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)
2192 #include "sched_idletask.c"
2193 #include "sched_fair.c"
2194 #include "sched_rt.c"
2195 #include "sched_autogroup.c"
2196 #include "sched_stoptask.c"
2197 #ifdef CONFIG_SCHED_DEBUG
2198 # include "sched_debug.c"
2201 void sched_set_stop_task(int cpu, struct task_struct *stop)
2203 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2204 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2208 * Make it appear like a SCHED_FIFO task, its something
2209 * userspace knows about and won't get confused about.
2211 * Also, it will make PI more or less work without too
2212 * much confusion -- but then, stop work should not
2213 * rely on PI working anyway.
2215 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2217 stop->sched_class = &stop_sched_class;
2220 cpu_rq(cpu)->stop = stop;
2224 * Reset it back to a normal scheduling class so that
2225 * it can die in pieces.
2227 old_stop->sched_class = &rt_sched_class;
2232 * __normal_prio - return the priority that is based on the static prio
2234 static inline int __normal_prio(struct task_struct *p)
2236 return p->static_prio;
2240 * Calculate the expected normal priority: i.e. priority
2241 * without taking RT-inheritance into account. Might be
2242 * boosted by interactivity modifiers. Changes upon fork,
2243 * setprio syscalls, and whenever the interactivity
2244 * estimator recalculates.
2246 static inline int normal_prio(struct task_struct *p)
2250 if (task_has_rt_policy(p))
2251 prio = MAX_RT_PRIO-1 - p->rt_priority;
2253 prio = __normal_prio(p);
2258 * Calculate the current priority, i.e. the priority
2259 * taken into account by the scheduler. This value might
2260 * be boosted by RT tasks, or might be boosted by
2261 * interactivity modifiers. Will be RT if the task got
2262 * RT-boosted. If not then it returns p->normal_prio.
2264 static int effective_prio(struct task_struct *p)
2266 p->normal_prio = normal_prio(p);
2268 * If we are RT tasks or we were boosted to RT priority,
2269 * keep the priority unchanged. Otherwise, update priority
2270 * to the normal priority:
2272 if (!rt_prio(p->prio))
2273 return p->normal_prio;
2278 * task_curr - is this task currently executing on a CPU?
2279 * @p: the task in question.
2281 inline int task_curr(const struct task_struct *p)
2283 return cpu_curr(task_cpu(p)) == p;
2286 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2287 const struct sched_class *prev_class,
2290 if (prev_class != p->sched_class) {
2291 if (prev_class->switched_from)
2292 prev_class->switched_from(rq, p);
2293 p->sched_class->switched_to(rq, p);
2294 } else if (oldprio != p->prio)
2295 p->sched_class->prio_changed(rq, p, oldprio);
2298 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2300 const struct sched_class *class;
2302 if (p->sched_class == rq->curr->sched_class) {
2303 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2305 for_each_class(class) {
2306 if (class == rq->curr->sched_class)
2308 if (class == p->sched_class) {
2309 resched_task(rq->curr);
2316 * A queue event has occurred, and we're going to schedule. In
2317 * this case, we can save a useless back to back clock update.
2319 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2320 rq->skip_clock_update = 1;
2325 * Is this task likely cache-hot:
2328 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2332 if (p->sched_class != &fair_sched_class)
2335 if (unlikely(p->policy == SCHED_IDLE))
2339 * Buddy candidates are cache hot:
2341 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2342 (&p->se == cfs_rq_of(&p->se)->next ||
2343 &p->se == cfs_rq_of(&p->se)->last))
2346 if (sysctl_sched_migration_cost == -1)
2348 if (sysctl_sched_migration_cost == 0)
2351 delta = now - p->se.exec_start;
2353 return delta < (s64)sysctl_sched_migration_cost;
2356 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2358 #ifdef CONFIG_SCHED_DEBUG
2360 * We should never call set_task_cpu() on a blocked task,
2361 * ttwu() will sort out the placement.
2363 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2364 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2366 #ifdef CONFIG_LOCKDEP
2368 * The caller should hold either p->pi_lock or rq->lock, when changing
2369 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2371 * sched_move_task() holds both and thus holding either pins the cgroup,
2374 * Furthermore, all task_rq users should acquire both locks, see
2377 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2378 lockdep_is_held(&task_rq(p)->lock)));
2382 trace_sched_migrate_task(p, new_cpu);
2384 if (task_cpu(p) != new_cpu) {
2385 p->se.nr_migrations++;
2386 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2389 __set_task_cpu(p, new_cpu);
2392 struct migration_arg {
2393 struct task_struct *task;
2397 static int migration_cpu_stop(void *data);
2400 * wait_task_inactive - wait for a thread to unschedule.
2402 * If @match_state is nonzero, it's the @p->state value just checked and
2403 * not expected to change. If it changes, i.e. @p might have woken up,
2404 * then return zero. When we succeed in waiting for @p to be off its CPU,
2405 * we return a positive number (its total switch count). If a second call
2406 * a short while later returns the same number, the caller can be sure that
2407 * @p has remained unscheduled the whole time.
2409 * The caller must ensure that the task *will* unschedule sometime soon,
2410 * else this function might spin for a *long* time. This function can't
2411 * be called with interrupts off, or it may introduce deadlock with
2412 * smp_call_function() if an IPI is sent by the same process we are
2413 * waiting to become inactive.
2415 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2417 unsigned long flags;
2424 * We do the initial early heuristics without holding
2425 * any task-queue locks at all. We'll only try to get
2426 * the runqueue lock when things look like they will
2432 * If the task is actively running on another CPU
2433 * still, just relax and busy-wait without holding
2436 * NOTE! Since we don't hold any locks, it's not
2437 * even sure that "rq" stays as the right runqueue!
2438 * But we don't care, since "task_running()" will
2439 * return false if the runqueue has changed and p
2440 * is actually now running somewhere else!
2442 while (task_running(rq, p)) {
2443 if (match_state && unlikely(p->state != match_state))
2449 * Ok, time to look more closely! We need the rq
2450 * lock now, to be *sure*. If we're wrong, we'll
2451 * just go back and repeat.
2453 rq = task_rq_lock(p, &flags);
2454 trace_sched_wait_task(p);
2455 running = task_running(rq, p);
2458 if (!match_state || p->state == match_state)
2459 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2460 task_rq_unlock(rq, p, &flags);
2463 * If it changed from the expected state, bail out now.
2465 if (unlikely(!ncsw))
2469 * Was it really running after all now that we
2470 * checked with the proper locks actually held?
2472 * Oops. Go back and try again..
2474 if (unlikely(running)) {
2480 * It's not enough that it's not actively running,
2481 * it must be off the runqueue _entirely_, and not
2484 * So if it was still runnable (but just not actively
2485 * running right now), it's preempted, and we should
2486 * yield - it could be a while.
2488 if (unlikely(on_rq)) {
2489 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2491 set_current_state(TASK_UNINTERRUPTIBLE);
2492 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2497 * Ahh, all good. It wasn't running, and it wasn't
2498 * runnable, which means that it will never become
2499 * running in the future either. We're all done!
2508 * kick_process - kick a running thread to enter/exit the kernel
2509 * @p: the to-be-kicked thread
2511 * Cause a process which is running on another CPU to enter
2512 * kernel-mode, without any delay. (to get signals handled.)
2514 * NOTE: this function doesn't have to take the runqueue lock,
2515 * because all it wants to ensure is that the remote task enters
2516 * the kernel. If the IPI races and the task has been migrated
2517 * to another CPU then no harm is done and the purpose has been
2520 void kick_process(struct task_struct *p)
2526 if ((cpu != smp_processor_id()) && task_curr(p))
2527 smp_send_reschedule(cpu);
2530 EXPORT_SYMBOL_GPL(kick_process);
2531 #endif /* CONFIG_SMP */
2535 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2537 static int select_fallback_rq(int cpu, struct task_struct *p)
2540 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2542 /* Look for allowed, online CPU in same node. */
2543 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2544 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
2547 /* Any allowed, online CPU? */
2548 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
2549 if (dest_cpu < nr_cpu_ids)
2552 /* No more Mr. Nice Guy. */
2553 dest_cpu = cpuset_cpus_allowed_fallback(p);
2555 * Don't tell them about moving exiting tasks or
2556 * kernel threads (both mm NULL), since they never
2559 if (p->mm && printk_ratelimit()) {
2560 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2561 task_pid_nr(p), p->comm, cpu);
2568 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2571 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2573 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2576 * In order not to call set_task_cpu() on a blocking task we need
2577 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2580 * Since this is common to all placement strategies, this lives here.
2582 * [ this allows ->select_task() to simply return task_cpu(p) and
2583 * not worry about this generic constraint ]
2585 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
2587 cpu = select_fallback_rq(task_cpu(p), p);
2592 static void update_avg(u64 *avg, u64 sample)
2594 s64 diff = sample - *avg;
2600 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2602 #ifdef CONFIG_SCHEDSTATS
2603 struct rq *rq = this_rq();
2606 int this_cpu = smp_processor_id();
2608 if (cpu == this_cpu) {
2609 schedstat_inc(rq, ttwu_local);
2610 schedstat_inc(p, se.statistics.nr_wakeups_local);
2612 struct sched_domain *sd;
2614 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2616 for_each_domain(this_cpu, sd) {
2617 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2618 schedstat_inc(sd, ttwu_wake_remote);
2625 if (wake_flags & WF_MIGRATED)
2626 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2628 #endif /* CONFIG_SMP */
2630 schedstat_inc(rq, ttwu_count);
2631 schedstat_inc(p, se.statistics.nr_wakeups);
2633 if (wake_flags & WF_SYNC)
2634 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2636 #endif /* CONFIG_SCHEDSTATS */
2639 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2641 activate_task(rq, p, en_flags);
2644 /* if a worker is waking up, notify workqueue */
2645 if (p->flags & PF_WQ_WORKER)
2646 wq_worker_waking_up(p, cpu_of(rq));
2650 * Mark the task runnable and perform wakeup-preemption.
2653 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2655 trace_sched_wakeup(p, true);
2656 check_preempt_curr(rq, p, wake_flags);
2658 p->state = TASK_RUNNING;
2660 if (p->sched_class->task_woken)
2661 p->sched_class->task_woken(rq, p);
2663 if (rq->idle_stamp) {
2664 u64 delta = rq->clock - rq->idle_stamp;
2665 u64 max = 2*sysctl_sched_migration_cost;
2670 update_avg(&rq->avg_idle, delta);
2677 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2680 if (p->sched_contributes_to_load)
2681 rq->nr_uninterruptible--;
2684 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2685 ttwu_do_wakeup(rq, p, wake_flags);
2689 * Called in case the task @p isn't fully descheduled from its runqueue,
2690 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2691 * since all we need to do is flip p->state to TASK_RUNNING, since
2692 * the task is still ->on_rq.
2694 static int ttwu_remote(struct task_struct *p, int wake_flags)
2699 rq = __task_rq_lock(p);
2701 ttwu_do_wakeup(rq, p, wake_flags);
2704 __task_rq_unlock(rq);
2710 static void sched_ttwu_pending(void)
2712 struct rq *rq = this_rq();
2713 struct llist_node *llist = llist_del_all(&rq->wake_list);
2714 struct task_struct *p;
2716 raw_spin_lock(&rq->lock);
2719 p = llist_entry(llist, struct task_struct, wake_entry);
2720 llist = llist_next(llist);
2721 ttwu_do_activate(rq, p, 0);
2724 raw_spin_unlock(&rq->lock);
2727 void scheduler_ipi(void)
2729 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2733 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2734 * traditionally all their work was done from the interrupt return
2735 * path. Now that we actually do some work, we need to make sure
2738 * Some archs already do call them, luckily irq_enter/exit nest
2741 * Arguably we should visit all archs and update all handlers,
2742 * however a fair share of IPIs are still resched only so this would
2743 * somewhat pessimize the simple resched case.
2746 sched_ttwu_pending();
2749 * Check if someone kicked us for doing the nohz idle load balance.
2751 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
2752 this_rq()->idle_balance = 1;
2753 raise_softirq_irqoff(SCHED_SOFTIRQ);
2758 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2760 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
2761 smp_send_reschedule(cpu);
2764 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2765 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2770 rq = __task_rq_lock(p);
2772 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2773 ttwu_do_wakeup(rq, p, wake_flags);
2776 __task_rq_unlock(rq);
2781 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2782 #endif /* CONFIG_SMP */
2784 static void ttwu_queue(struct task_struct *p, int cpu)
2786 struct rq *rq = cpu_rq(cpu);
2788 #if defined(CONFIG_SMP)
2789 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2790 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2791 ttwu_queue_remote(p, cpu);
2796 raw_spin_lock(&rq->lock);
2797 ttwu_do_activate(rq, p, 0);
2798 raw_spin_unlock(&rq->lock);
2802 * try_to_wake_up - wake up a thread
2803 * @p: the thread to be awakened
2804 * @state: the mask of task states that can be woken
2805 * @wake_flags: wake modifier flags (WF_*)
2807 * Put it on the run-queue if it's not already there. The "current"
2808 * thread is always on the run-queue (except when the actual
2809 * re-schedule is in progress), and as such you're allowed to do
2810 * the simpler "current->state = TASK_RUNNING" to mark yourself
2811 * runnable without the overhead of this.
2813 * Returns %true if @p was woken up, %false if it was already running
2814 * or @state didn't match @p's state.
2817 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2819 unsigned long flags;
2820 int cpu, success = 0;
2823 raw_spin_lock_irqsave(&p->pi_lock, flags);
2824 if (!(p->state & state))
2827 success = 1; /* we're going to change ->state */
2830 if (p->on_rq && ttwu_remote(p, wake_flags))
2835 * If the owning (remote) cpu is still in the middle of schedule() with
2836 * this task as prev, wait until its done referencing the task.
2839 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2841 * In case the architecture enables interrupts in
2842 * context_switch(), we cannot busy wait, since that
2843 * would lead to deadlocks when an interrupt hits and
2844 * tries to wake up @prev. So bail and do a complete
2847 if (ttwu_activate_remote(p, wake_flags))
2854 * Pairs with the smp_wmb() in finish_lock_switch().
2858 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2859 p->state = TASK_WAKING;
2861 if (p->sched_class->task_waking)
2862 p->sched_class->task_waking(p);
2864 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2865 if (task_cpu(p) != cpu) {
2866 wake_flags |= WF_MIGRATED;
2867 set_task_cpu(p, cpu);
2869 #endif /* CONFIG_SMP */
2873 ttwu_stat(p, cpu, wake_flags);
2875 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2881 * try_to_wake_up_local - try to wake up a local task with rq lock held
2882 * @p: the thread to be awakened
2884 * Put @p on the run-queue if it's not already there. The caller must
2885 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2888 static void try_to_wake_up_local(struct task_struct *p)
2890 struct rq *rq = task_rq(p);
2892 BUG_ON(rq != this_rq());
2893 BUG_ON(p == current);
2894 lockdep_assert_held(&rq->lock);
2896 if (!raw_spin_trylock(&p->pi_lock)) {
2897 raw_spin_unlock(&rq->lock);
2898 raw_spin_lock(&p->pi_lock);
2899 raw_spin_lock(&rq->lock);
2902 if (!(p->state & TASK_NORMAL))
2906 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2908 ttwu_do_wakeup(rq, p, 0);
2909 ttwu_stat(p, smp_processor_id(), 0);
2911 raw_spin_unlock(&p->pi_lock);
2915 * wake_up_process - Wake up a specific process
2916 * @p: The process to be woken up.
2918 * Attempt to wake up the nominated process and move it to the set of runnable
2919 * processes. Returns 1 if the process was woken up, 0 if it was already
2922 * It may be assumed that this function implies a write memory barrier before
2923 * changing the task state if and only if any tasks are woken up.
2925 int wake_up_process(struct task_struct *p)
2927 WARN_ON(task_is_stopped_or_traced(p));
2928 return try_to_wake_up(p, TASK_NORMAL, 0);
2930 EXPORT_SYMBOL(wake_up_process);
2932 int wake_up_state(struct task_struct *p, unsigned int state)
2934 return try_to_wake_up(p, state, 0);
2938 * Perform scheduler related setup for a newly forked process p.
2939 * p is forked by current.
2941 * __sched_fork() is basic setup used by init_idle() too:
2943 static void __sched_fork(struct task_struct *p)
2948 p->se.exec_start = 0;
2949 p->se.sum_exec_runtime = 0;
2950 p->se.prev_sum_exec_runtime = 0;
2951 p->se.nr_migrations = 0;
2953 INIT_LIST_HEAD(&p->se.group_node);
2955 #ifdef CONFIG_SCHEDSTATS
2956 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2959 INIT_LIST_HEAD(&p->rt.run_list);
2961 #ifdef CONFIG_PREEMPT_NOTIFIERS
2962 INIT_HLIST_HEAD(&p->preempt_notifiers);
2967 * fork()/clone()-time setup:
2969 void sched_fork(struct task_struct *p)
2971 unsigned long flags;
2972 int cpu = get_cpu();
2976 * We mark the process as running here. This guarantees that
2977 * nobody will actually run it, and a signal or other external
2978 * event cannot wake it up and insert it on the runqueue either.
2980 p->state = TASK_RUNNING;
2983 * Make sure we do not leak PI boosting priority to the child.
2985 p->prio = current->normal_prio;
2988 * Revert to default priority/policy on fork if requested.
2990 if (unlikely(p->sched_reset_on_fork)) {
2991 if (task_has_rt_policy(p)) {
2992 p->policy = SCHED_NORMAL;
2993 p->static_prio = NICE_TO_PRIO(0);
2995 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2996 p->static_prio = NICE_TO_PRIO(0);
2998 p->prio = p->normal_prio = __normal_prio(p);
3002 * We don't need the reset flag anymore after the fork. It has
3003 * fulfilled its duty:
3005 p->sched_reset_on_fork = 0;
3008 if (!rt_prio(p->prio))
3009 p->sched_class = &fair_sched_class;
3011 if (p->sched_class->task_fork)
3012 p->sched_class->task_fork(p);
3015 * The child is not yet in the pid-hash so no cgroup attach races,
3016 * and the cgroup is pinned to this child due to cgroup_fork()
3017 * is ran before sched_fork().
3019 * Silence PROVE_RCU.
3021 raw_spin_lock_irqsave(&p->pi_lock, flags);
3022 set_task_cpu(p, cpu);
3023 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3025 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3026 if (likely(sched_info_on()))
3027 memset(&p->sched_info, 0, sizeof(p->sched_info));
3029 #if defined(CONFIG_SMP)
3032 #ifdef CONFIG_PREEMPT_COUNT
3033 /* Want to start with kernel preemption disabled. */
3034 task_thread_info(p)->preempt_count = 1;
3037 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3044 * wake_up_new_task - wake up a newly created task for the first time.
3046 * This function will do some initial scheduler statistics housekeeping
3047 * that must be done for every newly created context, then puts the task
3048 * on the runqueue and wakes it.
3050 void wake_up_new_task(struct task_struct *p)
3052 unsigned long flags;
3055 raw_spin_lock_irqsave(&p->pi_lock, flags);
3058 * Fork balancing, do it here and not earlier because:
3059 * - cpus_allowed can change in the fork path
3060 * - any previously selected cpu might disappear through hotplug
3062 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3065 rq = __task_rq_lock(p);
3066 activate_task(rq, p, 0);
3068 trace_sched_wakeup_new(p, true);
3069 check_preempt_curr(rq, p, WF_FORK);
3071 if (p->sched_class->task_woken)
3072 p->sched_class->task_woken(rq, p);
3074 task_rq_unlock(rq, p, &flags);
3077 #ifdef CONFIG_PREEMPT_NOTIFIERS
3080 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3081 * @notifier: notifier struct to register
3083 void preempt_notifier_register(struct preempt_notifier *notifier)
3085 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3087 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3090 * preempt_notifier_unregister - no longer interested in preemption notifications
3091 * @notifier: notifier struct to unregister
3093 * This is safe to call from within a preemption notifier.
3095 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3097 hlist_del(¬ifier->link);
3099 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3101 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3103 struct preempt_notifier *notifier;
3104 struct hlist_node *node;
3106 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3107 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3111 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3112 struct task_struct *next)
3114 struct preempt_notifier *notifier;
3115 struct hlist_node *node;
3117 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3118 notifier->ops->sched_out(notifier, next);
3121 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3123 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3128 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3129 struct task_struct *next)
3133 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3136 * prepare_task_switch - prepare to switch tasks
3137 * @rq: the runqueue preparing to switch
3138 * @prev: the current task that is being switched out
3139 * @next: the task we are going to switch to.
