686 if (p->sched_contributes_to_load)
2687 rq->nr_uninterruptible--;
2690 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2691 ttwu_do_wakeup(rq, p, wake_flags);
2695 * Called in case the task @p isn't fully descheduled from its runqueue,
2696 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2697 * since all we need to do is flip p->state to TASK_RUNNING, since
2698 * the task is still ->on_rq.
2700 static int ttwu_remote(struct task_struct *p, int wake_flags)
2705 rq = __task_rq_lock(p);
2707 ttwu_do_wakeup(rq, p, wake_flags);
2710 __task_rq_unlock(rq);
2716 static void sched_ttwu_pending(void)
2718 struct rq *rq = this_rq();
2719 struct llist_node *llist = llist_del_all(&rq->wake_list);
2720 struct task_struct *p;
2722 raw_spin_lock(&rq->lock);
2725 p = llist_entry(llist, struct task_struct, wake_entry);
2726 llist = llist_next(llist);
2727 ttwu_do_activate(rq, p, 0);
2730 raw_spin_unlock(&rq->lock);
2733 void scheduler_ipi(void)
2735 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2739 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2740 * traditionally all their work was done from the interrupt return
2741 * path. Now that we actually do some work, we need to make sure
2744 * Some archs already do call them, luckily irq_enter/exit nest
2747 * Arguably we should visit all archs and update all handlers,
2748 * however a fair share of IPIs are still resched only so this would
2749 * somewhat pessimize the simple resched case.
2752 sched_ttwu_pending();
2755 * Check if someone kicked us for doing the nohz idle load balance.
2757 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
2758 this_rq()->idle_balance = 1;
2759 raise_softirq_irqoff(SCHED_SOFTIRQ);
2764 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2766 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
2767 smp_send_reschedule(cpu);
2770 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2771 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2776 rq = __task_rq_lock(p);
2778 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2779 ttwu_do_wakeup(rq, p, wake_flags);
2782 __task_rq_unlock(rq);
2787 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2788 #endif /* CONFIG_SMP */
2790 static void ttwu_queue(struct task_struct *p, int cpu)
2792 struct rq *rq = cpu_rq(cpu);
2794 #if defined(CONFIG_SMP)
2795 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2796 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2797 ttwu_queue_remote(p, cpu);
2802 raw_spin_lock(&rq->lock);
2803 ttwu_do_activate(rq, p, 0);
2804 raw_spin_unlock(&rq->lock);
2808 * try_to_wake_up - wake up a thread
2809 * @p: the thread to be awakened
2810 * @state: the mask of task states that can be woken
2811 * @wake_flags: wake modifier flags (WF_*)
2813 * Put it on the run-queue if it's not already there. The "current"
2814 * thread is always on the run-queue (except when the actual
2815 * re-schedule is in progress), and as such you're allowed to do
2816 * the simpler "current->state = TASK_RUNNING" to mark yourself
2817 * runnable without the overhead of this.
2819 * Returns %true if @p was woken up, %false if it was already running
2820 * or @state didn't match @p's state.
2823 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2825 unsigned long flags;
2826 int cpu, success = 0;
2829 raw_spin_lock_irqsave(&p->pi_lock, flags);
2830 if (!(p->state & state))
2833 success = 1; /* we're going to change ->state */
2836 if (p->on_rq && ttwu_remote(p, wake_flags))
2841 * If the owning (remote) cpu is still in the middle of schedule() with
2842 * this task as prev, wait until its done referencing the task.
2845 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2847 * In case the architecture enables interrupts in
2848 * context_switch(), we cannot busy wait, since that
2849 * would lead to deadlocks when an interrupt hits and
2850 * tries to wake up @prev. So bail and do a complete
2853 if (ttwu_activate_remote(p, wake_flags))
2860 * Pairs with the smp_wmb() in finish_lock_switch().
2864 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2865 p->state = TASK_WAKING;
2867 if (p->sched_class->task_waking)
2868 p->sched_class->task_waking(p);
2870 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2871 if (task_cpu(p) != cpu) {
2872 wake_flags |= WF_MIGRATED;
2873 set_task_cpu(p, cpu);
2875 #endif /* CONFIG_SMP */
2879 ttwu_stat(p, cpu, wake_flags);
2881 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2887 * try_to_wake_up_local - try to wake up a local task with rq lock held
2888 * @p: the thread to be awakened
2890 * Put @p on the run-queue if it's not already there. The caller must
2891 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2894 static void try_to_wake_up_local(struct task_struct *p)
2896 struct rq *rq = task_rq(p);
2898 if (WARN_ON_ONCE(rq != this_rq()) ||
2899 WARN_ON_ONCE(p == current))
2902 lockdep_assert_held(&rq->lock);
2904 if (!raw_spin_trylock(&p->pi_lock)) {
2905 raw_spin_unlock(&rq->lock);
2906 raw_spin_lock(&p->pi_lock);
2907 raw_spin_lock(&rq->lock);
2910 if (!(p->state & TASK_NORMAL))
2914 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2916 ttwu_do_wakeup(rq, p, 0);
2917 ttwu_stat(p, smp_processor_id(), 0);
2919 raw_spin_unlock(&p->pi_lock);
2923 * wake_up_process - Wake up a specific process
2924 * @p: The process to be woken up.
2926 * Attempt to wake up the nominated process and move it to the set of runnable
2927 * processes. Returns 1 if the process was woken up, 0 if it was already
2930 * It may be assumed that this function implies a write memory barrier before
2931 * changing the task state if and only if any tasks are woken up.
2933 int wake_up_process(struct task_struct *p)
2935 WARN_ON(task_is_stopped_or_traced(p));
2936 return try_to_wake_up(p, TASK_NORMAL, 0);
2938 EXPORT_SYMBOL(wake_up_process);
2940 int wake_up_state(struct task_struct *p, unsigned int state)
2942 return try_to_wake_up(p, state, 0);
2946 * Perform scheduler related setup for a newly forked process p.
2947 * p is forked by current.
2949 * __sched_fork() is basic setup used by init_idle() too:
2951 static void __sched_fork(struct task_struct *p)
2956 p->se.exec_start = 0;
2957 p->se.sum_exec_runtime = 0;
2958 p->se.prev_sum_exec_runtime = 0;
2959 p->se.nr_migrations = 0;
2961 INIT_LIST_HEAD(&p->se.group_node);
2963 #ifdef CONFIG_SCHEDSTATS
2964 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2967 INIT_LIST_HEAD(&p->rt.run_list);
2969 #ifdef CONFIG_PREEMPT_NOTIFIERS
2970 INIT_HLIST_HEAD(&p->preempt_notifiers);
2975 * fork()/clone()-time setup:
2977 void sched_fork(struct task_struct *p)
2979 unsigned long flags;
2980 int cpu = get_cpu();
2984 * We mark the process as running here. This guarantees that
2985 * nobody will actually run it, and a signal or other external
2986 * event cannot wake it up and insert it on the runqueue either.
2988 p->state = TASK_RUNNING;
2991 * Make sure we do not leak PI boosting priority to the child.
2993 p->prio = current->normal_prio;
2996 * Revert to default priority/policy on fork if requested.
2998 if (unlikely(p->sched_reset_on_fork)) {
2999 if (task_has_rt_policy(p)) {
3000 p->policy = SCHED_NORMAL;
3001 p->static_prio = NICE_TO_PRIO(0);
3003 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3004 p->static_prio = NICE_TO_PRIO(0);
3006 p->prio = p->normal_prio = __normal_prio(p);
3010 * We don't need the reset flag anymore after the fork. It has
3011 * fulfilled its duty:
3013 p->sched_reset_on_fork = 0;
3016 if (!rt_prio(p->prio))
3017 p->sched_class = &fair_sched_class;
3019 if (p->sched_class->task_fork)
3020 p->sched_class->task_fork(p);
3023 * The child is not yet in the pid-hash so no cgroup attach races,
3024 * and the cgroup is pinned to this child due to cgroup_fork()
3025 * is ran before sched_fork().
3027 * Silence PROVE_RCU.
3029 raw_spin_lock_irqsave(&p->pi_lock, flags);
3030 set_task_cpu(p, cpu);
3031 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3033 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3034 if (likely(sched_info_on()))
3035 memset(&p->sched_info, 0, sizeof(p->sched_info));
3037 #if defined(CONFIG_SMP)
3040 #ifdef CONFIG_PREEMPT_COUNT
3041 /* Want to start with kernel preemption disabled. */
3042 task_thread_info(p)->preempt_count = 1;
3045 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3052 * wake_up_new_task - wake up a newly created task for the first time.
3054 * This function will do some initial scheduler statistics housekeeping
3055 * that must be done for every newly created context, then puts the task
3056 * on the runqueue and wakes it.
3058 void wake_up_new_task(struct task_struct *p)
3060 unsigned long flags;
3063 raw_spin_lock_irqsave(&p->pi_lock, flags);
3066 * Fork balancing, do it here and not earlier because:
3067 * - cpus_allowed can change in the fork path
3068 * - any previously selected cpu might disappear through hotplug
3070 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3073 rq = __task_rq_lock(p);
3074 activate_task(rq, p, 0);
3076 trace_sched_wakeup_new(p, true);
3077 check_preempt_curr(rq, p, WF_FORK);
3079 if (p->sched_class->task_woken)
3080 p->sched_class->task_woken(rq, p);
3082 task_rq_unlock(rq, p, &flags);
3085 #ifdef CONFIG_PREEMPT_NOTIFIERS
3088 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3089 * @notifier: notifier struct to register
3091 void preempt_notifier_register(struct preempt_notifier *notifier)
3093 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3095 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3098 * preempt_notifier_unregister - no longer interested in preemption notifications
3099 * @notifier: notifier struct to unregister
3101 * This is safe to call from within a preemption notifier.
3103 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3105 hlist_del(¬ifier->link);
3107 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3109 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3111 struct preempt_notifier *notifier;
3112 struct hlist_node *node;
3114 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3115 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3119 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3120 struct task_struct *next)
3122 struct preempt_notifier *notifier;
3123 struct hlist_node *node;
3125 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3126 notifier->ops->sched_out(notifier, next);
3129 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3131 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3136 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3137 struct task_struct *next)
3141 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3144 * prepare_task_switch - prepare to switch tasks
3145 * @rq: the runqueue preparing to switch
3146 * @prev: the current task that is being switched out
3147 * @next: the task we are going to switch to.
3149 * This is called with the rq lock held and interrupts off. It must
3150 * be paired with a subsequent finish_task_switch after the context
3153 * prepare_task_switch sets up locking and calls architecture specific
3157 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3158 struct task_struct *next)
3160 sched_info_switch(prev, next);
3161 perf_event_task_sched_out(prev, next);
3162 fire_sched_out_preempt_notifiers(prev, next);
3163 prepare_lock_switch(rq, next);
3164 prepare_arch_switch(next);
3165 trace_sched_switch(prev, next);
3169 * finish_task_switch - clean up after a task-switch
3170 * @rq: runqueue associated with task-switch
3171 * @prev: the thread we just switched away from.
3173 * finish_task_switch must be called after the context switch, paired
3174 * with a prepare_task_switch call before the context switch.
3175 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3176 * and do any other architecture-specific cleanup actions.
3178 * Note that we may have delayed dropping an mm in context_switch(). If
3179 * so, we finish that here outside of the runqueue lock. (Doing it
3180 * with the lock held can cause deadlocks; see schedule() for
3183 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3184 __releases(rq->lock)
3186 struct mm_struct *mm = rq->prev_mm;
3192 * A task struct has one reference for the use as "current".
3193 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3194 * schedule one last time. The schedule call will never return, and
3195 * the scheduled task must drop that reference.
3197 * We must observe prev->state before clearing prev->on_cpu (in
3198 * finish_lock_switch), otherwise a concurrent wakeup can get prev
3199 * running on another CPU and we could rave with its RUNNING -> DEAD
3200 * transition, resulting in a double drop.
3202 prev_state = prev->state;
3203 finish_arch_switch(prev);
3204 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3205 local_irq_disable();
3206 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3207 perf_event_task_sched_in(prev, current);
3208 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3210 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3211 finish_lock_switch(rq, prev);
3213 fire_sched_in_preempt_notifiers(current);
3216 if (unlikely(prev_state == TASK_DEAD)) {
3218 * Remove function-return probe instances associated with this
3219 * task and put them back on the free list.
3221 kprobe_flush_task(prev);
3222 put_task_struct(prev);
3228 /* assumes rq->lock is held */
3229 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3231 if (prev->sched_class->pre_schedule)
3232 prev->sched_class->pre_schedule(rq, prev);
3235 /* rq->lock is NOT held, but preemption is disabled */
3236 static inline void post_schedule(struct rq *rq)
3238 if (rq->post_schedule) {
3239 unsigned long flags;
3241 raw_spin_lock_irqsave(&rq->lock, flags);
3242 if (rq->curr->sched_class->post_schedule)
3243 rq->curr->sched_class->post_schedule(rq);
3244 raw_spin_unlock_irqrestore(&rq->lock, flags);
3246 rq->post_schedule = 0;
3252 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3256 static inline void post_schedule(struct rq *rq)
3263 * schedule_tail - first thing a freshly forked thread must call.
3264 * @prev: the thread we just switched away from.
3266 asmlinkage void schedule_tail(struct task_struct *prev)
3267 __releases(rq->lock)
3269 struct rq *rq = this_rq();
3271 finish_task_switch(rq, prev);
3274 * FIXME: do we need to worry about rq being invalidated by the
3279 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3280 /* In this case, finish_task_switch does not reenable preemption */
3283 if (current->set_child_tid)
3284 put_user(task_pid_vnr(current), current->set_child_tid);
3288 * context_switch - switch to the new MM and the new
3289 * thread's register state.
3292 context_switch(struct rq *rq, struct task_struct *prev,
3293 struct task_struct *next)
3295 struct mm_struct *mm, *oldmm;
3297 prepare_task_switch(rq, prev, next);
3300 oldmm = prev->active_mm;
3302 * For paravirt, this is coupled with an exit in switch_to to
3303 * combine the page table reload and the switch backend into
3306 arch_start_context_switch(prev);
3309 next->active_mm = oldmm;
3310 atomic_inc(&oldmm->mm_count);
3311 enter_lazy_tlb(oldmm, next);
3313 switch_mm(oldmm, mm, next);
3316 prev->active_mm = NULL;
3317 rq->prev_mm = oldmm;
3320 * Since the runqueue lock will be released by the next
3321 * task (which is an invalid locking op but in the case
3322 * of the scheduler it's an obvious special-case), so we
3323 * do an early lockdep release here:
3325 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3326 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3329 /* Here we just switch the register state and the stack. */
3330 switch_to(prev, next, prev);
3334 * this_rq must be evaluated again because prev may have moved
3335 * CPUs since it called schedule(), thus the 'rq' on its stack
3336 * frame will be invalid.
3338 finish_task_switch(this_rq(), prev);
3342 * nr_running, nr_uninterruptible and nr_context_switches:
3344 * externally visible scheduler statistics: current number of runnable
3345 * threads, current number of uninterruptible-sleeping threads, total
3346 * number of context switches performed since bootup.
3348 unsigned long nr_running(void)
3350 unsigned long i, sum = 0;
3352 for_each_online_cpu(i)
3353 sum += cpu_rq(i)->nr_running;
3358 unsigned long nr_uninterruptible(void)
3360 unsigned long i, sum = 0;
3362 for_each_possible_cpu(i)
3363 sum += cpu_rq(i)->nr_uninterruptible;
3366 * Since we read the counters lockless, it might be slightly
3367 * inaccurate. Do not allow it to go below zero though:
3369 if (unlikely((long)sum < 0))
3375 unsigned long long nr_context_switches(void)
3378 unsigned long long sum = 0;
3380 for_each_possible_cpu(i)
3381 sum += cpu_rq(i)->nr_switches;
3386 unsigned long nr_iowait(void)
3388 unsigned long i, sum = 0;
3390 for_each_possible_cpu(i)
3391 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3396 unsigned long nr_iowait_cpu(int cpu)
3398 struct rq *this = cpu_rq(cpu);
3399 return atomic_read(&this->nr_iowait);
3402 unsigned long this_cpu_load(void)
3404 struct rq *this = this_rq();
3405 return this->cpu_load[0];
3410 * Global load-average calculations
3412 * We take a distributed and async approach to calculating the global load-avg
3413 * in order to minimize overhead.
3415 * The global load average is an exponentially decaying average of nr_running +
3416 * nr_uninterruptible.
3418 * Once every LOAD_FREQ:
3421 * for_each_possible_cpu(cpu)
3422 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
3424 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
3426 * Due to a number of reasons the above turns in the mess below:
3428 * - for_each_possible_cpu() is prohibitively expensive on machines with
3429 * serious number of cpus, therefore we need to take a distributed approach
3430 * to calculating nr_active.
3432 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
3433 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
3435 * So assuming nr_active := 0 when we start out -- true per definition, we
3436 * can simply take per-cpu deltas and fold those into a global accumulate
3437 * to obtain the same result. See calc_load_fold_active().
3439 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
3440 * across the machine, we assume 10 ticks is sufficient time for every
3441 * cpu to have completed this task.
3443 * This places an upper-bound on the IRQ-off latency of the machine. Then
3444 * again, being late doesn't loose the delta, just wrecks the sample.
3446 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
3447 * this would add another cross-cpu cacheline miss and atomic operation
3448 * to the wakeup path. Instead we increment on whatever cpu the task ran
3449 * when it went into uninterruptible state and decrement on whatever cpu
3450 * did the wakeup. This means that only the sum of nr_uninterruptible over
3451 * all cpus yields the correct result.
3453 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
3456 /* Variables and functions for calc_load */
3457 static atomic_long_t calc_load_tasks;
3458 static unsigned long calc_load_update;
3459 unsigned long avenrun[3];
3460 EXPORT_SYMBOL(avenrun); /* should be removed */
3463 * get_avenrun - get the load average array
3464 * @loads: pointer to dest load array
3465 * @offset: offset to add
3466 * @shift: shift count to shift the result left
3468 * These values are estimates at best, so no need for locking.
3470 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3472 loads[0] = (avenrun[0] + offset) << shift;
3473 loads[1] = (avenrun[1] + offset) << shift;
3474 loads[2] = (avenrun[2] + offset) << shift;
3477 static long calc_load_fold_active(struct rq *this_rq)
3479 long nr_active, delta = 0;
3481 nr_active = this_rq->nr_running;
3482 nr_active += (long) this_rq->nr_uninterruptible;
3484 if (nr_active != this_rq->calc_load_active) {
3485 delta = nr_active - this_rq->calc_load_active;
3486 this_rq->calc_load_active = nr_active;
3493 * a1 = a0 * e + a * (1 - e)
3495 static unsigned long
3496 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3499 load += active * (FIXED_1 - exp);
3500 load += 1UL << (FSHIFT - 1);
3501 return load >> FSHIFT;
3506 * Handle NO_HZ for the global load-average.
3508 * Since the above described distributed algorithm to compute the global
3509 * load-average relies on per-cpu sampling from the tick, it is affected by
3512 * The basic idea is to fold the nr_active delta into a global idle-delta upon
3513 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
3514 * when we read the global state.
3516 * Obviously reality has to ruin such a delightfully simple scheme:
3518 * - When we go NO_HZ idle during the window, we can negate our sample
3519 * contribution, causing under-accounting.
