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/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.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>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 raw_spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 raw_spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 raw_spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 raw_spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_CGROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 struct cgroup_subsys_state css;
248 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* schedulable entities of this group on each cpu */
250 struct sched_entity **se;
251 /* runqueue "owned" by this group on each cpu */
252 struct cfs_rq **cfs_rq;
253 unsigned long shares;
256 #ifdef CONFIG_RT_GROUP_SCHED
257 struct sched_rt_entity **rt_se;
258 struct rt_rq **rt_rq;
260 struct rt_bandwidth rt_bandwidth;
264 struct list_head list;
266 struct task_group *parent;
267 struct list_head siblings;
268 struct list_head children;
271 #define root_task_group init_task_group
273 /* task_group_lock serializes add/remove of task groups and also changes to
274 * a task group's cpu shares.
276 static DEFINE_SPINLOCK(task_group_lock);
278 #ifdef CONFIG_FAIR_GROUP_SCHED
281 static int root_task_group_empty(void)
283 return list_empty(&root_task_group.children);
287 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
290 * A weight of 0 or 1 can cause arithmetics problems.
291 * A weight of a cfs_rq is the sum of weights of which entities
292 * are queued on this cfs_rq, so a weight of a entity should not be
293 * too large, so as the shares value of a task group.
294 * (The default weight is 1024 - so there's no practical
295 * limitation from this.)
298 #define MAX_SHARES (1UL << 18)
300 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
303 /* Default task group.
304 * Every task in system belong to this group at bootup.
306 struct task_group init_task_group;
308 /* return group to which a task belongs */
309 static inline struct task_group *task_group(struct task_struct *p)
311 struct task_group *tg;
313 #ifdef CONFIG_CGROUP_SCHED
314 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
315 struct task_group, css);
317 tg = &init_task_group;
322 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
323 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
327 p->se.parent = task_group(p)->se[cpu];
330 #ifdef CONFIG_RT_GROUP_SCHED
331 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
332 p->rt.parent = task_group(p)->rt_se[cpu];
338 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
339 static inline struct task_group *task_group(struct task_struct *p)
344 #endif /* CONFIG_CGROUP_SCHED */
346 /* CFS-related fields in a runqueue */
348 struct load_weight load;
349 unsigned long nr_running;
354 struct rb_root tasks_timeline;
355 struct rb_node *rb_leftmost;
357 struct list_head tasks;
358 struct list_head *balance_iterator;
361 * 'curr' points to currently running entity on this cfs_rq.
362 * It is set to NULL otherwise (i.e when none are currently running).
364 struct sched_entity *curr, *next, *last;
366 unsigned int nr_spread_over;
368 #ifdef CONFIG_FAIR_GROUP_SCHED
369 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
372 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
373 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
374 * (like users, containers etc.)
376 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
377 * list is used during load balance.
379 struct list_head leaf_cfs_rq_list;
380 struct task_group *tg; /* group that "owns" this runqueue */
384 * the part of load.weight contributed by tasks
386 unsigned long task_weight;
389 * h_load = weight * f(tg)
391 * Where f(tg) is the recursive weight fraction assigned to
394 unsigned long h_load;
397 * this cpu's part of tg->shares
399 unsigned long shares;
402 * load.weight at the time we set shares
404 unsigned long rq_weight;
409 /* Real-Time classes' related field in a runqueue: */
411 struct rt_prio_array active;
412 unsigned long rt_nr_running;
413 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
415 int curr; /* highest queued rt task prio */
417 int next; /* next highest */
422 unsigned long rt_nr_migratory;
423 unsigned long rt_nr_total;
425 struct plist_head pushable_tasks;
430 /* Nests inside the rq lock: */
431 raw_spinlock_t rt_runtime_lock;
433 #ifdef CONFIG_RT_GROUP_SCHED
434 unsigned long rt_nr_boosted;
437 struct list_head leaf_rt_rq_list;
438 struct task_group *tg;
445 * We add the notion of a root-domain which will be used to define per-domain
446 * variables. Each exclusive cpuset essentially defines an island domain by
447 * fully partitioning the member cpus from any other cpuset. Whenever a new
448 * exclusive cpuset is created, we also create and attach a new root-domain
455 cpumask_var_t online;
458 * The "RT overload" flag: it gets set if a CPU has more than
459 * one runnable RT task.
461 cpumask_var_t rto_mask;
464 struct cpupri cpupri;
469 * By default the system creates a single root-domain with all cpus as
470 * members (mimicking the global state we have today).
472 static struct root_domain def_root_domain;
477 * This is the main, per-CPU runqueue data structure.
479 * Locking rule: those places that want to lock multiple runqueues
480 * (such as the load balancing or the thread migration code), lock
481 * acquire operations must be ordered by ascending &runqueue.
488 * nr_running and cpu_load should be in the same cacheline because
489 * remote CPUs use both these fields when doing load calculation.
491 unsigned long nr_running;
492 #define CPU_LOAD_IDX_MAX 5
493 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
496 unsigned char in_nohz_recently;
498 unsigned int skip_clock_update;
500 /* capture load from *all* tasks on this cpu: */
501 struct load_weight load;
502 unsigned long nr_load_updates;
508 #ifdef CONFIG_FAIR_GROUP_SCHED
509 /* list of leaf cfs_rq on this cpu: */
510 struct list_head leaf_cfs_rq_list;
512 #ifdef CONFIG_RT_GROUP_SCHED
513 struct list_head leaf_rt_rq_list;
517 * This is part of a global counter where only the total sum
518 * over all CPUs matters. A task can increase this counter on
519 * one CPU and if it got migrated afterwards it may decrease
520 * it on another CPU. Always updated under the runqueue lock:
522 unsigned long nr_uninterruptible;
524 struct task_struct *curr, *idle;
525 unsigned long next_balance;
526 struct mm_struct *prev_mm;
533 struct root_domain *rd;
534 struct sched_domain *sd;
536 unsigned char idle_at_tick;
537 /* For active balancing */
541 /* cpu of this runqueue: */
545 unsigned long avg_load_per_task;
547 struct task_struct *migration_thread;
548 struct list_head migration_queue;
556 /* calc_load related fields */
557 unsigned long calc_load_update;
558 long calc_load_active;
560 #ifdef CONFIG_SCHED_HRTICK
562 int hrtick_csd_pending;
563 struct call_single_data hrtick_csd;
565 struct hrtimer hrtick_timer;
568 #ifdef CONFIG_SCHEDSTATS
570 struct sched_info rq_sched_info;
571 unsigned long long rq_cpu_time;
572 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
574 /* sys_sched_yield() stats */
575 unsigned int yld_count;
577 /* schedule() stats */
578 unsigned int sched_switch;
579 unsigned int sched_count;
580 unsigned int sched_goidle;
582 /* try_to_wake_up() stats */
583 unsigned int ttwu_count;
584 unsigned int ttwu_local;
587 unsigned int bkl_count;
591 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
594 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
596 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
599 * A queue event has occurred, and we're going to schedule. In
600 * this case, we can save a useless back to back clock update.
602 if (test_tsk_need_resched(p))
603 rq->skip_clock_update = 1;
606 static inline int cpu_of(struct rq *rq)
615 #define rcu_dereference_check_sched_domain(p) \
616 rcu_dereference_check((p), \
617 rcu_read_lock_sched_held() || \
618 lockdep_is_held(&sched_domains_mutex))
621 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
622 * See detach_destroy_domains: synchronize_sched for details.
624 * The domain tree of any CPU may only be accessed from within
625 * preempt-disabled sections.
627 #define for_each_domain(cpu, __sd) \
628 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
630 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
631 #define this_rq() (&__get_cpu_var(runqueues))
632 #define task_rq(p) cpu_rq(task_cpu(p))
633 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
634 #define raw_rq() (&__raw_get_cpu_var(runqueues))
636 inline void update_rq_clock(struct rq *rq)
638 if (!rq->skip_clock_update)
639 rq->clock = sched_clock_cpu(cpu_of(rq));
643 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
645 #ifdef CONFIG_SCHED_DEBUG
646 # define const_debug __read_mostly
648 # define const_debug static const
653 * @cpu: the processor in question.
655 * Returns true if the current cpu runqueue is locked.
656 * This interface allows printk to be called with the runqueue lock
657 * held and know whether or not it is OK to wake up the klogd.
659 int runqueue_is_locked(int cpu)
661 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
665 * Debugging: various feature bits
668 #define SCHED_FEAT(name, enabled) \
669 __SCHED_FEAT_##name ,
672 #include "sched_features.h"
677 #define SCHED_FEAT(name, enabled) \
678 (1UL << __SCHED_FEAT_##name) * enabled |
680 const_debug unsigned int sysctl_sched_features =
681 #include "sched_features.h"
686 #ifdef CONFIG_SCHED_DEBUG
687 #define SCHED_FEAT(name, enabled) \
690 static __read_mostly char *sched_feat_names[] = {
691 #include "sched_features.h"
697 static int sched_feat_show(struct seq_file *m, void *v)
701 for (i = 0; sched_feat_names[i]; i++) {
702 if (!(sysctl_sched_features & (1UL << i)))
704 seq_printf(m, "%s ", sched_feat_names[i]);
712 sched_feat_write(struct file *filp, const char __user *ubuf,
713 size_t cnt, loff_t *ppos)
723 if (copy_from_user(&buf, ubuf, cnt))
728 if (strncmp(buf, "NO_", 3) == 0) {
733 for (i = 0; sched_feat_names[i]; i++) {
734 int len = strlen(sched_feat_names[i]);
736 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
738 sysctl_sched_features &= ~(1UL << i);
740 sysctl_sched_features |= (1UL << i);
745 if (!sched_feat_names[i])
753 static int sched_feat_open(struct inode *inode, struct file *filp)
755 return single_open(filp, sched_feat_show, NULL);
758 static const struct file_operations sched_feat_fops = {
759 .open = sched_feat_open,
760 .write = sched_feat_write,
763 .release = single_release,
766 static __init int sched_init_debug(void)
768 debugfs_create_file("sched_features", 0644, NULL, NULL,
773 late_initcall(sched_init_debug);
777 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
780 * Number of tasks to iterate in a single balance run.
781 * Limited because this is done with IRQs disabled.
783 const_debug unsigned int sysctl_sched_nr_migrate = 32;
786 * ratelimit for updating the group shares.
789 unsigned int sysctl_sched_shares_ratelimit = 250000;
790 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
793 * Inject some fuzzyness into changing the per-cpu group shares
794 * this avoids remote rq-locks at the expense of fairness.
797 unsigned int sysctl_sched_shares_thresh = 4;
800 * period over which we average the RT time consumption, measured
805 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
808 * period over which we measure -rt task cpu usage in us.
811 unsigned int sysctl_sched_rt_period = 1000000;
813 static __read_mostly int scheduler_running;
816 * part of the period that we allow rt tasks to run in us.
819 int sysctl_sched_rt_runtime = 950000;
821 static inline u64 global_rt_period(void)
823 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
826 static inline u64 global_rt_runtime(void)
828 if (sysctl_sched_rt_runtime < 0)
831 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
834 #ifndef prepare_arch_switch
835 # define prepare_arch_switch(next) do { } while (0)
837 #ifndef finish_arch_switch
838 # define finish_arch_switch(prev) do { } while (0)
841 static inline int task_current(struct rq *rq, struct task_struct *p)
843 return rq->curr == p;
846 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
847 static inline int task_running(struct rq *rq, struct task_struct *p)
849 return task_current(rq, p);
852 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
856 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
858 #ifdef CONFIG_DEBUG_SPINLOCK
859 /* this is a valid case when another task releases the spinlock */
860 rq->lock.owner = current;
863 * If we are tracking spinlock dependencies then we have to
864 * fix up the runqueue lock - which gets 'carried over' from
867 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
869 raw_spin_unlock_irq(&rq->lock);
872 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
873 static inline int task_running(struct rq *rq, struct task_struct *p)
878 return task_current(rq, p);
882 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
886 * We can optimise this out completely for !SMP, because the
887 * SMP rebalancing from interrupt is the only thing that cares
892 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
893 raw_spin_unlock_irq(&rq->lock);
895 raw_spin_unlock(&rq->lock);
899 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
903 * After ->oncpu is cleared, the task can be moved to a different CPU.
904 * We must ensure this doesn't happen until the switch is completely
910 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
914 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
917 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
920 static inline int task_is_waking(struct task_struct *p)
922 return unlikely(p->state == TASK_WAKING);
926 * __task_rq_lock - lock the runqueue a given task resides on.
927 * Must be called interrupts disabled.
929 static inline struct rq *__task_rq_lock(struct task_struct *p)
936 raw_spin_lock(&rq->lock);
937 if (likely(rq == task_rq(p)))
939 raw_spin_unlock(&rq->lock);
944 * task_rq_lock - lock the runqueue a given task resides on and disable
945 * interrupts. Note the ordering: we can safely lookup the task_rq without
946 * explicitly disabling preemption.
948 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
954 local_irq_save(*flags);
956 raw_spin_lock(&rq->lock);
957 if (likely(rq == task_rq(p)))
959 raw_spin_unlock_irqrestore(&rq->lock, *flags);
963 void task_rq_unlock_wait(struct task_struct *p)
965 struct rq *rq = task_rq(p);
967 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
968 raw_spin_unlock_wait(&rq->lock);
971 static void __task_rq_unlock(struct rq *rq)
974 raw_spin_unlock(&rq->lock);
977 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
980 raw_spin_unlock_irqrestore(&rq->lock, *flags);
984 * this_rq_lock - lock this runqueue and disable interrupts.
986 static struct rq *this_rq_lock(void)
993 raw_spin_lock(&rq->lock);
998 #ifdef CONFIG_SCHED_HRTICK
1000 * Use HR-timers to deliver accurate preemption points.
1002 * Its all a bit involved since we cannot program an hrt while holding the
1003 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1006 * When we get rescheduled we reprogram the hrtick_timer outside of the
1012 * - enabled by features
1013 * - hrtimer is actually high res
1015 static inline int hrtick_enabled(struct rq *rq)
1017 if (!sched_feat(HRTICK))
1019 if (!cpu_active(cpu_of(rq)))
1021 return hrtimer_is_hres_active(&rq->hrtick_timer);
1024 static void hrtick_clear(struct rq *rq)
1026 if (hrtimer_active(&rq->hrtick_timer))
1027 hrtimer_cancel(&rq->hrtick_timer);
1031 * High-resolution timer tick.
1032 * Runs from hardirq context with interrupts disabled.
1034 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1036 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1038 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1040 raw_spin_lock(&rq->lock);
1041 update_rq_clock(rq);
1042 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1043 raw_spin_unlock(&rq->lock);
1045 return HRTIMER_NORESTART;
1050 * called from hardirq (IPI) context
1052 static void __hrtick_start(void *arg)
1054 struct rq *rq = arg;
1056 raw_spin_lock(&rq->lock);
1057 hrtimer_restart(&rq->hrtick_timer);
1058 rq->hrtick_csd_pending = 0;
1059 raw_spin_unlock(&rq->lock);
1063 * Called to set the hrtick timer state.
1065 * called with rq->lock held and irqs disabled
1067 static void hrtick_start(struct rq *rq, u64 delay)
1069 struct hrtimer *timer = &rq->hrtick_timer;
1070 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1072 hrtimer_set_expires(timer, time);
1074 if (rq == this_rq()) {
1075 hrtimer_restart(timer);
1076 } else if (!rq->hrtick_csd_pending) {
1077 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1078 rq->hrtick_csd_pending = 1;
1083 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1085 int cpu = (int)(long)hcpu;
1088 case CPU_UP_CANCELED:
1089 case CPU_UP_CANCELED_FROZEN:
1090 case CPU_DOWN_PREPARE:
1091 case CPU_DOWN_PREPARE_FROZEN:
1093 case CPU_DEAD_FROZEN:
1094 hrtick_clear(cpu_rq(cpu));
1101 static __init void init_hrtick(void)
1103 hotcpu_notifier(hotplug_hrtick, 0);
1107 * Called to set the hrtick timer state.
1109 * called with rq->lock held and irqs disabled
1111 static void hrtick_start(struct rq *rq, u64 delay)
1113 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1114 HRTIMER_MODE_REL_PINNED, 0);
1117 static inline void init_hrtick(void)
1120 #endif /* CONFIG_SMP */
1122 static void init_rq_hrtick(struct rq *rq)
1125 rq->hrtick_csd_pending = 0;
1127 rq->hrtick_csd.flags = 0;
1128 rq->hrtick_csd.func = __hrtick_start;
1129 rq->hrtick_csd.info = rq;
1132 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1133 rq->hrtick_timer.function = hrtick;
1135 #else /* CONFIG_SCHED_HRTICK */
1136 static inline void hrtick_clear(struct rq *rq)
1140 static inline void init_rq_hrtick(struct rq *rq)
1144 static inline void init_hrtick(void)
1147 #endif /* CONFIG_SCHED_HRTICK */
1150 * resched_task - mark a task 'to be rescheduled now'.
1152 * On UP this means the setting of the need_resched flag, on SMP it
1153 * might also involve a cross-CPU call to trigger the scheduler on
1158 #ifndef tsk_is_polling
1159 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1162 static void resched_task(struct task_struct *p)
1166 assert_raw_spin_locked(&task_rq(p)->lock);
1168 if (test_tsk_need_resched(p))
1171 set_tsk_need_resched(p);
1174 if (cpu == smp_processor_id())
1177 /* NEED_RESCHED must be visible before we test polling */
1179 if (!tsk_is_polling(p))
1180 smp_send_reschedule(cpu);
1183 static void resched_cpu(int cpu)
1185 struct rq *rq = cpu_rq(cpu);
1186 unsigned long flags;
1188 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1190 resched_task(cpu_curr(cpu));
1191 raw_spin_unlock_irqrestore(&rq->lock, flags);
1196 * When add_timer_on() enqueues a timer into the timer wheel of an
1197 * idle CPU then this timer might expire before the next timer event
1198 * which is scheduled to wake up that CPU. In case of a completely
1199 * idle system the next event might even be infinite time into the
1200 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1201 * leaves the inner idle loop so the newly added timer is taken into
1202 * account when the CPU goes back to idle and evaluates the timer
1203 * wheel for the next timer event.
1205 void wake_up_idle_cpu(int cpu)
1207 struct rq *rq = cpu_rq(cpu);
1209 if (cpu == smp_processor_id())
1213 * This is safe, as this function is called with the timer
1214 * wheel base lock of (cpu) held. When the CPU is on the way
1215 * to idle and has not yet set rq->curr to idle then it will
1216 * be serialized on the timer wheel base lock and take the new
1217 * timer into account automatically.
1219 if (rq->curr != rq->idle)
1223 * We can set TIF_RESCHED on the idle task of the other CPU
1224 * lockless. The worst case is that the other CPU runs the
1225 * idle task through an additional NOOP schedule()
1227 set_tsk_need_resched(rq->idle);
1229 /* NEED_RESCHED must be visible before we test polling */
1231 if (!tsk_is_polling(rq->idle))
1232 smp_send_reschedule(cpu);
1235 int nohz_ratelimit(int cpu)
1237 struct rq *rq = cpu_rq(cpu);
1238 u64 diff = rq->clock - rq->nohz_stamp;
1240 rq->nohz_stamp = rq->clock;
1242 return diff < (NSEC_PER_SEC / HZ) >> 1;
1245 #endif /* CONFIG_NO_HZ */
1247 static u64 sched_avg_period(void)
1249 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1252 static void sched_avg_update(struct rq *rq)
1254 s64 period = sched_avg_period();
1256 while ((s64)(rq->clock - rq->age_stamp) > period) {
1257 rq->age_stamp += period;
1262 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1264 rq->rt_avg += rt_delta;
1265 sched_avg_update(rq);
1268 #else /* !CONFIG_SMP */
1269 static void resched_task(struct task_struct *p)
1271 assert_raw_spin_locked(&task_rq(p)->lock);
1272 set_tsk_need_resched(p);
1275 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1278 #endif /* CONFIG_SMP */
1280 #if BITS_PER_LONG == 32
1281 # define WMULT_CONST (~0UL)
1283 # define WMULT_CONST (1UL << 32)
1286 #define WMULT_SHIFT 32
1289 * Shift right and round:
1291 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1294 * delta *= weight / lw
1296 static unsigned long
1297 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1298 struct load_weight *lw)
1302 if (!lw->inv_weight) {
1303 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1306 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1310 tmp = (u64)delta_exec * weight;
1312 * Check whether we'd overflow the 64-bit multiplication:
1314 if (unlikely(tmp > WMULT_CONST))
1315 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1318 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1320 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1323 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1329 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1336 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1337 * of tasks with abnormal "nice" values across CPUs the contribution that
1338 * each task makes to its run queue's load is weighted according to its
1339 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1340 * scaled version of the new time slice allocation that they receive on time
1344 #define WEIGHT_IDLEPRIO 3
1345 #define WMULT_IDLEPRIO 1431655765
1348 * Nice levels are multiplicative, with a gentle 10% change for every
1349 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1350 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1351 * that remained on nice 0.
1353 * The "10% effect" is relative and cumulative: from _any_ nice level,
1354 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1355 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1356 * If a task goes up by ~10% and another task goes down by ~10% then
1357 * the relative distance between them is ~25%.)
1359 static const int prio_to_weight[40] = {
1360 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1361 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1362 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1363 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1364 /* 0 */ 1024, 820, 655, 526, 423,
1365 /* 5 */ 335, 272, 215, 172, 137,
1366 /* 10 */ 110, 87, 70, 56, 45,
1367 /* 15 */ 36, 29, 23, 18, 15,
1371 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1373 * In cases where the weight does not change often, we can use the
1374 * precalculated inverse to speed up arithmetics by turning divisions
1375 * into multiplications:
1377 static const u32 prio_to_wmult[40] = {
1378 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1379 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1380 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1381 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1382 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1383 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1384 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1385 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1388 /* Time spent by the tasks of the cpu accounting group executing in ... */
1389 enum cpuacct_stat_index {
1390 CPUACCT_STAT_USER, /* ... user mode */
1391 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1393 CPUACCT_STAT_NSTATS,
1396 #ifdef CONFIG_CGROUP_CPUACCT
1397 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1398 static void cpuacct_update_stats(struct task_struct *tsk,
1399 enum cpuacct_stat_index idx, cputime_t val);
1401 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1402 static inline void cpuacct_update_stats(struct task_struct *tsk,
1403 enum cpuacct_stat_index idx, cputime_t val) {}
1406 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1408 update_load_add(&rq->load, load);
1411 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1413 update_load_sub(&rq->load, load);
1416 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1417 typedef int (*tg_visitor)(struct task_group *, void *);
1420 * Iterate the full tree, calling @down when first entering a node and @up when
1421 * leaving it for the final time.
1423 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1425 struct task_group *parent, *child;
1429 parent = &root_task_group;
1431 ret = (*down)(parent, data);
1434 list_for_each_entry_rcu(child, &parent->children, siblings) {
1441 ret = (*up)(parent, data);
1446 parent = parent->parent;
1455 static int tg_nop(struct task_group *tg, void *data)
1462 /* Used instead of source_load when we know the type == 0 */
1463 static unsigned long weighted_cpuload(const int cpu)
1465 return cpu_rq(cpu)->load.weight;
1469 * Return a low guess at the load of a migration-source cpu weighted
1470 * according to the scheduling class and "nice" value.
