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 */
2837 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2838 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2839 * in smp_cond_load_acquire() below.
2841 * sched_ttwu_pending() try_to_wake_up()
2842 * [S] p->on_rq = 1; [L] P->state
2843 * UNLOCK rq->lock -----.
2847 * LOCK rq->lock -----'
2851 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2853 * Pairs with the UNLOCK+LOCK on rq->lock from the
2854 * last wakeup of our task and the schedule that got our task
2858 if (p->on_rq && ttwu_remote(p, wake_flags))
2863 * If the owning (remote) cpu is still in the middle of schedule() with
2864 * this task as prev, wait until its done referencing the task.
2867 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2869 * In case the architecture enables interrupts in
2870 * context_switch(), we cannot busy wait, since that
2871 * would lead to deadlocks when an interrupt hits and
2872 * tries to wake up @prev. So bail and do a complete
2875 if (ttwu_activate_remote(p, wake_flags))
2882 * Pairs with the smp_wmb() in finish_lock_switch().
2886 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2887 p->state = TASK_WAKING;
2889 if (p->sched_class->task_waking)
2890 p->sched_class->task_waking(p);
2892 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2893 if (task_cpu(p) != cpu) {
2894 wake_flags |= WF_MIGRATED;
2895 set_task_cpu(p, cpu);
2897 #endif /* CONFIG_SMP */
2901 ttwu_stat(p, cpu, wake_flags);
2903 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2909 * try_to_wake_up_local - try to wake up a local task with rq lock held
2910 * @p: the thread to be awakened
2912 * Put @p on the run-queue if it's not already there. The caller must
2913 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2916 static void try_to_wake_up_local(struct task_struct *p)
2918 struct rq *rq = task_rq(p);
2920 if (WARN_ON_ONCE(rq != this_rq()) ||
2921 WARN_ON_ONCE(p == current))
2924 lockdep_assert_held(&rq->lock);
2926 if (!raw_spin_trylock(&p->pi_lock)) {
2927 raw_spin_unlock(&rq->lock);
2928 raw_spin_lock(&p->pi_lock);
2929 raw_spin_lock(&rq->lock);
2932 if (!(p->state & TASK_NORMAL))
2936 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2938 ttwu_do_wakeup(rq, p, 0);
2939 ttwu_stat(p, smp_processor_id(), 0);
2941 raw_spin_unlock(&p->pi_lock);
2945 * wake_up_process - Wake up a specific process
2946 * @p: The process to be woken up.
2948 * Attempt to wake up the nominated process and move it to the set of runnable
2949 * processes. Returns 1 if the process was woken up, 0 if it was already
2952 * It may be assumed that this function implies a write memory barrier before
2953 * changing the task state if and only if any tasks are woken up.
2955 int wake_up_process(struct task_struct *p)
2957 return try_to_wake_up(p, TASK_NORMAL, 0);
2959 EXPORT_SYMBOL(wake_up_process);
2961 int wake_up_state(struct task_struct *p, unsigned int state)
2963 return try_to_wake_up(p, state, 0);
2967 * Perform scheduler related setup for a newly forked process p.
2968 * p is forked by current.
2970 * __sched_fork() is basic setup used by init_idle() too:
2972 static void __sched_fork(struct task_struct *p)
2977 p->se.exec_start = 0;
2978 p->se.sum_exec_runtime = 0;
2979 p->se.prev_sum_exec_runtime = 0;
2980 p->se.nr_migrations = 0;
2982 INIT_LIST_HEAD(&p->se.group_node);
2984 #ifdef CONFIG_SCHEDSTATS
2985 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2988 INIT_LIST_HEAD(&p->rt.run_list);
2990 #ifdef CONFIG_PREEMPT_NOTIFIERS
2991 INIT_HLIST_HEAD(&p->preempt_notifiers);
2996 * fork()/clone()-time setup:
2998 void sched_fork(struct task_struct *p)
3000 unsigned long flags;
3001 int cpu = get_cpu();
3005 * We mark the process as running here. This guarantees that
3006 * nobody will actually run it, and a signal or other external
3007 * event cannot wake it up and insert it on the runqueue either.
3009 p->state = TASK_RUNNING;
3012 * Make sure we do not leak PI boosting priority to the child.
3014 p->prio = current->normal_prio;
3017 * Revert to default priority/policy on fork if requested.
3019 if (unlikely(p->sched_reset_on_fork)) {
3020 if (task_has_rt_policy(p)) {
3021 p->policy = SCHED_NORMAL;
3022 p->static_prio = NICE_TO_PRIO(0);
3024 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3025 p->static_prio = NICE_TO_PRIO(0);
3027 p->prio = p->normal_prio = __normal_prio(p);
3031 * We don't need the reset flag anymore after the fork. It has
3032 * fulfilled its duty:
3034 p->sched_reset_on_fork = 0;
3037 if (!rt_prio(p->prio))
3038 p->sched_class = &fair_sched_class;
3040 if (p->sched_class->task_fork)
3041 p->sched_class->task_fork(p);
3044 * The child is not yet in the pid-hash so no cgroup attach races,
3045 * and the cgroup is pinned to this child due to cgroup_fork()
3046 * is ran before sched_fork().
3048 * Silence PROVE_RCU.
3050 raw_spin_lock_irqsave(&p->pi_lock, flags);
3051 set_task_cpu(p, cpu);
3052 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3054 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3055 if (likely(sched_info_on()))
3056 memset(&p->sched_info, 0, sizeof(p->sched_info));
3058 #if defined(CONFIG_SMP)
3061 #ifdef CONFIG_PREEMPT_COUNT
3062 /* Want to start with kernel preemption disabled. */
3063 task_thread_info(p)->preempt_count = 1;
3066 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3073 * wake_up_new_task - wake up a newly created task for the first time.
3075 * This function will do some initial scheduler statistics housekeeping
3076 * that must be done for every newly created context, then puts the task
3077 * on the runqueue and wakes it.
3079 void wake_up_new_task(struct task_struct *p)
3081 unsigned long flags;
3084 raw_spin_lock_irqsave(&p->pi_lock, flags);
3087 * Fork balancing, do it here and not earlier because:
3088 * - cpus_allowed can change in the fork path
3089 * - any previously selected cpu might disappear through hotplug
3091 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3094 rq = __task_rq_lock(p);
3095 activate_task(rq, p, 0);
3097 trace_sched_wakeup_new(p, true);
3098 check_preempt_curr(rq, p, WF_FORK);
3100 if (p->sched_class->task_woken)
3101 p->sched_class->task_woken(rq, p);
3103 task_rq_unlock(rq, p, &flags);
3106 #ifdef CONFIG_PREEMPT_NOTIFIERS
3109 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3110 * @notifier: notifier struct to register
3112 void preempt_notifier_register(struct preempt_notifier *notifier)
3114 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3116 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3119 * preempt_notifier_unregister - no longer interested in preemption notifications
3120 * @notifier: notifier struct to unregister
3122 * This is safe to call from within a preemption notifier.
3124 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3126 hlist_del(¬ifier->link);
3128 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3130 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3132 struct preempt_notifier *notifier;
3133 struct hlist_node *node;
3135 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3136 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3140 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3141 struct task_struct *next)
3143 struct preempt_notifier *notifier;
3144 struct hlist_node *node;
3146 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3147 notifier->ops->sched_out(notifier, next);
3150 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3152 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3157 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3158 struct task_struct *next)
3162 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3165 * prepare_task_switch - prepare to switch tasks
3166 * @rq: the runqueue preparing to switch
3167 * @prev: the current task that is being switched out
3168 * @next: the task we are going to switch to.
3170 * This is called with the rq lock held and interrupts off. It must
3171 * be paired with a subsequent finish_task_switch after the context
3174 * prepare_task_switch sets up locking and calls architecture specific
3178 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3179 struct task_struct *next)
3181 sched_info_switch(prev, next);
3182 perf_event_task_sched_out(prev, next);
3183 fire_sched_out_preempt_notifiers(prev, next);
3184 prepare_lock_switch(rq, next);
3185 prepare_arch_switch(next);
3186 trace_sched_switch(prev, next);
3190 * finish_task_switch - clean up after a task-switch
3191 * @rq: runqueue associated with task-switch
3192 * @prev: the thread we just switched away from.
3194 * finish_task_switch must be called after the context switch, paired
3195 * with a prepare_task_switch call before the context switch.
3196 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3197 * and do any other architecture-specific cleanup actions.
3199 * Note that we may have delayed dropping an mm in context_switch(). If
3200 * so, we finish that here outside of the runqueue lock. (Doing it
3201 * with the lock held can cause deadlocks; see schedule() for
3204 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3205 __releases(rq->lock)
3207 struct mm_struct *mm = rq->prev_mm;
3213 * A task struct has one reference for the use as "current".
3214 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3215 * schedule one last time. The schedule call will never return, and
3216 * the scheduled task must drop that reference.
3218 * We must observe prev->state before clearing prev->on_cpu (in
3219 * finish_lock_switch), otherwise a concurrent wakeup can get prev
3220 * running on another CPU and we could rave with its RUNNING -> DEAD
3221 * transition, resulting in a double drop.
3223 prev_state = prev->state;
3224 finish_arch_switch(prev);
3225 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3226 local_irq_disable();
3227 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3228 perf_event_task_sched_in(prev, current);
3229 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3231 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3232 finish_lock_switch(rq, prev);
3234 fire_sched_in_preempt_notifiers(current);
3237 if (unlikely(prev_state == TASK_DEAD)) {
3239 * Remove function-return probe instances associated with this
3240 * task and put them back on the free list.
3242 kprobe_flush_task(prev);
3243 put_task_struct(prev);
3249 /* assumes rq->lock is held */
3250 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3252 if (prev->sched_class->pre_schedule)
3253 prev->sched_class->pre_schedule(rq, prev);
3256 /* rq->lock is NOT held, but preemption is disabled */
3257 static inline void post_schedule(struct rq *rq)
3259 if (rq->post_schedule) {
3260 unsigned long flags;
3262 raw_spin_lock_irqsave(&rq->lock, flags);
3263 if (rq->curr->sched_class->post_schedule)
3264 rq->curr->sched_class->post_schedule(rq);
3265 raw_spin_unlock_irqrestore(&rq->lock, flags);
3267 rq->post_schedule = 0;
3273 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3277 static inline void post_schedule(struct rq *rq)
3284 * schedule_tail - first thing a freshly forked thread must call.
3285 * @prev: the thread we just switched away from.
3287 asmlinkage void schedule_tail(struct task_struct *prev)
3288 __releases(rq->lock)
3290 struct rq *rq = this_rq();
3292 finish_task_switch(rq, prev);
3295 * FIXME: do we need to worry about rq being invalidated by the
3300 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3301 /* In this case, finish_task_switch does not reenable preemption */
3304 if (current->set_child_tid)
3305 put_user(task_pid_vnr(current), current->set_child_tid);
3309 * context_switch - switch to the new MM and the new
3310 * thread's register state.
3313 context_switch(struct rq *rq, struct task_struct *prev,
3314 struct task_struct *next)
3316 struct mm_struct *mm, *oldmm;
3318 prepare_task_switch(rq, prev, next);
3321 oldmm = prev->active_mm;
3323 * For paravirt, this is coupled with an exit in switch_to to
3324 * combine the page table reload and the switch backend into
3327 arch_start_context_switch(prev);
3330 next->active_mm = oldmm;
3331 atomic_inc(&oldmm->mm_count);
3332 enter_lazy_tlb(oldmm, next);
3334 switch_mm(oldmm, mm, next);
3337 prev->active_mm = NULL;
3338 rq->prev_mm = oldmm;
3341 * Since the runqueue lock will be released by the next
3342 * task (which is an invalid locking op but in the case
3343 * of the scheduler it's an obvious special-case), so we
3344 * do an early lockdep release here:
3346 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3347 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3350 /* Here we just switch the register state and the stack. */
3351 switch_to(prev, next, prev);
3355 * this_rq must be evaluated again because prev may have moved
3356 * CPUs since it called schedule(), thus the 'rq' on its stack
3357 * frame will be invalid.
3359 finish_task_switch(this_rq(), prev);
3363 * nr_running, nr_uninterruptible and nr_context_switches:
3365 * externally visible scheduler statistics: current number of runnable
3366 * threads, current number of uninterruptible-sleeping threads, total
3367 * number of context switches performed since bootup.
3369 unsigned long nr_running(void)
3371 unsigned long i, sum = 0;
3373 for_each_online_cpu(i)
3374 sum += cpu_rq(i)->nr_running;
3379 unsigned long nr_uninterruptible(void)
3381 unsigned long i, sum = 0;
3383 for_each_possible_cpu(i)
3384 sum += cpu_rq(i)->nr_uninterruptible;
3387 * Since we read the counters lockless, it might be slightly
3388 * inaccurate. Do not allow it to go below zero though:
3390 if (unlikely((long)sum < 0))
3396 unsigned long long nr_context_switches(void)
3399 unsigned long long sum = 0;
3401 for_each_possible_cpu(i)
3402 sum += cpu_rq(i)->nr_switches;
3407 unsigned long nr_iowait(void)
3409 unsigned long i, sum = 0;
3411 for_each_possible_cpu(i)
3412 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3417 unsigned long nr_iowait_cpu(int cpu)
3419 struct rq *this = cpu_rq(cpu);
3420 return atomic_read(&this->nr_iowait);
3423 unsigned long this_cpu_load(void)
3425 struct rq *this = this_rq();
3426 return this->cpu_load[0];
3431 * Global load-average calculations
3433 * We take a distributed and async approach to calculating the global load-avg
3434 * in order to minimize overhead.
3436 * The global load average is an exponentially decaying average of nr_running +
3437 * nr_uninterruptible.
3439 * Once every LOAD_FREQ:
3442 * for_each_possible_cpu(cpu)
3443 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
3445 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
3447 * Due to a number of reasons the above turns in the mess below:
3449 * - for_each_possible_cpu() is prohibitively expensive on machines with
3450 * serious number of cpus, therefore we need to take a distributed approach
3451 * to calculating nr_active.
3453 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
3454 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
3456 * So assuming nr_active := 0 when we start out -- true per definition, we
3457 * can simply take per-cpu deltas and fold those into a global accumulate
3458 * to obtain the same result. See calc_load_fold_active().
3460 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
3461 * across the machine, we assume 10 ticks is sufficient time for every
3462 * cpu to have completed this task.
3464 * This places an upper-bound on the IRQ-off latency of the machine. Then
3465 * again, being late doesn't loose the delta, just wrecks the sample.
3467 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
3468 * this would add another cross-cpu cacheline miss and atomic operation
3469 * to the wakeup path. Instead we increment on whatever cpu the task ran
3470 * when it went into uninterruptible state and decrement on whatever cpu
3471 * did the wakeup. This means that only the sum of nr_uninterruptible over
3472 * all cpus yields the correct result.
3474 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
3477 /* Variables and functions for calc_load */
3478 static atomic_long_t calc_load_tasks;
3479 static unsigned long calc_load_update;
3480 unsigned long avenrun[3];
3481 EXPORT_SYMBOL(avenrun); /* should be removed */
3484 * get_avenrun - get the load average array
3485 * @loads: pointer to dest load array
3486 * @offset: offset to add
3487 * @shift: shift count to shift the result left
3489 * These values are estimates at best, so no need for locking.
3491 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3493 loads[0] = (avenrun[0] + offset) << shift;
3494 loads[1] = (avenrun[1] + offset) << shift;
3495 loads[2] = (avenrun[2] + offset) << shift;
3498 static long calc_load_fold_active(struct rq *this_rq)
3500 long nr_active, delta = 0;
3502 nr_active = this_rq->nr_running;
3503 nr_active += (long) this_rq->nr_uninterruptible;
3505 if (nr_active != this_rq->calc_load_active) {
3506 delta = nr_active - this_rq->calc_load_active;
3507 this_rq->calc_load_active = nr_active;
3514 * a1 = a0 * e + a * (1 - e)
3516 static unsigned long
3517 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3519 unsigned long newload;
3521 newload = load * exp + active * (FIXED_1 - exp);
3523 newload += FIXED_1-1;
3525 return newload / FIXED_1;
3530 * Handle NO_HZ for the global load-average.
3532 * Since the above described distributed algorithm to compute the global
3533 * load-average relies on per-cpu sampling from the tick, it is affected by
3536 * The basic idea is to fold the nr_active delta into a global idle-delta upon
3537 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
3538 * when we read the global state.
3540 * Obviously reality has to ruin such a delightfully simple scheme:
3542 * - When we go NO_HZ idle during the window, we can negate our sample
3543 * contribution, causing under-accounting.
3545 * We avoid this by keeping two idle-delta counters and flipping them
3546 * when the window starts, thus separating old and new NO_HZ load.
3548 * The only trick is the slight shift in index flip for read vs write.
3552 * |-|-----------|-|-----------|-|-----------|-|
3553 * r:0 0 1 1 0 0 1 1 0
3554 * w:0 1 1 0 0 1 1 0 0
3556 * This ensures we'll fold the old idle contribution in this window while
3557 * accumlating the new one.
3559 * - When we wake up from NO_HZ idle during the window, we push up our
3560 * contribution, since we effectively move our sample point to a known
3563 * This is solved by pushing the window forward, and thus skipping the
3564 * sample, for this cpu (effectively using the idle-delta for this cpu which
3565 * was in effect at the time the window opened). This also solves the issue
3566 * of having to deal with a cpu having been in NOHZ idle for multiple
3567 * LOAD_FREQ intervals.
3569 * When making the ILB scale, we should try to pull this in as well.
3571 static atomic_long_t calc_load_idle[2];
3572 static int calc_load_idx;
3574 static inline int calc_load_write_idx(void)
3576 int idx = calc_load_idx;
3579 * See calc_global_nohz(), if we observe the new index, we also
3580 * need to observe the new update time.
3585 * If the folding window started, make sure we start writing in the
3588 if (!time_before(jiffies, calc_load_update))
3594 static inline int calc_load_read_idx(void)
3596 return calc_load_idx & 1;
3599 void calc_load_enter_idle(void)
3601 struct rq *this_rq = this_rq();
3605 * We're going into NOHZ mode, if there's any pending delta, fold it
3606 * into the pending idle delta.
3608 delta = calc_load_fold_active(this_rq);
3610 int idx = calc_load_write_idx();
3611 atomic_long_add(delta, &calc_load_idle[idx]);
3615 void calc_load_exit_idle(void)
3617 struct rq *this_rq = this_rq();
3620 * If we're still before the sample window, we're done.
3622 if (time_before(jiffies, this_rq->calc_load_update))
3626 * We woke inside or after the sample window, this means we're already
3627 * accounted through the nohz accounting, so skip the entire deal and
3628 * sync up for the next window.
3630 this_rq->calc_load_update = calc_load_update;
3631 if (time_before(jiffies, this_rq->calc_load_update + 10))
3632 this_rq->calc_load_update += LOAD_FREQ;
3635 static long calc_load_fold_idle(void)
3637 int idx = calc_load_read_idx();
3640 if (atomic_long_read(&calc_load_idle[idx]))
3641 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
3647 * fixed_power_int - compute: x^n, in O(log n) time
3649 * @x: base of the power
3650 * @frac_bits: fractional bits of @x
3651 * @n: power to raise @x to.
