4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 raw_spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 raw_spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 raw_spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 raw_spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_CGROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 struct cgroup_subsys_state css;
248 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* schedulable entities of this group on each cpu */
250 struct sched_entity **se;
251 /* runqueue "owned" by this group on each cpu */
252 struct cfs_rq **cfs_rq;
253 unsigned long shares;
256 #ifdef CONFIG_RT_GROUP_SCHED
257 struct sched_rt_entity **rt_se;
258 struct rt_rq **rt_rq;
260 struct rt_bandwidth rt_bandwidth;
264 struct list_head list;
266 struct task_group *parent;
267 struct list_head siblings;
268 struct list_head children;
271 #define root_task_group init_task_group
273 /* task_group_lock serializes add/remove of task groups and also changes to
274 * a task group's cpu shares.
276 static DEFINE_SPINLOCK(task_group_lock);
278 #ifdef CONFIG_FAIR_GROUP_SCHED
281 static int root_task_group_empty(void)
283 return list_empty(&root_task_group.children);
287 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
290 * A weight of 0 or 1 can cause arithmetics problems.
291 * A weight of a cfs_rq is the sum of weights of which entities
292 * are queued on this cfs_rq, so a weight of a entity should not be
293 * too large, so as the shares value of a task group.
294 * (The default weight is 1024 - so there's no practical
295 * limitation from this.)
298 #define MAX_SHARES (1UL << 18)
300 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
303 /* Default task group.
304 * Every task in system belong to this group at bootup.
306 struct task_group init_task_group;
308 /* return group to which a task belongs */
309 static inline struct task_group *task_group(struct task_struct *p)
311 struct task_group *tg;
313 #ifdef CONFIG_CGROUP_SCHED
314 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
315 struct task_group, css);
317 tg = &init_task_group;
322 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
323 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
327 p->se.parent = task_group(p)->se[cpu];
330 #ifdef CONFIG_RT_GROUP_SCHED
331 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
332 p->rt.parent = task_group(p)->rt_se[cpu];
338 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
339 static inline struct task_group *task_group(struct task_struct *p)
344 #endif /* CONFIG_CGROUP_SCHED */
346 /* CFS-related fields in a runqueue */
348 struct load_weight load;
349 unsigned long nr_running;
354 struct rb_root tasks_timeline;
355 struct rb_node *rb_leftmost;
357 struct list_head tasks;
358 struct list_head *balance_iterator;
361 * 'curr' points to currently running entity on this cfs_rq.
362 * It is set to NULL otherwise (i.e when none are currently running).
364 struct sched_entity *curr, *next, *last;
366 unsigned int nr_spread_over;
368 #ifdef CONFIG_FAIR_GROUP_SCHED
369 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
372 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
373 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
374 * (like users, containers etc.)
376 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
377 * list is used during load balance.
379 struct list_head leaf_cfs_rq_list;
380 struct task_group *tg; /* group that "owns" this runqueue */
384 * the part of load.weight contributed by tasks
386 unsigned long task_weight;
389 * h_load = weight * f(tg)
391 * Where f(tg) is the recursive weight fraction assigned to
394 unsigned long h_load;
397 * this cpu's part of tg->shares
399 unsigned long shares;
402 * load.weight at the time we set shares
404 unsigned long rq_weight;
409 /* Real-Time classes' related field in a runqueue: */
411 struct rt_prio_array active;
412 unsigned long rt_nr_running;
413 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
415 int curr; /* highest queued rt task prio */
417 int next; /* next highest */
422 unsigned long rt_nr_migratory;
423 unsigned long rt_nr_total;
425 struct plist_head pushable_tasks;
430 /* Nests inside the rq lock: */
431 raw_spinlock_t rt_runtime_lock;
433 #ifdef CONFIG_RT_GROUP_SCHED
434 unsigned long rt_nr_boosted;
437 struct list_head leaf_rt_rq_list;
438 struct task_group *tg;
445 * We add the notion of a root-domain which will be used to define per-domain
446 * variables. Each exclusive cpuset essentially defines an island domain by
447 * fully partitioning the member cpus from any other cpuset. Whenever a new
448 * exclusive cpuset is created, we also create and attach a new root-domain
455 cpumask_var_t online;
458 * The "RT overload" flag: it gets set if a CPU has more than
459 * one runnable RT task.
461 cpumask_var_t rto_mask;
464 struct cpupri cpupri;
469 * By default the system creates a single root-domain with all cpus as
470 * members (mimicking the global state we have today).
472 static struct root_domain def_root_domain;
477 * This is the main, per-CPU runqueue data structure.
479 * Locking rule: those places that want to lock multiple runqueues
480 * (such as the load balancing or the thread migration code), lock
481 * acquire operations must be ordered by ascending &runqueue.
488 * nr_running and cpu_load should be in the same cacheline because
489 * remote CPUs use both these fields when doing load calculation.
491 unsigned long nr_running;
492 #define CPU_LOAD_IDX_MAX 5
493 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
496 unsigned char in_nohz_recently;
498 unsigned int skip_clock_update;
500 /* capture load from *all* tasks on this cpu: */
501 struct load_weight load;
502 unsigned long nr_load_updates;
508 #ifdef CONFIG_FAIR_GROUP_SCHED
509 /* list of leaf cfs_rq on this cpu: */
510 struct list_head leaf_cfs_rq_list;
512 #ifdef CONFIG_RT_GROUP_SCHED
513 struct list_head leaf_rt_rq_list;
517 * This is part of a global counter where only the total sum
518 * over all CPUs matters. A task can increase this counter on
519 * one CPU and if it got migrated afterwards it may decrease
520 * it on another CPU. Always updated under the runqueue lock:
522 unsigned long nr_uninterruptible;
524 struct task_struct *curr, *idle;
525 unsigned long next_balance;
526 struct mm_struct *prev_mm;
533 struct root_domain *rd;
534 struct sched_domain *sd;
536 unsigned char idle_at_tick;
537 /* For active balancing */
541 /* cpu of this runqueue: */
545 unsigned long avg_load_per_task;
547 struct task_struct *migration_thread;
548 struct list_head migration_queue;
556 /* calc_load related fields */
557 unsigned long calc_load_update;
558 long calc_load_active;
560 #ifdef CONFIG_SCHED_HRTICK
562 int hrtick_csd_pending;
563 struct call_single_data hrtick_csd;
565 struct hrtimer hrtick_timer;
568 #ifdef CONFIG_SCHEDSTATS
570 struct sched_info rq_sched_info;
571 unsigned long long rq_cpu_time;
572 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
574 /* sys_sched_yield() stats */
575 unsigned int yld_count;
577 /* schedule() stats */
578 unsigned int sched_switch;
579 unsigned int sched_count;
580 unsigned int sched_goidle;
582 /* try_to_wake_up() stats */
583 unsigned int ttwu_count;
584 unsigned int ttwu_local;
587 unsigned int bkl_count;
591 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
594 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
596 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
599 * A queue event has occurred, and we're going to schedule. In
600 * this case, we can save a useless back to back clock update.
602 if (test_tsk_need_resched(p))
603 rq->skip_clock_update = 1;
606 static inline int cpu_of(struct rq *rq)
615 #define rcu_dereference_check_sched_domain(p) \
616 rcu_dereference_check((p), \
617 rcu_read_lock_sched_held() || \
618 lockdep_is_held(&sched_domains_mutex))
621 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
622 * See detach_destroy_domains: synchronize_sched for details.
624 * The domain tree of any CPU may only be accessed from within
625 * preempt-disabled sections.
627 #define for_each_domain(cpu, __sd) \
628 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
630 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
631 #define this_rq() (&__get_cpu_var(runqueues))
632 #define task_rq(p) cpu_rq(task_cpu(p))
633 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
634 #define raw_rq() (&__raw_get_cpu_var(runqueues))
636 inline void update_rq_clock(struct rq *rq)
638 if (!rq->skip_clock_update)
639 rq->clock = sched_clock_cpu(cpu_of(rq));
643 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
645 #ifdef CONFIG_SCHED_DEBUG
646 # define const_debug __read_mostly
648 # define const_debug static const
653 * @cpu: the processor in question.
655 * Returns true if the current cpu runqueue is locked.
656 * This interface allows printk to be called with the runqueue lock
657 * held and know whether or not it is OK to wake up the klogd.
659 int runqueue_is_locked(int cpu)
661 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
665 * Debugging: various feature bits
668 #define SCHED_FEAT(name, enabled) \
669 __SCHED_FEAT_##name ,
672 #include "sched_features.h"
677 #define SCHED_FEAT(name, enabled) \
678 (1UL << __SCHED_FEAT_##name) * enabled |
680 const_debug unsigned int sysctl_sched_features =
681 #include "sched_features.h"
686 #ifdef CONFIG_SCHED_DEBUG
687 #define SCHED_FEAT(name, enabled) \
690 static __read_mostly char *sched_feat_names[] = {
691 #include "sched_features.h"
697 static int sched_feat_show(struct seq_file *m, void *v)
701 for (i = 0; sched_feat_names[i]; i++) {
702 if (!(sysctl_sched_features & (1UL << i)))
704 seq_printf(m, "%s ", sched_feat_names[i]);
712 sched_feat_write(struct file *filp, const char __user *ubuf,
713 size_t cnt, loff_t *ppos)
723 if (copy_from_user(&buf, ubuf, cnt))
728 if (strncmp(buf, "NO_", 3) == 0) {
733 for (i = 0; sched_feat_names[i]; i++) {
734 int len = strlen(sched_feat_names[i]);
736 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
738 sysctl_sched_features &= ~(1UL << i);
740 sysctl_sched_features |= (1UL << i);
745 if (!sched_feat_names[i])
753 static int sched_feat_open(struct inode *inode, struct file *filp)
755 return single_open(filp, sched_feat_show, NULL);
758 static const struct file_operations sched_feat_fops = {
759 .open = sched_feat_open,
760 .write = sched_feat_write,
763 .release = single_release,
766 static __init int sched_init_debug(void)
768 debugfs_create_file("sched_features", 0644, NULL, NULL,
773 late_initcall(sched_init_debug);
777 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
780 * Number of tasks to iterate in a single balance run.
781 * Limited because this is done with IRQs disabled.
783 const_debug unsigned int sysctl_sched_nr_migrate = 32;
786 * ratelimit for updating the group shares.
789 unsigned int sysctl_sched_shares_ratelimit = 250000;
790 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
793 * Inject some fuzzyness into changing the per-cpu group shares
794 * this avoids remote rq-locks at the expense of fairness.
797 unsigned int sysctl_sched_shares_thresh = 4;
800 * period over which we average the RT time consumption, measured
805 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
808 * period over which we measure -rt task cpu usage in us.
811 unsigned int sysctl_sched_rt_period = 1000000;
813 static __read_mostly int scheduler_running;
816 * part of the period that we allow rt tasks to run in us.
819 int sysctl_sched_rt_runtime = 950000;
821 static inline u64 global_rt_period(void)
823 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
826 static inline u64 global_rt_runtime(void)
828 if (sysctl_sched_rt_runtime < 0)
831 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
834 #ifndef prepare_arch_switch
835 # define prepare_arch_switch(next) do { } while (0)
837 #ifndef finish_arch_switch
838 # define finish_arch_switch(prev) do { } while (0)
841 static inline int task_current(struct rq *rq, struct task_struct *p)
843 return rq->curr == p;
846 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
847 static inline int task_running(struct rq *rq, struct task_struct *p)
849 return task_current(rq, p);
852 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
856 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
858 #ifdef CONFIG_DEBUG_SPINLOCK
859 /* this is a valid case when another task releases the spinlock */
860 rq->lock.owner = current;
863 * If we are tracking spinlock dependencies then we have to
864 * fix up the runqueue lock - which gets 'carried over' from
867 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
869 raw_spin_unlock_irq(&rq->lock);
872 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
873 static inline int task_running(struct rq *rq, struct task_struct *p)
878 return task_current(rq, p);
882 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
886 * We can optimise this out completely for !SMP, because the
887 * SMP rebalancing from interrupt is the only thing that cares
892 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
893 raw_spin_unlock_irq(&rq->lock);
895 raw_spin_unlock(&rq->lock);
899 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
903 * After ->oncpu is cleared, the task can be moved to a different CPU.
904 * We must ensure this doesn't happen until the switch is completely
910 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
914 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
917 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
920 static inline int task_is_waking(struct task_struct *p)
922 return unlikely(p->state == TASK_WAKING);
926 * __task_rq_lock - lock the runqueue a given task resides on.
927 * Must be called interrupts disabled.
929 static inline struct rq *__task_rq_lock(struct task_struct *p)
936 raw_spin_lock(&rq->lock);
937 if (likely(rq == task_rq(p)))
939 raw_spin_unlock(&rq->lock);
944 * task_rq_lock - lock the runqueue a given task resides on and disable
945 * interrupts. Note the ordering: we can safely lookup the task_rq without
946 * explicitly disabling preemption.
948 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
954 local_irq_save(*flags);
956 raw_spin_lock(&rq->lock);
957 if (likely(rq == task_rq(p)))
959 raw_spin_unlock_irqrestore(&rq->lock, *flags);
963 void task_rq_unlock_wait(struct task_struct *p)
965 struct rq *rq = task_rq(p);
967 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
968 raw_spin_unlock_wait(&rq->lock);
971 static void __task_rq_unlock(struct rq *rq)
974 raw_spin_unlock(&rq->lock);
977 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
980 raw_spin_unlock_irqrestore(&rq->lock, *flags);
984 * this_rq_lock - lock this runqueue and disable interrupts.
986 static struct rq *this_rq_lock(void)
993 raw_spin_lock(&rq->lock);
998 #ifdef CONFIG_SCHED_HRTICK
1000 * Use HR-timers to deliver accurate preemption points.
1002 * Its all a bit involved since we cannot program an hrt while holding the
1003 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1006 * When we get rescheduled we reprogram the hrtick_timer outside of the
1012 * - enabled by features
1013 * - hrtimer is actually high res
1015 static inline int hrtick_enabled(struct rq *rq)
1017 if (!sched_feat(HRTICK))
1019 if (!cpu_active(cpu_of(rq)))
1021 return hrtimer_is_hres_active(&rq->hrtick_timer);
1024 static void hrtick_clear(struct rq *rq)
1026 if (hrtimer_active(&rq->hrtick_timer))
1027 hrtimer_cancel(&rq->hrtick_timer);
1031 * High-resolution timer tick.
1032 * Runs from hardirq context with interrupts disabled.
1034 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1036 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1038 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1040 raw_spin_lock(&rq->lock);
1041 update_rq_clock(rq);
1042 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1043 raw_spin_unlock(&rq->lock);
1045 return HRTIMER_NORESTART;
1050 * called from hardirq (IPI) context
1052 static void __hrtick_start(void *arg)
1054 struct rq *rq = arg;
1056 raw_spin_lock(&rq->lock);
1057 hrtimer_restart(&rq->hrtick_timer);
1058 rq->hrtick_csd_pending = 0;
1059 raw_spin_unlock(&rq->lock);
1063 * Called to set the hrtick timer state.
1065 * called with rq->lock held and irqs disabled
1067 static void hrtick_start(struct rq *rq, u64 delay)
1069 struct hrtimer *timer = &rq->hrtick_timer;
1070 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1072 hrtimer_set_expires(timer, time);
1074 if (rq == this_rq()) {
1075 hrtimer_restart(timer);
1076 } else if (!rq->hrtick_csd_pending) {
1077 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1078 rq->hrtick_csd_pending = 1;
1083 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1085 int cpu = (int)(long)hcpu;
1088 case CPU_UP_CANCELED:
1089 case CPU_UP_CANCELED_FROZEN:
1090 case CPU_DOWN_PREPARE:
1091 case CPU_DOWN_PREPARE_FROZEN:
1093 case CPU_DEAD_FROZEN:
1094 hrtick_clear(cpu_rq(cpu));
1101 static __init void init_hrtick(void)
1103 hotcpu_notifier(hotplug_hrtick, 0);
1107 * Called to set the hrtick timer state.
1109 * called with rq->lock held and irqs disabled
1111 static void hrtick_start(struct rq *rq, u64 delay)
1113 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1114 HRTIMER_MODE_REL_PINNED, 0);
1117 static inline void init_hrtick(void)
1120 #endif /* CONFIG_SMP */
1122 static void init_rq_hrtick(struct rq *rq)
1125 rq->hrtick_csd_pending = 0;
1127 rq->hrtick_csd.flags = 0;
1128 rq->hrtick_csd.func = __hrtick_start;
1129 rq->hrtick_csd.info = rq;
1132 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1133 rq->hrtick_timer.function = hrtick;
1135 #else /* CONFIG_SCHED_HRTICK */
1136 static inline void hrtick_clear(struct rq *rq)
1140 static inline void init_rq_hrtick(struct rq *rq)
1144 static inline void init_hrtick(void)
1147 #endif /* CONFIG_SCHED_HRTICK */
1150 * resched_task - mark a task 'to be rescheduled now'.
1152 * On UP this means the setting of the need_resched flag, on SMP it
1153 * might also involve a cross-CPU call to trigger the scheduler on
1158 #ifndef tsk_is_polling
1159 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1162 static void resched_task(struct task_struct *p)
1166 assert_raw_spin_locked(&task_rq(p)->lock);
1168 if (test_tsk_need_resched(p))
1171 set_tsk_need_resched(p);
1174 if (cpu == smp_processor_id())
1177 /* NEED_RESCHED must be visible before we test polling */
1179 if (!tsk_is_polling(p))
1180 smp_send_reschedule(cpu);
1183 static void resched_cpu(int cpu)
1185 struct rq *rq = cpu_rq(cpu);
1186 unsigned long flags;
1188 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1190 resched_task(cpu_curr(cpu));
1191 raw_spin_unlock_irqrestore(&rq->lock, flags);
1196 * When add_timer_on() enqueues a timer into the timer wheel of an
1197 * idle CPU then this timer might expire before the next timer event
1198 * which is scheduled to wake up that CPU. In case of a completely
1199 * idle system the next event might even be infinite time into the
1200 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1201 * leaves the inner idle loop so the newly added timer is taken into
1202 * account when the CPU goes back to idle and evaluates the timer
1203 * wheel for the next timer event.
1205 void wake_up_idle_cpu(int cpu)
1207 struct rq *rq = cpu_rq(cpu);
1209 if (cpu == smp_processor_id())
1213 * This is safe, as this function is called with the timer
1214 * wheel base lock of (cpu) held. When the CPU is on the way
1215 * to idle and has not yet set rq->curr to idle then it will
1216 * be serialized on the timer wheel base lock and take the new
1217 * timer into account automatically.
1219 if (rq->curr != rq->idle)
1223 * We can set TIF_RESCHED on the idle task of the other CPU
1224 * lockless. The worst case is that the other CPU runs the
1225 * idle task through an additional NOOP schedule()
1227 set_tsk_need_resched(rq->idle);
1229 /* NEED_RESCHED must be visible before we test polling */
1231 if (!tsk_is_polling(rq->idle))
1232 smp_send_reschedule(cpu);
1235 int nohz_ratelimit(int cpu)
1237 struct rq *rq = cpu_rq(cpu);
1238 u64 diff = rq->clock - rq->nohz_stamp;
1240 rq->nohz_stamp = rq->clock;
1242 return diff < (NSEC_PER_SEC / HZ) >> 1;
1245 #endif /* CONFIG_NO_HZ */
1247 static u64 sched_avg_period(void)
1249 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1252 static void sched_avg_update(struct rq *rq)
1254 s64 period = sched_avg_period();
1256 while ((s64)(rq->clock - rq->age_stamp) > period) {
1257 rq->age_stamp += period;
1262 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1264 rq->rt_avg += rt_delta;
1265 sched_avg_update(rq);
1268 #else /* !CONFIG_SMP */
1269 static void resched_task(struct task_struct *p)
1271 assert_raw_spin_locked(&task_rq(p)->lock);
1272 set_tsk_need_resched(p);
1275 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1278 #endif /* CONFIG_SMP */
1280 #if BITS_PER_LONG == 32
1281 # define WMULT_CONST (~0UL)
1283 # define WMULT_CONST (1UL << 32)
1286 #define WMULT_SHIFT 32
1289 * Shift right and round:
1291 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1294 * delta *= weight / lw
1296 static unsigned long
1297 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1298 struct load_weight *lw)
1302 if (!lw->inv_weight) {
1303 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1306 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1310 tmp = (u64)delta_exec * weight;
1312 * Check whether we'd overflow the 64-bit multiplication:
1314 if (unlikely(tmp > WMULT_CONST))
1315 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1318 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1320 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1323 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1329 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1336 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1337 * of tasks with abnormal "nice" values across CPUs the contribution that
1338 * each task makes to its run queue's load is weighted according to its
1339 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1340 * scaled version of the new time slice allocation that they receive on time
1344 #define WEIGHT_IDLEPRIO 3
1345 #define WMULT_IDLEPRIO 1431655765
1348 * Nice levels are multiplicative, with a gentle 10% change for every
1349 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1350 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1351 * that remained on nice 0.
1353 * The "10% effect" is relative and cumulative: from _any_ nice level,
1354 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1355 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1356 * If a task goes up by ~10% and another task goes down by ~10% then
1357 * the relative distance between them is ~25%.)
1359 static const int prio_to_weight[40] = {
1360 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1361 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1362 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1363 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1364 /* 0 */ 1024, 820, 655, 526, 423,
1365 /* 5 */ 335, 272, 215, 172, 137,
1366 /* 10 */ 110, 87, 70, 56, 45,
1367 /* 15 */ 36, 29, 23, 18, 15,
1371 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1373 * In cases where the weight does not change often, we can use the
1374 * precalculated inverse to speed up arithmetics by turning divisions
1375 * into multiplications:
1377 static const u32 prio_to_wmult[40] = {
1378 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1379 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1380 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1381 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1382 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1383 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1384 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1385 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1388 /* Time spent by the tasks of the cpu accounting group executing in ... */
1389 enum cpuacct_stat_index {
1390 CPUACCT_STAT_USER, /* ... user mode */
1391 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1393 CPUACCT_STAT_NSTATS,
1396 #ifdef CONFIG_CGROUP_CPUACCT
1397 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1398 static void cpuacct_update_stats(struct task_struct *tsk,
1399 enum cpuacct_stat_index idx, cputime_t val);
1401 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1402 static inline void cpuacct_update_stats(struct task_struct *tsk,
1403 enum cpuacct_stat_index idx, cputime_t val) {}
1406 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1408 update_load_add(&rq->load, load);
1411 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1413 update_load_sub(&rq->load, load);
1416 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1417 typedef int (*tg_visitor)(struct task_group *, void *);
1420 * Iterate the full tree, calling @down when first entering a node and @up when
1421 * leaving it for the final time.
1423 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1425 struct task_group *parent, *child;
1429 parent = &root_task_group;
1431 ret = (*down)(parent, data);
1434 list_for_each_entry_rcu(child, &parent->children, siblings) {
1441 ret = (*up)(parent, data);
1446 parent = parent->parent;
1455 static int tg_nop(struct task_group *tg, void *data)
1462 /* Used instead of source_load when we know the type == 0 */
1463 static unsigned long weighted_cpuload(const int cpu)
1465 return cpu_rq(cpu)->load.weight;
1469 * Return a low guess at the load of a migration-source cpu weighted
1470 * according to the scheduling class and "nice" value.
