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 against
918 * ttwu() so that task_cpu() reports a stable number.
920 * We need to make an exception for PF_STARTING tasks because the fork
921 * path might require task_rq_lock() to work, eg. it can call
922 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
924 static inline int task_is_waking(struct task_struct *p)
926 return unlikely((p->state == TASK_WAKING) && !(p->flags & PF_STARTING));
930 * __task_rq_lock - lock the runqueue a given task resides on.
931 * Must be called interrupts disabled.
933 static inline struct rq *__task_rq_lock(struct task_struct *p)
939 while (task_is_waking(p))
942 raw_spin_lock(&rq->lock);
943 if (likely(rq == task_rq(p) && !task_is_waking(p)))
945 raw_spin_unlock(&rq->lock);
950 * task_rq_lock - lock the runqueue a given task resides on and disable
951 * interrupts. Note the ordering: we can safely lookup the task_rq without
952 * explicitly disabling preemption.
954 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
960 while (task_is_waking(p))
962 local_irq_save(*flags);
964 raw_spin_lock(&rq->lock);
965 if (likely(rq == task_rq(p) && !task_is_waking(p)))
967 raw_spin_unlock_irqrestore(&rq->lock, *flags);
971 void task_rq_unlock_wait(struct task_struct *p)
973 struct rq *rq = task_rq(p);
975 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
976 raw_spin_unlock_wait(&rq->lock);
979 static void __task_rq_unlock(struct rq *rq)
982 raw_spin_unlock(&rq->lock);
985 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
988 raw_spin_unlock_irqrestore(&rq->lock, *flags);
992 * this_rq_lock - lock this runqueue and disable interrupts.
994 static struct rq *this_rq_lock(void)
1001 raw_spin_lock(&rq->lock);
1006 #ifdef CONFIG_SCHED_HRTICK
1008 * Use HR-timers to deliver accurate preemption points.
1010 * Its all a bit involved since we cannot program an hrt while holding the
1011 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1014 * When we get rescheduled we reprogram the hrtick_timer outside of the
1020 * - enabled by features
1021 * - hrtimer is actually high res
1023 static inline int hrtick_enabled(struct rq *rq)
1025 if (!sched_feat(HRTICK))
1027 if (!cpu_active(cpu_of(rq)))
1029 return hrtimer_is_hres_active(&rq->hrtick_timer);
1032 static void hrtick_clear(struct rq *rq)
1034 if (hrtimer_active(&rq->hrtick_timer))
1035 hrtimer_cancel(&rq->hrtick_timer);
1039 * High-resolution timer tick.
1040 * Runs from hardirq context with interrupts disabled.
1042 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1044 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1046 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1048 raw_spin_lock(&rq->lock);
1049 update_rq_clock(rq);
1050 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1051 raw_spin_unlock(&rq->lock);
1053 return HRTIMER_NORESTART;
1058 * called from hardirq (IPI) context
1060 static void __hrtick_start(void *arg)
1062 struct rq *rq = arg;
1064 raw_spin_lock(&rq->lock);
1065 hrtimer_restart(&rq->hrtick_timer);
1066 rq->hrtick_csd_pending = 0;
1067 raw_spin_unlock(&rq->lock);
1071 * Called to set the hrtick timer state.
1073 * called with rq->lock held and irqs disabled
1075 static void hrtick_start(struct rq *rq, u64 delay)
1077 struct hrtimer *timer = &rq->hrtick_timer;
1078 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1080 hrtimer_set_expires(timer, time);
1082 if (rq == this_rq()) {
1083 hrtimer_restart(timer);
1084 } else if (!rq->hrtick_csd_pending) {
1085 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1086 rq->hrtick_csd_pending = 1;
1091 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1093 int cpu = (int)(long)hcpu;
1096 case CPU_UP_CANCELED:
1097 case CPU_UP_CANCELED_FROZEN:
1098 case CPU_DOWN_PREPARE:
1099 case CPU_DOWN_PREPARE_FROZEN:
1101 case CPU_DEAD_FROZEN:
1102 hrtick_clear(cpu_rq(cpu));
1109 static __init void init_hrtick(void)
1111 hotcpu_notifier(hotplug_hrtick, 0);
1115 * Called to set the hrtick timer state.
1117 * called with rq->lock held and irqs disabled
1119 static void hrtick_start(struct rq *rq, u64 delay)
1121 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1122 HRTIMER_MODE_REL_PINNED, 0);
1125 static inline void init_hrtick(void)
1128 #endif /* CONFIG_SMP */
1130 static void init_rq_hrtick(struct rq *rq)
1133 rq->hrtick_csd_pending = 0;
1135 rq->hrtick_csd.flags = 0;
1136 rq->hrtick_csd.func = __hrtick_start;
1137 rq->hrtick_csd.info = rq;
1140 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1141 rq->hrtick_timer.function = hrtick;
1143 #else /* CONFIG_SCHED_HRTICK */
1144 static inline void hrtick_clear(struct rq *rq)
1148 static inline void init_rq_hrtick(struct rq *rq)
1152 static inline void init_hrtick(void)
1155 #endif /* CONFIG_SCHED_HRTICK */
1158 * resched_task - mark a task 'to be rescheduled now'.
1160 * On UP this means the setting of the need_resched flag, on SMP it
1161 * might also involve a cross-CPU call to trigger the scheduler on
1166 #ifndef tsk_is_polling
1167 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1170 static void resched_task(struct task_struct *p)
1174 assert_raw_spin_locked(&task_rq(p)->lock);
1176 if (test_tsk_need_resched(p))
1179 set_tsk_need_resched(p);
1182 if (cpu == smp_processor_id())
1185 /* NEED_RESCHED must be visible before we test polling */
1187 if (!tsk_is_polling(p))
1188 smp_send_reschedule(cpu);
1191 static void resched_cpu(int cpu)
1193 struct rq *rq = cpu_rq(cpu);
1194 unsigned long flags;
1196 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1198 resched_task(cpu_curr(cpu));
1199 raw_spin_unlock_irqrestore(&rq->lock, flags);
1204 * When add_timer_on() enqueues a timer into the timer wheel of an
1205 * idle CPU then this timer might expire before the next timer event
1206 * which is scheduled to wake up that CPU. In case of a completely
1207 * idle system the next event might even be infinite time into the
1208 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1209 * leaves the inner idle loop so the newly added timer is taken into
1210 * account when the CPU goes back to idle and evaluates the timer
1211 * wheel for the next timer event.
1213 void wake_up_idle_cpu(int cpu)
1215 struct rq *rq = cpu_rq(cpu);
1217 if (cpu == smp_processor_id())
1221 * This is safe, as this function is called with the timer
1222 * wheel base lock of (cpu) held. When the CPU is on the way
1223 * to idle and has not yet set rq->curr to idle then it will
1224 * be serialized on the timer wheel base lock and take the new
1225 * timer into account automatically.
1227 if (rq->curr != rq->idle)
1231 * We can set TIF_RESCHED on the idle task of the other CPU
1232 * lockless. The worst case is that the other CPU runs the
1233 * idle task through an additional NOOP schedule()
1235 set_tsk_need_resched(rq->idle);
1237 /* NEED_RESCHED must be visible before we test polling */
1239 if (!tsk_is_polling(rq->idle))
1240 smp_send_reschedule(cpu);
1243 int nohz_ratelimit(int cpu)
1245 struct rq *rq = cpu_rq(cpu);
1246 u64 diff = rq->clock - rq->nohz_stamp;
1248 rq->nohz_stamp = rq->clock;
1250 return diff < (NSEC_PER_SEC / HZ) >> 1;
1253 #endif /* CONFIG_NO_HZ */
1255 static u64 sched_avg_period(void)
1257 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1260 static void sched_avg_update(struct rq *rq)
1262 s64 period = sched_avg_period();
1264 while ((s64)(rq->clock - rq->age_stamp) > period) {
1265 rq->age_stamp += period;
1270 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1272 rq->rt_avg += rt_delta;
1273 sched_avg_update(rq);
1276 #else /* !CONFIG_SMP */
1277 static void resched_task(struct task_struct *p)
1279 assert_raw_spin_locked(&task_rq(p)->lock);
1280 set_tsk_need_resched(p);
1283 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1286 #endif /* CONFIG_SMP */
1288 #if BITS_PER_LONG == 32
1289 # define WMULT_CONST (~0UL)
1291 # define WMULT_CONST (1UL << 32)
1294 #define WMULT_SHIFT 32
1297 * Shift right and round:
1299 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1302 * delta *= weight / lw
1304 static unsigned long
1305 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1306 struct load_weight *lw)
1310 if (!lw->inv_weight) {
1311 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1314 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1318 tmp = (u64)delta_exec * weight;
1320 * Check whether we'd overflow the 64-bit multiplication:
1322 if (unlikely(tmp > WMULT_CONST))
1323 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1326 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1328 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1331 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1337 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1344 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1345 * of tasks with abnormal "nice" values across CPUs the contribution that
1346 * each task makes to its run queue's load is weighted according to its
1347 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1348 * scaled version of the new time slice allocation that they receive on time
1352 #define WEIGHT_IDLEPRIO 3
1353 #define WMULT_IDLEPRIO 1431655765
1356 * Nice levels are multiplicative, with a gentle 10% change for every
1357 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1358 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1359 * that remained on nice 0.
1361 * The "10% effect" is relative and cumulative: from _any_ nice level,
1362 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1363 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1364 * If a task goes up by ~10% and another task goes down by ~10% then
1365 * the relative distance between them is ~25%.)
1367 static const int prio_to_weight[40] = {
1368 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1369 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1370 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1371 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1372 /* 0 */ 1024, 820, 655, 526, 423,
1373 /* 5 */ 335, 272, 215, 172, 137,
1374 /* 10 */ 110, 87, 70, 56, 45,
1375 /* 15 */ 36, 29, 23, 18, 15,
1379 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1381 * In cases where the weight does not change often, we can use the
1382 * precalculated inverse to speed up arithmetics by turning divisions
1383 * into multiplications:
1385 static const u32 prio_to_wmult[40] = {
1386 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1387 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1388 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1389 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1390 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1391 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1392 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1393 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1396 /* Time spent by the tasks of the cpu accounting group executing in ... */
1397 enum cpuacct_stat_index {
1398 CPUACCT_STAT_USER, /* ... user mode */
1399 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1401 CPUACCT_STAT_NSTATS,
1404 #ifdef CONFIG_CGROUP_CPUACCT
1405 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1406 static void cpuacct_update_stats(struct task_struct *tsk,
1407 enum cpuacct_stat_index idx, cputime_t val);
1409 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1410 static inline void cpuacct_update_stats(struct task_struct *tsk,
1411 enum cpuacct_stat_index idx, cputime_t val) {}
1414 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1416 update_load_add(&rq->load, load);
1419 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1421 update_load_sub(&rq->load, load);
1424 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1425 typedef int (*tg_visitor)(struct task_group *, void *);
1428 * Iterate the full tree, calling @down when first entering a node and @up when
1429 * leaving it for the final time.
1431 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1433 struct task_group *parent, *child;
1437 parent = &root_task_group;
1439 ret = (*down)(parent, data);
1442 list_for_each_entry_rcu(child, &parent->children, siblings) {
1449 ret = (*up)(parent, data);
1454 parent = parent->parent;
1463 static int tg_nop(struct task_group *tg, void *data)
1470 /* Used instead of source_load when we know the type == 0 */
1471 static unsigned long weighted_cpuload(const int cpu)
1473 return cpu_rq(cpu)->load.weight;
1477 * Return a low guess at the load of a migration-source cpu weighted
1478 * according to the scheduling class and "nice" value.
1480 * We want to under-estimate the load of migration sources, to
1481 * balance conservatively.
1483 static unsigned long source_load(int cpu, int type)
1485 struct rq *rq = cpu_rq(cpu);
1486 unsigned long total = weighted_cpuload(cpu);
1488 if (type == 0 || !sched_feat(LB_BIAS))
1491 return min(rq->cpu_load[type-1], total);
1495 * Return a high guess at the load of a migration-target cpu weighted
1496 * according to the scheduling class and "nice" value.
1498 static unsigned long target_load(int cpu, int type)
1500 struct rq *rq = cpu_rq(cpu);
1501 unsigned long total = weighted_cpuload(cpu);
1503 if (type == 0 || !sched_feat(LB_BIAS))
1506 return max(rq->cpu_load[type-1], total);
1509 static struct sched_group *group_of(int cpu)
1511 struct sched_domain *sd = rcu_dereference_sched(cpu_rq(cpu)->sd);
1519 static unsigned long power_of(int cpu)
1521 struct sched_group *group = group_of(cpu);
1524 return SCHED_LOAD_SCALE;
1526 return group->cpu_power;
1529 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1531 static unsigned long cpu_avg_load_per_task(int cpu)
1533 struct rq *rq = cpu_rq(cpu);
1534 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1537 rq->avg_load_per_task = rq->load.weight / nr_running;
1539 rq->avg_load_per_task = 0;
1541 return rq->avg_load_per_task;
1544 #ifdef CONFIG_FAIR_GROUP_SCHED
1546 static __read_mostly unsigned long *update_shares_data;
1548 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1551 * Calculate and set the cpu's group shares.
1553 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1554 unsigned long sd_shares,
1555 unsigned long sd_rq_weight,
1556 unsigned long *usd_rq_weight)
1558 unsigned long shares, rq_weight;
1561 rq_weight = usd_rq_weight[cpu];
1564 rq_weight = NICE_0_LOAD;
1568 * \Sum_j shares_j * rq_weight_i
1569 * shares_i = -----------------------------
1570 * \Sum_j rq_weight_j
1572 shares = (sd_shares * rq_weight) / sd_rq_weight;
1573 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1575 if (abs(shares - tg->se[cpu]->load.weight) >
1576 sysctl_sched_shares_thresh) {
1577 struct rq *rq = cpu_rq(cpu);
1578 unsigned long flags;
1580 raw_spin_lock_irqsave(&rq->lock, flags);
1581 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1582 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1583 __set_se_shares(tg->se[cpu], shares);
1584 raw_spin_unlock_irqrestore(&rq->lock, flags);
1589 * Re-compute the task group their per cpu shares over the given domain.
1590 * This needs to be done in a bottom-up fashion because the rq weight of a
1591 * parent group depends on the shares of its child groups.
1593 static int tg_shares_up(struct task_group *tg, void *data)
1595 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1596 unsigned long *usd_rq_weight;
1597 struct sched_domain *sd = data;
1598 unsigned long flags;
1604 local_irq_save(flags);
1605 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1607 for_each_cpu(i, sched_domain_span(sd)) {
1608 weight = tg->cfs_rq[i]->load.weight;
1609 usd_rq_weight[i] = weight;
1611 rq_weight += weight;
1613 * If there are currently no tasks on the cpu pretend there
1614 * is one of average load so that when a new task gets to
1615 * run here it will not get delayed by group starvation.
1618 weight = NICE_0_LOAD;
1620 sum_weight += weight;
1621 shares += tg->cfs_rq[i]->shares;
1625 rq_weight = sum_weight;
1627 if ((!shares && rq_weight) || shares > tg->shares)
1628 shares = tg->shares;
1630 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1631 shares = tg->shares;
1633 for_each_cpu(i, sched_domain_span(sd))
1634 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1636 local_irq_restore(flags);
1642 * Compute the cpu's hierarchical load factor for each task group.
1643 * This needs to be done in a top-down fashion because the load of a child
1644 * group is a fraction of its parents load.
1646 static int tg_load_down(struct task_group *tg, void *data)
1649 long cpu = (long)data;
1652 load = cpu_rq(cpu)->load.weight;
1654 load = tg->parent->cfs_rq[cpu]->h_load;
1655 load *= tg->cfs_rq[cpu]->shares;
1656 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1659 tg->cfs_rq[cpu]->h_load = load;
1664 static void update_shares(struct sched_domain *sd)
1669 if (root_task_group_empty())
1672 now = cpu_clock(raw_smp_processor_id());
1673 elapsed = now - sd->last_update;
1675 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1676 sd->last_update = now;
1677 walk_tg_tree(tg_nop, tg_shares_up, sd);
1681 static void update_h_load(long cpu)
1683 if (root_task_group_empty())
1686 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1691 static inline void update_shares(struct sched_domain *sd)
1697 #ifdef CONFIG_PREEMPT
1699 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1702 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1703 * way at the expense of forcing extra atomic operations in all
1704 * invocations. This assures that the double_lock is acquired using the
1705 * same underlying policy as the spinlock_t on this architecture, which
1706 * reduces latency compared to the unfair variant below. However, it
1707 * also adds more overhead and therefore may reduce throughput.
1709 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1710 __releases(this_rq->lock)
1711 __acquires(busiest->lock)
1712 __acquires(this_rq->lock)
1714 raw_spin_unlock(&this_rq->lock);
1715 double_rq_lock(this_rq, busiest);
1722 * Unfair double_lock_balance: Optimizes throughput at the expense of
1723 * latency by eliminating extra atomic operations when the locks are
1724 * already in proper order on entry. This favors lower cpu-ids and will
1725 * grant the double lock to lower cpus over higher ids under contention,
1726 * regardless of entry order into the function.
1728 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1729 __releases(this_rq->lock)
1730 __acquires(busiest->lock)
1731 __acquires(this_rq->lock)
1735 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1736 if (busiest < this_rq) {
1737 raw_spin_unlock(&this_rq->lock);
1738 raw_spin_lock(&busiest->lock);
1739 raw_spin_lock_nested(&this_rq->lock,
1740 SINGLE_DEPTH_NESTING);
1743 raw_spin_lock_nested(&busiest->lock,
1744 SINGLE_DEPTH_NESTING);
1749 #endif /* CONFIG_PREEMPT */
1752 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1754 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1756 if (unlikely(!irqs_disabled())) {
1757 /* printk() doesn't work good under rq->lock */
1758 raw_spin_unlock(&this_rq->lock);
1762 return _double_lock_balance(this_rq, busiest);
1765 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1766 __releases(busiest->lock)
1768 raw_spin_unlock(&busiest->lock);
1769 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1773 * double_rq_lock - safely lock two runqueues
1775 * Note this does not disable interrupts like task_rq_lock,
1776 * you need to do so manually before calling.
1778 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1779 __acquires(rq1->lock)
1780 __acquires(rq2->lock)
1782 BUG_ON(!irqs_disabled());
1784 raw_spin_lock(&rq1->lock);
1785 __acquire(rq2->lock); /* Fake it out ;) */
1788 raw_spin_lock(&rq1->lock);
1789 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1791 raw_spin_lock(&rq2->lock);
1792 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1798 * double_rq_unlock - safely unlock two runqueues
1800 * Note this does not restore interrupts like task_rq_unlock,
1801 * you need to do so manually after calling.
1803 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1804 __releases(rq1->lock)
1805 __releases(rq2->lock)
1807 raw_spin_unlock(&rq1->lock);
1809 raw_spin_unlock(&rq2->lock);
1811 __release(rq2->lock);
1816 #ifdef CONFIG_FAIR_GROUP_SCHED
1817 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1820 cfs_rq->shares = shares;
1825 static void calc_load_account_active(struct rq *this_rq);
1826 static void update_sysctl(void);
1827 static int get_update_sysctl_factor(void);
1829 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1831 set_task_rq(p, cpu);
1834 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1835 * successfuly executed on another CPU. We must ensure that updates of
1836 * per-task data have been completed by this moment.
1839 task_thread_info(p)->cpu = cpu;
1843 static const struct sched_class rt_sched_class;
1845 #define sched_class_highest (&rt_sched_class)
1846 #define for_each_class(class) \
1847 for (class = sched_class_highest; class; class = class->next)
1849 #include "sched_stats.h"
1851 static void inc_nr_running(struct rq *rq)
1856 static void dec_nr_running(struct rq *rq)
1861 static void set_load_weight(struct task_struct *p)
1863 if (task_has_rt_policy(p)) {
1864 p->se.load.weight = prio_to_weight[0] * 2;
1865 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1870 * SCHED_IDLE tasks get minimal weight:
1872 if (p->policy == SCHED_IDLE) {
1873 p->se.load.weight = WEIGHT_IDLEPRIO;
1874 p->se.load.inv_weight = WMULT_IDLEPRIO;
1878 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1879 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1882 static void update_avg(u64 *avg, u64 sample)
1884 s64 diff = sample - *avg;
1889 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1891 update_rq_clock(rq);
1892 sched_info_queued(p);
1893 p->sched_class->enqueue_task(rq, p, wakeup, head);
1897 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1899 update_rq_clock(rq);
1900 sched_info_dequeued(p);
1901 p->sched_class->dequeue_task(rq, p, sleep);
1906 * activate_task - move a task to the runqueue.
1908 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1910 if (task_contributes_to_load(p))
1911 rq->nr_uninterruptible--;
1913 enqueue_task(rq, p, wakeup, false);
1918 * deactivate_task - remove a task from the runqueue.
1920 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1922 if (task_contributes_to_load(p))
1923 rq->nr_uninterruptible++;
1925 dequeue_task(rq, p, sleep);
1929 #include "sched_idletask.c"
1930 #include "sched_fair.c"
1931 #include "sched_rt.c"
1932 #ifdef CONFIG_SCHED_DEBUG
1933 # include "sched_debug.c"
1937 * __normal_prio - return the priority that is based on the static prio
1939 static inline int __normal_prio(struct task_struct *p)
1941 return p->static_prio;
1945 * Calculate the expected normal priority: i.e. priority
1946 * without taking RT-inheritance into account. Might be
1947 * boosted by interactivity modifiers. Changes upon fork,
1948 * setprio syscalls, and whenever the interactivity
1949 * estimator recalculates.
