4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
74 #include <linux/init_task.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
79 #ifdef CONFIG_PARAVIRT
80 #include <asm/paravirt.h>
83 #include "sched_cpupri.h"
84 #include "workqueue_sched.h"
85 #include "sched_autogroup.h"
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/sched.h>
91 * Convert user-nice values [ -20 ... 0 ... 19 ]
92 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
95 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
96 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
97 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
100 * 'User priority' is the nice value converted to something we
101 * can work with better when scaling various scheduler parameters,
102 * it's a [ 0 ... 39 ] range.
104 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
105 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
106 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
109 * Helpers for converting nanosecond timing to jiffy resolution
111 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
113 #define NICE_0_LOAD SCHED_LOAD_SCALE
114 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
117 * These are the 'tuning knobs' of the scheduler:
119 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
120 * Timeslices get refilled after they expire.
122 #define DEF_TIMESLICE (100 * HZ / 1000)
125 * single value that denotes runtime == period, ie unlimited time.
127 #define RUNTIME_INF ((u64)~0ULL)
129 static inline int rt_policy(int policy)
131 if (policy == SCHED_FIFO || policy == SCHED_RR)
136 static inline int task_has_rt_policy(struct task_struct *p)
138 return rt_policy(p->policy);
142 * This is the priority-queue data structure of the RT scheduling class:
144 struct rt_prio_array {
145 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
146 struct list_head queue[MAX_RT_PRIO];
149 struct rt_bandwidth {
150 /* nests inside the rq lock: */
151 raw_spinlock_t rt_runtime_lock;
154 struct hrtimer rt_period_timer;
157 static struct rt_bandwidth def_rt_bandwidth;
159 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
161 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
163 struct rt_bandwidth *rt_b =
164 container_of(timer, struct rt_bandwidth, rt_period_timer);
170 now = hrtimer_cb_get_time(timer);
171 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
176 idle = do_sched_rt_period_timer(rt_b, overrun);
179 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
183 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
185 rt_b->rt_period = ns_to_ktime(period);
186 rt_b->rt_runtime = runtime;
188 raw_spin_lock_init(&rt_b->rt_runtime_lock);
190 hrtimer_init(&rt_b->rt_period_timer,
191 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
192 rt_b->rt_period_timer.function = sched_rt_period_timer;
195 static inline int rt_bandwidth_enabled(void)
197 return sysctl_sched_rt_runtime >= 0;
200 static void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
203 ktime_t soft, hard, now;
206 if (hrtimer_active(period_timer))
209 now = hrtimer_cb_get_time(period_timer);
210 hrtimer_forward(period_timer, now, period);
212 soft = hrtimer_get_softexpires(period_timer);
213 hard = hrtimer_get_expires(period_timer);
214 delta = ktime_to_ns(ktime_sub(hard, soft));
215 __hrtimer_start_range_ns(period_timer, soft, delta,
216 HRTIMER_MODE_ABS_PINNED, 0);
220 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 raw_spin_lock(&rt_b->rt_runtime_lock);
229 start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
230 raw_spin_unlock(&rt_b->rt_runtime_lock);
233 #ifdef CONFIG_RT_GROUP_SCHED
234 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
236 hrtimer_cancel(&rt_b->rt_period_timer);
241 * sched_domains_mutex serializes calls to init_sched_domains,
242 * detach_destroy_domains and partition_sched_domains.
244 static DEFINE_MUTEX(sched_domains_mutex);
246 #ifdef CONFIG_CGROUP_SCHED
248 #include <linux/cgroup.h>
252 static LIST_HEAD(task_groups);
254 struct cfs_bandwidth {
255 #ifdef CONFIG_CFS_BANDWIDTH
259 s64 hierarchal_quota;
262 int idle, timer_active;
263 struct hrtimer period_timer, slack_timer;
264 struct list_head throttled_cfs_rq;
267 int nr_periods, nr_throttled;
272 /* task group related information */
274 struct cgroup_subsys_state css;
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity **se;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq **cfs_rq;
281 unsigned long shares;
283 atomic_t load_weight;
286 #ifdef CONFIG_RT_GROUP_SCHED
287 struct sched_rt_entity **rt_se;
288 struct rt_rq **rt_rq;
290 struct rt_bandwidth rt_bandwidth;
294 struct list_head list;
296 struct task_group *parent;
297 struct list_head siblings;
298 struct list_head children;
300 #ifdef CONFIG_SCHED_AUTOGROUP
301 struct autogroup *autogroup;
304 struct cfs_bandwidth cfs_bandwidth;
307 /* task_group_lock serializes the addition/removal of task groups */
308 static DEFINE_SPINLOCK(task_group_lock);
310 #ifdef CONFIG_FAIR_GROUP_SCHED
312 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
315 * A weight of 0 or 1 can cause arithmetics problems.
316 * A weight of a cfs_rq is the sum of weights of which entities
317 * are queued on this cfs_rq, so a weight of a entity should not be
318 * too large, so as the shares value of a task group.
319 * (The default weight is 1024 - so there's no practical
320 * limitation from this.)
322 #define MIN_SHARES (1UL << 1)
323 #define MAX_SHARES (1UL << 18)
325 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
328 /* Default task group.
329 * Every task in system belong to this group at bootup.
331 struct task_group root_task_group;
333 #endif /* CONFIG_CGROUP_SCHED */
335 /* CFS-related fields in a runqueue */
337 struct load_weight load;
338 unsigned long nr_running, h_nr_running;
343 u64 min_vruntime_copy;
346 struct rb_root tasks_timeline;
347 struct rb_node *rb_leftmost;
349 struct list_head tasks;
350 struct list_head *balance_iterator;
353 * 'curr' points to currently running entity on this cfs_rq.
354 * It is set to NULL otherwise (i.e when none are currently running).
356 struct sched_entity *curr, *next, *last, *skip;
358 #ifdef CONFIG_SCHED_DEBUG
359 unsigned int nr_spread_over;
362 #ifdef CONFIG_FAIR_GROUP_SCHED
363 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
366 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
367 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
368 * (like users, containers etc.)
370 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
371 * list is used during load balance.
374 struct list_head leaf_cfs_rq_list;
375 struct task_group *tg; /* group that "owns" this runqueue */
379 * the part of load.weight contributed by tasks
381 unsigned long task_weight;
384 * h_load = weight * f(tg)
386 * Where f(tg) is the recursive weight fraction assigned to
389 unsigned long h_load;
392 * Maintaining per-cpu shares distribution for group scheduling
394 * load_stamp is the last time we updated the load average
395 * load_last is the last time we updated the load average and saw load
396 * load_unacc_exec_time is currently unaccounted execution time
400 u64 load_stamp, load_last, load_unacc_exec_time;
402 unsigned long load_contribution;
404 #ifdef CONFIG_CFS_BANDWIDTH
407 s64 runtime_remaining;
409 u64 throttled_timestamp;
410 int throttled, throttle_count;
411 struct list_head throttled_list;
416 #ifdef CONFIG_FAIR_GROUP_SCHED
417 #ifdef CONFIG_CFS_BANDWIDTH
418 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
420 return &tg->cfs_bandwidth;
423 static inline u64 default_cfs_period(void);
424 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
425 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
427 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
429 struct cfs_bandwidth *cfs_b =
430 container_of(timer, struct cfs_bandwidth, slack_timer);
431 do_sched_cfs_slack_timer(cfs_b);
433 return HRTIMER_NORESTART;
436 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
438 struct cfs_bandwidth *cfs_b =
439 container_of(timer, struct cfs_bandwidth, period_timer);
445 now = hrtimer_cb_get_time(timer);
446 overrun = hrtimer_forward(timer, now, cfs_b->period);
451 idle = do_sched_cfs_period_timer(cfs_b, overrun);
454 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
457 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
459 raw_spin_lock_init(&cfs_b->lock);
461 cfs_b->quota = RUNTIME_INF;
462 cfs_b->period = ns_to_ktime(default_cfs_period());
464 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
465 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
466 cfs_b->period_timer.function = sched_cfs_period_timer;
467 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
468 cfs_b->slack_timer.function = sched_cfs_slack_timer;
471 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
473 cfs_rq->runtime_enabled = 0;
474 INIT_LIST_HEAD(&cfs_rq->throttled_list);
477 /* requires cfs_b->lock, may release to reprogram timer */
478 static void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
481 * The timer may be active because we're trying to set a new bandwidth
482 * period or because we're racing with the tear-down path
483 * (timer_active==0 becomes visible before the hrtimer call-back
484 * terminates). In either case we ensure that it's re-programmed
486 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
487 raw_spin_unlock(&cfs_b->lock);
488 /* ensure cfs_b->lock is available while we wait */
489 hrtimer_cancel(&cfs_b->period_timer);
491 raw_spin_lock(&cfs_b->lock);
492 /* if someone else restarted the timer then we're done */
493 if (cfs_b->timer_active)
497 cfs_b->timer_active = 1;
498 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
501 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
503 hrtimer_cancel(&cfs_b->period_timer);
504 hrtimer_cancel(&cfs_b->slack_timer);
507 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
508 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
509 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
511 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
515 #endif /* CONFIG_CFS_BANDWIDTH */
516 #endif /* CONFIG_FAIR_GROUP_SCHED */
518 /* Real-Time classes' related field in a runqueue: */
520 struct rt_prio_array active;
521 unsigned long rt_nr_running;
522 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
524 int curr; /* highest queued rt task prio */
526 int next; /* next highest */
531 unsigned long rt_nr_migratory;
532 unsigned long rt_nr_total;
534 struct plist_head pushable_tasks;
539 /* Nests inside the rq lock: */
540 raw_spinlock_t rt_runtime_lock;
542 #ifdef CONFIG_RT_GROUP_SCHED
543 unsigned long rt_nr_boosted;
546 struct list_head leaf_rt_rq_list;
547 struct task_group *tg;
554 * We add the notion of a root-domain which will be used to define per-domain
555 * variables. Each exclusive cpuset essentially defines an island domain by
556 * fully partitioning the member cpus from any other cpuset. Whenever a new
557 * exclusive cpuset is created, we also create and attach a new root-domain
566 cpumask_var_t online;
569 * The "RT overload" flag: it gets set if a CPU has more than
570 * one runnable RT task.
572 cpumask_var_t rto_mask;
573 struct cpupri cpupri;
577 * By default the system creates a single root-domain with all cpus as
578 * members (mimicking the global state we have today).
580 static struct root_domain def_root_domain;
582 #endif /* CONFIG_SMP */
585 * This is the main, per-CPU runqueue data structure.
587 * Locking rule: those places that want to lock multiple runqueues
588 * (such as the load balancing or the thread migration code), lock
589 * acquire operations must be ordered by ascending &runqueue.
596 * nr_running and cpu_load should be in the same cacheline because
597 * remote CPUs use both these fields when doing load calculation.
599 unsigned long nr_running;
600 #define CPU_LOAD_IDX_MAX 5
601 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
602 unsigned long last_load_update_tick;
605 unsigned char nohz_balance_kick;
607 int skip_clock_update;
609 /* capture load from *all* tasks on this cpu: */
610 struct load_weight load;
611 unsigned long nr_load_updates;
617 #ifdef CONFIG_FAIR_GROUP_SCHED
618 /* list of leaf cfs_rq on this cpu: */
619 struct list_head leaf_cfs_rq_list;
621 #ifdef CONFIG_RT_GROUP_SCHED
622 struct list_head leaf_rt_rq_list;
626 * This is part of a global counter where only the total sum
627 * over all CPUs matters. A task can increase this counter on
628 * one CPU and if it got migrated afterwards it may decrease
629 * it on another CPU. Always updated under the runqueue lock:
631 unsigned long nr_uninterruptible;
633 struct task_struct *curr, *idle, *stop;
634 unsigned long next_balance;
635 struct mm_struct *prev_mm;
643 struct root_domain *rd;
644 struct sched_domain *sd;
646 unsigned long cpu_power;
648 unsigned char idle_balance;
649 /* For active balancing */
653 struct cpu_stop_work active_balance_work;
654 /* cpu of this runqueue: */
664 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
667 #ifdef CONFIG_PARAVIRT
670 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
671 u64 prev_steal_time_rq;
674 /* calc_load related fields */
675 unsigned long calc_load_update;
676 long calc_load_active;
678 #ifdef CONFIG_SCHED_HRTICK
680 int hrtick_csd_pending;
681 struct call_single_data hrtick_csd;
683 struct hrtimer hrtick_timer;
686 #ifdef CONFIG_SCHEDSTATS
688 struct sched_info rq_sched_info;
689 unsigned long long rq_cpu_time;
690 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
692 /* sys_sched_yield() stats */
693 unsigned int yld_count;
695 /* schedule() stats */
696 unsigned int sched_switch;
697 unsigned int sched_count;
698 unsigned int sched_goidle;
700 /* try_to_wake_up() stats */
701 unsigned int ttwu_count;
702 unsigned int ttwu_local;
706 struct llist_head wake_list;
710 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
713 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
715 static inline int cpu_of(struct rq *rq)
724 #define rcu_dereference_check_sched_domain(p) \
725 rcu_dereference_check((p), \
726 lockdep_is_held(&sched_domains_mutex))
729 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
730 * See detach_destroy_domains: synchronize_sched for details.
732 * The domain tree of any CPU may only be accessed from within
733 * preempt-disabled sections.
735 #define for_each_domain(cpu, __sd) \
736 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
738 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
739 #define this_rq() (&__get_cpu_var(runqueues))
740 #define task_rq(p) cpu_rq(task_cpu(p))
741 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
742 #define raw_rq() (&__raw_get_cpu_var(runqueues))
744 #ifdef CONFIG_CGROUP_SCHED
747 * Return the group to which this tasks belongs.
749 * We cannot use task_subsys_state() and friends because the cgroup
750 * subsystem changes that value before the cgroup_subsys::attach() method
751 * is called, therefore we cannot pin it and might observe the wrong value.
753 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
754 * core changes this before calling sched_move_task().
756 * Instead we use a 'copy' which is updated from sched_move_task() while
757 * holding both task_struct::pi_lock and rq::lock.
759 static inline struct task_group *task_group(struct task_struct *p)
761 return p->sched_task_group;
764 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
765 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
767 #ifdef CONFIG_FAIR_GROUP_SCHED
768 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
769 p->se.parent = task_group(p)->se[cpu];
772 #ifdef CONFIG_RT_GROUP_SCHED
773 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
774 p->rt.parent = task_group(p)->rt_se[cpu];
778 #else /* CONFIG_CGROUP_SCHED */
780 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
781 static inline struct task_group *task_group(struct task_struct *p)
786 #endif /* CONFIG_CGROUP_SCHED */
788 static void update_rq_clock_task(struct rq *rq, s64 delta);
790 static void update_rq_clock(struct rq *rq)
794 if (rq->skip_clock_update > 0)
797 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
799 update_rq_clock_task(rq, delta);
803 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
805 #ifdef CONFIG_SCHED_DEBUG
806 # define const_debug __read_mostly
808 # define const_debug static const
812 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
813 * @cpu: the processor in question.
815 * This interface allows printk to be called with the runqueue lock
816 * held and know whether or not it is OK to wake up the klogd.
818 int runqueue_is_locked(int cpu)
820 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
824 * Debugging: various feature bits
827 #define SCHED_FEAT(name, enabled) \
828 __SCHED_FEAT_##name ,
831 #include "sched_features.h"
836 #define SCHED_FEAT(name, enabled) \
837 (1UL << __SCHED_FEAT_##name) * enabled |
839 const_debug unsigned int sysctl_sched_features =
840 #include "sched_features.h"
845 #ifdef CONFIG_SCHED_DEBUG
846 #define SCHED_FEAT(name, enabled) \
849 static __read_mostly char *sched_feat_names[] = {
850 #include "sched_features.h"
856 static int sched_feat_show(struct seq_file *m, void *v)
860 for (i = 0; sched_feat_names[i]; i++) {
861 if (!(sysctl_sched_features & (1UL << i)))
863 seq_printf(m, "%s ", sched_feat_names[i]);
871 sched_feat_write(struct file *filp, const char __user *ubuf,
872 size_t cnt, loff_t *ppos)
882 if (copy_from_user(&buf, ubuf, cnt))
888 if (strncmp(cmp, "NO_", 3) == 0) {
893 for (i = 0; sched_feat_names[i]; i++) {
894 if (strcmp(cmp, sched_feat_names[i]) == 0) {
896 sysctl_sched_features &= ~(1UL << i);
898 sysctl_sched_features |= (1UL << i);
903 if (!sched_feat_names[i])
911 static int sched_feat_open(struct inode *inode, struct file *filp)
913 return single_open(filp, sched_feat_show, NULL);
916 static const struct file_operations sched_feat_fops = {
917 .open = sched_feat_open,
918 .write = sched_feat_write,
921 .release = single_release,
924 static __init int sched_init_debug(void)
926 debugfs_create_file("sched_features", 0644, NULL, NULL,
931 late_initcall(sched_init_debug);
935 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
938 * Number of tasks to iterate in a single balance run.
939 * Limited because this is done with IRQs disabled.
941 const_debug unsigned int sysctl_sched_nr_migrate = 32;
944 * period over which we average the RT time consumption, measured
949 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
952 * period over which we measure -rt task cpu usage in us.
955 unsigned int sysctl_sched_rt_period = 1000000;
957 static __read_mostly int scheduler_running;
960 * part of the period that we allow rt tasks to run in us.
963 int sysctl_sched_rt_runtime = 950000;
965 static inline u64 global_rt_period(void)
967 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
970 static inline u64 global_rt_runtime(void)
972 if (sysctl_sched_rt_runtime < 0)
975 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
978 #ifndef prepare_arch_switch
979 # define prepare_arch_switch(next) do { } while (0)
981 #ifndef finish_arch_switch
982 # define finish_arch_switch(prev) do { } while (0)
985 static inline int task_current(struct rq *rq, struct task_struct *p)
987 return rq->curr == p;
990 static inline int task_running(struct rq *rq, struct task_struct *p)
995 return task_current(rq, p);
999 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1000 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1004 * We can optimise this out completely for !SMP, because the
1005 * SMP rebalancing from interrupt is the only thing that cares
1012 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1016 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1017 * We must ensure this doesn't happen until the switch is completely
1023 #ifdef CONFIG_DEBUG_SPINLOCK
1024 /* this is a valid case when another task releases the spinlock */
1025 rq->lock.owner = current;
1028 * If we are tracking spinlock dependencies then we have to
1029 * fix up the runqueue lock - which gets 'carried over' from
1030 * prev into current:
1032 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1034 raw_spin_unlock_irq(&rq->lock);
1037 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1038 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1042 * We can optimise this out completely for !SMP, because the
1043 * SMP rebalancing from interrupt is the only thing that cares
1048 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1049 raw_spin_unlock_irq(&rq->lock);
1051 raw_spin_unlock(&rq->lock);
1055 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1059 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1060 * We must ensure this doesn't happen until the switch is completely
1066 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1070 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1073 * __task_rq_lock - lock the rq @p resides on.
1075 static inline struct rq *__task_rq_lock(struct task_struct *p)
1076 __acquires(rq->lock)
1080 lockdep_assert_held(&p->pi_lock);
1084 raw_spin_lock(&rq->lock);
1085 if (likely(rq == task_rq(p)))
1087 raw_spin_unlock(&rq->lock);
1092 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1094 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1095 __acquires(p->pi_lock)
1096 __acquires(rq->lock)
1101 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1103 raw_spin_lock(&rq->lock);
1104 if (likely(rq == task_rq(p)))
1106 raw_spin_unlock(&rq->lock);
1107 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1111 static void __task_rq_unlock(struct rq *rq)
1112 __releases(rq->lock)
1114 raw_spin_unlock(&rq->lock);
1118 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1119 __releases(rq->lock)
1120 __releases(p->pi_lock)
1122 raw_spin_unlock(&rq->lock);
1123 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1127 * this_rq_lock - lock this runqueue and disable interrupts.
1129 static struct rq *this_rq_lock(void)
1130 __acquires(rq->lock)
1134 local_irq_disable();
1136 raw_spin_lock(&rq->lock);
1141 #ifdef CONFIG_SCHED_HRTICK
1143 * Use HR-timers to deliver accurate preemption points.
1145 * Its all a bit involved since we cannot program an hrt while holding the
1146 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1149 * When we get rescheduled we reprogram the hrtick_timer outside of the
1155 * - enabled by features
1156 * - hrtimer is actually high res
1158 static inline int hrtick_enabled(struct rq *rq)
1160 if (!sched_feat(HRTICK))
1162 if (!cpu_active(cpu_of(rq)))
1164 return hrtimer_is_hres_active(&rq->hrtick_timer);
1167 static void hrtick_clear(struct rq *rq)
1169 if (hrtimer_active(&rq->hrtick_timer))
1170 hrtimer_cancel(&rq->hrtick_timer);
1174 * High-resolution timer tick.
1175 * Runs from hardirq context with interrupts disabled.
1177 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1179 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1181 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1183 raw_spin_lock(&rq->lock);
1184 update_rq_clock(rq);
1185 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1186 raw_spin_unlock(&rq->lock);
1188 return HRTIMER_NORESTART;
1193 * called from hardirq (IPI) context
1195 static void __hrtick_start(void *arg)
1197 struct rq *rq = arg;
1199 raw_spin_lock(&rq->lock);
1200 hrtimer_restart(&rq->hrtick_timer);
1201 rq->hrtick_csd_pending = 0;
1202 raw_spin_unlock(&rq->lock);
1206 * Called to set the hrtick timer state.
1208 * called with rq->lock held and irqs disabled
1210 static void hrtick_start(struct rq *rq, u64 delay)
1212 struct hrtimer *timer = &rq->hrtick_timer;
1213 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1215 hrtimer_set_expires(timer, time);
1217 if (rq == this_rq()) {
1218 hrtimer_restart(timer);
1219 } else if (!rq->hrtick_csd_pending) {
1220 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1221 rq->hrtick_csd_pending = 1;
1226 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1228 int cpu = (int)(long)hcpu;
1231 case CPU_UP_CANCELED:
1232 case CPU_UP_CANCELED_FROZEN:
1233 case CPU_DOWN_PREPARE:
1234 case CPU_DOWN_PREPARE_FROZEN:
1236 case CPU_DEAD_FROZEN:
1237 hrtick_clear(cpu_rq(cpu));
1244 static __init void init_hrtick(void)
1246 hotcpu_notifier(hotplug_hrtick, 0);
1250 * Called to set the hrtick timer state.
1252 * called with rq->lock held and irqs disabled
1254 static void hrtick_start(struct rq *rq, u64 delay)
1256 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1257 HRTIMER_MODE_REL_PINNED, 0);
1260 static inline void init_hrtick(void)
1263 #endif /* CONFIG_SMP */
1265 static void init_rq_hrtick(struct rq *rq)
1268 rq->hrtick_csd_pending = 0;
1270 rq->hrtick_csd.flags = 0;
1271 rq->hrtick_csd.func = __hrtick_start;
1272 rq->hrtick_csd.info = rq;
1275 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1276 rq->hrtick_timer.function = hrtick;
1278 #else /* CONFIG_SCHED_HRTICK */
1279 static inline void hrtick_clear(struct rq *rq)
1283 static inline void init_rq_hrtick(struct rq *rq)
1287 static inline void init_hrtick(void)
1290 #endif /* CONFIG_SCHED_HRTICK */
1293 * resched_task - mark a task 'to be rescheduled now'.
1295 * On UP this means the setting of the need_resched flag, on SMP it
1296 * might also involve a cross-CPU call to trigger the scheduler on
1301 #ifndef tsk_is_polling
1302 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1305 static void resched_task(struct task_struct *p)
1309 assert_raw_spin_locked(&task_rq(p)->lock);
1311 if (test_tsk_need_resched(p))
1314 set_tsk_need_resched(p);
1317 if (cpu == smp_processor_id())
1320 /* NEED_RESCHED must be visible before we test polling */
1322 if (!tsk_is_polling(p))
1323 smp_send_reschedule(cpu);
1326 static void resched_cpu(int cpu)
1328 struct rq *rq = cpu_rq(cpu);
1329 unsigned long flags;
1331 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1333 resched_task(cpu_curr(cpu));
1334 raw_spin_unlock_irqrestore(&rq->lock, flags);
1339 * In the semi idle case, use the nearest busy cpu for migrating timers
1340 * from an idle cpu. This is good for power-savings.
1342 * We don't do similar optimization for completely idle system, as
1343 * selecting an idle cpu will add more delays to the timers than intended
1344 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1346 int get_nohz_timer_target(void)
1348 int cpu = smp_processor_id();
1350 struct sched_domain *sd;
1353 for_each_domain(cpu, sd) {
1354 for_each_cpu(i, sched_domain_span(sd)) {
1366 * When add_timer_on() enqueues a timer into the timer wheel of an
1367 * idle CPU then this timer might expire before the next timer event
1368 * which is scheduled to wake up that CPU. In case of a completely
1369 * idle system the next event might even be infinite time into the
1370 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1371 * leaves the inner idle loop so the newly added timer is taken into
1372 * account when the CPU goes back to idle and evaluates the timer
1373 * wheel for the next timer event.
1375 void wake_up_idle_cpu(int cpu)
1377 struct rq *rq = cpu_rq(cpu);
1379 if (cpu == smp_processor_id())
1383 * This is safe, as this function is called with the timer
1384 * wheel base lock of (cpu) held. When the CPU is on the way
1385 * to idle and has not yet set rq->curr to idle then it will
1386 * be serialized on the timer wheel base lock and take the new
1387 * timer into account automatically.