3521 * We avoid this by keeping two idle-delta counters and flipping them
3522 * when the window starts, thus separating old and new NO_HZ load.
3524 * The only trick is the slight shift in index flip for read vs write.
3528 * |-|-----------|-|-----------|-|-----------|-|
3529 * r:0 0 1 1 0 0 1 1 0
3530 * w:0 1 1 0 0 1 1 0 0
3532 * This ensures we'll fold the old idle contribution in this window while
3533 * accumlating the new one.
3535 * - When we wake up from NO_HZ idle during the window, we push up our
3536 * contribution, since we effectively move our sample point to a known
3539 * This is solved by pushing the window forward, and thus skipping the
3540 * sample, for this cpu (effectively using the idle-delta for this cpu which
3541 * was in effect at the time the window opened). This also solves the issue
3542 * of having to deal with a cpu having been in NOHZ idle for multiple
3543 * LOAD_FREQ intervals.
3545 * When making the ILB scale, we should try to pull this in as well.
3547 static atomic_long_t calc_load_idle[2];
3548 static int calc_load_idx;
3550 static inline int calc_load_write_idx(void)
3552 int idx = calc_load_idx;
3555 * See calc_global_nohz(), if we observe the new index, we also
3556 * need to observe the new update time.
3561 * If the folding window started, make sure we start writing in the
3564 if (!time_before(jiffies, calc_load_update))
3570 static inline int calc_load_read_idx(void)
3572 return calc_load_idx & 1;
3575 void calc_load_enter_idle(void)
3577 struct rq *this_rq = this_rq();
3581 * We're going into NOHZ mode, if there's any pending delta, fold it
3582 * into the pending idle delta.
3584 delta = calc_load_fold_active(this_rq);
3586 int idx = calc_load_write_idx();
3587 atomic_long_add(delta, &calc_load_idle[idx]);
3591 void calc_load_exit_idle(void)
3593 struct rq *this_rq = this_rq();
3596 * If we're still before the sample window, we're done.
3598 if (time_before(jiffies, this_rq->calc_load_update))
3602 * We woke inside or after the sample window, this means we're already
3603 * accounted through the nohz accounting, so skip the entire deal and
3604 * sync up for the next window.
3606 this_rq->calc_load_update = calc_load_update;
3607 if (time_before(jiffies, this_rq->calc_load_update + 10))
3608 this_rq->calc_load_update += LOAD_FREQ;
3611 static long calc_load_fold_idle(void)
3613 int idx = calc_load_read_idx();
3616 if (atomic_long_read(&calc_load_idle[idx]))
3617 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
3623 * fixed_power_int - compute: x^n, in O(log n) time
3625 * @x: base of the power
3626 * @frac_bits: fractional bits of @x
3627 * @n: power to raise @x to.
3629 * By exploiting the relation between the definition of the natural power
3630 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3631 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3632 * (where: n_i \elem {0, 1}, the binary vector representing n),
3633 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3634 * of course trivially computable in O(log_2 n), the length of our binary
3637 static unsigned long
3638 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3640 unsigned long result = 1UL << frac_bits;
3645 result += 1UL << (frac_bits - 1);
3646 result >>= frac_bits;
3652 x += 1UL << (frac_bits - 1);
3660 * a1 = a0 * e + a * (1 - e)
3662 * a2 = a1 * e + a * (1 - e)
3663 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3664 * = a0 * e^2 + a * (1 - e) * (1 + e)
3666 * a3 = a2 * e + a * (1 - e)
3667 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3668 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3672 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3673 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3674 * = a0 * e^n + a * (1 - e^n)
3676 * [1] application of the geometric series:
3679 * S_n := \Sum x^i = -------------
3682 static unsigned long
3683 calc_load_n(unsigned long load, unsigned long exp,
3684 unsigned long active, unsigned int n)
3687 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3691 * NO_HZ can leave us missing all per-cpu ticks calling
3692 * calc_load_account_active(), but since an idle CPU folds its delta into
3693 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3694 * in the pending idle delta if our idle period crossed a load cycle boundary.
3696 * Once we've updated the global active value, we need to apply the exponential
3697 * weights adjusted to the number of cycles missed.
3699 static void calc_global_nohz(void)
3701 long delta, active, n;
3703 if (!time_before(jiffies, calc_load_update + 10)) {
3705 * Catch-up, fold however many we are behind still
3707 delta = jiffies - calc_load_update - 10;
3708 n = 1 + (delta / LOAD_FREQ);
3710 active = atomic_long_read(&calc_load_tasks);
3711 active = active > 0 ? active * FIXED_1 : 0;
3713 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3714 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3715 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3717 calc_load_update += n * LOAD_FREQ;
3721 * Flip the idle index...
3723 * Make sure we first write the new time then flip the index, so that
3724 * calc_load_write_idx() will see the new time when it reads the new
3725 * index, this avoids a double flip messing things up.
3730 #else /* !CONFIG_NO_HZ */
3732 static inline long calc_load_fold_idle(void) { return 0; }
3733 static inline void calc_global_nohz(void) { }
3735 #endif /* CONFIG_NO_HZ */
3738 * calc_load - update the avenrun load estimates 10 ticks after the
3739 * CPUs have updated calc_load_tasks.
3741 void calc_global_load(unsigned long ticks)
3745 if (time_before(jiffies, calc_load_update + 10))
3749 * Fold the 'old' idle-delta to include all NO_HZ cpus.
3751 delta = calc_load_fold_idle();
3753 atomic_long_add(delta, &calc_load_tasks);
3755 active = atomic_long_read(&calc_load_tasks);
3756 active = active > 0 ? active * FIXED_1 : 0;
3758 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3759 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3760 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3762 calc_load_update += LOAD_FREQ;
3765 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
3771 * Called from update_cpu_load() to periodically update this CPU's
3774 static void calc_load_account_active(struct rq *this_rq)
3778 if (time_before(jiffies, this_rq->calc_load_update))
3781 delta = calc_load_fold_active(this_rq);
3783 atomic_long_add(delta, &calc_load_tasks);
3785 this_rq->calc_load_update += LOAD_FREQ;
3789 * End of global load-average stuff
3793 * The exact cpuload at various idx values, calculated at every tick would be
3794 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3796 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3797 * on nth tick when cpu may be busy, then we have:
3798 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3799 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3801 * decay_load_missed() below does efficient calculation of
3802 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3803 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3805 * The calculation is approximated on a 128 point scale.
3806 * degrade_zero_ticks is the number of ticks after which load at any
3807 * particular idx is approximated to be zero.
3808 * degrade_factor is a precomputed table, a row for each load idx.
3809 * Each column corresponds to degradation factor for a power of two ticks,
3810 * based on 128 point scale.
3812 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3813 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3815 * With this power of 2 load factors, we can degrade the load n times
3816 * by looking at 1 bits in n and doing as many mult/shift instead of
3817 * n mult/shifts needed by the exact degradation.
3819 #define DEGRADE_SHIFT 7
3820 static const unsigned char
3821 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3822 static const unsigned char
3823 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3824 {0, 0, 0, 0, 0, 0, 0, 0},
3825 {64, 32, 8, 0, 0, 0, 0, 0},
3826 {96, 72, 40, 12, 1, 0, 0},
3827 {112, 98, 75, 43, 15, 1, 0},
3828 {120, 112, 98, 76, 45, 16, 2} };
3831 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3832 * would be when CPU is idle and so we just decay the old load without
3833 * adding any new load.
3835 static unsigned long
3836 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3840 if (!missed_updates)
3843 if (missed_updates >= degrade_zero_ticks[idx])
3847 return load >> missed_updates;
3849 while (missed_updates) {
3850 if (missed_updates % 2)
3851 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3853 missed_updates >>= 1;
3860 * Update rq->cpu_load[] statistics. This function is usually called every
3861 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3862 * every tick. We fix it up based on jiffies.
3864 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
3865 unsigned long pending_updates)
3869 this_rq->nr_load_updates++;
3871 /* Update our load: */
3872 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3873 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3874 unsigned long old_load, new_load;
3876 /* scale is effectively 1 << i now, and >> i divides by scale */
3878 old_load = this_rq->cpu_load[i];
3879 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3880 new_load = this_load;
3882 * Round up the averaging division if load is increasing. This
3883 * prevents us from getting stuck on 9 if the load is 10, for
3886 if (new_load > old_load)
3887 new_load += scale - 1;
3889 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3892 sched_avg_update(this_rq);
3897 * There is no sane way to deal with nohz on smp when using jiffies because the
3898 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
3899 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
3901 * Therefore we cannot use the delta approach from the regular tick since that
3902 * would seriously skew the load calculation. However we'll make do for those
3903 * updates happening while idle (nohz_idle_balance) or coming out of idle
3904 * (tick_nohz_idle_exit).
3906 * This means we might still be one tick off for nohz periods.
3910 * Called from nohz_idle_balance() to update the load ratings before doing the
3913 static void update_idle_cpu_load(struct rq *this_rq)
3915 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3916 unsigned long load = this_rq->load.weight;
3917 unsigned long pending_updates;
3920 * bail if there's load or we're actually up-to-date.
3922 if (load || curr_jiffies == this_rq->last_load_update_tick)
3925 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3926 this_rq->last_load_update_tick = curr_jiffies;
3928 __update_cpu_load(this_rq, load, pending_updates);
3932 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
3934 void update_cpu_load_nohz(void)
3936 struct rq *this_rq = this_rq();
3937 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3938 unsigned long pending_updates;
3940 if (curr_jiffies == this_rq->last_load_update_tick)
3943 raw_spin_lock(&this_rq->lock);
3944 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3945 if (pending_updates) {
3946 this_rq->last_load_update_tick = curr_jiffies;
3948 * We were idle, this means load 0, the current load might be
3949 * !0 due to remote wakeups and the sort.
3951 __update_cpu_load(this_rq, 0, pending_updates);
3953 raw_spin_unlock(&this_rq->lock);
3955 #endif /* CONFIG_NO_HZ */
3958 * Called from scheduler_tick()
3960 static void update_cpu_load_active(struct rq *this_rq)
3963 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
3965 this_rq->last_load_update_tick = jiffies;
3966 __update_cpu_load(this_rq, this_rq->load.weight, 1);
3968 calc_load_account_active(this_rq);
3974 * sched_exec - execve() is a valuable balancing opportunity, because at
3975 * this point the task has the smallest effective memory and cache footprint.
3977 void sched_exec(void)
3979 struct task_struct *p = current;
3980 unsigned long flags;
3983 raw_spin_lock_irqsave(&p->pi_lock, flags);
3984 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3985 if (dest_cpu == smp_processor_id())
3988 if (likely(cpu_active(dest_cpu))) {
3989 struct migration_arg arg = { p, dest_cpu };
3991 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3992 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3996 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4001 DEFINE_PER_CPU(struct kernel_stat, kstat);
4003 EXPORT_PER_CPU_SYMBOL(kstat);
4006 * Return any ns on the sched_clock that have not yet been accounted in
4007 * @p in case that task is currently running.
4009 * Called with task_rq_lock() held on @rq.
4011 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4015 if (task_current(rq, p)) {
4016 update_rq_clock(rq);
4017 ns = rq->clock_task - p->se.exec_start;
4025 unsigned long long task_delta_exec(struct task_struct *p)
4027 unsigned long flags;
4031 rq = task_rq_lock(p, &flags);
4032 ns = do_task_delta_exec(p, rq);
4033 task_rq_unlock(rq, p, &flags);
4039 * Return accounted runtime for the task.
4040 * In case the task is currently running, return the runtime plus current's
4041 * pending runtime that have not been accounted yet.
4043 unsigned long long task_sched_runtime(struct task_struct *p)
4045 unsigned long flags;
4049 rq = task_rq_lock(p, &flags);
4050 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4051 task_rq_unlock(rq, p, &flags);
4057 * Account user cpu time to a process.
4058 * @p: the process that the cpu time gets accounted to
4059 * @cputime: the cpu time spent in user space since the last update
4060 * @cputime_scaled: cputime scaled by cpu frequency
4062 void account_user_time(struct task_struct *p, cputime_t cputime,
4063 cputime_t cputime_scaled)
4065 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4068 /* Add user time to process. */
4069 p->utime = cputime_add(p->utime, cputime);
4070 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4071 account_group_user_time(p, cputime);
4073 /* Add user time to cpustat. */
4074 tmp = cputime_to_cputime64(cputime);
4075 if (TASK_NICE(p) > 0)
4076 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4078 cpustat->user = cputime64_add(cpustat->user, tmp);
4080 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4081 /* Account for user time used */
4082 acct_update_integrals(p);
4086 * Account guest cpu time to a process.
4087 * @p: the process that the cpu time gets accounted to
4088 * @cputime: the cpu time spent in virtual machine since the last update
4089 * @cputime_scaled: cputime scaled by cpu frequency
4091 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4092 cputime_t cputime_scaled)
4095 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4097 tmp = cputime_to_cputime64(cputime);
4099 /* Add guest time to process. */
4100 p->utime = cputime_add(p->utime, cputime);
4101 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4102 account_group_user_time(p, cputime);
4103 p->gtime = cputime_add(p->gtime, cputime);
4105 /* Add guest time to cpustat. */
4106 if (TASK_NICE(p) > 0) {
4107 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4108 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
4110 cpustat->user = cputime64_add(cpustat->user, tmp);
4111 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4116 * Account system cpu time to a process and desired cpustat field
4117 * @p: the process that the cpu time gets accounted to
4118 * @cputime: the cpu time spent in kernel space since the last update
4119 * @cputime_scaled: cputime scaled by cpu frequency
4120 * @target_cputime64: pointer to cpustat field that has to be updated
4123 void __account_system_time(struct task_struct *p, cputime_t cputime,
4124 cputime_t cputime_scaled, cputime64_t *target_cputime64)
4126 cputime64_t tmp = cputime_to_cputime64(cputime);
4128 /* Add system time to process. */
4129 p->stime = cputime_add(p->stime, cputime);
4130 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4131 account_group_system_time(p, cputime);
4133 /* Add system time to cpustat. */
4134 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
4135 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4137 /* Account for system time used */
4138 acct_update_integrals(p);
4142 * Account system cpu time to a process.
4143 * @p: the process that the cpu time gets accounted to
4144 * @hardirq_offset: the offset to subtract from hardirq_count()
4145 * @cputime: the cpu time spent in kernel space since the last update
4146 * @cputime_scaled: cputime scaled by cpu frequency
4148 void account_system_time(struct task_struct *p, int hardirq_offset,
4149 cputime_t cputime, cputime_t cputime_scaled)
4151 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4152 cputime64_t *target_cputime64;
4154 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4155 account_guest_time(p, cputime, cputime_scaled);
4159 if (hardirq_count() - hardirq_offset)
4160 target_cputime64 = &cpustat->irq;
4161 else if (in_serving_softirq())
4162 target_cputime64 = &cpustat->softirq;
4164 target_cputime64 = &cpustat->system;
4166 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
4170 * Account for involuntary wait time.
4171 * @cputime: the cpu time spent in involuntary wait
4173 void account_steal_time(cputime_t cputime)
4175 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4176 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4178 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4182 * Account for idle time.
4183 * @cputime: the cpu time spent in idle wait
4185 void account_idle_time(cputime_t cputime)
4187 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4188 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4189 struct rq *rq = this_rq();
4191 if (atomic_read(&rq->nr_iowait) > 0)
4192 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4194 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4197 static __always_inline bool steal_account_process_tick(void)
4199 #ifdef CONFIG_PARAVIRT
4200 if (static_branch(¶virt_steal_enabled)) {
4203 steal = paravirt_steal_clock(smp_processor_id());
4204 steal -= this_rq()->prev_steal_time;
4206 st = steal_ticks(steal);
4207 this_rq()->prev_steal_time += st * TICK_NSEC;
4209 account_steal_time(st);
4216 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4218 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4220 * Account a tick to a process and cpustat
4221 * @p: the process that the cpu time gets accounted to
4222 * @user_tick: is the tick from userspace
4223 * @rq: the pointer to rq
4225 * Tick demultiplexing follows the order
4226 * - pending hardirq update
4227 * - pending softirq update
4231 * - check for guest_time
4232 * - else account as system_time
4234 * Check for hardirq is done both for system and user time as there is
4235 * no timer going off while we are on hardirq and hence we may never get an
4236 * opportunity to update it solely in system time.
4237 * p->stime and friends are only updated on system time and not on irq
4238 * softirq as those do not count in task exec_runtime any more.
4240 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4243 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4244 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
4245 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4247 if (steal_account_process_tick())
4250 if (irqtime_account_hi_update()) {
4251 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4252 } else if (irqtime_account_si_update()) {
4253 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4254 } else if (this_cpu_ksoftirqd() == p) {
4256 * ksoftirqd time do not get accounted in cpu_softirq_time.
4257 * So, we have to handle it separately here.
4258 * Also, p->stime needs to be updated for ksoftirqd.
4260 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4262 } else if (user_tick) {
4263 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4264 } else if (p == rq->idle) {
4265 account_idle_time(cputime_one_jiffy);
4266 } else if (p->flags & PF_VCPU) { /* System time or guest time */
4267 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
4269 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4274 static void irqtime_account_idle_ticks(int ticks)
4277 struct rq *rq = this_rq();
4279 for (i = 0; i < ticks; i++)
4280 irqtime_account_process_tick(current, 0, rq);
4282 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4283 static void irqtime_account_idle_ticks(int ticks) {}
4284 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4286 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4289 * Account a single tick of cpu time.
4290 * @p: the process that the cpu time gets accounted to
4291 * @user_tick: indicates if the tick is a user or a system tick
4293 void account_process_tick(struct task_struct *p, int user_tick)
4295 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4296 struct rq *rq = this_rq();
4298 if (sched_clock_irqtime) {
4299 irqtime_account_process_tick(p, user_tick, rq);
4303 if (steal_account_process_tick())
4307 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4308 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4309 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4312 account_idle_time(cputime_one_jiffy);
4316 * Account multiple ticks of steal time.
4317 * @p: the process from which the cpu time has been stolen
4318 * @ticks: number of stolen ticks
4320 void account_steal_ticks(unsigned long ticks)
4322 account_steal_time(jiffies_to_cputime(ticks));
4326 * Account multiple ticks of idle time.
4327 * @ticks: number of stolen ticks
4329 void account_idle_ticks(unsigned long ticks)
4332 if (sched_clock_irqtime) {
4333 irqtime_account_idle_ticks(ticks);
4337 account_idle_time(jiffies_to_cputime(ticks));
4343 * Use precise platform statistics if available:
4345 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4346 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4352 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4354 struct task_cputime cputime;
4356 thread_group_cputime(p, &cputime);
4358 *ut = cputime.utime;
4359 *st = cputime.stime;
4363 #ifndef nsecs_to_cputime
4364 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4367 static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total)
4369 u64 temp = (__force u64) rtime;
4371 temp *= (__force u64) utime;
4373 if (sizeof(cputime_t) == 4)
4374 temp = div_u64(temp, (__force u32) total);
4376 temp = div64_u64(temp, (__force u64) total);
4378 return (__force cputime_t) temp;
4381 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4383 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4386 * Use CFS's precise accounting:
4388 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4391 utime = scale_utime(utime, rtime, total);
4396 * Compare with previous values, to keep monotonicity:
4398 p->prev_utime = max(p->prev_utime, utime);
4399 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4401 *ut = p->prev_utime;
4402 *st = p->prev_stime;
4406 * Must be called with siglock held.
