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];
495 unsigned char in_nohz_recently;
497 /* capture load from *all* tasks on this cpu: */
498 struct load_weight load;
499 unsigned long nr_load_updates;
505 #ifdef CONFIG_FAIR_GROUP_SCHED
506 /* list of leaf cfs_rq on this cpu: */
507 struct list_head leaf_cfs_rq_list;
509 #ifdef CONFIG_RT_GROUP_SCHED
510 struct list_head leaf_rt_rq_list;
514 * This is part of a global counter where only the total sum
515 * over all CPUs matters. A task can increase this counter on
516 * one CPU and if it got migrated afterwards it may decrease
517 * it on another CPU. Always updated under the runqueue lock:
519 unsigned long nr_uninterruptible;
521 struct task_struct *curr, *idle;
522 unsigned long next_balance;
523 struct mm_struct *prev_mm;
530 struct root_domain *rd;
531 struct sched_domain *sd;
533 unsigned char idle_at_tick;
534 /* For active balancing */
538 /* cpu of this runqueue: */
542 unsigned long avg_load_per_task;
544 struct task_struct *migration_thread;
545 struct list_head migration_queue;
553 /* calc_load related fields */
554 unsigned long calc_load_update;
555 long calc_load_active;
557 #ifdef CONFIG_SCHED_HRTICK
559 int hrtick_csd_pending;
560 struct call_single_data hrtick_csd;
562 struct hrtimer hrtick_timer;
565 #ifdef CONFIG_SCHEDSTATS
567 struct sched_info rq_sched_info;
568 unsigned long long rq_cpu_time;
569 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
571 /* sys_sched_yield() stats */
572 unsigned int yld_count;
574 /* schedule() stats */
575 unsigned int sched_switch;
576 unsigned int sched_count;
577 unsigned int sched_goidle;
579 /* try_to_wake_up() stats */
580 unsigned int ttwu_count;
581 unsigned int ttwu_local;
584 unsigned int bkl_count;
588 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
591 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
593 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
596 static inline int cpu_of(struct rq *rq)
606 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
607 * See detach_destroy_domains: synchronize_sched for details.
609 * The domain tree of any CPU may only be accessed from within
610 * preempt-disabled sections.
612 #define for_each_domain(cpu, __sd) \
613 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
615 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
616 #define this_rq() (&__get_cpu_var(runqueues))
617 #define task_rq(p) cpu_rq(task_cpu(p))
618 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
619 #define raw_rq() (&__raw_get_cpu_var(runqueues))
621 inline void update_rq_clock(struct rq *rq)
623 rq->clock = sched_clock_cpu(cpu_of(rq));
627 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
629 #ifdef CONFIG_SCHED_DEBUG
630 # define const_debug __read_mostly
632 # define const_debug static const
637 * @cpu: the processor in question.
639 * Returns true if the current cpu runqueue is locked.
640 * This interface allows printk to be called with the runqueue lock
641 * held and know whether or not it is OK to wake up the klogd.