1472 * We want to under-estimate the load of migration sources, to
1473 * balance conservatively.
1475 static unsigned long source_load(int cpu, int type)
1477 struct rq *rq = cpu_rq(cpu);
1478 unsigned long total = weighted_cpuload(cpu);
1480 if (type == 0 || !sched_feat(LB_BIAS))
1483 return min(rq->cpu_load[type-1], total);
1487 * Return a high guess at the load of a migration-target cpu weighted
1488 * according to the scheduling class and "nice" value.
1490 static unsigned long target_load(int cpu, int type)
1492 struct rq *rq = cpu_rq(cpu);
1493 unsigned long total = weighted_cpuload(cpu);
1495 if (type == 0 || !sched_feat(LB_BIAS))
1498 return max(rq->cpu_load[type-1], total);
1501 static struct sched_group *group_of(int cpu)
1503 struct sched_domain *sd = rcu_dereference_sched(cpu_rq(cpu)->sd);
1511 static unsigned long power_of(int cpu)
1513 struct sched_group *group = group_of(cpu);
1516 return SCHED_LOAD_SCALE;
1518 return group->cpu_power;
1521 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1523 static unsigned long cpu_avg_load_per_task(int cpu)
1525 struct rq *rq = cpu_rq(cpu);
1526 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1529 rq->avg_load_per_task = rq->load.weight / nr_running;
1531 rq->avg_load_per_task = 0;
1533 return rq->avg_load_per_task;
1536 #ifdef CONFIG_FAIR_GROUP_SCHED
1538 static __read_mostly unsigned long __percpu *update_shares_data;
1540 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1543 * Calculate and set the cpu's group shares.
1545 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1546 unsigned long sd_shares,
1547 unsigned long sd_rq_weight,
1548 unsigned long *usd_rq_weight)
1550 unsigned long shares, rq_weight;
1553 rq_weight = usd_rq_weight[cpu];
1556 rq_weight = NICE_0_LOAD;
1560 * \Sum_j shares_j * rq_weight_i
1561 * shares_i = -----------------------------
1562 * \Sum_j rq_weight_j
1564 shares = (sd_shares * rq_weight) / sd_rq_weight;
1565 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1567 if (abs(shares - tg->se[cpu]->load.weight) >
1568 sysctl_sched_shares_thresh) {
1569 struct rq *rq = cpu_rq(cpu);
1570 unsigned long flags;
1572 raw_spin_lock_irqsave(&rq->lock, flags);
1573 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1574 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1575 __set_se_shares(tg->se[cpu], shares);
1576 raw_spin_unlock_irqrestore(&rq->lock, flags);
1581 * Re-compute the task group their per cpu shares over the given domain.
1582 * This needs to be done in a bottom-up fashion because the rq weight of a
1583 * parent group depends on the shares of its child groups.
1585 static int tg_shares_up(struct task_group *tg, void *data)
1587 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1588 unsigned long *usd_rq_weight;
1589 struct sched_domain *sd = data;
1590 unsigned long flags;
1596 local_irq_save(flags);
1597 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1599 for_each_cpu(i, sched_domain_span(sd)) {
1600 weight = tg->cfs_rq[i]->load.weight;
1601 usd_rq_weight[i] = weight;
1603 rq_weight += weight;
1605 * If there are currently no tasks on the cpu pretend there
1606 * is one of average load so that when a new task gets to
1607 * run here it will not get delayed by group starvation.
1610 weight = NICE_0_LOAD;
1612 sum_weight += weight;
1613 shares += tg->cfs_rq[i]->shares;
1617 rq_weight = sum_weight;
1619 if ((!shares && rq_weight) || shares > tg->shares)
1620 shares = tg->shares;
1622 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1623 shares = tg->shares;
1625 for_each_cpu(i, sched_domain_span(sd))
1626 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1628 local_irq_restore(flags);
1634 * Compute the cpu's hierarchical load factor for each task group.
1635 * This needs to be done in a top-down fashion because the load of a child
1636 * group is a fraction of its parents load.
1638 static int tg_load_down(struct task_group *tg, void *data)
1641 long cpu = (long)data;
1644 load = cpu_rq(cpu)->load.weight;
1646 load = tg->parent->cfs_rq[cpu]->h_load;
1647 load *= tg->cfs_rq[cpu]->shares;
1648 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1651 tg->cfs_rq[cpu]->h_load = load;
1656 static void update_shares(struct sched_domain *sd)
1661 if (root_task_group_empty())
1664 now = cpu_clock(raw_smp_processor_id());
1665 elapsed = now - sd->last_update;
1667 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1668 sd->last_update = now;
1669 walk_tg_tree(tg_nop, tg_shares_up, sd);
1673 static void update_h_load(long cpu)
1675 if (root_task_group_empty())
1678 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1683 static inline void update_shares(struct sched_domain *sd)
1689 #ifdef CONFIG_PREEMPT
1691 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1694 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1695 * way at the expense of forcing extra atomic operations in all
1696 * invocations. This assures that the double_lock is acquired using the
1697 * same underlying policy as the spinlock_t on this architecture, which
1698 * reduces latency compared to the unfair variant below. However, it
1699 * also adds more overhead and therefore may reduce throughput.
1701 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1702 __releases(this_rq->lock)
1703 __acquires(busiest->lock)
1704 __acquires(this_rq->lock)
1706 raw_spin_unlock(&this_rq->lock);
1707 double_rq_lock(this_rq, busiest);
1714 * Unfair double_lock_balance: Optimizes throughput at the expense of
1715 * latency by eliminating extra atomic operations when the locks are
1716 * already in proper order on entry. This favors lower cpu-ids and will
1717 * grant the double lock to lower cpus over higher ids under contention,
1718 * regardless of entry order into the function.
1720 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1721 __releases(this_rq->lock)
1722 __acquires(busiest->lock)
1723 __acquires(this_rq->lock)
1727 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1728 if (busiest < this_rq) {
1729 raw_spin_unlock(&this_rq->lock);
1730 raw_spin_lock(&busiest->lock);
1731 raw_spin_lock_nested(&this_rq->lock,
1732 SINGLE_DEPTH_NESTING);
1735 raw_spin_lock_nested(&busiest->lock,
1736 SINGLE_DEPTH_NESTING);
1741 #endif /* CONFIG_PREEMPT */
1744 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1746 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1748 if (unlikely(!irqs_disabled())) {
1749 /* printk() doesn't work good under rq->lock */
1750 raw_spin_unlock(&this_rq->lock);
1754 return _double_lock_balance(this_rq, busiest);
1757 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1758 __releases(busiest->lock)
1760 raw_spin_unlock(&busiest->lock);
1761 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1765 * double_rq_lock - safely lock two runqueues
1767 * Note this does not disable interrupts like task_rq_lock,
1768 * you need to do so manually before calling.
1770 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1771 __acquires(rq1->lock)
1772 __acquires(rq2->lock)
1774 BUG_ON(!irqs_disabled());
1776 raw_spin_lock(&rq1->lock);
1777 __acquire(rq2->lock); /* Fake it out ;) */
1780 raw_spin_lock(&rq1->lock);
1781 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1783 raw_spin_lock(&rq2->lock);
1784 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1790 * double_rq_unlock - safely unlock two runqueues
1792 * Note this does not restore interrupts like task_rq_unlock,
1793 * you need to do so manually after calling.
1795 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1796 __releases(rq1->lock)
1797 __releases(rq2->lock)
1799 raw_spin_unlock(&rq1->lock);
1801 raw_spin_unlock(&rq2->lock);
1803 __release(rq2->lock);
1808 #ifdef CONFIG_FAIR_GROUP_SCHED
1809 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1812 cfs_rq->shares = shares;
1817 static void calc_load_account_active(struct rq *this_rq);
1818 static void update_sysctl(void);
1819 static int get_update_sysctl_factor(void);
1821 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1823 set_task_rq(p, cpu);
1826 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1827 * successfuly executed on another CPU. We must ensure that updates of
1828 * per-task data have been completed by this moment.
1831 task_thread_info(p)->cpu = cpu;
1835 static const struct sched_class rt_sched_class;
1837 #define sched_class_highest (&rt_sched_class)
1838 #define for_each_class(class) \
1839 for (class = sched_class_highest; class; class = class->next)
1841 #include "sched_stats.h"
1843 static void inc_nr_running(struct rq *rq)
1848 static void dec_nr_running(struct rq *rq)
1853 static void set_load_weight(struct task_struct *p)
1855 if (task_has_rt_policy(p)) {
1856 p->se.load.weight = prio_to_weight[0] * 2;
1857 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1862 * SCHED_IDLE tasks get minimal weight:
1864 if (p->policy == SCHED_IDLE) {
1865 p->se.load.weight = WEIGHT_IDLEPRIO;
1866 p->se.load.inv_weight = WMULT_IDLEPRIO;
1870 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1871 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1874 static void update_avg(u64 *avg, u64 sample)
1876 s64 diff = sample - *avg;
1881 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1883 update_rq_clock(rq);
1884 sched_info_queued(p);
1885 p->sched_class->enqueue_task(rq, p, wakeup, head);
1889 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1891 update_rq_clock(rq);
1892 sched_info_dequeued(p);
1893 p->sched_class->dequeue_task(rq, p, sleep);
1898 * activate_task - move a task to the runqueue.
1900 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1902 if (task_contributes_to_load(p))
1903 rq->nr_uninterruptible--;
1905 enqueue_task(rq, p, wakeup, false);
1910 * deactivate_task - remove a task from the runqueue.
1912 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1914 if (task_contributes_to_load(p))
1915 rq->nr_uninterruptible++;
1917 dequeue_task(rq, p, sleep);
1921 #include "sched_idletask.c"
1922 #include "sched_fair.c"
1923 #include "sched_rt.c"
1924 #ifdef CONFIG_SCHED_DEBUG
1925 # include "sched_debug.c"
1929 * __normal_prio - return the priority that is based on the static prio
1931 static inline int __normal_prio(struct task_struct *p)
1933 return p->static_prio;
1937 * Calculate the expected normal priority: i.e. priority
1938 * without taking RT-inheritance into account. Might be
1939 * boosted by interactivity modifiers. Changes upon fork,
1940 * setprio syscalls, and whenever the interactivity
1941 * estimator recalculates.
1943 static inline int normal_prio(struct task_struct *p)
1947 if (task_has_rt_policy(p))
1948 prio = MAX_RT_PRIO-1 - p->rt_priority;
1950 prio = __normal_prio(p);
1955 * Calculate the current priority, i.e. the priority
1956 * taken into account by the scheduler. This value might
1957 * be boosted by RT tasks, or might be boosted by
1958 * interactivity modifiers. Will be RT if the task got
1959 * RT-boosted. If not then it returns p->normal_prio.
1961 static int effective_prio(struct task_struct *p)
1963 p->normal_prio = normal_prio(p);
1965 * If we are RT tasks or we were boosted to RT priority,
1966 * keep the priority unchanged. Otherwise, update priority
1967 * to the normal priority:
1969 if (!rt_prio(p->prio))
1970 return p->normal_prio;
1975 * task_curr - is this task currently executing on a CPU?
1976 * @p: the task in question.
1978 inline int task_curr(const struct task_struct *p)
1980 return cpu_curr(task_cpu(p)) == p;
1983 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1984 const struct sched_class *prev_class,
1985 int oldprio, int running)
1987 if (prev_class != p->sched_class) {
1988 if (prev_class->switched_from)
1989 prev_class->switched_from(rq, p, running);
1990 p->sched_class->switched_to(rq, p, running);
1992 p->sched_class->prio_changed(rq, p, oldprio, running);
1997 * Is this task likely cache-hot:
2000 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2004 if (p->sched_class != &fair_sched_class)
2008 * Buddy candidates are cache hot:
2010 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2011 (&p->se == cfs_rq_of(&p->se)->next ||
2012 &p->se == cfs_rq_of(&p->se)->last))
2015 if (sysctl_sched_migration_cost == -1)
2017 if (sysctl_sched_migration_cost == 0)
2020 delta = now - p->se.exec_start;
2022 return delta < (s64)sysctl_sched_migration_cost;
2025 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2027 #ifdef CONFIG_SCHED_DEBUG
2029 * We should never call set_task_cpu() on a blocked task,
2030 * ttwu() will sort out the placement.
2032 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2033 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2036 trace_sched_migrate_task(p, new_cpu);
2038 if (task_cpu(p) != new_cpu) {
2039 p->se.nr_migrations++;
2040 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2043 __set_task_cpu(p, new_cpu);
2046 struct migration_req {
2047 struct list_head list;
2049 struct task_struct *task;
2052 struct completion done;
2056 * The task's runqueue lock must be held.
2057 * Returns true if you have to wait for migration thread.
2060 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2062 struct rq *rq = task_rq(p);
2065 * If the task is not on a runqueue (and not running), then
2066 * the next wake-up will properly place the task.
2068 if (!p->se.on_rq && !task_running(rq, p))
2071 init_completion(&req->done);
2073 req->dest_cpu = dest_cpu;
2074 list_add(&req->list, &rq->migration_queue);
2080 * wait_task_context_switch - wait for a thread to complete at least one
2083 * @p must not be current.
2085 void wait_task_context_switch(struct task_struct *p)
2087 unsigned long nvcsw, nivcsw, flags;
2095 * The runqueue is assigned before the actual context
2096 * switch. We need to take the runqueue lock.
2098 * We could check initially without the lock but it is
2099 * very likely that we need to take the lock in every
2102 rq = task_rq_lock(p, &flags);
2103 running = task_running(rq, p);
2104 task_rq_unlock(rq, &flags);
2106 if (likely(!running))
2109 * The switch count is incremented before the actual
2110 * context switch. We thus wait for two switches to be
2111 * sure at least one completed.
2113 if ((p->nvcsw - nvcsw) > 1)
2115 if ((p->nivcsw - nivcsw) > 1)
2123 * wait_task_inactive - wait for a thread to unschedule.
2125 * If @match_state is nonzero, it's the @p->state value just checked and
2126 * not expected to change. If it changes, i.e. @p might have woken up,
2127 * then return zero. When we succeed in waiting for @p to be off its CPU,
2128 * we return a positive number (its total switch count). If a second call
2129 * a short while later returns the same number, the caller can be sure that
2130 * @p has remained unscheduled the whole time.
2132 * The caller must ensure that the task *will* unschedule sometime soon,
2133 * else this function might spin for a *long* time. This function can't
2134 * be called with interrupts off, or it may introduce deadlock with
2135 * smp_call_function() if an IPI is sent by the same process we are
2136 * waiting to become inactive.
2138 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2140 unsigned long flags;
2147 * We do the initial early heuristics without holding
2148 * any task-queue locks at all. We'll only try to get
2149 * the runqueue lock when things look like they will
2155 * If the task is actively running on another CPU
2156 * still, just relax and busy-wait without holding
2159 * NOTE! Since we don't hold any locks, it's not
2160 * even sure that "rq" stays as the right runqueue!
2161 * But we don't care, since "task_running()" will
2162 * return false if the runqueue has changed and p
2163 * is actually now running somewhere else!
2165 while (task_running(rq, p)) {
2166 if (match_state && unlikely(p->state != match_state))
2172 * Ok, time to look more closely! We need the rq
2173 * lock now, to be *sure*. If we're wrong, we'll
2174 * just go back and repeat.
2176 rq = task_rq_lock(p, &flags);
2177 trace_sched_wait_task(rq, p);
2178 running = task_running(rq, p);
2179 on_rq = p->se.on_rq;
2181 if (!match_state || p->state == match_state)
2182 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2183 task_rq_unlock(rq, &flags);
2186 * If it changed from the expected state, bail out now.
2188 if (unlikely(!ncsw))
2192 * Was it really running after all now that we
2193 * checked with the proper locks actually held?
2195 * Oops. Go back and try again..
2197 if (unlikely(running)) {
2203 * It's not enough that it's not actively running,
2204 * it must be off the runqueue _entirely_, and not
2207 * So if it was still runnable (but just not actively
2208 * running right now), it's preempted, and we should
2209 * yield - it could be a while.
2211 if (unlikely(on_rq)) {
2212 schedule_timeout_uninterruptible(1);
2217 * Ahh, all good. It wasn't running, and it wasn't
2218 * runnable, which means that it will never become
2219 * running in the future either. We're all done!
2228 * kick_process - kick a running thread to enter/exit the kernel
2229 * @p: the to-be-kicked thread
2231 * Cause a process which is running on another CPU to enter
2232 * kernel-mode, without any delay. (to get signals handled.)
2234 * NOTE: this function doesnt have to take the runqueue lock,
2235 * because all it wants to ensure is that the remote task enters
2236 * the kernel. If the IPI races and the task has been migrated
2237 * to another CPU then no harm is done and the purpose has been
2240 void kick_process(struct task_struct *p)
2246 if ((cpu != smp_processor_id()) && task_curr(p))
2247 smp_send_reschedule(cpu);
2250 EXPORT_SYMBOL_GPL(kick_process);
2251 #endif /* CONFIG_SMP */
2254 * task_oncpu_function_call - call a function on the cpu on which a task runs
2255 * @p: the task to evaluate
2256 * @func: the function to be called
2257 * @info: the function call argument
2259 * Calls the function @func when the task is currently running. This might
2260 * be on the current CPU, which just calls the function directly
2262 void task_oncpu_function_call(struct task_struct *p,
2263 void (*func) (void *info), void *info)
2270 smp_call_function_single(cpu, func, info, 1);
2276 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2278 static int select_fallback_rq(int cpu, struct task_struct *p)
2281 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2283 /* Look for allowed, online CPU in same node. */
2284 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2285 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2288 /* Any allowed, online CPU? */
2289 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2290 if (dest_cpu < nr_cpu_ids)
2293 /* No more Mr. Nice Guy. */
2294 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2295 dest_cpu = cpuset_cpus_allowed_fallback(p);
2297 * Don't tell them about moving exiting tasks or
2298 * kernel threads (both mm NULL), since they never
2301 if (p->mm && printk_ratelimit()) {
2302 printk(KERN_INFO "process %d (%s) no "
2303 "longer affine to cpu%d\n",
2304 task_pid_nr(p), p->comm, cpu);
2312 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2315 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2317 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2320 * In order not to call set_task_cpu() on a blocking task we need
2321 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2324 * Since this is common to all placement strategies, this lives here.
2326 * [ this allows ->select_task() to simply return task_cpu(p) and
2327 * not worry about this generic constraint ]
2329 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2331 cpu = select_fallback_rq(task_cpu(p), p);
2338 * try_to_wake_up - wake up a thread
2339 * @p: the to-be-woken-up thread
2340 * @state: the mask of task states that can be woken
2341 * @sync: do a synchronous wakeup?
2343 * Put it on the run-queue if it's not already there. The "current"
2344 * thread is always on the run-queue (except when the actual
2345 * re-schedule is in progress), and as such you're allowed to do
2346 * the simpler "current->state = TASK_RUNNING" to mark yourself
2347 * runnable without the overhead of this.
2349 * returns failure only if the task is already active.
2351 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2354 int cpu, orig_cpu, this_cpu, success = 0;
2355 unsigned long flags;
2358 this_cpu = get_cpu();
2361 rq = task_rq_lock(p, &flags);
2362 if (!(p->state & state))
2372 if (unlikely(task_running(rq, p)))
2376 * In order to handle concurrent wakeups and release the rq->lock
2377 * we put the task in TASK_WAKING state.
2379 * First fix up the nr_uninterruptible count:
2381 if (task_contributes_to_load(p))
2382 rq->nr_uninterruptible--;
2383 p->state = TASK_WAKING;
2385 if (p->sched_class->task_waking)
2386 p->sched_class->task_waking(rq, p);
2388 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2389 if (cpu != orig_cpu)
2390 set_task_cpu(p, cpu);
2391 __task_rq_unlock(rq);
2394 raw_spin_lock(&rq->lock);
2397 * We migrated the task without holding either rq->lock, however
2398 * since the task is not on the task list itself, nobody else
2399 * will try and migrate the task, hence the rq should match the
2400 * cpu we just moved it to.
2402 WARN_ON(task_cpu(p) != cpu);
2403 WARN_ON(p->state != TASK_WAKING);
2405 #ifdef CONFIG_SCHEDSTATS
2406 schedstat_inc(rq, ttwu_count);
2407 if (cpu == this_cpu)
2408 schedstat_inc(rq, ttwu_local);
2410 struct sched_domain *sd;
2411 for_each_domain(this_cpu, sd) {
2412 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2413 schedstat_inc(sd, ttwu_wake_remote);
2418 #endif /* CONFIG_SCHEDSTATS */
2421 #endif /* CONFIG_SMP */
2422 schedstat_inc(p, se.statistics.nr_wakeups);
2423 if (wake_flags & WF_SYNC)
2424 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2425 if (orig_cpu != cpu)
2426 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2427 if (cpu == this_cpu)
2428 schedstat_inc(p, se.statistics.nr_wakeups_local);
2430 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2431 activate_task(rq, p, 1);
2435 trace_sched_wakeup(rq, p, success);
2436 check_preempt_curr(rq, p, wake_flags);
2438 p->state = TASK_RUNNING;
2440 if (p->sched_class->task_woken)
2441 p->sched_class->task_woken(rq, p);
2443 if (unlikely(rq->idle_stamp)) {
2444 u64 delta = rq->clock - rq->idle_stamp;
2445 u64 max = 2*sysctl_sched_migration_cost;
2450 update_avg(&rq->avg_idle, delta);
2455 task_rq_unlock(rq, &flags);
2462 * wake_up_process - Wake up a specific process
2463 * @p: The process to be woken up.
2465 * Attempt to wake up the nominated process and move it to the set of runnable
2466 * processes. Returns 1 if the process was woken up, 0 if it was already
2469 * It may be assumed that this function implies a write memory barrier before
2470 * changing the task state if and only if any tasks are woken up.
2472 int wake_up_process(struct task_struct *p)
2474 return try_to_wake_up(p, TASK_ALL, 0);
2476 EXPORT_SYMBOL(wake_up_process);
2478 int wake_up_state(struct task_struct *p, unsigned int state)
2480 return try_to_wake_up(p, state, 0);
2484 * Perform scheduler related setup for a newly forked process p.
2485 * p is forked by current.
2487 * __sched_fork() is basic setup used by init_idle() too:
2489 static void __sched_fork(struct task_struct *p)
2491 p->se.exec_start = 0;
2492 p->se.sum_exec_runtime = 0;
2493 p->se.prev_sum_exec_runtime = 0;
2494 p->se.nr_migrations = 0;
2496 #ifdef CONFIG_SCHEDSTATS
2497 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2500 INIT_LIST_HEAD(&p->rt.run_list);
2502 INIT_LIST_HEAD(&p->se.group_node);
2504 #ifdef CONFIG_PREEMPT_NOTIFIERS
2505 INIT_HLIST_HEAD(&p->preempt_notifiers);
2510 * fork()/clone()-time setup:
2512 void sched_fork(struct task_struct *p, int clone_flags)
2514 int cpu = get_cpu();
2518 * We mark the process as running here. This guarantees that
2519 * nobody will actually run it, and a signal or other external
2520 * event cannot wake it up and insert it on the runqueue either.
