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 <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy)
126 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
131 static inline int task_has_rt_policy(struct task_struct *p)
133 return rt_policy(p->policy);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array {
140 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
141 struct list_head queue[MAX_RT_PRIO];
144 struct rt_bandwidth {
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock;
149 struct hrtimer rt_period_timer;
152 static struct rt_bandwidth def_rt_bandwidth;
154 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
156 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
158 struct rt_bandwidth *rt_b =
159 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 now = hrtimer_cb_get_time(timer);
166 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171 idle = do_sched_rt_period_timer(rt_b, overrun);
174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
180 rt_b->rt_period = ns_to_ktime(period);
181 rt_b->rt_runtime = runtime;
183 raw_spin_lock_init(&rt_b->rt_runtime_lock);
185 hrtimer_init(&rt_b->rt_period_timer,
186 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
187 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime >= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
199 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 if (hrtimer_active(&rt_b->rt_period_timer))
205 raw_spin_lock(&rt_b->rt_runtime_lock);
210 if (hrtimer_active(&rt_b->rt_period_timer))
213 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
214 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
216 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
217 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
218 delta = ktime_to_ns(ktime_sub(hard, soft));
219 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
220 HRTIMER_MODE_ABS_PINNED, 0);
222 raw_spin_unlock(&rt_b->rt_runtime_lock);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
228 hrtimer_cancel(&rt_b->rt_period_timer);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex);
238 #ifdef CONFIG_CGROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups);
246 /* task group related information */
248 struct cgroup_subsys_state css;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 /* schedulable entities of this group on each cpu */
252 struct sched_entity **se;
253 /* runqueue "owned" by this group on each cpu */
254 struct cfs_rq **cfs_rq;
255 unsigned long shares;
258 #ifdef CONFIG_RT_GROUP_SCHED
259 struct sched_rt_entity **rt_se;
260 struct rt_rq **rt_rq;
262 struct rt_bandwidth rt_bandwidth;
266 struct list_head list;
268 struct task_group *parent;
269 struct list_head siblings;
270 struct list_head children;
273 #define root_task_group init_task_group
275 /* task_group_lock serializes add/remove of task groups and also changes to
276 * a task group's cpu shares.
278 static DEFINE_SPINLOCK(task_group_lock);
280 #ifdef CONFIG_FAIR_GROUP_SCHED
283 static int root_task_group_empty(void)
285 return list_empty(&root_task_group.children);
289 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
292 * A weight of 0 or 1 can cause arithmetics problems.
293 * A weight of a cfs_rq is the sum of weights of which entities
294 * are queued on this cfs_rq, so a weight of a entity should not be
295 * too large, so as the shares value of a task group.
296 * (The default weight is 1024 - so there's no practical
297 * limitation from this.)
300 #define MAX_SHARES (1UL << 18)
302 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
305 /* Default task group.
306 * Every task in system belong to this group at bootup.
308 struct task_group init_task_group;
310 #endif /* CONFIG_CGROUP_SCHED */
312 /* CFS-related fields in a runqueue */
314 struct load_weight load;
315 unsigned long nr_running;
320 struct rb_root tasks_timeline;
321 struct rb_node *rb_leftmost;
323 struct list_head tasks;
324 struct list_head *balance_iterator;
327 * 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity *curr, *next, *last;
332 unsigned int nr_spread_over;
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
338 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
339 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
340 * (like users, containers etc.)
342 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
343 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list;
346 struct task_group *tg; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load;
363 * this cpu's part of tg->shares
365 unsigned long shares;
368 * load.weight at the time we set shares
370 unsigned long rq_weight;
375 /* Real-Time classes' related field in a runqueue: */
377 struct rt_prio_array active;
378 unsigned long rt_nr_running;
379 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 int curr; /* highest queued rt task prio */
383 int next; /* next highest */
388 unsigned long rt_nr_migratory;
389 unsigned long rt_nr_total;
391 struct plist_head pushable_tasks;
396 /* Nests inside the rq lock: */
397 raw_spinlock_t rt_runtime_lock;
399 #ifdef CONFIG_RT_GROUP_SCHED
400 unsigned long rt_nr_boosted;
403 struct list_head leaf_rt_rq_list;
404 struct task_group *tg;
411 * We add the notion of a root-domain which will be used to define per-domain
412 * variables. Each exclusive cpuset essentially defines an island domain by
413 * fully partitioning the member cpus from any other cpuset. Whenever a new
414 * exclusive cpuset is created, we also create and attach a new root-domain
421 cpumask_var_t online;
424 * The "RT overload" flag: it gets set if a CPU has more than
425 * one runnable RT task.
427 cpumask_var_t rto_mask;
430 struct cpupri cpupri;
435 * By default the system creates a single root-domain with all cpus as
436 * members (mimicking the global state we have today).
438 static struct root_domain def_root_domain;
443 * This is the main, per-CPU runqueue data structure.
445 * Locking rule: those places that want to lock multiple runqueues
446 * (such as the load balancing or the thread migration code), lock
447 * acquire operations must be ordered by ascending &runqueue.
454 * nr_running and cpu_load should be in the same cacheline because
455 * remote CPUs use both these fields when doing load calculation.
457 unsigned long nr_running;
458 #define CPU_LOAD_IDX_MAX 5
459 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
462 unsigned char in_nohz_recently;
464 unsigned int skip_clock_update;
466 /* capture load from *all* tasks on this cpu: */
467 struct load_weight load;
468 unsigned long nr_load_updates;
474 #ifdef CONFIG_FAIR_GROUP_SCHED
475 /* list of leaf cfs_rq on this cpu: */
476 struct list_head leaf_cfs_rq_list;
478 #ifdef CONFIG_RT_GROUP_SCHED
479 struct list_head leaf_rt_rq_list;
483 * This is part of a global counter where only the total sum
484 * over all CPUs matters. A task can increase this counter on
485 * one CPU and if it got migrated afterwards it may decrease
486 * it on another CPU. Always updated under the runqueue lock:
488 unsigned long nr_uninterruptible;
490 struct task_struct *curr, *idle;
491 unsigned long next_balance;
492 struct mm_struct *prev_mm;
499 struct root_domain *rd;
500 struct sched_domain *sd;
502 unsigned long cpu_power;
504 unsigned char idle_at_tick;
505 /* For active balancing */
509 struct cpu_stop_work active_balance_work;
510 /* cpu of this runqueue: */
514 unsigned long avg_load_per_task;
522 /* calc_load related fields */
523 unsigned long calc_load_update;
524 long calc_load_active;
526 #ifdef CONFIG_SCHED_HRTICK
528 int hrtick_csd_pending;
529 struct call_single_data hrtick_csd;
531 struct hrtimer hrtick_timer;
534 #ifdef CONFIG_SCHEDSTATS
536 struct sched_info rq_sched_info;
537 unsigned long long rq_cpu_time;
538 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
540 /* sys_sched_yield() stats */
541 unsigned int yld_count;
543 /* schedule() stats */
544 unsigned int sched_switch;
545 unsigned int sched_count;
546 unsigned int sched_goidle;
548 /* try_to_wake_up() stats */
549 unsigned int ttwu_count;
550 unsigned int ttwu_local;
553 unsigned int bkl_count;
557 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
560 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
562 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
565 * A queue event has occurred, and we're going to schedule. In
566 * this case, we can save a useless back to back clock update.
568 if (test_tsk_need_resched(p))
569 rq->skip_clock_update = 1;
572 static inline int cpu_of(struct rq *rq)
581 #define rcu_dereference_check_sched_domain(p) \
582 rcu_dereference_check((p), \
583 rcu_read_lock_sched_held() || \
584 lockdep_is_held(&sched_domains_mutex))
587 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
588 * See detach_destroy_domains: synchronize_sched for details.
590 * The domain tree of any CPU may only be accessed from within
591 * preempt-disabled sections.
593 #define for_each_domain(cpu, __sd) \
594 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
596 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
597 #define this_rq() (&__get_cpu_var(runqueues))
598 #define task_rq(p) cpu_rq(task_cpu(p))
599 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
600 #define raw_rq() (&__raw_get_cpu_var(runqueues))
602 #ifdef CONFIG_CGROUP_SCHED
605 * Return the group to which this tasks belongs.
607 * We use task_subsys_state_check() and extend the RCU verification
608 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
609 * holds that lock for each task it moves into the cgroup. Therefore
610 * by holding that lock, we pin the task to the current cgroup.
612 static inline struct task_group *task_group(struct task_struct *p)
614 struct cgroup_subsys_state *css;
616 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
617 lockdep_is_held(&task_rq(p)->lock));
618 return container_of(css, struct task_group, css);
621 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
622 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
624 #ifdef CONFIG_FAIR_GROUP_SCHED
625 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
626 p->se.parent = task_group(p)->se[cpu];
629 #ifdef CONFIG_RT_GROUP_SCHED
630 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
631 p->rt.parent = task_group(p)->rt_se[cpu];
635 #else /* CONFIG_CGROUP_SCHED */
637 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
638 static inline struct task_group *task_group(struct task_struct *p)
643 #endif /* CONFIG_CGROUP_SCHED */
645 inline void update_rq_clock(struct rq *rq)
647 if (!rq->skip_clock_update)
648 rq->clock = sched_clock_cpu(cpu_of(rq));
652 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
654 #ifdef CONFIG_SCHED_DEBUG
655 # define const_debug __read_mostly
657 # define const_debug static const
662 * @cpu: the processor in question.
664 * Returns true if the current cpu runqueue is locked.
665 * This interface allows printk to be called with the runqueue lock
666 * held and know whether or not it is OK to wake up the klogd.
668 int runqueue_is_locked(int cpu)
670 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
674 * Debugging: various feature bits
677 #define SCHED_FEAT(name, enabled) \
678 __SCHED_FEAT_##name ,
681 #include "sched_features.h"
686 #define SCHED_FEAT(name, enabled) \
687 (1UL << __SCHED_FEAT_##name) * enabled |
689 const_debug unsigned int sysctl_sched_features =
690 #include "sched_features.h"
695 #ifdef CONFIG_SCHED_DEBUG
696 #define SCHED_FEAT(name, enabled) \
699 static __read_mostly char *sched_feat_names[] = {
700 #include "sched_features.h"
706 static int sched_feat_show(struct seq_file *m, void *v)
710 for (i = 0; sched_feat_names[i]; i++) {
711 if (!(sysctl_sched_features & (1UL << i)))
713 seq_printf(m, "%s ", sched_feat_names[i]);
721 sched_feat_write(struct file *filp, const char __user *ubuf,
722 size_t cnt, loff_t *ppos)
732 if (copy_from_user(&buf, ubuf, cnt))
737 if (strncmp(buf, "NO_", 3) == 0) {
742 for (i = 0; sched_feat_names[i]; i++) {
743 int len = strlen(sched_feat_names[i]);
745 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
747 sysctl_sched_features &= ~(1UL << i);
749 sysctl_sched_features |= (1UL << i);
754 if (!sched_feat_names[i])
762 static int sched_feat_open(struct inode *inode, struct file *filp)
764 return single_open(filp, sched_feat_show, NULL);
767 static const struct file_operations sched_feat_fops = {
768 .open = sched_feat_open,
769 .write = sched_feat_write,
772 .release = single_release,
775 static __init int sched_init_debug(void)
777 debugfs_create_file("sched_features", 0644, NULL, NULL,
782 late_initcall(sched_init_debug);
786 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
789 * Number of tasks to iterate in a single balance run.
790 * Limited because this is done with IRQs disabled.
792 const_debug unsigned int sysctl_sched_nr_migrate = 32;
795 * ratelimit for updating the group shares.
798 unsigned int sysctl_sched_shares_ratelimit = 250000;
799 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
802 * Inject some fuzzyness into changing the per-cpu group shares
803 * this avoids remote rq-locks at the expense of fairness.
806 unsigned int sysctl_sched_shares_thresh = 4;
809 * period over which we average the RT time consumption, measured
814 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
817 * period over which we measure -rt task cpu usage in us.
820 unsigned int sysctl_sched_rt_period = 1000000;
822 static __read_mostly int scheduler_running;
825 * part of the period that we allow rt tasks to run in us.
828 int sysctl_sched_rt_runtime = 950000;
830 static inline u64 global_rt_period(void)
832 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
835 static inline u64 global_rt_runtime(void)
837 if (sysctl_sched_rt_runtime < 0)
840 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
843 #ifndef prepare_arch_switch
844 # define prepare_arch_switch(next) do { } while (0)
846 #ifndef finish_arch_switch
847 # define finish_arch_switch(prev) do { } while (0)
850 static inline int task_current(struct rq *rq, struct task_struct *p)
852 return rq->curr == p;
855 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
856 static inline int task_running(struct rq *rq, struct task_struct *p)
858 return task_current(rq, p);
861 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
865 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
867 #ifdef CONFIG_DEBUG_SPINLOCK
868 /* this is a valid case when another task releases the spinlock */
869 rq->lock.owner = current;
872 * If we are tracking spinlock dependencies then we have to
873 * fix up the runqueue lock - which gets 'carried over' from
876 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
878 raw_spin_unlock_irq(&rq->lock);
881 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
882 static inline int task_running(struct rq *rq, struct task_struct *p)
887 return task_current(rq, p);
891 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
895 * We can optimise this out completely for !SMP, because the
896 * SMP rebalancing from interrupt is the only thing that cares
901 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
902 raw_spin_unlock_irq(&rq->lock);
904 raw_spin_unlock(&rq->lock);
908 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
912 * After ->oncpu is cleared, the task can be moved to a different CPU.
913 * We must ensure this doesn't happen until the switch is completely
919 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
923 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
926 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
929 static inline int task_is_waking(struct task_struct *p)
931 return unlikely(p->state == TASK_WAKING);
935 * __task_rq_lock - lock the runqueue a given task resides on.
936 * Must be called interrupts disabled.
938 static inline struct rq *__task_rq_lock(struct task_struct *p)
945 raw_spin_lock(&rq->lock);
946 if (likely(rq == task_rq(p)))
948 raw_spin_unlock(&rq->lock);
953 * task_rq_lock - lock the runqueue a given task resides on and disable
954 * interrupts. Note the ordering: we can safely lookup the task_rq without
955 * explicitly disabling preemption.
957 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
963 local_irq_save(*flags);
965 raw_spin_lock(&rq->lock);
966 if (likely(rq == task_rq(p)))
968 raw_spin_unlock_irqrestore(&rq->lock, *flags);
972 static void __task_rq_unlock(struct rq *rq)
975 raw_spin_unlock(&rq->lock);
978 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
981 raw_spin_unlock_irqrestore(&rq->lock, *flags);
985 * this_rq_lock - lock this runqueue and disable interrupts.
987 static struct rq *this_rq_lock(void)
994 raw_spin_lock(&rq->lock);
999 #ifdef CONFIG_SCHED_HRTICK
1001 * Use HR-timers to deliver accurate preemption points.
1003 * Its all a bit involved since we cannot program an hrt while holding the
1004 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1007 * When we get rescheduled we reprogram the hrtick_timer outside of the
1013 * - enabled by features
1014 * - hrtimer is actually high res
1016 static inline int hrtick_enabled(struct rq *rq)
1018 if (!sched_feat(HRTICK))
1020 if (!cpu_active(cpu_of(rq)))
1022 return hrtimer_is_hres_active(&rq->hrtick_timer);
1025 static void hrtick_clear(struct rq *rq)
1027 if (hrtimer_active(&rq->hrtick_timer))
1028 hrtimer_cancel(&rq->hrtick_timer);
1032 * High-resolution timer tick.
1033 * Runs from hardirq context with interrupts disabled.
1035 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1037 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1039 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1041 raw_spin_lock(&rq->lock);
1042 update_rq_clock(rq);
1043 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1044 raw_spin_unlock(&rq->lock);
1046 return HRTIMER_NORESTART;
1051 * called from hardirq (IPI) context
1053 static void __hrtick_start(void *arg)
1055 struct rq *rq = arg;
1057 raw_spin_lock(&rq->lock);
1058 hrtimer_restart(&rq->hrtick_timer);
1059 rq->hrtick_csd_pending = 0;
1060 raw_spin_unlock(&rq->lock);
1064 * Called to set the hrtick timer state.
1066 * called with rq->lock held and irqs disabled
1068 static void hrtick_start(struct rq *rq, u64 delay)
1070 struct hrtimer *timer = &rq->hrtick_timer;
1071 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1073 hrtimer_set_expires(timer, time);
1075 if (rq == this_rq()) {
1076 hrtimer_restart(timer);
1077 } else if (!rq->hrtick_csd_pending) {
1078 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1079 rq->hrtick_csd_pending = 1;
1084 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1086 int cpu = (int)(long)hcpu;
1089 case CPU_UP_CANCELED:
1090 case CPU_UP_CANCELED_FROZEN:
1091 case CPU_DOWN_PREPARE:
1092 case CPU_DOWN_PREPARE_FROZEN:
1094 case CPU_DEAD_FROZEN:
1095 hrtick_clear(cpu_rq(cpu));
1102 static __init void init_hrtick(void)
1104 hotcpu_notifier(hotplug_hrtick, 0);
1108 * Called to set the hrtick timer state.
1110 * called with rq->lock held and irqs disabled
1112 static void hrtick_start(struct rq *rq, u64 delay)
1114 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1115 HRTIMER_MODE_REL_PINNED, 0);
1118 static inline void init_hrtick(void)
1121 #endif /* CONFIG_SMP */
1123 static void init_rq_hrtick(struct rq *rq)
1126 rq->hrtick_csd_pending = 0;
1128 rq->hrtick_csd.flags = 0;
1129 rq->hrtick_csd.func = __hrtick_start;
1130 rq->hrtick_csd.info = rq;
1133 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1134 rq->hrtick_timer.function = hrtick;
1136 #else /* CONFIG_SCHED_HRTICK */
1137 static inline void hrtick_clear(struct rq *rq)
1141 static inline void init_rq_hrtick(struct rq *rq)
1145 static inline void init_hrtick(void)
1148 #endif /* CONFIG_SCHED_HRTICK */
1151 * resched_task - mark a task 'to be rescheduled now'.
1153 * On UP this means the setting of the need_resched flag, on SMP it
1154 * might also involve a cross-CPU call to trigger the scheduler on
1159 #ifndef tsk_is_polling
1160 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1163 static void resched_task(struct task_struct *p)
1167 assert_raw_spin_locked(&task_rq(p)->lock);
1169 if (test_tsk_need_resched(p))
1172 set_tsk_need_resched(p);
1175 if (cpu == smp_processor_id())
1178 /* NEED_RESCHED must be visible before we test polling */
1180 if (!tsk_is_polling(p))
1181 smp_send_reschedule(cpu);
1184 static void resched_cpu(int cpu)
1186 struct rq *rq = cpu_rq(cpu);
1187 unsigned long flags;
1189 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1191 resched_task(cpu_curr(cpu));
1192 raw_spin_unlock_irqrestore(&rq->lock, flags);
1197 * When add_timer_on() enqueues a timer into the timer wheel of an
1198 * idle CPU then this timer might expire before the next timer event
1199 * which is scheduled to wake up that CPU. In case of a completely
1200 * idle system the next event might even be infinite time into the
1201 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1202 * leaves the inner idle loop so the newly added timer is taken into
1203 * account when the CPU goes back to idle and evaluates the timer
1204 * wheel for the next timer event.
1206 void wake_up_idle_cpu(int cpu)
1208 struct rq *rq = cpu_rq(cpu);
1210 if (cpu == smp_processor_id())
1214 * This is safe, as this function is called with the timer
1215 * wheel base lock of (cpu) held. When the CPU is on the way
1216 * to idle and has not yet set rq->curr to idle then it will
1217 * be serialized on the timer wheel base lock and take the new
1218 * timer into account automatically.
1220 if (rq->curr != rq->idle)
1224 * We can set TIF_RESCHED on the idle task of the other CPU
1225 * lockless. The worst case is that the other CPU runs the
1226 * idle task through an additional NOOP schedule()
1228 set_tsk_need_resched(rq->idle);
1230 /* NEED_RESCHED must be visible before we test polling */
1232 if (!tsk_is_polling(rq->idle))
1233 smp_send_reschedule(cpu);
1236 int nohz_ratelimit(int cpu)
1238 struct rq *rq = cpu_rq(cpu);
1239 u64 diff = rq->clock - rq->nohz_stamp;
1241 rq->nohz_stamp = rq->clock;
1243 return diff < (NSEC_PER_SEC / HZ) >> 1;
1246 #endif /* CONFIG_NO_HZ */
1248 static u64 sched_avg_period(void)
1250 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1253 static void sched_avg_update(struct rq *rq)
1255 s64 period = sched_avg_period();
1257 while ((s64)(rq->clock - rq->age_stamp) > period) {
1258 rq->age_stamp += period;
1263 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1265 rq->rt_avg += rt_delta;
1266 sched_avg_update(rq);
1269 #else /* !CONFIG_SMP */
1270 static void resched_task(struct task_struct *p)
1272 assert_raw_spin_locked(&task_rq(p)->lock);
1273 set_tsk_need_resched(p);
1276 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1279 #endif /* CONFIG_SMP */
1281 #if BITS_PER_LONG == 32
1282 # define WMULT_CONST (~0UL)
1284 # define WMULT_CONST (1UL << 32)
1287 #define WMULT_SHIFT 32
1290 * Shift right and round:
1292 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1295 * delta *= weight / lw
1297 static unsigned long
1298 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1299 struct load_weight *lw)
1303 if (!lw->inv_weight) {
1304 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1307 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1311 tmp = (u64)delta_exec * weight;
1313 * Check whether we'd overflow the 64-bit multiplication:
1315 if (unlikely(tmp > WMULT_CONST))
1316 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1319 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1321 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1324 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1330 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1337 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1338 * of tasks with abnormal "nice" values across CPUs the contribution that
1339 * each task makes to its run queue's load is weighted according to its
1340 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1341 * scaled version of the new time slice allocation that they receive on time
1345 #define WEIGHT_IDLEPRIO 3
1346 #define WMULT_IDLEPRIO 1431655765
1349 * Nice levels are multiplicative, with a gentle 10% change for every
1350 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1351 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1352 * that remained on nice 0.
1354 * The "10% effect" is relative and cumulative: from _any_ nice level,
1355 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1356 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1357 * If a task goes up by ~10% and another task goes down by ~10% then
1358 * the relative distance between them is ~25%.)