4408 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4410 struct signal_struct *sig = p->signal;
4411 struct task_cputime cputime;
4412 cputime_t rtime, utime, total;
4414 thread_group_cputime(p, &cputime);
4416 total = cputime_add(cputime.utime, cputime.stime);
4417 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4420 utime = scale_utime(cputime.utime, rtime, total);
4424 sig->prev_utime = max(sig->prev_utime, utime);
4425 sig->prev_stime = max(sig->prev_stime,
4426 cputime_sub(rtime, sig->prev_utime));
4428 *ut = sig->prev_utime;
4429 *st = sig->prev_stime;
4434 * This function gets called by the timer code, with HZ frequency.
4435 * We call it with interrupts disabled.
4437 void scheduler_tick(void)
4439 int cpu = smp_processor_id();
4440 struct rq *rq = cpu_rq(cpu);
4441 struct task_struct *curr = rq->curr;
4445 raw_spin_lock(&rq->lock);
4446 update_rq_clock(rq);
4447 update_cpu_load_active(rq);
4448 curr->sched_class->task_tick(rq, curr, 0);
4449 raw_spin_unlock(&rq->lock);
4451 perf_event_task_tick();
4454 rq->idle_balance = idle_cpu(cpu);
4455 trigger_load_balance(rq, cpu);
4459 notrace unsigned long get_parent_ip(unsigned long addr)
4461 if (in_lock_functions(addr)) {
4462 addr = CALLER_ADDR2;
4463 if (in_lock_functions(addr))
4464 addr = CALLER_ADDR3;
4469 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4470 defined(CONFIG_PREEMPT_TRACER))
4472 void __kprobes add_preempt_count(int val)
4474 #ifdef CONFIG_DEBUG_PREEMPT
4478 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4481 preempt_count() += val;
4482 #ifdef CONFIG_DEBUG_PREEMPT
4484 * Spinlock count overflowing soon?
4486 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4489 if (preempt_count() == val)
4490 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4492 EXPORT_SYMBOL(add_preempt_count);
4494 void __kprobes sub_preempt_count(int val)
4496 #ifdef CONFIG_DEBUG_PREEMPT
4500 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4503 * Is the spinlock portion underflowing?
4505 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4506 !(preempt_count() & PREEMPT_MASK)))
4510 if (preempt_count() == val)
4511 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4512 preempt_count() -= val;
4514 EXPORT_SYMBOL(sub_preempt_count);
4519 * Print scheduling while atomic bug:
4521 static noinline void __schedule_bug(struct task_struct *prev)
4523 struct pt_regs *regs = get_irq_regs();
4525 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4526 prev->comm, prev->pid, preempt_count());
4528 debug_show_held_locks(prev);
4530 if (irqs_disabled())
4531 print_irqtrace_events(prev);
4540 * Various schedule()-time debugging checks and statistics:
4542 static inline void schedule_debug(struct task_struct *prev)
4545 * Test if we are atomic. Since do_exit() needs to call into
4546 * schedule() atomically, we ignore that path for now.
4547 * Otherwise, whine if we are scheduling when we should not be.
4549 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4550 __schedule_bug(prev);
4553 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4555 schedstat_inc(this_rq(), sched_count);
4558 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4560 if (prev->on_rq || rq->skip_clock_update < 0)
4561 update_rq_clock(rq);
4562 prev->sched_class->put_prev_task(rq, prev);
4566 * Pick up the highest-prio task:
4568 static inline struct task_struct *
4569 pick_next_task(struct rq *rq)
4571 const struct sched_class *class;
4572 struct task_struct *p;
4575 * Optimization: we know that if all tasks are in
4576 * the fair class we can call that function directly:
4578 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4579 p = fair_sched_class.pick_next_task(rq);
4584 for_each_class(class) {
4585 p = class->pick_next_task(rq);
4590 BUG(); /* the idle class will always have a runnable task */
4594 * __schedule() is the main scheduler function.
4596 static void __sched __schedule(void)
4598 struct task_struct *prev, *next;
4599 unsigned long *switch_count;
4605 cpu = smp_processor_id();
4607 rcu_note_context_switch(cpu);
4610 schedule_debug(prev);
4612 if (sched_feat(HRTICK))
4615 raw_spin_lock_irq(&rq->lock);
4617 switch_count = &prev->nivcsw;
4618 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4619 if (unlikely(signal_pending_state(prev->state, prev))) {
4620 prev->state = TASK_RUNNING;
4622 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4626 * If a worker went to sleep, notify and ask workqueue
4627 * whether it wants to wake up a task to maintain
4630 if (prev->flags & PF_WQ_WORKER) {
4631 struct task_struct *to_wakeup;
4633 to_wakeup = wq_worker_sleeping(prev, cpu);
4635 try_to_wake_up_local(to_wakeup);
4638 switch_count = &prev->nvcsw;
4641 pre_schedule(rq, prev);
4643 if (unlikely(!rq->nr_running))
4644 idle_balance(cpu, rq);
4646 put_prev_task(rq, prev);
4647 next = pick_next_task(rq);
4648 clear_tsk_need_resched(prev);
4649 rq->skip_clock_update = 0;
4651 if (likely(prev != next)) {
4656 context_switch(rq, prev, next); /* unlocks the rq */
4658 * The context switch have flipped the stack from under us
4659 * and restored the local variables which were saved when
4660 * this task called schedule() in the past. prev == current
4661 * is still correct, but it can be moved to another cpu/rq.
4663 cpu = smp_processor_id();
4666 raw_spin_unlock_irq(&rq->lock);
4670 preempt_enable_no_resched();
4675 static inline void sched_submit_work(struct task_struct *tsk)
4680 * If we are going to sleep and we have plugged IO queued,
4681 * make sure to submit it to avoid deadlocks.
4683 if (blk_needs_flush_plug(tsk))
4684 blk_schedule_flush_plug(tsk);
4687 asmlinkage void __sched schedule(void)
4689 struct task_struct *tsk = current;
4691 sched_submit_work(tsk);
4694 EXPORT_SYMBOL(schedule);
4696 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4698 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4700 if (lock->owner != owner)
4704 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4705 * lock->owner still matches owner, if that fails, owner might
4706 * point to free()d memory, if it still matches, the rcu_read_lock()
4707 * ensures the memory stays valid.
4711 return owner->on_cpu;
4715 * Look out! "owner" is an entirely speculative pointer
4716 * access and not reliable.
4718 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4720 if (!sched_feat(OWNER_SPIN))
4724 while (owner_running(lock, owner)) {
4728 arch_mutex_cpu_relax();
4733 * We break out the loop above on need_resched() and when the
4734 * owner changed, which is a sign for heavy contention. Return
4735 * success only when lock->owner is NULL.
4737 return lock->owner == NULL;
4741 #ifdef CONFIG_PREEMPT
4743 * this is the entry point to schedule() from in-kernel preemption
4744 * off of preempt_enable. Kernel preemptions off return from interrupt
4745 * occur there and call schedule directly.
4747 asmlinkage void __sched notrace preempt_schedule(void)
4749 struct thread_info *ti = current_thread_info();
4752 * If there is a non-zero preempt_count or interrupts are disabled,
4753 * we do not want to preempt the current task. Just return..
4755 if (likely(ti->preempt_count || irqs_disabled()))
4759 add_preempt_count_notrace(PREEMPT_ACTIVE);
4761 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4764 * Check again in case we missed a preemption opportunity
4765 * between schedule and now.
4768 } while (need_resched());
4770 EXPORT_SYMBOL(preempt_schedule);
4773 * this is the entry point to schedule() from kernel preemption
4774 * off of irq context.
4775 * Note, that this is called and return with irqs disabled. This will
4776 * protect us against recursive calling from irq.
4778 asmlinkage void __sched preempt_schedule_irq(void)
4780 struct thread_info *ti = current_thread_info();
4782 /* Catch callers which need to be fixed */
4783 BUG_ON(ti->preempt_count || !irqs_disabled());
4786 add_preempt_count(PREEMPT_ACTIVE);
4789 local_irq_disable();
4790 sub_preempt_count(PREEMPT_ACTIVE);
4793 * Check again in case we missed a preemption opportunity
4794 * between schedule and now.
4797 } while (need_resched());
4800 #endif /* CONFIG_PREEMPT */
4802 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4805 return try_to_wake_up(curr->private, mode, wake_flags);
4807 EXPORT_SYMBOL(default_wake_function);
4810 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4811 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4812 * number) then we wake all the non-exclusive tasks and one exclusive task.
4814 * There are circumstances in which we can try to wake a task which has already
4815 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4816 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4818 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4819 int nr_exclusive, int wake_flags, void *key)
4821 wait_queue_t *curr, *next;
4823 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4824 unsigned flags = curr->flags;
4826 if (curr->func(curr, mode, wake_flags, key) &&
4827 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4833 * __wake_up - wake up threads blocked on a waitqueue.
4835 * @mode: which threads
4836 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4837 * @key: is directly passed to the wakeup function
4839 * It may be assumed that this function implies a write memory barrier before
4840 * changing the task state if and only if any tasks are woken up.
4842 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4843 int nr_exclusive, void *key)
4845 unsigned long flags;
4847 spin_lock_irqsave(&q->lock, flags);
4848 __wake_up_common(q, mode, nr_exclusive, 0, key);
4849 spin_unlock_irqrestore(&q->lock, flags);
4851 EXPORT_SYMBOL(__wake_up);
4854 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4856 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4858 __wake_up_common(q, mode, 1, 0, NULL);
4860 EXPORT_SYMBOL_GPL(__wake_up_locked);
4862 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4864 __wake_up_common(q, mode, 1, 0, key);
4866 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4869 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4871 * @mode: which threads
4872 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4873 * @key: opaque value to be passed to wakeup targets
4875 * The sync wakeup differs that the waker knows that it will schedule
4876 * away soon, so while the target thread will be woken up, it will not
4877 * be migrated to another CPU - ie. the two threads are 'synchronized'
4878 * with each other. This can prevent needless bouncing between CPUs.
4880 * On UP it can prevent extra preemption.
4882 * It may be assumed that this function implies a write memory barrier before
4883 * changing the task state if and only if any tasks are woken up.
4885 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4886 int nr_exclusive, void *key)
4888 unsigned long flags;
4889 int wake_flags = WF_SYNC;
4894 if (unlikely(!nr_exclusive))
4897 spin_lock_irqsave(&q->lock, flags);
4898 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4899 spin_unlock_irqrestore(&q->lock, flags);
4901 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4904 * __wake_up_sync - see __wake_up_sync_key()
4906 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4908 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4910 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4913 * complete: - signals a single thread waiting on this completion
4914 * @x: holds the state of this particular completion
4916 * This will wake up a single thread waiting on this completion. Threads will be
4917 * awakened in the same order in which they were queued.
4919 * See also complete_all(), wait_for_completion() and related routines.
4921 * It may be assumed that this function implies a write memory barrier before
4922 * changing the task state if and only if any tasks are woken up.
4924 void complete(struct completion *x)
4926 unsigned long flags;
4928 spin_lock_irqsave(&x->wait.lock, flags);
4930 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4931 spin_unlock_irqrestore(&x->wait.lock, flags);
4933 EXPORT_SYMBOL(complete);
4936 * complete_all: - signals all threads waiting on this completion
4937 * @x: holds the state of this particular completion
4939 * This will wake up all threads waiting on this particular completion event.
4941 * It may be assumed that this function implies a write memory barrier before
4942 * changing the task state if and only if any tasks are woken up.
4944 void complete_all(struct completion *x)
4946 unsigned long flags;
4948 spin_lock_irqsave(&x->wait.lock, flags);
4949 x->done += UINT_MAX/2;
4950 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4951 spin_unlock_irqrestore(&x->wait.lock, flags);
4953 EXPORT_SYMBOL(complete_all);
4955 static inline long __sched
4956 do_wait_for_common(struct completion *x, long timeout, int state)
4959 DECLARE_WAITQUEUE(wait, current);
4961 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4963 if (signal_pending_state(state, current)) {
4964 timeout = -ERESTARTSYS;
4967 __set_current_state(state);
4968 spin_unlock_irq(&x->wait.lock);
4969 timeout = schedule_timeout(timeout);
4970 spin_lock_irq(&x->wait.lock);
4971 } while (!x->done && timeout);
4972 __remove_wait_queue(&x->wait, &wait);
4977 return timeout ?: 1;
4981 wait_for_common(struct completion *x, long timeout, int state)
4985 spin_lock_irq(&x->wait.lock);
4986 timeout = do_wait_for_common(x, timeout, state);
4987 spin_unlock_irq(&x->wait.lock);
4992 * wait_for_completion: - waits for completion of a task
4993 * @x: holds the state of this particular completion
4995 * This waits to be signaled for completion of a specific task. It is NOT
4996 * interruptible and there is no timeout.
4998 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4999 * and interrupt capability. Also see complete().
5001 void __sched wait_for_completion(struct completion *x)
5003 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5005 EXPORT_SYMBOL(wait_for_completion);
5008 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5009 * @x: holds the state of this particular completion
5010 * @timeout: timeout value in jiffies
5012 * This waits for either a completion of a specific task to be signaled or for a
5013 * specified timeout to expire. The timeout is in jiffies. It is not
5016 * The return value is 0 if timed out, and positive (at least 1, or number of
5017 * jiffies left till timeout) if completed.
5019 unsigned long __sched
5020 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5022 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5024 EXPORT_SYMBOL(wait_for_completion_timeout);
5027 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5028 * @x: holds the state of this particular completion
5030 * This waits for completion of a specific task to be signaled. It is
5033 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
5035 int __sched wait_for_completion_interruptible(struct completion *x)
5037 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5038 if (t == -ERESTARTSYS)
5042 EXPORT_SYMBOL(wait_for_completion_interruptible);
5045 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5046 * @x: holds the state of this particular completion
5047 * @timeout: timeout value in jiffies
5049 * This waits for either a completion of a specific task to be signaled or for a
5050 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5052 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
5053 * positive (at least 1, or number of jiffies left till timeout) if completed.
5056 wait_for_completion_interruptible_timeout(struct completion *x,
5057 unsigned long timeout)
5059 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5061 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5064 * wait_for_completion_killable: - waits for completion of a task (killable)
5065 * @x: holds the state of this particular completion
5067 * This waits to be signaled for completion of a specific task. It can be
5068 * interrupted by a kill signal.
5070 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
5072 int __sched wait_for_completion_killable(struct completion *x)
5074 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5075 if (t == -ERESTARTSYS)
5079 EXPORT_SYMBOL(wait_for_completion_killable);
5082 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
5083 * @x: holds the state of this particular completion
5084 * @timeout: timeout value in jiffies
5086 * This waits for either a completion of a specific task to be
5087 * signaled or for a specified timeout to expire. It can be
5088 * interrupted by a kill signal. The timeout is in jiffies.
5090 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
5091 * positive (at least 1, or number of jiffies left till timeout) if completed.
5094 wait_for_completion_killable_timeout(struct completion *x,
5095 unsigned long timeout)
5097 return wait_for_common(x, timeout, TASK_KILLABLE);
5099 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
5102 * try_wait_for_completion - try to decrement a completion without blocking
5103 * @x: completion structure
5105 * Returns: 0 if a decrement cannot be done without blocking
5106 * 1 if a decrement succeeded.
5108 * If a completion is being used as a counting completion,
5109 * attempt to decrement the counter without blocking. This
5110 * enables us to avoid waiting if the resource the completion
5111 * is protecting is not available.
5113 bool try_wait_for_completion(struct completion *x)
5115 unsigned long flags;
5118 spin_lock_irqsave(&x->wait.lock, flags);
5123 spin_unlock_irqrestore(&x->wait.lock, flags);
5126 EXPORT_SYMBOL(try_wait_for_completion);
5129 * completion_done - Test to see if a completion has any waiters
5130 * @x: completion structure
5132 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5133 * 1 if there are no waiters.
5136 bool completion_done(struct completion *x)
5138 unsigned long flags;
5141 spin_lock_irqsave(&x->wait.lock, flags);
5144 spin_unlock_irqrestore(&x->wait.lock, flags);
5147 EXPORT_SYMBOL(completion_done);
5150 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5152 unsigned long flags;
5155 init_waitqueue_entry(&wait, current);
5157 __set_current_state(state);
5159 spin_lock_irqsave(&q->lock, flags);
5160 __add_wait_queue(q, &wait);
5161 spin_unlock(&q->lock);
5162 timeout = schedule_timeout(timeout);
5163 spin_lock_irq(&q->lock);
5164 __remove_wait_queue(q, &wait);
5165 spin_unlock_irqrestore(&q->lock, flags);
5170 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5172 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5174 EXPORT_SYMBOL(interruptible_sleep_on);
5177 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5179 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5181 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5183 void __sched sleep_on(wait_queue_head_t *q)
5185 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5187 EXPORT_SYMBOL(sleep_on);
5189 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5191 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5193 EXPORT_SYMBOL(sleep_on_timeout);
5195 #ifdef CONFIG_RT_MUTEXES
5198 * rt_mutex_setprio - set the current priority of a task
5200 * @prio: prio value (kernel-internal form)
5202 * This function changes the 'effective' priority of a task. It does
5203 * not touch ->normal_prio like __setscheduler().
5205 * Used by the rt_mutex code to implement priority inheritance logic.
5207 void rt_mutex_setprio(struct task_struct *p, int prio)
5209 int oldprio, on_rq, running;
5211 const struct sched_class *prev_class;
5213 BUG_ON(prio < 0 || prio > MAX_PRIO);
5215 rq = __task_rq_lock(p);
5217 trace_sched_pi_setprio(p, prio);
5219 prev_class = p->sched_class;
5221 running = task_current(rq, p);
5223 dequeue_task(rq, p, 0);
5225 p->sched_class->put_prev_task(rq, p);
5228 p->sched_class = &rt_sched_class;
5230 if (rt_prio(oldprio))
5232 p->sched_class = &fair_sched_class;
5238 p->sched_class->set_curr_task(rq);
5240 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
5242 check_class_changed(rq, p, prev_class, oldprio);
5243 __task_rq_unlock(rq);
5248 void set_user_nice(struct task_struct *p, long nice)
5250 int old_prio, delta, on_rq;
5251 unsigned long flags;
5254 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5257 * We have to be careful, if called from sys_setpriority(),
5258 * the task might be in the middle of scheduling on another CPU.
5260 rq = task_rq_lock(p, &flags);
5262 * The RT priorities are set via sched_setscheduler(), but we still
5263 * allow the 'normal' nice value to be set - but as expected
5264 * it wont have any effect on scheduling until the task is
5265 * SCHED_FIFO/SCHED_RR:
5267 if (task_has_rt_policy(p)) {
5268 p->static_prio = NICE_TO_PRIO(nice);
5273 dequeue_task(rq, p, 0);
5275 p->static_prio = NICE_TO_PRIO(nice);
5278 p->prio = effective_prio(p);
5279 delta = p->prio - old_prio;
5282 enqueue_task(rq, p, 0);
5284 * If the task increased its priority or is running and
5285 * lowered its priority, then reschedule its CPU:
5287 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5288 resched_task(rq->curr);
5291 task_rq_unlock(rq, p, &flags);
5293 EXPORT_SYMBOL(set_user_nice);
5296 * can_nice - check if a task can reduce its nice value
5300 int can_nice(const struct task_struct *p, const int nice)
5302 /* convert nice value [19,-20] to rlimit style value [1,40] */
5303 int nice_rlim = 20 - nice;
5305 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5306 capable(CAP_SYS_NICE));
5309 #ifdef __ARCH_WANT_SYS_NICE
5312 * sys_nice - change the priority of the current process.