2522 p->state = TASK_RUNNING;
2525 * Revert to default priority/policy on fork if requested.
2527 if (unlikely(p->sched_reset_on_fork)) {
2528 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2529 p->policy = SCHED_NORMAL;
2530 p->normal_prio = p->static_prio;
2533 if (PRIO_TO_NICE(p->static_prio) < 0) {
2534 p->static_prio = NICE_TO_PRIO(0);
2535 p->normal_prio = p->static_prio;
2540 * We don't need the reset flag anymore after the fork. It has
2541 * fulfilled its duty:
2543 p->sched_reset_on_fork = 0;
2547 * Make sure we do not leak PI boosting priority to the child.
2549 p->prio = current->normal_prio;
2551 if (!rt_prio(p->prio))
2552 p->sched_class = &fair_sched_class;
2554 if (p->sched_class->task_fork)
2555 p->sched_class->task_fork(p);
2557 set_task_cpu(p, cpu);
2559 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2560 if (likely(sched_info_on()))
2561 memset(&p->sched_info, 0, sizeof(p->sched_info));
2563 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2566 #ifdef CONFIG_PREEMPT
2567 /* Want to start with kernel preemption disabled. */
2568 task_thread_info(p)->preempt_count = 1;
2570 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2576 * wake_up_new_task - wake up a newly created task for the first time.
2578 * This function will do some initial scheduler statistics housekeeping
2579 * that must be done for every newly created context, then puts the task
2580 * on the runqueue and wakes it.
2582 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2584 unsigned long flags;
2586 int cpu __maybe_unused = get_cpu();
2589 rq = task_rq_lock(p, &flags);
2590 p->state = TASK_WAKING;
2593 * Fork balancing, do it here and not earlier because:
2594 * - cpus_allowed can change in the fork path
2595 * - any previously selected cpu might disappear through hotplug
2597 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2598 * without people poking at ->cpus_allowed.
2600 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2601 set_task_cpu(p, cpu);
2603 p->state = TASK_RUNNING;
2604 task_rq_unlock(rq, &flags);
2607 rq = task_rq_lock(p, &flags);
2608 activate_task(rq, p, 0);
2609 trace_sched_wakeup_new(rq, p, 1);
2610 check_preempt_curr(rq, p, WF_FORK);
2612 if (p->sched_class->task_woken)
2613 p->sched_class->task_woken(rq, p);
2615 task_rq_unlock(rq, &flags);
2619 #ifdef CONFIG_PREEMPT_NOTIFIERS
2622 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2623 * @notifier: notifier struct to register
2625 void preempt_notifier_register(struct preempt_notifier *notifier)
2627 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2629 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2632 * preempt_notifier_unregister - no longer interested in preemption notifications
2633 * @notifier: notifier struct to unregister
2635 * This is safe to call from within a preemption notifier.
2637 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2639 hlist_del(¬ifier->link);
2641 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2643 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2645 struct preempt_notifier *notifier;
2646 struct hlist_node *node;
2648 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2649 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2653 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2654 struct task_struct *next)
2656 struct preempt_notifier *notifier;
2657 struct hlist_node *node;
2659 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2660 notifier->ops->sched_out(notifier, next);
2663 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2665 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2670 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2671 struct task_struct *next)
2675 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2678 * prepare_task_switch - prepare to switch tasks
2679 * @rq: the runqueue preparing to switch
2680 * @prev: the current task that is being switched out
2681 * @next: the task we are going to switch to.
2683 * This is called with the rq lock held and interrupts off. It must
2684 * be paired with a subsequent finish_task_switch after the context
2687 * prepare_task_switch sets up locking and calls architecture specific
2691 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2692 struct task_struct *next)
2694 fire_sched_out_preempt_notifiers(prev, next);
2695 prepare_lock_switch(rq, next);
2696 prepare_arch_switch(next);
2700 * finish_task_switch - clean up after a task-switch
2701 * @rq: runqueue associated with task-switch
2702 * @prev: the thread we just switched away from.
2704 * finish_task_switch must be called after the context switch, paired
2705 * with a prepare_task_switch call before the context switch.
2706 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2707 * and do any other architecture-specific cleanup actions.
2709 * Note that we may have delayed dropping an mm in context_switch(). If
2710 * so, we finish that here outside of the runqueue lock. (Doing it
2711 * with the lock held can cause deadlocks; see schedule() for
2714 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2715 __releases(rq->lock)
2717 struct mm_struct *mm = rq->prev_mm;
2723 * A task struct has one reference for the use as "current".
2724 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2725 * schedule one last time. The schedule call will never return, and
2726 * the scheduled task must drop that reference.
2727 * The test for TASK_DEAD must occur while the runqueue locks are
2728 * still held, otherwise prev could be scheduled on another cpu, die
2729 * there before we look at prev->state, and then the reference would
2731 * Manfred Spraul <manfred@colorfullife.com>
2733 prev_state = prev->state;
2734 finish_arch_switch(prev);
2735 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2736 local_irq_disable();
2737 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2738 perf_event_task_sched_in(current);
2739 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2741 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2742 finish_lock_switch(rq, prev);
2744 fire_sched_in_preempt_notifiers(current);
2747 if (unlikely(prev_state == TASK_DEAD)) {
2749 * Remove function-return probe instances associated with this
2750 * task and put them back on the free list.
2752 kprobe_flush_task(prev);
2753 put_task_struct(prev);
2759 /* assumes rq->lock is held */
2760 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2762 if (prev->sched_class->pre_schedule)
2763 prev->sched_class->pre_schedule(rq, prev);
2766 /* rq->lock is NOT held, but preemption is disabled */
2767 static inline void post_schedule(struct rq *rq)
2769 if (rq->post_schedule) {
2770 unsigned long flags;
2772 raw_spin_lock_irqsave(&rq->lock, flags);
2773 if (rq->curr->sched_class->post_schedule)
2774 rq->curr->sched_class->post_schedule(rq);
2775 raw_spin_unlock_irqrestore(&rq->lock, flags);
2777 rq->post_schedule = 0;
2783 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2787 static inline void post_schedule(struct rq *rq)
2794 * schedule_tail - first thing a freshly forked thread must call.
2795 * @prev: the thread we just switched away from.
2797 asmlinkage void schedule_tail(struct task_struct *prev)
2798 __releases(rq->lock)
2800 struct rq *rq = this_rq();
2802 finish_task_switch(rq, prev);
2805 * FIXME: do we need to worry about rq being invalidated by the
2810 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2811 /* In this case, finish_task_switch does not reenable preemption */
2814 if (current->set_child_tid)
2815 put_user(task_pid_vnr(current), current->set_child_tid);
2819 * context_switch - switch to the new MM and the new
2820 * thread's register state.
2823 context_switch(struct rq *rq, struct task_struct *prev,
2824 struct task_struct *next)
2826 struct mm_struct *mm, *oldmm;
2828 prepare_task_switch(rq, prev, next);
2829 trace_sched_switch(rq, prev, next);
2831 oldmm = prev->active_mm;
2833 * For paravirt, this is coupled with an exit in switch_to to
2834 * combine the page table reload and the switch backend into
2837 arch_start_context_switch(prev);
2840 next->active_mm = oldmm;
2841 atomic_inc(&oldmm->mm_count);
2842 enter_lazy_tlb(oldmm, next);
2844 switch_mm(oldmm, mm, next);
2846 if (likely(!prev->mm)) {
2847 prev->active_mm = NULL;
2848 rq->prev_mm = oldmm;
2851 * Since the runqueue lock will be released by the next
2852 * task (which is an invalid locking op but in the case
2853 * of the scheduler it's an obvious special-case), so we
2854 * do an early lockdep release here:
2856 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2857 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2860 /* Here we just switch the register state and the stack. */
2861 switch_to(prev, next, prev);
2865 * this_rq must be evaluated again because prev may have moved
2866 * CPUs since it called schedule(), thus the 'rq' on its stack
2867 * frame will be invalid.
2869 finish_task_switch(this_rq(), prev);
2873 * nr_running, nr_uninterruptible and nr_context_switches:
2875 * externally visible scheduler statistics: current number of runnable
2876 * threads, current number of uninterruptible-sleeping threads, total
2877 * number of context switches performed since bootup.
2879 unsigned long nr_running(void)
2881 unsigned long i, sum = 0;
2883 for_each_online_cpu(i)
2884 sum += cpu_rq(i)->nr_running;
2889 unsigned long nr_uninterruptible(void)
2891 unsigned long i, sum = 0;
2893 for_each_possible_cpu(i)
2894 sum += cpu_rq(i)->nr_uninterruptible;
2897 * Since we read the counters lockless, it might be slightly
2898 * inaccurate. Do not allow it to go below zero though:
2900 if (unlikely((long)sum < 0))
2906 unsigned long long nr_context_switches(void)
2909 unsigned long long sum = 0;
2911 for_each_possible_cpu(i)
2912 sum += cpu_rq(i)->nr_switches;
2917 unsigned long nr_iowait(void)
2919 unsigned long i, sum = 0;
2921 for_each_possible_cpu(i)
2922 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2927 unsigned long nr_iowait_cpu(void)
2929 struct rq *this = this_rq();
2930 return atomic_read(&this->nr_iowait);
2933 unsigned long this_cpu_load(void)
2935 struct rq *this = this_rq();
2936 return this->cpu_load[0];
2940 /* Variables and functions for calc_load */
2941 static atomic_long_t calc_load_tasks;
2942 static unsigned long calc_load_update;
2943 unsigned long avenrun[3];
2944 EXPORT_SYMBOL(avenrun);
2947 * get_avenrun - get the load average array
2948 * @loads: pointer to dest load array
2949 * @offset: offset to add
2950 * @shift: shift count to shift the result left
2952 * These values are estimates at best, so no need for locking.
2954 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2956 loads[0] = (avenrun[0] + offset) << shift;
2957 loads[1] = (avenrun[1] + offset) << shift;
2958 loads[2] = (avenrun[2] + offset) << shift;
2961 static unsigned long
2962 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2965 load += active * (FIXED_1 - exp);
2966 return load >> FSHIFT;
2970 * calc_load - update the avenrun load estimates 10 ticks after the
2971 * CPUs have updated calc_load_tasks.
2973 void calc_global_load(void)
2975 unsigned long upd = calc_load_update + 10;
2978 if (time_before(jiffies, upd))
2981 active = atomic_long_read(&calc_load_tasks);
2982 active = active > 0 ? active * FIXED_1 : 0;
2984 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2985 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2986 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2988 calc_load_update += LOAD_FREQ;
2992 * Either called from update_cpu_load() or from a cpu going idle
2994 static void calc_load_account_active(struct rq *this_rq)
2996 long nr_active, delta;
2998 nr_active = this_rq->nr_running;
2999 nr_active += (long) this_rq->nr_uninterruptible;
3001 if (nr_active != this_rq->calc_load_active) {
3002 delta = nr_active - this_rq->calc_load_active;
3003 this_rq->calc_load_active = nr_active;
3004 atomic_long_add(delta, &calc_load_tasks);
3009 * Update rq->cpu_load[] statistics. This function is usually called every
3010 * scheduler tick (TICK_NSEC).
3012 static void update_cpu_load(struct rq *this_rq)
3014 unsigned long this_load = this_rq->load.weight;
3017 this_rq->nr_load_updates++;
3019 /* Update our load: */
3020 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3021 unsigned long old_load, new_load;
3023 /* scale is effectively 1 << i now, and >> i divides by scale */
3025 old_load = this_rq->cpu_load[i];
3026 new_load = this_load;
3028 * Round up the averaging division if load is increasing. This
3029 * prevents us from getting stuck on 9 if the load is 10, for
3032 if (new_load > old_load)
3033 new_load += scale-1;
3034 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3037 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3038 this_rq->calc_load_update += LOAD_FREQ;
3039 calc_load_account_active(this_rq);
3046 * sched_exec - execve() is a valuable balancing opportunity, because at
3047 * this point the task has the smallest effective memory and cache footprint.
3049 void sched_exec(void)
3051 struct task_struct *p = current;
3052 struct migration_req req;
3053 unsigned long flags;
3057 rq = task_rq_lock(p, &flags);
3058 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3059 if (dest_cpu == smp_processor_id())
3063 * select_task_rq() can race against ->cpus_allowed
3065 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3066 likely(cpu_active(dest_cpu)) &&
3067 migrate_task(p, dest_cpu, &req)) {
3068 /* Need to wait for migration thread (might exit: take ref). */
3069 struct task_struct *mt = rq->migration_thread;
3071 get_task_struct(mt);
3072 task_rq_unlock(rq, &flags);
3073 wake_up_process(mt);
3074 put_task_struct(mt);
3075 wait_for_completion(&req.done);
3080 task_rq_unlock(rq, &flags);
3085 DEFINE_PER_CPU(struct kernel_stat, kstat);
3087 EXPORT_PER_CPU_SYMBOL(kstat);
3090 * Return any ns on the sched_clock that have not yet been accounted in
3091 * @p in case that task is currently running.
3093 * Called with task_rq_lock() held on @rq.
3095 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3099 if (task_current(rq, p)) {
3100 update_rq_clock(rq);
3101 ns = rq->clock - p->se.exec_start;
3109 unsigned long long task_delta_exec(struct task_struct *p)
3111 unsigned long flags;
3115 rq = task_rq_lock(p, &flags);
3116 ns = do_task_delta_exec(p, rq);
3117 task_rq_unlock(rq, &flags);
3123 * Return accounted runtime for the task.
3124 * In case the task is currently running, return the runtime plus current's
3125 * pending runtime that have not been accounted yet.
3127 unsigned long long task_sched_runtime(struct task_struct *p)
3129 unsigned long flags;
3133 rq = task_rq_lock(p, &flags);
3134 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3135 task_rq_unlock(rq, &flags);
3141 * Return sum_exec_runtime for the thread group.
3142 * In case the task is currently running, return the sum plus current's
3143 * pending runtime that have not been accounted yet.
3145 * Note that the thread group might have other running tasks as well,
3146 * so the return value not includes other pending runtime that other
3147 * running tasks might have.
3149 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3151 struct task_cputime totals;
3152 unsigned long flags;
3156 rq = task_rq_lock(p, &flags);
3157 thread_group_cputime(p, &totals);
3158 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3159 task_rq_unlock(rq, &flags);
3165 * Account user cpu time to a process.
3166 * @p: the process that the cpu time gets accounted to
3167 * @cputime: the cpu time spent in user space since the last update
3168 * @cputime_scaled: cputime scaled by cpu frequency
3170 void account_user_time(struct task_struct *p, cputime_t cputime,
3171 cputime_t cputime_scaled)
3173 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3176 /* Add user time to process. */
3177 p->utime = cputime_add(p->utime, cputime);
3178 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3179 account_group_user_time(p, cputime);
3181 /* Add user time to cpustat. */
3182 tmp = cputime_to_cputime64(cputime);
3183 if (TASK_NICE(p) > 0)
3184 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3186 cpustat->user = cputime64_add(cpustat->user, tmp);
3188 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3189 /* Account for user time used */
3190 acct_update_integrals(p);
3194 * Account guest cpu time to a process.
3195 * @p: the process that the cpu time gets accounted to
3196 * @cputime: the cpu time spent in virtual machine since the last update
3197 * @cputime_scaled: cputime scaled by cpu frequency
3199 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3200 cputime_t cputime_scaled)
3203 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3205 tmp = cputime_to_cputime64(cputime);
3207 /* Add guest time to process. */
3208 p->utime = cputime_add(p->utime, cputime);
3209 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3210 account_group_user_time(p, cputime);
3211 p->gtime = cputime_add(p->gtime, cputime);
3213 /* Add guest time to cpustat. */
3214 if (TASK_NICE(p) > 0) {
3215 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3216 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3218 cpustat->user = cputime64_add(cpustat->user, tmp);
3219 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3224 * Account system cpu time to a process.
3225 * @p: the process that the cpu time gets accounted to
3226 * @hardirq_offset: the offset to subtract from hardirq_count()
3227 * @cputime: the cpu time spent in kernel space since the last update
3228 * @cputime_scaled: cputime scaled by cpu frequency
3230 void account_system_time(struct task_struct *p, int hardirq_offset,
3231 cputime_t cputime, cputime_t cputime_scaled)
3233 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3236 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3237 account_guest_time(p, cputime, cputime_scaled);
3241 /* Add system time to process. */
3242 p->stime = cputime_add(p->stime, cputime);
3243 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3244 account_group_system_time(p, cputime);
3246 /* Add system time to cpustat. */
3247 tmp = cputime_to_cputime64(cputime);
3248 if (hardirq_count() - hardirq_offset)
3249 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3250 else if (softirq_count())
3251 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3253 cpustat->system = cputime64_add(cpustat->system, tmp);
3255 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3257 /* Account for system time used */
3258 acct_update_integrals(p);
3262 * Account for involuntary wait time.
3263 * @steal: the cpu time spent in involuntary wait
3265 void account_steal_time(cputime_t cputime)
3267 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3268 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3270 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3274 * Account for idle time.
3275 * @cputime: the cpu time spent in idle wait
3277 void account_idle_time(cputime_t cputime)
3279 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3280 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3281 struct rq *rq = this_rq();
3283 if (atomic_read(&rq->nr_iowait) > 0)
3284 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3286 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3289 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3292 * Account a single tick of cpu time.
3293 * @p: the process that the cpu time gets accounted to
3294 * @user_tick: indicates if the tick is a user or a system tick
3296 void account_process_tick(struct task_struct *p, int user_tick)
3298 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3299 struct rq *rq = this_rq();
3302 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3303 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3304 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3307 account_idle_time(cputime_one_jiffy);
3311 * Account multiple ticks of steal time.
3312 * @p: the process from which the cpu time has been stolen
3313 * @ticks: number of stolen ticks
3315 void account_steal_ticks(unsigned long ticks)
3317 account_steal_time(jiffies_to_cputime(ticks));
3321 * Account multiple ticks of idle time.
3322 * @ticks: number of stolen ticks
3324 void account_idle_ticks(unsigned long ticks)
3326 account_idle_time(jiffies_to_cputime(ticks));
3332 * Use precise platform statistics if available:
3334 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3335 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3341 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3343 struct task_cputime cputime;
3345 thread_group_cputime(p, &cputime);
3347 *ut = cputime.utime;
3348 *st = cputime.stime;
3352 #ifndef nsecs_to_cputime
3353 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3356 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3358 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3361 * Use CFS's precise accounting:
3363 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3368 temp = (u64)(rtime * utime);
3369 do_div(temp, total);
3370 utime = (cputime_t)temp;
3375 * Compare with previous values, to keep monotonicity:
3377 p->prev_utime = max(p->prev_utime, utime);
3378 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3380 *ut = p->prev_utime;
3381 *st = p->prev_stime;
3385 * Must be called with siglock held.
3387 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3389 struct signal_struct *sig = p->signal;
3390 struct task_cputime cputime;
3391 cputime_t rtime, utime, total;
3393 thread_group_cputime(p, &cputime);
3395 total = cputime_add(cputime.utime, cputime.stime);
3396 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3401 temp = (u64)(rtime * cputime.utime);
3402 do_div(temp, total);
3403 utime = (cputime_t)temp;
3407 sig->prev_utime = max(sig->prev_utime, utime);
3408 sig->prev_stime = max(sig->prev_stime,
3409 cputime_sub(rtime, sig->prev_utime));
3411 *ut = sig->prev_utime;
3412 *st = sig->prev_stime;
3417 * This function gets called by the timer code, with HZ frequency.
3418 * We call it with interrupts disabled.
3420 * It also gets called by the fork code, when changing the parent's
3423 void scheduler_tick(void)
3425 int cpu = smp_processor_id();
3426 struct rq *rq = cpu_rq(cpu);
3427 struct task_struct *curr = rq->curr;
3431 raw_spin_lock(&rq->lock);
3432 update_rq_clock(rq);
3433 update_cpu_load(rq);
3434 curr->sched_class->task_tick(rq, curr, 0);
3435 raw_spin_unlock(&rq->lock);
3437 perf_event_task_tick(curr);
3440 rq->idle_at_tick = idle_cpu(cpu);
3441 trigger_load_balance(rq, cpu);
3445 notrace unsigned long get_parent_ip(unsigned long addr)
3447 if (in_lock_functions(addr)) {
3448 addr = CALLER_ADDR2;
3449 if (in_lock_functions(addr))
3450 addr = CALLER_ADDR3;
3455 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3456 defined(CONFIG_PREEMPT_TRACER))
3458 void __kprobes add_preempt_count(int val)
3460 #ifdef CONFIG_DEBUG_PREEMPT
3464 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3467 preempt_count() += val;
3468 #ifdef CONFIG_DEBUG_PREEMPT
3470 * Spinlock count overflowing soon?
3472 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3475 if (preempt_count() == val)
3476 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3478 EXPORT_SYMBOL(add_preempt_count);
3480 void __kprobes sub_preempt_count(int val)
3482 #ifdef CONFIG_DEBUG_PREEMPT
3486 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3489 * Is the spinlock portion underflowing?
3491 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3492 !(preempt_count() & PREEMPT_MASK)))
3496 if (preempt_count() == val)
3497 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3498 preempt_count() -= val;
3500 EXPORT_SYMBOL(sub_preempt_count);
3505 * Print scheduling while atomic bug:
3507 static noinline void __schedule_bug(struct task_struct *prev)
3509 struct pt_regs *regs = get_irq_regs();
3511 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3512 prev->comm, prev->pid, preempt_count());
3514 debug_show_held_locks(prev);
3516 if (irqs_disabled())
3517 print_irqtrace_events(prev);
3526 * Various schedule()-time debugging checks and statistics:
3528 static inline void schedule_debug(struct task_struct *prev)
3531 * Test if we are atomic. Since do_exit() needs to call into
3532 * schedule() atomically, we ignore that path for now.
3533 * Otherwise, whine if we are scheduling when we should not be.
3535 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3536 __schedule_bug(prev);
3538 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3540 schedstat_inc(this_rq(), sched_count);
3541 #ifdef CONFIG_SCHEDSTATS
3542 if (unlikely(prev->lock_depth >= 0)) {
3543 schedstat_inc(this_rq(), bkl_count);
3544 schedstat_inc(prev, sched_info.bkl_count);
3549 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3552 update_rq_clock(rq);
3553 rq->skip_clock_update = 0;
3554 prev->sched_class->put_prev_task(rq, prev);
3558 * Pick up the highest-prio task:
3560 static inline struct task_struct *
3561 pick_next_task(struct rq *rq)
3563 const struct sched_class *class;
3564 struct task_struct *p;
3567 * Optimization: we know that if all tasks are in
3568 * the fair class we can call that function directly:
3570 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3571 p = fair_sched_class.pick_next_task(rq);
3576 class = sched_class_highest;
3578 p = class->pick_next_task(rq);
3582 * Will never be NULL as the idle class always
3583 * returns a non-NULL p:
3585 class = class->next;
3590 * schedule() is the main scheduler function.