1472 * We want to under-estimate the load of migration sources, to
1473 * balance conservatively.
1475 static unsigned long source_load(int cpu, int type)
1477 struct rq *rq = cpu_rq(cpu);
1478 unsigned long total = weighted_cpuload(cpu);
1480 if (type == 0 || !sched_feat(LB_BIAS))
1483 return min(rq->cpu_load[type-1], total);
1487 * Return a high guess at the load of a migration-target cpu weighted
1488 * according to the scheduling class and "nice" value.
1490 static unsigned long target_load(int cpu, int type)
1492 struct rq *rq = cpu_rq(cpu);
1493 unsigned long total = weighted_cpuload(cpu);
1495 if (type == 0 || !sched_feat(LB_BIAS))
1498 return max(rq->cpu_load[type-1], total);
1501 static struct sched_group *group_of(int cpu)
1503 struct sched_domain *sd = rcu_dereference_sched(cpu_rq(cpu)->sd);
1511 static unsigned long power_of(int cpu)
1513 struct sched_group *group = group_of(cpu);
1516 return SCHED_LOAD_SCALE;
1518 return group->cpu_power;
1521 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1523 static unsigned long cpu_avg_load_per_task(int cpu)
1525 struct rq *rq = cpu_rq(cpu);
1526 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1529 rq->avg_load_per_task = rq->load.weight / nr_running;
1531 rq->avg_load_per_task = 0;
1533 return rq->avg_load_per_task;
1536 #ifdef CONFIG_FAIR_GROUP_SCHED
1538 static __read_mostly unsigned long __percpu *update_shares_data;
1540 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1543 * Calculate and set the cpu's group shares.
1545 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1546 unsigned long sd_shares,
1547 unsigned long sd_rq_weight,
1548 unsigned long *usd_rq_weight)
1550 unsigned long shares, rq_weight;
1553 rq_weight = usd_rq_weight[cpu];
1556 rq_weight = NICE_0_LOAD;
1560 * \Sum_j shares_j * rq_weight_i
1561 * shares_i = -----------------------------
1562 * \Sum_j rq_weight_j
1564 shares = (sd_shares * rq_weight) / sd_rq_weight;
1565 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1567 if (abs(shares - tg->se[cpu]->load.weight) >
1568 sysctl_sched_shares_thresh) {
1569 struct rq *rq = cpu_rq(cpu);
1570 unsigned long flags;
1572 raw_spin_lock_irqsave(&rq->lock, flags);
1573 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1574 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1575 __set_se_shares(tg->se[cpu], shares);
1576 raw_spin_unlock_irqrestore(&rq->lock, flags);
1581 * Re-compute the task group their per cpu shares over the given domain.
1582 * This needs to be done in a bottom-up fashion because the rq weight of a
1583 * parent group depends on the shares of its child groups.
1585 static int tg_shares_up(struct task_group *tg, void *data)
1587 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1588 unsigned long *usd_rq_weight;
1589 struct sched_domain *sd = data;
1590 unsigned long flags;
1596 local_irq_save(flags);
1597 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1599 for_each_cpu(i, sched_domain_span(sd)) {
1600 weight = tg->cfs_rq[i]->load.weight;
1601 usd_rq_weight[i] = weight;
1603 rq_weight += weight;
1605 * If there are currently no tasks on the cpu pretend there
1606 * is one of average load so that when a new task gets to
1607 * run here it will not get delayed by group starvation.
1610 weight = NICE_0_LOAD;
1612 sum_weight += weight;
1613 shares += tg->cfs_rq[i]->shares;
1617 rq_weight = sum_weight;
1619 if ((!shares && rq_weight) || shares > tg->shares)
1620 shares = tg->shares;
1622 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1623 shares = tg->shares;
1625 for_each_cpu(i, sched_domain_span(sd))
1626 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1628 local_irq_restore(flags);
1634 * Compute the cpu's hierarchical load factor for each task group.
1635 * This needs to be done in a top-down fashion because the load of a child
1636 * group is a fraction of its parents load.
1638 static int tg_load_down(struct task_group *tg, void *data)
1641 long cpu = (long)data;
1644 load = cpu_rq(cpu)->load.weight;
1646 load = tg->parent->cfs_rq[cpu]->h_load;
1647 load *= tg->cfs_rq[cpu]->shares;
1648 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1651 tg->cfs_rq[cpu]->h_load = load;
1656 static void update_shares(struct sched_domain *sd)
1661 if (root_task_group_empty())
1664 now = cpu_clock(raw_smp_processor_id());
1665 elapsed = now - sd->last_update;
1667 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1668 sd->last_update = now;
1669 walk_tg_tree(tg_nop, tg_shares_up, sd);
1673 static void update_h_load(long cpu)
1675 if (root_task_group_empty())
1678 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1683 static inline void update_shares(struct sched_domain *sd)
1689 #ifdef CONFIG_PREEMPT
1691 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1694 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1695 * way at the expense of forcing extra atomic operations in all
1696 * invocations. This assures that the double_lock is acquired using the
1697 * same underlying policy as the spinlock_t on this architecture, which
1698 * reduces latency compared to the unfair variant below. However, it
1699 * also adds more overhead and therefore may reduce throughput.
1701 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1702 __releases(this_rq->lock)
1703 __acquires(busiest->lock)
1704 __acquires(this_rq->lock)
1706 raw_spin_unlock(&this_rq->lock);
1707 double_rq_lock(this_rq, busiest);
1714 * Unfair double_lock_balance: Optimizes throughput at the expense of
1715 * latency by eliminating extra atomic operations when the locks are
1716 * already in proper order on entry. This favors lower cpu-ids and will
1717 * grant the double lock to lower cpus over higher ids under contention,
1718 * regardless of entry order into the function.
1720 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1721 __releases(this_rq->lock)
1722 __acquires(busiest->lock)
1723 __acquires(this_rq->lock)
1727 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1728 if (busiest < this_rq) {
1729 raw_spin_unlock(&this_rq->lock);
1730 raw_spin_lock(&busiest->lock);
1731 raw_spin_lock_nested(&this_rq->lock,
1732 SINGLE_DEPTH_NESTING);
1735 raw_spin_lock_nested(&busiest->lock,
1736 SINGLE_DEPTH_NESTING);
1741 #endif /* CONFIG_PREEMPT */
1744 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1746 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1748 if (unlikely(!irqs_disabled())) {
1749 /* printk() doesn't work good under rq->lock */
1750 raw_spin_unlock(&this_rq->lock);
1754 return _double_lock_balance(this_rq, busiest);
1757 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1758 __releases(busiest->lock)
1760 raw_spin_unlock(&busiest->lock);
1761 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1765 * double_rq_lock - safely lock two runqueues
1767 * Note this does not disable interrupts like task_rq_lock,
1768 * you need to do so manually before calling.
1770 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1771 __acquires(rq1->lock)
1772 __acquires(rq2->lock)
1774 BUG_ON(!irqs_disabled());
1776 raw_spin_lock(&rq1->lock);
1777 __acquire(rq2->lock); /* Fake it out ;) */
1780 raw_spin_lock(&rq1->lock);
1781 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1783 raw_spin_lock(&rq2->lock);
1784 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1790 * double_rq_unlock - safely unlock two runqueues
1792 * Note this does not restore interrupts like task_rq_unlock,
1793 * you need to do so manually after calling.
1795 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1796 __releases(rq1->lock)
1797 __releases(rq2->lock)
1799 raw_spin_unlock(&rq1->lock);
1801 raw_spin_unlock(&rq2->lock);
1803 __release(rq2->lock);
1808 #ifdef CONFIG_FAIR_GROUP_SCHED
1809 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1812 cfs_rq->shares = shares;
1817 static void calc_load_account_active(struct rq *this_rq);
1818 static void update_sysctl(void);
1819 static int get_update_sysctl_factor(void);
1821 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1823 set_task_rq(p, cpu);
1826 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1827 * successfuly executed on another CPU. We must ensure that updates of
1828 * per-task data have been completed by this moment.
1831 task_thread_info(p)->cpu = cpu;
1835 static const struct sched_class rt_sched_class;
1837 #define sched_class_highest (&rt_sched_class)
1838 #define for_each_class(class) \
1839 for (class = sched_class_highest; class; class = class->next)
1841 #include "sched_stats.h"
1843 static void inc_nr_running(struct rq *rq)
1848 static void dec_nr_running(struct rq *rq)
1853 static void set_load_weight(struct task_struct *p)
1855 if (task_has_rt_policy(p)) {
1856 p->se.load.weight = prio_to_weight[0] * 2;
1857 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1862 * SCHED_IDLE tasks get minimal weight:
1864 if (p->policy == SCHED_IDLE) {
1865 p->se.load.weight = WEIGHT_IDLEPRIO;
1866 p->se.load.inv_weight = WMULT_IDLEPRIO;
1870 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1871 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1874 static void update_avg(u64 *avg, u64 sample)
1876 s64 diff = sample - *avg;
1880 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1882 update_rq_clock(rq);
1883 sched_info_queued(p);
1884 p->sched_class->enqueue_task(rq, p, flags);
1888 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1890 update_rq_clock(rq);
1891 sched_info_dequeued(p);
1892 p->sched_class->dequeue_task(rq, p, flags);
1897 * activate_task - move a task to the runqueue.
1899 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1901 if (task_contributes_to_load(p))
1902 rq->nr_uninterruptible--;
1904 enqueue_task(rq, p, flags);
1909 * deactivate_task - remove a task from the runqueue.
1911 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1913 if (task_contributes_to_load(p))
1914 rq->nr_uninterruptible++;
1916 dequeue_task(rq, p, flags);
1920 #include "sched_idletask.c"
1921 #include "sched_fair.c"
1922 #include "sched_rt.c"
1923 #ifdef CONFIG_SCHED_DEBUG
1924 # include "sched_debug.c"
1928 * __normal_prio - return the priority that is based on the static prio
1930 static inline int __normal_prio(struct task_struct *p)
1932 return p->static_prio;
1936 * Calculate the expected normal priority: i.e. priority
1937 * without taking RT-inheritance into account. Might be
1938 * boosted by interactivity modifiers. Changes upon fork,
1939 * setprio syscalls, and whenever the interactivity
1940 * estimator recalculates.
1942 static inline int normal_prio(struct task_struct *p)
1946 if (task_has_rt_policy(p))
1947 prio = MAX_RT_PRIO-1 - p->rt_priority;
1949 prio = __normal_prio(p);
1954 * Calculate the current priority, i.e. the priority
1955 * taken into account by the scheduler. This value might
1956 * be boosted by RT tasks, or might be boosted by
1957 * interactivity modifiers. Will be RT if the task got
1958 * RT-boosted. If not then it returns p->normal_prio.
1960 static int effective_prio(struct task_struct *p)
1962 p->normal_prio = normal_prio(p);
1964 * If we are RT tasks or we were boosted to RT priority,
1965 * keep the priority unchanged. Otherwise, update priority
1966 * to the normal priority:
1968 if (!rt_prio(p->prio))
1969 return p->normal_prio;
1974 * task_curr - is this task currently executing on a CPU?
1975 * @p: the task in question.
1977 inline int task_curr(const struct task_struct *p)
1979 return cpu_curr(task_cpu(p)) == p;
1982 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1983 const struct sched_class *prev_class,
1984 int oldprio, int running)
1986 if (prev_class != p->sched_class) {
1987 if (prev_class->switched_from)
1988 prev_class->switched_from(rq, p, running);
1989 p->sched_class->switched_to(rq, p, running);
1991 p->sched_class->prio_changed(rq, p, oldprio, running);
1996 * Is this task likely cache-hot:
1999 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2003 if (p->sched_class != &fair_sched_class)
2007 * Buddy candidates are cache hot:
2009 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2010 (&p->se == cfs_rq_of(&p->se)->next ||
2011 &p->se == cfs_rq_of(&p->se)->last))
2014 if (sysctl_sched_migration_cost == -1)
2016 if (sysctl_sched_migration_cost == 0)
2019 delta = now - p->se.exec_start;
2021 return delta < (s64)sysctl_sched_migration_cost;
2024 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2026 #ifdef CONFIG_SCHED_DEBUG
2028 * We should never call set_task_cpu() on a blocked task,
2029 * ttwu() will sort out the placement.
2031 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2032 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2035 trace_sched_migrate_task(p, new_cpu);
2037 if (task_cpu(p) != new_cpu) {
2038 p->se.nr_migrations++;
2039 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2042 __set_task_cpu(p, new_cpu);
2045 struct migration_req {
2046 struct list_head list;
2048 struct task_struct *task;
2051 struct completion done;
2055 * The task's runqueue lock must be held.
2056 * Returns true if you have to wait for migration thread.
2059 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2061 struct rq *rq = task_rq(p);
2064 * If the task is not on a runqueue (and not running), then
2065 * the next wake-up will properly place the task.
2067 if (!p->se.on_rq && !task_running(rq, p))
2070 init_completion(&req->done);
2072 req->dest_cpu = dest_cpu;
2073 list_add(&req->list, &rq->migration_queue);
2079 * wait_task_context_switch - wait for a thread to complete at least one
2082 * @p must not be current.
2084 void wait_task_context_switch(struct task_struct *p)
2086 unsigned long nvcsw, nivcsw, flags;
2094 * The runqueue is assigned before the actual context
2095 * switch. We need to take the runqueue lock.
2097 * We could check initially without the lock but it is
2098 * very likely that we need to take the lock in every
2101 rq = task_rq_lock(p, &flags);
2102 running = task_running(rq, p);
2103 task_rq_unlock(rq, &flags);
2105 if (likely(!running))
2108 * The switch count is incremented before the actual
2109 * context switch. We thus wait for two switches to be
2110 * sure at least one completed.
2112 if ((p->nvcsw - nvcsw) > 1)
2114 if ((p->nivcsw - nivcsw) > 1)
2122 * wait_task_inactive - wait for a thread to unschedule.
2124 * If @match_state is nonzero, it's the @p->state value just checked and
2125 * not expected to change. If it changes, i.e. @p might have woken up,
2126 * then return zero. When we succeed in waiting for @p to be off its CPU,
2127 * we return a positive number (its total switch count). If a second call
2128 * a short while later returns the same number, the caller can be sure that
2129 * @p has remained unscheduled the whole time.
2131 * The caller must ensure that the task *will* unschedule sometime soon,
2132 * else this function might spin for a *long* time. This function can't
2133 * be called with interrupts off, or it may introduce deadlock with
2134 * smp_call_function() if an IPI is sent by the same process we are
2135 * waiting to become inactive.
2137 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2139 unsigned long flags;
2146 * We do the initial early heuristics without holding
2147 * any task-queue locks at all. We'll only try to get
2148 * the runqueue lock when things look like they will
2154 * If the task is actively running on another CPU
2155 * still, just relax and busy-wait without holding
2158 * NOTE! Since we don't hold any locks, it's not
2159 * even sure that "rq" stays as the right runqueue!
2160 * But we don't care, since "task_running()" will
2161 * return false if the runqueue has changed and p
2162 * is actually now running somewhere else!
2164 while (task_running(rq, p)) {
2165 if (match_state && unlikely(p->state != match_state))
2171 * Ok, time to look more closely! We need the rq
2172 * lock now, to be *sure*. If we're wrong, we'll
2173 * just go back and repeat.
2175 rq = task_rq_lock(p, &flags);
2176 trace_sched_wait_task(rq, p);
2177 running = task_running(rq, p);
2178 on_rq = p->se.on_rq;
2180 if (!match_state || p->state == match_state)
2181 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2182 task_rq_unlock(rq, &flags);
2185 * If it changed from the expected state, bail out now.
2187 if (unlikely(!ncsw))
2191 * Was it really running after all now that we
2192 * checked with the proper locks actually held?
2194 * Oops. Go back and try again..
2196 if (unlikely(running)) {
2202 * It's not enough that it's not actively running,
2203 * it must be off the runqueue _entirely_, and not
2206 * So if it was still runnable (but just not actively
2207 * running right now), it's preempted, and we should
2208 * yield - it could be a while.
2210 if (unlikely(on_rq)) {
2211 schedule_timeout_uninterruptible(1);
2216 * Ahh, all good. It wasn't running, and it wasn't
2217 * runnable, which means that it will never become
2218 * running in the future either. We're all done!
2227 * kick_process - kick a running thread to enter/exit the kernel
2228 * @p: the to-be-kicked thread
2230 * Cause a process which is running on another CPU to enter
2231 * kernel-mode, without any delay. (to get signals handled.)
2233 * NOTE: this function doesnt have to take the runqueue lock,
2234 * because all it wants to ensure is that the remote task enters
2235 * the kernel. If the IPI races and the task has been migrated
2236 * to another CPU then no harm is done and the purpose has been
2239 void kick_process(struct task_struct *p)
2245 if ((cpu != smp_processor_id()) && task_curr(p))
2246 smp_send_reschedule(cpu);
2249 EXPORT_SYMBOL_GPL(kick_process);
2250 #endif /* CONFIG_SMP */
2253 * task_oncpu_function_call - call a function on the cpu on which a task runs
2254 * @p: the task to evaluate
2255 * @func: the function to be called
2256 * @info: the function call argument
2258 * Calls the function @func when the task is currently running. This might
2259 * be on the current CPU, which just calls the function directly
2261 void task_oncpu_function_call(struct task_struct *p,
2262 void (*func) (void *info), void *info)
2269 smp_call_function_single(cpu, func, info, 1);
2275 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2277 static int select_fallback_rq(int cpu, struct task_struct *p)
2280 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2282 /* Look for allowed, online CPU in same node. */
2283 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2284 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2287 /* Any allowed, online CPU? */
2288 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2289 if (dest_cpu < nr_cpu_ids)
2292 /* No more Mr. Nice Guy. */
2293 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2294 dest_cpu = cpuset_cpus_allowed_fallback(p);
2296 * Don't tell them about moving exiting tasks or
2297 * kernel threads (both mm NULL), since they never
2300 if (p->mm && printk_ratelimit()) {
2301 printk(KERN_INFO "process %d (%s) no "
2302 "longer affine to cpu%d\n",
2303 task_pid_nr(p), p->comm, cpu);
2311 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2314 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2316 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2319 * In order not to call set_task_cpu() on a blocking task we need
2320 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2323 * Since this is common to all placement strategies, this lives here.
2325 * [ this allows ->select_task() to simply return task_cpu(p) and
2326 * not worry about this generic constraint ]
2328 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2330 cpu = select_fallback_rq(task_cpu(p), p);
2337 * try_to_wake_up - wake up a thread
2338 * @p: the to-be-woken-up thread
2339 * @state: the mask of task states that can be woken
2340 * @sync: do a synchronous wakeup?
2342 * Put it on the run-queue if it's not already there. The "current"
2343 * thread is always on the run-queue (except when the actual
2344 * re-schedule is in progress), and as such you're allowed to do
2345 * the simpler "current->state = TASK_RUNNING" to mark yourself
2346 * runnable without the overhead of this.
2348 * returns failure only if the task is already active.
2350 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2353 int cpu, orig_cpu, this_cpu, success = 0;
2354 unsigned long flags;
2355 unsigned long en_flags = ENQUEUE_WAKEUP;
2358 this_cpu = get_cpu();
2361 rq = task_rq_lock(p, &flags);
2362 if (!(p->state & state))
2372 if (unlikely(task_running(rq, p)))
2376 * In order to handle concurrent wakeups and release the rq->lock
2377 * we put the task in TASK_WAKING state.
2379 * First fix up the nr_uninterruptible count:
2381 if (task_contributes_to_load(p)) {
2382 if (likely(cpu_online(orig_cpu)))
2383 rq->nr_uninterruptible--;
2385 this_rq()->nr_uninterruptible--;
2387 p->state = TASK_WAKING;
2389 if (p->sched_class->task_waking) {
2390 p->sched_class->task_waking(rq, p);
2391 en_flags |= ENQUEUE_WAKING;
2394 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2395 if (cpu != orig_cpu)
2396 set_task_cpu(p, cpu);
2397 __task_rq_unlock(rq);
2400 raw_spin_lock(&rq->lock);
2403 * We migrated the task without holding either rq->lock, however
2404 * since the task is not on the task list itself, nobody else
2405 * will try and migrate the task, hence the rq should match the
2406 * cpu we just moved it to.
2408 WARN_ON(task_cpu(p) != cpu);
2409 WARN_ON(p->state != TASK_WAKING);
2411 #ifdef CONFIG_SCHEDSTATS
2412 schedstat_inc(rq, ttwu_count);
2413 if (cpu == this_cpu)
2414 schedstat_inc(rq, ttwu_local);
2416 struct sched_domain *sd;
2417 for_each_domain(this_cpu, sd) {
2418 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2419 schedstat_inc(sd, ttwu_wake_remote);
2424 #endif /* CONFIG_SCHEDSTATS */
2427 #endif /* CONFIG_SMP */
2428 schedstat_inc(p, se.statistics.nr_wakeups);
2429 if (wake_flags & WF_SYNC)
2430 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2431 if (orig_cpu != cpu)
2432 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2433 if (cpu == this_cpu)
2434 schedstat_inc(p, se.statistics.nr_wakeups_local);
2436 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2437 activate_task(rq, p, en_flags);
2441 trace_sched_wakeup(rq, p, success);
2442 check_preempt_curr(rq, p, wake_flags);
2444 p->state = TASK_RUNNING;
2446 if (p->sched_class->task_woken)
2447 p->sched_class->task_woken(rq, p);
2449 if (unlikely(rq->idle_stamp)) {
2450 u64 delta = rq->clock - rq->idle_stamp;
2451 u64 max = 2*sysctl_sched_migration_cost;
2456 update_avg(&rq->avg_idle, delta);
2461 task_rq_unlock(rq, &flags);
2468 * wake_up_process - Wake up a specific process
2469 * @p: The process to be woken up.
2471 * Attempt to wake up the nominated process and move it to the set of runnable
2472 * processes. Returns 1 if the process was woken up, 0 if it was already
2475 * It may be assumed that this function implies a write memory barrier before
2476 * changing the task state if and only if any tasks are woken up.
2478 int wake_up_process(struct task_struct *p)
2480 return try_to_wake_up(p, TASK_ALL, 0);
2482 EXPORT_SYMBOL(wake_up_process);
2484 int wake_up_state(struct task_struct *p, unsigned int state)
2486 return try_to_wake_up(p, state, 0);
2490 * Perform scheduler related setup for a newly forked process p.
2491 * p is forked by current.
2493 * __sched_fork() is basic setup used by init_idle() too:
2495 static void __sched_fork(struct task_struct *p)
2497 p->se.exec_start = 0;
2498 p->se.sum_exec_runtime = 0;
2499 p->se.prev_sum_exec_runtime = 0;
2500 p->se.nr_migrations = 0;
2502 #ifdef CONFIG_SCHEDSTATS
2503 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2506 INIT_LIST_HEAD(&p->rt.run_list);
2508 INIT_LIST_HEAD(&p->se.group_node);
2510 #ifdef CONFIG_PREEMPT_NOTIFIERS
2511 INIT_HLIST_HEAD(&p->preempt_notifiers);
2516 * fork()/clone()-time setup:
2518 void sched_fork(struct task_struct *p, int clone_flags)
2520 int cpu = get_cpu();
2524 * We mark the process as running here. This guarantees that
2525 * nobody will actually run it, and a signal or other external
2526 * event cannot wake it up and insert it on the runqueue either.