5313 * @increment: priority increment
5315 * sys_setpriority is a more generic, but much slower function that
5316 * does similar things.
5318 SYSCALL_DEFINE1(nice, int, increment)
5323 * Setpriority might change our priority at the same moment.
5324 * We don't have to worry. Conceptually one call occurs first
5325 * and we have a single winner.
5327 if (increment < -40)
5332 nice = TASK_NICE(current) + increment;
5338 if (increment < 0 && !can_nice(current, nice))
5341 retval = security_task_setnice(current, nice);
5345 set_user_nice(current, nice);
5352 * task_prio - return the priority value of a given task.
5353 * @p: the task in question.
5355 * This is the priority value as seen by users in /proc.
5356 * RT tasks are offset by -200. Normal tasks are centered
5357 * around 0, value goes from -16 to +15.
5359 int task_prio(const struct task_struct *p)
5361 return p->prio - MAX_RT_PRIO;
5365 * task_nice - return the nice value of a given task.
5366 * @p: the task in question.
5368 int task_nice(const struct task_struct *p)
5370 return TASK_NICE(p);
5372 EXPORT_SYMBOL(task_nice);
5375 * idle_cpu - is a given cpu idle currently?
5376 * @cpu: the processor in question.
5378 int idle_cpu(int cpu)
5380 struct rq *rq = cpu_rq(cpu);
5382 if (rq->curr != rq->idle)
5389 if (!llist_empty(&rq->wake_list))
5397 * idle_task - return the idle task for a given cpu.
5398 * @cpu: the processor in question.
5400 struct task_struct *idle_task(int cpu)
5402 return cpu_rq(cpu)->idle;
5406 * find_process_by_pid - find a process with a matching PID value.
5407 * @pid: the pid in question.
5409 static struct task_struct *find_process_by_pid(pid_t pid)
5411 return pid ? find_task_by_vpid(pid) : current;
5414 /* Actually do priority change: must hold rq lock. */
5416 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5419 p->rt_priority = prio;
5420 p->normal_prio = normal_prio(p);
5421 /* we are holding p->pi_lock already */
5422 p->prio = rt_mutex_getprio(p);
5423 if (rt_prio(p->prio))
5424 p->sched_class = &rt_sched_class;
5426 p->sched_class = &fair_sched_class;
5431 * check the target process has a UID that matches the current process's
5433 static bool check_same_owner(struct task_struct *p)
5435 const struct cred *cred = current_cred(), *pcred;
5439 pcred = __task_cred(p);
5440 if (cred->user->user_ns == pcred->user->user_ns)
5441 match = (cred->euid == pcred->euid ||
5442 cred->euid == pcred->uid);
5449 static int __sched_setscheduler(struct task_struct *p, int policy,
5450 const struct sched_param *param, bool user)
5452 int retval, oldprio, oldpolicy = -1, on_rq, running;
5453 unsigned long flags;
5454 const struct sched_class *prev_class;
5458 /* may grab non-irq protected spin_locks */
5459 BUG_ON(in_interrupt());
5461 /* double check policy once rq lock held */
5463 reset_on_fork = p->sched_reset_on_fork;
5464 policy = oldpolicy = p->policy;
5466 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5467 policy &= ~SCHED_RESET_ON_FORK;
5469 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5470 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5471 policy != SCHED_IDLE)
5476 * Valid priorities for SCHED_FIFO and SCHED_RR are
5477 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5478 * SCHED_BATCH and SCHED_IDLE is 0.
5480 if (param->sched_priority < 0 ||
5481 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5482 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5484 if (rt_policy(policy) != (param->sched_priority != 0))
5488 * Allow unprivileged RT tasks to decrease priority:
5490 if (user && !capable(CAP_SYS_NICE)) {
5491 if (rt_policy(policy)) {
5492 unsigned long rlim_rtprio =
5493 task_rlimit(p, RLIMIT_RTPRIO);
5495 /* can't set/change the rt policy */
5496 if (policy != p->policy && !rlim_rtprio)
5499 /* can't increase priority */
5500 if (param->sched_priority > p->rt_priority &&
5501 param->sched_priority > rlim_rtprio)
5506 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5507 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5509 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5510 if (!can_nice(p, TASK_NICE(p)))
5514 /* can't change other user's priorities */
5515 if (!check_same_owner(p))
5518 /* Normal users shall not reset the sched_reset_on_fork flag */
5519 if (p->sched_reset_on_fork && !reset_on_fork)
5524 retval = security_task_setscheduler(p);
5530 * make sure no PI-waiters arrive (or leave) while we are
5531 * changing the priority of the task:
5533 * To be able to change p->policy safely, the appropriate
5534 * runqueue lock must be held.
5536 rq = task_rq_lock(p, &flags);
5539 * Changing the policy of the stop threads its a very bad idea
5541 if (p == rq->stop) {
5542 task_rq_unlock(rq, p, &flags);
5547 * If not changing anything there's no need to proceed further:
5549 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5550 param->sched_priority == p->rt_priority))) {
5552 __task_rq_unlock(rq);
5553 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5557 #ifdef CONFIG_RT_GROUP_SCHED
5560 * Do not allow realtime tasks into groups that have no runtime
5563 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5564 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5565 !task_group_is_autogroup(task_group(p))) {
5566 task_rq_unlock(rq, p, &flags);
5572 /* recheck policy now with rq lock held */
5573 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5574 policy = oldpolicy = -1;
5575 task_rq_unlock(rq, p, &flags);
5579 running = task_current(rq, p);
5581 deactivate_task(rq, p, 0);
5583 p->sched_class->put_prev_task(rq, p);
5585 p->sched_reset_on_fork = reset_on_fork;
5588 prev_class = p->sched_class;
5589 __setscheduler(rq, p, policy, param->sched_priority);
5592 p->sched_class->set_curr_task(rq);
5594 activate_task(rq, p, 0);
5596 check_class_changed(rq, p, prev_class, oldprio);
5597 task_rq_unlock(rq, p, &flags);
5599 rt_mutex_adjust_pi(p);
5605 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5606 * @p: the task in question.
5607 * @policy: new policy.
5608 * @param: structure containing the new RT priority.
5610 * NOTE that the task may be already dead.
5612 int sched_setscheduler(struct task_struct *p, int policy,
5613 const struct sched_param *param)
5615 return __sched_setscheduler(p, policy, param, true);
5617 EXPORT_SYMBOL_GPL(sched_setscheduler);
5620 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5621 * @p: the task in question.
5622 * @policy: new policy.
5623 * @param: structure containing the new RT priority.
5625 * Just like sched_setscheduler, only don't bother checking if the
5626 * current context has permission. For example, this is needed in
5627 * stop_machine(): we create temporary high priority worker threads,
5628 * but our caller might not have that capability.
5630 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5631 const struct sched_param *param)
5633 return __sched_setscheduler(p, policy, param, false);
5637 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5639 struct sched_param lparam;
5640 struct task_struct *p;
5643 if (!param || pid < 0)
5645 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5650 p = find_process_by_pid(pid);
5652 retval = sched_setscheduler(p, policy, &lparam);
5659 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5660 * @pid: the pid in question.
5661 * @policy: new policy.
5662 * @param: structure containing the new RT priority.
5664 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5665 struct sched_param __user *, param)
5667 /* negative values for policy are not valid */
5671 return do_sched_setscheduler(pid, policy, param);
5675 * sys_sched_setparam - set/change the RT priority of a thread
5676 * @pid: the pid in question.
5677 * @param: structure containing the new RT priority.
5679 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5681 return do_sched_setscheduler(pid, -1, param);
5685 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5686 * @pid: the pid in question.
5688 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5690 struct task_struct *p;
5698 p = find_process_by_pid(pid);
5700 retval = security_task_getscheduler(p);
5703 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5710 * sys_sched_getparam - get the RT priority of a thread
5711 * @pid: the pid in question.
5712 * @param: structure containing the RT priority.
5714 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5716 struct sched_param lp;
5717 struct task_struct *p;
5720 if (!param || pid < 0)
5724 p = find_process_by_pid(pid);
5729 retval = security_task_getscheduler(p);
5733 lp.sched_priority = p->rt_priority;
5737 * This one might sleep, we cannot do it with a spinlock held ...
5739 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5748 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5750 cpumask_var_t cpus_allowed, new_mask;
5751 struct task_struct *p;
5757 p = find_process_by_pid(pid);
5764 /* Prevent p going away */
5768 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5772 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5774 goto out_free_cpus_allowed;
5777 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5780 retval = security_task_setscheduler(p);
5784 cpuset_cpus_allowed(p, cpus_allowed);
5785 cpumask_and(new_mask, in_mask, cpus_allowed);
5787 retval = set_cpus_allowed_ptr(p, new_mask);
5790 cpuset_cpus_allowed(p, cpus_allowed);
5791 if (!cpumask_subset(new_mask, cpus_allowed)) {
5793 * We must have raced with a concurrent cpuset
5794 * update. Just reset the cpus_allowed to the
5795 * cpuset's cpus_allowed
5797 cpumask_copy(new_mask, cpus_allowed);
5802 free_cpumask_var(new_mask);
5803 out_free_cpus_allowed:
5804 free_cpumask_var(cpus_allowed);
5811 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5812 struct cpumask *new_mask)
5814 if (len < cpumask_size())
5815 cpumask_clear(new_mask);
5816 else if (len > cpumask_size())
5817 len = cpumask_size();
5819 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5823 * sys_sched_setaffinity - set the cpu affinity of a process
5824 * @pid: pid of the process
5825 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5826 * @user_mask_ptr: user-space pointer to the new cpu mask
5828 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5829 unsigned long __user *, user_mask_ptr)
5831 cpumask_var_t new_mask;
5834 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5837 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5839 retval = sched_setaffinity(pid, new_mask);
5840 free_cpumask_var(new_mask);
5844 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5846 struct task_struct *p;
5847 unsigned long flags;
5854 p = find_process_by_pid(pid);
5858 retval = security_task_getscheduler(p);
5862 raw_spin_lock_irqsave(&p->pi_lock, flags);
5863 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5864 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5874 * sys_sched_getaffinity - get the cpu affinity of a process
5875 * @pid: pid of the process
5876 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5877 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5879 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5880 unsigned long __user *, user_mask_ptr)
5885 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5887 if (len & (sizeof(unsigned long)-1))
5890 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5893 ret = sched_getaffinity(pid, mask);
5895 size_t retlen = min_t(size_t, len, cpumask_size());
5897 if (copy_to_user(user_mask_ptr, mask, retlen))
5902 free_cpumask_var(mask);
5908 * sys_sched_yield - yield the current processor to other threads.
5910 * This function yields the current CPU to other tasks. If there are no
5911 * other threads running on this CPU then this function will return.
5913 SYSCALL_DEFINE0(sched_yield)
5915 struct rq *rq = this_rq_lock();
5917 schedstat_inc(rq, yld_count);
5918 current->sched_class->yield_task(rq);
5921 * Since we are going to call schedule() anyway, there's
5922 * no need to preempt or enable interrupts:
5924 __release(rq->lock);
5925 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5926 do_raw_spin_unlock(&rq->lock);
5927 preempt_enable_no_resched();
5934 static inline int should_resched(void)
5936 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5939 static void __cond_resched(void)
5941 add_preempt_count(PREEMPT_ACTIVE);
5943 sub_preempt_count(PREEMPT_ACTIVE);
5946 int __sched _cond_resched(void)
5948 if (should_resched()) {
5954 EXPORT_SYMBOL(_cond_resched);
5957 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5958 * call schedule, and on return reacquire the lock.
5960 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5961 * operations here to prevent schedule() from being called twice (once via
5962 * spin_unlock(), once by hand).
5964 int __cond_resched_lock(spinlock_t *lock)
5966 int resched = should_resched();
5969 lockdep_assert_held(lock);
5971 if (spin_needbreak(lock) || resched) {
5982 EXPORT_SYMBOL(__cond_resched_lock);
5984 int __sched __cond_resched_softirq(void)
5986 BUG_ON(!in_softirq());
5988 if (should_resched()) {
5996 EXPORT_SYMBOL(__cond_resched_softirq);
5999 * yield - yield the current processor to other threads.
6001 * This is a shortcut for kernel-space yielding - it marks the
6002 * thread runnable and calls sys_sched_yield().
6004 void __sched yield(void)
6006 set_current_state(TASK_RUNNING);
6009 EXPORT_SYMBOL(yield);
6012 * yield_to - yield the current processor to another thread in
6013 * your thread group, or accelerate that thread toward the
6014 * processor it's on.
6016 * @preempt: whether task preemption is allowed or not
6018 * It's the caller's job to ensure that the target task struct
6019 * can't go away on us before we can do any checks.
6021 * Returns true if we indeed boosted the target task.
6023 bool __sched yield_to(struct task_struct *p, bool preempt)
6025 struct task_struct *curr = current;
6026 struct rq *rq, *p_rq;
6027 unsigned long flags;
6030 local_irq_save(flags);
6035 double_rq_lock(rq, p_rq);
6036 while (task_rq(p) != p_rq) {
6037 double_rq_unlock(rq, p_rq);
6041 if (!curr->sched_class->yield_to_task)
6044 if (curr->sched_class != p->sched_class)
6047 if (task_running(p_rq, p) || p->state)
6050 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
6052 schedstat_inc(rq, yld_count);
6054 * Make p's CPU reschedule; pick_next_entity takes care of
6057 if (preempt && rq != p_rq)
6058 resched_task(p_rq->curr);
6062 double_rq_unlock(rq, p_rq);
6063 local_irq_restore(flags);
6070 EXPORT_SYMBOL_GPL(yield_to);
6073 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6074 * that process accounting knows that this is a task in IO wait state.
6076 void __sched io_schedule(void)
6078 struct rq *rq = raw_rq();
6080 delayacct_blkio_start();
6081 atomic_inc(&rq->nr_iowait);
6082 blk_flush_plug(current);
6083 current->in_iowait = 1;
6085 current->in_iowait = 0;
6086 atomic_dec(&rq->nr_iowait);
6087 delayacct_blkio_end();
6089 EXPORT_SYMBOL(io_schedule);
6091 long __sched io_schedule_timeout(long timeout)
6093 struct rq *rq = raw_rq();
6096 delayacct_blkio_start();
6097 atomic_inc(&rq->nr_iowait);
6098 blk_flush_plug(current);
6099 current->in_iowait = 1;
6100 ret = schedule_timeout(timeout);
6101 current->in_iowait = 0;
6102 atomic_dec(&rq->nr_iowait);
6103 delayacct_blkio_end();
6108 * sys_sched_get_priority_max - return maximum RT priority.
6109 * @policy: scheduling class.
6111 * this syscall returns the maximum rt_priority that can be used
6112 * by a given scheduling class.
6114 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6121 ret = MAX_USER_RT_PRIO-1;
6133 * sys_sched_get_priority_min - return minimum RT priority.
6134 * @policy: scheduling class.
6136 * this syscall returns the minimum rt_priority that can be used
6137 * by a given scheduling class.
6139 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6157 * sys_sched_rr_get_interval - return the default timeslice of a process.
6158 * @pid: pid of the process.
6159 * @interval: userspace pointer to the timeslice value.
6161 * this syscall writes the default timeslice value of a given process
6162 * into the user-space timespec buffer. A value of '0' means infinity.
6164 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6165 struct timespec __user *, interval)
6167 struct task_struct *p;
6168 unsigned int time_slice;
6169 unsigned long flags;
6179 p = find_process_by_pid(pid);
6183 retval = security_task_getscheduler(p);
6187 rq = task_rq_lock(p, &flags);
6188 time_slice = p->sched_class->get_rr_interval(rq, p);
6189 task_rq_unlock(rq, p, &flags);
6192 jiffies_to_timespec(time_slice, &t);
6193 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6201 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6203 void sched_show_task(struct task_struct *p)
6205 unsigned long free = 0;
6208 state = p->state ? __ffs(p->state) + 1 : 0;
6209 printk(KERN_INFO "%-15.15s %c", p->comm,
6210 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6211 #if BITS_PER_LONG == 32
6212 if (state == TASK_RUNNING)
6213 printk(KERN_CONT " running ");
6215 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6217 if (state == TASK_RUNNING)
6218 printk(KERN_CONT " running task ");
6220 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6222 #ifdef CONFIG_DEBUG_STACK_USAGE
6223 free = stack_not_used(p);
6225 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6226 task_pid_nr(p), task_pid_nr(p->real_parent),
6227 (unsigned long)task_thread_info(p)->flags);
6229 show_stack(p, NULL);
6232 void show_state_filter(unsigned long state_filter)
6234 struct task_struct *g, *p;
6236 #if BITS_PER_LONG == 32
6238 " task PC stack pid father\n");
6241 " task PC stack pid father\n");
6244 do_each_thread(g, p) {
6246 * reset the NMI-timeout, listing all files on a slow
6247 * console might take a lot of time:
6249 touch_nmi_watchdog();
6250 if (!state_filter || (p->state & state_filter))
6252 } while_each_thread(g, p);
6254 touch_all_softlockup_watchdogs();
6256 #ifdef CONFIG_SCHED_DEBUG
6257 sysrq_sched_debug_show();
6261 * Only show locks if all tasks are dumped:
6264 debug_show_all_locks();
6267 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6269 idle->sched_class = &idle_sched_class;
6273 * init_idle - set up an idle thread for a given CPU
6274 * @idle: task in question
6275 * @cpu: cpu the idle task belongs to
6277 * NOTE: this function does not set the idle thread's NEED_RESCHED
6278 * flag, to make booting more robust.
6280 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6282 struct rq *rq = cpu_rq(cpu);
6283 unsigned long flags;
6285 raw_spin_lock_irqsave(&rq->lock, flags);
6288 idle->state = TASK_RUNNING;
6289 idle->se.exec_start = sched_clock();
6291 do_set_cpus_allowed(idle, cpumask_of(cpu));
6293 * We're having a chicken and egg problem, even though we are
6294 * holding rq->lock, the cpu isn't yet set to this cpu so the
6295 * lockdep check in task_group() will fail.
6297 * Similar case to sched_fork(). / Alternatively we could
6298 * use task_rq_lock() here and obtain the other rq->lock.
