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>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
82 #include "sched_cpupri.h"
83 #include "workqueue_sched.h"
84 #include "sched_autogroup.h"
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/sched.h>
90 * Convert user-nice values [ -20 ... 0 ... 19 ]
91 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
94 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
95 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
96 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
99 * 'User priority' is the nice value converted to something we
100 * can work with better when scaling various scheduler parameters,
101 * it's a [ 0 ... 39 ] range.
103 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
104 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
105 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
108 * Helpers for converting nanosecond timing to jiffy resolution
110 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
112 #define NICE_0_LOAD SCHED_LOAD_SCALE
113 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
116 * These are the 'tuning knobs' of the scheduler:
118 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
119 * Timeslices get refilled after they expire.
121 #define DEF_TIMESLICE (100 * HZ / 1000)
124 * single value that denotes runtime == period, ie unlimited time.
126 #define RUNTIME_INF ((u64)~0ULL)
128 static inline int rt_policy(int policy)
130 if (policy == SCHED_FIFO || policy == SCHED_RR)
135 static inline int task_has_rt_policy(struct task_struct *p)
137 return rt_policy(p->policy);
141 * This is the priority-queue data structure of the RT scheduling class:
143 struct rt_prio_array {
144 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
145 struct list_head queue[MAX_RT_PRIO];
148 struct rt_bandwidth {
149 /* nests inside the rq lock: */
150 raw_spinlock_t rt_runtime_lock;
153 struct hrtimer rt_period_timer;
156 static struct rt_bandwidth def_rt_bandwidth;
158 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
160 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
162 struct rt_bandwidth *rt_b =
163 container_of(timer, struct rt_bandwidth, rt_period_timer);
169 now = hrtimer_cb_get_time(timer);
170 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
175 idle = do_sched_rt_period_timer(rt_b, overrun);
178 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
182 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
184 rt_b->rt_period = ns_to_ktime(period);
185 rt_b->rt_runtime = runtime;
187 raw_spin_lock_init(&rt_b->rt_runtime_lock);
189 hrtimer_init(&rt_b->rt_period_timer,
190 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
191 rt_b->rt_period_timer.function = sched_rt_period_timer;
194 static inline int rt_bandwidth_enabled(void)
196 return sysctl_sched_rt_runtime >= 0;
199 static void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
202 ktime_t soft, hard, now;
205 if (hrtimer_active(period_timer))
208 now = hrtimer_cb_get_time(period_timer);
209 hrtimer_forward(period_timer, now, period);
211 soft = hrtimer_get_softexpires(period_timer);
212 hard = hrtimer_get_expires(period_timer);
213 delta = ktime_to_ns(ktime_sub(hard, soft));
214 __hrtimer_start_range_ns(period_timer, soft, delta,
215 HRTIMER_MODE_ABS_PINNED, 0);
219 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
221 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
224 if (hrtimer_active(&rt_b->rt_period_timer))
227 raw_spin_lock(&rt_b->rt_runtime_lock);
228 start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
229 raw_spin_unlock(&rt_b->rt_runtime_lock);
232 #ifdef CONFIG_RT_GROUP_SCHED
233 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
235 hrtimer_cancel(&rt_b->rt_period_timer);
240 * sched_domains_mutex serializes calls to init_sched_domains,
241 * detach_destroy_domains and partition_sched_domains.
243 static DEFINE_MUTEX(sched_domains_mutex);
245 #ifdef CONFIG_CGROUP_SCHED
247 #include <linux/cgroup.h>
251 static LIST_HEAD(task_groups);
253 struct cfs_bandwidth {
254 #ifdef CONFIG_CFS_BANDWIDTH
258 s64 hierarchal_quota;
261 int idle, timer_active;
262 struct hrtimer period_timer;
263 struct list_head throttled_cfs_rq;
268 /* task group related information */
270 struct cgroup_subsys_state css;
272 #ifdef CONFIG_FAIR_GROUP_SCHED
273 /* schedulable entities of this group on each cpu */
274 struct sched_entity **se;
275 /* runqueue "owned" by this group on each cpu */
276 struct cfs_rq **cfs_rq;
277 unsigned long shares;
279 atomic_t load_weight;
282 #ifdef CONFIG_RT_GROUP_SCHED
283 struct sched_rt_entity **rt_se;
284 struct rt_rq **rt_rq;
286 struct rt_bandwidth rt_bandwidth;
290 struct list_head list;
292 struct task_group *parent;
293 struct list_head siblings;
294 struct list_head children;
296 #ifdef CONFIG_SCHED_AUTOGROUP
297 struct autogroup *autogroup;
300 struct cfs_bandwidth cfs_bandwidth;
303 /* task_group_lock serializes the addition/removal of task groups */
304 static DEFINE_SPINLOCK(task_group_lock);
306 #ifdef CONFIG_FAIR_GROUP_SCHED
308 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
311 * A weight of 0 or 1 can cause arithmetics problems.
312 * A weight of a cfs_rq is the sum of weights of which entities
313 * are queued on this cfs_rq, so a weight of a entity should not be
314 * too large, so as the shares value of a task group.
315 * (The default weight is 1024 - so there's no practical
316 * limitation from this.)
318 #define MIN_SHARES (1UL << 1)
319 #define MAX_SHARES (1UL << 18)
321 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
324 /* Default task group.
325 * Every task in system belong to this group at bootup.
327 struct task_group root_task_group;
329 #endif /* CONFIG_CGROUP_SCHED */
331 /* CFS-related fields in a runqueue */
333 struct load_weight load;
334 unsigned long nr_running, h_nr_running;
339 u64 min_vruntime_copy;
342 struct rb_root tasks_timeline;
343 struct rb_node *rb_leftmost;
345 struct list_head tasks;
346 struct list_head *balance_iterator;
349 * 'curr' points to currently running entity on this cfs_rq.
350 * It is set to NULL otherwise (i.e when none are currently running).
352 struct sched_entity *curr, *next, *last, *skip;
354 #ifdef CONFIG_SCHED_DEBUG
355 unsigned int nr_spread_over;
358 #ifdef CONFIG_FAIR_GROUP_SCHED
359 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
362 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
363 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
364 * (like users, containers etc.)
366 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
367 * list is used during load balance.
370 struct list_head leaf_cfs_rq_list;
371 struct task_group *tg; /* group that "owns" this runqueue */
375 * the part of load.weight contributed by tasks
377 unsigned long task_weight;
380 * h_load = weight * f(tg)
382 * Where f(tg) is the recursive weight fraction assigned to
385 unsigned long h_load;
388 * Maintaining per-cpu shares distribution for group scheduling
390 * load_stamp is the last time we updated the load average
391 * load_last is the last time we updated the load average and saw load
392 * load_unacc_exec_time is currently unaccounted execution time
396 u64 load_stamp, load_last, load_unacc_exec_time;
398 unsigned long load_contribution;
400 #ifdef CONFIG_CFS_BANDWIDTH
403 s64 runtime_remaining;
405 int throttled, throttle_count;
406 struct list_head throttled_list;
411 #ifdef CONFIG_FAIR_GROUP_SCHED
412 #ifdef CONFIG_CFS_BANDWIDTH
413 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
415 return &tg->cfs_bandwidth;
418 static inline u64 default_cfs_period(void);
419 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
421 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
423 struct cfs_bandwidth *cfs_b =
424 container_of(timer, struct cfs_bandwidth, period_timer);
430 now = hrtimer_cb_get_time(timer);
431 overrun = hrtimer_forward(timer, now, cfs_b->period);
436 idle = do_sched_cfs_period_timer(cfs_b, overrun);
439 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
442 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
444 raw_spin_lock_init(&cfs_b->lock);
446 cfs_b->quota = RUNTIME_INF;
447 cfs_b->period = ns_to_ktime(default_cfs_period());
449 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
450 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
451 cfs_b->period_timer.function = sched_cfs_period_timer;
454 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
456 cfs_rq->runtime_enabled = 0;
457 INIT_LIST_HEAD(&cfs_rq->throttled_list);
460 /* requires cfs_b->lock, may release to reprogram timer */
461 static void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
464 * The timer may be active because we're trying to set a new bandwidth
465 * period or because we're racing with the tear-down path
466 * (timer_active==0 becomes visible before the hrtimer call-back
467 * terminates). In either case we ensure that it's re-programmed
469 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
470 raw_spin_unlock(&cfs_b->lock);
471 /* ensure cfs_b->lock is available while we wait */
472 hrtimer_cancel(&cfs_b->period_timer);
474 raw_spin_lock(&cfs_b->lock);
475 /* if someone else restarted the timer then we're done */
476 if (cfs_b->timer_active)
480 cfs_b->timer_active = 1;
481 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
484 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
486 hrtimer_cancel(&cfs_b->period_timer);
489 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
490 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
491 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
493 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
497 #endif /* CONFIG_CFS_BANDWIDTH */
498 #endif /* CONFIG_FAIR_GROUP_SCHED */
500 /* Real-Time classes' related field in a runqueue: */
502 struct rt_prio_array active;
503 unsigned long rt_nr_running;
504 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
506 int curr; /* highest queued rt task prio */
508 int next; /* next highest */
513 unsigned long rt_nr_migratory;
514 unsigned long rt_nr_total;
516 struct plist_head pushable_tasks;
521 /* Nests inside the rq lock: */
522 raw_spinlock_t rt_runtime_lock;
524 #ifdef CONFIG_RT_GROUP_SCHED
525 unsigned long rt_nr_boosted;
528 struct list_head leaf_rt_rq_list;
529 struct task_group *tg;
536 * We add the notion of a root-domain which will be used to define per-domain
537 * variables. Each exclusive cpuset essentially defines an island domain by
538 * fully partitioning the member cpus from any other cpuset. Whenever a new
539 * exclusive cpuset is created, we also create and attach a new root-domain
548 cpumask_var_t online;
551 * The "RT overload" flag: it gets set if a CPU has more than
552 * one runnable RT task.
554 cpumask_var_t rto_mask;
555 struct cpupri cpupri;
559 * By default the system creates a single root-domain with all cpus as
560 * members (mimicking the global state we have today).
562 static struct root_domain def_root_domain;
564 #endif /* CONFIG_SMP */
567 * This is the main, per-CPU runqueue data structure.
569 * Locking rule: those places that want to lock multiple runqueues
570 * (such as the load balancing or the thread migration code), lock
571 * acquire operations must be ordered by ascending &runqueue.
578 * nr_running and cpu_load should be in the same cacheline because
579 * remote CPUs use both these fields when doing load calculation.
581 unsigned long nr_running;
582 #define CPU_LOAD_IDX_MAX 5
583 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
584 unsigned long last_load_update_tick;
587 unsigned char nohz_balance_kick;
589 int skip_clock_update;
591 /* capture load from *all* tasks on this cpu: */
592 struct load_weight load;
593 unsigned long nr_load_updates;
599 #ifdef CONFIG_FAIR_GROUP_SCHED
600 /* list of leaf cfs_rq on this cpu: */
601 struct list_head leaf_cfs_rq_list;
603 #ifdef CONFIG_RT_GROUP_SCHED
604 struct list_head leaf_rt_rq_list;
608 * This is part of a global counter where only the total sum
609 * over all CPUs matters. A task can increase this counter on
610 * one CPU and if it got migrated afterwards it may decrease
611 * it on another CPU. Always updated under the runqueue lock:
613 unsigned long nr_uninterruptible;
615 struct task_struct *curr, *idle, *stop;
616 unsigned long next_balance;
617 struct mm_struct *prev_mm;
625 struct root_domain *rd;
626 struct sched_domain *sd;
628 unsigned long cpu_power;
630 unsigned char idle_at_tick;
631 /* For active balancing */
635 struct cpu_stop_work active_balance_work;
636 /* cpu of this runqueue: */
646 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
649 #ifdef CONFIG_PARAVIRT
652 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
653 u64 prev_steal_time_rq;
656 /* calc_load related fields */
657 unsigned long calc_load_update;
658 long calc_load_active;
660 #ifdef CONFIG_SCHED_HRTICK
662 int hrtick_csd_pending;
663 struct call_single_data hrtick_csd;
665 struct hrtimer hrtick_timer;
668 #ifdef CONFIG_SCHEDSTATS
670 struct sched_info rq_sched_info;
671 unsigned long long rq_cpu_time;
672 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
674 /* sys_sched_yield() stats */
675 unsigned int yld_count;
677 /* schedule() stats */
678 unsigned int sched_switch;
679 unsigned int sched_count;
680 unsigned int sched_goidle;
682 /* try_to_wake_up() stats */
683 unsigned int ttwu_count;
684 unsigned int ttwu_local;
688 struct task_struct *wake_list;
692 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
695 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
697 static inline int cpu_of(struct rq *rq)
706 #define rcu_dereference_check_sched_domain(p) \
707 rcu_dereference_check((p), \
708 lockdep_is_held(&sched_domains_mutex))
711 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
712 * See detach_destroy_domains: synchronize_sched for details.
714 * The domain tree of any CPU may only be accessed from within
715 * preempt-disabled sections.
717 #define for_each_domain(cpu, __sd) \
718 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
720 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
721 #define this_rq() (&__get_cpu_var(runqueues))
722 #define task_rq(p) cpu_rq(task_cpu(p))
723 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
724 #define raw_rq() (&__raw_get_cpu_var(runqueues))
726 #ifdef CONFIG_CGROUP_SCHED
729 * Return the group to which this tasks belongs.
731 * We use task_subsys_state_check() and extend the RCU verification with
732 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
733 * task it moves into the cgroup. Therefore by holding either of those locks,
734 * we pin the task to the current cgroup.
736 static inline struct task_group *task_group(struct task_struct *p)
738 struct task_group *tg;
739 struct cgroup_subsys_state *css;
741 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
742 lockdep_is_held(&p->pi_lock) ||
743 lockdep_is_held(&task_rq(p)->lock));
744 tg = container_of(css, struct task_group, css);
746 return autogroup_task_group(p, tg);
749 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
750 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
752 #ifdef CONFIG_FAIR_GROUP_SCHED
753 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
754 p->se.parent = task_group(p)->se[cpu];
757 #ifdef CONFIG_RT_GROUP_SCHED
758 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
759 p->rt.parent = task_group(p)->rt_se[cpu];
763 #else /* CONFIG_CGROUP_SCHED */
765 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
766 static inline struct task_group *task_group(struct task_struct *p)
771 #endif /* CONFIG_CGROUP_SCHED */
773 static void update_rq_clock_task(struct rq *rq, s64 delta);
775 static void update_rq_clock(struct rq *rq)
779 if (rq->skip_clock_update > 0)
782 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
784 update_rq_clock_task(rq, delta);
788 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
790 #ifdef CONFIG_SCHED_DEBUG
791 # define const_debug __read_mostly
793 # define const_debug static const
797 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
798 * @cpu: the processor in question.
800 * This interface allows printk to be called with the runqueue lock
801 * held and know whether or not it is OK to wake up the klogd.
803 int runqueue_is_locked(int cpu)
805 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
809 * Debugging: various feature bits
812 #define SCHED_FEAT(name, enabled) \
813 __SCHED_FEAT_##name ,
816 #include "sched_features.h"
821 #define SCHED_FEAT(name, enabled) \
822 (1UL << __SCHED_FEAT_##name) * enabled |
824 const_debug unsigned int sysctl_sched_features =
825 #include "sched_features.h"
830 #ifdef CONFIG_SCHED_DEBUG
831 #define SCHED_FEAT(name, enabled) \
834 static __read_mostly char *sched_feat_names[] = {
835 #include "sched_features.h"
841 static int sched_feat_show(struct seq_file *m, void *v)
845 for (i = 0; sched_feat_names[i]; i++) {
846 if (!(sysctl_sched_features & (1UL << i)))
848 seq_printf(m, "%s ", sched_feat_names[i]);
856 sched_feat_write(struct file *filp, const char __user *ubuf,
857 size_t cnt, loff_t *ppos)
867 if (copy_from_user(&buf, ubuf, cnt))
873 if (strncmp(cmp, "NO_", 3) == 0) {
878 for (i = 0; sched_feat_names[i]; i++) {
879 if (strcmp(cmp, sched_feat_names[i]) == 0) {
881 sysctl_sched_features &= ~(1UL << i);
883 sysctl_sched_features |= (1UL << i);
888 if (!sched_feat_names[i])
896 static int sched_feat_open(struct inode *inode, struct file *filp)
898 return single_open(filp, sched_feat_show, NULL);
901 static const struct file_operations sched_feat_fops = {
902 .open = sched_feat_open,
903 .write = sched_feat_write,
906 .release = single_release,
909 static __init int sched_init_debug(void)
911 debugfs_create_file("sched_features", 0644, NULL, NULL,
916 late_initcall(sched_init_debug);
920 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
923 * Number of tasks to iterate in a single balance run.
924 * Limited because this is done with IRQs disabled.
926 const_debug unsigned int sysctl_sched_nr_migrate = 32;
929 * period over which we average the RT time consumption, measured
934 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
937 * period over which we measure -rt task cpu usage in us.
940 unsigned int sysctl_sched_rt_period = 1000000;
942 static __read_mostly int scheduler_running;
945 * part of the period that we allow rt tasks to run in us.
948 int sysctl_sched_rt_runtime = 950000;
950 static inline u64 global_rt_period(void)
952 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
955 static inline u64 global_rt_runtime(void)
957 if (sysctl_sched_rt_runtime < 0)
960 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
963 #ifndef prepare_arch_switch
964 # define prepare_arch_switch(next) do { } while (0)
966 #ifndef finish_arch_switch
967 # define finish_arch_switch(prev) do { } while (0)
970 static inline int task_current(struct rq *rq, struct task_struct *p)
972 return rq->curr == p;
975 static inline int task_running(struct rq *rq, struct task_struct *p)
980 return task_current(rq, p);
984 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
985 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
989 * We can optimise this out completely for !SMP, because the
990 * SMP rebalancing from interrupt is the only thing that cares
997 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1001 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1002 * We must ensure this doesn't happen until the switch is completely
1008 #ifdef CONFIG_DEBUG_SPINLOCK
1009 /* this is a valid case when another task releases the spinlock */
1010 rq->lock.owner = current;
1013 * If we are tracking spinlock dependencies then we have to
1014 * fix up the runqueue lock - which gets 'carried over' from
1015 * prev into current:
1017 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1019 raw_spin_unlock_irq(&rq->lock);
1022 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1023 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1027 * We can optimise this out completely for !SMP, because the
1028 * SMP rebalancing from interrupt is the only thing that cares
1033 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1034 raw_spin_unlock_irq(&rq->lock);
1036 raw_spin_unlock(&rq->lock);
1040 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1044 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1045 * We must ensure this doesn't happen until the switch is completely
1051 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1055 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1058 * __task_rq_lock - lock the rq @p resides on.
1060 static inline struct rq *__task_rq_lock(struct task_struct *p)
1061 __acquires(rq->lock)
1065 lockdep_assert_held(&p->pi_lock);
1069 raw_spin_lock(&rq->lock);
1070 if (likely(rq == task_rq(p)))
1072 raw_spin_unlock(&rq->lock);
1077 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1079 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1080 __acquires(p->pi_lock)
1081 __acquires(rq->lock)
1086 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1088 raw_spin_lock(&rq->lock);
1089 if (likely(rq == task_rq(p)))
1091 raw_spin_unlock(&rq->lock);
1092 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1096 static void __task_rq_unlock(struct rq *rq)
1097 __releases(rq->lock)
1099 raw_spin_unlock(&rq->lock);
1103 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1104 __releases(rq->lock)
1105 __releases(p->pi_lock)
1107 raw_spin_unlock(&rq->lock);
1108 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1112 * this_rq_lock - lock this runqueue and disable interrupts.
1114 static struct rq *this_rq_lock(void)
1115 __acquires(rq->lock)
1119 local_irq_disable();
1121 raw_spin_lock(&rq->lock);
1126 #ifdef CONFIG_SCHED_HRTICK
1128 * Use HR-timers to deliver accurate preemption points.
1130 * Its all a bit involved since we cannot program an hrt while holding the
1131 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1134 * When we get rescheduled we reprogram the hrtick_timer outside of the
1140 * - enabled by features
1141 * - hrtimer is actually high res
1143 static inline int hrtick_enabled(struct rq *rq)
1145 if (!sched_feat(HRTICK))
1147 if (!cpu_active(cpu_of(rq)))
1149 return hrtimer_is_hres_active(&rq->hrtick_timer);
1152 static void hrtick_clear(struct rq *rq)
1154 if (hrtimer_active(&rq->hrtick_timer))
1155 hrtimer_cancel(&rq->hrtick_timer);
1159 * High-resolution timer tick.
1160 * Runs from hardirq context with interrupts disabled.
1162 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1164 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1166 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1168 raw_spin_lock(&rq->lock);
1169 update_rq_clock(rq);
1170 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1171 raw_spin_unlock(&rq->lock);
1173 return HRTIMER_NORESTART;
1178 * called from hardirq (IPI) context
1180 static void __hrtick_start(void *arg)
1182 struct rq *rq = arg;
1184 raw_spin_lock(&rq->lock);
1185 hrtimer_restart(&rq->hrtick_timer);
1186 rq->hrtick_csd_pending = 0;
1187 raw_spin_unlock(&rq->lock);
1191 * Called to set the hrtick timer state.
1193 * called with rq->lock held and irqs disabled
1195 static void hrtick_start(struct rq *rq, u64 delay)
1197 struct hrtimer *timer = &rq->hrtick_timer;
1198 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1200 hrtimer_set_expires(timer, time);
1202 if (rq == this_rq()) {
1203 hrtimer_restart(timer);
1204 } else if (!rq->hrtick_csd_pending) {
1205 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1206 rq->hrtick_csd_pending = 1;
1211 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1213 int cpu = (int)(long)hcpu;
1216 case CPU_UP_CANCELED:
1217 case CPU_UP_CANCELED_FROZEN:
1218 case CPU_DOWN_PREPARE:
1219 case CPU_DOWN_PREPARE_FROZEN:
1221 case CPU_DEAD_FROZEN:
1222 hrtick_clear(cpu_rq(cpu));
1229 static __init void init_hrtick(void)
1231 hotcpu_notifier(hotplug_hrtick, 0);
1235 * Called to set the hrtick timer state.
1237 * called with rq->lock held and irqs disabled
1239 static void hrtick_start(struct rq *rq, u64 delay)
1241 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1242 HRTIMER_MODE_REL_PINNED, 0);
1245 static inline void init_hrtick(void)
1248 #endif /* CONFIG_SMP */
1250 static void init_rq_hrtick(struct rq *rq)
1253 rq->hrtick_csd_pending = 0;
1255 rq->hrtick_csd.flags = 0;
1256 rq->hrtick_csd.func = __hrtick_start;
1257 rq->hrtick_csd.info = rq;
1260 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1261 rq->hrtick_timer.function = hrtick;
1263 #else /* CONFIG_SCHED_HRTICK */
1264 static inline void hrtick_clear(struct rq *rq)
1268 static inline void init_rq_hrtick(struct rq *rq)
1272 static inline void init_hrtick(void)
1275 #endif /* CONFIG_SCHED_HRTICK */
1278 * resched_task - mark a task 'to be rescheduled now'.
1280 * On UP this means the setting of the need_resched flag, on SMP it
1281 * might also involve a cross-CPU call to trigger the scheduler on
1286 #ifndef tsk_is_polling
1287 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1290 static void resched_task(struct task_struct *p)
1294 assert_raw_spin_locked(&task_rq(p)->lock);
1296 if (test_tsk_need_resched(p))
1299 set_tsk_need_resched(p);
1302 if (cpu == smp_processor_id())
1305 /* NEED_RESCHED must be visible before we test polling */
1307 if (!tsk_is_polling(p))
1308 smp_send_reschedule(cpu);
1311 static void resched_cpu(int cpu)
1313 struct rq *rq = cpu_rq(cpu);
1314 unsigned long flags;
1316 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1318 resched_task(cpu_curr(cpu));
1319 raw_spin_unlock_irqrestore(&rq->lock, flags);
1324 * In the semi idle case, use the nearest busy cpu for migrating timers
1325 * from an idle cpu. This is good for power-savings.
1327 * We don't do similar optimization for completely idle system, as
1328 * selecting an idle cpu will add more delays to the timers than intended
1329 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1331 int get_nohz_timer_target(void)
1333 int cpu = smp_processor_id();
1335 struct sched_domain *sd;
1338 for_each_domain(cpu, sd) {
1339 for_each_cpu(i, sched_domain_span(sd)) {
1351 * When add_timer_on() enqueues a timer into the timer wheel of an
1352 * idle CPU then this timer might expire before the next timer event
1353 * which is scheduled to wake up that CPU. In case of a completely
1354 * idle system the next event might even be infinite time into the
1355 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1356 * leaves the inner idle loop so the newly added timer is taken into
1357 * account when the CPU goes back to idle and evaluates the timer
1358 * wheel for the next timer event.