1951 static inline int normal_prio(struct task_struct *p)
1955 if (task_has_rt_policy(p))
1956 prio = MAX_RT_PRIO-1 - p->rt_priority;
1958 prio = __normal_prio(p);
1963 * Calculate the current priority, i.e. the priority
1964 * taken into account by the scheduler. This value might
1965 * be boosted by RT tasks, or might be boosted by
1966 * interactivity modifiers. Will be RT if the task got
1967 * RT-boosted. If not then it returns p->normal_prio.
1969 static int effective_prio(struct task_struct *p)
1971 p->normal_prio = normal_prio(p);
1973 * If we are RT tasks or we were boosted to RT priority,
1974 * keep the priority unchanged. Otherwise, update priority
1975 * to the normal priority:
1977 if (!rt_prio(p->prio))
1978 return p->normal_prio;
1983 * task_curr - is this task currently executing on a CPU?
1984 * @p: the task in question.
1986 inline int task_curr(const struct task_struct *p)
1988 return cpu_curr(task_cpu(p)) == p;
1991 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1992 const struct sched_class *prev_class,
1993 int oldprio, int running)
1995 if (prev_class != p->sched_class) {
1996 if (prev_class->switched_from)
1997 prev_class->switched_from(rq, p, running);
1998 p->sched_class->switched_to(rq, p, running);
2000 p->sched_class->prio_changed(rq, p, oldprio, running);
2005 * Is this task likely cache-hot:
2008 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2012 if (p->sched_class != &fair_sched_class)
2016 * Buddy candidates are cache hot:
2018 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2019 (&p->se == cfs_rq_of(&p->se)->next ||
2020 &p->se == cfs_rq_of(&p->se)->last))
2023 if (sysctl_sched_migration_cost == -1)
2025 if (sysctl_sched_migration_cost == 0)
2028 delta = now - p->se.exec_start;
2030 return delta < (s64)sysctl_sched_migration_cost;
2033 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2035 #ifdef CONFIG_SCHED_DEBUG
2037 * We should never call set_task_cpu() on a blocked task,
2038 * ttwu() will sort out the placement.
2040 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2041 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2044 trace_sched_migrate_task(p, new_cpu);
2046 if (task_cpu(p) != new_cpu) {
2047 p->se.nr_migrations++;
2048 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2051 __set_task_cpu(p, new_cpu);
2054 struct migration_req {
2055 struct list_head list;
2057 struct task_struct *task;
2060 struct completion done;
2064 * The task's runqueue lock must be held.
2065 * Returns true if you have to wait for migration thread.
2068 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2070 struct rq *rq = task_rq(p);
2073 * If the task is not on a runqueue (and not running), then
2074 * the next wake-up will properly place the task.
2076 if (!p->se.on_rq && !task_running(rq, p))
2079 init_completion(&req->done);
2081 req->dest_cpu = dest_cpu;
2082 list_add(&req->list, &rq->migration_queue);
2088 * wait_task_context_switch - wait for a thread to complete at least one
2091 * @p must not be current.
2093 void wait_task_context_switch(struct task_struct *p)
2095 unsigned long nvcsw, nivcsw, flags;
2103 * The runqueue is assigned before the actual context
2104 * switch. We need to take the runqueue lock.
2106 * We could check initially without the lock but it is
2107 * very likely that we need to take the lock in every
2110 rq = task_rq_lock(p, &flags);
2111 running = task_running(rq, p);
2112 task_rq_unlock(rq, &flags);
2114 if (likely(!running))
2117 * The switch count is incremented before the actual
2118 * context switch. We thus wait for two switches to be
2119 * sure at least one completed.
2121 if ((p->nvcsw - nvcsw) > 1)
2123 if ((p->nivcsw - nivcsw) > 1)
2131 * wait_task_inactive - wait for a thread to unschedule.
2133 * If @match_state is nonzero, it's the @p->state value just checked and
2134 * not expected to change. If it changes, i.e. @p might have woken up,
2135 * then return zero. When we succeed in waiting for @p to be off its CPU,
2136 * we return a positive number (its total switch count). If a second call
2137 * a short while later returns the same number, the caller can be sure that
2138 * @p has remained unscheduled the whole time.
2140 * The caller must ensure that the task *will* unschedule sometime soon,
2141 * else this function might spin for a *long* time. This function can't
2142 * be called with interrupts off, or it may introduce deadlock with
2143 * smp_call_function() if an IPI is sent by the same process we are
2144 * waiting to become inactive.
2146 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2148 unsigned long flags;
2155 * We do the initial early heuristics without holding
2156 * any task-queue locks at all. We'll only try to get
2157 * the runqueue lock when things look like they will
2163 * If the task is actively running on another CPU
2164 * still, just relax and busy-wait without holding
2167 * NOTE! Since we don't hold any locks, it's not
2168 * even sure that "rq" stays as the right runqueue!
2169 * But we don't care, since "task_running()" will
2170 * return false if the runqueue has changed and p
2171 * is actually now running somewhere else!
2173 while (task_running(rq, p)) {
2174 if (match_state && unlikely(p->state != match_state))
2180 * Ok, time to look more closely! We need the rq
2181 * lock now, to be *sure*. If we're wrong, we'll
2182 * just go back and repeat.
2184 rq = task_rq_lock(p, &flags);
2185 trace_sched_wait_task(rq, p);
2186 running = task_running(rq, p);
2187 on_rq = p->se.on_rq;
2189 if (!match_state || p->state == match_state)
2190 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2191 task_rq_unlock(rq, &flags);
2194 * If it changed from the expected state, bail out now.
2196 if (unlikely(!ncsw))
2200 * Was it really running after all now that we
2201 * checked with the proper locks actually held?
2203 * Oops. Go back and try again..
2205 if (unlikely(running)) {
2211 * It's not enough that it's not actively running,
2212 * it must be off the runqueue _entirely_, and not
2215 * So if it was still runnable (but just not actively
2216 * running right now), it's preempted, and we should
2217 * yield - it could be a while.
2219 if (unlikely(on_rq)) {
2220 schedule_timeout_uninterruptible(1);
2225 * Ahh, all good. It wasn't running, and it wasn't
2226 * runnable, which means that it will never become
2227 * running in the future either. We're all done!
2236 * kick_process - kick a running thread to enter/exit the kernel
2237 * @p: the to-be-kicked thread
2239 * Cause a process which is running on another CPU to enter
2240 * kernel-mode, without any delay. (to get signals handled.)
2242 * NOTE: this function doesnt have to take the runqueue lock,
2243 * because all it wants to ensure is that the remote task enters
2244 * the kernel. If the IPI races and the task has been migrated
2245 * to another CPU then no harm is done and the purpose has been
2248 void kick_process(struct task_struct *p)
2254 if ((cpu != smp_processor_id()) && task_curr(p))
2255 smp_send_reschedule(cpu);
2258 EXPORT_SYMBOL_GPL(kick_process);
2259 #endif /* CONFIG_SMP */
2262 * task_oncpu_function_call - call a function on the cpu on which a task runs
2263 * @p: the task to evaluate
2264 * @func: the function to be called
2265 * @info: the function call argument
2267 * Calls the function @func when the task is currently running. This might
2268 * be on the current CPU, which just calls the function directly
2270 void task_oncpu_function_call(struct task_struct *p,
2271 void (*func) (void *info), void *info)
2278 smp_call_function_single(cpu, func, info, 1);
2283 static int select_fallback_rq(int cpu, struct task_struct *p)
2286 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2288 /* Look for allowed, online CPU in same node. */
2289 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2290 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2293 /* Any allowed, online CPU? */
2294 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2295 if (dest_cpu < nr_cpu_ids)
2298 /* No more Mr. Nice Guy. */
2299 if (dest_cpu >= nr_cpu_ids) {
2301 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2303 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2306 * Don't tell them about moving exiting tasks or
2307 * kernel threads (both mm NULL), since they never
2310 if (p->mm && printk_ratelimit()) {
2311 printk(KERN_INFO "process %d (%s) no "
2312 "longer affine to cpu%d\n",
2313 task_pid_nr(p), p->comm, cpu);
2321 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2322 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2325 * exec: is unstable, retry loop
2326 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2329 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2331 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2334 * In order not to call set_task_cpu() on a blocking task we need
2335 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2338 * Since this is common to all placement strategies, this lives here.
2340 * [ this allows ->select_task() to simply return task_cpu(p) and
2341 * not worry about this generic constraint ]
2343 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2345 cpu = select_fallback_rq(task_cpu(p), p);
2352 * try_to_wake_up - wake up a thread
2353 * @p: the to-be-woken-up thread
2354 * @state: the mask of task states that can be woken
2355 * @sync: do a synchronous wakeup?
2357 * Put it on the run-queue if it's not already there. The "current"
2358 * thread is always on the run-queue (except when the actual
2359 * re-schedule is in progress), and as such you're allowed to do
2360 * the simpler "current->state = TASK_RUNNING" to mark yourself
2361 * runnable without the overhead of this.
2363 * returns failure only if the task is already active.
2365 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2368 int cpu, orig_cpu, this_cpu, success = 0;
2369 unsigned long flags;
2372 if (!sched_feat(SYNC_WAKEUPS))
2373 wake_flags &= ~WF_SYNC;
2375 this_cpu = get_cpu();
2378 rq = task_rq_lock(p, &flags);
2379 if (!(p->state & state))
2389 if (unlikely(task_running(rq, p)))
2393 * In order to handle concurrent wakeups and release the rq->lock
2394 * we put the task in TASK_WAKING state.
2396 * First fix up the nr_uninterruptible count:
2398 if (task_contributes_to_load(p))
2399 rq->nr_uninterruptible--;
2400 p->state = TASK_WAKING;
2402 if (p->sched_class->task_waking)
2403 p->sched_class->task_waking(rq, p);
2405 __task_rq_unlock(rq);
2407 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2408 if (cpu != orig_cpu) {
2410 * Since we migrate the task without holding any rq->lock,
2411 * we need to be careful with task_rq_lock(), since that
2412 * might end up locking an invalid rq.
2414 set_task_cpu(p, cpu);
2418 raw_spin_lock(&rq->lock);
2421 * We migrated the task without holding either rq->lock, however
2422 * since the task is not on the task list itself, nobody else
2423 * will try and migrate the task, hence the rq should match the
2424 * cpu we just moved it to.
2426 WARN_ON(task_cpu(p) != cpu);
2427 WARN_ON(p->state != TASK_WAKING);
2429 #ifdef CONFIG_SCHEDSTATS
2430 schedstat_inc(rq, ttwu_count);
2431 if (cpu == this_cpu)
2432 schedstat_inc(rq, ttwu_local);
2434 struct sched_domain *sd;
2435 for_each_domain(this_cpu, sd) {
2436 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2437 schedstat_inc(sd, ttwu_wake_remote);
2442 #endif /* CONFIG_SCHEDSTATS */
2445 #endif /* CONFIG_SMP */
2446 schedstat_inc(p, se.statistics.nr_wakeups);
2447 if (wake_flags & WF_SYNC)
2448 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2449 if (orig_cpu != cpu)
2450 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2451 if (cpu == this_cpu)
2452 schedstat_inc(p, se.statistics.nr_wakeups_local);
2454 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2455 activate_task(rq, p, 1);
2459 trace_sched_wakeup(rq, p, success);
2460 check_preempt_curr(rq, p, wake_flags);
2462 p->state = TASK_RUNNING;
2464 if (p->sched_class->task_woken)
2465 p->sched_class->task_woken(rq, p);
2467 if (unlikely(rq->idle_stamp)) {
2468 u64 delta = rq->clock - rq->idle_stamp;
2469 u64 max = 2*sysctl_sched_migration_cost;
2474 update_avg(&rq->avg_idle, delta);
2479 task_rq_unlock(rq, &flags);
2486 * wake_up_process - Wake up a specific process
2487 * @p: The process to be woken up.
2489 * Attempt to wake up the nominated process and move it to the set of runnable
2490 * processes. Returns 1 if the process was woken up, 0 if it was already
2493 * It may be assumed that this function implies a write memory barrier before
2494 * changing the task state if and only if any tasks are woken up.
2496 int wake_up_process(struct task_struct *p)
2498 return try_to_wake_up(p, TASK_ALL, 0);
2500 EXPORT_SYMBOL(wake_up_process);
2502 int wake_up_state(struct task_struct *p, unsigned int state)
2504 return try_to_wake_up(p, state, 0);
2508 * Perform scheduler related setup for a newly forked process p.
2509 * p is forked by current.
2511 * __sched_fork() is basic setup used by init_idle() too:
2513 static void __sched_fork(struct task_struct *p)
2515 p->se.exec_start = 0;
2516 p->se.sum_exec_runtime = 0;
2517 p->se.prev_sum_exec_runtime = 0;
2518 p->se.nr_migrations = 0;
2520 #ifdef CONFIG_SCHEDSTATS
2521 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2524 INIT_LIST_HEAD(&p->rt.run_list);
2526 INIT_LIST_HEAD(&p->se.group_node);
2528 #ifdef CONFIG_PREEMPT_NOTIFIERS
2529 INIT_HLIST_HEAD(&p->preempt_notifiers);
2534 * fork()/clone()-time setup:
2536 void sched_fork(struct task_struct *p, int clone_flags)
2538 int cpu = get_cpu();
2542 * We mark the process as waking here. This guarantees that
2543 * nobody will actually run it, and a signal or other external
2544 * event cannot wake it up and insert it on the runqueue either.
2546 p->state = TASK_WAKING;
2549 * Revert to default priority/policy on fork if requested.
2551 if (unlikely(p->sched_reset_on_fork)) {
2552 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2553 p->policy = SCHED_NORMAL;
2554 p->normal_prio = p->static_prio;
2557 if (PRIO_TO_NICE(p->static_prio) < 0) {
2558 p->static_prio = NICE_TO_PRIO(0);
2559 p->normal_prio = p->static_prio;
2564 * We don't need the reset flag anymore after the fork. It has
2565 * fulfilled its duty:
2567 p->sched_reset_on_fork = 0;
2571 * Make sure we do not leak PI boosting priority to the child.
2573 p->prio = current->normal_prio;
2575 if (!rt_prio(p->prio))
2576 p->sched_class = &fair_sched_class;
2578 if (p->sched_class->task_fork)
2579 p->sched_class->task_fork(p);
2581 set_task_cpu(p, cpu);
2583 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2584 if (likely(sched_info_on()))
2585 memset(&p->sched_info, 0, sizeof(p->sched_info));
2587 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2590 #ifdef CONFIG_PREEMPT
2591 /* Want to start with kernel preemption disabled. */
2592 task_thread_info(p)->preempt_count = 1;
2594 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2600 * wake_up_new_task - wake up a newly created task for the first time.
2602 * This function will do some initial scheduler statistics housekeeping
2603 * that must be done for every newly created context, then puts the task
2604 * on the runqueue and wakes it.
2606 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2608 unsigned long flags;
2610 int cpu = get_cpu();
2614 * Fork balancing, do it here and not earlier because:
2615 * - cpus_allowed can change in the fork path
2616 * - any previously selected cpu might disappear through hotplug
2618 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2619 * ->cpus_allowed is stable, we have preemption disabled, meaning
2620 * cpu_online_mask is stable.
2622 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2623 set_task_cpu(p, cpu);
2627 * Since the task is not on the rq and we still have TASK_WAKING set
2628 * nobody else will migrate this task.
2631 raw_spin_lock_irqsave(&rq->lock, flags);
2633 BUG_ON(p->state != TASK_WAKING);
2634 p->state = TASK_RUNNING;
2635 activate_task(rq, p, 0);
2636 trace_sched_wakeup_new(rq, p, 1);
2637 check_preempt_curr(rq, p, WF_FORK);
2639 if (p->sched_class->task_woken)
2640 p->sched_class->task_woken(rq, p);
2642 task_rq_unlock(rq, &flags);
2646 #ifdef CONFIG_PREEMPT_NOTIFIERS
2649 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2650 * @notifier: notifier struct to register
2652 void preempt_notifier_register(struct preempt_notifier *notifier)
2654 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2656 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2659 * preempt_notifier_unregister - no longer interested in preemption notifications
2660 * @notifier: notifier struct to unregister
2662 * This is safe to call from within a preemption notifier.
2664 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2666 hlist_del(¬ifier->link);
2668 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2670 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2672 struct preempt_notifier *notifier;
2673 struct hlist_node *node;
2675 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2676 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2680 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2681 struct task_struct *next)
2683 struct preempt_notifier *notifier;
2684 struct hlist_node *node;
2686 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2687 notifier->ops->sched_out(notifier, next);
2690 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2692 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2697 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2698 struct task_struct *next)
2702 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2705 * prepare_task_switch - prepare to switch tasks
2706 * @rq: the runqueue preparing to switch
2707 * @prev: the current task that is being switched out
2708 * @next: the task we are going to switch to.
2710 * This is called with the rq lock held and interrupts off. It must
2711 * be paired with a subsequent finish_task_switch after the context
2714 * prepare_task_switch sets up locking and calls architecture specific
2718 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2719 struct task_struct *next)
2721 fire_sched_out_preempt_notifiers(prev, next);
2722 prepare_lock_switch(rq, next);
2723 prepare_arch_switch(next);
2727 * finish_task_switch - clean up after a task-switch
2728 * @rq: runqueue associated with task-switch
2729 * @prev: the thread we just switched away from.
2731 * finish_task_switch must be called after the context switch, paired
2732 * with a prepare_task_switch call before the context switch.
2733 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2734 * and do any other architecture-specific cleanup actions.
2736 * Note that we may have delayed dropping an mm in context_switch(). If
2737 * so, we finish that here outside of the runqueue lock. (Doing it
2738 * with the lock held can cause deadlocks; see schedule() for
2741 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2742 __releases(rq->lock)
2744 struct mm_struct *mm = rq->prev_mm;
2750 * A task struct has one reference for the use as "current".
2751 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2752 * schedule one last time. The schedule call will never return, and
2753 * the scheduled task must drop that reference.
2754 * The test for TASK_DEAD must occur while the runqueue locks are
2755 * still held, otherwise prev could be scheduled on another cpu, die
2756 * there before we look at prev->state, and then the reference would
2758 * Manfred Spraul <manfred@colorfullife.com>
2760 prev_state = prev->state;
2761 finish_arch_switch(prev);
2762 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2763 local_irq_disable();
2764 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2765 perf_event_task_sched_in(current);
2766 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2768 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2769 finish_lock_switch(rq, prev);
2771 fire_sched_in_preempt_notifiers(current);
2774 if (unlikely(prev_state == TASK_DEAD)) {
2776 * Remove function-return probe instances associated with this
2777 * task and put them back on the free list.
2779 kprobe_flush_task(prev);
2780 put_task_struct(prev);
2786 /* assumes rq->lock is held */
2787 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2789 if (prev->sched_class->pre_schedule)
2790 prev->sched_class->pre_schedule(rq, prev);
2793 /* rq->lock is NOT held, but preemption is disabled */
2794 static inline void post_schedule(struct rq *rq)
2796 if (rq->post_schedule) {
2797 unsigned long flags;
2799 raw_spin_lock_irqsave(&rq->lock, flags);
2800 if (rq->curr->sched_class->post_schedule)
2801 rq->curr->sched_class->post_schedule(rq);
2802 raw_spin_unlock_irqrestore(&rq->lock, flags);
2804 rq->post_schedule = 0;
2810 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2814 static inline void post_schedule(struct rq *rq)
2821 * schedule_tail - first thing a freshly forked thread must call.
2822 * @prev: the thread we just switched away from.
2824 asmlinkage void schedule_tail(struct task_struct *prev)
2825 __releases(rq->lock)
2827 struct rq *rq = this_rq();
2829 finish_task_switch(rq, prev);
2832 * FIXME: do we need to worry about rq being invalidated by the
2837 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2838 /* In this case, finish_task_switch does not reenable preemption */
2841 if (current->set_child_tid)
2842 put_user(task_pid_vnr(current), current->set_child_tid);
2846 * context_switch - switch to the new MM and the new
2847 * thread's register state.
2850 context_switch(struct rq *rq, struct task_struct *prev,
2851 struct task_struct *next)
2853 struct mm_struct *mm, *oldmm;
2855 prepare_task_switch(rq, prev, next);
2856 trace_sched_switch(rq, prev, next);
2858 oldmm = prev->active_mm;
2860 * For paravirt, this is coupled with an exit in switch_to to
2861 * combine the page table reload and the switch backend into
2864 arch_start_context_switch(prev);
2867 next->active_mm = oldmm;
2868 atomic_inc(&oldmm->mm_count);
2869 enter_lazy_tlb(oldmm, next);
2871 switch_mm(oldmm, mm, next);
2873 if (likely(!prev->mm)) {
2874 prev->active_mm = NULL;
2875 rq->prev_mm = oldmm;
2878 * Since the runqueue lock will be released by the next
2879 * task (which is an invalid locking op but in the case
2880 * of the scheduler it's an obvious special-case), so we
2881 * do an early lockdep release here:
2883 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2884 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2887 /* Here we just switch the register state and the stack. */
2888 switch_to(prev, next, prev);
2892 * this_rq must be evaluated again because prev may have moved
2893 * CPUs since it called schedule(), thus the 'rq' on its stack
2894 * frame will be invalid.
2896 finish_task_switch(this_rq(), prev);
2900 * nr_running, nr_uninterruptible and nr_context_switches:
2902 * externally visible scheduler statistics: current number of runnable
2903 * threads, current number of uninterruptible-sleeping threads, total
2904 * number of context switches performed since bootup.