6303 __set_task_cpu(idle, cpu);
6306 rq->curr = rq->idle = idle;
6307 #if defined(CONFIG_SMP)
6310 raw_spin_unlock_irqrestore(&rq->lock, flags);
6312 /* Set the preempt count _outside_ the spinlocks! */
6313 task_thread_info(idle)->preempt_count = 0;
6316 * The idle tasks have their own, simple scheduling class:
6318 idle->sched_class = &idle_sched_class;
6319 ftrace_graph_init_idle_task(idle, cpu);
6320 #if defined(CONFIG_SMP)
6321 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6326 * Increase the granularity value when there are more CPUs,
6327 * because with more CPUs the 'effective latency' as visible
6328 * to users decreases. But the relationship is not linear,
6329 * so pick a second-best guess by going with the log2 of the
6332 * This idea comes from the SD scheduler of Con Kolivas:
6334 static int get_update_sysctl_factor(void)
6336 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6337 unsigned int factor;
6339 switch (sysctl_sched_tunable_scaling) {
6340 case SCHED_TUNABLESCALING_NONE:
6343 case SCHED_TUNABLESCALING_LINEAR:
6346 case SCHED_TUNABLESCALING_LOG:
6348 factor = 1 + ilog2(cpus);
6355 static void update_sysctl(void)
6357 unsigned int factor = get_update_sysctl_factor();
6359 #define SET_SYSCTL(name) \
6360 (sysctl_##name = (factor) * normalized_sysctl_##name)
6361 SET_SYSCTL(sched_min_granularity);
6362 SET_SYSCTL(sched_latency);
6363 SET_SYSCTL(sched_wakeup_granularity);
6367 static inline void sched_init_granularity(void)
6373 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6375 if (p->sched_class && p->sched_class->set_cpus_allowed)
6376 p->sched_class->set_cpus_allowed(p, new_mask);
6378 cpumask_copy(&p->cpus_allowed, new_mask);
6379 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6383 * This is how migration works:
6385 * 1) we invoke migration_cpu_stop() on the target CPU using
6387 * 2) stopper starts to run (implicitly forcing the migrated thread
6389 * 3) it checks whether the migrated task is still in the wrong runqueue.
6390 * 4) if it's in the wrong runqueue then the migration thread removes
6391 * it and puts it into the right queue.
6392 * 5) stopper completes and stop_one_cpu() returns and the migration
6397 * Change a given task's CPU affinity. Migrate the thread to a
6398 * proper CPU and schedule it away if the CPU it's executing on
6399 * is removed from the allowed bitmask.
6401 * NOTE: the caller must have a valid reference to the task, the
6402 * task must not exit() & deallocate itself prematurely. The
6403 * call is not atomic; no spinlocks may be held.
6405 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6407 unsigned long flags;
6409 unsigned int dest_cpu;
6412 rq = task_rq_lock(p, &flags);
6414 if (cpumask_equal(&p->cpus_allowed, new_mask))
6417 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6422 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6427 do_set_cpus_allowed(p, new_mask);
6429 /* Can the task run on the task's current CPU? If so, we're done */
6430 if (cpumask_test_cpu(task_cpu(p), new_mask))
6433 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6435 struct migration_arg arg = { p, dest_cpu };
6436 /* Need help from migration thread: drop lock and wait. */
6437 task_rq_unlock(rq, p, &flags);
6438 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6439 tlb_migrate_finish(p->mm);
6443 task_rq_unlock(rq, p, &flags);
6447 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6450 * Move (not current) task off this cpu, onto dest cpu. We're doing
6451 * this because either it can't run here any more (set_cpus_allowed()
6452 * away from this CPU, or CPU going down), or because we're
6453 * attempting to rebalance this task on exec (sched_exec).
6455 * So we race with normal scheduler movements, but that's OK, as long
6456 * as the task is no longer on this CPU.
6458 * Returns non-zero if task was successfully migrated.
6460 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6462 struct rq *rq_dest, *rq_src;
6465 if (unlikely(!cpu_active(dest_cpu)))
6468 rq_src = cpu_rq(src_cpu);
6469 rq_dest = cpu_rq(dest_cpu);
6471 raw_spin_lock(&p->pi_lock);
6472 double_rq_lock(rq_src, rq_dest);
6473 /* Already moved. */
6474 if (task_cpu(p) != src_cpu)
6476 /* Affinity changed (again). */
6477 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
6481 * If we're not on a rq, the next wake-up will ensure we're
6485 deactivate_task(rq_src, p, 0);
6486 set_task_cpu(p, dest_cpu);
6487 activate_task(rq_dest, p, 0);
6488 check_preempt_curr(rq_dest, p, 0);
6493 double_rq_unlock(rq_src, rq_dest);
6494 raw_spin_unlock(&p->pi_lock);
6499 * migration_cpu_stop - this will be executed by a highprio stopper thread
6500 * and performs thread migration by bumping thread off CPU then
6501 * 'pushing' onto another runqueue.
6503 static int migration_cpu_stop(void *data)
6505 struct migration_arg *arg = data;
6508 * The original target cpu might have gone down and we might
6509 * be on another cpu but it doesn't matter.
6511 local_irq_disable();
6512 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6517 #ifdef CONFIG_HOTPLUG_CPU
6520 * Ensures that the idle task is using init_mm right before its cpu goes
6523 void idle_task_exit(void)
6525 struct mm_struct *mm = current->active_mm;
6527 BUG_ON(cpu_online(smp_processor_id()));
6530 switch_mm(mm, &init_mm, current);
6535 * While a dead CPU has no uninterruptible tasks queued at this point,
6536 * it might still have a nonzero ->nr_uninterruptible counter, because
6537 * for performance reasons the counter is not stricly tracking tasks to
6538 * their home CPUs. So we just add the counter to another CPU's counter,
6539 * to keep the global sum constant after CPU-down:
6541 static void migrate_nr_uninterruptible(struct rq *rq_src)
6543 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6545 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6546 rq_src->nr_uninterruptible = 0;
6550 * remove the tasks which were accounted by rq from calc_load_tasks.
6552 static void calc_global_load_remove(struct rq *rq)
6554 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6555 rq->calc_load_active = 0;
6558 #ifdef CONFIG_CFS_BANDWIDTH
6559 static void unthrottle_offline_cfs_rqs(struct rq *rq)
6561 struct cfs_rq *cfs_rq;
6563 for_each_leaf_cfs_rq(rq, cfs_rq) {
6564 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6566 if (!cfs_rq->runtime_enabled)
6570 * clock_task is not advancing so we just need to make sure
6571 * there's some valid quota amount
6573 cfs_rq->runtime_remaining = cfs_b->quota;
6574 if (cfs_rq_throttled(cfs_rq))
6575 unthrottle_cfs_rq(cfs_rq);
6581 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6582 * try_to_wake_up()->select_task_rq().
6584 * Called with rq->lock held even though we'er in stop_machine() and
6585 * there's no concurrency possible, we hold the required locks anyway
6586 * because of lock validation efforts.
6588 static void migrate_tasks(unsigned int dead_cpu)
6590 struct rq *rq = cpu_rq(dead_cpu);
6591 struct task_struct *next, *stop = rq->stop;
6595 * Fudge the rq selection such that the below task selection loop
6596 * doesn't get stuck on the currently eligible stop task.
6598 * We're currently inside stop_machine() and the rq is either stuck
6599 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6600 * either way we should never end up calling schedule() until we're
6607 * There's this thread running, bail when that's the only
6610 if (rq->nr_running == 1)
6613 next = pick_next_task(rq);
6615 next->sched_class->put_prev_task(rq, next);
6617 /* Find suitable destination for @next, with force if needed. */
6618 dest_cpu = select_fallback_rq(dead_cpu, next);
6619 raw_spin_unlock(&rq->lock);
6621 __migrate_task(next, dead_cpu, dest_cpu);
6623 raw_spin_lock(&rq->lock);
6629 #endif /* CONFIG_HOTPLUG_CPU */
6631 #if !defined(CONFIG_HOTPLUG_CPU) || !defined(CONFIG_CFS_BANDWIDTH)
6632 static void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6635 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6637 static struct ctl_table sd_ctl_dir[] = {
6639 .procname = "sched_domain",
6645 static struct ctl_table sd_ctl_root[] = {
6647 .procname = "kernel",
6649 .child = sd_ctl_dir,
6654 static struct ctl_table *sd_alloc_ctl_entry(int n)
6656 struct ctl_table *entry =
6657 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6662 static void sd_free_ctl_entry(struct ctl_table **tablep)
6664 struct ctl_table *entry;
6667 * In the intermediate directories, both the child directory and
6668 * procname are dynamically allocated and could fail but the mode
6669 * will always be set. In the lowest directory the names are
6670 * static strings and all have proc handlers.
6672 for (entry = *tablep; entry->mode; entry++) {
6674 sd_free_ctl_entry(&entry->child);
6675 if (entry->proc_handler == NULL)
6676 kfree(entry->procname);
6683 static int min_load_idx = 0;
6684 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
6687 set_table_entry(struct ctl_table *entry,
6688 const char *procname, void *data, int maxlen,
6689 mode_t mode, proc_handler *proc_handler,
6692 entry->procname = procname;
6694 entry->maxlen = maxlen;
6696 entry->proc_handler = proc_handler;
6699 entry->extra1 = &min_load_idx;
6700 entry->extra2 = &max_load_idx;
6704 static struct ctl_table *
6705 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6707 struct ctl_table *table = sd_alloc_ctl_entry(13);
6712 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6713 sizeof(long), 0644, proc_doulongvec_minmax, false);
6714 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6715 sizeof(long), 0644, proc_doulongvec_minmax, false);
6716 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6717 sizeof(int), 0644, proc_dointvec_minmax, true);
6718 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6719 sizeof(int), 0644, proc_dointvec_minmax, true);
6720 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6721 sizeof(int), 0644, proc_dointvec_minmax, true);
6722 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6723 sizeof(int), 0644, proc_dointvec_minmax, true);
6724 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6725 sizeof(int), 0644, proc_dointvec_minmax, true);
6726 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6727 sizeof(int), 0644, proc_dointvec_minmax, false);
6728 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6729 sizeof(int), 0644, proc_dointvec_minmax, false);
6730 set_table_entry(&table[9], "cache_nice_tries",
6731 &sd->cache_nice_tries,
6732 sizeof(int), 0644, proc_dointvec_minmax, false);
6733 set_table_entry(&table[10], "flags", &sd->flags,
6734 sizeof(int), 0644, proc_dointvec_minmax, false);
6735 set_table_entry(&table[11], "name", sd->name,
6736 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
6737 /* &table[12] is terminator */
6742 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6744 struct ctl_table *entry, *table;
6745 struct sched_domain *sd;
6746 int domain_num = 0, i;
6749 for_each_domain(cpu, sd)
6751 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6756 for_each_domain(cpu, sd) {
6757 snprintf(buf, 32, "domain%d", i);
6758 entry->procname = kstrdup(buf, GFP_KERNEL);
6760 entry->child = sd_alloc_ctl_domain_table(sd);
6767 static struct ctl_table_header *sd_sysctl_header;
6768 static void register_sched_domain_sysctl(void)
6770 int i, cpu_num = num_possible_cpus();
6771 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6774 WARN_ON(sd_ctl_dir[0].child);
6775 sd_ctl_dir[0].child = entry;
6780 for_each_possible_cpu(i) {
6781 snprintf(buf, 32, "cpu%d", i);
6782 entry->procname = kstrdup(buf, GFP_KERNEL);
6784 entry->child = sd_alloc_ctl_cpu_table(i);
6788 WARN_ON(sd_sysctl_header);
6789 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6792 /* may be called multiple times per register */
6793 static void unregister_sched_domain_sysctl(void)
6795 if (sd_sysctl_header)
6796 unregister_sysctl_table(sd_sysctl_header);
6797 sd_sysctl_header = NULL;
6798 if (sd_ctl_dir[0].child)
6799 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6802 static void register_sched_domain_sysctl(void)
6805 static void unregister_sched_domain_sysctl(void)
6810 static void set_rq_online(struct rq *rq)
6813 const struct sched_class *class;
6815 cpumask_set_cpu(rq->cpu, rq->rd->online);
6818 for_each_class(class) {
6819 if (class->rq_online)
6820 class->rq_online(rq);
6825 static void set_rq_offline(struct rq *rq)
6828 const struct sched_class *class;
6830 for_each_class(class) {
6831 if (class->rq_offline)
6832 class->rq_offline(rq);
6835 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6841 * migration_call - callback that gets triggered when a CPU is added.
6842 * Here we can start up the necessary migration thread for the new CPU.
6844 static int __cpuinit
6845 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6847 int cpu = (long)hcpu;
6848 unsigned long flags;
6849 struct rq *rq = cpu_rq(cpu);
6851 switch (action & ~CPU_TASKS_FROZEN) {
6853 case CPU_UP_PREPARE:
6854 rq->calc_load_update = calc_load_update;
6858 /* Update our root-domain */
6859 raw_spin_lock_irqsave(&rq->lock, flags);
6861 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6865 raw_spin_unlock_irqrestore(&rq->lock, flags);
6868 #ifdef CONFIG_HOTPLUG_CPU
6870 sched_ttwu_pending();
6871 /* Update our root-domain */
6872 raw_spin_lock_irqsave(&rq->lock, flags);
6874 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6878 BUG_ON(rq->nr_running != 1); /* the migration thread */
6879 raw_spin_unlock_irqrestore(&rq->lock, flags);
6881 migrate_nr_uninterruptible(rq);
6882 calc_global_load_remove(rq);
6887 update_max_interval();
6893 * Register at high priority so that task migration (migrate_all_tasks)
6894 * happens before everything else. This has to be lower priority than
6895 * the notifier in the perf_event subsystem, though.
6897 static struct notifier_block __cpuinitdata migration_notifier = {
6898 .notifier_call = migration_call,
6899 .priority = CPU_PRI_MIGRATION,
6902 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6903 unsigned long action, void *hcpu)
6905 switch (action & ~CPU_TASKS_FROZEN) {
6907 case CPU_DOWN_FAILED:
6908 set_cpu_active((long)hcpu, true);
6915 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6916 unsigned long action, void *hcpu)
6918 switch (action & ~CPU_TASKS_FROZEN) {
6919 case CPU_DOWN_PREPARE:
6920 set_cpu_active((long)hcpu, false);
6927 static int __init migration_init(void)
6929 void *cpu = (void *)(long)smp_processor_id();
6932 /* Initialize migration for the boot CPU */
6933 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6934 BUG_ON(err == NOTIFY_BAD);
6935 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6936 register_cpu_notifier(&migration_notifier);
6938 /* Register cpu active notifiers */
6939 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6940 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6944 early_initcall(migration_init);
6949 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6951 #ifdef CONFIG_SCHED_DEBUG
6953 static __read_mostly int sched_domain_debug_enabled;
6955 static int __init sched_domain_debug_setup(char *str)
6957 sched_domain_debug_enabled = 1;
6961 early_param("sched_debug", sched_domain_debug_setup);
6963 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6964 struct cpumask *groupmask)
6966 struct sched_group *group = sd->groups;
6969 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6970 cpumask_clear(groupmask);
6972 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6974 if (!(sd->flags & SD_LOAD_BALANCE)) {
6975 printk("does not load-balance\n");
6977 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6982 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6984 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6985 printk(KERN_ERR "ERROR: domain->span does not contain "
6988 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6989 printk(KERN_ERR "ERROR: domain->groups does not contain"
6993 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6997 printk(KERN_ERR "ERROR: group is NULL\n");
7001 if (!group->sgp->power) {
7002 printk(KERN_CONT "\n");
7003 printk(KERN_ERR "ERROR: domain->cpu_power not "
7008 if (!cpumask_weight(sched_group_cpus(group))) {
7009 printk(KERN_CONT "\n");
7010 printk(KERN_ERR "ERROR: empty group\n");
7014 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7015 printk(KERN_CONT "\n");
7016 printk(KERN_ERR "ERROR: repeated CPUs\n");
7020 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7022 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7024 printk(KERN_CONT " %s", str);
7025 if (group->sgp->power != SCHED_POWER_SCALE) {
7026 printk(KERN_CONT " (cpu_power = %d)",
7030 group = group->next;
7031 } while (group != sd->groups);
7032 printk(KERN_CONT "\n");
7034 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7035 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7038 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7039 printk(KERN_ERR "ERROR: parent span is not a superset "
7040 "of domain->span\n");
7044 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7048 if (!sched_domain_debug_enabled)
7052 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7056 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7059 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
7067 #else /* !CONFIG_SCHED_DEBUG */
7068 # define sched_domain_debug(sd, cpu) do { } while (0)
7069 #endif /* CONFIG_SCHED_DEBUG */
7071 static int sd_degenerate(struct sched_domain *sd)
7073 if (cpumask_weight(sched_domain_span(sd)) == 1)
7076 /* Following flags need at least 2 groups */
7077 if (sd->flags & (SD_LOAD_BALANCE |
7078 SD_BALANCE_NEWIDLE |
7082 SD_SHARE_PKG_RESOURCES)) {
7083 if (sd->groups != sd->groups->next)
7087 /* Following flags don't use groups */
7088 if (sd->flags & (SD_WAKE_AFFINE))
7095 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7097 unsigned long cflags = sd->flags, pflags = parent->flags;
7099 if (sd_degenerate(parent))
7102 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7105 /* Flags needing groups don't count if only 1 group in parent */
7106 if (parent->groups == parent->groups->next) {
7107 pflags &= ~(SD_LOAD_BALANCE |
7108 SD_BALANCE_NEWIDLE |
7112 SD_SHARE_PKG_RESOURCES);
7113 if (nr_node_ids == 1)
7114 pflags &= ~SD_SERIALIZE;
7116 if (~cflags & pflags)
7122 static void free_rootdomain(struct rcu_head *rcu)
7124 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
7126 cpupri_cleanup(&rd->cpupri);
7127 free_cpumask_var(rd->rto_mask);
7128 free_cpumask_var(rd->online);
7129 free_cpumask_var(rd->span);
7133 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7135 struct root_domain *old_rd = NULL;
7136 unsigned long flags;
7138 raw_spin_lock_irqsave(&rq->lock, flags);
7143 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7146 cpumask_clear_cpu(rq->cpu, old_rd->span);
7149 * If we dont want to free the old_rt yet then
7150 * set old_rd to NULL to skip the freeing later
7153 if (!atomic_dec_and_test(&old_rd->refcount))
7157 atomic_inc(&rd->refcount);
7160 cpumask_set_cpu(rq->cpu, rd->span);
7161 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7164 raw_spin_unlock_irqrestore(&rq->lock, flags);
7167 call_rcu_sched(&old_rd->rcu, free_rootdomain);
7170 static int init_rootdomain(struct root_domain *rd)
7172 memset(rd, 0, sizeof(*rd));
7174 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7176 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7178 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7181 if (cpupri_init(&rd->cpupri) != 0)
7186 free_cpumask_var(rd->rto_mask);
7188 free_cpumask_var(rd->online);
7190 free_cpumask_var(rd->span);
7195 static void init_defrootdomain(void)
7197 init_rootdomain(&def_root_domain);
7199 atomic_set(&def_root_domain.refcount, 1);
7202 static struct root_domain *alloc_rootdomain(void)
7204 struct root_domain *rd;
7206 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7210 if (init_rootdomain(rd) != 0) {
7218 static void free_sched_groups(struct sched_group *sg, int free_sgp)
7220 struct sched_group *tmp, *first;
7229 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
7234 } while (sg != first);
7237 static void free_sched_domain(struct rcu_head *rcu)
7239 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
7242 * If its an overlapping domain it has private groups, iterate and
7245 if (sd->flags & SD_OVERLAP) {
7246 free_sched_groups(sd->groups, 1);
7247 } else if (atomic_dec_and_test(&sd->groups->ref)) {
7248 kfree(sd->groups->sgp);
7254 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
7256 call_rcu(&sd->rcu, free_sched_domain);
7259 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
7261 for (; sd; sd = sd->parent)
7262 destroy_sched_domain(sd, cpu);
7266 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7267 * hold the hotplug lock.