3141 * This is called with the rq lock held and interrupts off. It must
3142 * be paired with a subsequent finish_task_switch after the context
3145 * prepare_task_switch sets up locking and calls architecture specific
3149 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3150 struct task_struct *next)
3152 sched_info_switch(prev, next);
3153 perf_event_task_sched_out(prev, next);
3154 fire_sched_out_preempt_notifiers(prev, next);
3155 prepare_lock_switch(rq, next);
3156 prepare_arch_switch(next);
3157 trace_sched_switch(prev, next);
3161 * finish_task_switch - clean up after a task-switch
3162 * @rq: runqueue associated with task-switch
3163 * @prev: the thread we just switched away from.
3165 * finish_task_switch must be called after the context switch, paired
3166 * with a prepare_task_switch call before the context switch.
3167 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3168 * and do any other architecture-specific cleanup actions.
3170 * Note that we may have delayed dropping an mm in context_switch(). If
3171 * so, we finish that here outside of the runqueue lock. (Doing it
3172 * with the lock held can cause deadlocks; see schedule() for
3175 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3176 __releases(rq->lock)
3178 struct mm_struct *mm = rq->prev_mm;
3184 * A task struct has one reference for the use as "current".
3185 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3186 * schedule one last time. The schedule call will never return, and
3187 * the scheduled task must drop that reference.
3188 * The test for TASK_DEAD must occur while the runqueue locks are
3189 * still held, otherwise prev could be scheduled on another cpu, die
3190 * there before we look at prev->state, and then the reference would
3192 * Manfred Spraul <manfred@colorfullife.com>
3194 prev_state = prev->state;
3195 finish_arch_switch(prev);
3196 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3197 local_irq_disable();
3198 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3199 perf_event_task_sched_in(prev, current);
3200 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3202 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3203 finish_lock_switch(rq, prev);
3205 fire_sched_in_preempt_notifiers(current);
3208 if (unlikely(prev_state == TASK_DEAD)) {
3210 * Remove function-return probe instances associated with this
3211 * task and put them back on the free list.
3213 kprobe_flush_task(prev);
3214 put_task_struct(prev);
3220 /* assumes rq->lock is held */
3221 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3223 if (prev->sched_class->pre_schedule)
3224 prev->sched_class->pre_schedule(rq, prev);
3227 /* rq->lock is NOT held, but preemption is disabled */
3228 static inline void post_schedule(struct rq *rq)
3230 if (rq->post_schedule) {
3231 unsigned long flags;
3233 raw_spin_lock_irqsave(&rq->lock, flags);
3234 if (rq->curr->sched_class->post_schedule)
3235 rq->curr->sched_class->post_schedule(rq);
3236 raw_spin_unlock_irqrestore(&rq->lock, flags);
3238 rq->post_schedule = 0;
3244 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3248 static inline void post_schedule(struct rq *rq)
3255 * schedule_tail - first thing a freshly forked thread must call.
3256 * @prev: the thread we just switched away from.
3258 asmlinkage void schedule_tail(struct task_struct *prev)
3259 __releases(rq->lock)
3261 struct rq *rq = this_rq();
3263 finish_task_switch(rq, prev);
3266 * FIXME: do we need to worry about rq being invalidated by the
3271 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3272 /* In this case, finish_task_switch does not reenable preemption */
3275 if (current->set_child_tid)
3276 put_user(task_pid_vnr(current), current->set_child_tid);
3280 * context_switch - switch to the new MM and the new
3281 * thread's register state.
3284 context_switch(struct rq *rq, struct task_struct *prev,
3285 struct task_struct *next)
3287 struct mm_struct *mm, *oldmm;
3289 prepare_task_switch(rq, prev, next);
3292 oldmm = prev->active_mm;
3294 * For paravirt, this is coupled with an exit in switch_to to
3295 * combine the page table reload and the switch backend into
3298 arch_start_context_switch(prev);
3301 next->active_mm = oldmm;
3302 atomic_inc(&oldmm->mm_count);
3303 enter_lazy_tlb(oldmm, next);
3305 switch_mm(oldmm, mm, next);
3308 prev->active_mm = NULL;
3309 rq->prev_mm = oldmm;
3312 * Since the runqueue lock will be released by the next
3313 * task (which is an invalid locking op but in the case
3314 * of the scheduler it's an obvious special-case), so we
3315 * do an early lockdep release here:
3317 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3318 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3321 /* Here we just switch the register state and the stack. */
3322 switch_to(prev, next, prev);
3326 * this_rq must be evaluated again because prev may have moved
3327 * CPUs since it called schedule(), thus the 'rq' on its stack
3328 * frame will be invalid.
3330 finish_task_switch(this_rq(), prev);
3334 * nr_running, nr_uninterruptible and nr_context_switches:
3336 * externally visible scheduler statistics: current number of runnable
3337 * threads, current number of uninterruptible-sleeping threads, total
3338 * number of context switches performed since bootup.
3340 unsigned long nr_running(void)
3342 unsigned long i, sum = 0;
3344 for_each_online_cpu(i)
3345 sum += cpu_rq(i)->nr_running;
3350 unsigned long nr_uninterruptible(void)
3352 unsigned long i, sum = 0;
3354 for_each_possible_cpu(i)
3355 sum += cpu_rq(i)->nr_uninterruptible;
3358 * Since we read the counters lockless, it might be slightly
3359 * inaccurate. Do not allow it to go below zero though:
3361 if (unlikely((long)sum < 0))
3367 unsigned long long nr_context_switches(void)
3370 unsigned long long sum = 0;
3372 for_each_possible_cpu(i)
3373 sum += cpu_rq(i)->nr_switches;
3378 unsigned long nr_iowait(void)
3380 unsigned long i, sum = 0;
3382 for_each_possible_cpu(i)
3383 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3388 unsigned long nr_iowait_cpu(int cpu)
3390 struct rq *this = cpu_rq(cpu);
3391 return atomic_read(&this->nr_iowait);
3394 unsigned long this_cpu_load(void)
3396 struct rq *this = this_rq();
3397 return this->cpu_load[0];
3402 * Global load-average calculations
3404 * We take a distributed and async approach to calculating the global load-avg
3405 * in order to minimize overhead.
3407 * The global load average is an exponentially decaying average of nr_running +
3408 * nr_uninterruptible.
3410 * Once every LOAD_FREQ:
3413 * for_each_possible_cpu(cpu)
3414 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
3416 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
3418 * Due to a number of reasons the above turns in the mess below:
3420 * - for_each_possible_cpu() is prohibitively expensive on machines with
3421 * serious number of cpus, therefore we need to take a distributed approach
3422 * to calculating nr_active.
3424 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
3425 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
3427 * So assuming nr_active := 0 when we start out -- true per definition, we
3428 * can simply take per-cpu deltas and fold those into a global accumulate
3429 * to obtain the same result. See calc_load_fold_active().
3431 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
3432 * across the machine, we assume 10 ticks is sufficient time for every
3433 * cpu to have completed this task.
3435 * This places an upper-bound on the IRQ-off latency of the machine. Then
3436 * again, being late doesn't loose the delta, just wrecks the sample.
3438 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
3439 * this would add another cross-cpu cacheline miss and atomic operation
3440 * to the wakeup path. Instead we increment on whatever cpu the task ran
3441 * when it went into uninterruptible state and decrement on whatever cpu
3442 * did the wakeup. This means that only the sum of nr_uninterruptible over
3443 * all cpus yields the correct result.
3445 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
3448 /* Variables and functions for calc_load */
3449 static atomic_long_t calc_load_tasks;
3450 static unsigned long calc_load_update;
3451 unsigned long avenrun[3];
3452 EXPORT_SYMBOL(avenrun); /* should be removed */
3455 * get_avenrun - get the load average array
3456 * @loads: pointer to dest load array
3457 * @offset: offset to add
3458 * @shift: shift count to shift the result left
3460 * These values are estimates at best, so no need for locking.
3462 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3464 loads[0] = (avenrun[0] + offset) << shift;
3465 loads[1] = (avenrun[1] + offset) << shift;
3466 loads[2] = (avenrun[2] + offset) << shift;
3469 static long calc_load_fold_active(struct rq *this_rq)
3471 long nr_active, delta = 0;
3473 nr_active = this_rq->nr_running;
3474 nr_active += (long) this_rq->nr_uninterruptible;
3476 if (nr_active != this_rq->calc_load_active) {
3477 delta = nr_active - this_rq->calc_load_active;
3478 this_rq->calc_load_active = nr_active;
3485 * a1 = a0 * e + a * (1 - e)
3487 static unsigned long
3488 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3491 load += active * (FIXED_1 - exp);
3492 load += 1UL << (FSHIFT - 1);
3493 return load >> FSHIFT;
3498 * Handle NO_HZ for the global load-average.
3500 * Since the above described distributed algorithm to compute the global
3501 * load-average relies on per-cpu sampling from the tick, it is affected by
3504 * The basic idea is to fold the nr_active delta into a global idle-delta upon
3505 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
3506 * when we read the global state.
3508 * Obviously reality has to ruin such a delightfully simple scheme:
3510 * - When we go NO_HZ idle during the window, we can negate our sample
3511 * contribution, causing under-accounting.
3513 * We avoid this by keeping two idle-delta counters and flipping them
3514 * when the window starts, thus separating old and new NO_HZ load.
3516 * The only trick is the slight shift in index flip for read vs write.
3520 * |-|-----------|-|-----------|-|-----------|-|
3521 * r:0 0 1 1 0 0 1 1 0
3522 * w:0 1 1 0 0 1 1 0 0
3524 * This ensures we'll fold the old idle contribution in this window while
3525 * accumlating the new one.
3527 * - When we wake up from NO_HZ idle during the window, we push up our
3528 * contribution, since we effectively move our sample point to a known
3531 * This is solved by pushing the window forward, and thus skipping the
3532 * sample, for this cpu (effectively using the idle-delta for this cpu which
3533 * was in effect at the time the window opened). This also solves the issue
3534 * of having to deal with a cpu having been in NOHZ idle for multiple
3535 * LOAD_FREQ intervals.
3537 * When making the ILB scale, we should try to pull this in as well.
3539 static atomic_long_t calc_load_idle[2];
3540 static int calc_load_idx;
3542 static inline int calc_load_write_idx(void)
3544 int idx = calc_load_idx;
3547 * See calc_global_nohz(), if we observe the new index, we also
3548 * need to observe the new update time.
3553 * If the folding window started, make sure we start writing in the
3556 if (!time_before(jiffies, calc_load_update))
3562 static inline int calc_load_read_idx(void)
3564 return calc_load_idx & 1;
3567 void calc_load_enter_idle(void)
3569 struct rq *this_rq = this_rq();
3573 * We're going into NOHZ mode, if there's any pending delta, fold it
3574 * into the pending idle delta.
3576 delta = calc_load_fold_active(this_rq);
3578 int idx = calc_load_write_idx();
3579 atomic_long_add(delta, &calc_load_idle[idx]);
3583 void calc_load_exit_idle(void)
3585 struct rq *this_rq = this_rq();
3588 * If we're still before the sample window, we're done.
3590 if (time_before(jiffies, this_rq->calc_load_update))
3594 * We woke inside or after the sample window, this means we're already
3595 * accounted through the nohz accounting, so skip the entire deal and
3596 * sync up for the next window.
3598 this_rq->calc_load_update = calc_load_update;
3599 if (time_before(jiffies, this_rq->calc_load_update + 10))
3600 this_rq->calc_load_update += LOAD_FREQ;
3603 static long calc_load_fold_idle(void)
3605 int idx = calc_load_read_idx();
3608 if (atomic_long_read(&calc_load_idle[idx]))
3609 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
3615 * fixed_power_int - compute: x^n, in O(log n) time
3617 * @x: base of the power
3618 * @frac_bits: fractional bits of @x
3619 * @n: power to raise @x to.
3621 * By exploiting the relation between the definition of the natural power
3622 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3623 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3624 * (where: n_i \elem {0, 1}, the binary vector representing n),
3625 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3626 * of course trivially computable in O(log_2 n), the length of our binary
3629 static unsigned long
3630 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3632 unsigned long result = 1UL << frac_bits;
3637 result += 1UL << (frac_bits - 1);
3638 result >>= frac_bits;
3644 x += 1UL << (frac_bits - 1);
3652 * a1 = a0 * e + a * (1 - e)
3654 * a2 = a1 * e + a * (1 - e)
3655 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3656 * = a0 * e^2 + a * (1 - e) * (1 + e)
3658 * a3 = a2 * e + a * (1 - e)
3659 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3660 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3664 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3665 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3666 * = a0 * e^n + a * (1 - e^n)
3668 * [1] application of the geometric series:
3671 * S_n := \Sum x^i = -------------
3674 static unsigned long
3675 calc_load_n(unsigned long load, unsigned long exp,
3676 unsigned long active, unsigned int n)
3679 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3683 * NO_HZ can leave us missing all per-cpu ticks calling
3684 * calc_load_account_active(), but since an idle CPU folds its delta into
3685 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3686 * in the pending idle delta if our idle period crossed a load cycle boundary.
3688 * Once we've updated the global active value, we need to apply the exponential
3689 * weights adjusted to the number of cycles missed.
3691 static void calc_global_nohz(void)
3693 long delta, active, n;
3695 if (!time_before(jiffies, calc_load_update + 10)) {
3697 * Catch-up, fold however many we are behind still
3699 delta = jiffies - calc_load_update - 10;
3700 n = 1 + (delta / LOAD_FREQ);
3702 active = atomic_long_read(&calc_load_tasks);
3703 active = active > 0 ? active * FIXED_1 : 0;
3705 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3706 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3707 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3709 calc_load_update += n * LOAD_FREQ;
3713 * Flip the idle index...
3715 * Make sure we first write the new time then flip the index, so that
3716 * calc_load_write_idx() will see the new time when it reads the new
3717 * index, this avoids a double flip messing things up.
3722 #else /* !CONFIG_NO_HZ */
3724 static inline long calc_load_fold_idle(void) { return 0; }
3725 static inline void calc_global_nohz(void) { }
3727 #endif /* CONFIG_NO_HZ */
3730 * calc_load - update the avenrun load estimates 10 ticks after the
3731 * CPUs have updated calc_load_tasks.
3733 void calc_global_load(unsigned long ticks)
3737 if (time_before(jiffies, calc_load_update + 10))
3741 * Fold the 'old' idle-delta to include all NO_HZ cpus.
3743 delta = calc_load_fold_idle();
3745 atomic_long_add(delta, &calc_load_tasks);
3747 active = atomic_long_read(&calc_load_tasks);
3748 active = active > 0 ? active * FIXED_1 : 0;
3750 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3751 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3752 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3754 calc_load_update += LOAD_FREQ;
3757 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
3763 * Called from update_cpu_load() to periodically update this CPU's
3766 static void calc_load_account_active(struct rq *this_rq)
3770 if (time_before(jiffies, this_rq->calc_load_update))
3773 delta = calc_load_fold_active(this_rq);
3775 atomic_long_add(delta, &calc_load_tasks);
3777 this_rq->calc_load_update += LOAD_FREQ;
3781 * End of global load-average stuff
3785 * The exact cpuload at various idx values, calculated at every tick would be
3786 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3788 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3789 * on nth tick when cpu may be busy, then we have:
3790 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3791 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3793 * decay_load_missed() below does efficient calculation of
3794 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3795 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3797 * The calculation is approximated on a 128 point scale.
3798 * degrade_zero_ticks is the number of ticks after which load at any
3799 * particular idx is approximated to be zero.
3800 * degrade_factor is a precomputed table, a row for each load idx.
3801 * Each column corresponds to degradation factor for a power of two ticks,
3802 * based on 128 point scale.
3804 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3805 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3807 * With this power of 2 load factors, we can degrade the load n times
3808 * by looking at 1 bits in n and doing as many mult/shift instead of
3809 * n mult/shifts needed by the exact degradation.
3811 #define DEGRADE_SHIFT 7
3812 static const unsigned char
3813 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3814 static const unsigned char
3815 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3816 {0, 0, 0, 0, 0, 0, 0, 0},
3817 {64, 32, 8, 0, 0, 0, 0, 0},
3818 {96, 72, 40, 12, 1, 0, 0},
3819 {112, 98, 75, 43, 15, 1, 0},
3820 {120, 112, 98, 76, 45, 16, 2} };
3823 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3824 * would be when CPU is idle and so we just decay the old load without
3825 * adding any new load.
3827 static unsigned long
3828 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3832 if (!missed_updates)
3835 if (missed_updates >= degrade_zero_ticks[idx])
3839 return load >> missed_updates;
3841 while (missed_updates) {
3842 if (missed_updates % 2)
3843 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3845 missed_updates >>= 1;
3852 * Update rq->cpu_load[] statistics. This function is usually called every
3853 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3854 * every tick. We fix it up based on jiffies.
3856 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
3857 unsigned long pending_updates)
3861 this_rq->nr_load_updates++;
3863 /* Update our load: */
3864 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3865 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3866 unsigned long old_load, new_load;
3868 /* scale is effectively 1 << i now, and >> i divides by scale */
3870 old_load = this_rq->cpu_load[i];
3871 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3872 new_load = this_load;
3874 * Round up the averaging division if load is increasing. This
3875 * prevents us from getting stuck on 9 if the load is 10, for
3878 if (new_load > old_load)
3879 new_load += scale - 1;
3881 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3884 sched_avg_update(this_rq);
3889 * There is no sane way to deal with nohz on smp when using jiffies because the
3890 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
3891 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
3893 * Therefore we cannot use the delta approach from the regular tick since that
3894 * would seriously skew the load calculation. However we'll make do for those
3895 * updates happening while idle (nohz_idle_balance) or coming out of idle
3896 * (tick_nohz_idle_exit).
3898 * This means we might still be one tick off for nohz periods.
3902 * Called from nohz_idle_balance() to update the load ratings before doing the
3905 static void update_idle_cpu_load(struct rq *this_rq)
3907 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3908 unsigned long load = this_rq->load.weight;
3909 unsigned long pending_updates;
3912 * bail if there's load or we're actually up-to-date.
3914 if (load || curr_jiffies == this_rq->last_load_update_tick)
3917 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3918 this_rq->last_load_update_tick = curr_jiffies;
3920 __update_cpu_load(this_rq, load, pending_updates);
3924 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
3926 void update_cpu_load_nohz(void)
3928 struct rq *this_rq = this_rq();
3929 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3930 unsigned long pending_updates;
3932 if (curr_jiffies == this_rq->last_load_update_tick)
3935 raw_spin_lock(&this_rq->lock);
3936 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3937 if (pending_updates) {
3938 this_rq->last_load_update_tick = curr_jiffies;
3940 * We were idle, this means load 0, the current load might be
3941 * !0 due to remote wakeups and the sort.
3943 __update_cpu_load(this_rq, 0, pending_updates);
3945 raw_spin_unlock(&this_rq->lock);
3947 #endif /* CONFIG_NO_HZ */
3950 * Called from scheduler_tick()
3952 static void update_cpu_load_active(struct rq *this_rq)
3955 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
3957 this_rq->last_load_update_tick = jiffies;
3958 __update_cpu_load(this_rq, this_rq->load.weight, 1);
3960 calc_load_account_active(this_rq);
3966 * sched_exec - execve() is a valuable balancing opportunity, because at
3967 * this point the task has the smallest effective memory and cache footprint.
3969 void sched_exec(void)
3971 struct task_struct *p = current;
3972 unsigned long flags;
3975 raw_spin_lock_irqsave(&p->pi_lock, flags);
3976 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3977 if (dest_cpu == smp_processor_id())
3980 if (likely(cpu_active(dest_cpu))) {
3981 struct migration_arg arg = { p, dest_cpu };
3983 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3984 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3988 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3993 DEFINE_PER_CPU(struct kernel_stat, kstat);
3995 EXPORT_PER_CPU_SYMBOL(kstat);
3998 * Return any ns on the sched_clock that have not yet been accounted in
3999 * @p in case that task is currently running.
4001 * Called with task_rq_lock() held on @rq.
4003 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4007 if (task_current(rq, p)) {
4008 update_rq_clock(rq);
4009 ns = rq->clock_task - p->se.exec_start;
4017 unsigned long long task_delta_exec(struct task_struct *p)
4019 unsigned long flags;
4023 rq = task_rq_lock(p, &flags);
4024 ns = do_task_delta_exec(p, rq);
4025 task_rq_unlock(rq, p, &flags);
4031 * Return accounted runtime for the task.