2906 unsigned long nr_running(void)
2908 unsigned long i, sum = 0;
2910 for_each_online_cpu(i)
2911 sum += cpu_rq(i)->nr_running;
2916 unsigned long nr_uninterruptible(void)
2918 unsigned long i, sum = 0;
2920 for_each_possible_cpu(i)
2921 sum += cpu_rq(i)->nr_uninterruptible;
2924 * Since we read the counters lockless, it might be slightly
2925 * inaccurate. Do not allow it to go below zero though:
2927 if (unlikely((long)sum < 0))
2933 unsigned long long nr_context_switches(void)
2936 unsigned long long sum = 0;
2938 for_each_possible_cpu(i)
2939 sum += cpu_rq(i)->nr_switches;
2944 unsigned long nr_iowait(void)
2946 unsigned long i, sum = 0;
2948 for_each_possible_cpu(i)
2949 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2954 unsigned long nr_iowait_cpu(void)
2956 struct rq *this = this_rq();
2957 return atomic_read(&this->nr_iowait);
2960 unsigned long this_cpu_load(void)
2962 struct rq *this = this_rq();
2963 return this->cpu_load[0];
2967 /* Variables and functions for calc_load */
2968 static atomic_long_t calc_load_tasks;
2969 static unsigned long calc_load_update;
2970 unsigned long avenrun[3];
2971 EXPORT_SYMBOL(avenrun);
2974 * get_avenrun - get the load average array
2975 * @loads: pointer to dest load array
2976 * @offset: offset to add
2977 * @shift: shift count to shift the result left
2979 * These values are estimates at best, so no need for locking.
2981 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2983 loads[0] = (avenrun[0] + offset) << shift;
2984 loads[1] = (avenrun[1] + offset) << shift;
2985 loads[2] = (avenrun[2] + offset) << shift;
2988 static unsigned long
2989 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2992 load += active * (FIXED_1 - exp);
2993 return load >> FSHIFT;
2997 * calc_load - update the avenrun load estimates 10 ticks after the
2998 * CPUs have updated calc_load_tasks.
3000 void calc_global_load(void)
3002 unsigned long upd = calc_load_update + 10;
3005 if (time_before(jiffies, upd))
3008 active = atomic_long_read(&calc_load_tasks);
3009 active = active > 0 ? active * FIXED_1 : 0;
3011 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3012 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3013 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3015 calc_load_update += LOAD_FREQ;
3019 * Either called from update_cpu_load() or from a cpu going idle
3021 static void calc_load_account_active(struct rq *this_rq)
3023 long nr_active, delta;
3025 nr_active = this_rq->nr_running;
3026 nr_active += (long) this_rq->nr_uninterruptible;
3028 if (nr_active != this_rq->calc_load_active) {
3029 delta = nr_active - this_rq->calc_load_active;
3030 this_rq->calc_load_active = nr_active;
3031 atomic_long_add(delta, &calc_load_tasks);
3036 * Update rq->cpu_load[] statistics. This function is usually called every
3037 * scheduler tick (TICK_NSEC).
3039 static void update_cpu_load(struct rq *this_rq)
3041 unsigned long this_load = this_rq->load.weight;
3044 this_rq->nr_load_updates++;
3046 /* Update our load: */
3047 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3048 unsigned long old_load, new_load;
3050 /* scale is effectively 1 << i now, and >> i divides by scale */
3052 old_load = this_rq->cpu_load[i];
3053 new_load = this_load;
3055 * Round up the averaging division if load is increasing. This
3056 * prevents us from getting stuck on 9 if the load is 10, for
3059 if (new_load > old_load)
3060 new_load += scale-1;
3061 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3064 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3065 this_rq->calc_load_update += LOAD_FREQ;
3066 calc_load_account_active(this_rq);
3073 * sched_exec - execve() is a valuable balancing opportunity, because at
3074 * this point the task has the smallest effective memory and cache footprint.
3076 void sched_exec(void)
3078 struct task_struct *p = current;
3079 struct migration_req req;
3080 int dest_cpu, this_cpu;
3081 unsigned long flags;
3085 this_cpu = get_cpu();
3086 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3087 if (dest_cpu == this_cpu) {
3092 rq = task_rq_lock(p, &flags);
3096 * select_task_rq() can race against ->cpus_allowed
3098 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3099 || unlikely(!cpu_active(dest_cpu))) {
3100 task_rq_unlock(rq, &flags);
3104 /* force the process onto the specified CPU */
3105 if (migrate_task(p, dest_cpu, &req)) {
3106 /* Need to wait for migration thread (might exit: take ref). */
3107 struct task_struct *mt = rq->migration_thread;
3109 get_task_struct(mt);
3110 task_rq_unlock(rq, &flags);
3111 wake_up_process(mt);
3112 put_task_struct(mt);
3113 wait_for_completion(&req.done);
3117 task_rq_unlock(rq, &flags);
3122 DEFINE_PER_CPU(struct kernel_stat, kstat);
3124 EXPORT_PER_CPU_SYMBOL(kstat);
3127 * Return any ns on the sched_clock that have not yet been accounted in
3128 * @p in case that task is currently running.
3130 * Called with task_rq_lock() held on @rq.
3132 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3136 if (task_current(rq, p)) {
3137 update_rq_clock(rq);
3138 ns = rq->clock - p->se.exec_start;
3146 unsigned long long task_delta_exec(struct task_struct *p)
3148 unsigned long flags;
3152 rq = task_rq_lock(p, &flags);
3153 ns = do_task_delta_exec(p, rq);
3154 task_rq_unlock(rq, &flags);
3160 * Return accounted runtime for the task.
3161 * In case the task is currently running, return the runtime plus current's
3162 * pending runtime that have not been accounted yet.
3164 unsigned long long task_sched_runtime(struct task_struct *p)
3166 unsigned long flags;
3170 rq = task_rq_lock(p, &flags);
3171 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3172 task_rq_unlock(rq, &flags);
3178 * Return sum_exec_runtime for the thread group.
3179 * In case the task is currently running, return the sum plus current's
3180 * pending runtime that have not been accounted yet.
3182 * Note that the thread group might have other running tasks as well,
3183 * so the return value not includes other pending runtime that other
3184 * running tasks might have.
3186 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3188 struct task_cputime totals;
3189 unsigned long flags;
3193 rq = task_rq_lock(p, &flags);
3194 thread_group_cputime(p, &totals);
3195 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3196 task_rq_unlock(rq, &flags);
3202 * Account user cpu time to a process.
3203 * @p: the process that the cpu time gets accounted to
3204 * @cputime: the cpu time spent in user space since the last update
3205 * @cputime_scaled: cputime scaled by cpu frequency
3207 void account_user_time(struct task_struct *p, cputime_t cputime,
3208 cputime_t cputime_scaled)
3210 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3213 /* Add user 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);
3218 /* Add user time to cpustat. */
3219 tmp = cputime_to_cputime64(cputime);
3220 if (TASK_NICE(p) > 0)
3221 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3223 cpustat->user = cputime64_add(cpustat->user, tmp);
3225 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3226 /* Account for user time used */
3227 acct_update_integrals(p);
3231 * Account guest cpu time to a process.
3232 * @p: the process that the cpu time gets accounted to
3233 * @cputime: the cpu time spent in virtual machine since the last update
3234 * @cputime_scaled: cputime scaled by cpu frequency
3236 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3237 cputime_t cputime_scaled)
3240 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3242 tmp = cputime_to_cputime64(cputime);
3244 /* Add guest time to process. */
3245 p->utime = cputime_add(p->utime, cputime);
3246 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3247 account_group_user_time(p, cputime);
3248 p->gtime = cputime_add(p->gtime, cputime);
3250 /* Add guest time to cpustat. */
3251 if (TASK_NICE(p) > 0) {
3252 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3253 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3255 cpustat->user = cputime64_add(cpustat->user, tmp);
3256 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3261 * Account system cpu time to a process.
3262 * @p: the process that the cpu time gets accounted to
3263 * @hardirq_offset: the offset to subtract from hardirq_count()
3264 * @cputime: the cpu time spent in kernel space since the last update
3265 * @cputime_scaled: cputime scaled by cpu frequency
3267 void account_system_time(struct task_struct *p, int hardirq_offset,
3268 cputime_t cputime, cputime_t cputime_scaled)
3270 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3273 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3274 account_guest_time(p, cputime, cputime_scaled);
3278 /* Add system time to process. */
3279 p->stime = cputime_add(p->stime, cputime);
3280 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3281 account_group_system_time(p, cputime);
3283 /* Add system time to cpustat. */
3284 tmp = cputime_to_cputime64(cputime);
3285 if (hardirq_count() - hardirq_offset)
3286 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3287 else if (softirq_count())
3288 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3290 cpustat->system = cputime64_add(cpustat->system, tmp);
3292 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3294 /* Account for system time used */
3295 acct_update_integrals(p);
3299 * Account for involuntary wait time.
3300 * @steal: the cpu time spent in involuntary wait
3302 void account_steal_time(cputime_t cputime)
3304 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3305 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3307 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3311 * Account for idle time.
3312 * @cputime: the cpu time spent in idle wait
3314 void account_idle_time(cputime_t cputime)
3316 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3317 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3318 struct rq *rq = this_rq();
3320 if (atomic_read(&rq->nr_iowait) > 0)
3321 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3323 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3326 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3329 * Account a single tick of cpu time.
3330 * @p: the process that the cpu time gets accounted to
3331 * @user_tick: indicates if the tick is a user or a system tick
3333 void account_process_tick(struct task_struct *p, int user_tick)
3335 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3336 struct rq *rq = this_rq();
3339 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3340 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3341 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3344 account_idle_time(cputime_one_jiffy);
3348 * Account multiple ticks of steal time.
3349 * @p: the process from which the cpu time has been stolen
3350 * @ticks: number of stolen ticks
3352 void account_steal_ticks(unsigned long ticks)
3354 account_steal_time(jiffies_to_cputime(ticks));
3358 * Account multiple ticks of idle time.
3359 * @ticks: number of stolen ticks
3361 void account_idle_ticks(unsigned long ticks)
3363 account_idle_time(jiffies_to_cputime(ticks));
3369 * Use precise platform statistics if available:
3371 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3372 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3378 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3380 struct task_cputime cputime;
3382 thread_group_cputime(p, &cputime);
3384 *ut = cputime.utime;
3385 *st = cputime.stime;
3389 #ifndef nsecs_to_cputime
3390 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3393 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3395 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3398 * Use CFS's precise accounting:
3400 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3405 temp = (u64)(rtime * utime);
3406 do_div(temp, total);
3407 utime = (cputime_t)temp;
3412 * Compare with previous values, to keep monotonicity:
3414 p->prev_utime = max(p->prev_utime, utime);
3415 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3417 *ut = p->prev_utime;
3418 *st = p->prev_stime;
3422 * Must be called with siglock held.
3424 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3426 struct signal_struct *sig = p->signal;
3427 struct task_cputime cputime;
3428 cputime_t rtime, utime, total;
3430 thread_group_cputime(p, &cputime);
3432 total = cputime_add(cputime.utime, cputime.stime);
3433 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3438 temp = (u64)(rtime * cputime.utime);
3439 do_div(temp, total);
3440 utime = (cputime_t)temp;
3444 sig->prev_utime = max(sig->prev_utime, utime);
3445 sig->prev_stime = max(sig->prev_stime,
3446 cputime_sub(rtime, sig->prev_utime));
3448 *ut = sig->prev_utime;
3449 *st = sig->prev_stime;
3454 * This function gets called by the timer code, with HZ frequency.
3455 * We call it with interrupts disabled.
3457 * It also gets called by the fork code, when changing the parent's
3460 void scheduler_tick(void)
3462 int cpu = smp_processor_id();
3463 struct rq *rq = cpu_rq(cpu);
3464 struct task_struct *curr = rq->curr;
3468 raw_spin_lock(&rq->lock);
3469 update_rq_clock(rq);
3470 update_cpu_load(rq);
3471 curr->sched_class->task_tick(rq, curr, 0);
3472 raw_spin_unlock(&rq->lock);
3474 perf_event_task_tick(curr);
3477 rq->idle_at_tick = idle_cpu(cpu);
3478 trigger_load_balance(rq, cpu);
3482 notrace unsigned long get_parent_ip(unsigned long addr)
3484 if (in_lock_functions(addr)) {
3485 addr = CALLER_ADDR2;
3486 if (in_lock_functions(addr))
3487 addr = CALLER_ADDR3;
3492 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3493 defined(CONFIG_PREEMPT_TRACER))
3495 void __kprobes add_preempt_count(int val)
3497 #ifdef CONFIG_DEBUG_PREEMPT
3501 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3504 preempt_count() += val;
3505 #ifdef CONFIG_DEBUG_PREEMPT
3507 * Spinlock count overflowing soon?
3509 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3512 if (preempt_count() == val)
3513 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3515 EXPORT_SYMBOL(add_preempt_count);
3517 void __kprobes sub_preempt_count(int val)
3519 #ifdef CONFIG_DEBUG_PREEMPT
3523 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3526 * Is the spinlock portion underflowing?
3528 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3529 !(preempt_count() & PREEMPT_MASK)))
3533 if (preempt_count() == val)
3534 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3535 preempt_count() -= val;
3537 EXPORT_SYMBOL(sub_preempt_count);
3542 * Print scheduling while atomic bug:
3544 static noinline void __schedule_bug(struct task_struct *prev)
3546 struct pt_regs *regs = get_irq_regs();
3548 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3549 prev->comm, prev->pid, preempt_count());
3551 debug_show_held_locks(prev);
3553 if (irqs_disabled())
3554 print_irqtrace_events(prev);
3563 * Various schedule()-time debugging checks and statistics:
3565 static inline void schedule_debug(struct task_struct *prev)
3568 * Test if we are atomic. Since do_exit() needs to call into
3569 * schedule() atomically, we ignore that path for now.
3570 * Otherwise, whine if we are scheduling when we should not be.
3572 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3573 __schedule_bug(prev);
3575 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3577 schedstat_inc(this_rq(), sched_count);
3578 #ifdef CONFIG_SCHEDSTATS
3579 if (unlikely(prev->lock_depth >= 0)) {
3580 schedstat_inc(this_rq(), bkl_count);
3581 schedstat_inc(prev, sched_info.bkl_count);
3586 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3589 update_rq_clock(rq);
3590 rq->skip_clock_update = 0;
3591 prev->sched_class->put_prev_task(rq, prev);
3595 * Pick up the highest-prio task:
3597 static inline struct task_struct *
3598 pick_next_task(struct rq *rq)
3600 const struct sched_class *class;
3601 struct task_struct *p;
3604 * Optimization: we know that if all tasks are in
3605 * the fair class we can call that function directly:
3607 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3608 p = fair_sched_class.pick_next_task(rq);
3613 class = sched_class_highest;
3615 p = class->pick_next_task(rq);
3619 * Will never be NULL as the idle class always
3620 * returns a non-NULL p:
3622 class = class->next;
3627 * schedule() is the main scheduler function.
3629 asmlinkage void __sched schedule(void)
3631 struct task_struct *prev, *next;
3632 unsigned long *switch_count;
3638 cpu = smp_processor_id();
3642 switch_count = &prev->nivcsw;
3644 release_kernel_lock(prev);
3645 need_resched_nonpreemptible:
3647 schedule_debug(prev);
3649 if (sched_feat(HRTICK))
3652 raw_spin_lock_irq(&rq->lock);
3653 clear_tsk_need_resched(prev);
3655 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3656 if (unlikely(signal_pending_state(prev->state, prev)))
3657 prev->state = TASK_RUNNING;
3659 deactivate_task(rq, prev, 1);
3660 switch_count = &prev->nvcsw;
3663 pre_schedule(rq, prev);
3665 if (unlikely(!rq->nr_running))
3666 idle_balance(cpu, rq);
3668 put_prev_task(rq, prev);
3669 next = pick_next_task(rq);
3671 if (likely(prev != next)) {
3672 sched_info_switch(prev, next);
3673 perf_event_task_sched_out(prev, next);
3679 context_switch(rq, prev, next); /* unlocks the rq */
3681 * the context switch might have flipped the stack from under
3682 * us, hence refresh the local variables.
3684 cpu = smp_processor_id();
3687 raw_spin_unlock_irq(&rq->lock);
3691 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3693 switch_count = &prev->nivcsw;
3694 goto need_resched_nonpreemptible;
3697 preempt_enable_no_resched();
3701 EXPORT_SYMBOL(schedule);
3703 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3705 * Look out! "owner" is an entirely speculative pointer
3706 * access and not reliable.
3708 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3713 if (!sched_feat(OWNER_SPIN))
3716 #ifdef CONFIG_DEBUG_PAGEALLOC
3718 * Need to access the cpu field knowing that
3719 * DEBUG_PAGEALLOC could have unmapped it if
3720 * the mutex owner just released it and exited.
3722 if (probe_kernel_address(&owner->cpu, cpu))
3729 * Even if the access succeeded (likely case),
3730 * the cpu field may no longer be valid.
3732 if (cpu >= nr_cpumask_bits)
3736 * We need to validate that we can do a
3737 * get_cpu() and that we have the percpu area.
3739 if (!cpu_online(cpu))
3746 * Owner changed, break to re-assess state.
3748 if (lock->owner != owner)
3752 * Is that owner really running on that cpu?
3754 if (task_thread_info(rq->curr) != owner || need_resched())
3764 #ifdef CONFIG_PREEMPT
3766 * this is the entry point to schedule() from in-kernel preemption
3767 * off of preempt_enable. Kernel preemptions off return from interrupt
3768 * occur there and call schedule directly.
3770 asmlinkage void __sched preempt_schedule(void)
3772 struct thread_info *ti = current_thread_info();
3775 * If there is a non-zero preempt_count or interrupts are disabled,
3776 * we do not want to preempt the current task. Just return..
3778 if (likely(ti->preempt_count || irqs_disabled()))
3782 add_preempt_count(PREEMPT_ACTIVE);
3784 sub_preempt_count(PREEMPT_ACTIVE);
3787 * Check again in case we missed a preemption opportunity
3788 * between schedule and now.
3791 } while (need_resched());
3793 EXPORT_SYMBOL(preempt_schedule);
3796 * this is the entry point to schedule() from kernel preemption
3797 * off of irq context.
3798 * Note, that this is called and return with irqs disabled. This will
3799 * protect us against recursive calling from irq.
3801 asmlinkage void __sched preempt_schedule_irq(void)
3803 struct thread_info *ti = current_thread_info();
3805 /* Catch callers which need to be fixed */
3806 BUG_ON(ti->preempt_count || !irqs_disabled());
3809 add_preempt_count(PREEMPT_ACTIVE);
3812 local_irq_disable();
3813 sub_preempt_count(PREEMPT_ACTIVE);
3816 * Check again in case we missed a preemption opportunity
3817 * between schedule and now.
3820 } while (need_resched());
3823 #endif /* CONFIG_PREEMPT */
3825 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3828 return try_to_wake_up(curr->private, mode, wake_flags);
3830 EXPORT_SYMBOL(default_wake_function);
3833 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3834 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3835 * number) then we wake all the non-exclusive tasks and one exclusive task.
3837 * There are circumstances in which we can try to wake a task which has already
3838 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3839 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3841 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3842 int nr_exclusive, int wake_flags, void *key)
3844 wait_queue_t *curr, *next;
3846 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3847 unsigned flags = curr->flags;
3849 if (curr->func(curr, mode, wake_flags, key) &&
3850 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3856 * __wake_up - wake up threads blocked on a waitqueue.
3858 * @mode: which threads
3859 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3860 * @key: is directly passed to the wakeup function
3862 * It may be assumed that this function implies a write memory barrier before
3863 * changing the task state if and only if any tasks are woken up.
3865 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3866 int nr_exclusive, void *key)
3868 unsigned long flags;
3870 spin_lock_irqsave(&q->lock, flags);
3871 __wake_up_common(q, mode, nr_exclusive, 0, key);
3872 spin_unlock_irqrestore(&q->lock, flags);
3874 EXPORT_SYMBOL(__wake_up);
3877 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3879 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3881 __wake_up_common(q, mode, 1, 0, NULL);
3884 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3886 __wake_up_common(q, mode, 1, 0, key);
3890 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3892 * @mode: which threads
3893 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3894 * @key: opaque value to be passed to wakeup targets
3896 * The sync wakeup differs that the waker knows that it will schedule
3897 * away soon, so while the target thread will be woken up, it will not
3898 * be migrated to another CPU - ie. the two threads are 'synchronized'
3899 * with each other. This can prevent needless bouncing between CPUs.
3901 * On UP it can prevent extra preemption.
3903 * It may be assumed that this function implies a write memory barrier before
3904 * changing the task state if and only if any tasks are woken up.
3906 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3907 int nr_exclusive, void *key)
3909 unsigned long flags;
3910 int wake_flags = WF_SYNC;
3915 if (unlikely(!nr_exclusive))
3918 spin_lock_irqsave(&q->lock, flags);
3919 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3920 spin_unlock_irqrestore(&q->lock, flags);
3922 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3925 * __wake_up_sync - see __wake_up_sync_key()
3927 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3929 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3931 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3934 * complete: - signals a single thread waiting on this completion
3935 * @x: holds the state of this particular completion
3937 * This will wake up a single thread waiting on this completion. Threads will be
3938 * awakened in the same order in which they were queued.
3940 * See also complete_all(), wait_for_completion() and related routines.
3942 * It may be assumed that this function implies a write memory barrier before
3943 * changing the task state if and only if any tasks are woken up.