2528 p->state = TASK_RUNNING;
2531 * Revert to default priority/policy on fork if requested.
2533 if (unlikely(p->sched_reset_on_fork)) {
2534 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2535 p->policy = SCHED_NORMAL;
2536 p->normal_prio = p->static_prio;
2539 if (PRIO_TO_NICE(p->static_prio) < 0) {
2540 p->static_prio = NICE_TO_PRIO(0);
2541 p->normal_prio = p->static_prio;
2546 * We don't need the reset flag anymore after the fork. It has
2547 * fulfilled its duty:
2549 p->sched_reset_on_fork = 0;
2553 * Make sure we do not leak PI boosting priority to the child.
2555 p->prio = current->normal_prio;
2557 if (!rt_prio(p->prio))
2558 p->sched_class = &fair_sched_class;
2560 if (p->sched_class->task_fork)
2561 p->sched_class->task_fork(p);
2563 set_task_cpu(p, cpu);
2565 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2566 if (likely(sched_info_on()))
2567 memset(&p->sched_info, 0, sizeof(p->sched_info));
2569 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2572 #ifdef CONFIG_PREEMPT
2573 /* Want to start with kernel preemption disabled. */
2574 task_thread_info(p)->preempt_count = 1;
2576 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2582 * wake_up_new_task - wake up a newly created task for the first time.
2584 * This function will do some initial scheduler statistics housekeeping
2585 * that must be done for every newly created context, then puts the task
2586 * on the runqueue and wakes it.
2588 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2590 unsigned long flags;
2592 int cpu __maybe_unused = get_cpu();
2595 rq = task_rq_lock(p, &flags);
2596 p->state = TASK_WAKING;
2599 * Fork balancing, do it here and not earlier because:
2600 * - cpus_allowed can change in the fork path
2601 * - any previously selected cpu might disappear through hotplug
2603 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2604 * without people poking at ->cpus_allowed.
2606 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2607 set_task_cpu(p, cpu);
2609 p->state = TASK_RUNNING;
2610 task_rq_unlock(rq, &flags);
2613 rq = task_rq_lock(p, &flags);
2614 activate_task(rq, p, 0);
2615 trace_sched_wakeup_new(rq, p, 1);
2616 check_preempt_curr(rq, p, WF_FORK);
2618 if (p->sched_class->task_woken)
2619 p->sched_class->task_woken(rq, p);
2621 task_rq_unlock(rq, &flags);
2625 #ifdef CONFIG_PREEMPT_NOTIFIERS
2628 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2629 * @notifier: notifier struct to register
2631 void preempt_notifier_register(struct preempt_notifier *notifier)
2633 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2635 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2638 * preempt_notifier_unregister - no longer interested in preemption notifications
2639 * @notifier: notifier struct to unregister
2641 * This is safe to call from within a preemption notifier.
2643 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2645 hlist_del(¬ifier->link);
2647 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2649 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2651 struct preempt_notifier *notifier;
2652 struct hlist_node *node;
2654 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2655 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2659 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2660 struct task_struct *next)
2662 struct preempt_notifier *notifier;
2663 struct hlist_node *node;
2665 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2666 notifier->ops->sched_out(notifier, next);
2669 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2671 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2676 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2677 struct task_struct *next)
2681 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2684 * prepare_task_switch - prepare to switch tasks
2685 * @rq: the runqueue preparing to switch
2686 * @prev: the current task that is being switched out
2687 * @next: the task we are going to switch to.
2689 * This is called with the rq lock held and interrupts off. It must
2690 * be paired with a subsequent finish_task_switch after the context
2693 * prepare_task_switch sets up locking and calls architecture specific
2697 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2698 struct task_struct *next)
2700 fire_sched_out_preempt_notifiers(prev, next);
2701 prepare_lock_switch(rq, next);
2702 prepare_arch_switch(next);
2706 * finish_task_switch - clean up after a task-switch
2707 * @rq: runqueue associated with task-switch
2708 * @prev: the thread we just switched away from.
2710 * finish_task_switch must be called after the context switch, paired
2711 * with a prepare_task_switch call before the context switch.
2712 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2713 * and do any other architecture-specific cleanup actions.
2715 * Note that we may have delayed dropping an mm in context_switch(). If
2716 * so, we finish that here outside of the runqueue lock. (Doing it
2717 * with the lock held can cause deadlocks; see schedule() for
2720 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2721 __releases(rq->lock)
2723 struct mm_struct *mm = rq->prev_mm;
2729 * A task struct has one reference for the use as "current".
2730 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2731 * schedule one last time. The schedule call will never return, and
2732 * the scheduled task must drop that reference.
2733 * The test for TASK_DEAD must occur while the runqueue locks are
2734 * still held, otherwise prev could be scheduled on another cpu, die
2735 * there before we look at prev->state, and then the reference would
2737 * Manfred Spraul <manfred@colorfullife.com>
2739 prev_state = prev->state;
2740 finish_arch_switch(prev);
2741 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2742 local_irq_disable();
2743 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2744 perf_event_task_sched_in(current);
2745 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2747 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2748 finish_lock_switch(rq, prev);
2750 fire_sched_in_preempt_notifiers(current);
2753 if (unlikely(prev_state == TASK_DEAD)) {
2755 * Remove function-return probe instances associated with this
2756 * task and put them back on the free list.
2758 kprobe_flush_task(prev);
2759 put_task_struct(prev);
2765 /* assumes rq->lock is held */
2766 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2768 if (prev->sched_class->pre_schedule)
2769 prev->sched_class->pre_schedule(rq, prev);
2772 /* rq->lock is NOT held, but preemption is disabled */
2773 static inline void post_schedule(struct rq *rq)
2775 if (rq->post_schedule) {
2776 unsigned long flags;
2778 raw_spin_lock_irqsave(&rq->lock, flags);
2779 if (rq->curr->sched_class->post_schedule)
2780 rq->curr->sched_class->post_schedule(rq);
2781 raw_spin_unlock_irqrestore(&rq->lock, flags);
2783 rq->post_schedule = 0;
2789 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2793 static inline void post_schedule(struct rq *rq)
2800 * schedule_tail - first thing a freshly forked thread must call.
2801 * @prev: the thread we just switched away from.
2803 asmlinkage void schedule_tail(struct task_struct *prev)
2804 __releases(rq->lock)
2806 struct rq *rq = this_rq();
2808 finish_task_switch(rq, prev);
2811 * FIXME: do we need to worry about rq being invalidated by the
2816 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2817 /* In this case, finish_task_switch does not reenable preemption */
2820 if (current->set_child_tid)
2821 put_user(task_pid_vnr(current), current->set_child_tid);
2825 * context_switch - switch to the new MM and the new
2826 * thread's register state.
2829 context_switch(struct rq *rq, struct task_struct *prev,
2830 struct task_struct *next)
2832 struct mm_struct *mm, *oldmm;
2834 prepare_task_switch(rq, prev, next);
2835 trace_sched_switch(rq, prev, next);
2837 oldmm = prev->active_mm;
2839 * For paravirt, this is coupled with an exit in switch_to to
2840 * combine the page table reload and the switch backend into
2843 arch_start_context_switch(prev);
2846 next->active_mm = oldmm;
2847 atomic_inc(&oldmm->mm_count);
2848 enter_lazy_tlb(oldmm, next);
2850 switch_mm(oldmm, mm, next);
2852 if (likely(!prev->mm)) {
2853 prev->active_mm = NULL;
2854 rq->prev_mm = oldmm;
2857 * Since the runqueue lock will be released by the next
2858 * task (which is an invalid locking op but in the case
2859 * of the scheduler it's an obvious special-case), so we
2860 * do an early lockdep release here:
2862 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2863 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2866 /* Here we just switch the register state and the stack. */
2867 switch_to(prev, next, prev);
2871 * this_rq must be evaluated again because prev may have moved
2872 * CPUs since it called schedule(), thus the 'rq' on its stack
2873 * frame will be invalid.
2875 finish_task_switch(this_rq(), prev);
2879 * nr_running, nr_uninterruptible and nr_context_switches:
2881 * externally visible scheduler statistics: current number of runnable
2882 * threads, current number of uninterruptible-sleeping threads, total
2883 * number of context switches performed since bootup.
2885 unsigned long nr_running(void)
2887 unsigned long i, sum = 0;
2889 for_each_online_cpu(i)
2890 sum += cpu_rq(i)->nr_running;
2895 unsigned long nr_uninterruptible(void)
2897 unsigned long i, sum = 0;
2899 for_each_possible_cpu(i)
2900 sum += cpu_rq(i)->nr_uninterruptible;
2903 * Since we read the counters lockless, it might be slightly
2904 * inaccurate. Do not allow it to go below zero though:
2906 if (unlikely((long)sum < 0))
2912 unsigned long long nr_context_switches(void)
2915 unsigned long long sum = 0;
2917 for_each_possible_cpu(i)
2918 sum += cpu_rq(i)->nr_switches;
2923 unsigned long nr_iowait(void)
2925 unsigned long i, sum = 0;
2927 for_each_possible_cpu(i)
2928 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2933 unsigned long nr_iowait_cpu(void)
2935 struct rq *this = this_rq();
2936 return atomic_read(&this->nr_iowait);
2939 unsigned long this_cpu_load(void)
2941 struct rq *this = this_rq();
2942 return this->cpu_load[0];
2946 /* Variables and functions for calc_load */
2947 static atomic_long_t calc_load_tasks;
2948 static unsigned long calc_load_update;
2949 unsigned long avenrun[3];
2950 EXPORT_SYMBOL(avenrun);
2953 * get_avenrun - get the load average array
2954 * @loads: pointer to dest load array
2955 * @offset: offset to add
2956 * @shift: shift count to shift the result left
2958 * These values are estimates at best, so no need for locking.
2960 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2962 loads[0] = (avenrun[0] + offset) << shift;
2963 loads[1] = (avenrun[1] + offset) << shift;
2964 loads[2] = (avenrun[2] + offset) << shift;
2967 static unsigned long
2968 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2971 load += active * (FIXED_1 - exp);
2972 return load >> FSHIFT;
2976 * calc_load - update the avenrun load estimates 10 ticks after the
2977 * CPUs have updated calc_load_tasks.
2979 void calc_global_load(void)
2981 unsigned long upd = calc_load_update + 10;
2984 if (time_before(jiffies, upd))
2987 active = atomic_long_read(&calc_load_tasks);
2988 active = active > 0 ? active * FIXED_1 : 0;
2990 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2991 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2992 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2994 calc_load_update += LOAD_FREQ;
2998 * Either called from update_cpu_load() or from a cpu going idle
3000 static void calc_load_account_active(struct rq *this_rq)
3002 long nr_active, delta;
3004 nr_active = this_rq->nr_running;
3005 nr_active += (long) this_rq->nr_uninterruptible;
3007 if (nr_active != this_rq->calc_load_active) {
3008 delta = nr_active - this_rq->calc_load_active;
3009 this_rq->calc_load_active = nr_active;
3010 atomic_long_add(delta, &calc_load_tasks);
3015 * Update rq->cpu_load[] statistics. This function is usually called every
3016 * scheduler tick (TICK_NSEC).
3018 static void update_cpu_load(struct rq *this_rq)
3020 unsigned long this_load = this_rq->load.weight;
3023 this_rq->nr_load_updates++;
3025 /* Update our load: */
3026 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3027 unsigned long old_load, new_load;
3029 /* scale is effectively 1 << i now, and >> i divides by scale */
3031 old_load = this_rq->cpu_load[i];
3032 new_load = this_load;
3034 * Round up the averaging division if load is increasing. This
3035 * prevents us from getting stuck on 9 if the load is 10, for
3038 if (new_load > old_load)
3039 new_load += scale-1;
3040 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3043 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3044 this_rq->calc_load_update += LOAD_FREQ;
3045 calc_load_account_active(this_rq);
3052 * sched_exec - execve() is a valuable balancing opportunity, because at
3053 * this point the task has the smallest effective memory and cache footprint.
3055 void sched_exec(void)
3057 struct task_struct *p = current;
3058 struct migration_req req;
3059 unsigned long flags;
3063 rq = task_rq_lock(p, &flags);
3064 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3065 if (dest_cpu == smp_processor_id())
3069 * select_task_rq() can race against ->cpus_allowed
3071 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3072 likely(cpu_active(dest_cpu)) &&
3073 migrate_task(p, dest_cpu, &req)) {
3074 /* Need to wait for migration thread (might exit: take ref). */
3075 struct task_struct *mt = rq->migration_thread;
3077 get_task_struct(mt);
3078 task_rq_unlock(rq, &flags);
3079 wake_up_process(mt);
3080 put_task_struct(mt);
3081 wait_for_completion(&req.done);
3086 task_rq_unlock(rq, &flags);
3091 DEFINE_PER_CPU(struct kernel_stat, kstat);
3093 EXPORT_PER_CPU_SYMBOL(kstat);
3096 * Return any ns on the sched_clock that have not yet been accounted in
3097 * @p in case that task is currently running.
3099 * Called with task_rq_lock() held on @rq.
3101 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3105 if (task_current(rq, p)) {
3106 update_rq_clock(rq);
3107 ns = rq->clock - p->se.exec_start;
3115 unsigned long long task_delta_exec(struct task_struct *p)
3117 unsigned long flags;
3121 rq = task_rq_lock(p, &flags);
3122 ns = do_task_delta_exec(p, rq);
3123 task_rq_unlock(rq, &flags);
3129 * Return accounted runtime for the task.
3130 * In case the task is currently running, return the runtime plus current's
3131 * pending runtime that have not been accounted yet.
3133 unsigned long long task_sched_runtime(struct task_struct *p)
3135 unsigned long flags;
3139 rq = task_rq_lock(p, &flags);
3140 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3141 task_rq_unlock(rq, &flags);
3147 * Return sum_exec_runtime for the thread group.
3148 * In case the task is currently running, return the sum plus current's
3149 * pending runtime that have not been accounted yet.
3151 * Note that the thread group might have other running tasks as well,
3152 * so the return value not includes other pending runtime that other
3153 * running tasks might have.
3155 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3157 struct task_cputime totals;
3158 unsigned long flags;
3162 rq = task_rq_lock(p, &flags);
3163 thread_group_cputime(p, &totals);
3164 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3165 task_rq_unlock(rq, &flags);
3171 * Account user cpu time to a process.
3172 * @p: the process that the cpu time gets accounted to
3173 * @cputime: the cpu time spent in user space since the last update
3174 * @cputime_scaled: cputime scaled by cpu frequency
3176 void account_user_time(struct task_struct *p, cputime_t cputime,
3177 cputime_t cputime_scaled)
3179 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3182 /* Add user time to process. */
3183 p->utime = cputime_add(p->utime, cputime);
3184 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3185 account_group_user_time(p, cputime);
3187 /* Add user time to cpustat. */
3188 tmp = cputime_to_cputime64(cputime);
3189 if (TASK_NICE(p) > 0)
3190 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3192 cpustat->user = cputime64_add(cpustat->user, tmp);
3194 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3195 /* Account for user time used */
3196 acct_update_integrals(p);
3200 * Account guest cpu time to a process.
3201 * @p: the process that the cpu time gets accounted to
3202 * @cputime: the cpu time spent in virtual machine since the last update
3203 * @cputime_scaled: cputime scaled by cpu frequency
3205 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3206 cputime_t cputime_scaled)
3209 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3211 tmp = cputime_to_cputime64(cputime);
3213 /* Add guest time to process. */
3214 p->utime = cputime_add(p->utime, cputime);
3215 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3216 account_group_user_time(p, cputime);
3217 p->gtime = cputime_add(p->gtime, cputime);
3219 /* Add guest time to cpustat. */
3220 if (TASK_NICE(p) > 0) {
3221 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3222 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3224 cpustat->user = cputime64_add(cpustat->user, tmp);
3225 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3230 * Account system cpu time to a process.
3231 * @p: the process that the cpu time gets accounted to
3232 * @hardirq_offset: the offset to subtract from hardirq_count()
3233 * @cputime: the cpu time spent in kernel space since the last update
3234 * @cputime_scaled: cputime scaled by cpu frequency
3236 void account_system_time(struct task_struct *p, int hardirq_offset,
3237 cputime_t cputime, cputime_t cputime_scaled)
3239 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3242 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3243 account_guest_time(p, cputime, cputime_scaled);
3247 /* Add system time to process. */
3248 p->stime = cputime_add(p->stime, cputime);
3249 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3250 account_group_system_time(p, cputime);
3252 /* Add system time to cpustat. */
3253 tmp = cputime_to_cputime64(cputime);
3254 if (hardirq_count() - hardirq_offset)
3255 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3256 else if (softirq_count())
3257 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3259 cpustat->system = cputime64_add(cpustat->system, tmp);
3261 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3263 /* Account for system time used */
3264 acct_update_integrals(p);
3268 * Account for involuntary wait time.
3269 * @steal: the cpu time spent in involuntary wait
3271 void account_steal_time(cputime_t cputime)
3273 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3274 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3276 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3280 * Account for idle time.
3281 * @cputime: the cpu time spent in idle wait
3283 void account_idle_time(cputime_t cputime)
3285 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3286 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3287 struct rq *rq = this_rq();
3289 if (atomic_read(&rq->nr_iowait) > 0)
3290 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3292 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3295 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3298 * Account a single tick of cpu time.
3299 * @p: the process that the cpu time gets accounted to
3300 * @user_tick: indicates if the tick is a user or a system tick
3302 void account_process_tick(struct task_struct *p, int user_tick)
3304 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3305 struct rq *rq = this_rq();
3308 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3309 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3310 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3313 account_idle_time(cputime_one_jiffy);
3317 * Account multiple ticks of steal time.
3318 * @p: the process from which the cpu time has been stolen
3319 * @ticks: number of stolen ticks
3321 void account_steal_ticks(unsigned long ticks)
3323 account_steal_time(jiffies_to_cputime(ticks));
3327 * Account multiple ticks of idle time.
3328 * @ticks: number of stolen ticks
3330 void account_idle_ticks(unsigned long ticks)
3332 account_idle_time(jiffies_to_cputime(ticks));
3338 * Use precise platform statistics if available:
3340 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3341 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3347 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3349 struct task_cputime cputime;
3351 thread_group_cputime(p, &cputime);
3353 *ut = cputime.utime;
3354 *st = cputime.stime;
3358 #ifndef nsecs_to_cputime
3359 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3362 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3364 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3367 * Use CFS's precise accounting:
3369 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3374 temp = (u64)(rtime * utime);
3375 do_div(temp, total);
3376 utime = (cputime_t)temp;
3381 * Compare with previous values, to keep monotonicity:
3383 p->prev_utime = max(p->prev_utime, utime);
3384 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3386 *ut = p->prev_utime;
3387 *st = p->prev_stime;
3391 * Must be called with siglock held.
3393 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3395 struct signal_struct *sig = p->signal;
3396 struct task_cputime cputime;
3397 cputime_t rtime, utime, total;
3399 thread_group_cputime(p, &cputime);
3401 total = cputime_add(cputime.utime, cputime.stime);
3402 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3407 temp = (u64)(rtime * cputime.utime);
3408 do_div(temp, total);
3409 utime = (cputime_t)temp;
3413 sig->prev_utime = max(sig->prev_utime, utime);
3414 sig->prev_stime = max(sig->prev_stime,
3415 cputime_sub(rtime, sig->prev_utime));
3417 *ut = sig->prev_utime;
3418 *st = sig->prev_stime;
3423 * This function gets called by the timer code, with HZ frequency.
3424 * We call it with interrupts disabled.
3426 * It also gets called by the fork code, when changing the parent's
3429 void scheduler_tick(void)
3431 int cpu = smp_processor_id();
3432 struct rq *rq = cpu_rq(cpu);
3433 struct task_struct *curr = rq->curr;
3437 raw_spin_lock(&rq->lock);
3438 update_rq_clock(rq);
3439 update_cpu_load(rq);
3440 curr->sched_class->task_tick(rq, curr, 0);
3441 raw_spin_unlock(&rq->lock);
3443 perf_event_task_tick(curr);
3446 rq->idle_at_tick = idle_cpu(cpu);
3447 trigger_load_balance(rq, cpu);
3451 notrace unsigned long get_parent_ip(unsigned long addr)
3453 if (in_lock_functions(addr)) {
3454 addr = CALLER_ADDR2;
3455 if (in_lock_functions(addr))
3456 addr = CALLER_ADDR3;
3461 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3462 defined(CONFIG_PREEMPT_TRACER))
3464 void __kprobes add_preempt_count(int val)
3466 #ifdef CONFIG_DEBUG_PREEMPT
3470 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3473 preempt_count() += val;
3474 #ifdef CONFIG_DEBUG_PREEMPT
3476 * Spinlock count overflowing soon?
3478 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3481 if (preempt_count() == val)
3482 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3484 EXPORT_SYMBOL(add_preempt_count);
3486 void __kprobes sub_preempt_count(int val)
3488 #ifdef CONFIG_DEBUG_PREEMPT
3492 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3495 * Is the spinlock portion underflowing?
3497 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3498 !(preempt_count() & PREEMPT_MASK)))
3502 if (preempt_count() == val)
3503 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3504 preempt_count() -= val;
3506 EXPORT_SYMBOL(sub_preempt_count);
3511 * Print scheduling while atomic bug:
3513 static noinline void __schedule_bug(struct task_struct *prev)
3515 struct pt_regs *regs = get_irq_regs();
3517 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3518 prev->comm, prev->pid, preempt_count());
3520 debug_show_held_locks(prev);
3522 if (irqs_disabled())
3523 print_irqtrace_events(prev);
3532 * Various schedule()-time debugging checks and statistics:
3534 static inline void schedule_debug(struct task_struct *prev)
3537 * Test if we are atomic. Since do_exit() needs to call into
3538 * schedule() atomically, we ignore that path for now.
3539 * Otherwise, whine if we are scheduling when we should not be.
3541 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3542 __schedule_bug(prev);
3544 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3546 schedstat_inc(this_rq(), sched_count);
3547 #ifdef CONFIG_SCHEDSTATS
3548 if (unlikely(prev->lock_depth >= 0)) {
3549 schedstat_inc(this_rq(), bkl_count);
3550 schedstat_inc(prev, sched_info.bkl_count);
3555 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3558 update_rq_clock(rq);
3559 rq->skip_clock_update = 0;
3560 prev->sched_class->put_prev_task(rq, prev);
3564 * Pick up the highest-prio task:
3566 static inline struct task_struct *
3567 pick_next_task(struct rq *rq)
3569 const struct sched_class *class;
3570 struct task_struct *p;
3573 * Optimization: we know that if all tasks are in
3574 * the fair class we can call that function directly:
3576 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3577 p = fair_sched_class.pick_next_task(rq);
3582 class = sched_class_highest;
3584 p = class->pick_next_task(rq);
3588 * Will never be NULL as the idle class always
3589 * returns a non-NULL p:
3591 class = class->next;
3596 * schedule() is the main scheduler function.