643 int runqueue_is_locked(int cpu)
645 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
649 * Debugging: various feature bits
652 #define SCHED_FEAT(name, enabled) \
653 __SCHED_FEAT_##name ,
656 #include "sched_features.h"
661 #define SCHED_FEAT(name, enabled) \
662 (1UL << __SCHED_FEAT_##name) * enabled |
664 const_debug unsigned int sysctl_sched_features =
665 #include "sched_features.h"
670 #ifdef CONFIG_SCHED_DEBUG
671 #define SCHED_FEAT(name, enabled) \
674 static __read_mostly char *sched_feat_names[] = {
675 #include "sched_features.h"
681 static int sched_feat_show(struct seq_file *m, void *v)
685 for (i = 0; sched_feat_names[i]; i++) {
686 if (!(sysctl_sched_features & (1UL << i)))
688 seq_printf(m, "%s ", sched_feat_names[i]);
696 sched_feat_write(struct file *filp, const char __user *ubuf,
697 size_t cnt, loff_t *ppos)
707 if (copy_from_user(&buf, ubuf, cnt))
712 if (strncmp(buf, "NO_", 3) == 0) {
717 for (i = 0; sched_feat_names[i]; i++) {
718 int len = strlen(sched_feat_names[i]);
720 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
722 sysctl_sched_features &= ~(1UL << i);
724 sysctl_sched_features |= (1UL << i);
729 if (!sched_feat_names[i])
737 static int sched_feat_open(struct inode *inode, struct file *filp)
739 return single_open(filp, sched_feat_show, NULL);
742 static const struct file_operations sched_feat_fops = {
743 .open = sched_feat_open,
744 .write = sched_feat_write,
747 .release = single_release,
750 static __init int sched_init_debug(void)
752 debugfs_create_file("sched_features", 0644, NULL, NULL,
757 late_initcall(sched_init_debug);
761 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
764 * Number of tasks to iterate in a single balance run.
765 * Limited because this is done with IRQs disabled.
767 const_debug unsigned int sysctl_sched_nr_migrate = 32;
770 * ratelimit for updating the group shares.
773 unsigned int sysctl_sched_shares_ratelimit = 250000;
774 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
777 * Inject some fuzzyness into changing the per-cpu group shares
778 * this avoids remote rq-locks at the expense of fairness.
781 unsigned int sysctl_sched_shares_thresh = 4;
784 * period over which we average the RT time consumption, measured
789 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
792 * period over which we measure -rt task cpu usage in us.
795 unsigned int sysctl_sched_rt_period = 1000000;
797 static __read_mostly int scheduler_running;
800 * part of the period that we allow rt tasks to run in us.
803 int sysctl_sched_rt_runtime = 950000;
805 static inline u64 global_rt_period(void)
807 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
810 static inline u64 global_rt_runtime(void)
812 if (sysctl_sched_rt_runtime < 0)
815 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
818 #ifndef prepare_arch_switch
819 # define prepare_arch_switch(next) do { } while (0)
821 #ifndef finish_arch_switch
822 # define finish_arch_switch(prev) do { } while (0)
825 static inline int task_current(struct rq *rq, struct task_struct *p)
827 return rq->curr == p;
830 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
831 static inline int task_running(struct rq *rq, struct task_struct *p)
833 return task_current(rq, p);
836 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
840 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
842 #ifdef CONFIG_DEBUG_SPINLOCK
843 /* this is a valid case when another task releases the spinlock */
844 rq->lock.owner = current;
847 * If we are tracking spinlock dependencies then we have to
848 * fix up the runqueue lock - which gets 'carried over' from
851 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
853 raw_spin_unlock_irq(&rq->lock);
856 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
857 static inline int task_running(struct rq *rq, struct task_struct *p)
862 return task_current(rq, p);
866 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
870 * We can optimise this out completely for !SMP, because the
871 * SMP rebalancing from interrupt is the only thing that cares
876 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
877 raw_spin_unlock_irq(&rq->lock);
879 raw_spin_unlock(&rq->lock);
883 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
887 * After ->oncpu is cleared, the task can be moved to a different CPU.
888 * We must ensure this doesn't happen until the switch is completely
894 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
898 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
901 * __task_rq_lock - lock the runqueue a given task resides on.
902 * Must be called interrupts disabled.
904 static inline struct rq *__task_rq_lock(struct task_struct *p)
908 struct rq *rq = task_rq(p);
909 raw_spin_lock(&rq->lock);
910 if (likely(rq == task_rq(p)))
912 raw_spin_unlock(&rq->lock);
917 * task_rq_lock - lock the runqueue a given task resides on and disable
918 * interrupts. Note the ordering: we can safely lookup the task_rq without
919 * explicitly disabling preemption.