3945 void complete(struct completion *x)
3947 unsigned long flags;
3949 spin_lock_irqsave(&x->wait.lock, flags);
3951 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3952 spin_unlock_irqrestore(&x->wait.lock, flags);
3954 EXPORT_SYMBOL(complete);
3957 * complete_all: - signals all threads waiting on this completion
3958 * @x: holds the state of this particular completion
3960 * This will wake up all threads waiting on this particular completion event.
3962 * It may be assumed that this function implies a write memory barrier before
3963 * changing the task state if and only if any tasks are woken up.
3965 void complete_all(struct completion *x)
3967 unsigned long flags;
3969 spin_lock_irqsave(&x->wait.lock, flags);
3970 x->done += UINT_MAX/2;
3971 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3972 spin_unlock_irqrestore(&x->wait.lock, flags);
3974 EXPORT_SYMBOL(complete_all);
3976 static inline long __sched
3977 do_wait_for_common(struct completion *x, long timeout, int state)
3980 DECLARE_WAITQUEUE(wait, current);
3982 wait.flags |= WQ_FLAG_EXCLUSIVE;
3983 __add_wait_queue_tail(&x->wait, &wait);
3985 if (signal_pending_state(state, current)) {
3986 timeout = -ERESTARTSYS;
3989 __set_current_state(state);
3990 spin_unlock_irq(&x->wait.lock);
3991 timeout = schedule_timeout(timeout);
3992 spin_lock_irq(&x->wait.lock);
3993 } while (!x->done && timeout);
3994 __remove_wait_queue(&x->wait, &wait);
3999 return timeout ?: 1;
4003 wait_for_common(struct completion *x, long timeout, int state)
4007 spin_lock_irq(&x->wait.lock);
4008 timeout = do_wait_for_common(x, timeout, state);
4009 spin_unlock_irq(&x->wait.lock);
4014 * wait_for_completion: - waits for completion of a task
4015 * @x: holds the state of this particular completion
4017 * This waits to be signaled for completion of a specific task. It is NOT
4018 * interruptible and there is no timeout.
4020 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4021 * and interrupt capability. Also see complete().
4023 void __sched wait_for_completion(struct completion *x)
4025 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4027 EXPORT_SYMBOL(wait_for_completion);
4030 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4031 * @x: holds the state of this particular completion
4032 * @timeout: timeout value in jiffies
4034 * This waits for either a completion of a specific task to be signaled or for a
4035 * specified timeout to expire. The timeout is in jiffies. It is not
4038 unsigned long __sched
4039 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4041 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4043 EXPORT_SYMBOL(wait_for_completion_timeout);
4046 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4047 * @x: holds the state of this particular completion
4049 * This waits for completion of a specific task to be signaled. It is
4052 int __sched wait_for_completion_interruptible(struct completion *x)
4054 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4055 if (t == -ERESTARTSYS)
4059 EXPORT_SYMBOL(wait_for_completion_interruptible);
4062 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4063 * @x: holds the state of this particular completion
4064 * @timeout: timeout value in jiffies
4066 * This waits for either a completion of a specific task to be signaled or for a
4067 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4069 unsigned long __sched
4070 wait_for_completion_interruptible_timeout(struct completion *x,
4071 unsigned long timeout)
4073 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4075 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4078 * wait_for_completion_killable: - waits for completion of a task (killable)
4079 * @x: holds the state of this particular completion
4081 * This waits to be signaled for completion of a specific task. It can be
4082 * interrupted by a kill signal.
4084 int __sched wait_for_completion_killable(struct completion *x)
4086 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4087 if (t == -ERESTARTSYS)
4091 EXPORT_SYMBOL(wait_for_completion_killable);
4094 * try_wait_for_completion - try to decrement a completion without blocking
4095 * @x: completion structure
4097 * Returns: 0 if a decrement cannot be done without blocking
4098 * 1 if a decrement succeeded.
4100 * If a completion is being used as a counting completion,
4101 * attempt to decrement the counter without blocking. This
4102 * enables us to avoid waiting if the resource the completion
4103 * is protecting is not available.
4105 bool try_wait_for_completion(struct completion *x)
4107 unsigned long flags;
4110 spin_lock_irqsave(&x->wait.lock, flags);
4115 spin_unlock_irqrestore(&x->wait.lock, flags);
4118 EXPORT_SYMBOL(try_wait_for_completion);
4121 * completion_done - Test to see if a completion has any waiters
4122 * @x: completion structure
4124 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4125 * 1 if there are no waiters.
4128 bool completion_done(struct completion *x)
4130 unsigned long flags;
4133 spin_lock_irqsave(&x->wait.lock, flags);
4136 spin_unlock_irqrestore(&x->wait.lock, flags);
4139 EXPORT_SYMBOL(completion_done);
4142 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4144 unsigned long flags;
4147 init_waitqueue_entry(&wait, current);
4149 __set_current_state(state);
4151 spin_lock_irqsave(&q->lock, flags);
4152 __add_wait_queue(q, &wait);
4153 spin_unlock(&q->lock);
4154 timeout = schedule_timeout(timeout);
4155 spin_lock_irq(&q->lock);
4156 __remove_wait_queue(q, &wait);
4157 spin_unlock_irqrestore(&q->lock, flags);
4162 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4164 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4166 EXPORT_SYMBOL(interruptible_sleep_on);
4169 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4171 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4173 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4175 void __sched sleep_on(wait_queue_head_t *q)
4177 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4179 EXPORT_SYMBOL(sleep_on);
4181 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4183 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4185 EXPORT_SYMBOL(sleep_on_timeout);
4187 #ifdef CONFIG_RT_MUTEXES
4190 * rt_mutex_setprio - set the current priority of a task
4192 * @prio: prio value (kernel-internal form)
4194 * This function changes the 'effective' priority of a task. It does
4195 * not touch ->normal_prio like __setscheduler().
4197 * Used by the rt_mutex code to implement priority inheritance logic.
4199 void rt_mutex_setprio(struct task_struct *p, int prio)
4201 unsigned long flags;
4202 int oldprio, on_rq, running;
4204 const struct sched_class *prev_class;
4206 BUG_ON(prio < 0 || prio > MAX_PRIO);
4208 rq = task_rq_lock(p, &flags);
4211 prev_class = p->sched_class;
4212 on_rq = p->se.on_rq;
4213 running = task_current(rq, p);
4215 dequeue_task(rq, p, 0);
4217 p->sched_class->put_prev_task(rq, p);
4220 p->sched_class = &rt_sched_class;
4222 p->sched_class = &fair_sched_class;
4227 p->sched_class->set_curr_task(rq);
4229 enqueue_task(rq, p, 0, oldprio < prio);
4231 check_class_changed(rq, p, prev_class, oldprio, running);
4233 task_rq_unlock(rq, &flags);
4238 void set_user_nice(struct task_struct *p, long nice)
4240 int old_prio, delta, on_rq;
4241 unsigned long flags;
4244 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4247 * We have to be careful, if called from sys_setpriority(),
4248 * the task might be in the middle of scheduling on another CPU.
4250 rq = task_rq_lock(p, &flags);
4252 * The RT priorities are set via sched_setscheduler(), but we still
4253 * allow the 'normal' nice value to be set - but as expected
4254 * it wont have any effect on scheduling until the task is
4255 * SCHED_FIFO/SCHED_RR:
4257 if (task_has_rt_policy(p)) {
4258 p->static_prio = NICE_TO_PRIO(nice);
4261 on_rq = p->se.on_rq;
4263 dequeue_task(rq, p, 0);
4265 p->static_prio = NICE_TO_PRIO(nice);
4268 p->prio = effective_prio(p);
4269 delta = p->prio - old_prio;
4272 enqueue_task(rq, p, 0, false);
4274 * If the task increased its priority or is running and
4275 * lowered its priority, then reschedule its CPU:
4277 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4278 resched_task(rq->curr);
4281 task_rq_unlock(rq, &flags);
4283 EXPORT_SYMBOL(set_user_nice);
4286 * can_nice - check if a task can reduce its nice value
4290 int can_nice(const struct task_struct *p, const int nice)
4292 /* convert nice value [19,-20] to rlimit style value [1,40] */
4293 int nice_rlim = 20 - nice;
4295 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4296 capable(CAP_SYS_NICE));
4299 #ifdef __ARCH_WANT_SYS_NICE
4302 * sys_nice - change the priority of the current process.
4303 * @increment: priority increment
4305 * sys_setpriority is a more generic, but much slower function that
4306 * does similar things.
4308 SYSCALL_DEFINE1(nice, int, increment)
4313 * Setpriority might change our priority at the same moment.
4314 * We don't have to worry. Conceptually one call occurs first
4315 * and we have a single winner.
4317 if (increment < -40)
4322 nice = TASK_NICE(current) + increment;
4328 if (increment < 0 && !can_nice(current, nice))
4331 retval = security_task_setnice(current, nice);
4335 set_user_nice(current, nice);
4342 * task_prio - return the priority value of a given task.
4343 * @p: the task in question.
4345 * This is the priority value as seen by users in /proc.
4346 * RT tasks are offset by -200. Normal tasks are centered
4347 * around 0, value goes from -16 to +15.
4349 int task_prio(const struct task_struct *p)
4351 return p->prio - MAX_RT_PRIO;
4355 * task_nice - return the nice value of a given task.
4356 * @p: the task in question.
4358 int task_nice(const struct task_struct *p)
4360 return TASK_NICE(p);
4362 EXPORT_SYMBOL(task_nice);
4365 * idle_cpu - is a given cpu idle currently?
4366 * @cpu: the processor in question.
4368 int idle_cpu(int cpu)
4370 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4374 * idle_task - return the idle task for a given cpu.
4375 * @cpu: the processor in question.
4377 struct task_struct *idle_task(int cpu)
4379 return cpu_rq(cpu)->idle;
4383 * find_process_by_pid - find a process with a matching PID value.
4384 * @pid: the pid in question.
4386 static struct task_struct *find_process_by_pid(pid_t pid)
4388 return pid ? find_task_by_vpid(pid) : current;
4391 /* Actually do priority change: must hold rq lock. */
4393 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4395 BUG_ON(p->se.on_rq);
4398 p->rt_priority = prio;
4399 p->normal_prio = normal_prio(p);
4400 /* we are holding p->pi_lock already */
4401 p->prio = rt_mutex_getprio(p);
4402 if (rt_prio(p->prio))
4403 p->sched_class = &rt_sched_class;
4405 p->sched_class = &fair_sched_class;
4410 * check the target process has a UID that matches the current process's
4412 static bool check_same_owner(struct task_struct *p)
4414 const struct cred *cred = current_cred(), *pcred;
4418 pcred = __task_cred(p);
4419 match = (cred->euid == pcred->euid ||
4420 cred->euid == pcred->uid);
4425 static int __sched_setscheduler(struct task_struct *p, int policy,
4426 struct sched_param *param, bool user)
4428 int retval, oldprio, oldpolicy = -1, on_rq, running;
4429 unsigned long flags;
4430 const struct sched_class *prev_class;
4434 /* may grab non-irq protected spin_locks */
4435 BUG_ON(in_interrupt());
4437 /* double check policy once rq lock held */
4439 reset_on_fork = p->sched_reset_on_fork;
4440 policy = oldpolicy = p->policy;
4442 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4443 policy &= ~SCHED_RESET_ON_FORK;
4445 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4446 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4447 policy != SCHED_IDLE)
4452 * Valid priorities for SCHED_FIFO and SCHED_RR are
4453 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4454 * SCHED_BATCH and SCHED_IDLE is 0.
4456 if (param->sched_priority < 0 ||
4457 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4458 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4460 if (rt_policy(policy) != (param->sched_priority != 0))
4464 * Allow unprivileged RT tasks to decrease priority:
4466 if (user && !capable(CAP_SYS_NICE)) {
4467 if (rt_policy(policy)) {
4468 unsigned long rlim_rtprio;
4470 if (!lock_task_sighand(p, &flags))
4472 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4473 unlock_task_sighand(p, &flags);
4475 /* can't set/change the rt policy */
4476 if (policy != p->policy && !rlim_rtprio)
4479 /* can't increase priority */
4480 if (param->sched_priority > p->rt_priority &&
4481 param->sched_priority > rlim_rtprio)
4485 * Like positive nice levels, dont allow tasks to
4486 * move out of SCHED_IDLE either:
4488 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4491 /* can't change other user's priorities */
4492 if (!check_same_owner(p))
4495 /* Normal users shall not reset the sched_reset_on_fork flag */
4496 if (p->sched_reset_on_fork && !reset_on_fork)
4501 #ifdef CONFIG_RT_GROUP_SCHED
4503 * Do not allow realtime tasks into groups that have no runtime
4506 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4507 task_group(p)->rt_bandwidth.rt_runtime == 0)
4511 retval = security_task_setscheduler(p, policy, param);
4517 * make sure no PI-waiters arrive (or leave) while we are
4518 * changing the priority of the task:
4520 raw_spin_lock_irqsave(&p->pi_lock, flags);
4522 * To be able to change p->policy safely, the apropriate
4523 * runqueue lock must be held.
4525 rq = __task_rq_lock(p);
4526 /* recheck policy now with rq lock held */
4527 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4528 policy = oldpolicy = -1;
4529 __task_rq_unlock(rq);
4530 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4533 on_rq = p->se.on_rq;
4534 running = task_current(rq, p);
4536 deactivate_task(rq, p, 0);
4538 p->sched_class->put_prev_task(rq, p);
4540 p->sched_reset_on_fork = reset_on_fork;
4543 prev_class = p->sched_class;
4544 __setscheduler(rq, p, policy, param->sched_priority);
4547 p->sched_class->set_curr_task(rq);
4549 activate_task(rq, p, 0);
4551 check_class_changed(rq, p, prev_class, oldprio, running);
4553 __task_rq_unlock(rq);
4554 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4556 rt_mutex_adjust_pi(p);
4562 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4563 * @p: the task in question.
4564 * @policy: new policy.
4565 * @param: structure containing the new RT priority.
4567 * NOTE that the task may be already dead.
4569 int sched_setscheduler(struct task_struct *p, int policy,
4570 struct sched_param *param)
4572 return __sched_setscheduler(p, policy, param, true);
4574 EXPORT_SYMBOL_GPL(sched_setscheduler);
4577 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4578 * @p: the task in question.
4579 * @policy: new policy.
4580 * @param: structure containing the new RT priority.
4582 * Just like sched_setscheduler, only don't bother checking if the
4583 * current context has permission. For example, this is needed in
4584 * stop_machine(): we create temporary high priority worker threads,
4585 * but our caller might not have that capability.
4587 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4588 struct sched_param *param)
4590 return __sched_setscheduler(p, policy, param, false);
4594 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4596 struct sched_param lparam;
4597 struct task_struct *p;
4600 if (!param || pid < 0)
4602 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4607 p = find_process_by_pid(pid);
4609 retval = sched_setscheduler(p, policy, &lparam);
4616 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4617 * @pid: the pid in question.
4618 * @policy: new policy.
4619 * @param: structure containing the new RT priority.
4621 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4622 struct sched_param __user *, param)
4624 /* negative values for policy are not valid */
4628 return do_sched_setscheduler(pid, policy, param);
4632 * sys_sched_setparam - set/change the RT priority of a thread
4633 * @pid: the pid in question.
4634 * @param: structure containing the new RT priority.
4636 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4638 return do_sched_setscheduler(pid, -1, param);
4642 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4643 * @pid: the pid in question.
4645 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4647 struct task_struct *p;
4655 p = find_process_by_pid(pid);
4657 retval = security_task_getscheduler(p);
4660 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4667 * sys_sched_getparam - get the RT priority of a thread
4668 * @pid: the pid in question.
4669 * @param: structure containing the RT priority.
4671 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4673 struct sched_param lp;
4674 struct task_struct *p;
4677 if (!param || pid < 0)
4681 p = find_process_by_pid(pid);
4686 retval = security_task_getscheduler(p);
4690 lp.sched_priority = p->rt_priority;
4694 * This one might sleep, we cannot do it with a spinlock held ...
4696 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4705 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4707 cpumask_var_t cpus_allowed, new_mask;
4708 struct task_struct *p;
4714 p = find_process_by_pid(pid);
4721 /* Prevent p going away */
4725 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4729 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4731 goto out_free_cpus_allowed;
4734 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4737 retval = security_task_setscheduler(p, 0, NULL);
4741 cpuset_cpus_allowed(p, cpus_allowed);
4742 cpumask_and(new_mask, in_mask, cpus_allowed);
4744 retval = set_cpus_allowed_ptr(p, new_mask);
4747 cpuset_cpus_allowed(p, cpus_allowed);
4748 if (!cpumask_subset(new_mask, cpus_allowed)) {
4750 * We must have raced with a concurrent cpuset
4751 * update. Just reset the cpus_allowed to the
4752 * cpuset's cpus_allowed
4754 cpumask_copy(new_mask, cpus_allowed);
4759 free_cpumask_var(new_mask);
4760 out_free_cpus_allowed:
4761 free_cpumask_var(cpus_allowed);
4768 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4769 struct cpumask *new_mask)
4771 if (len < cpumask_size())
4772 cpumask_clear(new_mask);
4773 else if (len > cpumask_size())
4774 len = cpumask_size();
4776 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4780 * sys_sched_setaffinity - set the cpu affinity of a process
4781 * @pid: pid of the process
4782 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4783 * @user_mask_ptr: user-space pointer to the new cpu mask
4785 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4786 unsigned long __user *, user_mask_ptr)
4788 cpumask_var_t new_mask;
4791 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4794 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4796 retval = sched_setaffinity(pid, new_mask);
4797 free_cpumask_var(new_mask);
4801 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4803 struct task_struct *p;
4804 unsigned long flags;
4812 p = find_process_by_pid(pid);
4816 retval = security_task_getscheduler(p);
4820 rq = task_rq_lock(p, &flags);
4821 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4822 task_rq_unlock(rq, &flags);
4832 * sys_sched_getaffinity - get the cpu affinity of a process
4833 * @pid: pid of the process
4834 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4835 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4837 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4838 unsigned long __user *, user_mask_ptr)
4843 if (len < cpumask_size())
4846 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4849 ret = sched_getaffinity(pid, mask);
4851 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
4854 ret = cpumask_size();
4856 free_cpumask_var(mask);
4862 * sys_sched_yield - yield the current processor to other threads.
4864 * This function yields the current CPU to other tasks. If there are no
4865 * other threads running on this CPU then this function will return.
4867 SYSCALL_DEFINE0(sched_yield)
4869 struct rq *rq = this_rq_lock();
4871 schedstat_inc(rq, yld_count);
4872 current->sched_class->yield_task(rq);
4875 * Since we are going to call schedule() anyway, there's
4876 * no need to preempt or enable interrupts:
4878 __release(rq->lock);
4879 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4880 do_raw_spin_unlock(&rq->lock);
4881 preempt_enable_no_resched();
4888 static inline int should_resched(void)
4890 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4893 static void __cond_resched(void)
4895 add_preempt_count(PREEMPT_ACTIVE);
4897 sub_preempt_count(PREEMPT_ACTIVE);
4900 int __sched _cond_resched(void)
4902 if (should_resched()) {
4908 EXPORT_SYMBOL(_cond_resched);
4911 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4912 * call schedule, and on return reacquire the lock.
4914 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4915 * operations here to prevent schedule() from being called twice (once via
4916 * spin_unlock(), once by hand).
4918 int __cond_resched_lock(spinlock_t *lock)
4920 int resched = should_resched();
4923 lockdep_assert_held(lock);
4925 if (spin_needbreak(lock) || resched) {
4936 EXPORT_SYMBOL(__cond_resched_lock);
4938 int __sched __cond_resched_softirq(void)
4940 BUG_ON(!in_softirq());
4942 if (should_resched()) {
4950 EXPORT_SYMBOL(__cond_resched_softirq);
4953 * yield - yield the current processor to other threads.
4955 * This is a shortcut for kernel-space yielding - it marks the
4956 * thread runnable and calls sys_sched_yield().
4958 void __sched yield(void)
4960 set_current_state(TASK_RUNNING);
4963 EXPORT_SYMBOL(yield);
4966 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4967 * that process accounting knows that this is a task in IO wait state.
4969 void __sched io_schedule(void)
4971 struct rq *rq = raw_rq();
4973 delayacct_blkio_start();
4974 atomic_inc(&rq->nr_iowait);
4975 current->in_iowait = 1;
4977 current->in_iowait = 0;
4978 atomic_dec(&rq->nr_iowait);
4979 delayacct_blkio_end();
4981 EXPORT_SYMBOL(io_schedule);
4983 long __sched io_schedule_timeout(long timeout)
4985 struct rq *rq = raw_rq();
4988 delayacct_blkio_start();
4989 atomic_inc(&rq->nr_iowait);
4990 current->in_iowait = 1;
4991 ret = schedule_timeout(timeout);
4992 current->in_iowait = 0;
4993 atomic_dec(&rq->nr_iowait);
4994 delayacct_blkio_end();
4999 * sys_sched_get_priority_max - return maximum RT priority.
5000 * @policy: scheduling class.
5002 * this syscall returns the maximum rt_priority that can be used
5003 * by a given scheduling class.
5005 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5012 ret = MAX_USER_RT_PRIO-1;
5024 * sys_sched_get_priority_min - return minimum RT priority.
5025 * @policy: scheduling class.
5027 * this syscall returns the minimum rt_priority that can be used
5028 * by a given scheduling class.
5030 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5048 * sys_sched_rr_get_interval - return the default timeslice of a process.
5049 * @pid: pid of the process.
5050 * @interval: userspace pointer to the timeslice value.
5052 * this syscall writes the default timeslice value of a given process
5053 * into the user-space timespec buffer. A value of '0' means infinity.