3598 asmlinkage void __sched schedule(void)
3600 struct task_struct *prev, *next;
3601 unsigned long *switch_count;
3607 cpu = smp_processor_id();
3611 switch_count = &prev->nivcsw;
3613 release_kernel_lock(prev);
3614 need_resched_nonpreemptible:
3616 schedule_debug(prev);
3618 if (sched_feat(HRTICK))
3621 raw_spin_lock_irq(&rq->lock);
3622 clear_tsk_need_resched(prev);
3624 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3625 if (unlikely(signal_pending_state(prev->state, prev)))
3626 prev->state = TASK_RUNNING;
3628 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3629 switch_count = &prev->nvcsw;
3632 pre_schedule(rq, prev);
3634 if (unlikely(!rq->nr_running))
3635 idle_balance(cpu, rq);
3637 put_prev_task(rq, prev);
3638 next = pick_next_task(rq);
3640 if (likely(prev != next)) {
3641 sched_info_switch(prev, next);
3642 perf_event_task_sched_out(prev, next);
3648 context_switch(rq, prev, next); /* unlocks the rq */
3650 * the context switch might have flipped the stack from under
3651 * us, hence refresh the local variables.
3653 cpu = smp_processor_id();
3656 raw_spin_unlock_irq(&rq->lock);
3660 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3662 switch_count = &prev->nivcsw;
3663 goto need_resched_nonpreemptible;
3666 preempt_enable_no_resched();
3670 EXPORT_SYMBOL(schedule);
3672 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3674 * Look out! "owner" is an entirely speculative pointer
3675 * access and not reliable.
3677 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3682 if (!sched_feat(OWNER_SPIN))
3685 #ifdef CONFIG_DEBUG_PAGEALLOC
3687 * Need to access the cpu field knowing that
3688 * DEBUG_PAGEALLOC could have unmapped it if
3689 * the mutex owner just released it and exited.
3691 if (probe_kernel_address(&owner->cpu, cpu))
3698 * Even if the access succeeded (likely case),
3699 * the cpu field may no longer be valid.
3701 if (cpu >= nr_cpumask_bits)
3705 * We need to validate that we can do a
3706 * get_cpu() and that we have the percpu area.
3708 if (!cpu_online(cpu))
3715 * Owner changed, break to re-assess state.
3717 if (lock->owner != owner)
3721 * Is that owner really running on that cpu?
3723 if (task_thread_info(rq->curr) != owner || need_resched())
3733 #ifdef CONFIG_PREEMPT
3735 * this is the entry point to schedule() from in-kernel preemption
3736 * off of preempt_enable. Kernel preemptions off return from interrupt
3737 * occur there and call schedule directly.
3739 asmlinkage void __sched preempt_schedule(void)
3741 struct thread_info *ti = current_thread_info();
3744 * If there is a non-zero preempt_count or interrupts are disabled,
3745 * we do not want to preempt the current task. Just return..
3747 if (likely(ti->preempt_count || irqs_disabled()))
3751 add_preempt_count(PREEMPT_ACTIVE);
3753 sub_preempt_count(PREEMPT_ACTIVE);
3756 * Check again in case we missed a preemption opportunity
3757 * between schedule and now.
3760 } while (need_resched());
3762 EXPORT_SYMBOL(preempt_schedule);
3765 * this is the entry point to schedule() from kernel preemption
3766 * off of irq context.
3767 * Note, that this is called and return with irqs disabled. This will
3768 * protect us against recursive calling from irq.
3770 asmlinkage void __sched preempt_schedule_irq(void)
3772 struct thread_info *ti = current_thread_info();
3774 /* Catch callers which need to be fixed */
3775 BUG_ON(ti->preempt_count || !irqs_disabled());
3778 add_preempt_count(PREEMPT_ACTIVE);
3781 local_irq_disable();
3782 sub_preempt_count(PREEMPT_ACTIVE);
3785 * Check again in case we missed a preemption opportunity
3786 * between schedule and now.
3789 } while (need_resched());
3792 #endif /* CONFIG_PREEMPT */
3794 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3797 return try_to_wake_up(curr->private, mode, wake_flags);
3799 EXPORT_SYMBOL(default_wake_function);
3802 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3803 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3804 * number) then we wake all the non-exclusive tasks and one exclusive task.
3806 * There are circumstances in which we can try to wake a task which has already
3807 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3808 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3810 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3811 int nr_exclusive, int wake_flags, void *key)
3813 wait_queue_t *curr, *next;
3815 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3816 unsigned flags = curr->flags;
3818 if (curr->func(curr, mode, wake_flags, key) &&
3819 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3825 * __wake_up - wake up threads blocked on a waitqueue.
3827 * @mode: which threads
3828 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3829 * @key: is directly passed to the wakeup function
3831 * It may be assumed that this function implies a write memory barrier before
3832 * changing the task state if and only if any tasks are woken up.
3834 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3835 int nr_exclusive, void *key)
3837 unsigned long flags;
3839 spin_lock_irqsave(&q->lock, flags);
3840 __wake_up_common(q, mode, nr_exclusive, 0, key);
3841 spin_unlock_irqrestore(&q->lock, flags);
3843 EXPORT_SYMBOL(__wake_up);
3846 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3848 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3850 __wake_up_common(q, mode, 1, 0, NULL);
3853 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3855 __wake_up_common(q, mode, 1, 0, key);
3859 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3861 * @mode: which threads
3862 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3863 * @key: opaque value to be passed to wakeup targets
3865 * The sync wakeup differs that the waker knows that it will schedule
3866 * away soon, so while the target thread will be woken up, it will not
3867 * be migrated to another CPU - ie. the two threads are 'synchronized'
3868 * with each other. This can prevent needless bouncing between CPUs.
3870 * On UP it can prevent extra preemption.
3872 * It may be assumed that this function implies a write memory barrier before
3873 * changing the task state if and only if any tasks are woken up.
3875 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3876 int nr_exclusive, void *key)
3878 unsigned long flags;
3879 int wake_flags = WF_SYNC;
3884 if (unlikely(!nr_exclusive))
3887 spin_lock_irqsave(&q->lock, flags);
3888 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3889 spin_unlock_irqrestore(&q->lock, flags);
3891 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3894 * __wake_up_sync - see __wake_up_sync_key()
3896 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3898 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3900 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3903 * complete: - signals a single thread waiting on this completion
3904 * @x: holds the state of this particular completion
3906 * This will wake up a single thread waiting on this completion. Threads will be
3907 * awakened in the same order in which they were queued.
3909 * See also complete_all(), wait_for_completion() and related routines.
3911 * It may be assumed that this function implies a write memory barrier before
3912 * changing the task state if and only if any tasks are woken up.
3914 void complete(struct completion *x)
3916 unsigned long flags;
3918 spin_lock_irqsave(&x->wait.lock, flags);
3920 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3921 spin_unlock_irqrestore(&x->wait.lock, flags);
3923 EXPORT_SYMBOL(complete);
3926 * complete_all: - signals all threads waiting on this completion
3927 * @x: holds the state of this particular completion
3929 * This will wake up all threads waiting on this particular completion event.
3931 * It may be assumed that this function implies a write memory barrier before
3932 * changing the task state if and only if any tasks are woken up.
3934 void complete_all(struct completion *x)
3936 unsigned long flags;
3938 spin_lock_irqsave(&x->wait.lock, flags);
3939 x->done += UINT_MAX/2;
3940 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3941 spin_unlock_irqrestore(&x->wait.lock, flags);
3943 EXPORT_SYMBOL(complete_all);
3945 static inline long __sched
3946 do_wait_for_common(struct completion *x, long timeout, int state)
3949 DECLARE_WAITQUEUE(wait, current);
3951 wait.flags |= WQ_FLAG_EXCLUSIVE;
3952 __add_wait_queue_tail(&x->wait, &wait);
3954 if (signal_pending_state(state, current)) {
3955 timeout = -ERESTARTSYS;
3958 __set_current_state(state);
3959 spin_unlock_irq(&x->wait.lock);
3960 timeout = schedule_timeout(timeout);
3961 spin_lock_irq(&x->wait.lock);
3962 } while (!x->done && timeout);
3963 __remove_wait_queue(&x->wait, &wait);
3968 return timeout ?: 1;
3972 wait_for_common(struct completion *x, long timeout, int state)
3976 spin_lock_irq(&x->wait.lock);
3977 timeout = do_wait_for_common(x, timeout, state);
3978 spin_unlock_irq(&x->wait.lock);
3983 * wait_for_completion: - waits for completion of a task
3984 * @x: holds the state of this particular completion
3986 * This waits to be signaled for completion of a specific task. It is NOT
3987 * interruptible and there is no timeout.
3989 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3990 * and interrupt capability. Also see complete().
3992 void __sched wait_for_completion(struct completion *x)
3994 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3996 EXPORT_SYMBOL(wait_for_completion);
3999 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4000 * @x: holds the state of this particular completion
4001 * @timeout: timeout value in jiffies
4003 * This waits for either a completion of a specific task to be signaled or for a
4004 * specified timeout to expire. The timeout is in jiffies. It is not
4007 unsigned long __sched
4008 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4010 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4012 EXPORT_SYMBOL(wait_for_completion_timeout);
4015 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4016 * @x: holds the state of this particular completion
4018 * This waits for completion of a specific task to be signaled. It is
4021 int __sched wait_for_completion_interruptible(struct completion *x)
4023 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4024 if (t == -ERESTARTSYS)
4028 EXPORT_SYMBOL(wait_for_completion_interruptible);
4031 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4032 * @x: holds the state of this particular completion
4033 * @timeout: timeout value in jiffies
4035 * This waits for either a completion of a specific task to be signaled or for a
4036 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4038 unsigned long __sched
4039 wait_for_completion_interruptible_timeout(struct completion *x,
4040 unsigned long timeout)
4042 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4044 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4047 * wait_for_completion_killable: - waits for completion of a task (killable)
4048 * @x: holds the state of this particular completion
4050 * This waits to be signaled for completion of a specific task. It can be
4051 * interrupted by a kill signal.
4053 int __sched wait_for_completion_killable(struct completion *x)
4055 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4056 if (t == -ERESTARTSYS)
4060 EXPORT_SYMBOL(wait_for_completion_killable);
4063 * try_wait_for_completion - try to decrement a completion without blocking
4064 * @x: completion structure
4066 * Returns: 0 if a decrement cannot be done without blocking
4067 * 1 if a decrement succeeded.
4069 * If a completion is being used as a counting completion,
4070 * attempt to decrement the counter without blocking. This
4071 * enables us to avoid waiting if the resource the completion
4072 * is protecting is not available.
4074 bool try_wait_for_completion(struct completion *x)
4076 unsigned long flags;
4079 spin_lock_irqsave(&x->wait.lock, flags);
4084 spin_unlock_irqrestore(&x->wait.lock, flags);
4087 EXPORT_SYMBOL(try_wait_for_completion);
4090 * completion_done - Test to see if a completion has any waiters
4091 * @x: completion structure
4093 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4094 * 1 if there are no waiters.
4097 bool completion_done(struct completion *x)
4099 unsigned long flags;
4102 spin_lock_irqsave(&x->wait.lock, flags);
4105 spin_unlock_irqrestore(&x->wait.lock, flags);
4108 EXPORT_SYMBOL(completion_done);
4111 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4113 unsigned long flags;
4116 init_waitqueue_entry(&wait, current);
4118 __set_current_state(state);
4120 spin_lock_irqsave(&q->lock, flags);
4121 __add_wait_queue(q, &wait);
4122 spin_unlock(&q->lock);
4123 timeout = schedule_timeout(timeout);
4124 spin_lock_irq(&q->lock);
4125 __remove_wait_queue(q, &wait);
4126 spin_unlock_irqrestore(&q->lock, flags);
4131 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4133 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4135 EXPORT_SYMBOL(interruptible_sleep_on);
4138 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4140 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4142 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4144 void __sched sleep_on(wait_queue_head_t *q)
4146 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4148 EXPORT_SYMBOL(sleep_on);
4150 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4152 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4154 EXPORT_SYMBOL(sleep_on_timeout);
4156 #ifdef CONFIG_RT_MUTEXES
4159 * rt_mutex_setprio - set the current priority of a task
4161 * @prio: prio value (kernel-internal form)
4163 * This function changes the 'effective' priority of a task. It does
4164 * not touch ->normal_prio like __setscheduler().
4166 * Used by the rt_mutex code to implement priority inheritance logic.
4168 void rt_mutex_setprio(struct task_struct *p, int prio)
4170 unsigned long flags;
4171 int oldprio, on_rq, running;
4173 const struct sched_class *prev_class;
4175 BUG_ON(prio < 0 || prio > MAX_PRIO);
4177 rq = task_rq_lock(p, &flags);
4180 prev_class = p->sched_class;
4181 on_rq = p->se.on_rq;
4182 running = task_current(rq, p);
4184 dequeue_task(rq, p, 0);
4186 p->sched_class->put_prev_task(rq, p);
4189 p->sched_class = &rt_sched_class;
4191 p->sched_class = &fair_sched_class;
4196 p->sched_class->set_curr_task(rq);
4198 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4200 check_class_changed(rq, p, prev_class, oldprio, running);
4202 task_rq_unlock(rq, &flags);
4207 void set_user_nice(struct task_struct *p, long nice)
4209 int old_prio, delta, on_rq;
4210 unsigned long flags;
4213 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4216 * We have to be careful, if called from sys_setpriority(),
4217 * the task might be in the middle of scheduling on another CPU.
4219 rq = task_rq_lock(p, &flags);
4221 * The RT priorities are set via sched_setscheduler(), but we still
4222 * allow the 'normal' nice value to be set - but as expected
4223 * it wont have any effect on scheduling until the task is
4224 * SCHED_FIFO/SCHED_RR:
4226 if (task_has_rt_policy(p)) {
4227 p->static_prio = NICE_TO_PRIO(nice);
4230 on_rq = p->se.on_rq;
4232 dequeue_task(rq, p, 0);
4234 p->static_prio = NICE_TO_PRIO(nice);
4237 p->prio = effective_prio(p);
4238 delta = p->prio - old_prio;
4241 enqueue_task(rq, p, 0);
4243 * If the task increased its priority or is running and
4244 * lowered its priority, then reschedule its CPU:
4246 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4247 resched_task(rq->curr);
4250 task_rq_unlock(rq, &flags);
4252 EXPORT_SYMBOL(set_user_nice);
4255 * can_nice - check if a task can reduce its nice value
4259 int can_nice(const struct task_struct *p, const int nice)
4261 /* convert nice value [19,-20] to rlimit style value [1,40] */
4262 int nice_rlim = 20 - nice;
4264 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4265 capable(CAP_SYS_NICE));
4268 #ifdef __ARCH_WANT_SYS_NICE
4271 * sys_nice - change the priority of the current process.
4272 * @increment: priority increment
4274 * sys_setpriority is a more generic, but much slower function that
4275 * does similar things.
4277 SYSCALL_DEFINE1(nice, int, increment)
4282 * Setpriority might change our priority at the same moment.
4283 * We don't have to worry. Conceptually one call occurs first
4284 * and we have a single winner.
4286 if (increment < -40)
4291 nice = TASK_NICE(current) + increment;
4297 if (increment < 0 && !can_nice(current, nice))
4300 retval = security_task_setnice(current, nice);
4304 set_user_nice(current, nice);
4311 * task_prio - return the priority value of a given task.
4312 * @p: the task in question.
4314 * This is the priority value as seen by users in /proc.
4315 * RT tasks are offset by -200. Normal tasks are centered
4316 * around 0, value goes from -16 to +15.
4318 int task_prio(const struct task_struct *p)
4320 return p->prio - MAX_RT_PRIO;
4324 * task_nice - return the nice value of a given task.
4325 * @p: the task in question.
4327 int task_nice(const struct task_struct *p)
4329 return TASK_NICE(p);
4331 EXPORT_SYMBOL(task_nice);
4334 * idle_cpu - is a given cpu idle currently?
4335 * @cpu: the processor in question.
4337 int idle_cpu(int cpu)
4339 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4343 * idle_task - return the idle task for a given cpu.
4344 * @cpu: the processor in question.
4346 struct task_struct *idle_task(int cpu)
4348 return cpu_rq(cpu)->idle;
4352 * find_process_by_pid - find a process with a matching PID value.
4353 * @pid: the pid in question.
4355 static struct task_struct *find_process_by_pid(pid_t pid)
4357 return pid ? find_task_by_vpid(pid) : current;
4360 /* Actually do priority change: must hold rq lock. */
4362 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4364 BUG_ON(p->se.on_rq);
4367 p->rt_priority = prio;
4368 p->normal_prio = normal_prio(p);
4369 /* we are holding p->pi_lock already */
4370 p->prio = rt_mutex_getprio(p);
4371 if (rt_prio(p->prio))
4372 p->sched_class = &rt_sched_class;
4374 p->sched_class = &fair_sched_class;
4379 * check the target process has a UID that matches the current process's
4381 static bool check_same_owner(struct task_struct *p)
4383 const struct cred *cred = current_cred(), *pcred;
4387 pcred = __task_cred(p);
4388 match = (cred->euid == pcred->euid ||
4389 cred->euid == pcred->uid);
4394 static int __sched_setscheduler(struct task_struct *p, int policy,
4395 struct sched_param *param, bool user)
4397 int retval, oldprio, oldpolicy = -1, on_rq, running;
4398 unsigned long flags;
4399 const struct sched_class *prev_class;
4403 /* may grab non-irq protected spin_locks */
4404 BUG_ON(in_interrupt());
4406 /* double check policy once rq lock held */
4408 reset_on_fork = p->sched_reset_on_fork;
4409 policy = oldpolicy = p->policy;
4411 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4412 policy &= ~SCHED_RESET_ON_FORK;
4414 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4415 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4416 policy != SCHED_IDLE)
4421 * Valid priorities for SCHED_FIFO and SCHED_RR are
4422 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4423 * SCHED_BATCH and SCHED_IDLE is 0.
4425 if (param->sched_priority < 0 ||
4426 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4427 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4429 if (rt_policy(policy) != (param->sched_priority != 0))
4433 * Allow unprivileged RT tasks to decrease priority:
4435 if (user && !capable(CAP_SYS_NICE)) {
4436 if (rt_policy(policy)) {
4437 unsigned long rlim_rtprio;
4439 if (!lock_task_sighand(p, &flags))
4441 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4442 unlock_task_sighand(p, &flags);
4444 /* can't set/change the rt policy */
4445 if (policy != p->policy && !rlim_rtprio)
4448 /* can't increase priority */
4449 if (param->sched_priority > p->rt_priority &&
4450 param->sched_priority > rlim_rtprio)
4454 * Like positive nice levels, dont allow tasks to
4455 * move out of SCHED_IDLE either:
4457 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4460 /* can't change other user's priorities */
4461 if (!check_same_owner(p))
4464 /* Normal users shall not reset the sched_reset_on_fork flag */
4465 if (p->sched_reset_on_fork && !reset_on_fork)
4470 #ifdef CONFIG_RT_GROUP_SCHED
4472 * Do not allow realtime tasks into groups that have no runtime
4475 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4476 task_group(p)->rt_bandwidth.rt_runtime == 0)
4480 retval = security_task_setscheduler(p, policy, param);
4486 * make sure no PI-waiters arrive (or leave) while we are
4487 * changing the priority of the task:
4489 raw_spin_lock_irqsave(&p->pi_lock, flags);
4491 * To be able to change p->policy safely, the apropriate
4492 * runqueue lock must be held.
4494 rq = __task_rq_lock(p);
4495 /* recheck policy now with rq lock held */
4496 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4497 policy = oldpolicy = -1;
4498 __task_rq_unlock(rq);
4499 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4502 on_rq = p->se.on_rq;
4503 running = task_current(rq, p);
4505 deactivate_task(rq, p, 0);
4507 p->sched_class->put_prev_task(rq, p);
4509 p->sched_reset_on_fork = reset_on_fork;
4512 prev_class = p->sched_class;
4513 __setscheduler(rq, p, policy, param->sched_priority);
4516 p->sched_class->set_curr_task(rq);
4518 activate_task(rq, p, 0);
4520 check_class_changed(rq, p, prev_class, oldprio, running);
4522 __task_rq_unlock(rq);
4523 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4525 rt_mutex_adjust_pi(p);
4531 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4532 * @p: the task in question.
4533 * @policy: new policy.
4534 * @param: structure containing the new RT priority.
4536 * NOTE that the task may be already dead.
4538 int sched_setscheduler(struct task_struct *p, int policy,
4539 struct sched_param *param)
4541 return __sched_setscheduler(p, policy, param, true);
4543 EXPORT_SYMBOL_GPL(sched_setscheduler);
4546 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4547 * @p: the task in question.
4548 * @policy: new policy.
4549 * @param: structure containing the new RT priority.
4551 * Just like sched_setscheduler, only don't bother checking if the
4552 * current context has permission. For example, this is needed in
4553 * stop_machine(): we create temporary high priority worker threads,
4554 * but our caller might not have that capability.
4556 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4557 struct sched_param *param)
4559 return __sched_setscheduler(p, policy, param, false);
4563 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4565 struct sched_param lparam;
4566 struct task_struct *p;
4569 if (!param || pid < 0)
4571 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4576 p = find_process_by_pid(pid);
4578 retval = sched_setscheduler(p, policy, &lparam);
4585 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4586 * @pid: the pid in question.
4587 * @policy: new policy.
4588 * @param: structure containing the new RT priority.
4590 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4591 struct sched_param __user *, param)
4593 /* negative values for policy are not valid */
4597 return do_sched_setscheduler(pid, policy, param);
4601 * sys_sched_setparam - set/change the RT priority of a thread
4602 * @pid: the pid in question.
4603 * @param: structure containing the new RT priority.
4605 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4607 return do_sched_setscheduler(pid, -1, param);
4611 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4612 * @pid: the pid in question.
4614 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4616 struct task_struct *p;
4624 p = find_process_by_pid(pid);
4626 retval = security_task_getscheduler(p);
4629 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4636 * sys_sched_getparam - get the RT priority of a thread
4637 * @pid: the pid in question.
4638 * @param: structure containing the RT priority.
4640 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4642 struct sched_param lp;
4643 struct task_struct *p;
4646 if (!param || pid < 0)
4650 p = find_process_by_pid(pid);
4655 retval = security_task_getscheduler(p);
4659 lp.sched_priority = p->rt_priority;
4663 * This one might sleep, we cannot do it with a spinlock held ...