921 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
927 local_irq_save(*flags);
929 raw_spin_lock(&rq->lock);
930 if (likely(rq == task_rq(p)))
932 raw_spin_unlock_irqrestore(&rq->lock, *flags);
936 void task_rq_unlock_wait(struct task_struct *p)
938 struct rq *rq = task_rq(p);
940 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
941 raw_spin_unlock_wait(&rq->lock);
944 static void __task_rq_unlock(struct rq *rq)
947 raw_spin_unlock(&rq->lock);
950 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
953 raw_spin_unlock_irqrestore(&rq->lock, *flags);
957 * this_rq_lock - lock this runqueue and disable interrupts.
959 static struct rq *this_rq_lock(void)
966 raw_spin_lock(&rq->lock);
971 #ifdef CONFIG_SCHED_HRTICK
973 * Use HR-timers to deliver accurate preemption points.
975 * Its all a bit involved since we cannot program an hrt while holding the
976 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
979 * When we get rescheduled we reprogram the hrtick_timer outside of the
985 * - enabled by features
986 * - hrtimer is actually high res
988 static inline int hrtick_enabled(struct rq *rq)
990 if (!sched_feat(HRTICK))
992 if (!cpu_active(cpu_of(rq)))
994 return hrtimer_is_hres_active(&rq->hrtick_timer);
997 static void hrtick_clear(struct rq *rq)
999 if (hrtimer_active(&rq->hrtick_timer))
1000 hrtimer_cancel(&rq->hrtick_timer);
1004 * High-resolution timer tick.
1005 * Runs from hardirq context with interrupts disabled.
1007 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1009 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1011 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1013 raw_spin_lock(&rq->lock);
1014 update_rq_clock(rq);
1015 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1016 raw_spin_unlock(&rq->lock);
1018 return HRTIMER_NORESTART;
1023 * called from hardirq (IPI) context
1025 static void __hrtick_start(void *arg)
1027 struct rq *rq = arg;
1029 raw_spin_lock(&rq->lock);
1030 hrtimer_restart(&rq->hrtick_timer);
1031 rq->hrtick_csd_pending = 0;
1032 raw_spin_unlock(&rq->lock);
1036 * Called to set the hrtick timer state.
1038 * called with rq->lock held and irqs disabled
1040 static void hrtick_start(struct rq *rq, u64 delay)
1042 struct hrtimer *timer = &rq->hrtick_timer;
1043 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1045 hrtimer_set_expires(timer, time);
1047 if (rq == this_rq()) {
1048 hrtimer_restart(timer);
1049 } else if (!rq->hrtick_csd_pending) {
1050 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1051 rq->hrtick_csd_pending = 1;
1056 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1058 int cpu = (int)(long)hcpu;
1061 case CPU_UP_CANCELED:
1062 case CPU_UP_CANCELED_FROZEN:
1063 case CPU_DOWN_PREPARE:
1064 case CPU_DOWN_PREPARE_FROZEN:
1066 case CPU_DEAD_FROZEN:
1067 hrtick_clear(cpu_rq(cpu));
1074 static __init void init_hrtick(void)
1076 hotcpu_notifier(hotplug_hrtick, 0);
1080 * Called to set the hrtick timer state.
1082 * called with rq->lock held and irqs disabled
1084 static void hrtick_start(struct rq *rq, u64 delay)
1086 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1087 HRTIMER_MODE_REL_PINNED, 0);
1090 static inline void init_hrtick(void)
1093 #endif /* CONFIG_SMP */
1095 static void init_rq_hrtick(struct rq *rq)
1098 rq->hrtick_csd_pending = 0;
1100 rq->hrtick_csd.flags = 0;
1101 rq->hrtick_csd.func = __hrtick_start;
1102 rq->hrtick_csd.info = rq;
1105 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1106 rq->hrtick_timer.function = hrtick;
1108 #else /* CONFIG_SCHED_HRTICK */
1109 static inline void hrtick_clear(struct rq *rq)
1113 static inline void init_rq_hrtick(struct rq *rq)
1117 static inline void init_hrtick(void)
1120 #endif /* CONFIG_SCHED_HRTICK */
1123 * resched_task - mark a task 'to be rescheduled now'.