3653 * By exploiting the relation between the definition of the natural power
3654 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3655 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3656 * (where: n_i \elem {0, 1}, the binary vector representing n),
3657 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3658 * of course trivially computable in O(log_2 n), the length of our binary
3661 static unsigned long
3662 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3664 unsigned long result = 1UL << frac_bits;
3669 result += 1UL << (frac_bits - 1);
3670 result >>= frac_bits;
3676 x += 1UL << (frac_bits - 1);
3684 * a1 = a0 * e + a * (1 - e)
3686 * a2 = a1 * e + a * (1 - e)
3687 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3688 * = a0 * e^2 + a * (1 - e) * (1 + e)
3690 * a3 = a2 * e + a * (1 - e)
3691 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3692 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3696 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3697 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3698 * = a0 * e^n + a * (1 - e^n)
3700 * [1] application of the geometric series:
3703 * S_n := \Sum x^i = -------------
3706 static unsigned long
3707 calc_load_n(unsigned long load, unsigned long exp,
3708 unsigned long active, unsigned int n)
3711 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3715 * NO_HZ can leave us missing all per-cpu ticks calling
3716 * calc_load_account_active(), but since an idle CPU folds its delta into
3717 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3718 * in the pending idle delta if our idle period crossed a load cycle boundary.
3720 * Once we've updated the global active value, we need to apply the exponential
3721 * weights adjusted to the number of cycles missed.
3723 static void calc_global_nohz(void)
3725 long delta, active, n;
3727 if (!time_before(jiffies, calc_load_update + 10)) {
3729 * Catch-up, fold however many we are behind still
3731 delta = jiffies - calc_load_update - 10;
3732 n = 1 + (delta / LOAD_FREQ);
3734 active = atomic_long_read(&calc_load_tasks);
3735 active = active > 0 ? active * FIXED_1 : 0;
3737 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3738 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3739 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3741 calc_load_update += n * LOAD_FREQ;
3745 * Flip the idle index...
3747 * Make sure we first write the new time then flip the index, so that
3748 * calc_load_write_idx() will see the new time when it reads the new
3749 * index, this avoids a double flip messing things up.
3754 #else /* !CONFIG_NO_HZ */
3756 static inline long calc_load_fold_idle(void) { return 0; }
3757 static inline void calc_global_nohz(void) { }
3759 #endif /* CONFIG_NO_HZ */
3762 * calc_load - update the avenrun load estimates 10 ticks after the
3763 * CPUs have updated calc_load_tasks.
3765 void calc_global_load(unsigned long ticks)
3769 if (time_before(jiffies, calc_load_update + 10))
3773 * Fold the 'old' idle-delta to include all NO_HZ cpus.
3775 delta = calc_load_fold_idle();
3777 atomic_long_add(delta, &calc_load_tasks);
3779 active = atomic_long_read(&calc_load_tasks);
3780 active = active > 0 ? active * FIXED_1 : 0;
3782 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3783 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3784 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3786 calc_load_update += LOAD_FREQ;
3789 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
3795 * Called from update_cpu_load() to periodically update this CPU's
3798 static void calc_load_account_active(struct rq *this_rq)
3802 if (time_before(jiffies, this_rq->calc_load_update))
3805 delta = calc_load_fold_active(this_rq);
3807 atomic_long_add(delta, &calc_load_tasks);
3809 this_rq->calc_load_update += LOAD_FREQ;
3813 * End of global load-average stuff
3817 * The exact cpuload at various idx values, calculated at every tick would be
3818 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3820 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3821 * on nth tick when cpu may be busy, then we have:
3822 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3823 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3825 * decay_load_missed() below does efficient calculation of
3826 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3827 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3829 * The calculation is approximated on a 128 point scale.
3830 * degrade_zero_ticks is the number of ticks after which load at any
3831 * particular idx is approximated to be zero.
3832 * degrade_factor is a precomputed table, a row for each load idx.
3833 * Each column corresponds to degradation factor for a power of two ticks,
3834 * based on 128 point scale.
3836 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3837 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3839 * With this power of 2 load factors, we can degrade the load n times
3840 * by looking at 1 bits in n and doing as many mult/shift instead of
3841 * n mult/shifts needed by the exact degradation.
3843 #define DEGRADE_SHIFT 7
3844 static const unsigned char
3845 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3846 static const unsigned char
3847 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3848 {0, 0, 0, 0, 0, 0, 0, 0},
3849 {64, 32, 8, 0, 0, 0, 0, 0},
3850 {96, 72, 40, 12, 1, 0, 0},
3851 {112, 98, 75, 43, 15, 1, 0},
3852 {120, 112, 98, 76, 45, 16, 2} };
3855 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3856 * would be when CPU is idle and so we just decay the old load without
3857 * adding any new load.
3859 static unsigned long
3860 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3864 if (!missed_updates)
3867 if (missed_updates >= degrade_zero_ticks[idx])
3871 return load >> missed_updates;
3873 while (missed_updates) {
3874 if (missed_updates % 2)
3875 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3877 missed_updates >>= 1;
3884 * Update rq->cpu_load[] statistics. This function is usually called every
3885 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3886 * every tick. We fix it up based on jiffies.
3888 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
3889 unsigned long pending_updates)
3893 this_rq->nr_load_updates++;
3895 /* Update our load: */
3896 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3897 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3898 unsigned long old_load, new_load;
3900 /* scale is effectively 1 << i now, and >> i divides by scale */
3902 old_load = this_rq->cpu_load[i];
3903 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3904 new_load = this_load;
3906 * Round up the averaging division if load is increasing. This
3907 * prevents us from getting stuck on 9 if the load is 10, for
3910 if (new_load > old_load)
3911 new_load += scale - 1;
3913 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3916 sched_avg_update(this_rq);
3921 * There is no sane way to deal with nohz on smp when using jiffies because the
3922 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
3923 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
3925 * Therefore we cannot use the delta approach from the regular tick since that
3926 * would seriously skew the load calculation. However we'll make do for those
3927 * updates happening while idle (nohz_idle_balance) or coming out of idle
3928 * (tick_nohz_idle_exit).
3930 * This means we might still be one tick off for nohz periods.
3934 * Called from nohz_idle_balance() to update the load ratings before doing the
3937 static void update_idle_cpu_load(struct rq *this_rq)
3939 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3940 unsigned long load = this_rq->load.weight;
3941 unsigned long pending_updates;
3944 * bail if there's load or we're actually up-to-date.
3946 if (load || curr_jiffies == this_rq->last_load_update_tick)
3949 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3950 this_rq->last_load_update_tick = curr_jiffies;
3952 __update_cpu_load(this_rq, load, pending_updates);
3956 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
3958 void update_cpu_load_nohz(void)
3960 struct rq *this_rq = this_rq();
3961 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3962 unsigned long pending_updates;
3964 if (curr_jiffies == this_rq->last_load_update_tick)
3967 raw_spin_lock(&this_rq->lock);
3968 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3969 if (pending_updates) {
3970 this_rq->last_load_update_tick = curr_jiffies;
3972 * We were idle, this means load 0, the current load might be
3973 * !0 due to remote wakeups and the sort.
3975 __update_cpu_load(this_rq, 0, pending_updates);
3977 raw_spin_unlock(&this_rq->lock);
3979 #endif /* CONFIG_NO_HZ */
3982 * Called from scheduler_tick()
3984 static void update_cpu_load_active(struct rq *this_rq)
3987 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
3989 this_rq->last_load_update_tick = jiffies;
3990 __update_cpu_load(this_rq, this_rq->load.weight, 1);
3992 calc_load_account_active(this_rq);
3998 * sched_exec - execve() is a valuable balancing opportunity, because at
3999 * this point the task has the smallest effective memory and cache footprint.
4001 void sched_exec(void)
4003 struct task_struct *p = current;
4004 unsigned long flags;
4007 raw_spin_lock_irqsave(&p->pi_lock, flags);
4008 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
4009 if (dest_cpu == smp_processor_id())
4012 if (likely(cpu_active(dest_cpu))) {
4013 struct migration_arg arg = { p, dest_cpu };
4015 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4016 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
4020 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4025 DEFINE_PER_CPU(struct kernel_stat, kstat);
4027 EXPORT_PER_CPU_SYMBOL(kstat);
4030 * Return any ns on the sched_clock that have not yet been accounted in
4031 * @p in case that task is currently running.
4033 * Called with task_rq_lock() held on @rq.
4035 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4039 if (task_current(rq, p)) {
4040 update_rq_clock(rq);
4041 ns = rq->clock_task - p->se.exec_start;
4049 unsigned long long task_delta_exec(struct task_struct *p)
4051 unsigned long flags;
4055 rq = task_rq_lock(p, &flags);
4056 ns = do_task_delta_exec(p, rq);
4057 task_rq_unlock(rq, p, &flags);
4063 * Return accounted runtime for the task.
4064 * In case the task is currently running, return the runtime plus current's
4065 * pending runtime that have not been accounted yet.
4067 unsigned long long task_sched_runtime(struct task_struct *p)
4069 unsigned long flags;
4073 rq = task_rq_lock(p, &flags);
4074 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4075 task_rq_unlock(rq, p, &flags);
4081 * Account user cpu time to a process.
4082 * @p: the process that the cpu time gets accounted to
4083 * @cputime: the cpu time spent in user space since the last update
4084 * @cputime_scaled: cputime scaled by cpu frequency
4086 void account_user_time(struct task_struct *p, cputime_t cputime,
4087 cputime_t cputime_scaled)
4089 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4092 /* Add user time to process. */
4093 p->utime = cputime_add(p->utime, cputime);
4094 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4095 account_group_user_time(p, cputime);
4097 /* Add user time to cpustat. */
4098 tmp = cputime_to_cputime64(cputime);
4099 if (TASK_NICE(p) > 0)
4100 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4102 cpustat->user = cputime64_add(cpustat->user, tmp);
4104 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4105 /* Account for user time used */
4106 acct_update_integrals(p);
4110 * Account guest cpu time to a process.
4111 * @p: the process that the cpu time gets accounted to
4112 * @cputime: the cpu time spent in virtual machine since the last update
4113 * @cputime_scaled: cputime scaled by cpu frequency
4115 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4116 cputime_t cputime_scaled)
4119 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4121 tmp = cputime_to_cputime64(cputime);
4123 /* Add guest time to process. */
4124 p->utime = cputime_add(p->utime, cputime);
4125 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4126 account_group_user_time(p, cputime);
4127 p->gtime = cputime_add(p->gtime, cputime);
4129 /* Add guest time to cpustat. */
4130 if (TASK_NICE(p) > 0) {
4131 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4132 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
4134 cpustat->user = cputime64_add(cpustat->user, tmp);
4135 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4140 * Account system cpu time to a process and desired cpustat field
4141 * @p: the process that the cpu time gets accounted to
4142 * @cputime: the cpu time spent in kernel space since the last update
4143 * @cputime_scaled: cputime scaled by cpu frequency
4144 * @target_cputime64: pointer to cpustat field that has to be updated
4147 void __account_system_time(struct task_struct *p, cputime_t cputime,
4148 cputime_t cputime_scaled, cputime64_t *target_cputime64)
4150 cputime64_t tmp = cputime_to_cputime64(cputime);
4152 /* Add system time to process. */
4153 p->stime = cputime_add(p->stime, cputime);
4154 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4155 account_group_system_time(p, cputime);
4157 /* Add system time to cpustat. */
4158 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
4159 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4161 /* Account for system time used */
4162 acct_update_integrals(p);
4166 * Account system cpu time to a process.
4167 * @p: the process that the cpu time gets accounted to
4168 * @hardirq_offset: the offset to subtract from hardirq_count()
4169 * @cputime: the cpu time spent in kernel space since the last update
4170 * @cputime_scaled: cputime scaled by cpu frequency
4172 void account_system_time(struct task_struct *p, int hardirq_offset,
4173 cputime_t cputime, cputime_t cputime_scaled)
4175 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4176 cputime64_t *target_cputime64;
4178 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4179 account_guest_time(p, cputime, cputime_scaled);
4183 if (hardirq_count() - hardirq_offset)
4184 target_cputime64 = &cpustat->irq;
4185 else if (in_serving_softirq())
4186 target_cputime64 = &cpustat->softirq;
4188 target_cputime64 = &cpustat->system;
4190 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
4194 * Account for involuntary wait time.
4195 * @cputime: the cpu time spent in involuntary wait
4197 void account_steal_time(cputime_t cputime)
4199 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4200 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4202 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4206 * Account for idle time.
4207 * @cputime: the cpu time spent in idle wait
4209 void account_idle_time(cputime_t cputime)
4211 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4212 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4213 struct rq *rq = this_rq();
4215 if (atomic_read(&rq->nr_iowait) > 0)
4216 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4218 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4221 static __always_inline bool steal_account_process_tick(void)
4223 #ifdef CONFIG_PARAVIRT
4224 if (static_branch(¶virt_steal_enabled)) {
4227 steal = paravirt_steal_clock(smp_processor_id());
4228 steal -= this_rq()->prev_steal_time;
4230 st = steal_ticks(steal);
4231 this_rq()->prev_steal_time += st * TICK_NSEC;
4233 account_steal_time(st);
4240 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4242 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4244 * Account a tick to a process and cpustat
4245 * @p: the process that the cpu time gets accounted to
4246 * @user_tick: is the tick from userspace
4247 * @rq: the pointer to rq
4249 * Tick demultiplexing follows the order
4250 * - pending hardirq update
4251 * - pending softirq update
4255 * - check for guest_time
4256 * - else account as system_time
4258 * Check for hardirq is done both for system and user time as there is
4259 * no timer going off while we are on hardirq and hence we may never get an
4260 * opportunity to update it solely in system time.
4261 * p->stime and friends are only updated on system time and not on irq
4262 * softirq as those do not count in task exec_runtime any more.
4264 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4267 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4268 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
4269 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4271 if (steal_account_process_tick())
4274 if (irqtime_account_hi_update()) {
4275 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4276 } else if (irqtime_account_si_update()) {
4277 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4278 } else if (this_cpu_ksoftirqd() == p) {
4280 * ksoftirqd time do not get accounted in cpu_softirq_time.
4281 * So, we have to handle it separately here.
4282 * Also, p->stime needs to be updated for ksoftirqd.
4284 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4286 } else if (user_tick) {
4287 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4288 } else if (p == rq->idle) {
4289 account_idle_time(cputime_one_jiffy);
4290 } else if (p->flags & PF_VCPU) { /* System time or guest time */
4291 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
4293 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4298 static void irqtime_account_idle_ticks(int ticks)
4301 struct rq *rq = this_rq();
4303 for (i = 0; i < ticks; i++)
4304 irqtime_account_process_tick(current, 0, rq);
4306 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4307 static void irqtime_account_idle_ticks(int ticks) {}
4308 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4310 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4313 * Account a single tick of cpu time.
4314 * @p: the process that the cpu time gets accounted to
4315 * @user_tick: indicates if the tick is a user or a system tick
4317 void account_process_tick(struct task_struct *p, int user_tick)
4319 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4320 struct rq *rq = this_rq();
4322 if (sched_clock_irqtime) {
4323 irqtime_account_process_tick(p, user_tick, rq);
4327 if (steal_account_process_tick())
4331 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4332 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4333 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4336 account_idle_time(cputime_one_jiffy);
4340 * Account multiple ticks of steal time.
4341 * @p: the process from which the cpu time has been stolen
4342 * @ticks: number of stolen ticks
4344 void account_steal_ticks(unsigned long ticks)
4346 account_steal_time(jiffies_to_cputime(ticks));
4350 * Account multiple ticks of idle time.
4351 * @ticks: number of stolen ticks
4353 void account_idle_ticks(unsigned long ticks)
4356 if (sched_clock_irqtime) {
4357 irqtime_account_idle_ticks(ticks);
4361 account_idle_time(jiffies_to_cputime(ticks));
4367 * Use precise platform statistics if available:
4369 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4370 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4376 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4378 struct task_cputime cputime;
4380 thread_group_cputime(p, &cputime);
4382 *ut = cputime.utime;
4383 *st = cputime.stime;
4387 #ifndef nsecs_to_cputime
4388 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4391 static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total)
4393 u64 temp = (__force u64) rtime;
4395 temp *= (__force u64) utime;
4397 if (sizeof(cputime_t) == 4)
4398 temp = div_u64(temp, (__force u32) total);
4400 temp = div64_u64(temp, (__force u64) total);
4402 return (__force cputime_t) temp;
4405 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4407 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4410 * Use CFS's precise accounting:
4412 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4415 utime = scale_utime(utime, rtime, total);
4420 * Compare with previous values, to keep monotonicity:
4422 p->prev_utime = max(p->prev_utime, utime);
4423 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4425 *ut = p->prev_utime;
4426 *st = p->prev_stime;
4430 * Must be called with siglock held.
4432 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4434 struct signal_struct *sig = p->signal;
4435 struct task_cputime cputime;
4436 cputime_t rtime, utime, total;
4438 thread_group_cputime(p, &cputime);
4440 total = cputime_add(cputime.utime, cputime.stime);
4441 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4444 utime = scale_utime(cputime.utime, rtime, total);
4448 sig->prev_utime = max(sig->prev_utime, utime);
4449 sig->prev_stime = max(sig->prev_stime,
4450 cputime_sub(rtime, sig->prev_utime));
4452 *ut = sig->prev_utime;
4453 *st = sig->prev_stime;
4458 * This function gets called by the timer code, with HZ frequency.
4459 * We call it with interrupts disabled.
4461 void scheduler_tick(void)
4463 int cpu = smp_processor_id();
4464 struct rq *rq = cpu_rq(cpu);
4465 struct task_struct *curr = rq->curr;
4469 raw_spin_lock(&rq->lock);
4470 update_rq_clock(rq);
4471 update_cpu_load_active(rq);
4472 curr->sched_class->task_tick(rq, curr, 0);
4473 raw_spin_unlock(&rq->lock);
4475 perf_event_task_tick();
4478 rq->idle_balance = idle_cpu(cpu);
4479 trigger_load_balance(rq, cpu);
4483 notrace unsigned long get_parent_ip(unsigned long addr)
4485 if (in_lock_functions(addr)) {
4486 addr = CALLER_ADDR2;
4487 if (in_lock_functions(addr))
4488 addr = CALLER_ADDR3;
4493 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4494 defined(CONFIG_PREEMPT_TRACER))
4496 void __kprobes add_preempt_count(int val)
4498 #ifdef CONFIG_DEBUG_PREEMPT
4502 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4505 preempt_count() += val;
4506 #ifdef CONFIG_DEBUG_PREEMPT
4508 * Spinlock count overflowing soon?
4510 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4513 if (preempt_count() == val)
4514 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4516 EXPORT_SYMBOL(add_preempt_count);
4518 void __kprobes sub_preempt_count(int val)
4520 #ifdef CONFIG_DEBUG_PREEMPT
4524 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4527 * Is the spinlock portion underflowing?