1360 static const int prio_to_weight[40] = {
1361 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1362 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1363 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1364 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1365 /* 0 */ 1024, 820, 655, 526, 423,
1366 /* 5 */ 335, 272, 215, 172, 137,
1367 /* 10 */ 110, 87, 70, 56, 45,
1368 /* 15 */ 36, 29, 23, 18, 15,
1372 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1374 * In cases where the weight does not change often, we can use the
1375 * precalculated inverse to speed up arithmetics by turning divisions
1376 * into multiplications:
1378 static const u32 prio_to_wmult[40] = {
1379 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1380 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1381 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1382 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1383 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1384 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1385 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1386 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1389 /* Time spent by the tasks of the cpu accounting group executing in ... */
1390 enum cpuacct_stat_index {
1391 CPUACCT_STAT_USER, /* ... user mode */
1392 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1394 CPUACCT_STAT_NSTATS,
1397 #ifdef CONFIG_CGROUP_CPUACCT
1398 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1399 static void cpuacct_update_stats(struct task_struct *tsk,
1400 enum cpuacct_stat_index idx, cputime_t val);
1402 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1403 static inline void cpuacct_update_stats(struct task_struct *tsk,
1404 enum cpuacct_stat_index idx, cputime_t val) {}
1407 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1409 update_load_add(&rq->load, load);
1412 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1414 update_load_sub(&rq->load, load);
1417 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1418 typedef int (*tg_visitor)(struct task_group *, void *);
1421 * Iterate the full tree, calling @down when first entering a node and @up when
1422 * leaving it for the final time.
1424 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1426 struct task_group *parent, *child;
1430 parent = &root_task_group;
1432 ret = (*down)(parent, data);
1435 list_for_each_entry_rcu(child, &parent->children, siblings) {
1442 ret = (*up)(parent, data);
1447 parent = parent->parent;
1456 static int tg_nop(struct task_group *tg, void *data)
1463 /* Used instead of source_load when we know the type == 0 */
1464 static unsigned long weighted_cpuload(const int cpu)
1466 return cpu_rq(cpu)->load.weight;
1470 * Return a low guess at the load of a migration-source cpu weighted
1471 * according to the scheduling class and "nice" value.
1473 * We want to under-estimate the load of migration sources, to
1474 * balance conservatively.
1476 static unsigned long source_load(int cpu, int type)
1478 struct rq *rq = cpu_rq(cpu);
1479 unsigned long total = weighted_cpuload(cpu);
1481 if (type == 0 || !sched_feat(LB_BIAS))
1484 return min(rq->cpu_load[type-1], total);
1488 * Return a high guess at the load of a migration-target cpu weighted
1489 * according to the scheduling class and "nice" value.
1491 static unsigned long target_load(int cpu, int type)
1493 struct rq *rq = cpu_rq(cpu);
1494 unsigned long total = weighted_cpuload(cpu);
1496 if (type == 0 || !sched_feat(LB_BIAS))
1499 return max(rq->cpu_load[type-1], total);
1502 static unsigned long power_of(int cpu)
1504 return cpu_rq(cpu)->cpu_power;
1507 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1509 static unsigned long cpu_avg_load_per_task(int cpu)
1511 struct rq *rq = cpu_rq(cpu);
1512 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1515 rq->avg_load_per_task = rq->load.weight / nr_running;
1517 rq->avg_load_per_task = 0;
1519 return rq->avg_load_per_task;
1522 #ifdef CONFIG_FAIR_GROUP_SCHED
1524 static __read_mostly unsigned long __percpu *update_shares_data;
1526 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1529 * Calculate and set the cpu's group shares.
1531 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1532 unsigned long sd_shares,
1533 unsigned long sd_rq_weight,
1534 unsigned long *usd_rq_weight)
1536 unsigned long shares, rq_weight;
1539 rq_weight = usd_rq_weight[cpu];
1542 rq_weight = NICE_0_LOAD;
1546 * \Sum_j shares_j * rq_weight_i
1547 * shares_i = -----------------------------
1548 * \Sum_j rq_weight_j
1550 shares = (sd_shares * rq_weight) / sd_rq_weight;
1551 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1553 if (abs(shares - tg->se[cpu]->load.weight) >
1554 sysctl_sched_shares_thresh) {
1555 struct rq *rq = cpu_rq(cpu);
1556 unsigned long flags;
1558 raw_spin_lock_irqsave(&rq->lock, flags);
1559 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1560 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1561 __set_se_shares(tg->se[cpu], shares);
1562 raw_spin_unlock_irqrestore(&rq->lock, flags);
1567 * Re-compute the task group their per cpu shares over the given domain.
1568 * This needs to be done in a bottom-up fashion because the rq weight of a
1569 * parent group depends on the shares of its child groups.
1571 static int tg_shares_up(struct task_group *tg, void *data)
1573 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1574 unsigned long *usd_rq_weight;
1575 struct sched_domain *sd = data;
1576 unsigned long flags;
1582 local_irq_save(flags);
1583 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1585 for_each_cpu(i, sched_domain_span(sd)) {
1586 weight = tg->cfs_rq[i]->load.weight;
1587 usd_rq_weight[i] = weight;
1589 rq_weight += weight;
1591 * If there are currently no tasks on the cpu pretend there
1592 * is one of average load so that when a new task gets to
1593 * run here it will not get delayed by group starvation.
1596 weight = NICE_0_LOAD;
1598 sum_weight += weight;
1599 shares += tg->cfs_rq[i]->shares;
1603 rq_weight = sum_weight;
1605 if ((!shares && rq_weight) || shares > tg->shares)
1606 shares = tg->shares;
1608 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1609 shares = tg->shares;
1611 for_each_cpu(i, sched_domain_span(sd))
1612 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1614 local_irq_restore(flags);
1620 * Compute the cpu's hierarchical load factor for each task group.
1621 * This needs to be done in a top-down fashion because the load of a child
1622 * group is a fraction of its parents load.
1624 static int tg_load_down(struct task_group *tg, void *data)
1627 long cpu = (long)data;
1630 load = cpu_rq(cpu)->load.weight;
1632 load = tg->parent->cfs_rq[cpu]->h_load;
1633 load *= tg->cfs_rq[cpu]->shares;
1634 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1637 tg->cfs_rq[cpu]->h_load = load;
1642 static void update_shares(struct sched_domain *sd)
1647 if (root_task_group_empty())
1650 now = cpu_clock(raw_smp_processor_id());
1651 elapsed = now - sd->last_update;
1653 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1654 sd->last_update = now;
1655 walk_tg_tree(tg_nop, tg_shares_up, sd);
1659 static void update_h_load(long cpu)
1661 if (root_task_group_empty())
1664 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1669 static inline void update_shares(struct sched_domain *sd)
1675 #ifdef CONFIG_PREEMPT
1677 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1680 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1681 * way at the expense of forcing extra atomic operations in all
1682 * invocations. This assures that the double_lock is acquired using the
1683 * same underlying policy as the spinlock_t on this architecture, which
1684 * reduces latency compared to the unfair variant below. However, it
1685 * also adds more overhead and therefore may reduce throughput.
1687 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1688 __releases(this_rq->lock)
1689 __acquires(busiest->lock)
1690 __acquires(this_rq->lock)
1692 raw_spin_unlock(&this_rq->lock);
1693 double_rq_lock(this_rq, busiest);
1700 * Unfair double_lock_balance: Optimizes throughput at the expense of
1701 * latency by eliminating extra atomic operations when the locks are
1702 * already in proper order on entry. This favors lower cpu-ids and will
1703 * grant the double lock to lower cpus over higher ids under contention,
1704 * regardless of entry order into the function.
1706 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1707 __releases(this_rq->lock)
1708 __acquires(busiest->lock)
1709 __acquires(this_rq->lock)
1713 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1714 if (busiest < this_rq) {
1715 raw_spin_unlock(&this_rq->lock);
1716 raw_spin_lock(&busiest->lock);
1717 raw_spin_lock_nested(&this_rq->lock,
1718 SINGLE_DEPTH_NESTING);
1721 raw_spin_lock_nested(&busiest->lock,
1722 SINGLE_DEPTH_NESTING);
1727 #endif /* CONFIG_PREEMPT */
1730 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1732 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1734 if (unlikely(!irqs_disabled())) {
1735 /* printk() doesn't work good under rq->lock */
1736 raw_spin_unlock(&this_rq->lock);
1740 return _double_lock_balance(this_rq, busiest);
1743 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1744 __releases(busiest->lock)
1746 raw_spin_unlock(&busiest->lock);
1747 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1751 * double_rq_lock - safely lock two runqueues
1753 * Note this does not disable interrupts like task_rq_lock,
1754 * you need to do so manually before calling.
1756 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1757 __acquires(rq1->lock)
1758 __acquires(rq2->lock)
1760 BUG_ON(!irqs_disabled());
1762 raw_spin_lock(&rq1->lock);
1763 __acquire(rq2->lock); /* Fake it out ;) */
1766 raw_spin_lock(&rq1->lock);
1767 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1769 raw_spin_lock(&rq2->lock);
1770 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1776 * double_rq_unlock - safely unlock two runqueues
1778 * Note this does not restore interrupts like task_rq_unlock,
1779 * you need to do so manually after calling.
1781 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1782 __releases(rq1->lock)
1783 __releases(rq2->lock)
1785 raw_spin_unlock(&rq1->lock);
1787 raw_spin_unlock(&rq2->lock);
1789 __release(rq2->lock);
1794 #ifdef CONFIG_FAIR_GROUP_SCHED
1795 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1798 cfs_rq->shares = shares;
1803 static void calc_load_account_idle(struct rq *this_rq);
1804 static void update_sysctl(void);
1805 static int get_update_sysctl_factor(void);
1807 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1809 set_task_rq(p, cpu);
1812 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1813 * successfuly executed on another CPU. We must ensure that updates of
1814 * per-task data have been completed by this moment.
1817 task_thread_info(p)->cpu = cpu;
1821 static const struct sched_class rt_sched_class;
1823 #define sched_class_highest (&rt_sched_class)
1824 #define for_each_class(class) \
1825 for (class = sched_class_highest; class; class = class->next)
1827 #include "sched_stats.h"
1829 static void inc_nr_running(struct rq *rq)
1834 static void dec_nr_running(struct rq *rq)
1839 static void set_load_weight(struct task_struct *p)
1841 if (task_has_rt_policy(p)) {
1842 p->se.load.weight = 0;
1843 p->se.load.inv_weight = WMULT_CONST;
1848 * SCHED_IDLE tasks get minimal weight:
1850 if (p->policy == SCHED_IDLE) {
1851 p->se.load.weight = WEIGHT_IDLEPRIO;
1852 p->se.load.inv_weight = WMULT_IDLEPRIO;
1856 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1857 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1860 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1862 update_rq_clock(rq);
1863 sched_info_queued(p);
1864 p->sched_class->enqueue_task(rq, p, flags);
1868 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1870 update_rq_clock(rq);
1871 sched_info_dequeued(p);
1872 p->sched_class->dequeue_task(rq, p, flags);
1877 * activate_task - move a task to the runqueue.
1879 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1881 if (task_contributes_to_load(p))
1882 rq->nr_uninterruptible--;
1884 enqueue_task(rq, p, flags);
1889 * deactivate_task - remove a task from the runqueue.
1891 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1893 if (task_contributes_to_load(p))
1894 rq->nr_uninterruptible++;
1896 dequeue_task(rq, p, flags);
1900 #include "sched_idletask.c"
1901 #include "sched_fair.c"
1902 #include "sched_rt.c"
1903 #ifdef CONFIG_SCHED_DEBUG
1904 # include "sched_debug.c"
1908 * __normal_prio - return the priority that is based on the static prio
1910 static inline int __normal_prio(struct task_struct *p)
1912 return p->static_prio;
1916 * Calculate the expected normal priority: i.e. priority
1917 * without taking RT-inheritance into account. Might be
1918 * boosted by interactivity modifiers. Changes upon fork,
1919 * setprio syscalls, and whenever the interactivity
1920 * estimator recalculates.
1922 static inline int normal_prio(struct task_struct *p)
1926 if (task_has_rt_policy(p))
1927 prio = MAX_RT_PRIO-1 - p->rt_priority;
1929 prio = __normal_prio(p);
1934 * Calculate the current priority, i.e. the priority
1935 * taken into account by the scheduler. This value might
1936 * be boosted by RT tasks, or might be boosted by
1937 * interactivity modifiers. Will be RT if the task got
1938 * RT-boosted. If not then it returns p->normal_prio.
1940 static int effective_prio(struct task_struct *p)
1942 p->normal_prio = normal_prio(p);
1944 * If we are RT tasks or we were boosted to RT priority,
1945 * keep the priority unchanged. Otherwise, update priority
1946 * to the normal priority:
1948 if (!rt_prio(p->prio))
1949 return p->normal_prio;
1954 * task_curr - is this task currently executing on a CPU?
1955 * @p: the task in question.
1957 inline int task_curr(const struct task_struct *p)
1959 return cpu_curr(task_cpu(p)) == p;
1962 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1963 const struct sched_class *prev_class,
1964 int oldprio, int running)
1966 if (prev_class != p->sched_class) {
1967 if (prev_class->switched_from)
1968 prev_class->switched_from(rq, p, running);
1969 p->sched_class->switched_to(rq, p, running);
1971 p->sched_class->prio_changed(rq, p, oldprio, running);
1976 * Is this task likely cache-hot:
1979 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1983 if (p->sched_class != &fair_sched_class)
1987 * Buddy candidates are cache hot:
1989 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
1990 (&p->se == cfs_rq_of(&p->se)->next ||
1991 &p->se == cfs_rq_of(&p->se)->last))
1994 if (sysctl_sched_migration_cost == -1)
1996 if (sysctl_sched_migration_cost == 0)
1999 delta = now - p->se.exec_start;
2001 return delta < (s64)sysctl_sched_migration_cost;
2004 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2006 #ifdef CONFIG_SCHED_DEBUG
2008 * We should never call set_task_cpu() on a blocked task,
2009 * ttwu() will sort out the placement.
2011 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2012 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2015 trace_sched_migrate_task(p, new_cpu);
2017 if (task_cpu(p) != new_cpu) {
2018 p->se.nr_migrations++;
2019 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2022 __set_task_cpu(p, new_cpu);
2025 struct migration_arg {
2026 struct task_struct *task;
2030 static int migration_cpu_stop(void *data);
2033 * The task's runqueue lock must be held.
2034 * Returns true if you have to wait for migration thread.
2036 static bool migrate_task(struct task_struct *p, int dest_cpu)
2038 struct rq *rq = task_rq(p);
2041 * If the task is not on a runqueue (and not running), then
2042 * the next wake-up will properly place the task.
2044 return p->se.on_rq || task_running(rq, p);
2048 * wait_task_inactive - wait for a thread to unschedule.
2050 * If @match_state is nonzero, it's the @p->state value just checked and
2051 * not expected to change. If it changes, i.e. @p might have woken up,
2052 * then return zero. When we succeed in waiting for @p to be off its CPU,
2053 * we return a positive number (its total switch count). If a second call
2054 * a short while later returns the same number, the caller can be sure that
2055 * @p has remained unscheduled the whole time.
2057 * The caller must ensure that the task *will* unschedule sometime soon,
2058 * else this function might spin for a *long* time. This function can't
2059 * be called with interrupts off, or it may introduce deadlock with
2060 * smp_call_function() if an IPI is sent by the same process we are
2061 * waiting to become inactive.
2063 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2065 unsigned long flags;
2072 * We do the initial early heuristics without holding
2073 * any task-queue locks at all. We'll only try to get
2074 * the runqueue lock when things look like they will
2080 * If the task is actively running on another CPU
2081 * still, just relax and busy-wait without holding
2084 * NOTE! Since we don't hold any locks, it's not
2085 * even sure that "rq" stays as the right runqueue!
2086 * But we don't care, since "task_running()" will
2087 * return false if the runqueue has changed and p
2088 * is actually now running somewhere else!
2090 while (task_running(rq, p)) {
2091 if (match_state && unlikely(p->state != match_state))
2097 * Ok, time to look more closely! We need the rq
2098 * lock now, to be *sure*. If we're wrong, we'll
2099 * just go back and repeat.
2101 rq = task_rq_lock(p, &flags);
2102 trace_sched_wait_task(p);
2103 running = task_running(rq, p);
2104 on_rq = p->se.on_rq;
2106 if (!match_state || p->state == match_state)
2107 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2108 task_rq_unlock(rq, &flags);
2111 * If it changed from the expected state, bail out now.
2113 if (unlikely(!ncsw))
2117 * Was it really running after all now that we
2118 * checked with the proper locks actually held?
2120 * Oops. Go back and try again..
2122 if (unlikely(running)) {
2128 * It's not enough that it's not actively running,
2129 * it must be off the runqueue _entirely_, and not
2132 * So if it was still runnable (but just not actively
2133 * running right now), it's preempted, and we should
2134 * yield - it could be a while.
2136 if (unlikely(on_rq)) {
2137 schedule_timeout_uninterruptible(1);
2142 * Ahh, all good. It wasn't running, and it wasn't
2143 * runnable, which means that it will never become
2144 * running in the future either. We're all done!
2153 * kick_process - kick a running thread to enter/exit the kernel
2154 * @p: the to-be-kicked thread
2156 * Cause a process which is running on another CPU to enter
2157 * kernel-mode, without any delay. (to get signals handled.)
2159 * NOTE: this function doesnt have to take the runqueue lock,
2160 * because all it wants to ensure is that the remote task enters
2161 * the kernel. If the IPI races and the task has been migrated
2162 * to another CPU then no harm is done and the purpose has been
2165 void kick_process(struct task_struct *p)
2171 if ((cpu != smp_processor_id()) && task_curr(p))
2172 smp_send_reschedule(cpu);
2175 EXPORT_SYMBOL_GPL(kick_process);
2176 #endif /* CONFIG_SMP */
2179 * task_oncpu_function_call - call a function on the cpu on which a task runs
2180 * @p: the task to evaluate
2181 * @func: the function to be called
2182 * @info: the function call argument
2184 * Calls the function @func when the task is currently running. This might
2185 * be on the current CPU, which just calls the function directly
2187 void task_oncpu_function_call(struct task_struct *p,
2188 void (*func) (void *info), void *info)
2195 smp_call_function_single(cpu, func, info, 1);
2201 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2203 static int select_fallback_rq(int cpu, struct task_struct *p)
2206 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2208 /* Look for allowed, online CPU in same node. */
2209 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2210 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2213 /* Any allowed, online CPU? */
2214 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2215 if (dest_cpu < nr_cpu_ids)
2218 /* No more Mr. Nice Guy. */
2219 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2220 dest_cpu = cpuset_cpus_allowed_fallback(p);
2222 * Don't tell them about moving exiting tasks or
2223 * kernel threads (both mm NULL), since they never
2226 if (p->mm && printk_ratelimit()) {
2227 printk(KERN_INFO "process %d (%s) no "
2228 "longer affine to cpu%d\n",
2229 task_pid_nr(p), p->comm, cpu);
2237 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2240 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2242 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2245 * In order not to call set_task_cpu() on a blocking task we need
2246 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2249 * Since this is common to all placement strategies, this lives here.
2251 * [ this allows ->select_task() to simply return task_cpu(p) and
2252 * not worry about this generic constraint ]
2254 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2256 cpu = select_fallback_rq(task_cpu(p), p);
2261 static void update_avg(u64 *avg, u64 sample)
2263 s64 diff = sample - *avg;
2268 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2269 bool is_sync, bool is_migrate, bool is_local,
2270 unsigned long en_flags)
2272 schedstat_inc(p, se.statistics.nr_wakeups);
2274 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2276 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2278 schedstat_inc(p, se.statistics.nr_wakeups_local);
2280 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2282 activate_task(rq, p, en_flags);
2285 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2286 int wake_flags, bool success)
2288 trace_sched_wakeup(p, success);
2289 check_preempt_curr(rq, p, wake_flags);
2291 p->state = TASK_RUNNING;
2293 if (p->sched_class->task_woken)
2294 p->sched_class->task_woken(rq, p);
2296 if (unlikely(rq->idle_stamp)) {
2297 u64 delta = rq->clock - rq->idle_stamp;
2298 u64 max = 2*sysctl_sched_migration_cost;
2303 update_avg(&rq->avg_idle, delta);
2307 /* if a worker is waking up, notify workqueue */
2308 if ((p->flags & PF_WQ_WORKER) && success)
2309 wq_worker_waking_up(p, cpu_of(rq));
2313 * try_to_wake_up - wake up a thread
2314 * @p: the thread to be awakened
2315 * @state: the mask of task states that can be woken
2316 * @wake_flags: wake modifier flags (WF_*)
2318 * Put it on the run-queue if it's not already there. The "current"
2319 * thread is always on the run-queue (except when the actual
2320 * re-schedule is in progress), and as such you're allowed to do
2321 * the simpler "current->state = TASK_RUNNING" to mark yourself
2322 * runnable without the overhead of this.
2324 * Returns %true if @p was woken up, %false if it was already running
2325 * or @state didn't match @p's state.
2327 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2330 int cpu, orig_cpu, this_cpu, success = 0;
2331 unsigned long flags;
2332 unsigned long en_flags = ENQUEUE_WAKEUP;
2335 this_cpu = get_cpu();
2338 rq = task_rq_lock(p, &flags);
2339 if (!(p->state & state))
2349 if (unlikely(task_running(rq, p)))
2353 * In order to handle concurrent wakeups and release the rq->lock
2354 * we put the task in TASK_WAKING state.
2356 * First fix up the nr_uninterruptible count:
2358 if (task_contributes_to_load(p)) {
2359 if (likely(cpu_online(orig_cpu)))
2360 rq->nr_uninterruptible--;
2362 this_rq()->nr_uninterruptible--;
2364 p->state = TASK_WAKING;
2366 if (p->sched_class->task_waking) {
2367 p->sched_class->task_waking(rq, p);
2368 en_flags |= ENQUEUE_WAKING;
2371 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2372 if (cpu != orig_cpu)
2373 set_task_cpu(p, cpu);
2374 __task_rq_unlock(rq);
2377 raw_spin_lock(&rq->lock);
2380 * We migrated the task without holding either rq->lock, however
2381 * since the task is not on the task list itself, nobody else
2382 * will try and migrate the task, hence the rq should match the
2383 * cpu we just moved it to.
2385 WARN_ON(task_cpu(p) != cpu);
2386 WARN_ON(p->state != TASK_WAKING);
2388 #ifdef CONFIG_SCHEDSTATS
2389 schedstat_inc(rq, ttwu_count);
2390 if (cpu == this_cpu)
2391 schedstat_inc(rq, ttwu_local);
2393 struct sched_domain *sd;
2394 for_each_domain(this_cpu, sd) {
2395 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2396 schedstat_inc(sd, ttwu_wake_remote);
2401 #endif /* CONFIG_SCHEDSTATS */
2404 #endif /* CONFIG_SMP */
2405 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2406 cpu == this_cpu, en_flags);
2409 ttwu_post_activation(p, rq, wake_flags, success);
2411 task_rq_unlock(rq, &flags);
2418 * try_to_wake_up_local - try to wake up a local task with rq lock held
2419 * @p: the thread to be awakened
2421 * Put @p on the run-queue if it's not alredy there. The caller must
2422 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2423 * the current task. this_rq() stays locked over invocation.