1125 * On UP this means the setting of the need_resched flag, on SMP it
1126 * might also involve a cross-CPU call to trigger the scheduler on
1131 #ifndef tsk_is_polling
1132 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1135 static void resched_task(struct task_struct *p)
1139 assert_raw_spin_locked(&task_rq(p)->lock);
1141 if (test_tsk_need_resched(p))
1144 set_tsk_need_resched(p);
1147 if (cpu == smp_processor_id())
1150 /* NEED_RESCHED must be visible before we test polling */
1152 if (!tsk_is_polling(p))
1153 smp_send_reschedule(cpu);
1156 static void resched_cpu(int cpu)
1158 struct rq *rq = cpu_rq(cpu);
1159 unsigned long flags;
1161 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1163 resched_task(cpu_curr(cpu));
1164 raw_spin_unlock_irqrestore(&rq->lock, flags);
1169 * When add_timer_on() enqueues a timer into the timer wheel of an
1170 * idle CPU then this timer might expire before the next timer event
1171 * which is scheduled to wake up that CPU. In case of a completely
1172 * idle system the next event might even be infinite time into the
1173 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1174 * leaves the inner idle loop so the newly added timer is taken into
1175 * account when the CPU goes back to idle and evaluates the timer
1176 * wheel for the next timer event.
1178 void wake_up_idle_cpu(int cpu)
1180 struct rq *rq = cpu_rq(cpu);
1182 if (cpu == smp_processor_id())
1186 * This is safe, as this function is called with the timer
1187 * wheel base lock of (cpu) held. When the CPU is on the way
1188 * to idle and has not yet set rq->curr to idle then it will
1189 * be serialized on the timer wheel base lock and take the new
1190 * timer into account automatically.
1192 if (rq->curr != rq->idle)
1196 * We can set TIF_RESCHED on the idle task of the other CPU
1197 * lockless. The worst case is that the other CPU runs the
1198 * idle task through an additional NOOP schedule()
1200 set_tsk_need_resched(rq->idle);
1202 /* NEED_RESCHED must be visible before we test polling */
1204 if (!tsk_is_polling(rq->idle))
1205 smp_send_reschedule(cpu);
1207 #endif /* CONFIG_NO_HZ */
1209 static u64 sched_avg_period(void)
1211 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1214 static void sched_avg_update(struct rq *rq)
1216 s64 period = sched_avg_period();
1218 while ((s64)(rq->clock - rq->age_stamp) > period) {
1219 rq->age_stamp += period;
1224 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1226 rq->rt_avg += rt_delta;
1227 sched_avg_update(rq);
1230 #else /* !CONFIG_SMP */
1231 static void resched_task(struct task_struct *p)
1233 assert_raw_spin_locked(&task_rq(p)->lock);
1234 set_tsk_need_resched(p);
1237 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1240 #endif /* CONFIG_SMP */
1242 #if BITS_PER_LONG == 32
1243 # define WMULT_CONST (~0UL)
1245 # define WMULT_CONST (1UL << 32)
1248 #define WMULT_SHIFT 32
1251 * Shift right and round:
1253 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1256 * delta *= weight / lw
1258 static unsigned long
1259 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1260 struct load_weight *lw)
1264 if (!lw->inv_weight) {
1265 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1268 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1272 tmp = (u64)delta_exec * weight;
1274 * Check whether we'd overflow the 64-bit multiplication:
1276 if (unlikely(tmp > WMULT_CONST))
1277 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1280 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1282 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1285 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1291 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1298 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1299 * of tasks with abnormal "nice" values across CPUs the contribution that
1300 * each task makes to its run queue's load is weighted according to its
1301 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1302 * scaled version of the new time slice allocation that they receive on time
1306 #define WEIGHT_IDLEPRIO 3
1307 #define WMULT_IDLEPRIO 1431655765
1310 * Nice levels are multiplicative, with a gentle 10% change for every
1311 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1312 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1313 * that remained on nice 0.