4665 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4674 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4676 cpumask_var_t cpus_allowed, new_mask;
4677 struct task_struct *p;
4683 p = find_process_by_pid(pid);
4690 /* Prevent p going away */
4694 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4698 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4700 goto out_free_cpus_allowed;
4703 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4706 retval = security_task_setscheduler(p, 0, NULL);
4710 cpuset_cpus_allowed(p, cpus_allowed);
4711 cpumask_and(new_mask, in_mask, cpus_allowed);
4713 retval = set_cpus_allowed_ptr(p, new_mask);
4716 cpuset_cpus_allowed(p, cpus_allowed);
4717 if (!cpumask_subset(new_mask, cpus_allowed)) {
4719 * We must have raced with a concurrent cpuset
4720 * update. Just reset the cpus_allowed to the
4721 * cpuset's cpus_allowed
4723 cpumask_copy(new_mask, cpus_allowed);
4728 free_cpumask_var(new_mask);
4729 out_free_cpus_allowed:
4730 free_cpumask_var(cpus_allowed);
4737 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4738 struct cpumask *new_mask)
4740 if (len < cpumask_size())
4741 cpumask_clear(new_mask);
4742 else if (len > cpumask_size())
4743 len = cpumask_size();
4745 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4749 * sys_sched_setaffinity - set the cpu affinity of a process
4750 * @pid: pid of the process
4751 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4752 * @user_mask_ptr: user-space pointer to the new cpu mask
4754 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4755 unsigned long __user *, user_mask_ptr)
4757 cpumask_var_t new_mask;
4760 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4763 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4765 retval = sched_setaffinity(pid, new_mask);
4766 free_cpumask_var(new_mask);
4770 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4772 struct task_struct *p;
4773 unsigned long flags;
4781 p = find_process_by_pid(pid);
4785 retval = security_task_getscheduler(p);
4789 rq = task_rq_lock(p, &flags);
4790 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4791 task_rq_unlock(rq, &flags);
4801 * sys_sched_getaffinity - get the cpu affinity of a process
4802 * @pid: pid of the process
4803 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4804 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4806 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4807 unsigned long __user *, user_mask_ptr)
4812 if (len < nr_cpu_ids)
4814 if (len & (sizeof(unsigned long)-1))
4817 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4820 ret = sched_getaffinity(pid, mask);
4822 size_t retlen = min_t(size_t, len, cpumask_size());
4824 if (copy_to_user(user_mask_ptr, mask, retlen))
4829 free_cpumask_var(mask);
4835 * sys_sched_yield - yield the current processor to other threads.
4837 * This function yields the current CPU to other tasks. If there are no
4838 * other threads running on this CPU then this function will return.
4840 SYSCALL_DEFINE0(sched_yield)
4842 struct rq *rq = this_rq_lock();
4844 schedstat_inc(rq, yld_count);
4845 current->sched_class->yield_task(rq);
4848 * Since we are going to call schedule() anyway, there's
4849 * no need to preempt or enable interrupts:
4851 __release(rq->lock);
4852 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4853 do_raw_spin_unlock(&rq->lock);
4854 preempt_enable_no_resched();
4861 static inline int should_resched(void)
4863 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4866 static void __cond_resched(void)
4868 add_preempt_count(PREEMPT_ACTIVE);
4870 sub_preempt_count(PREEMPT_ACTIVE);
4873 int __sched _cond_resched(void)
4875 if (should_resched()) {
4881 EXPORT_SYMBOL(_cond_resched);
4884 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4885 * call schedule, and on return reacquire the lock.
4887 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4888 * operations here to prevent schedule() from being called twice (once via
4889 * spin_unlock(), once by hand).
4891 int __cond_resched_lock(spinlock_t *lock)
4893 int resched = should_resched();
4896 lockdep_assert_held(lock);
4898 if (spin_needbreak(lock) || resched) {
4909 EXPORT_SYMBOL(__cond_resched_lock);
4911 int __sched __cond_resched_softirq(void)
4913 BUG_ON(!in_softirq());
4915 if (should_resched()) {
4923 EXPORT_SYMBOL(__cond_resched_softirq);
4926 * yield - yield the current processor to other threads.
4928 * This is a shortcut for kernel-space yielding - it marks the
4929 * thread runnable and calls sys_sched_yield().
4931 void __sched yield(void)
4933 set_current_state(TASK_RUNNING);
4936 EXPORT_SYMBOL(yield);
4939 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4940 * that process accounting knows that this is a task in IO wait state.
4942 void __sched io_schedule(void)
4944 struct rq *rq = raw_rq();
4946 delayacct_blkio_start();
4947 atomic_inc(&rq->nr_iowait);
4948 current->in_iowait = 1;
4950 current->in_iowait = 0;
4951 atomic_dec(&rq->nr_iowait);
4952 delayacct_blkio_end();
4954 EXPORT_SYMBOL(io_schedule);
4956 long __sched io_schedule_timeout(long timeout)
4958 struct rq *rq = raw_rq();
4961 delayacct_blkio_start();
4962 atomic_inc(&rq->nr_iowait);
4963 current->in_iowait = 1;
4964 ret = schedule_timeout(timeout);
4965 current->in_iowait = 0;
4966 atomic_dec(&rq->nr_iowait);
4967 delayacct_blkio_end();
4972 * sys_sched_get_priority_max - return maximum RT priority.
4973 * @policy: scheduling class.
4975 * this syscall returns the maximum rt_priority that can be used
4976 * by a given scheduling class.
4978 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4985 ret = MAX_USER_RT_PRIO-1;
4997 * sys_sched_get_priority_min - return minimum RT priority.
4998 * @policy: scheduling class.
5000 * this syscall returns the minimum rt_priority that can be used
5001 * by a given scheduling class.
5003 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5021 * sys_sched_rr_get_interval - return the default timeslice of a process.
5022 * @pid: pid of the process.
5023 * @interval: userspace pointer to the timeslice value.
5025 * this syscall writes the default timeslice value of a given process
5026 * into the user-space timespec buffer. A value of '0' means infinity.
5028 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5029 struct timespec __user *, interval)
5031 struct task_struct *p;
5032 unsigned int time_slice;
5033 unsigned long flags;
5043 p = find_process_by_pid(pid);
5047 retval = security_task_getscheduler(p);
5051 rq = task_rq_lock(p, &flags);
5052 time_slice = p->sched_class->get_rr_interval(rq, p);
5053 task_rq_unlock(rq, &flags);
5056 jiffies_to_timespec(time_slice, &t);
5057 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5065 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5067 void sched_show_task(struct task_struct *p)
5069 unsigned long free = 0;
5072 state = p->state ? __ffs(p->state) + 1 : 0;
5073 printk(KERN_INFO "%-13.13s %c", p->comm,
5074 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5075 #if BITS_PER_LONG == 32
5076 if (state == TASK_RUNNING)
5077 printk(KERN_CONT " running ");
5079 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5081 if (state == TASK_RUNNING)
5082 printk(KERN_CONT " running task ");
5084 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5086 #ifdef CONFIG_DEBUG_STACK_USAGE
5087 free = stack_not_used(p);
5089 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5090 task_pid_nr(p), task_pid_nr(p->real_parent),
5091 (unsigned long)task_thread_info(p)->flags);
5093 show_stack(p, NULL);
5096 void show_state_filter(unsigned long state_filter)
5098 struct task_struct *g, *p;
5100 #if BITS_PER_LONG == 32
5102 " task PC stack pid father\n");
5105 " task PC stack pid father\n");
5107 read_lock(&tasklist_lock);
5108 do_each_thread(g, p) {
5110 * reset the NMI-timeout, listing all files on a slow
5111 * console might take alot of time:
5113 touch_nmi_watchdog();
5114 if (!state_filter || (p->state & state_filter))
5116 } while_each_thread(g, p);
5118 touch_all_softlockup_watchdogs();
5120 #ifdef CONFIG_SCHED_DEBUG
5121 sysrq_sched_debug_show();
5123 read_unlock(&tasklist_lock);
5125 * Only show locks if all tasks are dumped:
5128 debug_show_all_locks();
5131 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5133 idle->sched_class = &idle_sched_class;
5137 * init_idle - set up an idle thread for a given CPU
5138 * @idle: task in question
5139 * @cpu: cpu the idle task belongs to
5141 * NOTE: this function does not set the idle thread's NEED_RESCHED
5142 * flag, to make booting more robust.
5144 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5146 struct rq *rq = cpu_rq(cpu);
5147 unsigned long flags;
5149 raw_spin_lock_irqsave(&rq->lock, flags);
5152 idle->state = TASK_RUNNING;
5153 idle->se.exec_start = sched_clock();
5155 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5156 __set_task_cpu(idle, cpu);
5158 rq->curr = rq->idle = idle;
5159 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5162 raw_spin_unlock_irqrestore(&rq->lock, flags);
5164 /* Set the preempt count _outside_ the spinlocks! */
5165 #if defined(CONFIG_PREEMPT)
5166 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5168 task_thread_info(idle)->preempt_count = 0;
5171 * The idle tasks have their own, simple scheduling class:
5173 idle->sched_class = &idle_sched_class;
5174 ftrace_graph_init_task(idle);
5178 * In a system that switches off the HZ timer nohz_cpu_mask
5179 * indicates which cpus entered this state. This is used
5180 * in the rcu update to wait only for active cpus. For system
5181 * which do not switch off the HZ timer nohz_cpu_mask should
5182 * always be CPU_BITS_NONE.
5184 cpumask_var_t nohz_cpu_mask;
5187 * Increase the granularity value when there are more CPUs,
5188 * because with more CPUs the 'effective latency' as visible
5189 * to users decreases. But the relationship is not linear,
5190 * so pick a second-best guess by going with the log2 of the
5193 * This idea comes from the SD scheduler of Con Kolivas:
5195 static int get_update_sysctl_factor(void)
5197 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5198 unsigned int factor;
5200 switch (sysctl_sched_tunable_scaling) {
5201 case SCHED_TUNABLESCALING_NONE:
5204 case SCHED_TUNABLESCALING_LINEAR:
5207 case SCHED_TUNABLESCALING_LOG:
5209 factor = 1 + ilog2(cpus);
5216 static void update_sysctl(void)
5218 unsigned int factor = get_update_sysctl_factor();
5220 #define SET_SYSCTL(name) \
5221 (sysctl_##name = (factor) * normalized_sysctl_##name)
5222 SET_SYSCTL(sched_min_granularity);
5223 SET_SYSCTL(sched_latency);
5224 SET_SYSCTL(sched_wakeup_granularity);
5225 SET_SYSCTL(sched_shares_ratelimit);
5229 static inline void sched_init_granularity(void)
5236 * This is how migration works:
5238 * 1) we queue a struct migration_req structure in the source CPU's
5239 * runqueue and wake up that CPU's migration thread.
5240 * 2) we down() the locked semaphore => thread blocks.
5241 * 3) migration thread wakes up (implicitly it forces the migrated
5242 * thread off the CPU)
5243 * 4) it gets the migration request and checks whether the migrated
5244 * task is still in the wrong runqueue.
5245 * 5) if it's in the wrong runqueue then the migration thread removes
5246 * it and puts it into the right queue.
5247 * 6) migration thread up()s the semaphore.
5248 * 7) we wake up and the migration is done.
5252 * Change a given task's CPU affinity. Migrate the thread to a
5253 * proper CPU and schedule it away if the CPU it's executing on
5254 * is removed from the allowed bitmask.
5256 * NOTE: the caller must have a valid reference to the task, the
5257 * task must not exit() & deallocate itself prematurely. The
5258 * call is not atomic; no spinlocks may be held.
5260 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5262 struct migration_req req;
5263 unsigned long flags;
5268 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5269 * drop the rq->lock and still rely on ->cpus_allowed.
5272 while (task_is_waking(p))
5274 rq = task_rq_lock(p, &flags);
5275 if (task_is_waking(p)) {
5276 task_rq_unlock(rq, &flags);
5280 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5285 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5286 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5291 if (p->sched_class->set_cpus_allowed)
5292 p->sched_class->set_cpus_allowed(p, new_mask);
5294 cpumask_copy(&p->cpus_allowed, new_mask);
5295 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5298 /* Can the task run on the task's current CPU? If so, we're done */
5299 if (cpumask_test_cpu(task_cpu(p), new_mask))
5302 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5303 /* Need help from migration thread: drop lock and wait. */
5304 struct task_struct *mt = rq->migration_thread;
5306 get_task_struct(mt);
5307 task_rq_unlock(rq, &flags);
5308 wake_up_process(rq->migration_thread);
5309 put_task_struct(mt);
5310 wait_for_completion(&req.done);
5311 tlb_migrate_finish(p->mm);
5315 task_rq_unlock(rq, &flags);
5319 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5322 * Move (not current) task off this cpu, onto dest cpu. We're doing
5323 * this because either it can't run here any more (set_cpus_allowed()
5324 * away from this CPU, or CPU going down), or because we're
5325 * attempting to rebalance this task on exec (sched_exec).
5327 * So we race with normal scheduler movements, but that's OK, as long
5328 * as the task is no longer on this CPU.
5330 * Returns non-zero if task was successfully migrated.
5332 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5334 struct rq *rq_dest, *rq_src;
5337 if (unlikely(!cpu_active(dest_cpu)))
5340 rq_src = cpu_rq(src_cpu);
5341 rq_dest = cpu_rq(dest_cpu);
5343 double_rq_lock(rq_src, rq_dest);
5344 /* Already moved. */
5345 if (task_cpu(p) != src_cpu)
5347 /* Affinity changed (again). */
5348 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5352 * If we're not on a rq, the next wake-up will ensure we're
5356 deactivate_task(rq_src, p, 0);
5357 set_task_cpu(p, dest_cpu);
5358 activate_task(rq_dest, p, 0);
5359 check_preempt_curr(rq_dest, p, 0);
5364 double_rq_unlock(rq_src, rq_dest);
5368 #define RCU_MIGRATION_IDLE 0
5369 #define RCU_MIGRATION_NEED_QS 1
5370 #define RCU_MIGRATION_GOT_QS 2
5371 #define RCU_MIGRATION_MUST_SYNC 3
5374 * migration_thread - this is a highprio system thread that performs
5375 * thread migration by bumping thread off CPU then 'pushing' onto
5378 static int migration_thread(void *data)
5381 int cpu = (long)data;
5385 BUG_ON(rq->migration_thread != current);
5387 set_current_state(TASK_INTERRUPTIBLE);
5388 while (!kthread_should_stop()) {
5389 struct migration_req *req;
5390 struct list_head *head;
5392 raw_spin_lock_irq(&rq->lock);
5394 if (cpu_is_offline(cpu)) {
5395 raw_spin_unlock_irq(&rq->lock);
5399 if (rq->active_balance) {
5400 active_load_balance(rq, cpu);
5401 rq->active_balance = 0;
5404 head = &rq->migration_queue;
5406 if (list_empty(head)) {
5407 raw_spin_unlock_irq(&rq->lock);
5409 set_current_state(TASK_INTERRUPTIBLE);
5412 req = list_entry(head->next, struct migration_req, list);
5413 list_del_init(head->next);
5415 if (req->task != NULL) {
5416 raw_spin_unlock(&rq->lock);
5417 __migrate_task(req->task, cpu, req->dest_cpu);
5418 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5419 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5420 raw_spin_unlock(&rq->lock);
5422 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5423 raw_spin_unlock(&rq->lock);
5424 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5428 complete(&req->done);
5430 __set_current_state(TASK_RUNNING);
5435 #ifdef CONFIG_HOTPLUG_CPU
5437 * Figure out where task on dead CPU should go, use force if necessary.
5439 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5441 struct rq *rq = cpu_rq(dead_cpu);
5442 int needs_cpu, uninitialized_var(dest_cpu);
5443 unsigned long flags;
5445 local_irq_save(flags);
5447 raw_spin_lock(&rq->lock);
5448 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5450 dest_cpu = select_fallback_rq(dead_cpu, p);
5451 raw_spin_unlock(&rq->lock);
5453 * It can only fail if we race with set_cpus_allowed(),
5454 * in the racer should migrate the task anyway.
5457 __migrate_task(p, dead_cpu, dest_cpu);
5458 local_irq_restore(flags);
5462 * While a dead CPU has no uninterruptible tasks queued at this point,
5463 * it might still have a nonzero ->nr_uninterruptible counter, because
5464 * for performance reasons the counter is not stricly tracking tasks to
5465 * their home CPUs. So we just add the counter to another CPU's counter,
5466 * to keep the global sum constant after CPU-down:
5468 static void migrate_nr_uninterruptible(struct rq *rq_src)
5470 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5471 unsigned long flags;
5473 local_irq_save(flags);
5474 double_rq_lock(rq_src, rq_dest);
5475 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5476 rq_src->nr_uninterruptible = 0;
5477 double_rq_unlock(rq_src, rq_dest);
5478 local_irq_restore(flags);
5481 /* Run through task list and migrate tasks from the dead cpu. */
5482 static void migrate_live_tasks(int src_cpu)
5484 struct task_struct *p, *t;
5486 read_lock(&tasklist_lock);
5488 do_each_thread(t, p) {
5492 if (task_cpu(p) == src_cpu)
5493 move_task_off_dead_cpu(src_cpu, p);
5494 } while_each_thread(t, p);
5496 read_unlock(&tasklist_lock);
5500 * Schedules idle task to be the next runnable task on current CPU.
5501 * It does so by boosting its priority to highest possible.
5502 * Used by CPU offline code.
5504 void sched_idle_next(void)
5506 int this_cpu = smp_processor_id();
5507 struct rq *rq = cpu_rq(this_cpu);
5508 struct task_struct *p = rq->idle;
5509 unsigned long flags;
5511 /* cpu has to be offline */
5512 BUG_ON(cpu_online(this_cpu));
5515 * Strictly not necessary since rest of the CPUs are stopped by now
5516 * and interrupts disabled on the current cpu.
5518 raw_spin_lock_irqsave(&rq->lock, flags);
5520 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5522 activate_task(rq, p, 0);
5524 raw_spin_unlock_irqrestore(&rq->lock, flags);
5528 * Ensures that the idle task is using init_mm right before its cpu goes
5531 void idle_task_exit(void)
5533 struct mm_struct *mm = current->active_mm;
5535 BUG_ON(cpu_online(smp_processor_id()));
5538 switch_mm(mm, &init_mm, current);
5542 /* called under rq->lock with disabled interrupts */
5543 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5545 struct rq *rq = cpu_rq(dead_cpu);
5547 /* Must be exiting, otherwise would be on tasklist. */
5548 BUG_ON(!p->exit_state);
5550 /* Cannot have done final schedule yet: would have vanished. */
5551 BUG_ON(p->state == TASK_DEAD);
5556 * Drop lock around migration; if someone else moves it,
5557 * that's OK. No task can be added to this CPU, so iteration is
5560 raw_spin_unlock_irq(&rq->lock);
5561 move_task_off_dead_cpu(dead_cpu, p);
5562 raw_spin_lock_irq(&rq->lock);
5567 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5568 static void migrate_dead_tasks(unsigned int dead_cpu)
5570 struct rq *rq = cpu_rq(dead_cpu);
5571 struct task_struct *next;
5574 if (!rq->nr_running)
5576 next = pick_next_task(rq);
5579 next->sched_class->put_prev_task(rq, next);
5580 migrate_dead(dead_cpu, next);
5586 * remove the tasks which were accounted by rq from calc_load_tasks.
5588 static void calc_global_load_remove(struct rq *rq)
5590 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5591 rq->calc_load_active = 0;
5593 #endif /* CONFIG_HOTPLUG_CPU */
5595 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5597 static struct ctl_table sd_ctl_dir[] = {
5599 .procname = "sched_domain",
5605 static struct ctl_table sd_ctl_root[] = {
5607 .procname = "kernel",
5609 .child = sd_ctl_dir,
5614 static struct ctl_table *sd_alloc_ctl_entry(int n)
5616 struct ctl_table *entry =
5617 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5622 static void sd_free_ctl_entry(struct ctl_table **tablep)
5624 struct ctl_table *entry;
5627 * In the intermediate directories, both the child directory and
5628 * procname are dynamically allocated and could fail but the mode
5629 * will always be set. In the lowest directory the names are
5630 * static strings and all have proc handlers.
5632 for (entry = *tablep; entry->mode; entry++) {
5634 sd_free_ctl_entry(&entry->child);
5635 if (entry->proc_handler == NULL)
5636 kfree(entry->procname);
5644 set_table_entry(struct ctl_table *entry,
5645 const char *procname, void *data, int maxlen,
5646 mode_t mode, proc_handler *proc_handler)
5648 entry->procname = procname;
5650 entry->maxlen = maxlen;
5652 entry->proc_handler = proc_handler;
5655 static struct ctl_table *
5656 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5658 struct ctl_table *table = sd_alloc_ctl_entry(13);
5663 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5664 sizeof(long), 0644, proc_doulongvec_minmax);
5665 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5666 sizeof(long), 0644, proc_doulongvec_minmax);
5667 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5668 sizeof(int), 0644, proc_dointvec_minmax);
5669 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5670 sizeof(int), 0644, proc_dointvec_minmax);
5671 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5672 sizeof(int), 0644, proc_dointvec_minmax);
5673 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5674 sizeof(int), 0644, proc_dointvec_minmax);
5675 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5676 sizeof(int), 0644, proc_dointvec_minmax);
5677 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5678 sizeof(int), 0644, proc_dointvec_minmax);
5679 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5680 sizeof(int), 0644, proc_dointvec_minmax);
5681 set_table_entry(&table[9], "cache_nice_tries",
5682 &sd->cache_nice_tries,
5683 sizeof(int), 0644, proc_dointvec_minmax);
5684 set_table_entry(&table[10], "flags", &sd->flags,
5685 sizeof(int), 0644, proc_dointvec_minmax);
5686 set_table_entry(&table[11], "name", sd->name,
5687 CORENAME_MAX_SIZE, 0444, proc_dostring);
5688 /* &table[12] is terminator */
5693 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5695 struct ctl_table *entry, *table;
5696 struct sched_domain *sd;
5697 int domain_num = 0, i;
5700 for_each_domain(cpu, sd)
5702 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5707 for_each_domain(cpu, sd) {
5708 snprintf(buf, 32, "domain%d", i);
5709 entry->procname = kstrdup(buf, GFP_KERNEL);
5711 entry->child = sd_alloc_ctl_domain_table(sd);
5718 static struct ctl_table_header *sd_sysctl_header;
5719 static void register_sched_domain_sysctl(void)
5721 int i, cpu_num = num_possible_cpus();
5722 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5725 WARN_ON(sd_ctl_dir[0].child);
5726 sd_ctl_dir[0].child = entry;
5731 for_each_possible_cpu(i) {
5732 snprintf(buf, 32, "cpu%d", i);
5733 entry->procname = kstrdup(buf, GFP_KERNEL);
5735 entry->child = sd_alloc_ctl_cpu_table(i);
5739 WARN_ON(sd_sysctl_header);
5740 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5743 /* may be called multiple times per register */
5744 static void unregister_sched_domain_sysctl(void)
5746 if (sd_sysctl_header)
5747 unregister_sysctl_table(sd_sysctl_header);
5748 sd_sysctl_header = NULL;
5749 if (sd_ctl_dir[0].child)
5750 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5753 static void register_sched_domain_sysctl(void)
5756 static void unregister_sched_domain_sysctl(void)
5761 static void set_rq_online(struct rq *rq)
5764 const struct sched_class *class;
5766 cpumask_set_cpu(rq->cpu, rq->rd->online);
5769 for_each_class(class) {
5770 if (class->rq_online)
5771 class->rq_online(rq);
5776 static void set_rq_offline(struct rq *rq)
5779 const struct sched_class *class;
5781 for_each_class(class) {
5782 if (class->rq_offline)
5783 class->rq_offline(rq);
5786 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5792 * migration_call - callback that gets triggered when a CPU is added.