2425 static void try_to_wake_up_local(struct task_struct *p)
2427 struct rq *rq = task_rq(p);
2428 bool success = false;
2430 BUG_ON(rq != this_rq());
2431 BUG_ON(p == current);
2432 lockdep_assert_held(&rq->lock);
2434 if (!(p->state & TASK_NORMAL))
2438 if (likely(!task_running(rq, p))) {
2439 schedstat_inc(rq, ttwu_count);
2440 schedstat_inc(rq, ttwu_local);
2442 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2445 ttwu_post_activation(p, rq, 0, success);
2449 * wake_up_process - Wake up a specific process
2450 * @p: The process to be woken up.
2452 * Attempt to wake up the nominated process and move it to the set of runnable
2453 * processes. Returns 1 if the process was woken up, 0 if it was already
2456 * It may be assumed that this function implies a write memory barrier before
2457 * changing the task state if and only if any tasks are woken up.
2459 int wake_up_process(struct task_struct *p)
2461 return try_to_wake_up(p, TASK_ALL, 0);
2463 EXPORT_SYMBOL(wake_up_process);
2465 int wake_up_state(struct task_struct *p, unsigned int state)
2467 return try_to_wake_up(p, state, 0);
2471 * Perform scheduler related setup for a newly forked process p.
2472 * p is forked by current.
2474 * __sched_fork() is basic setup used by init_idle() too:
2476 static void __sched_fork(struct task_struct *p)
2478 p->se.exec_start = 0;
2479 p->se.sum_exec_runtime = 0;
2480 p->se.prev_sum_exec_runtime = 0;
2481 p->se.nr_migrations = 0;
2483 #ifdef CONFIG_SCHEDSTATS
2484 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2487 INIT_LIST_HEAD(&p->rt.run_list);
2489 INIT_LIST_HEAD(&p->se.group_node);
2491 #ifdef CONFIG_PREEMPT_NOTIFIERS
2492 INIT_HLIST_HEAD(&p->preempt_notifiers);
2497 * fork()/clone()-time setup:
2499 void sched_fork(struct task_struct *p, int clone_flags)
2501 int cpu = get_cpu();
2505 * We mark the process as running here. This guarantees that
2506 * nobody will actually run it, and a signal or other external
2507 * event cannot wake it up and insert it on the runqueue either.
2509 p->state = TASK_RUNNING;
2512 * Revert to default priority/policy on fork if requested.
2514 if (unlikely(p->sched_reset_on_fork)) {
2515 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2516 p->policy = SCHED_NORMAL;
2517 p->normal_prio = p->static_prio;
2520 if (PRIO_TO_NICE(p->static_prio) < 0) {
2521 p->static_prio = NICE_TO_PRIO(0);
2522 p->normal_prio = p->static_prio;
2527 * We don't need the reset flag anymore after the fork. It has
2528 * fulfilled its duty:
2530 p->sched_reset_on_fork = 0;
2534 * Make sure we do not leak PI boosting priority to the child.
2536 p->prio = current->normal_prio;
2538 if (!rt_prio(p->prio))
2539 p->sched_class = &fair_sched_class;
2541 if (p->sched_class->task_fork)
2542 p->sched_class->task_fork(p);
2544 set_task_cpu(p, cpu);
2546 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2547 if (likely(sched_info_on()))
2548 memset(&p->sched_info, 0, sizeof(p->sched_info));
2550 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2553 #ifdef CONFIG_PREEMPT
2554 /* Want to start with kernel preemption disabled. */
2555 task_thread_info(p)->preempt_count = 1;
2557 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2563 * wake_up_new_task - wake up a newly created task for the first time.
2565 * This function will do some initial scheduler statistics housekeeping
2566 * that must be done for every newly created context, then puts the task
2567 * on the runqueue and wakes it.
2569 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2571 unsigned long flags;
2573 int cpu __maybe_unused = get_cpu();
2576 rq = task_rq_lock(p, &flags);
2577 p->state = TASK_WAKING;
2580 * Fork balancing, do it here and not earlier because:
2581 * - cpus_allowed can change in the fork path
2582 * - any previously selected cpu might disappear through hotplug
2584 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2585 * without people poking at ->cpus_allowed.
2587 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2588 set_task_cpu(p, cpu);
2590 p->state = TASK_RUNNING;
2591 task_rq_unlock(rq, &flags);
2594 rq = task_rq_lock(p, &flags);
2595 activate_task(rq, p, 0);
2596 trace_sched_wakeup_new(p, 1);
2597 check_preempt_curr(rq, p, WF_FORK);
2599 if (p->sched_class->task_woken)
2600 p->sched_class->task_woken(rq, p);
2602 task_rq_unlock(rq, &flags);
2606 #ifdef CONFIG_PREEMPT_NOTIFIERS
2609 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2610 * @notifier: notifier struct to register
2612 void preempt_notifier_register(struct preempt_notifier *notifier)
2614 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2616 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2619 * preempt_notifier_unregister - no longer interested in preemption notifications
2620 * @notifier: notifier struct to unregister
2622 * This is safe to call from within a preemption notifier.
2624 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2626 hlist_del(¬ifier->link);
2628 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2630 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2632 struct preempt_notifier *notifier;
2633 struct hlist_node *node;
2635 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2636 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2640 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2641 struct task_struct *next)
2643 struct preempt_notifier *notifier;
2644 struct hlist_node *node;
2646 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2647 notifier->ops->sched_out(notifier, next);
2650 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2652 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2657 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2658 struct task_struct *next)
2662 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2665 * prepare_task_switch - prepare to switch tasks
2666 * @rq: the runqueue preparing to switch
2667 * @prev: the current task that is being switched out
2668 * @next: the task we are going to switch to.
2670 * This is called with the rq lock held and interrupts off. It must
2671 * be paired with a subsequent finish_task_switch after the context
2674 * prepare_task_switch sets up locking and calls architecture specific
2678 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2679 struct task_struct *next)
2681 fire_sched_out_preempt_notifiers(prev, next);
2682 prepare_lock_switch(rq, next);
2683 prepare_arch_switch(next);
2687 * finish_task_switch - clean up after a task-switch
2688 * @rq: runqueue associated with task-switch
2689 * @prev: the thread we just switched away from.
2691 * finish_task_switch must be called after the context switch, paired
2692 * with a prepare_task_switch call before the context switch.
2693 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2694 * and do any other architecture-specific cleanup actions.
2696 * Note that we may have delayed dropping an mm in context_switch(). If
2697 * so, we finish that here outside of the runqueue lock. (Doing it
2698 * with the lock held can cause deadlocks; see schedule() for
2701 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2702 __releases(rq->lock)
2704 struct mm_struct *mm = rq->prev_mm;
2710 * A task struct has one reference for the use as "current".
2711 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2712 * schedule one last time. The schedule call will never return, and
2713 * the scheduled task must drop that reference.
2714 * The test for TASK_DEAD must occur while the runqueue locks are
2715 * still held, otherwise prev could be scheduled on another cpu, die
2716 * there before we look at prev->state, and then the reference would
2718 * Manfred Spraul <manfred@colorfullife.com>
2720 prev_state = prev->state;
2721 finish_arch_switch(prev);
2722 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2723 local_irq_disable();
2724 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2725 perf_event_task_sched_in(current);
2726 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2728 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2729 finish_lock_switch(rq, prev);
2731 fire_sched_in_preempt_notifiers(current);
2734 if (unlikely(prev_state == TASK_DEAD)) {
2736 * Remove function-return probe instances associated with this
2737 * task and put them back on the free list.
2739 kprobe_flush_task(prev);
2740 put_task_struct(prev);
2746 /* assumes rq->lock is held */
2747 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2749 if (prev->sched_class->pre_schedule)
2750 prev->sched_class->pre_schedule(rq, prev);
2753 /* rq->lock is NOT held, but preemption is disabled */
2754 static inline void post_schedule(struct rq *rq)
2756 if (rq->post_schedule) {
2757 unsigned long flags;
2759 raw_spin_lock_irqsave(&rq->lock, flags);
2760 if (rq->curr->sched_class->post_schedule)
2761 rq->curr->sched_class->post_schedule(rq);
2762 raw_spin_unlock_irqrestore(&rq->lock, flags);
2764 rq->post_schedule = 0;
2770 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2774 static inline void post_schedule(struct rq *rq)
2781 * schedule_tail - first thing a freshly forked thread must call.
2782 * @prev: the thread we just switched away from.
2784 asmlinkage void schedule_tail(struct task_struct *prev)
2785 __releases(rq->lock)
2787 struct rq *rq = this_rq();
2789 finish_task_switch(rq, prev);
2792 * FIXME: do we need to worry about rq being invalidated by the
2797 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2798 /* In this case, finish_task_switch does not reenable preemption */
2801 if (current->set_child_tid)
2802 put_user(task_pid_vnr(current), current->set_child_tid);
2806 * context_switch - switch to the new MM and the new
2807 * thread's register state.
2810 context_switch(struct rq *rq, struct task_struct *prev,
2811 struct task_struct *next)
2813 struct mm_struct *mm, *oldmm;
2815 prepare_task_switch(rq, prev, next);
2816 trace_sched_switch(prev, next);
2818 oldmm = prev->active_mm;
2820 * For paravirt, this is coupled with an exit in switch_to to
2821 * combine the page table reload and the switch backend into
2824 arch_start_context_switch(prev);
2827 next->active_mm = oldmm;
2828 atomic_inc(&oldmm->mm_count);
2829 enter_lazy_tlb(oldmm, next);
2831 switch_mm(oldmm, mm, next);
2833 if (likely(!prev->mm)) {
2834 prev->active_mm = NULL;
2835 rq->prev_mm = oldmm;
2838 * Since the runqueue lock will be released by the next
2839 * task (which is an invalid locking op but in the case
2840 * of the scheduler it's an obvious special-case), so we
2841 * do an early lockdep release here:
2843 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2844 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2847 /* Here we just switch the register state and the stack. */
2848 switch_to(prev, next, prev);
2852 * this_rq must be evaluated again because prev may have moved
2853 * CPUs since it called schedule(), thus the 'rq' on its stack
2854 * frame will be invalid.
2856 finish_task_switch(this_rq(), prev);
2860 * nr_running, nr_uninterruptible and nr_context_switches:
2862 * externally visible scheduler statistics: current number of runnable
2863 * threads, current number of uninterruptible-sleeping threads, total
2864 * number of context switches performed since bootup.
2866 unsigned long nr_running(void)
2868 unsigned long i, sum = 0;
2870 for_each_online_cpu(i)
2871 sum += cpu_rq(i)->nr_running;
2876 unsigned long nr_uninterruptible(void)
2878 unsigned long i, sum = 0;
2880 for_each_possible_cpu(i)
2881 sum += cpu_rq(i)->nr_uninterruptible;
2884 * Since we read the counters lockless, it might be slightly
2885 * inaccurate. Do not allow it to go below zero though:
2887 if (unlikely((long)sum < 0))
2893 unsigned long long nr_context_switches(void)
2896 unsigned long long sum = 0;
2898 for_each_possible_cpu(i)
2899 sum += cpu_rq(i)->nr_switches;
2904 unsigned long nr_iowait(void)
2906 unsigned long i, sum = 0;
2908 for_each_possible_cpu(i)
2909 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2914 unsigned long nr_iowait_cpu(void)
2916 struct rq *this = this_rq();
2917 return atomic_read(&this->nr_iowait);
2920 unsigned long this_cpu_load(void)
2922 struct rq *this = this_rq();
2923 return this->cpu_load[0];
2927 /* Variables and functions for calc_load */
2928 static atomic_long_t calc_load_tasks;
2929 static unsigned long calc_load_update;
2930 unsigned long avenrun[3];
2931 EXPORT_SYMBOL(avenrun);
2933 static long calc_load_fold_active(struct rq *this_rq)
2935 long nr_active, delta = 0;
2937 nr_active = this_rq->nr_running;
2938 nr_active += (long) this_rq->nr_uninterruptible;
2940 if (nr_active != this_rq->calc_load_active) {
2941 delta = nr_active - this_rq->calc_load_active;
2942 this_rq->calc_load_active = nr_active;
2950 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2952 * When making the ILB scale, we should try to pull this in as well.
2954 static atomic_long_t calc_load_tasks_idle;
2956 static void calc_load_account_idle(struct rq *this_rq)
2960 delta = calc_load_fold_active(this_rq);
2962 atomic_long_add(delta, &calc_load_tasks_idle);
2965 static long calc_load_fold_idle(void)
2970 * Its got a race, we don't care...
2972 if (atomic_long_read(&calc_load_tasks_idle))
2973 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2978 static void calc_load_account_idle(struct rq *this_rq)
2982 static inline long calc_load_fold_idle(void)
2989 * get_avenrun - get the load average array
2990 * @loads: pointer to dest load array
2991 * @offset: offset to add
2992 * @shift: shift count to shift the result left
2994 * These values are estimates at best, so no need for locking.
2996 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2998 loads[0] = (avenrun[0] + offset) << shift;
2999 loads[1] = (avenrun[1] + offset) << shift;
3000 loads[2] = (avenrun[2] + offset) << shift;
3003 static unsigned long
3004 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3007 load += active * (FIXED_1 - exp);
3008 return load >> FSHIFT;
3012 * calc_load - update the avenrun load estimates 10 ticks after the
3013 * CPUs have updated calc_load_tasks.
3015 void calc_global_load(void)
3017 unsigned long upd = calc_load_update + 10;
3020 if (time_before(jiffies, upd))
3023 active = atomic_long_read(&calc_load_tasks);
3024 active = active > 0 ? active * FIXED_1 : 0;
3026 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3027 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3028 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3030 calc_load_update += LOAD_FREQ;
3034 * Called from update_cpu_load() to periodically update this CPU's
3037 static void calc_load_account_active(struct rq *this_rq)
3041 if (time_before(jiffies, this_rq->calc_load_update))
3044 delta = calc_load_fold_active(this_rq);
3045 delta += calc_load_fold_idle();
3047 atomic_long_add(delta, &calc_load_tasks);
3049 this_rq->calc_load_update += LOAD_FREQ;
3053 * Update rq->cpu_load[] statistics. This function is usually called every
3054 * scheduler tick (TICK_NSEC).
3056 static void update_cpu_load(struct rq *this_rq)
3058 unsigned long this_load = this_rq->load.weight;
3061 this_rq->nr_load_updates++;
3063 /* Update our load: */
3064 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3065 unsigned long old_load, new_load;
3067 /* scale is effectively 1 << i now, and >> i divides by scale */
3069 old_load = this_rq->cpu_load[i];
3070 new_load = this_load;
3072 * Round up the averaging division if load is increasing. This
3073 * prevents us from getting stuck on 9 if the load is 10, for
3076 if (new_load > old_load)
3077 new_load += scale-1;
3078 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3081 calc_load_account_active(this_rq);
3087 * sched_exec - execve() is a valuable balancing opportunity, because at
3088 * this point the task has the smallest effective memory and cache footprint.
3090 void sched_exec(void)
3092 struct task_struct *p = current;
3093 unsigned long flags;
3097 rq = task_rq_lock(p, &flags);
3098 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3099 if (dest_cpu == smp_processor_id())
3103 * select_task_rq() can race against ->cpus_allowed
3105 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3106 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3107 struct migration_arg arg = { p, dest_cpu };
3109 task_rq_unlock(rq, &flags);
3110 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3114 task_rq_unlock(rq, &flags);
3119 DEFINE_PER_CPU(struct kernel_stat, kstat);
3121 EXPORT_PER_CPU_SYMBOL(kstat);
3124 * Return any ns on the sched_clock that have not yet been accounted in
3125 * @p in case that task is currently running.
3127 * Called with task_rq_lock() held on @rq.
3129 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3133 if (task_current(rq, p)) {
3134 update_rq_clock(rq);
3135 ns = rq->clock - p->se.exec_start;
3143 unsigned long long task_delta_exec(struct task_struct *p)
3145 unsigned long flags;
3149 rq = task_rq_lock(p, &flags);
3150 ns = do_task_delta_exec(p, rq);
3151 task_rq_unlock(rq, &flags);
3157 * Return accounted runtime for the task.
3158 * In case the task is currently running, return the runtime plus current's
3159 * pending runtime that have not been accounted yet.
3161 unsigned long long task_sched_runtime(struct task_struct *p)
3163 unsigned long flags;
3167 rq = task_rq_lock(p, &flags);
3168 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3169 task_rq_unlock(rq, &flags);
3175 * Return sum_exec_runtime for the thread group.
3176 * In case the task is currently running, return the sum plus current's
3177 * pending runtime that have not been accounted yet.
3179 * Note that the thread group might have other running tasks as well,
3180 * so the return value not includes other pending runtime that other
3181 * running tasks might have.
3183 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3185 struct task_cputime totals;
3186 unsigned long flags;
3190 rq = task_rq_lock(p, &flags);
3191 thread_group_cputime(p, &totals);
3192 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3193 task_rq_unlock(rq, &flags);
3199 * Account user cpu time to a process.
3200 * @p: the process that the cpu time gets accounted to
3201 * @cputime: the cpu time spent in user space since the last update
3202 * @cputime_scaled: cputime scaled by cpu frequency
3204 void account_user_time(struct task_struct *p, cputime_t cputime,
3205 cputime_t cputime_scaled)
3207 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3210 /* Add user time to process. */
3211 p->utime = cputime_add(p->utime, cputime);
3212 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3213 account_group_user_time(p, cputime);
3215 /* Add user time to cpustat. */
3216 tmp = cputime_to_cputime64(cputime);
3217 if (TASK_NICE(p) > 0)
3218 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3220 cpustat->user = cputime64_add(cpustat->user, tmp);
3222 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3223 /* Account for user time used */
3224 acct_update_integrals(p);
3228 * Account guest cpu time to a process.
3229 * @p: the process that the cpu time gets accounted to
3230 * @cputime: the cpu time spent in virtual machine since the last update
3231 * @cputime_scaled: cputime scaled by cpu frequency
3233 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3234 cputime_t cputime_scaled)
3237 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3239 tmp = cputime_to_cputime64(cputime);
3241 /* Add guest time to process. */
3242 p->utime = cputime_add(p->utime, cputime);
3243 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3244 account_group_user_time(p, cputime);
3245 p->gtime = cputime_add(p->gtime, cputime);
3247 /* Add guest time to cpustat. */
3248 if (TASK_NICE(p) > 0) {
3249 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3250 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3252 cpustat->user = cputime64_add(cpustat->user, tmp);
3253 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3258 * Account system cpu time to a process.
3259 * @p: the process that the cpu time gets accounted to
3260 * @hardirq_offset: the offset to subtract from hardirq_count()
3261 * @cputime: the cpu time spent in kernel space since the last update
3262 * @cputime_scaled: cputime scaled by cpu frequency
3264 void account_system_time(struct task_struct *p, int hardirq_offset,
3265 cputime_t cputime, cputime_t cputime_scaled)
3267 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3270 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3271 account_guest_time(p, cputime, cputime_scaled);
3275 /* Add system time to process. */
3276 p->stime = cputime_add(p->stime, cputime);
3277 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3278 account_group_system_time(p, cputime);
3280 /* Add system time to cpustat. */
3281 tmp = cputime_to_cputime64(cputime);
3282 if (hardirq_count() - hardirq_offset)
3283 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3284 else if (softirq_count())
3285 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3287 cpustat->system = cputime64_add(cpustat->system, tmp);
3289 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3291 /* Account for system time used */
3292 acct_update_integrals(p);
3296 * Account for involuntary wait time.
3297 * @steal: the cpu time spent in involuntary wait
3299 void account_steal_time(cputime_t cputime)
3301 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3302 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3304 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3308 * Account for idle time.
3309 * @cputime: the cpu time spent in idle wait
3311 void account_idle_time(cputime_t cputime)
3313 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3314 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3315 struct rq *rq = this_rq();
3317 if (atomic_read(&rq->nr_iowait) > 0)
3318 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3320 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3323 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3326 * Account a single tick of cpu time.
3327 * @p: the process that the cpu time gets accounted to
3328 * @user_tick: indicates if the tick is a user or a system tick
3330 void account_process_tick(struct task_struct *p, int user_tick)
3332 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3333 struct rq *rq = this_rq();
3336 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3337 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3338 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3341 account_idle_time(cputime_one_jiffy);
3345 * Account multiple ticks of steal time.
3346 * @p: the process from which the cpu time has been stolen
3347 * @ticks: number of stolen ticks
3349 void account_steal_ticks(unsigned long ticks)
3351 account_steal_time(jiffies_to_cputime(ticks));
3355 * Account multiple ticks of idle time.
3356 * @ticks: number of stolen ticks
3358 void account_idle_ticks(unsigned long ticks)
3360 account_idle_time(jiffies_to_cputime(ticks));
3366 * Use precise platform statistics if available:
3368 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3369 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3375 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3377 struct task_cputime cputime;
3379 thread_group_cputime(p, &cputime);
3381 *ut = cputime.utime;
3382 *st = cputime.stime;
3386 #ifndef nsecs_to_cputime
3387 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3390 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3392 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3395 * Use CFS's precise accounting:
3397 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3402 temp = (u64)(rtime * utime);
3403 do_div(temp, total);
3404 utime = (cputime_t)temp;
3409 * Compare with previous values, to keep monotonicity:
3411 p->prev_utime = max(p->prev_utime, utime);
3412 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3414 *ut = p->prev_utime;
3415 *st = p->prev_stime;
3419 * Must be called with siglock held.
3421 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3423 struct signal_struct *sig = p->signal;
3424 struct task_cputime cputime;
3425 cputime_t rtime, utime, total;
3427 thread_group_cputime(p, &cputime);
3429 total = cputime_add(cputime.utime, cputime.stime);
3430 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3435 temp = (u64)(rtime * cputime.utime);
3436 do_div(temp, total);
3437 utime = (cputime_t)temp;
3441 sig->prev_utime = max(sig->prev_utime, utime);
3442 sig->prev_stime = max(sig->prev_stime,
3443 cputime_sub(rtime, sig->prev_utime));
3445 *ut = sig->prev_utime;
3446 *st = sig->prev_stime;
3451 * This function gets called by the timer code, with HZ frequency.
3452 * We call it with interrupts disabled.
3454 * It also gets called by the fork code, when changing the parent's
3457 void scheduler_tick(void)
3459 int cpu = smp_processor_id();
3460 struct rq *rq = cpu_rq(cpu);
3461 struct task_struct *curr = rq->curr;
3465 raw_spin_lock(&rq->lock);
3466 update_rq_clock(rq);
3467 update_cpu_load(rq);
3468 curr->sched_class->task_tick(rq, curr, 0);
3469 raw_spin_unlock(&rq->lock);
3471 perf_event_task_tick(curr);
3474 rq->idle_at_tick = idle_cpu(cpu);
3475 trigger_load_balance(rq, cpu);
3479 notrace unsigned long get_parent_ip(unsigned long addr)
3481 if (in_lock_functions(addr)) {
3482 addr = CALLER_ADDR2;
3483 if (in_lock_functions(addr))
3484 addr = CALLER_ADDR3;
3489 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3490 defined(CONFIG_PREEMPT_TRACER))
3492 void __kprobes add_preempt_count(int val)
3494 #ifdef CONFIG_DEBUG_PREEMPT
3498 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3501 preempt_count() += val;
3502 #ifdef CONFIG_DEBUG_PREEMPT
3504 * Spinlock count overflowing soon?