5055 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5056 struct timespec __user *, interval)
5058 struct task_struct *p;
5059 unsigned int time_slice;
5060 unsigned long flags;
5070 p = find_process_by_pid(pid);
5074 retval = security_task_getscheduler(p);
5078 rq = task_rq_lock(p, &flags);
5079 time_slice = p->sched_class->get_rr_interval(rq, p);
5080 task_rq_unlock(rq, &flags);
5083 jiffies_to_timespec(time_slice, &t);
5084 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5092 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5094 void sched_show_task(struct task_struct *p)
5096 unsigned long free = 0;
5099 state = p->state ? __ffs(p->state) + 1 : 0;
5100 printk(KERN_INFO "%-13.13s %c", p->comm,
5101 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5102 #if BITS_PER_LONG == 32
5103 if (state == TASK_RUNNING)
5104 printk(KERN_CONT " running ");
5106 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5108 if (state == TASK_RUNNING)
5109 printk(KERN_CONT " running task ");
5111 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5113 #ifdef CONFIG_DEBUG_STACK_USAGE
5114 free = stack_not_used(p);
5116 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5117 task_pid_nr(p), task_pid_nr(p->real_parent),
5118 (unsigned long)task_thread_info(p)->flags);
5120 show_stack(p, NULL);
5123 void show_state_filter(unsigned long state_filter)
5125 struct task_struct *g, *p;
5127 #if BITS_PER_LONG == 32
5129 " task PC stack pid father\n");
5132 " task PC stack pid father\n");
5134 read_lock(&tasklist_lock);
5135 do_each_thread(g, p) {
5137 * reset the NMI-timeout, listing all files on a slow
5138 * console might take alot of time:
5140 touch_nmi_watchdog();
5141 if (!state_filter || (p->state & state_filter))
5143 } while_each_thread(g, p);
5145 touch_all_softlockup_watchdogs();
5147 #ifdef CONFIG_SCHED_DEBUG
5148 sysrq_sched_debug_show();
5150 read_unlock(&tasklist_lock);
5152 * Only show locks if all tasks are dumped:
5155 debug_show_all_locks();
5158 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5160 idle->sched_class = &idle_sched_class;
5164 * init_idle - set up an idle thread for a given CPU
5165 * @idle: task in question
5166 * @cpu: cpu the idle task belongs to
5168 * NOTE: this function does not set the idle thread's NEED_RESCHED
5169 * flag, to make booting more robust.
5171 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5173 struct rq *rq = cpu_rq(cpu);
5174 unsigned long flags;
5176 raw_spin_lock_irqsave(&rq->lock, flags);
5179 idle->state = TASK_RUNNING;
5180 idle->se.exec_start = sched_clock();
5182 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5183 __set_task_cpu(idle, cpu);
5185 rq->curr = rq->idle = idle;
5186 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5189 raw_spin_unlock_irqrestore(&rq->lock, flags);
5191 /* Set the preempt count _outside_ the spinlocks! */
5192 #if defined(CONFIG_PREEMPT)
5193 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5195 task_thread_info(idle)->preempt_count = 0;
5198 * The idle tasks have their own, simple scheduling class:
5200 idle->sched_class = &idle_sched_class;
5201 ftrace_graph_init_task(idle);
5205 * In a system that switches off the HZ timer nohz_cpu_mask
5206 * indicates which cpus entered this state. This is used
5207 * in the rcu update to wait only for active cpus. For system
5208 * which do not switch off the HZ timer nohz_cpu_mask should
5209 * always be CPU_BITS_NONE.
5211 cpumask_var_t nohz_cpu_mask;
5214 * Increase the granularity value when there are more CPUs,
5215 * because with more CPUs the 'effective latency' as visible
5216 * to users decreases. But the relationship is not linear,
5217 * so pick a second-best guess by going with the log2 of the
5220 * This idea comes from the SD scheduler of Con Kolivas:
5222 static int get_update_sysctl_factor(void)
5224 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5225 unsigned int factor;
5227 switch (sysctl_sched_tunable_scaling) {
5228 case SCHED_TUNABLESCALING_NONE:
5231 case SCHED_TUNABLESCALING_LINEAR:
5234 case SCHED_TUNABLESCALING_LOG:
5236 factor = 1 + ilog2(cpus);
5243 static void update_sysctl(void)
5245 unsigned int factor = get_update_sysctl_factor();
5247 #define SET_SYSCTL(name) \
5248 (sysctl_##name = (factor) * normalized_sysctl_##name)
5249 SET_SYSCTL(sched_min_granularity);
5250 SET_SYSCTL(sched_latency);
5251 SET_SYSCTL(sched_wakeup_granularity);
5252 SET_SYSCTL(sched_shares_ratelimit);
5256 static inline void sched_init_granularity(void)
5263 * This is how migration works:
5265 * 1) we queue a struct migration_req structure in the source CPU's
5266 * runqueue and wake up that CPU's migration thread.
5267 * 2) we down() the locked semaphore => thread blocks.
5268 * 3) migration thread wakes up (implicitly it forces the migrated
5269 * thread off the CPU)
5270 * 4) it gets the migration request and checks whether the migrated
5271 * task is still in the wrong runqueue.
5272 * 5) if it's in the wrong runqueue then the migration thread removes
5273 * it and puts it into the right queue.
5274 * 6) migration thread up()s the semaphore.
5275 * 7) we wake up and the migration is done.
5279 * Change a given task's CPU affinity. Migrate the thread to a
5280 * proper CPU and schedule it away if the CPU it's executing on
5281 * is removed from the allowed bitmask.
5283 * NOTE: the caller must have a valid reference to the task, the
5284 * task must not exit() & deallocate itself prematurely. The
5285 * call is not atomic; no spinlocks may be held.
5287 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5289 struct migration_req req;
5290 unsigned long flags;
5294 rq = task_rq_lock(p, &flags);
5296 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5301 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5302 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5307 if (p->sched_class->set_cpus_allowed)
5308 p->sched_class->set_cpus_allowed(p, new_mask);
5310 cpumask_copy(&p->cpus_allowed, new_mask);
5311 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5314 /* Can the task run on the task's current CPU? If so, we're done */
5315 if (cpumask_test_cpu(task_cpu(p), new_mask))
5318 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5319 /* Need help from migration thread: drop lock and wait. */
5320 struct task_struct *mt = rq->migration_thread;
5322 get_task_struct(mt);
5323 task_rq_unlock(rq, &flags);
5324 wake_up_process(rq->migration_thread);
5325 put_task_struct(mt);
5326 wait_for_completion(&req.done);
5327 tlb_migrate_finish(p->mm);
5331 task_rq_unlock(rq, &flags);
5335 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5338 * Move (not current) task off this cpu, onto dest cpu. We're doing
5339 * this because either it can't run here any more (set_cpus_allowed()
5340 * away from this CPU, or CPU going down), or because we're
5341 * attempting to rebalance this task on exec (sched_exec).
5343 * So we race with normal scheduler movements, but that's OK, as long
5344 * as the task is no longer on this CPU.
5346 * Returns non-zero if task was successfully migrated.
5348 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5350 struct rq *rq_dest, *rq_src;
5353 if (unlikely(!cpu_active(dest_cpu)))
5356 rq_src = cpu_rq(src_cpu);
5357 rq_dest = cpu_rq(dest_cpu);
5359 double_rq_lock(rq_src, rq_dest);
5360 /* Already moved. */
5361 if (task_cpu(p) != src_cpu)
5363 /* Affinity changed (again). */
5364 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5368 * If we're not on a rq, the next wake-up will ensure we're
5372 deactivate_task(rq_src, p, 0);
5373 set_task_cpu(p, dest_cpu);
5374 activate_task(rq_dest, p, 0);
5375 check_preempt_curr(rq_dest, p, 0);
5380 double_rq_unlock(rq_src, rq_dest);
5384 #define RCU_MIGRATION_IDLE 0
5385 #define RCU_MIGRATION_NEED_QS 1
5386 #define RCU_MIGRATION_GOT_QS 2
5387 #define RCU_MIGRATION_MUST_SYNC 3
5390 * migration_thread - this is a highprio system thread that performs
5391 * thread migration by bumping thread off CPU then 'pushing' onto
5394 static int migration_thread(void *data)
5397 int cpu = (long)data;
5401 BUG_ON(rq->migration_thread != current);
5403 set_current_state(TASK_INTERRUPTIBLE);
5404 while (!kthread_should_stop()) {
5405 struct migration_req *req;
5406 struct list_head *head;
5408 raw_spin_lock_irq(&rq->lock);
5410 if (cpu_is_offline(cpu)) {
5411 raw_spin_unlock_irq(&rq->lock);
5415 if (rq->active_balance) {
5416 active_load_balance(rq, cpu);
5417 rq->active_balance = 0;
5420 head = &rq->migration_queue;
5422 if (list_empty(head)) {
5423 raw_spin_unlock_irq(&rq->lock);
5425 set_current_state(TASK_INTERRUPTIBLE);
5428 req = list_entry(head->next, struct migration_req, list);
5429 list_del_init(head->next);
5431 if (req->task != NULL) {
5432 raw_spin_unlock(&rq->lock);
5433 __migrate_task(req->task, cpu, req->dest_cpu);
5434 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5435 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5436 raw_spin_unlock(&rq->lock);
5438 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5439 raw_spin_unlock(&rq->lock);
5440 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5444 complete(&req->done);
5446 __set_current_state(TASK_RUNNING);
5451 #ifdef CONFIG_HOTPLUG_CPU
5453 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5457 local_irq_disable();
5458 ret = __migrate_task(p, src_cpu, dest_cpu);
5464 * Figure out where task on dead CPU should go, use force if necessary.
5466 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5471 dest_cpu = select_fallback_rq(dead_cpu, p);
5473 /* It can have affinity changed while we were choosing. */
5474 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
5479 * While a dead CPU has no uninterruptible tasks queued at this point,
5480 * it might still have a nonzero ->nr_uninterruptible counter, because
5481 * for performance reasons the counter is not stricly tracking tasks to
5482 * their home CPUs. So we just add the counter to another CPU's counter,
5483 * to keep the global sum constant after CPU-down:
5485 static void migrate_nr_uninterruptible(struct rq *rq_src)
5487 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5488 unsigned long flags;
5490 local_irq_save(flags);
5491 double_rq_lock(rq_src, rq_dest);
5492 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5493 rq_src->nr_uninterruptible = 0;
5494 double_rq_unlock(rq_src, rq_dest);
5495 local_irq_restore(flags);
5498 /* Run through task list and migrate tasks from the dead cpu. */
5499 static void migrate_live_tasks(int src_cpu)
5501 struct task_struct *p, *t;
5503 read_lock(&tasklist_lock);
5505 do_each_thread(t, p) {
5509 if (task_cpu(p) == src_cpu)
5510 move_task_off_dead_cpu(src_cpu, p);
5511 } while_each_thread(t, p);
5513 read_unlock(&tasklist_lock);
5517 * Schedules idle task to be the next runnable task on current CPU.
5518 * It does so by boosting its priority to highest possible.
5519 * Used by CPU offline code.
5521 void sched_idle_next(void)
5523 int this_cpu = smp_processor_id();
5524 struct rq *rq = cpu_rq(this_cpu);
5525 struct task_struct *p = rq->idle;
5526 unsigned long flags;
5528 /* cpu has to be offline */
5529 BUG_ON(cpu_online(this_cpu));
5532 * Strictly not necessary since rest of the CPUs are stopped by now
5533 * and interrupts disabled on the current cpu.
5535 raw_spin_lock_irqsave(&rq->lock, flags);
5537 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5539 activate_task(rq, p, 0);
5541 raw_spin_unlock_irqrestore(&rq->lock, flags);
5545 * Ensures that the idle task is using init_mm right before its cpu goes
5548 void idle_task_exit(void)
5550 struct mm_struct *mm = current->active_mm;
5552 BUG_ON(cpu_online(smp_processor_id()));
5555 switch_mm(mm, &init_mm, current);
5559 /* called under rq->lock with disabled interrupts */
5560 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5562 struct rq *rq = cpu_rq(dead_cpu);
5564 /* Must be exiting, otherwise would be on tasklist. */
5565 BUG_ON(!p->exit_state);
5567 /* Cannot have done final schedule yet: would have vanished. */
5568 BUG_ON(p->state == TASK_DEAD);
5573 * Drop lock around migration; if someone else moves it,
5574 * that's OK. No task can be added to this CPU, so iteration is
5577 raw_spin_unlock_irq(&rq->lock);
5578 move_task_off_dead_cpu(dead_cpu, p);
5579 raw_spin_lock_irq(&rq->lock);
5584 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5585 static void migrate_dead_tasks(unsigned int dead_cpu)
5587 struct rq *rq = cpu_rq(dead_cpu);
5588 struct task_struct *next;
5591 if (!rq->nr_running)
5593 next = pick_next_task(rq);
5596 next->sched_class->put_prev_task(rq, next);
5597 migrate_dead(dead_cpu, next);
5603 * remove the tasks which were accounted by rq from calc_load_tasks.
5605 static void calc_global_load_remove(struct rq *rq)
5607 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5608 rq->calc_load_active = 0;
5610 #endif /* CONFIG_HOTPLUG_CPU */
5612 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5614 static struct ctl_table sd_ctl_dir[] = {
5616 .procname = "sched_domain",
5622 static struct ctl_table sd_ctl_root[] = {
5624 .procname = "kernel",
5626 .child = sd_ctl_dir,
5631 static struct ctl_table *sd_alloc_ctl_entry(int n)
5633 struct ctl_table *entry =
5634 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5639 static void sd_free_ctl_entry(struct ctl_table **tablep)
5641 struct ctl_table *entry;
5644 * In the intermediate directories, both the child directory and
5645 * procname are dynamically allocated and could fail but the mode
5646 * will always be set. In the lowest directory the names are
5647 * static strings and all have proc handlers.
5649 for (entry = *tablep; entry->mode; entry++) {
5651 sd_free_ctl_entry(&entry->child);
5652 if (entry->proc_handler == NULL)
5653 kfree(entry->procname);
5661 set_table_entry(struct ctl_table *entry,
5662 const char *procname, void *data, int maxlen,
5663 mode_t mode, proc_handler *proc_handler)
5665 entry->procname = procname;
5667 entry->maxlen = maxlen;
5669 entry->proc_handler = proc_handler;
5672 static struct ctl_table *
5673 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5675 struct ctl_table *table = sd_alloc_ctl_entry(13);
5680 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5681 sizeof(long), 0644, proc_doulongvec_minmax);
5682 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5683 sizeof(long), 0644, proc_doulongvec_minmax);
5684 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5685 sizeof(int), 0644, proc_dointvec_minmax);
5686 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5687 sizeof(int), 0644, proc_dointvec_minmax);
5688 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5689 sizeof(int), 0644, proc_dointvec_minmax);
5690 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5691 sizeof(int), 0644, proc_dointvec_minmax);
5692 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5693 sizeof(int), 0644, proc_dointvec_minmax);
5694 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5695 sizeof(int), 0644, proc_dointvec_minmax);
5696 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5697 sizeof(int), 0644, proc_dointvec_minmax);
5698 set_table_entry(&table[9], "cache_nice_tries",
5699 &sd->cache_nice_tries,
5700 sizeof(int), 0644, proc_dointvec_minmax);
5701 set_table_entry(&table[10], "flags", &sd->flags,
5702 sizeof(int), 0644, proc_dointvec_minmax);
5703 set_table_entry(&table[11], "name", sd->name,
5704 CORENAME_MAX_SIZE, 0444, proc_dostring);
5705 /* &table[12] is terminator */
5710 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5712 struct ctl_table *entry, *table;
5713 struct sched_domain *sd;
5714 int domain_num = 0, i;
5717 for_each_domain(cpu, sd)
5719 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5724 for_each_domain(cpu, sd) {
5725 snprintf(buf, 32, "domain%d", i);
5726 entry->procname = kstrdup(buf, GFP_KERNEL);
5728 entry->child = sd_alloc_ctl_domain_table(sd);
5735 static struct ctl_table_header *sd_sysctl_header;
5736 static void register_sched_domain_sysctl(void)
5738 int i, cpu_num = num_possible_cpus();
5739 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5742 WARN_ON(sd_ctl_dir[0].child);
5743 sd_ctl_dir[0].child = entry;
5748 for_each_possible_cpu(i) {
5749 snprintf(buf, 32, "cpu%d", i);
5750 entry->procname = kstrdup(buf, GFP_KERNEL);
5752 entry->child = sd_alloc_ctl_cpu_table(i);
5756 WARN_ON(sd_sysctl_header);
5757 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5760 /* may be called multiple times per register */
5761 static void unregister_sched_domain_sysctl(void)
5763 if (sd_sysctl_header)
5764 unregister_sysctl_table(sd_sysctl_header);
5765 sd_sysctl_header = NULL;
5766 if (sd_ctl_dir[0].child)
5767 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5770 static void register_sched_domain_sysctl(void)
5773 static void unregister_sched_domain_sysctl(void)
5778 static void set_rq_online(struct rq *rq)
5781 const struct sched_class *class;
5783 cpumask_set_cpu(rq->cpu, rq->rd->online);
5786 for_each_class(class) {
5787 if (class->rq_online)
5788 class->rq_online(rq);
5793 static void set_rq_offline(struct rq *rq)
5796 const struct sched_class *class;
5798 for_each_class(class) {
5799 if (class->rq_offline)
5800 class->rq_offline(rq);
5803 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5809 * migration_call - callback that gets triggered when a CPU is added.
5810 * Here we can start up the necessary migration thread for the new CPU.
5812 static int __cpuinit
5813 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5815 struct task_struct *p;
5816 int cpu = (long)hcpu;
5817 unsigned long flags;
5822 case CPU_UP_PREPARE:
5823 case CPU_UP_PREPARE_FROZEN:
5824 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5827 kthread_bind(p, cpu);
5828 /* Must be high prio: stop_machine expects to yield to it. */
5829 rq = task_rq_lock(p, &flags);
5830 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5831 task_rq_unlock(rq, &flags);
5833 cpu_rq(cpu)->migration_thread = p;
5834 rq->calc_load_update = calc_load_update;
5838 case CPU_ONLINE_FROZEN:
5839 /* Strictly unnecessary, as first user will wake it. */
5840 wake_up_process(cpu_rq(cpu)->migration_thread);
5842 /* Update our root-domain */
5844 raw_spin_lock_irqsave(&rq->lock, flags);
5846 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5850 raw_spin_unlock_irqrestore(&rq->lock, flags);
5853 #ifdef CONFIG_HOTPLUG_CPU
5854 case CPU_UP_CANCELED:
5855 case CPU_UP_CANCELED_FROZEN:
5856 if (!cpu_rq(cpu)->migration_thread)
5858 /* Unbind it from offline cpu so it can run. Fall thru. */
5859 kthread_bind(cpu_rq(cpu)->migration_thread,
5860 cpumask_any(cpu_online_mask));
5861 kthread_stop(cpu_rq(cpu)->migration_thread);
5862 put_task_struct(cpu_rq(cpu)->migration_thread);
5863 cpu_rq(cpu)->migration_thread = NULL;
5867 case CPU_DEAD_FROZEN:
5868 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5869 migrate_live_tasks(cpu);
5871 kthread_stop(rq->migration_thread);
5872 put_task_struct(rq->migration_thread);
5873 rq->migration_thread = NULL;
5874 /* Idle task back to normal (off runqueue, low prio) */
5875 raw_spin_lock_irq(&rq->lock);
5876 deactivate_task(rq, rq->idle, 0);
5877 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5878 rq->idle->sched_class = &idle_sched_class;
5879 migrate_dead_tasks(cpu);
5880 raw_spin_unlock_irq(&rq->lock);
5882 migrate_nr_uninterruptible(rq);
5883 BUG_ON(rq->nr_running != 0);
5884 calc_global_load_remove(rq);
5886 * No need to migrate the tasks: it was best-effort if
5887 * they didn't take sched_hotcpu_mutex. Just wake up
5890 raw_spin_lock_irq(&rq->lock);
5891 while (!list_empty(&rq->migration_queue)) {
5892 struct migration_req *req;
5894 req = list_entry(rq->migration_queue.next,
5895 struct migration_req, list);
5896 list_del_init(&req->list);
5897 raw_spin_unlock_irq(&rq->lock);
5898 complete(&req->done);
5899 raw_spin_lock_irq(&rq->lock);
5901 raw_spin_unlock_irq(&rq->lock);
5905 case CPU_DYING_FROZEN:
5906 /* Update our root-domain */
5908 raw_spin_lock_irqsave(&rq->lock, flags);
5910 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5913 raw_spin_unlock_irqrestore(&rq->lock, flags);
5921 * Register at high priority so that task migration (migrate_all_tasks)
5922 * happens before everything else. This has to be lower priority than
5923 * the notifier in the perf_event subsystem, though.