1360 void wake_up_idle_cpu(int cpu)
1362 struct rq *rq = cpu_rq(cpu);
1364 if (cpu == smp_processor_id())
1368 * This is safe, as this function is called with the timer
1369 * wheel base lock of (cpu) held. When the CPU is on the way
1370 * to idle and has not yet set rq->curr to idle then it will
1371 * be serialized on the timer wheel base lock and take the new
1372 * timer into account automatically.
1374 if (rq->curr != rq->idle)
1378 * We can set TIF_RESCHED on the idle task of the other CPU
1379 * lockless. The worst case is that the other CPU runs the
1380 * idle task through an additional NOOP schedule()
1382 set_tsk_need_resched(rq->idle);
1384 /* NEED_RESCHED must be visible before we test polling */
1386 if (!tsk_is_polling(rq->idle))
1387 smp_send_reschedule(cpu);
1390 #endif /* CONFIG_NO_HZ */
1392 static u64 sched_avg_period(void)
1394 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1397 static void sched_avg_update(struct rq *rq)
1399 s64 period = sched_avg_period();
1401 while ((s64)(rq->clock - rq->age_stamp) > period) {
1403 * Inline assembly required to prevent the compiler
1404 * optimising this loop into a divmod call.
1405 * See __iter_div_u64_rem() for another example of this.
1407 asm("" : "+rm" (rq->age_stamp));
1408 rq->age_stamp += period;
1413 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1415 rq->rt_avg += rt_delta;
1416 sched_avg_update(rq);
1419 #else /* !CONFIG_SMP */
1420 static void resched_task(struct task_struct *p)
1422 assert_raw_spin_locked(&task_rq(p)->lock);
1423 set_tsk_need_resched(p);
1426 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1430 static void sched_avg_update(struct rq *rq)
1433 #endif /* CONFIG_SMP */
1435 #if BITS_PER_LONG == 32
1436 # define WMULT_CONST (~0UL)
1438 # define WMULT_CONST (1UL << 32)
1441 #define WMULT_SHIFT 32
1444 * Shift right and round:
1446 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1449 * delta *= weight / lw
1451 static unsigned long
1452 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1453 struct load_weight *lw)
1458 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1459 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1460 * 2^SCHED_LOAD_RESOLUTION.
1462 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1463 tmp = (u64)delta_exec * scale_load_down(weight);
1465 tmp = (u64)delta_exec;
1467 if (!lw->inv_weight) {
1468 unsigned long w = scale_load_down(lw->weight);
1470 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1472 else if (unlikely(!w))
1473 lw->inv_weight = WMULT_CONST;
1475 lw->inv_weight = WMULT_CONST / w;
1479 * Check whether we'd overflow the 64-bit multiplication:
1481 if (unlikely(tmp > WMULT_CONST))
1482 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1485 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1487 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1490 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1496 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1502 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1509 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1510 * of tasks with abnormal "nice" values across CPUs the contribution that
1511 * each task makes to its run queue's load is weighted according to its
1512 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1513 * scaled version of the new time slice allocation that they receive on time
1517 #define WEIGHT_IDLEPRIO 3
1518 #define WMULT_IDLEPRIO 1431655765
1521 * Nice levels are multiplicative, with a gentle 10% change for every
1522 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1523 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1524 * that remained on nice 0.
1526 * The "10% effect" is relative and cumulative: from _any_ nice level,
1527 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1528 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1529 * If a task goes up by ~10% and another task goes down by ~10% then
1530 * the relative distance between them is ~25%.)
1532 static const int prio_to_weight[40] = {
1533 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1534 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1535 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1536 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1537 /* 0 */ 1024, 820, 655, 526, 423,
1538 /* 5 */ 335, 272, 215, 172, 137,
1539 /* 10 */ 110, 87, 70, 56, 45,
1540 /* 15 */ 36, 29, 23, 18, 15,
1544 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1546 * In cases where the weight does not change often, we can use the
1547 * precalculated inverse to speed up arithmetics by turning divisions
1548 * into multiplications:
1550 static const u32 prio_to_wmult[40] = {
1551 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1552 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1553 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1554 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1555 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1556 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1557 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1558 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1561 /* Time spent by the tasks of the cpu accounting group executing in ... */
1562 enum cpuacct_stat_index {
1563 CPUACCT_STAT_USER, /* ... user mode */
1564 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1566 CPUACCT_STAT_NSTATS,
1569 #ifdef CONFIG_CGROUP_CPUACCT
1570 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1571 static void cpuacct_update_stats(struct task_struct *tsk,
1572 enum cpuacct_stat_index idx, cputime_t val);
1574 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1575 static inline void cpuacct_update_stats(struct task_struct *tsk,
1576 enum cpuacct_stat_index idx, cputime_t val) {}
1579 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1581 update_load_add(&rq->load, load);
1584 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1586 update_load_sub(&rq->load, load);
1589 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1590 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1591 typedef int (*tg_visitor)(struct task_group *, void *);
1594 * Iterate task_group tree rooted at *from, calling @down when first entering a
1595 * node and @up when leaving it for the final time.
1597 * Caller must hold rcu_lock or sufficient equivalent.
1599 static int walk_tg_tree_from(struct task_group *from,
1600 tg_visitor down, tg_visitor up, void *data)
1602 struct task_group *parent, *child;
1608 ret = (*down)(parent, data);
1611 list_for_each_entry_rcu(child, &parent->children, siblings) {
1618 ret = (*up)(parent, data);
1619 if (ret || parent == from)
1623 parent = parent->parent;
1631 * Iterate the full tree, calling @down when first entering a node and @up when
1632 * leaving it for the final time.
1634 * Caller must hold rcu_lock or sufficient equivalent.
1637 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1639 return walk_tg_tree_from(&root_task_group, down, up, data);
1642 static int tg_nop(struct task_group *tg, void *data)
1649 /* Used instead of source_load when we know the type == 0 */
1650 static unsigned long weighted_cpuload(const int cpu)
1652 return cpu_rq(cpu)->load.weight;
1656 * Return a low guess at the load of a migration-source cpu weighted
1657 * according to the scheduling class and "nice" value.
1659 * We want to under-estimate the load of migration sources, to
1660 * balance conservatively.
1662 static unsigned long source_load(int cpu, int type)
1664 struct rq *rq = cpu_rq(cpu);
1665 unsigned long total = weighted_cpuload(cpu);
1667 if (type == 0 || !sched_feat(LB_BIAS))
1670 return min(rq->cpu_load[type-1], total);
1674 * Return a high guess at the load of a migration-target cpu weighted
1675 * according to the scheduling class and "nice" value.
1677 static unsigned long target_load(int cpu, int type)
1679 struct rq *rq = cpu_rq(cpu);
1680 unsigned long total = weighted_cpuload(cpu);
1682 if (type == 0 || !sched_feat(LB_BIAS))
1685 return max(rq->cpu_load[type-1], total);
1688 static unsigned long power_of(int cpu)
1690 return cpu_rq(cpu)->cpu_power;
1693 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1695 static unsigned long cpu_avg_load_per_task(int cpu)
1697 struct rq *rq = cpu_rq(cpu);
1698 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1701 return rq->load.weight / nr_running;
1706 #ifdef CONFIG_PREEMPT
1708 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1711 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1712 * way at the expense of forcing extra atomic operations in all
1713 * invocations. This assures that the double_lock is acquired using the
1714 * same underlying policy as the spinlock_t on this architecture, which
1715 * reduces latency compared to the unfair variant below. However, it
1716 * also adds more overhead and therefore may reduce throughput.
1718 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1719 __releases(this_rq->lock)
1720 __acquires(busiest->lock)
1721 __acquires(this_rq->lock)
1723 raw_spin_unlock(&this_rq->lock);
1724 double_rq_lock(this_rq, busiest);
1731 * Unfair double_lock_balance: Optimizes throughput at the expense of
1732 * latency by eliminating extra atomic operations when the locks are
1733 * already in proper order on entry. This favors lower cpu-ids and will
1734 * grant the double lock to lower cpus over higher ids under contention,
1735 * regardless of entry order into the function.
1737 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1738 __releases(this_rq->lock)
1739 __acquires(busiest->lock)
1740 __acquires(this_rq->lock)
1744 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1745 if (busiest < this_rq) {
1746 raw_spin_unlock(&this_rq->lock);
1747 raw_spin_lock(&busiest->lock);
1748 raw_spin_lock_nested(&this_rq->lock,
1749 SINGLE_DEPTH_NESTING);
1752 raw_spin_lock_nested(&busiest->lock,
1753 SINGLE_DEPTH_NESTING);
1758 #endif /* CONFIG_PREEMPT */
1761 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1763 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1765 if (unlikely(!irqs_disabled())) {
1766 /* printk() doesn't work good under rq->lock */
1767 raw_spin_unlock(&this_rq->lock);
1771 return _double_lock_balance(this_rq, busiest);
1774 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1775 __releases(busiest->lock)
1777 raw_spin_unlock(&busiest->lock);
1778 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1782 * double_rq_lock - safely lock two runqueues
1784 * Note this does not disable interrupts like task_rq_lock,
1785 * you need to do so manually before calling.
1787 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1788 __acquires(rq1->lock)
1789 __acquires(rq2->lock)
1791 BUG_ON(!irqs_disabled());
1793 raw_spin_lock(&rq1->lock);
1794 __acquire(rq2->lock); /* Fake it out ;) */
1797 raw_spin_lock(&rq1->lock);
1798 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1800 raw_spin_lock(&rq2->lock);
1801 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1807 * double_rq_unlock - safely unlock two runqueues
1809 * Note this does not restore interrupts like task_rq_unlock,
1810 * you need to do so manually after calling.
1812 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1813 __releases(rq1->lock)
1814 __releases(rq2->lock)
1816 raw_spin_unlock(&rq1->lock);
1818 raw_spin_unlock(&rq2->lock);
1820 __release(rq2->lock);
1823 #else /* CONFIG_SMP */
1826 * double_rq_lock - safely lock two runqueues
1828 * Note this does not disable interrupts like task_rq_lock,
1829 * you need to do so manually before calling.
1831 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1832 __acquires(rq1->lock)
1833 __acquires(rq2->lock)
1835 BUG_ON(!irqs_disabled());
1837 raw_spin_lock(&rq1->lock);
1838 __acquire(rq2->lock); /* Fake it out ;) */
1842 * double_rq_unlock - safely unlock two runqueues
1844 * Note this does not restore interrupts like task_rq_unlock,
1845 * you need to do so manually after calling.
1847 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1848 __releases(rq1->lock)
1849 __releases(rq2->lock)
1852 raw_spin_unlock(&rq1->lock);
1853 __release(rq2->lock);
1858 static void calc_load_account_idle(struct rq *this_rq);
1859 static void update_sysctl(void);
1860 static int get_update_sysctl_factor(void);
1861 static void update_cpu_load(struct rq *this_rq);
1863 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1865 set_task_rq(p, cpu);
1868 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1869 * successfuly executed on another CPU. We must ensure that updates of
1870 * per-task data have been completed by this moment.
1873 task_thread_info(p)->cpu = cpu;
1877 static const struct sched_class rt_sched_class;
1879 #define sched_class_highest (&stop_sched_class)
1880 #define for_each_class(class) \
1881 for (class = sched_class_highest; class; class = class->next)
1883 #include "sched_stats.h"
1885 static void inc_nr_running(struct rq *rq)
1890 static void dec_nr_running(struct rq *rq)
1895 static void set_load_weight(struct task_struct *p)
1897 int prio = p->static_prio - MAX_RT_PRIO;
1898 struct load_weight *load = &p->se.load;
1901 * SCHED_IDLE tasks get minimal weight:
1903 if (p->policy == SCHED_IDLE) {
1904 load->weight = scale_load(WEIGHT_IDLEPRIO);
1905 load->inv_weight = WMULT_IDLEPRIO;
1909 load->weight = scale_load(prio_to_weight[prio]);
1910 load->inv_weight = prio_to_wmult[prio];
1913 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1915 update_rq_clock(rq);
1916 sched_info_queued(p);
1917 p->sched_class->enqueue_task(rq, p, flags);
1920 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1922 update_rq_clock(rq);
1923 sched_info_dequeued(p);
1924 p->sched_class->dequeue_task(rq, p, flags);
1928 * activate_task - move a task to the runqueue.
1930 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1932 if (task_contributes_to_load(p))
1933 rq->nr_uninterruptible--;
1935 enqueue_task(rq, p, flags);
1939 * deactivate_task - remove a task from the runqueue.
1941 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1943 if (task_contributes_to_load(p))
1944 rq->nr_uninterruptible++;
1946 dequeue_task(rq, p, flags);
1949 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1952 * There are no locks covering percpu hardirq/softirq time.
1953 * They are only modified in account_system_vtime, on corresponding CPU
1954 * with interrupts disabled. So, writes are safe.
1955 * They are read and saved off onto struct rq in update_rq_clock().
1956 * This may result in other CPU reading this CPU's irq time and can
1957 * race with irq/account_system_vtime on this CPU. We would either get old
1958 * or new value with a side effect of accounting a slice of irq time to wrong
1959 * task when irq is in progress while we read rq->clock. That is a worthy
1960 * compromise in place of having locks on each irq in account_system_time.
1962 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1963 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1965 static DEFINE_PER_CPU(u64, irq_start_time);
1966 static int sched_clock_irqtime;
1968 void enable_sched_clock_irqtime(void)
1970 sched_clock_irqtime = 1;
1973 void disable_sched_clock_irqtime(void)
1975 sched_clock_irqtime = 0;
1978 #ifndef CONFIG_64BIT
1979 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1981 static inline void irq_time_write_begin(void)
1983 __this_cpu_inc(irq_time_seq.sequence);
1987 static inline void irq_time_write_end(void)
1990 __this_cpu_inc(irq_time_seq.sequence);
1993 static inline u64 irq_time_read(int cpu)
1999 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
2000 irq_time = per_cpu(cpu_softirq_time, cpu) +
2001 per_cpu(cpu_hardirq_time, cpu);
2002 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
2006 #else /* CONFIG_64BIT */
2007 static inline void irq_time_write_begin(void)
2011 static inline void irq_time_write_end(void)
2015 static inline u64 irq_time_read(int cpu)
2017 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
2019 #endif /* CONFIG_64BIT */
2022 * Called before incrementing preempt_count on {soft,}irq_enter
2023 * and before decrementing preempt_count on {soft,}irq_exit.
2025 void account_system_vtime(struct task_struct *curr)
2027 unsigned long flags;
2031 if (!sched_clock_irqtime)
2034 local_irq_save(flags);
2036 cpu = smp_processor_id();
2037 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2038 __this_cpu_add(irq_start_time, delta);
2040 irq_time_write_begin();
2042 * We do not account for softirq time from ksoftirqd here.
2043 * We want to continue accounting softirq time to ksoftirqd thread
2044 * in that case, so as not to confuse scheduler with a special task
2045 * that do not consume any time, but still wants to run.
2047 if (hardirq_count())
2048 __this_cpu_add(cpu_hardirq_time, delta);
2049 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
2050 __this_cpu_add(cpu_softirq_time, delta);
2052 irq_time_write_end();
2053 local_irq_restore(flags);
2055 EXPORT_SYMBOL_GPL(account_system_vtime);
2057 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2059 #ifdef CONFIG_PARAVIRT
2060 static inline u64 steal_ticks(u64 steal)
2062 if (unlikely(steal > NSEC_PER_SEC))
2063 return div_u64(steal, TICK_NSEC);
2065 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
2069 static void update_rq_clock_task(struct rq *rq, s64 delta)
2072 * In theory, the compile should just see 0 here, and optimize out the call
2073 * to sched_rt_avg_update. But I don't trust it...
2075 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2076 s64 steal = 0, irq_delta = 0;
2078 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2079 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2082 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2083 * this case when a previous update_rq_clock() happened inside a
2084 * {soft,}irq region.
2086 * When this happens, we stop ->clock_task and only update the
2087 * prev_irq_time stamp to account for the part that fit, so that a next
2088 * update will consume the rest. This ensures ->clock_task is
2091 * It does however cause some slight miss-attribution of {soft,}irq
2092 * time, a more accurate solution would be to update the irq_time using
2093 * the current rq->clock timestamp, except that would require using
2096 if (irq_delta > delta)
2099 rq->prev_irq_time += irq_delta;
2102 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2103 if (static_branch((¶virt_steal_rq_enabled))) {
2106 steal = paravirt_steal_clock(cpu_of(rq));
2107 steal -= rq->prev_steal_time_rq;
2109 if (unlikely(steal > delta))
2112 st = steal_ticks(steal);
2113 steal = st * TICK_NSEC;
2115 rq->prev_steal_time_rq += steal;
2121 rq->clock_task += delta;
2123 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2124 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2125 sched_rt_avg_update(rq, irq_delta + steal);
2129 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2130 static int irqtime_account_hi_update(void)
2132 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2133 unsigned long flags;
2137 local_irq_save(flags);
2138 latest_ns = this_cpu_read(cpu_hardirq_time);
2139 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2141 local_irq_restore(flags);
2145 static int irqtime_account_si_update(void)
2147 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2148 unsigned long flags;
2152 local_irq_save(flags);
2153 latest_ns = this_cpu_read(cpu_softirq_time);
2154 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2156 local_irq_restore(flags);
2160 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2162 #define sched_clock_irqtime (0)
2166 #include "sched_idletask.c"
2167 #include "sched_fair.c"
2168 #include "sched_rt.c"
2169 #include "sched_autogroup.c"
2170 #include "sched_stoptask.c"
2171 #ifdef CONFIG_SCHED_DEBUG
2172 # include "sched_debug.c"
2175 void sched_set_stop_task(int cpu, struct task_struct *stop)
2177 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2178 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2182 * Make it appear like a SCHED_FIFO task, its something
2183 * userspace knows about and won't get confused about.
2185 * Also, it will make PI more or less work without too
2186 * much confusion -- but then, stop work should not
2187 * rely on PI working anyway.
2189 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2191 stop->sched_class = &stop_sched_class;
2194 cpu_rq(cpu)->stop = stop;
2198 * Reset it back to a normal scheduling class so that
2199 * it can die in pieces.
2201 old_stop->sched_class = &rt_sched_class;
2206 * __normal_prio - return the priority that is based on the static prio
2208 static inline int __normal_prio(struct task_struct *p)
2210 return p->static_prio;
2214 * Calculate the expected normal priority: i.e. priority
2215 * without taking RT-inheritance into account. Might be
2216 * boosted by interactivity modifiers. Changes upon fork,
2217 * setprio syscalls, and whenever the interactivity
2218 * estimator recalculates.
2220 static inline int normal_prio(struct task_struct *p)
2224 if (task_has_rt_policy(p))
2225 prio = MAX_RT_PRIO-1 - p->rt_priority;
2227 prio = __normal_prio(p);
2232 * Calculate the current priority, i.e. the priority
2233 * taken into account by the scheduler. This value might
2234 * be boosted by RT tasks, or might be boosted by
2235 * interactivity modifiers. Will be RT if the task got
2236 * RT-boosted. If not then it returns p->normal_prio.
2238 static int effective_prio(struct task_struct *p)
2240 p->normal_prio = normal_prio(p);
2242 * If we are RT tasks or we were boosted to RT priority,
2243 * keep the priority unchanged. Otherwise, update priority
2244 * to the normal priority:
2246 if (!rt_prio(p->prio))
2247 return p->normal_prio;
2252 * task_curr - is this task currently executing on a CPU?
2253 * @p: the task in question.
2255 inline int task_curr(const struct task_struct *p)
2257 return cpu_curr(task_cpu(p)) == p;
2260 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2261 const struct sched_class *prev_class,
2264 if (prev_class != p->sched_class) {
2265 if (prev_class->switched_from)
2266 prev_class->switched_from(rq, p);
2267 p->sched_class->switched_to(rq, p);
2268 } else if (oldprio != p->prio)
2269 p->sched_class->prio_changed(rq, p, oldprio);
2272 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2274 const struct sched_class *class;
2276 if (p->sched_class == rq->curr->sched_class) {
2277 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2279 for_each_class(class) {
2280 if (class == rq->curr->sched_class)
2282 if (class == p->sched_class) {
2283 resched_task(rq->curr);
2290 * A queue event has occurred, and we're going to schedule. In
2291 * this case, we can save a useless back to back clock update.
2293 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2294 rq->skip_clock_update = 1;
2299 * Is this task likely cache-hot:
2302 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2306 if (p->sched_class != &fair_sched_class)
2309 if (unlikely(p->policy == SCHED_IDLE))
2313 * Buddy candidates are cache hot:
2315 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2316 (&p->se == cfs_rq_of(&p->se)->next ||
2317 &p->se == cfs_rq_of(&p->se)->last))
2320 if (sysctl_sched_migration_cost == -1)
2322 if (sysctl_sched_migration_cost == 0)
2325 delta = now - p->se.exec_start;
2327 return delta < (s64)sysctl_sched_migration_cost;
2330 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2332 #ifdef CONFIG_SCHED_DEBUG
2334 * We should never call set_task_cpu() on a blocked task,
2335 * ttwu() will sort out the placement.
2337 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2338 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2340 #ifdef CONFIG_LOCKDEP
2342 * The caller should hold either p->pi_lock or rq->lock, when changing
2343 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2345 * sched_move_task() holds both and thus holding either pins the cgroup,
2346 * see set_task_rq().
2348 * Furthermore, all task_rq users should acquire both locks, see
2351 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2352 lockdep_is_held(&task_rq(p)->lock)));
2356 trace_sched_migrate_task(p, new_cpu);
2358 if (task_cpu(p) != new_cpu) {
2359 p->se.nr_migrations++;
2360 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2363 __set_task_cpu(p, new_cpu);
2366 struct migration_arg {
2367 struct task_struct *task;
2371 static int migration_cpu_stop(void *data);
2374 * wait_task_inactive - wait for a thread to unschedule.
2376 * If @match_state is nonzero, it's the @p->state value just checked and
2377 * not expected to change. If it changes, i.e. @p might have woken up,
2378 * then return zero. When we succeed in waiting for @p to be off its CPU,
2379 * we return a positive number (its total switch count). If a second call
2380 * a short while later returns the same number, the caller can be sure that
2381 * @p has remained unscheduled the whole time.
2383 * The caller must ensure that the task *will* unschedule sometime soon,
2384 * else this function might spin for a *long* time. This function can't
2385 * be called with interrupts off, or it may introduce deadlock with
2386 * smp_call_function() if an IPI is sent by the same process we are
2387 * waiting to become inactive.
2389 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2391 unsigned long flags;
2398 * We do the initial early heuristics without holding
2399 * any task-queue locks at all. We'll only try to get
2400 * the runqueue lock when things look like they will
2406 * If the task is actively running on another CPU
2407 * still, just relax and busy-wait without holding
2410 * NOTE! Since we don't hold any locks, it's not
2411 * even sure that "rq" stays as the right runqueue!
2412 * But we don't care, since "task_running()" will
2413 * return false if the runqueue has changed and p
2414 * is actually now running somewhere else!
2416 while (task_running(rq, p)) {
2417 if (match_state && unlikely(p->state != match_state))
2423 * Ok, time to look more closely! We need the rq
2424 * lock now, to be *sure*. If we're wrong, we'll
2425 * just go back and repeat.
2427 rq = task_rq_lock(p, &flags);
2428 trace_sched_wait_task(p);
2429 running = task_running(rq, p);
2432 if (!match_state || p->state == match_state)
2433 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2434 task_rq_unlock(rq, p, &flags);
2437 * If it changed from the expected state, bail out now.
2439 if (unlikely(!ncsw))
2443 * Was it really running after all now that we
2444 * checked with the proper locks actually held?
2446 * Oops. Go back and try again..
2448 if (unlikely(running)) {
2454 * It's not enough that it's not actively running,
2455 * it must be off the runqueue _entirely_, and not
2458 * So if it was still runnable (but just not actively
2459 * running right now), it's preempted, and we should
2460 * yield - it could be a while.
2462 if (unlikely(on_rq)) {
2463 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2465 set_current_state(TASK_UNINTERRUPTIBLE);
2466 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2471 * Ahh, all good. It wasn't running, and it wasn't
2472 * runnable, which means that it will never become
2473 * running in the future either. We're all done!
2482 * kick_process - kick a running thread to enter/exit the kernel
2483 * @p: the to-be-kicked thread
2485 * Cause a process which is running on another CPU to enter
2486 * kernel-mode, without any delay. (to get signals handled.)
2488 * NOTE: this function doesn't have to take the runqueue lock,
2489 * because all it wants to ensure is that the remote task enters
2490 * the kernel. If the IPI races and the task has been migrated
2491 * to another CPU then no harm is done and the purpose has been
2494 void kick_process(struct task_struct *p)
2500 if ((cpu != smp_processor_id()) && task_curr(p))
2501 smp_send_reschedule(cpu);
2504 EXPORT_SYMBOL_GPL(kick_process);
2505 #endif /* CONFIG_SMP */
2509 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2511 static int select_fallback_rq(int cpu, struct task_struct *p)
2514 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2516 /* Look for allowed, online CPU in same node. */
2517 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2518 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2521 /* Any allowed, online CPU? */
2522 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2523 if (dest_cpu < nr_cpu_ids)
2526 /* No more Mr. Nice Guy. */
2527 dest_cpu = cpuset_cpus_allowed_fallback(p);
2529 * Don't tell them about moving exiting tasks or
2530 * kernel threads (both mm NULL), since they never
2533 if (p->mm && printk_ratelimit()) {
2534 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2535 task_pid_nr(p), p->comm, cpu);
2542 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2545 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2547 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2550 * In order not to call set_task_cpu() on a blocking task we need
2551 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2554 * Since this is common to all placement strategies, this lives here.