5793 * Here we can start up the necessary migration thread for the new CPU.
5795 static int __cpuinit
5796 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5798 struct task_struct *p;
5799 int cpu = (long)hcpu;
5800 unsigned long flags;
5805 case CPU_UP_PREPARE:
5806 case CPU_UP_PREPARE_FROZEN:
5807 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5810 kthread_bind(p, cpu);
5811 /* Must be high prio: stop_machine expects to yield to it. */
5812 rq = task_rq_lock(p, &flags);
5813 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5814 task_rq_unlock(rq, &flags);
5816 cpu_rq(cpu)->migration_thread = p;
5817 rq->calc_load_update = calc_load_update;
5821 case CPU_ONLINE_FROZEN:
5822 /* Strictly unnecessary, as first user will wake it. */
5823 wake_up_process(cpu_rq(cpu)->migration_thread);
5825 /* Update our root-domain */
5827 raw_spin_lock_irqsave(&rq->lock, flags);
5829 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5833 raw_spin_unlock_irqrestore(&rq->lock, flags);
5836 #ifdef CONFIG_HOTPLUG_CPU
5837 case CPU_UP_CANCELED:
5838 case CPU_UP_CANCELED_FROZEN:
5839 if (!cpu_rq(cpu)->migration_thread)
5841 /* Unbind it from offline cpu so it can run. Fall thru. */
5842 kthread_bind(cpu_rq(cpu)->migration_thread,
5843 cpumask_any(cpu_online_mask));
5844 kthread_stop(cpu_rq(cpu)->migration_thread);
5845 put_task_struct(cpu_rq(cpu)->migration_thread);
5846 cpu_rq(cpu)->migration_thread = NULL;
5850 case CPU_DEAD_FROZEN:
5851 migrate_live_tasks(cpu);
5853 kthread_stop(rq->migration_thread);
5854 put_task_struct(rq->migration_thread);
5855 rq->migration_thread = NULL;
5856 /* Idle task back to normal (off runqueue, low prio) */
5857 raw_spin_lock_irq(&rq->lock);
5858 deactivate_task(rq, rq->idle, 0);
5859 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5860 rq->idle->sched_class = &idle_sched_class;
5861 migrate_dead_tasks(cpu);
5862 raw_spin_unlock_irq(&rq->lock);
5863 migrate_nr_uninterruptible(rq);
5864 BUG_ON(rq->nr_running != 0);
5865 calc_global_load_remove(rq);
5867 * No need to migrate the tasks: it was best-effort if
5868 * they didn't take sched_hotcpu_mutex. Just wake up
5871 raw_spin_lock_irq(&rq->lock);
5872 while (!list_empty(&rq->migration_queue)) {
5873 struct migration_req *req;
5875 req = list_entry(rq->migration_queue.next,
5876 struct migration_req, list);
5877 list_del_init(&req->list);
5878 raw_spin_unlock_irq(&rq->lock);
5879 complete(&req->done);
5880 raw_spin_lock_irq(&rq->lock);
5882 raw_spin_unlock_irq(&rq->lock);
5886 case CPU_DYING_FROZEN:
5887 /* Update our root-domain */
5889 raw_spin_lock_irqsave(&rq->lock, flags);
5891 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5894 raw_spin_unlock_irqrestore(&rq->lock, flags);
5902 * Register at high priority so that task migration (migrate_all_tasks)
5903 * happens before everything else. This has to be lower priority than
5904 * the notifier in the perf_event subsystem, though.
5906 static struct notifier_block __cpuinitdata migration_notifier = {
5907 .notifier_call = migration_call,
5911 static int __init migration_init(void)
5913 void *cpu = (void *)(long)smp_processor_id();
5916 /* Start one for the boot CPU: */
5917 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5918 BUG_ON(err == NOTIFY_BAD);
5919 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5920 register_cpu_notifier(&migration_notifier);
5924 early_initcall(migration_init);
5929 #ifdef CONFIG_SCHED_DEBUG
5931 static __read_mostly int sched_domain_debug_enabled;
5933 static int __init sched_domain_debug_setup(char *str)
5935 sched_domain_debug_enabled = 1;
5939 early_param("sched_debug", sched_domain_debug_setup);
5941 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5942 struct cpumask *groupmask)
5944 struct sched_group *group = sd->groups;
5947 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5948 cpumask_clear(groupmask);
5950 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5952 if (!(sd->flags & SD_LOAD_BALANCE)) {
5953 printk("does not load-balance\n");
5955 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5960 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5962 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5963 printk(KERN_ERR "ERROR: domain->span does not contain "
5966 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5967 printk(KERN_ERR "ERROR: domain->groups does not contain"
5971 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5975 printk(KERN_ERR "ERROR: group is NULL\n");
5979 if (!group->cpu_power) {
5980 printk(KERN_CONT "\n");
5981 printk(KERN_ERR "ERROR: domain->cpu_power not "
5986 if (!cpumask_weight(sched_group_cpus(group))) {
5987 printk(KERN_CONT "\n");
5988 printk(KERN_ERR "ERROR: empty group\n");
5992 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5993 printk(KERN_CONT "\n");
5994 printk(KERN_ERR "ERROR: repeated CPUs\n");
5998 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6000 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6002 printk(KERN_CONT " %s", str);
6003 if (group->cpu_power != SCHED_LOAD_SCALE) {
6004 printk(KERN_CONT " (cpu_power = %d)",
6008 group = group->next;
6009 } while (group != sd->groups);
6010 printk(KERN_CONT "\n");
6012 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6013 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6016 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6017 printk(KERN_ERR "ERROR: parent span is not a superset "
6018 "of domain->span\n");
6022 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6024 cpumask_var_t groupmask;
6027 if (!sched_domain_debug_enabled)
6031 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6035 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6037 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6038 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6043 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6050 free_cpumask_var(groupmask);
6052 #else /* !CONFIG_SCHED_DEBUG */
6053 # define sched_domain_debug(sd, cpu) do { } while (0)
6054 #endif /* CONFIG_SCHED_DEBUG */
6056 static int sd_degenerate(struct sched_domain *sd)
6058 if (cpumask_weight(sched_domain_span(sd)) == 1)
6061 /* Following flags need at least 2 groups */
6062 if (sd->flags & (SD_LOAD_BALANCE |
6063 SD_BALANCE_NEWIDLE |
6067 SD_SHARE_PKG_RESOURCES)) {
6068 if (sd->groups != sd->groups->next)
6072 /* Following flags don't use groups */
6073 if (sd->flags & (SD_WAKE_AFFINE))
6080 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6082 unsigned long cflags = sd->flags, pflags = parent->flags;
6084 if (sd_degenerate(parent))
6087 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6090 /* Flags needing groups don't count if only 1 group in parent */
6091 if (parent->groups == parent->groups->next) {
6092 pflags &= ~(SD_LOAD_BALANCE |
6093 SD_BALANCE_NEWIDLE |
6097 SD_SHARE_PKG_RESOURCES);
6098 if (nr_node_ids == 1)
6099 pflags &= ~SD_SERIALIZE;
6101 if (~cflags & pflags)
6107 static void free_rootdomain(struct root_domain *rd)
6109 synchronize_sched();
6111 cpupri_cleanup(&rd->cpupri);
6113 free_cpumask_var(rd->rto_mask);
6114 free_cpumask_var(rd->online);
6115 free_cpumask_var(rd->span);
6119 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6121 struct root_domain *old_rd = NULL;
6122 unsigned long flags;
6124 raw_spin_lock_irqsave(&rq->lock, flags);
6129 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6132 cpumask_clear_cpu(rq->cpu, old_rd->span);
6135 * If we dont want to free the old_rt yet then
6136 * set old_rd to NULL to skip the freeing later
6139 if (!atomic_dec_and_test(&old_rd->refcount))
6143 atomic_inc(&rd->refcount);
6146 cpumask_set_cpu(rq->cpu, rd->span);
6147 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6150 raw_spin_unlock_irqrestore(&rq->lock, flags);
6153 free_rootdomain(old_rd);
6156 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6158 gfp_t gfp = GFP_KERNEL;
6160 memset(rd, 0, sizeof(*rd));
6165 if (!alloc_cpumask_var(&rd->span, gfp))
6167 if (!alloc_cpumask_var(&rd->online, gfp))
6169 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6172 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6177 free_cpumask_var(rd->rto_mask);
6179 free_cpumask_var(rd->online);
6181 free_cpumask_var(rd->span);
6186 static void init_defrootdomain(void)
6188 init_rootdomain(&def_root_domain, true);
6190 atomic_set(&def_root_domain.refcount, 1);
6193 static struct root_domain *alloc_rootdomain(void)
6195 struct root_domain *rd;
6197 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6201 if (init_rootdomain(rd, false) != 0) {
6210 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6211 * hold the hotplug lock.
6214 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6216 struct rq *rq = cpu_rq(cpu);
6217 struct sched_domain *tmp;
6219 /* Remove the sched domains which do not contribute to scheduling. */
6220 for (tmp = sd; tmp; ) {
6221 struct sched_domain *parent = tmp->parent;
6225 if (sd_parent_degenerate(tmp, parent)) {
6226 tmp->parent = parent->parent;
6228 parent->parent->child = tmp;
6233 if (sd && sd_degenerate(sd)) {
6239 sched_domain_debug(sd, cpu);
6241 rq_attach_root(rq, rd);
6242 rcu_assign_pointer(rq->sd, sd);
6245 /* cpus with isolated domains */
6246 static cpumask_var_t cpu_isolated_map;
6248 /* Setup the mask of cpus configured for isolated domains */
6249 static int __init isolated_cpu_setup(char *str)
6251 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6252 cpulist_parse(str, cpu_isolated_map);
6256 __setup("isolcpus=", isolated_cpu_setup);
6259 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6260 * to a function which identifies what group(along with sched group) a CPU
6261 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6262 * (due to the fact that we keep track of groups covered with a struct cpumask).
6264 * init_sched_build_groups will build a circular linked list of the groups
6265 * covered by the given span, and will set each group's ->cpumask correctly,
6266 * and ->cpu_power to 0.
6269 init_sched_build_groups(const struct cpumask *span,
6270 const struct cpumask *cpu_map,
6271 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6272 struct sched_group **sg,
6273 struct cpumask *tmpmask),
6274 struct cpumask *covered, struct cpumask *tmpmask)
6276 struct sched_group *first = NULL, *last = NULL;
6279 cpumask_clear(covered);
6281 for_each_cpu(i, span) {
6282 struct sched_group *sg;
6283 int group = group_fn(i, cpu_map, &sg, tmpmask);
6286 if (cpumask_test_cpu(i, covered))
6289 cpumask_clear(sched_group_cpus(sg));
6292 for_each_cpu(j, span) {
6293 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6296 cpumask_set_cpu(j, covered);
6297 cpumask_set_cpu(j, sched_group_cpus(sg));
6308 #define SD_NODES_PER_DOMAIN 16
6313 * find_next_best_node - find the next node to include in a sched_domain
6314 * @node: node whose sched_domain we're building
6315 * @used_nodes: nodes already in the sched_domain
6317 * Find the next node to include in a given scheduling domain. Simply
6318 * finds the closest node not already in the @used_nodes map.
6320 * Should use nodemask_t.
6322 static int find_next_best_node(int node, nodemask_t *used_nodes)
6324 int i, n, val, min_val, best_node = 0;
6328 for (i = 0; i < nr_node_ids; i++) {
6329 /* Start at @node */
6330 n = (node + i) % nr_node_ids;
6332 if (!nr_cpus_node(n))
6335 /* Skip already used nodes */
6336 if (node_isset(n, *used_nodes))
6339 /* Simple min distance search */
6340 val = node_distance(node, n);
6342 if (val < min_val) {
6348 node_set(best_node, *used_nodes);
6353 * sched_domain_node_span - get a cpumask for a node's sched_domain
6354 * @node: node whose cpumask we're constructing
6355 * @span: resulting cpumask
6357 * Given a node, construct a good cpumask for its sched_domain to span. It
6358 * should be one that prevents unnecessary balancing, but also spreads tasks
6361 static void sched_domain_node_span(int node, struct cpumask *span)
6363 nodemask_t used_nodes;
6366 cpumask_clear(span);
6367 nodes_clear(used_nodes);
6369 cpumask_or(span, span, cpumask_of_node(node));
6370 node_set(node, used_nodes);
6372 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6373 int next_node = find_next_best_node(node, &used_nodes);
6375 cpumask_or(span, span, cpumask_of_node(next_node));
6378 #endif /* CONFIG_NUMA */
6380 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6383 * The cpus mask in sched_group and sched_domain hangs off the end.
6385 * ( See the the comments in include/linux/sched.h:struct sched_group
6386 * and struct sched_domain. )
6388 struct static_sched_group {
6389 struct sched_group sg;
6390 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6393 struct static_sched_domain {
6394 struct sched_domain sd;
6395 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6401 cpumask_var_t domainspan;
6402 cpumask_var_t covered;
6403 cpumask_var_t notcovered;
6405 cpumask_var_t nodemask;
6406 cpumask_var_t this_sibling_map;
6407 cpumask_var_t this_core_map;
6408 cpumask_var_t send_covered;
6409 cpumask_var_t tmpmask;
6410 struct sched_group **sched_group_nodes;
6411 struct root_domain *rd;
6415 sa_sched_groups = 0,
6420 sa_this_sibling_map,
6422 sa_sched_group_nodes,
6432 * SMT sched-domains:
6434 #ifdef CONFIG_SCHED_SMT
6435 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6436 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6439 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6440 struct sched_group **sg, struct cpumask *unused)
6443 *sg = &per_cpu(sched_groups, cpu).sg;
6446 #endif /* CONFIG_SCHED_SMT */
6449 * multi-core sched-domains:
6451 #ifdef CONFIG_SCHED_MC
6452 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6453 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6454 #endif /* CONFIG_SCHED_MC */
6456 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6458 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6459 struct sched_group **sg, struct cpumask *mask)
6463 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6464 group = cpumask_first(mask);
6466 *sg = &per_cpu(sched_group_core, group).sg;
6469 #elif defined(CONFIG_SCHED_MC)
6471 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6472 struct sched_group **sg, struct cpumask *unused)
6475 *sg = &per_cpu(sched_group_core, cpu).sg;
6480 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6481 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6484 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6485 struct sched_group **sg, struct cpumask *mask)
6488 #ifdef CONFIG_SCHED_MC
6489 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6490 group = cpumask_first(mask);
6491 #elif defined(CONFIG_SCHED_SMT)
6492 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6493 group = cpumask_first(mask);
6498 *sg = &per_cpu(sched_group_phys, group).sg;
6504 * The init_sched_build_groups can't handle what we want to do with node
6505 * groups, so roll our own. Now each node has its own list of groups which
6506 * gets dynamically allocated.
6508 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6509 static struct sched_group ***sched_group_nodes_bycpu;
6511 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6512 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6514 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6515 struct sched_group **sg,
6516 struct cpumask *nodemask)
6520 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6521 group = cpumask_first(nodemask);
6524 *sg = &per_cpu(sched_group_allnodes, group).sg;
6528 static void init_numa_sched_groups_power(struct sched_group *group_head)
6530 struct sched_group *sg = group_head;
6536 for_each_cpu(j, sched_group_cpus(sg)) {
6537 struct sched_domain *sd;
6539 sd = &per_cpu(phys_domains, j).sd;
6540 if (j != group_first_cpu(sd->groups)) {
6542 * Only add "power" once for each
6548 sg->cpu_power += sd->groups->cpu_power;
6551 } while (sg != group_head);
6554 static int build_numa_sched_groups(struct s_data *d,
6555 const struct cpumask *cpu_map, int num)
6557 struct sched_domain *sd;
6558 struct sched_group *sg, *prev;
6561 cpumask_clear(d->covered);
6562 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6563 if (cpumask_empty(d->nodemask)) {
6564 d->sched_group_nodes[num] = NULL;
6568 sched_domain_node_span(num, d->domainspan);
6569 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6571 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6574 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6578 d->sched_group_nodes[num] = sg;
6580 for_each_cpu(j, d->nodemask) {
6581 sd = &per_cpu(node_domains, j).sd;
6586 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6588 cpumask_or(d->covered, d->covered, d->nodemask);
6591 for (j = 0; j < nr_node_ids; j++) {
6592 n = (num + j) % nr_node_ids;
6593 cpumask_complement(d->notcovered, d->covered);
6594 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6595 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6596 if (cpumask_empty(d->tmpmask))
6598 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6599 if (cpumask_empty(d->tmpmask))
6601 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6605 "Can not alloc domain group for node %d\n", j);
6609 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6610 sg->next = prev->next;
6611 cpumask_or(d->covered, d->covered, d->tmpmask);
6618 #endif /* CONFIG_NUMA */
6621 /* Free memory allocated for various sched_group structures */
6622 static void free_sched_groups(const struct cpumask *cpu_map,
6623 struct cpumask *nodemask)
6627 for_each_cpu(cpu, cpu_map) {
6628 struct sched_group **sched_group_nodes
6629 = sched_group_nodes_bycpu[cpu];
6631 if (!sched_group_nodes)
6634 for (i = 0; i < nr_node_ids; i++) {
6635 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6637 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6638 if (cpumask_empty(nodemask))
6648 if (oldsg != sched_group_nodes[i])
6651 kfree(sched_group_nodes);
6652 sched_group_nodes_bycpu[cpu] = NULL;
6655 #else /* !CONFIG_NUMA */
6656 static void free_sched_groups(const struct cpumask *cpu_map,
6657 struct cpumask *nodemask)
6660 #endif /* CONFIG_NUMA */
6663 * Initialize sched groups cpu_power.
6665 * cpu_power indicates the capacity of sched group, which is used while
6666 * distributing the load between different sched groups in a sched domain.
6667 * Typically cpu_power for all the groups in a sched domain will be same unless
6668 * there are asymmetries in the topology. If there are asymmetries, group
6669 * having more cpu_power will pickup more load compared to the group having
6672 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6674 struct sched_domain *child;
6675 struct sched_group *group;
6679 WARN_ON(!sd || !sd->groups);
6681 if (cpu != group_first_cpu(sd->groups))
6686 sd->groups->cpu_power = 0;
6689 power = SCHED_LOAD_SCALE;
6690 weight = cpumask_weight(sched_domain_span(sd));
6692 * SMT siblings share the power of a single core.
6693 * Usually multiple threads get a better yield out of
6694 * that one core than a single thread would have,
6695 * reflect that in sd->smt_gain.
6697 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6698 power *= sd->smt_gain;
6700 power >>= SCHED_LOAD_SHIFT;
6702 sd->groups->cpu_power += power;
6707 * Add cpu_power of each child group to this groups cpu_power.
6709 group = child->groups;
6711 sd->groups->cpu_power += group->cpu_power;
6712 group = group->next;
6713 } while (group != child->groups);
6717 * Initializers for schedule domains
6718 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6721 #ifdef CONFIG_SCHED_DEBUG
6722 # define SD_INIT_NAME(sd, type) sd->name = #type
6724 # define SD_INIT_NAME(sd, type) do { } while (0)
6727 #define SD_INIT(sd, type) sd_init_##type(sd)
6729 #define SD_INIT_FUNC(type) \
6730 static noinline void sd_init_##type(struct sched_domain *sd) \
6732 memset(sd, 0, sizeof(*sd)); \
6733 *sd = SD_##type##_INIT; \
6734 sd->level = SD_LV_##type; \
6735 SD_INIT_NAME(sd, type); \
6740 SD_INIT_FUNC(ALLNODES)
6743 #ifdef CONFIG_SCHED_SMT
6744 SD_INIT_FUNC(SIBLING)
6746 #ifdef CONFIG_SCHED_MC
6750 static int default_relax_domain_level = -1;
6752 static int __init setup_relax_domain_level(char *str)
6756 val = simple_strtoul(str, NULL, 0);
6757 if (val < SD_LV_MAX)
6758 default_relax_domain_level = val;
6762 __setup("relax_domain_level=", setup_relax_domain_level);
6764 static void set_domain_attribute(struct sched_domain *sd,
6765 struct sched_domain_attr *attr)
6769 if (!attr || attr->relax_domain_level < 0) {
6770 if (default_relax_domain_level < 0)
6773 request = default_relax_domain_level;
6775 request = attr->relax_domain_level;
6776 if (request < sd->level) {
6777 /* turn off idle balance on this domain */
6778 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6780 /* turn on idle balance on this domain */
6781 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6785 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6786 const struct cpumask *cpu_map)
6789 case sa_sched_groups:
6790 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6791 d->sched_group_nodes = NULL;
6793 free_rootdomain(d->rd); /* fall through */
6795 free_cpumask_var(d->tmpmask); /* fall through */
6796 case sa_send_covered:
6797 free_cpumask_var(d->send_covered); /* fall through */
6798 case sa_this_core_map:
6799 free_cpumask_var(d->this_core_map); /* fall through */
6800 case sa_this_sibling_map:
6801 free_cpumask_var(d->this_sibling_map); /* fall through */
6803 free_cpumask_var(d->nodemask); /* fall through */
6804 case sa_sched_group_nodes:
6806 kfree(d->sched_group_nodes); /* fall through */
6808 free_cpumask_var(d->notcovered); /* fall through */
6810 free_cpumask_var(d->covered); /* fall through */
6812 free_cpumask_var(d->domainspan); /* fall through */
6819 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6820 const struct cpumask *cpu_map)
6823 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6825 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6826 return sa_domainspan;
6827 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6829 /* Allocate the per-node list of sched groups */
6830 d->sched_group_nodes = kcalloc(nr_node_ids,
6831 sizeof(struct sched_group *), GFP_KERNEL);
6832 if (!d->sched_group_nodes) {
6833 printk(KERN_WARNING "Can not alloc sched group node list\n");
6834 return sa_notcovered;
6836 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6838 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6839 return sa_sched_group_nodes;
6840 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6842 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6843 return sa_this_sibling_map;
6844 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6845 return sa_this_core_map;
6846 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6847 return sa_send_covered;
6848 d->rd = alloc_rootdomain();
6850 printk(KERN_WARNING "Cannot alloc root domain\n");
6853 return sa_rootdomain;
6856 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6857 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6859 struct sched_domain *sd = NULL;
6861 struct sched_domain *parent;
6864 if (cpumask_weight(cpu_map) >
6865 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6866 sd = &per_cpu(allnodes_domains, i).sd;
6867 SD_INIT(sd, ALLNODES);
6868 set_domain_attribute(sd, attr);
6869 cpumask_copy(sched_domain_span(sd), cpu_map);
6870 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6875 sd = &per_cpu(node_domains, i).sd;
6877 set_domain_attribute(sd, attr);
6878 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6879 sd->parent = parent;
6882 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6887 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6888 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6889 struct sched_domain *parent, int i)
6891 struct sched_domain *sd;
6892 sd = &per_cpu(phys_domains, i).sd;
6894 set_domain_attribute(sd, attr);
6895 cpumask_copy(sched_domain_span(sd), d->nodemask);
6896 sd->parent = parent;
6899 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6903 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6904 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6905 struct sched_domain *parent, int i)
6907 struct sched_domain *sd = parent;
6908 #ifdef CONFIG_SCHED_MC
6909 sd = &per_cpu(core_domains, i).sd;
6911 set_domain_attribute(sd, attr);
6912 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6913 sd->parent = parent;
6915 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6920 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6921 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6922 struct sched_domain *parent, int i)
6924 struct sched_domain *sd = parent;
6925 #ifdef CONFIG_SCHED_SMT
6926 sd = &per_cpu(cpu_domains, i).sd;
6927 SD_INIT(sd, SIBLING);
6928 set_domain_attribute(sd, attr);
6929 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6930 sd->parent = parent;
6932 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6937 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6938 const struct cpumask *cpu_map, int cpu)
6941 #ifdef CONFIG_SCHED_SMT
6942 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6943 cpumask_and(d->this_sibling_map, cpu_map,
6944 topology_thread_cpumask(cpu));
6945 if (cpu == cpumask_first(d->this_sibling_map))
6946 init_sched_build_groups(d->this_sibling_map, cpu_map,
6948 d->send_covered, d->tmpmask);
6951 #ifdef CONFIG_SCHED_MC
6952 case SD_LV_MC: /* set up multi-core groups */
6953 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6954 if (cpu == cpumask_first(d->this_core_map))
6955 init_sched_build_groups(d->this_core_map, cpu_map,
6957 d->send_covered, d->tmpmask);
6960 case SD_LV_CPU: /* set up physical groups */
6961 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
6962 if (!cpumask_empty(d->nodemask))
6963 init_sched_build_groups(d->nodemask, cpu_map,
6965 d->send_covered, d->tmpmask);
6968 case SD_LV_ALLNODES:
6969 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
6970 d->send_covered, d->tmpmask);
6979 * Build sched domains for a given set of cpus and attach the sched domains
6980 * to the individual cpus
6982 static int __build_sched_domains(const struct cpumask *cpu_map,
6983 struct sched_domain_attr *attr)
6985 enum s_alloc alloc_state = sa_none;
6987 struct sched_domain *sd;
6993 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6994 if (alloc_state != sa_rootdomain)
6996 alloc_state = sa_sched_groups;
6999 * Set up domains for cpus specified by the cpu_map.