5925 static struct notifier_block __cpuinitdata migration_notifier = {
5926 .notifier_call = migration_call,
5930 static int __init migration_init(void)
5932 void *cpu = (void *)(long)smp_processor_id();
5935 /* Start one for the boot CPU: */
5936 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5937 BUG_ON(err == NOTIFY_BAD);
5938 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5939 register_cpu_notifier(&migration_notifier);
5943 early_initcall(migration_init);
5948 #ifdef CONFIG_SCHED_DEBUG
5950 static __read_mostly int sched_domain_debug_enabled;
5952 static int __init sched_domain_debug_setup(char *str)
5954 sched_domain_debug_enabled = 1;
5958 early_param("sched_debug", sched_domain_debug_setup);
5960 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5961 struct cpumask *groupmask)
5963 struct sched_group *group = sd->groups;
5966 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5967 cpumask_clear(groupmask);
5969 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5971 if (!(sd->flags & SD_LOAD_BALANCE)) {
5972 printk("does not load-balance\n");
5974 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5979 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5981 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5982 printk(KERN_ERR "ERROR: domain->span does not contain "
5985 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5986 printk(KERN_ERR "ERROR: domain->groups does not contain"
5990 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5994 printk(KERN_ERR "ERROR: group is NULL\n");
5998 if (!group->cpu_power) {
5999 printk(KERN_CONT "\n");
6000 printk(KERN_ERR "ERROR: domain->cpu_power not "
6005 if (!cpumask_weight(sched_group_cpus(group))) {
6006 printk(KERN_CONT "\n");
6007 printk(KERN_ERR "ERROR: empty group\n");
6011 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6012 printk(KERN_CONT "\n");
6013 printk(KERN_ERR "ERROR: repeated CPUs\n");
6017 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6019 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6021 printk(KERN_CONT " %s", str);
6022 if (group->cpu_power != SCHED_LOAD_SCALE) {
6023 printk(KERN_CONT " (cpu_power = %d)",
6027 group = group->next;
6028 } while (group != sd->groups);
6029 printk(KERN_CONT "\n");
6031 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6032 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6035 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6036 printk(KERN_ERR "ERROR: parent span is not a superset "
6037 "of domain->span\n");
6041 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6043 cpumask_var_t groupmask;
6046 if (!sched_domain_debug_enabled)
6050 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6054 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6056 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6057 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6062 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6069 free_cpumask_var(groupmask);
6071 #else /* !CONFIG_SCHED_DEBUG */
6072 # define sched_domain_debug(sd, cpu) do { } while (0)
6073 #endif /* CONFIG_SCHED_DEBUG */
6075 static int sd_degenerate(struct sched_domain *sd)
6077 if (cpumask_weight(sched_domain_span(sd)) == 1)
6080 /* Following flags need at least 2 groups */
6081 if (sd->flags & (SD_LOAD_BALANCE |
6082 SD_BALANCE_NEWIDLE |
6086 SD_SHARE_PKG_RESOURCES)) {
6087 if (sd->groups != sd->groups->next)
6091 /* Following flags don't use groups */
6092 if (sd->flags & (SD_WAKE_AFFINE))
6099 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6101 unsigned long cflags = sd->flags, pflags = parent->flags;
6103 if (sd_degenerate(parent))
6106 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6109 /* Flags needing groups don't count if only 1 group in parent */
6110 if (parent->groups == parent->groups->next) {
6111 pflags &= ~(SD_LOAD_BALANCE |
6112 SD_BALANCE_NEWIDLE |
6116 SD_SHARE_PKG_RESOURCES);
6117 if (nr_node_ids == 1)
6118 pflags &= ~SD_SERIALIZE;
6120 if (~cflags & pflags)
6126 static void free_rootdomain(struct root_domain *rd)
6128 synchronize_sched();
6130 cpupri_cleanup(&rd->cpupri);
6132 free_cpumask_var(rd->rto_mask);
6133 free_cpumask_var(rd->online);
6134 free_cpumask_var(rd->span);
6138 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6140 struct root_domain *old_rd = NULL;
6141 unsigned long flags;
6143 raw_spin_lock_irqsave(&rq->lock, flags);
6148 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6151 cpumask_clear_cpu(rq->cpu, old_rd->span);
6154 * If we dont want to free the old_rt yet then
6155 * set old_rd to NULL to skip the freeing later
6158 if (!atomic_dec_and_test(&old_rd->refcount))
6162 atomic_inc(&rd->refcount);
6165 cpumask_set_cpu(rq->cpu, rd->span);
6166 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6169 raw_spin_unlock_irqrestore(&rq->lock, flags);
6172 free_rootdomain(old_rd);
6175 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6177 gfp_t gfp = GFP_KERNEL;
6179 memset(rd, 0, sizeof(*rd));
6184 if (!alloc_cpumask_var(&rd->span, gfp))
6186 if (!alloc_cpumask_var(&rd->online, gfp))
6188 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6191 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6196 free_cpumask_var(rd->rto_mask);
6198 free_cpumask_var(rd->online);
6200 free_cpumask_var(rd->span);
6205 static void init_defrootdomain(void)
6207 init_rootdomain(&def_root_domain, true);
6209 atomic_set(&def_root_domain.refcount, 1);
6212 static struct root_domain *alloc_rootdomain(void)
6214 struct root_domain *rd;
6216 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6220 if (init_rootdomain(rd, false) != 0) {
6229 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6230 * hold the hotplug lock.
6233 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6235 struct rq *rq = cpu_rq(cpu);
6236 struct sched_domain *tmp;
6238 /* Remove the sched domains which do not contribute to scheduling. */
6239 for (tmp = sd; tmp; ) {
6240 struct sched_domain *parent = tmp->parent;
6244 if (sd_parent_degenerate(tmp, parent)) {
6245 tmp->parent = parent->parent;
6247 parent->parent->child = tmp;
6252 if (sd && sd_degenerate(sd)) {
6258 sched_domain_debug(sd, cpu);
6260 rq_attach_root(rq, rd);
6261 rcu_assign_pointer(rq->sd, sd);
6264 /* cpus with isolated domains */
6265 static cpumask_var_t cpu_isolated_map;
6267 /* Setup the mask of cpus configured for isolated domains */
6268 static int __init isolated_cpu_setup(char *str)
6270 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6271 cpulist_parse(str, cpu_isolated_map);
6275 __setup("isolcpus=", isolated_cpu_setup);
6278 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6279 * to a function which identifies what group(along with sched group) a CPU
6280 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6281 * (due to the fact that we keep track of groups covered with a struct cpumask).
6283 * init_sched_build_groups will build a circular linked list of the groups
6284 * covered by the given span, and will set each group's ->cpumask correctly,
6285 * and ->cpu_power to 0.
6288 init_sched_build_groups(const struct cpumask *span,
6289 const struct cpumask *cpu_map,
6290 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6291 struct sched_group **sg,
6292 struct cpumask *tmpmask),
6293 struct cpumask *covered, struct cpumask *tmpmask)
6295 struct sched_group *first = NULL, *last = NULL;
6298 cpumask_clear(covered);
6300 for_each_cpu(i, span) {
6301 struct sched_group *sg;
6302 int group = group_fn(i, cpu_map, &sg, tmpmask);
6305 if (cpumask_test_cpu(i, covered))
6308 cpumask_clear(sched_group_cpus(sg));
6311 for_each_cpu(j, span) {
6312 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6315 cpumask_set_cpu(j, covered);
6316 cpumask_set_cpu(j, sched_group_cpus(sg));
6327 #define SD_NODES_PER_DOMAIN 16
6332 * find_next_best_node - find the next node to include in a sched_domain
6333 * @node: node whose sched_domain we're building
6334 * @used_nodes: nodes already in the sched_domain
6336 * Find the next node to include in a given scheduling domain. Simply
6337 * finds the closest node not already in the @used_nodes map.
6339 * Should use nodemask_t.
6341 static int find_next_best_node(int node, nodemask_t *used_nodes)
6343 int i, n, val, min_val, best_node = 0;
6347 for (i = 0; i < nr_node_ids; i++) {
6348 /* Start at @node */
6349 n = (node + i) % nr_node_ids;
6351 if (!nr_cpus_node(n))
6354 /* Skip already used nodes */
6355 if (node_isset(n, *used_nodes))
6358 /* Simple min distance search */
6359 val = node_distance(node, n);
6361 if (val < min_val) {
6367 node_set(best_node, *used_nodes);
6372 * sched_domain_node_span - get a cpumask for a node's sched_domain
6373 * @node: node whose cpumask we're constructing
6374 * @span: resulting cpumask
6376 * Given a node, construct a good cpumask for its sched_domain to span. It
6377 * should be one that prevents unnecessary balancing, but also spreads tasks
6380 static void sched_domain_node_span(int node, struct cpumask *span)
6382 nodemask_t used_nodes;
6385 cpumask_clear(span);
6386 nodes_clear(used_nodes);
6388 cpumask_or(span, span, cpumask_of_node(node));
6389 node_set(node, used_nodes);
6391 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6392 int next_node = find_next_best_node(node, &used_nodes);
6394 cpumask_or(span, span, cpumask_of_node(next_node));
6397 #endif /* CONFIG_NUMA */
6399 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6402 * The cpus mask in sched_group and sched_domain hangs off the end.
6404 * ( See the the comments in include/linux/sched.h:struct sched_group
6405 * and struct sched_domain. )
6407 struct static_sched_group {
6408 struct sched_group sg;
6409 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6412 struct static_sched_domain {
6413 struct sched_domain sd;
6414 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6420 cpumask_var_t domainspan;
6421 cpumask_var_t covered;
6422 cpumask_var_t notcovered;
6424 cpumask_var_t nodemask;
6425 cpumask_var_t this_sibling_map;
6426 cpumask_var_t this_core_map;
6427 cpumask_var_t send_covered;
6428 cpumask_var_t tmpmask;
6429 struct sched_group **sched_group_nodes;
6430 struct root_domain *rd;
6434 sa_sched_groups = 0,
6439 sa_this_sibling_map,
6441 sa_sched_group_nodes,
6451 * SMT sched-domains:
6453 #ifdef CONFIG_SCHED_SMT
6454 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6455 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6458 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6459 struct sched_group **sg, struct cpumask *unused)
6462 *sg = &per_cpu(sched_groups, cpu).sg;
6465 #endif /* CONFIG_SCHED_SMT */
6468 * multi-core sched-domains:
6470 #ifdef CONFIG_SCHED_MC
6471 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6472 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6473 #endif /* CONFIG_SCHED_MC */
6475 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6477 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6478 struct sched_group **sg, struct cpumask *mask)
6482 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6483 group = cpumask_first(mask);
6485 *sg = &per_cpu(sched_group_core, group).sg;
6488 #elif defined(CONFIG_SCHED_MC)
6490 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6491 struct sched_group **sg, struct cpumask *unused)
6494 *sg = &per_cpu(sched_group_core, cpu).sg;
6499 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6500 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6503 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6504 struct sched_group **sg, struct cpumask *mask)
6507 #ifdef CONFIG_SCHED_MC
6508 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6509 group = cpumask_first(mask);
6510 #elif defined(CONFIG_SCHED_SMT)
6511 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6512 group = cpumask_first(mask);
6517 *sg = &per_cpu(sched_group_phys, group).sg;
6523 * The init_sched_build_groups can't handle what we want to do with node
6524 * groups, so roll our own. Now each node has its own list of groups which
6525 * gets dynamically allocated.
6527 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6528 static struct sched_group ***sched_group_nodes_bycpu;
6530 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6531 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6533 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6534 struct sched_group **sg,
6535 struct cpumask *nodemask)
6539 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6540 group = cpumask_first(nodemask);
6543 *sg = &per_cpu(sched_group_allnodes, group).sg;
6547 static void init_numa_sched_groups_power(struct sched_group *group_head)
6549 struct sched_group *sg = group_head;
6555 for_each_cpu(j, sched_group_cpus(sg)) {
6556 struct sched_domain *sd;
6558 sd = &per_cpu(phys_domains, j).sd;
6559 if (j != group_first_cpu(sd->groups)) {
6561 * Only add "power" once for each
6567 sg->cpu_power += sd->groups->cpu_power;
6570 } while (sg != group_head);
6573 static int build_numa_sched_groups(struct s_data *d,
6574 const struct cpumask *cpu_map, int num)
6576 struct sched_domain *sd;
6577 struct sched_group *sg, *prev;
6580 cpumask_clear(d->covered);
6581 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6582 if (cpumask_empty(d->nodemask)) {
6583 d->sched_group_nodes[num] = NULL;
6587 sched_domain_node_span(num, d->domainspan);
6588 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6590 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6593 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6597 d->sched_group_nodes[num] = sg;
6599 for_each_cpu(j, d->nodemask) {
6600 sd = &per_cpu(node_domains, j).sd;
6605 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6607 cpumask_or(d->covered, d->covered, d->nodemask);
6610 for (j = 0; j < nr_node_ids; j++) {
6611 n = (num + j) % nr_node_ids;
6612 cpumask_complement(d->notcovered, d->covered);
6613 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6614 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6615 if (cpumask_empty(d->tmpmask))
6617 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6618 if (cpumask_empty(d->tmpmask))
6620 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6624 "Can not alloc domain group for node %d\n", j);
6628 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6629 sg->next = prev->next;
6630 cpumask_or(d->covered, d->covered, d->tmpmask);
6637 #endif /* CONFIG_NUMA */
6640 /* Free memory allocated for various sched_group structures */
6641 static void free_sched_groups(const struct cpumask *cpu_map,
6642 struct cpumask *nodemask)
6646 for_each_cpu(cpu, cpu_map) {
6647 struct sched_group **sched_group_nodes
6648 = sched_group_nodes_bycpu[cpu];
6650 if (!sched_group_nodes)
6653 for (i = 0; i < nr_node_ids; i++) {
6654 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6656 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6657 if (cpumask_empty(nodemask))
6667 if (oldsg != sched_group_nodes[i])
6670 kfree(sched_group_nodes);
6671 sched_group_nodes_bycpu[cpu] = NULL;
6674 #else /* !CONFIG_NUMA */
6675 static void free_sched_groups(const struct cpumask *cpu_map,
6676 struct cpumask *nodemask)
6679 #endif /* CONFIG_NUMA */
6682 * Initialize sched groups cpu_power.
6684 * cpu_power indicates the capacity of sched group, which is used while
6685 * distributing the load between different sched groups in a sched domain.
6686 * Typically cpu_power for all the groups in a sched domain will be same unless
6687 * there are asymmetries in the topology. If there are asymmetries, group
6688 * having more cpu_power will pickup more load compared to the group having
6691 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6693 struct sched_domain *child;
6694 struct sched_group *group;
6698 WARN_ON(!sd || !sd->groups);
6700 if (cpu != group_first_cpu(sd->groups))
6705 sd->groups->cpu_power = 0;
6708 power = SCHED_LOAD_SCALE;
6709 weight = cpumask_weight(sched_domain_span(sd));
6711 * SMT siblings share the power of a single core.
6712 * Usually multiple threads get a better yield out of
6713 * that one core than a single thread would have,
6714 * reflect that in sd->smt_gain.
6716 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6717 power *= sd->smt_gain;
6719 power >>= SCHED_LOAD_SHIFT;
6721 sd->groups->cpu_power += power;
6726 * Add cpu_power of each child group to this groups cpu_power.
6728 group = child->groups;
6730 sd->groups->cpu_power += group->cpu_power;
6731 group = group->next;
6732 } while (group != child->groups);
6736 * Initializers for schedule domains
6737 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6740 #ifdef CONFIG_SCHED_DEBUG
6741 # define SD_INIT_NAME(sd, type) sd->name = #type
6743 # define SD_INIT_NAME(sd, type) do { } while (0)
6746 #define SD_INIT(sd, type) sd_init_##type(sd)
6748 #define SD_INIT_FUNC(type) \
6749 static noinline void sd_init_##type(struct sched_domain *sd) \
6751 memset(sd, 0, sizeof(*sd)); \
6752 *sd = SD_##type##_INIT; \
6753 sd->level = SD_LV_##type; \
6754 SD_INIT_NAME(sd, type); \
6759 SD_INIT_FUNC(ALLNODES)
6762 #ifdef CONFIG_SCHED_SMT
6763 SD_INIT_FUNC(SIBLING)
6765 #ifdef CONFIG_SCHED_MC
6769 static int default_relax_domain_level = -1;
6771 static int __init setup_relax_domain_level(char *str)
6775 val = simple_strtoul(str, NULL, 0);
6776 if (val < SD_LV_MAX)
6777 default_relax_domain_level = val;
6781 __setup("relax_domain_level=", setup_relax_domain_level);
6783 static void set_domain_attribute(struct sched_domain *sd,
6784 struct sched_domain_attr *attr)
6788 if (!attr || attr->relax_domain_level < 0) {
6789 if (default_relax_domain_level < 0)
6792 request = default_relax_domain_level;
6794 request = attr->relax_domain_level;
6795 if (request < sd->level) {
6796 /* turn off idle balance on this domain */
6797 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6799 /* turn on idle balance on this domain */
6800 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6804 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6805 const struct cpumask *cpu_map)
6808 case sa_sched_groups:
6809 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6810 d->sched_group_nodes = NULL;
6812 free_rootdomain(d->rd); /* fall through */
6814 free_cpumask_var(d->tmpmask); /* fall through */
6815 case sa_send_covered:
6816 free_cpumask_var(d->send_covered); /* fall through */
6817 case sa_this_core_map:
6818 free_cpumask_var(d->this_core_map); /* fall through */
6819 case sa_this_sibling_map:
6820 free_cpumask_var(d->this_sibling_map); /* fall through */
6822 free_cpumask_var(d->nodemask); /* fall through */
6823 case sa_sched_group_nodes:
6825 kfree(d->sched_group_nodes); /* fall through */
6827 free_cpumask_var(d->notcovered); /* fall through */
6829 free_cpumask_var(d->covered); /* fall through */
6831 free_cpumask_var(d->domainspan); /* fall through */
6838 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6839 const struct cpumask *cpu_map)
6842 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6844 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6845 return sa_domainspan;
6846 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6848 /* Allocate the per-node list of sched groups */
6849 d->sched_group_nodes = kcalloc(nr_node_ids,
6850 sizeof(struct sched_group *), GFP_KERNEL);
6851 if (!d->sched_group_nodes) {
6852 printk(KERN_WARNING "Can not alloc sched group node list\n");
6853 return sa_notcovered;
6855 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6857 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6858 return sa_sched_group_nodes;
6859 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6861 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6862 return sa_this_sibling_map;
6863 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6864 return sa_this_core_map;
6865 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6866 return sa_send_covered;
6867 d->rd = alloc_rootdomain();
6869 printk(KERN_WARNING "Cannot alloc root domain\n");
6872 return sa_rootdomain;
6875 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6876 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6878 struct sched_domain *sd = NULL;
6880 struct sched_domain *parent;
6883 if (cpumask_weight(cpu_map) >
6884 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6885 sd = &per_cpu(allnodes_domains, i).sd;
6886 SD_INIT(sd, ALLNODES);
6887 set_domain_attribute(sd, attr);
6888 cpumask_copy(sched_domain_span(sd), cpu_map);
6889 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6894 sd = &per_cpu(node_domains, i).sd;
6896 set_domain_attribute(sd, attr);
6897 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6898 sd->parent = parent;
6901 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6906 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6907 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6908 struct sched_domain *parent, int i)
6910 struct sched_domain *sd;
6911 sd = &per_cpu(phys_domains, i).sd;
6913 set_domain_attribute(sd, attr);
6914 cpumask_copy(sched_domain_span(sd), d->nodemask);
6915 sd->parent = parent;
6918 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6922 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6923 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6924 struct sched_domain *parent, int i)
6926 struct sched_domain *sd = parent;
6927 #ifdef CONFIG_SCHED_MC
6928 sd = &per_cpu(core_domains, i).sd;
6930 set_domain_attribute(sd, attr);
6931 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6932 sd->parent = parent;
6934 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6939 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6940 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6941 struct sched_domain *parent, int i)
6943 struct sched_domain *sd = parent;
6944 #ifdef CONFIG_SCHED_SMT
6945 sd = &per_cpu(cpu_domains, i).sd;
6946 SD_INIT(sd, SIBLING);
6947 set_domain_attribute(sd, attr);
6948 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6949 sd->parent = parent;
6951 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6956 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6957 const struct cpumask *cpu_map, int cpu)
6960 #ifdef CONFIG_SCHED_SMT
6961 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6962 cpumask_and(d->this_sibling_map, cpu_map,
6963 topology_thread_cpumask(cpu));
6964 if (cpu == cpumask_first(d->this_sibling_map))
6965 init_sched_build_groups(d->this_sibling_map, cpu_map,
6967 d->send_covered, d->tmpmask);
6970 #ifdef CONFIG_SCHED_MC
6971 case SD_LV_MC: /* set up multi-core groups */
6972 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6973 if (cpu == cpumask_first(d->this_core_map))
6974 init_sched_build_groups(d->this_core_map, cpu_map,
6976 d->send_covered, d->tmpmask);
6979 case SD_LV_CPU: /* set up physical groups */
6980 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
6981 if (!cpumask_empty(d->nodemask))
6982 init_sched_build_groups(d->nodemask, cpu_map,
6984 d->send_covered, d->tmpmask);
6987 case SD_LV_ALLNODES:
6988 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
6989 d->send_covered, d->tmpmask);
6998 * Build sched domains for a given set of cpus and attach the sched domains
6999 * to the individual cpus
7001 static int __build_sched_domains(const struct cpumask *cpu_map,
7002 struct sched_domain_attr *attr)
7004 enum s_alloc alloc_state = sa_none;
7006 struct sched_domain *sd;
7012 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7013 if (alloc_state != sa_rootdomain)
7015 alloc_state = sa_sched_groups;
7018 * Set up domains for cpus specified by the cpu_map.