2556 * [ this allows ->select_task() to simply return task_cpu(p) and
2557 * not worry about this generic constraint ]
2559 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2561 cpu = select_fallback_rq(task_cpu(p), p);
2566 static void update_avg(u64 *avg, u64 sample)
2568 s64 diff = sample - *avg;
2574 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2576 #ifdef CONFIG_SCHEDSTATS
2577 struct rq *rq = this_rq();
2580 int this_cpu = smp_processor_id();
2582 if (cpu == this_cpu) {
2583 schedstat_inc(rq, ttwu_local);
2584 schedstat_inc(p, se.statistics.nr_wakeups_local);
2586 struct sched_domain *sd;
2588 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2590 for_each_domain(this_cpu, sd) {
2591 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2592 schedstat_inc(sd, ttwu_wake_remote);
2599 if (wake_flags & WF_MIGRATED)
2600 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2602 #endif /* CONFIG_SMP */
2604 schedstat_inc(rq, ttwu_count);
2605 schedstat_inc(p, se.statistics.nr_wakeups);
2607 if (wake_flags & WF_SYNC)
2608 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2610 #endif /* CONFIG_SCHEDSTATS */
2613 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2615 activate_task(rq, p, en_flags);
2618 /* if a worker is waking up, notify workqueue */
2619 if (p->flags & PF_WQ_WORKER)
2620 wq_worker_waking_up(p, cpu_of(rq));
2624 * Mark the task runnable and perform wakeup-preemption.
2627 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2629 trace_sched_wakeup(p, true);
2630 check_preempt_curr(rq, p, wake_flags);
2632 p->state = TASK_RUNNING;
2634 if (p->sched_class->task_woken)
2635 p->sched_class->task_woken(rq, p);
2637 if (rq->idle_stamp) {
2638 u64 delta = rq->clock - rq->idle_stamp;
2639 u64 max = 2*sysctl_sched_migration_cost;
2644 update_avg(&rq->avg_idle, delta);
2651 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2654 if (p->sched_contributes_to_load)
2655 rq->nr_uninterruptible--;
2658 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2659 ttwu_do_wakeup(rq, p, wake_flags);
2663 * Called in case the task @p isn't fully descheduled from its runqueue,
2664 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2665 * since all we need to do is flip p->state to TASK_RUNNING, since
2666 * the task is still ->on_rq.
2668 static int ttwu_remote(struct task_struct *p, int wake_flags)
2673 rq = __task_rq_lock(p);
2675 ttwu_do_wakeup(rq, p, wake_flags);
2678 __task_rq_unlock(rq);
2684 static void sched_ttwu_do_pending(struct task_struct *list)
2686 struct rq *rq = this_rq();
2688 raw_spin_lock(&rq->lock);
2691 struct task_struct *p = list;
2692 list = list->wake_entry;
2693 ttwu_do_activate(rq, p, 0);
2696 raw_spin_unlock(&rq->lock);
2699 #ifdef CONFIG_HOTPLUG_CPU
2701 static void sched_ttwu_pending(void)
2703 struct rq *rq = this_rq();
2704 struct task_struct *list = xchg(&rq->wake_list, NULL);
2709 sched_ttwu_do_pending(list);
2712 #endif /* CONFIG_HOTPLUG_CPU */
2714 void scheduler_ipi(void)
2716 struct rq *rq = this_rq();
2717 struct task_struct *list = xchg(&rq->wake_list, NULL);
2723 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2724 * traditionally all their work was done from the interrupt return
2725 * path. Now that we actually do some work, we need to make sure
2728 * Some archs already do call them, luckily irq_enter/exit nest
2731 * Arguably we should visit all archs and update all handlers,
2732 * however a fair share of IPIs are still resched only so this would
2733 * somewhat pessimize the simple resched case.
2736 sched_ttwu_do_pending(list);
2740 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2742 struct rq *rq = cpu_rq(cpu);
2743 struct task_struct *next = rq->wake_list;
2746 struct task_struct *old = next;
2748 p->wake_entry = next;
2749 next = cmpxchg(&rq->wake_list, old, p);
2755 smp_send_reschedule(cpu);
2758 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2759 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2764 rq = __task_rq_lock(p);
2766 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2767 ttwu_do_wakeup(rq, p, wake_flags);
2770 __task_rq_unlock(rq);
2775 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2776 #endif /* CONFIG_SMP */
2778 static void ttwu_queue(struct task_struct *p, int cpu)
2780 struct rq *rq = cpu_rq(cpu);
2782 #if defined(CONFIG_SMP)
2783 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2784 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2785 ttwu_queue_remote(p, cpu);
2790 raw_spin_lock(&rq->lock);
2791 ttwu_do_activate(rq, p, 0);
2792 raw_spin_unlock(&rq->lock);
2796 * try_to_wake_up - wake up a thread
2797 * @p: the thread to be awakened
2798 * @state: the mask of task states that can be woken
2799 * @wake_flags: wake modifier flags (WF_*)
2801 * Put it on the run-queue if it's not already there. The "current"
2802 * thread is always on the run-queue (except when the actual
2803 * re-schedule is in progress), and as such you're allowed to do
2804 * the simpler "current->state = TASK_RUNNING" to mark yourself
2805 * runnable without the overhead of this.
2807 * Returns %true if @p was woken up, %false if it was already running
2808 * or @state didn't match @p's state.
2811 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2813 unsigned long flags;
2814 int cpu, success = 0;
2817 raw_spin_lock_irqsave(&p->pi_lock, flags);
2818 if (!(p->state & state))
2821 success = 1; /* we're going to change ->state */
2824 if (p->on_rq && ttwu_remote(p, wake_flags))
2829 * If the owning (remote) cpu is still in the middle of schedule() with
2830 * this task as prev, wait until its done referencing the task.
2833 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2835 * In case the architecture enables interrupts in
2836 * context_switch(), we cannot busy wait, since that
2837 * would lead to deadlocks when an interrupt hits and
2838 * tries to wake up @prev. So bail and do a complete
2841 if (ttwu_activate_remote(p, wake_flags))
2848 * Pairs with the smp_wmb() in finish_lock_switch().
2852 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2853 p->state = TASK_WAKING;
2855 if (p->sched_class->task_waking)
2856 p->sched_class->task_waking(p);
2858 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2859 if (task_cpu(p) != cpu) {
2860 wake_flags |= WF_MIGRATED;
2861 set_task_cpu(p, cpu);
2863 #endif /* CONFIG_SMP */
2867 ttwu_stat(p, cpu, wake_flags);
2869 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2875 * try_to_wake_up_local - try to wake up a local task with rq lock held
2876 * @p: the thread to be awakened
2878 * Put @p on the run-queue if it's not already there. The caller must
2879 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2882 static void try_to_wake_up_local(struct task_struct *p)
2884 struct rq *rq = task_rq(p);
2886 BUG_ON(rq != this_rq());
2887 BUG_ON(p == current);
2888 lockdep_assert_held(&rq->lock);
2890 if (!raw_spin_trylock(&p->pi_lock)) {
2891 raw_spin_unlock(&rq->lock);
2892 raw_spin_lock(&p->pi_lock);
2893 raw_spin_lock(&rq->lock);
2896 if (!(p->state & TASK_NORMAL))
2900 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2902 ttwu_do_wakeup(rq, p, 0);
2903 ttwu_stat(p, smp_processor_id(), 0);
2905 raw_spin_unlock(&p->pi_lock);
2909 * wake_up_process - Wake up a specific process
2910 * @p: The process to be woken up.
2912 * Attempt to wake up the nominated process and move it to the set of runnable
2913 * processes. Returns 1 if the process was woken up, 0 if it was already
2916 * It may be assumed that this function implies a write memory barrier before
2917 * changing the task state if and only if any tasks are woken up.
2919 int wake_up_process(struct task_struct *p)
2921 return try_to_wake_up(p, TASK_ALL, 0);
2923 EXPORT_SYMBOL(wake_up_process);
2925 int wake_up_state(struct task_struct *p, unsigned int state)
2927 return try_to_wake_up(p, state, 0);
2931 * Perform scheduler related setup for a newly forked process p.
2932 * p is forked by current.
2934 * __sched_fork() is basic setup used by init_idle() too:
2936 static void __sched_fork(struct task_struct *p)
2941 p->se.exec_start = 0;
2942 p->se.sum_exec_runtime = 0;
2943 p->se.prev_sum_exec_runtime = 0;
2944 p->se.nr_migrations = 0;
2946 INIT_LIST_HEAD(&p->se.group_node);
2948 #ifdef CONFIG_SCHEDSTATS
2949 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2952 INIT_LIST_HEAD(&p->rt.run_list);
2954 #ifdef CONFIG_PREEMPT_NOTIFIERS
2955 INIT_HLIST_HEAD(&p->preempt_notifiers);
2960 * fork()/clone()-time setup:
2962 void sched_fork(struct task_struct *p)
2964 unsigned long flags;
2965 int cpu = get_cpu();
2969 * We mark the process as running here. This guarantees that
2970 * nobody will actually run it, and a signal or other external
2971 * event cannot wake it up and insert it on the runqueue either.
2973 p->state = TASK_RUNNING;
2976 * Make sure we do not leak PI boosting priority to the child.
2978 p->prio = current->normal_prio;
2981 * Revert to default priority/policy on fork if requested.
2983 if (unlikely(p->sched_reset_on_fork)) {
2984 if (task_has_rt_policy(p)) {
2985 p->policy = SCHED_NORMAL;
2986 p->static_prio = NICE_TO_PRIO(0);
2988 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2989 p->static_prio = NICE_TO_PRIO(0);
2991 p->prio = p->normal_prio = __normal_prio(p);
2995 * We don't need the reset flag anymore after the fork. It has
2996 * fulfilled its duty:
2998 p->sched_reset_on_fork = 0;
3001 if (!rt_prio(p->prio))
3002 p->sched_class = &fair_sched_class;
3004 if (p->sched_class->task_fork)
3005 p->sched_class->task_fork(p);
3008 * The child is not yet in the pid-hash so no cgroup attach races,
3009 * and the cgroup is pinned to this child due to cgroup_fork()
3010 * is ran before sched_fork().
3012 * Silence PROVE_RCU.
3014 raw_spin_lock_irqsave(&p->pi_lock, flags);
3015 set_task_cpu(p, cpu);
3016 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3018 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3019 if (likely(sched_info_on()))
3020 memset(&p->sched_info, 0, sizeof(p->sched_info));
3022 #if defined(CONFIG_SMP)
3025 #ifdef CONFIG_PREEMPT_COUNT
3026 /* Want to start with kernel preemption disabled. */
3027 task_thread_info(p)->preempt_count = 1;
3030 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3037 * wake_up_new_task - wake up a newly created task for the first time.
3039 * This function will do some initial scheduler statistics housekeeping
3040 * that must be done for every newly created context, then puts the task
3041 * on the runqueue and wakes it.
3043 void wake_up_new_task(struct task_struct *p)
3045 unsigned long flags;
3048 raw_spin_lock_irqsave(&p->pi_lock, flags);
3051 * Fork balancing, do it here and not earlier because:
3052 * - cpus_allowed can change in the fork path
3053 * - any previously selected cpu might disappear through hotplug
3055 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3058 rq = __task_rq_lock(p);
3059 activate_task(rq, p, 0);
3061 trace_sched_wakeup_new(p, true);
3062 check_preempt_curr(rq, p, WF_FORK);
3064 if (p->sched_class->task_woken)
3065 p->sched_class->task_woken(rq, p);
3067 task_rq_unlock(rq, p, &flags);
3070 #ifdef CONFIG_PREEMPT_NOTIFIERS
3073 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3074 * @notifier: notifier struct to register
3076 void preempt_notifier_register(struct preempt_notifier *notifier)
3078 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3080 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3083 * preempt_notifier_unregister - no longer interested in preemption notifications
3084 * @notifier: notifier struct to unregister
3086 * This is safe to call from within a preemption notifier.
3088 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3090 hlist_del(¬ifier->link);
3092 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3094 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3096 struct preempt_notifier *notifier;
3097 struct hlist_node *node;
3099 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3100 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3104 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3105 struct task_struct *next)
3107 struct preempt_notifier *notifier;
3108 struct hlist_node *node;
3110 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3111 notifier->ops->sched_out(notifier, next);
3114 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3116 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3121 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3122 struct task_struct *next)
3126 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3129 * prepare_task_switch - prepare to switch tasks
3130 * @rq: the runqueue preparing to switch
3131 * @prev: the current task that is being switched out
3132 * @next: the task we are going to switch to.
3134 * This is called with the rq lock held and interrupts off. It must
3135 * be paired with a subsequent finish_task_switch after the context
3138 * prepare_task_switch sets up locking and calls architecture specific
3142 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3143 struct task_struct *next)
3145 sched_info_switch(prev, next);
3146 perf_event_task_sched_out(prev, next);
3147 fire_sched_out_preempt_notifiers(prev, next);
3148 prepare_lock_switch(rq, next);
3149 prepare_arch_switch(next);
3150 trace_sched_switch(prev, next);
3154 * finish_task_switch - clean up after a task-switch
3155 * @rq: runqueue associated with task-switch
3156 * @prev: the thread we just switched away from.
3158 * finish_task_switch must be called after the context switch, paired
3159 * with a prepare_task_switch call before the context switch.
3160 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3161 * and do any other architecture-specific cleanup actions.
3163 * Note that we may have delayed dropping an mm in context_switch(). If
3164 * so, we finish that here outside of the runqueue lock. (Doing it
3165 * with the lock held can cause deadlocks; see schedule() for
3168 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3169 __releases(rq->lock)
3171 struct mm_struct *mm = rq->prev_mm;
3177 * A task struct has one reference for the use as "current".
3178 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3179 * schedule one last time. The schedule call will never return, and
3180 * the scheduled task must drop that reference.
3181 * The test for TASK_DEAD must occur while the runqueue locks are
3182 * still held, otherwise prev could be scheduled on another cpu, die
3183 * there before we look at prev->state, and then the reference would
3185 * Manfred Spraul <manfred@colorfullife.com>
3187 prev_state = prev->state;
3188 finish_arch_switch(prev);
3189 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3190 local_irq_disable();
3191 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3192 perf_event_task_sched_in(current);
3193 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3195 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3196 finish_lock_switch(rq, prev);
3198 fire_sched_in_preempt_notifiers(current);
3201 if (unlikely(prev_state == TASK_DEAD)) {
3203 * Remove function-return probe instances associated with this
3204 * task and put them back on the free list.
3206 kprobe_flush_task(prev);
3207 put_task_struct(prev);
3213 /* assumes rq->lock is held */
3214 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3216 if (prev->sched_class->pre_schedule)
3217 prev->sched_class->pre_schedule(rq, prev);
3220 /* rq->lock is NOT held, but preemption is disabled */
3221 static inline void post_schedule(struct rq *rq)
3223 if (rq->post_schedule) {
3224 unsigned long flags;
3226 raw_spin_lock_irqsave(&rq->lock, flags);
3227 if (rq->curr->sched_class->post_schedule)
3228 rq->curr->sched_class->post_schedule(rq);
3229 raw_spin_unlock_irqrestore(&rq->lock, flags);
3231 rq->post_schedule = 0;
3237 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3241 static inline void post_schedule(struct rq *rq)
3248 * schedule_tail - first thing a freshly forked thread must call.
3249 * @prev: the thread we just switched away from.
3251 asmlinkage void schedule_tail(struct task_struct *prev)
3252 __releases(rq->lock)
3254 struct rq *rq = this_rq();
3256 finish_task_switch(rq, prev);
3259 * FIXME: do we need to worry about rq being invalidated by the
3264 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3265 /* In this case, finish_task_switch does not reenable preemption */
3268 if (current->set_child_tid)
3269 put_user(task_pid_vnr(current), current->set_child_tid);
3273 * context_switch - switch to the new MM and the new
3274 * thread's register state.
3277 context_switch(struct rq *rq, struct task_struct *prev,
3278 struct task_struct *next)
3280 struct mm_struct *mm, *oldmm;
3282 prepare_task_switch(rq, prev, next);
3285 oldmm = prev->active_mm;
3287 * For paravirt, this is coupled with an exit in switch_to to
3288 * combine the page table reload and the switch backend into
3291 arch_start_context_switch(prev);
3294 next->active_mm = oldmm;
3295 atomic_inc(&oldmm->mm_count);
3296 enter_lazy_tlb(oldmm, next);
3298 switch_mm(oldmm, mm, next);
3301 prev->active_mm = NULL;
3302 rq->prev_mm = oldmm;
3305 * Since the runqueue lock will be released by the next
3306 * task (which is an invalid locking op but in the case
3307 * of the scheduler it's an obvious special-case), so we
3308 * do an early lockdep release here:
3310 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3311 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3314 /* Here we just switch the register state and the stack. */
3315 switch_to(prev, next, prev);
3319 * this_rq must be evaluated again because prev may have moved
3320 * CPUs since it called schedule(), thus the 'rq' on its stack
3321 * frame will be invalid.
3323 finish_task_switch(this_rq(), prev);
3327 * nr_running, nr_uninterruptible and nr_context_switches:
3329 * externally visible scheduler statistics: current number of runnable
3330 * threads, current number of uninterruptible-sleeping threads, total
3331 * number of context switches performed since bootup.
3333 unsigned long nr_running(void)
3335 unsigned long i, sum = 0;
3337 for_each_online_cpu(i)
3338 sum += cpu_rq(i)->nr_running;
3343 unsigned long nr_uninterruptible(void)
3345 unsigned long i, sum = 0;
3347 for_each_possible_cpu(i)
3348 sum += cpu_rq(i)->nr_uninterruptible;
3351 * Since we read the counters lockless, it might be slightly
3352 * inaccurate. Do not allow it to go below zero though:
3354 if (unlikely((long)sum < 0))
3360 unsigned long long nr_context_switches(void)
3363 unsigned long long sum = 0;
3365 for_each_possible_cpu(i)
3366 sum += cpu_rq(i)->nr_switches;
3371 unsigned long nr_iowait(void)
3373 unsigned long i, sum = 0;
3375 for_each_possible_cpu(i)
3376 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3381 unsigned long nr_iowait_cpu(int cpu)
3383 struct rq *this = cpu_rq(cpu);
3384 return atomic_read(&this->nr_iowait);
3387 unsigned long this_cpu_load(void)
3389 struct rq *this = this_rq();
3390 return this->cpu_load[0];
3394 /* Variables and functions for calc_load */
3395 static atomic_long_t calc_load_tasks;
3396 static unsigned long calc_load_update;
3397 unsigned long avenrun[3];
3398 EXPORT_SYMBOL(avenrun);
3400 static long calc_load_fold_active(struct rq *this_rq)
3402 long nr_active, delta = 0;
3404 nr_active = this_rq->nr_running;
3405 nr_active += (long) this_rq->nr_uninterruptible;
3407 if (nr_active != this_rq->calc_load_active) {
3408 delta = nr_active - this_rq->calc_load_active;
3409 this_rq->calc_load_active = nr_active;
3415 static unsigned long
3416 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3419 load += active * (FIXED_1 - exp);
3420 load += 1UL << (FSHIFT - 1);
3421 return load >> FSHIFT;
3426 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3428 * When making the ILB scale, we should try to pull this in as well.
3430 static atomic_long_t calc_load_tasks_idle;
3432 static void calc_load_account_idle(struct rq *this_rq)
3436 delta = calc_load_fold_active(this_rq);
3438 atomic_long_add(delta, &calc_load_tasks_idle);
3441 static long calc_load_fold_idle(void)
3446 * Its got a race, we don't care...
3448 if (atomic_long_read(&calc_load_tasks_idle))
3449 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3455 * fixed_power_int - compute: x^n, in O(log n) time
3457 * @x: base of the power
3458 * @frac_bits: fractional bits of @x
3459 * @n: power to raise @x to.
3461 * By exploiting the relation between the definition of the natural power
3462 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3463 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3464 * (where: n_i \elem {0, 1}, the binary vector representing n),
3465 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3466 * of course trivially computable in O(log_2 n), the length of our binary
3469 static unsigned long
3470 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3472 unsigned long result = 1UL << frac_bits;
3477 result += 1UL << (frac_bits - 1);
3478 result >>= frac_bits;
3484 x += 1UL << (frac_bits - 1);
3492 * a1 = a0 * e + a * (1 - e)
3494 * a2 = a1 * e + a * (1 - e)
3495 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3496 * = a0 * e^2 + a * (1 - e) * (1 + e)
3498 * a3 = a2 * e + a * (1 - e)
3499 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3500 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3504 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3505 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3506 * = a0 * e^n + a * (1 - e^n)
3508 * [1] application of the geometric series:
3511 * S_n := \Sum x^i = -------------
3514 static unsigned long
3515 calc_load_n(unsigned long load, unsigned long exp,
3516 unsigned long active, unsigned int n)
3519 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3523 * NO_HZ can leave us missing all per-cpu ticks calling
3524 * calc_load_account_active(), but since an idle CPU folds its delta into
3525 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3526 * in the pending idle delta if our idle period crossed a load cycle boundary.
3528 * Once we've updated the global active value, we need to apply the exponential
3529 * weights adjusted to the number of cycles missed.
3531 static void calc_global_nohz(unsigned long ticks)
3533 long delta, active, n;
3535 if (time_before(jiffies, calc_load_update))
3539 * If we crossed a calc_load_update boundary, make sure to fold
3540 * any pending idle changes, the respective CPUs might have
3541 * missed the tick driven calc_load_account_active() update
3544 delta = calc_load_fold_idle();
3546 atomic_long_add(delta, &calc_load_tasks);
3549 * If we were idle for multiple load cycles, apply them.
3551 if (ticks >= LOAD_FREQ) {
3552 n = ticks / LOAD_FREQ;
3554 active = atomic_long_read(&calc_load_tasks);
3555 active = active > 0 ? active * FIXED_1 : 0;
3557 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3558 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3559 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3561 calc_load_update += n * LOAD_FREQ;
3565 * Its possible the remainder of the above division also crosses
3566 * a LOAD_FREQ period, the regular check in calc_global_load()
3567 * which comes after this will take care of that.
3569 * Consider us being 11 ticks before a cycle completion, and us
3570 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3571 * age us 4 cycles, and the test in calc_global_load() will
3572 * pick up the final one.
3576 static void calc_load_account_idle(struct rq *this_rq)
3580 static inline long calc_load_fold_idle(void)
3585 static void calc_global_nohz(unsigned long ticks)
3591 * get_avenrun - get the load average array
3592 * @loads: pointer to dest load array
3593 * @offset: offset to add
3594 * @shift: shift count to shift the result left
3596 * These values are estimates at best, so no need for locking.
3598 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3600 loads[0] = (avenrun[0] + offset) << shift;
3601 loads[1] = (avenrun[1] + offset) << shift;
3602 loads[2] = (avenrun[2] + offset) << shift;
3606 * calc_load - update the avenrun load estimates 10 ticks after the
3607 * CPUs have updated calc_load_tasks.
3609 void calc_global_load(unsigned long ticks)
3613 calc_global_nohz(ticks);
3615 if (time_before(jiffies, calc_load_update + 10))
3618 active = atomic_long_read(&calc_load_tasks);
3619 active = active > 0 ? active * FIXED_1 : 0;
3621 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3622 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3623 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3625 calc_load_update += LOAD_FREQ;
3629 * Called from update_cpu_load() to periodically update this CPU's
3632 static void calc_load_account_active(struct rq *this_rq)
3636 if (time_before(jiffies, this_rq->calc_load_update))
3639 delta = calc_load_fold_active(this_rq);
3640 delta += calc_load_fold_idle();
3642 atomic_long_add(delta, &calc_load_tasks);
3644 this_rq->calc_load_update += LOAD_FREQ;
3648 * The exact cpuload at various idx values, calculated at every tick would be
3649 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3651 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3652 * on nth tick when cpu may be busy, then we have:
3653 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3654 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3656 * decay_load_missed() below does efficient calculation of
3657 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3658 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3660 * The calculation is approximated on a 128 point scale.
3661 * degrade_zero_ticks is the number of ticks after which load at any
3662 * particular idx is approximated to be zero.
3663 * degrade_factor is a precomputed table, a row for each load idx.
3664 * Each column corresponds to degradation factor for a power of two ticks,
3665 * based on 128 point scale.
3667 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3668 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3670 * With this power of 2 load factors, we can degrade the load n times
3671 * by looking at 1 bits in n and doing as many mult/shift instead of
3672 * n mult/shifts needed by the exact degradation.