4032 * In case the task is currently running, return the runtime plus current's
4033 * pending runtime that have not been accounted yet.
4035 unsigned long long task_sched_runtime(struct task_struct *p)
4037 unsigned long flags;
4041 rq = task_rq_lock(p, &flags);
4042 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4043 task_rq_unlock(rq, p, &flags);
4049 * Account user cpu time to a process.
4050 * @p: the process that the cpu time gets accounted to
4051 * @cputime: the cpu time spent in user space since the last update
4052 * @cputime_scaled: cputime scaled by cpu frequency
4054 void account_user_time(struct task_struct *p, cputime_t cputime,
4055 cputime_t cputime_scaled)
4057 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4060 /* Add user time to process. */
4061 p->utime = cputime_add(p->utime, cputime);
4062 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4063 account_group_user_time(p, cputime);
4065 /* Add user time to cpustat. */
4066 tmp = cputime_to_cputime64(cputime);
4067 if (TASK_NICE(p) > 0)
4068 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4070 cpustat->user = cputime64_add(cpustat->user, tmp);
4072 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4073 /* Account for user time used */
4074 acct_update_integrals(p);
4078 * Account guest cpu time to a process.
4079 * @p: the process that the cpu time gets accounted to
4080 * @cputime: the cpu time spent in virtual machine since the last update
4081 * @cputime_scaled: cputime scaled by cpu frequency
4083 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4084 cputime_t cputime_scaled)
4087 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4089 tmp = cputime_to_cputime64(cputime);
4091 /* Add guest time to process. */
4092 p->utime = cputime_add(p->utime, cputime);
4093 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4094 account_group_user_time(p, cputime);
4095 p->gtime = cputime_add(p->gtime, cputime);
4097 /* Add guest time to cpustat. */
4098 if (TASK_NICE(p) > 0) {
4099 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4100 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
4102 cpustat->user = cputime64_add(cpustat->user, tmp);
4103 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4108 * Account system cpu time to a process and desired cpustat field
4109 * @p: the process that the cpu time gets accounted to
4110 * @cputime: the cpu time spent in kernel space since the last update
4111 * @cputime_scaled: cputime scaled by cpu frequency
4112 * @target_cputime64: pointer to cpustat field that has to be updated
4115 void __account_system_time(struct task_struct *p, cputime_t cputime,
4116 cputime_t cputime_scaled, cputime64_t *target_cputime64)
4118 cputime64_t tmp = cputime_to_cputime64(cputime);
4120 /* Add system time to process. */
4121 p->stime = cputime_add(p->stime, cputime);
4122 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4123 account_group_system_time(p, cputime);
4125 /* Add system time to cpustat. */
4126 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
4127 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4129 /* Account for system time used */
4130 acct_update_integrals(p);
4134 * Account system cpu time to a process.
4135 * @p: the process that the cpu time gets accounted to
4136 * @hardirq_offset: the offset to subtract from hardirq_count()
4137 * @cputime: the cpu time spent in kernel space since the last update
4138 * @cputime_scaled: cputime scaled by cpu frequency
4140 void account_system_time(struct task_struct *p, int hardirq_offset,
4141 cputime_t cputime, cputime_t cputime_scaled)
4143 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4144 cputime64_t *target_cputime64;
4146 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4147 account_guest_time(p, cputime, cputime_scaled);
4151 if (hardirq_count() - hardirq_offset)
4152 target_cputime64 = &cpustat->irq;
4153 else if (in_serving_softirq())
4154 target_cputime64 = &cpustat->softirq;
4156 target_cputime64 = &cpustat->system;
4158 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
4162 * Account for involuntary wait time.
4163 * @cputime: the cpu time spent in involuntary wait
4165 void account_steal_time(cputime_t cputime)
4167 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4168 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4170 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4174 * Account for idle time.
4175 * @cputime: the cpu time spent in idle wait
4177 void account_idle_time(cputime_t cputime)
4179 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4180 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4181 struct rq *rq = this_rq();
4183 if (atomic_read(&rq->nr_iowait) > 0)
4184 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4186 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4189 static __always_inline bool steal_account_process_tick(void)
4191 #ifdef CONFIG_PARAVIRT
4192 if (static_branch(¶virt_steal_enabled)) {
4195 steal = paravirt_steal_clock(smp_processor_id());
4196 steal -= this_rq()->prev_steal_time;
4198 st = steal_ticks(steal);
4199 this_rq()->prev_steal_time += st * TICK_NSEC;
4201 account_steal_time(st);
4208 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4210 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4212 * Account a tick to a process and cpustat
4213 * @p: the process that the cpu time gets accounted to
4214 * @user_tick: is the tick from userspace
4215 * @rq: the pointer to rq
4217 * Tick demultiplexing follows the order
4218 * - pending hardirq update
4219 * - pending softirq update
4223 * - check for guest_time
4224 * - else account as system_time
4226 * Check for hardirq is done both for system and user time as there is
4227 * no timer going off while we are on hardirq and hence we may never get an
4228 * opportunity to update it solely in system time.
4229 * p->stime and friends are only updated on system time and not on irq
4230 * softirq as those do not count in task exec_runtime any more.
4232 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4235 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4236 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
4237 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4239 if (steal_account_process_tick())
4242 if (irqtime_account_hi_update()) {
4243 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4244 } else if (irqtime_account_si_update()) {
4245 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4246 } else if (this_cpu_ksoftirqd() == p) {
4248 * ksoftirqd time do not get accounted in cpu_softirq_time.
4249 * So, we have to handle it separately here.
4250 * Also, p->stime needs to be updated for ksoftirqd.
4252 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4254 } else if (user_tick) {
4255 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4256 } else if (p == rq->idle) {
4257 account_idle_time(cputime_one_jiffy);
4258 } else if (p->flags & PF_VCPU) { /* System time or guest time */
4259 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
4261 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4266 static void irqtime_account_idle_ticks(int ticks)
4269 struct rq *rq = this_rq();
4271 for (i = 0; i < ticks; i++)
4272 irqtime_account_process_tick(current, 0, rq);
4274 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4275 static void irqtime_account_idle_ticks(int ticks) {}
4276 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4278 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4281 * Account a single tick of cpu time.
4282 * @p: the process that the cpu time gets accounted to
4283 * @user_tick: indicates if the tick is a user or a system tick
4285 void account_process_tick(struct task_struct *p, int user_tick)
4287 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4288 struct rq *rq = this_rq();
4290 if (sched_clock_irqtime) {
4291 irqtime_account_process_tick(p, user_tick, rq);
4295 if (steal_account_process_tick())
4299 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4300 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4301 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4304 account_idle_time(cputime_one_jiffy);
4308 * Account multiple ticks of steal time.
4309 * @p: the process from which the cpu time has been stolen
4310 * @ticks: number of stolen ticks
4312 void account_steal_ticks(unsigned long ticks)
4314 account_steal_time(jiffies_to_cputime(ticks));
4318 * Account multiple ticks of idle time.
4319 * @ticks: number of stolen ticks
4321 void account_idle_ticks(unsigned long ticks)
4324 if (sched_clock_irqtime) {
4325 irqtime_account_idle_ticks(ticks);
4329 account_idle_time(jiffies_to_cputime(ticks));
4335 * Use precise platform statistics if available:
4337 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4338 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4344 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4346 struct task_cputime cputime;
4348 thread_group_cputime(p, &cputime);
4350 *ut = cputime.utime;
4351 *st = cputime.stime;
4355 #ifndef nsecs_to_cputime
4356 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4359 static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total)
4361 u64 temp = (__force u64) rtime;
4363 temp *= (__force u64) utime;
4365 if (sizeof(cputime_t) == 4)
4366 temp = div_u64(temp, (__force u32) total);
4368 temp = div64_u64(temp, (__force u64) total);
4370 return (__force cputime_t) temp;
4373 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4375 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4378 * Use CFS's precise accounting:
4380 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4383 utime = scale_utime(utime, rtime, total);
4388 * Compare with previous values, to keep monotonicity:
4390 p->prev_utime = max(p->prev_utime, utime);
4391 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4393 *ut = p->prev_utime;
4394 *st = p->prev_stime;
4398 * Must be called with siglock held.
4400 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4402 struct signal_struct *sig = p->signal;
4403 struct task_cputime cputime;
4404 cputime_t rtime, utime, total;
4406 thread_group_cputime(p, &cputime);
4408 total = cputime_add(cputime.utime, cputime.stime);
4409 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4412 utime = scale_utime(cputime.utime, rtime, total);
4416 sig->prev_utime = max(sig->prev_utime, utime);
4417 sig->prev_stime = max(sig->prev_stime,
4418 cputime_sub(rtime, sig->prev_utime));
4420 *ut = sig->prev_utime;
4421 *st = sig->prev_stime;
4426 * This function gets called by the timer code, with HZ frequency.
4427 * We call it with interrupts disabled.
4429 void scheduler_tick(void)
4431 int cpu = smp_processor_id();
4432 struct rq *rq = cpu_rq(cpu);
4433 struct task_struct *curr = rq->curr;
4437 raw_spin_lock(&rq->lock);
4438 update_rq_clock(rq);
4439 update_cpu_load_active(rq);
4440 curr->sched_class->task_tick(rq, curr, 0);
4441 raw_spin_unlock(&rq->lock);
4443 perf_event_task_tick();
4446 rq->idle_balance = idle_cpu(cpu);
4447 trigger_load_balance(rq, cpu);
4451 notrace unsigned long get_parent_ip(unsigned long addr)
4453 if (in_lock_functions(addr)) {
4454 addr = CALLER_ADDR2;
4455 if (in_lock_functions(addr))
4456 addr = CALLER_ADDR3;
4461 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4462 defined(CONFIG_PREEMPT_TRACER))
4464 void __kprobes add_preempt_count(int val)
4466 #ifdef CONFIG_DEBUG_PREEMPT
4470 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4473 preempt_count() += val;
4474 #ifdef CONFIG_DEBUG_PREEMPT
4476 * Spinlock count overflowing soon?
4478 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4481 if (preempt_count() == val)
4482 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4484 EXPORT_SYMBOL(add_preempt_count);
4486 void __kprobes sub_preempt_count(int val)
4488 #ifdef CONFIG_DEBUG_PREEMPT
4492 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4495 * Is the spinlock portion underflowing?
4497 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4498 !(preempt_count() & PREEMPT_MASK)))
4502 if (preempt_count() == val)
4503 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4504 preempt_count() -= val;
4506 EXPORT_SYMBOL(sub_preempt_count);
4511 * Print scheduling while atomic bug:
4513 static noinline void __schedule_bug(struct task_struct *prev)
4515 struct pt_regs *regs = get_irq_regs();
4517 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4518 prev->comm, prev->pid, preempt_count());
4520 debug_show_held_locks(prev);
4522 if (irqs_disabled())
4523 print_irqtrace_events(prev);
4532 * Various schedule()-time debugging checks and statistics:
4534 static inline void schedule_debug(struct task_struct *prev)
4537 * Test if we are atomic. Since do_exit() needs to call into
4538 * schedule() atomically, we ignore that path for now.
4539 * Otherwise, whine if we are scheduling when we should not be.
4541 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4542 __schedule_bug(prev);
4545 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4547 schedstat_inc(this_rq(), sched_count);
4550 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4552 if (prev->on_rq || rq->skip_clock_update < 0)
4553 update_rq_clock(rq);
4554 prev->sched_class->put_prev_task(rq, prev);
4558 * Pick up the highest-prio task:
4560 static inline struct task_struct *
4561 pick_next_task(struct rq *rq)
4563 const struct sched_class *class;
4564 struct task_struct *p;
4567 * Optimization: we know that if all tasks are in
4568 * the fair class we can call that function directly:
4570 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4571 p = fair_sched_class.pick_next_task(rq);
4576 for_each_class(class) {
4577 p = class->pick_next_task(rq);
4582 BUG(); /* the idle class will always have a runnable task */
4586 * __schedule() is the main scheduler function.
4588 static void __sched __schedule(void)
4590 struct task_struct *prev, *next;
4591 unsigned long *switch_count;
4597 cpu = smp_processor_id();
4599 rcu_note_context_switch(cpu);
4602 schedule_debug(prev);
4604 if (sched_feat(HRTICK))
4607 raw_spin_lock_irq(&rq->lock);
4609 switch_count = &prev->nivcsw;
4610 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4611 if (unlikely(signal_pending_state(prev->state, prev))) {
4612 prev->state = TASK_RUNNING;
4614 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4618 * If a worker went to sleep, notify and ask workqueue
4619 * whether it wants to wake up a task to maintain
4622 if (prev->flags & PF_WQ_WORKER) {
4623 struct task_struct *to_wakeup;
4625 to_wakeup = wq_worker_sleeping(prev, cpu);
4627 try_to_wake_up_local(to_wakeup);
4630 switch_count = &prev->nvcsw;
4633 pre_schedule(rq, prev);
4635 if (unlikely(!rq->nr_running))
4636 idle_balance(cpu, rq);
4638 put_prev_task(rq, prev);
4639 next = pick_next_task(rq);
4640 clear_tsk_need_resched(prev);
4641 rq->skip_clock_update = 0;
4643 if (likely(prev != next)) {
4648 context_switch(rq, prev, next); /* unlocks the rq */
4650 * The context switch have flipped the stack from under us
4651 * and restored the local variables which were saved when
4652 * this task called schedule() in the past. prev == current
4653 * is still correct, but it can be moved to another cpu/rq.
4655 cpu = smp_processor_id();
4658 raw_spin_unlock_irq(&rq->lock);
4662 preempt_enable_no_resched();
4667 static inline void sched_submit_work(struct task_struct *tsk)
4672 * If we are going to sleep and we have plugged IO queued,
4673 * make sure to submit it to avoid deadlocks.
4675 if (blk_needs_flush_plug(tsk))
4676 blk_schedule_flush_plug(tsk);
4679 asmlinkage void __sched schedule(void)
4681 struct task_struct *tsk = current;
4683 sched_submit_work(tsk);
4686 EXPORT_SYMBOL(schedule);
4688 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4690 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4692 if (lock->owner != owner)
4696 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4697 * lock->owner still matches owner, if that fails, owner might
4698 * point to free()d memory, if it still matches, the rcu_read_lock()
4699 * ensures the memory stays valid.
4703 return owner->on_cpu;
4707 * Look out! "owner" is an entirely speculative pointer
4708 * access and not reliable.
4710 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4712 if (!sched_feat(OWNER_SPIN))
4716 while (owner_running(lock, owner)) {
4720 arch_mutex_cpu_relax();
4725 * We break out the loop above on need_resched() and when the
4726 * owner changed, which is a sign for heavy contention. Return
4727 * success only when lock->owner is NULL.
4729 return lock->owner == NULL;
4733 #ifdef CONFIG_PREEMPT
4735 * this is the entry point to schedule() from in-kernel preemption
4736 * off of preempt_enable. Kernel preemptions off return from interrupt
4737 * occur there and call schedule directly.
4739 asmlinkage void __sched notrace preempt_schedule(void)
4741 struct thread_info *ti = current_thread_info();
4744 * If there is a non-zero preempt_count or interrupts are disabled,
4745 * we do not want to preempt the current task. Just return..
4747 if (likely(ti->preempt_count || irqs_disabled()))
4751 add_preempt_count_notrace(PREEMPT_ACTIVE);
4753 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4756 * Check again in case we missed a preemption opportunity
4757 * between schedule and now.
4760 } while (need_resched());
4762 EXPORT_SYMBOL(preempt_schedule);
4765 * this is the entry point to schedule() from kernel preemption
4766 * off of irq context.
4767 * Note, that this is called and return with irqs disabled. This will
4768 * protect us against recursive calling from irq.
4770 asmlinkage void __sched preempt_schedule_irq(void)
4772 struct thread_info *ti = current_thread_info();
4774 /* Catch callers which need to be fixed */
4775 BUG_ON(ti->preempt_count || !irqs_disabled());
4778 add_preempt_count(PREEMPT_ACTIVE);
4781 local_irq_disable();
4782 sub_preempt_count(PREEMPT_ACTIVE);
4785 * Check again in case we missed a preemption opportunity
4786 * between schedule and now.
4789 } while (need_resched());
4792 #endif /* CONFIG_PREEMPT */
4794 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4797 return try_to_wake_up(curr->private, mode, wake_flags);
4799 EXPORT_SYMBOL(default_wake_function);
4802 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4803 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4804 * number) then we wake all the non-exclusive tasks and one exclusive task.
4806 * There are circumstances in which we can try to wake a task which has already
4807 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4808 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4810 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4811 int nr_exclusive, int wake_flags, void *key)
4813 wait_queue_t *curr, *next;
4815 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4816 unsigned flags = curr->flags;
4818 if (curr->func(curr, mode, wake_flags, key) &&
4819 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4825 * __wake_up - wake up threads blocked on a waitqueue.
4827 * @mode: which threads
4828 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4829 * @key: is directly passed to the wakeup function
4831 * It may be assumed that this function implies a write memory barrier before
4832 * changing the task state if and only if any tasks are woken up.
4834 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4835 int nr_exclusive, void *key)
4837 unsigned long flags;
4839 spin_lock_irqsave(&q->lock, flags);
4840 __wake_up_common(q, mode, nr_exclusive, 0, key);
4841 spin_unlock_irqrestore(&q->lock, flags);
4843 EXPORT_SYMBOL(__wake_up);
4846 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4848 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4850 __wake_up_common(q, mode, 1, 0, NULL);
4852 EXPORT_SYMBOL_GPL(__wake_up_locked);
4854 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4856 __wake_up_common(q, mode, 1, 0, key);
4858 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4861 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4863 * @mode: which threads
4864 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4865 * @key: opaque value to be passed to wakeup targets
4867 * The sync wakeup differs that the waker knows that it will schedule
4868 * away soon, so while the target thread will be woken up, it will not
4869 * be migrated to another CPU - ie. the two threads are 'synchronized'
4870 * with each other. This can prevent needless bouncing between CPUs.
4872 * On UP it can prevent extra preemption.
4874 * It may be assumed that this function implies a write memory barrier before
4875 * changing the task state if and only if any tasks are woken up.
4877 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4878 int nr_exclusive, void *key)
4880 unsigned long flags;
4881 int wake_flags = WF_SYNC;
4886 if (unlikely(!nr_exclusive))
4889 spin_lock_irqsave(&q->lock, flags);
4890 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4891 spin_unlock_irqrestore(&q->lock, flags);
4893 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4896 * __wake_up_sync - see __wake_up_sync_key()
4898 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4900 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4902 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4905 * complete: - signals a single thread waiting on this completion
4906 * @x: holds the state of this particular completion
4908 * This will wake up a single thread waiting on this completion. Threads will be
4909 * awakened in the same order in which they were queued.
4911 * See also complete_all(), wait_for_completion() and related routines.
4913 * It may be assumed that this function implies a write memory barrier before
4914 * changing the task state if and only if any tasks are woken up.
4916 void complete(struct completion *x)
4918 unsigned long flags;
4920 spin_lock_irqsave(&x->wait.lock, flags);
4922 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4923 spin_unlock_irqrestore(&x->wait.lock, flags);
4925 EXPORT_SYMBOL(complete);
4928 * complete_all: - signals all threads waiting on this completion
4929 * @x: holds the state of this particular completion
4931 * This will wake up all threads waiting on this particular completion event.
4933 * It may be assumed that this function implies a write memory barrier before
4934 * changing the task state if and only if any tasks are woken up.
4936 void complete_all(struct completion *x)
4938 unsigned long flags;
4940 spin_lock_irqsave(&x->wait.lock, flags);
4941 x->done += UINT_MAX/2;
4942 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4943 spin_unlock_irqrestore(&x->wait.lock, flags);
4945 EXPORT_SYMBOL(complete_all);
4947 static inline long __sched
4948 do_wait_for_common(struct completion *x, long timeout, int state)
4951 DECLARE_WAITQUEUE(wait, current);
4953 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4955 if (signal_pending_state(state, current)) {
4956 timeout = -ERESTARTSYS;
4959 __set_current_state(state);
4960 spin_unlock_irq(&x->wait.lock);
4961 timeout = schedule_timeout(timeout);
4962 spin_lock_irq(&x->wait.lock);
4963 } while (!x->done && timeout);
4964 __remove_wait_queue(&x->wait, &wait);
4969 return timeout ?: 1;
4973 wait_for_common(struct completion *x, long timeout, int state)
4977 spin_lock_irq(&x->wait.lock);
4978 timeout = do_wait_for_common(x, timeout, state);
4979 spin_unlock_irq(&x->wait.lock);
4984 * wait_for_completion: - waits for completion of a task
4985 * @x: holds the state of this particular completion
4987 * This waits to be signaled for completion of a specific task. It is NOT
4988 * interruptible and there is no timeout.