1389 if (rq->curr != rq->idle)
1393 * We can set TIF_RESCHED on the idle task of the other CPU
1394 * lockless. The worst case is that the other CPU runs the
1395 * idle task through an additional NOOP schedule()
1397 set_tsk_need_resched(rq->idle);
1399 /* NEED_RESCHED must be visible before we test polling */
1401 if (!tsk_is_polling(rq->idle))
1402 smp_send_reschedule(cpu);
1405 static inline bool got_nohz_idle_kick(void)
1407 return idle_cpu(smp_processor_id()) && this_rq()->nohz_balance_kick;
1410 #else /* CONFIG_NO_HZ */
1412 static inline bool got_nohz_idle_kick(void)
1417 #endif /* CONFIG_NO_HZ */
1419 static u64 sched_avg_period(void)
1421 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1424 static void sched_avg_update(struct rq *rq)
1426 s64 period = sched_avg_period();
1428 while ((s64)(rq->clock - rq->age_stamp) > period) {
1430 * Inline assembly required to prevent the compiler
1431 * optimising this loop into a divmod call.
1432 * See __iter_div_u64_rem() for another example of this.
1434 asm("" : "+rm" (rq->age_stamp));
1435 rq->age_stamp += period;
1440 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1442 rq->rt_avg += rt_delta;
1443 sched_avg_update(rq);
1446 #else /* !CONFIG_SMP */
1447 static void resched_task(struct task_struct *p)
1449 assert_raw_spin_locked(&task_rq(p)->lock);
1450 set_tsk_need_resched(p);
1453 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1457 static void sched_avg_update(struct rq *rq)
1460 #endif /* CONFIG_SMP */
1462 #if BITS_PER_LONG == 32
1463 # define WMULT_CONST (~0UL)
1465 # define WMULT_CONST (1UL << 32)
1468 #define WMULT_SHIFT 32
1471 * Shift right and round:
1473 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1476 * delta *= weight / lw
1478 static unsigned long
1479 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1480 struct load_weight *lw)
1485 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1486 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1487 * 2^SCHED_LOAD_RESOLUTION.
1489 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1490 tmp = (u64)delta_exec * scale_load_down(weight);
1492 tmp = (u64)delta_exec;
1494 if (!lw->inv_weight) {
1495 unsigned long w = scale_load_down(lw->weight);
1497 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1499 else if (unlikely(!w))
1500 lw->inv_weight = WMULT_CONST;
1502 lw->inv_weight = WMULT_CONST / w;
1506 * Check whether we'd overflow the 64-bit multiplication:
1508 if (unlikely(tmp > WMULT_CONST))
1509 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1512 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1514 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1517 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1523 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1529 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1536 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1537 * of tasks with abnormal "nice" values across CPUs the contribution that
1538 * each task makes to its run queue's load is weighted according to its
1539 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1540 * scaled version of the new time slice allocation that they receive on time
1544 #define WEIGHT_IDLEPRIO 3
1545 #define WMULT_IDLEPRIO 1431655765
1548 * Nice levels are multiplicative, with a gentle 10% change for every
1549 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1550 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1551 * that remained on nice 0.
1553 * The "10% effect" is relative and cumulative: from _any_ nice level,
1554 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1555 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1556 * If a task goes up by ~10% and another task goes down by ~10% then
1557 * the relative distance between them is ~25%.)
1559 static const int prio_to_weight[40] = {
1560 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1561 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1562 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1563 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1564 /* 0 */ 1024, 820, 655, 526, 423,
1565 /* 5 */ 335, 272, 215, 172, 137,
1566 /* 10 */ 110, 87, 70, 56, 45,
1567 /* 15 */ 36, 29, 23, 18, 15,
1571 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1573 * In cases where the weight does not change often, we can use the
1574 * precalculated inverse to speed up arithmetics by turning divisions
1575 * into multiplications:
1577 static const u32 prio_to_wmult[40] = {
1578 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1579 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1580 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1581 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1582 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1583 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1584 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1585 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1588 /* Time spent by the tasks of the cpu accounting group executing in ... */
1589 enum cpuacct_stat_index {
1590 CPUACCT_STAT_USER, /* ... user mode */
1591 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1593 CPUACCT_STAT_NSTATS,
1596 #ifdef CONFIG_CGROUP_CPUACCT
1597 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1598 static void cpuacct_update_stats(struct task_struct *tsk,
1599 enum cpuacct_stat_index idx, cputime_t val);
1601 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1602 static inline void cpuacct_update_stats(struct task_struct *tsk,
1603 enum cpuacct_stat_index idx, cputime_t val) {}
1606 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1608 update_load_add(&rq->load, load);
1611 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1613 update_load_sub(&rq->load, load);
1616 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1617 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1618 typedef int (*tg_visitor)(struct task_group *, void *);
1621 * Iterate task_group tree rooted at *from, calling @down when first entering a
1622 * node and @up when leaving it for the final time.
1624 * Caller must hold rcu_lock or sufficient equivalent.
1626 static int walk_tg_tree_from(struct task_group *from,
1627 tg_visitor down, tg_visitor up, void *data)
1629 struct task_group *parent, *child;
1635 ret = (*down)(parent, data);
1638 list_for_each_entry_rcu(child, &parent->children, siblings) {
1645 ret = (*up)(parent, data);
1646 if (ret || parent == from)
1650 parent = parent->parent;
1658 * Iterate the full tree, calling @down when first entering a node and @up when
1659 * leaving it for the final time.
1661 * Caller must hold rcu_lock or sufficient equivalent.
1664 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1666 return walk_tg_tree_from(&root_task_group, down, up, data);
1669 static int tg_nop(struct task_group *tg, void *data)
1676 /* Used instead of source_load when we know the type == 0 */
1677 static unsigned long weighted_cpuload(const int cpu)
1679 return cpu_rq(cpu)->load.weight;
1683 * Return a low guess at the load of a migration-source cpu weighted
1684 * according to the scheduling class and "nice" value.
1686 * We want to under-estimate the load of migration sources, to
1687 * balance conservatively.
1689 static unsigned long source_load(int cpu, int type)
1691 struct rq *rq = cpu_rq(cpu);
1692 unsigned long total = weighted_cpuload(cpu);
1694 if (type == 0 || !sched_feat(LB_BIAS))
1697 return min(rq->cpu_load[type-1], total);
1701 * Return a high guess at the load of a migration-target cpu weighted
1702 * according to the scheduling class and "nice" value.
1704 static unsigned long target_load(int cpu, int type)
1706 struct rq *rq = cpu_rq(cpu);
1707 unsigned long total = weighted_cpuload(cpu);
1709 if (type == 0 || !sched_feat(LB_BIAS))
1712 return max(rq->cpu_load[type-1], total);
1715 static unsigned long power_of(int cpu)
1717 return cpu_rq(cpu)->cpu_power;
1720 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1722 static unsigned long cpu_avg_load_per_task(int cpu)
1724 struct rq *rq = cpu_rq(cpu);
1725 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1728 return rq->load.weight / nr_running;
1733 #ifdef CONFIG_PREEMPT
1735 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1738 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1739 * way at the expense of forcing extra atomic operations in all
1740 * invocations. This assures that the double_lock is acquired using the
1741 * same underlying policy as the spinlock_t on this architecture, which
1742 * reduces latency compared to the unfair variant below. However, it
1743 * also adds more overhead and therefore may reduce throughput.
1745 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1746 __releases(this_rq->lock)
1747 __acquires(busiest->lock)
1748 __acquires(this_rq->lock)
1750 raw_spin_unlock(&this_rq->lock);
1751 double_rq_lock(this_rq, busiest);
1758 * Unfair double_lock_balance: Optimizes throughput at the expense of
1759 * latency by eliminating extra atomic operations when the locks are
1760 * already in proper order on entry. This favors lower cpu-ids and will
1761 * grant the double lock to lower cpus over higher ids under contention,
1762 * regardless of entry order into the function.
1764 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1765 __releases(this_rq->lock)
1766 __acquires(busiest->lock)
1767 __acquires(this_rq->lock)
1771 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1772 if (busiest < this_rq) {
1773 raw_spin_unlock(&this_rq->lock);
1774 raw_spin_lock(&busiest->lock);
1775 raw_spin_lock_nested(&this_rq->lock,
1776 SINGLE_DEPTH_NESTING);
1779 raw_spin_lock_nested(&busiest->lock,
1780 SINGLE_DEPTH_NESTING);
1785 #endif /* CONFIG_PREEMPT */
1788 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1790 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1792 if (unlikely(!irqs_disabled())) {
1793 /* printk() doesn't work good under rq->lock */
1794 raw_spin_unlock(&this_rq->lock);
1798 return _double_lock_balance(this_rq, busiest);
1801 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1802 __releases(busiest->lock)
1804 raw_spin_unlock(&busiest->lock);
1805 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1809 * double_rq_lock - safely lock two runqueues
1811 * Note this does not disable interrupts like task_rq_lock,
1812 * you need to do so manually before calling.
1814 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1815 __acquires(rq1->lock)
1816 __acquires(rq2->lock)
1818 BUG_ON(!irqs_disabled());
1820 raw_spin_lock(&rq1->lock);
1821 __acquire(rq2->lock); /* Fake it out ;) */
1824 raw_spin_lock(&rq1->lock);
1825 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1827 raw_spin_lock(&rq2->lock);
1828 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1834 * double_rq_unlock - safely unlock two runqueues
1836 * Note this does not restore interrupts like task_rq_unlock,
1837 * you need to do so manually after calling.
1839 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1840 __releases(rq1->lock)
1841 __releases(rq2->lock)
1843 raw_spin_unlock(&rq1->lock);
1845 raw_spin_unlock(&rq2->lock);
1847 __release(rq2->lock);
1850 #else /* CONFIG_SMP */
1853 * double_rq_lock - safely lock two runqueues
1855 * Note this does not disable interrupts like task_rq_lock,
1856 * you need to do so manually before calling.
1858 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1859 __acquires(rq1->lock)
1860 __acquires(rq2->lock)
1862 BUG_ON(!irqs_disabled());
1864 raw_spin_lock(&rq1->lock);
1865 __acquire(rq2->lock); /* Fake it out ;) */
1869 * double_rq_unlock - safely unlock two runqueues
1871 * Note this does not restore interrupts like task_rq_unlock,
1872 * you need to do so manually after calling.
1874 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1875 __releases(rq1->lock)
1876 __releases(rq2->lock)
1879 raw_spin_unlock(&rq1->lock);
1880 __release(rq2->lock);
1885 static void update_sysctl(void);
1886 static int get_update_sysctl_factor(void);
1887 static void update_idle_cpu_load(struct rq *this_rq);
1889 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1891 set_task_rq(p, cpu);
1894 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1895 * successfully executed on another CPU. We must ensure that updates of
1896 * per-task data have been completed by this moment.
1899 task_thread_info(p)->cpu = cpu;
1903 static const struct sched_class rt_sched_class;
1905 #define sched_class_highest (&stop_sched_class)
1906 #define for_each_class(class) \
1907 for (class = sched_class_highest; class; class = class->next)
1909 #include "sched_stats.h"
1911 static void inc_nr_running(struct rq *rq)
1916 static void dec_nr_running(struct rq *rq)
1921 static void set_load_weight(struct task_struct *p)
1923 int prio = p->static_prio - MAX_RT_PRIO;
1924 struct load_weight *load = &p->se.load;
1927 * SCHED_IDLE tasks get minimal weight:
1929 if (p->policy == SCHED_IDLE) {
1930 load->weight = scale_load(WEIGHT_IDLEPRIO);
1931 load->inv_weight = WMULT_IDLEPRIO;
1935 load->weight = scale_load(prio_to_weight[prio]);
1936 load->inv_weight = prio_to_wmult[prio];
1939 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1941 update_rq_clock(rq);
1942 sched_info_queued(p);
1943 p->sched_class->enqueue_task(rq, p, flags);
1946 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1948 update_rq_clock(rq);
1949 sched_info_dequeued(p);
1950 p->sched_class->dequeue_task(rq, p, flags);
1954 * activate_task - move a task to the runqueue.
1956 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1958 if (task_contributes_to_load(p))
1959 rq->nr_uninterruptible--;
1961 enqueue_task(rq, p, flags);
1965 * deactivate_task - remove a task from the runqueue.
1967 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1969 if (task_contributes_to_load(p))
1970 rq->nr_uninterruptible++;
1972 dequeue_task(rq, p, flags);
1975 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1978 * There are no locks covering percpu hardirq/softirq time.
1979 * They are only modified in account_system_vtime, on corresponding CPU
1980 * with interrupts disabled. So, writes are safe.
1981 * They are read and saved off onto struct rq in update_rq_clock().
1982 * This may result in other CPU reading this CPU's irq time and can
1983 * race with irq/account_system_vtime on this CPU. We would either get old
1984 * or new value with a side effect of accounting a slice of irq time to wrong
1985 * task when irq is in progress while we read rq->clock. That is a worthy
1986 * compromise in place of having locks on each irq in account_system_time.
1988 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1989 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1991 static DEFINE_PER_CPU(u64, irq_start_time);
1992 static int sched_clock_irqtime;
1994 void enable_sched_clock_irqtime(void)
1996 sched_clock_irqtime = 1;
1999 void disable_sched_clock_irqtime(void)
2001 sched_clock_irqtime = 0;
2004 #ifndef CONFIG_64BIT
2005 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
2007 static inline void irq_time_write_begin(void)
2009 __this_cpu_inc(irq_time_seq.sequence);
2013 static inline void irq_time_write_end(void)
2016 __this_cpu_inc(irq_time_seq.sequence);
2019 static inline u64 irq_time_read(int cpu)
2025 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
2026 irq_time = per_cpu(cpu_softirq_time, cpu) +
2027 per_cpu(cpu_hardirq_time, cpu);
2028 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
2032 #else /* CONFIG_64BIT */
2033 static inline void irq_time_write_begin(void)
2037 static inline void irq_time_write_end(void)
2041 static inline u64 irq_time_read(int cpu)
2043 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
2045 #endif /* CONFIG_64BIT */
2048 * Called before incrementing preempt_count on {soft,}irq_enter
2049 * and before decrementing preempt_count on {soft,}irq_exit.
2051 void account_system_vtime(struct task_struct *curr)
2053 unsigned long flags;
2057 if (!sched_clock_irqtime)
2060 local_irq_save(flags);
2062 cpu = smp_processor_id();
2063 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2064 __this_cpu_add(irq_start_time, delta);
2066 irq_time_write_begin();
2068 * We do not account for softirq time from ksoftirqd here.
2069 * We want to continue accounting softirq time to ksoftirqd thread
2070 * in that case, so as not to confuse scheduler with a special task
2071 * that do not consume any time, but still wants to run.
2073 if (hardirq_count())
2074 __this_cpu_add(cpu_hardirq_time, delta);
2075 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
2076 __this_cpu_add(cpu_softirq_time, delta);
2078 irq_time_write_end();
2079 local_irq_restore(flags);
2081 EXPORT_SYMBOL_GPL(account_system_vtime);
2083 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2085 #ifdef CONFIG_PARAVIRT
2086 static inline u64 steal_ticks(u64 steal)
2088 if (unlikely(steal > NSEC_PER_SEC))
2089 return div_u64(steal, TICK_NSEC);
2091 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
2095 static void update_rq_clock_task(struct rq *rq, s64 delta)
2098 * In theory, the compile should just see 0 here, and optimize out the call
2099 * to sched_rt_avg_update. But I don't trust it...
2101 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2102 s64 steal = 0, irq_delta = 0;
2104 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2105 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2108 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2109 * this case when a previous update_rq_clock() happened inside a
2110 * {soft,}irq region.
2112 * When this happens, we stop ->clock_task and only update the
2113 * prev_irq_time stamp to account for the part that fit, so that a next
2114 * update will consume the rest. This ensures ->clock_task is
2117 * It does however cause some slight miss-attribution of {soft,}irq
2118 * time, a more accurate solution would be to update the irq_time using
2119 * the current rq->clock timestamp, except that would require using
2122 if (irq_delta > delta)
2125 rq->prev_irq_time += irq_delta;
2128 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2129 if (static_branch((¶virt_steal_rq_enabled))) {
2132 steal = paravirt_steal_clock(cpu_of(rq));
2133 steal -= rq->prev_steal_time_rq;
2135 if (unlikely(steal > delta))
2138 st = steal_ticks(steal);
2139 steal = st * TICK_NSEC;
2141 rq->prev_steal_time_rq += steal;
2147 rq->clock_task += delta;
2149 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2150 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2151 sched_rt_avg_update(rq, irq_delta + steal);
2155 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2156 static int irqtime_account_hi_update(void)
2158 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2159 unsigned long flags;
2163 local_irq_save(flags);
2164 latest_ns = this_cpu_read(cpu_hardirq_time);
2165 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2167 local_irq_restore(flags);
2171 static int irqtime_account_si_update(void)
2173 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2174 unsigned long flags;
2178 local_irq_save(flags);
2179 latest_ns = this_cpu_read(cpu_softirq_time);
2180 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2182 local_irq_restore(flags);
2186 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2188 #define sched_clock_irqtime (0)
2193 static void unthrottle_offline_cfs_rqs(struct rq *rq);
2196 #include "sched_idletask.c"
2197 #include "sched_fair.c"
2198 #include "sched_rt.c"
2199 #include "sched_autogroup.c"
2200 #include "sched_stoptask.c"
2201 #ifdef CONFIG_SCHED_DEBUG
2202 # include "sched_debug.c"
2205 void sched_set_stop_task(int cpu, struct task_struct *stop)
2207 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2208 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2212 * Make it appear like a SCHED_FIFO task, its something
2213 * userspace knows about and won't get confused about.
2215 * Also, it will make PI more or less work without too
2216 * much confusion -- but then, stop work should not
2217 * rely on PI working anyway.
2219 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2221 stop->sched_class = &stop_sched_class;
2224 cpu_rq(cpu)->stop = stop;
2228 * Reset it back to a normal scheduling class so that
2229 * it can die in pieces.
2231 old_stop->sched_class = &rt_sched_class;
2236 * __normal_prio - return the priority that is based on the static prio
2238 static inline int __normal_prio(struct task_struct *p)
2240 return p->static_prio;
2244 * Calculate the expected normal priority: i.e. priority
2245 * without taking RT-inheritance into account. Might be
2246 * boosted by interactivity modifiers. Changes upon fork,
2247 * setprio syscalls, and whenever the interactivity
2248 * estimator recalculates.
2250 static inline int normal_prio(struct task_struct *p)
2254 if (task_has_rt_policy(p))
2255 prio = MAX_RT_PRIO-1 - p->rt_priority;
2257 prio = __normal_prio(p);
2262 * Calculate the current priority, i.e. the priority
2263 * taken into account by the scheduler. This value might
2264 * be boosted by RT tasks, or might be boosted by
2265 * interactivity modifiers. Will be RT if the task got
2266 * RT-boosted. If not then it returns p->normal_prio.
2268 static int effective_prio(struct task_struct *p)
2270 p->normal_prio = normal_prio(p);
2272 * If we are RT tasks or we were boosted to RT priority,
2273 * keep the priority unchanged. Otherwise, update priority
2274 * to the normal priority:
2276 if (!rt_prio(p->prio))
2277 return p->normal_prio;
2282 * task_curr - is this task currently executing on a CPU?
2283 * @p: the task in question.
2285 inline int task_curr(const struct task_struct *p)
2287 return cpu_curr(task_cpu(p)) == p;
2290 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2291 const struct sched_class *prev_class,
2294 if (prev_class != p->sched_class) {
2295 if (prev_class->switched_from)
2296 prev_class->switched_from(rq, p);
2297 p->sched_class->switched_to(rq, p);
2298 } else if (oldprio != p->prio)
2299 p->sched_class->prio_changed(rq, p, oldprio);
2302 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2304 const struct sched_class *class;
2306 if (p->sched_class == rq->curr->sched_class) {
2307 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2309 for_each_class(class) {
2310 if (class == rq->curr->sched_class)
2312 if (class == p->sched_class) {
2313 resched_task(rq->curr);
2320 * A queue event has occurred, and we're going to schedule. In
2321 * this case, we can save a useless back to back clock update.
2323 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2324 rq->skip_clock_update = 1;
2329 * Is this task likely cache-hot:
2332 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2336 if (p->sched_class != &fair_sched_class)
2339 if (unlikely(p->policy == SCHED_IDLE))
2343 * Buddy candidates are cache hot:
2345 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2346 (&p->se == cfs_rq_of(&p->se)->next ||
2347 &p->se == cfs_rq_of(&p->se)->last))
2350 if (sysctl_sched_migration_cost == -1)
2352 if (sysctl_sched_migration_cost == 0)
2355 delta = now - p->se.exec_start;
2357 return delta < (s64)sysctl_sched_migration_cost;
2360 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2362 #ifdef CONFIG_SCHED_DEBUG
2364 * We should never call set_task_cpu() on a blocked task,
2365 * ttwu() will sort out the placement.
2367 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2368 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2370 #ifdef CONFIG_LOCKDEP
2372 * The caller should hold either p->pi_lock or rq->lock, when changing
2373 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2375 * sched_move_task() holds both and thus holding either pins the cgroup,
2378 * Furthermore, all task_rq users should acquire both locks, see
2381 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2382 lockdep_is_held(&task_rq(p)->lock)));
2386 trace_sched_migrate_task(p, new_cpu);
2388 if (task_cpu(p) != new_cpu) {
2389 p->se.nr_migrations++;
2390 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2393 __set_task_cpu(p, new_cpu);
2396 struct migration_arg {
2397 struct task_struct *task;
2401 static int migration_cpu_stop(void *data);
2404 * wait_task_inactive - wait for a thread to unschedule.
2406 * If @match_state is nonzero, it's the @p->state value just checked and
2407 * not expected to change. If it changes, i.e. @p might have woken up,
2408 * then return zero. When we succeed in waiting for @p to be off its CPU,
2409 * we return a positive number (its total switch count). If a second call
2410 * a short while later returns the same number, the caller can be sure that
2411 * @p has remained unscheduled the whole time.
2413 * The caller must ensure that the task *will* unschedule sometime soon,
2414 * else this function might spin for a *long* time. This function can't
2415 * be called with interrupts off, or it may introduce deadlock with
2416 * smp_call_function() if an IPI is sent by the same process we are
2417 * waiting to become inactive.
2419 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2421 unsigned long flags;
2428 * We do the initial early heuristics without holding
2429 * any task-queue locks at all. We'll only try to get
2430 * the runqueue lock when things look like they will
2436 * If the task is actively running on another CPU
2437 * still, just relax and busy-wait without holding
2440 * NOTE! Since we don't hold any locks, it's not
2441 * even sure that "rq" stays as the right runqueue!
2442 * But we don't care, since "task_running()" will
2443 * return false if the runqueue has changed and p
2444 * is actually now running somewhere else!
2446 while (task_running(rq, p)) {
2447 if (match_state && unlikely(p->state != match_state))
2453 * Ok, time to look more closely! We need the rq
2454 * lock now, to be *sure*. If we're wrong, we'll
2455 * just go back and repeat.
2457 rq = task_rq_lock(p, &flags);
2458 trace_sched_wait_task(p);
2459 running = task_running(rq, p);
2462 if (!match_state || p->state == match_state)
2463 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2464 task_rq_unlock(rq, p, &flags);
2467 * If it changed from the expected state, bail out now.
2469 if (unlikely(!ncsw))
2473 * Was it really running after all now that we
2474 * checked with the proper locks actually held?
2476 * Oops. Go back and try again..
2478 if (unlikely(running)) {
2484 * It's not enough that it's not actively running,
2485 * it must be off the runqueue _entirely_, and not
2488 * So if it was still runnable (but just not actively
2489 * running right now), it's preempted, and we should
2490 * yield - it could be a while.
2492 if (unlikely(on_rq)) {
2493 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2495 set_current_state(TASK_UNINTERRUPTIBLE);
2496 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2501 * Ahh, all good. It wasn't running, and it wasn't
2502 * runnable, which means that it will never become
2503 * running in the future either. We're all done!
2512 * kick_process - kick a running thread to enter/exit the kernel
2513 * @p: the to-be-kicked thread
2515 * Cause a process which is running on another CPU to enter
2516 * kernel-mode, without any delay. (to get signals handled.)
2518 * NOTE: this function doesn't have to take the runqueue lock,
2519 * because all it wants to ensure is that the remote task enters
2520 * the kernel. If the IPI races and the task has been migrated
2521 * to another CPU then no harm is done and the purpose has been
2524 void kick_process(struct task_struct *p)
2530 if ((cpu != smp_processor_id()) && task_curr(p))
2531 smp_send_reschedule(cpu);
2534 EXPORT_SYMBOL_GPL(kick_process);
2535 #endif /* CONFIG_SMP */
2539 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2541 static int select_fallback_rq(int cpu, struct task_struct *p)
2544 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2546 /* Look for allowed, online CPU in same node. */
2547 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2548 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
2551 /* Any allowed, online CPU? */
2552 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
2553 if (dest_cpu < nr_cpu_ids)
2556 /* No more Mr. Nice Guy. */
2557 dest_cpu = cpuset_cpus_allowed_fallback(p);
2559 * Don't tell them about moving exiting tasks or
2560 * kernel threads (both mm NULL), since they never
2563 if (p->mm && printk_ratelimit()) {
2564 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2565 task_pid_nr(p), p->comm, cpu);
2572 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2575 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2577 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2580 * In order not to call set_task_cpu() on a blocking task we need
2581 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2584 * Since this is common to all placement strategies, this lives here.
2586 * [ this allows ->select_task() to simply return task_cpu(p) and
2587 * not worry about this generic constraint ]
2589 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
2591 cpu = select_fallback_rq(task_cpu(p), p);
2596 static void update_avg(u64 *avg, u64 sample)
2598 s64 diff = sample - *avg;
2604 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2606 #ifdef CONFIG_SCHEDSTATS
2607 struct rq *rq = this_rq();
2610 int this_cpu = smp_processor_id();
2612 if (cpu == this_cpu) {
2613 schedstat_inc(rq, ttwu_local);
2614 schedstat_inc(p, se.statistics.nr_wakeups_local);
2616 struct sched_domain *sd;
2618 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2620 for_each_domain(this_cpu, sd) {
2621 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2622 schedstat_inc(sd, ttwu_wake_remote);
2629 if (wake_flags & WF_MIGRATED)
2630 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2632 #endif /* CONFIG_SMP */
2634 schedstat_inc(rq, ttwu_count);
2635 schedstat_inc(p, se.statistics.nr_wakeups);
2637 if (wake_flags & WF_SYNC)
2638 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2640 #endif /* CONFIG_SCHEDSTATS */
2643 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2645 activate_task(rq, p, en_flags);
2648 /* if a worker is waking up, notify workqueue */
2649 if (p->flags & PF_WQ_WORKER)
2650 wq_worker_waking_up(p, cpu_of(rq));
2654 * Mark the task runnable and perform wakeup-preemption.