4529 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4530 !(preempt_count() & PREEMPT_MASK)))
4534 if (preempt_count() == val)
4535 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4536 preempt_count() -= val;
4538 EXPORT_SYMBOL(sub_preempt_count);
4543 * Print scheduling while atomic bug:
4545 static noinline void __schedule_bug(struct task_struct *prev)
4547 struct pt_regs *regs = get_irq_regs();
4549 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4550 prev->comm, prev->pid, preempt_count());
4552 debug_show_held_locks(prev);
4554 if (irqs_disabled())
4555 print_irqtrace_events(prev);
4564 * Various schedule()-time debugging checks and statistics:
4566 static inline void schedule_debug(struct task_struct *prev)
4569 * Test if we are atomic. Since do_exit() needs to call into
4570 * schedule() atomically, we ignore that path for now.
4571 * Otherwise, whine if we are scheduling when we should not be.
4573 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4574 __schedule_bug(prev);
4577 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4579 schedstat_inc(this_rq(), sched_count);
4582 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4584 if (prev->on_rq || rq->skip_clock_update < 0)
4585 update_rq_clock(rq);
4586 prev->sched_class->put_prev_task(rq, prev);
4590 * Pick up the highest-prio task:
4592 static inline struct task_struct *
4593 pick_next_task(struct rq *rq)
4595 const struct sched_class *class;
4596 struct task_struct *p;
4599 * Optimization: we know that if all tasks are in
4600 * the fair class we can call that function directly:
4602 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4603 p = fair_sched_class.pick_next_task(rq);
4608 for_each_class(class) {
4609 p = class->pick_next_task(rq);
4614 BUG(); /* the idle class will always have a runnable task */
4618 * __schedule() is the main scheduler function.
4620 static void __sched __schedule(void)
4622 struct task_struct *prev, *next;
4623 unsigned long *switch_count;
4629 cpu = smp_processor_id();
4631 rcu_note_context_switch(cpu);
4634 schedule_debug(prev);
4636 if (sched_feat(HRTICK))
4639 raw_spin_lock_irq(&rq->lock);
4641 switch_count = &prev->nivcsw;
4642 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4643 if (unlikely(signal_pending_state(prev->state, prev))) {
4644 prev->state = TASK_RUNNING;
4646 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4650 * If a worker went to sleep, notify and ask workqueue
4651 * whether it wants to wake up a task to maintain
4654 if (prev->flags & PF_WQ_WORKER) {
4655 struct task_struct *to_wakeup;
4657 to_wakeup = wq_worker_sleeping(prev, cpu);
4659 try_to_wake_up_local(to_wakeup);
4662 switch_count = &prev->nvcsw;
4665 pre_schedule(rq, prev);
4667 if (unlikely(!rq->nr_running))
4668 idle_balance(cpu, rq);
4670 put_prev_task(rq, prev);
4671 next = pick_next_task(rq);
4672 clear_tsk_need_resched(prev);
4673 rq->skip_clock_update = 0;
4675 if (likely(prev != next)) {
4680 context_switch(rq, prev, next); /* unlocks the rq */
4682 * The context switch have flipped the stack from under us
4683 * and restored the local variables which were saved when
4684 * this task called schedule() in the past. prev == current
4685 * is still correct, but it can be moved to another cpu/rq.
4687 cpu = smp_processor_id();
4690 raw_spin_unlock_irq(&rq->lock);
4694 preempt_enable_no_resched();
4699 static inline void sched_submit_work(struct task_struct *tsk)
4704 * If we are going to sleep and we have plugged IO queued,
4705 * make sure to submit it to avoid deadlocks.
4707 if (blk_needs_flush_plug(tsk))
4708 blk_schedule_flush_plug(tsk);
4711 asmlinkage void __sched schedule(void)
4713 struct task_struct *tsk = current;
4715 sched_submit_work(tsk);
4718 EXPORT_SYMBOL(schedule);
4720 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4722 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4724 if (lock->owner != owner)
4728 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4729 * lock->owner still matches owner, if that fails, owner might
4730 * point to free()d memory, if it still matches, the rcu_read_lock()
4731 * ensures the memory stays valid.
4735 return owner->on_cpu;
4739 * Look out! "owner" is an entirely speculative pointer
4740 * access and not reliable.
4742 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4744 if (!sched_feat(OWNER_SPIN))
4748 while (owner_running(lock, owner)) {
4752 arch_mutex_cpu_relax();
4757 * We break out the loop above on need_resched() and when the
4758 * owner changed, which is a sign for heavy contention. Return
4759 * success only when lock->owner is NULL.
4761 return lock->owner == NULL;
4765 #ifdef CONFIG_PREEMPT
4767 * this is the entry point to schedule() from in-kernel preemption
4768 * off of preempt_enable. Kernel preemptions off return from interrupt
4769 * occur there and call schedule directly.
4771 asmlinkage void __sched notrace preempt_schedule(void)
4773 struct thread_info *ti = current_thread_info();
4776 * If there is a non-zero preempt_count or interrupts are disabled,
4777 * we do not want to preempt the current task. Just return..
4779 if (likely(ti->preempt_count || irqs_disabled()))
4783 add_preempt_count_notrace(PREEMPT_ACTIVE);
4785 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4788 * Check again in case we missed a preemption opportunity
4789 * between schedule and now.
4792 } while (need_resched());
4794 EXPORT_SYMBOL(preempt_schedule);
4797 * this is the entry point to schedule() from kernel preemption
4798 * off of irq context.
4799 * Note, that this is called and return with irqs disabled. This will
4800 * protect us against recursive calling from irq.
4802 asmlinkage void __sched preempt_schedule_irq(void)
4804 struct thread_info *ti = current_thread_info();
4806 /* Catch callers which need to be fixed */
4807 BUG_ON(ti->preempt_count || !irqs_disabled());
4810 add_preempt_count(PREEMPT_ACTIVE);
4813 local_irq_disable();
4814 sub_preempt_count(PREEMPT_ACTIVE);
4817 * Check again in case we missed a preemption opportunity
4818 * between schedule and now.
4821 } while (need_resched());
4824 #endif /* CONFIG_PREEMPT */
4826 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4829 return try_to_wake_up(curr->private, mode, wake_flags);
4831 EXPORT_SYMBOL(default_wake_function);
4834 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4835 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4836 * number) then we wake all the non-exclusive tasks and one exclusive task.
4838 * There are circumstances in which we can try to wake a task which has already
4839 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4840 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4842 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4843 int nr_exclusive, int wake_flags, void *key)
4845 wait_queue_t *curr, *next;
4847 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4848 unsigned flags = curr->flags;
4850 if (curr->func(curr, mode, wake_flags, key) &&
4851 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4857 * __wake_up - wake up threads blocked on a waitqueue.
4859 * @mode: which threads
4860 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4861 * @key: is directly passed to the wakeup function
4863 * It may be assumed that this function implies a write memory barrier before
4864 * changing the task state if and only if any tasks are woken up.
4866 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4867 int nr_exclusive, void *key)
4869 unsigned long flags;
4871 spin_lock_irqsave(&q->lock, flags);
4872 __wake_up_common(q, mode, nr_exclusive, 0, key);
4873 spin_unlock_irqrestore(&q->lock, flags);
4875 EXPORT_SYMBOL(__wake_up);
4878 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4880 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4882 __wake_up_common(q, mode, 1, 0, NULL);
4884 EXPORT_SYMBOL_GPL(__wake_up_locked);
4886 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4888 __wake_up_common(q, mode, 1, 0, key);
4890 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4893 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4895 * @mode: which threads
4896 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4897 * @key: opaque value to be passed to wakeup targets
4899 * The sync wakeup differs that the waker knows that it will schedule
4900 * away soon, so while the target thread will be woken up, it will not
4901 * be migrated to another CPU - ie. the two threads are 'synchronized'
4902 * with each other. This can prevent needless bouncing between CPUs.
4904 * On UP it can prevent extra preemption.
4906 * It may be assumed that this function implies a write memory barrier before
4907 * changing the task state if and only if any tasks are woken up.
4909 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4910 int nr_exclusive, void *key)
4912 unsigned long flags;
4913 int wake_flags = WF_SYNC;
4918 if (unlikely(!nr_exclusive))
4921 spin_lock_irqsave(&q->lock, flags);
4922 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4923 spin_unlock_irqrestore(&q->lock, flags);
4925 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4928 * __wake_up_sync - see __wake_up_sync_key()
4930 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4932 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4934 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4937 * complete: - signals a single thread waiting on this completion
4938 * @x: holds the state of this particular completion
4940 * This will wake up a single thread waiting on this completion. Threads will be
4941 * awakened in the same order in which they were queued.
4943 * See also complete_all(), wait_for_completion() and related routines.
4945 * It may be assumed that this function implies a write memory barrier before
4946 * changing the task state if and only if any tasks are woken up.
4948 void complete(struct completion *x)
4950 unsigned long flags;
4952 spin_lock_irqsave(&x->wait.lock, flags);
4954 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4955 spin_unlock_irqrestore(&x->wait.lock, flags);
4957 EXPORT_SYMBOL(complete);
4960 * complete_all: - signals all threads waiting on this completion
4961 * @x: holds the state of this particular completion
4963 * This will wake up all threads waiting on this particular completion event.
4965 * It may be assumed that this function implies a write memory barrier before
4966 * changing the task state if and only if any tasks are woken up.
4968 void complete_all(struct completion *x)
4970 unsigned long flags;
4972 spin_lock_irqsave(&x->wait.lock, flags);
4973 x->done += UINT_MAX/2;
4974 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4975 spin_unlock_irqrestore(&x->wait.lock, flags);
4977 EXPORT_SYMBOL(complete_all);
4979 static inline long __sched
4980 do_wait_for_common(struct completion *x, long timeout, int state)
4983 DECLARE_WAITQUEUE(wait, current);
4985 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4987 if (signal_pending_state(state, current)) {
4988 timeout = -ERESTARTSYS;
4991 __set_current_state(state);
4992 spin_unlock_irq(&x->wait.lock);
4993 timeout = schedule_timeout(timeout);
4994 spin_lock_irq(&x->wait.lock);
4995 } while (!x->done && timeout);
4996 __remove_wait_queue(&x->wait, &wait);
5001 return timeout ?: 1;
5005 wait_for_common(struct completion *x, long timeout, int state)
5009 spin_lock_irq(&x->wait.lock);
5010 timeout = do_wait_for_common(x, timeout, state);
5011 spin_unlock_irq(&x->wait.lock);
5016 * wait_for_completion: - waits for completion of a task
5017 * @x: holds the state of this particular completion
5019 * This waits to be signaled for completion of a specific task. It is NOT
5020 * interruptible and there is no timeout.
5022 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5023 * and interrupt capability. Also see complete().
5025 void __sched wait_for_completion(struct completion *x)
5027 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5029 EXPORT_SYMBOL(wait_for_completion);
5032 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5033 * @x: holds the state of this particular completion
5034 * @timeout: timeout value in jiffies
5036 * This waits for either a completion of a specific task to be signaled or for a
5037 * specified timeout to expire. The timeout is in jiffies. It is not
5040 * The return value is 0 if timed out, and positive (at least 1, or number of
5041 * jiffies left till timeout) if completed.
5043 unsigned long __sched
5044 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5046 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5048 EXPORT_SYMBOL(wait_for_completion_timeout);
5051 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5052 * @x: holds the state of this particular completion
5054 * This waits for completion of a specific task to be signaled. It is
5057 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
5059 int __sched wait_for_completion_interruptible(struct completion *x)
5061 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5062 if (t == -ERESTARTSYS)
5066 EXPORT_SYMBOL(wait_for_completion_interruptible);
5069 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5070 * @x: holds the state of this particular completion
5071 * @timeout: timeout value in jiffies
5073 * This waits for either a completion of a specific task to be signaled or for a
5074 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5076 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
5077 * positive (at least 1, or number of jiffies left till timeout) if completed.
5080 wait_for_completion_interruptible_timeout(struct completion *x,
5081 unsigned long timeout)
5083 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5085 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5088 * wait_for_completion_killable: - waits for completion of a task (killable)
5089 * @x: holds the state of this particular completion
5091 * This waits to be signaled for completion of a specific task. It can be
5092 * interrupted by a kill signal.
5094 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
5096 int __sched wait_for_completion_killable(struct completion *x)
5098 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5099 if (t == -ERESTARTSYS)
5103 EXPORT_SYMBOL(wait_for_completion_killable);
5106 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
5107 * @x: holds the state of this particular completion
5108 * @timeout: timeout value in jiffies
5110 * This waits for either a completion of a specific task to be
5111 * signaled or for a specified timeout to expire. It can be
5112 * interrupted by a kill signal. The timeout is in jiffies.
5114 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
5115 * positive (at least 1, or number of jiffies left till timeout) if completed.
5118 wait_for_completion_killable_timeout(struct completion *x,
5119 unsigned long timeout)
5121 return wait_for_common(x, timeout, TASK_KILLABLE);
5123 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
5126 * try_wait_for_completion - try to decrement a completion without blocking
5127 * @x: completion structure
5129 * Returns: 0 if a decrement cannot be done without blocking
5130 * 1 if a decrement succeeded.
5132 * If a completion is being used as a counting completion,
5133 * attempt to decrement the counter without blocking. This
5134 * enables us to avoid waiting if the resource the completion
5135 * is protecting is not available.
5137 bool try_wait_for_completion(struct completion *x)
5139 unsigned long flags;
5142 spin_lock_irqsave(&x->wait.lock, flags);
5147 spin_unlock_irqrestore(&x->wait.lock, flags);
5150 EXPORT_SYMBOL(try_wait_for_completion);
5153 * completion_done - Test to see if a completion has any waiters
5154 * @x: completion structure
5156 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5157 * 1 if there are no waiters.
5160 bool completion_done(struct completion *x)
5162 unsigned long flags;
5165 spin_lock_irqsave(&x->wait.lock, flags);
5168 spin_unlock_irqrestore(&x->wait.lock, flags);
5171 EXPORT_SYMBOL(completion_done);
5174 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5176 unsigned long flags;
5179 init_waitqueue_entry(&wait, current);
5181 __set_current_state(state);
5183 spin_lock_irqsave(&q->lock, flags);
5184 __add_wait_queue(q, &wait);
5185 spin_unlock(&q->lock);
5186 timeout = schedule_timeout(timeout);
5187 spin_lock_irq(&q->lock);
5188 __remove_wait_queue(q, &wait);
5189 spin_unlock_irqrestore(&q->lock, flags);
5194 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5196 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5198 EXPORT_SYMBOL(interruptible_sleep_on);
5201 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5203 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5205 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5207 void __sched sleep_on(wait_queue_head_t *q)
5209 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5211 EXPORT_SYMBOL(sleep_on);
5213 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5215 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5217 EXPORT_SYMBOL(sleep_on_timeout);
5219 #ifdef CONFIG_RT_MUTEXES
5222 * rt_mutex_setprio - set the current priority of a task
5224 * @prio: prio value (kernel-internal form)
5226 * This function changes the 'effective' priority of a task. It does
5227 * not touch ->normal_prio like __setscheduler().
5229 * Used by the rt_mutex code to implement priority inheritance logic.
5231 void rt_mutex_setprio(struct task_struct *p, int prio)
5233 int oldprio, on_rq, running;
5235 const struct sched_class *prev_class;
5237 BUG_ON(prio < 0 || prio > MAX_PRIO);
5239 rq = __task_rq_lock(p);
5241 trace_sched_pi_setprio(p, prio);
5243 prev_class = p->sched_class;
5245 running = task_current(rq, p);
5247 dequeue_task(rq, p, 0);
5249 p->sched_class->put_prev_task(rq, p);
5252 p->sched_class = &rt_sched_class;
5254 if (rt_prio(oldprio))
5256 p->sched_class = &fair_sched_class;
5262 p->sched_class->set_curr_task(rq);
5264 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
5266 check_class_changed(rq, p, prev_class, oldprio);
5267 __task_rq_unlock(rq);
5272 void set_user_nice(struct task_struct *p, long nice)
5274 int old_prio, delta, on_rq;
5275 unsigned long flags;
5278 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5281 * We have to be careful, if called from sys_setpriority(),
5282 * the task might be in the middle of scheduling on another CPU.
5284 rq = task_rq_lock(p, &flags);
5286 * The RT priorities are set via sched_setscheduler(), but we still
5287 * allow the 'normal' nice value to be set - but as expected
5288 * it wont have any effect on scheduling until the task is
5289 * SCHED_FIFO/SCHED_RR:
5291 if (task_has_rt_policy(p)) {
5292 p->static_prio = NICE_TO_PRIO(nice);
5297 dequeue_task(rq, p, 0);
5299 p->static_prio = NICE_TO_PRIO(nice);
5302 p->prio = effective_prio(p);
5303 delta = p->prio - old_prio;
5306 enqueue_task(rq, p, 0);
5308 * If the task increased its priority or is running and
5309 * lowered its priority, then reschedule its CPU:
5311 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5312 resched_task(rq->curr);
5315 task_rq_unlock(rq, p, &flags);
5317 EXPORT_SYMBOL(set_user_nice);
5320 * can_nice - check if a task can reduce its nice value
5324 int can_nice(const struct task_struct *p, const int nice)
5326 /* convert nice value [19,-20] to rlimit style value [1,40] */
5327 int nice_rlim = 20 - nice;
5329 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5330 capable(CAP_SYS_NICE));
5333 #ifdef __ARCH_WANT_SYS_NICE
5336 * sys_nice - change the priority of the current process.
5337 * @increment: priority increment
5339 * sys_setpriority is a more generic, but much slower function that
5340 * does similar things.
5342 SYSCALL_DEFINE1(nice, int, increment)
5347 * Setpriority might change our priority at the same moment.
5348 * We don't have to worry. Conceptually one call occurs first
5349 * and we have a single winner.
5351 if (increment < -40)
5356 nice = TASK_NICE(current) + increment;
5362 if (increment < 0 && !can_nice(current, nice))
5365 retval = security_task_setnice(current, nice);
5369 set_user_nice(current, nice);
5376 * task_prio - return the priority value of a given task.
5377 * @p: the task in question.
5379 * This is the priority value as seen by users in /proc.
5380 * RT tasks are offset by -200. Normal tasks are centered
5381 * around 0, value goes from -16 to +15.
5383 int task_prio(const struct task_struct *p)
5385 return p->prio - MAX_RT_PRIO;
5389 * task_nice - return the nice value of a given task.
5390 * @p: the task in question.
5392 int task_nice(const struct task_struct *p)
5394 return TASK_NICE(p);
5396 EXPORT_SYMBOL(task_nice);
5399 * idle_cpu - is a given cpu idle currently?
5400 * @cpu: the processor in question.
5402 int idle_cpu(int cpu)
5404 struct rq *rq = cpu_rq(cpu);
5406 if (rq->curr != rq->idle)
5413 if (!llist_empty(&rq->wake_list))
5421 * idle_task - return the idle task for a given cpu.
5422 * @cpu: the processor in question.