3674 #define DEGRADE_SHIFT 7
3675 static const unsigned char
3676 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3677 static const unsigned char
3678 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3679 {0, 0, 0, 0, 0, 0, 0, 0},
3680 {64, 32, 8, 0, 0, 0, 0, 0},
3681 {96, 72, 40, 12, 1, 0, 0},
3682 {112, 98, 75, 43, 15, 1, 0},
3683 {120, 112, 98, 76, 45, 16, 2} };
3686 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3687 * would be when CPU is idle and so we just decay the old load without
3688 * adding any new load.
3690 static unsigned long
3691 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3695 if (!missed_updates)
3698 if (missed_updates >= degrade_zero_ticks[idx])
3702 return load >> missed_updates;
3704 while (missed_updates) {
3705 if (missed_updates % 2)
3706 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3708 missed_updates >>= 1;
3715 * Update rq->cpu_load[] statistics. This function is usually called every
3716 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3717 * every tick. We fix it up based on jiffies.
3719 static void update_cpu_load(struct rq *this_rq)
3721 unsigned long this_load = this_rq->load.weight;
3722 unsigned long curr_jiffies = jiffies;
3723 unsigned long pending_updates;
3726 this_rq->nr_load_updates++;
3728 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3729 if (curr_jiffies == this_rq->last_load_update_tick)
3732 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3733 this_rq->last_load_update_tick = curr_jiffies;
3735 /* Update our load: */
3736 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3737 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3738 unsigned long old_load, new_load;
3740 /* scale is effectively 1 << i now, and >> i divides by scale */
3742 old_load = this_rq->cpu_load[i];
3743 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3744 new_load = this_load;
3746 * Round up the averaging division if load is increasing. This
3747 * prevents us from getting stuck on 9 if the load is 10, for
3750 if (new_load > old_load)
3751 new_load += scale - 1;
3753 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3756 sched_avg_update(this_rq);
3759 static void update_cpu_load_active(struct rq *this_rq)
3761 update_cpu_load(this_rq);
3763 calc_load_account_active(this_rq);
3769 * sched_exec - execve() is a valuable balancing opportunity, because at
3770 * this point the task has the smallest effective memory and cache footprint.
3772 void sched_exec(void)
3774 struct task_struct *p = current;
3775 unsigned long flags;
3778 raw_spin_lock_irqsave(&p->pi_lock, flags);
3779 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3780 if (dest_cpu == smp_processor_id())
3783 if (likely(cpu_active(dest_cpu))) {
3784 struct migration_arg arg = { p, dest_cpu };
3786 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3787 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3791 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3796 DEFINE_PER_CPU(struct kernel_stat, kstat);
3798 EXPORT_PER_CPU_SYMBOL(kstat);
3801 * Return any ns on the sched_clock that have not yet been accounted in
3802 * @p in case that task is currently running.
3804 * Called with task_rq_lock() held on @rq.
3806 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3810 if (task_current(rq, p)) {
3811 update_rq_clock(rq);
3812 ns = rq->clock_task - p->se.exec_start;
3820 unsigned long long task_delta_exec(struct task_struct *p)
3822 unsigned long flags;
3826 rq = task_rq_lock(p, &flags);
3827 ns = do_task_delta_exec(p, rq);
3828 task_rq_unlock(rq, p, &flags);
3834 * Return accounted runtime for the task.
3835 * In case the task is currently running, return the runtime plus current's
3836 * pending runtime that have not been accounted yet.
3838 unsigned long long task_sched_runtime(struct task_struct *p)
3840 unsigned long flags;
3844 rq = task_rq_lock(p, &flags);
3845 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3846 task_rq_unlock(rq, p, &flags);
3852 * Return sum_exec_runtime for the thread group.
3853 * In case the task is currently running, return the sum plus current's
3854 * pending runtime that have not been accounted yet.
3856 * Note that the thread group might have other running tasks as well,
3857 * so the return value not includes other pending runtime that other
3858 * running tasks might have.
3860 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3862 struct task_cputime totals;
3863 unsigned long flags;
3867 rq = task_rq_lock(p, &flags);
3868 thread_group_cputime(p, &totals);
3869 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3870 task_rq_unlock(rq, p, &flags);
3876 * Account user cpu time to a process.
3877 * @p: the process that the cpu time gets accounted to
3878 * @cputime: the cpu time spent in user space since the last update
3879 * @cputime_scaled: cputime scaled by cpu frequency
3881 void account_user_time(struct task_struct *p, cputime_t cputime,
3882 cputime_t cputime_scaled)
3884 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3887 /* Add user time to process. */
3888 p->utime = cputime_add(p->utime, cputime);
3889 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3890 account_group_user_time(p, cputime);
3892 /* Add user time to cpustat. */
3893 tmp = cputime_to_cputime64(cputime);
3894 if (TASK_NICE(p) > 0)
3895 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3897 cpustat->user = cputime64_add(cpustat->user, tmp);
3899 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3900 /* Account for user time used */
3901 acct_update_integrals(p);
3905 * Account guest cpu time to a process.
3906 * @p: the process that the cpu time gets accounted to
3907 * @cputime: the cpu time spent in virtual machine since the last update
3908 * @cputime_scaled: cputime scaled by cpu frequency
3910 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3911 cputime_t cputime_scaled)
3914 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3916 tmp = cputime_to_cputime64(cputime);
3918 /* Add guest time to process. */
3919 p->utime = cputime_add(p->utime, cputime);
3920 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3921 account_group_user_time(p, cputime);
3922 p->gtime = cputime_add(p->gtime, cputime);
3924 /* Add guest time to cpustat. */
3925 if (TASK_NICE(p) > 0) {
3926 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3927 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3929 cpustat->user = cputime64_add(cpustat->user, tmp);
3930 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3935 * Account system cpu time to a process and desired cpustat field
3936 * @p: the process that the cpu time gets accounted to
3937 * @cputime: the cpu time spent in kernel space since the last update
3938 * @cputime_scaled: cputime scaled by cpu frequency
3939 * @target_cputime64: pointer to cpustat field that has to be updated
3942 void __account_system_time(struct task_struct *p, cputime_t cputime,
3943 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3945 cputime64_t tmp = cputime_to_cputime64(cputime);
3947 /* Add system time to process. */
3948 p->stime = cputime_add(p->stime, cputime);
3949 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3950 account_group_system_time(p, cputime);
3952 /* Add system time to cpustat. */
3953 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3954 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3956 /* Account for system time used */
3957 acct_update_integrals(p);
3961 * Account system cpu time to a process.
3962 * @p: the process that the cpu time gets accounted to
3963 * @hardirq_offset: the offset to subtract from hardirq_count()
3964 * @cputime: the cpu time spent in kernel space since the last update
3965 * @cputime_scaled: cputime scaled by cpu frequency
3967 void account_system_time(struct task_struct *p, int hardirq_offset,
3968 cputime_t cputime, cputime_t cputime_scaled)
3970 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3971 cputime64_t *target_cputime64;
3973 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3974 account_guest_time(p, cputime, cputime_scaled);
3978 if (hardirq_count() - hardirq_offset)
3979 target_cputime64 = &cpustat->irq;
3980 else if (in_serving_softirq())
3981 target_cputime64 = &cpustat->softirq;
3983 target_cputime64 = &cpustat->system;
3985 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3989 * Account for involuntary wait time.
3990 * @cputime: the cpu time spent in involuntary wait
3992 void account_steal_time(cputime_t cputime)
3994 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3995 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3997 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4001 * Account for idle time.
4002 * @cputime: the cpu time spent in idle wait
4004 void account_idle_time(cputime_t cputime)
4006 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4007 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4008 struct rq *rq = this_rq();
4010 if (atomic_read(&rq->nr_iowait) > 0)
4011 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4013 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4016 static __always_inline bool steal_account_process_tick(void)
4018 #ifdef CONFIG_PARAVIRT
4019 if (static_branch(¶virt_steal_enabled)) {
4022 steal = paravirt_steal_clock(smp_processor_id());
4023 steal -= this_rq()->prev_steal_time;
4025 st = steal_ticks(steal);
4026 this_rq()->prev_steal_time += st * TICK_NSEC;
4028 account_steal_time(st);
4035 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4037 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4039 * Account a tick to a process and cpustat
4040 * @p: the process that the cpu time gets accounted to
4041 * @user_tick: is the tick from userspace
4042 * @rq: the pointer to rq
4044 * Tick demultiplexing follows the order
4045 * - pending hardirq update
4046 * - pending softirq update
4050 * - check for guest_time
4051 * - else account as system_time
4053 * Check for hardirq is done both for system and user time as there is
4054 * no timer going off while we are on hardirq and hence we may never get an
4055 * opportunity to update it solely in system time.
4056 * p->stime and friends are only updated on system time and not on irq
4057 * softirq as those do not count in task exec_runtime any more.
4059 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4062 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4063 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
4064 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4066 if (steal_account_process_tick())
4069 if (irqtime_account_hi_update()) {
4070 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4071 } else if (irqtime_account_si_update()) {
4072 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4073 } else if (this_cpu_ksoftirqd() == p) {
4075 * ksoftirqd time do not get accounted in cpu_softirq_time.
4076 * So, we have to handle it separately here.
4077 * Also, p->stime needs to be updated for ksoftirqd.
4079 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4081 } else if (user_tick) {
4082 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4083 } else if (p == rq->idle) {
4084 account_idle_time(cputime_one_jiffy);
4085 } else if (p->flags & PF_VCPU) { /* System time or guest time */
4086 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
4088 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4093 static void irqtime_account_idle_ticks(int ticks)
4096 struct rq *rq = this_rq();
4098 for (i = 0; i < ticks; i++)
4099 irqtime_account_process_tick(current, 0, rq);
4101 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4102 static void irqtime_account_idle_ticks(int ticks) {}
4103 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4105 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4108 * Account a single tick of cpu time.
4109 * @p: the process that the cpu time gets accounted to
4110 * @user_tick: indicates if the tick is a user or a system tick
4112 void account_process_tick(struct task_struct *p, int user_tick)
4114 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4115 struct rq *rq = this_rq();
4117 if (sched_clock_irqtime) {
4118 irqtime_account_process_tick(p, user_tick, rq);
4122 if (steal_account_process_tick())
4126 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4127 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4128 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4131 account_idle_time(cputime_one_jiffy);
4135 * Account multiple ticks of steal time.
4136 * @p: the process from which the cpu time has been stolen
4137 * @ticks: number of stolen ticks
4139 void account_steal_ticks(unsigned long ticks)
4141 account_steal_time(jiffies_to_cputime(ticks));
4145 * Account multiple ticks of idle time.
4146 * @ticks: number of stolen ticks
4148 void account_idle_ticks(unsigned long ticks)
4151 if (sched_clock_irqtime) {
4152 irqtime_account_idle_ticks(ticks);
4156 account_idle_time(jiffies_to_cputime(ticks));
4162 * Use precise platform statistics if available:
4164 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4165 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4171 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4173 struct task_cputime cputime;
4175 thread_group_cputime(p, &cputime);
4177 *ut = cputime.utime;
4178 *st = cputime.stime;
4182 #ifndef nsecs_to_cputime
4183 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4186 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4188 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4191 * Use CFS's precise accounting:
4193 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4199 do_div(temp, total);
4200 utime = (cputime_t)temp;
4205 * Compare with previous values, to keep monotonicity:
4207 p->prev_utime = max(p->prev_utime, utime);
4208 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4210 *ut = p->prev_utime;
4211 *st = p->prev_stime;
4215 * Must be called with siglock held.
4217 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4219 struct signal_struct *sig = p->signal;
4220 struct task_cputime cputime;
4221 cputime_t rtime, utime, total;
4223 thread_group_cputime(p, &cputime);
4225 total = cputime_add(cputime.utime, cputime.stime);
4226 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4231 temp *= cputime.utime;
4232 do_div(temp, total);
4233 utime = (cputime_t)temp;
4237 sig->prev_utime = max(sig->prev_utime, utime);
4238 sig->prev_stime = max(sig->prev_stime,
4239 cputime_sub(rtime, sig->prev_utime));
4241 *ut = sig->prev_utime;
4242 *st = sig->prev_stime;
4247 * This function gets called by the timer code, with HZ frequency.
4248 * We call it with interrupts disabled.
4250 void scheduler_tick(void)
4252 int cpu = smp_processor_id();
4253 struct rq *rq = cpu_rq(cpu);
4254 struct task_struct *curr = rq->curr;
4258 raw_spin_lock(&rq->lock);
4259 update_rq_clock(rq);
4260 update_cpu_load_active(rq);
4261 curr->sched_class->task_tick(rq, curr, 0);
4262 raw_spin_unlock(&rq->lock);
4264 perf_event_task_tick();
4267 rq->idle_at_tick = idle_cpu(cpu);
4268 trigger_load_balance(rq, cpu);
4272 notrace unsigned long get_parent_ip(unsigned long addr)
4274 if (in_lock_functions(addr)) {
4275 addr = CALLER_ADDR2;
4276 if (in_lock_functions(addr))
4277 addr = CALLER_ADDR3;
4282 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4283 defined(CONFIG_PREEMPT_TRACER))
4285 void __kprobes add_preempt_count(int val)
4287 #ifdef CONFIG_DEBUG_PREEMPT
4291 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4294 preempt_count() += val;
4295 #ifdef CONFIG_DEBUG_PREEMPT
4297 * Spinlock count overflowing soon?
4299 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4302 if (preempt_count() == val)
4303 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4305 EXPORT_SYMBOL(add_preempt_count);
4307 void __kprobes sub_preempt_count(int val)
4309 #ifdef CONFIG_DEBUG_PREEMPT
4313 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4316 * Is the spinlock portion underflowing?
4318 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4319 !(preempt_count() & PREEMPT_MASK)))
4323 if (preempt_count() == val)
4324 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4325 preempt_count() -= val;
4327 EXPORT_SYMBOL(sub_preempt_count);
4332 * Print scheduling while atomic bug:
4334 static noinline void __schedule_bug(struct task_struct *prev)
4336 struct pt_regs *regs = get_irq_regs();
4338 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4339 prev->comm, prev->pid, preempt_count());
4341 debug_show_held_locks(prev);
4343 if (irqs_disabled())
4344 print_irqtrace_events(prev);
4353 * Various schedule()-time debugging checks and statistics:
4355 static inline void schedule_debug(struct task_struct *prev)
4358 * Test if we are atomic. Since do_exit() needs to call into
4359 * schedule() atomically, we ignore that path for now.
4360 * Otherwise, whine if we are scheduling when we should not be.
4362 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4363 __schedule_bug(prev);
4365 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4367 schedstat_inc(this_rq(), sched_count);
4370 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4372 if (prev->on_rq || rq->skip_clock_update < 0)
4373 update_rq_clock(rq);
4374 prev->sched_class->put_prev_task(rq, prev);
4378 * Pick up the highest-prio task:
4380 static inline struct task_struct *
4381 pick_next_task(struct rq *rq)
4383 const struct sched_class *class;
4384 struct task_struct *p;
4387 * Optimization: we know that if all tasks are in
4388 * the fair class we can call that function directly:
4390 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4391 p = fair_sched_class.pick_next_task(rq);
4396 for_each_class(class) {
4397 p = class->pick_next_task(rq);
4402 BUG(); /* the idle class will always have a runnable task */
4406 * schedule() is the main scheduler function.
4408 asmlinkage void __sched schedule(void)
4410 struct task_struct *prev, *next;
4411 unsigned long *switch_count;
4417 cpu = smp_processor_id();
4419 rcu_note_context_switch(cpu);
4422 schedule_debug(prev);
4424 if (sched_feat(HRTICK))
4427 raw_spin_lock_irq(&rq->lock);
4429 switch_count = &prev->nivcsw;
4430 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4431 if (unlikely(signal_pending_state(prev->state, prev))) {
4432 prev->state = TASK_RUNNING;
4434 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4438 * If a worker went to sleep, notify and ask workqueue
4439 * whether it wants to wake up a task to maintain
4442 if (prev->flags & PF_WQ_WORKER) {
4443 struct task_struct *to_wakeup;
4445 to_wakeup = wq_worker_sleeping(prev, cpu);
4447 try_to_wake_up_local(to_wakeup);
4451 * If we are going to sleep and we have plugged IO
4452 * queued, make sure to submit it to avoid deadlocks.
4454 if (blk_needs_flush_plug(prev)) {
4455 raw_spin_unlock(&rq->lock);
4456 blk_schedule_flush_plug(prev);
4457 raw_spin_lock(&rq->lock);
4460 switch_count = &prev->nvcsw;
4463 pre_schedule(rq, prev);
4465 if (unlikely(!rq->nr_running))
4466 idle_balance(cpu, rq);
4468 put_prev_task(rq, prev);
4469 next = pick_next_task(rq);
4470 clear_tsk_need_resched(prev);
4471 rq->skip_clock_update = 0;
4473 if (likely(prev != next)) {
4478 context_switch(rq, prev, next); /* unlocks the rq */
4480 * The context switch have flipped the stack from under us
4481 * and restored the local variables which were saved when
4482 * this task called schedule() in the past. prev == current
4483 * is still correct, but it can be moved to another cpu/rq.
4485 cpu = smp_processor_id();
4488 raw_spin_unlock_irq(&rq->lock);
4492 preempt_enable_no_resched();
4496 EXPORT_SYMBOL(schedule);
4498 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4500 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4502 if (lock->owner != owner)
4506 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4507 * lock->owner still matches owner, if that fails, owner might
4508 * point to free()d memory, if it still matches, the rcu_read_lock()
4509 * ensures the memory stays valid.
4513 return owner->on_cpu;
4517 * Look out! "owner" is an entirely speculative pointer
4518 * access and not reliable.
4520 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4522 if (!sched_feat(OWNER_SPIN))
4526 while (owner_running(lock, owner)) {
4530 arch_mutex_cpu_relax();
4535 * We break out the loop above on need_resched() and when the
4536 * owner changed, which is a sign for heavy contention. Return
4537 * success only when lock->owner is NULL.
4539 return lock->owner == NULL;
4543 #ifdef CONFIG_PREEMPT
4545 * this is the entry point to schedule() from in-kernel preemption
4546 * off of preempt_enable. Kernel preemptions off return from interrupt
4547 * occur there and call schedule directly.
4549 asmlinkage void __sched notrace preempt_schedule(void)
4551 struct thread_info *ti = current_thread_info();
4554 * If there is a non-zero preempt_count or interrupts are disabled,
4555 * we do not want to preempt the current task. Just return..
4557 if (likely(ti->preempt_count || irqs_disabled()))
4561 add_preempt_count_notrace(PREEMPT_ACTIVE);
4563 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4566 * Check again in case we missed a preemption opportunity
4567 * between schedule and now.
4570 } while (need_resched());
4572 EXPORT_SYMBOL(preempt_schedule);
4575 * this is the entry point to schedule() from kernel preemption
4576 * off of irq context.
4577 * Note, that this is called and return with irqs disabled. This will
4578 * protect us against recursive calling from irq.
4580 asmlinkage void __sched preempt_schedule_irq(void)
4582 struct thread_info *ti = current_thread_info();
4584 /* Catch callers which need to be fixed */
4585 BUG_ON(ti->preempt_count || !irqs_disabled());
4588 add_preempt_count(PREEMPT_ACTIVE);
4591 local_irq_disable();
4592 sub_preempt_count(PREEMPT_ACTIVE);
4595 * Check again in case we missed a preemption opportunity
4596 * between schedule and now.
4599 } while (need_resched());
4602 #endif /* CONFIG_PREEMPT */
4604 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4607 return try_to_wake_up(curr->private, mode, wake_flags);
4609 EXPORT_SYMBOL(default_wake_function);
4612 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4613 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4614 * number) then we wake all the non-exclusive tasks and one exclusive task.
4616 * There are circumstances in which we can try to wake a task which has already
4617 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4618 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4620 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4621 int nr_exclusive, int wake_flags, void *key)
4623 wait_queue_t *curr, *next;
4625 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4626 unsigned flags = curr->flags;
4628 if (curr->func(curr, mode, wake_flags, key) &&
4629 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4635 * __wake_up - wake up threads blocked on a waitqueue.
4637 * @mode: which threads
4638 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4639 * @key: is directly passed to the wakeup function
4641 * It may be assumed that this function implies a write memory barrier before
4642 * changing the task state if and only if any tasks are woken up.
4644 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4645 int nr_exclusive, void *key)
4647 unsigned long flags;
4649 spin_lock_irqsave(&q->lock, flags);
4650 __wake_up_common(q, mode, nr_exclusive, 0, key);
4651 spin_unlock_irqrestore(&q->lock, flags);
4653 EXPORT_SYMBOL(__wake_up);
4656 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4658 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4660 __wake_up_common(q, mode, 1, 0, NULL);
4662 EXPORT_SYMBOL_GPL(__wake_up_locked);
4664 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4666 __wake_up_common(q, mode, 1, 0, key);
4668 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4671 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4673 * @mode: which threads
4674 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4675 * @key: opaque value to be passed to wakeup targets
4677 * The sync wakeup differs that the waker knows that it will schedule
4678 * away soon, so while the target thread will be woken up, it will not
4679 * be migrated to another CPU - ie. the two threads are 'synchronized'
4680 * with each other. This can prevent needless bouncing between CPUs.
4682 * On UP it can prevent extra preemption.
4684 * It may be assumed that this function implies a write memory barrier before
4685 * changing the task state if and only if any tasks are woken up.
4687 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4688 int nr_exclusive, void *key)
4690 unsigned long flags;
4691 int wake_flags = WF_SYNC;
4696 if (unlikely(!nr_exclusive))
4699 spin_lock_irqsave(&q->lock, flags);
4700 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4701 spin_unlock_irqrestore(&q->lock, flags);
4703 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4706 * __wake_up_sync - see __wake_up_sync_key()
4708 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4710 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4712 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4715 * complete: - signals a single thread waiting on this completion
4716 * @x: holds the state of this particular completion
4718 * This will wake up a single thread waiting on this completion. Threads will be
4719 * awakened in the same order in which they were queued.
4721 * See also complete_all(), wait_for_completion() and related routines.
4723 * It may be assumed that this function implies a write memory barrier before
4724 * changing the task state if and only if any tasks are woken up.
4726 void complete(struct completion *x)
4728 unsigned long flags;
4730 spin_lock_irqsave(&x->wait.lock, flags);
4732 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4733 spin_unlock_irqrestore(&x->wait.lock, flags);
4735 EXPORT_SYMBOL(complete);
4738 * complete_all: - signals all threads waiting on this completion
4739 * @x: holds the state of this particular completion
4741 * This will wake up all threads waiting on this particular completion event.
4743 * It may be assumed that this function implies a write memory barrier before
4744 * changing the task state if and only if any tasks are woken up.
4746 void complete_all(struct completion *x)
4748 unsigned long flags;
4750 spin_lock_irqsave(&x->wait.lock, flags);
4751 x->done += UINT_MAX/2;
4752 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4753 spin_unlock_irqrestore(&x->wait.lock, flags);
4755 EXPORT_SYMBOL(complete_all);
4757 static inline long __sched
4758 do_wait_for_common(struct completion *x, long timeout, int state)
4761 DECLARE_WAITQUEUE(wait, current);
4763 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4765 if (signal_pending_state(state, current)) {
4766 timeout = -ERESTARTSYS;
4769 __set_current_state(state);
4770 spin_unlock_irq(&x->wait.lock);
4771 timeout = schedule_timeout(timeout);
4772 spin_lock_irq(&x->wait.lock);
4773 } while (!x->done && timeout);
4774 __remove_wait_queue(&x->wait, &wait);
4779 return timeout ?: 1;
4783 wait_for_common(struct completion *x, long timeout, int state)
4787 spin_lock_irq(&x->wait.lock);
4788 timeout = do_wait_for_common(x, timeout, state);
4789 spin_unlock_irq(&x->wait.lock);
4794 * wait_for_completion: - waits for completion of a task
4795 * @x: holds the state of this particular completion
4797 * This waits to be signaled for completion of a specific task. It is NOT
4798 * interruptible and there is no timeout.
4800 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4801 * and interrupt capability. Also see complete().
4803 void __sched wait_for_completion(struct completion *x)
4805 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4807 EXPORT_SYMBOL(wait_for_completion);
4810 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4811 * @x: holds the state of this particular completion
4812 * @timeout: timeout value in jiffies
4814 * This waits for either a completion of a specific task to be signaled or for a
4815 * specified timeout to expire. The timeout is in jiffies. It is not
4818 unsigned long __sched
4819 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4821 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4823 EXPORT_SYMBOL(wait_for_completion_timeout);
4826 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4827 * @x: holds the state of this particular completion
4829 * This waits for completion of a specific task to be signaled. It is
4832 int __sched wait_for_completion_interruptible(struct completion *x)
4834 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4835 if (t == -ERESTARTSYS)
4839 EXPORT_SYMBOL(wait_for_completion_interruptible);
4842 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4843 * @x: holds the state of this particular completion
4844 * @timeout: timeout value in jiffies
4846 * This waits for either a completion of a specific task to be signaled or for a
4847 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4850 wait_for_completion_interruptible_timeout(struct completion *x,
4851 unsigned long timeout)
4853 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4855 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4858 * wait_for_completion_killable: - waits for completion of a task (killable)
4859 * @x: holds the state of this particular completion
4861 * This waits to be signaled for completion of a specific task. It can be
4862 * interrupted by a kill signal.