1315 * The "10% effect" is relative and cumulative: from _any_ nice level,
1316 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1317 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1318 * If a task goes up by ~10% and another task goes down by ~10% then
1319 * the relative distance between them is ~25%.)
1321 static const int prio_to_weight[40] = {
1322 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1323 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1324 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1325 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1326 /* 0 */ 1024, 820, 655, 526, 423,
1327 /* 5 */ 335, 272, 215, 172, 137,
1328 /* 10 */ 110, 87, 70, 56, 45,
1329 /* 15 */ 36, 29, 23, 18, 15,
1333 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1335 * In cases where the weight does not change often, we can use the
1336 * precalculated inverse to speed up arithmetics by turning divisions
1337 * into multiplications:
1339 static const u32 prio_to_wmult[40] = {
1340 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1341 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1342 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1343 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1344 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1345 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1346 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1347 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1350 /* Time spent by the tasks of the cpu accounting group executing in ... */
1351 enum cpuacct_stat_index {
1352 CPUACCT_STAT_USER, /* ... user mode */
1353 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1355 CPUACCT_STAT_NSTATS,
1358 #ifdef CONFIG_CGROUP_CPUACCT
1359 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1360 static void cpuacct_update_stats(struct task_struct *tsk,
1361 enum cpuacct_stat_index idx, cputime_t val);
1363 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1364 static inline void cpuacct_update_stats(struct task_struct *tsk,
1365 enum cpuacct_stat_index idx, cputime_t val) {}
1368 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1370 update_load_add(&rq->load, load);
1373 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1375 update_load_sub(&rq->load, load);
1378 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1379 typedef int (*tg_visitor)(struct task_group *, void *);
1382 * Iterate the full tree, calling @down when first entering a node and @up when
1383 * leaving it for the final time.
1385 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1387 struct task_group *parent, *child;
1391 parent = &root_task_group;
1393 ret = (*down)(parent, data);
1396 list_for_each_entry_rcu(child, &parent->children, siblings) {
1403 ret = (*up)(parent, data);
1408 parent = parent->parent;
1417 static int tg_nop(struct task_group *tg, void *data)
1424 /* Used instead of source_load when we know the type == 0 */
1425 static unsigned long weighted_cpuload(const int cpu)
1427 return cpu_rq(cpu)->load.weight;
1431 * Return a low guess at the load of a migration-source cpu weighted
1432 * according to the scheduling class and "nice" value.
1434 * We want to under-estimate the load of migration sources, to
1435 * balance conservatively.
1437 static unsigned long source_load(int cpu, int type)
1439 struct rq *rq = cpu_rq(cpu);
1440 unsigned long total = weighted_cpuload(cpu);
1442 if (type == 0 || !sched_feat(LB_BIAS))
1445 return min(rq->cpu_load[type-1], total);
1449 * Return a high guess at the load of a migration-target cpu weighted
1450 * according to the scheduling class and "nice" value.
1452 static unsigned long target_load(int cpu, int type)
1454 struct rq *rq = cpu_rq(cpu);
1455 unsigned long total = weighted_cpuload(cpu);
1457 if (type == 0 || !sched_feat(LB_BIAS))
1460 return max(rq->cpu_load[type-1], total);
1463 static struct sched_group *group_of(int cpu)
1465 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1473 static unsigned long power_of(int cpu)
1475 struct sched_group *group = group_of(cpu);
1478 return SCHED_LOAD_SCALE;
1480 return group->cpu_power;
1483 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1485 static unsigned long cpu_avg_load_per_task(int cpu)
1487 struct rq *rq = cpu_rq(cpu);
1488 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1491 rq->avg_load_per_task = rq->load.weight / nr_running;
1493 rq->avg_load_per_task = 0;
1495 return rq->avg_load_per_task;
1498 #ifdef CONFIG_FAIR_GROUP_SCHED
1500 static __read_mostly unsigned long *update_shares_data;
1502 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1505 * Calculate and set the cpu's group shares.