5424 struct task_struct *idle_task(int cpu)
5426 return cpu_rq(cpu)->idle;
5430 * find_process_by_pid - find a process with a matching PID value.
5431 * @pid: the pid in question.
5433 static struct task_struct *find_process_by_pid(pid_t pid)
5435 return pid ? find_task_by_vpid(pid) : current;
5438 /* Actually do priority change: must hold rq lock. */
5440 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5443 p->rt_priority = prio;
5444 p->normal_prio = normal_prio(p);
5445 /* we are holding p->pi_lock already */
5446 p->prio = rt_mutex_getprio(p);
5447 if (rt_prio(p->prio))
5448 p->sched_class = &rt_sched_class;
5450 p->sched_class = &fair_sched_class;
5455 * check the target process has a UID that matches the current process's
5457 static bool check_same_owner(struct task_struct *p)
5459 const struct cred *cred = current_cred(), *pcred;
5463 pcred = __task_cred(p);
5464 if (cred->user->user_ns == pcred->user->user_ns)
5465 match = (cred->euid == pcred->euid ||
5466 cred->euid == pcred->uid);
5473 static int __sched_setscheduler(struct task_struct *p, int policy,
5474 const struct sched_param *param, bool user)
5476 int retval, oldprio, oldpolicy = -1, on_rq, running;
5477 unsigned long flags;
5478 const struct sched_class *prev_class;
5482 /* may grab non-irq protected spin_locks */
5483 BUG_ON(in_interrupt());
5485 /* double check policy once rq lock held */
5487 reset_on_fork = p->sched_reset_on_fork;
5488 policy = oldpolicy = p->policy;
5490 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5491 policy &= ~SCHED_RESET_ON_FORK;
5493 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5494 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5495 policy != SCHED_IDLE)
5500 * Valid priorities for SCHED_FIFO and SCHED_RR are
5501 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5502 * SCHED_BATCH and SCHED_IDLE is 0.
5504 if (param->sched_priority < 0 ||
5505 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5506 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5508 if (rt_policy(policy) != (param->sched_priority != 0))
5512 * Allow unprivileged RT tasks to decrease priority:
5514 if (user && !capable(CAP_SYS_NICE)) {
5515 if (rt_policy(policy)) {
5516 unsigned long rlim_rtprio =
5517 task_rlimit(p, RLIMIT_RTPRIO);
5519 /* can't set/change the rt policy */
5520 if (policy != p->policy && !rlim_rtprio)
5523 /* can't increase priority */
5524 if (param->sched_priority > p->rt_priority &&
5525 param->sched_priority > rlim_rtprio)
5530 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5531 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5533 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5534 if (!can_nice(p, TASK_NICE(p)))
5538 /* can't change other user's priorities */
5539 if (!check_same_owner(p))
5542 /* Normal users shall not reset the sched_reset_on_fork flag */
5543 if (p->sched_reset_on_fork && !reset_on_fork)
5548 retval = security_task_setscheduler(p);
5554 * make sure no PI-waiters arrive (or leave) while we are
5555 * changing the priority of the task:
5557 * To be able to change p->policy safely, the appropriate
5558 * runqueue lock must be held.
5560 rq = task_rq_lock(p, &flags);
5563 * Changing the policy of the stop threads its a very bad idea
5565 if (p == rq->stop) {
5566 task_rq_unlock(rq, p, &flags);
5571 * If not changing anything there's no need to proceed further:
5573 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5574 param->sched_priority == p->rt_priority))) {
5576 __task_rq_unlock(rq);
5577 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5581 #ifdef CONFIG_RT_GROUP_SCHED
5584 * Do not allow realtime tasks into groups that have no runtime
5587 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5588 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5589 !task_group_is_autogroup(task_group(p))) {
5590 task_rq_unlock(rq, p, &flags);
5596 /* recheck policy now with rq lock held */
5597 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5598 policy = oldpolicy = -1;
5599 task_rq_unlock(rq, p, &flags);
5603 running = task_current(rq, p);
5605 deactivate_task(rq, p, 0);
5607 p->sched_class->put_prev_task(rq, p);
5609 p->sched_reset_on_fork = reset_on_fork;
5612 prev_class = p->sched_class;
5613 __setscheduler(rq, p, policy, param->sched_priority);
5616 p->sched_class->set_curr_task(rq);
5618 activate_task(rq, p, 0);
5620 check_class_changed(rq, p, prev_class, oldprio);
5621 task_rq_unlock(rq, p, &flags);
5623 rt_mutex_adjust_pi(p);
5629 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5630 * @p: the task in question.
5631 * @policy: new policy.
5632 * @param: structure containing the new RT priority.
5634 * NOTE that the task may be already dead.
5636 int sched_setscheduler(struct task_struct *p, int policy,
5637 const struct sched_param *param)
5639 return __sched_setscheduler(p, policy, param, true);
5641 EXPORT_SYMBOL_GPL(sched_setscheduler);
5644 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5645 * @p: the task in question.
5646 * @policy: new policy.
5647 * @param: structure containing the new RT priority.
5649 * Just like sched_setscheduler, only don't bother checking if the
5650 * current context has permission. For example, this is needed in
5651 * stop_machine(): we create temporary high priority worker threads,
5652 * but our caller might not have that capability.
5654 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5655 const struct sched_param *param)
5657 return __sched_setscheduler(p, policy, param, false);
5661 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5663 struct sched_param lparam;
5664 struct task_struct *p;
5667 if (!param || pid < 0)
5669 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5674 p = find_process_by_pid(pid);
5676 retval = sched_setscheduler(p, policy, &lparam);
5683 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5684 * @pid: the pid in question.
5685 * @policy: new policy.
5686 * @param: structure containing the new RT priority.
5688 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5689 struct sched_param __user *, param)
5691 /* negative values for policy are not valid */
5695 return do_sched_setscheduler(pid, policy, param);
5699 * sys_sched_setparam - set/change the RT priority of a thread
5700 * @pid: the pid in question.
5701 * @param: structure containing the new RT priority.
5703 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5705 return do_sched_setscheduler(pid, -1, param);
5709 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5710 * @pid: the pid in question.
5712 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5714 struct task_struct *p;
5722 p = find_process_by_pid(pid);
5724 retval = security_task_getscheduler(p);
5727 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5734 * sys_sched_getparam - get the RT priority of a thread
5735 * @pid: the pid in question.
5736 * @param: structure containing the RT priority.
5738 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5740 struct sched_param lp;
5741 struct task_struct *p;
5744 if (!param || pid < 0)
5748 p = find_process_by_pid(pid);
5753 retval = security_task_getscheduler(p);
5757 lp.sched_priority = p->rt_priority;
5761 * This one might sleep, we cannot do it with a spinlock held ...
5763 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5772 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5774 cpumask_var_t cpus_allowed, new_mask;
5775 struct task_struct *p;
5781 p = find_process_by_pid(pid);
5788 /* Prevent p going away */
5792 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5796 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5798 goto out_free_cpus_allowed;
5801 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5804 retval = security_task_setscheduler(p);
5808 cpuset_cpus_allowed(p, cpus_allowed);
5809 cpumask_and(new_mask, in_mask, cpus_allowed);
5811 retval = set_cpus_allowed_ptr(p, new_mask);
5814 cpuset_cpus_allowed(p, cpus_allowed);
5815 if (!cpumask_subset(new_mask, cpus_allowed)) {
5817 * We must have raced with a concurrent cpuset
5818 * update. Just reset the cpus_allowed to the
5819 * cpuset's cpus_allowed
5821 cpumask_copy(new_mask, cpus_allowed);
5826 free_cpumask_var(new_mask);
5827 out_free_cpus_allowed:
5828 free_cpumask_var(cpus_allowed);
5835 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5836 struct cpumask *new_mask)
5838 if (len < cpumask_size())
5839 cpumask_clear(new_mask);
5840 else if (len > cpumask_size())
5841 len = cpumask_size();
5843 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5847 * sys_sched_setaffinity - set the cpu affinity of a process
5848 * @pid: pid of the process
5849 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5850 * @user_mask_ptr: user-space pointer to the new cpu mask
5852 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5853 unsigned long __user *, user_mask_ptr)
5855 cpumask_var_t new_mask;
5858 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5861 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5863 retval = sched_setaffinity(pid, new_mask);
5864 free_cpumask_var(new_mask);
5868 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5870 struct task_struct *p;
5871 unsigned long flags;
5878 p = find_process_by_pid(pid);
5882 retval = security_task_getscheduler(p);
5886 raw_spin_lock_irqsave(&p->pi_lock, flags);
5887 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5888 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5898 * sys_sched_getaffinity - get the cpu affinity of a process
5899 * @pid: pid of the process
5900 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5901 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5903 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5904 unsigned long __user *, user_mask_ptr)
5909 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5911 if (len & (sizeof(unsigned long)-1))
5914 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5917 ret = sched_getaffinity(pid, mask);
5919 size_t retlen = min_t(size_t, len, cpumask_size());
5921 if (copy_to_user(user_mask_ptr, mask, retlen))
5926 free_cpumask_var(mask);
5932 * sys_sched_yield - yield the current processor to other threads.
5934 * This function yields the current CPU to other tasks. If there are no
5935 * other threads running on this CPU then this function will return.
5937 SYSCALL_DEFINE0(sched_yield)
5939 struct rq *rq = this_rq_lock();
5941 schedstat_inc(rq, yld_count);
5942 current->sched_class->yield_task(rq);
5945 * Since we are going to call schedule() anyway, there's
5946 * no need to preempt or enable interrupts:
5948 __release(rq->lock);
5949 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5950 do_raw_spin_unlock(&rq->lock);
5951 preempt_enable_no_resched();
5958 static inline int should_resched(void)
5960 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5963 static void __cond_resched(void)
5965 add_preempt_count(PREEMPT_ACTIVE);
5967 sub_preempt_count(PREEMPT_ACTIVE);
5970 int __sched _cond_resched(void)
5972 if (should_resched()) {
5978 EXPORT_SYMBOL(_cond_resched);
5981 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5982 * call schedule, and on return reacquire the lock.
5984 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5985 * operations here to prevent schedule() from being called twice (once via
5986 * spin_unlock(), once by hand).
5988 int __cond_resched_lock(spinlock_t *lock)
5990 int resched = should_resched();
5993 lockdep_assert_held(lock);
5995 if (spin_needbreak(lock) || resched) {
6006 EXPORT_SYMBOL(__cond_resched_lock);
6008 int __sched __cond_resched_softirq(void)
6010 BUG_ON(!in_softirq());
6012 if (should_resched()) {
6020 EXPORT_SYMBOL(__cond_resched_softirq);
6023 * yield - yield the current processor to other threads.
6025 * This is a shortcut for kernel-space yielding - it marks the
6026 * thread runnable and calls sys_sched_yield().
6028 void __sched yield(void)
6030 set_current_state(TASK_RUNNING);
6033 EXPORT_SYMBOL(yield);
6036 * yield_to - yield the current processor to another thread in
6037 * your thread group, or accelerate that thread toward the
6038 * processor it's on.
6040 * @preempt: whether task preemption is allowed or not
6042 * It's the caller's job to ensure that the target task struct
6043 * can't go away on us before we can do any checks.
6045 * Returns true if we indeed boosted the target task.
6047 bool __sched yield_to(struct task_struct *p, bool preempt)
6049 struct task_struct *curr = current;
6050 struct rq *rq, *p_rq;
6051 unsigned long flags;
6054 local_irq_save(flags);
6059 double_rq_lock(rq, p_rq);
6060 while (task_rq(p) != p_rq) {
6061 double_rq_unlock(rq, p_rq);
6065 if (!curr->sched_class->yield_to_task)
6068 if (curr->sched_class != p->sched_class)
6071 if (task_running(p_rq, p) || p->state)
6074 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
6076 schedstat_inc(rq, yld_count);
6078 * Make p's CPU reschedule; pick_next_entity takes care of
6081 if (preempt && rq != p_rq)
6082 resched_task(p_rq->curr);
6086 double_rq_unlock(rq, p_rq);
6087 local_irq_restore(flags);
6094 EXPORT_SYMBOL_GPL(yield_to);
6097 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6098 * that process accounting knows that this is a task in IO wait state.
6100 void __sched io_schedule(void)
6102 struct rq *rq = raw_rq();
6104 delayacct_blkio_start();
6105 atomic_inc(&rq->nr_iowait);
6106 blk_flush_plug(current);
6107 current->in_iowait = 1;
6109 current->in_iowait = 0;
6110 atomic_dec(&rq->nr_iowait);
6111 delayacct_blkio_end();
6113 EXPORT_SYMBOL(io_schedule);
6115 long __sched io_schedule_timeout(long timeout)
6117 struct rq *rq = raw_rq();
6120 delayacct_blkio_start();
6121 atomic_inc(&rq->nr_iowait);
6122 blk_flush_plug(current);
6123 current->in_iowait = 1;
6124 ret = schedule_timeout(timeout);
6125 current->in_iowait = 0;
6126 atomic_dec(&rq->nr_iowait);
6127 delayacct_blkio_end();
6132 * sys_sched_get_priority_max - return maximum RT priority.
6133 * @policy: scheduling class.
6135 * this syscall returns the maximum rt_priority that can be used
6136 * by a given scheduling class.
6138 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6145 ret = MAX_USER_RT_PRIO-1;
6157 * sys_sched_get_priority_min - return minimum RT priority.
6158 * @policy: scheduling class.
6160 * this syscall returns the minimum rt_priority that can be used
6161 * by a given scheduling class.
6163 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6181 * sys_sched_rr_get_interval - return the default timeslice of a process.
6182 * @pid: pid of the process.
6183 * @interval: userspace pointer to the timeslice value.
6185 * this syscall writes the default timeslice value of a given process
6186 * into the user-space timespec buffer. A value of '0' means infinity.
6188 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6189 struct timespec __user *, interval)
6191 struct task_struct *p;
6192 unsigned int time_slice;
6193 unsigned long flags;
6203 p = find_process_by_pid(pid);
6207 retval = security_task_getscheduler(p);
6211 rq = task_rq_lock(p, &flags);
6212 time_slice = p->sched_class->get_rr_interval(rq, p);
6213 task_rq_unlock(rq, p, &flags);
6216 jiffies_to_timespec(time_slice, &t);
6217 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6225 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6227 void sched_show_task(struct task_struct *p)
6229 unsigned long free = 0;
6232 state = p->state ? __ffs(p->state) + 1 : 0;
6233 printk(KERN_INFO "%-15.15s %c", p->comm,
6234 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6235 #if BITS_PER_LONG == 32
6236 if (state == TASK_RUNNING)
6237 printk(KERN_CONT " running ");
6239 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6241 if (state == TASK_RUNNING)
6242 printk(KERN_CONT " running task ");
6244 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6246 #ifdef CONFIG_DEBUG_STACK_USAGE
6247 free = stack_not_used(p);
6249 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6250 task_pid_nr(p), task_pid_nr(p->real_parent),
6251 (unsigned long)task_thread_info(p)->flags);
6253 show_stack(p, NULL);
6256 void show_state_filter(unsigned long state_filter)
6258 struct task_struct *g, *p;
6260 #if BITS_PER_LONG == 32
6262 " task PC stack pid father\n");
6265 " task PC stack pid father\n");
6268 do_each_thread(g, p) {
6270 * reset the NMI-timeout, listing all files on a slow
6271 * console might take a lot of time:
6272 * Also, reset softlockup watchdogs on all CPUs, because
6273 * another CPU might be blocked waiting for us to process
6276 touch_nmi_watchdog();
6277 touch_all_softlockup_watchdogs();
6278 if (!state_filter || (p->state & state_filter))
6280 } while_each_thread(g, p);
6282 #ifdef CONFIG_SCHED_DEBUG
6283 sysrq_sched_debug_show();
6287 * Only show locks if all tasks are dumped:
6290 debug_show_all_locks();
6293 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6295 idle->sched_class = &idle_sched_class;
6299 * init_idle - set up an idle thread for a given CPU
6300 * @idle: task in question
6301 * @cpu: cpu the idle task belongs to
6303 * NOTE: this function does not set the idle thread's NEED_RESCHED
6304 * flag, to make booting more robust.
6306 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6308 struct rq *rq = cpu_rq(cpu);
6309 unsigned long flags;
6311 raw_spin_lock_irqsave(&rq->lock, flags);
6314 idle->state = TASK_RUNNING;
6315 idle->se.exec_start = sched_clock();
6317 do_set_cpus_allowed(idle, cpumask_of(cpu));
6319 * We're having a chicken and egg problem, even though we are
6320 * holding rq->lock, the cpu isn't yet set to this cpu so the
6321 * lockdep check in task_group() will fail.
6323 * Similar case to sched_fork(). / Alternatively we could
6324 * use task_rq_lock() here and obtain the other rq->lock.
6329 __set_task_cpu(idle, cpu);
6332 rq->curr = rq->idle = idle;
6333 #if defined(CONFIG_SMP)
6336 raw_spin_unlock_irqrestore(&rq->lock, flags);
6338 /* Set the preempt count _outside_ the spinlocks! */
6339 task_thread_info(idle)->preempt_count = 0;
6342 * The idle tasks have their own, simple scheduling class:
6344 idle->sched_class = &idle_sched_class;
6345 ftrace_graph_init_idle_task(idle, cpu);
6346 #if defined(CONFIG_SMP)
6347 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6352 * Increase the granularity value when there are more CPUs,
6353 * because with more CPUs the 'effective latency' as visible
6354 * to users decreases. But the relationship is not linear,
6355 * so pick a second-best guess by going with the log2 of the
6358 * This idea comes from the SD scheduler of Con Kolivas:
6360 static int get_update_sysctl_factor(void)
6362 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6363 unsigned int factor;
6365 switch (sysctl_sched_tunable_scaling) {
6366 case SCHED_TUNABLESCALING_NONE:
6369 case SCHED_TUNABLESCALING_LINEAR:
6372 case SCHED_TUNABLESCALING_LOG:
6374 factor = 1 + ilog2(cpus);
6381 static void update_sysctl(void)
6383 unsigned int factor = get_update_sysctl_factor();
6385 #define SET_SYSCTL(name) \
6386 (sysctl_##name = (factor) * normalized_sysctl_##name)
6387 SET_SYSCTL(sched_min_granularity);
6388 SET_SYSCTL(sched_latency);
6389 SET_SYSCTL(sched_wakeup_granularity);
6393 static inline void sched_init_granularity(void)
6399 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6401 if (p->sched_class && p->sched_class->set_cpus_allowed)
6402 p->sched_class->set_cpus_allowed(p, new_mask);
6404 cpumask_copy(&p->cpus_allowed, new_mask);
6405 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6409 * This is how migration works:
6411 * 1) we invoke migration_cpu_stop() on the target CPU using
6413 * 2) stopper starts to run (implicitly forcing the migrated thread
6415 * 3) it checks whether the migrated task is still in the wrong runqueue.
6416 * 4) if it's in the wrong runqueue then the migration thread removes
6417 * it and puts it into the right queue.