4864 int __sched wait_for_completion_killable(struct completion *x)
4866 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4867 if (t == -ERESTARTSYS)
4871 EXPORT_SYMBOL(wait_for_completion_killable);
4874 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4875 * @x: holds the state of this particular completion
4876 * @timeout: timeout value in jiffies
4878 * This waits for either a completion of a specific task to be
4879 * signaled or for a specified timeout to expire. It can be
4880 * interrupted by a kill signal. The timeout is in jiffies.
4883 wait_for_completion_killable_timeout(struct completion *x,
4884 unsigned long timeout)
4886 return wait_for_common(x, timeout, TASK_KILLABLE);
4888 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4891 * try_wait_for_completion - try to decrement a completion without blocking
4892 * @x: completion structure
4894 * Returns: 0 if a decrement cannot be done without blocking
4895 * 1 if a decrement succeeded.
4897 * If a completion is being used as a counting completion,
4898 * attempt to decrement the counter without blocking. This
4899 * enables us to avoid waiting if the resource the completion
4900 * is protecting is not available.
4902 bool try_wait_for_completion(struct completion *x)
4904 unsigned long flags;
4907 spin_lock_irqsave(&x->wait.lock, flags);
4912 spin_unlock_irqrestore(&x->wait.lock, flags);
4915 EXPORT_SYMBOL(try_wait_for_completion);
4918 * completion_done - Test to see if a completion has any waiters
4919 * @x: completion structure
4921 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4922 * 1 if there are no waiters.
4925 bool completion_done(struct completion *x)
4927 unsigned long flags;
4930 spin_lock_irqsave(&x->wait.lock, flags);
4933 spin_unlock_irqrestore(&x->wait.lock, flags);
4936 EXPORT_SYMBOL(completion_done);
4939 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4941 unsigned long flags;
4944 init_waitqueue_entry(&wait, current);
4946 __set_current_state(state);
4948 spin_lock_irqsave(&q->lock, flags);
4949 __add_wait_queue(q, &wait);
4950 spin_unlock(&q->lock);
4951 timeout = schedule_timeout(timeout);
4952 spin_lock_irq(&q->lock);
4953 __remove_wait_queue(q, &wait);
4954 spin_unlock_irqrestore(&q->lock, flags);
4959 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4961 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4963 EXPORT_SYMBOL(interruptible_sleep_on);
4966 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4968 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4970 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4972 void __sched sleep_on(wait_queue_head_t *q)
4974 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4976 EXPORT_SYMBOL(sleep_on);
4978 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4980 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4982 EXPORT_SYMBOL(sleep_on_timeout);
4984 #ifdef CONFIG_RT_MUTEXES
4987 * rt_mutex_setprio - set the current priority of a task
4989 * @prio: prio value (kernel-internal form)
4991 * This function changes the 'effective' priority of a task. It does
4992 * not touch ->normal_prio like __setscheduler().
4994 * Used by the rt_mutex code to implement priority inheritance logic.
4996 void rt_mutex_setprio(struct task_struct *p, int prio)
4998 int oldprio, on_rq, running;
5000 const struct sched_class *prev_class;
5002 BUG_ON(prio < 0 || prio > MAX_PRIO);
5004 rq = __task_rq_lock(p);
5006 trace_sched_pi_setprio(p, prio);
5008 prev_class = p->sched_class;
5010 running = task_current(rq, p);
5012 dequeue_task(rq, p, 0);
5014 p->sched_class->put_prev_task(rq, p);
5017 p->sched_class = &rt_sched_class;
5019 p->sched_class = &fair_sched_class;
5024 p->sched_class->set_curr_task(rq);
5026 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
5028 check_class_changed(rq, p, prev_class, oldprio);
5029 __task_rq_unlock(rq);
5034 void set_user_nice(struct task_struct *p, long nice)
5036 int old_prio, delta, on_rq;
5037 unsigned long flags;
5040 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5043 * We have to be careful, if called from sys_setpriority(),
5044 * the task might be in the middle of scheduling on another CPU.
5046 rq = task_rq_lock(p, &flags);
5048 * The RT priorities are set via sched_setscheduler(), but we still
5049 * allow the 'normal' nice value to be set - but as expected
5050 * it wont have any effect on scheduling until the task is
5051 * SCHED_FIFO/SCHED_RR:
5053 if (task_has_rt_policy(p)) {
5054 p->static_prio = NICE_TO_PRIO(nice);
5059 dequeue_task(rq, p, 0);
5061 p->static_prio = NICE_TO_PRIO(nice);
5064 p->prio = effective_prio(p);
5065 delta = p->prio - old_prio;
5068 enqueue_task(rq, p, 0);
5070 * If the task increased its priority or is running and
5071 * lowered its priority, then reschedule its CPU:
5073 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5074 resched_task(rq->curr);
5077 task_rq_unlock(rq, p, &flags);
5079 EXPORT_SYMBOL(set_user_nice);
5082 * can_nice - check if a task can reduce its nice value
5086 int can_nice(const struct task_struct *p, const int nice)
5088 /* convert nice value [19,-20] to rlimit style value [1,40] */
5089 int nice_rlim = 20 - nice;
5091 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5092 capable(CAP_SYS_NICE));
5095 #ifdef __ARCH_WANT_SYS_NICE
5098 * sys_nice - change the priority of the current process.
5099 * @increment: priority increment
5101 * sys_setpriority is a more generic, but much slower function that
5102 * does similar things.
5104 SYSCALL_DEFINE1(nice, int, increment)
5109 * Setpriority might change our priority at the same moment.
5110 * We don't have to worry. Conceptually one call occurs first
5111 * and we have a single winner.
5113 if (increment < -40)
5118 nice = TASK_NICE(current) + increment;
5124 if (increment < 0 && !can_nice(current, nice))
5127 retval = security_task_setnice(current, nice);
5131 set_user_nice(current, nice);
5138 * task_prio - return the priority value of a given task.
5139 * @p: the task in question.
5141 * This is the priority value as seen by users in /proc.
5142 * RT tasks are offset by -200. Normal tasks are centered
5143 * around 0, value goes from -16 to +15.
5145 int task_prio(const struct task_struct *p)
5147 return p->prio - MAX_RT_PRIO;
5151 * task_nice - return the nice value of a given task.
5152 * @p: the task in question.
5154 int task_nice(const struct task_struct *p)
5156 return TASK_NICE(p);
5158 EXPORT_SYMBOL(task_nice);
5161 * idle_cpu - is a given cpu idle currently?
5162 * @cpu: the processor in question.
5164 int idle_cpu(int cpu)
5166 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5170 * idle_task - return the idle task for a given cpu.
5171 * @cpu: the processor in question.
5173 struct task_struct *idle_task(int cpu)
5175 return cpu_rq(cpu)->idle;
5179 * find_process_by_pid - find a process with a matching PID value.
5180 * @pid: the pid in question.
5182 static struct task_struct *find_process_by_pid(pid_t pid)
5184 return pid ? find_task_by_vpid(pid) : current;
5187 /* Actually do priority change: must hold rq lock. */
5189 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5192 p->rt_priority = prio;
5193 p->normal_prio = normal_prio(p);
5194 /* we are holding p->pi_lock already */
5195 p->prio = rt_mutex_getprio(p);
5196 if (rt_prio(p->prio))
5197 p->sched_class = &rt_sched_class;
5199 p->sched_class = &fair_sched_class;
5204 * check the target process has a UID that matches the current process's
5206 static bool check_same_owner(struct task_struct *p)
5208 const struct cred *cred = current_cred(), *pcred;
5212 pcred = __task_cred(p);
5213 if (cred->user->user_ns == pcred->user->user_ns)
5214 match = (cred->euid == pcred->euid ||
5215 cred->euid == pcred->uid);
5222 static int __sched_setscheduler(struct task_struct *p, int policy,
5223 const struct sched_param *param, bool user)
5225 int retval, oldprio, oldpolicy = -1, on_rq, running;
5226 unsigned long flags;
5227 const struct sched_class *prev_class;
5231 /* may grab non-irq protected spin_locks */
5232 BUG_ON(in_interrupt());
5234 /* double check policy once rq lock held */
5236 reset_on_fork = p->sched_reset_on_fork;
5237 policy = oldpolicy = p->policy;
5239 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5240 policy &= ~SCHED_RESET_ON_FORK;
5242 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5243 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5244 policy != SCHED_IDLE)
5249 * Valid priorities for SCHED_FIFO and SCHED_RR are
5250 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5251 * SCHED_BATCH and SCHED_IDLE is 0.
5253 if (param->sched_priority < 0 ||
5254 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5255 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5257 if (rt_policy(policy) != (param->sched_priority != 0))
5261 * Allow unprivileged RT tasks to decrease priority:
5263 if (user && !capable(CAP_SYS_NICE)) {
5264 if (rt_policy(policy)) {
5265 unsigned long rlim_rtprio =
5266 task_rlimit(p, RLIMIT_RTPRIO);
5268 /* can't set/change the rt policy */
5269 if (policy != p->policy && !rlim_rtprio)
5272 /* can't increase priority */
5273 if (param->sched_priority > p->rt_priority &&
5274 param->sched_priority > rlim_rtprio)
5279 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5280 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5282 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5283 if (!can_nice(p, TASK_NICE(p)))
5287 /* can't change other user's priorities */
5288 if (!check_same_owner(p))
5291 /* Normal users shall not reset the sched_reset_on_fork flag */
5292 if (p->sched_reset_on_fork && !reset_on_fork)
5297 retval = security_task_setscheduler(p);
5303 * make sure no PI-waiters arrive (or leave) while we are
5304 * changing the priority of the task:
5306 * To be able to change p->policy safely, the appropriate
5307 * runqueue lock must be held.
5309 rq = task_rq_lock(p, &flags);
5312 * Changing the policy of the stop threads its a very bad idea
5314 if (p == rq->stop) {
5315 task_rq_unlock(rq, p, &flags);
5320 * If not changing anything there's no need to proceed further:
5322 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5323 param->sched_priority == p->rt_priority))) {
5325 __task_rq_unlock(rq);
5326 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5330 #ifdef CONFIG_RT_GROUP_SCHED
5333 * Do not allow realtime tasks into groups that have no runtime
5336 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5337 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5338 !task_group_is_autogroup(task_group(p))) {
5339 task_rq_unlock(rq, p, &flags);
5345 /* recheck policy now with rq lock held */
5346 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5347 policy = oldpolicy = -1;
5348 task_rq_unlock(rq, p, &flags);
5352 running = task_current(rq, p);
5354 deactivate_task(rq, p, 0);
5356 p->sched_class->put_prev_task(rq, p);
5358 p->sched_reset_on_fork = reset_on_fork;
5361 prev_class = p->sched_class;
5362 __setscheduler(rq, p, policy, param->sched_priority);
5365 p->sched_class->set_curr_task(rq);
5367 activate_task(rq, p, 0);
5369 check_class_changed(rq, p, prev_class, oldprio);
5370 task_rq_unlock(rq, p, &flags);
5372 rt_mutex_adjust_pi(p);
5378 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5379 * @p: the task in question.
5380 * @policy: new policy.
5381 * @param: structure containing the new RT priority.
5383 * NOTE that the task may be already dead.
5385 int sched_setscheduler(struct task_struct *p, int policy,
5386 const struct sched_param *param)
5388 return __sched_setscheduler(p, policy, param, true);
5390 EXPORT_SYMBOL_GPL(sched_setscheduler);
5393 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5394 * @p: the task in question.
5395 * @policy: new policy.
5396 * @param: structure containing the new RT priority.
5398 * Just like sched_setscheduler, only don't bother checking if the
5399 * current context has permission. For example, this is needed in
5400 * stop_machine(): we create temporary high priority worker threads,
5401 * but our caller might not have that capability.
5403 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5404 const struct sched_param *param)
5406 return __sched_setscheduler(p, policy, param, false);
5410 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5412 struct sched_param lparam;
5413 struct task_struct *p;
5416 if (!param || pid < 0)
5418 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5423 p = find_process_by_pid(pid);
5425 retval = sched_setscheduler(p, policy, &lparam);
5432 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5433 * @pid: the pid in question.
5434 * @policy: new policy.
5435 * @param: structure containing the new RT priority.
5437 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5438 struct sched_param __user *, param)
5440 /* negative values for policy are not valid */
5444 return do_sched_setscheduler(pid, policy, param);
5448 * sys_sched_setparam - set/change the RT priority of a thread
5449 * @pid: the pid in question.
5450 * @param: structure containing the new RT priority.
5452 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5454 return do_sched_setscheduler(pid, -1, param);
5458 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5459 * @pid: the pid in question.
5461 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5463 struct task_struct *p;
5471 p = find_process_by_pid(pid);
5473 retval = security_task_getscheduler(p);
5476 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5483 * sys_sched_getparam - get the RT priority of a thread
5484 * @pid: the pid in question.
5485 * @param: structure containing the RT priority.
5487 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5489 struct sched_param lp;
5490 struct task_struct *p;
5493 if (!param || pid < 0)
5497 p = find_process_by_pid(pid);
5502 retval = security_task_getscheduler(p);
5506 lp.sched_priority = p->rt_priority;
5510 * This one might sleep, we cannot do it with a spinlock held ...
5512 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5521 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5523 cpumask_var_t cpus_allowed, new_mask;
5524 struct task_struct *p;
5530 p = find_process_by_pid(pid);
5537 /* Prevent p going away */
5541 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5545 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5547 goto out_free_cpus_allowed;
5550 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5553 retval = security_task_setscheduler(p);
5557 cpuset_cpus_allowed(p, cpus_allowed);
5558 cpumask_and(new_mask, in_mask, cpus_allowed);
5560 retval = set_cpus_allowed_ptr(p, new_mask);
5563 cpuset_cpus_allowed(p, cpus_allowed);
5564 if (!cpumask_subset(new_mask, cpus_allowed)) {
5566 * We must have raced with a concurrent cpuset
5567 * update. Just reset the cpus_allowed to the
5568 * cpuset's cpus_allowed
5570 cpumask_copy(new_mask, cpus_allowed);
5575 free_cpumask_var(new_mask);
5576 out_free_cpus_allowed:
5577 free_cpumask_var(cpus_allowed);
5584 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5585 struct cpumask *new_mask)
5587 if (len < cpumask_size())
5588 cpumask_clear(new_mask);
5589 else if (len > cpumask_size())
5590 len = cpumask_size();
5592 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5596 * sys_sched_setaffinity - set the cpu affinity of a process
5597 * @pid: pid of the process
5598 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5599 * @user_mask_ptr: user-space pointer to the new cpu mask
5601 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5602 unsigned long __user *, user_mask_ptr)
5604 cpumask_var_t new_mask;
5607 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5610 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5612 retval = sched_setaffinity(pid, new_mask);
5613 free_cpumask_var(new_mask);
5617 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5619 struct task_struct *p;
5620 unsigned long flags;
5627 p = find_process_by_pid(pid);
5631 retval = security_task_getscheduler(p);
5635 raw_spin_lock_irqsave(&p->pi_lock, flags);
5636 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5637 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5647 * sys_sched_getaffinity - get the cpu affinity of a process
5648 * @pid: pid of the process
5649 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5650 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5652 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5653 unsigned long __user *, user_mask_ptr)
5658 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5660 if (len & (sizeof(unsigned long)-1))
5663 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5666 ret = sched_getaffinity(pid, mask);
5668 size_t retlen = min_t(size_t, len, cpumask_size());
5670 if (copy_to_user(user_mask_ptr, mask, retlen))
5675 free_cpumask_var(mask);
5681 * sys_sched_yield - yield the current processor to other threads.
5683 * This function yields the current CPU to other tasks. If there are no
5684 * other threads running on this CPU then this function will return.
5686 SYSCALL_DEFINE0(sched_yield)
5688 struct rq *rq = this_rq_lock();
5690 schedstat_inc(rq, yld_count);
5691 current->sched_class->yield_task(rq);
5694 * Since we are going to call schedule() anyway, there's
5695 * no need to preempt or enable interrupts:
5697 __release(rq->lock);
5698 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5699 do_raw_spin_unlock(&rq->lock);
5700 preempt_enable_no_resched();
5707 static inline int should_resched(void)
5709 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5712 static void __cond_resched(void)
5714 add_preempt_count(PREEMPT_ACTIVE);
5716 sub_preempt_count(PREEMPT_ACTIVE);
5719 int __sched _cond_resched(void)
5721 if (should_resched()) {
5727 EXPORT_SYMBOL(_cond_resched);
5730 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5731 * call schedule, and on return reacquire the lock.
5733 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5734 * operations here to prevent schedule() from being called twice (once via
5735 * spin_unlock(), once by hand).
5737 int __cond_resched_lock(spinlock_t *lock)
5739 int resched = should_resched();
5742 lockdep_assert_held(lock);
5744 if (spin_needbreak(lock) || resched) {
5755 EXPORT_SYMBOL(__cond_resched_lock);
5757 int __sched __cond_resched_softirq(void)
5759 BUG_ON(!in_softirq());
5761 if (should_resched()) {
5769 EXPORT_SYMBOL(__cond_resched_softirq);
5772 * yield - yield the current processor to other threads.
5774 * This is a shortcut for kernel-space yielding - it marks the
5775 * thread runnable and calls sys_sched_yield().
5777 void __sched yield(void)
5779 set_current_state(TASK_RUNNING);
5782 EXPORT_SYMBOL(yield);
5785 * yield_to - yield the current processor to another thread in
5786 * your thread group, or accelerate that thread toward the
5787 * processor it's on.
5789 * @preempt: whether task preemption is allowed or not
5791 * It's the caller's job to ensure that the target task struct
5792 * can't go away on us before we can do any checks.
5794 * Returns true if we indeed boosted the target task.
5796 bool __sched yield_to(struct task_struct *p, bool preempt)
5798 struct task_struct *curr = current;
5799 struct rq *rq, *p_rq;
5800 unsigned long flags;
5803 local_irq_save(flags);
5808 double_rq_lock(rq, p_rq);
5809 while (task_rq(p) != p_rq) {
5810 double_rq_unlock(rq, p_rq);
5814 if (!curr->sched_class->yield_to_task)
5817 if (curr->sched_class != p->sched_class)
5820 if (task_running(p_rq, p) || p->state)
5823 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5825 schedstat_inc(rq, yld_count);
5827 * Make p's CPU reschedule; pick_next_entity takes care of
5830 if (preempt && rq != p_rq)
5831 resched_task(p_rq->curr);
5835 double_rq_unlock(rq, p_rq);
5836 local_irq_restore(flags);
5843 EXPORT_SYMBOL_GPL(yield_to);
5846 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5847 * that process accounting knows that this is a task in IO wait state.
5849 void __sched io_schedule(void)
5851 struct rq *rq = raw_rq();
5853 delayacct_blkio_start();
5854 atomic_inc(&rq->nr_iowait);
5855 blk_flush_plug(current);
5856 current->in_iowait = 1;
5858 current->in_iowait = 0;
5859 atomic_dec(&rq->nr_iowait);
5860 delayacct_blkio_end();
5862 EXPORT_SYMBOL(io_schedule);
5864 long __sched io_schedule_timeout(long timeout)
5866 struct rq *rq = raw_rq();
5869 delayacct_blkio_start();
5870 atomic_inc(&rq->nr_iowait);
5871 blk_flush_plug(current);
5872 current->in_iowait = 1;
5873 ret = schedule_timeout(timeout);
5874 current->in_iowait = 0;
5875 atomic_dec(&rq->nr_iowait);
5876 delayacct_blkio_end();
5881 * sys_sched_get_priority_max - return maximum RT priority.
5882 * @policy: scheduling class.
5884 * this syscall returns the maximum rt_priority that can be used
5885 * by a given scheduling class.
5887 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5894 ret = MAX_USER_RT_PRIO-1;
5906 * sys_sched_get_priority_min - return minimum RT priority.
5907 * @policy: scheduling class.
5909 * this syscall returns the minimum rt_priority that can be used
5910 * by a given scheduling class.
5912 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5930 * sys_sched_rr_get_interval - return the default timeslice of a process.
5931 * @pid: pid of the process.
5932 * @interval: userspace pointer to the timeslice value.
5934 * this syscall writes the default timeslice value of a given process
5935 * into the user-space timespec buffer. A value of '0' means infinity.
5937 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5938 struct timespec __user *, interval)
5940 struct task_struct *p;
5941 unsigned int time_slice;
5942 unsigned long flags;
5952 p = find_process_by_pid(pid);
5956 retval = security_task_getscheduler(p);
5960 rq = task_rq_lock(p, &flags);
5961 time_slice = p->sched_class->get_rr_interval(rq, p);
5962 task_rq_unlock(rq, p, &flags);
5965 jiffies_to_timespec(time_slice, &t);
5966 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5974 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5976 void sched_show_task(struct task_struct *p)
5978 unsigned long free = 0;
5981 state = p->state ? __ffs(p->state) + 1 : 0;
5982 printk(KERN_INFO "%-15.15s %c", p->comm,
5983 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5984 #if BITS_PER_LONG == 32
5985 if (state == TASK_RUNNING)
5986 printk(KERN_CONT " running ");
5988 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5990 if (state == TASK_RUNNING)
5991 printk(KERN_CONT " running task ");
5993 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5995 #ifdef CONFIG_DEBUG_STACK_USAGE
5996 free = stack_not_used(p);
5998 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5999 task_pid_nr(p), task_pid_nr(p->real_parent),
6000 (unsigned long)task_thread_info(p)->flags);
6002 show_stack(p, NULL);
6005 void show_state_filter(unsigned long state_filter)
6007 struct task_struct *g, *p;
6009 #if BITS_PER_LONG == 32
6011 " task PC stack pid father\n");
6014 " task PC stack pid father\n");
6016 read_lock(&tasklist_lock);
6017 do_each_thread(g, p) {
6019 * reset the NMI-timeout, listing all files on a slow
6020 * console might take a lot of time:
6022 touch_nmi_watchdog();
6023 if (!state_filter || (p->state & state_filter))
6025 } while_each_thread(g, p);
6027 touch_all_softlockup_watchdogs();
6029 #ifdef CONFIG_SCHED_DEBUG
6030 sysrq_sched_debug_show();
6032 read_unlock(&tasklist_lock);
6034 * Only show locks if all tasks are dumped:
6037 debug_show_all_locks();
6040 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6042 idle->sched_class = &idle_sched_class;
6046 * init_idle - set up an idle thread for a given CPU
6047 * @idle: task in question
6048 * @cpu: cpu the idle task belongs to
6050 * NOTE: this function does not set the idle thread's NEED_RESCHED
6051 * flag, to make booting more robust.
6053 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6055 struct rq *rq = cpu_rq(cpu);
6056 unsigned long flags;
6058 raw_spin_lock_irqsave(&rq->lock, flags);
6061 idle->state = TASK_RUNNING;
6062 idle->se.exec_start = sched_clock();
6064 do_set_cpus_allowed(idle, cpumask_of(cpu));
6066 * We're having a chicken and egg problem, even though we are
6067 * holding rq->lock, the cpu isn't yet set to this cpu so the
6068 * lockdep check in task_group() will fail.
6070 * Similar case to sched_fork(). / Alternatively we could
6071 * use task_rq_lock() here and obtain the other rq->lock.
6076 __set_task_cpu(idle, cpu);
6079 rq->curr = rq->idle = idle;
6080 #if defined(CONFIG_SMP)
6083 raw_spin_unlock_irqrestore(&rq->lock, flags);
6085 /* Set the preempt count _outside_ the spinlocks! */
6086 task_thread_info(idle)->preempt_count = 0;
6089 * The idle tasks have their own, simple scheduling class:
6091 idle->sched_class = &idle_sched_class;
6092 ftrace_graph_init_idle_task(idle, cpu);
6096 * In a system that switches off the HZ timer nohz_cpu_mask
6097 * indicates which cpus entered this state. This is used
6098 * in the rcu update to wait only for active cpus. For system
6099 * which do not switch off the HZ timer nohz_cpu_mask should
6100 * always be CPU_BITS_NONE.
6102 cpumask_var_t nohz_cpu_mask;
6105 * Increase the granularity value when there are more CPUs,
6106 * because with more CPUs the 'effective latency' as visible
6107 * to users decreases. But the relationship is not linear,
6108 * so pick a second-best guess by going with the log2 of the
6111 * This idea comes from the SD scheduler of Con Kolivas:
6113 static int get_update_sysctl_factor(void)
6115 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6116 unsigned int factor;
6118 switch (sysctl_sched_tunable_scaling) {
6119 case SCHED_TUNABLESCALING_NONE:
6122 case SCHED_TUNABLESCALING_LINEAR:
6125 case SCHED_TUNABLESCALING_LOG:
6127 factor = 1 + ilog2(cpus);
6134 static void update_sysctl(void)
6136 unsigned int factor = get_update_sysctl_factor();
6138 #define SET_SYSCTL(name) \
6139 (sysctl_##name = (factor) * normalized_sysctl_##name)
6140 SET_SYSCTL(sched_min_granularity);
6141 SET_SYSCTL(sched_latency);
6142 SET_SYSCTL(sched_wakeup_granularity);
6146 static inline void sched_init_granularity(void)
6152 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6154 if (p->sched_class && p->sched_class->set_cpus_allowed)
6155 p->sched_class->set_cpus_allowed(p, new_mask);
6157 cpumask_copy(&p->cpus_allowed, new_mask);
6158 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6163 * This is how migration works:
6165 * 1) we invoke migration_cpu_stop() on the target CPU using
6167 * 2) stopper starts to run (implicitly forcing the migrated thread
6169 * 3) it checks whether the migrated task is still in the wrong runqueue.