1507 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1508 unsigned long sd_shares,
1509 unsigned long sd_rq_weight,
1510 unsigned long *usd_rq_weight)
1512 unsigned long shares, rq_weight;
1515 rq_weight = usd_rq_weight[cpu];
1518 rq_weight = NICE_0_LOAD;
1522 * \Sum_j shares_j * rq_weight_i
1523 * shares_i = -----------------------------
1524 * \Sum_j rq_weight_j
1526 shares = (sd_shares * rq_weight) / sd_rq_weight;
1527 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1529 if (abs(shares - tg->se[cpu]->load.weight) >
1530 sysctl_sched_shares_thresh) {
1531 struct rq *rq = cpu_rq(cpu);
1532 unsigned long flags;
1534 raw_spin_lock_irqsave(&rq->lock, flags);
1535 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1536 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1537 __set_se_shares(tg->se[cpu], shares);
1538 raw_spin_unlock_irqrestore(&rq->lock, flags);
1543 * Re-compute the task group their per cpu shares over the given domain.
1544 * This needs to be done in a bottom-up fashion because the rq weight of a
1545 * parent group depends on the shares of its child groups.
1547 static int tg_shares_up(struct task_group *tg, void *data)
1549 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1550 unsigned long *usd_rq_weight;
1551 struct sched_domain *sd = data;
1552 unsigned long flags;
1558 local_irq_save(flags);
1559 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1561 for_each_cpu(i, sched_domain_span(sd)) {
1562 weight = tg->cfs_rq[i]->load.weight;
1563 usd_rq_weight[i] = weight;
1565 rq_weight += weight;
1567 * If there are currently no tasks on the cpu pretend there
1568 * is one of average load so that when a new task gets to
1569 * run here it will not get delayed by group starvation.
1572 weight = NICE_0_LOAD;
1574 sum_weight += weight;
1575 shares += tg->cfs_rq[i]->shares;
1579 rq_weight = sum_weight;
1581 if ((!shares && rq_weight) || shares > tg->shares)
1582 shares = tg->shares;
1584 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1585 shares = tg->shares;
1587 for_each_cpu(i, sched_domain_span(sd))
1588 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1590 local_irq_restore(flags);
1596 * Compute the cpu's hierarchical load factor for each task group.
1597 * This needs to be done in a top-down fashion because the load of a child
1598 * group is a fraction of its parents load.
1600 static int tg_load_down(struct task_group *tg, void *data)
1603 long cpu = (long)data;
1606 load = cpu_rq(cpu)->load.weight;
1608 load = tg->parent->cfs_rq[cpu]->h_load;
1609 load *= tg->cfs_rq[cpu]->shares;
1610 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1613 tg->cfs_rq[cpu]->h_load = load;
1618 static void update_shares(struct sched_domain *sd)
1623 if (root_task_group_empty())
1626 now = cpu_clock(raw_smp_processor_id());
1627 elapsed = now - sd->last_update;
1629 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1630 sd->last_update = now;
1631 walk_tg_tree(tg_nop, tg_shares_up, sd);
1635 static void update_h_load(long cpu)
1637 if (root_task_group_empty())
1640 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1645 static inline void update_shares(struct sched_domain *sd)
1651 #ifdef CONFIG_PREEMPT
1653 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1656 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1657 * way at the expense of forcing extra atomic operations in all
1658 * invocations. This assures that the double_lock is acquired using the
1659 * same underlying policy as the spinlock_t on this architecture, which
1660 * reduces latency compared to the unfair variant below. However, it
1661 * also adds more overhead and therefore may reduce throughput.