3592 asmlinkage void __sched schedule(void)
3594 struct task_struct *prev, *next;
3595 unsigned long *switch_count;
3601 cpu = smp_processor_id();
3605 switch_count = &prev->nivcsw;
3607 release_kernel_lock(prev);
3608 need_resched_nonpreemptible:
3610 schedule_debug(prev);
3612 if (sched_feat(HRTICK))
3615 raw_spin_lock_irq(&rq->lock);
3616 clear_tsk_need_resched(prev);
3618 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3619 if (unlikely(signal_pending_state(prev->state, prev)))
3620 prev->state = TASK_RUNNING;
3622 deactivate_task(rq, prev, 1);
3623 switch_count = &prev->nvcsw;
3626 pre_schedule(rq, prev);
3628 if (unlikely(!rq->nr_running))
3629 idle_balance(cpu, rq);
3631 put_prev_task(rq, prev);
3632 next = pick_next_task(rq);
3634 if (likely(prev != next)) {
3635 sched_info_switch(prev, next);
3636 perf_event_task_sched_out(prev, next);
3642 context_switch(rq, prev, next); /* unlocks the rq */
3644 * the context switch might have flipped the stack from under
3645 * us, hence refresh the local variables.
3647 cpu = smp_processor_id();
3650 raw_spin_unlock_irq(&rq->lock);
3654 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3656 switch_count = &prev->nivcsw;
3657 goto need_resched_nonpreemptible;
3660 preempt_enable_no_resched();
3664 EXPORT_SYMBOL(schedule);
3666 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3668 * Look out! "owner" is an entirely speculative pointer
3669 * access and not reliable.
3671 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3676 if (!sched_feat(OWNER_SPIN))
3679 #ifdef CONFIG_DEBUG_PAGEALLOC
3681 * Need to access the cpu field knowing that
3682 * DEBUG_PAGEALLOC could have unmapped it if
3683 * the mutex owner just released it and exited.
3685 if (probe_kernel_address(&owner->cpu, cpu))
3692 * Even if the access succeeded (likely case),
3693 * the cpu field may no longer be valid.
3695 if (cpu >= nr_cpumask_bits)
3699 * We need to validate that we can do a
3700 * get_cpu() and that we have the percpu area.
3702 if (!cpu_online(cpu))
3709 * Owner changed, break to re-assess state.
3711 if (lock->owner != owner)
3715 * Is that owner really running on that cpu?
3717 if (task_thread_info(rq->curr) != owner || need_resched())
3727 #ifdef CONFIG_PREEMPT
3729 * this is the entry point to schedule() from in-kernel preemption
3730 * off of preempt_enable. Kernel preemptions off return from interrupt
3731 * occur there and call schedule directly.
3733 asmlinkage void __sched preempt_schedule(void)
3735 struct thread_info *ti = current_thread_info();
3738 * If there is a non-zero preempt_count or interrupts are disabled,
3739 * we do not want to preempt the current task. Just return..
3741 if (likely(ti->preempt_count || irqs_disabled()))
3745 add_preempt_count(PREEMPT_ACTIVE);
3747 sub_preempt_count(PREEMPT_ACTIVE);
3750 * Check again in case we missed a preemption opportunity
3751 * between schedule and now.
3754 } while (need_resched());
3756 EXPORT_SYMBOL(preempt_schedule);
3759 * this is the entry point to schedule() from kernel preemption
3760 * off of irq context.
3761 * Note, that this is called and return with irqs disabled. This will
3762 * protect us against recursive calling from irq.
3764 asmlinkage void __sched preempt_schedule_irq(void)
3766 struct thread_info *ti = current_thread_info();
3768 /* Catch callers which need to be fixed */
3769 BUG_ON(ti->preempt_count || !irqs_disabled());
3772 add_preempt_count(PREEMPT_ACTIVE);
3775 local_irq_disable();
3776 sub_preempt_count(PREEMPT_ACTIVE);
3779 * Check again in case we missed a preemption opportunity
3780 * between schedule and now.
3783 } while (need_resched());
3786 #endif /* CONFIG_PREEMPT */
3788 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3791 return try_to_wake_up(curr->private, mode, wake_flags);
3793 EXPORT_SYMBOL(default_wake_function);
3796 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3797 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3798 * number) then we wake all the non-exclusive tasks and one exclusive task.
3800 * There are circumstances in which we can try to wake a task which has already
3801 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3802 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3804 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3805 int nr_exclusive, int wake_flags, void *key)
3807 wait_queue_t *curr, *next;
3809 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3810 unsigned flags = curr->flags;
3812 if (curr->func(curr, mode, wake_flags, key) &&
3813 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3819 * __wake_up - wake up threads blocked on a waitqueue.
3821 * @mode: which threads
3822 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3823 * @key: is directly passed to the wakeup function
3825 * It may be assumed that this function implies a write memory barrier before
3826 * changing the task state if and only if any tasks are woken up.
3828 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3829 int nr_exclusive, void *key)
3831 unsigned long flags;
3833 spin_lock_irqsave(&q->lock, flags);
3834 __wake_up_common(q, mode, nr_exclusive, 0, key);
3835 spin_unlock_irqrestore(&q->lock, flags);
3837 EXPORT_SYMBOL(__wake_up);
3840 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3842 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3844 __wake_up_common(q, mode, 1, 0, NULL);
3847 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3849 __wake_up_common(q, mode, 1, 0, key);
3853 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3855 * @mode: which threads
3856 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3857 * @key: opaque value to be passed to wakeup targets
3859 * The sync wakeup differs that the waker knows that it will schedule
3860 * away soon, so while the target thread will be woken up, it will not
3861 * be migrated to another CPU - ie. the two threads are 'synchronized'
3862 * with each other. This can prevent needless bouncing between CPUs.
3864 * On UP it can prevent extra preemption.
3866 * It may be assumed that this function implies a write memory barrier before
3867 * changing the task state if and only if any tasks are woken up.
3869 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3870 int nr_exclusive, void *key)
3872 unsigned long flags;
3873 int wake_flags = WF_SYNC;
3878 if (unlikely(!nr_exclusive))
3881 spin_lock_irqsave(&q->lock, flags);
3882 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3883 spin_unlock_irqrestore(&q->lock, flags);
3885 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3888 * __wake_up_sync - see __wake_up_sync_key()
3890 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3892 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3894 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3897 * complete: - signals a single thread waiting on this completion
3898 * @x: holds the state of this particular completion
3900 * This will wake up a single thread waiting on this completion. Threads will be
3901 * awakened in the same order in which they were queued.
3903 * See also complete_all(), wait_for_completion() and related routines.
3905 * It may be assumed that this function implies a write memory barrier before
3906 * changing the task state if and only if any tasks are woken up.
3908 void complete(struct completion *x)
3910 unsigned long flags;
3912 spin_lock_irqsave(&x->wait.lock, flags);
3914 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3915 spin_unlock_irqrestore(&x->wait.lock, flags);
3917 EXPORT_SYMBOL(complete);
3920 * complete_all: - signals all threads waiting on this completion
3921 * @x: holds the state of this particular completion
3923 * This will wake up all threads waiting on this particular completion event.
3925 * It may be assumed that this function implies a write memory barrier before
3926 * changing the task state if and only if any tasks are woken up.
3928 void complete_all(struct completion *x)
3930 unsigned long flags;
3932 spin_lock_irqsave(&x->wait.lock, flags);
3933 x->done += UINT_MAX/2;
3934 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3935 spin_unlock_irqrestore(&x->wait.lock, flags);
3937 EXPORT_SYMBOL(complete_all);
3939 static inline long __sched
3940 do_wait_for_common(struct completion *x, long timeout, int state)
3943 DECLARE_WAITQUEUE(wait, current);
3945 wait.flags |= WQ_FLAG_EXCLUSIVE;
3946 __add_wait_queue_tail(&x->wait, &wait);
3948 if (signal_pending_state(state, current)) {
3949 timeout = -ERESTARTSYS;
3952 __set_current_state(state);
3953 spin_unlock_irq(&x->wait.lock);
3954 timeout = schedule_timeout(timeout);
3955 spin_lock_irq(&x->wait.lock);
3956 } while (!x->done && timeout);
3957 __remove_wait_queue(&x->wait, &wait);
3962 return timeout ?: 1;
3966 wait_for_common(struct completion *x, long timeout, int state)
3970 spin_lock_irq(&x->wait.lock);
3971 timeout = do_wait_for_common(x, timeout, state);
3972 spin_unlock_irq(&x->wait.lock);
3977 * wait_for_completion: - waits for completion of a task
3978 * @x: holds the state of this particular completion
3980 * This waits to be signaled for completion of a specific task. It is NOT
3981 * interruptible and there is no timeout.
3983 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3984 * and interrupt capability. Also see complete().
3986 void __sched wait_for_completion(struct completion *x)
3988 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3990 EXPORT_SYMBOL(wait_for_completion);
3993 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3994 * @x: holds the state of this particular completion
3995 * @timeout: timeout value in jiffies
3997 * This waits for either a completion of a specific task to be signaled or for a
3998 * specified timeout to expire. The timeout is in jiffies. It is not
4001 unsigned long __sched
4002 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4004 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4006 EXPORT_SYMBOL(wait_for_completion_timeout);
4009 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4010 * @x: holds the state of this particular completion
4012 * This waits for completion of a specific task to be signaled. It is
4015 int __sched wait_for_completion_interruptible(struct completion *x)
4017 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4018 if (t == -ERESTARTSYS)
4022 EXPORT_SYMBOL(wait_for_completion_interruptible);
4025 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4026 * @x: holds the state of this particular completion
4027 * @timeout: timeout value in jiffies
4029 * This waits for either a completion of a specific task to be signaled or for a
4030 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4032 unsigned long __sched
4033 wait_for_completion_interruptible_timeout(struct completion *x,
4034 unsigned long timeout)
4036 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4038 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4041 * wait_for_completion_killable: - waits for completion of a task (killable)
4042 * @x: holds the state of this particular completion
4044 * This waits to be signaled for completion of a specific task. It can be
4045 * interrupted by a kill signal.
4047 int __sched wait_for_completion_killable(struct completion *x)
4049 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4050 if (t == -ERESTARTSYS)
4054 EXPORT_SYMBOL(wait_for_completion_killable);
4057 * try_wait_for_completion - try to decrement a completion without blocking
4058 * @x: completion structure
4060 * Returns: 0 if a decrement cannot be done without blocking
4061 * 1 if a decrement succeeded.
4063 * If a completion is being used as a counting completion,
4064 * attempt to decrement the counter without blocking. This
4065 * enables us to avoid waiting if the resource the completion
4066 * is protecting is not available.
4068 bool try_wait_for_completion(struct completion *x)
4070 unsigned long flags;
4073 spin_lock_irqsave(&x->wait.lock, flags);
4078 spin_unlock_irqrestore(&x->wait.lock, flags);
4081 EXPORT_SYMBOL(try_wait_for_completion);
4084 * completion_done - Test to see if a completion has any waiters
4085 * @x: completion structure
4087 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4088 * 1 if there are no waiters.
4091 bool completion_done(struct completion *x)
4093 unsigned long flags;
4096 spin_lock_irqsave(&x->wait.lock, flags);
4099 spin_unlock_irqrestore(&x->wait.lock, flags);
4102 EXPORT_SYMBOL(completion_done);
4105 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4107 unsigned long flags;
4110 init_waitqueue_entry(&wait, current);
4112 __set_current_state(state);
4114 spin_lock_irqsave(&q->lock, flags);
4115 __add_wait_queue(q, &wait);
4116 spin_unlock(&q->lock);
4117 timeout = schedule_timeout(timeout);
4118 spin_lock_irq(&q->lock);
4119 __remove_wait_queue(q, &wait);
4120 spin_unlock_irqrestore(&q->lock, flags);
4125 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4127 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4129 EXPORT_SYMBOL(interruptible_sleep_on);
4132 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4134 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4136 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4138 void __sched sleep_on(wait_queue_head_t *q)
4140 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4142 EXPORT_SYMBOL(sleep_on);
4144 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4146 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4148 EXPORT_SYMBOL(sleep_on_timeout);
4150 #ifdef CONFIG_RT_MUTEXES
4153 * rt_mutex_setprio - set the current priority of a task
4155 * @prio: prio value (kernel-internal form)
4157 * This function changes the 'effective' priority of a task. It does
4158 * not touch ->normal_prio like __setscheduler().
4160 * Used by the rt_mutex code to implement priority inheritance logic.
4162 void rt_mutex_setprio(struct task_struct *p, int prio)
4164 unsigned long flags;
4165 int oldprio, on_rq, running;
4167 const struct sched_class *prev_class;
4169 BUG_ON(prio < 0 || prio > MAX_PRIO);
4171 rq = task_rq_lock(p, &flags);
4174 prev_class = p->sched_class;
4175 on_rq = p->se.on_rq;
4176 running = task_current(rq, p);
4178 dequeue_task(rq, p, 0);
4180 p->sched_class->put_prev_task(rq, p);
4183 p->sched_class = &rt_sched_class;
4185 p->sched_class = &fair_sched_class;
4190 p->sched_class->set_curr_task(rq);
4192 enqueue_task(rq, p, 0, oldprio < prio);
4194 check_class_changed(rq, p, prev_class, oldprio, running);
4196 task_rq_unlock(rq, &flags);
4201 void set_user_nice(struct task_struct *p, long nice)
4203 int old_prio, delta, on_rq;
4204 unsigned long flags;
4207 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4210 * We have to be careful, if called from sys_setpriority(),
4211 * the task might be in the middle of scheduling on another CPU.
4213 rq = task_rq_lock(p, &flags);
4215 * The RT priorities are set via sched_setscheduler(), but we still
4216 * allow the 'normal' nice value to be set - but as expected
4217 * it wont have any effect on scheduling until the task is
4218 * SCHED_FIFO/SCHED_RR:
4220 if (task_has_rt_policy(p)) {
4221 p->static_prio = NICE_TO_PRIO(nice);
4224 on_rq = p->se.on_rq;
4226 dequeue_task(rq, p, 0);
4228 p->static_prio = NICE_TO_PRIO(nice);
4231 p->prio = effective_prio(p);
4232 delta = p->prio - old_prio;
4235 enqueue_task(rq, p, 0, false);
4237 * If the task increased its priority or is running and
4238 * lowered its priority, then reschedule its CPU:
4240 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4241 resched_task(rq->curr);
4244 task_rq_unlock(rq, &flags);
4246 EXPORT_SYMBOL(set_user_nice);
4249 * can_nice - check if a task can reduce its nice value
4253 int can_nice(const struct task_struct *p, const int nice)
4255 /* convert nice value [19,-20] to rlimit style value [1,40] */
4256 int nice_rlim = 20 - nice;
4258 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4259 capable(CAP_SYS_NICE));
4262 #ifdef __ARCH_WANT_SYS_NICE
4265 * sys_nice - change the priority of the current process.
4266 * @increment: priority increment
4268 * sys_setpriority is a more generic, but much slower function that
4269 * does similar things.
4271 SYSCALL_DEFINE1(nice, int, increment)
4276 * Setpriority might change our priority at the same moment.
4277 * We don't have to worry. Conceptually one call occurs first
4278 * and we have a single winner.
4280 if (increment < -40)
4285 nice = TASK_NICE(current) + increment;
4291 if (increment < 0 && !can_nice(current, nice))
4294 retval = security_task_setnice(current, nice);
4298 set_user_nice(current, nice);
4305 * task_prio - return the priority value of a given task.
4306 * @p: the task in question.
4308 * This is the priority value as seen by users in /proc.
4309 * RT tasks are offset by -200. Normal tasks are centered
4310 * around 0, value goes from -16 to +15.
4312 int task_prio(const struct task_struct *p)
4314 return p->prio - MAX_RT_PRIO;
4318 * task_nice - return the nice value of a given task.
4319 * @p: the task in question.
4321 int task_nice(const struct task_struct *p)
4323 return TASK_NICE(p);
4325 EXPORT_SYMBOL(task_nice);
4328 * idle_cpu - is a given cpu idle currently?
4329 * @cpu: the processor in question.
4331 int idle_cpu(int cpu)
4333 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4337 * idle_task - return the idle task for a given cpu.
4338 * @cpu: the processor in question.
4340 struct task_struct *idle_task(int cpu)
4342 return cpu_rq(cpu)->idle;
4346 * find_process_by_pid - find a process with a matching PID value.
4347 * @pid: the pid in question.
4349 static struct task_struct *find_process_by_pid(pid_t pid)
4351 return pid ? find_task_by_vpid(pid) : current;
4354 /* Actually do priority change: must hold rq lock. */
4356 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4358 BUG_ON(p->se.on_rq);
4361 p->rt_priority = prio;
4362 p->normal_prio = normal_prio(p);
4363 /* we are holding p->pi_lock already */
4364 p->prio = rt_mutex_getprio(p);
4365 if (rt_prio(p->prio))
4366 p->sched_class = &rt_sched_class;
4368 p->sched_class = &fair_sched_class;
4373 * check the target process has a UID that matches the current process's
4375 static bool check_same_owner(struct task_struct *p)
4377 const struct cred *cred = current_cred(), *pcred;
4381 pcred = __task_cred(p);
4382 match = (cred->euid == pcred->euid ||
4383 cred->euid == pcred->uid);
4388 static int __sched_setscheduler(struct task_struct *p, int policy,
4389 struct sched_param *param, bool user)
4391 int retval, oldprio, oldpolicy = -1, on_rq, running;
4392 unsigned long flags;
4393 const struct sched_class *prev_class;
4397 /* may grab non-irq protected spin_locks */
4398 BUG_ON(in_interrupt());
4400 /* double check policy once rq lock held */
4402 reset_on_fork = p->sched_reset_on_fork;
4403 policy = oldpolicy = p->policy;
4405 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4406 policy &= ~SCHED_RESET_ON_FORK;
4408 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4409 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4410 policy != SCHED_IDLE)
4415 * Valid priorities for SCHED_FIFO and SCHED_RR are
4416 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4417 * SCHED_BATCH and SCHED_IDLE is 0.
4419 if (param->sched_priority < 0 ||
4420 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4421 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4423 if (rt_policy(policy) != (param->sched_priority != 0))
4427 * Allow unprivileged RT tasks to decrease priority:
4429 if (user && !capable(CAP_SYS_NICE)) {
4430 if (rt_policy(policy)) {
4431 unsigned long rlim_rtprio;
4433 if (!lock_task_sighand(p, &flags))
4435 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4436 unlock_task_sighand(p, &flags);
4438 /* can't set/change the rt policy */
4439 if (policy != p->policy && !rlim_rtprio)
4442 /* can't increase priority */
4443 if (param->sched_priority > p->rt_priority &&
4444 param->sched_priority > rlim_rtprio)
4448 * Like positive nice levels, dont allow tasks to
4449 * move out of SCHED_IDLE either:
4451 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4454 /* can't change other user's priorities */
4455 if (!check_same_owner(p))
4458 /* Normal users shall not reset the sched_reset_on_fork flag */
4459 if (p->sched_reset_on_fork && !reset_on_fork)
4464 #ifdef CONFIG_RT_GROUP_SCHED
4466 * Do not allow realtime tasks into groups that have no runtime
4469 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4470 task_group(p)->rt_bandwidth.rt_runtime == 0)
4474 retval = security_task_setscheduler(p, policy, param);
4480 * make sure no PI-waiters arrive (or leave) while we are
4481 * changing the priority of the task:
4483 raw_spin_lock_irqsave(&p->pi_lock, flags);
4485 * To be able to change p->policy safely, the apropriate
4486 * runqueue lock must be held.
4488 rq = __task_rq_lock(p);
4489 /* recheck policy now with rq lock held */
4490 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4491 policy = oldpolicy = -1;
4492 __task_rq_unlock(rq);
4493 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4496 on_rq = p->se.on_rq;
4497 running = task_current(rq, p);
4499 deactivate_task(rq, p, 0);
4501 p->sched_class->put_prev_task(rq, p);
4503 p->sched_reset_on_fork = reset_on_fork;
4506 prev_class = p->sched_class;
4507 __setscheduler(rq, p, policy, param->sched_priority);
4510 p->sched_class->set_curr_task(rq);
4512 activate_task(rq, p, 0);
4514 check_class_changed(rq, p, prev_class, oldprio, running);
4516 __task_rq_unlock(rq);
4517 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4519 rt_mutex_adjust_pi(p);
4525 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4526 * @p: the task in question.
4527 * @policy: new policy.
4528 * @param: structure containing the new RT priority.
4530 * NOTE that the task may be already dead.
4532 int sched_setscheduler(struct task_struct *p, int policy,
4533 struct sched_param *param)
4535 return __sched_setscheduler(p, policy, param, true);
4537 EXPORT_SYMBOL_GPL(sched_setscheduler);
4540 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4541 * @p: the task in question.
4542 * @policy: new policy.
4543 * @param: structure containing the new RT priority.
4545 * Just like sched_setscheduler, only don't bother checking if the
4546 * current context has permission. For example, this is needed in
4547 * stop_machine(): we create temporary high priority worker threads,
4548 * but our caller might not have that capability.
4550 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4551 struct sched_param *param)
4553 return __sched_setscheduler(p, policy, param, false);
4557 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4559 struct sched_param lparam;
4560 struct task_struct *p;
4563 if (!param || pid < 0)
4565 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4570 p = find_process_by_pid(pid);
4572 retval = sched_setscheduler(p, policy, &lparam);
4579 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4580 * @pid: the pid in question.
4581 * @policy: new policy.
4582 * @param: structure containing the new RT priority.
4584 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4585 struct sched_param __user *, param)
4587 /* negative values for policy are not valid */
4591 return do_sched_setscheduler(pid, policy, param);
4595 * sys_sched_setparam - set/change the RT priority of a thread
4596 * @pid: the pid in question.
4597 * @param: structure containing the new RT priority.
4599 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4601 return do_sched_setscheduler(pid, -1, param);
4605 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4606 * @pid: the pid in question.
4608 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4610 struct task_struct *p;
4618 p = find_process_by_pid(pid);
4620 retval = security_task_getscheduler(p);
4623 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4630 * sys_sched_getparam - get the RT priority of a thread
4631 * @pid: the pid in question.
4632 * @param: structure containing the RT priority.
4634 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4636 struct sched_param lp;
4637 struct task_struct *p;
4640 if (!param || pid < 0)
4644 p = find_process_by_pid(pid);
4649 retval = security_task_getscheduler(p);
4653 lp.sched_priority = p->rt_priority;
4657 * This one might sleep, we cannot do it with a spinlock held ...