2657 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2659 trace_sched_wakeup(p, true);
2660 check_preempt_curr(rq, p, wake_flags);
2662 p->state = TASK_RUNNING;
2664 if (p->sched_class->task_woken)
2665 p->sched_class->task_woken(rq, p);
2667 if (rq->idle_stamp) {
2668 u64 delta = rq->clock - rq->idle_stamp;
2669 u64 max = 2*sysctl_sched_migration_cost;
2674 update_avg(&rq->avg_idle, delta);
2681 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2684 if (p->sched_contributes_to_load)
2685 rq->nr_uninterruptible--;
2688 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2689 ttwu_do_wakeup(rq, p, wake_flags);
2693 * Called in case the task @p isn't fully descheduled from its runqueue,
2694 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2695 * since all we need to do is flip p->state to TASK_RUNNING, since
2696 * the task is still ->on_rq.
2698 static int ttwu_remote(struct task_struct *p, int wake_flags)
2703 rq = __task_rq_lock(p);
2705 ttwu_do_wakeup(rq, p, wake_flags);
2708 __task_rq_unlock(rq);
2714 static void sched_ttwu_pending(void)
2716 struct rq *rq = this_rq();
2717 struct llist_node *llist = llist_del_all(&rq->wake_list);
2718 struct task_struct *p;
2720 raw_spin_lock(&rq->lock);
2723 p = llist_entry(llist, struct task_struct, wake_entry);
2724 llist = llist_next(llist);
2725 ttwu_do_activate(rq, p, 0);
2728 raw_spin_unlock(&rq->lock);
2731 void scheduler_ipi(void)
2733 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2737 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2738 * traditionally all their work was done from the interrupt return
2739 * path. Now that we actually do some work, we need to make sure
2742 * Some archs already do call them, luckily irq_enter/exit nest
2745 * Arguably we should visit all archs and update all handlers,
2746 * however a fair share of IPIs are still resched only so this would
2747 * somewhat pessimize the simple resched case.
2750 sched_ttwu_pending();
2753 * Check if someone kicked us for doing the nohz idle load balance.
2755 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
2756 this_rq()->idle_balance = 1;
2757 raise_softirq_irqoff(SCHED_SOFTIRQ);
2762 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2764 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
2765 smp_send_reschedule(cpu);
2768 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2769 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2774 rq = __task_rq_lock(p);
2776 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2777 ttwu_do_wakeup(rq, p, wake_flags);
2780 __task_rq_unlock(rq);
2785 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2786 #endif /* CONFIG_SMP */
2788 static void ttwu_queue(struct task_struct *p, int cpu)
2790 struct rq *rq = cpu_rq(cpu);
2792 #if defined(CONFIG_SMP)
2793 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2794 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2795 ttwu_queue_remote(p, cpu);
2800 raw_spin_lock(&rq->lock);
2801 ttwu_do_activate(rq, p, 0);
2802 raw_spin_unlock(&rq->lock);
2806 * try_to_wake_up - wake up a thread
2807 * @p: the thread to be awakened
2808 * @state: the mask of task states that can be woken
2809 * @wake_flags: wake modifier flags (WF_*)
2811 * Put it on the run-queue if it's not already there. The "current"
2812 * thread is always on the run-queue (except when the actual
2813 * re-schedule is in progress), and as such you're allowed to do
2814 * the simpler "current->state = TASK_RUNNING" to mark yourself
2815 * runnable without the overhead of this.
2817 * Returns %true if @p was woken up, %false if it was already running
2818 * or @state didn't match @p's state.
2821 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2823 unsigned long flags;
2824 int cpu, success = 0;
2827 raw_spin_lock_irqsave(&p->pi_lock, flags);
2828 if (!(p->state & state))
2831 success = 1; /* we're going to change ->state */
2834 if (p->on_rq && ttwu_remote(p, wake_flags))
2839 * If the owning (remote) cpu is still in the middle of schedule() with
2840 * this task as prev, wait until its done referencing the task.
2843 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2845 * In case the architecture enables interrupts in
2846 * context_switch(), we cannot busy wait, since that
2847 * would lead to deadlocks when an interrupt hits and
2848 * tries to wake up @prev. So bail and do a complete
2851 if (ttwu_activate_remote(p, wake_flags))
2858 * Pairs with the smp_wmb() in finish_lock_switch().
2862 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2863 p->state = TASK_WAKING;
2865 if (p->sched_class->task_waking)
2866 p->sched_class->task_waking(p);
2868 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2869 if (task_cpu(p) != cpu) {
2870 wake_flags |= WF_MIGRATED;
2871 set_task_cpu(p, cpu);
2873 #endif /* CONFIG_SMP */
2877 ttwu_stat(p, cpu, wake_flags);
2879 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2885 * try_to_wake_up_local - try to wake up a local task with rq lock held
2886 * @p: the thread to be awakened
2888 * Put @p on the run-queue if it's not already there. The caller must
2889 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2892 static void try_to_wake_up_local(struct task_struct *p)
2894 struct rq *rq = task_rq(p);
2896 if (WARN_ON_ONCE(rq != this_rq()) ||
2897 WARN_ON_ONCE(p == current))
2900 lockdep_assert_held(&rq->lock);
2902 if (!raw_spin_trylock(&p->pi_lock)) {
2903 raw_spin_unlock(&rq->lock);
2904 raw_spin_lock(&p->pi_lock);
2905 raw_spin_lock(&rq->lock);
2908 if (!(p->state & TASK_NORMAL))
2912 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2914 ttwu_do_wakeup(rq, p, 0);
2915 ttwu_stat(p, smp_processor_id(), 0);
2917 raw_spin_unlock(&p->pi_lock);
2921 * wake_up_process - Wake up a specific process
2922 * @p: The process to be woken up.
2924 * Attempt to wake up the nominated process and move it to the set of runnable
2925 * processes. Returns 1 if the process was woken up, 0 if it was already
2928 * It may be assumed that this function implies a write memory barrier before
2929 * changing the task state if and only if any tasks are woken up.
2931 int wake_up_process(struct task_struct *p)
2933 WARN_ON(task_is_stopped_or_traced(p));
2934 return try_to_wake_up(p, TASK_NORMAL, 0);
2936 EXPORT_SYMBOL(wake_up_process);
2938 int wake_up_state(struct task_struct *p, unsigned int state)
2940 return try_to_wake_up(p, state, 0);
2944 * Perform scheduler related setup for a newly forked process p.
2945 * p is forked by current.
2947 * __sched_fork() is basic setup used by init_idle() too:
2949 static void __sched_fork(struct task_struct *p)
2954 p->se.exec_start = 0;
2955 p->se.sum_exec_runtime = 0;
2956 p->se.prev_sum_exec_runtime = 0;
2957 p->se.nr_migrations = 0;
2959 INIT_LIST_HEAD(&p->se.group_node);
2961 #ifdef CONFIG_SCHEDSTATS
2962 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2965 INIT_LIST_HEAD(&p->rt.run_list);
2967 #ifdef CONFIG_PREEMPT_NOTIFIERS
2968 INIT_HLIST_HEAD(&p->preempt_notifiers);
2973 * fork()/clone()-time setup:
2975 void sched_fork(struct task_struct *p)
2977 unsigned long flags;
2978 int cpu = get_cpu();
2982 * We mark the process as running here. This guarantees that
2983 * nobody will actually run it, and a signal or other external
2984 * event cannot wake it up and insert it on the runqueue either.
2986 p->state = TASK_RUNNING;
2989 * Make sure we do not leak PI boosting priority to the child.
2991 p->prio = current->normal_prio;
2994 * Revert to default priority/policy on fork if requested.
2996 if (unlikely(p->sched_reset_on_fork)) {
2997 if (task_has_rt_policy(p)) {
2998 p->policy = SCHED_NORMAL;
2999 p->static_prio = NICE_TO_PRIO(0);
3001 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3002 p->static_prio = NICE_TO_PRIO(0);
3004 p->prio = p->normal_prio = __normal_prio(p);
3008 * We don't need the reset flag anymore after the fork. It has
3009 * fulfilled its duty:
3011 p->sched_reset_on_fork = 0;
3014 if (!rt_prio(p->prio))
3015 p->sched_class = &fair_sched_class;
3017 if (p->sched_class->task_fork)
3018 p->sched_class->task_fork(p);
3021 * The child is not yet in the pid-hash so no cgroup attach races,
3022 * and the cgroup is pinned to this child due to cgroup_fork()
3023 * is ran before sched_fork().
3025 * Silence PROVE_RCU.
3027 raw_spin_lock_irqsave(&p->pi_lock, flags);
3028 set_task_cpu(p, cpu);
3029 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3031 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3032 if (likely(sched_info_on()))
3033 memset(&p->sched_info, 0, sizeof(p->sched_info));
3035 #if defined(CONFIG_SMP)
3038 #ifdef CONFIG_PREEMPT_COUNT
3039 /* Want to start with kernel preemption disabled. */
3040 task_thread_info(p)->preempt_count = 1;
3043 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3050 * wake_up_new_task - wake up a newly created task for the first time.
3052 * This function will do some initial scheduler statistics housekeeping
3053 * that must be done for every newly created context, then puts the task
3054 * on the runqueue and wakes it.
3056 void wake_up_new_task(struct task_struct *p)
3058 unsigned long flags;
3061 raw_spin_lock_irqsave(&p->pi_lock, flags);
3064 * Fork balancing, do it here and not earlier because:
3065 * - cpus_allowed can change in the fork path
3066 * - any previously selected cpu might disappear through hotplug
3068 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3071 rq = __task_rq_lock(p);
3072 activate_task(rq, p, 0);
3074 trace_sched_wakeup_new(p, true);
3075 check_preempt_curr(rq, p, WF_FORK);
3077 if (p->sched_class->task_woken)
3078 p->sched_class->task_woken(rq, p);
3080 task_rq_unlock(rq, p, &flags);
3083 #ifdef CONFIG_PREEMPT_NOTIFIERS
3086 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3087 * @notifier: notifier struct to register
3089 void preempt_notifier_register(struct preempt_notifier *notifier)
3091 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3093 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3096 * preempt_notifier_unregister - no longer interested in preemption notifications
3097 * @notifier: notifier struct to unregister
3099 * This is safe to call from within a preemption notifier.
3101 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3103 hlist_del(¬ifier->link);
3105 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3107 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3109 struct preempt_notifier *notifier;
3110 struct hlist_node *node;
3112 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3113 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3117 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3118 struct task_struct *next)
3120 struct preempt_notifier *notifier;
3121 struct hlist_node *node;
3123 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3124 notifier->ops->sched_out(notifier, next);
3127 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3129 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3134 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3135 struct task_struct *next)
3139 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3142 * prepare_task_switch - prepare to switch tasks
3143 * @rq: the runqueue preparing to switch
3144 * @prev: the current task that is being switched out
3145 * @next: the task we are going to switch to.
3147 * This is called with the rq lock held and interrupts off. It must
3148 * be paired with a subsequent finish_task_switch after the context
3151 * prepare_task_switch sets up locking and calls architecture specific
3155 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3156 struct task_struct *next)
3158 sched_info_switch(prev, next);
3159 perf_event_task_sched_out(prev, next);
3160 fire_sched_out_preempt_notifiers(prev, next);
3161 prepare_lock_switch(rq, next);
3162 prepare_arch_switch(next);
3163 trace_sched_switch(prev, next);
3167 * finish_task_switch - clean up after a task-switch
3168 * @rq: runqueue associated with task-switch
3169 * @prev: the thread we just switched away from.
3171 * finish_task_switch must be called after the context switch, paired
3172 * with a prepare_task_switch call before the context switch.
3173 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3174 * and do any other architecture-specific cleanup actions.
3176 * Note that we may have delayed dropping an mm in context_switch(). If
3177 * so, we finish that here outside of the runqueue lock. (Doing it
3178 * with the lock held can cause deadlocks; see schedule() for
3181 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3182 __releases(rq->lock)
3184 struct mm_struct *mm = rq->prev_mm;
3190 * A task struct has one reference for the use as "current".
3191 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3192 * schedule one last time. The schedule call will never return, and
3193 * the scheduled task must drop that reference.
3194 * The test for TASK_DEAD must occur while the runqueue locks are
3195 * still held, otherwise prev could be scheduled on another cpu, die
3196 * there before we look at prev->state, and then the reference would
3198 * Manfred Spraul <manfred@colorfullife.com>
3200 prev_state = prev->state;
3201 finish_arch_switch(prev);
3202 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3203 local_irq_disable();
3204 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3205 perf_event_task_sched_in(prev, current);
3206 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3208 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3209 finish_lock_switch(rq, prev);
3211 fire_sched_in_preempt_notifiers(current);
3214 if (unlikely(prev_state == TASK_DEAD)) {
3216 * Remove function-return probe instances associated with this
3217 * task and put them back on the free list.
3219 kprobe_flush_task(prev);
3220 put_task_struct(prev);
3226 /* assumes rq->lock is held */
3227 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3229 if (prev->sched_class->pre_schedule)
3230 prev->sched_class->pre_schedule(rq, prev);
3233 /* rq->lock is NOT held, but preemption is disabled */
3234 static inline void post_schedule(struct rq *rq)
3236 if (rq->post_schedule) {
3237 unsigned long flags;
3239 raw_spin_lock_irqsave(&rq->lock, flags);
3240 if (rq->curr->sched_class->post_schedule)
3241 rq->curr->sched_class->post_schedule(rq);
3242 raw_spin_unlock_irqrestore(&rq->lock, flags);
3244 rq->post_schedule = 0;
3250 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3254 static inline void post_schedule(struct rq *rq)
3261 * schedule_tail - first thing a freshly forked thread must call.
3262 * @prev: the thread we just switched away from.
3264 asmlinkage void schedule_tail(struct task_struct *prev)
3265 __releases(rq->lock)
3267 struct rq *rq = this_rq();
3269 finish_task_switch(rq, prev);
3272 * FIXME: do we need to worry about rq being invalidated by the
3277 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3278 /* In this case, finish_task_switch does not reenable preemption */
3281 if (current->set_child_tid)
3282 put_user(task_pid_vnr(current), current->set_child_tid);
3286 * context_switch - switch to the new MM and the new
3287 * thread's register state.
3290 context_switch(struct rq *rq, struct task_struct *prev,
3291 struct task_struct *next)
3293 struct mm_struct *mm, *oldmm;
3295 prepare_task_switch(rq, prev, next);
3298 oldmm = prev->active_mm;
3300 * For paravirt, this is coupled with an exit in switch_to to
3301 * combine the page table reload and the switch backend into
3304 arch_start_context_switch(prev);
3307 next->active_mm = oldmm;
3308 atomic_inc(&oldmm->mm_count);
3309 enter_lazy_tlb(oldmm, next);
3311 switch_mm(oldmm, mm, next);
3314 prev->active_mm = NULL;
3315 rq->prev_mm = oldmm;
3318 * Since the runqueue lock will be released by the next
3319 * task (which is an invalid locking op but in the case
3320 * of the scheduler it's an obvious special-case), so we
3321 * do an early lockdep release here:
3323 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3324 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3327 /* Here we just switch the register state and the stack. */
3328 switch_to(prev, next, prev);
3332 * this_rq must be evaluated again because prev may have moved
3333 * CPUs since it called schedule(), thus the 'rq' on its stack
3334 * frame will be invalid.
3336 finish_task_switch(this_rq(), prev);
3340 * nr_running, nr_uninterruptible and nr_context_switches:
3342 * externally visible scheduler statistics: current number of runnable
3343 * threads, current number of uninterruptible-sleeping threads, total
3344 * number of context switches performed since bootup.
3346 unsigned long nr_running(void)
3348 unsigned long i, sum = 0;
3350 for_each_online_cpu(i)
3351 sum += cpu_rq(i)->nr_running;
3356 unsigned long nr_uninterruptible(void)
3358 unsigned long i, sum = 0;
3360 for_each_possible_cpu(i)
3361 sum += cpu_rq(i)->nr_uninterruptible;
3364 * Since we read the counters lockless, it might be slightly
3365 * inaccurate. Do not allow it to go below zero though:
3367 if (unlikely((long)sum < 0))
3373 unsigned long long nr_context_switches(void)
3376 unsigned long long sum = 0;
3378 for_each_possible_cpu(i)
3379 sum += cpu_rq(i)->nr_switches;
3384 unsigned long nr_iowait(void)
3386 unsigned long i, sum = 0;
3388 for_each_possible_cpu(i)
3389 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3394 unsigned long nr_iowait_cpu(int cpu)
3396 struct rq *this = cpu_rq(cpu);
3397 return atomic_read(&this->nr_iowait);
3400 unsigned long this_cpu_load(void)
3402 struct rq *this = this_rq();
3403 return this->cpu_load[0];
3408 * Global load-average calculations
3410 * We take a distributed and async approach to calculating the global load-avg
3411 * in order to minimize overhead.
3413 * The global load average is an exponentially decaying average of nr_running +
3414 * nr_uninterruptible.
3416 * Once every LOAD_FREQ:
3419 * for_each_possible_cpu(cpu)
3420 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
3422 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
3424 * Due to a number of reasons the above turns in the mess below:
3426 * - for_each_possible_cpu() is prohibitively expensive on machines with
3427 * serious number of cpus, therefore we need to take a distributed approach
3428 * to calculating nr_active.
3430 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
3431 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
3433 * So assuming nr_active := 0 when we start out -- true per definition, we
3434 * can simply take per-cpu deltas and fold those into a global accumulate
3435 * to obtain the same result. See calc_load_fold_active().
3437 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
3438 * across the machine, we assume 10 ticks is sufficient time for every
3439 * cpu to have completed this task.
3441 * This places an upper-bound on the IRQ-off latency of the machine. Then
3442 * again, being late doesn't loose the delta, just wrecks the sample.
3444 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
3445 * this would add another cross-cpu cacheline miss and atomic operation
3446 * to the wakeup path. Instead we increment on whatever cpu the task ran
3447 * when it went into uninterruptible state and decrement on whatever cpu
3448 * did the wakeup. This means that only the sum of nr_uninterruptible over
3449 * all cpus yields the correct result.
3451 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
3454 /* Variables and functions for calc_load */
3455 static atomic_long_t calc_load_tasks;
3456 static unsigned long calc_load_update;
3457 unsigned long avenrun[3];
3458 EXPORT_SYMBOL(avenrun); /* should be removed */
3461 * get_avenrun - get the load average array
3462 * @loads: pointer to dest load array
3463 * @offset: offset to add
3464 * @shift: shift count to shift the result left
3466 * These values are estimates at best, so no need for locking.
3468 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3470 loads[0] = (avenrun[0] + offset) << shift;
3471 loads[1] = (avenrun[1] + offset) << shift;
3472 loads[2] = (avenrun[2] + offset) << shift;
3475 static long calc_load_fold_active(struct rq *this_rq)
3477 long nr_active, delta = 0;
3479 nr_active = this_rq->nr_running;
3480 nr_active += (long) this_rq->nr_uninterruptible;
3482 if (nr_active != this_rq->calc_load_active) {
3483 delta = nr_active - this_rq->calc_load_active;
3484 this_rq->calc_load_active = nr_active;
3491 * a1 = a0 * e + a * (1 - e)
3493 static unsigned long
3494 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3497 load += active * (FIXED_1 - exp);
3498 load += 1UL << (FSHIFT - 1);
3499 return load >> FSHIFT;
3504 * Handle NO_HZ for the global load-average.
3506 * Since the above described distributed algorithm to compute the global
3507 * load-average relies on per-cpu sampling from the tick, it is affected by
3510 * The basic idea is to fold the nr_active delta into a global idle-delta upon
3511 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
3512 * when we read the global state.
3514 * Obviously reality has to ruin such a delightfully simple scheme:
3516 * - When we go NO_HZ idle during the window, we can negate our sample
3517 * contribution, causing under-accounting.
3519 * We avoid this by keeping two idle-delta counters and flipping them
3520 * when the window starts, thus separating old and new NO_HZ load.
3522 * The only trick is the slight shift in index flip for read vs write.
3526 * |-|-----------|-|-----------|-|-----------|-|
3527 * r:0 0 1 1 0 0 1 1 0
3528 * w:0 1 1 0 0 1 1 0 0
3530 * This ensures we'll fold the old idle contribution in this window while
3531 * accumlating the new one.
3533 * - When we wake up from NO_HZ idle during the window, we push up our
3534 * contribution, since we effectively move our sample point to a known
3537 * This is solved by pushing the window forward, and thus skipping the
3538 * sample, for this cpu (effectively using the idle-delta for this cpu which
3539 * was in effect at the time the window opened). This also solves the issue
3540 * of having to deal with a cpu having been in NOHZ idle for multiple
3541 * LOAD_FREQ intervals.
3543 * When making the ILB scale, we should try to pull this in as well.
3545 static atomic_long_t calc_load_idle[2];
3546 static int calc_load_idx;
3548 static inline int calc_load_write_idx(void)
3550 int idx = calc_load_idx;
3553 * See calc_global_nohz(), if we observe the new index, we also
3554 * need to observe the new update time.
3559 * If the folding window started, make sure we start writing in the
3562 if (!time_before(jiffies, calc_load_update))
3568 static inline int calc_load_read_idx(void)
3570 return calc_load_idx & 1;
3573 void calc_load_enter_idle(void)
3575 struct rq *this_rq = this_rq();
3579 * We're going into NOHZ mode, if there's any pending delta, fold it
3580 * into the pending idle delta.
3582 delta = calc_load_fold_active(this_rq);
3584 int idx = calc_load_write_idx();
3585 atomic_long_add(delta, &calc_load_idle[idx]);
3589 void calc_load_exit_idle(void)
3591 struct rq *this_rq = this_rq();
3594 * If we're still before the sample window, we're done.
3596 if (time_before(jiffies, this_rq->calc_load_update))
3600 * We woke inside or after the sample window, this means we're already
3601 * accounted through the nohz accounting, so skip the entire deal and
3602 * sync up for the next window.
3604 this_rq->calc_load_update = calc_load_update;
3605 if (time_before(jiffies, this_rq->calc_load_update + 10))
3606 this_rq->calc_load_update += LOAD_FREQ;
3609 static long calc_load_fold_idle(void)
3611 int idx = calc_load_read_idx();
3614 if (atomic_long_read(&calc_load_idle[idx]))
3615 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
3621 * fixed_power_int - compute: x^n, in O(log n) time
3623 * @x: base of the power
3624 * @frac_bits: fractional bits of @x
3625 * @n: power to raise @x to.
3627 * By exploiting the relation between the definition of the natural power
3628 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3629 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3630 * (where: n_i \elem {0, 1}, the binary vector representing n),
3631 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3632 * of course trivially computable in O(log_2 n), the length of our binary
3635 static unsigned long
3636 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3638 unsigned long result = 1UL << frac_bits;
3643 result += 1UL << (frac_bits - 1);
3644 result >>= frac_bits;
3650 x += 1UL << (frac_bits - 1);
3658 * a1 = a0 * e + a * (1 - e)
3660 * a2 = a1 * e + a * (1 - e)
3661 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3662 * = a0 * e^2 + a * (1 - e) * (1 + e)
3664 * a3 = a2 * e + a * (1 - e)
3665 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3666 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3670 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3671 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3672 * = a0 * e^n + a * (1 - e^n)
3674 * [1] application of the geometric series:
3677 * S_n := \Sum x^i = -------------
3680 static unsigned long
3681 calc_load_n(unsigned long load, unsigned long exp,
3682 unsigned long active, unsigned int n)
3685 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3689 * NO_HZ can leave us missing all per-cpu ticks calling
3690 * calc_load_account_active(), but since an idle CPU folds its delta into
3691 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3692 * in the pending idle delta if our idle period crossed a load cycle boundary.
3694 * Once we've updated the global active value, we need to apply the exponential
3695 * weights adjusted to the number of cycles missed.
3697 static void calc_global_nohz(void)
3699 long delta, active, n;
3701 if (!time_before(jiffies, calc_load_update + 10)) {
3703 * Catch-up, fold however many we are behind still
3705 delta = jiffies - calc_load_update - 10;
3706 n = 1 + (delta / LOAD_FREQ);
3708 active = atomic_long_read(&calc_load_tasks);
3709 active = active > 0 ? active * FIXED_1 : 0;
3711 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3712 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3713 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3715 calc_load_update += n * LOAD_FREQ;
3719 * Flip the idle index...
3721 * Make sure we first write the new time then flip the index, so that
3722 * calc_load_write_idx() will see the new time when it reads the new
3723 * index, this avoids a double flip messing things up.
3728 #else /* !CONFIG_NO_HZ */
3730 static inline long calc_load_fold_idle(void) { return 0; }
3731 static inline void calc_global_nohz(void) { }
3733 #endif /* CONFIG_NO_HZ */
3736 * calc_load - update the avenrun load estimates 10 ticks after the
3737 * CPUs have updated calc_load_tasks.
3739 void calc_global_load(unsigned long ticks)
3743 if (time_before(jiffies, calc_load_update + 10))
3747 * Fold the 'old' idle-delta to include all NO_HZ cpus.