1663 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1664 __releases(this_rq->lock)
1665 __acquires(busiest->lock)
1666 __acquires(this_rq->lock)
1668 raw_spin_unlock(&this_rq->lock);
1669 double_rq_lock(this_rq, busiest);
1676 * Unfair double_lock_balance: Optimizes throughput at the expense of
1677 * latency by eliminating extra atomic operations when the locks are
1678 * already in proper order on entry. This favors lower cpu-ids and will
1679 * grant the double lock to lower cpus over higher ids under contention,
1680 * regardless of entry order into the function.
1682 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1683 __releases(this_rq->lock)
1684 __acquires(busiest->lock)
1685 __acquires(this_rq->lock)
1689 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1690 if (busiest < this_rq) {
1691 raw_spin_unlock(&this_rq->lock);
1692 raw_spin_lock(&busiest->lock);
1693 raw_spin_lock_nested(&this_rq->lock,
1694 SINGLE_DEPTH_NESTING);
1697 raw_spin_lock_nested(&busiest->lock,
1698 SINGLE_DEPTH_NESTING);
1703 #endif /* CONFIG_PREEMPT */
1706 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1708 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1710 if (unlikely(!irqs_disabled())) {
1711 /* printk() doesn't work good under rq->lock */
1712 raw_spin_unlock(&this_rq->lock);
1716 return _double_lock_balance(this_rq, busiest);
1719 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1720 __releases(busiest->lock)
1722 raw_spin_unlock(&busiest->lock);
1723 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1727 * double_rq_lock - safely lock two runqueues
1729 * Note this does not disable interrupts like task_rq_lock,
1730 * you need to do so manually before calling.
1732 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1733 __acquires(rq1->lock)
1734 __acquires(rq2->lock)
1736 BUG_ON(!irqs_disabled());
1738 raw_spin_lock(&rq1->lock);
1739 __acquire(rq2->lock); /* Fake it out ;) */
1742 raw_spin_lock(&rq1->lock);
1743 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1745 raw_spin_lock(&rq2->lock);
1746 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1749 update_rq_clock(rq1);
1750 update_rq_clock(rq2);
1754 * double_rq_unlock - safely unlock two runqueues
1756 * Note this does not restore interrupts like task_rq_unlock,
1757 * you need to do so manually after calling.
1759 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1760 __releases(rq1->lock)
1761 __releases(rq2->lock)
1763 raw_spin_unlock(&rq1->lock);
1765 raw_spin_unlock(&rq2->lock);
1767 __release(rq2->lock);
1772 #ifdef CONFIG_FAIR_GROUP_SCHED
1773 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1776 cfs_rq->shares = shares;
1781 static void calc_load_account_active(struct rq *this_rq);
1782 static void update_sysctl(void);
1783 static int get_update_sysctl_factor(void);
1785 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1787 set_task_rq(p, cpu);
1790 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1791 * successfuly executed on another CPU. We must ensure that updates of
1792 * per-task data have been completed by this moment.
1795 task_thread_info(p)->cpu = cpu;
1799 static const struct sched_class rt_sched_class;
1801 #define sched_class_highest (&rt_sched_class)
1802 #define for_each_class(class) \
1803 for (class = sched_class_highest; class; class = class->next)
1805 #include "sched_stats.h"
1807 static void inc_nr_running(struct rq *rq)
1812 static void dec_nr_running(struct rq *rq)
1817 static void set_load_weight(struct task_struct *p)
1819 if (task_has_rt_policy(p)) {
1820 p->se.load.weight = prio_to_weight[0] * 2;
1821 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1826 * SCHED_IDLE tasks get minimal weight:
1828 if (p->policy == SCHED_IDLE) {
1829 p->se.load.weight = WEIGHT_IDLEPRIO;
1830 p->se.load.inv_weight = WMULT_IDLEPRIO;
1834 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1835 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1838 static void update_avg(u64 *avg, u64 sample)
1840 s64 diff = sample - *avg;
1845 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1848 p->se.start_runtime = p->se.sum_exec_runtime;
1850 sched_info_queued(p);
1851 p->sched_class->enqueue_task(rq, p, wakeup, head);
1855 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1858 if (p->se.last_wakeup) {
1859 update_avg(&p->se.avg_overlap,
1860 p->se.sum_exec_runtime - p->se.last_wakeup);
1861 p->se.last_wakeup = 0;
1863 update_avg(&p->se.avg_wakeup,
1864 sysctl_sched_wakeup_granularity);
1868 sched_info_dequeued(p);
1869 p->sched_class->dequeue_task(rq, p, sleep);
1874 * activate_task - move a task to the runqueue.