6418 * 5) stopper completes and stop_one_cpu() returns and the migration
6423 * Change a given task's CPU affinity. Migrate the thread to a
6424 * proper CPU and schedule it away if the CPU it's executing on
6425 * is removed from the allowed bitmask.
6427 * NOTE: the caller must have a valid reference to the task, the
6428 * task must not exit() & deallocate itself prematurely. The
6429 * call is not atomic; no spinlocks may be held.
6431 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6433 unsigned long flags;
6435 unsigned int dest_cpu;
6438 rq = task_rq_lock(p, &flags);
6440 if (cpumask_equal(&p->cpus_allowed, new_mask))
6443 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6448 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6453 do_set_cpus_allowed(p, new_mask);
6455 /* Can the task run on the task's current CPU? If so, we're done */
6456 if (cpumask_test_cpu(task_cpu(p), new_mask))
6459 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6461 struct migration_arg arg = { p, dest_cpu };
6462 /* Need help from migration thread: drop lock and wait. */
6463 task_rq_unlock(rq, p, &flags);
6464 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6465 tlb_migrate_finish(p->mm);
6469 task_rq_unlock(rq, p, &flags);
6473 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6476 * Move (not current) task off this cpu, onto dest cpu. We're doing
6477 * this because either it can't run here any more (set_cpus_allowed()
6478 * away from this CPU, or CPU going down), or because we're
6479 * attempting to rebalance this task on exec (sched_exec).
6481 * So we race with normal scheduler movements, but that's OK, as long
6482 * as the task is no longer on this CPU.
6484 * Returns non-zero if task was successfully migrated.
6486 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6488 struct rq *rq_dest, *rq_src;
6491 if (unlikely(!cpu_active(dest_cpu)))
6494 rq_src = cpu_rq(src_cpu);
6495 rq_dest = cpu_rq(dest_cpu);
6497 raw_spin_lock(&p->pi_lock);
6498 double_rq_lock(rq_src, rq_dest);
6499 /* Already moved. */
6500 if (task_cpu(p) != src_cpu)
6502 /* Affinity changed (again). */
6503 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
6507 * If we're not on a rq, the next wake-up will ensure we're
6511 deactivate_task(rq_src, p, 0);
6512 set_task_cpu(p, dest_cpu);
6513 activate_task(rq_dest, p, 0);
6514 check_preempt_curr(rq_dest, p, 0);
6519 double_rq_unlock(rq_src, rq_dest);
6520 raw_spin_unlock(&p->pi_lock);
6525 * migration_cpu_stop - this will be executed by a highprio stopper thread
6526 * and performs thread migration by bumping thread off CPU then
6527 * 'pushing' onto another runqueue.
6529 static int migration_cpu_stop(void *data)
6531 struct migration_arg *arg = data;
6534 * The original target cpu might have gone down and we might
6535 * be on another cpu but it doesn't matter.
6537 local_irq_disable();
6538 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6543 #ifdef CONFIG_HOTPLUG_CPU
6546 * Ensures that the idle task is using init_mm right before its cpu goes
6549 void idle_task_exit(void)
6551 struct mm_struct *mm = current->active_mm;
6553 BUG_ON(cpu_online(smp_processor_id()));
6556 switch_mm(mm, &init_mm, current);
6561 * While a dead CPU has no uninterruptible tasks queued at this point,
6562 * it might still have a nonzero ->nr_uninterruptible counter, because
6563 * for performance reasons the counter is not stricly tracking tasks to
6564 * their home CPUs. So we just add the counter to another CPU's counter,
6565 * to keep the global sum constant after CPU-down:
6567 static void migrate_nr_uninterruptible(struct rq *rq_src)
6569 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6571 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6572 rq_src->nr_uninterruptible = 0;
6576 * remove the tasks which were accounted by rq from calc_load_tasks.
6578 static void calc_global_load_remove(struct rq *rq)
6580 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6581 rq->calc_load_active = 0;
6584 #ifdef CONFIG_CFS_BANDWIDTH
6585 static void unthrottle_offline_cfs_rqs(struct rq *rq)
6587 struct cfs_rq *cfs_rq;
6589 for_each_leaf_cfs_rq(rq, cfs_rq) {
6590 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6592 if (!cfs_rq->runtime_enabled)
6596 * clock_task is not advancing so we just need to make sure
6597 * there's some valid quota amount
6599 cfs_rq->runtime_remaining = cfs_b->quota;
6600 if (cfs_rq_throttled(cfs_rq))
6601 unthrottle_cfs_rq(cfs_rq);
6607 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6608 * try_to_wake_up()->select_task_rq().
6610 * Called with rq->lock held even though we'er in stop_machine() and
6611 * there's no concurrency possible, we hold the required locks anyway
6612 * because of lock validation efforts.
6614 static void migrate_tasks(unsigned int dead_cpu)
6616 struct rq *rq = cpu_rq(dead_cpu);
6617 struct task_struct *next, *stop = rq->stop;
6621 * Fudge the rq selection such that the below task selection loop
6622 * doesn't get stuck on the currently eligible stop task.
6624 * We're currently inside stop_machine() and the rq is either stuck
6625 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6626 * either way we should never end up calling schedule() until we're
6633 * There's this thread running, bail when that's the only
6636 if (rq->nr_running == 1)
6639 next = pick_next_task(rq);
6641 next->sched_class->put_prev_task(rq, next);
6643 /* Find suitable destination for @next, with force if needed. */
6644 dest_cpu = select_fallback_rq(dead_cpu, next);
6645 raw_spin_unlock(&rq->lock);
6647 __migrate_task(next, dead_cpu, dest_cpu);
6649 raw_spin_lock(&rq->lock);
6655 #endif /* CONFIG_HOTPLUG_CPU */
6657 #if !defined(CONFIG_HOTPLUG_CPU) || !defined(CONFIG_CFS_BANDWIDTH)
6658 static void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6661 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6663 static struct ctl_table sd_ctl_dir[] = {
6665 .procname = "sched_domain",
6671 static struct ctl_table sd_ctl_root[] = {
6673 .procname = "kernel",
6675 .child = sd_ctl_dir,
6680 static struct ctl_table *sd_alloc_ctl_entry(int n)
6682 struct ctl_table *entry =
6683 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6688 static void sd_free_ctl_entry(struct ctl_table **tablep)
6690 struct ctl_table *entry;
6693 * In the intermediate directories, both the child directory and
6694 * procname are dynamically allocated and could fail but the mode
6695 * will always be set. In the lowest directory the names are
6696 * static strings and all have proc handlers.
6698 for (entry = *tablep; entry->mode; entry++) {
6700 sd_free_ctl_entry(&entry->child);
6701 if (entry->proc_handler == NULL)
6702 kfree(entry->procname);
6709 static int min_load_idx = 0;
6710 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
6713 set_table_entry(struct ctl_table *entry,
6714 const char *procname, void *data, int maxlen,
6715 mode_t mode, proc_handler *proc_handler,
6718 entry->procname = procname;
6720 entry->maxlen = maxlen;
6722 entry->proc_handler = proc_handler;
6725 entry->extra1 = &min_load_idx;
6726 entry->extra2 = &max_load_idx;
6730 static struct ctl_table *
6731 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6733 struct ctl_table *table = sd_alloc_ctl_entry(13);
6738 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6739 sizeof(long), 0644, proc_doulongvec_minmax, false);
6740 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6741 sizeof(long), 0644, proc_doulongvec_minmax, false);
6742 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6743 sizeof(int), 0644, proc_dointvec_minmax, true);
6744 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6745 sizeof(int), 0644, proc_dointvec_minmax, true);
6746 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6747 sizeof(int), 0644, proc_dointvec_minmax, true);
6748 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6749 sizeof(int), 0644, proc_dointvec_minmax, true);
6750 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6751 sizeof(int), 0644, proc_dointvec_minmax, true);
6752 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6753 sizeof(int), 0644, proc_dointvec_minmax, false);
6754 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6755 sizeof(int), 0644, proc_dointvec_minmax, false);
6756 set_table_entry(&table[9], "cache_nice_tries",
6757 &sd->cache_nice_tries,
6758 sizeof(int), 0644, proc_dointvec_minmax, false);
6759 set_table_entry(&table[10], "flags", &sd->flags,
6760 sizeof(int), 0644, proc_dointvec_minmax, false);
6761 set_table_entry(&table[11], "name", sd->name,
6762 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
6763 /* &table[12] is terminator */
6768 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6770 struct ctl_table *entry, *table;
6771 struct sched_domain *sd;
6772 int domain_num = 0, i;
6775 for_each_domain(cpu, sd)
6777 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6782 for_each_domain(cpu, sd) {
6783 snprintf(buf, 32, "domain%d", i);
6784 entry->procname = kstrdup(buf, GFP_KERNEL);
6786 entry->child = sd_alloc_ctl_domain_table(sd);
6793 static struct ctl_table_header *sd_sysctl_header;
6794 static void register_sched_domain_sysctl(void)
6796 int i, cpu_num = num_possible_cpus();
6797 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6800 WARN_ON(sd_ctl_dir[0].child);
6801 sd_ctl_dir[0].child = entry;
6806 for_each_possible_cpu(i) {
6807 snprintf(buf, 32, "cpu%d", i);
6808 entry->procname = kstrdup(buf, GFP_KERNEL);
6810 entry->child = sd_alloc_ctl_cpu_table(i);
6814 WARN_ON(sd_sysctl_header);
6815 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6818 /* may be called multiple times per register */
6819 static void unregister_sched_domain_sysctl(void)
6821 if (sd_sysctl_header)
6822 unregister_sysctl_table(sd_sysctl_header);
6823 sd_sysctl_header = NULL;
6824 if (sd_ctl_dir[0].child)
6825 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6828 static void register_sched_domain_sysctl(void)
6831 static void unregister_sched_domain_sysctl(void)
6836 static void set_rq_online(struct rq *rq)
6839 const struct sched_class *class;
6841 cpumask_set_cpu(rq->cpu, rq->rd->online);
6844 for_each_class(class) {
6845 if (class->rq_online)
6846 class->rq_online(rq);
6851 static void set_rq_offline(struct rq *rq)
6854 const struct sched_class *class;
6856 for_each_class(class) {
6857 if (class->rq_offline)
6858 class->rq_offline(rq);
6861 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6867 * migration_call - callback that gets triggered when a CPU is added.
6868 * Here we can start up the necessary migration thread for the new CPU.
6870 static int __cpuinit
6871 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6873 int cpu = (long)hcpu;
6874 unsigned long flags;
6875 struct rq *rq = cpu_rq(cpu);
6877 switch (action & ~CPU_TASKS_FROZEN) {
6879 case CPU_UP_PREPARE:
6880 rq->calc_load_update = calc_load_update;
6884 /* Update our root-domain */
6885 raw_spin_lock_irqsave(&rq->lock, flags);
6887 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6891 raw_spin_unlock_irqrestore(&rq->lock, flags);
6894 #ifdef CONFIG_HOTPLUG_CPU
6896 sched_ttwu_pending();
6897 /* Update our root-domain */
6898 raw_spin_lock_irqsave(&rq->lock, flags);
6900 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6904 BUG_ON(rq->nr_running != 1); /* the migration thread */
6905 raw_spin_unlock_irqrestore(&rq->lock, flags);
6907 migrate_nr_uninterruptible(rq);
6908 calc_global_load_remove(rq);
6913 update_max_interval();
6919 * Register at high priority so that task migration (migrate_all_tasks)
6920 * happens before everything else. This has to be lower priority than
6921 * the notifier in the perf_event subsystem, though.
6923 static struct notifier_block __cpuinitdata migration_notifier = {
6924 .notifier_call = migration_call,
6925 .priority = CPU_PRI_MIGRATION,
6928 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6929 unsigned long action, void *hcpu)
6931 switch (action & ~CPU_TASKS_FROZEN) {
6933 case CPU_DOWN_FAILED:
6934 set_cpu_active((long)hcpu, true);
6941 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6942 unsigned long action, void *hcpu)
6944 switch (action & ~CPU_TASKS_FROZEN) {
6945 case CPU_DOWN_PREPARE:
6946 set_cpu_active((long)hcpu, false);
6953 static int __init migration_init(void)
6955 void *cpu = (void *)(long)smp_processor_id();
6958 /* Initialize migration for the boot CPU */
6959 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6960 BUG_ON(err == NOTIFY_BAD);
6961 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6962 register_cpu_notifier(&migration_notifier);
6964 /* Register cpu active notifiers */
6965 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6966 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6970 early_initcall(migration_init);
6975 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6977 #ifdef CONFIG_SCHED_DEBUG
6979 static __read_mostly int sched_domain_debug_enabled;
6981 static int __init sched_domain_debug_setup(char *str)
6983 sched_domain_debug_enabled = 1;
6987 early_param("sched_debug", sched_domain_debug_setup);
6989 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6990 struct cpumask *groupmask)
6992 struct sched_group *group = sd->groups;
6995 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6996 cpumask_clear(groupmask);
6998 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7000 if (!(sd->flags & SD_LOAD_BALANCE)) {
7001 printk("does not load-balance\n");
7003 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7008 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7010 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7011 printk(KERN_ERR "ERROR: domain->span does not contain "
7014 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7015 printk(KERN_ERR "ERROR: domain->groups does not contain"
7019 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7023 printk(KERN_ERR "ERROR: group is NULL\n");
7027 if (!group->sgp->power) {
7028 printk(KERN_CONT "\n");
7029 printk(KERN_ERR "ERROR: domain->cpu_power not "
7034 if (!cpumask_weight(sched_group_cpus(group))) {
7035 printk(KERN_CONT "\n");
7036 printk(KERN_ERR "ERROR: empty group\n");
7040 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7041 printk(KERN_CONT "\n");
7042 printk(KERN_ERR "ERROR: repeated CPUs\n");
7046 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7048 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7050 printk(KERN_CONT " %s", str);
7051 if (group->sgp->power != SCHED_POWER_SCALE) {
7052 printk(KERN_CONT " (cpu_power = %d)",
7056 group = group->next;
7057 } while (group != sd->groups);
7058 printk(KERN_CONT "\n");
7060 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7061 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7064 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7065 printk(KERN_ERR "ERROR: parent span is not a superset "
7066 "of domain->span\n");
7070 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7074 if (!sched_domain_debug_enabled)
7078 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7082 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7085 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
7093 #else /* !CONFIG_SCHED_DEBUG */
7094 # define sched_domain_debug(sd, cpu) do { } while (0)
7095 #endif /* CONFIG_SCHED_DEBUG */
7097 static int sd_degenerate(struct sched_domain *sd)
7099 if (cpumask_weight(sched_domain_span(sd)) == 1)
7102 /* Following flags need at least 2 groups */
7103 if (sd->flags & (SD_LOAD_BALANCE |
7104 SD_BALANCE_NEWIDLE |
7108 SD_SHARE_PKG_RESOURCES)) {
7109 if (sd->groups != sd->groups->next)
7113 /* Following flags don't use groups */
7114 if (sd->flags & (SD_WAKE_AFFINE))
7121 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7123 unsigned long cflags = sd->flags, pflags = parent->flags;
7125 if (sd_degenerate(parent))
7128 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7131 /* Flags needing groups don't count if only 1 group in parent */
7132 if (parent->groups == parent->groups->next) {
7133 pflags &= ~(SD_LOAD_BALANCE |
7134 SD_BALANCE_NEWIDLE |
7138 SD_SHARE_PKG_RESOURCES);
7139 if (nr_node_ids == 1)
7140 pflags &= ~SD_SERIALIZE;
7142 if (~cflags & pflags)
7148 static void free_rootdomain(struct rcu_head *rcu)
7150 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
7152 cpupri_cleanup(&rd->cpupri);
7153 free_cpumask_var(rd->rto_mask);
7154 free_cpumask_var(rd->online);
7155 free_cpumask_var(rd->span);
7159 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7161 struct root_domain *old_rd = NULL;
7162 unsigned long flags;
7164 raw_spin_lock_irqsave(&rq->lock, flags);
7169 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7172 cpumask_clear_cpu(rq->cpu, old_rd->span);
7175 * If we dont want to free the old_rt yet then
7176 * set old_rd to NULL to skip the freeing later
7179 if (!atomic_dec_and_test(&old_rd->refcount))
7183 atomic_inc(&rd->refcount);
7186 cpumask_set_cpu(rq->cpu, rd->span);
7187 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7190 raw_spin_unlock_irqrestore(&rq->lock, flags);
7193 call_rcu_sched(&old_rd->rcu, free_rootdomain);
7196 static int init_rootdomain(struct root_domain *rd)
7198 memset(rd, 0, sizeof(*rd));
7200 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
7202 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
7204 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7207 if (cpupri_init(&rd->cpupri) != 0)
7212 free_cpumask_var(rd->rto_mask);
7214 free_cpumask_var(rd->online);
7216 free_cpumask_var(rd->span);
7221 static void init_defrootdomain(void)
7223 init_rootdomain(&def_root_domain);
7225 atomic_set(&def_root_domain.refcount, 1);
7228 static struct root_domain *alloc_rootdomain(void)
7230 struct root_domain *rd;
7232 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7236 if (init_rootdomain(rd) != 0) {
7244 static void free_sched_groups(struct sched_group *sg, int free_sgp)
7246 struct sched_group *tmp, *first;
7255 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
7260 } while (sg != first);
7263 static void free_sched_domain(struct rcu_head *rcu)
7265 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
7268 * If its an overlapping domain it has private groups, iterate and
7271 if (sd->flags & SD_OVERLAP) {
7272 free_sched_groups(sd->groups, 1);
7273 } else if (atomic_dec_and_test(&sd->groups->ref)) {
7274 kfree(sd->groups->sgp);
7280 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
7282 call_rcu(&sd->rcu, free_sched_domain);
7285 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
7287 for (; sd; sd = sd->parent)
7288 destroy_sched_domain(sd, cpu);
7292 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7293 * hold the hotplug lock.
7296 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7298 struct rq *rq = cpu_rq(cpu);
7299 struct sched_domain *tmp;
7301 /* Remove the sched domains which do not contribute to scheduling. */
7302 for (tmp = sd; tmp; ) {
7303 struct sched_domain *parent = tmp->parent;
7307 if (sd_parent_degenerate(tmp, parent)) {
7308 tmp->parent = parent->parent;
7310 parent->parent->child = tmp;
7311 destroy_sched_domain(parent, cpu);
7316 if (sd && sd_degenerate(sd)) {
7319 destroy_sched_domain(tmp, cpu);
7324 sched_domain_debug(sd, cpu);
7326 rq_attach_root(rq, rd);
7328 rcu_assign_pointer(rq->sd, sd);
7329 destroy_sched_domains(tmp, cpu);
7332 /* cpus with isolated domains */
7333 static cpumask_var_t cpu_isolated_map;
7335 /* Setup the mask of cpus configured for isolated domains */
7336 static int __init isolated_cpu_setup(char *str)
7338 alloc_bootmem_cpumask_var(&cpu_isolated_map);
7339 cpulist_parse(str, cpu_isolated_map);
7343 __setup("isolcpus=", isolated_cpu_setup);
7348 * find_next_best_node - find the next node to include in a sched_domain
7349 * @node: node whose sched_domain we're building
7350 * @used_nodes: nodes already in the sched_domain
7352 * Find the next node to include in a given scheduling domain. Simply
7353 * finds the closest node not already in the @used_nodes map.