3749 delta = calc_load_fold_idle();
3751 atomic_long_add(delta, &calc_load_tasks);
3753 active = atomic_long_read(&calc_load_tasks);
3754 active = active > 0 ? active * FIXED_1 : 0;
3756 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3757 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3758 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3760 calc_load_update += LOAD_FREQ;
3763 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
3769 * Called from update_cpu_load() to periodically update this CPU's
3772 static void calc_load_account_active(struct rq *this_rq)
3776 if (time_before(jiffies, this_rq->calc_load_update))
3779 delta = calc_load_fold_active(this_rq);
3781 atomic_long_add(delta, &calc_load_tasks);
3783 this_rq->calc_load_update += LOAD_FREQ;
3787 * End of global load-average stuff
3791 * The exact cpuload at various idx values, calculated at every tick would be
3792 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3794 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3795 * on nth tick when cpu may be busy, then we have:
3796 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3797 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3799 * decay_load_missed() below does efficient calculation of
3800 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3801 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3803 * The calculation is approximated on a 128 point scale.
3804 * degrade_zero_ticks is the number of ticks after which load at any
3805 * particular idx is approximated to be zero.
3806 * degrade_factor is a precomputed table, a row for each load idx.
3807 * Each column corresponds to degradation factor for a power of two ticks,
3808 * based on 128 point scale.
3810 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3811 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3813 * With this power of 2 load factors, we can degrade the load n times
3814 * by looking at 1 bits in n and doing as many mult/shift instead of
3815 * n mult/shifts needed by the exact degradation.
3817 #define DEGRADE_SHIFT 7
3818 static const unsigned char
3819 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3820 static const unsigned char
3821 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3822 {0, 0, 0, 0, 0, 0, 0, 0},
3823 {64, 32, 8, 0, 0, 0, 0, 0},
3824 {96, 72, 40, 12, 1, 0, 0},
3825 {112, 98, 75, 43, 15, 1, 0},
3826 {120, 112, 98, 76, 45, 16, 2} };
3829 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3830 * would be when CPU is idle and so we just decay the old load without
3831 * adding any new load.
3833 static unsigned long
3834 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3838 if (!missed_updates)
3841 if (missed_updates >= degrade_zero_ticks[idx])
3845 return load >> missed_updates;
3847 while (missed_updates) {
3848 if (missed_updates % 2)
3849 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3851 missed_updates >>= 1;
3858 * Update rq->cpu_load[] statistics. This function is usually called every
3859 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3860 * every tick. We fix it up based on jiffies.
3862 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
3863 unsigned long pending_updates)
3867 this_rq->nr_load_updates++;
3869 /* Update our load: */
3870 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3871 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3872 unsigned long old_load, new_load;
3874 /* scale is effectively 1 << i now, and >> i divides by scale */
3876 old_load = this_rq->cpu_load[i];
3877 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3878 new_load = this_load;
3880 * Round up the averaging division if load is increasing. This
3881 * prevents us from getting stuck on 9 if the load is 10, for
3884 if (new_load > old_load)
3885 new_load += scale - 1;
3887 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3890 sched_avg_update(this_rq);
3895 * There is no sane way to deal with nohz on smp when using jiffies because the
3896 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
3897 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
3899 * Therefore we cannot use the delta approach from the regular tick since that
3900 * would seriously skew the load calculation. However we'll make do for those
3901 * updates happening while idle (nohz_idle_balance) or coming out of idle
3902 * (tick_nohz_idle_exit).
3904 * This means we might still be one tick off for nohz periods.
3908 * Called from nohz_idle_balance() to update the load ratings before doing the
3911 static void update_idle_cpu_load(struct rq *this_rq)
3913 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3914 unsigned long load = this_rq->load.weight;
3915 unsigned long pending_updates;
3918 * bail if there's load or we're actually up-to-date.
3920 if (load || curr_jiffies == this_rq->last_load_update_tick)
3923 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3924 this_rq->last_load_update_tick = curr_jiffies;
3926 __update_cpu_load(this_rq, load, pending_updates);
3930 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
3932 void update_cpu_load_nohz(void)
3934 struct rq *this_rq = this_rq();
3935 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3936 unsigned long pending_updates;
3938 if (curr_jiffies == this_rq->last_load_update_tick)
3941 raw_spin_lock(&this_rq->lock);
3942 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3943 if (pending_updates) {
3944 this_rq->last_load_update_tick = curr_jiffies;
3946 * We were idle, this means load 0, the current load might be
3947 * !0 due to remote wakeups and the sort.
3949 __update_cpu_load(this_rq, 0, pending_updates);
3951 raw_spin_unlock(&this_rq->lock);
3953 #endif /* CONFIG_NO_HZ */
3956 * Called from scheduler_tick()
3958 static void update_cpu_load_active(struct rq *this_rq)
3961 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
3963 this_rq->last_load_update_tick = jiffies;
3964 __update_cpu_load(this_rq, this_rq->load.weight, 1);
3966 calc_load_account_active(this_rq);
3972 * sched_exec - execve() is a valuable balancing opportunity, because at
3973 * this point the task has the smallest effective memory and cache footprint.
3975 void sched_exec(void)
3977 struct task_struct *p = current;
3978 unsigned long flags;
3981 raw_spin_lock_irqsave(&p->pi_lock, flags);
3982 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3983 if (dest_cpu == smp_processor_id())
3986 if (likely(cpu_active(dest_cpu))) {
3987 struct migration_arg arg = { p, dest_cpu };
3989 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3990 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3994 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3999 DEFINE_PER_CPU(struct kernel_stat, kstat);
4001 EXPORT_PER_CPU_SYMBOL(kstat);
4004 * Return any ns on the sched_clock that have not yet been accounted in
4005 * @p in case that task is currently running.
4007 * Called with task_rq_lock() held on @rq.
4009 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4013 if (task_current(rq, p)) {
4014 update_rq_clock(rq);
4015 ns = rq->clock_task - p->se.exec_start;
4023 unsigned long long task_delta_exec(struct task_struct *p)
4025 unsigned long flags;
4029 rq = task_rq_lock(p, &flags);
4030 ns = do_task_delta_exec(p, rq);
4031 task_rq_unlock(rq, p, &flags);
4037 * Return accounted runtime for the task.
4038 * In case the task is currently running, return the runtime plus current's
4039 * pending runtime that have not been accounted yet.
4041 unsigned long long task_sched_runtime(struct task_struct *p)
4043 unsigned long flags;
4047 rq = task_rq_lock(p, &flags);
4048 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4049 task_rq_unlock(rq, p, &flags);
4055 * Account user cpu time to a process.
4056 * @p: the process that the cpu time gets accounted to
4057 * @cputime: the cpu time spent in user space since the last update
4058 * @cputime_scaled: cputime scaled by cpu frequency
4060 void account_user_time(struct task_struct *p, cputime_t cputime,
4061 cputime_t cputime_scaled)
4063 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4066 /* Add user time to process. */
4067 p->utime = cputime_add(p->utime, cputime);
4068 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4069 account_group_user_time(p, cputime);
4071 /* Add user time to cpustat. */
4072 tmp = cputime_to_cputime64(cputime);
4073 if (TASK_NICE(p) > 0)
4074 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4076 cpustat->user = cputime64_add(cpustat->user, tmp);
4078 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4079 /* Account for user time used */
4080 acct_update_integrals(p);
4084 * Account guest cpu time to a process.
4085 * @p: the process that the cpu time gets accounted to
4086 * @cputime: the cpu time spent in virtual machine since the last update
4087 * @cputime_scaled: cputime scaled by cpu frequency
4089 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4090 cputime_t cputime_scaled)
4093 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4095 tmp = cputime_to_cputime64(cputime);
4097 /* Add guest time to process. */
4098 p->utime = cputime_add(p->utime, cputime);
4099 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4100 account_group_user_time(p, cputime);
4101 p->gtime = cputime_add(p->gtime, cputime);
4103 /* Add guest time to cpustat. */
4104 if (TASK_NICE(p) > 0) {
4105 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4106 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
4108 cpustat->user = cputime64_add(cpustat->user, tmp);
4109 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4114 * Account system cpu time to a process and desired cpustat field
4115 * @p: the process that the cpu time gets accounted to
4116 * @cputime: the cpu time spent in kernel space since the last update
4117 * @cputime_scaled: cputime scaled by cpu frequency
4118 * @target_cputime64: pointer to cpustat field that has to be updated
4121 void __account_system_time(struct task_struct *p, cputime_t cputime,
4122 cputime_t cputime_scaled, cputime64_t *target_cputime64)
4124 cputime64_t tmp = cputime_to_cputime64(cputime);
4126 /* Add system time to process. */
4127 p->stime = cputime_add(p->stime, cputime);
4128 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4129 account_group_system_time(p, cputime);
4131 /* Add system time to cpustat. */
4132 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
4133 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4135 /* Account for system time used */
4136 acct_update_integrals(p);
4140 * Account system cpu time to a process.
4141 * @p: the process that the cpu time gets accounted to
4142 * @hardirq_offset: the offset to subtract from hardirq_count()
4143 * @cputime: the cpu time spent in kernel space since the last update
4144 * @cputime_scaled: cputime scaled by cpu frequency
4146 void account_system_time(struct task_struct *p, int hardirq_offset,
4147 cputime_t cputime, cputime_t cputime_scaled)
4149 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4150 cputime64_t *target_cputime64;
4152 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4153 account_guest_time(p, cputime, cputime_scaled);
4157 if (hardirq_count() - hardirq_offset)
4158 target_cputime64 = &cpustat->irq;
4159 else if (in_serving_softirq())
4160 target_cputime64 = &cpustat->softirq;
4162 target_cputime64 = &cpustat->system;
4164 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
4168 * Account for involuntary wait time.
4169 * @cputime: the cpu time spent in involuntary wait
4171 void account_steal_time(cputime_t cputime)
4173 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4174 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4176 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4180 * Account for idle time.
4181 * @cputime: the cpu time spent in idle wait
4183 void account_idle_time(cputime_t cputime)
4185 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4186 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4187 struct rq *rq = this_rq();
4189 if (atomic_read(&rq->nr_iowait) > 0)
4190 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4192 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4195 static __always_inline bool steal_account_process_tick(void)
4197 #ifdef CONFIG_PARAVIRT
4198 if (static_branch(¶virt_steal_enabled)) {
4201 steal = paravirt_steal_clock(smp_processor_id());
4202 steal -= this_rq()->prev_steal_time;
4204 st = steal_ticks(steal);
4205 this_rq()->prev_steal_time += st * TICK_NSEC;
4207 account_steal_time(st);
4214 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4216 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4218 * Account a tick to a process and cpustat
4219 * @p: the process that the cpu time gets accounted to
4220 * @user_tick: is the tick from userspace
4221 * @rq: the pointer to rq
4223 * Tick demultiplexing follows the order
4224 * - pending hardirq update
4225 * - pending softirq update
4229 * - check for guest_time
4230 * - else account as system_time
4232 * Check for hardirq is done both for system and user time as there is
4233 * no timer going off while we are on hardirq and hence we may never get an
4234 * opportunity to update it solely in system time.
4235 * p->stime and friends are only updated on system time and not on irq
4236 * softirq as those do not count in task exec_runtime any more.
4238 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4241 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4242 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
4243 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4245 if (steal_account_process_tick())
4248 if (irqtime_account_hi_update()) {
4249 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4250 } else if (irqtime_account_si_update()) {
4251 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4252 } else if (this_cpu_ksoftirqd() == p) {
4254 * ksoftirqd time do not get accounted in cpu_softirq_time.
4255 * So, we have to handle it separately here.
4256 * Also, p->stime needs to be updated for ksoftirqd.
4258 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4260 } else if (user_tick) {
4261 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4262 } else if (p == rq->idle) {
4263 account_idle_time(cputime_one_jiffy);
4264 } else if (p->flags & PF_VCPU) { /* System time or guest time */
4265 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
4267 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4272 static void irqtime_account_idle_ticks(int ticks)
4275 struct rq *rq = this_rq();
4277 for (i = 0; i < ticks; i++)
4278 irqtime_account_process_tick(current, 0, rq);
4280 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4281 static void irqtime_account_idle_ticks(int ticks) {}
4282 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4284 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4287 * Account a single tick of cpu time.
4288 * @p: the process that the cpu time gets accounted to
4289 * @user_tick: indicates if the tick is a user or a system tick
4291 void account_process_tick(struct task_struct *p, int user_tick)
4293 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4294 struct rq *rq = this_rq();
4296 if (sched_clock_irqtime) {
4297 irqtime_account_process_tick(p, user_tick, rq);
4301 if (steal_account_process_tick())
4305 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4306 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4307 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4310 account_idle_time(cputime_one_jiffy);
4314 * Account multiple ticks of steal time.
4315 * @p: the process from which the cpu time has been stolen
4316 * @ticks: number of stolen ticks
4318 void account_steal_ticks(unsigned long ticks)
4320 account_steal_time(jiffies_to_cputime(ticks));
4324 * Account multiple ticks of idle time.
4325 * @ticks: number of stolen ticks
4327 void account_idle_ticks(unsigned long ticks)
4330 if (sched_clock_irqtime) {
4331 irqtime_account_idle_ticks(ticks);
4335 account_idle_time(jiffies_to_cputime(ticks));
4341 * Use precise platform statistics if available:
4343 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4344 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4350 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4352 struct task_cputime cputime;
4354 thread_group_cputime(p, &cputime);
4356 *ut = cputime.utime;
4357 *st = cputime.stime;
4361 #ifndef nsecs_to_cputime
4362 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4365 static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total)
4367 u64 temp = (__force u64) rtime;
4369 temp *= (__force u64) utime;
4371 if (sizeof(cputime_t) == 4)
4372 temp = div_u64(temp, (__force u32) total);
4374 temp = div64_u64(temp, (__force u64) total);
4376 return (__force cputime_t) temp;
4379 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4381 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4384 * Use CFS's precise accounting:
4386 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4389 utime = scale_utime(utime, rtime, total);
4394 * Compare with previous values, to keep monotonicity:
4396 p->prev_utime = max(p->prev_utime, utime);
4397 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4399 *ut = p->prev_utime;
4400 *st = p->prev_stime;
4404 * Must be called with siglock held.
4406 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4408 struct signal_struct *sig = p->signal;
4409 struct task_cputime cputime;
4410 cputime_t rtime, utime, total;
4412 thread_group_cputime(p, &cputime);
4414 total = cputime_add(cputime.utime, cputime.stime);
4415 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4418 utime = scale_utime(cputime.utime, rtime, total);
4422 sig->prev_utime = max(sig->prev_utime, utime);
4423 sig->prev_stime = max(sig->prev_stime,
4424 cputime_sub(rtime, sig->prev_utime));
4426 *ut = sig->prev_utime;
4427 *st = sig->prev_stime;
4432 * This function gets called by the timer code, with HZ frequency.
4433 * We call it with interrupts disabled.
4435 void scheduler_tick(void)
4437 int cpu = smp_processor_id();
4438 struct rq *rq = cpu_rq(cpu);
4439 struct task_struct *curr = rq->curr;
4443 raw_spin_lock(&rq->lock);
4444 update_rq_clock(rq);
4445 update_cpu_load_active(rq);
4446 curr->sched_class->task_tick(rq, curr, 0);
4447 raw_spin_unlock(&rq->lock);
4449 perf_event_task_tick();
4452 rq->idle_balance = idle_cpu(cpu);
4453 trigger_load_balance(rq, cpu);
4457 notrace unsigned long get_parent_ip(unsigned long addr)
4459 if (in_lock_functions(addr)) {
4460 addr = CALLER_ADDR2;
4461 if (in_lock_functions(addr))
4462 addr = CALLER_ADDR3;
4467 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4468 defined(CONFIG_PREEMPT_TRACER))
4470 void __kprobes add_preempt_count(int val)
4472 #ifdef CONFIG_DEBUG_PREEMPT
4476 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4479 preempt_count() += val;
4480 #ifdef CONFIG_DEBUG_PREEMPT
4482 * Spinlock count overflowing soon?
4484 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4487 if (preempt_count() == val)
4488 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4490 EXPORT_SYMBOL(add_preempt_count);
4492 void __kprobes sub_preempt_count(int val)
4494 #ifdef CONFIG_DEBUG_PREEMPT
4498 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4501 * Is the spinlock portion underflowing?
4503 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4504 !(preempt_count() & PREEMPT_MASK)))
4508 if (preempt_count() == val)
4509 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4510 preempt_count() -= val;
4512 EXPORT_SYMBOL(sub_preempt_count);
4517 * Print scheduling while atomic bug:
4519 static noinline void __schedule_bug(struct task_struct *prev)
4521 struct pt_regs *regs = get_irq_regs();
4523 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4524 prev->comm, prev->pid, preempt_count());
4526 debug_show_held_locks(prev);
4528 if (irqs_disabled())
4529 print_irqtrace_events(prev);
4538 * Various schedule()-time debugging checks and statistics:
4540 static inline void schedule_debug(struct task_struct *prev)
4543 * Test if we are atomic. Since do_exit() needs to call into
4544 * schedule() atomically, we ignore that path for now.
4545 * Otherwise, whine if we are scheduling when we should not be.
4547 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4548 __schedule_bug(prev);
4551 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4553 schedstat_inc(this_rq(), sched_count);
4556 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4558 if (prev->on_rq || rq->skip_clock_update < 0)
4559 update_rq_clock(rq);
4560 prev->sched_class->put_prev_task(rq, prev);
4564 * Pick up the highest-prio task:
4566 static inline struct task_struct *
4567 pick_next_task(struct rq *rq)
4569 const struct sched_class *class;
4570 struct task_struct *p;
4573 * Optimization: we know that if all tasks are in
4574 * the fair class we can call that function directly:
4576 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4577 p = fair_sched_class.pick_next_task(rq);
4582 for_each_class(class) {
4583 p = class->pick_next_task(rq);
4588 BUG(); /* the idle class will always have a runnable task */
4592 * __schedule() is the main scheduler function.
4594 static void __sched __schedule(void)
4596 struct task_struct *prev, *next;
4597 unsigned long *switch_count;
4603 cpu = smp_processor_id();
4605 rcu_note_context_switch(cpu);
4608 schedule_debug(prev);
4610 if (sched_feat(HRTICK))
4613 raw_spin_lock_irq(&rq->lock);
4615 switch_count = &prev->nivcsw;
4616 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4617 if (unlikely(signal_pending_state(prev->state, prev))) {
4618 prev->state = TASK_RUNNING;
4620 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4624 * If a worker went to sleep, notify and ask workqueue
4625 * whether it wants to wake up a task to maintain
4628 if (prev->flags & PF_WQ_WORKER) {
4629 struct task_struct *to_wakeup;
4631 to_wakeup = wq_worker_sleeping(prev, cpu);
4633 try_to_wake_up_local(to_wakeup);
4636 switch_count = &prev->nvcsw;
4639 pre_schedule(rq, prev);
4641 if (unlikely(!rq->nr_running))
4642 idle_balance(cpu, rq);
4644 put_prev_task(rq, prev);
4645 next = pick_next_task(rq);
4646 clear_tsk_need_resched(prev);
4647 rq->skip_clock_update = 0;
4649 if (likely(prev != next)) {
4654 context_switch(rq, prev, next); /* unlocks the rq */
4656 * The context switch have flipped the stack from under us
4657 * and restored the local variables which were saved when
4658 * this task called schedule() in the past. prev == current
4659 * is still correct, but it can be moved to another cpu/rq.
4661 cpu = smp_processor_id();
4664 raw_spin_unlock_irq(&rq->lock);
4668 preempt_enable_no_resched();
4673 static inline void sched_submit_work(struct task_struct *tsk)
4678 * If we are going to sleep and we have plugged IO queued,
4679 * make sure to submit it to avoid deadlocks.
4681 if (blk_needs_flush_plug(tsk))
4682 blk_schedule_flush_plug(tsk);
4685 asmlinkage void __sched schedule(void)
4687 struct task_struct *tsk = current;
4689 sched_submit_work(tsk);
4692 EXPORT_SYMBOL(schedule);
4694 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4696 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4698 if (lock->owner != owner)
4702 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4703 * lock->owner still matches owner, if that fails, owner might
4704 * point to free()d memory, if it still matches, the rcu_read_lock()
4705 * ensures the memory stays valid.
4709 return owner->on_cpu;
4713 * Look out! "owner" is an entirely speculative pointer
4714 * access and not reliable.
4716 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4718 if (!sched_feat(OWNER_SPIN))
4722 while (owner_running(lock, owner)) {
4726 arch_mutex_cpu_relax();
4731 * We break out the loop above on need_resched() and when the
4732 * owner changed, which is a sign for heavy contention. Return
4733 * success only when lock->owner is NULL.
4735 return lock->owner == NULL;
4739 #ifdef CONFIG_PREEMPT
4741 * this is the entry point to schedule() from in-kernel preemption
4742 * off of preempt_enable. Kernel preemptions off return from interrupt
4743 * occur there and call schedule directly.
4745 asmlinkage void __sched notrace preempt_schedule(void)
4747 struct thread_info *ti = current_thread_info();
4750 * If there is a non-zero preempt_count or interrupts are disabled,
4751 * we do not want to preempt the current task. Just return..
4753 if (likely(ti->preempt_count || irqs_disabled()))
4757 add_preempt_count_notrace(PREEMPT_ACTIVE);
4759 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4762 * Check again in case we missed a preemption opportunity
4763 * between schedule and now.
4766 } while (need_resched());
4768 EXPORT_SYMBOL(preempt_schedule);
4771 * this is the entry point to schedule() from kernel preemption
4772 * off of irq context.
4773 * Note, that this is called and return with irqs disabled. This will
4774 * protect us against recursive calling from irq.
4776 asmlinkage void __sched preempt_schedule_irq(void)
4778 struct thread_info *ti = current_thread_info();
4780 /* Catch callers which need to be fixed */
4781 BUG_ON(ti->preempt_count || !irqs_disabled());
4784 add_preempt_count(PREEMPT_ACTIVE);
4787 local_irq_disable();
4788 sub_preempt_count(PREEMPT_ACTIVE);
4791 * Check again in case we missed a preemption opportunity
4792 * between schedule and now.
4795 } while (need_resched());
4798 #endif /* CONFIG_PREEMPT */
4800 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4803 return try_to_wake_up(curr->private, mode, wake_flags);
4805 EXPORT_SYMBOL(default_wake_function);
4808 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4809 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4810 * number) then we wake all the non-exclusive tasks and one exclusive task.
4812 * There are circumstances in which we can try to wake a task which has already
4813 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4814 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4816 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4817 int nr_exclusive, int wake_flags, void *key)
4819 wait_queue_t *curr, *next;
4821 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4822 unsigned flags = curr->flags;
4824 if (curr->func(curr, mode, wake_flags, key) &&
4825 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4831 * __wake_up - wake up threads blocked on a waitqueue.
4833 * @mode: which threads
4834 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4835 * @key: is directly passed to the wakeup function
4837 * It may be assumed that this function implies a write memory barrier before
4838 * changing the task state if and only if any tasks are woken up.
4840 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4841 int nr_exclusive, void *key)
4843 unsigned long flags;
4845 spin_lock_irqsave(&q->lock, flags);
4846 __wake_up_common(q, mode, nr_exclusive, 0, key);
4847 spin_unlock_irqrestore(&q->lock, flags);
4849 EXPORT_SYMBOL(__wake_up);
4852 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4854 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4856 __wake_up_common(q, mode, 1, 0, NULL);
4858 EXPORT_SYMBOL_GPL(__wake_up_locked);
4860 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4862 __wake_up_common(q, mode, 1, 0, key);
4864 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4867 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4869 * @mode: which threads
4870 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4871 * @key: opaque value to be passed to wakeup targets
4873 * The sync wakeup differs that the waker knows that it will schedule
4874 * away soon, so while the target thread will be woken up, it will not
4875 * be migrated to another CPU - ie. the two threads are 'synchronized'
4876 * with each other. This can prevent needless bouncing between CPUs.
4878 * On UP it can prevent extra preemption.
4880 * It may be assumed that this function implies a write memory barrier before
4881 * changing the task state if and only if any tasks are woken up.
4883 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4884 int nr_exclusive, void *key)
4886 unsigned long flags;
4887 int wake_flags = WF_SYNC;
4892 if (unlikely(!nr_exclusive))
4895 spin_lock_irqsave(&q->lock, flags);
4896 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4897 spin_unlock_irqrestore(&q->lock, flags);
4899 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4902 * __wake_up_sync - see __wake_up_sync_key()
4904 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4906 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4908 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4911 * complete: - signals a single thread waiting on this completion
4912 * @x: holds the state of this particular completion
4914 * This will wake up a single thread waiting on this completion. Threads will be
4915 * awakened in the same order in which they were queued.
4917 * See also complete_all(), wait_for_completion() and related routines.
4919 * It may be assumed that this function implies a write memory barrier before
4920 * changing the task state if and only if any tasks are woken up.
4922 void complete(struct completion *x)
4924 unsigned long flags;
4926 spin_lock_irqsave(&x->wait.lock, flags);
4928 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4929 spin_unlock_irqrestore(&x->wait.lock, flags);
4931 EXPORT_SYMBOL(complete);
4934 * complete_all: - signals all threads waiting on this completion
4935 * @x: holds the state of this particular completion
4937 * This will wake up all threads waiting on this particular completion event.
4939 * It may be assumed that this function implies a write memory barrier before
4940 * changing the task state if and only if any tasks are woken up.