7270 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7272 struct rq *rq = cpu_rq(cpu);
7273 struct sched_domain *tmp;
7275 /* Remove the sched domains which do not contribute to scheduling. */
7276 for (tmp = sd; tmp; ) {
7277 struct sched_domain *parent = tmp->parent;
7281 if (sd_parent_degenerate(tmp, parent)) {
7282 tmp->parent = parent->parent;
7284 parent->parent->child = tmp;
7285 destroy_sched_domain(parent, cpu);
7290 if (sd && sd_degenerate(sd)) {
7293 destroy_sched_domain(tmp, cpu);
7298 sched_domain_debug(sd, cpu);
7300 rq_attach_root(rq, rd);
7302 rcu_assign_pointer(rq->sd, sd);
7303 destroy_sched_domains(tmp, cpu);
7306 /* cpus with isolated domains */
7307 static cpumask_var_t cpu_isolated_map;
7309 /* Setup the mask of cpus configured for isolated domains */
7310 static int __init isolated_cpu_setup(char *str)
7312 alloc_bootmem_cpumask_var(&cpu_isolated_map);
7313 cpulist_parse(str, cpu_isolated_map);
7317 __setup("isolcpus=", isolated_cpu_setup);
7322 * find_next_best_node - find the next node to include in a sched_domain
7323 * @node: node whose sched_domain we're building
7324 * @used_nodes: nodes already in the sched_domain
7326 * Find the next node to include in a given scheduling domain. Simply
7327 * finds the closest node not already in the @used_nodes map.
7329 * Should use nodemask_t.
7331 static int find_next_best_node(int node, nodemask_t *used_nodes)
7333 int i, n, val, min_val, best_node = -1;
7337 for (i = 0; i < nr_node_ids; i++) {
7338 /* Start at @node */
7339 n = (node + i) % nr_node_ids;
7341 if (!nr_cpus_node(n))
7344 /* Skip already used nodes */
7345 if (node_isset(n, *used_nodes))
7348 /* Simple min distance search */
7349 val = node_distance(node, n);
7351 if (val < min_val) {
7357 if (best_node != -1)
7358 node_set(best_node, *used_nodes);
7363 * sched_domain_node_span - get a cpumask for a node's sched_domain
7364 * @node: node whose cpumask we're constructing
7365 * @span: resulting cpumask
7367 * Given a node, construct a good cpumask for its sched_domain to span. It
7368 * should be one that prevents unnecessary balancing, but also spreads tasks
7371 static void sched_domain_node_span(int node, struct cpumask *span)
7373 nodemask_t used_nodes;
7376 cpumask_clear(span);
7377 nodes_clear(used_nodes);
7379 cpumask_or(span, span, cpumask_of_node(node));
7380 node_set(node, used_nodes);
7382 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7383 int next_node = find_next_best_node(node, &used_nodes);
7386 cpumask_or(span, span, cpumask_of_node(next_node));
7390 static const struct cpumask *cpu_node_mask(int cpu)
7392 lockdep_assert_held(&sched_domains_mutex);
7394 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7396 return sched_domains_tmpmask;
7399 static const struct cpumask *cpu_allnodes_mask(int cpu)
7401 return cpu_possible_mask;
7403 #endif /* CONFIG_NUMA */
7405 static const struct cpumask *cpu_cpu_mask(int cpu)
7407 return cpumask_of_node(cpu_to_node(cpu));
7410 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7413 struct sched_domain **__percpu sd;
7414 struct sched_group **__percpu sg;
7415 struct sched_group_power **__percpu sgp;
7419 struct sched_domain ** __percpu sd;
7420 struct root_domain *rd;
7430 struct sched_domain_topology_level;
7432 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7433 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7435 #define SDTL_OVERLAP 0x01
7437 struct sched_domain_topology_level {
7438 sched_domain_init_f init;
7439 sched_domain_mask_f mask;
7441 struct sd_data data;
7445 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7447 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7448 const struct cpumask *span = sched_domain_span(sd);
7449 struct cpumask *covered = sched_domains_tmpmask;
7450 struct sd_data *sdd = sd->private;
7451 struct sched_domain *child;
7454 cpumask_clear(covered);
7456 for_each_cpu(i, span) {
7457 struct cpumask *sg_span;
7459 if (cpumask_test_cpu(i, covered))
7462 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7463 GFP_KERNEL, cpu_to_node(i));
7468 sg_span = sched_group_cpus(sg);
7470 child = *per_cpu_ptr(sdd->sd, i);
7472 child = child->child;
7473 cpumask_copy(sg_span, sched_domain_span(child));
7475 cpumask_set_cpu(i, sg_span);
7477 cpumask_or(covered, covered, sg_span);
7479 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7480 atomic_inc(&sg->sgp->ref);
7482 if (cpumask_test_cpu(cpu, sg_span))
7492 sd->groups = groups;
7497 free_sched_groups(first, 0);
7502 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7504 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7505 struct sched_domain *child = sd->child;
7508 cpu = cpumask_first(sched_domain_span(child));
7511 *sg = *per_cpu_ptr(sdd->sg, cpu);
7512 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7513 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7520 * build_sched_groups will build a circular linked list of the groups
7521 * covered by the given span, and will set each group's ->cpumask correctly,
7522 * and ->cpu_power to 0.
7524 * Assumes the sched_domain tree is fully constructed
7527 build_sched_groups(struct sched_domain *sd, int cpu)
7529 struct sched_group *first = NULL, *last = NULL;
7530 struct sd_data *sdd = sd->private;
7531 const struct cpumask *span = sched_domain_span(sd);
7532 struct cpumask *covered;
7535 get_group(cpu, sdd, &sd->groups);
7536 atomic_inc(&sd->groups->ref);
7538 if (cpu != cpumask_first(sched_domain_span(sd)))
7541 lockdep_assert_held(&sched_domains_mutex);
7542 covered = sched_domains_tmpmask;
7544 cpumask_clear(covered);
7546 for_each_cpu(i, span) {
7547 struct sched_group *sg;
7548 int group = get_group(i, sdd, &sg);
7551 if (cpumask_test_cpu(i, covered))
7554 cpumask_clear(sched_group_cpus(sg));
7557 for_each_cpu(j, span) {
7558 if (get_group(j, sdd, NULL) != group)
7561 cpumask_set_cpu(j, covered);
7562 cpumask_set_cpu(j, sched_group_cpus(sg));
7577 * Initialize sched groups cpu_power.
7579 * cpu_power indicates the capacity of sched group, which is used while
7580 * distributing the load between different sched groups in a sched domain.
7581 * Typically cpu_power for all the groups in a sched domain will be same unless
7582 * there are asymmetries in the topology. If there are asymmetries, group
7583 * having more cpu_power will pickup more load compared to the group having
7586 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7588 struct sched_group *sg = sd->groups;
7590 WARN_ON(!sd || !sg);
7593 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7595 } while (sg != sd->groups);
7597 if (cpu != group_first_cpu(sg))
7600 update_group_power(sd, cpu);
7604 * Initializers for schedule domains
7605 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7608 #ifdef CONFIG_SCHED_DEBUG
7609 # define SD_INIT_NAME(sd, type) sd->name = #type
7611 # define SD_INIT_NAME(sd, type) do { } while (0)
7614 #define SD_INIT_FUNC(type) \
7615 static noinline struct sched_domain * \
7616 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7618 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7619 *sd = SD_##type##_INIT; \
7620 SD_INIT_NAME(sd, type); \
7621 sd->private = &tl->data; \
7627 SD_INIT_FUNC(ALLNODES)
7630 #ifdef CONFIG_SCHED_SMT
7631 SD_INIT_FUNC(SIBLING)
7633 #ifdef CONFIG_SCHED_MC
7636 #ifdef CONFIG_SCHED_BOOK
7640 static int default_relax_domain_level = -1;
7641 int sched_domain_level_max;
7643 static int __init setup_relax_domain_level(char *str)
7645 if (kstrtoint(str, 0, &default_relax_domain_level))
7646 pr_warn("Unable to set relax_domain_level\n");
7650 __setup("relax_domain_level=", setup_relax_domain_level);
7652 static void set_domain_attribute(struct sched_domain *sd,
7653 struct sched_domain_attr *attr)
7657 if (!attr || attr->relax_domain_level < 0) {
7658 if (default_relax_domain_level < 0)
7661 request = default_relax_domain_level;
7663 request = attr->relax_domain_level;
7664 if (request < sd->level) {
7665 /* turn off idle balance on this domain */
7666 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7668 /* turn on idle balance on this domain */
7669 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7673 static void __sdt_free(const struct cpumask *cpu_map);
7674 static int __sdt_alloc(const struct cpumask *cpu_map);
7676 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7677 const struct cpumask *cpu_map)
7681 if (!atomic_read(&d->rd->refcount))
7682 free_rootdomain(&d->rd->rcu); /* fall through */
7684 free_percpu(d->sd); /* fall through */
7686 __sdt_free(cpu_map); /* fall through */
7692 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7693 const struct cpumask *cpu_map)
7695 memset(d, 0, sizeof(*d));
7697 if (__sdt_alloc(cpu_map))
7698 return sa_sd_storage;
7699 d->sd = alloc_percpu(struct sched_domain *);
7701 return sa_sd_storage;
7702 d->rd = alloc_rootdomain();
7705 return sa_rootdomain;
7709 * NULL the sd_data elements we've used to build the sched_domain and
7710 * sched_group structure so that the subsequent __free_domain_allocs()
7711 * will not free the data we're using.
7713 static void claim_allocations(int cpu, struct sched_domain *sd)
7715 struct sd_data *sdd = sd->private;
7717 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7718 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7720 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7721 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7723 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7724 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7727 #ifdef CONFIG_SCHED_SMT
7728 static const struct cpumask *cpu_smt_mask(int cpu)
7730 return topology_thread_cpumask(cpu);
7735 * Topology list, bottom-up.
7737 static struct sched_domain_topology_level default_topology[] = {
7738 #ifdef CONFIG_SCHED_SMT
7739 { sd_init_SIBLING, cpu_smt_mask, },
7741 #ifdef CONFIG_SCHED_MC
7742 { sd_init_MC, cpu_coregroup_mask, },
7744 #ifdef CONFIG_SCHED_BOOK
7745 { sd_init_BOOK, cpu_book_mask, },
7747 { sd_init_CPU, cpu_cpu_mask, },
7749 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7750 { sd_init_ALLNODES, cpu_allnodes_mask, },
7755 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7757 static int __sdt_alloc(const struct cpumask *cpu_map)
7759 struct sched_domain_topology_level *tl;
7762 for (tl = sched_domain_topology; tl->init; tl++) {
7763 struct sd_data *sdd = &tl->data;
7765 sdd->sd = alloc_percpu(struct sched_domain *);
7769 sdd->sg = alloc_percpu(struct sched_group *);
7773 sdd->sgp = alloc_percpu(struct sched_group_power *);
7777 for_each_cpu(j, cpu_map) {
7778 struct sched_domain *sd;
7779 struct sched_group *sg;
7780 struct sched_group_power *sgp;
7782 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7783 GFP_KERNEL, cpu_to_node(j));
7787 *per_cpu_ptr(sdd->sd, j) = sd;
7789 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7790 GFP_KERNEL, cpu_to_node(j));
7794 *per_cpu_ptr(sdd->sg, j) = sg;
7796 sgp = kzalloc_node(sizeof(struct sched_group_power),
7797 GFP_KERNEL, cpu_to_node(j));
7801 *per_cpu_ptr(sdd->sgp, j) = sgp;
7808 static void __sdt_free(const struct cpumask *cpu_map)
7810 struct sched_domain_topology_level *tl;
7813 for (tl = sched_domain_topology; tl->init; tl++) {
7814 struct sd_data *sdd = &tl->data;
7816 for_each_cpu(j, cpu_map) {
7817 struct sched_domain *sd;
7820 sd = *per_cpu_ptr(sdd->sd, j);
7821 if (sd && (sd->flags & SD_OVERLAP))
7822 free_sched_groups(sd->groups, 0);
7823 kfree(*per_cpu_ptr(sdd->sd, j));
7827 kfree(*per_cpu_ptr(sdd->sg, j));
7829 kfree(*per_cpu_ptr(sdd->sgp, j));
7831 free_percpu(sdd->sd);
7833 free_percpu(sdd->sg);
7835 free_percpu(sdd->sgp);
7840 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7841 struct s_data *d, const struct cpumask *cpu_map,
7842 struct sched_domain_attr *attr, struct sched_domain *child,
7845 struct sched_domain *sd = tl->init(tl, cpu);
7849 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7851 sd->level = child->level + 1;
7852 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7856 set_domain_attribute(sd, attr);
7862 * Build sched domains for a given set of cpus and attach the sched domains
7863 * to the individual cpus
7865 static int build_sched_domains(const struct cpumask *cpu_map,
7866 struct sched_domain_attr *attr)
7868 enum s_alloc alloc_state = sa_none;
7869 struct sched_domain *sd;
7871 int i, ret = -ENOMEM;
7873 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7874 if (alloc_state != sa_rootdomain)
7877 /* Set up domains for cpus specified by the cpu_map. */
7878 for_each_cpu(i, cpu_map) {
7879 struct sched_domain_topology_level *tl;
7882 for (tl = sched_domain_topology; tl->init; tl++) {
7883 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7884 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7885 sd->flags |= SD_OVERLAP;
7886 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7893 *per_cpu_ptr(d.sd, i) = sd;
7896 /* Build the groups for the domains */
7897 for_each_cpu(i, cpu_map) {
7898 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7899 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7900 if (sd->flags & SD_OVERLAP) {
7901 if (build_overlap_sched_groups(sd, i))
7904 if (build_sched_groups(sd, i))
7910 /* Calculate CPU power for physical packages and nodes */
7911 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7912 if (!cpumask_test_cpu(i, cpu_map))
7915 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7916 claim_allocations(i, sd);
7917 init_sched_groups_power(i, sd);
7921 /* Attach the domains */
7923 for_each_cpu(i, cpu_map) {
7924 sd = *per_cpu_ptr(d.sd, i);
7925 cpu_attach_domain(sd, d.rd, i);
7931 __free_domain_allocs(&d, alloc_state, cpu_map);
7935 static cpumask_var_t *doms_cur; /* current sched domains */
7936 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7937 static struct sched_domain_attr *dattr_cur;
7938 /* attribues of custom domains in 'doms_cur' */
7941 * Special case: If a kmalloc of a doms_cur partition (array of
7942 * cpumask) fails, then fallback to a single sched domain,
7943 * as determined by the single cpumask fallback_doms.
7945 static cpumask_var_t fallback_doms;
7948 * arch_update_cpu_topology lets virtualized architectures update the
7949 * cpu core maps. It is supposed to return 1 if the topology changed
7950 * or 0 if it stayed the same.
7952 int __attribute__((weak)) arch_update_cpu_topology(void)
7957 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7960 cpumask_var_t *doms;
7962 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7965 for (i = 0; i < ndoms; i++) {
7966 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7967 free_sched_domains(doms, i);
7974 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7977 for (i = 0; i < ndoms; i++)
7978 free_cpumask_var(doms[i]);
7983 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7984 * For now this just excludes isolated cpus, but could be used to
7985 * exclude other special cases in the future.
7987 static int init_sched_domains(const struct cpumask *cpu_map)
7991 arch_update_cpu_topology();
7993 doms_cur = alloc_sched_domains(ndoms_cur);
7995 doms_cur = &fallback_doms;
7996 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7998 err = build_sched_domains(doms_cur[0], NULL);
7999 register_sched_domain_sysctl();
8005 * Detach sched domains from a group of cpus specified in cpu_map
8006 * These cpus will now be attached to the NULL domain
8008 static void detach_destroy_domains(const struct cpumask *cpu_map)
8013 for_each_cpu(i, cpu_map)
8014 cpu_attach_domain(NULL, &def_root_domain, i);
8018 /* handle null as "default" */
8019 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8020 struct sched_domain_attr *new, int idx_new)
8022 struct sched_domain_attr tmp;
8029 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8030 new ? (new + idx_new) : &tmp,
8031 sizeof(struct sched_domain_attr));
8035 * Partition sched domains as specified by the 'ndoms_new'
8036 * cpumasks in the array doms_new[] of cpumasks. This compares
8037 * doms_new[] to the current sched domain partitioning, doms_cur[].
8038 * It destroys each deleted domain and builds each new domain.
8040 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
8041 * The masks don't intersect (don't overlap.) We should setup one
8042 * sched domain for each mask. CPUs not in any of the cpumasks will
8043 * not be load balanced. If the same cpumask appears both in the
8044 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8047 * The passed in 'doms_new' should be allocated using
8048 * alloc_sched_domains. This routine takes ownership of it and will
8049 * free_sched_domains it when done with it. If the caller failed the
8050 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
8051 * and partition_sched_domains() will fallback to the single partition
8052 * 'fallback_doms', it also forces the domains to be rebuilt.
8054 * If doms_new == NULL it will be replaced with cpu_online_mask.
8055 * ndoms_new == 0 is a special case for destroying existing domains,
8056 * and it will not create the default domain.
8058 * Call with hotplug lock held
8060 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
8061 struct sched_domain_attr *dattr_new)
8066 mutex_lock(&sched_domains_mutex);
8068 /* always unregister in case we don't destroy any domains */
8069 unregister_sched_domain_sysctl();
8071 /* Let architecture update cpu core mappings. */
8072 new_topology = arch_update_cpu_topology();
8074 n = doms_new ? ndoms_new : 0;
8076 /* Destroy deleted domains */
8077 for (i = 0; i < ndoms_cur; i++) {
8078 for (j = 0; j < n && !new_topology; j++) {
8079 if (cpumask_equal(doms_cur[i], doms_new[j])
8080 && dattrs_equal(dattr_cur, i, dattr_new, j))
8083 /* no match - a current sched domain not in new doms_new[] */
8084 detach_destroy_domains(doms_cur[i]);
8089 if (doms_new == NULL) {
8091 doms_new = &fallback_doms;
8092 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
8093 WARN_ON_ONCE(dattr_new);
8096 /* Build new domains */
8097 for (i = 0; i < ndoms_new; i++) {
8098 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8099 if (cpumask_equal(doms_new[i], doms_cur[j])
8100 && dattrs_equal(dattr_new, i, dattr_cur, j))
8103 /* no match - add a new doms_new */
8104 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
8109 /* Remember the new sched domains */
8110 if (doms_cur != &fallback_doms)
8111 free_sched_domains(doms_cur, ndoms_cur);
8112 kfree(dattr_cur); /* kfree(NULL) is safe */
8113 doms_cur = doms_new;
8114 dattr_cur = dattr_new;
8115 ndoms_cur = ndoms_new;
8117 register_sched_domain_sysctl();
8119 mutex_unlock(&sched_domains_mutex);
8122 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8123 static void reinit_sched_domains(void)
8127 /* Destroy domains first to force the rebuild */
8128 partition_sched_domains(0, NULL, NULL);
8130 rebuild_sched_domains();
8134 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8136 unsigned int level = 0;
8138 if (sscanf(buf, "%u", &level) != 1)
8142 * level is always be positive so don't check for
8143 * level < POWERSAVINGS_BALANCE_NONE which is 0
8144 * What happens on 0 or 1 byte write,
8145 * need to check for count as well?