6170 * 4) if it's in the wrong runqueue then the migration thread removes
6171 * it and puts it into the right queue.
6172 * 5) stopper completes and stop_one_cpu() returns and the migration
6177 * Change a given task's CPU affinity. Migrate the thread to a
6178 * proper CPU and schedule it away if the CPU it's executing on
6179 * is removed from the allowed bitmask.
6181 * NOTE: the caller must have a valid reference to the task, the
6182 * task must not exit() & deallocate itself prematurely. The
6183 * call is not atomic; no spinlocks may be held.
6185 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6187 unsigned long flags;
6189 unsigned int dest_cpu;
6192 rq = task_rq_lock(p, &flags);
6194 if (cpumask_equal(&p->cpus_allowed, new_mask))
6197 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6202 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6207 do_set_cpus_allowed(p, new_mask);
6209 /* Can the task run on the task's current CPU? If so, we're done */
6210 if (cpumask_test_cpu(task_cpu(p), new_mask))
6213 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6215 struct migration_arg arg = { p, dest_cpu };
6216 /* Need help from migration thread: drop lock and wait. */
6217 task_rq_unlock(rq, p, &flags);
6218 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6219 tlb_migrate_finish(p->mm);
6223 task_rq_unlock(rq, p, &flags);
6227 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6230 * Move (not current) task off this cpu, onto dest cpu. We're doing
6231 * this because either it can't run here any more (set_cpus_allowed()
6232 * away from this CPU, or CPU going down), or because we're
6233 * attempting to rebalance this task on exec (sched_exec).
6235 * So we race with normal scheduler movements, but that's OK, as long
6236 * as the task is no longer on this CPU.
6238 * Returns non-zero if task was successfully migrated.
6240 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6242 struct rq *rq_dest, *rq_src;
6245 if (unlikely(!cpu_active(dest_cpu)))
6248 rq_src = cpu_rq(src_cpu);
6249 rq_dest = cpu_rq(dest_cpu);
6251 raw_spin_lock(&p->pi_lock);
6252 double_rq_lock(rq_src, rq_dest);
6253 /* Already moved. */
6254 if (task_cpu(p) != src_cpu)
6256 /* Affinity changed (again). */
6257 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6261 * If we're not on a rq, the next wake-up will ensure we're
6265 deactivate_task(rq_src, p, 0);
6266 set_task_cpu(p, dest_cpu);
6267 activate_task(rq_dest, p, 0);
6268 check_preempt_curr(rq_dest, p, 0);
6273 double_rq_unlock(rq_src, rq_dest);
6274 raw_spin_unlock(&p->pi_lock);
6279 * migration_cpu_stop - this will be executed by a highprio stopper thread
6280 * and performs thread migration by bumping thread off CPU then
6281 * 'pushing' onto another runqueue.
6283 static int migration_cpu_stop(void *data)
6285 struct migration_arg *arg = data;
6288 * The original target cpu might have gone down and we might
6289 * be on another cpu but it doesn't matter.
6291 local_irq_disable();
6292 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6297 #ifdef CONFIG_HOTPLUG_CPU
6300 * Ensures that the idle task is using init_mm right before its cpu goes
6303 void idle_task_exit(void)
6305 struct mm_struct *mm = current->active_mm;
6307 BUG_ON(cpu_online(smp_processor_id()));
6310 switch_mm(mm, &init_mm, current);
6315 * While a dead CPU has no uninterruptible tasks queued at this point,
6316 * it might still have a nonzero ->nr_uninterruptible counter, because
6317 * for performance reasons the counter is not stricly tracking tasks to
6318 * their home CPUs. So we just add the counter to another CPU's counter,
6319 * to keep the global sum constant after CPU-down:
6321 static void migrate_nr_uninterruptible(struct rq *rq_src)
6323 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6325 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6326 rq_src->nr_uninterruptible = 0;
6330 * remove the tasks which were accounted by rq from calc_load_tasks.
6332 static void calc_global_load_remove(struct rq *rq)
6334 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6335 rq->calc_load_active = 0;
6338 #ifdef CONFIG_CFS_BANDWIDTH
6339 static void unthrottle_offline_cfs_rqs(struct rq *rq)
6341 struct cfs_rq *cfs_rq;
6343 for_each_leaf_cfs_rq(rq, cfs_rq) {
6344 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6346 if (!cfs_rq->runtime_enabled)
6350 * clock_task is not advancing so we just need to make sure
6351 * there's some valid quota amount
6353 cfs_rq->runtime_remaining = cfs_b->quota;
6354 if (cfs_rq_throttled(cfs_rq))
6355 unthrottle_cfs_rq(cfs_rq);
6359 static void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6363 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6364 * try_to_wake_up()->select_task_rq().
6366 * Called with rq->lock held even though we'er in stop_machine() and
6367 * there's no concurrency possible, we hold the required locks anyway
6368 * because of lock validation efforts.
6370 static void migrate_tasks(unsigned int dead_cpu)
6372 struct rq *rq = cpu_rq(dead_cpu);
6373 struct task_struct *next, *stop = rq->stop;
6377 * Fudge the rq selection such that the below task selection loop
6378 * doesn't get stuck on the currently eligible stop task.
6380 * We're currently inside stop_machine() and the rq is either stuck
6381 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6382 * either way we should never end up calling schedule() until we're
6387 /* Ensure any throttled groups are reachable by pick_next_task */
6388 unthrottle_offline_cfs_rqs(rq);
6392 * There's this thread running, bail when that's the only
6395 if (rq->nr_running == 1)
6398 next = pick_next_task(rq);
6400 next->sched_class->put_prev_task(rq, next);
6402 /* Find suitable destination for @next, with force if needed. */
6403 dest_cpu = select_fallback_rq(dead_cpu, next);
6404 raw_spin_unlock(&rq->lock);
6406 __migrate_task(next, dead_cpu, dest_cpu);
6408 raw_spin_lock(&rq->lock);
6414 #endif /* CONFIG_HOTPLUG_CPU */
6416 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6418 static struct ctl_table sd_ctl_dir[] = {
6420 .procname = "sched_domain",
6426 static struct ctl_table sd_ctl_root[] = {
6428 .procname = "kernel",
6430 .child = sd_ctl_dir,
6435 static struct ctl_table *sd_alloc_ctl_entry(int n)
6437 struct ctl_table *entry =
6438 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6443 static void sd_free_ctl_entry(struct ctl_table **tablep)
6445 struct ctl_table *entry;
6448 * In the intermediate directories, both the child directory and
6449 * procname are dynamically allocated and could fail but the mode
6450 * will always be set. In the lowest directory the names are
6451 * static strings and all have proc handlers.
6453 for (entry = *tablep; entry->mode; entry++) {
6455 sd_free_ctl_entry(&entry->child);
6456 if (entry->proc_handler == NULL)
6457 kfree(entry->procname);
6465 set_table_entry(struct ctl_table *entry,
6466 const char *procname, void *data, int maxlen,
6467 mode_t mode, proc_handler *proc_handler)
6469 entry->procname = procname;
6471 entry->maxlen = maxlen;
6473 entry->proc_handler = proc_handler;
6476 static struct ctl_table *
6477 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6479 struct ctl_table *table = sd_alloc_ctl_entry(13);
6484 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6485 sizeof(long), 0644, proc_doulongvec_minmax);
6486 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6487 sizeof(long), 0644, proc_doulongvec_minmax);
6488 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6489 sizeof(int), 0644, proc_dointvec_minmax);
6490 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6491 sizeof(int), 0644, proc_dointvec_minmax);
6492 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6493 sizeof(int), 0644, proc_dointvec_minmax);
6494 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6495 sizeof(int), 0644, proc_dointvec_minmax);
6496 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6497 sizeof(int), 0644, proc_dointvec_minmax);
6498 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6499 sizeof(int), 0644, proc_dointvec_minmax);
6500 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6501 sizeof(int), 0644, proc_dointvec_minmax);
6502 set_table_entry(&table[9], "cache_nice_tries",
6503 &sd->cache_nice_tries,
6504 sizeof(int), 0644, proc_dointvec_minmax);
6505 set_table_entry(&table[10], "flags", &sd->flags,
6506 sizeof(int), 0644, proc_dointvec_minmax);
6507 set_table_entry(&table[11], "name", sd->name,
6508 CORENAME_MAX_SIZE, 0444, proc_dostring);
6509 /* &table[12] is terminator */
6514 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6516 struct ctl_table *entry, *table;
6517 struct sched_domain *sd;
6518 int domain_num = 0, i;
6521 for_each_domain(cpu, sd)
6523 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6528 for_each_domain(cpu, sd) {
6529 snprintf(buf, 32, "domain%d", i);
6530 entry->procname = kstrdup(buf, GFP_KERNEL);
6532 entry->child = sd_alloc_ctl_domain_table(sd);
6539 static struct ctl_table_header *sd_sysctl_header;
6540 static void register_sched_domain_sysctl(void)
6542 int i, cpu_num = num_possible_cpus();
6543 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6546 WARN_ON(sd_ctl_dir[0].child);
6547 sd_ctl_dir[0].child = entry;
6552 for_each_possible_cpu(i) {
6553 snprintf(buf, 32, "cpu%d", i);
6554 entry->procname = kstrdup(buf, GFP_KERNEL);
6556 entry->child = sd_alloc_ctl_cpu_table(i);
6560 WARN_ON(sd_sysctl_header);
6561 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6564 /* may be called multiple times per register */
6565 static void unregister_sched_domain_sysctl(void)
6567 if (sd_sysctl_header)
6568 unregister_sysctl_table(sd_sysctl_header);
6569 sd_sysctl_header = NULL;
6570 if (sd_ctl_dir[0].child)
6571 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6574 static void register_sched_domain_sysctl(void)
6577 static void unregister_sched_domain_sysctl(void)
6582 static void set_rq_online(struct rq *rq)
6585 const struct sched_class *class;
6587 cpumask_set_cpu(rq->cpu, rq->rd->online);
6590 for_each_class(class) {
6591 if (class->rq_online)
6592 class->rq_online(rq);
6597 static void set_rq_offline(struct rq *rq)
6600 const struct sched_class *class;
6602 for_each_class(class) {
6603 if (class->rq_offline)
6604 class->rq_offline(rq);
6607 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6613 * migration_call - callback that gets triggered when a CPU is added.
6614 * Here we can start up the necessary migration thread for the new CPU.
6616 static int __cpuinit
6617 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6619 int cpu = (long)hcpu;
6620 unsigned long flags;
6621 struct rq *rq = cpu_rq(cpu);
6623 switch (action & ~CPU_TASKS_FROZEN) {
6625 case CPU_UP_PREPARE:
6626 rq->calc_load_update = calc_load_update;
6630 /* Update our root-domain */
6631 raw_spin_lock_irqsave(&rq->lock, flags);
6633 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6637 raw_spin_unlock_irqrestore(&rq->lock, flags);
6640 #ifdef CONFIG_HOTPLUG_CPU
6642 sched_ttwu_pending();
6643 /* Update our root-domain */
6644 raw_spin_lock_irqsave(&rq->lock, flags);
6646 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6650 BUG_ON(rq->nr_running != 1); /* the migration thread */
6651 raw_spin_unlock_irqrestore(&rq->lock, flags);
6653 migrate_nr_uninterruptible(rq);
6654 calc_global_load_remove(rq);
6659 update_max_interval();
6665 * Register at high priority so that task migration (migrate_all_tasks)
6666 * happens before everything else. This has to be lower priority than
6667 * the notifier in the perf_event subsystem, though.
6669 static struct notifier_block __cpuinitdata migration_notifier = {
6670 .notifier_call = migration_call,
6671 .priority = CPU_PRI_MIGRATION,
6674 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6675 unsigned long action, void *hcpu)
6677 switch (action & ~CPU_TASKS_FROZEN) {
6679 case CPU_DOWN_FAILED:
6680 set_cpu_active((long)hcpu, true);
6687 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6688 unsigned long action, void *hcpu)
6690 switch (action & ~CPU_TASKS_FROZEN) {
6691 case CPU_DOWN_PREPARE:
6692 set_cpu_active((long)hcpu, false);
6699 static int __init migration_init(void)
6701 void *cpu = (void *)(long)smp_processor_id();
6704 /* Initialize migration for the boot CPU */
6705 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6706 BUG_ON(err == NOTIFY_BAD);
6707 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6708 register_cpu_notifier(&migration_notifier);
6710 /* Register cpu active notifiers */
6711 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6712 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6716 early_initcall(migration_init);
6721 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6723 #ifdef CONFIG_SCHED_DEBUG
6725 static __read_mostly int sched_domain_debug_enabled;
6727 static int __init sched_domain_debug_setup(char *str)
6729 sched_domain_debug_enabled = 1;
6733 early_param("sched_debug", sched_domain_debug_setup);
6735 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6736 struct cpumask *groupmask)
6738 struct sched_group *group = sd->groups;
6741 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6742 cpumask_clear(groupmask);
6744 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6746 if (!(sd->flags & SD_LOAD_BALANCE)) {
6747 printk("does not load-balance\n");
6749 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6754 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6756 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6757 printk(KERN_ERR "ERROR: domain->span does not contain "
6760 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6761 printk(KERN_ERR "ERROR: domain->groups does not contain"
6765 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6769 printk(KERN_ERR "ERROR: group is NULL\n");
6773 if (!group->sgp->power) {
6774 printk(KERN_CONT "\n");
6775 printk(KERN_ERR "ERROR: domain->cpu_power not "
6780 if (!cpumask_weight(sched_group_cpus(group))) {
6781 printk(KERN_CONT "\n");
6782 printk(KERN_ERR "ERROR: empty group\n");
6786 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6787 printk(KERN_CONT "\n");
6788 printk(KERN_ERR "ERROR: repeated CPUs\n");
6792 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6794 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6796 printk(KERN_CONT " %s", str);
6797 if (group->sgp->power != SCHED_POWER_SCALE) {
6798 printk(KERN_CONT " (cpu_power = %d)",
6802 group = group->next;
6803 } while (group != sd->groups);
6804 printk(KERN_CONT "\n");
6806 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6807 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6810 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6811 printk(KERN_ERR "ERROR: parent span is not a superset "
6812 "of domain->span\n");
6816 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6820 if (!sched_domain_debug_enabled)
6824 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6828 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6831 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6839 #else /* !CONFIG_SCHED_DEBUG */
6840 # define sched_domain_debug(sd, cpu) do { } while (0)
6841 #endif /* CONFIG_SCHED_DEBUG */
6843 static int sd_degenerate(struct sched_domain *sd)
6845 if (cpumask_weight(sched_domain_span(sd)) == 1)
6848 /* Following flags need at least 2 groups */
6849 if (sd->flags & (SD_LOAD_BALANCE |
6850 SD_BALANCE_NEWIDLE |
6854 SD_SHARE_PKG_RESOURCES)) {
6855 if (sd->groups != sd->groups->next)
6859 /* Following flags don't use groups */
6860 if (sd->flags & (SD_WAKE_AFFINE))
6867 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6869 unsigned long cflags = sd->flags, pflags = parent->flags;
6871 if (sd_degenerate(parent))
6874 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6877 /* Flags needing groups don't count if only 1 group in parent */
6878 if (parent->groups == parent->groups->next) {
6879 pflags &= ~(SD_LOAD_BALANCE |
6880 SD_BALANCE_NEWIDLE |
6884 SD_SHARE_PKG_RESOURCES);
6885 if (nr_node_ids == 1)
6886 pflags &= ~SD_SERIALIZE;
6888 if (~cflags & pflags)
6894 static void free_rootdomain(struct rcu_head *rcu)
6896 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6898 cpupri_cleanup(&rd->cpupri);
6899 free_cpumask_var(rd->rto_mask);
6900 free_cpumask_var(rd->online);
6901 free_cpumask_var(rd->span);
6905 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6907 struct root_domain *old_rd = NULL;
6908 unsigned long flags;
6910 raw_spin_lock_irqsave(&rq->lock, flags);
6915 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6918 cpumask_clear_cpu(rq->cpu, old_rd->span);
6921 * If we dont want to free the old_rt yet then
6922 * set old_rd to NULL to skip the freeing later
6925 if (!atomic_dec_and_test(&old_rd->refcount))
6929 atomic_inc(&rd->refcount);
6932 cpumask_set_cpu(rq->cpu, rd->span);
6933 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6936 raw_spin_unlock_irqrestore(&rq->lock, flags);
6939 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6942 static int init_rootdomain(struct root_domain *rd)
6944 memset(rd, 0, sizeof(*rd));
6946 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6948 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6950 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6953 if (cpupri_init(&rd->cpupri) != 0)
6958 free_cpumask_var(rd->rto_mask);
6960 free_cpumask_var(rd->online);
6962 free_cpumask_var(rd->span);
6967 static void init_defrootdomain(void)
6969 init_rootdomain(&def_root_domain);
6971 atomic_set(&def_root_domain.refcount, 1);
6974 static struct root_domain *alloc_rootdomain(void)
6976 struct root_domain *rd;
6978 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6982 if (init_rootdomain(rd) != 0) {
6990 static void free_sched_groups(struct sched_group *sg, int free_sgp)
6992 struct sched_group *tmp, *first;
7001 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
7006 } while (sg != first);
7009 static void free_sched_domain(struct rcu_head *rcu)
7011 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
7014 * If its an overlapping domain it has private groups, iterate and
7017 if (sd->flags & SD_OVERLAP) {
7018 free_sched_groups(sd->groups, 1);
7019 } else if (atomic_dec_and_test(&sd->groups->ref)) {
7020 kfree(sd->groups->sgp);
7026 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
7028 call_rcu(&sd->rcu, free_sched_domain);
7031 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
7033 for (; sd; sd = sd->parent)
7034 destroy_sched_domain(sd, cpu);
7038 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7039 * hold the hotplug lock.
7042 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7044 struct rq *rq = cpu_rq(cpu);
7045 struct sched_domain *tmp;
7047 /* Remove the sched domains which do not contribute to scheduling. */
7048 for (tmp = sd; tmp; ) {
7049 struct sched_domain *parent = tmp->parent;
7053 if (sd_parent_degenerate(tmp, parent)) {
7054 tmp->parent = parent->parent;
7056 parent->parent->child = tmp;
7057 destroy_sched_domain(parent, cpu);
7062 if (sd && sd_degenerate(sd)) {
7065 destroy_sched_domain(tmp, cpu);
7070 sched_domain_debug(sd, cpu);
7072 rq_attach_root(rq, rd);
7074 rcu_assign_pointer(rq->sd, sd);
7075 destroy_sched_domains(tmp, cpu);
7078 /* cpus with isolated domains */
7079 static cpumask_var_t cpu_isolated_map;
7081 /* Setup the mask of cpus configured for isolated domains */
7082 static int __init isolated_cpu_setup(char *str)
7084 alloc_bootmem_cpumask_var(&cpu_isolated_map);
7085 cpulist_parse(str, cpu_isolated_map);
7089 __setup("isolcpus=", isolated_cpu_setup);
7091 #define SD_NODES_PER_DOMAIN 16
7096 * find_next_best_node - find the next node to include in a sched_domain
7097 * @node: node whose sched_domain we're building
7098 * @used_nodes: nodes already in the sched_domain
7100 * Find the next node to include in a given scheduling domain. Simply
7101 * finds the closest node not already in the @used_nodes map.
7103 * Should use nodemask_t.
7105 static int find_next_best_node(int node, nodemask_t *used_nodes)
7107 int i, n, val, min_val, best_node = -1;
7111 for (i = 0; i < nr_node_ids; i++) {
7112 /* Start at @node */
7113 n = (node + i) % nr_node_ids;
7115 if (!nr_cpus_node(n))
7118 /* Skip already used nodes */
7119 if (node_isset(n, *used_nodes))
7122 /* Simple min distance search */
7123 val = node_distance(node, n);
7125 if (val < min_val) {
7131 if (best_node != -1)
7132 node_set(best_node, *used_nodes);
7137 * sched_domain_node_span - get a cpumask for a node's sched_domain
7138 * @node: node whose cpumask we're constructing
7139 * @span: resulting cpumask
7141 * Given a node, construct a good cpumask for its sched_domain to span. It
7142 * should be one that prevents unnecessary balancing, but also spreads tasks
7145 static void sched_domain_node_span(int node, struct cpumask *span)
7147 nodemask_t used_nodes;
7150 cpumask_clear(span);
7151 nodes_clear(used_nodes);
7153 cpumask_or(span, span, cpumask_of_node(node));
7154 node_set(node, used_nodes);
7156 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7157 int next_node = find_next_best_node(node, &used_nodes);
7160 cpumask_or(span, span, cpumask_of_node(next_node));
7164 static const struct cpumask *cpu_node_mask(int cpu)
7166 lockdep_assert_held(&sched_domains_mutex);
7168 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7170 return sched_domains_tmpmask;
7173 static const struct cpumask *cpu_allnodes_mask(int cpu)
7175 return cpu_possible_mask;
7177 #endif /* CONFIG_NUMA */
7179 static const struct cpumask *cpu_cpu_mask(int cpu)
7181 return cpumask_of_node(cpu_to_node(cpu));
7184 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7187 struct sched_domain **__percpu sd;
7188 struct sched_group **__percpu sg;
7189 struct sched_group_power **__percpu sgp;
7193 struct sched_domain ** __percpu sd;
7194 struct root_domain *rd;
7204 struct sched_domain_topology_level;
7206 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7207 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7209 #define SDTL_OVERLAP 0x01
7211 struct sched_domain_topology_level {
7212 sched_domain_init_f init;
7213 sched_domain_mask_f mask;
7215 struct sd_data data;
7219 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7221 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7222 const struct cpumask *span = sched_domain_span(sd);
7223 struct cpumask *covered = sched_domains_tmpmask;
7224 struct sd_data *sdd = sd->private;
7225 struct sched_domain *child;
7228 cpumask_clear(covered);
7230 for_each_cpu(i, span) {
7231 struct cpumask *sg_span;
7233 if (cpumask_test_cpu(i, covered))
7236 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7237 GFP_KERNEL, cpu_to_node(i));
7242 sg_span = sched_group_cpus(sg);
7244 child = *per_cpu_ptr(sdd->sd, i);
7246 child = child->child;
7247 cpumask_copy(sg_span, sched_domain_span(child));
7249 cpumask_set_cpu(i, sg_span);
7251 cpumask_or(covered, covered, sg_span);
7253 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7254 atomic_inc(&sg->sgp->ref);
7256 if (cpumask_test_cpu(cpu, sg_span))
7266 sd->groups = groups;
7271 free_sched_groups(first, 0);
7276 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7278 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7279 struct sched_domain *child = sd->child;
7282 cpu = cpumask_first(sched_domain_span(child));
7285 *sg = *per_cpu_ptr(sdd->sg, cpu);
7286 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7287 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7294 * build_sched_groups will build a circular linked list of the groups
7295 * covered by the given span, and will set each group's ->cpumask correctly,
7296 * and ->cpu_power to 0.
7298 * Assumes the sched_domain tree is fully constructed
7301 build_sched_groups(struct sched_domain *sd, int cpu)
7303 struct sched_group *first = NULL, *last = NULL;
7304 struct sd_data *sdd = sd->private;
7305 const struct cpumask *span = sched_domain_span(sd);
7306 struct cpumask *covered;
7309 get_group(cpu, sdd, &sd->groups);
7310 atomic_inc(&sd->groups->ref);
7312 if (cpu != cpumask_first(sched_domain_span(sd)))
7315 lockdep_assert_held(&sched_domains_mutex);
7316 covered = sched_domains_tmpmask;
7318 cpumask_clear(covered);
7320 for_each_cpu(i, span) {
7321 struct sched_group *sg;
7322 int group = get_group(i, sdd, &sg);
7325 if (cpumask_test_cpu(i, covered))
7328 cpumask_clear(sched_group_cpus(sg));
7331 for_each_cpu(j, span) {
7332 if (get_group(j, sdd, NULL) != group)
7335 cpumask_set_cpu(j, covered);
7336 cpumask_set_cpu(j, sched_group_cpus(sg));
7351 * Initialize sched groups cpu_power.
7353 * cpu_power indicates the capacity of sched group, which is used while
7354 * distributing the load between different sched groups in a sched domain.