4990 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4991 * and interrupt capability. Also see complete().
4993 void __sched wait_for_completion(struct completion *x)
4995 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4997 EXPORT_SYMBOL(wait_for_completion);
5000 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5001 * @x: holds the state of this particular completion
5002 * @timeout: timeout value in jiffies
5004 * This waits for either a completion of a specific task to be signaled or for a
5005 * specified timeout to expire. The timeout is in jiffies. It is not
5008 * The return value is 0 if timed out, and positive (at least 1, or number of
5009 * jiffies left till timeout) if completed.
5011 unsigned long __sched
5012 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5014 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5016 EXPORT_SYMBOL(wait_for_completion_timeout);
5019 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5020 * @x: holds the state of this particular completion
5022 * This waits for completion of a specific task to be signaled. It is
5025 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
5027 int __sched wait_for_completion_interruptible(struct completion *x)
5029 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5030 if (t == -ERESTARTSYS)
5034 EXPORT_SYMBOL(wait_for_completion_interruptible);
5037 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5038 * @x: holds the state of this particular completion
5039 * @timeout: timeout value in jiffies
5041 * This waits for either a completion of a specific task to be signaled or for a
5042 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5044 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
5045 * positive (at least 1, or number of jiffies left till timeout) if completed.
5048 wait_for_completion_interruptible_timeout(struct completion *x,
5049 unsigned long timeout)
5051 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5053 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5056 * wait_for_completion_killable: - waits for completion of a task (killable)
5057 * @x: holds the state of this particular completion
5059 * This waits to be signaled for completion of a specific task. It can be
5060 * interrupted by a kill signal.
5062 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
5064 int __sched wait_for_completion_killable(struct completion *x)
5066 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5067 if (t == -ERESTARTSYS)
5071 EXPORT_SYMBOL(wait_for_completion_killable);
5074 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
5075 * @x: holds the state of this particular completion
5076 * @timeout: timeout value in jiffies
5078 * This waits for either a completion of a specific task to be
5079 * signaled or for a specified timeout to expire. It can be
5080 * interrupted by a kill signal. The timeout is in jiffies.
5082 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
5083 * positive (at least 1, or number of jiffies left till timeout) if completed.
5086 wait_for_completion_killable_timeout(struct completion *x,
5087 unsigned long timeout)
5089 return wait_for_common(x, timeout, TASK_KILLABLE);
5091 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
5094 * try_wait_for_completion - try to decrement a completion without blocking
5095 * @x: completion structure
5097 * Returns: 0 if a decrement cannot be done without blocking
5098 * 1 if a decrement succeeded.
5100 * If a completion is being used as a counting completion,
5101 * attempt to decrement the counter without blocking. This
5102 * enables us to avoid waiting if the resource the completion
5103 * is protecting is not available.
5105 bool try_wait_for_completion(struct completion *x)
5107 unsigned long flags;
5110 spin_lock_irqsave(&x->wait.lock, flags);
5115 spin_unlock_irqrestore(&x->wait.lock, flags);
5118 EXPORT_SYMBOL(try_wait_for_completion);
5121 * completion_done - Test to see if a completion has any waiters
5122 * @x: completion structure
5124 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5125 * 1 if there are no waiters.
5128 bool completion_done(struct completion *x)
5130 unsigned long flags;
5133 spin_lock_irqsave(&x->wait.lock, flags);
5136 spin_unlock_irqrestore(&x->wait.lock, flags);
5139 EXPORT_SYMBOL(completion_done);
5142 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5144 unsigned long flags;
5147 init_waitqueue_entry(&wait, current);
5149 __set_current_state(state);
5151 spin_lock_irqsave(&q->lock, flags);
5152 __add_wait_queue(q, &wait);
5153 spin_unlock(&q->lock);
5154 timeout = schedule_timeout(timeout);
5155 spin_lock_irq(&q->lock);
5156 __remove_wait_queue(q, &wait);
5157 spin_unlock_irqrestore(&q->lock, flags);
5162 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5164 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5166 EXPORT_SYMBOL(interruptible_sleep_on);
5169 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5171 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5173 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5175 void __sched sleep_on(wait_queue_head_t *q)
5177 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5179 EXPORT_SYMBOL(sleep_on);
5181 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5183 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5185 EXPORT_SYMBOL(sleep_on_timeout);
5187 #ifdef CONFIG_RT_MUTEXES
5190 * rt_mutex_setprio - set the current priority of a task
5192 * @prio: prio value (kernel-internal form)
5194 * This function changes the 'effective' priority of a task. It does
5195 * not touch ->normal_prio like __setscheduler().
5197 * Used by the rt_mutex code to implement priority inheritance logic.
5199 void rt_mutex_setprio(struct task_struct *p, int prio)
5201 int oldprio, on_rq, running;
5203 const struct sched_class *prev_class;
5205 BUG_ON(prio < 0 || prio > MAX_PRIO);
5207 rq = __task_rq_lock(p);
5209 trace_sched_pi_setprio(p, prio);
5211 prev_class = p->sched_class;
5213 running = task_current(rq, p);
5215 dequeue_task(rq, p, 0);
5217 p->sched_class->put_prev_task(rq, p);
5220 p->sched_class = &rt_sched_class;
5222 p->sched_class = &fair_sched_class;
5227 p->sched_class->set_curr_task(rq);
5229 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
5231 check_class_changed(rq, p, prev_class, oldprio);
5232 __task_rq_unlock(rq);
5237 void set_user_nice(struct task_struct *p, long nice)
5239 int old_prio, delta, on_rq;
5240 unsigned long flags;
5243 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5246 * We have to be careful, if called from sys_setpriority(),
5247 * the task might be in the middle of scheduling on another CPU.
5249 rq = task_rq_lock(p, &flags);
5251 * The RT priorities are set via sched_setscheduler(), but we still
5252 * allow the 'normal' nice value to be set - but as expected
5253 * it wont have any effect on scheduling until the task is
5254 * SCHED_FIFO/SCHED_RR:
5256 if (task_has_rt_policy(p)) {
5257 p->static_prio = NICE_TO_PRIO(nice);
5262 dequeue_task(rq, p, 0);
5264 p->static_prio = NICE_TO_PRIO(nice);
5267 p->prio = effective_prio(p);
5268 delta = p->prio - old_prio;
5271 enqueue_task(rq, p, 0);
5273 * If the task increased its priority or is running and
5274 * lowered its priority, then reschedule its CPU:
5276 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5277 resched_task(rq->curr);
5280 task_rq_unlock(rq, p, &flags);
5282 EXPORT_SYMBOL(set_user_nice);
5285 * can_nice - check if a task can reduce its nice value
5289 int can_nice(const struct task_struct *p, const int nice)
5291 /* convert nice value [19,-20] to rlimit style value [1,40] */
5292 int nice_rlim = 20 - nice;
5294 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5295 capable(CAP_SYS_NICE));
5297 EXPORT_SYMBOL_GPL(can_nice);
5299 #ifdef __ARCH_WANT_SYS_NICE
5302 * sys_nice - change the priority of the current process.
5303 * @increment: priority increment
5305 * sys_setpriority is a more generic, but much slower function that
5306 * does similar things.
5308 SYSCALL_DEFINE1(nice, int, increment)
5313 * Setpriority might change our priority at the same moment.
5314 * We don't have to worry. Conceptually one call occurs first
5315 * and we have a single winner.
5317 if (increment < -40)
5322 nice = TASK_NICE(current) + increment;
5328 if (increment < 0 && !can_nice(current, nice))
5331 retval = security_task_setnice(current, nice);
5335 set_user_nice(current, nice);
5342 * task_prio - return the priority value of a given task.
5343 * @p: the task in question.
5345 * This is the priority value as seen by users in /proc.
5346 * RT tasks are offset by -200. Normal tasks are centered
5347 * around 0, value goes from -16 to +15.
5349 int task_prio(const struct task_struct *p)
5351 return p->prio - MAX_RT_PRIO;
5355 * task_nice - return the nice value of a given task.
5356 * @p: the task in question.
5358 int task_nice(const struct task_struct *p)
5360 return TASK_NICE(p);
5362 EXPORT_SYMBOL(task_nice);
5365 * idle_cpu - is a given cpu idle currently?
5366 * @cpu: the processor in question.
5368 int idle_cpu(int cpu)
5370 struct rq *rq = cpu_rq(cpu);
5372 if (rq->curr != rq->idle)
5379 if (!llist_empty(&rq->wake_list))
5387 * idle_task - return the idle task for a given cpu.
5388 * @cpu: the processor in question.
5390 struct task_struct *idle_task(int cpu)
5392 return cpu_rq(cpu)->idle;
5396 * find_process_by_pid - find a process with a matching PID value.
5397 * @pid: the pid in question.
5399 static struct task_struct *find_process_by_pid(pid_t pid)
5401 return pid ? find_task_by_vpid(pid) : current;
5404 /* Actually do priority change: must hold rq lock. */
5406 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5409 p->rt_priority = prio;
5410 p->normal_prio = normal_prio(p);
5411 /* we are holding p->pi_lock already */
5412 p->prio = rt_mutex_getprio(p);
5413 if (rt_prio(p->prio))
5414 p->sched_class = &rt_sched_class;
5416 p->sched_class = &fair_sched_class;
5421 * check the target process has a UID that matches the current process's
5423 static bool check_same_owner(struct task_struct *p)
5425 const struct cred *cred = current_cred(), *pcred;
5429 pcred = __task_cred(p);
5430 if (cred->user->user_ns == pcred->user->user_ns)
5431 match = (cred->euid == pcred->euid ||
5432 cred->euid == pcred->uid);
5439 static int __sched_setscheduler(struct task_struct *p, int policy,
5440 const struct sched_param *param, bool user)
5442 int retval, oldprio, oldpolicy = -1, on_rq, running;
5443 unsigned long flags;
5444 const struct sched_class *prev_class;
5448 /* may grab non-irq protected spin_locks */
5449 BUG_ON(in_interrupt());
5451 /* double check policy once rq lock held */
5453 reset_on_fork = p->sched_reset_on_fork;
5454 policy = oldpolicy = p->policy;
5456 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5457 policy &= ~SCHED_RESET_ON_FORK;
5459 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5460 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5461 policy != SCHED_IDLE)
5466 * Valid priorities for SCHED_FIFO and SCHED_RR are
5467 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5468 * SCHED_BATCH and SCHED_IDLE is 0.
5470 if (param->sched_priority < 0 ||
5471 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5472 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5474 if (rt_policy(policy) != (param->sched_priority != 0))
5478 * Allow unprivileged RT tasks to decrease priority:
5480 if (user && !capable(CAP_SYS_NICE)) {
5481 if (rt_policy(policy)) {
5482 unsigned long rlim_rtprio =
5483 task_rlimit(p, RLIMIT_RTPRIO);
5485 /* can't set/change the rt policy */
5486 if (policy != p->policy && !rlim_rtprio)
5489 /* can't increase priority */
5490 if (param->sched_priority > p->rt_priority &&
5491 param->sched_priority > rlim_rtprio)
5496 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5497 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5499 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5500 if (!can_nice(p, TASK_NICE(p)))
5504 /* can't change other user's priorities */
5505 if (!check_same_owner(p))
5508 /* Normal users shall not reset the sched_reset_on_fork flag */
5509 if (p->sched_reset_on_fork && !reset_on_fork)
5514 retval = security_task_setscheduler(p);
5520 * make sure no PI-waiters arrive (or leave) while we are
5521 * changing the priority of the task:
5523 * To be able to change p->policy safely, the appropriate
5524 * runqueue lock must be held.
5526 rq = task_rq_lock(p, &flags);
5529 * Changing the policy of the stop threads its a very bad idea
5531 if (p == rq->stop) {
5532 task_rq_unlock(rq, p, &flags);
5537 * If not changing anything there's no need to proceed further:
5539 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5540 param->sched_priority == p->rt_priority))) {
5542 __task_rq_unlock(rq);
5543 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5547 #ifdef CONFIG_RT_GROUP_SCHED
5550 * Do not allow realtime tasks into groups that have no runtime
5553 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5554 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5555 !task_group_is_autogroup(task_group(p))) {
5556 task_rq_unlock(rq, p, &flags);
5562 /* recheck policy now with rq lock held */
5563 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5564 policy = oldpolicy = -1;
5565 task_rq_unlock(rq, p, &flags);
5569 running = task_current(rq, p);
5571 deactivate_task(rq, p, 0);
5573 p->sched_class->put_prev_task(rq, p);
5575 p->sched_reset_on_fork = reset_on_fork;
5578 prev_class = p->sched_class;
5579 __setscheduler(rq, p, policy, param->sched_priority);
5582 p->sched_class->set_curr_task(rq);
5584 activate_task(rq, p, 0);
5586 check_class_changed(rq, p, prev_class, oldprio);
5587 task_rq_unlock(rq, p, &flags);
5589 rt_mutex_adjust_pi(p);
5595 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5596 * @p: the task in question.
5597 * @policy: new policy.
5598 * @param: structure containing the new RT priority.
5600 * NOTE that the task may be already dead.
5602 int sched_setscheduler(struct task_struct *p, int policy,
5603 const struct sched_param *param)
5605 return __sched_setscheduler(p, policy, param, true);
5607 EXPORT_SYMBOL_GPL(sched_setscheduler);
5610 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5611 * @p: the task in question.
5612 * @policy: new policy.
5613 * @param: structure containing the new RT priority.
5615 * Just like sched_setscheduler, only don't bother checking if the
5616 * current context has permission. For example, this is needed in
5617 * stop_machine(): we create temporary high priority worker threads,
5618 * but our caller might not have that capability.
5620 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5621 const struct sched_param *param)
5623 return __sched_setscheduler(p, policy, param, false);
5627 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5629 struct sched_param lparam;
5630 struct task_struct *p;
5633 if (!param || pid < 0)
5635 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5640 p = find_process_by_pid(pid);
5642 retval = sched_setscheduler(p, policy, &lparam);
5649 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5650 * @pid: the pid in question.
5651 * @policy: new policy.
5652 * @param: structure containing the new RT priority.
5654 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5655 struct sched_param __user *, param)
5657 /* negative values for policy are not valid */
5661 return do_sched_setscheduler(pid, policy, param);
5665 * sys_sched_setparam - set/change the RT priority of a thread
5666 * @pid: the pid in question.
5667 * @param: structure containing the new RT priority.
5669 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5671 return do_sched_setscheduler(pid, -1, param);
5675 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5676 * @pid: the pid in question.
5678 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5680 struct task_struct *p;
5688 p = find_process_by_pid(pid);
5690 retval = security_task_getscheduler(p);
5693 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5700 * sys_sched_getparam - get the RT priority of a thread
5701 * @pid: the pid in question.
5702 * @param: structure containing the RT priority.
5704 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5706 struct sched_param lp;
5707 struct task_struct *p;
5710 if (!param || pid < 0)
5714 p = find_process_by_pid(pid);
5719 retval = security_task_getscheduler(p);
5723 lp.sched_priority = p->rt_priority;
5727 * This one might sleep, we cannot do it with a spinlock held ...
5729 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5738 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5740 cpumask_var_t cpus_allowed, new_mask;
5741 struct task_struct *p;
5747 p = find_process_by_pid(pid);
5754 /* Prevent p going away */
5758 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5762 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5764 goto out_free_cpus_allowed;
5767 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5770 retval = security_task_setscheduler(p);
5774 cpuset_cpus_allowed(p, cpus_allowed);
5775 cpumask_and(new_mask, in_mask, cpus_allowed);
5777 retval = set_cpus_allowed_ptr(p, new_mask);
5780 cpuset_cpus_allowed(p, cpus_allowed);
5781 if (!cpumask_subset(new_mask, cpus_allowed)) {
5783 * We must have raced with a concurrent cpuset
5784 * update. Just reset the cpus_allowed to the
5785 * cpuset's cpus_allowed
5787 cpumask_copy(new_mask, cpus_allowed);
5792 free_cpumask_var(new_mask);
5793 out_free_cpus_allowed:
5794 free_cpumask_var(cpus_allowed);
5801 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5802 struct cpumask *new_mask)
5804 if (len < cpumask_size())
5805 cpumask_clear(new_mask);
5806 else if (len > cpumask_size())
5807 len = cpumask_size();
5809 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5813 * sys_sched_setaffinity - set the cpu affinity of a process
5814 * @pid: pid of the process
5815 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5816 * @user_mask_ptr: user-space pointer to the new cpu mask
5818 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5819 unsigned long __user *, user_mask_ptr)
5821 cpumask_var_t new_mask;
5824 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5827 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5829 retval = sched_setaffinity(pid, new_mask);
5830 free_cpumask_var(new_mask);
5834 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5836 struct task_struct *p;
5837 unsigned long flags;
5844 p = find_process_by_pid(pid);
5848 retval = security_task_getscheduler(p);
5852 raw_spin_lock_irqsave(&p->pi_lock, flags);
5853 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5854 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5864 * sys_sched_getaffinity - get the cpu affinity of a process
5865 * @pid: pid of the process
5866 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5867 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5869 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5870 unsigned long __user *, user_mask_ptr)
5875 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5877 if (len & (sizeof(unsigned long)-1))
5880 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5883 ret = sched_getaffinity(pid, mask);
5885 size_t retlen = min_t(size_t, len, cpumask_size());
5887 if (copy_to_user(user_mask_ptr, mask, retlen))
5892 free_cpumask_var(mask);
5898 * sys_sched_yield - yield the current processor to other threads.
5900 * This function yields the current CPU to other tasks. If there are no
5901 * other threads running on this CPU then this function will return.
5903 SYSCALL_DEFINE0(sched_yield)
5905 struct rq *rq = this_rq_lock();
5907 schedstat_inc(rq, yld_count);
5908 current->sched_class->yield_task(rq);
5911 * Since we are going to call schedule() anyway, there's
5912 * no need to preempt or enable interrupts:
5914 __release(rq->lock);
5915 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5916 do_raw_spin_unlock(&rq->lock);
5917 preempt_enable_no_resched();
5924 static inline int should_resched(void)
5926 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5929 static void __cond_resched(void)
5931 add_preempt_count(PREEMPT_ACTIVE);
5933 sub_preempt_count(PREEMPT_ACTIVE);
5936 int __sched _cond_resched(void)
5938 if (should_resched()) {
5944 EXPORT_SYMBOL(_cond_resched);
5947 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5948 * call schedule, and on return reacquire the lock.
5950 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5951 * operations here to prevent schedule() from being called twice (once via
5952 * spin_unlock(), once by hand).
5954 int __cond_resched_lock(spinlock_t *lock)
5956 int resched = should_resched();
5959 lockdep_assert_held(lock);
5961 if (spin_needbreak(lock) || resched) {
5972 EXPORT_SYMBOL(__cond_resched_lock);
5974 int __sched __cond_resched_softirq(void)
5976 BUG_ON(!in_softirq());
5978 if (should_resched()) {
5986 EXPORT_SYMBOL(__cond_resched_softirq);
5989 * yield - yield the current processor to other threads.
5991 * This is a shortcut for kernel-space yielding - it marks the
5992 * thread runnable and calls sys_sched_yield().
5994 void __sched yield(void)
5996 set_current_state(TASK_RUNNING);
5999 EXPORT_SYMBOL(yield);
6002 * yield_to - yield the current processor to another thread in
6003 * your thread group, or accelerate that thread toward the
6004 * processor it's on.
6006 * @preempt: whether task preemption is allowed or not
6008 * It's the caller's job to ensure that the target task struct
6009 * can't go away on us before we can do any checks.
6011 * Returns true if we indeed boosted the target task.
6013 bool __sched yield_to(struct task_struct *p, bool preempt)
6015 struct task_struct *curr = current;
6016 struct rq *rq, *p_rq;
6017 unsigned long flags;
6020 local_irq_save(flags);
6025 double_rq_lock(rq, p_rq);
6026 while (task_rq(p) != p_rq) {
6027 double_rq_unlock(rq, p_rq);
6031 if (!curr->sched_class->yield_to_task)
6034 if (curr->sched_class != p->sched_class)
6037 if (task_running(p_rq, p) || p->state)
6040 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
6042 schedstat_inc(rq, yld_count);
6044 * Make p's CPU reschedule; pick_next_entity takes care of
6047 if (preempt && rq != p_rq)
6048 resched_task(p_rq->curr);
6052 double_rq_unlock(rq, p_rq);
6053 local_irq_restore(flags);
6060 EXPORT_SYMBOL_GPL(yield_to);
6063 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6064 * that process accounting knows that this is a task in IO wait state.