4942 void complete_all(struct completion *x)
4944 unsigned long flags;
4946 spin_lock_irqsave(&x->wait.lock, flags);
4947 x->done += UINT_MAX/2;
4948 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4949 spin_unlock_irqrestore(&x->wait.lock, flags);
4951 EXPORT_SYMBOL(complete_all);
4953 static inline long __sched
4954 do_wait_for_common(struct completion *x, long timeout, int state)
4957 DECLARE_WAITQUEUE(wait, current);
4959 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4961 if (signal_pending_state(state, current)) {
4962 timeout = -ERESTARTSYS;
4965 __set_current_state(state);
4966 spin_unlock_irq(&x->wait.lock);
4967 timeout = schedule_timeout(timeout);
4968 spin_lock_irq(&x->wait.lock);
4969 } while (!x->done && timeout);
4970 __remove_wait_queue(&x->wait, &wait);
4975 return timeout ?: 1;
4979 wait_for_common(struct completion *x, long timeout, int state)
4983 spin_lock_irq(&x->wait.lock);
4984 timeout = do_wait_for_common(x, timeout, state);
4985 spin_unlock_irq(&x->wait.lock);
4990 * wait_for_completion: - waits for completion of a task
4991 * @x: holds the state of this particular completion
4993 * This waits to be signaled for completion of a specific task. It is NOT
4994 * interruptible and there is no timeout.
4996 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4997 * and interrupt capability. Also see complete().
4999 void __sched wait_for_completion(struct completion *x)
5001 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5003 EXPORT_SYMBOL(wait_for_completion);
5006 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5007 * @x: holds the state of this particular completion
5008 * @timeout: timeout value in jiffies
5010 * This waits for either a completion of a specific task to be signaled or for a
5011 * specified timeout to expire. The timeout is in jiffies. It is not
5014 * The return value is 0 if timed out, and positive (at least 1, or number of
5015 * jiffies left till timeout) if completed.
5017 unsigned long __sched
5018 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5020 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5022 EXPORT_SYMBOL(wait_for_completion_timeout);
5025 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5026 * @x: holds the state of this particular completion
5028 * This waits for completion of a specific task to be signaled. It is
5031 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
5033 int __sched wait_for_completion_interruptible(struct completion *x)
5035 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5036 if (t == -ERESTARTSYS)
5040 EXPORT_SYMBOL(wait_for_completion_interruptible);
5043 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5044 * @x: holds the state of this particular completion
5045 * @timeout: timeout value in jiffies
5047 * This waits for either a completion of a specific task to be signaled or for a
5048 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5050 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
5051 * positive (at least 1, or number of jiffies left till timeout) if completed.
5054 wait_for_completion_interruptible_timeout(struct completion *x,
5055 unsigned long timeout)
5057 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5059 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5062 * wait_for_completion_killable: - waits for completion of a task (killable)
5063 * @x: holds the state of this particular completion
5065 * This waits to be signaled for completion of a specific task. It can be
5066 * interrupted by a kill signal.
5068 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
5070 int __sched wait_for_completion_killable(struct completion *x)
5072 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5073 if (t == -ERESTARTSYS)
5077 EXPORT_SYMBOL(wait_for_completion_killable);
5080 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
5081 * @x: holds the state of this particular completion
5082 * @timeout: timeout value in jiffies
5084 * This waits for either a completion of a specific task to be
5085 * signaled or for a specified timeout to expire. It can be
5086 * interrupted by a kill signal. The timeout is in jiffies.
5088 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
5089 * positive (at least 1, or number of jiffies left till timeout) if completed.
5092 wait_for_completion_killable_timeout(struct completion *x,
5093 unsigned long timeout)
5095 return wait_for_common(x, timeout, TASK_KILLABLE);
5097 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
5100 * try_wait_for_completion - try to decrement a completion without blocking
5101 * @x: completion structure
5103 * Returns: 0 if a decrement cannot be done without blocking
5104 * 1 if a decrement succeeded.
5106 * If a completion is being used as a counting completion,
5107 * attempt to decrement the counter without blocking. This
5108 * enables us to avoid waiting if the resource the completion
5109 * is protecting is not available.
5111 bool try_wait_for_completion(struct completion *x)
5113 unsigned long flags;
5116 spin_lock_irqsave(&x->wait.lock, flags);
5121 spin_unlock_irqrestore(&x->wait.lock, flags);
5124 EXPORT_SYMBOL(try_wait_for_completion);
5127 * completion_done - Test to see if a completion has any waiters
5128 * @x: completion structure
5130 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5131 * 1 if there are no waiters.
5134 bool completion_done(struct completion *x)
5136 unsigned long flags;
5139 spin_lock_irqsave(&x->wait.lock, flags);
5142 spin_unlock_irqrestore(&x->wait.lock, flags);
5145 EXPORT_SYMBOL(completion_done);
5148 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5150 unsigned long flags;
5153 init_waitqueue_entry(&wait, current);
5155 __set_current_state(state);
5157 spin_lock_irqsave(&q->lock, flags);
5158 __add_wait_queue(q, &wait);
5159 spin_unlock(&q->lock);
5160 timeout = schedule_timeout(timeout);
5161 spin_lock_irq(&q->lock);
5162 __remove_wait_queue(q, &wait);
5163 spin_unlock_irqrestore(&q->lock, flags);
5168 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5170 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5172 EXPORT_SYMBOL(interruptible_sleep_on);
5175 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5177 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5179 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5181 void __sched sleep_on(wait_queue_head_t *q)
5183 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5185 EXPORT_SYMBOL(sleep_on);
5187 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5189 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5191 EXPORT_SYMBOL(sleep_on_timeout);
5193 #ifdef CONFIG_RT_MUTEXES
5196 * rt_mutex_setprio - set the current priority of a task
5198 * @prio: prio value (kernel-internal form)
5200 * This function changes the 'effective' priority of a task. It does
5201 * not touch ->normal_prio like __setscheduler().
5203 * Used by the rt_mutex code to implement priority inheritance logic.
5205 void rt_mutex_setprio(struct task_struct *p, int prio)
5207 int oldprio, on_rq, running;
5209 const struct sched_class *prev_class;
5211 BUG_ON(prio < 0 || prio > MAX_PRIO);
5213 rq = __task_rq_lock(p);
5215 trace_sched_pi_setprio(p, prio);
5217 prev_class = p->sched_class;
5219 running = task_current(rq, p);
5221 dequeue_task(rq, p, 0);
5223 p->sched_class->put_prev_task(rq, p);
5226 p->sched_class = &rt_sched_class;
5228 p->sched_class = &fair_sched_class;
5233 p->sched_class->set_curr_task(rq);
5235 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
5237 check_class_changed(rq, p, prev_class, oldprio);
5238 __task_rq_unlock(rq);
5243 void set_user_nice(struct task_struct *p, long nice)
5245 int old_prio, delta, on_rq;
5246 unsigned long flags;
5249 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5252 * We have to be careful, if called from sys_setpriority(),
5253 * the task might be in the middle of scheduling on another CPU.
5255 rq = task_rq_lock(p, &flags);
5257 * The RT priorities are set via sched_setscheduler(), but we still
5258 * allow the 'normal' nice value to be set - but as expected
5259 * it wont have any effect on scheduling until the task is
5260 * SCHED_FIFO/SCHED_RR:
5262 if (task_has_rt_policy(p)) {
5263 p->static_prio = NICE_TO_PRIO(nice);
5268 dequeue_task(rq, p, 0);
5270 p->static_prio = NICE_TO_PRIO(nice);
5273 p->prio = effective_prio(p);
5274 delta = p->prio - old_prio;
5277 enqueue_task(rq, p, 0);
5279 * If the task increased its priority or is running and
5280 * lowered its priority, then reschedule its CPU:
5282 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5283 resched_task(rq->curr);
5286 task_rq_unlock(rq, p, &flags);
5288 EXPORT_SYMBOL(set_user_nice);
5291 * can_nice - check if a task can reduce its nice value
5295 int can_nice(const struct task_struct *p, const int nice)
5297 /* convert nice value [19,-20] to rlimit style value [1,40] */
5298 int nice_rlim = 20 - nice;
5300 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5301 capable(CAP_SYS_NICE));
5304 #ifdef __ARCH_WANT_SYS_NICE
5307 * sys_nice - change the priority of the current process.
5308 * @increment: priority increment
5310 * sys_setpriority is a more generic, but much slower function that
5311 * does similar things.
5313 SYSCALL_DEFINE1(nice, int, increment)
5318 * Setpriority might change our priority at the same moment.
5319 * We don't have to worry. Conceptually one call occurs first
5320 * and we have a single winner.
5322 if (increment < -40)
5327 nice = TASK_NICE(current) + increment;
5333 if (increment < 0 && !can_nice(current, nice))
5336 retval = security_task_setnice(current, nice);
5340 set_user_nice(current, nice);
5347 * task_prio - return the priority value of a given task.
5348 * @p: the task in question.
5350 * This is the priority value as seen by users in /proc.
5351 * RT tasks are offset by -200. Normal tasks are centered
5352 * around 0, value goes from -16 to +15.
5354 int task_prio(const struct task_struct *p)
5356 return p->prio - MAX_RT_PRIO;
5360 * task_nice - return the nice value of a given task.
5361 * @p: the task in question.
5363 int task_nice(const struct task_struct *p)
5365 return TASK_NICE(p);
5367 EXPORT_SYMBOL(task_nice);
5370 * idle_cpu - is a given cpu idle currently?
5371 * @cpu: the processor in question.
5373 int idle_cpu(int cpu)
5375 struct rq *rq = cpu_rq(cpu);
5377 if (rq->curr != rq->idle)
5384 if (!llist_empty(&rq->wake_list))
5392 * idle_task - return the idle task for a given cpu.
5393 * @cpu: the processor in question.
5395 struct task_struct *idle_task(int cpu)
5397 return cpu_rq(cpu)->idle;
5401 * find_process_by_pid - find a process with a matching PID value.
5402 * @pid: the pid in question.
5404 static struct task_struct *find_process_by_pid(pid_t pid)
5406 return pid ? find_task_by_vpid(pid) : current;
5409 /* Actually do priority change: must hold rq lock. */
5411 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5414 p->rt_priority = prio;
5415 p->normal_prio = normal_prio(p);
5416 /* we are holding p->pi_lock already */
5417 p->prio = rt_mutex_getprio(p);
5418 if (rt_prio(p->prio))
5419 p->sched_class = &rt_sched_class;
5421 p->sched_class = &fair_sched_class;
5426 * check the target process has a UID that matches the current process's
5428 static bool check_same_owner(struct task_struct *p)
5430 const struct cred *cred = current_cred(), *pcred;
5434 pcred = __task_cred(p);
5435 if (cred->user->user_ns == pcred->user->user_ns)
5436 match = (cred->euid == pcred->euid ||
5437 cred->euid == pcred->uid);
5444 static int __sched_setscheduler(struct task_struct *p, int policy,
5445 const struct sched_param *param, bool user)
5447 int retval, oldprio, oldpolicy = -1, on_rq, running;
5448 unsigned long flags;
5449 const struct sched_class *prev_class;
5453 /* may grab non-irq protected spin_locks */
5454 BUG_ON(in_interrupt());
5456 /* double check policy once rq lock held */
5458 reset_on_fork = p->sched_reset_on_fork;
5459 policy = oldpolicy = p->policy;
5461 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5462 policy &= ~SCHED_RESET_ON_FORK;
5464 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5465 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5466 policy != SCHED_IDLE)
5471 * Valid priorities for SCHED_FIFO and SCHED_RR are
5472 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5473 * SCHED_BATCH and SCHED_IDLE is 0.
5475 if (param->sched_priority < 0 ||
5476 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5477 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5479 if (rt_policy(policy) != (param->sched_priority != 0))
5483 * Allow unprivileged RT tasks to decrease priority:
5485 if (user && !capable(CAP_SYS_NICE)) {
5486 if (rt_policy(policy)) {
5487 unsigned long rlim_rtprio =
5488 task_rlimit(p, RLIMIT_RTPRIO);
5490 /* can't set/change the rt policy */
5491 if (policy != p->policy && !rlim_rtprio)
5494 /* can't increase priority */
5495 if (param->sched_priority > p->rt_priority &&
5496 param->sched_priority > rlim_rtprio)
5501 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5502 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5504 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5505 if (!can_nice(p, TASK_NICE(p)))
5509 /* can't change other user's priorities */
5510 if (!check_same_owner(p))
5513 /* Normal users shall not reset the sched_reset_on_fork flag */
5514 if (p->sched_reset_on_fork && !reset_on_fork)
5519 retval = security_task_setscheduler(p);
5525 * make sure no PI-waiters arrive (or leave) while we are
5526 * changing the priority of the task:
5528 * To be able to change p->policy safely, the appropriate
5529 * runqueue lock must be held.
5531 rq = task_rq_lock(p, &flags);
5534 * Changing the policy of the stop threads its a very bad idea
5536 if (p == rq->stop) {
5537 task_rq_unlock(rq, p, &flags);
5542 * If not changing anything there's no need to proceed further:
5544 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5545 param->sched_priority == p->rt_priority))) {
5547 __task_rq_unlock(rq);
5548 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5552 #ifdef CONFIG_RT_GROUP_SCHED
5555 * Do not allow realtime tasks into groups that have no runtime
5558 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5559 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5560 !task_group_is_autogroup(task_group(p))) {
5561 task_rq_unlock(rq, p, &flags);
5567 /* recheck policy now with rq lock held */
5568 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5569 policy = oldpolicy = -1;
5570 task_rq_unlock(rq, p, &flags);
5574 running = task_current(rq, p);
5576 deactivate_task(rq, p, 0);
5578 p->sched_class->put_prev_task(rq, p);
5580 p->sched_reset_on_fork = reset_on_fork;
5583 prev_class = p->sched_class;
5584 __setscheduler(rq, p, policy, param->sched_priority);
5587 p->sched_class->set_curr_task(rq);
5589 activate_task(rq, p, 0);
5591 check_class_changed(rq, p, prev_class, oldprio);
5592 task_rq_unlock(rq, p, &flags);
5594 rt_mutex_adjust_pi(p);
5600 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5601 * @p: the task in question.
5602 * @policy: new policy.
5603 * @param: structure containing the new RT priority.
5605 * NOTE that the task may be already dead.
5607 int sched_setscheduler(struct task_struct *p, int policy,
5608 const struct sched_param *param)
5610 return __sched_setscheduler(p, policy, param, true);
5612 EXPORT_SYMBOL_GPL(sched_setscheduler);
5615 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5616 * @p: the task in question.
5617 * @policy: new policy.
5618 * @param: structure containing the new RT priority.
5620 * Just like sched_setscheduler, only don't bother checking if the
5621 * current context has permission. For example, this is needed in
5622 * stop_machine(): we create temporary high priority worker threads,
5623 * but our caller might not have that capability.
5625 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5626 const struct sched_param *param)
5628 return __sched_setscheduler(p, policy, param, false);
5632 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5634 struct sched_param lparam;
5635 struct task_struct *p;
5638 if (!param || pid < 0)
5640 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5645 p = find_process_by_pid(pid);
5647 retval = sched_setscheduler(p, policy, &lparam);
5654 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5655 * @pid: the pid in question.
5656 * @policy: new policy.
5657 * @param: structure containing the new RT priority.
5659 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5660 struct sched_param __user *, param)
5662 /* negative values for policy are not valid */
5666 return do_sched_setscheduler(pid, policy, param);
5670 * sys_sched_setparam - set/change the RT priority of a thread
5671 * @pid: the pid in question.
5672 * @param: structure containing the new RT priority.
5674 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5676 return do_sched_setscheduler(pid, -1, param);
5680 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5681 * @pid: the pid in question.
5683 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5685 struct task_struct *p;
5693 p = find_process_by_pid(pid);
5695 retval = security_task_getscheduler(p);
5698 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5705 * sys_sched_getparam - get the RT priority of a thread
5706 * @pid: the pid in question.
5707 * @param: structure containing the RT priority.
5709 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5711 struct sched_param lp;
5712 struct task_struct *p;
5715 if (!param || pid < 0)
5719 p = find_process_by_pid(pid);
5724 retval = security_task_getscheduler(p);
5728 lp.sched_priority = p->rt_priority;
5732 * This one might sleep, we cannot do it with a spinlock held ...
5734 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5743 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5745 cpumask_var_t cpus_allowed, new_mask;
5746 struct task_struct *p;
5752 p = find_process_by_pid(pid);
5759 /* Prevent p going away */
5763 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5767 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5769 goto out_free_cpus_allowed;
5772 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5775 retval = security_task_setscheduler(p);
5779 cpuset_cpus_allowed(p, cpus_allowed);
5780 cpumask_and(new_mask, in_mask, cpus_allowed);
5782 retval = set_cpus_allowed_ptr(p, new_mask);
5785 cpuset_cpus_allowed(p, cpus_allowed);
5786 if (!cpumask_subset(new_mask, cpus_allowed)) {
5788 * We must have raced with a concurrent cpuset
5789 * update. Just reset the cpus_allowed to the
5790 * cpuset's cpus_allowed
5792 cpumask_copy(new_mask, cpus_allowed);
5797 free_cpumask_var(new_mask);
5798 out_free_cpus_allowed:
5799 free_cpumask_var(cpus_allowed);
5806 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5807 struct cpumask *new_mask)
5809 if (len < cpumask_size())
5810 cpumask_clear(new_mask);
5811 else if (len > cpumask_size())
5812 len = cpumask_size();
5814 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5818 * sys_sched_setaffinity - set the cpu affinity of a process
5819 * @pid: pid of the process
5820 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5821 * @user_mask_ptr: user-space pointer to the new cpu mask
5823 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5824 unsigned long __user *, user_mask_ptr)
5826 cpumask_var_t new_mask;
5829 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5832 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5834 retval = sched_setaffinity(pid, new_mask);
5835 free_cpumask_var(new_mask);
5839 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5841 struct task_struct *p;
5842 unsigned long flags;
5849 p = find_process_by_pid(pid);
5853 retval = security_task_getscheduler(p);
5857 raw_spin_lock_irqsave(&p->pi_lock, flags);
5858 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5859 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5869 * sys_sched_getaffinity - get the cpu affinity of a process
5870 * @pid: pid of the process
5871 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5872 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5874 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5875 unsigned long __user *, user_mask_ptr)
5880 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5882 if (len & (sizeof(unsigned long)-1))
5885 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5888 ret = sched_getaffinity(pid, mask);
5890 size_t retlen = min_t(size_t, len, cpumask_size());
5892 if (copy_to_user(user_mask_ptr, mask, retlen))
5897 free_cpumask_var(mask);
5903 * sys_sched_yield - yield the current processor to other threads.
5905 * This function yields the current CPU to other tasks. If there are no
5906 * other threads running on this CPU then this function will return.
5908 SYSCALL_DEFINE0(sched_yield)
5910 struct rq *rq = this_rq_lock();
5912 schedstat_inc(rq, yld_count);
5913 current->sched_class->yield_task(rq);
5916 * Since we are going to call schedule() anyway, there's
5917 * no need to preempt or enable interrupts:
5919 __release(rq->lock);
5920 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5921 do_raw_spin_unlock(&rq->lock);
5922 preempt_enable_no_resched();
5929 static inline int should_resched(void)
5931 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5934 static void __cond_resched(void)
5936 add_preempt_count(PREEMPT_ACTIVE);
5938 sub_preempt_count(PREEMPT_ACTIVE);
5941 int __sched _cond_resched(void)
5943 if (should_resched()) {
5949 EXPORT_SYMBOL(_cond_resched);
5952 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5953 * call schedule, and on return reacquire the lock.
5955 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5956 * operations here to prevent schedule() from being called twice (once via
5957 * spin_unlock(), once by hand).
5959 int __cond_resched_lock(spinlock_t *lock)
5961 int resched = should_resched();
5964 lockdep_assert_held(lock);
5966 if (spin_needbreak(lock) || resched) {
5977 EXPORT_SYMBOL(__cond_resched_lock);
5979 int __sched __cond_resched_softirq(void)
5981 BUG_ON(!in_softirq());
5983 if (should_resched()) {
5991 EXPORT_SYMBOL(__cond_resched_softirq);
5994 * yield - yield the current processor to other threads.
5996 * This is a shortcut for kernel-space yielding - it marks the
5997 * thread runnable and calls sys_sched_yield().
5999 void __sched yield(void)
6001 set_current_state(TASK_RUNNING);
6004 EXPORT_SYMBOL(yield);
6007 * yield_to - yield the current processor to another thread in
6008 * your thread group, or accelerate that thread toward the
6009 * processor it's on.
6011 * @preempt: whether task preemption is allowed or not
6013 * It's the caller's job to ensure that the target task struct
6014 * can't go away on us before we can do any checks.
6016 * Returns true if we indeed boosted the target task.
6018 bool __sched yield_to(struct task_struct *p, bool preempt)
6020 struct task_struct *curr = current;
6021 struct rq *rq, *p_rq;
6022 unsigned long flags;
6025 local_irq_save(flags);
6030 double_rq_lock(rq, p_rq);
6031 while (task_rq(p) != p_rq) {
6032 double_rq_unlock(rq, p_rq);
6036 if (!curr->sched_class->yield_to_task)
6039 if (curr->sched_class != p->sched_class)
6042 if (task_running(p_rq, p) || p->state)
6045 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
6047 schedstat_inc(rq, yld_count);
6049 * Make p's CPU reschedule; pick_next_entity takes care of
6052 if (preempt && rq != p_rq)
6053 resched_task(p_rq->curr);
6057 double_rq_unlock(rq, p_rq);
6058 local_irq_restore(flags);
6065 EXPORT_SYMBOL_GPL(yield_to);
6068 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6069 * that process accounting knows that this is a task in IO wait state.
6071 void __sched io_schedule(void)
6073 struct rq *rq = raw_rq();
6075 delayacct_blkio_start();
6076 atomic_inc(&rq->nr_iowait);
6077 blk_flush_plug(current);
6078 current->in_iowait = 1;
6080 current->in_iowait = 0;
6081 atomic_dec(&rq->nr_iowait);
6082 delayacct_blkio_end();
6084 EXPORT_SYMBOL(io_schedule);
6086 long __sched io_schedule_timeout(long timeout)
6088 struct rq *rq = raw_rq();
6091 delayacct_blkio_start();
6092 atomic_inc(&rq->nr_iowait);
6093 blk_flush_plug(current);
6094 current->in_iowait = 1;
6095 ret = schedule_timeout(timeout);
6096 current->in_iowait = 0;
6097 atomic_dec(&rq->nr_iowait);
6098 delayacct_blkio_end();
6103 * sys_sched_get_priority_max - return maximum RT priority.
6104 * @policy: scheduling class.
6106 * this syscall returns the maximum rt_priority that can be used
6107 * by a given scheduling class.
6109 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6116 ret = MAX_USER_RT_PRIO-1;
6128 * sys_sched_get_priority_min - return minimum RT priority.
6129 * @policy: scheduling class.
6131 * this syscall returns the minimum rt_priority that can be used
6132 * by a given scheduling class.
6134 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6152 * sys_sched_rr_get_interval - return the default timeslice of a process.
6153 * @pid: pid of the process.
6154 * @interval: userspace pointer to the timeslice value.
6156 * this syscall writes the default timeslice value of a given process
6157 * into the user-space timespec buffer. A value of '0' means infinity.
6159 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6160 struct timespec __user *, interval)
6162 struct task_struct *p;
6163 unsigned int time_slice;
6164 unsigned long flags;
6174 p = find_process_by_pid(pid);
6178 retval = security_task_getscheduler(p);
6182 rq = task_rq_lock(p, &flags);
6183 time_slice = p->sched_class->get_rr_interval(rq, p);
6184 task_rq_unlock(rq, p, &flags);
6187 jiffies_to_timespec(time_slice, &t);
6188 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6196 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6198 void sched_show_task(struct task_struct *p)
6200 unsigned long free = 0;
6203 state = p->state ? __ffs(p->state) + 1 : 0;
6204 printk(KERN_INFO "%-15.15s %c", p->comm,
6205 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6206 #if BITS_PER_LONG == 32
6207 if (state == TASK_RUNNING)
6208 printk(KERN_CONT " running ");
6210 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6212 if (state == TASK_RUNNING)
6213 printk(KERN_CONT " running task ");
6215 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6217 #ifdef CONFIG_DEBUG_STACK_USAGE
6218 free = stack_not_used(p);
6220 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6221 task_pid_nr(p), task_pid_nr(p->real_parent),
6222 (unsigned long)task_thread_info(p)->flags);
6224 show_stack(p, NULL);
6227 void show_state_filter(unsigned long state_filter)
6229 struct task_struct *g, *p;
6231 #if BITS_PER_LONG == 32
6233 " task PC stack pid father\n");
6236 " task PC stack pid father\n");
6239 do_each_thread(g, p) {
6241 * reset the NMI-timeout, listing all files on a slow
6242 * console might take a lot of time:
6244 touch_nmi_watchdog();
6245 if (!state_filter || (p->state & state_filter))
6247 } while_each_thread(g, p);
6249 touch_all_softlockup_watchdogs();
6251 #ifdef CONFIG_SCHED_DEBUG
6252 sysrq_sched_debug_show();
6256 * Only show locks if all tasks are dumped:
6259 debug_show_all_locks();
6262 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6264 idle->sched_class = &idle_sched_class;
6268 * init_idle - set up an idle thread for a given CPU
6269 * @idle: task in question
6270 * @cpu: cpu the idle task belongs to
6272 * NOTE: this function does not set the idle thread's NEED_RESCHED
6273 * flag, to make booting more robust.
6275 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6277 struct rq *rq = cpu_rq(cpu);
6278 unsigned long flags;
6280 raw_spin_lock_irqsave(&rq->lock, flags);
6283 idle->state = TASK_RUNNING;
6284 idle->se.exec_start = sched_clock();
6286 do_set_cpus_allowed(idle, cpumask_of(cpu));
6288 * We're having a chicken and egg problem, even though we are
6289 * holding rq->lock, the cpu isn't yet set to this cpu so the
6290 * lockdep check in task_group() will fail.
6292 * Similar case to sched_fork(). / Alternatively we could
6293 * use task_rq_lock() here and obtain the other rq->lock.