8148 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8152 sched_smt_power_savings = level;
8154 sched_mc_power_savings = level;
8156 reinit_sched_domains();
8161 #ifdef CONFIG_SCHED_MC
8162 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8163 struct sysdev_class_attribute *attr,
8166 return sprintf(page, "%u\n", sched_mc_power_savings);
8168 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8169 struct sysdev_class_attribute *attr,
8170 const char *buf, size_t count)
8172 return sched_power_savings_store(buf, count, 0);
8174 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8175 sched_mc_power_savings_show,
8176 sched_mc_power_savings_store);
8179 #ifdef CONFIG_SCHED_SMT
8180 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8181 struct sysdev_class_attribute *attr,
8184 return sprintf(page, "%u\n", sched_smt_power_savings);
8186 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8187 struct sysdev_class_attribute *attr,
8188 const char *buf, size_t count)
8190 return sched_power_savings_store(buf, count, 1);
8192 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8193 sched_smt_power_savings_show,
8194 sched_smt_power_savings_store);
8197 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8201 #ifdef CONFIG_SCHED_SMT
8203 err = sysfs_create_file(&cls->kset.kobj,
8204 &attr_sched_smt_power_savings.attr);
8206 #ifdef CONFIG_SCHED_MC
8207 if (!err && mc_capable())
8208 err = sysfs_create_file(&cls->kset.kobj,
8209 &attr_sched_mc_power_savings.attr);
8213 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8215 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
8218 * Update cpusets according to cpu_active mask. If cpusets are
8219 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8220 * around partition_sched_domains().
8222 * If we come here as part of a suspend/resume, don't touch cpusets because we
8223 * want to restore it back to its original state upon resume anyway.
8225 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
8229 case CPU_ONLINE_FROZEN:
8230 case CPU_DOWN_FAILED_FROZEN:
8233 * num_cpus_frozen tracks how many CPUs are involved in suspend
8234 * resume sequence. As long as this is not the last online
8235 * operation in the resume sequence, just build a single sched
8236 * domain, ignoring cpusets.
8239 if (likely(num_cpus_frozen)) {
8240 partition_sched_domains(1, NULL, NULL);
8245 * This is the last CPU online operation. So fall through and
8246 * restore the original sched domains by considering the
8247 * cpuset configurations.
8251 case CPU_DOWN_FAILED:
8252 cpuset_update_active_cpus();
8260 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
8264 case CPU_DOWN_PREPARE:
8265 cpuset_update_active_cpus();
8267 case CPU_DOWN_PREPARE_FROZEN:
8269 partition_sched_domains(1, NULL, NULL);
8277 static int update_runtime(struct notifier_block *nfb,
8278 unsigned long action, void *hcpu)
8280 int cpu = (int)(long)hcpu;
8283 case CPU_DOWN_PREPARE:
8284 case CPU_DOWN_PREPARE_FROZEN:
8285 disable_runtime(cpu_rq(cpu));
8288 case CPU_DOWN_FAILED:
8289 case CPU_DOWN_FAILED_FROZEN:
8291 case CPU_ONLINE_FROZEN:
8292 enable_runtime(cpu_rq(cpu));
8300 void __init sched_init_smp(void)
8302 cpumask_var_t non_isolated_cpus;
8304 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8305 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8308 mutex_lock(&sched_domains_mutex);
8309 init_sched_domains(cpu_active_mask);
8310 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8311 if (cpumask_empty(non_isolated_cpus))
8312 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8313 mutex_unlock(&sched_domains_mutex);
8316 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
8317 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
8319 /* RT runtime code needs to handle some hotplug events */
8320 hotcpu_notifier(update_runtime, 0);
8324 /* Move init over to a non-isolated CPU */
8325 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8327 sched_init_granularity();
8328 free_cpumask_var(non_isolated_cpus);
8330 init_sched_rt_class();
8333 void __init sched_init_smp(void)
8335 sched_init_granularity();
8337 #endif /* CONFIG_SMP */
8339 const_debug unsigned int sysctl_timer_migration = 1;
8341 int in_sched_functions(unsigned long addr)
8343 return in_lock_functions(addr) ||
8344 (addr >= (unsigned long)__sched_text_start
8345 && addr < (unsigned long)__sched_text_end);
8348 static void init_cfs_rq(struct cfs_rq *cfs_rq)
8350 cfs_rq->tasks_timeline = RB_ROOT;
8351 INIT_LIST_HEAD(&cfs_rq->tasks);
8352 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8353 #ifndef CONFIG_64BIT
8354 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8358 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8360 struct rt_prio_array *array;
8363 array = &rt_rq->active;
8364 for (i = 0; i < MAX_RT_PRIO; i++) {
8365 INIT_LIST_HEAD(array->queue + i);
8366 __clear_bit(i, array->bitmap);
8368 /* delimiter for bitsearch: */
8369 __set_bit(MAX_RT_PRIO, array->bitmap);
8371 #if defined CONFIG_SMP
8372 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8373 rt_rq->highest_prio.next = MAX_RT_PRIO;
8374 rt_rq->rt_nr_migratory = 0;
8375 rt_rq->overloaded = 0;
8376 plist_head_init(&rt_rq->pushable_tasks);
8380 rt_rq->rt_throttled = 0;
8381 rt_rq->rt_runtime = 0;
8382 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8385 #ifdef CONFIG_FAIR_GROUP_SCHED
8386 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8387 struct sched_entity *se, int cpu,
8388 struct sched_entity *parent)
8390 struct rq *rq = cpu_rq(cpu);
8395 /* allow initial update_cfs_load() to truncate */
8396 cfs_rq->load_stamp = 1;
8398 init_cfs_rq_runtime(cfs_rq);
8400 tg->cfs_rq[cpu] = cfs_rq;
8403 /* se could be NULL for root_task_group */
8408 se->cfs_rq = &rq->cfs;
8410 se->cfs_rq = parent->my_q;
8413 update_load_set(&se->load, 0);
8414 se->parent = parent;
8418 #ifdef CONFIG_RT_GROUP_SCHED
8419 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8420 struct sched_rt_entity *rt_se, int cpu,
8421 struct sched_rt_entity *parent)
8423 struct rq *rq = cpu_rq(cpu);
8425 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8426 rt_rq->rt_nr_boosted = 0;
8430 tg->rt_rq[cpu] = rt_rq;
8431 tg->rt_se[cpu] = rt_se;
8437 rt_se->rt_rq = &rq->rt;
8439 rt_se->rt_rq = parent->my_q;
8441 rt_se->my_q = rt_rq;
8442 rt_se->parent = parent;
8443 INIT_LIST_HEAD(&rt_se->run_list);
8447 void __init sched_init(void)
8450 unsigned long alloc_size = 0, ptr;
8452 #ifdef CONFIG_FAIR_GROUP_SCHED
8453 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8455 #ifdef CONFIG_RT_GROUP_SCHED
8456 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8458 #ifdef CONFIG_CPUMASK_OFFSTACK
8459 alloc_size += num_possible_cpus() * cpumask_size();
8462 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8464 #ifdef CONFIG_FAIR_GROUP_SCHED
8465 root_task_group.se = (struct sched_entity **)ptr;
8466 ptr += nr_cpu_ids * sizeof(void **);
8468 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8469 ptr += nr_cpu_ids * sizeof(void **);
8471 #endif /* CONFIG_FAIR_GROUP_SCHED */
8472 #ifdef CONFIG_RT_GROUP_SCHED
8473 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8474 ptr += nr_cpu_ids * sizeof(void **);
8476 root_task_group.rt_rq = (struct rt_rq **)ptr;
8477 ptr += nr_cpu_ids * sizeof(void **);
8479 #endif /* CONFIG_RT_GROUP_SCHED */
8480 #ifdef CONFIG_CPUMASK_OFFSTACK
8481 for_each_possible_cpu(i) {
8482 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8483 ptr += cpumask_size();
8485 #endif /* CONFIG_CPUMASK_OFFSTACK */
8489 init_defrootdomain();
8492 init_rt_bandwidth(&def_rt_bandwidth,
8493 global_rt_period(), global_rt_runtime());
8495 #ifdef CONFIG_RT_GROUP_SCHED
8496 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8497 global_rt_period(), global_rt_runtime());
8498 #endif /* CONFIG_RT_GROUP_SCHED */
8500 #ifdef CONFIG_CGROUP_SCHED
8501 list_add(&root_task_group.list, &task_groups);
8502 INIT_LIST_HEAD(&root_task_group.children);
8503 autogroup_init(&init_task);
8504 #endif /* CONFIG_CGROUP_SCHED */
8506 for_each_possible_cpu(i) {
8510 raw_spin_lock_init(&rq->lock);
8512 rq->calc_load_active = 0;
8513 rq->calc_load_update = jiffies + LOAD_FREQ;
8514 init_cfs_rq(&rq->cfs);
8515 init_rt_rq(&rq->rt, rq);
8516 #ifdef CONFIG_FAIR_GROUP_SCHED
8517 root_task_group.shares = root_task_group_load;
8518 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8520 * How much cpu bandwidth does root_task_group get?
8522 * In case of task-groups formed thr' the cgroup filesystem, it
8523 * gets 100% of the cpu resources in the system. This overall
8524 * system cpu resource is divided among the tasks of
8525 * root_task_group and its child task-groups in a fair manner,
8526 * based on each entity's (task or task-group's) weight
8527 * (se->load.weight).
8529 * In other words, if root_task_group has 10 tasks of weight
8530 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8531 * then A0's share of the cpu resource is:
8533 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8535 * We achieve this by letting root_task_group's tasks sit
8536 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8538 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8539 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8540 #endif /* CONFIG_FAIR_GROUP_SCHED */
8542 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8543 #ifdef CONFIG_RT_GROUP_SCHED
8544 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8545 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8548 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8549 rq->cpu_load[j] = 0;
8551 rq->last_load_update_tick = jiffies;
8556 rq->cpu_power = SCHED_POWER_SCALE;
8557 rq->post_schedule = 0;
8558 rq->active_balance = 0;
8559 rq->next_balance = jiffies;
8564 rq->avg_idle = 2*sysctl_sched_migration_cost;
8565 rq_attach_root(rq, &def_root_domain);
8567 rq->nohz_balance_kick = 0;
8571 atomic_set(&rq->nr_iowait, 0);
8574 set_load_weight(&init_task);
8576 #ifdef CONFIG_PREEMPT_NOTIFIERS
8577 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8581 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8584 #ifdef CONFIG_RT_MUTEXES
8585 plist_head_init(&init_task.pi_waiters);
8589 * The boot idle thread does lazy MMU switching as well:
8591 atomic_inc(&init_mm.mm_count);
8592 enter_lazy_tlb(&init_mm, current);
8595 * Make us the idle thread. Technically, schedule() should not be
8596 * called from this thread, however somewhere below it might be,
8597 * but because we are the idle thread, we just pick up running again
8598 * when this runqueue becomes "idle".
8600 init_idle(current, smp_processor_id());
8602 calc_load_update = jiffies + LOAD_FREQ;
8605 * During early bootup we pretend to be a normal task:
8607 current->sched_class = &fair_sched_class;
8610 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8612 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8613 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8614 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8615 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8616 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8618 /* May be allocated at isolcpus cmdline parse time */
8619 if (cpu_isolated_map == NULL)
8620 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8623 scheduler_running = 1;
8626 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8627 static inline int preempt_count_equals(int preempt_offset)
8629 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8631 return (nested == preempt_offset);
8634 void __might_sleep(const char *file, int line, int preempt_offset)
8636 static unsigned long prev_jiffy; /* ratelimiting */
8638 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
8639 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8640 system_state != SYSTEM_RUNNING || oops_in_progress)
8642 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8644 prev_jiffy = jiffies;
8647 "BUG: sleeping function called from invalid context at %s:%d\n",
8650 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8651 in_atomic(), irqs_disabled(),
8652 current->pid, current->comm);
8654 debug_show_held_locks(current);
8655 if (irqs_disabled())
8656 print_irqtrace_events(current);
8659 EXPORT_SYMBOL(__might_sleep);
8662 #ifdef CONFIG_MAGIC_SYSRQ
8663 static void normalize_task(struct rq *rq, struct task_struct *p)
8665 const struct sched_class *prev_class = p->sched_class;
8666 int old_prio = p->prio;
8671 deactivate_task(rq, p, 0);
8672 __setscheduler(rq, p, SCHED_NORMAL, 0);
8674 activate_task(rq, p, 0);
8675 resched_task(rq->curr);
8678 check_class_changed(rq, p, prev_class, old_prio);
8681 void normalize_rt_tasks(void)
8683 struct task_struct *g, *p;
8684 unsigned long flags;
8687 read_lock_irqsave(&tasklist_lock, flags);
8688 do_each_thread(g, p) {
8690 * Only normalize user tasks:
8695 p->se.exec_start = 0;
8696 #ifdef CONFIG_SCHEDSTATS
8697 p->se.statistics.wait_start = 0;
8698 p->se.statistics.sleep_start = 0;
8699 p->se.statistics.block_start = 0;
8704 * Renice negative nice level userspace
8707 if (TASK_NICE(p) < 0 && p->mm)
8708 set_user_nice(p, 0);
8712 raw_spin_lock(&p->pi_lock);
8713 rq = __task_rq_lock(p);
8715 normalize_task(rq, p);
8717 __task_rq_unlock(rq);
8718 raw_spin_unlock(&p->pi_lock);
8719 } while_each_thread(g, p);
8721 read_unlock_irqrestore(&tasklist_lock, flags);
8724 #endif /* CONFIG_MAGIC_SYSRQ */
8726 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8728 * These functions are only useful for the IA64 MCA handling, or kdb.
8730 * They can only be called when the whole system has been
8731 * stopped - every CPU needs to be quiescent, and no scheduling
8732 * activity can take place. Using them for anything else would
8733 * be a serious bug, and as a result, they aren't even visible
8734 * under any other configuration.
8738 * curr_task - return the current task for a given cpu.
8739 * @cpu: the processor in question.
8741 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8743 struct task_struct *curr_task(int cpu)
8745 return cpu_curr(cpu);
8748 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8752 * set_curr_task - set the current task for a given cpu.
8753 * @cpu: the processor in question.
8754 * @p: the task pointer to set.
8756 * Description: This function must only be used when non-maskable interrupts
8757 * are serviced on a separate stack. It allows the architecture to switch the
8758 * notion of the current task on a cpu in a non-blocking manner. This function
8759 * must be called with all CPU's synchronized, and interrupts disabled, the
8760 * and caller must save the original value of the current task (see
8761 * curr_task() above) and restore that value before reenabling interrupts and
8762 * re-starting the system.
8764 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8766 void set_curr_task(int cpu, struct task_struct *p)
8773 #ifdef CONFIG_FAIR_GROUP_SCHED
8774 static void free_fair_sched_group(struct task_group *tg)
8778 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8780 for_each_possible_cpu(i) {
8782 kfree(tg->cfs_rq[i]);
8792 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8794 struct cfs_rq *cfs_rq;
8795 struct sched_entity *se;
8798 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8801 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8805 tg->shares = NICE_0_LOAD;
8807 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8809 for_each_possible_cpu(i) {
8810 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8811 GFP_KERNEL, cpu_to_node(i));
8815 se = kzalloc_node(sizeof(struct sched_entity),
8816 GFP_KERNEL, cpu_to_node(i));
8820 init_cfs_rq(cfs_rq);
8821 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8832 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8834 struct rq *rq = cpu_rq(cpu);
8835 unsigned long flags;
8838 * Only empty task groups can be destroyed; so we can speculatively
8839 * check on_list without danger of it being re-added.
8841 if (!tg->cfs_rq[cpu]->on_list)
8844 raw_spin_lock_irqsave(&rq->lock, flags);
8845 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8846 raw_spin_unlock_irqrestore(&rq->lock, flags);
8848 #else /* !CONFIG_FAIR_GROUP_SCHED */
8849 static inline void free_fair_sched_group(struct task_group *tg)
8854 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8859 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8862 #endif /* CONFIG_FAIR_GROUP_SCHED */
8864 #ifdef CONFIG_RT_GROUP_SCHED
8865 static void free_rt_sched_group(struct task_group *tg)
8870 destroy_rt_bandwidth(&tg->rt_bandwidth);
8872 for_each_possible_cpu(i) {
8874 kfree(tg->rt_rq[i]);
8876 kfree(tg->rt_se[i]);
8884 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8886 struct rt_rq *rt_rq;
8887 struct sched_rt_entity *rt_se;
8890 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8893 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8897 init_rt_bandwidth(&tg->rt_bandwidth,
8898 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8900 for_each_possible_cpu(i) {
8901 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8902 GFP_KERNEL, cpu_to_node(i));
8906 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8907 GFP_KERNEL, cpu_to_node(i));
8911 init_rt_rq(rt_rq, cpu_rq(i));
8912 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8913 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8923 #else /* !CONFIG_RT_GROUP_SCHED */
8924 static inline void free_rt_sched_group(struct task_group *tg)
8929 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8933 #endif /* CONFIG_RT_GROUP_SCHED */
8935 #ifdef CONFIG_CGROUP_SCHED
8936 static void free_sched_group(struct task_group *tg)
8938 free_fair_sched_group(tg);
8939 free_rt_sched_group(tg);
8944 /* allocate runqueue etc for a new task group */
8945 struct task_group *sched_create_group(struct task_group *parent)
8947 struct task_group *tg;
8948 unsigned long flags;
8950 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8952 return ERR_PTR(-ENOMEM);
8954 if (!alloc_fair_sched_group(tg, parent))
8957 if (!alloc_rt_sched_group(tg, parent))
8960 spin_lock_irqsave(&task_group_lock, flags);
8961 list_add_rcu(&tg->list, &task_groups);
8963 WARN_ON(!parent); /* root should already exist */
8965 tg->parent = parent;
8966 INIT_LIST_HEAD(&tg->children);
8967 list_add_rcu(&tg->siblings, &parent->children);
8968 spin_unlock_irqrestore(&task_group_lock, flags);
8973 free_sched_group(tg);
8974 return ERR_PTR(-ENOMEM);
8977 /* rcu callback to free various structures associated with a task group */
8978 static void free_sched_group_rcu(struct rcu_head *rhp)
8980 /* now it should be safe to free those cfs_rqs */
8981 free_sched_group(container_of(rhp, struct task_group, rcu));
8984 /* Destroy runqueue etc associated with a task group */
8985 void sched_destroy_group(struct task_group *tg)
8987 unsigned long flags;
8990 /* end participation in shares distribution */
8991 for_each_possible_cpu(i)
8992 unregister_fair_sched_group(tg, i);
8994 spin_lock_irqsave(&task_group_lock, flags);
8995 list_del_rcu(&tg->list);
8996 list_del_rcu(&tg->siblings);
8997 spin_unlock_irqrestore(&task_group_lock, flags);
8999 /* wait for possible concurrent references to cfs_rqs complete */
9000 call_rcu(&tg->rcu, free_sched_group_rcu);
9003 /* change task's runqueue when it moves between groups.
9004 * The caller of this function should have put the task in its new group
9005 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9006 * reflect its new group.