7355 * Should use nodemask_t.
7357 static int find_next_best_node(int node, nodemask_t *used_nodes)
7359 int i, n, val, min_val, best_node = -1;
7363 for (i = 0; i < nr_node_ids; i++) {
7364 /* Start at @node */
7365 n = (node + i) % nr_node_ids;
7367 if (!nr_cpus_node(n))
7370 /* Skip already used nodes */
7371 if (node_isset(n, *used_nodes))
7374 /* Simple min distance search */
7375 val = node_distance(node, n);
7377 if (val < min_val) {
7383 if (best_node != -1)
7384 node_set(best_node, *used_nodes);
7389 * sched_domain_node_span - get a cpumask for a node's sched_domain
7390 * @node: node whose cpumask we're constructing
7391 * @span: resulting cpumask
7393 * Given a node, construct a good cpumask for its sched_domain to span. It
7394 * should be one that prevents unnecessary balancing, but also spreads tasks
7397 static void sched_domain_node_span(int node, struct cpumask *span)
7399 nodemask_t used_nodes;
7402 cpumask_clear(span);
7403 nodes_clear(used_nodes);
7405 cpumask_or(span, span, cpumask_of_node(node));
7406 node_set(node, used_nodes);
7408 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7409 int next_node = find_next_best_node(node, &used_nodes);
7412 cpumask_or(span, span, cpumask_of_node(next_node));
7416 static const struct cpumask *cpu_node_mask(int cpu)
7418 lockdep_assert_held(&sched_domains_mutex);
7420 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7422 return sched_domains_tmpmask;
7425 static const struct cpumask *cpu_allnodes_mask(int cpu)
7427 return cpu_possible_mask;
7429 #endif /* CONFIG_NUMA */
7431 static const struct cpumask *cpu_cpu_mask(int cpu)
7433 return cpumask_of_node(cpu_to_node(cpu));
7436 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7439 struct sched_domain **__percpu sd;
7440 struct sched_group **__percpu sg;
7441 struct sched_group_power **__percpu sgp;
7445 struct sched_domain ** __percpu sd;
7446 struct root_domain *rd;
7456 struct sched_domain_topology_level;
7458 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7459 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7461 #define SDTL_OVERLAP 0x01
7463 struct sched_domain_topology_level {
7464 sched_domain_init_f init;
7465 sched_domain_mask_f mask;
7467 struct sd_data data;
7471 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7473 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7474 const struct cpumask *span = sched_domain_span(sd);
7475 struct cpumask *covered = sched_domains_tmpmask;
7476 struct sd_data *sdd = sd->private;
7477 struct sched_domain *child;
7480 cpumask_clear(covered);
7482 for_each_cpu(i, span) {
7483 struct cpumask *sg_span;
7485 if (cpumask_test_cpu(i, covered))
7488 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7489 GFP_KERNEL, cpu_to_node(i));
7494 sg_span = sched_group_cpus(sg);
7496 child = *per_cpu_ptr(sdd->sd, i);
7498 child = child->child;
7499 cpumask_copy(sg_span, sched_domain_span(child));
7501 cpumask_set_cpu(i, sg_span);
7503 cpumask_or(covered, covered, sg_span);
7505 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7506 atomic_inc(&sg->sgp->ref);
7508 if (cpumask_test_cpu(cpu, sg_span))
7518 sd->groups = groups;
7523 free_sched_groups(first, 0);
7528 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7530 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7531 struct sched_domain *child = sd->child;
7534 cpu = cpumask_first(sched_domain_span(child));
7537 *sg = *per_cpu_ptr(sdd->sg, cpu);
7538 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7539 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7546 * build_sched_groups will build a circular linked list of the groups
7547 * covered by the given span, and will set each group's ->cpumask correctly,
7548 * and ->cpu_power to 0.
7550 * Assumes the sched_domain tree is fully constructed
7553 build_sched_groups(struct sched_domain *sd, int cpu)
7555 struct sched_group *first = NULL, *last = NULL;
7556 struct sd_data *sdd = sd->private;
7557 const struct cpumask *span = sched_domain_span(sd);
7558 struct cpumask *covered;
7561 get_group(cpu, sdd, &sd->groups);
7562 atomic_inc(&sd->groups->ref);
7564 if (cpu != cpumask_first(sched_domain_span(sd)))
7567 lockdep_assert_held(&sched_domains_mutex);
7568 covered = sched_domains_tmpmask;
7570 cpumask_clear(covered);
7572 for_each_cpu(i, span) {
7573 struct sched_group *sg;
7574 int group = get_group(i, sdd, &sg);
7577 if (cpumask_test_cpu(i, covered))
7580 cpumask_clear(sched_group_cpus(sg));
7583 for_each_cpu(j, span) {
7584 if (get_group(j, sdd, NULL) != group)
7587 cpumask_set_cpu(j, covered);
7588 cpumask_set_cpu(j, sched_group_cpus(sg));
7603 * Initialize sched groups cpu_power.
7605 * cpu_power indicates the capacity of sched group, which is used while
7606 * distributing the load between different sched groups in a sched domain.
7607 * Typically cpu_power for all the groups in a sched domain will be same unless
7608 * there are asymmetries in the topology. If there are asymmetries, group
7609 * having more cpu_power will pickup more load compared to the group having
7612 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7614 struct sched_group *sg = sd->groups;
7616 WARN_ON(!sd || !sg);
7619 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7621 } while (sg != sd->groups);
7623 if (cpu != group_first_cpu(sg))
7626 update_group_power(sd, cpu);
7630 * Initializers for schedule domains
7631 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7634 #ifdef CONFIG_SCHED_DEBUG
7635 # define SD_INIT_NAME(sd, type) sd->name = #type
7637 # define SD_INIT_NAME(sd, type) do { } while (0)
7640 #define SD_INIT_FUNC(type) \
7641 static noinline struct sched_domain * \
7642 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7644 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7645 *sd = SD_##type##_INIT; \
7646 SD_INIT_NAME(sd, type); \
7647 sd->private = &tl->data; \
7653 SD_INIT_FUNC(ALLNODES)
7656 #ifdef CONFIG_SCHED_SMT
7657 SD_INIT_FUNC(SIBLING)
7659 #ifdef CONFIG_SCHED_MC
7662 #ifdef CONFIG_SCHED_BOOK
7666 static int default_relax_domain_level = -1;
7667 int sched_domain_level_max;
7669 static int __init setup_relax_domain_level(char *str)
7671 if (kstrtoint(str, 0, &default_relax_domain_level))
7672 pr_warn("Unable to set relax_domain_level\n");
7676 __setup("relax_domain_level=", setup_relax_domain_level);
7678 static void set_domain_attribute(struct sched_domain *sd,
7679 struct sched_domain_attr *attr)
7683 if (!attr || attr->relax_domain_level < 0) {
7684 if (default_relax_domain_level < 0)
7687 request = default_relax_domain_level;
7689 request = attr->relax_domain_level;
7690 if (request < sd->level) {
7691 /* turn off idle balance on this domain */
7692 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7694 /* turn on idle balance on this domain */
7695 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7699 static void __sdt_free(const struct cpumask *cpu_map);
7700 static int __sdt_alloc(const struct cpumask *cpu_map);
7702 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7703 const struct cpumask *cpu_map)
7707 if (!atomic_read(&d->rd->refcount))
7708 free_rootdomain(&d->rd->rcu); /* fall through */
7710 free_percpu(d->sd); /* fall through */
7712 __sdt_free(cpu_map); /* fall through */
7718 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7719 const struct cpumask *cpu_map)
7721 memset(d, 0, sizeof(*d));
7723 if (__sdt_alloc(cpu_map))
7724 return sa_sd_storage;
7725 d->sd = alloc_percpu(struct sched_domain *);
7727 return sa_sd_storage;
7728 d->rd = alloc_rootdomain();
7731 return sa_rootdomain;
7735 * NULL the sd_data elements we've used to build the sched_domain and
7736 * sched_group structure so that the subsequent __free_domain_allocs()
7737 * will not free the data we're using.
7739 static void claim_allocations(int cpu, struct sched_domain *sd)
7741 struct sd_data *sdd = sd->private;
7743 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7744 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7746 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7747 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7749 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7750 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7753 #ifdef CONFIG_SCHED_SMT
7754 static const struct cpumask *cpu_smt_mask(int cpu)
7756 return topology_thread_cpumask(cpu);
7761 * Topology list, bottom-up.
7763 static struct sched_domain_topology_level default_topology[] = {
7764 #ifdef CONFIG_SCHED_SMT
7765 { sd_init_SIBLING, cpu_smt_mask, },
7767 #ifdef CONFIG_SCHED_MC
7768 { sd_init_MC, cpu_coregroup_mask, },
7770 #ifdef CONFIG_SCHED_BOOK
7771 { sd_init_BOOK, cpu_book_mask, },
7773 { sd_init_CPU, cpu_cpu_mask, },
7775 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7776 { sd_init_ALLNODES, cpu_allnodes_mask, },
7781 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7783 static int __sdt_alloc(const struct cpumask *cpu_map)
7785 struct sched_domain_topology_level *tl;
7788 for (tl = sched_domain_topology; tl->init; tl++) {
7789 struct sd_data *sdd = &tl->data;
7791 sdd->sd = alloc_percpu(struct sched_domain *);
7795 sdd->sg = alloc_percpu(struct sched_group *);
7799 sdd->sgp = alloc_percpu(struct sched_group_power *);
7803 for_each_cpu(j, cpu_map) {
7804 struct sched_domain *sd;
7805 struct sched_group *sg;
7806 struct sched_group_power *sgp;
7808 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7809 GFP_KERNEL, cpu_to_node(j));
7813 *per_cpu_ptr(sdd->sd, j) = sd;
7815 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7816 GFP_KERNEL, cpu_to_node(j));
7820 *per_cpu_ptr(sdd->sg, j) = sg;
7822 sgp = kzalloc_node(sizeof(struct sched_group_power),
7823 GFP_KERNEL, cpu_to_node(j));
7827 *per_cpu_ptr(sdd->sgp, j) = sgp;
7834 static void __sdt_free(const struct cpumask *cpu_map)
7836 struct sched_domain_topology_level *tl;
7839 for (tl = sched_domain_topology; tl->init; tl++) {
7840 struct sd_data *sdd = &tl->data;
7842 for_each_cpu(j, cpu_map) {
7843 struct sched_domain *sd;
7846 sd = *per_cpu_ptr(sdd->sd, j);
7847 if (sd && (sd->flags & SD_OVERLAP))
7848 free_sched_groups(sd->groups, 0);
7849 kfree(*per_cpu_ptr(sdd->sd, j));
7853 kfree(*per_cpu_ptr(sdd->sg, j));
7855 kfree(*per_cpu_ptr(sdd->sgp, j));
7857 free_percpu(sdd->sd);
7859 free_percpu(sdd->sg);
7861 free_percpu(sdd->sgp);
7866 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7867 struct s_data *d, const struct cpumask *cpu_map,
7868 struct sched_domain_attr *attr, struct sched_domain *child,
7871 struct sched_domain *sd = tl->init(tl, cpu);
7875 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7877 sd->level = child->level + 1;
7878 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7882 set_domain_attribute(sd, attr);
7888 * Build sched domains for a given set of cpus and attach the sched domains
7889 * to the individual cpus
7891 static int build_sched_domains(const struct cpumask *cpu_map,
7892 struct sched_domain_attr *attr)
7894 enum s_alloc alloc_state = sa_none;
7895 struct sched_domain *sd;
7897 int i, ret = -ENOMEM;
7899 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7900 if (alloc_state != sa_rootdomain)
7903 /* Set up domains for cpus specified by the cpu_map. */
7904 for_each_cpu(i, cpu_map) {
7905 struct sched_domain_topology_level *tl;
7908 for (tl = sched_domain_topology; tl->init; tl++) {
7909 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7910 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7911 sd->flags |= SD_OVERLAP;
7912 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7919 *per_cpu_ptr(d.sd, i) = sd;
7922 /* Build the groups for the domains */
7923 for_each_cpu(i, cpu_map) {
7924 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7925 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7926 if (sd->flags & SD_OVERLAP) {
7927 if (build_overlap_sched_groups(sd, i))
7930 if (build_sched_groups(sd, i))
7936 /* Calculate CPU power for physical packages and nodes */
7937 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7938 if (!cpumask_test_cpu(i, cpu_map))
7941 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7942 claim_allocations(i, sd);
7943 init_sched_groups_power(i, sd);
7947 /* Attach the domains */
7949 for_each_cpu(i, cpu_map) {
7950 sd = *per_cpu_ptr(d.sd, i);
7951 cpu_attach_domain(sd, d.rd, i);
7957 __free_domain_allocs(&d, alloc_state, cpu_map);
7961 static cpumask_var_t *doms_cur; /* current sched domains */
7962 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7963 static struct sched_domain_attr *dattr_cur;
7964 /* attribues of custom domains in 'doms_cur' */
7967 * Special case: If a kmalloc of a doms_cur partition (array of
7968 * cpumask) fails, then fallback to a single sched domain,
7969 * as determined by the single cpumask fallback_doms.
7971 static cpumask_var_t fallback_doms;
7974 * arch_update_cpu_topology lets virtualized architectures update the
7975 * cpu core maps. It is supposed to return 1 if the topology changed
7976 * or 0 if it stayed the same.
7978 int __attribute__((weak)) arch_update_cpu_topology(void)
7983 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7986 cpumask_var_t *doms;
7988 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7991 for (i = 0; i < ndoms; i++) {
7992 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7993 free_sched_domains(doms, i);
8000 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
8003 for (i = 0; i < ndoms; i++)
8004 free_cpumask_var(doms[i]);
8009 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8010 * For now this just excludes isolated cpus, but could be used to
8011 * exclude other special cases in the future.
8013 static int init_sched_domains(const struct cpumask *cpu_map)
8017 arch_update_cpu_topology();
8019 doms_cur = alloc_sched_domains(ndoms_cur);
8021 doms_cur = &fallback_doms;
8022 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
8024 err = build_sched_domains(doms_cur[0], NULL);
8025 register_sched_domain_sysctl();
8031 * Detach sched domains from a group of cpus specified in cpu_map
8032 * These cpus will now be attached to the NULL domain
8034 static void detach_destroy_domains(const struct cpumask *cpu_map)
8039 for_each_cpu(i, cpu_map)
8040 cpu_attach_domain(NULL, &def_root_domain, i);
8044 /* handle null as "default" */
8045 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8046 struct sched_domain_attr *new, int idx_new)
8048 struct sched_domain_attr tmp;
8055 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8056 new ? (new + idx_new) : &tmp,
8057 sizeof(struct sched_domain_attr));
8061 * Partition sched domains as specified by the 'ndoms_new'
8062 * cpumasks in the array doms_new[] of cpumasks. This compares
8063 * doms_new[] to the current sched domain partitioning, doms_cur[].
8064 * It destroys each deleted domain and builds each new domain.
8066 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
8067 * The masks don't intersect (don't overlap.) We should setup one
8068 * sched domain for each mask. CPUs not in any of the cpumasks will
8069 * not be load balanced. If the same cpumask appears both in the
8070 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8073 * The passed in 'doms_new' should be allocated using
8074 * alloc_sched_domains. This routine takes ownership of it and will
8075 * free_sched_domains it when done with it. If the caller failed the
8076 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
8077 * and partition_sched_domains() will fallback to the single partition
8078 * 'fallback_doms', it also forces the domains to be rebuilt.
8080 * If doms_new == NULL it will be replaced with cpu_online_mask.
8081 * ndoms_new == 0 is a special case for destroying existing domains,
8082 * and it will not create the default domain.
8084 * Call with hotplug lock held
8086 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
8087 struct sched_domain_attr *dattr_new)
8092 mutex_lock(&sched_domains_mutex);
8094 /* always unregister in case we don't destroy any domains */
8095 unregister_sched_domain_sysctl();
8097 /* Let architecture update cpu core mappings. */
8098 new_topology = arch_update_cpu_topology();
8100 n = doms_new ? ndoms_new : 0;
8102 /* Destroy deleted domains */
8103 for (i = 0; i < ndoms_cur; i++) {
8104 for (j = 0; j < n && !new_topology; j++) {
8105 if (cpumask_equal(doms_cur[i], doms_new[j])
8106 && dattrs_equal(dattr_cur, i, dattr_new, j))
8109 /* no match - a current sched domain not in new doms_new[] */
8110 detach_destroy_domains(doms_cur[i]);
8115 if (doms_new == NULL) {
8117 doms_new = &fallback_doms;
8118 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
8119 WARN_ON_ONCE(dattr_new);
8122 /* Build new domains */
8123 for (i = 0; i < ndoms_new; i++) {
8124 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8125 if (cpumask_equal(doms_new[i], doms_cur[j])
8126 && dattrs_equal(dattr_new, i, dattr_cur, j))
8129 /* no match - add a new doms_new */
8130 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
8135 /* Remember the new sched domains */
8136 if (doms_cur != &fallback_doms)
8137 free_sched_domains(doms_cur, ndoms_cur);
8138 kfree(dattr_cur); /* kfree(NULL) is safe */
8139 doms_cur = doms_new;
8140 dattr_cur = dattr_new;
8141 ndoms_cur = ndoms_new;
8143 register_sched_domain_sysctl();
8145 mutex_unlock(&sched_domains_mutex);
8148 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8149 static void reinit_sched_domains(void)
8153 /* Destroy domains first to force the rebuild */
8154 partition_sched_domains(0, NULL, NULL);
8156 rebuild_sched_domains();
8160 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8162 unsigned int level = 0;
8164 if (sscanf(buf, "%u", &level) != 1)
8168 * level is always be positive so don't check for
8169 * level < POWERSAVINGS_BALANCE_NONE which is 0
8170 * What happens on 0 or 1 byte write,
8171 * need to check for count as well?
8174 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8178 sched_smt_power_savings = level;
8180 sched_mc_power_savings = level;
8182 reinit_sched_domains();
8187 #ifdef CONFIG_SCHED_MC
8188 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8189 struct sysdev_class_attribute *attr,
8192 return sprintf(page, "%u\n", sched_mc_power_savings);
8194 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8195 struct sysdev_class_attribute *attr,
8196 const char *buf, size_t count)
8198 return sched_power_savings_store(buf, count, 0);
8200 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8201 sched_mc_power_savings_show,
8202 sched_mc_power_savings_store);
8205 #ifdef CONFIG_SCHED_SMT
8206 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8207 struct sysdev_class_attribute *attr,
8210 return sprintf(page, "%u\n", sched_smt_power_savings);
8212 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8213 struct sysdev_class_attribute *attr,
8214 const char *buf, size_t count)
8216 return sched_power_savings_store(buf, count, 1);
8218 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8219 sched_smt_power_savings_show,
8220 sched_smt_power_savings_store);
8223 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8227 #ifdef CONFIG_SCHED_SMT
8229 err = sysfs_create_file(&cls->kset.kobj,
8230 &attr_sched_smt_power_savings.attr);
8232 #ifdef CONFIG_SCHED_MC
8233 if (!err && mc_capable())
8234 err = sysfs_create_file(&cls->kset.kobj,
8235 &attr_sched_mc_power_savings.attr);
8239 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8241 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
8244 * Update cpusets according to cpu_active mask. If cpusets are
8245 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8246 * around partition_sched_domains().