6298 __set_task_cpu(idle, cpu);
6301 rq->curr = rq->idle = idle;
6302 #if defined(CONFIG_SMP)
6305 raw_spin_unlock_irqrestore(&rq->lock, flags);
6307 /* Set the preempt count _outside_ the spinlocks! */
6308 task_thread_info(idle)->preempt_count = 0;
6311 * The idle tasks have their own, simple scheduling class:
6313 idle->sched_class = &idle_sched_class;
6314 ftrace_graph_init_idle_task(idle, cpu);
6315 #if defined(CONFIG_SMP)
6316 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6321 * Increase the granularity value when there are more CPUs,
6322 * because with more CPUs the 'effective latency' as visible
6323 * to users decreases. But the relationship is not linear,
6324 * so pick a second-best guess by going with the log2 of the
6327 * This idea comes from the SD scheduler of Con Kolivas:
6329 static int get_update_sysctl_factor(void)
6331 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6332 unsigned int factor;
6334 switch (sysctl_sched_tunable_scaling) {
6335 case SCHED_TUNABLESCALING_NONE:
6338 case SCHED_TUNABLESCALING_LINEAR:
6341 case SCHED_TUNABLESCALING_LOG:
6343 factor = 1 + ilog2(cpus);
6350 static void update_sysctl(void)
6352 unsigned int factor = get_update_sysctl_factor();
6354 #define SET_SYSCTL(name) \
6355 (sysctl_##name = (factor) * normalized_sysctl_##name)
6356 SET_SYSCTL(sched_min_granularity);
6357 SET_SYSCTL(sched_latency);
6358 SET_SYSCTL(sched_wakeup_granularity);
6362 static inline void sched_init_granularity(void)
6368 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6370 if (p->sched_class && p->sched_class->set_cpus_allowed)
6371 p->sched_class->set_cpus_allowed(p, new_mask);
6373 cpumask_copy(&p->cpus_allowed, new_mask);
6374 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6378 * This is how migration works:
6380 * 1) we invoke migration_cpu_stop() on the target CPU using
6382 * 2) stopper starts to run (implicitly forcing the migrated thread
6384 * 3) it checks whether the migrated task is still in the wrong runqueue.
6385 * 4) if it's in the wrong runqueue then the migration thread removes
6386 * it and puts it into the right queue.
6387 * 5) stopper completes and stop_one_cpu() returns and the migration
6392 * Change a given task's CPU affinity. Migrate the thread to a
6393 * proper CPU and schedule it away if the CPU it's executing on
6394 * is removed from the allowed bitmask.
6396 * NOTE: the caller must have a valid reference to the task, the
6397 * task must not exit() & deallocate itself prematurely. The
6398 * call is not atomic; no spinlocks may be held.
6400 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6402 unsigned long flags;
6404 unsigned int dest_cpu;
6407 rq = task_rq_lock(p, &flags);
6409 if (cpumask_equal(&p->cpus_allowed, new_mask))
6412 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6417 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6422 do_set_cpus_allowed(p, new_mask);
6424 /* Can the task run on the task's current CPU? If so, we're done */
6425 if (cpumask_test_cpu(task_cpu(p), new_mask))
6428 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6430 struct migration_arg arg = { p, dest_cpu };
6431 /* Need help from migration thread: drop lock and wait. */
6432 task_rq_unlock(rq, p, &flags);
6433 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6434 tlb_migrate_finish(p->mm);
6438 task_rq_unlock(rq, p, &flags);
6442 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6445 * Move (not current) task off this cpu, onto dest cpu. We're doing
6446 * this because either it can't run here any more (set_cpus_allowed()
6447 * away from this CPU, or CPU going down), or because we're
6448 * attempting to rebalance this task on exec (sched_exec).
6450 * So we race with normal scheduler movements, but that's OK, as long
6451 * as the task is no longer on this CPU.
6453 * Returns non-zero if task was successfully migrated.
6455 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6457 struct rq *rq_dest, *rq_src;
6460 if (unlikely(!cpu_active(dest_cpu)))
6463 rq_src = cpu_rq(src_cpu);
6464 rq_dest = cpu_rq(dest_cpu);
6466 raw_spin_lock(&p->pi_lock);
6467 double_rq_lock(rq_src, rq_dest);
6468 /* Already moved. */
6469 if (task_cpu(p) != src_cpu)
6471 /* Affinity changed (again). */
6472 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
6476 * If we're not on a rq, the next wake-up will ensure we're
6480 deactivate_task(rq_src, p, 0);
6481 set_task_cpu(p, dest_cpu);
6482 activate_task(rq_dest, p, 0);
6483 check_preempt_curr(rq_dest, p, 0);
6488 double_rq_unlock(rq_src, rq_dest);
6489 raw_spin_unlock(&p->pi_lock);
6494 * migration_cpu_stop - this will be executed by a highprio stopper thread
6495 * and performs thread migration by bumping thread off CPU then
6496 * 'pushing' onto another runqueue.
6498 static int migration_cpu_stop(void *data)
6500 struct migration_arg *arg = data;
6503 * The original target cpu might have gone down and we might
6504 * be on another cpu but it doesn't matter.
6506 local_irq_disable();
6507 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6512 #ifdef CONFIG_HOTPLUG_CPU
6515 * Ensures that the idle task is using init_mm right before its cpu goes
6518 void idle_task_exit(void)
6520 struct mm_struct *mm = current->active_mm;
6522 BUG_ON(cpu_online(smp_processor_id()));
6525 switch_mm(mm, &init_mm, current);
6530 * While a dead CPU has no uninterruptible tasks queued at this point,
6531 * it might still have a nonzero ->nr_uninterruptible counter, because
6532 * for performance reasons the counter is not stricly tracking tasks to
6533 * their home CPUs. So we just add the counter to another CPU's counter,
6534 * to keep the global sum constant after CPU-down:
6536 static void migrate_nr_uninterruptible(struct rq *rq_src)
6538 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6540 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6541 rq_src->nr_uninterruptible = 0;
6545 * remove the tasks which were accounted by rq from calc_load_tasks.
6547 static void calc_global_load_remove(struct rq *rq)
6549 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6550 rq->calc_load_active = 0;
6553 #ifdef CONFIG_CFS_BANDWIDTH
6554 static void unthrottle_offline_cfs_rqs(struct rq *rq)
6556 struct cfs_rq *cfs_rq;
6558 for_each_leaf_cfs_rq(rq, cfs_rq) {
6559 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6561 if (!cfs_rq->runtime_enabled)
6565 * clock_task is not advancing so we just need to make sure
6566 * there's some valid quota amount
6568 cfs_rq->runtime_remaining = cfs_b->quota;
6569 if (cfs_rq_throttled(cfs_rq))
6570 unthrottle_cfs_rq(cfs_rq);
6576 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6577 * try_to_wake_up()->select_task_rq().
6579 * Called with rq->lock held even though we'er in stop_machine() and
6580 * there's no concurrency possible, we hold the required locks anyway
6581 * because of lock validation efforts.
6583 static void migrate_tasks(unsigned int dead_cpu)
6585 struct rq *rq = cpu_rq(dead_cpu);
6586 struct task_struct *next, *stop = rq->stop;
6590 * Fudge the rq selection such that the below task selection loop
6591 * doesn't get stuck on the currently eligible stop task.
6593 * We're currently inside stop_machine() and the rq is either stuck
6594 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6595 * either way we should never end up calling schedule() until we're
6602 * There's this thread running, bail when that's the only
6605 if (rq->nr_running == 1)
6608 next = pick_next_task(rq);
6610 next->sched_class->put_prev_task(rq, next);
6612 /* Find suitable destination for @next, with force if needed. */
6613 dest_cpu = select_fallback_rq(dead_cpu, next);
6614 raw_spin_unlock(&rq->lock);
6616 __migrate_task(next, dead_cpu, dest_cpu);
6618 raw_spin_lock(&rq->lock);
6624 #endif /* CONFIG_HOTPLUG_CPU */
6626 #if !defined(CONFIG_HOTPLUG_CPU) || !defined(CONFIG_CFS_BANDWIDTH)
6627 static void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6630 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6632 static struct ctl_table sd_ctl_dir[] = {
6634 .procname = "sched_domain",
6640 static struct ctl_table sd_ctl_root[] = {
6642 .procname = "kernel",
6644 .child = sd_ctl_dir,
6649 static struct ctl_table *sd_alloc_ctl_entry(int n)
6651 struct ctl_table *entry =
6652 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6657 static void sd_free_ctl_entry(struct ctl_table **tablep)
6659 struct ctl_table *entry;
6662 * In the intermediate directories, both the child directory and
6663 * procname are dynamically allocated and could fail but the mode
6664 * will always be set. In the lowest directory the names are
6665 * static strings and all have proc handlers.
6667 for (entry = *tablep; entry->mode; entry++) {
6669 sd_free_ctl_entry(&entry->child);
6670 if (entry->proc_handler == NULL)
6671 kfree(entry->procname);
6678 static int min_load_idx = 0;
6679 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
6682 set_table_entry(struct ctl_table *entry,
6683 const char *procname, void *data, int maxlen,
6684 mode_t mode, proc_handler *proc_handler,
6687 entry->procname = procname;
6689 entry->maxlen = maxlen;
6691 entry->proc_handler = proc_handler;
6694 entry->extra1 = &min_load_idx;
6695 entry->extra2 = &max_load_idx;
6699 static struct ctl_table *
6700 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6702 struct ctl_table *table = sd_alloc_ctl_entry(13);
6707 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6708 sizeof(long), 0644, proc_doulongvec_minmax, false);
6709 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6710 sizeof(long), 0644, proc_doulongvec_minmax, false);
6711 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6712 sizeof(int), 0644, proc_dointvec_minmax, true);
6713 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6714 sizeof(int), 0644, proc_dointvec_minmax, true);
6715 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6716 sizeof(int), 0644, proc_dointvec_minmax, true);
6717 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6718 sizeof(int), 0644, proc_dointvec_minmax, true);
6719 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6720 sizeof(int), 0644, proc_dointvec_minmax, true);
6721 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6722 sizeof(int), 0644, proc_dointvec_minmax, false);
6723 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6724 sizeof(int), 0644, proc_dointvec_minmax, false);
6725 set_table_entry(&table[9], "cache_nice_tries",
6726 &sd->cache_nice_tries,
6727 sizeof(int), 0644, proc_dointvec_minmax, false);
6728 set_table_entry(&table[10], "flags", &sd->flags,
6729 sizeof(int), 0644, proc_dointvec_minmax, false);
6730 set_table_entry(&table[11], "name", sd->name,
6731 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
6732 /* &table[12] is terminator */
6737 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6739 struct ctl_table *entry, *table;
6740 struct sched_domain *sd;
6741 int domain_num = 0, i;
6744 for_each_domain(cpu, sd)
6746 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6751 for_each_domain(cpu, sd) {
6752 snprintf(buf, 32, "domain%d", i);
6753 entry->procname = kstrdup(buf, GFP_KERNEL);
6755 entry->child = sd_alloc_ctl_domain_table(sd);
6762 static struct ctl_table_header *sd_sysctl_header;
6763 static void register_sched_domain_sysctl(void)
6765 int i, cpu_num = num_possible_cpus();
6766 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6769 WARN_ON(sd_ctl_dir[0].child);
6770 sd_ctl_dir[0].child = entry;
6775 for_each_possible_cpu(i) {
6776 snprintf(buf, 32, "cpu%d", i);
6777 entry->procname = kstrdup(buf, GFP_KERNEL);
6779 entry->child = sd_alloc_ctl_cpu_table(i);
6783 WARN_ON(sd_sysctl_header);
6784 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6787 /* may be called multiple times per register */
6788 static void unregister_sched_domain_sysctl(void)
6790 if (sd_sysctl_header)
6791 unregister_sysctl_table(sd_sysctl_header);
6792 sd_sysctl_header = NULL;
6793 if (sd_ctl_dir[0].child)
6794 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6797 static void register_sched_domain_sysctl(void)
6800 static void unregister_sched_domain_sysctl(void)
6805 static void set_rq_online(struct rq *rq)
6808 const struct sched_class *class;
6810 cpumask_set_cpu(rq->cpu, rq->rd->online);
6813 for_each_class(class) {
6814 if (class->rq_online)
6815 class->rq_online(rq);
6820 static void set_rq_offline(struct rq *rq)
6823 const struct sched_class *class;
6825 for_each_class(class) {
6826 if (class->rq_offline)
6827 class->rq_offline(rq);
6830 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6836 * migration_call - callback that gets triggered when a CPU is added.
6837 * Here we can start up the necessary migration thread for the new CPU.
6839 static int __cpuinit
6840 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6842 int cpu = (long)hcpu;
6843 unsigned long flags;
6844 struct rq *rq = cpu_rq(cpu);
6846 switch (action & ~CPU_TASKS_FROZEN) {
6848 case CPU_UP_PREPARE:
6849 rq->calc_load_update = calc_load_update;
6853 /* Update our root-domain */
6854 raw_spin_lock_irqsave(&rq->lock, flags);
6856 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6860 raw_spin_unlock_irqrestore(&rq->lock, flags);
6863 #ifdef CONFIG_HOTPLUG_CPU
6865 sched_ttwu_pending();
6866 /* Update our root-domain */
6867 raw_spin_lock_irqsave(&rq->lock, flags);
6869 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6873 BUG_ON(rq->nr_running != 1); /* the migration thread */
6874 raw_spin_unlock_irqrestore(&rq->lock, flags);
6876 migrate_nr_uninterruptible(rq);
6877 calc_global_load_remove(rq);
6882 update_max_interval();
6888 * Register at high priority so that task migration (migrate_all_tasks)
6889 * happens before everything else. This has to be lower priority than
6890 * the notifier in the perf_event subsystem, though.
6892 static struct notifier_block __cpuinitdata migration_notifier = {
6893 .notifier_call = migration_call,
6894 .priority = CPU_PRI_MIGRATION,
6897 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6898 unsigned long action, void *hcpu)
6900 switch (action & ~CPU_TASKS_FROZEN) {
6902 case CPU_DOWN_FAILED:
6903 set_cpu_active((long)hcpu, true);
6910 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6911 unsigned long action, void *hcpu)
6913 switch (action & ~CPU_TASKS_FROZEN) {
6914 case CPU_DOWN_PREPARE:
6915 set_cpu_active((long)hcpu, false);
6922 static int __init migration_init(void)
6924 void *cpu = (void *)(long)smp_processor_id();
6927 /* Initialize migration for the boot CPU */
6928 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6929 BUG_ON(err == NOTIFY_BAD);
6930 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6931 register_cpu_notifier(&migration_notifier);
6933 /* Register cpu active notifiers */
6934 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6935 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6939 early_initcall(migration_init);
6944 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6946 #ifdef CONFIG_SCHED_DEBUG
6948 static __read_mostly int sched_domain_debug_enabled;
6950 static int __init sched_domain_debug_setup(char *str)
6952 sched_domain_debug_enabled = 1;
6956 early_param("sched_debug", sched_domain_debug_setup);
6958 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6959 struct cpumask *groupmask)
6961 struct sched_group *group = sd->groups;
6964 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6965 cpumask_clear(groupmask);
6967 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6969 if (!(sd->flags & SD_LOAD_BALANCE)) {
6970 printk("does not load-balance\n");
6972 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6977 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6979 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6980 printk(KERN_ERR "ERROR: domain->span does not contain "
6983 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6984 printk(KERN_ERR "ERROR: domain->groups does not contain"
6988 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6992 printk(KERN_ERR "ERROR: group is NULL\n");
6996 if (!group->sgp->power) {
6997 printk(KERN_CONT "\n");
6998 printk(KERN_ERR "ERROR: domain->cpu_power not "
7003 if (!cpumask_weight(sched_group_cpus(group))) {
7004 printk(KERN_CONT "\n");
7005 printk(KERN_ERR "ERROR: empty group\n");
7009 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7010 printk(KERN_CONT "\n");
7011 printk(KERN_ERR "ERROR: repeated CPUs\n");
7015 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7017 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7019 printk(KERN_CONT " %s", str);
7020 if (group->sgp->power != SCHED_POWER_SCALE) {
7021 printk(KERN_CONT " (cpu_power = %d)",
7025 group = group->next;
7026 } while (group != sd->groups);
7027 printk(KERN_CONT "\n");
7029 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7030 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7033 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7034 printk(KERN_ERR "ERROR: parent span is not a superset "
7035 "of domain->span\n");
7039 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7043 if (!sched_domain_debug_enabled)
7047 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7051 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7054 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
7062 #else /* !CONFIG_SCHED_DEBUG */
7063 # define sched_domain_debug(sd, cpu) do { } while (0)
7064 #endif /* CONFIG_SCHED_DEBUG */
7066 static int sd_degenerate(struct sched_domain *sd)
7068 if (cpumask_weight(sched_domain_span(sd)) == 1)
7071 /* Following flags need at least 2 groups */
7072 if (sd->flags & (SD_LOAD_BALANCE |
7073 SD_BALANCE_NEWIDLE |
7077 SD_SHARE_PKG_RESOURCES)) {
7078 if (sd->groups != sd->groups->next)
7082 /* Following flags don't use groups */
7083 if (sd->flags & (SD_WAKE_AFFINE))
7090 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7092 unsigned long cflags = sd->flags, pflags = parent->flags;
7094 if (sd_degenerate(parent))
7097 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7100 /* Flags needing groups don't count if only 1 group in parent */
7101 if (parent->groups == parent->groups->next) {
7102 pflags &= ~(SD_LOAD_BALANCE |
7103 SD_BALANCE_NEWIDLE |
7107 SD_SHARE_PKG_RESOURCES);
7108 if (nr_node_ids == 1)
7109 pflags &= ~SD_SERIALIZE;
7111 if (~cflags & pflags)
7117 static void free_rootdomain(struct rcu_head *rcu)
7119 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
7121 cpupri_cleanup(&rd->cpupri);
7122 free_cpumask_var(rd->rto_mask);
7123 free_cpumask_var(rd->online);
7124 free_cpumask_var(rd->span);
7128 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7130 struct root_domain *old_rd = NULL;
7131 unsigned long flags;
7133 raw_spin_lock_irqsave(&rq->lock, flags);
7138 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7141 cpumask_clear_cpu(rq->cpu, old_rd->span);
7144 * If we dont want to free the old_rt yet then
7145 * set old_rd to NULL to skip the freeing later
7148 if (!atomic_dec_and_test(&old_rd->refcount))
7152 atomic_inc(&rd->refcount);
7155 cpumask_set_cpu(rq->cpu, rd->span);
7156 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7159 raw_spin_unlock_irqrestore(&rq->lock, flags);
7162 call_rcu_sched(&old_rd->rcu, free_rootdomain);
7165 static int init_rootdomain(struct root_domain *rd)
7167 memset(rd, 0, sizeof(*rd));
7169 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7171 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7173 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7176 if (cpupri_init(&rd->cpupri) != 0)
7181 free_cpumask_var(rd->rto_mask);
7183 free_cpumask_var(rd->online);
7185 free_cpumask_var(rd->span);
7190 static void init_defrootdomain(void)
7192 init_rootdomain(&def_root_domain);
7194 atomic_set(&def_root_domain.refcount, 1);
7197 static struct root_domain *alloc_rootdomain(void)
7199 struct root_domain *rd;
7201 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7205 if (init_rootdomain(rd) != 0) {
7213 static void free_sched_groups(struct sched_group *sg, int free_sgp)
7215 struct sched_group *tmp, *first;
7224 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
7229 } while (sg != first);
7232 static void free_sched_domain(struct rcu_head *rcu)
7234 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
7237 * If its an overlapping domain it has private groups, iterate and
7240 if (sd->flags & SD_OVERLAP) {
7241 free_sched_groups(sd->groups, 1);
7242 } else if (atomic_dec_and_test(&sd->groups->ref)) {
7243 kfree(sd->groups->sgp);
7249 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
7251 call_rcu(&sd->rcu, free_sched_domain);
7254 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
7256 for (; sd; sd = sd->parent)
7257 destroy_sched_domain(sd, cpu);
7261 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7262 * hold the hotplug lock.
7265 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7267 struct rq *rq = cpu_rq(cpu);
7268 struct sched_domain *tmp;
7270 /* Remove the sched domains which do not contribute to scheduling. */
7271 for (tmp = sd; tmp; ) {
7272 struct sched_domain *parent = tmp->parent;
7276 if (sd_parent_degenerate(tmp, parent)) {
7277 tmp->parent = parent->parent;
7279 parent->parent->child = tmp;
7280 destroy_sched_domain(parent, cpu);
7285 if (sd && sd_degenerate(sd)) {
7288 destroy_sched_domain(tmp, cpu);
7293 sched_domain_debug(sd, cpu);
7295 rq_attach_root(rq, rd);
7297 rcu_assign_pointer(rq->sd, sd);
7298 destroy_sched_domains(tmp, cpu);
7301 /* cpus with isolated domains */
7302 static cpumask_var_t cpu_isolated_map;
7304 /* Setup the mask of cpus configured for isolated domains */
7305 static int __init isolated_cpu_setup(char *str)
7307 alloc_bootmem_cpumask_var(&cpu_isolated_map);
7308 cpulist_parse(str, cpu_isolated_map);
7312 __setup("isolcpus=", isolated_cpu_setup);
7317 * find_next_best_node - find the next node to include in a sched_domain
7318 * @node: node whose sched_domain we're building
7319 * @used_nodes: nodes already in the sched_domain
7321 * Find the next node to include in a given scheduling domain. Simply
7322 * finds the closest node not already in the @used_nodes map.
7324 * Should use nodemask_t.
7326 static int find_next_best_node(int node, nodemask_t *used_nodes)
7328 int i, n, val, min_val, best_node = -1;
7332 for (i = 0; i < nr_node_ids; i++) {
7333 /* Start at @node */
7334 n = (node + i) % nr_node_ids;
7336 if (!nr_cpus_node(n))
7339 /* Skip already used nodes */
7340 if (node_isset(n, *used_nodes))
7343 /* Simple min distance search */
7344 val = node_distance(node, n);
7346 if (val < min_val) {
7352 if (best_node != -1)
7353 node_set(best_node, *used_nodes);
7358 * sched_domain_node_span - get a cpumask for a node's sched_domain
7359 * @node: node whose cpumask we're constructing
7360 * @span: resulting cpumask
7362 * Given a node, construct a good cpumask for its sched_domain to span. It
7363 * should be one that prevents unnecessary balancing, but also spreads tasks
7366 static void sched_domain_node_span(int node, struct cpumask *span)
7368 nodemask_t used_nodes;
7371 cpumask_clear(span);
7372 nodes_clear(used_nodes);
7374 cpumask_or(span, span, cpumask_of_node(node));
7375 node_set(node, used_nodes);
7377 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7378 int next_node = find_next_best_node(node, &used_nodes);
7381 cpumask_or(span, span, cpumask_of_node(next_node));
7385 static const struct cpumask *cpu_node_mask(int cpu)
7387 lockdep_assert_held(&sched_domains_mutex);
7389 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7391 return sched_domains_tmpmask;
7394 static const struct cpumask *cpu_allnodes_mask(int cpu)
7396 return cpu_possible_mask;
7398 #endif /* CONFIG_NUMA */
7400 static const struct cpumask *cpu_cpu_mask(int cpu)
7402 return cpumask_of_node(cpu_to_node(cpu));
7405 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7408 struct sched_domain **__percpu sd;
7409 struct sched_group **__percpu sg;
7410 struct sched_group_power **__percpu sgp;
7414 struct sched_domain ** __percpu sd;
7415 struct root_domain *rd;
7425 struct sched_domain_topology_level;
7427 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7428 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7430 #define SDTL_OVERLAP 0x01
7432 struct sched_domain_topology_level {
7433 sched_domain_init_f init;
7434 sched_domain_mask_f mask;
7436 struct sd_data data;
7440 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7442 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7443 const struct cpumask *span = sched_domain_span(sd);
7444 struct cpumask *covered = sched_domains_tmpmask;
7445 struct sd_data *sdd = sd->private;
7446 struct sched_domain *child;
7449 cpumask_clear(covered);
7451 for_each_cpu(i, span) {
7452 struct cpumask *sg_span;
7454 if (cpumask_test_cpu(i, covered))
7457 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7458 GFP_KERNEL, cpu_to_node(i));
7463 sg_span = sched_group_cpus(sg);
7465 child = *per_cpu_ptr(sdd->sd, i);
7467 child = child->child;
7468 cpumask_copy(sg_span, sched_domain_span(child));
7470 cpumask_set_cpu(i, sg_span);
7472 cpumask_or(covered, covered, sg_span);
7474 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7475 atomic_inc(&sg->sgp->ref);
7477 if (cpumask_test_cpu(cpu, sg_span))
7487 sd->groups = groups;
7492 free_sched_groups(first, 0);
7497 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7499 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7500 struct sched_domain *child = sd->child;
7503 cpu = cpumask_first(sched_domain_span(child));
7506 *sg = *per_cpu_ptr(sdd->sg, cpu);
7507 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7508 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7515 * build_sched_groups will build a circular linked list of the groups
7516 * covered by the given span, and will set each group's ->cpumask correctly,
7517 * and ->cpu_power to 0.
7519 * Assumes the sched_domain tree is fully constructed
7522 build_sched_groups(struct sched_domain *sd, int cpu)
7524 struct sched_group *first = NULL, *last = NULL;
7525 struct sd_data *sdd = sd->private;
7526 const struct cpumask *span = sched_domain_span(sd);
7527 struct cpumask *covered;
7530 get_group(cpu, sdd, &sd->groups);
7531 atomic_inc(&sd->groups->ref);
7533 if (cpu != cpumask_first(sched_domain_span(sd)))
7536 lockdep_assert_held(&sched_domains_mutex);
7537 covered = sched_domains_tmpmask;
7539 cpumask_clear(covered);
7541 for_each_cpu(i, span) {
7542 struct sched_group *sg;
7543 int group = get_group(i, sdd, &sg);
7546 if (cpumask_test_cpu(i, covered))
7549 cpumask_clear(sched_group_cpus(sg));
7552 for_each_cpu(j, span) {
7553 if (get_group(j, sdd, NULL) != group)
7556 cpumask_set_cpu(j, covered);
7557 cpumask_set_cpu(j, sched_group_cpus(sg));
7572 * Initialize sched groups cpu_power.