6066 void __sched io_schedule(void)
6068 struct rq *rq = raw_rq();
6070 delayacct_blkio_start();
6071 atomic_inc(&rq->nr_iowait);
6072 blk_flush_plug(current);
6073 current->in_iowait = 1;
6075 current->in_iowait = 0;
6076 atomic_dec(&rq->nr_iowait);
6077 delayacct_blkio_end();
6079 EXPORT_SYMBOL(io_schedule);
6081 long __sched io_schedule_timeout(long timeout)
6083 struct rq *rq = raw_rq();
6086 delayacct_blkio_start();
6087 atomic_inc(&rq->nr_iowait);
6088 blk_flush_plug(current);
6089 current->in_iowait = 1;
6090 ret = schedule_timeout(timeout);
6091 current->in_iowait = 0;
6092 atomic_dec(&rq->nr_iowait);
6093 delayacct_blkio_end();
6098 * sys_sched_get_priority_max - return maximum RT priority.
6099 * @policy: scheduling class.
6101 * this syscall returns the maximum rt_priority that can be used
6102 * by a given scheduling class.
6104 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6111 ret = MAX_USER_RT_PRIO-1;
6123 * sys_sched_get_priority_min - return minimum RT priority.
6124 * @policy: scheduling class.
6126 * this syscall returns the minimum rt_priority that can be used
6127 * by a given scheduling class.
6129 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6147 * sys_sched_rr_get_interval - return the default timeslice of a process.
6148 * @pid: pid of the process.
6149 * @interval: userspace pointer to the timeslice value.
6151 * this syscall writes the default timeslice value of a given process
6152 * into the user-space timespec buffer. A value of '0' means infinity.
6154 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6155 struct timespec __user *, interval)
6157 struct task_struct *p;
6158 unsigned int time_slice;
6159 unsigned long flags;
6169 p = find_process_by_pid(pid);
6173 retval = security_task_getscheduler(p);
6177 rq = task_rq_lock(p, &flags);
6178 time_slice = p->sched_class->get_rr_interval(rq, p);
6179 task_rq_unlock(rq, p, &flags);
6182 jiffies_to_timespec(time_slice, &t);
6183 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6191 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6193 void sched_show_task(struct task_struct *p)
6195 unsigned long free = 0;
6198 state = p->state ? __ffs(p->state) + 1 : 0;
6199 printk(KERN_INFO "%-15.15s %c", p->comm,
6200 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6201 #if BITS_PER_LONG == 32
6202 if (state == TASK_RUNNING)
6203 printk(KERN_CONT " running ");
6205 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6207 if (state == TASK_RUNNING)
6208 printk(KERN_CONT " running task ");
6210 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6212 #ifdef CONFIG_DEBUG_STACK_USAGE
6213 free = stack_not_used(p);
6215 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6216 task_pid_nr(p), task_pid_nr(p->real_parent),
6217 (unsigned long)task_thread_info(p)->flags);
6219 show_stack(p, NULL);
6222 void show_state_filter(unsigned long state_filter)
6224 struct task_struct *g, *p;
6226 #if BITS_PER_LONG == 32
6228 " task PC stack pid father\n");
6231 " task PC stack pid father\n");
6234 do_each_thread(g, p) {
6236 * reset the NMI-timeout, listing all files on a slow
6237 * console might take a lot of time:
6239 touch_nmi_watchdog();
6240 if (!state_filter || (p->state & state_filter))
6242 } while_each_thread(g, p);
6244 touch_all_softlockup_watchdogs();
6246 #ifdef CONFIG_SCHED_DEBUG
6247 sysrq_sched_debug_show();
6251 * Only show locks if all tasks are dumped:
6254 debug_show_all_locks();
6257 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6259 idle->sched_class = &idle_sched_class;
6263 * init_idle - set up an idle thread for a given CPU
6264 * @idle: task in question
6265 * @cpu: cpu the idle task belongs to
6267 * NOTE: this function does not set the idle thread's NEED_RESCHED
6268 * flag, to make booting more robust.
6270 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6272 struct rq *rq = cpu_rq(cpu);
6273 unsigned long flags;
6275 raw_spin_lock_irqsave(&rq->lock, flags);
6278 idle->state = TASK_RUNNING;
6279 idle->se.exec_start = sched_clock();
6281 do_set_cpus_allowed(idle, cpumask_of(cpu));
6283 * We're having a chicken and egg problem, even though we are
6284 * holding rq->lock, the cpu isn't yet set to this cpu so the
6285 * lockdep check in task_group() will fail.
6287 * Similar case to sched_fork(). / Alternatively we could
6288 * use task_rq_lock() here and obtain the other rq->lock.
6293 __set_task_cpu(idle, cpu);
6296 rq->curr = rq->idle = idle;
6297 #if defined(CONFIG_SMP)
6300 raw_spin_unlock_irqrestore(&rq->lock, flags);
6302 /* Set the preempt count _outside_ the spinlocks! */
6303 task_thread_info(idle)->preempt_count = 0;
6306 * The idle tasks have their own, simple scheduling class:
6308 idle->sched_class = &idle_sched_class;
6309 ftrace_graph_init_idle_task(idle, cpu);
6310 #if defined(CONFIG_SMP)
6311 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6316 * Increase the granularity value when there are more CPUs,
6317 * because with more CPUs the 'effective latency' as visible
6318 * to users decreases. But the relationship is not linear,
6319 * so pick a second-best guess by going with the log2 of the
6322 * This idea comes from the SD scheduler of Con Kolivas:
6324 static int get_update_sysctl_factor(void)
6326 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6327 unsigned int factor;
6329 switch (sysctl_sched_tunable_scaling) {
6330 case SCHED_TUNABLESCALING_NONE:
6333 case SCHED_TUNABLESCALING_LINEAR:
6336 case SCHED_TUNABLESCALING_LOG:
6338 factor = 1 + ilog2(cpus);
6345 static void update_sysctl(void)
6347 unsigned int factor = get_update_sysctl_factor();
6349 #define SET_SYSCTL(name) \
6350 (sysctl_##name = (factor) * normalized_sysctl_##name)
6351 SET_SYSCTL(sched_min_granularity);
6352 SET_SYSCTL(sched_latency);
6353 SET_SYSCTL(sched_wakeup_granularity);
6357 static inline void sched_init_granularity(void)
6363 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6365 if (p->sched_class && p->sched_class->set_cpus_allowed)
6366 p->sched_class->set_cpus_allowed(p, new_mask);
6368 cpumask_copy(&p->cpus_allowed, new_mask);
6369 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6373 * This is how migration works:
6375 * 1) we invoke migration_cpu_stop() on the target CPU using
6377 * 2) stopper starts to run (implicitly forcing the migrated thread
6379 * 3) it checks whether the migrated task is still in the wrong runqueue.
6380 * 4) if it's in the wrong runqueue then the migration thread removes
6381 * it and puts it into the right queue.
6382 * 5) stopper completes and stop_one_cpu() returns and the migration
6387 * Change a given task's CPU affinity. Migrate the thread to a
6388 * proper CPU and schedule it away if the CPU it's executing on
6389 * is removed from the allowed bitmask.
6391 * NOTE: the caller must have a valid reference to the task, the
6392 * task must not exit() & deallocate itself prematurely. The
6393 * call is not atomic; no spinlocks may be held.
6395 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6397 unsigned long flags;
6399 unsigned int dest_cpu;
6402 rq = task_rq_lock(p, &flags);
6404 if (cpumask_equal(&p->cpus_allowed, new_mask))
6407 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6412 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6417 do_set_cpus_allowed(p, new_mask);
6419 /* Can the task run on the task's current CPU? If so, we're done */
6420 if (cpumask_test_cpu(task_cpu(p), new_mask))
6423 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6425 struct migration_arg arg = { p, dest_cpu };
6426 /* Need help from migration thread: drop lock and wait. */
6427 task_rq_unlock(rq, p, &flags);
6428 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6429 tlb_migrate_finish(p->mm);
6433 task_rq_unlock(rq, p, &flags);
6437 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6440 * Move (not current) task off this cpu, onto dest cpu. We're doing
6441 * this because either it can't run here any more (set_cpus_allowed()
6442 * away from this CPU, or CPU going down), or because we're
6443 * attempting to rebalance this task on exec (sched_exec).
6445 * So we race with normal scheduler movements, but that's OK, as long
6446 * as the task is no longer on this CPU.
6448 * Returns non-zero if task was successfully migrated.
6450 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6452 struct rq *rq_dest, *rq_src;
6455 if (unlikely(!cpu_active(dest_cpu)))
6458 rq_src = cpu_rq(src_cpu);
6459 rq_dest = cpu_rq(dest_cpu);
6461 raw_spin_lock(&p->pi_lock);
6462 double_rq_lock(rq_src, rq_dest);
6463 /* Already moved. */
6464 if (task_cpu(p) != src_cpu)
6466 /* Affinity changed (again). */
6467 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
6471 * If we're not on a rq, the next wake-up will ensure we're
6475 deactivate_task(rq_src, p, 0);
6476 set_task_cpu(p, dest_cpu);
6477 activate_task(rq_dest, p, 0);
6478 check_preempt_curr(rq_dest, p, 0);
6483 double_rq_unlock(rq_src, rq_dest);
6484 raw_spin_unlock(&p->pi_lock);
6489 * migration_cpu_stop - this will be executed by a highprio stopper thread
6490 * and performs thread migration by bumping thread off CPU then
6491 * 'pushing' onto another runqueue.
6493 static int migration_cpu_stop(void *data)
6495 struct migration_arg *arg = data;
6498 * The original target cpu might have gone down and we might
6499 * be on another cpu but it doesn't matter.
6501 local_irq_disable();
6502 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6507 #ifdef CONFIG_HOTPLUG_CPU
6510 * Ensures that the idle task is using init_mm right before its cpu goes
6513 void idle_task_exit(void)
6515 struct mm_struct *mm = current->active_mm;
6517 BUG_ON(cpu_online(smp_processor_id()));
6520 switch_mm(mm, &init_mm, current);
6525 * While a dead CPU has no uninterruptible tasks queued at this point,
6526 * it might still have a nonzero ->nr_uninterruptible counter, because
6527 * for performance reasons the counter is not stricly tracking tasks to
6528 * their home CPUs. So we just add the counter to another CPU's counter,
6529 * to keep the global sum constant after CPU-down:
6531 static void migrate_nr_uninterruptible(struct rq *rq_src)
6533 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6535 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6536 rq_src->nr_uninterruptible = 0;
6540 * remove the tasks which were accounted by rq from calc_load_tasks.
6542 static void calc_global_load_remove(struct rq *rq)
6544 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6545 rq->calc_load_active = 0;
6548 #ifdef CONFIG_CFS_BANDWIDTH
6549 static void unthrottle_offline_cfs_rqs(struct rq *rq)
6551 struct cfs_rq *cfs_rq;
6553 for_each_leaf_cfs_rq(rq, cfs_rq) {
6554 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6556 if (!cfs_rq->runtime_enabled)
6560 * clock_task is not advancing so we just need to make sure
6561 * there's some valid quota amount
6563 cfs_rq->runtime_remaining = cfs_b->quota;
6564 if (cfs_rq_throttled(cfs_rq))
6565 unthrottle_cfs_rq(cfs_rq);
6569 static void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6573 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6574 * try_to_wake_up()->select_task_rq().
6576 * Called with rq->lock held even though we'er in stop_machine() and
6577 * there's no concurrency possible, we hold the required locks anyway
6578 * because of lock validation efforts.
6580 static void migrate_tasks(unsigned int dead_cpu)
6582 struct rq *rq = cpu_rq(dead_cpu);
6583 struct task_struct *next, *stop = rq->stop;
6587 * Fudge the rq selection such that the below task selection loop
6588 * doesn't get stuck on the currently eligible stop task.
6590 * We're currently inside stop_machine() and the rq is either stuck
6591 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6592 * either way we should never end up calling schedule() until we're
6597 /* Ensure any throttled groups are reachable by pick_next_task */
6598 unthrottle_offline_cfs_rqs(rq);
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_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6628 static struct ctl_table sd_ctl_dir[] = {
6630 .procname = "sched_domain",
6636 static struct ctl_table sd_ctl_root[] = {
6638 .procname = "kernel",
6640 .child = sd_ctl_dir,
6645 static struct ctl_table *sd_alloc_ctl_entry(int n)
6647 struct ctl_table *entry =
6648 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6653 static void sd_free_ctl_entry(struct ctl_table **tablep)
6655 struct ctl_table *entry;
6658 * In the intermediate directories, both the child directory and
6659 * procname are dynamically allocated and could fail but the mode
6660 * will always be set. In the lowest directory the names are
6661 * static strings and all have proc handlers.
6663 for (entry = *tablep; entry->mode; entry++) {
6665 sd_free_ctl_entry(&entry->child);
6666 if (entry->proc_handler == NULL)
6667 kfree(entry->procname);
6675 set_table_entry(struct ctl_table *entry,
6676 const char *procname, void *data, int maxlen,
6677 mode_t mode, proc_handler *proc_handler)
6679 entry->procname = procname;
6681 entry->maxlen = maxlen;
6683 entry->proc_handler = proc_handler;
6686 static struct ctl_table *
6687 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6689 struct ctl_table *table = sd_alloc_ctl_entry(13);
6694 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6695 sizeof(long), 0644, proc_doulongvec_minmax);
6696 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6697 sizeof(long), 0644, proc_doulongvec_minmax);
6698 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6699 sizeof(int), 0644, proc_dointvec_minmax);
6700 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6701 sizeof(int), 0644, proc_dointvec_minmax);
6702 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6703 sizeof(int), 0644, proc_dointvec_minmax);
6704 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6705 sizeof(int), 0644, proc_dointvec_minmax);
6706 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6707 sizeof(int), 0644, proc_dointvec_minmax);
6708 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6709 sizeof(int), 0644, proc_dointvec_minmax);
6710 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6711 sizeof(int), 0644, proc_dointvec_minmax);
6712 set_table_entry(&table[9], "cache_nice_tries",
6713 &sd->cache_nice_tries,
6714 sizeof(int), 0644, proc_dointvec_minmax);
6715 set_table_entry(&table[10], "flags", &sd->flags,
6716 sizeof(int), 0644, proc_dointvec_minmax);
6717 set_table_entry(&table[11], "name", sd->name,
6718 CORENAME_MAX_SIZE, 0444, proc_dostring);
6719 /* &table[12] is terminator */
6724 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6726 struct ctl_table *entry, *table;
6727 struct sched_domain *sd;
6728 int domain_num = 0, i;
6731 for_each_domain(cpu, sd)
6733 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6738 for_each_domain(cpu, sd) {
6739 snprintf(buf, 32, "domain%d", i);
6740 entry->procname = kstrdup(buf, GFP_KERNEL);
6742 entry->child = sd_alloc_ctl_domain_table(sd);
6749 static struct ctl_table_header *sd_sysctl_header;
6750 static void register_sched_domain_sysctl(void)
6752 int i, cpu_num = num_possible_cpus();
6753 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6756 WARN_ON(sd_ctl_dir[0].child);
6757 sd_ctl_dir[0].child = entry;
6762 for_each_possible_cpu(i) {
6763 snprintf(buf, 32, "cpu%d", i);
6764 entry->procname = kstrdup(buf, GFP_KERNEL);
6766 entry->child = sd_alloc_ctl_cpu_table(i);
6770 WARN_ON(sd_sysctl_header);
6771 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6774 /* may be called multiple times per register */
6775 static void unregister_sched_domain_sysctl(void)
6777 if (sd_sysctl_header)
6778 unregister_sysctl_table(sd_sysctl_header);
6779 sd_sysctl_header = NULL;
6780 if (sd_ctl_dir[0].child)
6781 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6784 static void register_sched_domain_sysctl(void)
6787 static void unregister_sched_domain_sysctl(void)
6792 static void set_rq_online(struct rq *rq)
6795 const struct sched_class *class;
6797 cpumask_set_cpu(rq->cpu, rq->rd->online);
6800 for_each_class(class) {
6801 if (class->rq_online)
6802 class->rq_online(rq);
6807 static void set_rq_offline(struct rq *rq)
6810 const struct sched_class *class;
6812 for_each_class(class) {
6813 if (class->rq_offline)
6814 class->rq_offline(rq);
6817 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6823 * migration_call - callback that gets triggered when a CPU is added.
6824 * Here we can start up the necessary migration thread for the new CPU.
6826 static int __cpuinit
6827 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6829 int cpu = (long)hcpu;
6830 unsigned long flags;
6831 struct rq *rq = cpu_rq(cpu);
6833 switch (action & ~CPU_TASKS_FROZEN) {
6835 case CPU_UP_PREPARE:
6836 rq->calc_load_update = calc_load_update;
6840 /* Update our root-domain */
6841 raw_spin_lock_irqsave(&rq->lock, flags);
6843 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6847 raw_spin_unlock_irqrestore(&rq->lock, flags);
6850 #ifdef CONFIG_HOTPLUG_CPU
6852 sched_ttwu_pending();
6853 /* Update our root-domain */
6854 raw_spin_lock_irqsave(&rq->lock, flags);
6856 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6860 BUG_ON(rq->nr_running != 1); /* the migration thread */
6861 raw_spin_unlock_irqrestore(&rq->lock, flags);
6863 migrate_nr_uninterruptible(rq);
6864 calc_global_load_remove(rq);
6869 update_max_interval();
6875 * Register at high priority so that task migration (migrate_all_tasks)
6876 * happens before everything else. This has to be lower priority than
6877 * the notifier in the perf_event subsystem, though.