7001 for_each_cpu(i, cpu_map) {
7002 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7005 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7006 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7007 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7008 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7011 for_each_cpu(i, cpu_map) {
7012 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7013 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7016 /* Set up physical groups */
7017 for (i = 0; i < nr_node_ids; i++)
7018 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7021 /* Set up node groups */
7023 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7025 for (i = 0; i < nr_node_ids; i++)
7026 if (build_numa_sched_groups(&d, cpu_map, i))
7030 /* Calculate CPU power for physical packages and nodes */
7031 #ifdef CONFIG_SCHED_SMT
7032 for_each_cpu(i, cpu_map) {
7033 sd = &per_cpu(cpu_domains, i).sd;
7034 init_sched_groups_power(i, sd);
7037 #ifdef CONFIG_SCHED_MC
7038 for_each_cpu(i, cpu_map) {
7039 sd = &per_cpu(core_domains, i).sd;
7040 init_sched_groups_power(i, sd);
7044 for_each_cpu(i, cpu_map) {
7045 sd = &per_cpu(phys_domains, i).sd;
7046 init_sched_groups_power(i, sd);
7050 for (i = 0; i < nr_node_ids; i++)
7051 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7053 if (d.sd_allnodes) {
7054 struct sched_group *sg;
7056 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7058 init_numa_sched_groups_power(sg);
7062 /* Attach the domains */
7063 for_each_cpu(i, cpu_map) {
7064 #ifdef CONFIG_SCHED_SMT
7065 sd = &per_cpu(cpu_domains, i).sd;
7066 #elif defined(CONFIG_SCHED_MC)
7067 sd = &per_cpu(core_domains, i).sd;
7069 sd = &per_cpu(phys_domains, i).sd;
7071 cpu_attach_domain(sd, d.rd, i);
7074 d.sched_group_nodes = NULL; /* don't free this we still need it */
7075 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7079 __free_domain_allocs(&d, alloc_state, cpu_map);
7083 static int build_sched_domains(const struct cpumask *cpu_map)
7085 return __build_sched_domains(cpu_map, NULL);
7088 static cpumask_var_t *doms_cur; /* current sched domains */
7089 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7090 static struct sched_domain_attr *dattr_cur;
7091 /* attribues of custom domains in 'doms_cur' */
7094 * Special case: If a kmalloc of a doms_cur partition (array of
7095 * cpumask) fails, then fallback to a single sched domain,
7096 * as determined by the single cpumask fallback_doms.
7098 static cpumask_var_t fallback_doms;
7101 * arch_update_cpu_topology lets virtualized architectures update the
7102 * cpu core maps. It is supposed to return 1 if the topology changed
7103 * or 0 if it stayed the same.
7105 int __attribute__((weak)) arch_update_cpu_topology(void)
7110 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7113 cpumask_var_t *doms;
7115 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7118 for (i = 0; i < ndoms; i++) {
7119 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7120 free_sched_domains(doms, i);
7127 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7130 for (i = 0; i < ndoms; i++)
7131 free_cpumask_var(doms[i]);
7136 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7137 * For now this just excludes isolated cpus, but could be used to
7138 * exclude other special cases in the future.
7140 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7144 arch_update_cpu_topology();
7146 doms_cur = alloc_sched_domains(ndoms_cur);
7148 doms_cur = &fallback_doms;
7149 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7151 err = build_sched_domains(doms_cur[0]);
7152 register_sched_domain_sysctl();
7157 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7158 struct cpumask *tmpmask)
7160 free_sched_groups(cpu_map, tmpmask);
7164 * Detach sched domains from a group of cpus specified in cpu_map
7165 * These cpus will now be attached to the NULL domain
7167 static void detach_destroy_domains(const struct cpumask *cpu_map)
7169 /* Save because hotplug lock held. */
7170 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7173 for_each_cpu(i, cpu_map)
7174 cpu_attach_domain(NULL, &def_root_domain, i);
7175 synchronize_sched();
7176 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7179 /* handle null as "default" */
7180 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7181 struct sched_domain_attr *new, int idx_new)
7183 struct sched_domain_attr tmp;
7190 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7191 new ? (new + idx_new) : &tmp,
7192 sizeof(struct sched_domain_attr));
7196 * Partition sched domains as specified by the 'ndoms_new'
7197 * cpumasks in the array doms_new[] of cpumasks. This compares
7198 * doms_new[] to the current sched domain partitioning, doms_cur[].
7199 * It destroys each deleted domain and builds each new domain.
7201 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7202 * The masks don't intersect (don't overlap.) We should setup one
7203 * sched domain for each mask. CPUs not in any of the cpumasks will
7204 * not be load balanced. If the same cpumask appears both in the
7205 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7208 * The passed in 'doms_new' should be allocated using
7209 * alloc_sched_domains. This routine takes ownership of it and will
7210 * free_sched_domains it when done with it. If the caller failed the
7211 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7212 * and partition_sched_domains() will fallback to the single partition
7213 * 'fallback_doms', it also forces the domains to be rebuilt.
7215 * If doms_new == NULL it will be replaced with cpu_online_mask.
7216 * ndoms_new == 0 is a special case for destroying existing domains,
7217 * and it will not create the default domain.
7219 * Call with hotplug lock held
7221 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7222 struct sched_domain_attr *dattr_new)
7227 mutex_lock(&sched_domains_mutex);
7229 /* always unregister in case we don't destroy any domains */
7230 unregister_sched_domain_sysctl();
7232 /* Let architecture update cpu core mappings. */
7233 new_topology = arch_update_cpu_topology();
7235 n = doms_new ? ndoms_new : 0;
7237 /* Destroy deleted domains */
7238 for (i = 0; i < ndoms_cur; i++) {
7239 for (j = 0; j < n && !new_topology; j++) {
7240 if (cpumask_equal(doms_cur[i], doms_new[j])
7241 && dattrs_equal(dattr_cur, i, dattr_new, j))
7244 /* no match - a current sched domain not in new doms_new[] */
7245 detach_destroy_domains(doms_cur[i]);
7250 if (doms_new == NULL) {
7252 doms_new = &fallback_doms;
7253 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7254 WARN_ON_ONCE(dattr_new);
7257 /* Build new domains */
7258 for (i = 0; i < ndoms_new; i++) {
7259 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7260 if (cpumask_equal(doms_new[i], doms_cur[j])
7261 && dattrs_equal(dattr_new, i, dattr_cur, j))
7264 /* no match - add a new doms_new */
7265 __build_sched_domains(doms_new[i],
7266 dattr_new ? dattr_new + i : NULL);
7271 /* Remember the new sched domains */
7272 if (doms_cur != &fallback_doms)
7273 free_sched_domains(doms_cur, ndoms_cur);
7274 kfree(dattr_cur); /* kfree(NULL) is safe */
7275 doms_cur = doms_new;
7276 dattr_cur = dattr_new;
7277 ndoms_cur = ndoms_new;
7279 register_sched_domain_sysctl();
7281 mutex_unlock(&sched_domains_mutex);
7284 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7285 static void arch_reinit_sched_domains(void)
7289 /* Destroy domains first to force the rebuild */
7290 partition_sched_domains(0, NULL, NULL);
7292 rebuild_sched_domains();
7296 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7298 unsigned int level = 0;
7300 if (sscanf(buf, "%u", &level) != 1)
7304 * level is always be positive so don't check for
7305 * level < POWERSAVINGS_BALANCE_NONE which is 0
7306 * What happens on 0 or 1 byte write,
7307 * need to check for count as well?
7310 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7314 sched_smt_power_savings = level;
7316 sched_mc_power_savings = level;
7318 arch_reinit_sched_domains();
7323 #ifdef CONFIG_SCHED_MC
7324 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7325 struct sysdev_class_attribute *attr,
7328 return sprintf(page, "%u\n", sched_mc_power_savings);
7330 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7331 struct sysdev_class_attribute *attr,
7332 const char *buf, size_t count)
7334 return sched_power_savings_store(buf, count, 0);
7336 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7337 sched_mc_power_savings_show,
7338 sched_mc_power_savings_store);
7341 #ifdef CONFIG_SCHED_SMT
7342 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7343 struct sysdev_class_attribute *attr,
7346 return sprintf(page, "%u\n", sched_smt_power_savings);
7348 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7349 struct sysdev_class_attribute *attr,
7350 const char *buf, size_t count)
7352 return sched_power_savings_store(buf, count, 1);
7354 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7355 sched_smt_power_savings_show,
7356 sched_smt_power_savings_store);
7359 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7363 #ifdef CONFIG_SCHED_SMT
7365 err = sysfs_create_file(&cls->kset.kobj,
7366 &attr_sched_smt_power_savings.attr);
7368 #ifdef CONFIG_SCHED_MC
7369 if (!err && mc_capable())
7370 err = sysfs_create_file(&cls->kset.kobj,
7371 &attr_sched_mc_power_savings.attr);
7375 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7377 #ifndef CONFIG_CPUSETS
7379 * Add online and remove offline CPUs from the scheduler domains.
7380 * When cpusets are enabled they take over this function.
7382 static int update_sched_domains(struct notifier_block *nfb,
7383 unsigned long action, void *hcpu)
7387 case CPU_ONLINE_FROZEN:
7388 case CPU_DOWN_PREPARE:
7389 case CPU_DOWN_PREPARE_FROZEN:
7390 case CPU_DOWN_FAILED:
7391 case CPU_DOWN_FAILED_FROZEN:
7392 partition_sched_domains(1, NULL, NULL);
7401 static int update_runtime(struct notifier_block *nfb,
7402 unsigned long action, void *hcpu)
7404 int cpu = (int)(long)hcpu;
7407 case CPU_DOWN_PREPARE:
7408 case CPU_DOWN_PREPARE_FROZEN:
7409 disable_runtime(cpu_rq(cpu));
7412 case CPU_DOWN_FAILED:
7413 case CPU_DOWN_FAILED_FROZEN:
7415 case CPU_ONLINE_FROZEN:
7416 enable_runtime(cpu_rq(cpu));
7424 void __init sched_init_smp(void)
7426 cpumask_var_t non_isolated_cpus;
7428 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7429 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7431 #if defined(CONFIG_NUMA)
7432 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7434 BUG_ON(sched_group_nodes_bycpu == NULL);
7437 mutex_lock(&sched_domains_mutex);
7438 arch_init_sched_domains(cpu_active_mask);
7439 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7440 if (cpumask_empty(non_isolated_cpus))
7441 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7442 mutex_unlock(&sched_domains_mutex);
7445 #ifndef CONFIG_CPUSETS
7446 /* XXX: Theoretical race here - CPU may be hotplugged now */
7447 hotcpu_notifier(update_sched_domains, 0);
7450 /* RT runtime code needs to handle some hotplug events */
7451 hotcpu_notifier(update_runtime, 0);
7455 /* Move init over to a non-isolated CPU */
7456 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7458 sched_init_granularity();
7459 free_cpumask_var(non_isolated_cpus);
7461 init_sched_rt_class();
7464 void __init sched_init_smp(void)
7466 sched_init_granularity();
7468 #endif /* CONFIG_SMP */
7470 const_debug unsigned int sysctl_timer_migration = 1;
7472 int in_sched_functions(unsigned long addr)
7474 return in_lock_functions(addr) ||
7475 (addr >= (unsigned long)__sched_text_start
7476 && addr < (unsigned long)__sched_text_end);
7479 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7481 cfs_rq->tasks_timeline = RB_ROOT;
7482 INIT_LIST_HEAD(&cfs_rq->tasks);
7483 #ifdef CONFIG_FAIR_GROUP_SCHED
7486 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7489 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7491 struct rt_prio_array *array;
7494 array = &rt_rq->active;
7495 for (i = 0; i < MAX_RT_PRIO; i++) {
7496 INIT_LIST_HEAD(array->queue + i);
7497 __clear_bit(i, array->bitmap);
7499 /* delimiter for bitsearch: */
7500 __set_bit(MAX_RT_PRIO, array->bitmap);
7502 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7503 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7505 rt_rq->highest_prio.next = MAX_RT_PRIO;
7509 rt_rq->rt_nr_migratory = 0;
7510 rt_rq->overloaded = 0;
7511 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7515 rt_rq->rt_throttled = 0;
7516 rt_rq->rt_runtime = 0;
7517 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7519 #ifdef CONFIG_RT_GROUP_SCHED
7520 rt_rq->rt_nr_boosted = 0;
7525 #ifdef CONFIG_FAIR_GROUP_SCHED
7526 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7527 struct sched_entity *se, int cpu, int add,
7528 struct sched_entity *parent)
7530 struct rq *rq = cpu_rq(cpu);
7531 tg->cfs_rq[cpu] = cfs_rq;
7532 init_cfs_rq(cfs_rq, rq);
7535 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7538 /* se could be NULL for init_task_group */
7543 se->cfs_rq = &rq->cfs;
7545 se->cfs_rq = parent->my_q;
7548 se->load.weight = tg->shares;
7549 se->load.inv_weight = 0;
7550 se->parent = parent;
7554 #ifdef CONFIG_RT_GROUP_SCHED
7555 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7556 struct sched_rt_entity *rt_se, int cpu, int add,
7557 struct sched_rt_entity *parent)
7559 struct rq *rq = cpu_rq(cpu);
7561 tg->rt_rq[cpu] = rt_rq;
7562 init_rt_rq(rt_rq, rq);
7564 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7566 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7568 tg->rt_se[cpu] = rt_se;
7573 rt_se->rt_rq = &rq->rt;
7575 rt_se->rt_rq = parent->my_q;
7577 rt_se->my_q = rt_rq;
7578 rt_se->parent = parent;
7579 INIT_LIST_HEAD(&rt_se->run_list);
7583 void __init sched_init(void)
7586 unsigned long alloc_size = 0, ptr;
7588 #ifdef CONFIG_FAIR_GROUP_SCHED
7589 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7591 #ifdef CONFIG_RT_GROUP_SCHED
7592 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7594 #ifdef CONFIG_CPUMASK_OFFSTACK
7595 alloc_size += num_possible_cpus() * cpumask_size();
7598 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7600 #ifdef CONFIG_FAIR_GROUP_SCHED
7601 init_task_group.se = (struct sched_entity **)ptr;
7602 ptr += nr_cpu_ids * sizeof(void **);
7604 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7605 ptr += nr_cpu_ids * sizeof(void **);
7607 #endif /* CONFIG_FAIR_GROUP_SCHED */
7608 #ifdef CONFIG_RT_GROUP_SCHED
7609 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7610 ptr += nr_cpu_ids * sizeof(void **);
7612 init_task_group.rt_rq = (struct rt_rq **)ptr;
7613 ptr += nr_cpu_ids * sizeof(void **);
7615 #endif /* CONFIG_RT_GROUP_SCHED */
7616 #ifdef CONFIG_CPUMASK_OFFSTACK
7617 for_each_possible_cpu(i) {
7618 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7619 ptr += cpumask_size();
7621 #endif /* CONFIG_CPUMASK_OFFSTACK */
7625 init_defrootdomain();
7628 init_rt_bandwidth(&def_rt_bandwidth,
7629 global_rt_period(), global_rt_runtime());
7631 #ifdef CONFIG_RT_GROUP_SCHED
7632 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7633 global_rt_period(), global_rt_runtime());
7634 #endif /* CONFIG_RT_GROUP_SCHED */
7636 #ifdef CONFIG_CGROUP_SCHED
7637 list_add(&init_task_group.list, &task_groups);
7638 INIT_LIST_HEAD(&init_task_group.children);
7640 #endif /* CONFIG_CGROUP_SCHED */
7642 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7643 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7644 __alignof__(unsigned long));
7646 for_each_possible_cpu(i) {
7650 raw_spin_lock_init(&rq->lock);
7652 rq->calc_load_active = 0;
7653 rq->calc_load_update = jiffies + LOAD_FREQ;
7654 init_cfs_rq(&rq->cfs, rq);
7655 init_rt_rq(&rq->rt, rq);
7656 #ifdef CONFIG_FAIR_GROUP_SCHED
7657 init_task_group.shares = init_task_group_load;
7658 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7659 #ifdef CONFIG_CGROUP_SCHED
7661 * How much cpu bandwidth does init_task_group get?
7663 * In case of task-groups formed thr' the cgroup filesystem, it
7664 * gets 100% of the cpu resources in the system. This overall
7665 * system cpu resource is divided among the tasks of
7666 * init_task_group and its child task-groups in a fair manner,
7667 * based on each entity's (task or task-group's) weight
7668 * (se->load.weight).
7670 * In other words, if init_task_group has 10 tasks of weight
7671 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7672 * then A0's share of the cpu resource is:
7674 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7676 * We achieve this by letting init_task_group's tasks sit
7677 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7679 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7681 #endif /* CONFIG_FAIR_GROUP_SCHED */
7683 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7684 #ifdef CONFIG_RT_GROUP_SCHED
7685 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7686 #ifdef CONFIG_CGROUP_SCHED
7687 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7691 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7692 rq->cpu_load[j] = 0;
7696 rq->post_schedule = 0;
7697 rq->active_balance = 0;
7698 rq->next_balance = jiffies;
7702 rq->migration_thread = NULL;
7704 rq->avg_idle = 2*sysctl_sched_migration_cost;
7705 INIT_LIST_HEAD(&rq->migration_queue);
7706 rq_attach_root(rq, &def_root_domain);
7709 atomic_set(&rq->nr_iowait, 0);
7712 set_load_weight(&init_task);
7714 #ifdef CONFIG_PREEMPT_NOTIFIERS
7715 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7719 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7722 #ifdef CONFIG_RT_MUTEXES
7723 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7727 * The boot idle thread does lazy MMU switching as well:
7729 atomic_inc(&init_mm.mm_count);
7730 enter_lazy_tlb(&init_mm, current);
7733 * Make us the idle thread. Technically, schedule() should not be
7734 * called from this thread, however somewhere below it might be,
7735 * but because we are the idle thread, we just pick up running again
7736 * when this runqueue becomes "idle".
7738 init_idle(current, smp_processor_id());
7740 calc_load_update = jiffies + LOAD_FREQ;
7743 * During early bootup we pretend to be a normal task:
7745 current->sched_class = &fair_sched_class;
7747 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7748 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7751 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7752 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7754 /* May be allocated at isolcpus cmdline parse time */
7755 if (cpu_isolated_map == NULL)
7756 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7761 scheduler_running = 1;
7764 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7765 static inline int preempt_count_equals(int preempt_offset)
7767 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7769 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7772 void __might_sleep(const char *file, int line, int preempt_offset)
7775 static unsigned long prev_jiffy; /* ratelimiting */
7777 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7778 system_state != SYSTEM_RUNNING || oops_in_progress)
7780 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7782 prev_jiffy = jiffies;
7785 "BUG: sleeping function called from invalid context at %s:%d\n",
7788 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7789 in_atomic(), irqs_disabled(),
7790 current->pid, current->comm);
7792 debug_show_held_locks(current);
7793 if (irqs_disabled())
7794 print_irqtrace_events(current);
7798 EXPORT_SYMBOL(__might_sleep);
7801 #ifdef CONFIG_MAGIC_SYSRQ
7802 static void normalize_task(struct rq *rq, struct task_struct *p)
7806 on_rq = p->se.on_rq;
7808 deactivate_task(rq, p, 0);
7809 __setscheduler(rq, p, SCHED_NORMAL, 0);
7811 activate_task(rq, p, 0);
7812 resched_task(rq->curr);
7816 void normalize_rt_tasks(void)
7818 struct task_struct *g, *p;
7819 unsigned long flags;
7822 read_lock_irqsave(&tasklist_lock, flags);
7823 do_each_thread(g, p) {
7825 * Only normalize user tasks:
7830 p->se.exec_start = 0;
7831 #ifdef CONFIG_SCHEDSTATS
7832 p->se.statistics.wait_start = 0;
7833 p->se.statistics.sleep_start = 0;
7834 p->se.statistics.block_start = 0;
7839 * Renice negative nice level userspace
7842 if (TASK_NICE(p) < 0 && p->mm)
7843 set_user_nice(p, 0);
7847 raw_spin_lock(&p->pi_lock);
7848 rq = __task_rq_lock(p);
7850 normalize_task(rq, p);
7852 __task_rq_unlock(rq);
7853 raw_spin_unlock(&p->pi_lock);
7854 } while_each_thread(g, p);
7856 read_unlock_irqrestore(&tasklist_lock, flags);
7859 #endif /* CONFIG_MAGIC_SYSRQ */
7863 * These functions are only useful for the IA64 MCA handling.
7865 * They can only be called when the whole system has been
7866 * stopped - every CPU needs to be quiescent, and no scheduling
7867 * activity can take place. Using them for anything else would
7868 * be a serious bug, and as a result, they aren't even visible
7869 * under any other configuration.
7873 * curr_task - return the current task for a given cpu.
7874 * @cpu: the processor in question.