7574 * cpu_power indicates the capacity of sched group, which is used while
7575 * distributing the load between different sched groups in a sched domain.
7576 * Typically cpu_power for all the groups in a sched domain will be same unless
7577 * there are asymmetries in the topology. If there are asymmetries, group
7578 * having more cpu_power will pickup more load compared to the group having
7581 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7583 struct sched_group *sg = sd->groups;
7585 WARN_ON(!sd || !sg);
7588 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7590 } while (sg != sd->groups);
7592 if (cpu != group_first_cpu(sg))
7595 update_group_power(sd, cpu);
7599 * Initializers for schedule domains
7600 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7603 #ifdef CONFIG_SCHED_DEBUG
7604 # define SD_INIT_NAME(sd, type) sd->name = #type
7606 # define SD_INIT_NAME(sd, type) do { } while (0)
7609 #define SD_INIT_FUNC(type) \
7610 static noinline struct sched_domain * \
7611 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7613 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7614 *sd = SD_##type##_INIT; \
7615 SD_INIT_NAME(sd, type); \
7616 sd->private = &tl->data; \
7622 SD_INIT_FUNC(ALLNODES)
7625 #ifdef CONFIG_SCHED_SMT
7626 SD_INIT_FUNC(SIBLING)
7628 #ifdef CONFIG_SCHED_MC
7631 #ifdef CONFIG_SCHED_BOOK
7635 static int default_relax_domain_level = -1;
7636 int sched_domain_level_max;
7638 static int __init setup_relax_domain_level(char *str)
7640 if (kstrtoint(str, 0, &default_relax_domain_level))
7641 pr_warn("Unable to set relax_domain_level\n");
7645 __setup("relax_domain_level=", setup_relax_domain_level);
7647 static void set_domain_attribute(struct sched_domain *sd,
7648 struct sched_domain_attr *attr)
7652 if (!attr || attr->relax_domain_level < 0) {
7653 if (default_relax_domain_level < 0)
7656 request = default_relax_domain_level;
7658 request = attr->relax_domain_level;
7659 if (request < sd->level) {
7660 /* turn off idle balance on this domain */
7661 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7663 /* turn on idle balance on this domain */
7664 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7668 static void __sdt_free(const struct cpumask *cpu_map);
7669 static int __sdt_alloc(const struct cpumask *cpu_map);
7671 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7672 const struct cpumask *cpu_map)
7676 if (!atomic_read(&d->rd->refcount))
7677 free_rootdomain(&d->rd->rcu); /* fall through */
7679 free_percpu(d->sd); /* fall through */
7681 __sdt_free(cpu_map); /* fall through */
7687 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7688 const struct cpumask *cpu_map)
7690 memset(d, 0, sizeof(*d));
7692 if (__sdt_alloc(cpu_map))
7693 return sa_sd_storage;
7694 d->sd = alloc_percpu(struct sched_domain *);
7696 return sa_sd_storage;
7697 d->rd = alloc_rootdomain();
7700 return sa_rootdomain;
7704 * NULL the sd_data elements we've used to build the sched_domain and
7705 * sched_group structure so that the subsequent __free_domain_allocs()
7706 * will not free the data we're using.
7708 static void claim_allocations(int cpu, struct sched_domain *sd)
7710 struct sd_data *sdd = sd->private;
7712 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7713 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7715 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7716 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7718 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7719 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7722 #ifdef CONFIG_SCHED_SMT
7723 static const struct cpumask *cpu_smt_mask(int cpu)
7725 return topology_thread_cpumask(cpu);
7730 * Topology list, bottom-up.
7732 static struct sched_domain_topology_level default_topology[] = {
7733 #ifdef CONFIG_SCHED_SMT
7734 { sd_init_SIBLING, cpu_smt_mask, },
7736 #ifdef CONFIG_SCHED_MC
7737 { sd_init_MC, cpu_coregroup_mask, },
7739 #ifdef CONFIG_SCHED_BOOK
7740 { sd_init_BOOK, cpu_book_mask, },
7742 { sd_init_CPU, cpu_cpu_mask, },
7744 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7745 { sd_init_ALLNODES, cpu_allnodes_mask, },
7750 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7752 static int __sdt_alloc(const struct cpumask *cpu_map)
7754 struct sched_domain_topology_level *tl;
7757 for (tl = sched_domain_topology; tl->init; tl++) {
7758 struct sd_data *sdd = &tl->data;
7760 sdd->sd = alloc_percpu(struct sched_domain *);
7764 sdd->sg = alloc_percpu(struct sched_group *);
7768 sdd->sgp = alloc_percpu(struct sched_group_power *);
7772 for_each_cpu(j, cpu_map) {
7773 struct sched_domain *sd;
7774 struct sched_group *sg;
7775 struct sched_group_power *sgp;
7777 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7778 GFP_KERNEL, cpu_to_node(j));
7782 *per_cpu_ptr(sdd->sd, j) = sd;
7784 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7785 GFP_KERNEL, cpu_to_node(j));
7789 *per_cpu_ptr(sdd->sg, j) = sg;
7791 sgp = kzalloc_node(sizeof(struct sched_group_power),
7792 GFP_KERNEL, cpu_to_node(j));
7796 *per_cpu_ptr(sdd->sgp, j) = sgp;
7803 static void __sdt_free(const struct cpumask *cpu_map)
7805 struct sched_domain_topology_level *tl;
7808 for (tl = sched_domain_topology; tl->init; tl++) {
7809 struct sd_data *sdd = &tl->data;
7811 for_each_cpu(j, cpu_map) {
7812 struct sched_domain *sd;
7815 sd = *per_cpu_ptr(sdd->sd, j);
7816 if (sd && (sd->flags & SD_OVERLAP))
7817 free_sched_groups(sd->groups, 0);
7818 kfree(*per_cpu_ptr(sdd->sd, j));
7822 kfree(*per_cpu_ptr(sdd->sg, j));
7824 kfree(*per_cpu_ptr(sdd->sgp, j));
7826 free_percpu(sdd->sd);
7828 free_percpu(sdd->sg);
7830 free_percpu(sdd->sgp);
7835 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7836 struct s_data *d, const struct cpumask *cpu_map,
7837 struct sched_domain_attr *attr, struct sched_domain *child,
7840 struct sched_domain *sd = tl->init(tl, cpu);
7844 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7846 sd->level = child->level + 1;
7847 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7851 set_domain_attribute(sd, attr);
7857 * Build sched domains for a given set of cpus and attach the sched domains
7858 * to the individual cpus
7860 static int build_sched_domains(const struct cpumask *cpu_map,
7861 struct sched_domain_attr *attr)
7863 enum s_alloc alloc_state = sa_none;
7864 struct sched_domain *sd;
7866 int i, ret = -ENOMEM;
7868 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7869 if (alloc_state != sa_rootdomain)
7872 /* Set up domains for cpus specified by the cpu_map. */
7873 for_each_cpu(i, cpu_map) {
7874 struct sched_domain_topology_level *tl;
7877 for (tl = sched_domain_topology; tl->init; tl++) {
7878 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7879 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7880 sd->flags |= SD_OVERLAP;
7881 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7888 *per_cpu_ptr(d.sd, i) = sd;
7891 /* Build the groups for the domains */
7892 for_each_cpu(i, cpu_map) {
7893 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7894 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7895 if (sd->flags & SD_OVERLAP) {
7896 if (build_overlap_sched_groups(sd, i))
7899 if (build_sched_groups(sd, i))
7905 /* Calculate CPU power for physical packages and nodes */
7906 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7907 if (!cpumask_test_cpu(i, cpu_map))
7910 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7911 claim_allocations(i, sd);
7912 init_sched_groups_power(i, sd);
7916 /* Attach the domains */
7918 for_each_cpu(i, cpu_map) {
7919 sd = *per_cpu_ptr(d.sd, i);
7920 cpu_attach_domain(sd, d.rd, i);
7926 __free_domain_allocs(&d, alloc_state, cpu_map);
7930 static cpumask_var_t *doms_cur; /* current sched domains */
7931 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7932 static struct sched_domain_attr *dattr_cur;
7933 /* attribues of custom domains in 'doms_cur' */
7936 * Special case: If a kmalloc of a doms_cur partition (array of
7937 * cpumask) fails, then fallback to a single sched domain,
7938 * as determined by the single cpumask fallback_doms.
7940 static cpumask_var_t fallback_doms;
7943 * arch_update_cpu_topology lets virtualized architectures update the
7944 * cpu core maps. It is supposed to return 1 if the topology changed
7945 * or 0 if it stayed the same.
7947 int __attribute__((weak)) arch_update_cpu_topology(void)
7952 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7955 cpumask_var_t *doms;
7957 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7960 for (i = 0; i < ndoms; i++) {
7961 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7962 free_sched_domains(doms, i);
7969 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7972 for (i = 0; i < ndoms; i++)
7973 free_cpumask_var(doms[i]);
7978 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7979 * For now this just excludes isolated cpus, but could be used to
7980 * exclude other special cases in the future.
7982 static int init_sched_domains(const struct cpumask *cpu_map)
7986 arch_update_cpu_topology();
7988 doms_cur = alloc_sched_domains(ndoms_cur);
7990 doms_cur = &fallback_doms;
7991 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7993 err = build_sched_domains(doms_cur[0], NULL);
7994 register_sched_domain_sysctl();
8000 * Detach sched domains from a group of cpus specified in cpu_map
8001 * These cpus will now be attached to the NULL domain
8003 static void detach_destroy_domains(const struct cpumask *cpu_map)
8008 for_each_cpu(i, cpu_map)
8009 cpu_attach_domain(NULL, &def_root_domain, i);
8013 /* handle null as "default" */
8014 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8015 struct sched_domain_attr *new, int idx_new)
8017 struct sched_domain_attr tmp;
8024 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8025 new ? (new + idx_new) : &tmp,
8026 sizeof(struct sched_domain_attr));
8030 * Partition sched domains as specified by the 'ndoms_new'
8031 * cpumasks in the array doms_new[] of cpumasks. This compares
8032 * doms_new[] to the current sched domain partitioning, doms_cur[].
8033 * It destroys each deleted domain and builds each new domain.
8035 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
8036 * The masks don't intersect (don't overlap.) We should setup one
8037 * sched domain for each mask. CPUs not in any of the cpumasks will
8038 * not be load balanced. If the same cpumask appears both in the
8039 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8042 * The passed in 'doms_new' should be allocated using
8043 * alloc_sched_domains. This routine takes ownership of it and will
8044 * free_sched_domains it when done with it. If the caller failed the
8045 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
8046 * and partition_sched_domains() will fallback to the single partition
8047 * 'fallback_doms', it also forces the domains to be rebuilt.
8049 * If doms_new == NULL it will be replaced with cpu_online_mask.
8050 * ndoms_new == 0 is a special case for destroying existing domains,
8051 * and it will not create the default domain.
8053 * Call with hotplug lock held
8055 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
8056 struct sched_domain_attr *dattr_new)
8061 mutex_lock(&sched_domains_mutex);
8063 /* always unregister in case we don't destroy any domains */
8064 unregister_sched_domain_sysctl();
8066 /* Let architecture update cpu core mappings. */
8067 new_topology = arch_update_cpu_topology();
8069 n = doms_new ? ndoms_new : 0;
8071 /* Destroy deleted domains */
8072 for (i = 0; i < ndoms_cur; i++) {
8073 for (j = 0; j < n && !new_topology; j++) {
8074 if (cpumask_equal(doms_cur[i], doms_new[j])
8075 && dattrs_equal(dattr_cur, i, dattr_new, j))
8078 /* no match - a current sched domain not in new doms_new[] */
8079 detach_destroy_domains(doms_cur[i]);
8084 if (doms_new == NULL) {
8086 doms_new = &fallback_doms;
8087 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
8088 WARN_ON_ONCE(dattr_new);
8091 /* Build new domains */
8092 for (i = 0; i < ndoms_new; i++) {
8093 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8094 if (cpumask_equal(doms_new[i], doms_cur[j])
8095 && dattrs_equal(dattr_new, i, dattr_cur, j))
8098 /* no match - add a new doms_new */
8099 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
8104 /* Remember the new sched domains */
8105 if (doms_cur != &fallback_doms)
8106 free_sched_domains(doms_cur, ndoms_cur);
8107 kfree(dattr_cur); /* kfree(NULL) is safe */
8108 doms_cur = doms_new;
8109 dattr_cur = dattr_new;
8110 ndoms_cur = ndoms_new;
8112 register_sched_domain_sysctl();
8114 mutex_unlock(&sched_domains_mutex);
8117 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8118 static void reinit_sched_domains(void)
8122 /* Destroy domains first to force the rebuild */
8123 partition_sched_domains(0, NULL, NULL);
8125 rebuild_sched_domains();
8129 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8131 unsigned int level = 0;
8133 if (sscanf(buf, "%u", &level) != 1)
8137 * level is always be positive so don't check for
8138 * level < POWERSAVINGS_BALANCE_NONE which is 0
8139 * What happens on 0 or 1 byte write,
8140 * need to check for count as well?
8143 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8147 sched_smt_power_savings = level;
8149 sched_mc_power_savings = level;
8151 reinit_sched_domains();
8156 #ifdef CONFIG_SCHED_MC
8157 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8158 struct sysdev_class_attribute *attr,
8161 return sprintf(page, "%u\n", sched_mc_power_savings);
8163 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8164 struct sysdev_class_attribute *attr,
8165 const char *buf, size_t count)
8167 return sched_power_savings_store(buf, count, 0);
8169 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8170 sched_mc_power_savings_show,
8171 sched_mc_power_savings_store);
8174 #ifdef CONFIG_SCHED_SMT
8175 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8176 struct sysdev_class_attribute *attr,
8179 return sprintf(page, "%u\n", sched_smt_power_savings);
8181 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8182 struct sysdev_class_attribute *attr,
8183 const char *buf, size_t count)
8185 return sched_power_savings_store(buf, count, 1);
8187 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8188 sched_smt_power_savings_show,
8189 sched_smt_power_savings_store);
8192 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8196 #ifdef CONFIG_SCHED_SMT
8198 err = sysfs_create_file(&cls->kset.kobj,
8199 &attr_sched_smt_power_savings.attr);
8201 #ifdef CONFIG_SCHED_MC
8202 if (!err && mc_capable())
8203 err = sysfs_create_file(&cls->kset.kobj,
8204 &attr_sched_mc_power_savings.attr);
8208 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8210 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
8213 * Update cpusets according to cpu_active mask. If cpusets are
8214 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8215 * around partition_sched_domains().
8217 * If we come here as part of a suspend/resume, don't touch cpusets because we
8218 * want to restore it back to its original state upon resume anyway.
8220 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
8224 case CPU_ONLINE_FROZEN:
8225 case CPU_DOWN_FAILED_FROZEN:
8228 * num_cpus_frozen tracks how many CPUs are involved in suspend
8229 * resume sequence. As long as this is not the last online
8230 * operation in the resume sequence, just build a single sched
8231 * domain, ignoring cpusets.
8234 if (likely(num_cpus_frozen)) {
8235 partition_sched_domains(1, NULL, NULL);
8240 * This is the last CPU online operation. So fall through and
8241 * restore the original sched domains by considering the
8242 * cpuset configurations.
8246 case CPU_DOWN_FAILED:
8247 cpuset_update_active_cpus();
8255 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
8259 case CPU_DOWN_PREPARE:
8260 cpuset_update_active_cpus();
8262 case CPU_DOWN_PREPARE_FROZEN:
8264 partition_sched_domains(1, NULL, NULL);
8272 static int update_runtime(struct notifier_block *nfb,
8273 unsigned long action, void *hcpu)
8275 int cpu = (int)(long)hcpu;
8278 case CPU_DOWN_PREPARE:
8279 case CPU_DOWN_PREPARE_FROZEN:
8280 disable_runtime(cpu_rq(cpu));
8283 case CPU_DOWN_FAILED:
8284 case CPU_DOWN_FAILED_FROZEN:
8286 case CPU_ONLINE_FROZEN:
8287 enable_runtime(cpu_rq(cpu));
8295 void __init sched_init_smp(void)
8297 cpumask_var_t non_isolated_cpus;
8299 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8300 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8303 mutex_lock(&sched_domains_mutex);
8304 init_sched_domains(cpu_active_mask);
8305 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8306 if (cpumask_empty(non_isolated_cpus))
8307 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8308 mutex_unlock(&sched_domains_mutex);
8311 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
8312 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
8314 /* RT runtime code needs to handle some hotplug events */
8315 hotcpu_notifier(update_runtime, 0);
8319 /* Move init over to a non-isolated CPU */
8320 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8322 sched_init_granularity();
8323 free_cpumask_var(non_isolated_cpus);
8325 init_sched_rt_class();
8328 void __init sched_init_smp(void)
8330 sched_init_granularity();
8332 #endif /* CONFIG_SMP */
8334 const_debug unsigned int sysctl_timer_migration = 1;
8336 int in_sched_functions(unsigned long addr)
8338 return in_lock_functions(addr) ||
8339 (addr >= (unsigned long)__sched_text_start
8340 && addr < (unsigned long)__sched_text_end);
8343 static void init_cfs_rq(struct cfs_rq *cfs_rq)
8345 cfs_rq->tasks_timeline = RB_ROOT;
8346 INIT_LIST_HEAD(&cfs_rq->tasks);
8347 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8348 #ifndef CONFIG_64BIT
8349 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8353 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8355 struct rt_prio_array *array;
8358 array = &rt_rq->active;
8359 for (i = 0; i < MAX_RT_PRIO; i++) {
8360 INIT_LIST_HEAD(array->queue + i);
8361 __clear_bit(i, array->bitmap);
8363 /* delimiter for bitsearch: */
8364 __set_bit(MAX_RT_PRIO, array->bitmap);
8366 #if defined CONFIG_SMP
8367 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8368 rt_rq->highest_prio.next = MAX_RT_PRIO;
8369 rt_rq->rt_nr_migratory = 0;
8370 rt_rq->overloaded = 0;
8371 plist_head_init(&rt_rq->pushable_tasks);
8375 rt_rq->rt_throttled = 0;
8376 rt_rq->rt_runtime = 0;
8377 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8380 #ifdef CONFIG_FAIR_GROUP_SCHED
8381 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8382 struct sched_entity *se, int cpu,
8383 struct sched_entity *parent)
8385 struct rq *rq = cpu_rq(cpu);
8390 /* allow initial update_cfs_load() to truncate */
8391 cfs_rq->load_stamp = 1;
8393 init_cfs_rq_runtime(cfs_rq);
8395 tg->cfs_rq[cpu] = cfs_rq;
8398 /* se could be NULL for root_task_group */
8403 se->cfs_rq = &rq->cfs;
8405 se->cfs_rq = parent->my_q;
8408 update_load_set(&se->load, 0);
8409 se->parent = parent;
8413 #ifdef CONFIG_RT_GROUP_SCHED
8414 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8415 struct sched_rt_entity *rt_se, int cpu,
8416 struct sched_rt_entity *parent)
8418 struct rq *rq = cpu_rq(cpu);
8420 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8421 rt_rq->rt_nr_boosted = 0;
8425 tg->rt_rq[cpu] = rt_rq;
8426 tg->rt_se[cpu] = rt_se;
8432 rt_se->rt_rq = &rq->rt;
8434 rt_se->rt_rq = parent->my_q;
8436 rt_se->my_q = rt_rq;
8437 rt_se->parent = parent;
8438 INIT_LIST_HEAD(&rt_se->run_list);
8442 void __init sched_init(void)
8445 unsigned long alloc_size = 0, ptr;
8447 #ifdef CONFIG_FAIR_GROUP_SCHED
8448 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8450 #ifdef CONFIG_RT_GROUP_SCHED
8451 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8453 #ifdef CONFIG_CPUMASK_OFFSTACK
8454 alloc_size += num_possible_cpus() * cpumask_size();
8457 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8459 #ifdef CONFIG_FAIR_GROUP_SCHED
8460 root_task_group.se = (struct sched_entity **)ptr;
8461 ptr += nr_cpu_ids * sizeof(void **);
8463 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8464 ptr += nr_cpu_ids * sizeof(void **);
8466 #endif /* CONFIG_FAIR_GROUP_SCHED */
8467 #ifdef CONFIG_RT_GROUP_SCHED
8468 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8469 ptr += nr_cpu_ids * sizeof(void **);
8471 root_task_group.rt_rq = (struct rt_rq **)ptr;
8472 ptr += nr_cpu_ids * sizeof(void **);
8474 #endif /* CONFIG_RT_GROUP_SCHED */
8475 #ifdef CONFIG_CPUMASK_OFFSTACK
8476 for_each_possible_cpu(i) {
8477 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8478 ptr += cpumask_size();
8480 #endif /* CONFIG_CPUMASK_OFFSTACK */
8484 init_defrootdomain();
8487 init_rt_bandwidth(&def_rt_bandwidth,
8488 global_rt_period(), global_rt_runtime());
8490 #ifdef CONFIG_RT_GROUP_SCHED
8491 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8492 global_rt_period(), global_rt_runtime());
8493 #endif /* CONFIG_RT_GROUP_SCHED */
8495 #ifdef CONFIG_CGROUP_SCHED
8496 list_add(&root_task_group.list, &task_groups);
8497 INIT_LIST_HEAD(&root_task_group.children);
8498 autogroup_init(&init_task);
8499 #endif /* CONFIG_CGROUP_SCHED */
8501 for_each_possible_cpu(i) {
8505 raw_spin_lock_init(&rq->lock);
8507 rq->calc_load_active = 0;
8508 rq->calc_load_update = jiffies + LOAD_FREQ;
8509 init_cfs_rq(&rq->cfs);
8510 init_rt_rq(&rq->rt, rq);
8511 #ifdef CONFIG_FAIR_GROUP_SCHED
8512 root_task_group.shares = root_task_group_load;
8513 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8515 * How much cpu bandwidth does root_task_group get?
8517 * In case of task-groups formed thr' the cgroup filesystem, it
8518 * gets 100% of the cpu resources in the system. This overall
8519 * system cpu resource is divided among the tasks of
8520 * root_task_group and its child task-groups in a fair manner,
8521 * based on each entity's (task or task-group's) weight
8522 * (se->load.weight).
8524 * In other words, if root_task_group has 10 tasks of weight
8525 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8526 * then A0's share of the cpu resource is:
8528 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8530 * We achieve this by letting root_task_group's tasks sit
8531 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8533 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8534 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8535 #endif /* CONFIG_FAIR_GROUP_SCHED */
8537 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8538 #ifdef CONFIG_RT_GROUP_SCHED
8539 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8540 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8543 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8544 rq->cpu_load[j] = 0;
8546 rq->last_load_update_tick = jiffies;
8551 rq->cpu_power = SCHED_POWER_SCALE;
8552 rq->post_schedule = 0;
8553 rq->active_balance = 0;
8554 rq->next_balance = jiffies;
8559 rq->avg_idle = 2*sysctl_sched_migration_cost;
8560 rq_attach_root(rq, &def_root_domain);
8562 rq->nohz_balance_kick = 0;
8566 atomic_set(&rq->nr_iowait, 0);
8569 set_load_weight(&init_task);
8571 #ifdef CONFIG_PREEMPT_NOTIFIERS
8572 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8576 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8579 #ifdef CONFIG_RT_MUTEXES
8580 plist_head_init(&init_task.pi_waiters);
8584 * The boot idle thread does lazy MMU switching as well:
8586 atomic_inc(&init_mm.mm_count);
8587 enter_lazy_tlb(&init_mm, current);
8590 * Make us the idle thread. Technically, schedule() should not be
8591 * called from this thread, however somewhere below it might be,
8592 * but because we are the idle thread, we just pick up running again
8593 * when this runqueue becomes "idle".
8595 init_idle(current, smp_processor_id());
8597 calc_load_update = jiffies + LOAD_FREQ;
8600 * During early bootup we pretend to be a normal task:
8602 current->sched_class = &fair_sched_class;
8605 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8607 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8608 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8609 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8610 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8611 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8613 /* May be allocated at isolcpus cmdline parse time */
8614 if (cpu_isolated_map == NULL)
8615 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8618 scheduler_running = 1;
8621 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8622 static inline int preempt_count_equals(int preempt_offset)
8624 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8626 return (nested == preempt_offset);
8629 void __might_sleep(const char *file, int line, int preempt_offset)
8631 static unsigned long prev_jiffy; /* ratelimiting */
8633 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
8634 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8635 system_state != SYSTEM_RUNNING || oops_in_progress)
8637 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8639 prev_jiffy = jiffies;
8642 "BUG: sleeping function called from invalid context at %s:%d\n",
8645 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8646 in_atomic(), irqs_disabled(),
8647 current->pid, current->comm);
8649 debug_show_held_locks(current);
8650 if (irqs_disabled())
8651 print_irqtrace_events(current);
8654 EXPORT_SYMBOL(__might_sleep);
8657 #ifdef CONFIG_MAGIC_SYSRQ
8658 static void normalize_task(struct rq *rq, struct task_struct *p)
8660 const struct sched_class *prev_class = p->sched_class;
8661 int old_prio = p->prio;
8666 deactivate_task(rq, p, 0);
8667 __setscheduler(rq, p, SCHED_NORMAL, 0);
8669 activate_task(rq, p, 0);
8670 resched_task(rq->curr);
8673 check_class_changed(rq, p, prev_class, old_prio);
8676 void normalize_rt_tasks(void)
8678 struct task_struct *g, *p;
8679 unsigned long flags;
8682 read_lock_irqsave(&tasklist_lock, flags);
8683 do_each_thread(g, p) {
8685 * Only normalize user tasks:
8690 p->se.exec_start = 0;
8691 #ifdef CONFIG_SCHEDSTATS
8692 p->se.statistics.wait_start = 0;
8693 p->se.statistics.sleep_start = 0;
8694 p->se.statistics.block_start = 0;
8699 * Renice negative nice level userspace
8702 if (TASK_NICE(p) < 0 && p->mm)
8703 set_user_nice(p, 0);
8707 raw_spin_lock(&p->pi_lock);
8708 rq = __task_rq_lock(p);
8710 normalize_task(rq, p);
8712 __task_rq_unlock(rq);
8713 raw_spin_unlock(&p->pi_lock);
8714 } while_each_thread(g, p);
8716 read_unlock_irqrestore(&tasklist_lock, flags);
8719 #endif /* CONFIG_MAGIC_SYSRQ */
8721 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8723 * These functions are only useful for the IA64 MCA handling, or kdb.