6879 static struct notifier_block __cpuinitdata migration_notifier = {
6880 .notifier_call = migration_call,
6881 .priority = CPU_PRI_MIGRATION,
6884 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6885 unsigned long action, void *hcpu)
6887 switch (action & ~CPU_TASKS_FROZEN) {
6889 case CPU_DOWN_FAILED:
6890 set_cpu_active((long)hcpu, true);
6897 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6898 unsigned long action, void *hcpu)
6900 switch (action & ~CPU_TASKS_FROZEN) {
6901 case CPU_DOWN_PREPARE:
6902 set_cpu_active((long)hcpu, false);
6909 static int __init migration_init(void)
6911 void *cpu = (void *)(long)smp_processor_id();
6914 /* Initialize migration for the boot CPU */
6915 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6916 BUG_ON(err == NOTIFY_BAD);
6917 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6918 register_cpu_notifier(&migration_notifier);
6920 /* Register cpu active notifiers */
6921 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6922 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6926 early_initcall(migration_init);
6931 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6933 #ifdef CONFIG_SCHED_DEBUG
6935 static __read_mostly int sched_domain_debug_enabled;
6937 static int __init sched_domain_debug_setup(char *str)
6939 sched_domain_debug_enabled = 1;
6943 early_param("sched_debug", sched_domain_debug_setup);
6945 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6946 struct cpumask *groupmask)
6948 struct sched_group *group = sd->groups;
6951 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6952 cpumask_clear(groupmask);
6954 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6956 if (!(sd->flags & SD_LOAD_BALANCE)) {
6957 printk("does not load-balance\n");
6959 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6964 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6966 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6967 printk(KERN_ERR "ERROR: domain->span does not contain "
6970 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6971 printk(KERN_ERR "ERROR: domain->groups does not contain"
6975 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6979 printk(KERN_ERR "ERROR: group is NULL\n");
6983 if (!group->sgp->power) {
6984 printk(KERN_CONT "\n");
6985 printk(KERN_ERR "ERROR: domain->cpu_power not "
6990 if (!cpumask_weight(sched_group_cpus(group))) {
6991 printk(KERN_CONT "\n");
6992 printk(KERN_ERR "ERROR: empty group\n");
6996 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6997 printk(KERN_CONT "\n");
6998 printk(KERN_ERR "ERROR: repeated CPUs\n");
7002 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7004 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7006 printk(KERN_CONT " %s", str);
7007 if (group->sgp->power != SCHED_POWER_SCALE) {
7008 printk(KERN_CONT " (cpu_power = %d)",
7012 group = group->next;
7013 } while (group != sd->groups);
7014 printk(KERN_CONT "\n");
7016 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7017 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7020 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7021 printk(KERN_ERR "ERROR: parent span is not a superset "
7022 "of domain->span\n");
7026 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7030 if (!sched_domain_debug_enabled)
7034 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7038 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7041 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
7049 #else /* !CONFIG_SCHED_DEBUG */
7050 # define sched_domain_debug(sd, cpu) do { } while (0)
7051 #endif /* CONFIG_SCHED_DEBUG */
7053 static int sd_degenerate(struct sched_domain *sd)
7055 if (cpumask_weight(sched_domain_span(sd)) == 1)
7058 /* Following flags need at least 2 groups */
7059 if (sd->flags & (SD_LOAD_BALANCE |
7060 SD_BALANCE_NEWIDLE |
7064 SD_SHARE_PKG_RESOURCES)) {
7065 if (sd->groups != sd->groups->next)
7069 /* Following flags don't use groups */
7070 if (sd->flags & (SD_WAKE_AFFINE))
7077 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7079 unsigned long cflags = sd->flags, pflags = parent->flags;
7081 if (sd_degenerate(parent))
7084 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7087 /* Flags needing groups don't count if only 1 group in parent */
7088 if (parent->groups == parent->groups->next) {
7089 pflags &= ~(SD_LOAD_BALANCE |
7090 SD_BALANCE_NEWIDLE |
7094 SD_SHARE_PKG_RESOURCES);
7095 if (nr_node_ids == 1)
7096 pflags &= ~SD_SERIALIZE;
7098 if (~cflags & pflags)
7104 static void free_rootdomain(struct rcu_head *rcu)
7106 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
7108 cpupri_cleanup(&rd->cpupri);
7109 free_cpumask_var(rd->rto_mask);
7110 free_cpumask_var(rd->online);
7111 free_cpumask_var(rd->span);
7115 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7117 struct root_domain *old_rd = NULL;
7118 unsigned long flags;
7120 raw_spin_lock_irqsave(&rq->lock, flags);
7125 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7128 cpumask_clear_cpu(rq->cpu, old_rd->span);
7131 * If we dont want to free the old_rt yet then
7132 * set old_rd to NULL to skip the freeing later
7135 if (!atomic_dec_and_test(&old_rd->refcount))
7139 atomic_inc(&rd->refcount);
7142 cpumask_set_cpu(rq->cpu, rd->span);
7143 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7146 raw_spin_unlock_irqrestore(&rq->lock, flags);
7149 call_rcu_sched(&old_rd->rcu, free_rootdomain);
7152 static int init_rootdomain(struct root_domain *rd)
7154 memset(rd, 0, sizeof(*rd));
7156 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7158 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7160 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7163 if (cpupri_init(&rd->cpupri) != 0)
7168 free_cpumask_var(rd->rto_mask);
7170 free_cpumask_var(rd->online);
7172 free_cpumask_var(rd->span);
7177 static void init_defrootdomain(void)
7179 init_rootdomain(&def_root_domain);
7181 atomic_set(&def_root_domain.refcount, 1);
7184 static struct root_domain *alloc_rootdomain(void)
7186 struct root_domain *rd;
7188 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7192 if (init_rootdomain(rd) != 0) {
7200 static void free_sched_groups(struct sched_group *sg, int free_sgp)
7202 struct sched_group *tmp, *first;
7211 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
7216 } while (sg != first);
7219 static void free_sched_domain(struct rcu_head *rcu)
7221 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
7224 * If its an overlapping domain it has private groups, iterate and
7227 if (sd->flags & SD_OVERLAP) {
7228 free_sched_groups(sd->groups, 1);
7229 } else if (atomic_dec_and_test(&sd->groups->ref)) {
7230 kfree(sd->groups->sgp);
7236 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
7238 call_rcu(&sd->rcu, free_sched_domain);
7241 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
7243 for (; sd; sd = sd->parent)
7244 destroy_sched_domain(sd, cpu);
7248 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7249 * hold the hotplug lock.
7252 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7254 struct rq *rq = cpu_rq(cpu);
7255 struct sched_domain *tmp;
7257 /* Remove the sched domains which do not contribute to scheduling. */
7258 for (tmp = sd; tmp; ) {
7259 struct sched_domain *parent = tmp->parent;
7263 if (sd_parent_degenerate(tmp, parent)) {
7264 tmp->parent = parent->parent;
7266 parent->parent->child = tmp;
7267 destroy_sched_domain(parent, cpu);
7272 if (sd && sd_degenerate(sd)) {
7275 destroy_sched_domain(tmp, cpu);
7280 sched_domain_debug(sd, cpu);
7282 rq_attach_root(rq, rd);
7284 rcu_assign_pointer(rq->sd, sd);
7285 destroy_sched_domains(tmp, cpu);
7288 /* cpus with isolated domains */
7289 static cpumask_var_t cpu_isolated_map;
7291 /* Setup the mask of cpus configured for isolated domains */
7292 static int __init isolated_cpu_setup(char *str)
7294 alloc_bootmem_cpumask_var(&cpu_isolated_map);
7295 cpulist_parse(str, cpu_isolated_map);
7299 __setup("isolcpus=", isolated_cpu_setup);
7304 * find_next_best_node - find the next node to include in a sched_domain
7305 * @node: node whose sched_domain we're building
7306 * @used_nodes: nodes already in the sched_domain
7308 * Find the next node to include in a given scheduling domain. Simply
7309 * finds the closest node not already in the @used_nodes map.
7311 * Should use nodemask_t.
7313 static int find_next_best_node(int node, nodemask_t *used_nodes)
7315 int i, n, val, min_val, best_node = -1;
7319 for (i = 0; i < nr_node_ids; i++) {
7320 /* Start at @node */
7321 n = (node + i) % nr_node_ids;
7323 if (!nr_cpus_node(n))
7326 /* Skip already used nodes */
7327 if (node_isset(n, *used_nodes))
7330 /* Simple min distance search */
7331 val = node_distance(node, n);
7333 if (val < min_val) {
7339 if (best_node != -1)
7340 node_set(best_node, *used_nodes);
7345 * sched_domain_node_span - get a cpumask for a node's sched_domain
7346 * @node: node whose cpumask we're constructing
7347 * @span: resulting cpumask
7349 * Given a node, construct a good cpumask for its sched_domain to span. It
7350 * should be one that prevents unnecessary balancing, but also spreads tasks
7353 static void sched_domain_node_span(int node, struct cpumask *span)
7355 nodemask_t used_nodes;
7358 cpumask_clear(span);
7359 nodes_clear(used_nodes);
7361 cpumask_or(span, span, cpumask_of_node(node));
7362 node_set(node, used_nodes);
7364 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7365 int next_node = find_next_best_node(node, &used_nodes);
7368 cpumask_or(span, span, cpumask_of_node(next_node));
7372 static const struct cpumask *cpu_node_mask(int cpu)
7374 lockdep_assert_held(&sched_domains_mutex);
7376 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7378 return sched_domains_tmpmask;
7381 static const struct cpumask *cpu_allnodes_mask(int cpu)
7383 return cpu_possible_mask;
7385 #endif /* CONFIG_NUMA */
7387 static const struct cpumask *cpu_cpu_mask(int cpu)
7389 return cpumask_of_node(cpu_to_node(cpu));
7392 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7395 struct sched_domain **__percpu sd;
7396 struct sched_group **__percpu sg;
7397 struct sched_group_power **__percpu sgp;
7401 struct sched_domain ** __percpu sd;
7402 struct root_domain *rd;
7412 struct sched_domain_topology_level;
7414 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7415 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7417 #define SDTL_OVERLAP 0x01
7419 struct sched_domain_topology_level {
7420 sched_domain_init_f init;
7421 sched_domain_mask_f mask;
7423 struct sd_data data;
7427 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7429 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7430 const struct cpumask *span = sched_domain_span(sd);
7431 struct cpumask *covered = sched_domains_tmpmask;
7432 struct sd_data *sdd = sd->private;
7433 struct sched_domain *child;
7436 cpumask_clear(covered);
7438 for_each_cpu(i, span) {
7439 struct cpumask *sg_span;
7441 if (cpumask_test_cpu(i, covered))
7444 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7445 GFP_KERNEL, cpu_to_node(i));
7450 sg_span = sched_group_cpus(sg);
7452 child = *per_cpu_ptr(sdd->sd, i);
7454 child = child->child;
7455 cpumask_copy(sg_span, sched_domain_span(child));
7457 cpumask_set_cpu(i, sg_span);
7459 cpumask_or(covered, covered, sg_span);
7461 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7462 atomic_inc(&sg->sgp->ref);
7464 if (cpumask_test_cpu(cpu, sg_span))
7474 sd->groups = groups;
7479 free_sched_groups(first, 0);
7484 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7486 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7487 struct sched_domain *child = sd->child;
7490 cpu = cpumask_first(sched_domain_span(child));
7493 *sg = *per_cpu_ptr(sdd->sg, cpu);
7494 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7495 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7502 * build_sched_groups will build a circular linked list of the groups
7503 * covered by the given span, and will set each group's ->cpumask correctly,
7504 * and ->cpu_power to 0.
7506 * Assumes the sched_domain tree is fully constructed
7509 build_sched_groups(struct sched_domain *sd, int cpu)
7511 struct sched_group *first = NULL, *last = NULL;
7512 struct sd_data *sdd = sd->private;
7513 const struct cpumask *span = sched_domain_span(sd);
7514 struct cpumask *covered;
7517 get_group(cpu, sdd, &sd->groups);
7518 atomic_inc(&sd->groups->ref);
7520 if (cpu != cpumask_first(sched_domain_span(sd)))
7523 lockdep_assert_held(&sched_domains_mutex);
7524 covered = sched_domains_tmpmask;
7526 cpumask_clear(covered);
7528 for_each_cpu(i, span) {
7529 struct sched_group *sg;
7530 int group = get_group(i, sdd, &sg);
7533 if (cpumask_test_cpu(i, covered))
7536 cpumask_clear(sched_group_cpus(sg));
7539 for_each_cpu(j, span) {
7540 if (get_group(j, sdd, NULL) != group)
7543 cpumask_set_cpu(j, covered);
7544 cpumask_set_cpu(j, sched_group_cpus(sg));
7559 * Initialize sched groups cpu_power.
7561 * cpu_power indicates the capacity of sched group, which is used while
7562 * distributing the load between different sched groups in a sched domain.
7563 * Typically cpu_power for all the groups in a sched domain will be same unless
7564 * there are asymmetries in the topology. If there are asymmetries, group
7565 * having more cpu_power will pickup more load compared to the group having
7568 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7570 struct sched_group *sg = sd->groups;
7572 WARN_ON(!sd || !sg);
7575 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7577 } while (sg != sd->groups);
7579 if (cpu != group_first_cpu(sg))
7582 update_group_power(sd, cpu);
7586 * Initializers for schedule domains
7587 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7590 #ifdef CONFIG_SCHED_DEBUG
7591 # define SD_INIT_NAME(sd, type) sd->name = #type
7593 # define SD_INIT_NAME(sd, type) do { } while (0)
7596 #define SD_INIT_FUNC(type) \
7597 static noinline struct sched_domain * \
7598 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7600 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7601 *sd = SD_##type##_INIT; \
7602 SD_INIT_NAME(sd, type); \
7603 sd->private = &tl->data; \
7609 SD_INIT_FUNC(ALLNODES)
7612 #ifdef CONFIG_SCHED_SMT
7613 SD_INIT_FUNC(SIBLING)
7615 #ifdef CONFIG_SCHED_MC
7618 #ifdef CONFIG_SCHED_BOOK
7622 static int default_relax_domain_level = -1;
7623 int sched_domain_level_max;
7625 static int __init setup_relax_domain_level(char *str)
7627 if (kstrtoint(str, 0, &default_relax_domain_level))
7628 pr_warn("Unable to set relax_domain_level\n");
7632 __setup("relax_domain_level=", setup_relax_domain_level);
7634 static void set_domain_attribute(struct sched_domain *sd,
7635 struct sched_domain_attr *attr)
7639 if (!attr || attr->relax_domain_level < 0) {
7640 if (default_relax_domain_level < 0)
7643 request = default_relax_domain_level;
7645 request = attr->relax_domain_level;
7646 if (request < sd->level) {
7647 /* turn off idle balance on this domain */
7648 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7650 /* turn on idle balance on this domain */
7651 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7655 static void __sdt_free(const struct cpumask *cpu_map);
7656 static int __sdt_alloc(const struct cpumask *cpu_map);
7658 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7659 const struct cpumask *cpu_map)
7663 if (!atomic_read(&d->rd->refcount))
7664 free_rootdomain(&d->rd->rcu); /* fall through */
7666 free_percpu(d->sd); /* fall through */
7668 __sdt_free(cpu_map); /* fall through */
7674 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7675 const struct cpumask *cpu_map)
7677 memset(d, 0, sizeof(*d));
7679 if (__sdt_alloc(cpu_map))
7680 return sa_sd_storage;
7681 d->sd = alloc_percpu(struct sched_domain *);
7683 return sa_sd_storage;
7684 d->rd = alloc_rootdomain();
7687 return sa_rootdomain;
7691 * NULL the sd_data elements we've used to build the sched_domain and
7692 * sched_group structure so that the subsequent __free_domain_allocs()
7693 * will not free the data we're using.
7695 static void claim_allocations(int cpu, struct sched_domain *sd)
7697 struct sd_data *sdd = sd->private;
7699 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7700 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7702 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7703 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7705 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7706 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7709 #ifdef CONFIG_SCHED_SMT
7710 static const struct cpumask *cpu_smt_mask(int cpu)
7712 return topology_thread_cpumask(cpu);
7717 * Topology list, bottom-up.
7719 static struct sched_domain_topology_level default_topology[] = {
7720 #ifdef CONFIG_SCHED_SMT
7721 { sd_init_SIBLING, cpu_smt_mask, },
7723 #ifdef CONFIG_SCHED_MC
7724 { sd_init_MC, cpu_coregroup_mask, },
7726 #ifdef CONFIG_SCHED_BOOK
7727 { sd_init_BOOK, cpu_book_mask, },
7729 { sd_init_CPU, cpu_cpu_mask, },
7731 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7732 { sd_init_ALLNODES, cpu_allnodes_mask, },
7737 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7739 static int __sdt_alloc(const struct cpumask *cpu_map)
7741 struct sched_domain_topology_level *tl;
7744 for (tl = sched_domain_topology; tl->init; tl++) {
7745 struct sd_data *sdd = &tl->data;
7747 sdd->sd = alloc_percpu(struct sched_domain *);
7751 sdd->sg = alloc_percpu(struct sched_group *);
7755 sdd->sgp = alloc_percpu(struct sched_group_power *);
7759 for_each_cpu(j, cpu_map) {
7760 struct sched_domain *sd;
7761 struct sched_group *sg;
7762 struct sched_group_power *sgp;
7764 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7765 GFP_KERNEL, cpu_to_node(j));
7769 *per_cpu_ptr(sdd->sd, j) = sd;
7771 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7772 GFP_KERNEL, cpu_to_node(j));
7776 *per_cpu_ptr(sdd->sg, j) = sg;
7778 sgp = kzalloc_node(sizeof(struct sched_group_power),
7779 GFP_KERNEL, cpu_to_node(j));
7783 *per_cpu_ptr(sdd->sgp, j) = sgp;
7790 static void __sdt_free(const struct cpumask *cpu_map)
7792 struct sched_domain_topology_level *tl;
7795 for (tl = sched_domain_topology; tl->init; tl++) {
7796 struct sd_data *sdd = &tl->data;
7798 for_each_cpu(j, cpu_map) {
7799 struct sched_domain *sd;
7802 sd = *per_cpu_ptr(sdd->sd, j);
7803 if (sd && (sd->flags & SD_OVERLAP))
7804 free_sched_groups(sd->groups, 0);
7805 kfree(*per_cpu_ptr(sdd->sd, j));
7809 kfree(*per_cpu_ptr(sdd->sg, j));
7811 kfree(*per_cpu_ptr(sdd->sgp, j));
7813 free_percpu(sdd->sd);
7815 free_percpu(sdd->sg);
7817 free_percpu(sdd->sgp);
7822 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7823 struct s_data *d, const struct cpumask *cpu_map,
7824 struct sched_domain_attr *attr, struct sched_domain *child,
7827 struct sched_domain *sd = tl->init(tl, cpu);
7831 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7833 sd->level = child->level + 1;
7834 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7838 set_domain_attribute(sd, attr);
7844 * Build sched domains for a given set of cpus and attach the sched domains
7845 * to the individual cpus
7847 static int build_sched_domains(const struct cpumask *cpu_map,
7848 struct sched_domain_attr *attr)
7850 enum s_alloc alloc_state = sa_none;
7851 struct sched_domain *sd;
7853 int i, ret = -ENOMEM;
7855 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7856 if (alloc_state != sa_rootdomain)
7859 /* Set up domains for cpus specified by the cpu_map. */
7860 for_each_cpu(i, cpu_map) {
7861 struct sched_domain_topology_level *tl;
7864 for (tl = sched_domain_topology; tl->init; tl++) {
7865 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7866 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7867 sd->flags |= SD_OVERLAP;
7868 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7875 *per_cpu_ptr(d.sd, i) = sd;
7878 /* Build the groups for the domains */
7879 for_each_cpu(i, cpu_map) {
7880 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7881 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7882 if (sd->flags & SD_OVERLAP) {
7883 if (build_overlap_sched_groups(sd, i))
7886 if (build_sched_groups(sd, i))
7892 /* Calculate CPU power for physical packages and nodes */
7893 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7894 if (!cpumask_test_cpu(i, cpu_map))
7897 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7898 claim_allocations(i, sd);
7899 init_sched_groups_power(i, sd);
7903 /* Attach the domains */
7905 for_each_cpu(i, cpu_map) {
7906 sd = *per_cpu_ptr(d.sd, i);
7907 cpu_attach_domain(sd, d.rd, i);
7913 __free_domain_allocs(&d, alloc_state, cpu_map);
7917 static cpumask_var_t *doms_cur; /* current sched domains */
7918 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7919 static struct sched_domain_attr *dattr_cur;
7920 /* attribues of custom domains in 'doms_cur' */
7923 * Special case: If a kmalloc of a doms_cur partition (array of
7924 * cpumask) fails, then fallback to a single sched domain,
7925 * as determined by the single cpumask fallback_doms.
7927 static cpumask_var_t fallback_doms;
7930 * arch_update_cpu_topology lets virtualized architectures update the
7931 * cpu core maps. It is supposed to return 1 if the topology changed
7932 * or 0 if it stayed the same.
7934 int __attribute__((weak)) arch_update_cpu_topology(void)
7939 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7942 cpumask_var_t *doms;
7944 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7947 for (i = 0; i < ndoms; i++) {
7948 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7949 free_sched_domains(doms, i);
7956 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7959 for (i = 0; i < ndoms; i++)
7960 free_cpumask_var(doms[i]);
7965 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7966 * For now this just excludes isolated cpus, but could be used to
7967 * exclude other special cases in the future.
7969 static int init_sched_domains(const struct cpumask *cpu_map)
7973 arch_update_cpu_topology();
7975 doms_cur = alloc_sched_domains(ndoms_cur);
7977 doms_cur = &fallback_doms;
7978 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7980 err = build_sched_domains(doms_cur[0], NULL);
7981 register_sched_domain_sysctl();
7987 * Detach sched domains from a group of cpus specified in cpu_map
7988 * These cpus will now be attached to the NULL domain
7990 static void detach_destroy_domains(const struct cpumask *cpu_map)
7995 for_each_cpu(i, cpu_map)
7996 cpu_attach_domain(NULL, &def_root_domain, i);
8000 /* handle null as "default" */
8001 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8002 struct sched_domain_attr *new, int idx_new)
8004 struct sched_domain_attr tmp;
8011 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8012 new ? (new + idx_new) : &tmp,
8013 sizeof(struct sched_domain_attr));
8017 * Partition sched domains as specified by the 'ndoms_new'
8018 * cpumasks in the array doms_new[] of cpumasks. This compares
8019 * doms_new[] to the current sched domain partitioning, doms_cur[].
8020 * It destroys each deleted domain and builds each new domain.
8022 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
8023 * The masks don't intersect (don't overlap.) We should setup one
8024 * sched domain for each mask. CPUs not in any of the cpumasks will
8025 * not be load balanced. If the same cpumask appears both in the
8026 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8029 * The passed in 'doms_new' should be allocated using
8030 * alloc_sched_domains. This routine takes ownership of it and will
8031 * free_sched_domains it when done with it. If the caller failed the
8032 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
8033 * and partition_sched_domains() will fallback to the single partition
8034 * 'fallback_doms', it also forces the domains to be rebuilt.
8036 * If doms_new == NULL it will be replaced with cpu_online_mask.
8037 * ndoms_new == 0 is a special case for destroying existing domains,
8038 * and it will not create the default domain.