7020 for_each_cpu(i, cpu_map) {
7021 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7024 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7025 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7026 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7027 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7030 for_each_cpu(i, cpu_map) {
7031 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7032 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7035 /* Set up physical groups */
7036 for (i = 0; i < nr_node_ids; i++)
7037 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7040 /* Set up node groups */
7042 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7044 for (i = 0; i < nr_node_ids; i++)
7045 if (build_numa_sched_groups(&d, cpu_map, i))
7049 /* Calculate CPU power for physical packages and nodes */
7050 #ifdef CONFIG_SCHED_SMT
7051 for_each_cpu(i, cpu_map) {
7052 sd = &per_cpu(cpu_domains, i).sd;
7053 init_sched_groups_power(i, sd);
7056 #ifdef CONFIG_SCHED_MC
7057 for_each_cpu(i, cpu_map) {
7058 sd = &per_cpu(core_domains, i).sd;
7059 init_sched_groups_power(i, sd);
7063 for_each_cpu(i, cpu_map) {
7064 sd = &per_cpu(phys_domains, i).sd;
7065 init_sched_groups_power(i, sd);
7069 for (i = 0; i < nr_node_ids; i++)
7070 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7072 if (d.sd_allnodes) {
7073 struct sched_group *sg;
7075 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7077 init_numa_sched_groups_power(sg);
7081 /* Attach the domains */
7082 for_each_cpu(i, cpu_map) {
7083 #ifdef CONFIG_SCHED_SMT
7084 sd = &per_cpu(cpu_domains, i).sd;
7085 #elif defined(CONFIG_SCHED_MC)
7086 sd = &per_cpu(core_domains, i).sd;
7088 sd = &per_cpu(phys_domains, i).sd;
7090 cpu_attach_domain(sd, d.rd, i);
7093 d.sched_group_nodes = NULL; /* don't free this we still need it */
7094 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7098 __free_domain_allocs(&d, alloc_state, cpu_map);
7102 static int build_sched_domains(const struct cpumask *cpu_map)
7104 return __build_sched_domains(cpu_map, NULL);
7107 static cpumask_var_t *doms_cur; /* current sched domains */
7108 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7109 static struct sched_domain_attr *dattr_cur;
7110 /* attribues of custom domains in 'doms_cur' */
7113 * Special case: If a kmalloc of a doms_cur partition (array of
7114 * cpumask) fails, then fallback to a single sched domain,
7115 * as determined by the single cpumask fallback_doms.
7117 static cpumask_var_t fallback_doms;
7120 * arch_update_cpu_topology lets virtualized architectures update the
7121 * cpu core maps. It is supposed to return 1 if the topology changed
7122 * or 0 if it stayed the same.
7124 int __attribute__((weak)) arch_update_cpu_topology(void)
7129 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7132 cpumask_var_t *doms;
7134 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7137 for (i = 0; i < ndoms; i++) {
7138 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7139 free_sched_domains(doms, i);
7146 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7149 for (i = 0; i < ndoms; i++)
7150 free_cpumask_var(doms[i]);
7155 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7156 * For now this just excludes isolated cpus, but could be used to
7157 * exclude other special cases in the future.
7159 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7163 arch_update_cpu_topology();
7165 doms_cur = alloc_sched_domains(ndoms_cur);
7167 doms_cur = &fallback_doms;
7168 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7170 err = build_sched_domains(doms_cur[0]);
7171 register_sched_domain_sysctl();
7176 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7177 struct cpumask *tmpmask)
7179 free_sched_groups(cpu_map, tmpmask);
7183 * Detach sched domains from a group of cpus specified in cpu_map
7184 * These cpus will now be attached to the NULL domain
7186 static void detach_destroy_domains(const struct cpumask *cpu_map)
7188 /* Save because hotplug lock held. */
7189 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7192 for_each_cpu(i, cpu_map)
7193 cpu_attach_domain(NULL, &def_root_domain, i);
7194 synchronize_sched();
7195 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7198 /* handle null as "default" */
7199 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7200 struct sched_domain_attr *new, int idx_new)
7202 struct sched_domain_attr tmp;
7209 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7210 new ? (new + idx_new) : &tmp,
7211 sizeof(struct sched_domain_attr));
7215 * Partition sched domains as specified by the 'ndoms_new'
7216 * cpumasks in the array doms_new[] of cpumasks. This compares
7217 * doms_new[] to the current sched domain partitioning, doms_cur[].
7218 * It destroys each deleted domain and builds each new domain.
7220 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7221 * The masks don't intersect (don't overlap.) We should setup one
7222 * sched domain for each mask. CPUs not in any of the cpumasks will
7223 * not be load balanced. If the same cpumask appears both in the
7224 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7227 * The passed in 'doms_new' should be allocated using
7228 * alloc_sched_domains. This routine takes ownership of it and will
7229 * free_sched_domains it when done with it. If the caller failed the
7230 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7231 * and partition_sched_domains() will fallback to the single partition
7232 * 'fallback_doms', it also forces the domains to be rebuilt.
7234 * If doms_new == NULL it will be replaced with cpu_online_mask.
7235 * ndoms_new == 0 is a special case for destroying existing domains,
7236 * and it will not create the default domain.
7238 * Call with hotplug lock held
7240 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7241 struct sched_domain_attr *dattr_new)
7246 mutex_lock(&sched_domains_mutex);
7248 /* always unregister in case we don't destroy any domains */
7249 unregister_sched_domain_sysctl();
7251 /* Let architecture update cpu core mappings. */
7252 new_topology = arch_update_cpu_topology();
7254 n = doms_new ? ndoms_new : 0;
7256 /* Destroy deleted domains */
7257 for (i = 0; i < ndoms_cur; i++) {
7258 for (j = 0; j < n && !new_topology; j++) {
7259 if (cpumask_equal(doms_cur[i], doms_new[j])
7260 && dattrs_equal(dattr_cur, i, dattr_new, j))
7263 /* no match - a current sched domain not in new doms_new[] */
7264 detach_destroy_domains(doms_cur[i]);
7269 if (doms_new == NULL) {
7271 doms_new = &fallback_doms;
7272 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7273 WARN_ON_ONCE(dattr_new);
7276 /* Build new domains */
7277 for (i = 0; i < ndoms_new; i++) {
7278 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7279 if (cpumask_equal(doms_new[i], doms_cur[j])
7280 && dattrs_equal(dattr_new, i, dattr_cur, j))
7283 /* no match - add a new doms_new */
7284 __build_sched_domains(doms_new[i],
7285 dattr_new ? dattr_new + i : NULL);
7290 /* Remember the new sched domains */
7291 if (doms_cur != &fallback_doms)
7292 free_sched_domains(doms_cur, ndoms_cur);
7293 kfree(dattr_cur); /* kfree(NULL) is safe */
7294 doms_cur = doms_new;
7295 dattr_cur = dattr_new;
7296 ndoms_cur = ndoms_new;
7298 register_sched_domain_sysctl();
7300 mutex_unlock(&sched_domains_mutex);
7303 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7304 static void arch_reinit_sched_domains(void)
7308 /* Destroy domains first to force the rebuild */
7309 partition_sched_domains(0, NULL, NULL);
7311 rebuild_sched_domains();
7315 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7317 unsigned int level = 0;
7319 if (sscanf(buf, "%u", &level) != 1)
7323 * level is always be positive so don't check for
7324 * level < POWERSAVINGS_BALANCE_NONE which is 0
7325 * What happens on 0 or 1 byte write,
7326 * need to check for count as well?
7329 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7333 sched_smt_power_savings = level;
7335 sched_mc_power_savings = level;
7337 arch_reinit_sched_domains();
7342 #ifdef CONFIG_SCHED_MC
7343 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7346 return sprintf(page, "%u\n", sched_mc_power_savings);
7348 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7349 const char *buf, size_t count)
7351 return sched_power_savings_store(buf, count, 0);
7353 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7354 sched_mc_power_savings_show,
7355 sched_mc_power_savings_store);
7358 #ifdef CONFIG_SCHED_SMT
7359 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7362 return sprintf(page, "%u\n", sched_smt_power_savings);
7364 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7365 const char *buf, size_t count)
7367 return sched_power_savings_store(buf, count, 1);
7369 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7370 sched_smt_power_savings_show,
7371 sched_smt_power_savings_store);
7374 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7378 #ifdef CONFIG_SCHED_SMT
7380 err = sysfs_create_file(&cls->kset.kobj,
7381 &attr_sched_smt_power_savings.attr);
7383 #ifdef CONFIG_SCHED_MC
7384 if (!err && mc_capable())
7385 err = sysfs_create_file(&cls->kset.kobj,
7386 &attr_sched_mc_power_savings.attr);
7390 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7392 #ifndef CONFIG_CPUSETS
7394 * Add online and remove offline CPUs from the scheduler domains.
7395 * When cpusets are enabled they take over this function.
7397 static int update_sched_domains(struct notifier_block *nfb,
7398 unsigned long action, void *hcpu)
7402 case CPU_ONLINE_FROZEN:
7403 case CPU_DOWN_PREPARE:
7404 case CPU_DOWN_PREPARE_FROZEN:
7405 case CPU_DOWN_FAILED:
7406 case CPU_DOWN_FAILED_FROZEN:
7407 partition_sched_domains(1, NULL, NULL);
7416 static int update_runtime(struct notifier_block *nfb,
7417 unsigned long action, void *hcpu)
7419 int cpu = (int)(long)hcpu;
7422 case CPU_DOWN_PREPARE:
7423 case CPU_DOWN_PREPARE_FROZEN:
7424 disable_runtime(cpu_rq(cpu));
7427 case CPU_DOWN_FAILED:
7428 case CPU_DOWN_FAILED_FROZEN:
7430 case CPU_ONLINE_FROZEN:
7431 enable_runtime(cpu_rq(cpu));
7439 void __init sched_init_smp(void)
7441 cpumask_var_t non_isolated_cpus;
7443 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7444 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7446 #if defined(CONFIG_NUMA)
7447 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7449 BUG_ON(sched_group_nodes_bycpu == NULL);
7452 mutex_lock(&sched_domains_mutex);
7453 arch_init_sched_domains(cpu_active_mask);
7454 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7455 if (cpumask_empty(non_isolated_cpus))
7456 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7457 mutex_unlock(&sched_domains_mutex);
7460 #ifndef CONFIG_CPUSETS
7461 /* XXX: Theoretical race here - CPU may be hotplugged now */
7462 hotcpu_notifier(update_sched_domains, 0);
7465 /* RT runtime code needs to handle some hotplug events */
7466 hotcpu_notifier(update_runtime, 0);
7470 /* Move init over to a non-isolated CPU */
7471 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7473 sched_init_granularity();
7474 free_cpumask_var(non_isolated_cpus);
7476 init_sched_rt_class();
7479 void __init sched_init_smp(void)
7481 sched_init_granularity();
7483 #endif /* CONFIG_SMP */
7485 const_debug unsigned int sysctl_timer_migration = 1;
7487 int in_sched_functions(unsigned long addr)
7489 return in_lock_functions(addr) ||
7490 (addr >= (unsigned long)__sched_text_start
7491 && addr < (unsigned long)__sched_text_end);
7494 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7496 cfs_rq->tasks_timeline = RB_ROOT;
7497 INIT_LIST_HEAD(&cfs_rq->tasks);
7498 #ifdef CONFIG_FAIR_GROUP_SCHED
7501 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7504 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7506 struct rt_prio_array *array;
7509 array = &rt_rq->active;
7510 for (i = 0; i < MAX_RT_PRIO; i++) {
7511 INIT_LIST_HEAD(array->queue + i);
7512 __clear_bit(i, array->bitmap);
7514 /* delimiter for bitsearch: */
7515 __set_bit(MAX_RT_PRIO, array->bitmap);
7517 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7518 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7520 rt_rq->highest_prio.next = MAX_RT_PRIO;
7524 rt_rq->rt_nr_migratory = 0;
7525 rt_rq->overloaded = 0;
7526 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7530 rt_rq->rt_throttled = 0;
7531 rt_rq->rt_runtime = 0;
7532 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7534 #ifdef CONFIG_RT_GROUP_SCHED
7535 rt_rq->rt_nr_boosted = 0;
7540 #ifdef CONFIG_FAIR_GROUP_SCHED
7541 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7542 struct sched_entity *se, int cpu, int add,
7543 struct sched_entity *parent)
7545 struct rq *rq = cpu_rq(cpu);
7546 tg->cfs_rq[cpu] = cfs_rq;
7547 init_cfs_rq(cfs_rq, rq);
7550 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7553 /* se could be NULL for init_task_group */
7558 se->cfs_rq = &rq->cfs;
7560 se->cfs_rq = parent->my_q;
7563 se->load.weight = tg->shares;
7564 se->load.inv_weight = 0;
7565 se->parent = parent;
7569 #ifdef CONFIG_RT_GROUP_SCHED
7570 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7571 struct sched_rt_entity *rt_se, int cpu, int add,
7572 struct sched_rt_entity *parent)
7574 struct rq *rq = cpu_rq(cpu);
7576 tg->rt_rq[cpu] = rt_rq;
7577 init_rt_rq(rt_rq, rq);
7579 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7581 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7583 tg->rt_se[cpu] = rt_se;
7588 rt_se->rt_rq = &rq->rt;
7590 rt_se->rt_rq = parent->my_q;
7592 rt_se->my_q = rt_rq;
7593 rt_se->parent = parent;
7594 INIT_LIST_HEAD(&rt_se->run_list);
7598 void __init sched_init(void)
7601 unsigned long alloc_size = 0, ptr;
7603 #ifdef CONFIG_FAIR_GROUP_SCHED
7604 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7606 #ifdef CONFIG_RT_GROUP_SCHED
7607 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7609 #ifdef CONFIG_CPUMASK_OFFSTACK
7610 alloc_size += num_possible_cpus() * cpumask_size();
7613 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7615 #ifdef CONFIG_FAIR_GROUP_SCHED
7616 init_task_group.se = (struct sched_entity **)ptr;
7617 ptr += nr_cpu_ids * sizeof(void **);
7619 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7620 ptr += nr_cpu_ids * sizeof(void **);
7622 #endif /* CONFIG_FAIR_GROUP_SCHED */
7623 #ifdef CONFIG_RT_GROUP_SCHED
7624 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7625 ptr += nr_cpu_ids * sizeof(void **);
7627 init_task_group.rt_rq = (struct rt_rq **)ptr;
7628 ptr += nr_cpu_ids * sizeof(void **);
7630 #endif /* CONFIG_RT_GROUP_SCHED */
7631 #ifdef CONFIG_CPUMASK_OFFSTACK
7632 for_each_possible_cpu(i) {
7633 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7634 ptr += cpumask_size();
7636 #endif /* CONFIG_CPUMASK_OFFSTACK */
7640 init_defrootdomain();
7643 init_rt_bandwidth(&def_rt_bandwidth,
7644 global_rt_period(), global_rt_runtime());
7646 #ifdef CONFIG_RT_GROUP_SCHED
7647 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7648 global_rt_period(), global_rt_runtime());
7649 #endif /* CONFIG_RT_GROUP_SCHED */
7651 #ifdef CONFIG_CGROUP_SCHED
7652 list_add(&init_task_group.list, &task_groups);
7653 INIT_LIST_HEAD(&init_task_group.children);
7655 #endif /* CONFIG_CGROUP_SCHED */
7657 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7658 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7659 __alignof__(unsigned long));
7661 for_each_possible_cpu(i) {
7665 raw_spin_lock_init(&rq->lock);
7667 rq->calc_load_active = 0;
7668 rq->calc_load_update = jiffies + LOAD_FREQ;
7669 init_cfs_rq(&rq->cfs, rq);
7670 init_rt_rq(&rq->rt, rq);
7671 #ifdef CONFIG_FAIR_GROUP_SCHED
7672 init_task_group.shares = init_task_group_load;
7673 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7674 #ifdef CONFIG_CGROUP_SCHED
7676 * How much cpu bandwidth does init_task_group get?
7678 * In case of task-groups formed thr' the cgroup filesystem, it
7679 * gets 100% of the cpu resources in the system. This overall
7680 * system cpu resource is divided among the tasks of
7681 * init_task_group and its child task-groups in a fair manner,
7682 * based on each entity's (task or task-group's) weight
7683 * (se->load.weight).
7685 * In other words, if init_task_group has 10 tasks of weight
7686 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7687 * then A0's share of the cpu resource is:
7689 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7691 * We achieve this by letting init_task_group's tasks sit
7692 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7694 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7696 #endif /* CONFIG_FAIR_GROUP_SCHED */
7698 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7699 #ifdef CONFIG_RT_GROUP_SCHED
7700 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7701 #ifdef CONFIG_CGROUP_SCHED
7702 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7706 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7707 rq->cpu_load[j] = 0;
7711 rq->post_schedule = 0;
7712 rq->active_balance = 0;
7713 rq->next_balance = jiffies;
7717 rq->migration_thread = NULL;
7719 rq->avg_idle = 2*sysctl_sched_migration_cost;
7720 INIT_LIST_HEAD(&rq->migration_queue);
7721 rq_attach_root(rq, &def_root_domain);
7724 atomic_set(&rq->nr_iowait, 0);
7727 set_load_weight(&init_task);
7729 #ifdef CONFIG_PREEMPT_NOTIFIERS
7730 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7734 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7737 #ifdef CONFIG_RT_MUTEXES
7738 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7742 * The boot idle thread does lazy MMU switching as well:
7744 atomic_inc(&init_mm.mm_count);
7745 enter_lazy_tlb(&init_mm, current);
7748 * Make us the idle thread. Technically, schedule() should not be
7749 * called from this thread, however somewhere below it might be,
7750 * but because we are the idle thread, we just pick up running again
7751 * when this runqueue becomes "idle".
7753 init_idle(current, smp_processor_id());
7755 calc_load_update = jiffies + LOAD_FREQ;
7758 * During early bootup we pretend to be a normal task:
7760 current->sched_class = &fair_sched_class;
7762 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7763 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7766 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7767 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7769 /* May be allocated at isolcpus cmdline parse time */
7770 if (cpu_isolated_map == NULL)
7771 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7776 scheduler_running = 1;
7779 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7780 static inline int preempt_count_equals(int preempt_offset)
7782 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7784 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7787 void __might_sleep(const char *file, int line, int preempt_offset)
7790 static unsigned long prev_jiffy; /* ratelimiting */
7792 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7793 system_state != SYSTEM_RUNNING || oops_in_progress)
7795 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7797 prev_jiffy = jiffies;
7800 "BUG: sleeping function called from invalid context at %s:%d\n",
7803 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7804 in_atomic(), irqs_disabled(),
7805 current->pid, current->comm);
7807 debug_show_held_locks(current);
7808 if (irqs_disabled())
7809 print_irqtrace_events(current);
7813 EXPORT_SYMBOL(__might_sleep);
7816 #ifdef CONFIG_MAGIC_SYSRQ
7817 static void normalize_task(struct rq *rq, struct task_struct *p)
7821 on_rq = p->se.on_rq;
7823 deactivate_task(rq, p, 0);
7824 __setscheduler(rq, p, SCHED_NORMAL, 0);
7826 activate_task(rq, p, 0);
7827 resched_task(rq->curr);
7831 void normalize_rt_tasks(void)
7833 struct task_struct *g, *p;
7834 unsigned long flags;
7837 read_lock_irqsave(&tasklist_lock, flags);
7838 do_each_thread(g, p) {
7840 * Only normalize user tasks:
7845 p->se.exec_start = 0;
7846 #ifdef CONFIG_SCHEDSTATS
7847 p->se.statistics.wait_start = 0;
7848 p->se.statistics.sleep_start = 0;
7849 p->se.statistics.block_start = 0;
7854 * Renice negative nice level userspace
7857 if (TASK_NICE(p) < 0 && p->mm)
7858 set_user_nice(p, 0);
7862 raw_spin_lock(&p->pi_lock);
7863 rq = __task_rq_lock(p);
7865 normalize_task(rq, p);
7867 __task_rq_unlock(rq);
7868 raw_spin_unlock(&p->pi_lock);
7869 } while_each_thread(g, p);
7871 read_unlock_irqrestore(&tasklist_lock, flags);
7874 #endif /* CONFIG_MAGIC_SYSRQ */
7878 * These functions are only useful for the IA64 MCA handling.
7880 * They can only be called when the whole system has been
7881 * stopped - every CPU needs to be quiescent, and no scheduling
7882 * activity can take place. Using them for anything else would
7883 * be a serious bug, and as a result, they aren't even visible
7884 * under any other configuration.
7888 * curr_task - return the current task for a given cpu.
7889 * @cpu: the processor in question.
7891 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7893 struct task_struct *curr_task(int cpu)
7895 return cpu_curr(cpu);
7899 * set_curr_task - set the current task for a given cpu.
7900 * @cpu: the processor in question.
7901 * @p: the task pointer to set.
7903 * Description: This function must only be used when non-maskable interrupts
7904 * are serviced on a separate stack. It allows the architecture to switch the
7905 * notion of the current task on a cpu in a non-blocking manner. This function
7906 * must be called with all CPU's synchronized, and interrupts disabled, the
7907 * and caller must save the original value of the current task (see
7908 * curr_task() above) and restore that value before reenabling interrupts and
7909 * re-starting the system.