3506 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3509 if (preempt_count() == val)
3510 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3512 EXPORT_SYMBOL(add_preempt_count);
3514 void __kprobes sub_preempt_count(int val)
3516 #ifdef CONFIG_DEBUG_PREEMPT
3520 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3523 * Is the spinlock portion underflowing?
3525 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3526 !(preempt_count() & PREEMPT_MASK)))
3530 if (preempt_count() == val)
3531 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3532 preempt_count() -= val;
3534 EXPORT_SYMBOL(sub_preempt_count);
3539 * Print scheduling while atomic bug:
3541 static noinline void __schedule_bug(struct task_struct *prev)
3543 struct pt_regs *regs = get_irq_regs();
3545 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3546 prev->comm, prev->pid, preempt_count());
3548 debug_show_held_locks(prev);
3550 if (irqs_disabled())
3551 print_irqtrace_events(prev);
3560 * Various schedule()-time debugging checks and statistics:
3562 static inline void schedule_debug(struct task_struct *prev)
3565 * Test if we are atomic. Since do_exit() needs to call into
3566 * schedule() atomically, we ignore that path for now.
3567 * Otherwise, whine if we are scheduling when we should not be.
3569 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3570 __schedule_bug(prev);
3572 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3574 schedstat_inc(this_rq(), sched_count);
3575 #ifdef CONFIG_SCHEDSTATS
3576 if (unlikely(prev->lock_depth >= 0)) {
3577 schedstat_inc(this_rq(), bkl_count);
3578 schedstat_inc(prev, sched_info.bkl_count);
3583 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3586 update_rq_clock(rq);
3587 rq->skip_clock_update = 0;
3588 prev->sched_class->put_prev_task(rq, prev);
3592 * Pick up the highest-prio task:
3594 static inline struct task_struct *
3595 pick_next_task(struct rq *rq)
3597 const struct sched_class *class;
3598 struct task_struct *p;
3601 * Optimization: we know that if all tasks are in
3602 * the fair class we can call that function directly:
3604 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3605 p = fair_sched_class.pick_next_task(rq);
3610 class = sched_class_highest;
3612 p = class->pick_next_task(rq);
3616 * Will never be NULL as the idle class always
3617 * returns a non-NULL p:
3619 class = class->next;
3624 * schedule() is the main scheduler function.
3626 asmlinkage void __sched schedule(void)
3628 struct task_struct *prev, *next;
3629 unsigned long *switch_count;
3635 cpu = smp_processor_id();
3637 rcu_note_context_switch(cpu);
3639 switch_count = &prev->nivcsw;
3641 release_kernel_lock(prev);
3642 need_resched_nonpreemptible:
3644 schedule_debug(prev);
3646 if (sched_feat(HRTICK))
3649 raw_spin_lock_irq(&rq->lock);
3650 clear_tsk_need_resched(prev);
3652 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3653 if (unlikely(signal_pending_state(prev->state, prev))) {
3654 prev->state = TASK_RUNNING;
3657 * If a worker is going to sleep, notify and
3658 * ask workqueue whether it wants to wake up a
3659 * task to maintain concurrency. If so, wake
3662 if (prev->flags & PF_WQ_WORKER) {
3663 struct task_struct *to_wakeup;
3665 to_wakeup = wq_worker_sleeping(prev, cpu);
3667 try_to_wake_up_local(to_wakeup);
3669 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3671 switch_count = &prev->nvcsw;
3674 pre_schedule(rq, prev);
3676 if (unlikely(!rq->nr_running))
3677 idle_balance(cpu, rq);
3679 put_prev_task(rq, prev);
3680 next = pick_next_task(rq);
3682 if (likely(prev != next)) {
3683 sched_info_switch(prev, next);
3684 perf_event_task_sched_out(prev, next);
3690 context_switch(rq, prev, next); /* unlocks the rq */
3692 * the context switch might have flipped the stack from under
3693 * us, hence refresh the local variables.
3695 cpu = smp_processor_id();
3698 raw_spin_unlock_irq(&rq->lock);
3702 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3704 switch_count = &prev->nivcsw;
3705 goto need_resched_nonpreemptible;
3708 preempt_enable_no_resched();
3712 EXPORT_SYMBOL(schedule);
3714 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3716 * Look out! "owner" is an entirely speculative pointer
3717 * access and not reliable.
3719 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3724 if (!sched_feat(OWNER_SPIN))
3727 #ifdef CONFIG_DEBUG_PAGEALLOC
3729 * Need to access the cpu field knowing that
3730 * DEBUG_PAGEALLOC could have unmapped it if
3731 * the mutex owner just released it and exited.
3733 if (probe_kernel_address(&owner->cpu, cpu))
3740 * Even if the access succeeded (likely case),
3741 * the cpu field may no longer be valid.
3743 if (cpu >= nr_cpumask_bits)
3747 * We need to validate that we can do a
3748 * get_cpu() and that we have the percpu area.
3750 if (!cpu_online(cpu))
3757 * Owner changed, break to re-assess state.
3759 if (lock->owner != owner)
3763 * Is that owner really running on that cpu?
3765 if (task_thread_info(rq->curr) != owner || need_resched())
3775 #ifdef CONFIG_PREEMPT
3777 * this is the entry point to schedule() from in-kernel preemption
3778 * off of preempt_enable. Kernel preemptions off return from interrupt
3779 * occur there and call schedule directly.
3781 asmlinkage void __sched preempt_schedule(void)
3783 struct thread_info *ti = current_thread_info();
3786 * If there is a non-zero preempt_count or interrupts are disabled,
3787 * we do not want to preempt the current task. Just return..
3789 if (likely(ti->preempt_count || irqs_disabled()))
3793 add_preempt_count(PREEMPT_ACTIVE);
3795 sub_preempt_count(PREEMPT_ACTIVE);
3798 * Check again in case we missed a preemption opportunity
3799 * between schedule and now.
3802 } while (need_resched());
3804 EXPORT_SYMBOL(preempt_schedule);
3807 * this is the entry point to schedule() from kernel preemption
3808 * off of irq context.
3809 * Note, that this is called and return with irqs disabled. This will
3810 * protect us against recursive calling from irq.
3812 asmlinkage void __sched preempt_schedule_irq(void)
3814 struct thread_info *ti = current_thread_info();
3816 /* Catch callers which need to be fixed */
3817 BUG_ON(ti->preempt_count || !irqs_disabled());
3820 add_preempt_count(PREEMPT_ACTIVE);
3823 local_irq_disable();
3824 sub_preempt_count(PREEMPT_ACTIVE);
3827 * Check again in case we missed a preemption opportunity
3828 * between schedule and now.
3831 } while (need_resched());
3834 #endif /* CONFIG_PREEMPT */
3836 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3839 return try_to_wake_up(curr->private, mode, wake_flags);
3841 EXPORT_SYMBOL(default_wake_function);
3844 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3845 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3846 * number) then we wake all the non-exclusive tasks and one exclusive task.
3848 * There are circumstances in which we can try to wake a task which has already
3849 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3850 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3852 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3853 int nr_exclusive, int wake_flags, void *key)
3855 wait_queue_t *curr, *next;
3857 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3858 unsigned flags = curr->flags;
3860 if (curr->func(curr, mode, wake_flags, key) &&
3861 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3867 * __wake_up - wake up threads blocked on a waitqueue.
3869 * @mode: which threads
3870 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3871 * @key: is directly passed to the wakeup function
3873 * It may be assumed that this function implies a write memory barrier before
3874 * changing the task state if and only if any tasks are woken up.
3876 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3877 int nr_exclusive, void *key)
3879 unsigned long flags;
3881 spin_lock_irqsave(&q->lock, flags);
3882 __wake_up_common(q, mode, nr_exclusive, 0, key);
3883 spin_unlock_irqrestore(&q->lock, flags);
3885 EXPORT_SYMBOL(__wake_up);
3888 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3890 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3892 __wake_up_common(q, mode, 1, 0, NULL);
3894 EXPORT_SYMBOL_GPL(__wake_up_locked);
3896 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3898 __wake_up_common(q, mode, 1, 0, key);
3902 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3904 * @mode: which threads
3905 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3906 * @key: opaque value to be passed to wakeup targets
3908 * The sync wakeup differs that the waker knows that it will schedule
3909 * away soon, so while the target thread will be woken up, it will not
3910 * be migrated to another CPU - ie. the two threads are 'synchronized'
3911 * with each other. This can prevent needless bouncing between CPUs.
3913 * On UP it can prevent extra preemption.
3915 * It may be assumed that this function implies a write memory barrier before
3916 * changing the task state if and only if any tasks are woken up.
3918 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3919 int nr_exclusive, void *key)
3921 unsigned long flags;
3922 int wake_flags = WF_SYNC;
3927 if (unlikely(!nr_exclusive))
3930 spin_lock_irqsave(&q->lock, flags);
3931 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3932 spin_unlock_irqrestore(&q->lock, flags);
3934 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3937 * __wake_up_sync - see __wake_up_sync_key()
3939 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3941 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3943 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3946 * complete: - signals a single thread waiting on this completion
3947 * @x: holds the state of this particular completion
3949 * This will wake up a single thread waiting on this completion. Threads will be
3950 * awakened in the same order in which they were queued.
3952 * See also complete_all(), wait_for_completion() and related routines.
3954 * It may be assumed that this function implies a write memory barrier before
3955 * changing the task state if and only if any tasks are woken up.
3957 void complete(struct completion *x)
3959 unsigned long flags;
3961 spin_lock_irqsave(&x->wait.lock, flags);
3963 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3964 spin_unlock_irqrestore(&x->wait.lock, flags);
3966 EXPORT_SYMBOL(complete);
3969 * complete_all: - signals all threads waiting on this completion
3970 * @x: holds the state of this particular completion
3972 * This will wake up all threads waiting on this particular completion event.
3974 * It may be assumed that this function implies a write memory barrier before
3975 * changing the task state if and only if any tasks are woken up.
3977 void complete_all(struct completion *x)
3979 unsigned long flags;
3981 spin_lock_irqsave(&x->wait.lock, flags);
3982 x->done += UINT_MAX/2;
3983 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3984 spin_unlock_irqrestore(&x->wait.lock, flags);
3986 EXPORT_SYMBOL(complete_all);
3988 static inline long __sched
3989 do_wait_for_common(struct completion *x, long timeout, int state)
3992 DECLARE_WAITQUEUE(wait, current);
3994 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3996 if (signal_pending_state(state, current)) {
3997 timeout = -ERESTARTSYS;
4000 __set_current_state(state);
4001 spin_unlock_irq(&x->wait.lock);
4002 timeout = schedule_timeout(timeout);
4003 spin_lock_irq(&x->wait.lock);
4004 } while (!x->done && timeout);
4005 __remove_wait_queue(&x->wait, &wait);
4010 return timeout ?: 1;
4014 wait_for_common(struct completion *x, long timeout, int state)
4018 spin_lock_irq(&x->wait.lock);
4019 timeout = do_wait_for_common(x, timeout, state);
4020 spin_unlock_irq(&x->wait.lock);
4025 * wait_for_completion: - waits for completion of a task
4026 * @x: holds the state of this particular completion
4028 * This waits to be signaled for completion of a specific task. It is NOT
4029 * interruptible and there is no timeout.
4031 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4032 * and interrupt capability. Also see complete().
4034 void __sched wait_for_completion(struct completion *x)
4036 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4038 EXPORT_SYMBOL(wait_for_completion);
4041 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4042 * @x: holds the state of this particular completion
4043 * @timeout: timeout value in jiffies
4045 * This waits for either a completion of a specific task to be signaled or for a
4046 * specified timeout to expire. The timeout is in jiffies. It is not
4049 unsigned long __sched
4050 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4052 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4054 EXPORT_SYMBOL(wait_for_completion_timeout);
4057 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4058 * @x: holds the state of this particular completion
4060 * This waits for completion of a specific task to be signaled. It is
4063 int __sched wait_for_completion_interruptible(struct completion *x)
4065 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4066 if (t == -ERESTARTSYS)
4070 EXPORT_SYMBOL(wait_for_completion_interruptible);
4073 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4074 * @x: holds the state of this particular completion
4075 * @timeout: timeout value in jiffies
4077 * This waits for either a completion of a specific task to be signaled or for a
4078 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4080 unsigned long __sched
4081 wait_for_completion_interruptible_timeout(struct completion *x,
4082 unsigned long timeout)
4084 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4086 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4089 * wait_for_completion_killable: - waits for completion of a task (killable)
4090 * @x: holds the state of this particular completion
4092 * This waits to be signaled for completion of a specific task. It can be
4093 * interrupted by a kill signal.
4095 int __sched wait_for_completion_killable(struct completion *x)
4097 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4098 if (t == -ERESTARTSYS)
4102 EXPORT_SYMBOL(wait_for_completion_killable);
4105 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4106 * @x: holds the state of this particular completion
4107 * @timeout: timeout value in jiffies
4109 * This waits for either a completion of a specific task to be
4110 * signaled or for a specified timeout to expire. It can be
4111 * interrupted by a kill signal. The timeout is in jiffies.
4113 unsigned long __sched
4114 wait_for_completion_killable_timeout(struct completion *x,
4115 unsigned long timeout)
4117 return wait_for_common(x, timeout, TASK_KILLABLE);
4119 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4122 * try_wait_for_completion - try to decrement a completion without blocking
4123 * @x: completion structure
4125 * Returns: 0 if a decrement cannot be done without blocking
4126 * 1 if a decrement succeeded.
4128 * If a completion is being used as a counting completion,
4129 * attempt to decrement the counter without blocking. This
4130 * enables us to avoid waiting if the resource the completion
4131 * is protecting is not available.
4133 bool try_wait_for_completion(struct completion *x)
4135 unsigned long flags;
4138 spin_lock_irqsave(&x->wait.lock, flags);
4143 spin_unlock_irqrestore(&x->wait.lock, flags);
4146 EXPORT_SYMBOL(try_wait_for_completion);
4149 * completion_done - Test to see if a completion has any waiters
4150 * @x: completion structure
4152 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4153 * 1 if there are no waiters.
4156 bool completion_done(struct completion *x)
4158 unsigned long flags;
4161 spin_lock_irqsave(&x->wait.lock, flags);
4164 spin_unlock_irqrestore(&x->wait.lock, flags);
4167 EXPORT_SYMBOL(completion_done);
4170 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4172 unsigned long flags;
4175 init_waitqueue_entry(&wait, current);
4177 __set_current_state(state);
4179 spin_lock_irqsave(&q->lock, flags);
4180 __add_wait_queue(q, &wait);
4181 spin_unlock(&q->lock);
4182 timeout = schedule_timeout(timeout);
4183 spin_lock_irq(&q->lock);
4184 __remove_wait_queue(q, &wait);
4185 spin_unlock_irqrestore(&q->lock, flags);
4190 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4192 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4194 EXPORT_SYMBOL(interruptible_sleep_on);
4197 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4199 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4201 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4203 void __sched sleep_on(wait_queue_head_t *q)
4205 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4207 EXPORT_SYMBOL(sleep_on);
4209 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4211 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4213 EXPORT_SYMBOL(sleep_on_timeout);
4215 #ifdef CONFIG_RT_MUTEXES
4218 * rt_mutex_setprio - set the current priority of a task
4220 * @prio: prio value (kernel-internal form)
4222 * This function changes the 'effective' priority of a task. It does
4223 * not touch ->normal_prio like __setscheduler().
4225 * Used by the rt_mutex code to implement priority inheritance logic.
4227 void rt_mutex_setprio(struct task_struct *p, int prio)
4229 unsigned long flags;
4230 int oldprio, on_rq, running;
4232 const struct sched_class *prev_class;
4234 BUG_ON(prio < 0 || prio > MAX_PRIO);
4236 rq = task_rq_lock(p, &flags);
4239 prev_class = p->sched_class;
4240 on_rq = p->se.on_rq;
4241 running = task_current(rq, p);
4243 dequeue_task(rq, p, 0);
4245 p->sched_class->put_prev_task(rq, p);
4248 p->sched_class = &rt_sched_class;
4250 p->sched_class = &fair_sched_class;
4255 p->sched_class->set_curr_task(rq);
4257 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4259 check_class_changed(rq, p, prev_class, oldprio, running);
4261 task_rq_unlock(rq, &flags);
4266 void set_user_nice(struct task_struct *p, long nice)
4268 int old_prio, delta, on_rq;
4269 unsigned long flags;
4272 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4275 * We have to be careful, if called from sys_setpriority(),
4276 * the task might be in the middle of scheduling on another CPU.
4278 rq = task_rq_lock(p, &flags);
4280 * The RT priorities are set via sched_setscheduler(), but we still
4281 * allow the 'normal' nice value to be set - but as expected
4282 * it wont have any effect on scheduling until the task is
4283 * SCHED_FIFO/SCHED_RR:
4285 if (task_has_rt_policy(p)) {
4286 p->static_prio = NICE_TO_PRIO(nice);
4289 on_rq = p->se.on_rq;
4291 dequeue_task(rq, p, 0);
4293 p->static_prio = NICE_TO_PRIO(nice);
4296 p->prio = effective_prio(p);
4297 delta = p->prio - old_prio;
4300 enqueue_task(rq, p, 0);
4302 * If the task increased its priority or is running and
4303 * lowered its priority, then reschedule its CPU:
4305 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4306 resched_task(rq->curr);
4309 task_rq_unlock(rq, &flags);
4311 EXPORT_SYMBOL(set_user_nice);
4314 * can_nice - check if a task can reduce its nice value
4318 int can_nice(const struct task_struct *p, const int nice)
4320 /* convert nice value [19,-20] to rlimit style value [1,40] */
4321 int nice_rlim = 20 - nice;
4323 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4324 capable(CAP_SYS_NICE));
4327 #ifdef __ARCH_WANT_SYS_NICE
4330 * sys_nice - change the priority of the current process.
4331 * @increment: priority increment
4333 * sys_setpriority is a more generic, but much slower function that
4334 * does similar things.
4336 SYSCALL_DEFINE1(nice, int, increment)
4341 * Setpriority might change our priority at the same moment.
4342 * We don't have to worry. Conceptually one call occurs first
4343 * and we have a single winner.
4345 if (increment < -40)
4350 nice = TASK_NICE(current) + increment;
4356 if (increment < 0 && !can_nice(current, nice))
4359 retval = security_task_setnice(current, nice);
4363 set_user_nice(current, nice);
4370 * task_prio - return the priority value of a given task.
4371 * @p: the task in question.
4373 * This is the priority value as seen by users in /proc.
4374 * RT tasks are offset by -200. Normal tasks are centered
4375 * around 0, value goes from -16 to +15.
4377 int task_prio(const struct task_struct *p)
4379 return p->prio - MAX_RT_PRIO;
4383 * task_nice - return the nice value of a given task.
4384 * @p: the task in question.
4386 int task_nice(const struct task_struct *p)
4388 return TASK_NICE(p);
4390 EXPORT_SYMBOL(task_nice);
4393 * idle_cpu - is a given cpu idle currently?
4394 * @cpu: the processor in question.
4396 int idle_cpu(int cpu)
4398 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4402 * idle_task - return the idle task for a given cpu.
4403 * @cpu: the processor in question.
4405 struct task_struct *idle_task(int cpu)
4407 return cpu_rq(cpu)->idle;
4411 * find_process_by_pid - find a process with a matching PID value.
4412 * @pid: the pid in question.
4414 static struct task_struct *find_process_by_pid(pid_t pid)
4416 return pid ? find_task_by_vpid(pid) : current;
4419 /* Actually do priority change: must hold rq lock. */
4421 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4423 BUG_ON(p->se.on_rq);
4426 p->rt_priority = prio;
4427 p->normal_prio = normal_prio(p);
4428 /* we are holding p->pi_lock already */
4429 p->prio = rt_mutex_getprio(p);
4430 if (rt_prio(p->prio))
4431 p->sched_class = &rt_sched_class;
4433 p->sched_class = &fair_sched_class;
4438 * check the target process has a UID that matches the current process's
4440 static bool check_same_owner(struct task_struct *p)
4442 const struct cred *cred = current_cred(), *pcred;
4446 pcred = __task_cred(p);
4447 match = (cred->euid == pcred->euid ||
4448 cred->euid == pcred->uid);
4453 static int __sched_setscheduler(struct task_struct *p, int policy,
4454 struct sched_param *param, bool user)
4456 int retval, oldprio, oldpolicy = -1, on_rq, running;
4457 unsigned long flags;
4458 const struct sched_class *prev_class;
4462 /* may grab non-irq protected spin_locks */
4463 BUG_ON(in_interrupt());
4465 /* double check policy once rq lock held */
4467 reset_on_fork = p->sched_reset_on_fork;
4468 policy = oldpolicy = p->policy;
4470 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4471 policy &= ~SCHED_RESET_ON_FORK;
4473 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4474 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4475 policy != SCHED_IDLE)
4480 * Valid priorities for SCHED_FIFO and SCHED_RR are
4481 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4482 * SCHED_BATCH and SCHED_IDLE is 0.
4484 if (param->sched_priority < 0 ||
4485 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4486 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4488 if (rt_policy(policy) != (param->sched_priority != 0))
4492 * Allow unprivileged RT tasks to decrease priority:
4494 if (user && !capable(CAP_SYS_NICE)) {
4495 if (rt_policy(policy)) {
4496 unsigned long rlim_rtprio;
4498 if (!lock_task_sighand(p, &flags))
4500 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4501 unlock_task_sighand(p, &flags);
4503 /* can't set/change the rt policy */
4504 if (policy != p->policy && !rlim_rtprio)
4507 /* can't increase priority */
4508 if (param->sched_priority > p->rt_priority &&
4509 param->sched_priority > rlim_rtprio)
4513 * Like positive nice levels, dont allow tasks to
4514 * move out of SCHED_IDLE either:
4516 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4519 /* can't change other user's priorities */
4520 if (!check_same_owner(p))
4523 /* Normal users shall not reset the sched_reset_on_fork flag */
4524 if (p->sched_reset_on_fork && !reset_on_fork)
4529 retval = security_task_setscheduler(p, policy, param);
4535 * make sure no PI-waiters arrive (or leave) while we are
4536 * changing the priority of the task:
4538 raw_spin_lock_irqsave(&p->pi_lock, flags);
4540 * To be able to change p->policy safely, the apropriate
4541 * runqueue lock must be held.
4543 rq = __task_rq_lock(p);
4545 #ifdef CONFIG_RT_GROUP_SCHED
4548 * Do not allow realtime tasks into groups that have no runtime
4551 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4552 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4553 __task_rq_unlock(rq);
4554 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4560 /* recheck policy now with rq lock held */
4561 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4562 policy = oldpolicy = -1;
4563 __task_rq_unlock(rq);
4564 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4567 on_rq = p->se.on_rq;
4568 running = task_current(rq, p);
4570 deactivate_task(rq, p, 0);
4572 p->sched_class->put_prev_task(rq, p);
4574 p->sched_reset_on_fork = reset_on_fork;
4577 prev_class = p->sched_class;
4578 __setscheduler(rq, p, policy, param->sched_priority);
4581 p->sched_class->set_curr_task(rq);
4583 activate_task(rq, p, 0);
4585 check_class_changed(rq, p, prev_class, oldprio, running);
4587 __task_rq_unlock(rq);
4588 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4590 rt_mutex_adjust_pi(p);
4596 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4597 * @p: the task in question.