8040 * Call with hotplug lock held
8042 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
8043 struct sched_domain_attr *dattr_new)
8048 mutex_lock(&sched_domains_mutex);
8050 /* always unregister in case we don't destroy any domains */
8051 unregister_sched_domain_sysctl();
8053 /* Let architecture update cpu core mappings. */
8054 new_topology = arch_update_cpu_topology();
8056 n = doms_new ? ndoms_new : 0;
8058 /* Destroy deleted domains */
8059 for (i = 0; i < ndoms_cur; i++) {
8060 for (j = 0; j < n && !new_topology; j++) {
8061 if (cpumask_equal(doms_cur[i], doms_new[j])
8062 && dattrs_equal(dattr_cur, i, dattr_new, j))
8065 /* no match - a current sched domain not in new doms_new[] */
8066 detach_destroy_domains(doms_cur[i]);
8071 if (doms_new == NULL) {
8073 doms_new = &fallback_doms;
8074 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
8075 WARN_ON_ONCE(dattr_new);
8078 /* Build new domains */
8079 for (i = 0; i < ndoms_new; i++) {
8080 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8081 if (cpumask_equal(doms_new[i], doms_cur[j])
8082 && dattrs_equal(dattr_new, i, dattr_cur, j))
8085 /* no match - add a new doms_new */
8086 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
8091 /* Remember the new sched domains */
8092 if (doms_cur != &fallback_doms)
8093 free_sched_domains(doms_cur, ndoms_cur);
8094 kfree(dattr_cur); /* kfree(NULL) is safe */
8095 doms_cur = doms_new;
8096 dattr_cur = dattr_new;
8097 ndoms_cur = ndoms_new;
8099 register_sched_domain_sysctl();
8101 mutex_unlock(&sched_domains_mutex);
8104 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8105 static void reinit_sched_domains(void)
8109 /* Destroy domains first to force the rebuild */
8110 partition_sched_domains(0, NULL, NULL);
8112 rebuild_sched_domains();
8116 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8118 unsigned int level = 0;
8120 if (sscanf(buf, "%u", &level) != 1)
8124 * level is always be positive so don't check for
8125 * level < POWERSAVINGS_BALANCE_NONE which is 0
8126 * What happens on 0 or 1 byte write,
8127 * need to check for count as well?
8130 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8134 sched_smt_power_savings = level;
8136 sched_mc_power_savings = level;
8138 reinit_sched_domains();
8143 #ifdef CONFIG_SCHED_MC
8144 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8145 struct sysdev_class_attribute *attr,
8148 return sprintf(page, "%u\n", sched_mc_power_savings);
8150 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8151 struct sysdev_class_attribute *attr,
8152 const char *buf, size_t count)
8154 return sched_power_savings_store(buf, count, 0);
8156 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8157 sched_mc_power_savings_show,
8158 sched_mc_power_savings_store);
8161 #ifdef CONFIG_SCHED_SMT
8162 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8163 struct sysdev_class_attribute *attr,
8166 return sprintf(page, "%u\n", sched_smt_power_savings);
8168 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8169 struct sysdev_class_attribute *attr,
8170 const char *buf, size_t count)
8172 return sched_power_savings_store(buf, count, 1);
8174 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8175 sched_smt_power_savings_show,
8176 sched_smt_power_savings_store);
8179 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8183 #ifdef CONFIG_SCHED_SMT
8185 err = sysfs_create_file(&cls->kset.kobj,
8186 &attr_sched_smt_power_savings.attr);
8188 #ifdef CONFIG_SCHED_MC
8189 if (!err && mc_capable())
8190 err = sysfs_create_file(&cls->kset.kobj,
8191 &attr_sched_mc_power_savings.attr);
8195 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8197 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
8200 * Update cpusets according to cpu_active mask. If cpusets are
8201 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8202 * around partition_sched_domains().
8204 * If we come here as part of a suspend/resume, don't touch cpusets because we
8205 * want to restore it back to its original state upon resume anyway.
8207 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
8211 case CPU_ONLINE_FROZEN:
8212 case CPU_DOWN_FAILED_FROZEN:
8215 * num_cpus_frozen tracks how many CPUs are involved in suspend
8216 * resume sequence. As long as this is not the last online
8217 * operation in the resume sequence, just build a single sched
8218 * domain, ignoring cpusets.
8221 if (likely(num_cpus_frozen)) {
8222 partition_sched_domains(1, NULL, NULL);
8227 * This is the last CPU online operation. So fall through and
8228 * restore the original sched domains by considering the
8229 * cpuset configurations.
8233 case CPU_DOWN_FAILED:
8234 cpuset_update_active_cpus();
8242 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
8246 case CPU_DOWN_PREPARE:
8247 cpuset_update_active_cpus();
8249 case CPU_DOWN_PREPARE_FROZEN:
8251 partition_sched_domains(1, NULL, NULL);
8259 static int update_runtime(struct notifier_block *nfb,
8260 unsigned long action, void *hcpu)
8262 int cpu = (int)(long)hcpu;
8265 case CPU_DOWN_PREPARE:
8266 case CPU_DOWN_PREPARE_FROZEN:
8267 disable_runtime(cpu_rq(cpu));
8270 case CPU_DOWN_FAILED:
8271 case CPU_DOWN_FAILED_FROZEN:
8273 case CPU_ONLINE_FROZEN:
8274 enable_runtime(cpu_rq(cpu));
8282 void __init sched_init_smp(void)
8284 cpumask_var_t non_isolated_cpus;
8286 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8287 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8290 mutex_lock(&sched_domains_mutex);
8291 init_sched_domains(cpu_active_mask);
8292 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8293 if (cpumask_empty(non_isolated_cpus))
8294 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8295 mutex_unlock(&sched_domains_mutex);
8298 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
8299 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
8301 /* RT runtime code needs to handle some hotplug events */
8302 hotcpu_notifier(update_runtime, 0);
8306 /* Move init over to a non-isolated CPU */
8307 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8309 sched_init_granularity();
8310 free_cpumask_var(non_isolated_cpus);
8312 init_sched_rt_class();
8315 void __init sched_init_smp(void)
8317 sched_init_granularity();
8319 #endif /* CONFIG_SMP */
8321 const_debug unsigned int sysctl_timer_migration = 1;
8323 int in_sched_functions(unsigned long addr)
8325 return in_lock_functions(addr) ||
8326 (addr >= (unsigned long)__sched_text_start
8327 && addr < (unsigned long)__sched_text_end);
8330 static void init_cfs_rq(struct cfs_rq *cfs_rq)
8332 cfs_rq->tasks_timeline = RB_ROOT;
8333 INIT_LIST_HEAD(&cfs_rq->tasks);
8334 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8335 #ifndef CONFIG_64BIT
8336 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8340 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8342 struct rt_prio_array *array;
8345 array = &rt_rq->active;
8346 for (i = 0; i < MAX_RT_PRIO; i++) {
8347 INIT_LIST_HEAD(array->queue + i);
8348 __clear_bit(i, array->bitmap);
8350 /* delimiter for bitsearch: */
8351 __set_bit(MAX_RT_PRIO, array->bitmap);
8353 #if defined CONFIG_SMP
8354 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8355 rt_rq->highest_prio.next = MAX_RT_PRIO;
8356 rt_rq->rt_nr_migratory = 0;
8357 rt_rq->overloaded = 0;
8358 plist_head_init(&rt_rq->pushable_tasks);
8362 rt_rq->rt_throttled = 0;
8363 rt_rq->rt_runtime = 0;
8364 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8367 #ifdef CONFIG_FAIR_GROUP_SCHED
8368 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8369 struct sched_entity *se, int cpu,
8370 struct sched_entity *parent)
8372 struct rq *rq = cpu_rq(cpu);
8377 /* allow initial update_cfs_load() to truncate */
8378 cfs_rq->load_stamp = 1;
8380 init_cfs_rq_runtime(cfs_rq);
8382 tg->cfs_rq[cpu] = cfs_rq;
8385 /* se could be NULL for root_task_group */
8390 se->cfs_rq = &rq->cfs;
8392 se->cfs_rq = parent->my_q;
8395 update_load_set(&se->load, 0);
8396 se->parent = parent;
8400 #ifdef CONFIG_RT_GROUP_SCHED
8401 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8402 struct sched_rt_entity *rt_se, int cpu,
8403 struct sched_rt_entity *parent)
8405 struct rq *rq = cpu_rq(cpu);
8407 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8408 rt_rq->rt_nr_boosted = 0;
8412 tg->rt_rq[cpu] = rt_rq;
8413 tg->rt_se[cpu] = rt_se;
8419 rt_se->rt_rq = &rq->rt;
8421 rt_se->rt_rq = parent->my_q;
8423 rt_se->my_q = rt_rq;
8424 rt_se->parent = parent;
8425 INIT_LIST_HEAD(&rt_se->run_list);
8429 void __init sched_init(void)
8432 unsigned long alloc_size = 0, ptr;
8434 #ifdef CONFIG_FAIR_GROUP_SCHED
8435 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8437 #ifdef CONFIG_RT_GROUP_SCHED
8438 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8440 #ifdef CONFIG_CPUMASK_OFFSTACK
8441 alloc_size += num_possible_cpus() * cpumask_size();
8444 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8446 #ifdef CONFIG_FAIR_GROUP_SCHED
8447 root_task_group.se = (struct sched_entity **)ptr;
8448 ptr += nr_cpu_ids * sizeof(void **);
8450 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8451 ptr += nr_cpu_ids * sizeof(void **);
8453 #endif /* CONFIG_FAIR_GROUP_SCHED */
8454 #ifdef CONFIG_RT_GROUP_SCHED
8455 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8456 ptr += nr_cpu_ids * sizeof(void **);
8458 root_task_group.rt_rq = (struct rt_rq **)ptr;
8459 ptr += nr_cpu_ids * sizeof(void **);
8461 #endif /* CONFIG_RT_GROUP_SCHED */
8462 #ifdef CONFIG_CPUMASK_OFFSTACK
8463 for_each_possible_cpu(i) {
8464 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8465 ptr += cpumask_size();
8467 #endif /* CONFIG_CPUMASK_OFFSTACK */
8471 init_defrootdomain();
8474 init_rt_bandwidth(&def_rt_bandwidth,
8475 global_rt_period(), global_rt_runtime());
8477 #ifdef CONFIG_RT_GROUP_SCHED
8478 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8479 global_rt_period(), global_rt_runtime());
8480 #endif /* CONFIG_RT_GROUP_SCHED */
8482 #ifdef CONFIG_CGROUP_SCHED
8483 list_add(&root_task_group.list, &task_groups);
8484 INIT_LIST_HEAD(&root_task_group.children);
8485 autogroup_init(&init_task);
8486 #endif /* CONFIG_CGROUP_SCHED */
8488 for_each_possible_cpu(i) {
8492 raw_spin_lock_init(&rq->lock);
8494 rq->calc_load_active = 0;
8495 rq->calc_load_update = jiffies + LOAD_FREQ;
8496 init_cfs_rq(&rq->cfs);
8497 init_rt_rq(&rq->rt, rq);
8498 #ifdef CONFIG_FAIR_GROUP_SCHED
8499 root_task_group.shares = root_task_group_load;
8500 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8502 * How much cpu bandwidth does root_task_group get?
8504 * In case of task-groups formed thr' the cgroup filesystem, it
8505 * gets 100% of the cpu resources in the system. This overall
8506 * system cpu resource is divided among the tasks of
8507 * root_task_group and its child task-groups in a fair manner,
8508 * based on each entity's (task or task-group's) weight
8509 * (se->load.weight).
8511 * In other words, if root_task_group has 10 tasks of weight
8512 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8513 * then A0's share of the cpu resource is:
8515 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8517 * We achieve this by letting root_task_group's tasks sit
8518 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8520 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8521 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8522 #endif /* CONFIG_FAIR_GROUP_SCHED */
8524 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8525 #ifdef CONFIG_RT_GROUP_SCHED
8526 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8527 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8530 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8531 rq->cpu_load[j] = 0;
8533 rq->last_load_update_tick = jiffies;
8538 rq->cpu_power = SCHED_POWER_SCALE;
8539 rq->post_schedule = 0;
8540 rq->active_balance = 0;
8541 rq->next_balance = jiffies;
8546 rq->avg_idle = 2*sysctl_sched_migration_cost;
8547 rq_attach_root(rq, &def_root_domain);
8549 rq->nohz_balance_kick = 0;
8553 atomic_set(&rq->nr_iowait, 0);
8556 set_load_weight(&init_task);
8558 #ifdef CONFIG_PREEMPT_NOTIFIERS
8559 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8563 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8566 #ifdef CONFIG_RT_MUTEXES
8567 plist_head_init(&init_task.pi_waiters);
8571 * The boot idle thread does lazy MMU switching as well:
8573 atomic_inc(&init_mm.mm_count);
8574 enter_lazy_tlb(&init_mm, current);
8577 * Make us the idle thread. Technically, schedule() should not be
8578 * called from this thread, however somewhere below it might be,
8579 * but because we are the idle thread, we just pick up running again
8580 * when this runqueue becomes "idle".
8582 init_idle(current, smp_processor_id());
8584 calc_load_update = jiffies + LOAD_FREQ;
8587 * During early bootup we pretend to be a normal task:
8589 current->sched_class = &fair_sched_class;
8592 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8594 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8595 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8596 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8597 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8598 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8600 /* May be allocated at isolcpus cmdline parse time */
8601 if (cpu_isolated_map == NULL)
8602 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8605 scheduler_running = 1;
8608 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8609 static inline int preempt_count_equals(int preempt_offset)
8611 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8613 return (nested == preempt_offset);
8616 void __might_sleep(const char *file, int line, int preempt_offset)
8618 static unsigned long prev_jiffy; /* ratelimiting */
8620 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
8621 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8622 system_state != SYSTEM_RUNNING || oops_in_progress)
8624 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8626 prev_jiffy = jiffies;
8629 "BUG: sleeping function called from invalid context at %s:%d\n",
8632 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8633 in_atomic(), irqs_disabled(),
8634 current->pid, current->comm);
8636 debug_show_held_locks(current);
8637 if (irqs_disabled())
8638 print_irqtrace_events(current);
8641 EXPORT_SYMBOL(__might_sleep);
8644 #ifdef CONFIG_MAGIC_SYSRQ
8645 static void normalize_task(struct rq *rq, struct task_struct *p)
8647 const struct sched_class *prev_class = p->sched_class;
8648 int old_prio = p->prio;
8653 deactivate_task(rq, p, 0);
8654 __setscheduler(rq, p, SCHED_NORMAL, 0);
8656 activate_task(rq, p, 0);
8657 resched_task(rq->curr);
8660 check_class_changed(rq, p, prev_class, old_prio);
8663 void normalize_rt_tasks(void)
8665 struct task_struct *g, *p;
8666 unsigned long flags;
8669 read_lock_irqsave(&tasklist_lock, flags);
8670 do_each_thread(g, p) {
8672 * Only normalize user tasks:
8677 p->se.exec_start = 0;
8678 #ifdef CONFIG_SCHEDSTATS
8679 p->se.statistics.wait_start = 0;
8680 p->se.statistics.sleep_start = 0;
8681 p->se.statistics.block_start = 0;
8686 * Renice negative nice level userspace
8689 if (TASK_NICE(p) < 0 && p->mm)
8690 set_user_nice(p, 0);
8694 raw_spin_lock(&p->pi_lock);
8695 rq = __task_rq_lock(p);
8697 normalize_task(rq, p);
8699 __task_rq_unlock(rq);
8700 raw_spin_unlock(&p->pi_lock);
8701 } while_each_thread(g, p);
8703 read_unlock_irqrestore(&tasklist_lock, flags);
8706 #endif /* CONFIG_MAGIC_SYSRQ */
8708 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8710 * These functions are only useful for the IA64 MCA handling, or kdb.
8712 * They can only be called when the whole system has been
8713 * stopped - every CPU needs to be quiescent, and no scheduling
8714 * activity can take place. Using them for anything else would
8715 * be a serious bug, and as a result, they aren't even visible
8716 * under any other configuration.
8720 * curr_task - return the current task for a given cpu.
8721 * @cpu: the processor in question.
8723 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8725 struct task_struct *curr_task(int cpu)
8727 return cpu_curr(cpu);
8730 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8734 * set_curr_task - set the current task for a given cpu.
8735 * @cpu: the processor in question.
8736 * @p: the task pointer to set.
8738 * Description: This function must only be used when non-maskable interrupts
8739 * are serviced on a separate stack. It allows the architecture to switch the
8740 * notion of the current task on a cpu in a non-blocking manner. This function
8741 * must be called with all CPU's synchronized, and interrupts disabled, the
8742 * and caller must save the original value of the current task (see
8743 * curr_task() above) and restore that value before reenabling interrupts and
8744 * re-starting the system.
8746 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8748 void set_curr_task(int cpu, struct task_struct *p)
8755 #ifdef CONFIG_FAIR_GROUP_SCHED
8756 static void free_fair_sched_group(struct task_group *tg)
8760 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8762 for_each_possible_cpu(i) {
8764 kfree(tg->cfs_rq[i]);
8774 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8776 struct cfs_rq *cfs_rq;
8777 struct sched_entity *se;
8780 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8783 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8787 tg->shares = NICE_0_LOAD;
8789 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8791 for_each_possible_cpu(i) {
8792 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8793 GFP_KERNEL, cpu_to_node(i));
8797 se = kzalloc_node(sizeof(struct sched_entity),
8798 GFP_KERNEL, cpu_to_node(i));
8802 init_cfs_rq(cfs_rq);
8803 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8814 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8816 struct rq *rq = cpu_rq(cpu);
8817 unsigned long flags;
8820 * Only empty task groups can be destroyed; so we can speculatively
8821 * check on_list without danger of it being re-added.
8823 if (!tg->cfs_rq[cpu]->on_list)
8826 raw_spin_lock_irqsave(&rq->lock, flags);
8827 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8828 raw_spin_unlock_irqrestore(&rq->lock, flags);
8830 #else /* !CONFIG_FAIR_GROUP_SCHED */
8831 static inline void free_fair_sched_group(struct task_group *tg)
8836 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8841 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8844 #endif /* CONFIG_FAIR_GROUP_SCHED */
8846 #ifdef CONFIG_RT_GROUP_SCHED
8847 static void free_rt_sched_group(struct task_group *tg)
8852 destroy_rt_bandwidth(&tg->rt_bandwidth);
8854 for_each_possible_cpu(i) {
8856 kfree(tg->rt_rq[i]);
8858 kfree(tg->rt_se[i]);
8866 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8868 struct rt_rq *rt_rq;
8869 struct sched_rt_entity *rt_se;
8872 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8875 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8879 init_rt_bandwidth(&tg->rt_bandwidth,
8880 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8882 for_each_possible_cpu(i) {
8883 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8884 GFP_KERNEL, cpu_to_node(i));
8888 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8889 GFP_KERNEL, cpu_to_node(i));
8893 init_rt_rq(rt_rq, cpu_rq(i));
8894 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8895 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8905 #else /* !CONFIG_RT_GROUP_SCHED */
8906 static inline void free_rt_sched_group(struct task_group *tg)
8911 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8915 #endif /* CONFIG_RT_GROUP_SCHED */
8917 #ifdef CONFIG_CGROUP_SCHED
8918 static void free_sched_group(struct task_group *tg)
8920 free_fair_sched_group(tg);
8921 free_rt_sched_group(tg);
8926 /* allocate runqueue etc for a new task group */
8927 struct task_group *sched_create_group(struct task_group *parent)
8929 struct task_group *tg;
8930 unsigned long flags;
8932 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8934 return ERR_PTR(-ENOMEM);
8936 if (!alloc_fair_sched_group(tg, parent))
8939 if (!alloc_rt_sched_group(tg, parent))
8942 spin_lock_irqsave(&task_group_lock, flags);
8943 list_add_rcu(&tg->list, &task_groups);
8945 WARN_ON(!parent); /* root should already exist */
8947 tg->parent = parent;
8948 INIT_LIST_HEAD(&tg->children);
8949 list_add_rcu(&tg->siblings, &parent->children);
8950 spin_unlock_irqrestore(&task_group_lock, flags);
8955 free_sched_group(tg);
8956 return ERR_PTR(-ENOMEM);
8959 /* rcu callback to free various structures associated with a task group */
8960 static void free_sched_group_rcu(struct rcu_head *rhp)
8962 /* now it should be safe to free those cfs_rqs */
8963 free_sched_group(container_of(rhp, struct task_group, rcu));
8966 /* Destroy runqueue etc associated with a task group */
8967 void sched_destroy_group(struct task_group *tg)
8969 unsigned long flags;
8972 /* end participation in shares distribution */
8973 for_each_possible_cpu(i)
8974 unregister_fair_sched_group(tg, i);
8976 spin_lock_irqsave(&task_group_lock, flags);
8977 list_del_rcu(&tg->list);
8978 list_del_rcu(&tg->siblings);
8979 spin_unlock_irqrestore(&task_group_lock, flags);
8981 /* wait for possible concurrent references to cfs_rqs complete */
8982 call_rcu(&tg->rcu, free_sched_group_rcu);
8985 /* change task's runqueue when it moves between groups.