4659 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4668 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4670 cpumask_var_t cpus_allowed, new_mask;
4671 struct task_struct *p;
4677 p = find_process_by_pid(pid);
4684 /* Prevent p going away */
4688 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4692 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4694 goto out_free_cpus_allowed;
4697 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4700 retval = security_task_setscheduler(p, 0, NULL);
4704 cpuset_cpus_allowed(p, cpus_allowed);
4705 cpumask_and(new_mask, in_mask, cpus_allowed);
4707 retval = set_cpus_allowed_ptr(p, new_mask);
4710 cpuset_cpus_allowed(p, cpus_allowed);
4711 if (!cpumask_subset(new_mask, cpus_allowed)) {
4713 * We must have raced with a concurrent cpuset
4714 * update. Just reset the cpus_allowed to the
4715 * cpuset's cpus_allowed
4717 cpumask_copy(new_mask, cpus_allowed);
4722 free_cpumask_var(new_mask);
4723 out_free_cpus_allowed:
4724 free_cpumask_var(cpus_allowed);
4731 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4732 struct cpumask *new_mask)
4734 if (len < cpumask_size())
4735 cpumask_clear(new_mask);
4736 else if (len > cpumask_size())
4737 len = cpumask_size();
4739 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4743 * sys_sched_setaffinity - set the cpu affinity of a process
4744 * @pid: pid of the process
4745 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4746 * @user_mask_ptr: user-space pointer to the new cpu mask
4748 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4749 unsigned long __user *, user_mask_ptr)
4751 cpumask_var_t new_mask;
4754 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4757 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4759 retval = sched_setaffinity(pid, new_mask);
4760 free_cpumask_var(new_mask);
4764 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4766 struct task_struct *p;
4767 unsigned long flags;
4775 p = find_process_by_pid(pid);
4779 retval = security_task_getscheduler(p);
4783 rq = task_rq_lock(p, &flags);
4784 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4785 task_rq_unlock(rq, &flags);
4795 * sys_sched_getaffinity - get the cpu affinity of a process
4796 * @pid: pid of the process
4797 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4798 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4800 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4801 unsigned long __user *, user_mask_ptr)
4806 if (len < nr_cpu_ids)
4808 if (len & (sizeof(unsigned long)-1))
4811 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4814 ret = sched_getaffinity(pid, mask);
4816 size_t retlen = min_t(size_t, len, cpumask_size());
4818 if (copy_to_user(user_mask_ptr, mask, retlen))
4823 free_cpumask_var(mask);
4829 * sys_sched_yield - yield the current processor to other threads.
4831 * This function yields the current CPU to other tasks. If there are no
4832 * other threads running on this CPU then this function will return.
4834 SYSCALL_DEFINE0(sched_yield)
4836 struct rq *rq = this_rq_lock();
4838 schedstat_inc(rq, yld_count);
4839 current->sched_class->yield_task(rq);
4842 * Since we are going to call schedule() anyway, there's
4843 * no need to preempt or enable interrupts:
4845 __release(rq->lock);
4846 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4847 do_raw_spin_unlock(&rq->lock);
4848 preempt_enable_no_resched();
4855 static inline int should_resched(void)
4857 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4860 static void __cond_resched(void)
4862 add_preempt_count(PREEMPT_ACTIVE);
4864 sub_preempt_count(PREEMPT_ACTIVE);
4867 int __sched _cond_resched(void)
4869 if (should_resched()) {
4875 EXPORT_SYMBOL(_cond_resched);
4878 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4879 * call schedule, and on return reacquire the lock.
4881 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4882 * operations here to prevent schedule() from being called twice (once via
4883 * spin_unlock(), once by hand).
4885 int __cond_resched_lock(spinlock_t *lock)
4887 int resched = should_resched();
4890 lockdep_assert_held(lock);
4892 if (spin_needbreak(lock) || resched) {
4903 EXPORT_SYMBOL(__cond_resched_lock);
4905 int __sched __cond_resched_softirq(void)
4907 BUG_ON(!in_softirq());
4909 if (should_resched()) {
4917 EXPORT_SYMBOL(__cond_resched_softirq);
4920 * yield - yield the current processor to other threads.
4922 * This is a shortcut for kernel-space yielding - it marks the
4923 * thread runnable and calls sys_sched_yield().
4925 void __sched yield(void)
4927 set_current_state(TASK_RUNNING);
4930 EXPORT_SYMBOL(yield);
4933 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4934 * that process accounting knows that this is a task in IO wait state.
4936 void __sched io_schedule(void)
4938 struct rq *rq = raw_rq();
4940 delayacct_blkio_start();
4941 atomic_inc(&rq->nr_iowait);
4942 current->in_iowait = 1;
4944 current->in_iowait = 0;
4945 atomic_dec(&rq->nr_iowait);
4946 delayacct_blkio_end();
4948 EXPORT_SYMBOL(io_schedule);
4950 long __sched io_schedule_timeout(long timeout)
4952 struct rq *rq = raw_rq();
4955 delayacct_blkio_start();
4956 atomic_inc(&rq->nr_iowait);
4957 current->in_iowait = 1;
4958 ret = schedule_timeout(timeout);
4959 current->in_iowait = 0;
4960 atomic_dec(&rq->nr_iowait);
4961 delayacct_blkio_end();
4966 * sys_sched_get_priority_max - return maximum RT priority.
4967 * @policy: scheduling class.
4969 * this syscall returns the maximum rt_priority that can be used
4970 * by a given scheduling class.
4972 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4979 ret = MAX_USER_RT_PRIO-1;
4991 * sys_sched_get_priority_min - return minimum RT priority.
4992 * @policy: scheduling class.
4994 * this syscall returns the minimum rt_priority that can be used
4995 * by a given scheduling class.
4997 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5015 * sys_sched_rr_get_interval - return the default timeslice of a process.
5016 * @pid: pid of the process.
5017 * @interval: userspace pointer to the timeslice value.
5019 * this syscall writes the default timeslice value of a given process
5020 * into the user-space timespec buffer. A value of '0' means infinity.
5022 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5023 struct timespec __user *, interval)
5025 struct task_struct *p;
5026 unsigned int time_slice;
5027 unsigned long flags;
5037 p = find_process_by_pid(pid);
5041 retval = security_task_getscheduler(p);
5045 rq = task_rq_lock(p, &flags);
5046 time_slice = p->sched_class->get_rr_interval(rq, p);
5047 task_rq_unlock(rq, &flags);
5050 jiffies_to_timespec(time_slice, &t);
5051 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5059 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5061 void sched_show_task(struct task_struct *p)
5063 unsigned long free = 0;
5066 state = p->state ? __ffs(p->state) + 1 : 0;
5067 printk(KERN_INFO "%-13.13s %c", p->comm,
5068 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5069 #if BITS_PER_LONG == 32
5070 if (state == TASK_RUNNING)
5071 printk(KERN_CONT " running ");
5073 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5075 if (state == TASK_RUNNING)
5076 printk(KERN_CONT " running task ");
5078 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5080 #ifdef CONFIG_DEBUG_STACK_USAGE
5081 free = stack_not_used(p);
5083 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5084 task_pid_nr(p), task_pid_nr(p->real_parent),
5085 (unsigned long)task_thread_info(p)->flags);
5087 show_stack(p, NULL);
5090 void show_state_filter(unsigned long state_filter)
5092 struct task_struct *g, *p;
5094 #if BITS_PER_LONG == 32
5096 " task PC stack pid father\n");
5099 " task PC stack pid father\n");
5101 read_lock(&tasklist_lock);
5102 do_each_thread(g, p) {
5104 * reset the NMI-timeout, listing all files on a slow
5105 * console might take alot of time:
5107 touch_nmi_watchdog();
5108 if (!state_filter || (p->state & state_filter))
5110 } while_each_thread(g, p);
5112 touch_all_softlockup_watchdogs();
5114 #ifdef CONFIG_SCHED_DEBUG
5115 sysrq_sched_debug_show();
5117 read_unlock(&tasklist_lock);
5119 * Only show locks if all tasks are dumped:
5122 debug_show_all_locks();
5125 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5127 idle->sched_class = &idle_sched_class;
5131 * init_idle - set up an idle thread for a given CPU
5132 * @idle: task in question
5133 * @cpu: cpu the idle task belongs to
5135 * NOTE: this function does not set the idle thread's NEED_RESCHED
5136 * flag, to make booting more robust.
5138 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5140 struct rq *rq = cpu_rq(cpu);
5141 unsigned long flags;
5143 raw_spin_lock_irqsave(&rq->lock, flags);
5146 idle->state = TASK_RUNNING;
5147 idle->se.exec_start = sched_clock();
5149 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5150 __set_task_cpu(idle, cpu);
5152 rq->curr = rq->idle = idle;
5153 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5156 raw_spin_unlock_irqrestore(&rq->lock, flags);
5158 /* Set the preempt count _outside_ the spinlocks! */
5159 #if defined(CONFIG_PREEMPT)
5160 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5162 task_thread_info(idle)->preempt_count = 0;
5165 * The idle tasks have their own, simple scheduling class:
5167 idle->sched_class = &idle_sched_class;
5168 ftrace_graph_init_task(idle);
5172 * In a system that switches off the HZ timer nohz_cpu_mask
5173 * indicates which cpus entered this state. This is used
5174 * in the rcu update to wait only for active cpus. For system
5175 * which do not switch off the HZ timer nohz_cpu_mask should
5176 * always be CPU_BITS_NONE.
5178 cpumask_var_t nohz_cpu_mask;
5181 * Increase the granularity value when there are more CPUs,
5182 * because with more CPUs the 'effective latency' as visible
5183 * to users decreases. But the relationship is not linear,
5184 * so pick a second-best guess by going with the log2 of the
5187 * This idea comes from the SD scheduler of Con Kolivas:
5189 static int get_update_sysctl_factor(void)
5191 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5192 unsigned int factor;
5194 switch (sysctl_sched_tunable_scaling) {
5195 case SCHED_TUNABLESCALING_NONE:
5198 case SCHED_TUNABLESCALING_LINEAR:
5201 case SCHED_TUNABLESCALING_LOG:
5203 factor = 1 + ilog2(cpus);
5210 static void update_sysctl(void)
5212 unsigned int factor = get_update_sysctl_factor();
5214 #define SET_SYSCTL(name) \
5215 (sysctl_##name = (factor) * normalized_sysctl_##name)
5216 SET_SYSCTL(sched_min_granularity);
5217 SET_SYSCTL(sched_latency);
5218 SET_SYSCTL(sched_wakeup_granularity);
5219 SET_SYSCTL(sched_shares_ratelimit);
5223 static inline void sched_init_granularity(void)
5230 * This is how migration works:
5232 * 1) we queue a struct migration_req structure in the source CPU's
5233 * runqueue and wake up that CPU's migration thread.
5234 * 2) we down() the locked semaphore => thread blocks.
5235 * 3) migration thread wakes up (implicitly it forces the migrated
5236 * thread off the CPU)
5237 * 4) it gets the migration request and checks whether the migrated
5238 * task is still in the wrong runqueue.
5239 * 5) if it's in the wrong runqueue then the migration thread removes
5240 * it and puts it into the right queue.
5241 * 6) migration thread up()s the semaphore.
5242 * 7) we wake up and the migration is done.
5246 * Change a given task's CPU affinity. Migrate the thread to a
5247 * proper CPU and schedule it away if the CPU it's executing on
5248 * is removed from the allowed bitmask.
5250 * NOTE: the caller must have a valid reference to the task, the
5251 * task must not exit() & deallocate itself prematurely. The
5252 * call is not atomic; no spinlocks may be held.
5254 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5256 struct migration_req req;
5257 unsigned long flags;
5262 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5263 * drop the rq->lock and still rely on ->cpus_allowed.
5266 while (task_is_waking(p))
5268 rq = task_rq_lock(p, &flags);
5269 if (task_is_waking(p)) {
5270 task_rq_unlock(rq, &flags);
5274 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5279 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5280 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5285 if (p->sched_class->set_cpus_allowed)
5286 p->sched_class->set_cpus_allowed(p, new_mask);
5288 cpumask_copy(&p->cpus_allowed, new_mask);
5289 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5292 /* Can the task run on the task's current CPU? If so, we're done */
5293 if (cpumask_test_cpu(task_cpu(p), new_mask))
5296 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5297 /* Need help from migration thread: drop lock and wait. */
5298 struct task_struct *mt = rq->migration_thread;
5300 get_task_struct(mt);
5301 task_rq_unlock(rq, &flags);
5302 wake_up_process(rq->migration_thread);
5303 put_task_struct(mt);
5304 wait_for_completion(&req.done);
5305 tlb_migrate_finish(p->mm);
5309 task_rq_unlock(rq, &flags);
5313 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5316 * Move (not current) task off this cpu, onto dest cpu. We're doing
5317 * this because either it can't run here any more (set_cpus_allowed()
5318 * away from this CPU, or CPU going down), or because we're
5319 * attempting to rebalance this task on exec (sched_exec).
5321 * So we race with normal scheduler movements, but that's OK, as long
5322 * as the task is no longer on this CPU.
5324 * Returns non-zero if task was successfully migrated.
5326 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5328 struct rq *rq_dest, *rq_src;
5331 if (unlikely(!cpu_active(dest_cpu)))
5334 rq_src = cpu_rq(src_cpu);
5335 rq_dest = cpu_rq(dest_cpu);
5337 double_rq_lock(rq_src, rq_dest);
5338 /* Already moved. */
5339 if (task_cpu(p) != src_cpu)
5341 /* Affinity changed (again). */
5342 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5346 * If we're not on a rq, the next wake-up will ensure we're
5350 deactivate_task(rq_src, p, 0);
5351 set_task_cpu(p, dest_cpu);
5352 activate_task(rq_dest, p, 0);
5353 check_preempt_curr(rq_dest, p, 0);
5358 double_rq_unlock(rq_src, rq_dest);
5362 #define RCU_MIGRATION_IDLE 0
5363 #define RCU_MIGRATION_NEED_QS 1
5364 #define RCU_MIGRATION_GOT_QS 2
5365 #define RCU_MIGRATION_MUST_SYNC 3
5368 * migration_thread - this is a highprio system thread that performs
5369 * thread migration by bumping thread off CPU then 'pushing' onto
5372 static int migration_thread(void *data)
5375 int cpu = (long)data;
5379 BUG_ON(rq->migration_thread != current);
5381 set_current_state(TASK_INTERRUPTIBLE);
5382 while (!kthread_should_stop()) {
5383 struct migration_req *req;
5384 struct list_head *head;
5386 raw_spin_lock_irq(&rq->lock);
5388 if (cpu_is_offline(cpu)) {
5389 raw_spin_unlock_irq(&rq->lock);
5393 if (rq->active_balance) {
5394 active_load_balance(rq, cpu);
5395 rq->active_balance = 0;
5398 head = &rq->migration_queue;
5400 if (list_empty(head)) {
5401 raw_spin_unlock_irq(&rq->lock);
5403 set_current_state(TASK_INTERRUPTIBLE);
5406 req = list_entry(head->next, struct migration_req, list);
5407 list_del_init(head->next);
5409 if (req->task != NULL) {
5410 raw_spin_unlock(&rq->lock);
5411 __migrate_task(req->task, cpu, req->dest_cpu);
5412 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5413 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5414 raw_spin_unlock(&rq->lock);
5416 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5417 raw_spin_unlock(&rq->lock);
5418 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5422 complete(&req->done);
5424 __set_current_state(TASK_RUNNING);
5429 #ifdef CONFIG_HOTPLUG_CPU
5431 * Figure out where task on dead CPU should go, use force if necessary.
5433 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5435 struct rq *rq = cpu_rq(dead_cpu);
5436 int needs_cpu, uninitialized_var(dest_cpu);
5437 unsigned long flags;
5439 local_irq_save(flags);
5441 raw_spin_lock(&rq->lock);
5442 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5444 dest_cpu = select_fallback_rq(dead_cpu, p);
5445 raw_spin_unlock(&rq->lock);
5447 * It can only fail if we race with set_cpus_allowed(),
5448 * in the racer should migrate the task anyway.
5451 __migrate_task(p, dead_cpu, dest_cpu);
5452 local_irq_restore(flags);
5456 * While a dead CPU has no uninterruptible tasks queued at this point,
5457 * it might still have a nonzero ->nr_uninterruptible counter, because
5458 * for performance reasons the counter is not stricly tracking tasks to
5459 * their home CPUs. So we just add the counter to another CPU's counter,
5460 * to keep the global sum constant after CPU-down:
5462 static void migrate_nr_uninterruptible(struct rq *rq_src)
5464 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5465 unsigned long flags;
5467 local_irq_save(flags);
5468 double_rq_lock(rq_src, rq_dest);
5469 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5470 rq_src->nr_uninterruptible = 0;
5471 double_rq_unlock(rq_src, rq_dest);
5472 local_irq_restore(flags);
5475 /* Run through task list and migrate tasks from the dead cpu. */
5476 static void migrate_live_tasks(int src_cpu)
5478 struct task_struct *p, *t;
5480 read_lock(&tasklist_lock);
5482 do_each_thread(t, p) {
5486 if (task_cpu(p) == src_cpu)
5487 move_task_off_dead_cpu(src_cpu, p);
5488 } while_each_thread(t, p);
5490 read_unlock(&tasklist_lock);
5494 * Schedules idle task to be the next runnable task on current CPU.
5495 * It does so by boosting its priority to highest possible.
5496 * Used by CPU offline code.
5498 void sched_idle_next(void)
5500 int this_cpu = smp_processor_id();
5501 struct rq *rq = cpu_rq(this_cpu);
5502 struct task_struct *p = rq->idle;
5503 unsigned long flags;
5505 /* cpu has to be offline */
5506 BUG_ON(cpu_online(this_cpu));
5509 * Strictly not necessary since rest of the CPUs are stopped by now
5510 * and interrupts disabled on the current cpu.
5512 raw_spin_lock_irqsave(&rq->lock, flags);
5514 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5516 activate_task(rq, p, 0);
5518 raw_spin_unlock_irqrestore(&rq->lock, flags);
5522 * Ensures that the idle task is using init_mm right before its cpu goes
5525 void idle_task_exit(void)
5527 struct mm_struct *mm = current->active_mm;
5529 BUG_ON(cpu_online(smp_processor_id()));
5532 switch_mm(mm, &init_mm, current);
5536 /* called under rq->lock with disabled interrupts */
5537 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5539 struct rq *rq = cpu_rq(dead_cpu);
5541 /* Must be exiting, otherwise would be on tasklist. */
5542 BUG_ON(!p->exit_state);
5544 /* Cannot have done final schedule yet: would have vanished. */
5545 BUG_ON(p->state == TASK_DEAD);
5550 * Drop lock around migration; if someone else moves it,
5551 * that's OK. No task can be added to this CPU, so iteration is
5554 raw_spin_unlock_irq(&rq->lock);
5555 move_task_off_dead_cpu(dead_cpu, p);
5556 raw_spin_lock_irq(&rq->lock);
5561 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5562 static void migrate_dead_tasks(unsigned int dead_cpu)
5564 struct rq *rq = cpu_rq(dead_cpu);
5565 struct task_struct *next;
5568 if (!rq->nr_running)
5570 next = pick_next_task(rq);
5573 next->sched_class->put_prev_task(rq, next);
5574 migrate_dead(dead_cpu, next);
5580 * remove the tasks which were accounted by rq from calc_load_tasks.
5582 static void calc_global_load_remove(struct rq *rq)
5584 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5585 rq->calc_load_active = 0;
5587 #endif /* CONFIG_HOTPLUG_CPU */
5589 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5591 static struct ctl_table sd_ctl_dir[] = {
5593 .procname = "sched_domain",
5599 static struct ctl_table sd_ctl_root[] = {
5601 .procname = "kernel",
5603 .child = sd_ctl_dir,
5608 static struct ctl_table *sd_alloc_ctl_entry(int n)
5610 struct ctl_table *entry =
5611 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5616 static void sd_free_ctl_entry(struct ctl_table **tablep)
5618 struct ctl_table *entry;
5621 * In the intermediate directories, both the child directory and
5622 * procname are dynamically allocated and could fail but the mode
5623 * will always be set. In the lowest directory the names are
5624 * static strings and all have proc handlers.
5626 for (entry = *tablep; entry->mode; entry++) {
5628 sd_free_ctl_entry(&entry->child);
5629 if (entry->proc_handler == NULL)
5630 kfree(entry->procname);
5638 set_table_entry(struct ctl_table *entry,
5639 const char *procname, void *data, int maxlen,
5640 mode_t mode, proc_handler *proc_handler)
5642 entry->procname = procname;
5644 entry->maxlen = maxlen;
5646 entry->proc_handler = proc_handler;
5649 static struct ctl_table *
5650 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5652 struct ctl_table *table = sd_alloc_ctl_entry(13);
5657 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5658 sizeof(long), 0644, proc_doulongvec_minmax);
5659 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5660 sizeof(long), 0644, proc_doulongvec_minmax);
5661 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5662 sizeof(int), 0644, proc_dointvec_minmax);
5663 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5664 sizeof(int), 0644, proc_dointvec_minmax);
5665 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5666 sizeof(int), 0644, proc_dointvec_minmax);
5667 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5668 sizeof(int), 0644, proc_dointvec_minmax);
5669 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5670 sizeof(int), 0644, proc_dointvec_minmax);
5671 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5672 sizeof(int), 0644, proc_dointvec_minmax);
5673 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5674 sizeof(int), 0644, proc_dointvec_minmax);
5675 set_table_entry(&table[9], "cache_nice_tries",
5676 &sd->cache_nice_tries,
5677 sizeof(int), 0644, proc_dointvec_minmax);
5678 set_table_entry(&table[10], "flags", &sd->flags,
5679 sizeof(int), 0644, proc_dointvec_minmax);
5680 set_table_entry(&table[11], "name", sd->name,
5681 CORENAME_MAX_SIZE, 0444, proc_dostring);
5682 /* &table[12] is terminator */
5687 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5689 struct ctl_table *entry, *table;
5690 struct sched_domain *sd;
5691 int domain_num = 0, i;
5694 for_each_domain(cpu, sd)
5696 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5701 for_each_domain(cpu, sd) {
5702 snprintf(buf, 32, "domain%d", i);
5703 entry->procname = kstrdup(buf, GFP_KERNEL);
5705 entry->child = sd_alloc_ctl_domain_table(sd);
5712 static struct ctl_table_header *sd_sysctl_header;
5713 static void register_sched_domain_sysctl(void)
5715 int i, cpu_num = num_possible_cpus();
5716 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5719 WARN_ON(sd_ctl_dir[0].child);
5720 sd_ctl_dir[0].child = entry;
5725 for_each_possible_cpu(i) {
5726 snprintf(buf, 32, "cpu%d", i);
5727 entry->procname = kstrdup(buf, GFP_KERNEL);
5729 entry->child = sd_alloc_ctl_cpu_table(i);
5733 WARN_ON(sd_sysctl_header);
5734 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5737 /* may be called multiple times per register */
5738 static void unregister_sched_domain_sysctl(void)
5740 if (sd_sysctl_header)
5741 unregister_sysctl_table(sd_sysctl_header);
5742 sd_sysctl_header = NULL;
5743 if (sd_ctl_dir[0].child)
5744 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5747 static void register_sched_domain_sysctl(void)
5750 static void unregister_sched_domain_sysctl(void)
5755 static void set_rq_online(struct rq *rq)
5758 const struct sched_class *class;
5760 cpumask_set_cpu(rq->cpu, rq->rd->online);
5763 for_each_class(class) {
5764 if (class->rq_online)
5765 class->rq_online(rq);
5770 static void set_rq_offline(struct rq *rq)
5773 const struct sched_class *class;
5775 for_each_class(class) {
5776 if (class->rq_offline)
5777 class->rq_offline(rq);
5780 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5786 * migration_call - callback that gets triggered when a CPU is added.