4598 * @policy: new policy.
4599 * @param: structure containing the new RT priority.
4601 * NOTE that the task may be already dead.
4603 int sched_setscheduler(struct task_struct *p, int policy,
4604 struct sched_param *param)
4606 return __sched_setscheduler(p, policy, param, true);
4608 EXPORT_SYMBOL_GPL(sched_setscheduler);
4611 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4612 * @p: the task in question.
4613 * @policy: new policy.
4614 * @param: structure containing the new RT priority.
4616 * Just like sched_setscheduler, only don't bother checking if the
4617 * current context has permission. For example, this is needed in
4618 * stop_machine(): we create temporary high priority worker threads,
4619 * but our caller might not have that capability.
4621 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4622 struct sched_param *param)
4624 return __sched_setscheduler(p, policy, param, false);
4628 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4630 struct sched_param lparam;
4631 struct task_struct *p;
4634 if (!param || pid < 0)
4636 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4641 p = find_process_by_pid(pid);
4643 retval = sched_setscheduler(p, policy, &lparam);
4650 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4651 * @pid: the pid in question.
4652 * @policy: new policy.
4653 * @param: structure containing the new RT priority.
4655 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4656 struct sched_param __user *, param)
4658 /* negative values for policy are not valid */
4662 return do_sched_setscheduler(pid, policy, param);
4666 * sys_sched_setparam - set/change the RT priority of a thread
4667 * @pid: the pid in question.
4668 * @param: structure containing the new RT priority.
4670 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4672 return do_sched_setscheduler(pid, -1, param);
4676 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4677 * @pid: the pid in question.
4679 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4681 struct task_struct *p;
4689 p = find_process_by_pid(pid);
4691 retval = security_task_getscheduler(p);
4694 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4701 * sys_sched_getparam - get the RT priority of a thread
4702 * @pid: the pid in question.
4703 * @param: structure containing the RT priority.
4705 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4707 struct sched_param lp;
4708 struct task_struct *p;
4711 if (!param || pid < 0)
4715 p = find_process_by_pid(pid);
4720 retval = security_task_getscheduler(p);
4724 lp.sched_priority = p->rt_priority;
4728 * This one might sleep, we cannot do it with a spinlock held ...
4730 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4739 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4741 cpumask_var_t cpus_allowed, new_mask;
4742 struct task_struct *p;
4748 p = find_process_by_pid(pid);
4755 /* Prevent p going away */
4759 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4763 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4765 goto out_free_cpus_allowed;
4768 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4771 retval = security_task_setscheduler(p, 0, NULL);
4775 cpuset_cpus_allowed(p, cpus_allowed);
4776 cpumask_and(new_mask, in_mask, cpus_allowed);
4778 retval = set_cpus_allowed_ptr(p, new_mask);
4781 cpuset_cpus_allowed(p, cpus_allowed);
4782 if (!cpumask_subset(new_mask, cpus_allowed)) {
4784 * We must have raced with a concurrent cpuset
4785 * update. Just reset the cpus_allowed to the
4786 * cpuset's cpus_allowed
4788 cpumask_copy(new_mask, cpus_allowed);
4793 free_cpumask_var(new_mask);
4794 out_free_cpus_allowed:
4795 free_cpumask_var(cpus_allowed);
4802 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4803 struct cpumask *new_mask)
4805 if (len < cpumask_size())
4806 cpumask_clear(new_mask);
4807 else if (len > cpumask_size())
4808 len = cpumask_size();
4810 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4814 * sys_sched_setaffinity - set the cpu affinity of a process
4815 * @pid: pid of the process
4816 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4817 * @user_mask_ptr: user-space pointer to the new cpu mask
4819 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4820 unsigned long __user *, user_mask_ptr)
4822 cpumask_var_t new_mask;
4825 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4828 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4830 retval = sched_setaffinity(pid, new_mask);
4831 free_cpumask_var(new_mask);
4835 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4837 struct task_struct *p;
4838 unsigned long flags;
4846 p = find_process_by_pid(pid);
4850 retval = security_task_getscheduler(p);
4854 rq = task_rq_lock(p, &flags);
4855 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4856 task_rq_unlock(rq, &flags);
4866 * sys_sched_getaffinity - get the cpu affinity of a process
4867 * @pid: pid of the process
4868 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4869 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4871 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4872 unsigned long __user *, user_mask_ptr)
4877 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4879 if (len & (sizeof(unsigned long)-1))
4882 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4885 ret = sched_getaffinity(pid, mask);
4887 size_t retlen = min_t(size_t, len, cpumask_size());
4889 if (copy_to_user(user_mask_ptr, mask, retlen))
4894 free_cpumask_var(mask);
4900 * sys_sched_yield - yield the current processor to other threads.
4902 * This function yields the current CPU to other tasks. If there are no
4903 * other threads running on this CPU then this function will return.
4905 SYSCALL_DEFINE0(sched_yield)
4907 struct rq *rq = this_rq_lock();
4909 schedstat_inc(rq, yld_count);
4910 current->sched_class->yield_task(rq);
4913 * Since we are going to call schedule() anyway, there's
4914 * no need to preempt or enable interrupts:
4916 __release(rq->lock);
4917 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4918 do_raw_spin_unlock(&rq->lock);
4919 preempt_enable_no_resched();
4926 static inline int should_resched(void)
4928 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4931 static void __cond_resched(void)
4933 add_preempt_count(PREEMPT_ACTIVE);
4935 sub_preempt_count(PREEMPT_ACTIVE);
4938 int __sched _cond_resched(void)
4940 if (should_resched()) {
4946 EXPORT_SYMBOL(_cond_resched);
4949 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4950 * call schedule, and on return reacquire the lock.
4952 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4953 * operations here to prevent schedule() from being called twice (once via
4954 * spin_unlock(), once by hand).
4956 int __cond_resched_lock(spinlock_t *lock)
4958 int resched = should_resched();
4961 lockdep_assert_held(lock);
4963 if (spin_needbreak(lock) || resched) {
4974 EXPORT_SYMBOL(__cond_resched_lock);
4976 int __sched __cond_resched_softirq(void)
4978 BUG_ON(!in_softirq());
4980 if (should_resched()) {
4988 EXPORT_SYMBOL(__cond_resched_softirq);
4991 * yield - yield the current processor to other threads.
4993 * This is a shortcut for kernel-space yielding - it marks the
4994 * thread runnable and calls sys_sched_yield().
4996 void __sched yield(void)
4998 set_current_state(TASK_RUNNING);
5001 EXPORT_SYMBOL(yield);
5004 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5005 * that process accounting knows that this is a task in IO wait state.
5007 void __sched io_schedule(void)
5009 struct rq *rq = raw_rq();
5011 delayacct_blkio_start();
5012 atomic_inc(&rq->nr_iowait);
5013 current->in_iowait = 1;
5015 current->in_iowait = 0;
5016 atomic_dec(&rq->nr_iowait);
5017 delayacct_blkio_end();
5019 EXPORT_SYMBOL(io_schedule);
5021 long __sched io_schedule_timeout(long timeout)
5023 struct rq *rq = raw_rq();
5026 delayacct_blkio_start();
5027 atomic_inc(&rq->nr_iowait);
5028 current->in_iowait = 1;
5029 ret = schedule_timeout(timeout);
5030 current->in_iowait = 0;
5031 atomic_dec(&rq->nr_iowait);
5032 delayacct_blkio_end();
5037 * sys_sched_get_priority_max - return maximum RT priority.
5038 * @policy: scheduling class.
5040 * this syscall returns the maximum rt_priority that can be used
5041 * by a given scheduling class.
5043 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5050 ret = MAX_USER_RT_PRIO-1;
5062 * sys_sched_get_priority_min - return minimum RT priority.
5063 * @policy: scheduling class.
5065 * this syscall returns the minimum rt_priority that can be used
5066 * by a given scheduling class.
5068 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5086 * sys_sched_rr_get_interval - return the default timeslice of a process.
5087 * @pid: pid of the process.
5088 * @interval: userspace pointer to the timeslice value.
5090 * this syscall writes the default timeslice value of a given process
5091 * into the user-space timespec buffer. A value of '0' means infinity.
5093 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5094 struct timespec __user *, interval)
5096 struct task_struct *p;
5097 unsigned int time_slice;
5098 unsigned long flags;
5108 p = find_process_by_pid(pid);
5112 retval = security_task_getscheduler(p);
5116 rq = task_rq_lock(p, &flags);
5117 time_slice = p->sched_class->get_rr_interval(rq, p);
5118 task_rq_unlock(rq, &flags);
5121 jiffies_to_timespec(time_slice, &t);
5122 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5130 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5132 void sched_show_task(struct task_struct *p)
5134 unsigned long free = 0;
5137 state = p->state ? __ffs(p->state) + 1 : 0;
5138 printk(KERN_INFO "%-13.13s %c", p->comm,
5139 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5140 #if BITS_PER_LONG == 32
5141 if (state == TASK_RUNNING)
5142 printk(KERN_CONT " running ");
5144 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5146 if (state == TASK_RUNNING)
5147 printk(KERN_CONT " running task ");
5149 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5151 #ifdef CONFIG_DEBUG_STACK_USAGE
5152 free = stack_not_used(p);
5154 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5155 task_pid_nr(p), task_pid_nr(p->real_parent),
5156 (unsigned long)task_thread_info(p)->flags);
5158 show_stack(p, NULL);
5161 void show_state_filter(unsigned long state_filter)
5163 struct task_struct *g, *p;
5165 #if BITS_PER_LONG == 32
5167 " task PC stack pid father\n");
5170 " task PC stack pid father\n");
5172 read_lock(&tasklist_lock);
5173 do_each_thread(g, p) {
5175 * reset the NMI-timeout, listing all files on a slow
5176 * console might take alot of time:
5178 touch_nmi_watchdog();
5179 if (!state_filter || (p->state & state_filter))
5181 } while_each_thread(g, p);
5183 touch_all_softlockup_watchdogs();
5185 #ifdef CONFIG_SCHED_DEBUG
5186 sysrq_sched_debug_show();
5188 read_unlock(&tasklist_lock);
5190 * Only show locks if all tasks are dumped:
5193 debug_show_all_locks();
5196 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5198 idle->sched_class = &idle_sched_class;
5202 * init_idle - set up an idle thread for a given CPU
5203 * @idle: task in question
5204 * @cpu: cpu the idle task belongs to
5206 * NOTE: this function does not set the idle thread's NEED_RESCHED
5207 * flag, to make booting more robust.
5209 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5211 struct rq *rq = cpu_rq(cpu);
5212 unsigned long flags;
5214 raw_spin_lock_irqsave(&rq->lock, flags);
5217 idle->state = TASK_RUNNING;
5218 idle->se.exec_start = sched_clock();
5220 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5221 __set_task_cpu(idle, cpu);
5223 rq->curr = rq->idle = idle;
5224 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5227 raw_spin_unlock_irqrestore(&rq->lock, flags);
5229 /* Set the preempt count _outside_ the spinlocks! */
5230 #if defined(CONFIG_PREEMPT)
5231 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5233 task_thread_info(idle)->preempt_count = 0;
5236 * The idle tasks have their own, simple scheduling class:
5238 idle->sched_class = &idle_sched_class;
5239 ftrace_graph_init_task(idle);
5243 * In a system that switches off the HZ timer nohz_cpu_mask
5244 * indicates which cpus entered this state. This is used
5245 * in the rcu update to wait only for active cpus. For system
5246 * which do not switch off the HZ timer nohz_cpu_mask should
5247 * always be CPU_BITS_NONE.
5249 cpumask_var_t nohz_cpu_mask;
5252 * Increase the granularity value when there are more CPUs,
5253 * because with more CPUs the 'effective latency' as visible
5254 * to users decreases. But the relationship is not linear,
5255 * so pick a second-best guess by going with the log2 of the
5258 * This idea comes from the SD scheduler of Con Kolivas:
5260 static int get_update_sysctl_factor(void)
5262 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5263 unsigned int factor;
5265 switch (sysctl_sched_tunable_scaling) {
5266 case SCHED_TUNABLESCALING_NONE:
5269 case SCHED_TUNABLESCALING_LINEAR:
5272 case SCHED_TUNABLESCALING_LOG:
5274 factor = 1 + ilog2(cpus);
5281 static void update_sysctl(void)
5283 unsigned int factor = get_update_sysctl_factor();
5285 #define SET_SYSCTL(name) \
5286 (sysctl_##name = (factor) * normalized_sysctl_##name)
5287 SET_SYSCTL(sched_min_granularity);
5288 SET_SYSCTL(sched_latency);
5289 SET_SYSCTL(sched_wakeup_granularity);
5290 SET_SYSCTL(sched_shares_ratelimit);
5294 static inline void sched_init_granularity(void)
5301 * This is how migration works:
5303 * 1) we invoke migration_cpu_stop() on the target CPU using
5305 * 2) stopper starts to run (implicitly forcing the migrated thread
5307 * 3) it checks whether the migrated task is still in the wrong runqueue.
5308 * 4) if it's in the wrong runqueue then the migration thread removes
5309 * it and puts it into the right queue.
5310 * 5) stopper completes and stop_one_cpu() returns and the migration
5315 * Change a given task's CPU affinity. Migrate the thread to a
5316 * proper CPU and schedule it away if the CPU it's executing on
5317 * is removed from the allowed bitmask.
5319 * NOTE: the caller must have a valid reference to the task, the
5320 * task must not exit() & deallocate itself prematurely. The
5321 * call is not atomic; no spinlocks may be held.
5323 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5325 unsigned long flags;
5327 unsigned int dest_cpu;
5331 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5332 * drop the rq->lock and still rely on ->cpus_allowed.
5335 while (task_is_waking(p))
5337 rq = task_rq_lock(p, &flags);
5338 if (task_is_waking(p)) {
5339 task_rq_unlock(rq, &flags);
5343 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5348 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5349 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5354 if (p->sched_class->set_cpus_allowed)
5355 p->sched_class->set_cpus_allowed(p, new_mask);
5357 cpumask_copy(&p->cpus_allowed, new_mask);
5358 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5361 /* Can the task run on the task's current CPU? If so, we're done */
5362 if (cpumask_test_cpu(task_cpu(p), new_mask))
5365 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5366 if (migrate_task(p, dest_cpu)) {
5367 struct migration_arg arg = { p, dest_cpu };
5368 /* Need help from migration thread: drop lock and wait. */
5369 task_rq_unlock(rq, &flags);
5370 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5371 tlb_migrate_finish(p->mm);
5375 task_rq_unlock(rq, &flags);
5379 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5382 * Move (not current) task off this cpu, onto dest cpu. We're doing
5383 * this because either it can't run here any more (set_cpus_allowed()
5384 * away from this CPU, or CPU going down), or because we're
5385 * attempting to rebalance this task on exec (sched_exec).
5387 * So we race with normal scheduler movements, but that's OK, as long
5388 * as the task is no longer on this CPU.
5390 * Returns non-zero if task was successfully migrated.
5392 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5394 struct rq *rq_dest, *rq_src;
5397 if (unlikely(!cpu_active(dest_cpu)))
5400 rq_src = cpu_rq(src_cpu);
5401 rq_dest = cpu_rq(dest_cpu);
5403 double_rq_lock(rq_src, rq_dest);
5404 /* Already moved. */
5405 if (task_cpu(p) != src_cpu)
5407 /* Affinity changed (again). */
5408 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5412 * If we're not on a rq, the next wake-up will ensure we're
5416 deactivate_task(rq_src, p, 0);
5417 set_task_cpu(p, dest_cpu);
5418 activate_task(rq_dest, p, 0);
5419 check_preempt_curr(rq_dest, p, 0);
5424 double_rq_unlock(rq_src, rq_dest);
5429 * migration_cpu_stop - this will be executed by a highprio stopper thread
5430 * and performs thread migration by bumping thread off CPU then
5431 * 'pushing' onto another runqueue.
5433 static int migration_cpu_stop(void *data)
5435 struct migration_arg *arg = data;
5438 * The original target cpu might have gone down and we might
5439 * be on another cpu but it doesn't matter.
5441 local_irq_disable();
5442 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5447 #ifdef CONFIG_HOTPLUG_CPU
5449 * Figure out where task on dead CPU should go, use force if necessary.
5451 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5453 struct rq *rq = cpu_rq(dead_cpu);
5454 int needs_cpu, uninitialized_var(dest_cpu);
5455 unsigned long flags;
5457 local_irq_save(flags);
5459 raw_spin_lock(&rq->lock);
5460 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5462 dest_cpu = select_fallback_rq(dead_cpu, p);
5463 raw_spin_unlock(&rq->lock);
5465 * It can only fail if we race with set_cpus_allowed(),
5466 * in the racer should migrate the task anyway.
5469 __migrate_task(p, dead_cpu, dest_cpu);
5470 local_irq_restore(flags);
5474 * While a dead CPU has no uninterruptible tasks queued at this point,
5475 * it might still have a nonzero ->nr_uninterruptible counter, because
5476 * for performance reasons the counter is not stricly tracking tasks to
5477 * their home CPUs. So we just add the counter to another CPU's counter,
5478 * to keep the global sum constant after CPU-down:
5480 static void migrate_nr_uninterruptible(struct rq *rq_src)
5482 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5483 unsigned long flags;
5485 local_irq_save(flags);
5486 double_rq_lock(rq_src, rq_dest);
5487 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5488 rq_src->nr_uninterruptible = 0;
5489 double_rq_unlock(rq_src, rq_dest);
5490 local_irq_restore(flags);
5493 /* Run through task list and migrate tasks from the dead cpu. */
5494 static void migrate_live_tasks(int src_cpu)
5496 struct task_struct *p, *t;
5498 read_lock(&tasklist_lock);
5500 do_each_thread(t, p) {
5504 if (task_cpu(p) == src_cpu)
5505 move_task_off_dead_cpu(src_cpu, p);
5506 } while_each_thread(t, p);
5508 read_unlock(&tasklist_lock);
5512 * Schedules idle task to be the next runnable task on current CPU.
5513 * It does so by boosting its priority to highest possible.
5514 * Used by CPU offline code.
5516 void sched_idle_next(void)
5518 int this_cpu = smp_processor_id();
5519 struct rq *rq = cpu_rq(this_cpu);
5520 struct task_struct *p = rq->idle;
5521 unsigned long flags;
5523 /* cpu has to be offline */
5524 BUG_ON(cpu_online(this_cpu));
5527 * Strictly not necessary since rest of the CPUs are stopped by now
5528 * and interrupts disabled on the current cpu.
5530 raw_spin_lock_irqsave(&rq->lock, flags);
5532 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5534 activate_task(rq, p, 0);
5536 raw_spin_unlock_irqrestore(&rq->lock, flags);
5540 * Ensures that the idle task is using init_mm right before its cpu goes
5543 void idle_task_exit(void)
5545 struct mm_struct *mm = current->active_mm;
5547 BUG_ON(cpu_online(smp_processor_id()));
5550 switch_mm(mm, &init_mm, current);
5554 /* called under rq->lock with disabled interrupts */
5555 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5557 struct rq *rq = cpu_rq(dead_cpu);
5559 /* Must be exiting, otherwise would be on tasklist. */
5560 BUG_ON(!p->exit_state);
5562 /* Cannot have done final schedule yet: would have vanished. */
5563 BUG_ON(p->state == TASK_DEAD);
5568 * Drop lock around migration; if someone else moves it,
5569 * that's OK. No task can be added to this CPU, so iteration is
5572 raw_spin_unlock_irq(&rq->lock);
5573 move_task_off_dead_cpu(dead_cpu, p);
5574 raw_spin_lock_irq(&rq->lock);
5579 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5580 static void migrate_dead_tasks(unsigned int dead_cpu)
5582 struct rq *rq = cpu_rq(dead_cpu);
5583 struct task_struct *next;
5586 if (!rq->nr_running)
5588 next = pick_next_task(rq);
5591 next->sched_class->put_prev_task(rq, next);
5592 migrate_dead(dead_cpu, next);
5598 * remove the tasks which were accounted by rq from calc_load_tasks.
5600 static void calc_global_load_remove(struct rq *rq)
5602 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5603 rq->calc_load_active = 0;
5605 #endif /* CONFIG_HOTPLUG_CPU */
5607 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5609 static struct ctl_table sd_ctl_dir[] = {
5611 .procname = "sched_domain",
5617 static struct ctl_table sd_ctl_root[] = {
5619 .procname = "kernel",
5621 .child = sd_ctl_dir,
5626 static struct ctl_table *sd_alloc_ctl_entry(int n)
5628 struct ctl_table *entry =
5629 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5634 static void sd_free_ctl_entry(struct ctl_table **tablep)
5636 struct ctl_table *entry;
5639 * In the intermediate directories, both the child directory and
5640 * procname are dynamically allocated and could fail but the mode
5641 * will always be set. In the lowest directory the names are
5642 * static strings and all have proc handlers.