9008 void sched_move_task(struct task_struct *tsk)
9010 struct task_group *tg;
9012 unsigned long flags;
9015 rq = task_rq_lock(tsk, &flags);
9017 running = task_current(rq, tsk);
9021 dequeue_task(rq, tsk, 0);
9022 if (unlikely(running))
9023 tsk->sched_class->put_prev_task(rq, tsk);
9025 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
9026 lockdep_is_held(&tsk->sighand->siglock)),
9027 struct task_group, css);
9028 tg = autogroup_task_group(tsk, tg);
9029 tsk->sched_task_group = tg;
9031 #ifdef CONFIG_FAIR_GROUP_SCHED
9032 if (tsk->sched_class->task_move_group)
9033 tsk->sched_class->task_move_group(tsk, on_rq);
9036 set_task_rq(tsk, task_cpu(tsk));
9038 if (unlikely(running))
9039 tsk->sched_class->set_curr_task(rq);
9041 enqueue_task(rq, tsk, 0);
9043 task_rq_unlock(rq, tsk, &flags);
9045 #endif /* CONFIG_CGROUP_SCHED */
9047 #ifdef CONFIG_FAIR_GROUP_SCHED
9048 static DEFINE_MUTEX(shares_mutex);
9050 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9053 unsigned long flags;
9056 * We can't change the weight of the root cgroup.
9061 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9063 mutex_lock(&shares_mutex);
9064 if (tg->shares == shares)
9067 tg->shares = shares;
9068 for_each_possible_cpu(i) {
9069 struct rq *rq = cpu_rq(i);
9070 struct sched_entity *se;
9073 /* Propagate contribution to hierarchy */
9074 raw_spin_lock_irqsave(&rq->lock, flags);
9075 for_each_sched_entity(se)
9076 update_cfs_shares(group_cfs_rq(se));
9077 raw_spin_unlock_irqrestore(&rq->lock, flags);
9081 mutex_unlock(&shares_mutex);
9085 unsigned long sched_group_shares(struct task_group *tg)
9091 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
9092 static unsigned long to_ratio(u64 period, u64 runtime)
9094 if (runtime == RUNTIME_INF)
9097 return div64_u64(runtime << 20, period);
9101 #ifdef CONFIG_RT_GROUP_SCHED
9103 * Ensure that the real time constraints are schedulable.
9105 static DEFINE_MUTEX(rt_constraints_mutex);
9107 /* Must be called with tasklist_lock held */
9108 static inline int tg_has_rt_tasks(struct task_group *tg)
9110 struct task_struct *g, *p;
9113 * Autogroups do not have RT tasks; see autogroup_create().
9115 if (task_group_is_autogroup(tg))
9118 do_each_thread(g, p) {
9119 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9121 } while_each_thread(g, p);
9126 struct rt_schedulable_data {
9127 struct task_group *tg;
9132 static int tg_rt_schedulable(struct task_group *tg, void *data)
9134 struct rt_schedulable_data *d = data;
9135 struct task_group *child;
9136 unsigned long total, sum = 0;
9137 u64 period, runtime;
9139 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9140 runtime = tg->rt_bandwidth.rt_runtime;
9143 period = d->rt_period;
9144 runtime = d->rt_runtime;
9148 * Cannot have more runtime than the period.
9150 if (runtime > period && runtime != RUNTIME_INF)
9154 * Ensure we don't starve existing RT tasks.
9156 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9159 total = to_ratio(period, runtime);
9162 * Nobody can have more than the global setting allows.
9164 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9168 * The sum of our children's runtime should not exceed our own.
9170 list_for_each_entry_rcu(child, &tg->children, siblings) {
9171 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9172 runtime = child->rt_bandwidth.rt_runtime;
9174 if (child == d->tg) {
9175 period = d->rt_period;
9176 runtime = d->rt_runtime;
9179 sum += to_ratio(period, runtime);
9188 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9192 struct rt_schedulable_data data = {
9194 .rt_period = period,
9195 .rt_runtime = runtime,
9199 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
9205 static int tg_set_rt_bandwidth(struct task_group *tg,
9206 u64 rt_period, u64 rt_runtime)
9210 mutex_lock(&rt_constraints_mutex);
9211 read_lock(&tasklist_lock);
9212 err = __rt_schedulable(tg, rt_period, rt_runtime);
9216 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9217 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9218 tg->rt_bandwidth.rt_runtime = rt_runtime;
9220 for_each_possible_cpu(i) {
9221 struct rt_rq *rt_rq = tg->rt_rq[i];
9223 raw_spin_lock(&rt_rq->rt_runtime_lock);
9224 rt_rq->rt_runtime = rt_runtime;
9225 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9227 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9229 read_unlock(&tasklist_lock);
9230 mutex_unlock(&rt_constraints_mutex);
9235 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9237 u64 rt_runtime, rt_period;
9239 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9240 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9241 if (rt_runtime_us < 0)
9242 rt_runtime = RUNTIME_INF;
9244 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9247 long sched_group_rt_runtime(struct task_group *tg)
9251 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9254 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9255 do_div(rt_runtime_us, NSEC_PER_USEC);
9256 return rt_runtime_us;
9259 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9261 u64 rt_runtime, rt_period;
9263 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9264 rt_runtime = tg->rt_bandwidth.rt_runtime;
9269 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9272 long sched_group_rt_period(struct task_group *tg)
9276 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9277 do_div(rt_period_us, NSEC_PER_USEC);
9278 return rt_period_us;
9281 static int sched_rt_global_constraints(void)
9283 u64 runtime, period;
9286 if (sysctl_sched_rt_period <= 0)
9289 runtime = global_rt_runtime();
9290 period = global_rt_period();
9293 * Sanity check on the sysctl variables.
9295 if (runtime > period && runtime != RUNTIME_INF)
9298 mutex_lock(&rt_constraints_mutex);
9299 read_lock(&tasklist_lock);
9300 ret = __rt_schedulable(NULL, 0, 0);
9301 read_unlock(&tasklist_lock);
9302 mutex_unlock(&rt_constraints_mutex);
9307 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9309 /* Don't accept realtime tasks when there is no way for them to run */
9310 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9316 #else /* !CONFIG_RT_GROUP_SCHED */
9317 static int sched_rt_global_constraints(void)
9319 unsigned long flags;
9322 if (sysctl_sched_rt_period <= 0)
9326 * There's always some RT tasks in the root group
9327 * -- migration, kstopmachine etc..
9329 if (sysctl_sched_rt_runtime == 0)
9332 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9333 for_each_possible_cpu(i) {
9334 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9336 raw_spin_lock(&rt_rq->rt_runtime_lock);
9337 rt_rq->rt_runtime = global_rt_runtime();
9338 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9340 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9344 #endif /* CONFIG_RT_GROUP_SCHED */
9346 int sched_rt_handler(struct ctl_table *table, int write,
9347 void __user *buffer, size_t *lenp,
9351 int old_period, old_runtime;
9352 static DEFINE_MUTEX(mutex);
9355 old_period = sysctl_sched_rt_period;
9356 old_runtime = sysctl_sched_rt_runtime;
9358 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9360 if (!ret && write) {
9361 ret = sched_rt_global_constraints();
9363 sysctl_sched_rt_period = old_period;
9364 sysctl_sched_rt_runtime = old_runtime;
9366 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9367 def_rt_bandwidth.rt_period =
9368 ns_to_ktime(global_rt_period());
9371 mutex_unlock(&mutex);
9376 #ifdef CONFIG_CGROUP_SCHED
9378 /* return corresponding task_group object of a cgroup */
9379 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9381 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9382 struct task_group, css);
9385 static struct cgroup_subsys_state *
9386 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9388 struct task_group *tg, *parent;
9390 if (!cgrp->parent) {
9391 /* This is early initialization for the top cgroup */
9392 return &root_task_group.css;
9395 parent = cgroup_tg(cgrp->parent);
9396 tg = sched_create_group(parent);
9398 return ERR_PTR(-ENOMEM);
9404 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9406 struct task_group *tg = cgroup_tg(cgrp);
9408 sched_destroy_group(tg);
9412 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9414 #ifdef CONFIG_RT_GROUP_SCHED
9415 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9418 /* We don't support RT-tasks being in separate groups */
9419 if (tsk->sched_class != &fair_sched_class)
9426 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9428 sched_move_task(tsk);
9432 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9433 struct cgroup *old_cgrp, struct task_struct *task)
9436 * cgroup_exit() is called in the copy_process() failure path.
9437 * Ignore this case since the task hasn't ran yet, this avoids
9438 * trying to poke a half freed task state from generic code.
9440 if (!(task->flags & PF_EXITING))
9443 sched_move_task(task);
9446 #ifdef CONFIG_FAIR_GROUP_SCHED
9447 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9450 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9453 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9455 struct task_group *tg = cgroup_tg(cgrp);
9457 return (u64) scale_load_down(tg->shares);
9460 #ifdef CONFIG_CFS_BANDWIDTH
9461 static DEFINE_MUTEX(cfs_constraints_mutex);
9463 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9464 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9466 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9468 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9470 int i, ret = 0, runtime_enabled;
9471 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9473 if (tg == &root_task_group)
9477 * Ensure we have at some amount of bandwidth every period. This is
9478 * to prevent reaching a state of large arrears when throttled via
9479 * entity_tick() resulting in prolonged exit starvation.
9481 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9485 * Likewise, bound things on the otherside by preventing insane quota
9486 * periods. This also allows us to normalize in computing quota
9489 if (period > max_cfs_quota_period)
9492 mutex_lock(&cfs_constraints_mutex);
9493 ret = __cfs_schedulable(tg, period, quota);
9497 runtime_enabled = quota != RUNTIME_INF;
9498 raw_spin_lock_irq(&cfs_b->lock);
9499 cfs_b->period = ns_to_ktime(period);
9500 cfs_b->quota = quota;
9502 __refill_cfs_bandwidth_runtime(cfs_b);
9503 /* restart the period timer (if active) to handle new period expiry */
9504 if (runtime_enabled && cfs_b->timer_active) {
9505 /* force a reprogram */
9506 cfs_b->timer_active = 0;
9507 __start_cfs_bandwidth(cfs_b);
9509 raw_spin_unlock_irq(&cfs_b->lock);
9511 for_each_possible_cpu(i) {
9512 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9513 struct rq *rq = rq_of(cfs_rq);
9515 raw_spin_lock_irq(&rq->lock);
9516 cfs_rq->runtime_enabled = runtime_enabled;
9517 cfs_rq->runtime_remaining = 0;
9519 if (cfs_rq_throttled(cfs_rq))
9520 unthrottle_cfs_rq(cfs_rq);
9521 raw_spin_unlock_irq(&rq->lock);
9524 mutex_unlock(&cfs_constraints_mutex);
9529 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9533 period = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9534 if (cfs_quota_us < 0)
9535 quota = RUNTIME_INF;
9537 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9539 return tg_set_cfs_bandwidth(tg, period, quota);
9542 long tg_get_cfs_quota(struct task_group *tg)
9546 if (tg_cfs_bandwidth(tg)->quota == RUNTIME_INF)
9549 quota_us = tg_cfs_bandwidth(tg)->quota;
9550 do_div(quota_us, NSEC_PER_USEC);
9555 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9559 period = (u64)cfs_period_us * NSEC_PER_USEC;
9560 quota = tg_cfs_bandwidth(tg)->quota;
9565 return tg_set_cfs_bandwidth(tg, period, quota);
9568 long tg_get_cfs_period(struct task_group *tg)
9572 cfs_period_us = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9573 do_div(cfs_period_us, NSEC_PER_USEC);
9575 return cfs_period_us;
9578 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
9580 return tg_get_cfs_quota(cgroup_tg(cgrp));
9583 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
9586 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
9589 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
9591 return tg_get_cfs_period(cgroup_tg(cgrp));
9594 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9597 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
9600 struct cfs_schedulable_data {
9601 struct task_group *tg;
9606 * normalize group quota/period to be quota/max_period
9607 * note: units are usecs
9609 static u64 normalize_cfs_quota(struct task_group *tg,
9610 struct cfs_schedulable_data *d)
9618 period = tg_get_cfs_period(tg);
9619 quota = tg_get_cfs_quota(tg);
9622 /* note: these should typically be equivalent */
9623 if (quota == RUNTIME_INF || quota == -1)
9626 return to_ratio(period, quota);
9629 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9631 struct cfs_schedulable_data *d = data;
9632 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9633 s64 quota = 0, parent_quota = -1;
9636 quota = RUNTIME_INF;
9638 struct cfs_bandwidth *parent_b = tg_cfs_bandwidth(tg->parent);
9640 quota = normalize_cfs_quota(tg, d);
9641 parent_quota = parent_b->hierarchal_quota;
9644 * ensure max(child_quota) <= parent_quota, inherit when no
9647 if (quota == RUNTIME_INF)
9648 quota = parent_quota;
9649 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9652 cfs_b->hierarchal_quota = quota;
9657 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9660 struct cfs_schedulable_data data = {
9666 if (quota != RUNTIME_INF) {
9667 do_div(data.period, NSEC_PER_USEC);
9668 do_div(data.quota, NSEC_PER_USEC);
9672 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9678 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
9679 struct cgroup_map_cb *cb)
9681 struct task_group *tg = cgroup_tg(cgrp);
9682 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9684 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
9685 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
9686 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
9690 #endif /* CONFIG_CFS_BANDWIDTH */
9691 #endif /* CONFIG_FAIR_GROUP_SCHED */
9693 #ifdef CONFIG_RT_GROUP_SCHED
9694 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9697 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9700 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9702 return sched_group_rt_runtime(cgroup_tg(cgrp));
9705 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9708 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9711 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9713 return sched_group_rt_period(cgroup_tg(cgrp));
9715 #endif /* CONFIG_RT_GROUP_SCHED */
9717 static struct cftype cpu_files[] = {
9718 #ifdef CONFIG_FAIR_GROUP_SCHED
9721 .read_u64 = cpu_shares_read_u64,
9722 .write_u64 = cpu_shares_write_u64,
9725 #ifdef CONFIG_CFS_BANDWIDTH
9727 .name = "cfs_quota_us",
9728 .read_s64 = cpu_cfs_quota_read_s64,
9729 .write_s64 = cpu_cfs_quota_write_s64,
9732 .name = "cfs_period_us",
9733 .read_u64 = cpu_cfs_period_read_u64,
9734 .write_u64 = cpu_cfs_period_write_u64,
9738 .read_map = cpu_stats_show,
9741 #ifdef CONFIG_RT_GROUP_SCHED
9743 .name = "rt_runtime_us",
9744 .read_s64 = cpu_rt_runtime_read,
9745 .write_s64 = cpu_rt_runtime_write,
9748 .name = "rt_period_us",
9749 .read_u64 = cpu_rt_period_read_uint,
9750 .write_u64 = cpu_rt_period_write_uint,
9755 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9757 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9760 struct cgroup_subsys cpu_cgroup_subsys = {
9762 .create = cpu_cgroup_create,
9763 .destroy = cpu_cgroup_destroy,
9764 .can_attach_task = cpu_cgroup_can_attach_task,
9765 .attach_task = cpu_cgroup_attach_task,
9766 .exit = cpu_cgroup_exit,
9767 .populate = cpu_cgroup_populate,
9768 .subsys_id = cpu_cgroup_subsys_id,
9772 #endif /* CONFIG_CGROUP_SCHED */
9774 #ifdef CONFIG_CGROUP_CPUACCT
9777 * CPU accounting code for task groups.
9779 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9780 * (balbir@in.ibm.com).
9783 /* track cpu usage of a group of tasks and its child groups */
9785 struct cgroup_subsys_state css;
9786 /* cpuusage holds pointer to a u64-type object on every cpu */
9787 u64 __percpu *cpuusage;
9788 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9789 struct cpuacct *parent;
9792 struct cgroup_subsys cpuacct_subsys;
9794 /* return cpu accounting group corresponding to this container */
9795 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9797 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9798 struct cpuacct, css);
9801 /* return cpu accounting group to which this task belongs */
9802 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9804 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9805 struct cpuacct, css);
9808 /* create a new cpu accounting group */
9809 static struct cgroup_subsys_state *cpuacct_create(
9810 struct cgroup_subsys *ss, struct cgroup *cgrp)
9812 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9818 ca->cpuusage = alloc_percpu(u64);
9822 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9823 if (percpu_counter_init(&ca->cpustat[i], 0))
9824 goto out_free_counters;
9827 ca->parent = cgroup_ca(cgrp->parent);
9833 percpu_counter_destroy(&ca->cpustat[i]);
9834 free_percpu(ca->cpuusage);
9838 return ERR_PTR(-ENOMEM);
9841 /* destroy an existing cpu accounting group */
9843 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9845 struct cpuacct *ca = cgroup_ca(cgrp);
9848 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9849 percpu_counter_destroy(&ca->cpustat[i]);
9850 free_percpu(ca->cpuusage);
9854 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9856 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9859 #ifndef CONFIG_64BIT
9861 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9863 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9865 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9873 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9875 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9877 #ifndef CONFIG_64BIT
9879 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9881 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9883 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9889 /* return total cpu usage (in nanoseconds) of a group */
9890 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9892 struct cpuacct *ca = cgroup_ca(cgrp);
9893 u64 totalcpuusage = 0;
9896 for_each_present_cpu(i)
9897 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9899 return totalcpuusage;
9902 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9905 struct cpuacct *ca = cgroup_ca(cgrp);
9914 for_each_present_cpu(i)
9915 cpuacct_cpuusage_write(ca, i, 0);
9921 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9924 struct cpuacct *ca = cgroup_ca(cgroup);
9928 for_each_present_cpu(i) {
9929 percpu = cpuacct_cpuusage_read(ca, i);
9930 seq_printf(m, "%llu ", (unsigned long long) percpu);
9932 seq_printf(m, "\n");
9936 static const char *cpuacct_stat_desc[] = {
9937 [CPUACCT_STAT_USER] = "user",
9938 [CPUACCT_STAT_SYSTEM] = "system",
9941 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9942 struct cgroup_map_cb *cb)
9944 struct cpuacct *ca = cgroup_ca(cgrp);
9947 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9948 s64 val = percpu_counter_read(&ca->cpustat[i]);
9949 val = cputime64_to_clock_t(val);
9950 cb->fill(cb, cpuacct_stat_desc[i], val);
9955 static struct cftype files[] = {
9958 .read_u64 = cpuusage_read,
9959 .write_u64 = cpuusage_write,
9962 .name = "usage_percpu",
9963 .read_seq_string = cpuacct_percpu_seq_read,
9967 .read_map = cpuacct_stats_show,
9971 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9973 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9977 * charge this task's execution time to its accounting group.
9979 * called with rq->lock held.
9981 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9986 if (unlikely(!cpuacct_subsys.active))
9989 cpu = task_cpu(tsk);
9995 for (; ca; ca = ca->parent) {
9996 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9997 *cpuusage += cputime;
10004 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
10005 * in cputime_t units. As a result, cpuacct_update_stats calls
10006 * percpu_counter_add with values large enough to always overflow the
10007 * per cpu batch limit causing bad SMP scalability.
10009 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
10010 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
10011 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
10014 #define CPUACCT_BATCH \
10015 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
10017 #define CPUACCT_BATCH 0
10021 * Charge the system/user time to the task's accounting group.
10023 static void cpuacct_update_stats(struct task_struct *tsk,
10024 enum cpuacct_stat_index idx, cputime_t val)
10026 struct cpuacct *ca;
10027 int batch = CPUACCT_BATCH;
10029 if (unlikely(!cpuacct_subsys.active))
10036 __percpu_counter_add(&ca->cpustat[idx], val, batch);
10042 struct cgroup_subsys cpuacct_subsys = {
10044 .create = cpuacct_create,
10045 .destroy = cpuacct_destroy,
10046 .populate = cpuacct_populate,
10047 .subsys_id = cpuacct_subsys_id,
10049 #endif /* CONFIG_CGROUP_CPUACCT */