8248 * If we come here as part of a suspend/resume, don't touch cpusets because we
8249 * want to restore it back to its original state upon resume anyway.
8251 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
8255 case CPU_ONLINE_FROZEN:
8256 case CPU_DOWN_FAILED_FROZEN:
8259 * num_cpus_frozen tracks how many CPUs are involved in suspend
8260 * resume sequence. As long as this is not the last online
8261 * operation in the resume sequence, just build a single sched
8262 * domain, ignoring cpusets.
8265 if (likely(num_cpus_frozen)) {
8266 partition_sched_domains(1, NULL, NULL);
8271 * This is the last CPU online operation. So fall through and
8272 * restore the original sched domains by considering the
8273 * cpuset configurations.
8277 case CPU_DOWN_FAILED:
8278 cpuset_update_active_cpus();
8286 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
8290 case CPU_DOWN_PREPARE:
8291 cpuset_update_active_cpus();
8293 case CPU_DOWN_PREPARE_FROZEN:
8295 partition_sched_domains(1, NULL, NULL);
8303 static int update_runtime(struct notifier_block *nfb,
8304 unsigned long action, void *hcpu)
8306 int cpu = (int)(long)hcpu;
8309 case CPU_DOWN_PREPARE:
8310 case CPU_DOWN_PREPARE_FROZEN:
8311 disable_runtime(cpu_rq(cpu));
8314 case CPU_DOWN_FAILED:
8315 case CPU_DOWN_FAILED_FROZEN:
8317 case CPU_ONLINE_FROZEN:
8318 enable_runtime(cpu_rq(cpu));
8326 void __init sched_init_smp(void)
8328 cpumask_var_t non_isolated_cpus;
8330 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8331 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8334 mutex_lock(&sched_domains_mutex);
8335 init_sched_domains(cpu_active_mask);
8336 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8337 if (cpumask_empty(non_isolated_cpus))
8338 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8339 mutex_unlock(&sched_domains_mutex);
8342 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
8343 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
8345 /* RT runtime code needs to handle some hotplug events */
8346 hotcpu_notifier(update_runtime, 0);
8350 /* Move init over to a non-isolated CPU */
8351 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8353 sched_init_granularity();
8354 free_cpumask_var(non_isolated_cpus);
8356 init_sched_rt_class();
8359 void __init sched_init_smp(void)
8361 sched_init_granularity();
8363 #endif /* CONFIG_SMP */
8365 const_debug unsigned int sysctl_timer_migration = 1;
8367 int in_sched_functions(unsigned long addr)
8369 return in_lock_functions(addr) ||
8370 (addr >= (unsigned long)__sched_text_start
8371 && addr < (unsigned long)__sched_text_end);
8374 static void init_cfs_rq(struct cfs_rq *cfs_rq)
8376 cfs_rq->tasks_timeline = RB_ROOT;
8377 INIT_LIST_HEAD(&cfs_rq->tasks);
8378 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8379 #ifndef CONFIG_64BIT
8380 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8384 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8386 struct rt_prio_array *array;
8389 array = &rt_rq->active;
8390 for (i = 0; i < MAX_RT_PRIO; i++) {
8391 INIT_LIST_HEAD(array->queue + i);
8392 __clear_bit(i, array->bitmap);
8394 /* delimiter for bitsearch: */
8395 __set_bit(MAX_RT_PRIO, array->bitmap);
8397 #if defined CONFIG_SMP
8398 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8399 rt_rq->highest_prio.next = MAX_RT_PRIO;
8400 rt_rq->rt_nr_migratory = 0;
8401 rt_rq->overloaded = 0;
8402 plist_head_init(&rt_rq->pushable_tasks);
8406 rt_rq->rt_throttled = 0;
8407 rt_rq->rt_runtime = 0;
8408 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8411 #ifdef CONFIG_FAIR_GROUP_SCHED
8412 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8413 struct sched_entity *se, int cpu,
8414 struct sched_entity *parent)
8416 struct rq *rq = cpu_rq(cpu);
8421 /* allow initial update_cfs_load() to truncate */
8422 cfs_rq->load_stamp = 1;
8424 init_cfs_rq_runtime(cfs_rq);
8426 tg->cfs_rq[cpu] = cfs_rq;
8429 /* se could be NULL for root_task_group */
8434 se->cfs_rq = &rq->cfs;
8436 se->cfs_rq = parent->my_q;
8439 update_load_set(&se->load, 0);
8440 se->parent = parent;
8444 #ifdef CONFIG_RT_GROUP_SCHED
8445 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8446 struct sched_rt_entity *rt_se, int cpu,
8447 struct sched_rt_entity *parent)
8449 struct rq *rq = cpu_rq(cpu);
8451 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8452 rt_rq->rt_nr_boosted = 0;
8456 tg->rt_rq[cpu] = rt_rq;
8457 tg->rt_se[cpu] = rt_se;
8463 rt_se->rt_rq = &rq->rt;
8465 rt_se->rt_rq = parent->my_q;
8467 rt_se->my_q = rt_rq;
8468 rt_se->parent = parent;
8469 INIT_LIST_HEAD(&rt_se->run_list);
8473 void __init sched_init(void)
8476 unsigned long alloc_size = 0, ptr;
8478 #ifdef CONFIG_FAIR_GROUP_SCHED
8479 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8481 #ifdef CONFIG_RT_GROUP_SCHED
8482 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8484 #ifdef CONFIG_CPUMASK_OFFSTACK
8485 alloc_size += num_possible_cpus() * cpumask_size();
8488 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8490 #ifdef CONFIG_FAIR_GROUP_SCHED
8491 root_task_group.se = (struct sched_entity **)ptr;
8492 ptr += nr_cpu_ids * sizeof(void **);
8494 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8495 ptr += nr_cpu_ids * sizeof(void **);
8497 #endif /* CONFIG_FAIR_GROUP_SCHED */
8498 #ifdef CONFIG_RT_GROUP_SCHED
8499 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8500 ptr += nr_cpu_ids * sizeof(void **);
8502 root_task_group.rt_rq = (struct rt_rq **)ptr;
8503 ptr += nr_cpu_ids * sizeof(void **);
8505 #endif /* CONFIG_RT_GROUP_SCHED */
8506 #ifdef CONFIG_CPUMASK_OFFSTACK
8507 for_each_possible_cpu(i) {
8508 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8509 ptr += cpumask_size();
8511 #endif /* CONFIG_CPUMASK_OFFSTACK */
8515 init_defrootdomain();
8518 init_rt_bandwidth(&def_rt_bandwidth,
8519 global_rt_period(), global_rt_runtime());
8521 #ifdef CONFIG_RT_GROUP_SCHED
8522 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8523 global_rt_period(), global_rt_runtime());
8524 #endif /* CONFIG_RT_GROUP_SCHED */
8526 #ifdef CONFIG_CGROUP_SCHED
8527 list_add(&root_task_group.list, &task_groups);
8528 INIT_LIST_HEAD(&root_task_group.children);
8529 autogroup_init(&init_task);
8530 #endif /* CONFIG_CGROUP_SCHED */
8532 for_each_possible_cpu(i) {
8536 raw_spin_lock_init(&rq->lock);
8538 rq->calc_load_active = 0;
8539 rq->calc_load_update = jiffies + LOAD_FREQ;
8540 init_cfs_rq(&rq->cfs);
8541 init_rt_rq(&rq->rt, rq);
8542 #ifdef CONFIG_FAIR_GROUP_SCHED
8543 root_task_group.shares = root_task_group_load;
8544 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8546 * How much cpu bandwidth does root_task_group get?
8548 * In case of task-groups formed thr' the cgroup filesystem, it
8549 * gets 100% of the cpu resources in the system. This overall
8550 * system cpu resource is divided among the tasks of
8551 * root_task_group and its child task-groups in a fair manner,
8552 * based on each entity's (task or task-group's) weight
8553 * (se->load.weight).
8555 * In other words, if root_task_group has 10 tasks of weight
8556 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8557 * then A0's share of the cpu resource is:
8559 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8561 * We achieve this by letting root_task_group's tasks sit
8562 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8564 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8565 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8566 #endif /* CONFIG_FAIR_GROUP_SCHED */
8568 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8569 #ifdef CONFIG_RT_GROUP_SCHED
8570 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8571 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8574 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8575 rq->cpu_load[j] = 0;
8577 rq->last_load_update_tick = jiffies;
8582 rq->cpu_power = SCHED_POWER_SCALE;
8583 rq->post_schedule = 0;
8584 rq->active_balance = 0;
8585 rq->next_balance = jiffies;
8590 rq->avg_idle = 2*sysctl_sched_migration_cost;
8591 rq_attach_root(rq, &def_root_domain);
8593 rq->nohz_balance_kick = 0;
8597 atomic_set(&rq->nr_iowait, 0);
8600 set_load_weight(&init_task);
8602 #ifdef CONFIG_PREEMPT_NOTIFIERS
8603 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8607 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8610 #ifdef CONFIG_RT_MUTEXES
8611 plist_head_init(&init_task.pi_waiters);
8615 * The boot idle thread does lazy MMU switching as well:
8617 atomic_inc(&init_mm.mm_count);
8618 enter_lazy_tlb(&init_mm, current);
8621 * Make us the idle thread. Technically, schedule() should not be
8622 * called from this thread, however somewhere below it might be,
8623 * but because we are the idle thread, we just pick up running again
8624 * when this runqueue becomes "idle".
8626 init_idle(current, smp_processor_id());
8628 calc_load_update = jiffies + LOAD_FREQ;
8631 * During early bootup we pretend to be a normal task:
8633 current->sched_class = &fair_sched_class;
8636 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8638 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8639 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8640 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8641 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8642 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8644 /* May be allocated at isolcpus cmdline parse time */
8645 if (cpu_isolated_map == NULL)
8646 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8649 scheduler_running = 1;
8652 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8653 static inline int preempt_count_equals(int preempt_offset)
8655 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8657 return (nested == preempt_offset);
8660 void __might_sleep(const char *file, int line, int preempt_offset)
8662 static unsigned long prev_jiffy; /* ratelimiting */
8664 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
8665 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8666 system_state != SYSTEM_RUNNING || oops_in_progress)
8668 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8670 prev_jiffy = jiffies;
8673 "BUG: sleeping function called from invalid context at %s:%d\n",
8676 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8677 in_atomic(), irqs_disabled(),
8678 current->pid, current->comm);
8680 debug_show_held_locks(current);
8681 if (irqs_disabled())
8682 print_irqtrace_events(current);
8685 EXPORT_SYMBOL(__might_sleep);
8688 #ifdef CONFIG_MAGIC_SYSRQ
8689 static void normalize_task(struct rq *rq, struct task_struct *p)
8691 const struct sched_class *prev_class = p->sched_class;
8692 int old_prio = p->prio;
8697 deactivate_task(rq, p, 0);
8698 __setscheduler(rq, p, SCHED_NORMAL, 0);
8700 activate_task(rq, p, 0);
8701 resched_task(rq->curr);
8704 check_class_changed(rq, p, prev_class, old_prio);
8707 void normalize_rt_tasks(void)
8709 struct task_struct *g, *p;
8710 unsigned long flags;
8713 read_lock_irqsave(&tasklist_lock, flags);
8714 do_each_thread(g, p) {
8716 * Only normalize user tasks:
8721 p->se.exec_start = 0;
8722 #ifdef CONFIG_SCHEDSTATS
8723 p->se.statistics.wait_start = 0;
8724 p->se.statistics.sleep_start = 0;
8725 p->se.statistics.block_start = 0;
8730 * Renice negative nice level userspace
8733 if (TASK_NICE(p) < 0 && p->mm)
8734 set_user_nice(p, 0);
8738 raw_spin_lock(&p->pi_lock);
8739 rq = __task_rq_lock(p);
8741 normalize_task(rq, p);
8743 __task_rq_unlock(rq);
8744 raw_spin_unlock(&p->pi_lock);
8745 } while_each_thread(g, p);
8747 read_unlock_irqrestore(&tasklist_lock, flags);
8750 #endif /* CONFIG_MAGIC_SYSRQ */
8752 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8754 * These functions are only useful for the IA64 MCA handling, or kdb.
8756 * They can only be called when the whole system has been
8757 * stopped - every CPU needs to be quiescent, and no scheduling
8758 * activity can take place. Using them for anything else would
8759 * be a serious bug, and as a result, they aren't even visible
8760 * under any other configuration.
8764 * curr_task - return the current task for a given cpu.
8765 * @cpu: the processor in question.
8767 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8769 struct task_struct *curr_task(int cpu)
8771 return cpu_curr(cpu);
8774 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8778 * set_curr_task - set the current task for a given cpu.
8779 * @cpu: the processor in question.
8780 * @p: the task pointer to set.
8782 * Description: This function must only be used when non-maskable interrupts
8783 * are serviced on a separate stack. It allows the architecture to switch the
8784 * notion of the current task on a cpu in a non-blocking manner. This function
8785 * must be called with all CPU's synchronized, and interrupts disabled, the
8786 * and caller must save the original value of the current task (see
8787 * curr_task() above) and restore that value before reenabling interrupts and
8788 * re-starting the system.
8790 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8792 void set_curr_task(int cpu, struct task_struct *p)
8799 #ifdef CONFIG_FAIR_GROUP_SCHED
8800 static void free_fair_sched_group(struct task_group *tg)
8804 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8806 for_each_possible_cpu(i) {
8808 kfree(tg->cfs_rq[i]);
8818 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8820 struct cfs_rq *cfs_rq;
8821 struct sched_entity *se;
8824 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8827 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8831 tg->shares = NICE_0_LOAD;
8833 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8835 for_each_possible_cpu(i) {
8836 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8837 GFP_KERNEL, cpu_to_node(i));
8841 se = kzalloc_node(sizeof(struct sched_entity),
8842 GFP_KERNEL, cpu_to_node(i));
8846 init_cfs_rq(cfs_rq);
8847 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8858 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8860 struct rq *rq = cpu_rq(cpu);
8861 unsigned long flags;
8864 * Only empty task groups can be destroyed; so we can speculatively
8865 * check on_list without danger of it being re-added.
8867 if (!tg->cfs_rq[cpu]->on_list)
8870 raw_spin_lock_irqsave(&rq->lock, flags);
8871 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8872 raw_spin_unlock_irqrestore(&rq->lock, flags);
8874 #else /* !CONFIG_FAIR_GROUP_SCHED */
8875 static inline void free_fair_sched_group(struct task_group *tg)
8880 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8885 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8888 #endif /* CONFIG_FAIR_GROUP_SCHED */
8890 #ifdef CONFIG_RT_GROUP_SCHED
8891 static void free_rt_sched_group(struct task_group *tg)
8896 destroy_rt_bandwidth(&tg->rt_bandwidth);
8898 for_each_possible_cpu(i) {
8900 kfree(tg->rt_rq[i]);
8902 kfree(tg->rt_se[i]);
8910 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8912 struct rt_rq *rt_rq;
8913 struct sched_rt_entity *rt_se;
8916 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8919 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8923 init_rt_bandwidth(&tg->rt_bandwidth,
8924 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8926 for_each_possible_cpu(i) {
8927 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8928 GFP_KERNEL, cpu_to_node(i));
8932 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8933 GFP_KERNEL, cpu_to_node(i));
8937 init_rt_rq(rt_rq, cpu_rq(i));
8938 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8939 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8949 #else /* !CONFIG_RT_GROUP_SCHED */
8950 static inline void free_rt_sched_group(struct task_group *tg)
8955 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8959 #endif /* CONFIG_RT_GROUP_SCHED */
8961 #ifdef CONFIG_CGROUP_SCHED
8962 static void free_sched_group(struct task_group *tg)
8964 free_fair_sched_group(tg);
8965 free_rt_sched_group(tg);
8970 /* allocate runqueue etc for a new task group */
8971 struct task_group *sched_create_group(struct task_group *parent)
8973 struct task_group *tg;
8974 unsigned long flags;
8976 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8978 return ERR_PTR(-ENOMEM);
8980 if (!alloc_fair_sched_group(tg, parent))
8983 if (!alloc_rt_sched_group(tg, parent))
8986 spin_lock_irqsave(&task_group_lock, flags);
8987 list_add_rcu(&tg->list, &task_groups);
8989 WARN_ON(!parent); /* root should already exist */
8991 tg->parent = parent;
8992 INIT_LIST_HEAD(&tg->children);
8993 list_add_rcu(&tg->siblings, &parent->children);
8994 spin_unlock_irqrestore(&task_group_lock, flags);
8999 free_sched_group(tg);
9000 return ERR_PTR(-ENOMEM);
9003 /* rcu callback to free various structures associated with a task group */
9004 static void free_sched_group_rcu(struct rcu_head *rhp)
9006 /* now it should be safe to free those cfs_rqs */
9007 free_sched_group(container_of(rhp, struct task_group, rcu));
9010 /* Destroy runqueue etc associated with a task group */
9011 void sched_destroy_group(struct task_group *tg)
9013 unsigned long flags;
9016 /* end participation in shares distribution */
9017 for_each_possible_cpu(i)
9018 unregister_fair_sched_group(tg, i);
9020 spin_lock_irqsave(&task_group_lock, flags);
9021 list_del_rcu(&tg->list);
9022 list_del_rcu(&tg->siblings);
9023 spin_unlock_irqrestore(&task_group_lock, flags);
9025 /* wait for possible concurrent references to cfs_rqs complete */
9026 call_rcu(&tg->rcu, free_sched_group_rcu);
9029 /* change task's runqueue when it moves between groups.
9030 * The caller of this function should have put the task in its new group
9031 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9032 * reflect its new group.
9034 void sched_move_task(struct task_struct *tsk)
9036 struct task_group *tg;
9038 unsigned long flags;
9041 rq = task_rq_lock(tsk, &flags);
9043 running = task_current(rq, tsk);
9047 dequeue_task(rq, tsk, 0);
9048 if (unlikely(running))
9049 tsk->sched_class->put_prev_task(rq, tsk);
9051 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
9052 lockdep_is_held(&tsk->sighand->siglock)),
9053 struct task_group, css);
9054 tg = autogroup_task_group(tsk, tg);
9055 tsk->sched_task_group = tg;
9057 #ifdef CONFIG_FAIR_GROUP_SCHED
9058 if (tsk->sched_class->task_move_group)
9059 tsk->sched_class->task_move_group(tsk, on_rq);
9062 set_task_rq(tsk, task_cpu(tsk));
9064 if (unlikely(running))
9065 tsk->sched_class->set_curr_task(rq);
9067 enqueue_task(rq, tsk, 0);
9069 task_rq_unlock(rq, tsk, &flags);
9071 #endif /* CONFIG_CGROUP_SCHED */
9073 #ifdef CONFIG_FAIR_GROUP_SCHED
9074 static DEFINE_MUTEX(shares_mutex);
9076 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9079 unsigned long flags;
9082 * We can't change the weight of the root cgroup.