8986 * The caller of this function should have put the task in its new group
8987 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8988 * reflect its new group.
8990 void sched_move_task(struct task_struct *tsk)
8992 struct task_group *tg;
8994 unsigned long flags;
8997 rq = task_rq_lock(tsk, &flags);
8999 running = task_current(rq, tsk);
9003 dequeue_task(rq, tsk, 0);
9004 if (unlikely(running))
9005 tsk->sched_class->put_prev_task(rq, tsk);
9007 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
9008 lockdep_is_held(&tsk->sighand->siglock)),
9009 struct task_group, css);
9010 tg = autogroup_task_group(tsk, tg);
9011 tsk->sched_task_group = tg;
9013 #ifdef CONFIG_FAIR_GROUP_SCHED
9014 if (tsk->sched_class->task_move_group)
9015 tsk->sched_class->task_move_group(tsk, on_rq);
9018 set_task_rq(tsk, task_cpu(tsk));
9020 if (unlikely(running))
9021 tsk->sched_class->set_curr_task(rq);
9023 enqueue_task(rq, tsk, 0);
9025 task_rq_unlock(rq, tsk, &flags);
9027 #endif /* CONFIG_CGROUP_SCHED */
9029 #ifdef CONFIG_FAIR_GROUP_SCHED
9030 static DEFINE_MUTEX(shares_mutex);
9032 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9035 unsigned long flags;
9038 * We can't change the weight of the root cgroup.
9043 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9045 mutex_lock(&shares_mutex);
9046 if (tg->shares == shares)
9049 tg->shares = shares;
9050 for_each_possible_cpu(i) {
9051 struct rq *rq = cpu_rq(i);
9052 struct sched_entity *se;
9055 /* Propagate contribution to hierarchy */
9056 raw_spin_lock_irqsave(&rq->lock, flags);
9057 for_each_sched_entity(se)
9058 update_cfs_shares(group_cfs_rq(se));
9059 raw_spin_unlock_irqrestore(&rq->lock, flags);
9063 mutex_unlock(&shares_mutex);
9067 unsigned long sched_group_shares(struct task_group *tg)
9073 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
9074 static unsigned long to_ratio(u64 period, u64 runtime)
9076 if (runtime == RUNTIME_INF)
9079 return div64_u64(runtime << 20, period);
9083 #ifdef CONFIG_RT_GROUP_SCHED
9085 * Ensure that the real time constraints are schedulable.
9087 static DEFINE_MUTEX(rt_constraints_mutex);
9089 /* Must be called with tasklist_lock held */
9090 static inline int tg_has_rt_tasks(struct task_group *tg)
9092 struct task_struct *g, *p;
9094 do_each_thread(g, p) {
9095 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9097 } while_each_thread(g, p);
9102 struct rt_schedulable_data {
9103 struct task_group *tg;
9108 static int tg_rt_schedulable(struct task_group *tg, void *data)
9110 struct rt_schedulable_data *d = data;
9111 struct task_group *child;
9112 unsigned long total, sum = 0;
9113 u64 period, runtime;
9115 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9116 runtime = tg->rt_bandwidth.rt_runtime;
9119 period = d->rt_period;
9120 runtime = d->rt_runtime;
9124 * Cannot have more runtime than the period.
9126 if (runtime > period && runtime != RUNTIME_INF)
9130 * Ensure we don't starve existing RT tasks.
9132 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9135 total = to_ratio(period, runtime);
9138 * Nobody can have more than the global setting allows.
9140 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9144 * The sum of our children's runtime should not exceed our own.
9146 list_for_each_entry_rcu(child, &tg->children, siblings) {
9147 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9148 runtime = child->rt_bandwidth.rt_runtime;
9150 if (child == d->tg) {
9151 period = d->rt_period;
9152 runtime = d->rt_runtime;
9155 sum += to_ratio(period, runtime);
9164 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9168 struct rt_schedulable_data data = {
9170 .rt_period = period,
9171 .rt_runtime = runtime,
9175 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
9181 static int tg_set_rt_bandwidth(struct task_group *tg,
9182 u64 rt_period, u64 rt_runtime)
9186 mutex_lock(&rt_constraints_mutex);
9187 read_lock(&tasklist_lock);
9188 err = __rt_schedulable(tg, rt_period, rt_runtime);
9192 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9193 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9194 tg->rt_bandwidth.rt_runtime = rt_runtime;
9196 for_each_possible_cpu(i) {
9197 struct rt_rq *rt_rq = tg->rt_rq[i];
9199 raw_spin_lock(&rt_rq->rt_runtime_lock);
9200 rt_rq->rt_runtime = rt_runtime;
9201 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9203 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9205 read_unlock(&tasklist_lock);
9206 mutex_unlock(&rt_constraints_mutex);
9211 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9213 u64 rt_runtime, rt_period;
9215 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9216 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9217 if (rt_runtime_us < 0)
9218 rt_runtime = RUNTIME_INF;
9220 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9223 long sched_group_rt_runtime(struct task_group *tg)
9227 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9230 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9231 do_div(rt_runtime_us, NSEC_PER_USEC);
9232 return rt_runtime_us;
9235 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9237 u64 rt_runtime, rt_period;
9239 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9240 rt_runtime = tg->rt_bandwidth.rt_runtime;
9245 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9248 long sched_group_rt_period(struct task_group *tg)
9252 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9253 do_div(rt_period_us, NSEC_PER_USEC);
9254 return rt_period_us;
9257 static int sched_rt_global_constraints(void)
9259 u64 runtime, period;
9262 if (sysctl_sched_rt_period <= 0)
9265 runtime = global_rt_runtime();
9266 period = global_rt_period();
9269 * Sanity check on the sysctl variables.
9271 if (runtime > period && runtime != RUNTIME_INF)
9274 mutex_lock(&rt_constraints_mutex);
9275 read_lock(&tasklist_lock);
9276 ret = __rt_schedulable(NULL, 0, 0);
9277 read_unlock(&tasklist_lock);
9278 mutex_unlock(&rt_constraints_mutex);
9283 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9285 /* Don't accept realtime tasks when there is no way for them to run */
9286 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9292 #else /* !CONFIG_RT_GROUP_SCHED */
9293 static int sched_rt_global_constraints(void)
9295 unsigned long flags;
9298 if (sysctl_sched_rt_period <= 0)
9302 * There's always some RT tasks in the root group
9303 * -- migration, kstopmachine etc..
9305 if (sysctl_sched_rt_runtime == 0)
9308 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9309 for_each_possible_cpu(i) {
9310 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9312 raw_spin_lock(&rt_rq->rt_runtime_lock);
9313 rt_rq->rt_runtime = global_rt_runtime();
9314 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9316 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9320 #endif /* CONFIG_RT_GROUP_SCHED */
9322 int sched_rt_handler(struct ctl_table *table, int write,
9323 void __user *buffer, size_t *lenp,
9327 int old_period, old_runtime;
9328 static DEFINE_MUTEX(mutex);
9331 old_period = sysctl_sched_rt_period;
9332 old_runtime = sysctl_sched_rt_runtime;
9334 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9336 if (!ret && write) {
9337 ret = sched_rt_global_constraints();
9339 sysctl_sched_rt_period = old_period;
9340 sysctl_sched_rt_runtime = old_runtime;
9342 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9343 def_rt_bandwidth.rt_period =
9344 ns_to_ktime(global_rt_period());
9347 mutex_unlock(&mutex);
9352 #ifdef CONFIG_CGROUP_SCHED
9354 /* return corresponding task_group object of a cgroup */
9355 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9357 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9358 struct task_group, css);
9361 static struct cgroup_subsys_state *
9362 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9364 struct task_group *tg, *parent;
9366 if (!cgrp->parent) {
9367 /* This is early initialization for the top cgroup */
9368 return &root_task_group.css;
9371 parent = cgroup_tg(cgrp->parent);
9372 tg = sched_create_group(parent);
9374 return ERR_PTR(-ENOMEM);
9380 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9382 struct task_group *tg = cgroup_tg(cgrp);
9384 sched_destroy_group(tg);
9388 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9390 #ifdef CONFIG_RT_GROUP_SCHED
9391 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9394 /* We don't support RT-tasks being in separate groups */
9395 if (tsk->sched_class != &fair_sched_class)
9402 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9404 sched_move_task(tsk);
9408 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9409 struct cgroup *old_cgrp, struct task_struct *task)
9412 * cgroup_exit() is called in the copy_process() failure path.
9413 * Ignore this case since the task hasn't ran yet, this avoids
9414 * trying to poke a half freed task state from generic code.
9416 if (!(task->flags & PF_EXITING))
9419 sched_move_task(task);
9422 #ifdef CONFIG_FAIR_GROUP_SCHED
9423 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9426 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9429 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9431 struct task_group *tg = cgroup_tg(cgrp);
9433 return (u64) scale_load_down(tg->shares);
9436 #ifdef CONFIG_CFS_BANDWIDTH
9437 static DEFINE_MUTEX(cfs_constraints_mutex);
9439 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9440 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9442 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9444 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9446 int i, ret = 0, runtime_enabled;
9447 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9449 if (tg == &root_task_group)
9453 * Ensure we have at some amount of bandwidth every period. This is
9454 * to prevent reaching a state of large arrears when throttled via
9455 * entity_tick() resulting in prolonged exit starvation.
9457 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9461 * Likewise, bound things on the otherside by preventing insane quota
9462 * periods. This also allows us to normalize in computing quota
9465 if (period > max_cfs_quota_period)
9468 mutex_lock(&cfs_constraints_mutex);
9469 ret = __cfs_schedulable(tg, period, quota);
9473 runtime_enabled = quota != RUNTIME_INF;
9474 raw_spin_lock_irq(&cfs_b->lock);
9475 cfs_b->period = ns_to_ktime(period);
9476 cfs_b->quota = quota;
9478 __refill_cfs_bandwidth_runtime(cfs_b);
9479 /* restart the period timer (if active) to handle new period expiry */
9480 if (runtime_enabled && cfs_b->timer_active) {
9481 /* force a reprogram */
9482 cfs_b->timer_active = 0;
9483 __start_cfs_bandwidth(cfs_b);
9485 raw_spin_unlock_irq(&cfs_b->lock);
9487 for_each_possible_cpu(i) {
9488 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9489 struct rq *rq = rq_of(cfs_rq);
9491 raw_spin_lock_irq(&rq->lock);
9492 cfs_rq->runtime_enabled = runtime_enabled;
9493 cfs_rq->runtime_remaining = 0;
9495 if (cfs_rq_throttled(cfs_rq))
9496 unthrottle_cfs_rq(cfs_rq);
9497 raw_spin_unlock_irq(&rq->lock);
9500 mutex_unlock(&cfs_constraints_mutex);
9505 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9509 period = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9510 if (cfs_quota_us < 0)
9511 quota = RUNTIME_INF;
9513 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9515 return tg_set_cfs_bandwidth(tg, period, quota);
9518 long tg_get_cfs_quota(struct task_group *tg)
9522 if (tg_cfs_bandwidth(tg)->quota == RUNTIME_INF)
9525 quota_us = tg_cfs_bandwidth(tg)->quota;
9526 do_div(quota_us, NSEC_PER_USEC);
9531 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9535 period = (u64)cfs_period_us * NSEC_PER_USEC;
9536 quota = tg_cfs_bandwidth(tg)->quota;
9541 return tg_set_cfs_bandwidth(tg, period, quota);
9544 long tg_get_cfs_period(struct task_group *tg)
9548 cfs_period_us = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9549 do_div(cfs_period_us, NSEC_PER_USEC);
9551 return cfs_period_us;
9554 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
9556 return tg_get_cfs_quota(cgroup_tg(cgrp));
9559 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
9562 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
9565 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
9567 return tg_get_cfs_period(cgroup_tg(cgrp));
9570 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9573 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
9576 struct cfs_schedulable_data {
9577 struct task_group *tg;
9582 * normalize group quota/period to be quota/max_period
9583 * note: units are usecs
9585 static u64 normalize_cfs_quota(struct task_group *tg,
9586 struct cfs_schedulable_data *d)
9594 period = tg_get_cfs_period(tg);
9595 quota = tg_get_cfs_quota(tg);
9598 /* note: these should typically be equivalent */
9599 if (quota == RUNTIME_INF || quota == -1)
9602 return to_ratio(period, quota);
9605 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9607 struct cfs_schedulable_data *d = data;
9608 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9609 s64 quota = 0, parent_quota = -1;
9612 quota = RUNTIME_INF;
9614 struct cfs_bandwidth *parent_b = tg_cfs_bandwidth(tg->parent);
9616 quota = normalize_cfs_quota(tg, d);
9617 parent_quota = parent_b->hierarchal_quota;
9620 * ensure max(child_quota) <= parent_quota, inherit when no
9623 if (quota == RUNTIME_INF)
9624 quota = parent_quota;
9625 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9628 cfs_b->hierarchal_quota = quota;
9633 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9636 struct cfs_schedulable_data data = {
9642 if (quota != RUNTIME_INF) {
9643 do_div(data.period, NSEC_PER_USEC);
9644 do_div(data.quota, NSEC_PER_USEC);
9648 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9654 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
9655 struct cgroup_map_cb *cb)
9657 struct task_group *tg = cgroup_tg(cgrp);
9658 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9660 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
9661 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
9662 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
9666 #endif /* CONFIG_CFS_BANDWIDTH */
9667 #endif /* CONFIG_FAIR_GROUP_SCHED */
9669 #ifdef CONFIG_RT_GROUP_SCHED
9670 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9673 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9676 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9678 return sched_group_rt_runtime(cgroup_tg(cgrp));
9681 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9684 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9687 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9689 return sched_group_rt_period(cgroup_tg(cgrp));
9691 #endif /* CONFIG_RT_GROUP_SCHED */
9693 static struct cftype cpu_files[] = {
9694 #ifdef CONFIG_FAIR_GROUP_SCHED
9697 .read_u64 = cpu_shares_read_u64,
9698 .write_u64 = cpu_shares_write_u64,
9701 #ifdef CONFIG_CFS_BANDWIDTH
9703 .name = "cfs_quota_us",
9704 .read_s64 = cpu_cfs_quota_read_s64,
9705 .write_s64 = cpu_cfs_quota_write_s64,
9708 .name = "cfs_period_us",
9709 .read_u64 = cpu_cfs_period_read_u64,
9710 .write_u64 = cpu_cfs_period_write_u64,
9714 .read_map = cpu_stats_show,
9717 #ifdef CONFIG_RT_GROUP_SCHED
9719 .name = "rt_runtime_us",
9720 .read_s64 = cpu_rt_runtime_read,
9721 .write_s64 = cpu_rt_runtime_write,
9724 .name = "rt_period_us",
9725 .read_u64 = cpu_rt_period_read_uint,
9726 .write_u64 = cpu_rt_period_write_uint,
9731 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9733 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9736 struct cgroup_subsys cpu_cgroup_subsys = {
9738 .create = cpu_cgroup_create,
9739 .destroy = cpu_cgroup_destroy,
9740 .can_attach_task = cpu_cgroup_can_attach_task,
9741 .attach_task = cpu_cgroup_attach_task,
9742 .exit = cpu_cgroup_exit,
9743 .populate = cpu_cgroup_populate,
9744 .subsys_id = cpu_cgroup_subsys_id,
9748 #endif /* CONFIG_CGROUP_SCHED */
9750 #ifdef CONFIG_CGROUP_CPUACCT
9753 * CPU accounting code for task groups.
9755 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9756 * (balbir@in.ibm.com).
9759 /* track cpu usage of a group of tasks and its child groups */
9761 struct cgroup_subsys_state css;
9762 /* cpuusage holds pointer to a u64-type object on every cpu */
9763 u64 __percpu *cpuusage;
9764 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9765 struct cpuacct *parent;
9768 struct cgroup_subsys cpuacct_subsys;
9770 /* return cpu accounting group corresponding to this container */
9771 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9773 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9774 struct cpuacct, css);
9777 /* return cpu accounting group to which this task belongs */
9778 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9780 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9781 struct cpuacct, css);
9784 /* create a new cpu accounting group */
9785 static struct cgroup_subsys_state *cpuacct_create(
9786 struct cgroup_subsys *ss, struct cgroup *cgrp)
9788 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9794 ca->cpuusage = alloc_percpu(u64);
9798 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9799 if (percpu_counter_init(&ca->cpustat[i], 0))
9800 goto out_free_counters;
9803 ca->parent = cgroup_ca(cgrp->parent);
9809 percpu_counter_destroy(&ca->cpustat[i]);
9810 free_percpu(ca->cpuusage);
9814 return ERR_PTR(-ENOMEM);
9817 /* destroy an existing cpu accounting group */
9819 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9821 struct cpuacct *ca = cgroup_ca(cgrp);
9824 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9825 percpu_counter_destroy(&ca->cpustat[i]);
9826 free_percpu(ca->cpuusage);
9830 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9832 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9835 #ifndef CONFIG_64BIT
9837 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9839 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9841 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9849 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9851 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9853 #ifndef CONFIG_64BIT
9855 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9857 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9859 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9865 /* return total cpu usage (in nanoseconds) of a group */
9866 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9868 struct cpuacct *ca = cgroup_ca(cgrp);
9869 u64 totalcpuusage = 0;
9872 for_each_present_cpu(i)
9873 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9875 return totalcpuusage;
9878 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9881 struct cpuacct *ca = cgroup_ca(cgrp);
9890 for_each_present_cpu(i)
9891 cpuacct_cpuusage_write(ca, i, 0);
9897 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9900 struct cpuacct *ca = cgroup_ca(cgroup);
9904 for_each_present_cpu(i) {
9905 percpu = cpuacct_cpuusage_read(ca, i);
9906 seq_printf(m, "%llu ", (unsigned long long) percpu);
9908 seq_printf(m, "\n");
9912 static const char *cpuacct_stat_desc[] = {
9913 [CPUACCT_STAT_USER] = "user",
9914 [CPUACCT_STAT_SYSTEM] = "system",
9917 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9918 struct cgroup_map_cb *cb)
9920 struct cpuacct *ca = cgroup_ca(cgrp);
9923 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9924 s64 val = percpu_counter_read(&ca->cpustat[i]);
9925 val = cputime64_to_clock_t(val);
9926 cb->fill(cb, cpuacct_stat_desc[i], val);
9931 static struct cftype files[] = {
9934 .read_u64 = cpuusage_read,
9935 .write_u64 = cpuusage_write,
9938 .name = "usage_percpu",
9939 .read_seq_string = cpuacct_percpu_seq_read,
9943 .read_map = cpuacct_stats_show,
9947 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9949 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9953 * charge this task's execution time to its accounting group.
9955 * called with rq->lock held.
9957 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9962 if (unlikely(!cpuacct_subsys.active))
9965 cpu = task_cpu(tsk);
9971 for (; ca; ca = ca->parent) {
9972 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9973 *cpuusage += cputime;
9980 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9981 * in cputime_t units. As a result, cpuacct_update_stats calls
9982 * percpu_counter_add with values large enough to always overflow the
9983 * per cpu batch limit causing bad SMP scalability.
9985 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9986 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9987 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9990 #define CPUACCT_BATCH \
9991 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9993 #define CPUACCT_BATCH 0
9997 * Charge the system/user time to the task's accounting group.
9999 static void cpuacct_update_stats(struct task_struct *tsk,
10000 enum cpuacct_stat_index idx, cputime_t val)
10002 struct cpuacct *ca;
10003 int batch = CPUACCT_BATCH;
10005 if (unlikely(!cpuacct_subsys.active))
10012 __percpu_counter_add(&ca->cpustat[idx], val, batch);
10018 struct cgroup_subsys cpuacct_subsys = {
10020 .create = cpuacct_create,
10021 .destroy = cpuacct_destroy,
10022 .populate = cpuacct_populate,
10023 .subsys_id = cpuacct_subsys_id,
10025 #endif /* CONFIG_CGROUP_CPUACCT */