5787 * Here we can start up the necessary migration thread for the new CPU.
5789 static int __cpuinit
5790 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5792 struct task_struct *p;
5793 int cpu = (long)hcpu;
5794 unsigned long flags;
5799 case CPU_UP_PREPARE:
5800 case CPU_UP_PREPARE_FROZEN:
5801 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5804 kthread_bind(p, cpu);
5805 /* Must be high prio: stop_machine expects to yield to it. */
5806 rq = task_rq_lock(p, &flags);
5807 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5808 task_rq_unlock(rq, &flags);
5810 cpu_rq(cpu)->migration_thread = p;
5811 rq->calc_load_update = calc_load_update;
5815 case CPU_ONLINE_FROZEN:
5816 /* Strictly unnecessary, as first user will wake it. */
5817 wake_up_process(cpu_rq(cpu)->migration_thread);
5819 /* Update our root-domain */
5821 raw_spin_lock_irqsave(&rq->lock, flags);
5823 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5827 raw_spin_unlock_irqrestore(&rq->lock, flags);
5830 #ifdef CONFIG_HOTPLUG_CPU
5831 case CPU_UP_CANCELED:
5832 case CPU_UP_CANCELED_FROZEN:
5833 if (!cpu_rq(cpu)->migration_thread)
5835 /* Unbind it from offline cpu so it can run. Fall thru. */
5836 kthread_bind(cpu_rq(cpu)->migration_thread,
5837 cpumask_any(cpu_online_mask));
5838 kthread_stop(cpu_rq(cpu)->migration_thread);
5839 put_task_struct(cpu_rq(cpu)->migration_thread);
5840 cpu_rq(cpu)->migration_thread = NULL;
5844 case CPU_DEAD_FROZEN:
5845 migrate_live_tasks(cpu);
5847 kthread_stop(rq->migration_thread);
5848 put_task_struct(rq->migration_thread);
5849 rq->migration_thread = NULL;
5850 /* Idle task back to normal (off runqueue, low prio) */
5851 raw_spin_lock_irq(&rq->lock);
5852 deactivate_task(rq, rq->idle, 0);
5853 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5854 rq->idle->sched_class = &idle_sched_class;
5855 migrate_dead_tasks(cpu);
5856 raw_spin_unlock_irq(&rq->lock);
5857 migrate_nr_uninterruptible(rq);
5858 BUG_ON(rq->nr_running != 0);
5859 calc_global_load_remove(rq);
5861 * No need to migrate the tasks: it was best-effort if
5862 * they didn't take sched_hotcpu_mutex. Just wake up
5865 raw_spin_lock_irq(&rq->lock);
5866 while (!list_empty(&rq->migration_queue)) {
5867 struct migration_req *req;
5869 req = list_entry(rq->migration_queue.next,
5870 struct migration_req, list);
5871 list_del_init(&req->list);
5872 raw_spin_unlock_irq(&rq->lock);
5873 complete(&req->done);
5874 raw_spin_lock_irq(&rq->lock);
5876 raw_spin_unlock_irq(&rq->lock);
5880 case CPU_DYING_FROZEN:
5881 /* Update our root-domain */
5883 raw_spin_lock_irqsave(&rq->lock, flags);
5885 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5888 raw_spin_unlock_irqrestore(&rq->lock, flags);
5896 * Register at high priority so that task migration (migrate_all_tasks)
5897 * happens before everything else. This has to be lower priority than
5898 * the notifier in the perf_event subsystem, though.
5900 static struct notifier_block __cpuinitdata migration_notifier = {
5901 .notifier_call = migration_call,
5905 static int __init migration_init(void)
5907 void *cpu = (void *)(long)smp_processor_id();
5910 /* Start one for the boot CPU: */
5911 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5912 BUG_ON(err == NOTIFY_BAD);
5913 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5914 register_cpu_notifier(&migration_notifier);
5918 early_initcall(migration_init);
5923 #ifdef CONFIG_SCHED_DEBUG
5925 static __read_mostly int sched_domain_debug_enabled;
5927 static int __init sched_domain_debug_setup(char *str)
5929 sched_domain_debug_enabled = 1;
5933 early_param("sched_debug", sched_domain_debug_setup);
5935 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5936 struct cpumask *groupmask)
5938 struct sched_group *group = sd->groups;
5941 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5942 cpumask_clear(groupmask);
5944 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5946 if (!(sd->flags & SD_LOAD_BALANCE)) {
5947 printk("does not load-balance\n");
5949 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5954 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5956 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5957 printk(KERN_ERR "ERROR: domain->span does not contain "
5960 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5961 printk(KERN_ERR "ERROR: domain->groups does not contain"
5965 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5969 printk(KERN_ERR "ERROR: group is NULL\n");
5973 if (!group->cpu_power) {
5974 printk(KERN_CONT "\n");
5975 printk(KERN_ERR "ERROR: domain->cpu_power not "
5980 if (!cpumask_weight(sched_group_cpus(group))) {
5981 printk(KERN_CONT "\n");
5982 printk(KERN_ERR "ERROR: empty group\n");
5986 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5987 printk(KERN_CONT "\n");
5988 printk(KERN_ERR "ERROR: repeated CPUs\n");
5992 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5994 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5996 printk(KERN_CONT " %s", str);
5997 if (group->cpu_power != SCHED_LOAD_SCALE) {
5998 printk(KERN_CONT " (cpu_power = %d)",
6002 group = group->next;
6003 } while (group != sd->groups);
6004 printk(KERN_CONT "\n");
6006 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6007 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6010 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6011 printk(KERN_ERR "ERROR: parent span is not a superset "
6012 "of domain->span\n");
6016 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6018 cpumask_var_t groupmask;
6021 if (!sched_domain_debug_enabled)
6025 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6029 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6031 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6032 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6037 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6044 free_cpumask_var(groupmask);
6046 #else /* !CONFIG_SCHED_DEBUG */
6047 # define sched_domain_debug(sd, cpu) do { } while (0)
6048 #endif /* CONFIG_SCHED_DEBUG */
6050 static int sd_degenerate(struct sched_domain *sd)
6052 if (cpumask_weight(sched_domain_span(sd)) == 1)
6055 /* Following flags need at least 2 groups */
6056 if (sd->flags & (SD_LOAD_BALANCE |
6057 SD_BALANCE_NEWIDLE |
6061 SD_SHARE_PKG_RESOURCES)) {
6062 if (sd->groups != sd->groups->next)
6066 /* Following flags don't use groups */
6067 if (sd->flags & (SD_WAKE_AFFINE))
6074 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6076 unsigned long cflags = sd->flags, pflags = parent->flags;
6078 if (sd_degenerate(parent))
6081 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6084 /* Flags needing groups don't count if only 1 group in parent */
6085 if (parent->groups == parent->groups->next) {
6086 pflags &= ~(SD_LOAD_BALANCE |
6087 SD_BALANCE_NEWIDLE |
6091 SD_SHARE_PKG_RESOURCES);
6092 if (nr_node_ids == 1)
6093 pflags &= ~SD_SERIALIZE;
6095 if (~cflags & pflags)
6101 static void free_rootdomain(struct root_domain *rd)
6103 synchronize_sched();
6105 cpupri_cleanup(&rd->cpupri);
6107 free_cpumask_var(rd->rto_mask);
6108 free_cpumask_var(rd->online);
6109 free_cpumask_var(rd->span);
6113 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6115 struct root_domain *old_rd = NULL;
6116 unsigned long flags;
6118 raw_spin_lock_irqsave(&rq->lock, flags);
6123 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6126 cpumask_clear_cpu(rq->cpu, old_rd->span);
6129 * If we dont want to free the old_rt yet then
6130 * set old_rd to NULL to skip the freeing later
6133 if (!atomic_dec_and_test(&old_rd->refcount))
6137 atomic_inc(&rd->refcount);
6140 cpumask_set_cpu(rq->cpu, rd->span);
6141 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6144 raw_spin_unlock_irqrestore(&rq->lock, flags);
6147 free_rootdomain(old_rd);
6150 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6152 gfp_t gfp = GFP_KERNEL;
6154 memset(rd, 0, sizeof(*rd));
6159 if (!alloc_cpumask_var(&rd->span, gfp))
6161 if (!alloc_cpumask_var(&rd->online, gfp))
6163 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6166 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6171 free_cpumask_var(rd->rto_mask);
6173 free_cpumask_var(rd->online);
6175 free_cpumask_var(rd->span);
6180 static void init_defrootdomain(void)
6182 init_rootdomain(&def_root_domain, true);
6184 atomic_set(&def_root_domain.refcount, 1);
6187 static struct root_domain *alloc_rootdomain(void)
6189 struct root_domain *rd;
6191 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6195 if (init_rootdomain(rd, false) != 0) {
6204 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6205 * hold the hotplug lock.
6208 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6210 struct rq *rq = cpu_rq(cpu);
6211 struct sched_domain *tmp;
6213 /* Remove the sched domains which do not contribute to scheduling. */
6214 for (tmp = sd; tmp; ) {
6215 struct sched_domain *parent = tmp->parent;
6219 if (sd_parent_degenerate(tmp, parent)) {
6220 tmp->parent = parent->parent;
6222 parent->parent->child = tmp;
6227 if (sd && sd_degenerate(sd)) {
6233 sched_domain_debug(sd, cpu);
6235 rq_attach_root(rq, rd);
6236 rcu_assign_pointer(rq->sd, sd);
6239 /* cpus with isolated domains */
6240 static cpumask_var_t cpu_isolated_map;
6242 /* Setup the mask of cpus configured for isolated domains */
6243 static int __init isolated_cpu_setup(char *str)
6245 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6246 cpulist_parse(str, cpu_isolated_map);
6250 __setup("isolcpus=", isolated_cpu_setup);
6253 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6254 * to a function which identifies what group(along with sched group) a CPU
6255 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6256 * (due to the fact that we keep track of groups covered with a struct cpumask).
6258 * init_sched_build_groups will build a circular linked list of the groups
6259 * covered by the given span, and will set each group's ->cpumask correctly,
6260 * and ->cpu_power to 0.
6263 init_sched_build_groups(const struct cpumask *span,
6264 const struct cpumask *cpu_map,
6265 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6266 struct sched_group **sg,
6267 struct cpumask *tmpmask),
6268 struct cpumask *covered, struct cpumask *tmpmask)
6270 struct sched_group *first = NULL, *last = NULL;
6273 cpumask_clear(covered);
6275 for_each_cpu(i, span) {
6276 struct sched_group *sg;
6277 int group = group_fn(i, cpu_map, &sg, tmpmask);
6280 if (cpumask_test_cpu(i, covered))
6283 cpumask_clear(sched_group_cpus(sg));
6286 for_each_cpu(j, span) {
6287 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6290 cpumask_set_cpu(j, covered);
6291 cpumask_set_cpu(j, sched_group_cpus(sg));
6302 #define SD_NODES_PER_DOMAIN 16
6307 * find_next_best_node - find the next node to include in a sched_domain
6308 * @node: node whose sched_domain we're building
6309 * @used_nodes: nodes already in the sched_domain
6311 * Find the next node to include in a given scheduling domain. Simply
6312 * finds the closest node not already in the @used_nodes map.
6314 * Should use nodemask_t.
6316 static int find_next_best_node(int node, nodemask_t *used_nodes)
6318 int i, n, val, min_val, best_node = 0;
6322 for (i = 0; i < nr_node_ids; i++) {
6323 /* Start at @node */
6324 n = (node + i) % nr_node_ids;
6326 if (!nr_cpus_node(n))
6329 /* Skip already used nodes */
6330 if (node_isset(n, *used_nodes))
6333 /* Simple min distance search */
6334 val = node_distance(node, n);
6336 if (val < min_val) {
6342 node_set(best_node, *used_nodes);
6347 * sched_domain_node_span - get a cpumask for a node's sched_domain
6348 * @node: node whose cpumask we're constructing
6349 * @span: resulting cpumask
6351 * Given a node, construct a good cpumask for its sched_domain to span. It
6352 * should be one that prevents unnecessary balancing, but also spreads tasks
6355 static void sched_domain_node_span(int node, struct cpumask *span)
6357 nodemask_t used_nodes;
6360 cpumask_clear(span);
6361 nodes_clear(used_nodes);
6363 cpumask_or(span, span, cpumask_of_node(node));
6364 node_set(node, used_nodes);
6366 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6367 int next_node = find_next_best_node(node, &used_nodes);
6369 cpumask_or(span, span, cpumask_of_node(next_node));
6372 #endif /* CONFIG_NUMA */
6374 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6377 * The cpus mask in sched_group and sched_domain hangs off the end.
6379 * ( See the the comments in include/linux/sched.h:struct sched_group
6380 * and struct sched_domain. )
6382 struct static_sched_group {
6383 struct sched_group sg;
6384 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6387 struct static_sched_domain {
6388 struct sched_domain sd;
6389 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6395 cpumask_var_t domainspan;
6396 cpumask_var_t covered;
6397 cpumask_var_t notcovered;
6399 cpumask_var_t nodemask;
6400 cpumask_var_t this_sibling_map;
6401 cpumask_var_t this_core_map;
6402 cpumask_var_t send_covered;
6403 cpumask_var_t tmpmask;
6404 struct sched_group **sched_group_nodes;
6405 struct root_domain *rd;
6409 sa_sched_groups = 0,
6414 sa_this_sibling_map,
6416 sa_sched_group_nodes,
6426 * SMT sched-domains:
6428 #ifdef CONFIG_SCHED_SMT
6429 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6430 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6433 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6434 struct sched_group **sg, struct cpumask *unused)
6437 *sg = &per_cpu(sched_groups, cpu).sg;
6440 #endif /* CONFIG_SCHED_SMT */
6443 * multi-core sched-domains:
6445 #ifdef CONFIG_SCHED_MC
6446 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6447 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6448 #endif /* CONFIG_SCHED_MC */
6450 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6452 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6453 struct sched_group **sg, struct cpumask *mask)
6457 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6458 group = cpumask_first(mask);
6460 *sg = &per_cpu(sched_group_core, group).sg;
6463 #elif defined(CONFIG_SCHED_MC)
6465 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6466 struct sched_group **sg, struct cpumask *unused)
6469 *sg = &per_cpu(sched_group_core, cpu).sg;
6474 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6475 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6478 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6479 struct sched_group **sg, struct cpumask *mask)
6482 #ifdef CONFIG_SCHED_MC
6483 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6484 group = cpumask_first(mask);
6485 #elif defined(CONFIG_SCHED_SMT)
6486 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6487 group = cpumask_first(mask);
6492 *sg = &per_cpu(sched_group_phys, group).sg;
6498 * The init_sched_build_groups can't handle what we want to do with node
6499 * groups, so roll our own. Now each node has its own list of groups which
6500 * gets dynamically allocated.
6502 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6503 static struct sched_group ***sched_group_nodes_bycpu;
6505 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6506 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6508 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6509 struct sched_group **sg,
6510 struct cpumask *nodemask)
6514 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6515 group = cpumask_first(nodemask);
6518 *sg = &per_cpu(sched_group_allnodes, group).sg;
6522 static void init_numa_sched_groups_power(struct sched_group *group_head)
6524 struct sched_group *sg = group_head;
6530 for_each_cpu(j, sched_group_cpus(sg)) {
6531 struct sched_domain *sd;
6533 sd = &per_cpu(phys_domains, j).sd;
6534 if (j != group_first_cpu(sd->groups)) {
6536 * Only add "power" once for each
6542 sg->cpu_power += sd->groups->cpu_power;
6545 } while (sg != group_head);
6548 static int build_numa_sched_groups(struct s_data *d,
6549 const struct cpumask *cpu_map, int num)
6551 struct sched_domain *sd;
6552 struct sched_group *sg, *prev;
6555 cpumask_clear(d->covered);
6556 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6557 if (cpumask_empty(d->nodemask)) {
6558 d->sched_group_nodes[num] = NULL;
6562 sched_domain_node_span(num, d->domainspan);
6563 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6565 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6568 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6572 d->sched_group_nodes[num] = sg;
6574 for_each_cpu(j, d->nodemask) {
6575 sd = &per_cpu(node_domains, j).sd;
6580 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6582 cpumask_or(d->covered, d->covered, d->nodemask);
6585 for (j = 0; j < nr_node_ids; j++) {
6586 n = (num + j) % nr_node_ids;
6587 cpumask_complement(d->notcovered, d->covered);
6588 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6589 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6590 if (cpumask_empty(d->tmpmask))
6592 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6593 if (cpumask_empty(d->tmpmask))
6595 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6599 "Can not alloc domain group for node %d\n", j);
6603 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6604 sg->next = prev->next;
6605 cpumask_or(d->covered, d->covered, d->tmpmask);
6612 #endif /* CONFIG_NUMA */
6615 /* Free memory allocated for various sched_group structures */
6616 static void free_sched_groups(const struct cpumask *cpu_map,
6617 struct cpumask *nodemask)
6621 for_each_cpu(cpu, cpu_map) {
6622 struct sched_group **sched_group_nodes
6623 = sched_group_nodes_bycpu[cpu];
6625 if (!sched_group_nodes)
6628 for (i = 0; i < nr_node_ids; i++) {
6629 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6631 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6632 if (cpumask_empty(nodemask))
6642 if (oldsg != sched_group_nodes[i])
6645 kfree(sched_group_nodes);
6646 sched_group_nodes_bycpu[cpu] = NULL;
6649 #else /* !CONFIG_NUMA */
6650 static void free_sched_groups(const struct cpumask *cpu_map,
6651 struct cpumask *nodemask)
6654 #endif /* CONFIG_NUMA */
6657 * Initialize sched groups cpu_power.
6659 * cpu_power indicates the capacity of sched group, which is used while
6660 * distributing the load between different sched groups in a sched domain.
6661 * Typically cpu_power for all the groups in a sched domain will be same unless
6662 * there are asymmetries in the topology. If there are asymmetries, group
6663 * having more cpu_power will pickup more load compared to the group having
6666 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6668 struct sched_domain *child;
6669 struct sched_group *group;
6673 WARN_ON(!sd || !sd->groups);
6675 if (cpu != group_first_cpu(sd->groups))
6680 sd->groups->cpu_power = 0;
6683 power = SCHED_LOAD_SCALE;
6684 weight = cpumask_weight(sched_domain_span(sd));
6686 * SMT siblings share the power of a single core.
6687 * Usually multiple threads get a better yield out of
6688 * that one core than a single thread would have,
6689 * reflect that in sd->smt_gain.
6691 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6692 power *= sd->smt_gain;
6694 power >>= SCHED_LOAD_SHIFT;
6696 sd->groups->cpu_power += power;
6701 * Add cpu_power of each child group to this groups cpu_power.
6703 group = child->groups;
6705 sd->groups->cpu_power += group->cpu_power;
6706 group = group->next;
6707 } while (group != child->groups);
6711 * Initializers for schedule domains
6712 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6715 #ifdef CONFIG_SCHED_DEBUG
6716 # define SD_INIT_NAME(sd, type) sd->name = #type
6718 # define SD_INIT_NAME(sd, type) do { } while (0)
6721 #define SD_INIT(sd, type) sd_init_##type(sd)
6723 #define SD_INIT_FUNC(type) \
6724 static noinline void sd_init_##type(struct sched_domain *sd) \
6726 memset(sd, 0, sizeof(*sd)); \
6727 *sd = SD_##type##_INIT; \
6728 sd->level = SD_LV_##type; \
6729 SD_INIT_NAME(sd, type); \
6734 SD_INIT_FUNC(ALLNODES)
6737 #ifdef CONFIG_SCHED_SMT
6738 SD_INIT_FUNC(SIBLING)
6740 #ifdef CONFIG_SCHED_MC
6744 static int default_relax_domain_level = -1;
6746 static int __init setup_relax_domain_level(char *str)
6750 val = simple_strtoul(str, NULL, 0);
6751 if (val < SD_LV_MAX)
6752 default_relax_domain_level = val;
6756 __setup("relax_domain_level=", setup_relax_domain_level);
6758 static void set_domain_attribute(struct sched_domain *sd,
6759 struct sched_domain_attr *attr)
6763 if (!attr || attr->relax_domain_level < 0) {
6764 if (default_relax_domain_level < 0)
6767 request = default_relax_domain_level;
6769 request = attr->relax_domain_level;
6770 if (request < sd->level) {
6771 /* turn off idle balance on this domain */
6772 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6774 /* turn on idle balance on this domain */
6775 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6779 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6780 const struct cpumask *cpu_map)
6783 case sa_sched_groups:
6784 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6785 d->sched_group_nodes = NULL;
6787 free_rootdomain(d->rd); /* fall through */
6789 free_cpumask_var(d->tmpmask); /* fall through */
6790 case sa_send_covered:
6791 free_cpumask_var(d->send_covered); /* fall through */
6792 case sa_this_core_map:
6793 free_cpumask_var(d->this_core_map); /* fall through */
6794 case sa_this_sibling_map:
6795 free_cpumask_var(d->this_sibling_map); /* fall through */
6797 free_cpumask_var(d->nodemask); /* fall through */
6798 case sa_sched_group_nodes:
6800 kfree(d->sched_group_nodes); /* fall through */
6802 free_cpumask_var(d->notcovered); /* fall through */
6804 free_cpumask_var(d->covered); /* fall through */
6806 free_cpumask_var(d->domainspan); /* fall through */
6813 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6814 const struct cpumask *cpu_map)
6817 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6819 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6820 return sa_domainspan;
6821 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6823 /* Allocate the per-node list of sched groups */
6824 d->sched_group_nodes = kcalloc(nr_node_ids,
6825 sizeof(struct sched_group *), GFP_KERNEL);
6826 if (!d->sched_group_nodes) {
6827 printk(KERN_WARNING "Can not alloc sched group node list\n");
6828 return sa_notcovered;
6830 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6832 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6833 return sa_sched_group_nodes;
6834 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6836 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6837 return sa_this_sibling_map;
6838 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6839 return sa_this_core_map;
6840 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6841 return sa_send_covered;
6842 d->rd = alloc_rootdomain();
6844 printk(KERN_WARNING "Cannot alloc root domain\n");
6847 return sa_rootdomain;
6850 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6851 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6853 struct sched_domain *sd = NULL;
6855 struct sched_domain *parent;
6858 if (cpumask_weight(cpu_map) >
6859 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6860 sd = &per_cpu(allnodes_domains, i).sd;
6861 SD_INIT(sd, ALLNODES);
6862 set_domain_attribute(sd, attr);
6863 cpumask_copy(sched_domain_span(sd), cpu_map);
6864 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6869 sd = &per_cpu(node_domains, i).sd;
6871 set_domain_attribute(sd, attr);
6872 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6873 sd->parent = parent;
6876 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6881 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6882 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6883 struct sched_domain *parent, int i)
6885 struct sched_domain *sd;
6886 sd = &per_cpu(phys_domains, i).sd;
6888 set_domain_attribute(sd, attr);
6889 cpumask_copy(sched_domain_span(sd), d->nodemask);
6890 sd->parent = parent;
6893 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6897 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6898 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6899 struct sched_domain *parent, int i)
6901 struct sched_domain *sd = parent;
6902 #ifdef CONFIG_SCHED_MC
6903 sd = &per_cpu(core_domains, i).sd;
6905 set_domain_attribute(sd, attr);
6906 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6907 sd->parent = parent;
6909 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6914 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6915 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6916 struct sched_domain *parent, int i)
6918 struct sched_domain *sd = parent;
6919 #ifdef CONFIG_SCHED_SMT
6920 sd = &per_cpu(cpu_domains, i).sd;
6921 SD_INIT(sd, SIBLING);
6922 set_domain_attribute(sd, attr);
6923 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6924 sd->parent = parent;
6926 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6931 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6932 const struct cpumask *cpu_map, int cpu)
6935 #ifdef CONFIG_SCHED_SMT
6936 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6937 cpumask_and(d->this_sibling_map, cpu_map,
6938 topology_thread_cpumask(cpu));
6939 if (cpu == cpumask_first(d->this_sibling_map))
6940 init_sched_build_groups(d->this_sibling_map, cpu_map,
6942 d->send_covered, d->tmpmask);
6945 #ifdef CONFIG_SCHED_MC
6946 case SD_LV_MC: /* set up multi-core groups */
6947 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6948 if (cpu == cpumask_first(d->this_core_map))
6949 init_sched_build_groups(d->this_core_map, cpu_map,
6951 d->send_covered, d->tmpmask);
6954 case SD_LV_CPU: /* set up physical groups */
6955 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
6956 if (!cpumask_empty(d->nodemask))
6957 init_sched_build_groups(d->nodemask, cpu_map,
6959 d->send_covered, d->tmpmask);
6962 case SD_LV_ALLNODES:
6963 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
6964 d->send_covered, d->tmpmask);
6973 * Build sched domains for a given set of cpus and attach the sched domains
6974 * to the individual cpus
6976 static int __build_sched_domains(const struct cpumask *cpu_map,
6977 struct sched_domain_attr *attr)
6979 enum s_alloc alloc_state = sa_none;
6981 struct sched_domain *sd;
6987 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6988 if (alloc_state != sa_rootdomain)
6990 alloc_state = sa_sched_groups;
6993 * Set up domains for cpus specified by the cpu_map.