7355 * Typically cpu_power for all the groups in a sched domain will be same unless
7356 * there are asymmetries in the topology. If there are asymmetries, group
7357 * having more cpu_power will pickup more load compared to the group having
7360 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7362 struct sched_group *sg = sd->groups;
7364 WARN_ON(!sd || !sg);
7367 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7369 } while (sg != sd->groups);
7371 if (cpu != group_first_cpu(sg))
7374 update_group_power(sd, cpu);
7378 * Initializers for schedule domains
7379 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7382 #ifdef CONFIG_SCHED_DEBUG
7383 # define SD_INIT_NAME(sd, type) sd->name = #type
7385 # define SD_INIT_NAME(sd, type) do { } while (0)
7388 #define SD_INIT_FUNC(type) \
7389 static noinline struct sched_domain * \
7390 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7392 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7393 *sd = SD_##type##_INIT; \
7394 SD_INIT_NAME(sd, type); \
7395 sd->private = &tl->data; \
7401 SD_INIT_FUNC(ALLNODES)
7404 #ifdef CONFIG_SCHED_SMT
7405 SD_INIT_FUNC(SIBLING)
7407 #ifdef CONFIG_SCHED_MC
7410 #ifdef CONFIG_SCHED_BOOK
7414 static int default_relax_domain_level = -1;
7415 int sched_domain_level_max;
7417 static int __init setup_relax_domain_level(char *str)
7421 val = simple_strtoul(str, NULL, 0);
7422 if (val < sched_domain_level_max)
7423 default_relax_domain_level = val;
7427 __setup("relax_domain_level=", setup_relax_domain_level);
7429 static void set_domain_attribute(struct sched_domain *sd,
7430 struct sched_domain_attr *attr)
7434 if (!attr || attr->relax_domain_level < 0) {
7435 if (default_relax_domain_level < 0)
7438 request = default_relax_domain_level;
7440 request = attr->relax_domain_level;
7441 if (request < sd->level) {
7442 /* turn off idle balance on this domain */
7443 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7445 /* turn on idle balance on this domain */
7446 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7450 static void __sdt_free(const struct cpumask *cpu_map);
7451 static int __sdt_alloc(const struct cpumask *cpu_map);
7453 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7454 const struct cpumask *cpu_map)
7458 if (!atomic_read(&d->rd->refcount))
7459 free_rootdomain(&d->rd->rcu); /* fall through */
7461 free_percpu(d->sd); /* fall through */
7463 __sdt_free(cpu_map); /* fall through */
7469 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7470 const struct cpumask *cpu_map)
7472 memset(d, 0, sizeof(*d));
7474 if (__sdt_alloc(cpu_map))
7475 return sa_sd_storage;
7476 d->sd = alloc_percpu(struct sched_domain *);
7478 return sa_sd_storage;
7479 d->rd = alloc_rootdomain();
7482 return sa_rootdomain;
7486 * NULL the sd_data elements we've used to build the sched_domain and
7487 * sched_group structure so that the subsequent __free_domain_allocs()
7488 * will not free the data we're using.
7490 static void claim_allocations(int cpu, struct sched_domain *sd)
7492 struct sd_data *sdd = sd->private;
7494 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7495 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7497 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7498 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7500 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7501 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7504 #ifdef CONFIG_SCHED_SMT
7505 static const struct cpumask *cpu_smt_mask(int cpu)
7507 return topology_thread_cpumask(cpu);
7512 * Topology list, bottom-up.
7514 static struct sched_domain_topology_level default_topology[] = {
7515 #ifdef CONFIG_SCHED_SMT
7516 { sd_init_SIBLING, cpu_smt_mask, },
7518 #ifdef CONFIG_SCHED_MC
7519 { sd_init_MC, cpu_coregroup_mask, },
7521 #ifdef CONFIG_SCHED_BOOK
7522 { sd_init_BOOK, cpu_book_mask, },
7524 { sd_init_CPU, cpu_cpu_mask, },
7526 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7527 { sd_init_ALLNODES, cpu_allnodes_mask, },
7532 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7534 static int __sdt_alloc(const struct cpumask *cpu_map)
7536 struct sched_domain_topology_level *tl;
7539 for (tl = sched_domain_topology; tl->init; tl++) {
7540 struct sd_data *sdd = &tl->data;
7542 sdd->sd = alloc_percpu(struct sched_domain *);
7546 sdd->sg = alloc_percpu(struct sched_group *);
7550 sdd->sgp = alloc_percpu(struct sched_group_power *);
7554 for_each_cpu(j, cpu_map) {
7555 struct sched_domain *sd;
7556 struct sched_group *sg;
7557 struct sched_group_power *sgp;
7559 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7560 GFP_KERNEL, cpu_to_node(j));
7564 *per_cpu_ptr(sdd->sd, j) = sd;
7566 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7567 GFP_KERNEL, cpu_to_node(j));
7571 *per_cpu_ptr(sdd->sg, j) = sg;
7573 sgp = kzalloc_node(sizeof(struct sched_group_power),
7574 GFP_KERNEL, cpu_to_node(j));
7578 *per_cpu_ptr(sdd->sgp, j) = sgp;
7585 static void __sdt_free(const struct cpumask *cpu_map)
7587 struct sched_domain_topology_level *tl;
7590 for (tl = sched_domain_topology; tl->init; tl++) {
7591 struct sd_data *sdd = &tl->data;
7593 for_each_cpu(j, cpu_map) {
7594 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7595 if (sd && (sd->flags & SD_OVERLAP))
7596 free_sched_groups(sd->groups, 0);
7597 kfree(*per_cpu_ptr(sdd->sg, j));
7598 kfree(*per_cpu_ptr(sdd->sgp, j));
7600 free_percpu(sdd->sd);
7601 free_percpu(sdd->sg);
7602 free_percpu(sdd->sgp);
7606 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7607 struct s_data *d, const struct cpumask *cpu_map,
7608 struct sched_domain_attr *attr, struct sched_domain *child,
7611 struct sched_domain *sd = tl->init(tl, cpu);
7615 set_domain_attribute(sd, attr);
7616 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7618 sd->level = child->level + 1;
7619 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7628 * Build sched domains for a given set of cpus and attach the sched domains
7629 * to the individual cpus
7631 static int build_sched_domains(const struct cpumask *cpu_map,
7632 struct sched_domain_attr *attr)
7634 enum s_alloc alloc_state = sa_none;
7635 struct sched_domain *sd;
7637 int i, ret = -ENOMEM;
7639 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7640 if (alloc_state != sa_rootdomain)
7643 /* Set up domains for cpus specified by the cpu_map. */
7644 for_each_cpu(i, cpu_map) {
7645 struct sched_domain_topology_level *tl;
7648 for (tl = sched_domain_topology; tl->init; tl++) {
7649 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7650 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7651 sd->flags |= SD_OVERLAP;
7652 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7659 *per_cpu_ptr(d.sd, i) = sd;
7662 /* Build the groups for the domains */
7663 for_each_cpu(i, cpu_map) {
7664 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7665 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7666 if (sd->flags & SD_OVERLAP) {
7667 if (build_overlap_sched_groups(sd, i))
7670 if (build_sched_groups(sd, i))
7676 /* Calculate CPU power for physical packages and nodes */
7677 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7678 if (!cpumask_test_cpu(i, cpu_map))
7681 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7682 claim_allocations(i, sd);
7683 init_sched_groups_power(i, sd);
7687 /* Attach the domains */
7689 for_each_cpu(i, cpu_map) {
7690 sd = *per_cpu_ptr(d.sd, i);
7691 cpu_attach_domain(sd, d.rd, i);
7697 __free_domain_allocs(&d, alloc_state, cpu_map);
7701 static cpumask_var_t *doms_cur; /* current sched domains */
7702 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7703 static struct sched_domain_attr *dattr_cur;
7704 /* attribues of custom domains in 'doms_cur' */
7707 * Special case: If a kmalloc of a doms_cur partition (array of
7708 * cpumask) fails, then fallback to a single sched domain,
7709 * as determined by the single cpumask fallback_doms.
7711 static cpumask_var_t fallback_doms;
7714 * arch_update_cpu_topology lets virtualized architectures update the
7715 * cpu core maps. It is supposed to return 1 if the topology changed
7716 * or 0 if it stayed the same.
7718 int __attribute__((weak)) arch_update_cpu_topology(void)
7723 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7726 cpumask_var_t *doms;
7728 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7731 for (i = 0; i < ndoms; i++) {
7732 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7733 free_sched_domains(doms, i);
7740 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7743 for (i = 0; i < ndoms; i++)
7744 free_cpumask_var(doms[i]);
7749 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7750 * For now this just excludes isolated cpus, but could be used to
7751 * exclude other special cases in the future.
7753 static int init_sched_domains(const struct cpumask *cpu_map)
7757 arch_update_cpu_topology();
7759 doms_cur = alloc_sched_domains(ndoms_cur);
7761 doms_cur = &fallback_doms;
7762 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7764 err = build_sched_domains(doms_cur[0], NULL);
7765 register_sched_domain_sysctl();
7771 * Detach sched domains from a group of cpus specified in cpu_map
7772 * These cpus will now be attached to the NULL domain
7774 static void detach_destroy_domains(const struct cpumask *cpu_map)
7779 for_each_cpu(i, cpu_map)
7780 cpu_attach_domain(NULL, &def_root_domain, i);
7784 /* handle null as "default" */
7785 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7786 struct sched_domain_attr *new, int idx_new)
7788 struct sched_domain_attr tmp;
7795 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7796 new ? (new + idx_new) : &tmp,
7797 sizeof(struct sched_domain_attr));
7801 * Partition sched domains as specified by the 'ndoms_new'
7802 * cpumasks in the array doms_new[] of cpumasks. This compares
7803 * doms_new[] to the current sched domain partitioning, doms_cur[].
7804 * It destroys each deleted domain and builds each new domain.
7806 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7807 * The masks don't intersect (don't overlap.) We should setup one
7808 * sched domain for each mask. CPUs not in any of the cpumasks will
7809 * not be load balanced. If the same cpumask appears both in the
7810 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7813 * The passed in 'doms_new' should be allocated using
7814 * alloc_sched_domains. This routine takes ownership of it and will
7815 * free_sched_domains it when done with it. If the caller failed the
7816 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7817 * and partition_sched_domains() will fallback to the single partition
7818 * 'fallback_doms', it also forces the domains to be rebuilt.
7820 * If doms_new == NULL it will be replaced with cpu_online_mask.
7821 * ndoms_new == 0 is a special case for destroying existing domains,
7822 * and it will not create the default domain.
7824 * Call with hotplug lock held
7826 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7827 struct sched_domain_attr *dattr_new)
7832 mutex_lock(&sched_domains_mutex);
7834 /* always unregister in case we don't destroy any domains */
7835 unregister_sched_domain_sysctl();
7837 /* Let architecture update cpu core mappings. */
7838 new_topology = arch_update_cpu_topology();
7840 n = doms_new ? ndoms_new : 0;
7842 /* Destroy deleted domains */
7843 for (i = 0; i < ndoms_cur; i++) {
7844 for (j = 0; j < n && !new_topology; j++) {
7845 if (cpumask_equal(doms_cur[i], doms_new[j])
7846 && dattrs_equal(dattr_cur, i, dattr_new, j))
7849 /* no match - a current sched domain not in new doms_new[] */
7850 detach_destroy_domains(doms_cur[i]);
7855 if (doms_new == NULL) {
7857 doms_new = &fallback_doms;
7858 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7859 WARN_ON_ONCE(dattr_new);
7862 /* Build new domains */
7863 for (i = 0; i < ndoms_new; i++) {
7864 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7865 if (cpumask_equal(doms_new[i], doms_cur[j])
7866 && dattrs_equal(dattr_new, i, dattr_cur, j))
7869 /* no match - add a new doms_new */
7870 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7875 /* Remember the new sched domains */
7876 if (doms_cur != &fallback_doms)
7877 free_sched_domains(doms_cur, ndoms_cur);
7878 kfree(dattr_cur); /* kfree(NULL) is safe */
7879 doms_cur = doms_new;
7880 dattr_cur = dattr_new;
7881 ndoms_cur = ndoms_new;
7883 register_sched_domain_sysctl();
7885 mutex_unlock(&sched_domains_mutex);
7888 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7889 static void reinit_sched_domains(void)
7893 /* Destroy domains first to force the rebuild */
7894 partition_sched_domains(0, NULL, NULL);
7896 rebuild_sched_domains();
7900 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7902 unsigned int level = 0;
7904 if (sscanf(buf, "%u", &level) != 1)
7908 * level is always be positive so don't check for
7909 * level < POWERSAVINGS_BALANCE_NONE which is 0
7910 * What happens on 0 or 1 byte write,
7911 * need to check for count as well?
7914 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7918 sched_smt_power_savings = level;
7920 sched_mc_power_savings = level;
7922 reinit_sched_domains();
7927 #ifdef CONFIG_SCHED_MC
7928 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7929 struct sysdev_class_attribute *attr,
7932 return sprintf(page, "%u\n", sched_mc_power_savings);
7934 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7935 struct sysdev_class_attribute *attr,
7936 const char *buf, size_t count)
7938 return sched_power_savings_store(buf, count, 0);
7940 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7941 sched_mc_power_savings_show,
7942 sched_mc_power_savings_store);
7945 #ifdef CONFIG_SCHED_SMT
7946 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7947 struct sysdev_class_attribute *attr,
7950 return sprintf(page, "%u\n", sched_smt_power_savings);
7952 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7953 struct sysdev_class_attribute *attr,
7954 const char *buf, size_t count)
7956 return sched_power_savings_store(buf, count, 1);
7958 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7959 sched_smt_power_savings_show,
7960 sched_smt_power_savings_store);
7963 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7967 #ifdef CONFIG_SCHED_SMT
7969 err = sysfs_create_file(&cls->kset.kobj,
7970 &attr_sched_smt_power_savings.attr);
7972 #ifdef CONFIG_SCHED_MC
7973 if (!err && mc_capable())
7974 err = sysfs_create_file(&cls->kset.kobj,
7975 &attr_sched_mc_power_savings.attr);
7979 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7982 * Update cpusets according to cpu_active mask. If cpusets are
7983 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7984 * around partition_sched_domains().
7986 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7989 switch (action & ~CPU_TASKS_FROZEN) {
7991 case CPU_DOWN_FAILED:
7992 cpuset_update_active_cpus();
7999 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
8002 switch (action & ~CPU_TASKS_FROZEN) {
8003 case CPU_DOWN_PREPARE:
8004 cpuset_update_active_cpus();
8011 static int update_runtime(struct notifier_block *nfb,
8012 unsigned long action, void *hcpu)
8014 int cpu = (int)(long)hcpu;
8017 case CPU_DOWN_PREPARE:
8018 case CPU_DOWN_PREPARE_FROZEN:
8019 disable_runtime(cpu_rq(cpu));
8022 case CPU_DOWN_FAILED:
8023 case CPU_DOWN_FAILED_FROZEN:
8025 case CPU_ONLINE_FROZEN:
8026 enable_runtime(cpu_rq(cpu));
8034 void __init sched_init_smp(void)
8036 cpumask_var_t non_isolated_cpus;
8038 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8039 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8042 mutex_lock(&sched_domains_mutex);
8043 init_sched_domains(cpu_active_mask);
8044 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8045 if (cpumask_empty(non_isolated_cpus))
8046 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8047 mutex_unlock(&sched_domains_mutex);
8050 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
8051 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
8053 /* RT runtime code needs to handle some hotplug events */
8054 hotcpu_notifier(update_runtime, 0);
8058 /* Move init over to a non-isolated CPU */
8059 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8061 sched_init_granularity();
8062 free_cpumask_var(non_isolated_cpus);
8064 init_sched_rt_class();
8067 void __init sched_init_smp(void)
8069 sched_init_granularity();
8071 #endif /* CONFIG_SMP */
8073 const_debug unsigned int sysctl_timer_migration = 1;
8075 int in_sched_functions(unsigned long addr)
8077 return in_lock_functions(addr) ||
8078 (addr >= (unsigned long)__sched_text_start
8079 && addr < (unsigned long)__sched_text_end);
8082 static void init_cfs_rq(struct cfs_rq *cfs_rq)
8084 cfs_rq->tasks_timeline = RB_ROOT;
8085 INIT_LIST_HEAD(&cfs_rq->tasks);
8086 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8087 #ifndef CONFIG_64BIT
8088 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8092 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8094 struct rt_prio_array *array;
8097 array = &rt_rq->active;
8098 for (i = 0; i < MAX_RT_PRIO; i++) {
8099 INIT_LIST_HEAD(array->queue + i);
8100 __clear_bit(i, array->bitmap);
8102 /* delimiter for bitsearch: */
8103 __set_bit(MAX_RT_PRIO, array->bitmap);
8105 #if defined CONFIG_SMP
8106 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8107 rt_rq->highest_prio.next = MAX_RT_PRIO;
8108 rt_rq->rt_nr_migratory = 0;
8109 rt_rq->overloaded = 0;
8110 plist_head_init(&rt_rq->pushable_tasks);
8114 rt_rq->rt_throttled = 0;
8115 rt_rq->rt_runtime = 0;
8116 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8119 #ifdef CONFIG_FAIR_GROUP_SCHED
8120 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8121 struct sched_entity *se, int cpu,
8122 struct sched_entity *parent)
8124 struct rq *rq = cpu_rq(cpu);
8129 /* allow initial update_cfs_load() to truncate */
8130 cfs_rq->load_stamp = 1;
8132 init_cfs_rq_runtime(cfs_rq);
8134 tg->cfs_rq[cpu] = cfs_rq;
8137 /* se could be NULL for root_task_group */
8142 se->cfs_rq = &rq->cfs;
8144 se->cfs_rq = parent->my_q;
8147 update_load_set(&se->load, 0);
8148 se->parent = parent;
8152 #ifdef CONFIG_RT_GROUP_SCHED
8153 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8154 struct sched_rt_entity *rt_se, int cpu,
8155 struct sched_rt_entity *parent)
8157 struct rq *rq = cpu_rq(cpu);
8159 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8160 rt_rq->rt_nr_boosted = 0;
8164 tg->rt_rq[cpu] = rt_rq;
8165 tg->rt_se[cpu] = rt_se;
8171 rt_se->rt_rq = &rq->rt;
8173 rt_se->rt_rq = parent->my_q;
8175 rt_se->my_q = rt_rq;
8176 rt_se->parent = parent;
8177 INIT_LIST_HEAD(&rt_se->run_list);
8181 void __init sched_init(void)
8184 unsigned long alloc_size = 0, ptr;
8186 #ifdef CONFIG_FAIR_GROUP_SCHED
8187 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8189 #ifdef CONFIG_RT_GROUP_SCHED
8190 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8192 #ifdef CONFIG_CPUMASK_OFFSTACK
8193 alloc_size += num_possible_cpus() * cpumask_size();
8196 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8198 #ifdef CONFIG_FAIR_GROUP_SCHED
8199 root_task_group.se = (struct sched_entity **)ptr;
8200 ptr += nr_cpu_ids * sizeof(void **);
8202 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8203 ptr += nr_cpu_ids * sizeof(void **);
8205 #endif /* CONFIG_FAIR_GROUP_SCHED */
8206 #ifdef CONFIG_RT_GROUP_SCHED
8207 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8208 ptr += nr_cpu_ids * sizeof(void **);
8210 root_task_group.rt_rq = (struct rt_rq **)ptr;
8211 ptr += nr_cpu_ids * sizeof(void **);
8213 #endif /* CONFIG_RT_GROUP_SCHED */
8214 #ifdef CONFIG_CPUMASK_OFFSTACK
8215 for_each_possible_cpu(i) {
8216 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8217 ptr += cpumask_size();
8219 #endif /* CONFIG_CPUMASK_OFFSTACK */
8223 init_defrootdomain();
8226 init_rt_bandwidth(&def_rt_bandwidth,
8227 global_rt_period(), global_rt_runtime());
8229 #ifdef CONFIG_RT_GROUP_SCHED
8230 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8231 global_rt_period(), global_rt_runtime());
8232 #endif /* CONFIG_RT_GROUP_SCHED */
8234 #ifdef CONFIG_CGROUP_SCHED
8235 list_add(&root_task_group.list, &task_groups);
8236 INIT_LIST_HEAD(&root_task_group.children);
8237 autogroup_init(&init_task);
8238 #endif /* CONFIG_CGROUP_SCHED */
8240 for_each_possible_cpu(i) {
8244 raw_spin_lock_init(&rq->lock);
8246 rq->calc_load_active = 0;
8247 rq->calc_load_update = jiffies + LOAD_FREQ;
8248 init_cfs_rq(&rq->cfs);
8249 init_rt_rq(&rq->rt, rq);
8250 #ifdef CONFIG_FAIR_GROUP_SCHED
8251 root_task_group.shares = root_task_group_load;
8252 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8254 * How much cpu bandwidth does root_task_group get?
8256 * In case of task-groups formed thr' the cgroup filesystem, it
8257 * gets 100% of the cpu resources in the system. This overall
8258 * system cpu resource is divided among the tasks of
8259 * root_task_group and its child task-groups in a fair manner,
8260 * based on each entity's (task or task-group's) weight
8261 * (se->load.weight).
8263 * In other words, if root_task_group has 10 tasks of weight
8264 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8265 * then A0's share of the cpu resource is:
8267 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8269 * We achieve this by letting root_task_group's tasks sit
8270 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8272 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8273 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8274 #endif /* CONFIG_FAIR_GROUP_SCHED */
8276 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8277 #ifdef CONFIG_RT_GROUP_SCHED
8278 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8279 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8282 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8283 rq->cpu_load[j] = 0;
8285 rq->last_load_update_tick = jiffies;
8290 rq->cpu_power = SCHED_POWER_SCALE;
8291 rq->post_schedule = 0;
8292 rq->active_balance = 0;
8293 rq->next_balance = jiffies;
8298 rq->avg_idle = 2*sysctl_sched_migration_cost;
8299 rq_attach_root(rq, &def_root_domain);
8301 rq->nohz_balance_kick = 0;
8302 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8306 atomic_set(&rq->nr_iowait, 0);
8309 set_load_weight(&init_task);
8311 #ifdef CONFIG_PREEMPT_NOTIFIERS
8312 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8316 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8319 #ifdef CONFIG_RT_MUTEXES
8320 plist_head_init(&init_task.pi_waiters);
8324 * The boot idle thread does lazy MMU switching as well:
8326 atomic_inc(&init_mm.mm_count);
8327 enter_lazy_tlb(&init_mm, current);
8330 * Make us the idle thread. Technically, schedule() should not be
8331 * called from this thread, however somewhere below it might be,
8332 * but because we are the idle thread, we just pick up running again
8333 * when this runqueue becomes "idle".
8335 init_idle(current, smp_processor_id());
8337 calc_load_update = jiffies + LOAD_FREQ;
8340 * During early bootup we pretend to be a normal task:
8342 current->sched_class = &fair_sched_class;
8344 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8345 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8347 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8349 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8350 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8351 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8352 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8353 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8355 /* May be allocated at isolcpus cmdline parse time */
8356 if (cpu_isolated_map == NULL)
8357 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8360 scheduler_running = 1;
8363 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8364 static inline int preempt_count_equals(int preempt_offset)
8366 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8368 return (nested == preempt_offset);
8371 void __might_sleep(const char *file, int line, int preempt_offset)
8373 static unsigned long prev_jiffy; /* ratelimiting */
8375 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8376 system_state != SYSTEM_RUNNING || oops_in_progress)
8378 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8380 prev_jiffy = jiffies;
8383 "BUG: sleeping function called from invalid context at %s:%d\n",
8386 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8387 in_atomic(), irqs_disabled(),
8388 current->pid, current->comm);
8390 debug_show_held_locks(current);
8391 if (irqs_disabled())
8392 print_irqtrace_events(current);
8395 EXPORT_SYMBOL(__might_sleep);
8398 #ifdef CONFIG_MAGIC_SYSRQ
8399 static void normalize_task(struct rq *rq, struct task_struct *p)
8401 const struct sched_class *prev_class = p->sched_class;
8402 int old_prio = p->prio;
8407 deactivate_task(rq, p, 0);
8408 __setscheduler(rq, p, SCHED_NORMAL, 0);
8410 activate_task(rq, p, 0);
8411 resched_task(rq->curr);
8414 check_class_changed(rq, p, prev_class, old_prio);
8417 void normalize_rt_tasks(void)
8419 struct task_struct *g, *p;
8420 unsigned long flags;
8423 read_lock_irqsave(&tasklist_lock, flags);
8424 do_each_thread(g, p) {
8426 * Only normalize user tasks:
8431 p->se.exec_start = 0;
8432 #ifdef CONFIG_SCHEDSTATS
8433 p->se.statistics.wait_start = 0;
8434 p->se.statistics.sleep_start = 0;
8435 p->se.statistics.block_start = 0;
8440 * Renice negative nice level userspace
8443 if (TASK_NICE(p) < 0 && p->mm)
8444 set_user_nice(p, 0);
8448 raw_spin_lock(&p->pi_lock);
8449 rq = __task_rq_lock(p);
8451 normalize_task(rq, p);
8453 __task_rq_unlock(rq);
8454 raw_spin_unlock(&p->pi_lock);
8455 } while_each_thread(g, p);
8457 read_unlock_irqrestore(&tasklist_lock, flags);
8460 #endif /* CONFIG_MAGIC_SYSRQ */
8462 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8464 * These functions are only useful for the IA64 MCA handling, or kdb.