9087 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9089 mutex_lock(&shares_mutex);
9090 if (tg->shares == shares)
9093 tg->shares = shares;
9094 for_each_possible_cpu(i) {
9095 struct rq *rq = cpu_rq(i);
9096 struct sched_entity *se;
9099 /* Propagate contribution to hierarchy */
9100 raw_spin_lock_irqsave(&rq->lock, flags);
9101 for_each_sched_entity(se)
9102 update_cfs_shares(group_cfs_rq(se));
9103 raw_spin_unlock_irqrestore(&rq->lock, flags);
9107 mutex_unlock(&shares_mutex);
9111 unsigned long sched_group_shares(struct task_group *tg)
9117 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
9118 static unsigned long to_ratio(u64 period, u64 runtime)
9120 if (runtime == RUNTIME_INF)
9123 return div64_u64(runtime << 20, period);
9127 #ifdef CONFIG_RT_GROUP_SCHED
9129 * Ensure that the real time constraints are schedulable.
9131 static DEFINE_MUTEX(rt_constraints_mutex);
9133 /* Must be called with tasklist_lock held */
9134 static inline int tg_has_rt_tasks(struct task_group *tg)
9136 struct task_struct *g, *p;
9139 * Autogroups do not have RT tasks; see autogroup_create().
9141 if (task_group_is_autogroup(tg))
9144 do_each_thread(g, p) {
9145 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9147 } while_each_thread(g, p);
9152 struct rt_schedulable_data {
9153 struct task_group *tg;
9158 static int tg_rt_schedulable(struct task_group *tg, void *data)
9160 struct rt_schedulable_data *d = data;
9161 struct task_group *child;
9162 unsigned long total, sum = 0;
9163 u64 period, runtime;
9165 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9166 runtime = tg->rt_bandwidth.rt_runtime;
9169 period = d->rt_period;
9170 runtime = d->rt_runtime;
9174 * Cannot have more runtime than the period.
9176 if (runtime > period && runtime != RUNTIME_INF)
9180 * Ensure we don't starve existing RT tasks.
9182 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9185 total = to_ratio(period, runtime);
9188 * Nobody can have more than the global setting allows.
9190 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9194 * The sum of our children's runtime should not exceed our own.
9196 list_for_each_entry_rcu(child, &tg->children, siblings) {
9197 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9198 runtime = child->rt_bandwidth.rt_runtime;
9200 if (child == d->tg) {
9201 period = d->rt_period;
9202 runtime = d->rt_runtime;
9205 sum += to_ratio(period, runtime);
9214 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9218 struct rt_schedulable_data data = {
9220 .rt_period = period,
9221 .rt_runtime = runtime,
9225 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
9231 static int tg_set_rt_bandwidth(struct task_group *tg,
9232 u64 rt_period, u64 rt_runtime)
9236 mutex_lock(&rt_constraints_mutex);
9237 read_lock(&tasklist_lock);
9238 err = __rt_schedulable(tg, rt_period, rt_runtime);
9242 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9243 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9244 tg->rt_bandwidth.rt_runtime = rt_runtime;
9246 for_each_possible_cpu(i) {
9247 struct rt_rq *rt_rq = tg->rt_rq[i];
9249 raw_spin_lock(&rt_rq->rt_runtime_lock);
9250 rt_rq->rt_runtime = rt_runtime;
9251 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9253 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9255 read_unlock(&tasklist_lock);
9256 mutex_unlock(&rt_constraints_mutex);
9261 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9263 u64 rt_runtime, rt_period;
9265 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9266 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9267 if (rt_runtime_us < 0)
9268 rt_runtime = RUNTIME_INF;
9270 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9273 long sched_group_rt_runtime(struct task_group *tg)
9277 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9280 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9281 do_div(rt_runtime_us, NSEC_PER_USEC);
9282 return rt_runtime_us;
9285 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9287 u64 rt_runtime, rt_period;
9289 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9290 rt_runtime = tg->rt_bandwidth.rt_runtime;
9295 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9298 long sched_group_rt_period(struct task_group *tg)
9302 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9303 do_div(rt_period_us, NSEC_PER_USEC);
9304 return rt_period_us;
9307 static int sched_rt_global_constraints(void)
9309 u64 runtime, period;
9312 if (sysctl_sched_rt_period <= 0)
9315 runtime = global_rt_runtime();
9316 period = global_rt_period();
9319 * Sanity check on the sysctl variables.
9321 if (runtime > period && runtime != RUNTIME_INF)
9324 mutex_lock(&rt_constraints_mutex);
9325 read_lock(&tasklist_lock);
9326 ret = __rt_schedulable(NULL, 0, 0);
9327 read_unlock(&tasklist_lock);
9328 mutex_unlock(&rt_constraints_mutex);
9333 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9335 /* Don't accept realtime tasks when there is no way for them to run */
9336 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9342 #else /* !CONFIG_RT_GROUP_SCHED */
9343 static int sched_rt_global_constraints(void)
9345 unsigned long flags;
9348 if (sysctl_sched_rt_period <= 0)
9352 * There's always some RT tasks in the root group
9353 * -- migration, kstopmachine etc..
9355 if (sysctl_sched_rt_runtime == 0)
9358 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9359 for_each_possible_cpu(i) {
9360 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9362 raw_spin_lock(&rt_rq->rt_runtime_lock);
9363 rt_rq->rt_runtime = global_rt_runtime();
9364 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9366 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9370 #endif /* CONFIG_RT_GROUP_SCHED */
9372 int sched_rt_handler(struct ctl_table *table, int write,
9373 void __user *buffer, size_t *lenp,
9377 int old_period, old_runtime;
9378 static DEFINE_MUTEX(mutex);
9381 old_period = sysctl_sched_rt_period;
9382 old_runtime = sysctl_sched_rt_runtime;
9384 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9386 if (!ret && write) {
9387 ret = sched_rt_global_constraints();
9389 sysctl_sched_rt_period = old_period;
9390 sysctl_sched_rt_runtime = old_runtime;
9392 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9393 def_rt_bandwidth.rt_period =
9394 ns_to_ktime(global_rt_period());
9397 mutex_unlock(&mutex);
9402 #ifdef CONFIG_CGROUP_SCHED
9404 /* return corresponding task_group object of a cgroup */
9405 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9407 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9408 struct task_group, css);
9411 static struct cgroup_subsys_state *
9412 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9414 struct task_group *tg, *parent;
9416 if (!cgrp->parent) {
9417 /* This is early initialization for the top cgroup */
9418 return &root_task_group.css;
9421 parent = cgroup_tg(cgrp->parent);
9422 tg = sched_create_group(parent);
9424 return ERR_PTR(-ENOMEM);
9430 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9432 struct task_group *tg = cgroup_tg(cgrp);
9434 sched_destroy_group(tg);
9438 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9440 #ifdef CONFIG_RT_GROUP_SCHED
9441 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9444 /* We don't support RT-tasks being in separate groups */
9445 if (tsk->sched_class != &fair_sched_class)
9452 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9454 sched_move_task(tsk);
9458 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9459 struct cgroup *old_cgrp, struct task_struct *task)
9462 * cgroup_exit() is called in the copy_process() failure path.
9463 * Ignore this case since the task hasn't ran yet, this avoids
9464 * trying to poke a half freed task state from generic code.
9466 if (!(task->flags & PF_EXITING))
9469 sched_move_task(task);
9472 #ifdef CONFIG_FAIR_GROUP_SCHED
9473 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9476 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9479 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9481 struct task_group *tg = cgroup_tg(cgrp);
9483 return (u64) scale_load_down(tg->shares);
9486 #ifdef CONFIG_CFS_BANDWIDTH
9487 static DEFINE_MUTEX(cfs_constraints_mutex);
9489 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9490 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9492 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9494 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9496 int i, ret = 0, runtime_enabled;
9497 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9499 if (tg == &root_task_group)
9503 * Ensure we have at some amount of bandwidth every period. This is
9504 * to prevent reaching a state of large arrears when throttled via
9505 * entity_tick() resulting in prolonged exit starvation.
9507 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9511 * Likewise, bound things on the otherside by preventing insane quota
9512 * periods. This also allows us to normalize in computing quota
9515 if (period > max_cfs_quota_period)
9518 mutex_lock(&cfs_constraints_mutex);
9519 ret = __cfs_schedulable(tg, period, quota);
9523 runtime_enabled = quota != RUNTIME_INF;
9524 raw_spin_lock_irq(&cfs_b->lock);
9525 cfs_b->period = ns_to_ktime(period);
9526 cfs_b->quota = quota;
9528 __refill_cfs_bandwidth_runtime(cfs_b);
9529 /* restart the period timer (if active) to handle new period expiry */
9530 if (runtime_enabled && cfs_b->timer_active) {
9531 /* force a reprogram */
9532 cfs_b->timer_active = 0;
9533 __start_cfs_bandwidth(cfs_b);
9535 raw_spin_unlock_irq(&cfs_b->lock);
9537 for_each_possible_cpu(i) {
9538 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9539 struct rq *rq = rq_of(cfs_rq);
9541 raw_spin_lock_irq(&rq->lock);
9542 cfs_rq->runtime_enabled = runtime_enabled;
9543 cfs_rq->runtime_remaining = 0;
9545 if (cfs_rq_throttled(cfs_rq))
9546 unthrottle_cfs_rq(cfs_rq);
9547 raw_spin_unlock_irq(&rq->lock);
9550 mutex_unlock(&cfs_constraints_mutex);
9555 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9559 period = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9560 if (cfs_quota_us < 0)
9561 quota = RUNTIME_INF;
9563 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9565 return tg_set_cfs_bandwidth(tg, period, quota);
9568 long tg_get_cfs_quota(struct task_group *tg)
9572 if (tg_cfs_bandwidth(tg)->quota == RUNTIME_INF)
9575 quota_us = tg_cfs_bandwidth(tg)->quota;
9576 do_div(quota_us, NSEC_PER_USEC);
9581 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9585 period = (u64)cfs_period_us * NSEC_PER_USEC;
9586 quota = tg_cfs_bandwidth(tg)->quota;
9591 return tg_set_cfs_bandwidth(tg, period, quota);
9594 long tg_get_cfs_period(struct task_group *tg)
9598 cfs_period_us = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9599 do_div(cfs_period_us, NSEC_PER_USEC);
9601 return cfs_period_us;
9604 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
9606 return tg_get_cfs_quota(cgroup_tg(cgrp));
9609 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
9612 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
9615 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
9617 return tg_get_cfs_period(cgroup_tg(cgrp));
9620 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9623 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
9626 struct cfs_schedulable_data {
9627 struct task_group *tg;
9632 * normalize group quota/period to be quota/max_period
9633 * note: units are usecs
9635 static u64 normalize_cfs_quota(struct task_group *tg,
9636 struct cfs_schedulable_data *d)
9644 period = tg_get_cfs_period(tg);
9645 quota = tg_get_cfs_quota(tg);
9648 /* note: these should typically be equivalent */
9649 if (quota == RUNTIME_INF || quota == -1)
9652 return to_ratio(period, quota);
9655 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9657 struct cfs_schedulable_data *d = data;
9658 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9659 s64 quota = 0, parent_quota = -1;
9662 quota = RUNTIME_INF;
9664 struct cfs_bandwidth *parent_b = tg_cfs_bandwidth(tg->parent);
9666 quota = normalize_cfs_quota(tg, d);
9667 parent_quota = parent_b->hierarchal_quota;
9670 * ensure max(child_quota) <= parent_quota, inherit when no
9673 if (quota == RUNTIME_INF)
9674 quota = parent_quota;
9675 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9678 cfs_b->hierarchal_quota = quota;
9683 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9686 struct cfs_schedulable_data data = {
9692 if (quota != RUNTIME_INF) {
9693 do_div(data.period, NSEC_PER_USEC);
9694 do_div(data.quota, NSEC_PER_USEC);
9698 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9704 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
9705 struct cgroup_map_cb *cb)
9707 struct task_group *tg = cgroup_tg(cgrp);
9708 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9710 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
9711 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
9712 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
9716 #endif /* CONFIG_CFS_BANDWIDTH */
9717 #endif /* CONFIG_FAIR_GROUP_SCHED */
9719 #ifdef CONFIG_RT_GROUP_SCHED
9720 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9723 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9726 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9728 return sched_group_rt_runtime(cgroup_tg(cgrp));
9731 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9734 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9737 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9739 return sched_group_rt_period(cgroup_tg(cgrp));
9741 #endif /* CONFIG_RT_GROUP_SCHED */
9743 static struct cftype cpu_files[] = {
9744 #ifdef CONFIG_FAIR_GROUP_SCHED
9747 .read_u64 = cpu_shares_read_u64,
9748 .write_u64 = cpu_shares_write_u64,
9751 #ifdef CONFIG_CFS_BANDWIDTH
9753 .name = "cfs_quota_us",
9754 .read_s64 = cpu_cfs_quota_read_s64,
9755 .write_s64 = cpu_cfs_quota_write_s64,
9758 .name = "cfs_period_us",
9759 .read_u64 = cpu_cfs_period_read_u64,
9760 .write_u64 = cpu_cfs_period_write_u64,
9764 .read_map = cpu_stats_show,
9767 #ifdef CONFIG_RT_GROUP_SCHED
9769 .name = "rt_runtime_us",
9770 .read_s64 = cpu_rt_runtime_read,
9771 .write_s64 = cpu_rt_runtime_write,
9774 .name = "rt_period_us",
9775 .read_u64 = cpu_rt_period_read_uint,
9776 .write_u64 = cpu_rt_period_write_uint,
9781 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9783 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9786 struct cgroup_subsys cpu_cgroup_subsys = {
9788 .create = cpu_cgroup_create,
9789 .destroy = cpu_cgroup_destroy,
9790 .can_attach_task = cpu_cgroup_can_attach_task,
9791 .attach_task = cpu_cgroup_attach_task,
9792 .exit = cpu_cgroup_exit,
9793 .populate = cpu_cgroup_populate,
9794 .subsys_id = cpu_cgroup_subsys_id,
9798 #endif /* CONFIG_CGROUP_SCHED */
9800 #ifdef CONFIG_CGROUP_CPUACCT
9803 * CPU accounting code for task groups.
9805 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9806 * (balbir@in.ibm.com).
9809 /* track cpu usage of a group of tasks and its child groups */
9811 struct cgroup_subsys_state css;
9812 /* cpuusage holds pointer to a u64-type object on every cpu */
9813 u64 __percpu *cpuusage;
9814 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9815 struct cpuacct *parent;
9818 struct cgroup_subsys cpuacct_subsys;
9820 /* return cpu accounting group corresponding to this container */
9821 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9823 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9824 struct cpuacct, css);
9827 /* return cpu accounting group to which this task belongs */
9828 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9830 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9831 struct cpuacct, css);
9834 /* create a new cpu accounting group */
9835 static struct cgroup_subsys_state *cpuacct_create(
9836 struct cgroup_subsys *ss, struct cgroup *cgrp)
9838 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9844 ca->cpuusage = alloc_percpu(u64);
9848 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9849 if (percpu_counter_init(&ca->cpustat[i], 0))
9850 goto out_free_counters;
9853 ca->parent = cgroup_ca(cgrp->parent);
9859 percpu_counter_destroy(&ca->cpustat[i]);
9860 free_percpu(ca->cpuusage);
9864 return ERR_PTR(-ENOMEM);
9867 /* destroy an existing cpu accounting group */
9869 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9871 struct cpuacct *ca = cgroup_ca(cgrp);
9874 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9875 percpu_counter_destroy(&ca->cpustat[i]);
9876 free_percpu(ca->cpuusage);
9880 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9882 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9885 #ifndef CONFIG_64BIT
9887 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9889 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9891 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9899 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9901 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9903 #ifndef CONFIG_64BIT
9905 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9907 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9909 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9915 /* return total cpu usage (in nanoseconds) of a group */
9916 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9918 struct cpuacct *ca = cgroup_ca(cgrp);
9919 u64 totalcpuusage = 0;
9922 for_each_present_cpu(i)
9923 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9925 return totalcpuusage;
9928 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9931 struct cpuacct *ca = cgroup_ca(cgrp);
9940 for_each_present_cpu(i)
9941 cpuacct_cpuusage_write(ca, i, 0);
9947 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9950 struct cpuacct *ca = cgroup_ca(cgroup);
9954 for_each_present_cpu(i) {
9955 percpu = cpuacct_cpuusage_read(ca, i);
9956 seq_printf(m, "%llu ", (unsigned long long) percpu);
9958 seq_printf(m, "\n");
9962 static const char *cpuacct_stat_desc[] = {
9963 [CPUACCT_STAT_USER] = "user",
9964 [CPUACCT_STAT_SYSTEM] = "system",
9967 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9968 struct cgroup_map_cb *cb)
9970 struct cpuacct *ca = cgroup_ca(cgrp);
9973 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9974 s64 val = percpu_counter_read(&ca->cpustat[i]);
9975 val = cputime64_to_clock_t(val);
9976 cb->fill(cb, cpuacct_stat_desc[i], val);
9981 static struct cftype files[] = {
9984 .read_u64 = cpuusage_read,
9985 .write_u64 = cpuusage_write,
9988 .name = "usage_percpu",
9989 .read_seq_string = cpuacct_percpu_seq_read,
9993 .read_map = cpuacct_stats_show,
9997 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9999 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10003 * charge this task's execution time to its accounting group.
10005 * called with rq->lock held.
10007 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10009 struct cpuacct *ca;
10012 if (unlikely(!cpuacct_subsys.active))
10015 cpu = task_cpu(tsk);
10021 for (; ca; ca = ca->parent) {
10022 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10023 *cpuusage += cputime;
10030 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
10031 * in cputime_t units. As a result, cpuacct_update_stats calls
10032 * percpu_counter_add with values large enough to always overflow the
10033 * per cpu batch limit causing bad SMP scalability.
10035 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
10036 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
10037 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
10040 #define CPUACCT_BATCH \
10041 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
10043 #define CPUACCT_BATCH 0
10047 * Charge the system/user time to the task's accounting group.
10049 static void cpuacct_update_stats(struct task_struct *tsk,
10050 enum cpuacct_stat_index idx, cputime_t val)
10052 struct cpuacct *ca;
10053 int batch = CPUACCT_BATCH;
10055 if (unlikely(!cpuacct_subsys.active))
10062 __percpu_counter_add(&ca->cpustat[idx], val, batch);
10068 struct cgroup_subsys cpuacct_subsys = {
10070 .create = cpuacct_create,
10071 .destroy = cpuacct_destroy,
10072 .populate = cpuacct_populate,
10073 .subsys_id = cpuacct_subsys_id,
10075 #endif /* CONFIG_CGROUP_CPUACCT */