6995 for_each_cpu(i, cpu_map) {
6996 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
6999 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7000 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7001 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7002 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7005 for_each_cpu(i, cpu_map) {
7006 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7007 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7010 /* Set up physical groups */
7011 for (i = 0; i < nr_node_ids; i++)
7012 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7015 /* Set up node groups */
7017 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7019 for (i = 0; i < nr_node_ids; i++)
7020 if (build_numa_sched_groups(&d, cpu_map, i))
7024 /* Calculate CPU power for physical packages and nodes */
7025 #ifdef CONFIG_SCHED_SMT
7026 for_each_cpu(i, cpu_map) {
7027 sd = &per_cpu(cpu_domains, i).sd;
7028 init_sched_groups_power(i, sd);
7031 #ifdef CONFIG_SCHED_MC
7032 for_each_cpu(i, cpu_map) {
7033 sd = &per_cpu(core_domains, i).sd;
7034 init_sched_groups_power(i, sd);
7038 for_each_cpu(i, cpu_map) {
7039 sd = &per_cpu(phys_domains, i).sd;
7040 init_sched_groups_power(i, sd);
7044 for (i = 0; i < nr_node_ids; i++)
7045 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7047 if (d.sd_allnodes) {
7048 struct sched_group *sg;
7050 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7052 init_numa_sched_groups_power(sg);
7056 /* Attach the domains */
7057 for_each_cpu(i, cpu_map) {
7058 #ifdef CONFIG_SCHED_SMT
7059 sd = &per_cpu(cpu_domains, i).sd;
7060 #elif defined(CONFIG_SCHED_MC)
7061 sd = &per_cpu(core_domains, i).sd;
7063 sd = &per_cpu(phys_domains, i).sd;
7065 cpu_attach_domain(sd, d.rd, i);
7068 d.sched_group_nodes = NULL; /* don't free this we still need it */
7069 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7073 __free_domain_allocs(&d, alloc_state, cpu_map);
7077 static int build_sched_domains(const struct cpumask *cpu_map)
7079 return __build_sched_domains(cpu_map, NULL);
7082 static cpumask_var_t *doms_cur; /* current sched domains */
7083 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7084 static struct sched_domain_attr *dattr_cur;
7085 /* attribues of custom domains in 'doms_cur' */
7088 * Special case: If a kmalloc of a doms_cur partition (array of
7089 * cpumask) fails, then fallback to a single sched domain,
7090 * as determined by the single cpumask fallback_doms.
7092 static cpumask_var_t fallback_doms;
7095 * arch_update_cpu_topology lets virtualized architectures update the
7096 * cpu core maps. It is supposed to return 1 if the topology changed
7097 * or 0 if it stayed the same.
7099 int __attribute__((weak)) arch_update_cpu_topology(void)
7104 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7107 cpumask_var_t *doms;
7109 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7112 for (i = 0; i < ndoms; i++) {
7113 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7114 free_sched_domains(doms, i);
7121 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7124 for (i = 0; i < ndoms; i++)
7125 free_cpumask_var(doms[i]);
7130 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7131 * For now this just excludes isolated cpus, but could be used to
7132 * exclude other special cases in the future.
7134 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7138 arch_update_cpu_topology();
7140 doms_cur = alloc_sched_domains(ndoms_cur);
7142 doms_cur = &fallback_doms;
7143 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7145 err = build_sched_domains(doms_cur[0]);
7146 register_sched_domain_sysctl();
7151 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7152 struct cpumask *tmpmask)
7154 free_sched_groups(cpu_map, tmpmask);
7158 * Detach sched domains from a group of cpus specified in cpu_map
7159 * These cpus will now be attached to the NULL domain
7161 static void detach_destroy_domains(const struct cpumask *cpu_map)
7163 /* Save because hotplug lock held. */
7164 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7167 for_each_cpu(i, cpu_map)
7168 cpu_attach_domain(NULL, &def_root_domain, i);
7169 synchronize_sched();
7170 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7173 /* handle null as "default" */
7174 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7175 struct sched_domain_attr *new, int idx_new)
7177 struct sched_domain_attr tmp;
7184 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7185 new ? (new + idx_new) : &tmp,
7186 sizeof(struct sched_domain_attr));
7190 * Partition sched domains as specified by the 'ndoms_new'
7191 * cpumasks in the array doms_new[] of cpumasks. This compares
7192 * doms_new[] to the current sched domain partitioning, doms_cur[].
7193 * It destroys each deleted domain and builds each new domain.
7195 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7196 * The masks don't intersect (don't overlap.) We should setup one
7197 * sched domain for each mask. CPUs not in any of the cpumasks will
7198 * not be load balanced. If the same cpumask appears both in the
7199 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7202 * The passed in 'doms_new' should be allocated using
7203 * alloc_sched_domains. This routine takes ownership of it and will
7204 * free_sched_domains it when done with it. If the caller failed the
7205 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7206 * and partition_sched_domains() will fallback to the single partition
7207 * 'fallback_doms', it also forces the domains to be rebuilt.
7209 * If doms_new == NULL it will be replaced with cpu_online_mask.
7210 * ndoms_new == 0 is a special case for destroying existing domains,
7211 * and it will not create the default domain.
7213 * Call with hotplug lock held
7215 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7216 struct sched_domain_attr *dattr_new)
7221 mutex_lock(&sched_domains_mutex);
7223 /* always unregister in case we don't destroy any domains */
7224 unregister_sched_domain_sysctl();
7226 /* Let architecture update cpu core mappings. */
7227 new_topology = arch_update_cpu_topology();
7229 n = doms_new ? ndoms_new : 0;
7231 /* Destroy deleted domains */
7232 for (i = 0; i < ndoms_cur; i++) {
7233 for (j = 0; j < n && !new_topology; j++) {
7234 if (cpumask_equal(doms_cur[i], doms_new[j])
7235 && dattrs_equal(dattr_cur, i, dattr_new, j))
7238 /* no match - a current sched domain not in new doms_new[] */
7239 detach_destroy_domains(doms_cur[i]);
7244 if (doms_new == NULL) {
7246 doms_new = &fallback_doms;
7247 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7248 WARN_ON_ONCE(dattr_new);
7251 /* Build new domains */
7252 for (i = 0; i < ndoms_new; i++) {
7253 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7254 if (cpumask_equal(doms_new[i], doms_cur[j])
7255 && dattrs_equal(dattr_new, i, dattr_cur, j))
7258 /* no match - add a new doms_new */
7259 __build_sched_domains(doms_new[i],
7260 dattr_new ? dattr_new + i : NULL);
7265 /* Remember the new sched domains */
7266 if (doms_cur != &fallback_doms)
7267 free_sched_domains(doms_cur, ndoms_cur);
7268 kfree(dattr_cur); /* kfree(NULL) is safe */
7269 doms_cur = doms_new;
7270 dattr_cur = dattr_new;
7271 ndoms_cur = ndoms_new;
7273 register_sched_domain_sysctl();
7275 mutex_unlock(&sched_domains_mutex);
7278 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7279 static void arch_reinit_sched_domains(void)
7283 /* Destroy domains first to force the rebuild */
7284 partition_sched_domains(0, NULL, NULL);
7286 rebuild_sched_domains();
7290 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7292 unsigned int level = 0;
7294 if (sscanf(buf, "%u", &level) != 1)
7298 * level is always be positive so don't check for
7299 * level < POWERSAVINGS_BALANCE_NONE which is 0
7300 * What happens on 0 or 1 byte write,
7301 * need to check for count as well?
7304 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7308 sched_smt_power_savings = level;
7310 sched_mc_power_savings = level;
7312 arch_reinit_sched_domains();
7317 #ifdef CONFIG_SCHED_MC
7318 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7319 struct sysdev_class_attribute *attr,
7322 return sprintf(page, "%u\n", sched_mc_power_savings);
7324 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7325 struct sysdev_class_attribute *attr,
7326 const char *buf, size_t count)
7328 return sched_power_savings_store(buf, count, 0);
7330 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7331 sched_mc_power_savings_show,
7332 sched_mc_power_savings_store);
7335 #ifdef CONFIG_SCHED_SMT
7336 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7337 struct sysdev_class_attribute *attr,
7340 return sprintf(page, "%u\n", sched_smt_power_savings);
7342 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7343 struct sysdev_class_attribute *attr,
7344 const char *buf, size_t count)
7346 return sched_power_savings_store(buf, count, 1);
7348 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7349 sched_smt_power_savings_show,
7350 sched_smt_power_savings_store);
7353 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7357 #ifdef CONFIG_SCHED_SMT
7359 err = sysfs_create_file(&cls->kset.kobj,
7360 &attr_sched_smt_power_savings.attr);
7362 #ifdef CONFIG_SCHED_MC
7363 if (!err && mc_capable())
7364 err = sysfs_create_file(&cls->kset.kobj,
7365 &attr_sched_mc_power_savings.attr);
7369 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7371 #ifndef CONFIG_CPUSETS
7373 * Add online and remove offline CPUs from the scheduler domains.
7374 * When cpusets are enabled they take over this function.
7376 static int update_sched_domains(struct notifier_block *nfb,
7377 unsigned long action, void *hcpu)
7381 case CPU_ONLINE_FROZEN:
7382 case CPU_DOWN_PREPARE:
7383 case CPU_DOWN_PREPARE_FROZEN:
7384 case CPU_DOWN_FAILED:
7385 case CPU_DOWN_FAILED_FROZEN:
7386 partition_sched_domains(1, NULL, NULL);
7395 static int update_runtime(struct notifier_block *nfb,
7396 unsigned long action, void *hcpu)
7398 int cpu = (int)(long)hcpu;
7401 case CPU_DOWN_PREPARE:
7402 case CPU_DOWN_PREPARE_FROZEN:
7403 disable_runtime(cpu_rq(cpu));
7406 case CPU_DOWN_FAILED:
7407 case CPU_DOWN_FAILED_FROZEN:
7409 case CPU_ONLINE_FROZEN:
7410 enable_runtime(cpu_rq(cpu));
7418 void __init sched_init_smp(void)
7420 cpumask_var_t non_isolated_cpus;
7422 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7423 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7425 #if defined(CONFIG_NUMA)
7426 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7428 BUG_ON(sched_group_nodes_bycpu == NULL);
7431 mutex_lock(&sched_domains_mutex);
7432 arch_init_sched_domains(cpu_active_mask);
7433 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7434 if (cpumask_empty(non_isolated_cpus))
7435 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7436 mutex_unlock(&sched_domains_mutex);
7439 #ifndef CONFIG_CPUSETS
7440 /* XXX: Theoretical race here - CPU may be hotplugged now */
7441 hotcpu_notifier(update_sched_domains, 0);
7444 /* RT runtime code needs to handle some hotplug events */
7445 hotcpu_notifier(update_runtime, 0);
7449 /* Move init over to a non-isolated CPU */
7450 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7452 sched_init_granularity();
7453 free_cpumask_var(non_isolated_cpus);
7455 init_sched_rt_class();
7458 void __init sched_init_smp(void)
7460 sched_init_granularity();
7462 #endif /* CONFIG_SMP */
7464 const_debug unsigned int sysctl_timer_migration = 1;
7466 int in_sched_functions(unsigned long addr)
7468 return in_lock_functions(addr) ||
7469 (addr >= (unsigned long)__sched_text_start
7470 && addr < (unsigned long)__sched_text_end);
7473 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7475 cfs_rq->tasks_timeline = RB_ROOT;
7476 INIT_LIST_HEAD(&cfs_rq->tasks);
7477 #ifdef CONFIG_FAIR_GROUP_SCHED
7480 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7483 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7485 struct rt_prio_array *array;
7488 array = &rt_rq->active;
7489 for (i = 0; i < MAX_RT_PRIO; i++) {
7490 INIT_LIST_HEAD(array->queue + i);
7491 __clear_bit(i, array->bitmap);
7493 /* delimiter for bitsearch: */
7494 __set_bit(MAX_RT_PRIO, array->bitmap);
7496 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7497 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7499 rt_rq->highest_prio.next = MAX_RT_PRIO;
7503 rt_rq->rt_nr_migratory = 0;
7504 rt_rq->overloaded = 0;
7505 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7509 rt_rq->rt_throttled = 0;
7510 rt_rq->rt_runtime = 0;
7511 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7513 #ifdef CONFIG_RT_GROUP_SCHED
7514 rt_rq->rt_nr_boosted = 0;
7519 #ifdef CONFIG_FAIR_GROUP_SCHED
7520 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7521 struct sched_entity *se, int cpu, int add,
7522 struct sched_entity *parent)
7524 struct rq *rq = cpu_rq(cpu);
7525 tg->cfs_rq[cpu] = cfs_rq;
7526 init_cfs_rq(cfs_rq, rq);
7529 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7532 /* se could be NULL for init_task_group */
7537 se->cfs_rq = &rq->cfs;
7539 se->cfs_rq = parent->my_q;
7542 se->load.weight = tg->shares;
7543 se->load.inv_weight = 0;
7544 se->parent = parent;
7548 #ifdef CONFIG_RT_GROUP_SCHED
7549 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7550 struct sched_rt_entity *rt_se, int cpu, int add,
7551 struct sched_rt_entity *parent)
7553 struct rq *rq = cpu_rq(cpu);
7555 tg->rt_rq[cpu] = rt_rq;
7556 init_rt_rq(rt_rq, rq);
7558 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7560 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7562 tg->rt_se[cpu] = rt_se;
7567 rt_se->rt_rq = &rq->rt;
7569 rt_se->rt_rq = parent->my_q;
7571 rt_se->my_q = rt_rq;
7572 rt_se->parent = parent;
7573 INIT_LIST_HEAD(&rt_se->run_list);
7577 void __init sched_init(void)
7580 unsigned long alloc_size = 0, ptr;
7582 #ifdef CONFIG_FAIR_GROUP_SCHED
7583 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7585 #ifdef CONFIG_RT_GROUP_SCHED
7586 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7588 #ifdef CONFIG_CPUMASK_OFFSTACK
7589 alloc_size += num_possible_cpus() * cpumask_size();
7592 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7594 #ifdef CONFIG_FAIR_GROUP_SCHED
7595 init_task_group.se = (struct sched_entity **)ptr;
7596 ptr += nr_cpu_ids * sizeof(void **);
7598 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7599 ptr += nr_cpu_ids * sizeof(void **);
7601 #endif /* CONFIG_FAIR_GROUP_SCHED */
7602 #ifdef CONFIG_RT_GROUP_SCHED
7603 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7604 ptr += nr_cpu_ids * sizeof(void **);
7606 init_task_group.rt_rq = (struct rt_rq **)ptr;
7607 ptr += nr_cpu_ids * sizeof(void **);
7609 #endif /* CONFIG_RT_GROUP_SCHED */
7610 #ifdef CONFIG_CPUMASK_OFFSTACK
7611 for_each_possible_cpu(i) {
7612 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7613 ptr += cpumask_size();
7615 #endif /* CONFIG_CPUMASK_OFFSTACK */
7619 init_defrootdomain();
7622 init_rt_bandwidth(&def_rt_bandwidth,
7623 global_rt_period(), global_rt_runtime());
7625 #ifdef CONFIG_RT_GROUP_SCHED
7626 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7627 global_rt_period(), global_rt_runtime());
7628 #endif /* CONFIG_RT_GROUP_SCHED */
7630 #ifdef CONFIG_CGROUP_SCHED
7631 list_add(&init_task_group.list, &task_groups);
7632 INIT_LIST_HEAD(&init_task_group.children);
7634 #endif /* CONFIG_CGROUP_SCHED */
7636 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7637 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7638 __alignof__(unsigned long));
7640 for_each_possible_cpu(i) {
7644 raw_spin_lock_init(&rq->lock);
7646 rq->calc_load_active = 0;
7647 rq->calc_load_update = jiffies + LOAD_FREQ;
7648 init_cfs_rq(&rq->cfs, rq);
7649 init_rt_rq(&rq->rt, rq);
7650 #ifdef CONFIG_FAIR_GROUP_SCHED
7651 init_task_group.shares = init_task_group_load;
7652 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7653 #ifdef CONFIG_CGROUP_SCHED
7655 * How much cpu bandwidth does init_task_group get?
7657 * In case of task-groups formed thr' the cgroup filesystem, it
7658 * gets 100% of the cpu resources in the system. This overall
7659 * system cpu resource is divided among the tasks of
7660 * init_task_group and its child task-groups in a fair manner,
7661 * based on each entity's (task or task-group's) weight
7662 * (se->load.weight).
7664 * In other words, if init_task_group has 10 tasks of weight
7665 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7666 * then A0's share of the cpu resource is:
7668 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7670 * We achieve this by letting init_task_group's tasks sit
7671 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7673 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7675 #endif /* CONFIG_FAIR_GROUP_SCHED */
7677 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7678 #ifdef CONFIG_RT_GROUP_SCHED
7679 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7680 #ifdef CONFIG_CGROUP_SCHED
7681 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7685 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7686 rq->cpu_load[j] = 0;
7690 rq->post_schedule = 0;
7691 rq->active_balance = 0;
7692 rq->next_balance = jiffies;
7696 rq->migration_thread = NULL;
7698 rq->avg_idle = 2*sysctl_sched_migration_cost;
7699 INIT_LIST_HEAD(&rq->migration_queue);
7700 rq_attach_root(rq, &def_root_domain);
7703 atomic_set(&rq->nr_iowait, 0);
7706 set_load_weight(&init_task);
7708 #ifdef CONFIG_PREEMPT_NOTIFIERS
7709 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7713 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7716 #ifdef CONFIG_RT_MUTEXES
7717 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7721 * The boot idle thread does lazy MMU switching as well:
7723 atomic_inc(&init_mm.mm_count);
7724 enter_lazy_tlb(&init_mm, current);
7727 * Make us the idle thread. Technically, schedule() should not be
7728 * called from this thread, however somewhere below it might be,
7729 * but because we are the idle thread, we just pick up running again
7730 * when this runqueue becomes "idle".
7732 init_idle(current, smp_processor_id());
7734 calc_load_update = jiffies + LOAD_FREQ;
7737 * During early bootup we pretend to be a normal task:
7739 current->sched_class = &fair_sched_class;
7741 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7742 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7745 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7746 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7748 /* May be allocated at isolcpus cmdline parse time */
7749 if (cpu_isolated_map == NULL)
7750 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7755 scheduler_running = 1;
7758 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7759 static inline int preempt_count_equals(int preempt_offset)
7761 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7763 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7766 void __might_sleep(const char *file, int line, int preempt_offset)
7769 static unsigned long prev_jiffy; /* ratelimiting */
7771 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7772 system_state != SYSTEM_RUNNING || oops_in_progress)
7774 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7776 prev_jiffy = jiffies;
7779 "BUG: sleeping function called from invalid context at %s:%d\n",
7782 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7783 in_atomic(), irqs_disabled(),
7784 current->pid, current->comm);
7786 debug_show_held_locks(current);
7787 if (irqs_disabled())
7788 print_irqtrace_events(current);
7792 EXPORT_SYMBOL(__might_sleep);
7795 #ifdef CONFIG_MAGIC_SYSRQ
7796 static void normalize_task(struct rq *rq, struct task_struct *p)
7800 on_rq = p->se.on_rq;
7802 deactivate_task(rq, p, 0);
7803 __setscheduler(rq, p, SCHED_NORMAL, 0);
7805 activate_task(rq, p, 0);
7806 resched_task(rq->curr);
7810 void normalize_rt_tasks(void)
7812 struct task_struct *g, *p;
7813 unsigned long flags;
7816 read_lock_irqsave(&tasklist_lock, flags);
7817 do_each_thread(g, p) {
7819 * Only normalize user tasks:
7824 p->se.exec_start = 0;
7825 #ifdef CONFIG_SCHEDSTATS
7826 p->se.statistics.wait_start = 0;
7827 p->se.statistics.sleep_start = 0;
7828 p->se.statistics.block_start = 0;
7833 * Renice negative nice level userspace
7836 if (TASK_NICE(p) < 0 && p->mm)
7837 set_user_nice(p, 0);
7841 raw_spin_lock(&p->pi_lock);
7842 rq = __task_rq_lock(p);
7844 normalize_task(rq, p);
7846 __task_rq_unlock(rq);
7847 raw_spin_unlock(&p->pi_lock);
7848 } while_each_thread(g, p);
7850 read_unlock_irqrestore(&tasklist_lock, flags);
7853 #endif /* CONFIG_MAGIC_SYSRQ */
7857 * These functions are only useful for the IA64 MCA handling.
7859 * They can only be called when the whole system has been
7860 * stopped - every CPU needs to be quiescent, and no scheduling
7861 * activity can take place. Using them for anything else would
7862 * be a serious bug, and as a result, they aren't even visible
7863 * under any other configuration.
7867 * curr_task - return the current task for a given cpu.
7868 * @cpu: the processor in question.