5644 for (entry = *tablep; entry->mode; entry++) {
5646 sd_free_ctl_entry(&entry->child);
5647 if (entry->proc_handler == NULL)
5648 kfree(entry->procname);
5656 set_table_entry(struct ctl_table *entry,
5657 const char *procname, void *data, int maxlen,
5658 mode_t mode, proc_handler *proc_handler)
5660 entry->procname = procname;
5662 entry->maxlen = maxlen;
5664 entry->proc_handler = proc_handler;
5667 static struct ctl_table *
5668 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5670 struct ctl_table *table = sd_alloc_ctl_entry(13);
5675 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5676 sizeof(long), 0644, proc_doulongvec_minmax);
5677 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5678 sizeof(long), 0644, proc_doulongvec_minmax);
5679 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5680 sizeof(int), 0644, proc_dointvec_minmax);
5681 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5682 sizeof(int), 0644, proc_dointvec_minmax);
5683 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5684 sizeof(int), 0644, proc_dointvec_minmax);
5685 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5686 sizeof(int), 0644, proc_dointvec_minmax);
5687 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5688 sizeof(int), 0644, proc_dointvec_minmax);
5689 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5690 sizeof(int), 0644, proc_dointvec_minmax);
5691 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5692 sizeof(int), 0644, proc_dointvec_minmax);
5693 set_table_entry(&table[9], "cache_nice_tries",
5694 &sd->cache_nice_tries,
5695 sizeof(int), 0644, proc_dointvec_minmax);
5696 set_table_entry(&table[10], "flags", &sd->flags,
5697 sizeof(int), 0644, proc_dointvec_minmax);
5698 set_table_entry(&table[11], "name", sd->name,
5699 CORENAME_MAX_SIZE, 0444, proc_dostring);
5700 /* &table[12] is terminator */
5705 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5707 struct ctl_table *entry, *table;
5708 struct sched_domain *sd;
5709 int domain_num = 0, i;
5712 for_each_domain(cpu, sd)
5714 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5719 for_each_domain(cpu, sd) {
5720 snprintf(buf, 32, "domain%d", i);
5721 entry->procname = kstrdup(buf, GFP_KERNEL);
5723 entry->child = sd_alloc_ctl_domain_table(sd);
5730 static struct ctl_table_header *sd_sysctl_header;
5731 static void register_sched_domain_sysctl(void)
5733 int i, cpu_num = num_possible_cpus();
5734 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5737 WARN_ON(sd_ctl_dir[0].child);
5738 sd_ctl_dir[0].child = entry;
5743 for_each_possible_cpu(i) {
5744 snprintf(buf, 32, "cpu%d", i);
5745 entry->procname = kstrdup(buf, GFP_KERNEL);
5747 entry->child = sd_alloc_ctl_cpu_table(i);
5751 WARN_ON(sd_sysctl_header);
5752 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5755 /* may be called multiple times per register */
5756 static void unregister_sched_domain_sysctl(void)
5758 if (sd_sysctl_header)
5759 unregister_sysctl_table(sd_sysctl_header);
5760 sd_sysctl_header = NULL;
5761 if (sd_ctl_dir[0].child)
5762 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5765 static void register_sched_domain_sysctl(void)
5768 static void unregister_sched_domain_sysctl(void)
5773 static void set_rq_online(struct rq *rq)
5776 const struct sched_class *class;
5778 cpumask_set_cpu(rq->cpu, rq->rd->online);
5781 for_each_class(class) {
5782 if (class->rq_online)
5783 class->rq_online(rq);
5788 static void set_rq_offline(struct rq *rq)
5791 const struct sched_class *class;
5793 for_each_class(class) {
5794 if (class->rq_offline)
5795 class->rq_offline(rq);
5798 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5804 * migration_call - callback that gets triggered when a CPU is added.
5805 * Here we can start up the necessary migration thread for the new CPU.
5807 static int __cpuinit
5808 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5810 int cpu = (long)hcpu;
5811 unsigned long flags;
5812 struct rq *rq = cpu_rq(cpu);
5816 case CPU_UP_PREPARE:
5817 case CPU_UP_PREPARE_FROZEN:
5818 rq->calc_load_update = calc_load_update;
5822 case CPU_ONLINE_FROZEN:
5823 /* Update our root-domain */
5824 raw_spin_lock_irqsave(&rq->lock, flags);
5826 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5830 raw_spin_unlock_irqrestore(&rq->lock, flags);
5833 #ifdef CONFIG_HOTPLUG_CPU
5835 case CPU_DEAD_FROZEN:
5836 migrate_live_tasks(cpu);
5837 /* Idle task back to normal (off runqueue, low prio) */
5838 raw_spin_lock_irq(&rq->lock);
5839 deactivate_task(rq, rq->idle, 0);
5840 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5841 rq->idle->sched_class = &idle_sched_class;
5842 migrate_dead_tasks(cpu);
5843 raw_spin_unlock_irq(&rq->lock);
5844 migrate_nr_uninterruptible(rq);
5845 BUG_ON(rq->nr_running != 0);
5846 calc_global_load_remove(rq);
5850 case CPU_DYING_FROZEN:
5851 /* Update our root-domain */
5852 raw_spin_lock_irqsave(&rq->lock, flags);
5854 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5857 raw_spin_unlock_irqrestore(&rq->lock, flags);
5865 * Register at high priority so that task migration (migrate_all_tasks)
5866 * happens before everything else. This has to be lower priority than
5867 * the notifier in the perf_event subsystem, though.
5869 static struct notifier_block __cpuinitdata migration_notifier = {
5870 .notifier_call = migration_call,
5871 .priority = CPU_PRI_MIGRATION,
5874 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5875 unsigned long action, void *hcpu)
5877 switch (action & ~CPU_TASKS_FROZEN) {
5879 case CPU_DOWN_FAILED:
5880 set_cpu_active((long)hcpu, true);
5887 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5888 unsigned long action, void *hcpu)
5890 switch (action & ~CPU_TASKS_FROZEN) {
5891 case CPU_DOWN_PREPARE:
5892 set_cpu_active((long)hcpu, false);
5899 static int __init migration_init(void)
5901 void *cpu = (void *)(long)smp_processor_id();
5904 /* Initialize migration for the boot CPU */
5905 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5906 BUG_ON(err == NOTIFY_BAD);
5907 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5908 register_cpu_notifier(&migration_notifier);
5910 /* Register cpu active notifiers */
5911 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5912 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5916 early_initcall(migration_init);
5921 #ifdef CONFIG_SCHED_DEBUG
5923 static __read_mostly int sched_domain_debug_enabled;
5925 static int __init sched_domain_debug_setup(char *str)
5927 sched_domain_debug_enabled = 1;
5931 early_param("sched_debug", sched_domain_debug_setup);
5933 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5934 struct cpumask *groupmask)
5936 struct sched_group *group = sd->groups;
5939 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5940 cpumask_clear(groupmask);
5942 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5944 if (!(sd->flags & SD_LOAD_BALANCE)) {
5945 printk("does not load-balance\n");
5947 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5952 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5954 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5955 printk(KERN_ERR "ERROR: domain->span does not contain "
5958 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5959 printk(KERN_ERR "ERROR: domain->groups does not contain"
5963 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5967 printk(KERN_ERR "ERROR: group is NULL\n");
5971 if (!group->cpu_power) {
5972 printk(KERN_CONT "\n");
5973 printk(KERN_ERR "ERROR: domain->cpu_power not "
5978 if (!cpumask_weight(sched_group_cpus(group))) {
5979 printk(KERN_CONT "\n");
5980 printk(KERN_ERR "ERROR: empty group\n");
5984 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5985 printk(KERN_CONT "\n");
5986 printk(KERN_ERR "ERROR: repeated CPUs\n");
5990 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5992 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5994 printk(KERN_CONT " %s", str);
5995 if (group->cpu_power != SCHED_LOAD_SCALE) {
5996 printk(KERN_CONT " (cpu_power = %d)",
6000 group = group->next;
6001 } while (group != sd->groups);
6002 printk(KERN_CONT "\n");
6004 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6005 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6008 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6009 printk(KERN_ERR "ERROR: parent span is not a superset "
6010 "of domain->span\n");
6014 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6016 cpumask_var_t groupmask;
6019 if (!sched_domain_debug_enabled)
6023 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6027 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6029 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6030 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6035 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6042 free_cpumask_var(groupmask);
6044 #else /* !CONFIG_SCHED_DEBUG */
6045 # define sched_domain_debug(sd, cpu) do { } while (0)
6046 #endif /* CONFIG_SCHED_DEBUG */
6048 static int sd_degenerate(struct sched_domain *sd)
6050 if (cpumask_weight(sched_domain_span(sd)) == 1)
6053 /* Following flags need at least 2 groups */
6054 if (sd->flags & (SD_LOAD_BALANCE |
6055 SD_BALANCE_NEWIDLE |
6059 SD_SHARE_PKG_RESOURCES)) {
6060 if (sd->groups != sd->groups->next)
6064 /* Following flags don't use groups */
6065 if (sd->flags & (SD_WAKE_AFFINE))
6072 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6074 unsigned long cflags = sd->flags, pflags = parent->flags;
6076 if (sd_degenerate(parent))
6079 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6082 /* Flags needing groups don't count if only 1 group in parent */
6083 if (parent->groups == parent->groups->next) {
6084 pflags &= ~(SD_LOAD_BALANCE |
6085 SD_BALANCE_NEWIDLE |
6089 SD_SHARE_PKG_RESOURCES);
6090 if (nr_node_ids == 1)
6091 pflags &= ~SD_SERIALIZE;
6093 if (~cflags & pflags)
6099 static void free_rootdomain(struct root_domain *rd)
6101 synchronize_sched();
6103 cpupri_cleanup(&rd->cpupri);
6105 free_cpumask_var(rd->rto_mask);
6106 free_cpumask_var(rd->online);
6107 free_cpumask_var(rd->span);
6111 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6113 struct root_domain *old_rd = NULL;
6114 unsigned long flags;
6116 raw_spin_lock_irqsave(&rq->lock, flags);
6121 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6124 cpumask_clear_cpu(rq->cpu, old_rd->span);
6127 * If we dont want to free the old_rt yet then
6128 * set old_rd to NULL to skip the freeing later
6131 if (!atomic_dec_and_test(&old_rd->refcount))
6135 atomic_inc(&rd->refcount);
6138 cpumask_set_cpu(rq->cpu, rd->span);
6139 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6142 raw_spin_unlock_irqrestore(&rq->lock, flags);
6145 free_rootdomain(old_rd);
6148 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6150 gfp_t gfp = GFP_KERNEL;
6152 memset(rd, 0, sizeof(*rd));
6157 if (!alloc_cpumask_var(&rd->span, gfp))
6159 if (!alloc_cpumask_var(&rd->online, gfp))
6161 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6164 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6169 free_cpumask_var(rd->rto_mask);
6171 free_cpumask_var(rd->online);
6173 free_cpumask_var(rd->span);
6178 static void init_defrootdomain(void)
6180 init_rootdomain(&def_root_domain, true);
6182 atomic_set(&def_root_domain.refcount, 1);
6185 static struct root_domain *alloc_rootdomain(void)
6187 struct root_domain *rd;
6189 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6193 if (init_rootdomain(rd, false) != 0) {
6202 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6203 * hold the hotplug lock.
6206 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6208 struct rq *rq = cpu_rq(cpu);
6209 struct sched_domain *tmp;
6211 for (tmp = sd; tmp; tmp = tmp->parent)
6212 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6214 /* Remove the sched domains which do not contribute to scheduling. */
6215 for (tmp = sd; tmp; ) {
6216 struct sched_domain *parent = tmp->parent;
6220 if (sd_parent_degenerate(tmp, parent)) {
6221 tmp->parent = parent->parent;
6223 parent->parent->child = tmp;
6228 if (sd && sd_degenerate(sd)) {
6234 sched_domain_debug(sd, cpu);
6236 rq_attach_root(rq, rd);
6237 rcu_assign_pointer(rq->sd, sd);
6240 /* cpus with isolated domains */
6241 static cpumask_var_t cpu_isolated_map;
6243 /* Setup the mask of cpus configured for isolated domains */
6244 static int __init isolated_cpu_setup(char *str)
6246 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6247 cpulist_parse(str, cpu_isolated_map);
6251 __setup("isolcpus=", isolated_cpu_setup);
6254 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6255 * to a function which identifies what group(along with sched group) a CPU
6256 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6257 * (due to the fact that we keep track of groups covered with a struct cpumask).
6259 * init_sched_build_groups will build a circular linked list of the groups
6260 * covered by the given span, and will set each group's ->cpumask correctly,
6261 * and ->cpu_power to 0.
6264 init_sched_build_groups(const struct cpumask *span,
6265 const struct cpumask *cpu_map,
6266 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6267 struct sched_group **sg,
6268 struct cpumask *tmpmask),
6269 struct cpumask *covered, struct cpumask *tmpmask)
6271 struct sched_group *first = NULL, *last = NULL;
6274 cpumask_clear(covered);
6276 for_each_cpu(i, span) {
6277 struct sched_group *sg;
6278 int group = group_fn(i, cpu_map, &sg, tmpmask);
6281 if (cpumask_test_cpu(i, covered))
6284 cpumask_clear(sched_group_cpus(sg));
6287 for_each_cpu(j, span) {
6288 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6291 cpumask_set_cpu(j, covered);
6292 cpumask_set_cpu(j, sched_group_cpus(sg));
6303 #define SD_NODES_PER_DOMAIN 16
6308 * find_next_best_node - find the next node to include in a sched_domain
6309 * @node: node whose sched_domain we're building
6310 * @used_nodes: nodes already in the sched_domain
6312 * Find the next node to include in a given scheduling domain. Simply
6313 * finds the closest node not already in the @used_nodes map.
6315 * Should use nodemask_t.
6317 static int find_next_best_node(int node, nodemask_t *used_nodes)
6319 int i, n, val, min_val, best_node = 0;
6323 for (i = 0; i < nr_node_ids; i++) {
6324 /* Start at @node */
6325 n = (node + i) % nr_node_ids;
6327 if (!nr_cpus_node(n))
6330 /* Skip already used nodes */
6331 if (node_isset(n, *used_nodes))
6334 /* Simple min distance search */
6335 val = node_distance(node, n);
6337 if (val < min_val) {
6343 node_set(best_node, *used_nodes);
6348 * sched_domain_node_span - get a cpumask for a node's sched_domain
6349 * @node: node whose cpumask we're constructing
6350 * @span: resulting cpumask
6352 * Given a node, construct a good cpumask for its sched_domain to span. It
6353 * should be one that prevents unnecessary balancing, but also spreads tasks
6356 static void sched_domain_node_span(int node, struct cpumask *span)
6358 nodemask_t used_nodes;
6361 cpumask_clear(span);
6362 nodes_clear(used_nodes);
6364 cpumask_or(span, span, cpumask_of_node(node));
6365 node_set(node, used_nodes);
6367 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6368 int next_node = find_next_best_node(node, &used_nodes);
6370 cpumask_or(span, span, cpumask_of_node(next_node));
6373 #endif /* CONFIG_NUMA */
6375 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6378 * The cpus mask in sched_group and sched_domain hangs off the end.
6380 * ( See the the comments in include/linux/sched.h:struct sched_group
6381 * and struct sched_domain. )
6383 struct static_sched_group {
6384 struct sched_group sg;
6385 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6388 struct static_sched_domain {
6389 struct sched_domain sd;
6390 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6396 cpumask_var_t domainspan;
6397 cpumask_var_t covered;
6398 cpumask_var_t notcovered;
6400 cpumask_var_t nodemask;
6401 cpumask_var_t this_sibling_map;
6402 cpumask_var_t this_core_map;
6403 cpumask_var_t send_covered;
6404 cpumask_var_t tmpmask;
6405 struct sched_group **sched_group_nodes;
6406 struct root_domain *rd;
6410 sa_sched_groups = 0,
6415 sa_this_sibling_map,
6417 sa_sched_group_nodes,
6427 * SMT sched-domains:
6429 #ifdef CONFIG_SCHED_SMT
6430 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6431 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6434 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6435 struct sched_group **sg, struct cpumask *unused)
6438 *sg = &per_cpu(sched_groups, cpu).sg;
6441 #endif /* CONFIG_SCHED_SMT */
6444 * multi-core sched-domains:
6446 #ifdef CONFIG_SCHED_MC
6447 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6448 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6449 #endif /* CONFIG_SCHED_MC */
6451 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6453 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6454 struct sched_group **sg, struct cpumask *mask)
6458 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6459 group = cpumask_first(mask);
6461 *sg = &per_cpu(sched_group_core, group).sg;
6464 #elif defined(CONFIG_SCHED_MC)
6466 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6467 struct sched_group **sg, struct cpumask *unused)
6470 *sg = &per_cpu(sched_group_core, cpu).sg;
6475 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6476 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6479 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6480 struct sched_group **sg, struct cpumask *mask)
6483 #ifdef CONFIG_SCHED_MC
6484 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6485 group = cpumask_first(mask);
6486 #elif defined(CONFIG_SCHED_SMT)
6487 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6488 group = cpumask_first(mask);
6493 *sg = &per_cpu(sched_group_phys, group).sg;
6499 * The init_sched_build_groups can't handle what we want to do with node
6500 * groups, so roll our own. Now each node has its own list of groups which
6501 * gets dynamically allocated.
6503 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6504 static struct sched_group ***sched_group_nodes_bycpu;
6506 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6507 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6509 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6510 struct sched_group **sg,
6511 struct cpumask *nodemask)
6515 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6516 group = cpumask_first(nodemask);
6519 *sg = &per_cpu(sched_group_allnodes, group).sg;
6523 static void init_numa_sched_groups_power(struct sched_group *group_head)
6525 struct sched_group *sg = group_head;
6531 for_each_cpu(j, sched_group_cpus(sg)) {
6532 struct sched_domain *sd;
6534 sd = &per_cpu(phys_domains, j).sd;
6535 if (j != group_first_cpu(sd->groups)) {
6537 * Only add "power" once for each
6543 sg->cpu_power += sd->groups->cpu_power;
6546 } while (sg != group_head);
6549 static int build_numa_sched_groups(struct s_data *d,
6550 const struct cpumask *cpu_map, int num)
6552 struct sched_domain *sd;
6553 struct sched_group *sg, *prev;
6556 cpumask_clear(d->covered);
6557 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6558 if (cpumask_empty(d->nodemask)) {
6559 d->sched_group_nodes[num] = NULL;
6563 sched_domain_node_span(num, d->domainspan);
6564 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6566 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6569 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6573 d->sched_group_nodes[num] = sg;
6575 for_each_cpu(j, d->nodemask) {
6576 sd = &per_cpu(node_domains, j).sd;
6581 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6583 cpumask_or(d->covered, d->covered, d->nodemask);
6586 for (j = 0; j < nr_node_ids; j++) {
6587 n = (num + j) % nr_node_ids;
6588 cpumask_complement(d->notcovered, d->covered);
6589 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6590 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6591 if (cpumask_empty(d->tmpmask))
6593 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6594 if (cpumask_empty(d->tmpmask))
6596 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6600 "Can not alloc domain group for node %d\n", j);
6604 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6605 sg->next = prev->next;
6606 cpumask_or(d->covered, d->covered, d->tmpmask);
6613 #endif /* CONFIG_NUMA */
6616 /* Free memory allocated for various sched_group structures */
6617 static void free_sched_groups(const struct cpumask *cpu_map,
6618 struct cpumask *nodemask)
6622 for_each_cpu(cpu, cpu_map) {
6623 struct sched_group **sched_group_nodes
6624 = sched_group_nodes_bycpu[cpu];
6626 if (!sched_group_nodes)
6629 for (i = 0; i < nr_node_ids; i++) {
6630 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6632 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6633 if (cpumask_empty(nodemask))
6643 if (oldsg != sched_group_nodes[i])
6646 kfree(sched_group_nodes);
6647 sched_group_nodes_bycpu[cpu] = NULL;
6650 #else /* !CONFIG_NUMA */
6651 static void free_sched_groups(const struct cpumask *cpu_map,
6652 struct cpumask *nodemask)
6655 #endif /* CONFIG_NUMA */
6658 * Initialize sched groups cpu_power.
6660 * cpu_power indicates the capacity of sched group, which is used while
6661 * distributing the load between different sched groups in a sched domain.
6662 * Typically cpu_power for all the groups in a sched domain will be same unless
6663 * there are asymmetries in the topology. If there are asymmetries, group
6664 * having more cpu_power will pickup more load compared to the group having
6667 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6669 struct sched_domain *child;
6670 struct sched_group *group;
6674 WARN_ON(!sd || !sd->groups);
6676 if (cpu != group_first_cpu(sd->groups))
6681 sd->groups->cpu_power = 0;
6684 power = SCHED_LOAD_SCALE;
6685 weight = cpumask_weight(sched_domain_span(sd));
6687 * SMT siblings share the power of a single core.
6688 * Usually multiple threads get a better yield out of
6689 * that one core than a single thread would have,
6690 * reflect that in sd->smt_gain.
6692 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6693 power *= sd->smt_gain;
6695 power >>= SCHED_LOAD_SHIFT;
6697 sd->groups->cpu_power += power;
6702 * Add cpu_power of each child group to this groups cpu_power.
6704 group = child->groups;
6706 sd->groups->cpu_power += group->cpu_power;
6707 group = group->next;
6708 } while (group != child->groups);
6712 * Initializers for schedule domains
6713 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6716 #ifdef CONFIG_SCHED_DEBUG
6717 # define SD_INIT_NAME(sd, type) sd->name = #type
6719 # define SD_INIT_NAME(sd, type) do { } while (0)
6722 #define SD_INIT(sd, type) sd_init_##type(sd)
6724 #define SD_INIT_FUNC(type) \
6725 static noinline void sd_init_##type(struct sched_domain *sd) \
6727 memset(sd, 0, sizeof(*sd)); \
6728 *sd = SD_##type##_INIT; \
6729 sd->level = SD_LV_##type; \
6730 SD_INIT_NAME(sd, type); \
6735 SD_INIT_FUNC(ALLNODES)
6738 #ifdef CONFIG_SCHED_SMT
6739 SD_INIT_FUNC(SIBLING)
6741 #ifdef CONFIG_SCHED_MC
6745 static int default_relax_domain_level = -1;
6747 static int __init setup_relax_domain_level(char *str)
6751 val = simple_strtoul(str, NULL, 0);
6752 if (val < SD_LV_MAX)
6753 default_relax_domain_level = val;
6757 __setup("relax_domain_level=", setup_relax_domain_level);
6759 static void set_domain_attribute(struct sched_domain *sd,
6760 struct sched_domain_attr *attr)
6764 if (!attr || attr->relax_domain_level < 0) {
6765 if (default_relax_domain_level < 0)
6768 request = default_relax_domain_level;
6770 request = attr->relax_domain_level;
6771 if (request < sd->level) {
6772 /* turn off idle balance on this domain */
6773 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6775 /* turn on idle balance on this domain */
6776 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6780 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6781 const struct cpumask *cpu_map)
6784 case sa_sched_groups:
6785 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6786 d->sched_group_nodes = NULL;
6788 free_rootdomain(d->rd); /* fall through */
6790 free_cpumask_var(d->tmpmask); /* fall through */
6791 case sa_send_covered:
6792 free_cpumask_var(d->send_covered); /* fall through */
6793 case sa_this_core_map:
6794 free_cpumask_var(d->this_core_map); /* fall through */
6795 case sa_this_sibling_map:
6796 free_cpumask_var(d->this_sibling_map); /* fall through */
6798 free_cpumask_var(d->nodemask); /* fall through */
6799 case sa_sched_group_nodes:
6801 kfree(d->sched_group_nodes); /* fall through */
6803 free_cpumask_var(d->notcovered); /* fall through */
6805 free_cpumask_var(d->covered); /* fall through */
6807 free_cpumask_var(d->domainspan); /* fall through */
6814 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6815 const struct cpumask *cpu_map)
6818 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6820 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6821 return sa_domainspan;
6822 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6824 /* Allocate the per-node list of sched groups */
6825 d->sched_group_nodes = kcalloc(nr_node_ids,
6826 sizeof(struct sched_group *), GFP_KERNEL);
6827 if (!d->sched_group_nodes) {
6828 printk(KERN_WARNING "Can not alloc sched group node list\n");
6829 return sa_notcovered;
6831 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6833 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6834 return sa_sched_group_nodes;
6835 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6837 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6838 return sa_this_sibling_map;
6839 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6840 return sa_this_core_map;
6841 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6842 return sa_send_covered;
6843 d->rd = alloc_rootdomain();
6845 printk(KERN_WARNING "Cannot alloc root domain\n");
6848 return sa_rootdomain;
6851 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6852 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6854 struct sched_domain *sd = NULL;
6856 struct sched_domain *parent;
6859 if (cpumask_weight(cpu_map) >
6860 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6861 sd = &per_cpu(allnodes_domains, i).sd;
6862 SD_INIT(sd, ALLNODES);
6863 set_domain_attribute(sd, attr);
6864 cpumask_copy(sched_domain_span(sd), cpu_map);
6865 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6870 sd = &per_cpu(node_domains, i).sd;
6872 set_domain_attribute(sd, attr);
6873 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6874 sd->parent = parent;
6877 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6882 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6883 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6884 struct sched_domain *parent, int i)
6886 struct sched_domain *sd;
6887 sd = &per_cpu(phys_domains, i).sd;
6889 set_domain_attribute(sd, attr);
6890 cpumask_copy(sched_domain_span(sd), d->nodemask);
6891 sd->parent = parent;
6894 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6898 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6899 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6900 struct sched_domain *parent, int i)
6902 struct sched_domain *sd = parent;
6903 #ifdef CONFIG_SCHED_MC
6904 sd = &per_cpu(core_domains, i).sd;
6906 set_domain_attribute(sd, attr);
6907 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6908 sd->parent = parent;
6910 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6915 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6916 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6917 struct sched_domain *parent, int i)
6919 struct sched_domain *sd = parent;
6920 #ifdef CONFIG_SCHED_SMT
6921 sd = &per_cpu(cpu_domains, i).sd;
6922 SD_INIT(sd, SIBLING);
6923 set_domain_attribute(sd, attr);
6924 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6925 sd->parent = parent;
6927 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6932 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6933 const struct cpumask *cpu_map, int cpu)
6936 #ifdef CONFIG_SCHED_SMT
6937 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6938 cpumask_and(d->this_sibling_map, cpu_map,
6939 topology_thread_cpumask(cpu));
6940 if (cpu == cpumask_first(d->this_sibling_map))
6941 init_sched_build_groups(d->this_sibling_map, cpu_map,
6943 d->send_covered, d->tmpmask);
6946 #ifdef CONFIG_SCHED_MC
6947 case SD_LV_MC: /* set up multi-core groups */
6948 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6949 if (cpu == cpumask_first(d->this_core_map))
6950 init_sched_build_groups(d->this_core_map, cpu_map,
6952 d->send_covered, d->tmpmask);
6955 case SD_LV_CPU: /* set up physical groups */
6956 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
6957 if (!cpumask_empty(d->nodemask))
6958 init_sched_build_groups(d->nodemask, cpu_map,
6960 d->send_covered, d->tmpmask);
6963 case SD_LV_ALLNODES:
6964 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
6965 d->send_covered, d->tmpmask);
6974 * Build sched domains for a given set of cpus and attach the sched domains
6975 * to the individual cpus
6977 static int __build_sched_domains(const struct cpumask *cpu_map,
6978 struct sched_domain_attr *attr)
6980 enum s_alloc alloc_state = sa_none;
6982 struct sched_domain *sd;
6988 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6989 if (alloc_state != sa_rootdomain)
6991 alloc_state = sa_sched_groups;
6994 * Set up domains for cpus specified by the cpu_map.