8725 * They can only be called when the whole system has been
8726 * stopped - every CPU needs to be quiescent, and no scheduling
8727 * activity can take place. Using them for anything else would
8728 * be a serious bug, and as a result, they aren't even visible
8729 * under any other configuration.
8733 * curr_task - return the current task for a given cpu.
8734 * @cpu: the processor in question.
8736 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8738 struct task_struct *curr_task(int cpu)
8740 return cpu_curr(cpu);
8743 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8747 * set_curr_task - set the current task for a given cpu.
8748 * @cpu: the processor in question.
8749 * @p: the task pointer to set.
8751 * Description: This function must only be used when non-maskable interrupts
8752 * are serviced on a separate stack. It allows the architecture to switch the
8753 * notion of the current task on a cpu in a non-blocking manner. This function
8754 * must be called with all CPU's synchronized, and interrupts disabled, the
8755 * and caller must save the original value of the current task (see
8756 * curr_task() above) and restore that value before reenabling interrupts and
8757 * re-starting the system.
8759 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8761 void set_curr_task(int cpu, struct task_struct *p)
8768 #ifdef CONFIG_FAIR_GROUP_SCHED
8769 static void free_fair_sched_group(struct task_group *tg)
8773 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8775 for_each_possible_cpu(i) {
8777 kfree(tg->cfs_rq[i]);
8787 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8789 struct cfs_rq *cfs_rq;
8790 struct sched_entity *se;
8793 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8796 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8800 tg->shares = NICE_0_LOAD;
8802 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8804 for_each_possible_cpu(i) {
8805 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8806 GFP_KERNEL, cpu_to_node(i));
8810 se = kzalloc_node(sizeof(struct sched_entity),
8811 GFP_KERNEL, cpu_to_node(i));
8815 init_cfs_rq(cfs_rq);
8816 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8827 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8829 struct rq *rq = cpu_rq(cpu);
8830 unsigned long flags;
8833 * Only empty task groups can be destroyed; so we can speculatively
8834 * check on_list without danger of it being re-added.
8836 if (!tg->cfs_rq[cpu]->on_list)
8839 raw_spin_lock_irqsave(&rq->lock, flags);
8840 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8841 raw_spin_unlock_irqrestore(&rq->lock, flags);
8843 #else /* !CONFIG_FAIR_GROUP_SCHED */
8844 static inline void free_fair_sched_group(struct task_group *tg)
8849 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8854 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8857 #endif /* CONFIG_FAIR_GROUP_SCHED */
8859 #ifdef CONFIG_RT_GROUP_SCHED
8860 static void free_rt_sched_group(struct task_group *tg)
8865 destroy_rt_bandwidth(&tg->rt_bandwidth);
8867 for_each_possible_cpu(i) {
8869 kfree(tg->rt_rq[i]);
8871 kfree(tg->rt_se[i]);
8879 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8881 struct rt_rq *rt_rq;
8882 struct sched_rt_entity *rt_se;
8885 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8888 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8892 init_rt_bandwidth(&tg->rt_bandwidth,
8893 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8895 for_each_possible_cpu(i) {
8896 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8897 GFP_KERNEL, cpu_to_node(i));
8901 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8902 GFP_KERNEL, cpu_to_node(i));
8906 init_rt_rq(rt_rq, cpu_rq(i));
8907 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8908 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8918 #else /* !CONFIG_RT_GROUP_SCHED */
8919 static inline void free_rt_sched_group(struct task_group *tg)
8924 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8928 #endif /* CONFIG_RT_GROUP_SCHED */
8930 #ifdef CONFIG_CGROUP_SCHED
8931 static void free_sched_group(struct task_group *tg)
8933 free_fair_sched_group(tg);
8934 free_rt_sched_group(tg);
8939 /* allocate runqueue etc for a new task group */
8940 struct task_group *sched_create_group(struct task_group *parent)
8942 struct task_group *tg;
8943 unsigned long flags;
8945 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8947 return ERR_PTR(-ENOMEM);
8949 if (!alloc_fair_sched_group(tg, parent))
8952 if (!alloc_rt_sched_group(tg, parent))
8955 spin_lock_irqsave(&task_group_lock, flags);
8956 list_add_rcu(&tg->list, &task_groups);
8958 WARN_ON(!parent); /* root should already exist */
8960 tg->parent = parent;
8961 INIT_LIST_HEAD(&tg->children);
8962 list_add_rcu(&tg->siblings, &parent->children);
8963 spin_unlock_irqrestore(&task_group_lock, flags);
8968 free_sched_group(tg);
8969 return ERR_PTR(-ENOMEM);
8972 /* rcu callback to free various structures associated with a task group */
8973 static void free_sched_group_rcu(struct rcu_head *rhp)
8975 /* now it should be safe to free those cfs_rqs */
8976 free_sched_group(container_of(rhp, struct task_group, rcu));
8979 /* Destroy runqueue etc associated with a task group */
8980 void sched_destroy_group(struct task_group *tg)
8982 unsigned long flags;
8985 /* end participation in shares distribution */
8986 for_each_possible_cpu(i)
8987 unregister_fair_sched_group(tg, i);
8989 spin_lock_irqsave(&task_group_lock, flags);
8990 list_del_rcu(&tg->list);
8991 list_del_rcu(&tg->siblings);
8992 spin_unlock_irqrestore(&task_group_lock, flags);
8994 /* wait for possible concurrent references to cfs_rqs complete */
8995 call_rcu(&tg->rcu, free_sched_group_rcu);
8998 /* change task's runqueue when it moves between groups.
8999 * The caller of this function should have put the task in its new group
9000 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9001 * reflect its new group.
9003 void sched_move_task(struct task_struct *tsk)
9005 struct task_group *tg;
9007 unsigned long flags;
9010 rq = task_rq_lock(tsk, &flags);
9012 running = task_current(rq, tsk);
9016 dequeue_task(rq, tsk, 0);
9017 if (unlikely(running))
9018 tsk->sched_class->put_prev_task(rq, tsk);
9020 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
9021 lockdep_is_held(&tsk->sighand->siglock)),
9022 struct task_group, css);
9023 tg = autogroup_task_group(tsk, tg);
9024 tsk->sched_task_group = tg;
9026 #ifdef CONFIG_FAIR_GROUP_SCHED
9027 if (tsk->sched_class->task_move_group)
9028 tsk->sched_class->task_move_group(tsk, on_rq);
9031 set_task_rq(tsk, task_cpu(tsk));
9033 if (unlikely(running))
9034 tsk->sched_class->set_curr_task(rq);
9036 enqueue_task(rq, tsk, 0);
9038 task_rq_unlock(rq, tsk, &flags);
9040 #endif /* CONFIG_CGROUP_SCHED */
9042 #ifdef CONFIG_FAIR_GROUP_SCHED
9043 static DEFINE_MUTEX(shares_mutex);
9045 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9048 unsigned long flags;
9051 * We can't change the weight of the root cgroup.
9056 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9058 mutex_lock(&shares_mutex);
9059 if (tg->shares == shares)
9062 tg->shares = shares;
9063 for_each_possible_cpu(i) {
9064 struct rq *rq = cpu_rq(i);
9065 struct sched_entity *se;
9068 /* Propagate contribution to hierarchy */
9069 raw_spin_lock_irqsave(&rq->lock, flags);
9070 for_each_sched_entity(se)
9071 update_cfs_shares(group_cfs_rq(se));
9072 raw_spin_unlock_irqrestore(&rq->lock, flags);
9076 mutex_unlock(&shares_mutex);
9080 unsigned long sched_group_shares(struct task_group *tg)
9086 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
9087 static unsigned long to_ratio(u64 period, u64 runtime)
9089 if (runtime == RUNTIME_INF)
9092 return div64_u64(runtime << 20, period);
9096 #ifdef CONFIG_RT_GROUP_SCHED
9098 * Ensure that the real time constraints are schedulable.
9100 static DEFINE_MUTEX(rt_constraints_mutex);
9102 /* Must be called with tasklist_lock held */
9103 static inline int tg_has_rt_tasks(struct task_group *tg)
9105 struct task_struct *g, *p;
9107 do_each_thread(g, p) {
9108 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9110 } while_each_thread(g, p);
9115 struct rt_schedulable_data {
9116 struct task_group *tg;
9121 static int tg_rt_schedulable(struct task_group *tg, void *data)
9123 struct rt_schedulable_data *d = data;
9124 struct task_group *child;
9125 unsigned long total, sum = 0;
9126 u64 period, runtime;
9128 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9129 runtime = tg->rt_bandwidth.rt_runtime;
9132 period = d->rt_period;
9133 runtime = d->rt_runtime;
9137 * Cannot have more runtime than the period.
9139 if (runtime > period && runtime != RUNTIME_INF)
9143 * Ensure we don't starve existing RT tasks.
9145 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9148 total = to_ratio(period, runtime);
9151 * Nobody can have more than the global setting allows.
9153 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9157 * The sum of our children's runtime should not exceed our own.
9159 list_for_each_entry_rcu(child, &tg->children, siblings) {
9160 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9161 runtime = child->rt_bandwidth.rt_runtime;
9163 if (child == d->tg) {
9164 period = d->rt_period;
9165 runtime = d->rt_runtime;
9168 sum += to_ratio(period, runtime);
9177 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9181 struct rt_schedulable_data data = {
9183 .rt_period = period,
9184 .rt_runtime = runtime,
9188 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
9194 static int tg_set_rt_bandwidth(struct task_group *tg,
9195 u64 rt_period, u64 rt_runtime)
9199 mutex_lock(&rt_constraints_mutex);
9200 read_lock(&tasklist_lock);
9201 err = __rt_schedulable(tg, rt_period, rt_runtime);
9205 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9206 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9207 tg->rt_bandwidth.rt_runtime = rt_runtime;
9209 for_each_possible_cpu(i) {
9210 struct rt_rq *rt_rq = tg->rt_rq[i];
9212 raw_spin_lock(&rt_rq->rt_runtime_lock);
9213 rt_rq->rt_runtime = rt_runtime;
9214 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9216 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9218 read_unlock(&tasklist_lock);
9219 mutex_unlock(&rt_constraints_mutex);
9224 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9226 u64 rt_runtime, rt_period;
9228 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9229 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9230 if (rt_runtime_us < 0)
9231 rt_runtime = RUNTIME_INF;
9233 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9236 long sched_group_rt_runtime(struct task_group *tg)
9240 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9243 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9244 do_div(rt_runtime_us, NSEC_PER_USEC);
9245 return rt_runtime_us;
9248 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9250 u64 rt_runtime, rt_period;
9252 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9253 rt_runtime = tg->rt_bandwidth.rt_runtime;
9258 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9261 long sched_group_rt_period(struct task_group *tg)
9265 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9266 do_div(rt_period_us, NSEC_PER_USEC);
9267 return rt_period_us;
9270 static int sched_rt_global_constraints(void)
9272 u64 runtime, period;
9275 if (sysctl_sched_rt_period <= 0)
9278 runtime = global_rt_runtime();
9279 period = global_rt_period();
9282 * Sanity check on the sysctl variables.
9284 if (runtime > period && runtime != RUNTIME_INF)
9287 mutex_lock(&rt_constraints_mutex);
9288 read_lock(&tasklist_lock);
9289 ret = __rt_schedulable(NULL, 0, 0);
9290 read_unlock(&tasklist_lock);
9291 mutex_unlock(&rt_constraints_mutex);
9296 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9298 /* Don't accept realtime tasks when there is no way for them to run */
9299 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9305 #else /* !CONFIG_RT_GROUP_SCHED */
9306 static int sched_rt_global_constraints(void)
9308 unsigned long flags;
9311 if (sysctl_sched_rt_period <= 0)
9315 * There's always some RT tasks in the root group
9316 * -- migration, kstopmachine etc..
9318 if (sysctl_sched_rt_runtime == 0)
9321 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9322 for_each_possible_cpu(i) {
9323 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9325 raw_spin_lock(&rt_rq->rt_runtime_lock);
9326 rt_rq->rt_runtime = global_rt_runtime();
9327 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9329 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9333 #endif /* CONFIG_RT_GROUP_SCHED */
9335 int sched_rt_handler(struct ctl_table *table, int write,
9336 void __user *buffer, size_t *lenp,
9340 int old_period, old_runtime;
9341 static DEFINE_MUTEX(mutex);
9344 old_period = sysctl_sched_rt_period;
9345 old_runtime = sysctl_sched_rt_runtime;
9347 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9349 if (!ret && write) {
9350 ret = sched_rt_global_constraints();
9352 sysctl_sched_rt_period = old_period;
9353 sysctl_sched_rt_runtime = old_runtime;
9355 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9356 def_rt_bandwidth.rt_period =
9357 ns_to_ktime(global_rt_period());
9360 mutex_unlock(&mutex);
9365 #ifdef CONFIG_CGROUP_SCHED
9367 /* return corresponding task_group object of a cgroup */
9368 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9370 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9371 struct task_group, css);
9374 static struct cgroup_subsys_state *
9375 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9377 struct task_group *tg, *parent;
9379 if (!cgrp->parent) {
9380 /* This is early initialization for the top cgroup */
9381 return &root_task_group.css;
9384 parent = cgroup_tg(cgrp->parent);
9385 tg = sched_create_group(parent);
9387 return ERR_PTR(-ENOMEM);
9393 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9395 struct task_group *tg = cgroup_tg(cgrp);
9397 sched_destroy_group(tg);
9401 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9403 #ifdef CONFIG_RT_GROUP_SCHED
9404 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9407 /* We don't support RT-tasks being in separate groups */
9408 if (tsk->sched_class != &fair_sched_class)
9415 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9417 sched_move_task(tsk);
9421 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9422 struct cgroup *old_cgrp, struct task_struct *task)
9425 * cgroup_exit() is called in the copy_process() failure path.
9426 * Ignore this case since the task hasn't ran yet, this avoids
9427 * trying to poke a half freed task state from generic code.
9429 if (!(task->flags & PF_EXITING))
9432 sched_move_task(task);
9435 #ifdef CONFIG_FAIR_GROUP_SCHED
9436 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9439 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9442 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9444 struct task_group *tg = cgroup_tg(cgrp);
9446 return (u64) scale_load_down(tg->shares);
9449 #ifdef CONFIG_CFS_BANDWIDTH
9450 static DEFINE_MUTEX(cfs_constraints_mutex);
9452 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9453 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9455 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9457 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9459 int i, ret = 0, runtime_enabled;
9460 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9462 if (tg == &root_task_group)
9466 * Ensure we have at some amount of bandwidth every period. This is
9467 * to prevent reaching a state of large arrears when throttled via
9468 * entity_tick() resulting in prolonged exit starvation.
9470 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9474 * Likewise, bound things on the otherside by preventing insane quota
9475 * periods. This also allows us to normalize in computing quota
9478 if (period > max_cfs_quota_period)
9481 mutex_lock(&cfs_constraints_mutex);
9482 ret = __cfs_schedulable(tg, period, quota);
9486 runtime_enabled = quota != RUNTIME_INF;
9487 raw_spin_lock_irq(&cfs_b->lock);
9488 cfs_b->period = ns_to_ktime(period);
9489 cfs_b->quota = quota;
9491 __refill_cfs_bandwidth_runtime(cfs_b);
9492 /* restart the period timer (if active) to handle new period expiry */
9493 if (runtime_enabled && cfs_b->timer_active) {
9494 /* force a reprogram */
9495 cfs_b->timer_active = 0;
9496 __start_cfs_bandwidth(cfs_b);
9498 raw_spin_unlock_irq(&cfs_b->lock);
9500 for_each_possible_cpu(i) {
9501 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9502 struct rq *rq = rq_of(cfs_rq);
9504 raw_spin_lock_irq(&rq->lock);
9505 cfs_rq->runtime_enabled = runtime_enabled;
9506 cfs_rq->runtime_remaining = 0;
9508 if (cfs_rq_throttled(cfs_rq))
9509 unthrottle_cfs_rq(cfs_rq);
9510 raw_spin_unlock_irq(&rq->lock);
9513 mutex_unlock(&cfs_constraints_mutex);
9518 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9522 period = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9523 if (cfs_quota_us < 0)
9524 quota = RUNTIME_INF;
9526 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9528 return tg_set_cfs_bandwidth(tg, period, quota);
9531 long tg_get_cfs_quota(struct task_group *tg)
9535 if (tg_cfs_bandwidth(tg)->quota == RUNTIME_INF)
9538 quota_us = tg_cfs_bandwidth(tg)->quota;
9539 do_div(quota_us, NSEC_PER_USEC);
9544 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9548 period = (u64)cfs_period_us * NSEC_PER_USEC;
9549 quota = tg_cfs_bandwidth(tg)->quota;
9554 return tg_set_cfs_bandwidth(tg, period, quota);
9557 long tg_get_cfs_period(struct task_group *tg)
9561 cfs_period_us = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9562 do_div(cfs_period_us, NSEC_PER_USEC);
9564 return cfs_period_us;
9567 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
9569 return tg_get_cfs_quota(cgroup_tg(cgrp));
9572 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
9575 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
9578 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
9580 return tg_get_cfs_period(cgroup_tg(cgrp));
9583 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9586 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
9589 struct cfs_schedulable_data {
9590 struct task_group *tg;
9595 * normalize group quota/period to be quota/max_period
9596 * note: units are usecs
9598 static u64 normalize_cfs_quota(struct task_group *tg,
9599 struct cfs_schedulable_data *d)
9607 period = tg_get_cfs_period(tg);
9608 quota = tg_get_cfs_quota(tg);
9611 /* note: these should typically be equivalent */
9612 if (quota == RUNTIME_INF || quota == -1)
9615 return to_ratio(period, quota);
9618 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9620 struct cfs_schedulable_data *d = data;
9621 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9622 s64 quota = 0, parent_quota = -1;
9625 quota = RUNTIME_INF;
9627 struct cfs_bandwidth *parent_b = tg_cfs_bandwidth(tg->parent);
9629 quota = normalize_cfs_quota(tg, d);
9630 parent_quota = parent_b->hierarchal_quota;
9633 * ensure max(child_quota) <= parent_quota, inherit when no
9636 if (quota == RUNTIME_INF)
9637 quota = parent_quota;
9638 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9641 cfs_b->hierarchal_quota = quota;
9646 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9649 struct cfs_schedulable_data data = {
9655 if (quota != RUNTIME_INF) {
9656 do_div(data.period, NSEC_PER_USEC);
9657 do_div(data.quota, NSEC_PER_USEC);
9661 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9667 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
9668 struct cgroup_map_cb *cb)
9670 struct task_group *tg = cgroup_tg(cgrp);
9671 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9673 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
9674 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
9675 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
9679 #endif /* CONFIG_CFS_BANDWIDTH */
9680 #endif /* CONFIG_FAIR_GROUP_SCHED */
9682 #ifdef CONFIG_RT_GROUP_SCHED
9683 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9686 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9689 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9691 return sched_group_rt_runtime(cgroup_tg(cgrp));
9694 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9697 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9700 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9702 return sched_group_rt_period(cgroup_tg(cgrp));
9704 #endif /* CONFIG_RT_GROUP_SCHED */
9706 static struct cftype cpu_files[] = {
9707 #ifdef CONFIG_FAIR_GROUP_SCHED
9710 .read_u64 = cpu_shares_read_u64,
9711 .write_u64 = cpu_shares_write_u64,
9714 #ifdef CONFIG_CFS_BANDWIDTH
9716 .name = "cfs_quota_us",
9717 .read_s64 = cpu_cfs_quota_read_s64,
9718 .write_s64 = cpu_cfs_quota_write_s64,
9721 .name = "cfs_period_us",
9722 .read_u64 = cpu_cfs_period_read_u64,
9723 .write_u64 = cpu_cfs_period_write_u64,
9727 .read_map = cpu_stats_show,
9730 #ifdef CONFIG_RT_GROUP_SCHED
9732 .name = "rt_runtime_us",
9733 .read_s64 = cpu_rt_runtime_read,
9734 .write_s64 = cpu_rt_runtime_write,
9737 .name = "rt_period_us",
9738 .read_u64 = cpu_rt_period_read_uint,
9739 .write_u64 = cpu_rt_period_write_uint,
9744 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9746 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9749 struct cgroup_subsys cpu_cgroup_subsys = {
9751 .create = cpu_cgroup_create,
9752 .destroy = cpu_cgroup_destroy,
9753 .can_attach_task = cpu_cgroup_can_attach_task,
9754 .attach_task = cpu_cgroup_attach_task,
9755 .exit = cpu_cgroup_exit,
9756 .populate = cpu_cgroup_populate,
9757 .subsys_id = cpu_cgroup_subsys_id,
9761 #endif /* CONFIG_CGROUP_SCHED */
9763 #ifdef CONFIG_CGROUP_CPUACCT
9766 * CPU accounting code for task groups.
9768 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9769 * (balbir@in.ibm.com).
9772 /* track cpu usage of a group of tasks and its child groups */
9774 struct cgroup_subsys_state css;
9775 /* cpuusage holds pointer to a u64-type object on every cpu */
9776 u64 __percpu *cpuusage;
9777 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9778 struct cpuacct *parent;
9781 struct cgroup_subsys cpuacct_subsys;
9783 /* return cpu accounting group corresponding to this container */
9784 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9786 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9787 struct cpuacct, css);
9790 /* return cpu accounting group to which this task belongs */
9791 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9793 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9794 struct cpuacct, css);
9797 /* create a new cpu accounting group */
9798 static struct cgroup_subsys_state *cpuacct_create(
9799 struct cgroup_subsys *ss, struct cgroup *cgrp)
9801 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9807 ca->cpuusage = alloc_percpu(u64);
9811 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9812 if (percpu_counter_init(&ca->cpustat[i], 0))
9813 goto out_free_counters;
9816 ca->parent = cgroup_ca(cgrp->parent);
9822 percpu_counter_destroy(&ca->cpustat[i]);
9823 free_percpu(ca->cpuusage);
9827 return ERR_PTR(-ENOMEM);
9830 /* destroy an existing cpu accounting group */
9832 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9834 struct cpuacct *ca = cgroup_ca(cgrp);
9837 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9838 percpu_counter_destroy(&ca->cpustat[i]);
9839 free_percpu(ca->cpuusage);
9843 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9845 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9848 #ifndef CONFIG_64BIT
9850 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9852 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9854 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9862 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9864 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9866 #ifndef CONFIG_64BIT
9868 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9870 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9872 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9878 /* return total cpu usage (in nanoseconds) of a group */
9879 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9881 struct cpuacct *ca = cgroup_ca(cgrp);
9882 u64 totalcpuusage = 0;
9885 for_each_present_cpu(i)
9886 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9888 return totalcpuusage;
9891 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9894 struct cpuacct *ca = cgroup_ca(cgrp);
9903 for_each_present_cpu(i)
9904 cpuacct_cpuusage_write(ca, i, 0);
9910 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9913 struct cpuacct *ca = cgroup_ca(cgroup);
9917 for_each_present_cpu(i) {
9918 percpu = cpuacct_cpuusage_read(ca, i);
9919 seq_printf(m, "%llu ", (unsigned long long) percpu);
9921 seq_printf(m, "\n");
9925 static const char *cpuacct_stat_desc[] = {
9926 [CPUACCT_STAT_USER] = "user",
9927 [CPUACCT_STAT_SYSTEM] = "system",
9930 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9931 struct cgroup_map_cb *cb)
9933 struct cpuacct *ca = cgroup_ca(cgrp);
9936 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9937 s64 val = percpu_counter_read(&ca->cpustat[i]);
9938 val = cputime64_to_clock_t(val);
9939 cb->fill(cb, cpuacct_stat_desc[i], val);
9944 static struct cftype files[] = {
9947 .read_u64 = cpuusage_read,
9948 .write_u64 = cpuusage_write,
9951 .name = "usage_percpu",
9952 .read_seq_string = cpuacct_percpu_seq_read,
9956 .read_map = cpuacct_stats_show,
9960 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9962 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9966 * charge this task's execution time to its accounting group.
9968 * called with rq->lock held.
9970 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9975 if (unlikely(!cpuacct_subsys.active))
9978 cpu = task_cpu(tsk);
9984 for (; ca; ca = ca->parent) {
9985 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9986 *cpuusage += cputime;
9993 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9994 * in cputime_t units. As a result, cpuacct_update_stats calls
9995 * percpu_counter_add with values large enough to always overflow the
9996 * per cpu batch limit causing bad SMP scalability.
9998 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9999 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
10000 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
10003 #define CPUACCT_BATCH \
10004 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
10006 #define CPUACCT_BATCH 0
10010 * Charge the system/user time to the task's accounting group.
10012 static void cpuacct_update_stats(struct task_struct *tsk,
10013 enum cpuacct_stat_index idx, cputime_t val)
10015 struct cpuacct *ca;
10016 int batch = CPUACCT_BATCH;
10018 if (unlikely(!cpuacct_subsys.active))
10025 __percpu_counter_add(&ca->cpustat[idx], val, batch);
10031 struct cgroup_subsys cpuacct_subsys = {
10033 .create = cpuacct_create,
10034 .destroy = cpuacct_destroy,
10035 .populate = cpuacct_populate,
10036 .subsys_id = cpuacct_subsys_id,
10038 #endif /* CONFIG_CGROUP_CPUACCT */