7870 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7872 struct task_struct *curr_task(int cpu)
7874 return cpu_curr(cpu);
7878 * set_curr_task - set the current task for a given cpu.
7879 * @cpu: the processor in question.
7880 * @p: the task pointer to set.
7882 * Description: This function must only be used when non-maskable interrupts
7883 * are serviced on a separate stack. It allows the architecture to switch the
7884 * notion of the current task on a cpu in a non-blocking manner. This function
7885 * must be called with all CPU's synchronized, and interrupts disabled, the
7886 * and caller must save the original value of the current task (see
7887 * curr_task() above) and restore that value before reenabling interrupts and
7888 * re-starting the system.
7890 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7892 void set_curr_task(int cpu, struct task_struct *p)
7899 #ifdef CONFIG_FAIR_GROUP_SCHED
7900 static void free_fair_sched_group(struct task_group *tg)
7904 for_each_possible_cpu(i) {
7906 kfree(tg->cfs_rq[i]);
7916 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7918 struct cfs_rq *cfs_rq;
7919 struct sched_entity *se;
7923 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7926 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7930 tg->shares = NICE_0_LOAD;
7932 for_each_possible_cpu(i) {
7935 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7936 GFP_KERNEL, cpu_to_node(i));
7940 se = kzalloc_node(sizeof(struct sched_entity),
7941 GFP_KERNEL, cpu_to_node(i));
7945 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
7956 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7958 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7959 &cpu_rq(cpu)->leaf_cfs_rq_list);
7962 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7964 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7966 #else /* !CONFG_FAIR_GROUP_SCHED */
7967 static inline void free_fair_sched_group(struct task_group *tg)
7972 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7977 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7981 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7984 #endif /* CONFIG_FAIR_GROUP_SCHED */
7986 #ifdef CONFIG_RT_GROUP_SCHED
7987 static void free_rt_sched_group(struct task_group *tg)
7991 destroy_rt_bandwidth(&tg->rt_bandwidth);
7993 for_each_possible_cpu(i) {
7995 kfree(tg->rt_rq[i]);
7997 kfree(tg->rt_se[i]);
8005 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8007 struct rt_rq *rt_rq;
8008 struct sched_rt_entity *rt_se;
8012 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8015 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8019 init_rt_bandwidth(&tg->rt_bandwidth,
8020 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8022 for_each_possible_cpu(i) {
8025 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8026 GFP_KERNEL, cpu_to_node(i));
8030 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8031 GFP_KERNEL, cpu_to_node(i));
8035 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8046 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8048 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8049 &cpu_rq(cpu)->leaf_rt_rq_list);
8052 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8054 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8056 #else /* !CONFIG_RT_GROUP_SCHED */
8057 static inline void free_rt_sched_group(struct task_group *tg)
8062 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8067 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8071 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8074 #endif /* CONFIG_RT_GROUP_SCHED */
8076 #ifdef CONFIG_CGROUP_SCHED
8077 static void free_sched_group(struct task_group *tg)
8079 free_fair_sched_group(tg);
8080 free_rt_sched_group(tg);
8084 /* allocate runqueue etc for a new task group */
8085 struct task_group *sched_create_group(struct task_group *parent)
8087 struct task_group *tg;
8088 unsigned long flags;
8091 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8093 return ERR_PTR(-ENOMEM);
8095 if (!alloc_fair_sched_group(tg, parent))
8098 if (!alloc_rt_sched_group(tg, parent))
8101 spin_lock_irqsave(&task_group_lock, flags);
8102 for_each_possible_cpu(i) {
8103 register_fair_sched_group(tg, i);
8104 register_rt_sched_group(tg, i);
8106 list_add_rcu(&tg->list, &task_groups);
8108 WARN_ON(!parent); /* root should already exist */
8110 tg->parent = parent;
8111 INIT_LIST_HEAD(&tg->children);
8112 list_add_rcu(&tg->siblings, &parent->children);
8113 spin_unlock_irqrestore(&task_group_lock, flags);
8118 free_sched_group(tg);
8119 return ERR_PTR(-ENOMEM);
8122 /* rcu callback to free various structures associated with a task group */
8123 static void free_sched_group_rcu(struct rcu_head *rhp)
8125 /* now it should be safe to free those cfs_rqs */
8126 free_sched_group(container_of(rhp, struct task_group, rcu));
8129 /* Destroy runqueue etc associated with a task group */
8130 void sched_destroy_group(struct task_group *tg)
8132 unsigned long flags;
8135 spin_lock_irqsave(&task_group_lock, flags);
8136 for_each_possible_cpu(i) {
8137 unregister_fair_sched_group(tg, i);
8138 unregister_rt_sched_group(tg, i);
8140 list_del_rcu(&tg->list);
8141 list_del_rcu(&tg->siblings);
8142 spin_unlock_irqrestore(&task_group_lock, flags);
8144 /* wait for possible concurrent references to cfs_rqs complete */
8145 call_rcu(&tg->rcu, free_sched_group_rcu);
8148 /* change task's runqueue when it moves between groups.
8149 * The caller of this function should have put the task in its new group
8150 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8151 * reflect its new group.
8153 void sched_move_task(struct task_struct *tsk)
8156 unsigned long flags;
8159 rq = task_rq_lock(tsk, &flags);
8161 running = task_current(rq, tsk);
8162 on_rq = tsk->se.on_rq;
8165 dequeue_task(rq, tsk, 0);
8166 if (unlikely(running))
8167 tsk->sched_class->put_prev_task(rq, tsk);
8169 set_task_rq(tsk, task_cpu(tsk));
8171 #ifdef CONFIG_FAIR_GROUP_SCHED
8172 if (tsk->sched_class->moved_group)
8173 tsk->sched_class->moved_group(tsk, on_rq);
8176 if (unlikely(running))
8177 tsk->sched_class->set_curr_task(rq);
8179 enqueue_task(rq, tsk, 0, false);
8181 task_rq_unlock(rq, &flags);
8183 #endif /* CONFIG_CGROUP_SCHED */
8185 #ifdef CONFIG_FAIR_GROUP_SCHED
8186 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8188 struct cfs_rq *cfs_rq = se->cfs_rq;
8193 dequeue_entity(cfs_rq, se, 0);
8195 se->load.weight = shares;
8196 se->load.inv_weight = 0;
8199 enqueue_entity(cfs_rq, se, 0);
8202 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8204 struct cfs_rq *cfs_rq = se->cfs_rq;
8205 struct rq *rq = cfs_rq->rq;
8206 unsigned long flags;
8208 raw_spin_lock_irqsave(&rq->lock, flags);
8209 __set_se_shares(se, shares);
8210 raw_spin_unlock_irqrestore(&rq->lock, flags);
8213 static DEFINE_MUTEX(shares_mutex);
8215 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8218 unsigned long flags;
8221 * We can't change the weight of the root cgroup.
8226 if (shares < MIN_SHARES)
8227 shares = MIN_SHARES;
8228 else if (shares > MAX_SHARES)
8229 shares = MAX_SHARES;
8231 mutex_lock(&shares_mutex);
8232 if (tg->shares == shares)
8235 spin_lock_irqsave(&task_group_lock, flags);
8236 for_each_possible_cpu(i)
8237 unregister_fair_sched_group(tg, i);
8238 list_del_rcu(&tg->siblings);
8239 spin_unlock_irqrestore(&task_group_lock, flags);
8241 /* wait for any ongoing reference to this group to finish */
8242 synchronize_sched();
8245 * Now we are free to modify the group's share on each cpu
8246 * w/o tripping rebalance_share or load_balance_fair.
8248 tg->shares = shares;
8249 for_each_possible_cpu(i) {
8253 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8254 set_se_shares(tg->se[i], shares);
8258 * Enable load balance activity on this group, by inserting it back on
8259 * each cpu's rq->leaf_cfs_rq_list.
8261 spin_lock_irqsave(&task_group_lock, flags);
8262 for_each_possible_cpu(i)
8263 register_fair_sched_group(tg, i);
8264 list_add_rcu(&tg->siblings, &tg->parent->children);
8265 spin_unlock_irqrestore(&task_group_lock, flags);
8267 mutex_unlock(&shares_mutex);
8271 unsigned long sched_group_shares(struct task_group *tg)
8277 #ifdef CONFIG_RT_GROUP_SCHED
8279 * Ensure that the real time constraints are schedulable.
8281 static DEFINE_MUTEX(rt_constraints_mutex);
8283 static unsigned long to_ratio(u64 period, u64 runtime)
8285 if (runtime == RUNTIME_INF)
8288 return div64_u64(runtime << 20, period);
8291 /* Must be called with tasklist_lock held */
8292 static inline int tg_has_rt_tasks(struct task_group *tg)
8294 struct task_struct *g, *p;
8296 do_each_thread(g, p) {
8297 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8299 } while_each_thread(g, p);
8304 struct rt_schedulable_data {
8305 struct task_group *tg;
8310 static int tg_schedulable(struct task_group *tg, void *data)
8312 struct rt_schedulable_data *d = data;
8313 struct task_group *child;
8314 unsigned long total, sum = 0;
8315 u64 period, runtime;
8317 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8318 runtime = tg->rt_bandwidth.rt_runtime;
8321 period = d->rt_period;
8322 runtime = d->rt_runtime;
8326 * Cannot have more runtime than the period.
8328 if (runtime > period && runtime != RUNTIME_INF)
8332 * Ensure we don't starve existing RT tasks.
8334 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8337 total = to_ratio(period, runtime);
8340 * Nobody can have more than the global setting allows.
8342 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8346 * The sum of our children's runtime should not exceed our own.
8348 list_for_each_entry_rcu(child, &tg->children, siblings) {
8349 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8350 runtime = child->rt_bandwidth.rt_runtime;
8352 if (child == d->tg) {
8353 period = d->rt_period;
8354 runtime = d->rt_runtime;
8357 sum += to_ratio(period, runtime);
8366 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8368 struct rt_schedulable_data data = {
8370 .rt_period = period,
8371 .rt_runtime = runtime,
8374 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8377 static int tg_set_bandwidth(struct task_group *tg,
8378 u64 rt_period, u64 rt_runtime)
8382 mutex_lock(&rt_constraints_mutex);
8383 read_lock(&tasklist_lock);
8384 err = __rt_schedulable(tg, rt_period, rt_runtime);
8388 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8389 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8390 tg->rt_bandwidth.rt_runtime = rt_runtime;
8392 for_each_possible_cpu(i) {
8393 struct rt_rq *rt_rq = tg->rt_rq[i];
8395 raw_spin_lock(&rt_rq->rt_runtime_lock);
8396 rt_rq->rt_runtime = rt_runtime;
8397 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8399 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8401 read_unlock(&tasklist_lock);
8402 mutex_unlock(&rt_constraints_mutex);
8407 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8409 u64 rt_runtime, rt_period;
8411 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8412 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8413 if (rt_runtime_us < 0)
8414 rt_runtime = RUNTIME_INF;
8416 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8419 long sched_group_rt_runtime(struct task_group *tg)
8423 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8426 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8427 do_div(rt_runtime_us, NSEC_PER_USEC);
8428 return rt_runtime_us;
8431 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8433 u64 rt_runtime, rt_period;
8435 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8436 rt_runtime = tg->rt_bandwidth.rt_runtime;
8441 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8444 long sched_group_rt_period(struct task_group *tg)
8448 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8449 do_div(rt_period_us, NSEC_PER_USEC);
8450 return rt_period_us;
8453 static int sched_rt_global_constraints(void)
8455 u64 runtime, period;
8458 if (sysctl_sched_rt_period <= 0)
8461 runtime = global_rt_runtime();
8462 period = global_rt_period();
8465 * Sanity check on the sysctl variables.
8467 if (runtime > period && runtime != RUNTIME_INF)
8470 mutex_lock(&rt_constraints_mutex);
8471 read_lock(&tasklist_lock);
8472 ret = __rt_schedulable(NULL, 0, 0);
8473 read_unlock(&tasklist_lock);
8474 mutex_unlock(&rt_constraints_mutex);
8479 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8481 /* Don't accept realtime tasks when there is no way for them to run */
8482 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8488 #else /* !CONFIG_RT_GROUP_SCHED */
8489 static int sched_rt_global_constraints(void)
8491 unsigned long flags;
8494 if (sysctl_sched_rt_period <= 0)
8498 * There's always some RT tasks in the root group
8499 * -- migration, kstopmachine etc..
8501 if (sysctl_sched_rt_runtime == 0)
8504 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8505 for_each_possible_cpu(i) {
8506 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8508 raw_spin_lock(&rt_rq->rt_runtime_lock);
8509 rt_rq->rt_runtime = global_rt_runtime();
8510 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8512 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8516 #endif /* CONFIG_RT_GROUP_SCHED */
8518 int sched_rt_handler(struct ctl_table *table, int write,
8519 void __user *buffer, size_t *lenp,
8523 int old_period, old_runtime;
8524 static DEFINE_MUTEX(mutex);
8527 old_period = sysctl_sched_rt_period;
8528 old_runtime = sysctl_sched_rt_runtime;
8530 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8532 if (!ret && write) {
8533 ret = sched_rt_global_constraints();
8535 sysctl_sched_rt_period = old_period;
8536 sysctl_sched_rt_runtime = old_runtime;
8538 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8539 def_rt_bandwidth.rt_period =
8540 ns_to_ktime(global_rt_period());
8543 mutex_unlock(&mutex);
8548 #ifdef CONFIG_CGROUP_SCHED
8550 /* return corresponding task_group object of a cgroup */
8551 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8553 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8554 struct task_group, css);
8557 static struct cgroup_subsys_state *
8558 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8560 struct task_group *tg, *parent;
8562 if (!cgrp->parent) {
8563 /* This is early initialization for the top cgroup */
8564 return &init_task_group.css;
8567 parent = cgroup_tg(cgrp->parent);
8568 tg = sched_create_group(parent);
8570 return ERR_PTR(-ENOMEM);
8576 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8578 struct task_group *tg = cgroup_tg(cgrp);
8580 sched_destroy_group(tg);
8584 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8586 #ifdef CONFIG_RT_GROUP_SCHED
8587 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8590 /* We don't support RT-tasks being in separate groups */
8591 if (tsk->sched_class != &fair_sched_class)
8598 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8599 struct task_struct *tsk, bool threadgroup)
8601 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8605 struct task_struct *c;
8607 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8608 retval = cpu_cgroup_can_attach_task(cgrp, c);
8620 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8621 struct cgroup *old_cont, struct task_struct *tsk,
8624 sched_move_task(tsk);
8626 struct task_struct *c;
8628 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8635 #ifdef CONFIG_FAIR_GROUP_SCHED
8636 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8639 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8642 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8644 struct task_group *tg = cgroup_tg(cgrp);
8646 return (u64) tg->shares;
8648 #endif /* CONFIG_FAIR_GROUP_SCHED */
8650 #ifdef CONFIG_RT_GROUP_SCHED
8651 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8654 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8657 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8659 return sched_group_rt_runtime(cgroup_tg(cgrp));
8662 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8665 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8668 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8670 return sched_group_rt_period(cgroup_tg(cgrp));
8672 #endif /* CONFIG_RT_GROUP_SCHED */
8674 static struct cftype cpu_files[] = {
8675 #ifdef CONFIG_FAIR_GROUP_SCHED
8678 .read_u64 = cpu_shares_read_u64,
8679 .write_u64 = cpu_shares_write_u64,
8682 #ifdef CONFIG_RT_GROUP_SCHED
8684 .name = "rt_runtime_us",
8685 .read_s64 = cpu_rt_runtime_read,
8686 .write_s64 = cpu_rt_runtime_write,
8689 .name = "rt_period_us",
8690 .read_u64 = cpu_rt_period_read_uint,
8691 .write_u64 = cpu_rt_period_write_uint,
8696 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8698 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8701 struct cgroup_subsys cpu_cgroup_subsys = {
8703 .create = cpu_cgroup_create,
8704 .destroy = cpu_cgroup_destroy,
8705 .can_attach = cpu_cgroup_can_attach,
8706 .attach = cpu_cgroup_attach,
8707 .populate = cpu_cgroup_populate,
8708 .subsys_id = cpu_cgroup_subsys_id,
8712 #endif /* CONFIG_CGROUP_SCHED */
8714 #ifdef CONFIG_CGROUP_CPUACCT
8717 * CPU accounting code for task groups.
8719 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8720 * (balbir@in.ibm.com).
8723 /* track cpu usage of a group of tasks and its child groups */
8725 struct cgroup_subsys_state css;
8726 /* cpuusage holds pointer to a u64-type object on every cpu */
8727 u64 __percpu *cpuusage;
8728 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8729 struct cpuacct *parent;
8732 struct cgroup_subsys cpuacct_subsys;
8734 /* return cpu accounting group corresponding to this container */
8735 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8737 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8738 struct cpuacct, css);
8741 /* return cpu accounting group to which this task belongs */
8742 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8744 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8745 struct cpuacct, css);
8748 /* create a new cpu accounting group */
8749 static struct cgroup_subsys_state *cpuacct_create(
8750 struct cgroup_subsys *ss, struct cgroup *cgrp)
8752 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8758 ca->cpuusage = alloc_percpu(u64);
8762 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8763 if (percpu_counter_init(&ca->cpustat[i], 0))
8764 goto out_free_counters;
8767 ca->parent = cgroup_ca(cgrp->parent);
8773 percpu_counter_destroy(&ca->cpustat[i]);
8774 free_percpu(ca->cpuusage);
8778 return ERR_PTR(-ENOMEM);
8781 /* destroy an existing cpu accounting group */
8783 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8785 struct cpuacct *ca = cgroup_ca(cgrp);
8788 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8789 percpu_counter_destroy(&ca->cpustat[i]);
8790 free_percpu(ca->cpuusage);
8794 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8796 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8799 #ifndef CONFIG_64BIT
8801 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8803 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8805 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8813 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8815 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8817 #ifndef CONFIG_64BIT
8819 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8821 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8823 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8829 /* return total cpu usage (in nanoseconds) of a group */
8830 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8832 struct cpuacct *ca = cgroup_ca(cgrp);
8833 u64 totalcpuusage = 0;
8836 for_each_present_cpu(i)
8837 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8839 return totalcpuusage;
8842 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8845 struct cpuacct *ca = cgroup_ca(cgrp);
8854 for_each_present_cpu(i)
8855 cpuacct_cpuusage_write(ca, i, 0);
8861 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8864 struct cpuacct *ca = cgroup_ca(cgroup);
8868 for_each_present_cpu(i) {
8869 percpu = cpuacct_cpuusage_read(ca, i);
8870 seq_printf(m, "%llu ", (unsigned long long) percpu);
8872 seq_printf(m, "\n");
8876 static const char *cpuacct_stat_desc[] = {
8877 [CPUACCT_STAT_USER] = "user",
8878 [CPUACCT_STAT_SYSTEM] = "system",
8881 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8882 struct cgroup_map_cb *cb)
8884 struct cpuacct *ca = cgroup_ca(cgrp);
8887 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8888 s64 val = percpu_counter_read(&ca->cpustat[i]);
8889 val = cputime64_to_clock_t(val);
8890 cb->fill(cb, cpuacct_stat_desc[i], val);
8895 static struct cftype files[] = {
8898 .read_u64 = cpuusage_read,
8899 .write_u64 = cpuusage_write,
8902 .name = "usage_percpu",
8903 .read_seq_string = cpuacct_percpu_seq_read,
8907 .read_map = cpuacct_stats_show,
8911 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8913 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8917 * charge this task's execution time to its accounting group.
8919 * called with rq->lock held.
8921 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8926 if (unlikely(!cpuacct_subsys.active))
8929 cpu = task_cpu(tsk);
8935 for (; ca; ca = ca->parent) {
8936 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8937 *cpuusage += cputime;
8944 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8945 * in cputime_t units. As a result, cpuacct_update_stats calls
8946 * percpu_counter_add with values large enough to always overflow the
8947 * per cpu batch limit causing bad SMP scalability.
8949 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8950 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8951 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8954 #define CPUACCT_BATCH \
8955 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8957 #define CPUACCT_BATCH 0
8961 * Charge the system/user time to the task's accounting group.
8963 static void cpuacct_update_stats(struct task_struct *tsk,
8964 enum cpuacct_stat_index idx, cputime_t val)
8967 int batch = CPUACCT_BATCH;
8969 if (unlikely(!cpuacct_subsys.active))
8976 __percpu_counter_add(&ca->cpustat[idx], val, batch);
8982 struct cgroup_subsys cpuacct_subsys = {
8984 .create = cpuacct_create,
8985 .destroy = cpuacct_destroy,
8986 .populate = cpuacct_populate,
8987 .subsys_id = cpuacct_subsys_id,
8989 #endif /* CONFIG_CGROUP_CPUACCT */
8993 int rcu_expedited_torture_stats(char *page)
8997 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
8999 void synchronize_sched_expedited(void)
9002 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9004 #else /* #ifndef CONFIG_SMP */
9006 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9007 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9009 #define RCU_EXPEDITED_STATE_POST -2
9010 #define RCU_EXPEDITED_STATE_IDLE -1
9012 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9014 int rcu_expedited_torture_stats(char *page)
9019 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9020 for_each_online_cpu(cpu) {
9021 cnt += sprintf(&page[cnt], " %d:%d",
9022 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9024 cnt += sprintf(&page[cnt], "\n");
9027 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9029 static long synchronize_sched_expedited_count;
9032 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9033 * approach to force grace period to end quickly. This consumes
9034 * significant time on all CPUs, and is thus not recommended for
9035 * any sort of common-case code.
9037 * Note that it is illegal to call this function while holding any
9038 * lock that is acquired by a CPU-hotplug notifier. Failing to
9039 * observe this restriction will result in deadlock.
9041 void synchronize_sched_expedited(void)
9044 unsigned long flags;
9045 bool need_full_sync = 0;
9047 struct migration_req *req;
9051 smp_mb(); /* ensure prior mod happens before capturing snap. */
9052 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9054 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9056 if (trycount++ < 10)
9057 udelay(trycount * num_online_cpus());
9059 synchronize_sched();
9062 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9063 smp_mb(); /* ensure test happens before caller kfree */
9068 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9069 for_each_online_cpu(cpu) {
9071 req = &per_cpu(rcu_migration_req, cpu);
9072 init_completion(&req->done);
9074 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9075 raw_spin_lock_irqsave(&rq->lock, flags);
9076 list_add(&req->list, &rq->migration_queue);
9077 raw_spin_unlock_irqrestore(&rq->lock, flags);
9078 wake_up_process(rq->migration_thread);
9080 for_each_online_cpu(cpu) {
9081 rcu_expedited_state = cpu;
9082 req = &per_cpu(rcu_migration_req, cpu);
9084 wait_for_completion(&req->done);
9085 raw_spin_lock_irqsave(&rq->lock, flags);
9086 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9088 req->dest_cpu = RCU_MIGRATION_IDLE;
9089 raw_spin_unlock_irqrestore(&rq->lock, flags);
9091 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9092 synchronize_sched_expedited_count++;
9093 mutex_unlock(&rcu_sched_expedited_mutex);
9096 synchronize_sched();
9098 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9100 #endif /* #else #ifndef CONFIG_SMP */