6996 for_each_cpu(i, cpu_map) {
6997 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7000 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7001 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7002 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7003 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7006 for_each_cpu(i, cpu_map) {
7007 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7008 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7011 /* Set up physical groups */
7012 for (i = 0; i < nr_node_ids; i++)
7013 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7016 /* Set up node groups */
7018 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7020 for (i = 0; i < nr_node_ids; i++)
7021 if (build_numa_sched_groups(&d, cpu_map, i))
7025 /* Calculate CPU power for physical packages and nodes */
7026 #ifdef CONFIG_SCHED_SMT
7027 for_each_cpu(i, cpu_map) {
7028 sd = &per_cpu(cpu_domains, i).sd;
7029 init_sched_groups_power(i, sd);
7032 #ifdef CONFIG_SCHED_MC
7033 for_each_cpu(i, cpu_map) {
7034 sd = &per_cpu(core_domains, i).sd;
7035 init_sched_groups_power(i, sd);
7039 for_each_cpu(i, cpu_map) {
7040 sd = &per_cpu(phys_domains, i).sd;
7041 init_sched_groups_power(i, sd);
7045 for (i = 0; i < nr_node_ids; i++)
7046 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7048 if (d.sd_allnodes) {
7049 struct sched_group *sg;
7051 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7053 init_numa_sched_groups_power(sg);
7057 /* Attach the domains */
7058 for_each_cpu(i, cpu_map) {
7059 #ifdef CONFIG_SCHED_SMT
7060 sd = &per_cpu(cpu_domains, i).sd;
7061 #elif defined(CONFIG_SCHED_MC)
7062 sd = &per_cpu(core_domains, i).sd;
7064 sd = &per_cpu(phys_domains, i).sd;
7066 cpu_attach_domain(sd, d.rd, i);
7069 d.sched_group_nodes = NULL; /* don't free this we still need it */
7070 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7074 __free_domain_allocs(&d, alloc_state, cpu_map);
7078 static int build_sched_domains(const struct cpumask *cpu_map)
7080 return __build_sched_domains(cpu_map, NULL);
7083 static cpumask_var_t *doms_cur; /* current sched domains */
7084 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7085 static struct sched_domain_attr *dattr_cur;
7086 /* attribues of custom domains in 'doms_cur' */
7089 * Special case: If a kmalloc of a doms_cur partition (array of
7090 * cpumask) fails, then fallback to a single sched domain,
7091 * as determined by the single cpumask fallback_doms.
7093 static cpumask_var_t fallback_doms;
7096 * arch_update_cpu_topology lets virtualized architectures update the
7097 * cpu core maps. It is supposed to return 1 if the topology changed
7098 * or 0 if it stayed the same.
7100 int __attribute__((weak)) arch_update_cpu_topology(void)
7105 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7108 cpumask_var_t *doms;
7110 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7113 for (i = 0; i < ndoms; i++) {
7114 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7115 free_sched_domains(doms, i);
7122 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7125 for (i = 0; i < ndoms; i++)
7126 free_cpumask_var(doms[i]);
7131 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7132 * For now this just excludes isolated cpus, but could be used to
7133 * exclude other special cases in the future.
7135 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7139 arch_update_cpu_topology();
7141 doms_cur = alloc_sched_domains(ndoms_cur);
7143 doms_cur = &fallback_doms;
7144 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7146 err = build_sched_domains(doms_cur[0]);
7147 register_sched_domain_sysctl();
7152 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7153 struct cpumask *tmpmask)
7155 free_sched_groups(cpu_map, tmpmask);
7159 * Detach sched domains from a group of cpus specified in cpu_map
7160 * These cpus will now be attached to the NULL domain
7162 static void detach_destroy_domains(const struct cpumask *cpu_map)
7164 /* Save because hotplug lock held. */
7165 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7168 for_each_cpu(i, cpu_map)
7169 cpu_attach_domain(NULL, &def_root_domain, i);
7170 synchronize_sched();
7171 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7174 /* handle null as "default" */
7175 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7176 struct sched_domain_attr *new, int idx_new)
7178 struct sched_domain_attr tmp;
7185 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7186 new ? (new + idx_new) : &tmp,
7187 sizeof(struct sched_domain_attr));
7191 * Partition sched domains as specified by the 'ndoms_new'
7192 * cpumasks in the array doms_new[] of cpumasks. This compares
7193 * doms_new[] to the current sched domain partitioning, doms_cur[].
7194 * It destroys each deleted domain and builds each new domain.
7196 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7197 * The masks don't intersect (don't overlap.) We should setup one
7198 * sched domain for each mask. CPUs not in any of the cpumasks will
7199 * not be load balanced. If the same cpumask appears both in the
7200 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7203 * The passed in 'doms_new' should be allocated using
7204 * alloc_sched_domains. This routine takes ownership of it and will
7205 * free_sched_domains it when done with it. If the caller failed the
7206 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7207 * and partition_sched_domains() will fallback to the single partition
7208 * 'fallback_doms', it also forces the domains to be rebuilt.
7210 * If doms_new == NULL it will be replaced with cpu_online_mask.
7211 * ndoms_new == 0 is a special case for destroying existing domains,
7212 * and it will not create the default domain.
7214 * Call with hotplug lock held
7216 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7217 struct sched_domain_attr *dattr_new)
7222 mutex_lock(&sched_domains_mutex);
7224 /* always unregister in case we don't destroy any domains */
7225 unregister_sched_domain_sysctl();
7227 /* Let architecture update cpu core mappings. */
7228 new_topology = arch_update_cpu_topology();
7230 n = doms_new ? ndoms_new : 0;
7232 /* Destroy deleted domains */
7233 for (i = 0; i < ndoms_cur; i++) {
7234 for (j = 0; j < n && !new_topology; j++) {
7235 if (cpumask_equal(doms_cur[i], doms_new[j])
7236 && dattrs_equal(dattr_cur, i, dattr_new, j))
7239 /* no match - a current sched domain not in new doms_new[] */
7240 detach_destroy_domains(doms_cur[i]);
7245 if (doms_new == NULL) {
7247 doms_new = &fallback_doms;
7248 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7249 WARN_ON_ONCE(dattr_new);
7252 /* Build new domains */
7253 for (i = 0; i < ndoms_new; i++) {
7254 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7255 if (cpumask_equal(doms_new[i], doms_cur[j])
7256 && dattrs_equal(dattr_new, i, dattr_cur, j))
7259 /* no match - add a new doms_new */
7260 __build_sched_domains(doms_new[i],
7261 dattr_new ? dattr_new + i : NULL);
7266 /* Remember the new sched domains */
7267 if (doms_cur != &fallback_doms)
7268 free_sched_domains(doms_cur, ndoms_cur);
7269 kfree(dattr_cur); /* kfree(NULL) is safe */
7270 doms_cur = doms_new;
7271 dattr_cur = dattr_new;
7272 ndoms_cur = ndoms_new;
7274 register_sched_domain_sysctl();
7276 mutex_unlock(&sched_domains_mutex);
7279 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7280 static void arch_reinit_sched_domains(void)
7284 /* Destroy domains first to force the rebuild */
7285 partition_sched_domains(0, NULL, NULL);
7287 rebuild_sched_domains();
7291 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7293 unsigned int level = 0;
7295 if (sscanf(buf, "%u", &level) != 1)
7299 * level is always be positive so don't check for
7300 * level < POWERSAVINGS_BALANCE_NONE which is 0
7301 * What happens on 0 or 1 byte write,
7302 * need to check for count as well?
7305 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7309 sched_smt_power_savings = level;
7311 sched_mc_power_savings = level;
7313 arch_reinit_sched_domains();
7318 #ifdef CONFIG_SCHED_MC
7319 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7320 struct sysdev_class_attribute *attr,
7323 return sprintf(page, "%u\n", sched_mc_power_savings);
7325 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7326 struct sysdev_class_attribute *attr,
7327 const char *buf, size_t count)
7329 return sched_power_savings_store(buf, count, 0);
7331 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7332 sched_mc_power_savings_show,
7333 sched_mc_power_savings_store);
7336 #ifdef CONFIG_SCHED_SMT
7337 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7338 struct sysdev_class_attribute *attr,
7341 return sprintf(page, "%u\n", sched_smt_power_savings);
7343 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7344 struct sysdev_class_attribute *attr,
7345 const char *buf, size_t count)
7347 return sched_power_savings_store(buf, count, 1);
7349 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7350 sched_smt_power_savings_show,
7351 sched_smt_power_savings_store);
7354 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7358 #ifdef CONFIG_SCHED_SMT
7360 err = sysfs_create_file(&cls->kset.kobj,
7361 &attr_sched_smt_power_savings.attr);
7363 #ifdef CONFIG_SCHED_MC
7364 if (!err && mc_capable())
7365 err = sysfs_create_file(&cls->kset.kobj,
7366 &attr_sched_mc_power_savings.attr);
7370 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7373 * Update cpusets according to cpu_active mask. If cpusets are
7374 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7375 * around partition_sched_domains().
7377 static int __cpuexit cpuset_cpu_active(struct notifier_block *nfb,
7378 unsigned long action, void *hcpu)
7380 switch (action & ~CPU_TASKS_FROZEN) {
7382 case CPU_DOWN_FAILED:
7383 cpuset_update_active_cpus();
7390 static int __cpuexit cpuset_cpu_inactive(struct notifier_block *nfb,
7391 unsigned long action, void *hcpu)
7393 switch (action & ~CPU_TASKS_FROZEN) {
7394 case CPU_DOWN_PREPARE:
7395 cpuset_update_active_cpus();
7402 static int update_runtime(struct notifier_block *nfb,
7403 unsigned long action, void *hcpu)
7405 int cpu = (int)(long)hcpu;
7408 case CPU_DOWN_PREPARE:
7409 case CPU_DOWN_PREPARE_FROZEN:
7410 disable_runtime(cpu_rq(cpu));
7413 case CPU_DOWN_FAILED:
7414 case CPU_DOWN_FAILED_FROZEN:
7416 case CPU_ONLINE_FROZEN:
7417 enable_runtime(cpu_rq(cpu));
7425 void __init sched_init_smp(void)
7427 cpumask_var_t non_isolated_cpus;
7429 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7430 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7432 #if defined(CONFIG_NUMA)
7433 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7435 BUG_ON(sched_group_nodes_bycpu == NULL);
7438 mutex_lock(&sched_domains_mutex);
7439 arch_init_sched_domains(cpu_active_mask);
7440 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7441 if (cpumask_empty(non_isolated_cpus))
7442 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7443 mutex_unlock(&sched_domains_mutex);
7446 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7447 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7449 /* RT runtime code needs to handle some hotplug events */
7450 hotcpu_notifier(update_runtime, 0);
7454 /* Move init over to a non-isolated CPU */
7455 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7457 sched_init_granularity();
7458 free_cpumask_var(non_isolated_cpus);
7460 init_sched_rt_class();
7463 void __init sched_init_smp(void)
7465 sched_init_granularity();
7467 #endif /* CONFIG_SMP */
7469 const_debug unsigned int sysctl_timer_migration = 1;
7471 int in_sched_functions(unsigned long addr)
7473 return in_lock_functions(addr) ||
7474 (addr >= (unsigned long)__sched_text_start
7475 && addr < (unsigned long)__sched_text_end);
7478 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7480 cfs_rq->tasks_timeline = RB_ROOT;
7481 INIT_LIST_HEAD(&cfs_rq->tasks);
7482 #ifdef CONFIG_FAIR_GROUP_SCHED
7485 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7488 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7490 struct rt_prio_array *array;
7493 array = &rt_rq->active;
7494 for (i = 0; i < MAX_RT_PRIO; i++) {
7495 INIT_LIST_HEAD(array->queue + i);
7496 __clear_bit(i, array->bitmap);
7498 /* delimiter for bitsearch: */
7499 __set_bit(MAX_RT_PRIO, array->bitmap);
7501 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7502 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7504 rt_rq->highest_prio.next = MAX_RT_PRIO;
7508 rt_rq->rt_nr_migratory = 0;
7509 rt_rq->overloaded = 0;
7510 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7514 rt_rq->rt_throttled = 0;
7515 rt_rq->rt_runtime = 0;
7516 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7518 #ifdef CONFIG_RT_GROUP_SCHED
7519 rt_rq->rt_nr_boosted = 0;
7524 #ifdef CONFIG_FAIR_GROUP_SCHED
7525 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7526 struct sched_entity *se, int cpu, int add,
7527 struct sched_entity *parent)
7529 struct rq *rq = cpu_rq(cpu);
7530 tg->cfs_rq[cpu] = cfs_rq;
7531 init_cfs_rq(cfs_rq, rq);
7534 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7537 /* se could be NULL for init_task_group */
7542 se->cfs_rq = &rq->cfs;
7544 se->cfs_rq = parent->my_q;
7547 se->load.weight = tg->shares;
7548 se->load.inv_weight = 0;
7549 se->parent = parent;
7553 #ifdef CONFIG_RT_GROUP_SCHED
7554 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7555 struct sched_rt_entity *rt_se, int cpu, int add,
7556 struct sched_rt_entity *parent)
7558 struct rq *rq = cpu_rq(cpu);
7560 tg->rt_rq[cpu] = rt_rq;
7561 init_rt_rq(rt_rq, rq);
7563 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7565 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7567 tg->rt_se[cpu] = rt_se;
7572 rt_se->rt_rq = &rq->rt;
7574 rt_se->rt_rq = parent->my_q;
7576 rt_se->my_q = rt_rq;
7577 rt_se->parent = parent;
7578 INIT_LIST_HEAD(&rt_se->run_list);
7582 void __init sched_init(void)
7585 unsigned long alloc_size = 0, ptr;
7587 #ifdef CONFIG_FAIR_GROUP_SCHED
7588 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7590 #ifdef CONFIG_RT_GROUP_SCHED
7591 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7593 #ifdef CONFIG_CPUMASK_OFFSTACK
7594 alloc_size += num_possible_cpus() * cpumask_size();
7597 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7599 #ifdef CONFIG_FAIR_GROUP_SCHED
7600 init_task_group.se = (struct sched_entity **)ptr;
7601 ptr += nr_cpu_ids * sizeof(void **);
7603 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7604 ptr += nr_cpu_ids * sizeof(void **);
7606 #endif /* CONFIG_FAIR_GROUP_SCHED */
7607 #ifdef CONFIG_RT_GROUP_SCHED
7608 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7609 ptr += nr_cpu_ids * sizeof(void **);
7611 init_task_group.rt_rq = (struct rt_rq **)ptr;
7612 ptr += nr_cpu_ids * sizeof(void **);
7614 #endif /* CONFIG_RT_GROUP_SCHED */
7615 #ifdef CONFIG_CPUMASK_OFFSTACK
7616 for_each_possible_cpu(i) {
7617 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7618 ptr += cpumask_size();
7620 #endif /* CONFIG_CPUMASK_OFFSTACK */
7624 init_defrootdomain();
7627 init_rt_bandwidth(&def_rt_bandwidth,
7628 global_rt_period(), global_rt_runtime());
7630 #ifdef CONFIG_RT_GROUP_SCHED
7631 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7632 global_rt_period(), global_rt_runtime());
7633 #endif /* CONFIG_RT_GROUP_SCHED */
7635 #ifdef CONFIG_CGROUP_SCHED
7636 list_add(&init_task_group.list, &task_groups);
7637 INIT_LIST_HEAD(&init_task_group.children);
7639 #endif /* CONFIG_CGROUP_SCHED */
7641 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7642 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7643 __alignof__(unsigned long));
7645 for_each_possible_cpu(i) {
7649 raw_spin_lock_init(&rq->lock);
7651 rq->calc_load_active = 0;
7652 rq->calc_load_update = jiffies + LOAD_FREQ;
7653 init_cfs_rq(&rq->cfs, rq);
7654 init_rt_rq(&rq->rt, rq);
7655 #ifdef CONFIG_FAIR_GROUP_SCHED
7656 init_task_group.shares = init_task_group_load;
7657 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7658 #ifdef CONFIG_CGROUP_SCHED
7660 * How much cpu bandwidth does init_task_group get?
7662 * In case of task-groups formed thr' the cgroup filesystem, it
7663 * gets 100% of the cpu resources in the system. This overall
7664 * system cpu resource is divided among the tasks of
7665 * init_task_group and its child task-groups in a fair manner,
7666 * based on each entity's (task or task-group's) weight
7667 * (se->load.weight).
7669 * In other words, if init_task_group has 10 tasks of weight
7670 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7671 * then A0's share of the cpu resource is:
7673 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7675 * We achieve this by letting init_task_group's tasks sit
7676 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7678 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7680 #endif /* CONFIG_FAIR_GROUP_SCHED */
7682 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7683 #ifdef CONFIG_RT_GROUP_SCHED
7684 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7685 #ifdef CONFIG_CGROUP_SCHED
7686 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7690 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7691 rq->cpu_load[j] = 0;
7695 rq->cpu_power = SCHED_LOAD_SCALE;
7696 rq->post_schedule = 0;
7697 rq->active_balance = 0;
7698 rq->next_balance = jiffies;
7703 rq->avg_idle = 2*sysctl_sched_migration_cost;
7704 rq_attach_root(rq, &def_root_domain);
7707 atomic_set(&rq->nr_iowait, 0);
7710 set_load_weight(&init_task);
7712 #ifdef CONFIG_PREEMPT_NOTIFIERS
7713 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7717 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7720 #ifdef CONFIG_RT_MUTEXES
7721 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7725 * The boot idle thread does lazy MMU switching as well:
7727 atomic_inc(&init_mm.mm_count);
7728 enter_lazy_tlb(&init_mm, current);
7731 * Make us the idle thread. Technically, schedule() should not be
7732 * called from this thread, however somewhere below it might be,
7733 * but because we are the idle thread, we just pick up running again
7734 * when this runqueue becomes "idle".
7736 init_idle(current, smp_processor_id());
7738 calc_load_update = jiffies + LOAD_FREQ;
7741 * During early bootup we pretend to be a normal task:
7743 current->sched_class = &fair_sched_class;
7745 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7746 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7749 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7750 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7752 /* May be allocated at isolcpus cmdline parse time */
7753 if (cpu_isolated_map == NULL)
7754 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7759 scheduler_running = 1;
7762 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7763 static inline int preempt_count_equals(int preempt_offset)
7765 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7767 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7770 void __might_sleep(const char *file, int line, int preempt_offset)
7773 static unsigned long prev_jiffy; /* ratelimiting */
7775 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7776 system_state != SYSTEM_RUNNING || oops_in_progress)
7778 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7780 prev_jiffy = jiffies;
7783 "BUG: sleeping function called from invalid context at %s:%d\n",
7786 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7787 in_atomic(), irqs_disabled(),
7788 current->pid, current->comm);
7790 debug_show_held_locks(current);
7791 if (irqs_disabled())
7792 print_irqtrace_events(current);
7796 EXPORT_SYMBOL(__might_sleep);
7799 #ifdef CONFIG_MAGIC_SYSRQ
7800 static void normalize_task(struct rq *rq, struct task_struct *p)
7804 on_rq = p->se.on_rq;
7806 deactivate_task(rq, p, 0);
7807 __setscheduler(rq, p, SCHED_NORMAL, 0);
7809 activate_task(rq, p, 0);
7810 resched_task(rq->curr);
7814 void normalize_rt_tasks(void)
7816 struct task_struct *g, *p;
7817 unsigned long flags;
7820 read_lock_irqsave(&tasklist_lock, flags);
7821 do_each_thread(g, p) {
7823 * Only normalize user tasks:
7828 p->se.exec_start = 0;
7829 #ifdef CONFIG_SCHEDSTATS
7830 p->se.statistics.wait_start = 0;
7831 p->se.statistics.sleep_start = 0;
7832 p->se.statistics.block_start = 0;
7837 * Renice negative nice level userspace
7840 if (TASK_NICE(p) < 0 && p->mm)
7841 set_user_nice(p, 0);
7845 raw_spin_lock(&p->pi_lock);
7846 rq = __task_rq_lock(p);
7848 normalize_task(rq, p);
7850 __task_rq_unlock(rq);
7851 raw_spin_unlock(&p->pi_lock);
7852 } while_each_thread(g, p);
7854 read_unlock_irqrestore(&tasklist_lock, flags);
7857 #endif /* CONFIG_MAGIC_SYSRQ */
7859 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7861 * These functions are only useful for the IA64 MCA handling, or kdb.