7911 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7913 void set_curr_task(int cpu, struct task_struct *p)
7920 #ifdef CONFIG_FAIR_GROUP_SCHED
7921 static void free_fair_sched_group(struct task_group *tg)
7925 for_each_possible_cpu(i) {
7927 kfree(tg->cfs_rq[i]);
7937 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7939 struct cfs_rq *cfs_rq;
7940 struct sched_entity *se;
7944 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7947 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7951 tg->shares = NICE_0_LOAD;
7953 for_each_possible_cpu(i) {
7956 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7957 GFP_KERNEL, cpu_to_node(i));
7961 se = kzalloc_node(sizeof(struct sched_entity),
7962 GFP_KERNEL, cpu_to_node(i));
7966 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
7977 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7979 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7980 &cpu_rq(cpu)->leaf_cfs_rq_list);
7983 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7985 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7987 #else /* !CONFG_FAIR_GROUP_SCHED */
7988 static inline void free_fair_sched_group(struct task_group *tg)
7993 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7998 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8002 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8005 #endif /* CONFIG_FAIR_GROUP_SCHED */
8007 #ifdef CONFIG_RT_GROUP_SCHED
8008 static void free_rt_sched_group(struct task_group *tg)
8012 destroy_rt_bandwidth(&tg->rt_bandwidth);
8014 for_each_possible_cpu(i) {
8016 kfree(tg->rt_rq[i]);
8018 kfree(tg->rt_se[i]);
8026 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8028 struct rt_rq *rt_rq;
8029 struct sched_rt_entity *rt_se;
8033 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8036 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8040 init_rt_bandwidth(&tg->rt_bandwidth,
8041 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8043 for_each_possible_cpu(i) {
8046 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8047 GFP_KERNEL, cpu_to_node(i));
8051 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8052 GFP_KERNEL, cpu_to_node(i));
8056 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8067 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8069 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8070 &cpu_rq(cpu)->leaf_rt_rq_list);
8073 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8075 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8077 #else /* !CONFIG_RT_GROUP_SCHED */
8078 static inline void free_rt_sched_group(struct task_group *tg)
8083 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8088 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8092 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8095 #endif /* CONFIG_RT_GROUP_SCHED */
8097 #ifdef CONFIG_CGROUP_SCHED
8098 static void free_sched_group(struct task_group *tg)
8100 free_fair_sched_group(tg);
8101 free_rt_sched_group(tg);
8105 /* allocate runqueue etc for a new task group */
8106 struct task_group *sched_create_group(struct task_group *parent)
8108 struct task_group *tg;
8109 unsigned long flags;
8112 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8114 return ERR_PTR(-ENOMEM);
8116 if (!alloc_fair_sched_group(tg, parent))
8119 if (!alloc_rt_sched_group(tg, parent))
8122 spin_lock_irqsave(&task_group_lock, flags);
8123 for_each_possible_cpu(i) {
8124 register_fair_sched_group(tg, i);
8125 register_rt_sched_group(tg, i);
8127 list_add_rcu(&tg->list, &task_groups);
8129 WARN_ON(!parent); /* root should already exist */
8131 tg->parent = parent;
8132 INIT_LIST_HEAD(&tg->children);
8133 list_add_rcu(&tg->siblings, &parent->children);
8134 spin_unlock_irqrestore(&task_group_lock, flags);
8139 free_sched_group(tg);
8140 return ERR_PTR(-ENOMEM);
8143 /* rcu callback to free various structures associated with a task group */
8144 static void free_sched_group_rcu(struct rcu_head *rhp)
8146 /* now it should be safe to free those cfs_rqs */
8147 free_sched_group(container_of(rhp, struct task_group, rcu));
8150 /* Destroy runqueue etc associated with a task group */
8151 void sched_destroy_group(struct task_group *tg)
8153 unsigned long flags;
8156 spin_lock_irqsave(&task_group_lock, flags);
8157 for_each_possible_cpu(i) {
8158 unregister_fair_sched_group(tg, i);
8159 unregister_rt_sched_group(tg, i);
8161 list_del_rcu(&tg->list);
8162 list_del_rcu(&tg->siblings);
8163 spin_unlock_irqrestore(&task_group_lock, flags);
8165 /* wait for possible concurrent references to cfs_rqs complete */
8166 call_rcu(&tg->rcu, free_sched_group_rcu);
8169 /* change task's runqueue when it moves between groups.
8170 * The caller of this function should have put the task in its new group
8171 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8172 * reflect its new group.
8174 void sched_move_task(struct task_struct *tsk)
8177 unsigned long flags;
8180 rq = task_rq_lock(tsk, &flags);
8182 running = task_current(rq, tsk);
8183 on_rq = tsk->se.on_rq;
8186 dequeue_task(rq, tsk, 0);
8187 if (unlikely(running))
8188 tsk->sched_class->put_prev_task(rq, tsk);
8190 set_task_rq(tsk, task_cpu(tsk));
8192 #ifdef CONFIG_FAIR_GROUP_SCHED
8193 if (tsk->sched_class->moved_group)
8194 tsk->sched_class->moved_group(tsk, on_rq);
8197 if (unlikely(running))
8198 tsk->sched_class->set_curr_task(rq);
8200 enqueue_task(rq, tsk, 0, false);
8202 task_rq_unlock(rq, &flags);
8204 #endif /* CONFIG_CGROUP_SCHED */
8206 #ifdef CONFIG_FAIR_GROUP_SCHED
8207 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8209 struct cfs_rq *cfs_rq = se->cfs_rq;
8214 dequeue_entity(cfs_rq, se, 0);
8216 se->load.weight = shares;
8217 se->load.inv_weight = 0;
8220 enqueue_entity(cfs_rq, se, 0);
8223 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8225 struct cfs_rq *cfs_rq = se->cfs_rq;
8226 struct rq *rq = cfs_rq->rq;
8227 unsigned long flags;
8229 raw_spin_lock_irqsave(&rq->lock, flags);
8230 __set_se_shares(se, shares);
8231 raw_spin_unlock_irqrestore(&rq->lock, flags);
8234 static DEFINE_MUTEX(shares_mutex);
8236 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8239 unsigned long flags;
8242 * We can't change the weight of the root cgroup.
8247 if (shares < MIN_SHARES)
8248 shares = MIN_SHARES;
8249 else if (shares > MAX_SHARES)
8250 shares = MAX_SHARES;
8252 mutex_lock(&shares_mutex);
8253 if (tg->shares == shares)
8256 spin_lock_irqsave(&task_group_lock, flags);
8257 for_each_possible_cpu(i)
8258 unregister_fair_sched_group(tg, i);
8259 list_del_rcu(&tg->siblings);
8260 spin_unlock_irqrestore(&task_group_lock, flags);
8262 /* wait for any ongoing reference to this group to finish */
8263 synchronize_sched();
8266 * Now we are free to modify the group's share on each cpu
8267 * w/o tripping rebalance_share or load_balance_fair.
8269 tg->shares = shares;
8270 for_each_possible_cpu(i) {
8274 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8275 set_se_shares(tg->se[i], shares);
8279 * Enable load balance activity on this group, by inserting it back on
8280 * each cpu's rq->leaf_cfs_rq_list.
8282 spin_lock_irqsave(&task_group_lock, flags);
8283 for_each_possible_cpu(i)
8284 register_fair_sched_group(tg, i);
8285 list_add_rcu(&tg->siblings, &tg->parent->children);
8286 spin_unlock_irqrestore(&task_group_lock, flags);
8288 mutex_unlock(&shares_mutex);
8292 unsigned long sched_group_shares(struct task_group *tg)
8298 #ifdef CONFIG_RT_GROUP_SCHED
8300 * Ensure that the real time constraints are schedulable.
8302 static DEFINE_MUTEX(rt_constraints_mutex);
8304 static unsigned long to_ratio(u64 period, u64 runtime)
8306 if (runtime == RUNTIME_INF)
8309 return div64_u64(runtime << 20, period);
8312 /* Must be called with tasklist_lock held */
8313 static inline int tg_has_rt_tasks(struct task_group *tg)
8315 struct task_struct *g, *p;
8317 do_each_thread(g, p) {
8318 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8320 } while_each_thread(g, p);
8325 struct rt_schedulable_data {
8326 struct task_group *tg;
8331 static int tg_schedulable(struct task_group *tg, void *data)
8333 struct rt_schedulable_data *d = data;
8334 struct task_group *child;
8335 unsigned long total, sum = 0;
8336 u64 period, runtime;
8338 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8339 runtime = tg->rt_bandwidth.rt_runtime;
8342 period = d->rt_period;
8343 runtime = d->rt_runtime;
8347 * Cannot have more runtime than the period.
8349 if (runtime > period && runtime != RUNTIME_INF)
8353 * Ensure we don't starve existing RT tasks.
8355 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8358 total = to_ratio(period, runtime);
8361 * Nobody can have more than the global setting allows.
8363 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8367 * The sum of our children's runtime should not exceed our own.
8369 list_for_each_entry_rcu(child, &tg->children, siblings) {
8370 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8371 runtime = child->rt_bandwidth.rt_runtime;
8373 if (child == d->tg) {
8374 period = d->rt_period;
8375 runtime = d->rt_runtime;
8378 sum += to_ratio(period, runtime);
8387 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8389 struct rt_schedulable_data data = {
8391 .rt_period = period,
8392 .rt_runtime = runtime,
8395 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8398 static int tg_set_bandwidth(struct task_group *tg,
8399 u64 rt_period, u64 rt_runtime)
8403 mutex_lock(&rt_constraints_mutex);
8404 read_lock(&tasklist_lock);
8405 err = __rt_schedulable(tg, rt_period, rt_runtime);
8409 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8410 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8411 tg->rt_bandwidth.rt_runtime = rt_runtime;
8413 for_each_possible_cpu(i) {
8414 struct rt_rq *rt_rq = tg->rt_rq[i];
8416 raw_spin_lock(&rt_rq->rt_runtime_lock);
8417 rt_rq->rt_runtime = rt_runtime;
8418 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8420 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8422 read_unlock(&tasklist_lock);
8423 mutex_unlock(&rt_constraints_mutex);
8428 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8430 u64 rt_runtime, rt_period;
8432 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8433 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8434 if (rt_runtime_us < 0)
8435 rt_runtime = RUNTIME_INF;
8437 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8440 long sched_group_rt_runtime(struct task_group *tg)
8444 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8447 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8448 do_div(rt_runtime_us, NSEC_PER_USEC);
8449 return rt_runtime_us;
8452 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8454 u64 rt_runtime, rt_period;
8456 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8457 rt_runtime = tg->rt_bandwidth.rt_runtime;
8462 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8465 long sched_group_rt_period(struct task_group *tg)
8469 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8470 do_div(rt_period_us, NSEC_PER_USEC);
8471 return rt_period_us;
8474 static int sched_rt_global_constraints(void)
8476 u64 runtime, period;
8479 if (sysctl_sched_rt_period <= 0)
8482 runtime = global_rt_runtime();
8483 period = global_rt_period();
8486 * Sanity check on the sysctl variables.
8488 if (runtime > period && runtime != RUNTIME_INF)
8491 mutex_lock(&rt_constraints_mutex);
8492 read_lock(&tasklist_lock);
8493 ret = __rt_schedulable(NULL, 0, 0);
8494 read_unlock(&tasklist_lock);
8495 mutex_unlock(&rt_constraints_mutex);
8500 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8502 /* Don't accept realtime tasks when there is no way for them to run */
8503 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8509 #else /* !CONFIG_RT_GROUP_SCHED */
8510 static int sched_rt_global_constraints(void)
8512 unsigned long flags;
8515 if (sysctl_sched_rt_period <= 0)
8519 * There's always some RT tasks in the root group
8520 * -- migration, kstopmachine etc..
8522 if (sysctl_sched_rt_runtime == 0)
8525 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8526 for_each_possible_cpu(i) {
8527 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8529 raw_spin_lock(&rt_rq->rt_runtime_lock);
8530 rt_rq->rt_runtime = global_rt_runtime();
8531 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8533 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8537 #endif /* CONFIG_RT_GROUP_SCHED */
8539 int sched_rt_handler(struct ctl_table *table, int write,
8540 void __user *buffer, size_t *lenp,
8544 int old_period, old_runtime;
8545 static DEFINE_MUTEX(mutex);
8548 old_period = sysctl_sched_rt_period;
8549 old_runtime = sysctl_sched_rt_runtime;
8551 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8553 if (!ret && write) {
8554 ret = sched_rt_global_constraints();
8556 sysctl_sched_rt_period = old_period;
8557 sysctl_sched_rt_runtime = old_runtime;
8559 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8560 def_rt_bandwidth.rt_period =
8561 ns_to_ktime(global_rt_period());
8564 mutex_unlock(&mutex);
8569 #ifdef CONFIG_CGROUP_SCHED
8571 /* return corresponding task_group object of a cgroup */
8572 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8574 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8575 struct task_group, css);
8578 static struct cgroup_subsys_state *
8579 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8581 struct task_group *tg, *parent;
8583 if (!cgrp->parent) {
8584 /* This is early initialization for the top cgroup */
8585 return &init_task_group.css;
8588 parent = cgroup_tg(cgrp->parent);
8589 tg = sched_create_group(parent);
8591 return ERR_PTR(-ENOMEM);
8597 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8599 struct task_group *tg = cgroup_tg(cgrp);
8601 sched_destroy_group(tg);
8605 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8607 #ifdef CONFIG_RT_GROUP_SCHED
8608 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8611 /* We don't support RT-tasks being in separate groups */
8612 if (tsk->sched_class != &fair_sched_class)
8619 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8620 struct task_struct *tsk, bool threadgroup)
8622 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8626 struct task_struct *c;
8628 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8629 retval = cpu_cgroup_can_attach_task(cgrp, c);
8641 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8642 struct cgroup *old_cont, struct task_struct *tsk,
8645 sched_move_task(tsk);
8647 struct task_struct *c;
8649 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8656 #ifdef CONFIG_FAIR_GROUP_SCHED
8657 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8660 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8663 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8665 struct task_group *tg = cgroup_tg(cgrp);
8667 return (u64) tg->shares;
8669 #endif /* CONFIG_FAIR_GROUP_SCHED */
8671 #ifdef CONFIG_RT_GROUP_SCHED
8672 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8675 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8678 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8680 return sched_group_rt_runtime(cgroup_tg(cgrp));
8683 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8686 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8689 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8691 return sched_group_rt_period(cgroup_tg(cgrp));
8693 #endif /* CONFIG_RT_GROUP_SCHED */
8695 static struct cftype cpu_files[] = {
8696 #ifdef CONFIG_FAIR_GROUP_SCHED
8699 .read_u64 = cpu_shares_read_u64,
8700 .write_u64 = cpu_shares_write_u64,
8703 #ifdef CONFIG_RT_GROUP_SCHED
8705 .name = "rt_runtime_us",
8706 .read_s64 = cpu_rt_runtime_read,
8707 .write_s64 = cpu_rt_runtime_write,
8710 .name = "rt_period_us",
8711 .read_u64 = cpu_rt_period_read_uint,
8712 .write_u64 = cpu_rt_period_write_uint,
8717 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8719 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8722 struct cgroup_subsys cpu_cgroup_subsys = {
8724 .create = cpu_cgroup_create,
8725 .destroy = cpu_cgroup_destroy,
8726 .can_attach = cpu_cgroup_can_attach,
8727 .attach = cpu_cgroup_attach,
8728 .populate = cpu_cgroup_populate,
8729 .subsys_id = cpu_cgroup_subsys_id,
8733 #endif /* CONFIG_CGROUP_SCHED */
8735 #ifdef CONFIG_CGROUP_CPUACCT
8738 * CPU accounting code for task groups.
8740 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8741 * (balbir@in.ibm.com).
8744 /* track cpu usage of a group of tasks and its child groups */
8746 struct cgroup_subsys_state css;
8747 /* cpuusage holds pointer to a u64-type object on every cpu */
8749 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8750 struct cpuacct *parent;
8753 struct cgroup_subsys cpuacct_subsys;
8755 /* return cpu accounting group corresponding to this container */
8756 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8758 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8759 struct cpuacct, css);
8762 /* return cpu accounting group to which this task belongs */
8763 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8765 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8766 struct cpuacct, css);
8769 /* create a new cpu accounting group */
8770 static struct cgroup_subsys_state *cpuacct_create(
8771 struct cgroup_subsys *ss, struct cgroup *cgrp)
8773 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8779 ca->cpuusage = alloc_percpu(u64);
8783 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8784 if (percpu_counter_init(&ca->cpustat[i], 0))
8785 goto out_free_counters;
8788 ca->parent = cgroup_ca(cgrp->parent);
8794 percpu_counter_destroy(&ca->cpustat[i]);
8795 free_percpu(ca->cpuusage);
8799 return ERR_PTR(-ENOMEM);
8802 /* destroy an existing cpu accounting group */
8804 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8806 struct cpuacct *ca = cgroup_ca(cgrp);
8809 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8810 percpu_counter_destroy(&ca->cpustat[i]);
8811 free_percpu(ca->cpuusage);
8815 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8817 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8820 #ifndef CONFIG_64BIT
8822 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8824 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8826 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8834 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8836 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8838 #ifndef CONFIG_64BIT
8840 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8842 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8844 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8850 /* return total cpu usage (in nanoseconds) of a group */
8851 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8853 struct cpuacct *ca = cgroup_ca(cgrp);
8854 u64 totalcpuusage = 0;
8857 for_each_present_cpu(i)
8858 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8860 return totalcpuusage;
8863 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8866 struct cpuacct *ca = cgroup_ca(cgrp);
8875 for_each_present_cpu(i)
8876 cpuacct_cpuusage_write(ca, i, 0);
8882 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8885 struct cpuacct *ca = cgroup_ca(cgroup);
8889 for_each_present_cpu(i) {
8890 percpu = cpuacct_cpuusage_read(ca, i);
8891 seq_printf(m, "%llu ", (unsigned long long) percpu);
8893 seq_printf(m, "\n");
8897 static const char *cpuacct_stat_desc[] = {
8898 [CPUACCT_STAT_USER] = "user",
8899 [CPUACCT_STAT_SYSTEM] = "system",
8902 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8903 struct cgroup_map_cb *cb)
8905 struct cpuacct *ca = cgroup_ca(cgrp);
8908 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8909 s64 val = percpu_counter_read(&ca->cpustat[i]);
8910 val = cputime64_to_clock_t(val);
8911 cb->fill(cb, cpuacct_stat_desc[i], val);
8916 static struct cftype files[] = {
8919 .read_u64 = cpuusage_read,
8920 .write_u64 = cpuusage_write,
8923 .name = "usage_percpu",
8924 .read_seq_string = cpuacct_percpu_seq_read,
8928 .read_map = cpuacct_stats_show,
8932 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8934 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8938 * charge this task's execution time to its accounting group.
8940 * called with rq->lock held.
8942 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8947 if (unlikely(!cpuacct_subsys.active))
8950 cpu = task_cpu(tsk);
8956 for (; ca; ca = ca->parent) {
8957 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8958 *cpuusage += cputime;
8965 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8966 * in cputime_t units. As a result, cpuacct_update_stats calls
8967 * percpu_counter_add with values large enough to always overflow the
8968 * per cpu batch limit causing bad SMP scalability.
8970 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8971 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8972 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8975 #define CPUACCT_BATCH \
8976 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8978 #define CPUACCT_BATCH 0
8982 * Charge the system/user time to the task's accounting group.
8984 static void cpuacct_update_stats(struct task_struct *tsk,
8985 enum cpuacct_stat_index idx, cputime_t val)
8988 int batch = CPUACCT_BATCH;
8990 if (unlikely(!cpuacct_subsys.active))
8997 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9003 struct cgroup_subsys cpuacct_subsys = {
9005 .create = cpuacct_create,
9006 .destroy = cpuacct_destroy,
9007 .populate = cpuacct_populate,
9008 .subsys_id = cpuacct_subsys_id,
9010 #endif /* CONFIG_CGROUP_CPUACCT */
9014 int rcu_expedited_torture_stats(char *page)
9018 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9020 void synchronize_sched_expedited(void)
9023 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9025 #else /* #ifndef CONFIG_SMP */
9027 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9028 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9030 #define RCU_EXPEDITED_STATE_POST -2
9031 #define RCU_EXPEDITED_STATE_IDLE -1
9033 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9035 int rcu_expedited_torture_stats(char *page)
9040 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9041 for_each_online_cpu(cpu) {
9042 cnt += sprintf(&page[cnt], " %d:%d",
9043 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9045 cnt += sprintf(&page[cnt], "\n");
9048 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9050 static long synchronize_sched_expedited_count;
9053 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9054 * approach to force grace period to end quickly. This consumes
9055 * significant time on all CPUs, and is thus not recommended for
9056 * any sort of common-case code.
9058 * Note that it is illegal to call this function while holding any
9059 * lock that is acquired by a CPU-hotplug notifier. Failing to
9060 * observe this restriction will result in deadlock.
9062 void synchronize_sched_expedited(void)
9065 unsigned long flags;
9066 bool need_full_sync = 0;
9068 struct migration_req *req;
9072 smp_mb(); /* ensure prior mod happens before capturing snap. */
9073 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9075 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9077 if (trycount++ < 10)
9078 udelay(trycount * num_online_cpus());
9080 synchronize_sched();
9083 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9084 smp_mb(); /* ensure test happens before caller kfree */
9089 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9090 for_each_online_cpu(cpu) {
9092 req = &per_cpu(rcu_migration_req, cpu);
9093 init_completion(&req->done);
9095 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9096 raw_spin_lock_irqsave(&rq->lock, flags);
9097 list_add(&req->list, &rq->migration_queue);
9098 raw_spin_unlock_irqrestore(&rq->lock, flags);
9099 wake_up_process(rq->migration_thread);
9101 for_each_online_cpu(cpu) {
9102 rcu_expedited_state = cpu;
9103 req = &per_cpu(rcu_migration_req, cpu);
9105 wait_for_completion(&req->done);
9106 raw_spin_lock_irqsave(&rq->lock, flags);
9107 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9109 req->dest_cpu = RCU_MIGRATION_IDLE;
9110 raw_spin_unlock_irqrestore(&rq->lock, flags);
9112 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9113 synchronize_sched_expedited_count++;
9114 mutex_unlock(&rcu_sched_expedited_mutex);
9117 synchronize_sched();
9119 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9121 #endif /* #else #ifndef CONFIG_SMP */