7876 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7878 struct task_struct *curr_task(int cpu)
7880 return cpu_curr(cpu);
7884 * set_curr_task - set the current task for a given cpu.
7885 * @cpu: the processor in question.
7886 * @p: the task pointer to set.
7888 * Description: This function must only be used when non-maskable interrupts
7889 * are serviced on a separate stack. It allows the architecture to switch the
7890 * notion of the current task on a cpu in a non-blocking manner. This function
7891 * must be called with all CPU's synchronized, and interrupts disabled, the
7892 * and caller must save the original value of the current task (see
7893 * curr_task() above) and restore that value before reenabling interrupts and
7894 * re-starting the system.
7896 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7898 void set_curr_task(int cpu, struct task_struct *p)
7905 #ifdef CONFIG_FAIR_GROUP_SCHED
7906 static void free_fair_sched_group(struct task_group *tg)
7910 for_each_possible_cpu(i) {
7912 kfree(tg->cfs_rq[i]);
7922 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7924 struct cfs_rq *cfs_rq;
7925 struct sched_entity *se;
7929 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7932 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7936 tg->shares = NICE_0_LOAD;
7938 for_each_possible_cpu(i) {
7941 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7942 GFP_KERNEL, cpu_to_node(i));
7946 se = kzalloc_node(sizeof(struct sched_entity),
7947 GFP_KERNEL, cpu_to_node(i));
7951 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
7962 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7964 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7965 &cpu_rq(cpu)->leaf_cfs_rq_list);
7968 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7970 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7972 #else /* !CONFG_FAIR_GROUP_SCHED */
7973 static inline void free_fair_sched_group(struct task_group *tg)
7978 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7983 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7987 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7990 #endif /* CONFIG_FAIR_GROUP_SCHED */
7992 #ifdef CONFIG_RT_GROUP_SCHED
7993 static void free_rt_sched_group(struct task_group *tg)
7997 destroy_rt_bandwidth(&tg->rt_bandwidth);
7999 for_each_possible_cpu(i) {
8001 kfree(tg->rt_rq[i]);
8003 kfree(tg->rt_se[i]);
8011 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8013 struct rt_rq *rt_rq;
8014 struct sched_rt_entity *rt_se;
8018 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8021 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8025 init_rt_bandwidth(&tg->rt_bandwidth,
8026 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8028 for_each_possible_cpu(i) {
8031 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8032 GFP_KERNEL, cpu_to_node(i));
8036 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8037 GFP_KERNEL, cpu_to_node(i));
8041 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8052 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8054 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8055 &cpu_rq(cpu)->leaf_rt_rq_list);
8058 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8060 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8062 #else /* !CONFIG_RT_GROUP_SCHED */
8063 static inline void free_rt_sched_group(struct task_group *tg)
8068 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8073 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8077 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8080 #endif /* CONFIG_RT_GROUP_SCHED */
8082 #ifdef CONFIG_CGROUP_SCHED
8083 static void free_sched_group(struct task_group *tg)
8085 free_fair_sched_group(tg);
8086 free_rt_sched_group(tg);
8090 /* allocate runqueue etc for a new task group */
8091 struct task_group *sched_create_group(struct task_group *parent)
8093 struct task_group *tg;
8094 unsigned long flags;
8097 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8099 return ERR_PTR(-ENOMEM);
8101 if (!alloc_fair_sched_group(tg, parent))
8104 if (!alloc_rt_sched_group(tg, parent))
8107 spin_lock_irqsave(&task_group_lock, flags);
8108 for_each_possible_cpu(i) {
8109 register_fair_sched_group(tg, i);
8110 register_rt_sched_group(tg, i);
8112 list_add_rcu(&tg->list, &task_groups);
8114 WARN_ON(!parent); /* root should already exist */
8116 tg->parent = parent;
8117 INIT_LIST_HEAD(&tg->children);
8118 list_add_rcu(&tg->siblings, &parent->children);
8119 spin_unlock_irqrestore(&task_group_lock, flags);
8124 free_sched_group(tg);
8125 return ERR_PTR(-ENOMEM);
8128 /* rcu callback to free various structures associated with a task group */
8129 static void free_sched_group_rcu(struct rcu_head *rhp)
8131 /* now it should be safe to free those cfs_rqs */
8132 free_sched_group(container_of(rhp, struct task_group, rcu));
8135 /* Destroy runqueue etc associated with a task group */
8136 void sched_destroy_group(struct task_group *tg)
8138 unsigned long flags;
8141 spin_lock_irqsave(&task_group_lock, flags);
8142 for_each_possible_cpu(i) {
8143 unregister_fair_sched_group(tg, i);
8144 unregister_rt_sched_group(tg, i);
8146 list_del_rcu(&tg->list);
8147 list_del_rcu(&tg->siblings);
8148 spin_unlock_irqrestore(&task_group_lock, flags);
8150 /* wait for possible concurrent references to cfs_rqs complete */
8151 call_rcu(&tg->rcu, free_sched_group_rcu);
8154 /* change task's runqueue when it moves between groups.
8155 * The caller of this function should have put the task in its new group
8156 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8157 * reflect its new group.
8159 void sched_move_task(struct task_struct *tsk)
8162 unsigned long flags;
8165 rq = task_rq_lock(tsk, &flags);
8167 running = task_current(rq, tsk);
8168 on_rq = tsk->se.on_rq;
8171 dequeue_task(rq, tsk, 0);
8172 if (unlikely(running))
8173 tsk->sched_class->put_prev_task(rq, tsk);
8175 set_task_rq(tsk, task_cpu(tsk));
8177 #ifdef CONFIG_FAIR_GROUP_SCHED
8178 if (tsk->sched_class->moved_group)
8179 tsk->sched_class->moved_group(tsk, on_rq);
8182 if (unlikely(running))
8183 tsk->sched_class->set_curr_task(rq);
8185 enqueue_task(rq, tsk, 0);
8187 task_rq_unlock(rq, &flags);
8189 #endif /* CONFIG_CGROUP_SCHED */
8191 #ifdef CONFIG_FAIR_GROUP_SCHED
8192 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8194 struct cfs_rq *cfs_rq = se->cfs_rq;
8199 dequeue_entity(cfs_rq, se, 0);
8201 se->load.weight = shares;
8202 se->load.inv_weight = 0;
8205 enqueue_entity(cfs_rq, se, 0);
8208 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8210 struct cfs_rq *cfs_rq = se->cfs_rq;
8211 struct rq *rq = cfs_rq->rq;
8212 unsigned long flags;
8214 raw_spin_lock_irqsave(&rq->lock, flags);
8215 __set_se_shares(se, shares);
8216 raw_spin_unlock_irqrestore(&rq->lock, flags);
8219 static DEFINE_MUTEX(shares_mutex);
8221 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8224 unsigned long flags;
8227 * We can't change the weight of the root cgroup.
8232 if (shares < MIN_SHARES)
8233 shares = MIN_SHARES;
8234 else if (shares > MAX_SHARES)
8235 shares = MAX_SHARES;
8237 mutex_lock(&shares_mutex);
8238 if (tg->shares == shares)
8241 spin_lock_irqsave(&task_group_lock, flags);
8242 for_each_possible_cpu(i)
8243 unregister_fair_sched_group(tg, i);
8244 list_del_rcu(&tg->siblings);
8245 spin_unlock_irqrestore(&task_group_lock, flags);
8247 /* wait for any ongoing reference to this group to finish */
8248 synchronize_sched();
8251 * Now we are free to modify the group's share on each cpu
8252 * w/o tripping rebalance_share or load_balance_fair.
8254 tg->shares = shares;
8255 for_each_possible_cpu(i) {
8259 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8260 set_se_shares(tg->se[i], shares);
8264 * Enable load balance activity on this group, by inserting it back on
8265 * each cpu's rq->leaf_cfs_rq_list.
8267 spin_lock_irqsave(&task_group_lock, flags);
8268 for_each_possible_cpu(i)
8269 register_fair_sched_group(tg, i);
8270 list_add_rcu(&tg->siblings, &tg->parent->children);
8271 spin_unlock_irqrestore(&task_group_lock, flags);
8273 mutex_unlock(&shares_mutex);
8277 unsigned long sched_group_shares(struct task_group *tg)
8283 #ifdef CONFIG_RT_GROUP_SCHED
8285 * Ensure that the real time constraints are schedulable.
8287 static DEFINE_MUTEX(rt_constraints_mutex);
8289 static unsigned long to_ratio(u64 period, u64 runtime)
8291 if (runtime == RUNTIME_INF)
8294 return div64_u64(runtime << 20, period);
8297 /* Must be called with tasklist_lock held */
8298 static inline int tg_has_rt_tasks(struct task_group *tg)
8300 struct task_struct *g, *p;
8302 do_each_thread(g, p) {
8303 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8305 } while_each_thread(g, p);
8310 struct rt_schedulable_data {
8311 struct task_group *tg;
8316 static int tg_schedulable(struct task_group *tg, void *data)
8318 struct rt_schedulable_data *d = data;
8319 struct task_group *child;
8320 unsigned long total, sum = 0;
8321 u64 period, runtime;
8323 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8324 runtime = tg->rt_bandwidth.rt_runtime;
8327 period = d->rt_period;
8328 runtime = d->rt_runtime;
8332 * Cannot have more runtime than the period.
8334 if (runtime > period && runtime != RUNTIME_INF)
8338 * Ensure we don't starve existing RT tasks.
8340 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8343 total = to_ratio(period, runtime);
8346 * Nobody can have more than the global setting allows.
8348 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8352 * The sum of our children's runtime should not exceed our own.
8354 list_for_each_entry_rcu(child, &tg->children, siblings) {
8355 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8356 runtime = child->rt_bandwidth.rt_runtime;
8358 if (child == d->tg) {
8359 period = d->rt_period;
8360 runtime = d->rt_runtime;
8363 sum += to_ratio(period, runtime);
8372 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8374 struct rt_schedulable_data data = {
8376 .rt_period = period,
8377 .rt_runtime = runtime,
8380 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8383 static int tg_set_bandwidth(struct task_group *tg,
8384 u64 rt_period, u64 rt_runtime)
8388 mutex_lock(&rt_constraints_mutex);
8389 read_lock(&tasklist_lock);
8390 err = __rt_schedulable(tg, rt_period, rt_runtime);
8394 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8395 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8396 tg->rt_bandwidth.rt_runtime = rt_runtime;
8398 for_each_possible_cpu(i) {
8399 struct rt_rq *rt_rq = tg->rt_rq[i];
8401 raw_spin_lock(&rt_rq->rt_runtime_lock);
8402 rt_rq->rt_runtime = rt_runtime;
8403 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8405 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8407 read_unlock(&tasklist_lock);
8408 mutex_unlock(&rt_constraints_mutex);
8413 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8415 u64 rt_runtime, rt_period;
8417 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8418 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8419 if (rt_runtime_us < 0)
8420 rt_runtime = RUNTIME_INF;
8422 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8425 long sched_group_rt_runtime(struct task_group *tg)
8429 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8432 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8433 do_div(rt_runtime_us, NSEC_PER_USEC);
8434 return rt_runtime_us;
8437 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8439 u64 rt_runtime, rt_period;
8441 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8442 rt_runtime = tg->rt_bandwidth.rt_runtime;
8447 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8450 long sched_group_rt_period(struct task_group *tg)
8454 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8455 do_div(rt_period_us, NSEC_PER_USEC);
8456 return rt_period_us;
8459 static int sched_rt_global_constraints(void)
8461 u64 runtime, period;
8464 if (sysctl_sched_rt_period <= 0)
8467 runtime = global_rt_runtime();
8468 period = global_rt_period();
8471 * Sanity check on the sysctl variables.
8473 if (runtime > period && runtime != RUNTIME_INF)
8476 mutex_lock(&rt_constraints_mutex);
8477 read_lock(&tasklist_lock);
8478 ret = __rt_schedulable(NULL, 0, 0);
8479 read_unlock(&tasklist_lock);
8480 mutex_unlock(&rt_constraints_mutex);
8485 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8487 /* Don't accept realtime tasks when there is no way for them to run */
8488 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8494 #else /* !CONFIG_RT_GROUP_SCHED */
8495 static int sched_rt_global_constraints(void)
8497 unsigned long flags;
8500 if (sysctl_sched_rt_period <= 0)
8504 * There's always some RT tasks in the root group
8505 * -- migration, kstopmachine etc..
8507 if (sysctl_sched_rt_runtime == 0)
8510 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8511 for_each_possible_cpu(i) {
8512 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8514 raw_spin_lock(&rt_rq->rt_runtime_lock);
8515 rt_rq->rt_runtime = global_rt_runtime();
8516 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8518 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8522 #endif /* CONFIG_RT_GROUP_SCHED */
8524 int sched_rt_handler(struct ctl_table *table, int write,
8525 void __user *buffer, size_t *lenp,
8529 int old_period, old_runtime;
8530 static DEFINE_MUTEX(mutex);
8533 old_period = sysctl_sched_rt_period;
8534 old_runtime = sysctl_sched_rt_runtime;
8536 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8538 if (!ret && write) {
8539 ret = sched_rt_global_constraints();
8541 sysctl_sched_rt_period = old_period;
8542 sysctl_sched_rt_runtime = old_runtime;
8544 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8545 def_rt_bandwidth.rt_period =
8546 ns_to_ktime(global_rt_period());
8549 mutex_unlock(&mutex);
8554 #ifdef CONFIG_CGROUP_SCHED
8556 /* return corresponding task_group object of a cgroup */
8557 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8559 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8560 struct task_group, css);
8563 static struct cgroup_subsys_state *
8564 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8566 struct task_group *tg, *parent;
8568 if (!cgrp->parent) {
8569 /* This is early initialization for the top cgroup */
8570 return &init_task_group.css;
8573 parent = cgroup_tg(cgrp->parent);
8574 tg = sched_create_group(parent);
8576 return ERR_PTR(-ENOMEM);
8582 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8584 struct task_group *tg = cgroup_tg(cgrp);
8586 sched_destroy_group(tg);
8590 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8592 #ifdef CONFIG_RT_GROUP_SCHED
8593 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8596 /* We don't support RT-tasks being in separate groups */
8597 if (tsk->sched_class != &fair_sched_class)
8604 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8605 struct task_struct *tsk, bool threadgroup)
8607 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8611 struct task_struct *c;
8613 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8614 retval = cpu_cgroup_can_attach_task(cgrp, c);
8626 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8627 struct cgroup *old_cont, struct task_struct *tsk,
8630 sched_move_task(tsk);
8632 struct task_struct *c;
8634 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8641 #ifdef CONFIG_FAIR_GROUP_SCHED
8642 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8645 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8648 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8650 struct task_group *tg = cgroup_tg(cgrp);
8652 return (u64) tg->shares;
8654 #endif /* CONFIG_FAIR_GROUP_SCHED */
8656 #ifdef CONFIG_RT_GROUP_SCHED
8657 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8660 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8663 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8665 return sched_group_rt_runtime(cgroup_tg(cgrp));
8668 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8671 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8674 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8676 return sched_group_rt_period(cgroup_tg(cgrp));
8678 #endif /* CONFIG_RT_GROUP_SCHED */
8680 static struct cftype cpu_files[] = {
8681 #ifdef CONFIG_FAIR_GROUP_SCHED
8684 .read_u64 = cpu_shares_read_u64,
8685 .write_u64 = cpu_shares_write_u64,
8688 #ifdef CONFIG_RT_GROUP_SCHED
8690 .name = "rt_runtime_us",
8691 .read_s64 = cpu_rt_runtime_read,
8692 .write_s64 = cpu_rt_runtime_write,
8695 .name = "rt_period_us",
8696 .read_u64 = cpu_rt_period_read_uint,
8697 .write_u64 = cpu_rt_period_write_uint,
8702 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8704 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8707 struct cgroup_subsys cpu_cgroup_subsys = {
8709 .create = cpu_cgroup_create,
8710 .destroy = cpu_cgroup_destroy,
8711 .can_attach = cpu_cgroup_can_attach,
8712 .attach = cpu_cgroup_attach,
8713 .populate = cpu_cgroup_populate,
8714 .subsys_id = cpu_cgroup_subsys_id,
8718 #endif /* CONFIG_CGROUP_SCHED */
8720 #ifdef CONFIG_CGROUP_CPUACCT
8723 * CPU accounting code for task groups.
8725 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8726 * (balbir@in.ibm.com).
8729 /* track cpu usage of a group of tasks and its child groups */
8731 struct cgroup_subsys_state css;
8732 /* cpuusage holds pointer to a u64-type object on every cpu */
8733 u64 __percpu *cpuusage;
8734 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8735 struct cpuacct *parent;
8738 struct cgroup_subsys cpuacct_subsys;
8740 /* return cpu accounting group corresponding to this container */
8741 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8743 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8744 struct cpuacct, css);
8747 /* return cpu accounting group to which this task belongs */
8748 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8750 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8751 struct cpuacct, css);
8754 /* create a new cpu accounting group */
8755 static struct cgroup_subsys_state *cpuacct_create(
8756 struct cgroup_subsys *ss, struct cgroup *cgrp)
8758 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8764 ca->cpuusage = alloc_percpu(u64);
8768 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8769 if (percpu_counter_init(&ca->cpustat[i], 0))
8770 goto out_free_counters;
8773 ca->parent = cgroup_ca(cgrp->parent);
8779 percpu_counter_destroy(&ca->cpustat[i]);
8780 free_percpu(ca->cpuusage);
8784 return ERR_PTR(-ENOMEM);
8787 /* destroy an existing cpu accounting group */
8789 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8791 struct cpuacct *ca = cgroup_ca(cgrp);
8794 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8795 percpu_counter_destroy(&ca->cpustat[i]);
8796 free_percpu(ca->cpuusage);
8800 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8802 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8805 #ifndef CONFIG_64BIT
8807 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8809 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8811 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8819 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8821 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8823 #ifndef CONFIG_64BIT
8825 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8827 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8829 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8835 /* return total cpu usage (in nanoseconds) of a group */
8836 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8838 struct cpuacct *ca = cgroup_ca(cgrp);
8839 u64 totalcpuusage = 0;
8842 for_each_present_cpu(i)
8843 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8845 return totalcpuusage;
8848 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8851 struct cpuacct *ca = cgroup_ca(cgrp);
8860 for_each_present_cpu(i)
8861 cpuacct_cpuusage_write(ca, i, 0);
8867 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8870 struct cpuacct *ca = cgroup_ca(cgroup);
8874 for_each_present_cpu(i) {
8875 percpu = cpuacct_cpuusage_read(ca, i);
8876 seq_printf(m, "%llu ", (unsigned long long) percpu);
8878 seq_printf(m, "\n");
8882 static const char *cpuacct_stat_desc[] = {
8883 [CPUACCT_STAT_USER] = "user",
8884 [CPUACCT_STAT_SYSTEM] = "system",
8887 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8888 struct cgroup_map_cb *cb)
8890 struct cpuacct *ca = cgroup_ca(cgrp);
8893 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8894 s64 val = percpu_counter_read(&ca->cpustat[i]);
8895 val = cputime64_to_clock_t(val);
8896 cb->fill(cb, cpuacct_stat_desc[i], val);
8901 static struct cftype files[] = {
8904 .read_u64 = cpuusage_read,
8905 .write_u64 = cpuusage_write,
8908 .name = "usage_percpu",
8909 .read_seq_string = cpuacct_percpu_seq_read,
8913 .read_map = cpuacct_stats_show,
8917 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8919 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8923 * charge this task's execution time to its accounting group.
8925 * called with rq->lock held.
8927 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8932 if (unlikely(!cpuacct_subsys.active))
8935 cpu = task_cpu(tsk);
8941 for (; ca; ca = ca->parent) {
8942 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8943 *cpuusage += cputime;
8950 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8951 * in cputime_t units. As a result, cpuacct_update_stats calls
8952 * percpu_counter_add with values large enough to always overflow the
8953 * per cpu batch limit causing bad SMP scalability.
8955 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8956 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8957 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8960 #define CPUACCT_BATCH \
8961 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8963 #define CPUACCT_BATCH 0
8967 * Charge the system/user time to the task's accounting group.
8969 static void cpuacct_update_stats(struct task_struct *tsk,
8970 enum cpuacct_stat_index idx, cputime_t val)
8973 int batch = CPUACCT_BATCH;
8975 if (unlikely(!cpuacct_subsys.active))
8982 __percpu_counter_add(&ca->cpustat[idx], val, batch);
8988 struct cgroup_subsys cpuacct_subsys = {
8990 .create = cpuacct_create,
8991 .destroy = cpuacct_destroy,
8992 .populate = cpuacct_populate,
8993 .subsys_id = cpuacct_subsys_id,
8995 #endif /* CONFIG_CGROUP_CPUACCT */
8999 int rcu_expedited_torture_stats(char *page)
9003 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9005 void synchronize_sched_expedited(void)
9008 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9010 #else /* #ifndef CONFIG_SMP */
9012 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9013 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9015 #define RCU_EXPEDITED_STATE_POST -2
9016 #define RCU_EXPEDITED_STATE_IDLE -1
9018 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9020 int rcu_expedited_torture_stats(char *page)
9025 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9026 for_each_online_cpu(cpu) {
9027 cnt += sprintf(&page[cnt], " %d:%d",
9028 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9030 cnt += sprintf(&page[cnt], "\n");
9033 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9035 static long synchronize_sched_expedited_count;
9038 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9039 * approach to force grace period to end quickly. This consumes
9040 * significant time on all CPUs, and is thus not recommended for
9041 * any sort of common-case code.
9043 * Note that it is illegal to call this function while holding any
9044 * lock that is acquired by a CPU-hotplug notifier. Failing to
9045 * observe this restriction will result in deadlock.
9047 void synchronize_sched_expedited(void)
9050 unsigned long flags;
9051 bool need_full_sync = 0;
9053 struct migration_req *req;
9057 smp_mb(); /* ensure prior mod happens before capturing snap. */
9058 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9060 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9062 if (trycount++ < 10)
9063 udelay(trycount * num_online_cpus());
9065 synchronize_sched();
9068 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9069 smp_mb(); /* ensure test happens before caller kfree */
9074 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9075 for_each_online_cpu(cpu) {
9077 req = &per_cpu(rcu_migration_req, cpu);
9078 init_completion(&req->done);
9080 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9081 raw_spin_lock_irqsave(&rq->lock, flags);
9082 list_add(&req->list, &rq->migration_queue);
9083 raw_spin_unlock_irqrestore(&rq->lock, flags);
9084 wake_up_process(rq->migration_thread);
9086 for_each_online_cpu(cpu) {
9087 rcu_expedited_state = cpu;
9088 req = &per_cpu(rcu_migration_req, cpu);
9090 wait_for_completion(&req->done);
9091 raw_spin_lock_irqsave(&rq->lock, flags);
9092 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9094 req->dest_cpu = RCU_MIGRATION_IDLE;
9095 raw_spin_unlock_irqrestore(&rq->lock, flags);
9097 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9098 synchronize_sched_expedited_count++;
9099 mutex_unlock(&rcu_sched_expedited_mutex);
9102 synchronize_sched();
9104 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9106 #endif /* #else #ifndef CONFIG_SMP */