8466 * They can only be called when the whole system has been
8467 * stopped - every CPU needs to be quiescent, and no scheduling
8468 * activity can take place. Using them for anything else would
8469 * be a serious bug, and as a result, they aren't even visible
8470 * under any other configuration.
8474 * curr_task - return the current task for a given cpu.
8475 * @cpu: the processor in question.
8477 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8479 struct task_struct *curr_task(int cpu)
8481 return cpu_curr(cpu);
8484 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8488 * set_curr_task - set the current task for a given cpu.
8489 * @cpu: the processor in question.
8490 * @p: the task pointer to set.
8492 * Description: This function must only be used when non-maskable interrupts
8493 * are serviced on a separate stack. It allows the architecture to switch the
8494 * notion of the current task on a cpu in a non-blocking manner. This function
8495 * must be called with all CPU's synchronized, and interrupts disabled, the
8496 * and caller must save the original value of the current task (see
8497 * curr_task() above) and restore that value before reenabling interrupts and
8498 * re-starting the system.
8500 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8502 void set_curr_task(int cpu, struct task_struct *p)
8509 #ifdef CONFIG_FAIR_GROUP_SCHED
8510 static void free_fair_sched_group(struct task_group *tg)
8514 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8516 for_each_possible_cpu(i) {
8518 kfree(tg->cfs_rq[i]);
8528 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8530 struct cfs_rq *cfs_rq;
8531 struct sched_entity *se;
8534 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8537 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8541 tg->shares = NICE_0_LOAD;
8543 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8545 for_each_possible_cpu(i) {
8546 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8547 GFP_KERNEL, cpu_to_node(i));
8551 se = kzalloc_node(sizeof(struct sched_entity),
8552 GFP_KERNEL, cpu_to_node(i));
8556 init_cfs_rq(cfs_rq);
8557 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8568 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8570 struct rq *rq = cpu_rq(cpu);
8571 unsigned long flags;
8574 * Only empty task groups can be destroyed; so we can speculatively
8575 * check on_list without danger of it being re-added.
8577 if (!tg->cfs_rq[cpu]->on_list)
8580 raw_spin_lock_irqsave(&rq->lock, flags);
8581 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8582 raw_spin_unlock_irqrestore(&rq->lock, flags);
8584 #else /* !CONFIG_FAIR_GROUP_SCHED */
8585 static inline void free_fair_sched_group(struct task_group *tg)
8590 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8595 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8598 #endif /* CONFIG_FAIR_GROUP_SCHED */
8600 #ifdef CONFIG_RT_GROUP_SCHED
8601 static void free_rt_sched_group(struct task_group *tg)
8606 destroy_rt_bandwidth(&tg->rt_bandwidth);
8608 for_each_possible_cpu(i) {
8610 kfree(tg->rt_rq[i]);
8612 kfree(tg->rt_se[i]);
8620 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8622 struct rt_rq *rt_rq;
8623 struct sched_rt_entity *rt_se;
8626 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8629 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8633 init_rt_bandwidth(&tg->rt_bandwidth,
8634 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8636 for_each_possible_cpu(i) {
8637 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8638 GFP_KERNEL, cpu_to_node(i));
8642 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8643 GFP_KERNEL, cpu_to_node(i));
8647 init_rt_rq(rt_rq, cpu_rq(i));
8648 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8649 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8659 #else /* !CONFIG_RT_GROUP_SCHED */
8660 static inline void free_rt_sched_group(struct task_group *tg)
8665 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8669 #endif /* CONFIG_RT_GROUP_SCHED */
8671 #ifdef CONFIG_CGROUP_SCHED
8672 static void free_sched_group(struct task_group *tg)
8674 free_fair_sched_group(tg);
8675 free_rt_sched_group(tg);
8680 /* allocate runqueue etc for a new task group */
8681 struct task_group *sched_create_group(struct task_group *parent)
8683 struct task_group *tg;
8684 unsigned long flags;
8686 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8688 return ERR_PTR(-ENOMEM);
8690 if (!alloc_fair_sched_group(tg, parent))
8693 if (!alloc_rt_sched_group(tg, parent))
8696 spin_lock_irqsave(&task_group_lock, flags);
8697 list_add_rcu(&tg->list, &task_groups);
8699 WARN_ON(!parent); /* root should already exist */
8701 tg->parent = parent;
8702 INIT_LIST_HEAD(&tg->children);
8703 list_add_rcu(&tg->siblings, &parent->children);
8704 spin_unlock_irqrestore(&task_group_lock, flags);
8709 free_sched_group(tg);
8710 return ERR_PTR(-ENOMEM);
8713 /* rcu callback to free various structures associated with a task group */
8714 static void free_sched_group_rcu(struct rcu_head *rhp)
8716 /* now it should be safe to free those cfs_rqs */
8717 free_sched_group(container_of(rhp, struct task_group, rcu));
8720 /* Destroy runqueue etc associated with a task group */
8721 void sched_destroy_group(struct task_group *tg)
8723 unsigned long flags;
8726 /* end participation in shares distribution */
8727 for_each_possible_cpu(i)
8728 unregister_fair_sched_group(tg, i);
8730 spin_lock_irqsave(&task_group_lock, flags);
8731 list_del_rcu(&tg->list);
8732 list_del_rcu(&tg->siblings);
8733 spin_unlock_irqrestore(&task_group_lock, flags);
8735 /* wait for possible concurrent references to cfs_rqs complete */
8736 call_rcu(&tg->rcu, free_sched_group_rcu);
8739 /* change task's runqueue when it moves between groups.
8740 * The caller of this function should have put the task in its new group
8741 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8742 * reflect its new group.
8744 void sched_move_task(struct task_struct *tsk)
8747 unsigned long flags;
8750 rq = task_rq_lock(tsk, &flags);
8752 running = task_current(rq, tsk);
8756 dequeue_task(rq, tsk, 0);
8757 if (unlikely(running))
8758 tsk->sched_class->put_prev_task(rq, tsk);
8760 #ifdef CONFIG_FAIR_GROUP_SCHED
8761 if (tsk->sched_class->task_move_group)
8762 tsk->sched_class->task_move_group(tsk, on_rq);
8765 set_task_rq(tsk, task_cpu(tsk));
8767 if (unlikely(running))
8768 tsk->sched_class->set_curr_task(rq);
8770 enqueue_task(rq, tsk, 0);
8772 task_rq_unlock(rq, tsk, &flags);
8774 #endif /* CONFIG_CGROUP_SCHED */
8776 #ifdef CONFIG_FAIR_GROUP_SCHED
8777 static DEFINE_MUTEX(shares_mutex);
8779 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8782 unsigned long flags;
8785 * We can't change the weight of the root cgroup.
8790 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8792 mutex_lock(&shares_mutex);
8793 if (tg->shares == shares)
8796 tg->shares = shares;
8797 for_each_possible_cpu(i) {
8798 struct rq *rq = cpu_rq(i);
8799 struct sched_entity *se;
8802 /* Propagate contribution to hierarchy */
8803 raw_spin_lock_irqsave(&rq->lock, flags);
8804 for_each_sched_entity(se)
8805 update_cfs_shares(group_cfs_rq(se));
8806 raw_spin_unlock_irqrestore(&rq->lock, flags);
8810 mutex_unlock(&shares_mutex);
8814 unsigned long sched_group_shares(struct task_group *tg)
8820 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
8821 static unsigned long to_ratio(u64 period, u64 runtime)
8823 if (runtime == RUNTIME_INF)
8826 return div64_u64(runtime << 20, period);
8830 #ifdef CONFIG_RT_GROUP_SCHED
8832 * Ensure that the real time constraints are schedulable.
8834 static DEFINE_MUTEX(rt_constraints_mutex);
8836 /* Must be called with tasklist_lock held */
8837 static inline int tg_has_rt_tasks(struct task_group *tg)
8839 struct task_struct *g, *p;
8841 do_each_thread(g, p) {
8842 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8844 } while_each_thread(g, p);
8849 struct rt_schedulable_data {
8850 struct task_group *tg;
8855 static int tg_rt_schedulable(struct task_group *tg, void *data)
8857 struct rt_schedulable_data *d = data;
8858 struct task_group *child;
8859 unsigned long total, sum = 0;
8860 u64 period, runtime;
8862 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8863 runtime = tg->rt_bandwidth.rt_runtime;
8866 period = d->rt_period;
8867 runtime = d->rt_runtime;
8871 * Cannot have more runtime than the period.
8873 if (runtime > period && runtime != RUNTIME_INF)
8877 * Ensure we don't starve existing RT tasks.
8879 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8882 total = to_ratio(period, runtime);
8885 * Nobody can have more than the global setting allows.
8887 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8891 * The sum of our children's runtime should not exceed our own.
8893 list_for_each_entry_rcu(child, &tg->children, siblings) {
8894 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8895 runtime = child->rt_bandwidth.rt_runtime;
8897 if (child == d->tg) {
8898 period = d->rt_period;
8899 runtime = d->rt_runtime;
8902 sum += to_ratio(period, runtime);
8911 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8915 struct rt_schedulable_data data = {
8917 .rt_period = period,
8918 .rt_runtime = runtime,
8922 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8928 static int tg_set_rt_bandwidth(struct task_group *tg,
8929 u64 rt_period, u64 rt_runtime)
8933 mutex_lock(&rt_constraints_mutex);
8934 read_lock(&tasklist_lock);
8935 err = __rt_schedulable(tg, rt_period, rt_runtime);
8939 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8940 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8941 tg->rt_bandwidth.rt_runtime = rt_runtime;
8943 for_each_possible_cpu(i) {
8944 struct rt_rq *rt_rq = tg->rt_rq[i];
8946 raw_spin_lock(&rt_rq->rt_runtime_lock);
8947 rt_rq->rt_runtime = rt_runtime;
8948 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8950 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8952 read_unlock(&tasklist_lock);
8953 mutex_unlock(&rt_constraints_mutex);
8958 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8960 u64 rt_runtime, rt_period;
8962 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8963 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8964 if (rt_runtime_us < 0)
8965 rt_runtime = RUNTIME_INF;
8967 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8970 long sched_group_rt_runtime(struct task_group *tg)
8974 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8977 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8978 do_div(rt_runtime_us, NSEC_PER_USEC);
8979 return rt_runtime_us;
8982 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8984 u64 rt_runtime, rt_period;
8986 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8987 rt_runtime = tg->rt_bandwidth.rt_runtime;
8992 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8995 long sched_group_rt_period(struct task_group *tg)
8999 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9000 do_div(rt_period_us, NSEC_PER_USEC);
9001 return rt_period_us;
9004 static int sched_rt_global_constraints(void)
9006 u64 runtime, period;
9009 if (sysctl_sched_rt_period <= 0)
9012 runtime = global_rt_runtime();
9013 period = global_rt_period();
9016 * Sanity check on the sysctl variables.
9018 if (runtime > period && runtime != RUNTIME_INF)
9021 mutex_lock(&rt_constraints_mutex);
9022 read_lock(&tasklist_lock);
9023 ret = __rt_schedulable(NULL, 0, 0);
9024 read_unlock(&tasklist_lock);
9025 mutex_unlock(&rt_constraints_mutex);
9030 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9032 /* Don't accept realtime tasks when there is no way for them to run */
9033 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9039 #else /* !CONFIG_RT_GROUP_SCHED */
9040 static int sched_rt_global_constraints(void)
9042 unsigned long flags;
9045 if (sysctl_sched_rt_period <= 0)
9049 * There's always some RT tasks in the root group
9050 * -- migration, kstopmachine etc..
9052 if (sysctl_sched_rt_runtime == 0)
9055 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9056 for_each_possible_cpu(i) {
9057 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9059 raw_spin_lock(&rt_rq->rt_runtime_lock);
9060 rt_rq->rt_runtime = global_rt_runtime();
9061 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9063 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9067 #endif /* CONFIG_RT_GROUP_SCHED */
9069 int sched_rt_handler(struct ctl_table *table, int write,
9070 void __user *buffer, size_t *lenp,
9074 int old_period, old_runtime;
9075 static DEFINE_MUTEX(mutex);
9078 old_period = sysctl_sched_rt_period;
9079 old_runtime = sysctl_sched_rt_runtime;
9081 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9083 if (!ret && write) {
9084 ret = sched_rt_global_constraints();
9086 sysctl_sched_rt_period = old_period;
9087 sysctl_sched_rt_runtime = old_runtime;
9089 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9090 def_rt_bandwidth.rt_period =
9091 ns_to_ktime(global_rt_period());
9094 mutex_unlock(&mutex);
9099 #ifdef CONFIG_CGROUP_SCHED
9101 /* return corresponding task_group object of a cgroup */
9102 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9104 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9105 struct task_group, css);
9108 static struct cgroup_subsys_state *
9109 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9111 struct task_group *tg, *parent;
9113 if (!cgrp->parent) {
9114 /* This is early initialization for the top cgroup */
9115 return &root_task_group.css;
9118 parent = cgroup_tg(cgrp->parent);
9119 tg = sched_create_group(parent);
9121 return ERR_PTR(-ENOMEM);
9127 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9129 struct task_group *tg = cgroup_tg(cgrp);
9131 sched_destroy_group(tg);
9135 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9137 #ifdef CONFIG_RT_GROUP_SCHED
9138 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9141 /* We don't support RT-tasks being in separate groups */
9142 if (tsk->sched_class != &fair_sched_class)
9149 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9151 sched_move_task(tsk);
9155 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9156 struct cgroup *old_cgrp, struct task_struct *task)
9159 * cgroup_exit() is called in the copy_process() failure path.
9160 * Ignore this case since the task hasn't ran yet, this avoids
9161 * trying to poke a half freed task state from generic code.
9163 if (!(task->flags & PF_EXITING))
9166 sched_move_task(task);
9169 #ifdef CONFIG_FAIR_GROUP_SCHED
9170 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9173 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9176 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9178 struct task_group *tg = cgroup_tg(cgrp);
9180 return (u64) scale_load_down(tg->shares);
9183 #ifdef CONFIG_CFS_BANDWIDTH
9184 static DEFINE_MUTEX(cfs_constraints_mutex);
9186 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9187 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9189 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9191 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9193 int i, ret = 0, runtime_enabled;
9194 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9196 if (tg == &root_task_group)
9200 * Ensure we have at some amount of bandwidth every period. This is
9201 * to prevent reaching a state of large arrears when throttled via
9202 * entity_tick() resulting in prolonged exit starvation.
9204 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9208 * Likewise, bound things on the otherside by preventing insane quota
9209 * periods. This also allows us to normalize in computing quota
9212 if (period > max_cfs_quota_period)
9215 mutex_lock(&cfs_constraints_mutex);
9216 ret = __cfs_schedulable(tg, period, quota);
9220 runtime_enabled = quota != RUNTIME_INF;
9221 raw_spin_lock_irq(&cfs_b->lock);
9222 cfs_b->period = ns_to_ktime(period);
9223 cfs_b->quota = quota;
9225 __refill_cfs_bandwidth_runtime(cfs_b);
9226 /* restart the period timer (if active) to handle new period expiry */
9227 if (runtime_enabled && cfs_b->timer_active) {
9228 /* force a reprogram */
9229 cfs_b->timer_active = 0;
9230 __start_cfs_bandwidth(cfs_b);
9232 raw_spin_unlock_irq(&cfs_b->lock);
9234 for_each_possible_cpu(i) {
9235 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9236 struct rq *rq = rq_of(cfs_rq);
9238 raw_spin_lock_irq(&rq->lock);
9239 cfs_rq->runtime_enabled = runtime_enabled;
9240 cfs_rq->runtime_remaining = 0;
9242 if (cfs_rq_throttled(cfs_rq))
9243 unthrottle_cfs_rq(cfs_rq);
9244 raw_spin_unlock_irq(&rq->lock);
9247 mutex_unlock(&cfs_constraints_mutex);
9252 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9256 period = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9257 if (cfs_quota_us < 0)
9258 quota = RUNTIME_INF;
9260 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9262 return tg_set_cfs_bandwidth(tg, period, quota);
9265 long tg_get_cfs_quota(struct task_group *tg)
9269 if (tg_cfs_bandwidth(tg)->quota == RUNTIME_INF)
9272 quota_us = tg_cfs_bandwidth(tg)->quota;
9273 do_div(quota_us, NSEC_PER_USEC);
9278 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9282 period = (u64)cfs_period_us * NSEC_PER_USEC;
9283 quota = tg_cfs_bandwidth(tg)->quota;
9288 return tg_set_cfs_bandwidth(tg, period, quota);
9291 long tg_get_cfs_period(struct task_group *tg)
9295 cfs_period_us = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9296 do_div(cfs_period_us, NSEC_PER_USEC);
9298 return cfs_period_us;
9301 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
9303 return tg_get_cfs_quota(cgroup_tg(cgrp));
9306 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
9309 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
9312 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
9314 return tg_get_cfs_period(cgroup_tg(cgrp));
9317 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9320 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
9323 struct cfs_schedulable_data {
9324 struct task_group *tg;
9329 * normalize group quota/period to be quota/max_period
9330 * note: units are usecs
9332 static u64 normalize_cfs_quota(struct task_group *tg,
9333 struct cfs_schedulable_data *d)
9341 period = tg_get_cfs_period(tg);
9342 quota = tg_get_cfs_quota(tg);
9345 /* note: these should typically be equivalent */
9346 if (quota == RUNTIME_INF || quota == -1)
9349 return to_ratio(period, quota);
9352 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9354 struct cfs_schedulable_data *d = data;
9355 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9356 s64 quota = 0, parent_quota = -1;
9359 quota = RUNTIME_INF;
9361 struct cfs_bandwidth *parent_b = tg_cfs_bandwidth(tg->parent);
9363 quota = normalize_cfs_quota(tg, d);
9364 parent_quota = parent_b->hierarchal_quota;
9367 * ensure max(child_quota) <= parent_quota, inherit when no
9370 if (quota == RUNTIME_INF)
9371 quota = parent_quota;
9372 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9375 cfs_b->hierarchal_quota = quota;
9380 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9383 struct cfs_schedulable_data data = {
9389 if (quota != RUNTIME_INF) {
9390 do_div(data.period, NSEC_PER_USEC);
9391 do_div(data.quota, NSEC_PER_USEC);
9395 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9400 #endif /* CONFIG_CFS_BANDWIDTH */
9401 #endif /* CONFIG_FAIR_GROUP_SCHED */
9403 #ifdef CONFIG_RT_GROUP_SCHED
9404 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9407 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9410 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9412 return sched_group_rt_runtime(cgroup_tg(cgrp));
9415 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9418 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9421 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9423 return sched_group_rt_period(cgroup_tg(cgrp));
9425 #endif /* CONFIG_RT_GROUP_SCHED */
9427 static struct cftype cpu_files[] = {
9428 #ifdef CONFIG_FAIR_GROUP_SCHED
9431 .read_u64 = cpu_shares_read_u64,
9432 .write_u64 = cpu_shares_write_u64,
9435 #ifdef CONFIG_CFS_BANDWIDTH
9437 .name = "cfs_quota_us",
9438 .read_s64 = cpu_cfs_quota_read_s64,
9439 .write_s64 = cpu_cfs_quota_write_s64,
9442 .name = "cfs_period_us",
9443 .read_u64 = cpu_cfs_period_read_u64,
9444 .write_u64 = cpu_cfs_period_write_u64,
9447 #ifdef CONFIG_RT_GROUP_SCHED
9449 .name = "rt_runtime_us",
9450 .read_s64 = cpu_rt_runtime_read,
9451 .write_s64 = cpu_rt_runtime_write,
9454 .name = "rt_period_us",
9455 .read_u64 = cpu_rt_period_read_uint,
9456 .write_u64 = cpu_rt_period_write_uint,
9461 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9463 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9466 struct cgroup_subsys cpu_cgroup_subsys = {
9468 .create = cpu_cgroup_create,
9469 .destroy = cpu_cgroup_destroy,
9470 .can_attach_task = cpu_cgroup_can_attach_task,
9471 .attach_task = cpu_cgroup_attach_task,
9472 .exit = cpu_cgroup_exit,
9473 .populate = cpu_cgroup_populate,
9474 .subsys_id = cpu_cgroup_subsys_id,
9478 #endif /* CONFIG_CGROUP_SCHED */
9480 #ifdef CONFIG_CGROUP_CPUACCT
9483 * CPU accounting code for task groups.
9485 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9486 * (balbir@in.ibm.com).
9489 /* track cpu usage of a group of tasks and its child groups */
9491 struct cgroup_subsys_state css;
9492 /* cpuusage holds pointer to a u64-type object on every cpu */
9493 u64 __percpu *cpuusage;
9494 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9495 struct cpuacct *parent;
9498 struct cgroup_subsys cpuacct_subsys;
9500 /* return cpu accounting group corresponding to this container */
9501 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9503 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9504 struct cpuacct, css);
9507 /* return cpu accounting group to which this task belongs */
9508 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9510 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9511 struct cpuacct, css);
9514 /* create a new cpu accounting group */
9515 static struct cgroup_subsys_state *cpuacct_create(
9516 struct cgroup_subsys *ss, struct cgroup *cgrp)
9518 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9524 ca->cpuusage = alloc_percpu(u64);
9528 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9529 if (percpu_counter_init(&ca->cpustat[i], 0))
9530 goto out_free_counters;
9533 ca->parent = cgroup_ca(cgrp->parent);
9539 percpu_counter_destroy(&ca->cpustat[i]);
9540 free_percpu(ca->cpuusage);
9544 return ERR_PTR(-ENOMEM);
9547 /* destroy an existing cpu accounting group */
9549 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9551 struct cpuacct *ca = cgroup_ca(cgrp);
9554 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9555 percpu_counter_destroy(&ca->cpustat[i]);
9556 free_percpu(ca->cpuusage);
9560 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9562 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9565 #ifndef CONFIG_64BIT
9567 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9569 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9571 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9579 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9581 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9583 #ifndef CONFIG_64BIT
9585 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9587 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9589 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9595 /* return total cpu usage (in nanoseconds) of a group */
9596 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9598 struct cpuacct *ca = cgroup_ca(cgrp);
9599 u64 totalcpuusage = 0;
9602 for_each_present_cpu(i)
9603 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9605 return totalcpuusage;
9608 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9611 struct cpuacct *ca = cgroup_ca(cgrp);
9620 for_each_present_cpu(i)
9621 cpuacct_cpuusage_write(ca, i, 0);
9627 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9630 struct cpuacct *ca = cgroup_ca(cgroup);
9634 for_each_present_cpu(i) {
9635 percpu = cpuacct_cpuusage_read(ca, i);
9636 seq_printf(m, "%llu ", (unsigned long long) percpu);
9638 seq_printf(m, "\n");
9642 static const char *cpuacct_stat_desc[] = {
9643 [CPUACCT_STAT_USER] = "user",
9644 [CPUACCT_STAT_SYSTEM] = "system",
9647 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9648 struct cgroup_map_cb *cb)
9650 struct cpuacct *ca = cgroup_ca(cgrp);
9653 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9654 s64 val = percpu_counter_read(&ca->cpustat[i]);
9655 val = cputime64_to_clock_t(val);
9656 cb->fill(cb, cpuacct_stat_desc[i], val);
9661 static struct cftype files[] = {
9664 .read_u64 = cpuusage_read,
9665 .write_u64 = cpuusage_write,
9668 .name = "usage_percpu",
9669 .read_seq_string = cpuacct_percpu_seq_read,
9673 .read_map = cpuacct_stats_show,
9677 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9679 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9683 * charge this task's execution time to its accounting group.
9685 * called with rq->lock held.
9687 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9692 if (unlikely(!cpuacct_subsys.active))
9695 cpu = task_cpu(tsk);
9701 for (; ca; ca = ca->parent) {
9702 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9703 *cpuusage += cputime;
9710 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9711 * in cputime_t units. As a result, cpuacct_update_stats calls
9712 * percpu_counter_add with values large enough to always overflow the
9713 * per cpu batch limit causing bad SMP scalability.
9715 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9716 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9717 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9720 #define CPUACCT_BATCH \
9721 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9723 #define CPUACCT_BATCH 0
9727 * Charge the system/user time to the task's accounting group.
9729 static void cpuacct_update_stats(struct task_struct *tsk,
9730 enum cpuacct_stat_index idx, cputime_t val)
9733 int batch = CPUACCT_BATCH;
9735 if (unlikely(!cpuacct_subsys.active))
9742 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9748 struct cgroup_subsys cpuacct_subsys = {
9750 .create = cpuacct_create,
9751 .destroy = cpuacct_destroy,
9752 .populate = cpuacct_populate,
9753 .subsys_id = cpuacct_subsys_id,
9755 #endif /* CONFIG_CGROUP_CPUACCT */