7863 * They can only be called when the whole system has been
7864 * stopped - every CPU needs to be quiescent, and no scheduling
7865 * activity can take place. Using them for anything else would
7866 * be a serious bug, and as a result, they aren't even visible
7867 * under any other configuration.
7871 * curr_task - return the current task for a given cpu.
7872 * @cpu: the processor in question.
7874 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7876 struct task_struct *curr_task(int cpu)
7878 return cpu_curr(cpu);
7881 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7885 * set_curr_task - set the current task for a given cpu.
7886 * @cpu: the processor in question.
7887 * @p: the task pointer to set.
7889 * Description: This function must only be used when non-maskable interrupts
7890 * are serviced on a separate stack. It allows the architecture to switch the
7891 * notion of the current task on a cpu in a non-blocking manner. This function
7892 * must be called with all CPU's synchronized, and interrupts disabled, the
7893 * and caller must save the original value of the current task (see
7894 * curr_task() above) and restore that value before reenabling interrupts and
7895 * re-starting the system.
7897 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7899 void set_curr_task(int cpu, struct task_struct *p)
7906 #ifdef CONFIG_FAIR_GROUP_SCHED
7907 static void free_fair_sched_group(struct task_group *tg)
7911 for_each_possible_cpu(i) {
7913 kfree(tg->cfs_rq[i]);
7923 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7925 struct cfs_rq *cfs_rq;
7926 struct sched_entity *se;
7930 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7933 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7937 tg->shares = NICE_0_LOAD;
7939 for_each_possible_cpu(i) {
7942 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7943 GFP_KERNEL, cpu_to_node(i));
7947 se = kzalloc_node(sizeof(struct sched_entity),
7948 GFP_KERNEL, cpu_to_node(i));
7952 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
7963 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7965 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7966 &cpu_rq(cpu)->leaf_cfs_rq_list);
7969 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7971 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7973 #else /* !CONFG_FAIR_GROUP_SCHED */
7974 static inline void free_fair_sched_group(struct task_group *tg)
7979 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7984 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7988 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7991 #endif /* CONFIG_FAIR_GROUP_SCHED */
7993 #ifdef CONFIG_RT_GROUP_SCHED
7994 static void free_rt_sched_group(struct task_group *tg)
7998 destroy_rt_bandwidth(&tg->rt_bandwidth);
8000 for_each_possible_cpu(i) {
8002 kfree(tg->rt_rq[i]);
8004 kfree(tg->rt_se[i]);
8012 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8014 struct rt_rq *rt_rq;
8015 struct sched_rt_entity *rt_se;
8019 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8022 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8026 init_rt_bandwidth(&tg->rt_bandwidth,
8027 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8029 for_each_possible_cpu(i) {
8032 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8033 GFP_KERNEL, cpu_to_node(i));
8037 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8038 GFP_KERNEL, cpu_to_node(i));
8042 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8053 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8055 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8056 &cpu_rq(cpu)->leaf_rt_rq_list);
8059 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8061 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8063 #else /* !CONFIG_RT_GROUP_SCHED */
8064 static inline void free_rt_sched_group(struct task_group *tg)
8069 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8074 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8078 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8081 #endif /* CONFIG_RT_GROUP_SCHED */
8083 #ifdef CONFIG_CGROUP_SCHED
8084 static void free_sched_group(struct task_group *tg)
8086 free_fair_sched_group(tg);
8087 free_rt_sched_group(tg);
8091 /* allocate runqueue etc for a new task group */
8092 struct task_group *sched_create_group(struct task_group *parent)
8094 struct task_group *tg;
8095 unsigned long flags;
8098 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8100 return ERR_PTR(-ENOMEM);
8102 if (!alloc_fair_sched_group(tg, parent))
8105 if (!alloc_rt_sched_group(tg, parent))
8108 spin_lock_irqsave(&task_group_lock, flags);
8109 for_each_possible_cpu(i) {
8110 register_fair_sched_group(tg, i);
8111 register_rt_sched_group(tg, i);
8113 list_add_rcu(&tg->list, &task_groups);
8115 WARN_ON(!parent); /* root should already exist */
8117 tg->parent = parent;
8118 INIT_LIST_HEAD(&tg->children);
8119 list_add_rcu(&tg->siblings, &parent->children);
8120 spin_unlock_irqrestore(&task_group_lock, flags);
8125 free_sched_group(tg);
8126 return ERR_PTR(-ENOMEM);
8129 /* rcu callback to free various structures associated with a task group */
8130 static void free_sched_group_rcu(struct rcu_head *rhp)
8132 /* now it should be safe to free those cfs_rqs */
8133 free_sched_group(container_of(rhp, struct task_group, rcu));
8136 /* Destroy runqueue etc associated with a task group */
8137 void sched_destroy_group(struct task_group *tg)
8139 unsigned long flags;
8142 spin_lock_irqsave(&task_group_lock, flags);
8143 for_each_possible_cpu(i) {
8144 unregister_fair_sched_group(tg, i);
8145 unregister_rt_sched_group(tg, i);
8147 list_del_rcu(&tg->list);
8148 list_del_rcu(&tg->siblings);
8149 spin_unlock_irqrestore(&task_group_lock, flags);
8151 /* wait for possible concurrent references to cfs_rqs complete */
8152 call_rcu(&tg->rcu, free_sched_group_rcu);
8155 /* change task's runqueue when it moves between groups.
8156 * The caller of this function should have put the task in its new group
8157 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8158 * reflect its new group.
8160 void sched_move_task(struct task_struct *tsk)
8163 unsigned long flags;
8166 rq = task_rq_lock(tsk, &flags);
8168 running = task_current(rq, tsk);
8169 on_rq = tsk->se.on_rq;
8172 dequeue_task(rq, tsk, 0);
8173 if (unlikely(running))
8174 tsk->sched_class->put_prev_task(rq, tsk);
8176 set_task_rq(tsk, task_cpu(tsk));
8178 #ifdef CONFIG_FAIR_GROUP_SCHED
8179 if (tsk->sched_class->moved_group)
8180 tsk->sched_class->moved_group(tsk, on_rq);
8183 if (unlikely(running))
8184 tsk->sched_class->set_curr_task(rq);
8186 enqueue_task(rq, tsk, 0);
8188 task_rq_unlock(rq, &flags);
8190 #endif /* CONFIG_CGROUP_SCHED */
8192 #ifdef CONFIG_FAIR_GROUP_SCHED
8193 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8195 struct cfs_rq *cfs_rq = se->cfs_rq;
8200 dequeue_entity(cfs_rq, se, 0);
8202 se->load.weight = shares;
8203 se->load.inv_weight = 0;
8206 enqueue_entity(cfs_rq, se, 0);
8209 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8211 struct cfs_rq *cfs_rq = se->cfs_rq;
8212 struct rq *rq = cfs_rq->rq;
8213 unsigned long flags;
8215 raw_spin_lock_irqsave(&rq->lock, flags);
8216 __set_se_shares(se, shares);
8217 raw_spin_unlock_irqrestore(&rq->lock, flags);
8220 static DEFINE_MUTEX(shares_mutex);
8222 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8225 unsigned long flags;
8228 * We can't change the weight of the root cgroup.
8233 if (shares < MIN_SHARES)
8234 shares = MIN_SHARES;
8235 else if (shares > MAX_SHARES)
8236 shares = MAX_SHARES;
8238 mutex_lock(&shares_mutex);
8239 if (tg->shares == shares)
8242 spin_lock_irqsave(&task_group_lock, flags);
8243 for_each_possible_cpu(i)
8244 unregister_fair_sched_group(tg, i);
8245 list_del_rcu(&tg->siblings);
8246 spin_unlock_irqrestore(&task_group_lock, flags);
8248 /* wait for any ongoing reference to this group to finish */
8249 synchronize_sched();
8252 * Now we are free to modify the group's share on each cpu
8253 * w/o tripping rebalance_share or load_balance_fair.
8255 tg->shares = shares;
8256 for_each_possible_cpu(i) {
8260 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8261 set_se_shares(tg->se[i], shares);
8265 * Enable load balance activity on this group, by inserting it back on
8266 * each cpu's rq->leaf_cfs_rq_list.
8268 spin_lock_irqsave(&task_group_lock, flags);
8269 for_each_possible_cpu(i)
8270 register_fair_sched_group(tg, i);
8271 list_add_rcu(&tg->siblings, &tg->parent->children);
8272 spin_unlock_irqrestore(&task_group_lock, flags);
8274 mutex_unlock(&shares_mutex);
8278 unsigned long sched_group_shares(struct task_group *tg)
8284 #ifdef CONFIG_RT_GROUP_SCHED
8286 * Ensure that the real time constraints are schedulable.
8288 static DEFINE_MUTEX(rt_constraints_mutex);
8290 static unsigned long to_ratio(u64 period, u64 runtime)
8292 if (runtime == RUNTIME_INF)
8295 return div64_u64(runtime << 20, period);
8298 /* Must be called with tasklist_lock held */
8299 static inline int tg_has_rt_tasks(struct task_group *tg)
8301 struct task_struct *g, *p;
8303 do_each_thread(g, p) {
8304 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8306 } while_each_thread(g, p);
8311 struct rt_schedulable_data {
8312 struct task_group *tg;
8317 static int tg_schedulable(struct task_group *tg, void *data)
8319 struct rt_schedulable_data *d = data;
8320 struct task_group *child;
8321 unsigned long total, sum = 0;
8322 u64 period, runtime;
8324 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8325 runtime = tg->rt_bandwidth.rt_runtime;
8328 period = d->rt_period;
8329 runtime = d->rt_runtime;
8333 * Cannot have more runtime than the period.
8335 if (runtime > period && runtime != RUNTIME_INF)
8339 * Ensure we don't starve existing RT tasks.
8341 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8344 total = to_ratio(period, runtime);
8347 * Nobody can have more than the global setting allows.
8349 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8353 * The sum of our children's runtime should not exceed our own.
8355 list_for_each_entry_rcu(child, &tg->children, siblings) {
8356 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8357 runtime = child->rt_bandwidth.rt_runtime;
8359 if (child == d->tg) {
8360 period = d->rt_period;
8361 runtime = d->rt_runtime;
8364 sum += to_ratio(period, runtime);
8373 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8375 struct rt_schedulable_data data = {
8377 .rt_period = period,
8378 .rt_runtime = runtime,
8381 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8384 static int tg_set_bandwidth(struct task_group *tg,
8385 u64 rt_period, u64 rt_runtime)
8389 mutex_lock(&rt_constraints_mutex);
8390 read_lock(&tasklist_lock);
8391 err = __rt_schedulable(tg, rt_period, rt_runtime);
8395 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8396 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8397 tg->rt_bandwidth.rt_runtime = rt_runtime;
8399 for_each_possible_cpu(i) {
8400 struct rt_rq *rt_rq = tg->rt_rq[i];
8402 raw_spin_lock(&rt_rq->rt_runtime_lock);
8403 rt_rq->rt_runtime = rt_runtime;
8404 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8406 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8408 read_unlock(&tasklist_lock);
8409 mutex_unlock(&rt_constraints_mutex);
8414 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8416 u64 rt_runtime, rt_period;
8418 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8419 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8420 if (rt_runtime_us < 0)
8421 rt_runtime = RUNTIME_INF;
8423 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8426 long sched_group_rt_runtime(struct task_group *tg)
8430 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8433 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8434 do_div(rt_runtime_us, NSEC_PER_USEC);
8435 return rt_runtime_us;
8438 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8440 u64 rt_runtime, rt_period;
8442 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8443 rt_runtime = tg->rt_bandwidth.rt_runtime;
8448 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8451 long sched_group_rt_period(struct task_group *tg)
8455 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8456 do_div(rt_period_us, NSEC_PER_USEC);
8457 return rt_period_us;
8460 static int sched_rt_global_constraints(void)
8462 u64 runtime, period;
8465 if (sysctl_sched_rt_period <= 0)
8468 runtime = global_rt_runtime();
8469 period = global_rt_period();
8472 * Sanity check on the sysctl variables.
8474 if (runtime > period && runtime != RUNTIME_INF)
8477 mutex_lock(&rt_constraints_mutex);
8478 read_lock(&tasklist_lock);
8479 ret = __rt_schedulable(NULL, 0, 0);
8480 read_unlock(&tasklist_lock);
8481 mutex_unlock(&rt_constraints_mutex);
8486 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8488 /* Don't accept realtime tasks when there is no way for them to run */
8489 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8495 #else /* !CONFIG_RT_GROUP_SCHED */
8496 static int sched_rt_global_constraints(void)
8498 unsigned long flags;
8501 if (sysctl_sched_rt_period <= 0)
8505 * There's always some RT tasks in the root group
8506 * -- migration, kstopmachine etc..
8508 if (sysctl_sched_rt_runtime == 0)
8511 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8512 for_each_possible_cpu(i) {
8513 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8515 raw_spin_lock(&rt_rq->rt_runtime_lock);
8516 rt_rq->rt_runtime = global_rt_runtime();
8517 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8519 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8523 #endif /* CONFIG_RT_GROUP_SCHED */
8525 int sched_rt_handler(struct ctl_table *table, int write,
8526 void __user *buffer, size_t *lenp,
8530 int old_period, old_runtime;
8531 static DEFINE_MUTEX(mutex);
8534 old_period = sysctl_sched_rt_period;
8535 old_runtime = sysctl_sched_rt_runtime;
8537 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8539 if (!ret && write) {
8540 ret = sched_rt_global_constraints();
8542 sysctl_sched_rt_period = old_period;
8543 sysctl_sched_rt_runtime = old_runtime;
8545 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8546 def_rt_bandwidth.rt_period =
8547 ns_to_ktime(global_rt_period());
8550 mutex_unlock(&mutex);
8555 #ifdef CONFIG_CGROUP_SCHED
8557 /* return corresponding task_group object of a cgroup */
8558 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8560 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8561 struct task_group, css);
8564 static struct cgroup_subsys_state *
8565 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8567 struct task_group *tg, *parent;
8569 if (!cgrp->parent) {
8570 /* This is early initialization for the top cgroup */
8571 return &init_task_group.css;
8574 parent = cgroup_tg(cgrp->parent);
8575 tg = sched_create_group(parent);
8577 return ERR_PTR(-ENOMEM);
8583 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8585 struct task_group *tg = cgroup_tg(cgrp);
8587 sched_destroy_group(tg);
8591 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8593 #ifdef CONFIG_RT_GROUP_SCHED
8594 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8597 /* We don't support RT-tasks being in separate groups */
8598 if (tsk->sched_class != &fair_sched_class)
8605 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8606 struct task_struct *tsk, bool threadgroup)
8608 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8612 struct task_struct *c;
8614 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8615 retval = cpu_cgroup_can_attach_task(cgrp, c);
8627 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8628 struct cgroup *old_cont, struct task_struct *tsk,
8631 sched_move_task(tsk);
8633 struct task_struct *c;
8635 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8642 #ifdef CONFIG_FAIR_GROUP_SCHED
8643 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8646 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8649 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8651 struct task_group *tg = cgroup_tg(cgrp);
8653 return (u64) tg->shares;
8655 #endif /* CONFIG_FAIR_GROUP_SCHED */
8657 #ifdef CONFIG_RT_GROUP_SCHED
8658 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8661 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8664 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8666 return sched_group_rt_runtime(cgroup_tg(cgrp));
8669 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8672 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8675 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8677 return sched_group_rt_period(cgroup_tg(cgrp));
8679 #endif /* CONFIG_RT_GROUP_SCHED */
8681 static struct cftype cpu_files[] = {
8682 #ifdef CONFIG_FAIR_GROUP_SCHED
8685 .read_u64 = cpu_shares_read_u64,
8686 .write_u64 = cpu_shares_write_u64,
8689 #ifdef CONFIG_RT_GROUP_SCHED
8691 .name = "rt_runtime_us",
8692 .read_s64 = cpu_rt_runtime_read,
8693 .write_s64 = cpu_rt_runtime_write,
8696 .name = "rt_period_us",
8697 .read_u64 = cpu_rt_period_read_uint,
8698 .write_u64 = cpu_rt_period_write_uint,
8703 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8705 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8708 struct cgroup_subsys cpu_cgroup_subsys = {
8710 .create = cpu_cgroup_create,
8711 .destroy = cpu_cgroup_destroy,
8712 .can_attach = cpu_cgroup_can_attach,
8713 .attach = cpu_cgroup_attach,
8714 .populate = cpu_cgroup_populate,
8715 .subsys_id = cpu_cgroup_subsys_id,
8719 #endif /* CONFIG_CGROUP_SCHED */
8721 #ifdef CONFIG_CGROUP_CPUACCT
8724 * CPU accounting code for task groups.
8726 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8727 * (balbir@in.ibm.com).
8730 /* track cpu usage of a group of tasks and its child groups */
8732 struct cgroup_subsys_state css;
8733 /* cpuusage holds pointer to a u64-type object on every cpu */
8734 u64 __percpu *cpuusage;
8735 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8736 struct cpuacct *parent;
8739 struct cgroup_subsys cpuacct_subsys;
8741 /* return cpu accounting group corresponding to this container */
8742 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8744 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8745 struct cpuacct, css);
8748 /* return cpu accounting group to which this task belongs */
8749 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8751 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8752 struct cpuacct, css);
8755 /* create a new cpu accounting group */
8756 static struct cgroup_subsys_state *cpuacct_create(
8757 struct cgroup_subsys *ss, struct cgroup *cgrp)
8759 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8765 ca->cpuusage = alloc_percpu(u64);
8769 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8770 if (percpu_counter_init(&ca->cpustat[i], 0))
8771 goto out_free_counters;
8774 ca->parent = cgroup_ca(cgrp->parent);
8780 percpu_counter_destroy(&ca->cpustat[i]);
8781 free_percpu(ca->cpuusage);
8785 return ERR_PTR(-ENOMEM);
8788 /* destroy an existing cpu accounting group */
8790 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8792 struct cpuacct *ca = cgroup_ca(cgrp);
8795 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8796 percpu_counter_destroy(&ca->cpustat[i]);
8797 free_percpu(ca->cpuusage);
8801 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8803 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8806 #ifndef CONFIG_64BIT
8808 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8810 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8812 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8820 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8822 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8824 #ifndef CONFIG_64BIT
8826 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8828 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8830 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8836 /* return total cpu usage (in nanoseconds) of a group */
8837 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8839 struct cpuacct *ca = cgroup_ca(cgrp);
8840 u64 totalcpuusage = 0;
8843 for_each_present_cpu(i)
8844 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8846 return totalcpuusage;
8849 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8852 struct cpuacct *ca = cgroup_ca(cgrp);
8861 for_each_present_cpu(i)
8862 cpuacct_cpuusage_write(ca, i, 0);
8868 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8871 struct cpuacct *ca = cgroup_ca(cgroup);
8875 for_each_present_cpu(i) {
8876 percpu = cpuacct_cpuusage_read(ca, i);
8877 seq_printf(m, "%llu ", (unsigned long long) percpu);
8879 seq_printf(m, "\n");
8883 static const char *cpuacct_stat_desc[] = {
8884 [CPUACCT_STAT_USER] = "user",
8885 [CPUACCT_STAT_SYSTEM] = "system",
8888 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8889 struct cgroup_map_cb *cb)
8891 struct cpuacct *ca = cgroup_ca(cgrp);
8894 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8895 s64 val = percpu_counter_read(&ca->cpustat[i]);
8896 val = cputime64_to_clock_t(val);
8897 cb->fill(cb, cpuacct_stat_desc[i], val);
8902 static struct cftype files[] = {
8905 .read_u64 = cpuusage_read,
8906 .write_u64 = cpuusage_write,
8909 .name = "usage_percpu",
8910 .read_seq_string = cpuacct_percpu_seq_read,
8914 .read_map = cpuacct_stats_show,
8918 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8920 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8924 * charge this task's execution time to its accounting group.
8926 * called with rq->lock held.
8928 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8933 if (unlikely(!cpuacct_subsys.active))
8936 cpu = task_cpu(tsk);
8942 for (; ca; ca = ca->parent) {
8943 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8944 *cpuusage += cputime;
8951 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8952 * in cputime_t units. As a result, cpuacct_update_stats calls
8953 * percpu_counter_add with values large enough to always overflow the
8954 * per cpu batch limit causing bad SMP scalability.
8956 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8957 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8958 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8961 #define CPUACCT_BATCH \
8962 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8964 #define CPUACCT_BATCH 0
8968 * Charge the system/user time to the task's accounting group.
8970 static void cpuacct_update_stats(struct task_struct *tsk,
8971 enum cpuacct_stat_index idx, cputime_t val)
8974 int batch = CPUACCT_BATCH;
8976 if (unlikely(!cpuacct_subsys.active))
8983 __percpu_counter_add(&ca->cpustat[idx], val, batch);
8989 struct cgroup_subsys cpuacct_subsys = {
8991 .create = cpuacct_create,
8992 .destroy = cpuacct_destroy,
8993 .populate = cpuacct_populate,
8994 .subsys_id = cpuacct_subsys_id,
8996 #endif /* CONFIG_CGROUP_CPUACCT */
9000 void synchronize_sched_expedited(void)
9004 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9006 #else /* #ifndef CONFIG_SMP */
9008 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9010 static int synchronize_sched_expedited_cpu_stop(void *data)
9013 * There must be a full memory barrier on each affected CPU
9014 * between the time that try_stop_cpus() is called and the
9015 * time that it returns.
9017 * In the current initial implementation of cpu_stop, the
9018 * above condition is already met when the control reaches
9019 * this point and the following smp_mb() is not strictly
9020 * necessary. Do smp_mb() anyway for documentation and
9021 * robustness against future implementation changes.
9023 smp_mb(); /* See above comment block. */
9028 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9029 * approach to force grace period to end quickly. This consumes
9030 * significant time on all CPUs, and is thus not recommended for
9031 * any sort of common-case code.
9033 * Note that it is illegal to call this function while holding any
9034 * lock that is acquired by a CPU-hotplug notifier. Failing to
9035 * observe this restriction will result in deadlock.
9037 void synchronize_sched_expedited(void)
9039 int snap, trycount = 0;
9041 smp_mb(); /* ensure prior mod happens before capturing snap. */
9042 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9044 while (try_stop_cpus(cpu_online_mask,
9045 synchronize_sched_expedited_cpu_stop,
9048 if (trycount++ < 10)
9049 udelay(trycount * num_online_cpus());
9051 synchronize_sched();
9054 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9055 smp_mb(); /* ensure test happens before caller kfree */
9060 atomic_inc(&synchronize_sched_expedited_count);
9061 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9064 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9066 #endif /* #else #ifndef CONFIG_SMP */