1876 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1878 if (task_contributes_to_load(p))
1879 rq->nr_uninterruptible--;
1881 enqueue_task(rq, p, wakeup, false);
1886 * deactivate_task - remove a task from the runqueue.
1888 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1890 if (task_contributes_to_load(p))
1891 rq->nr_uninterruptible++;
1893 dequeue_task(rq, p, sleep);
1897 #include "sched_idletask.c"
1898 #include "sched_fair.c"
1899 #include "sched_rt.c"
1900 #ifdef CONFIG_SCHED_DEBUG
1901 # include "sched_debug.c"
1905 * __normal_prio - return the priority that is based on the static prio
1907 static inline int __normal_prio(struct task_struct *p)
1909 return p->static_prio;
1913 * Calculate the expected normal priority: i.e. priority
1914 * without taking RT-inheritance into account. Might be
1915 * boosted by interactivity modifiers. Changes upon fork,
1916 * setprio syscalls, and whenever the interactivity
1917 * estimator recalculates.
1919 static inline int normal_prio(struct task_struct *p)
1923 if (task_has_rt_policy(p))
1924 prio = MAX_RT_PRIO-1 - p->rt_priority;
1926 prio = __normal_prio(p);
1931 * Calculate the current priority, i.e. the priority
1932 * taken into account by the scheduler. This value might
1933 * be boosted by RT tasks, or might be boosted by
1934 * interactivity modifiers. Will be RT if the task got
1935 * RT-boosted. If not then it returns p->normal_prio.
1937 static int effective_prio(struct task_struct *p)
1939 p->normal_prio = normal_prio(p);
1941 * If we are RT tasks or we were boosted to RT priority,
1942 * keep the priority unchanged. Otherwise, update priority
1943 * to the normal priority:
1945 if (!rt_prio(p->prio))
1946 return p->normal_prio;
1951 * task_curr - is this task currently executing on a CPU?
1952 * @p: the task in question.
1954 inline int task_curr(const struct task_struct *p)
1956 return cpu_curr(task_cpu(p)) == p;
1959 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1960 const struct sched_class *prev_class,
1961 int oldprio, int running)
1963 if (prev_class != p->sched_class) {
1964 if (prev_class->switched_from)
1965 prev_class->switched_from(rq, p, running);
1966 p->sched_class->switched_to(rq, p, running);
1968 p->sched_class->prio_changed(rq, p, oldprio, running);
1973 * Is this task likely cache-hot:
1976 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1980 if (p->sched_class != &fair_sched_class)
1984 * Buddy candidates are cache hot:
1986 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
1987 (&p->se == cfs_rq_of(&p->se)->next ||
1988 &p->se == cfs_rq_of(&p->se)->last))
1991 if (sysctl_sched_migration_cost == -1)
1993 if (sysctl_sched_migration_cost == 0)
1996 delta = now - p->se.exec_start;
1998 return delta < (s64)sysctl_sched_migration_cost;
2001 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2003 #ifdef CONFIG_SCHED_DEBUG
2005 * We should never call set_task_cpu() on a blocked task,
2006 * ttwu() will sort out the placement.
2008 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2009 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2012 trace_sched_migrate_task(p, new_cpu);
2014 if (task_cpu(p) != new_cpu) {
2015 p->se.nr_migrations++;
2016 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2019 __set_task_cpu(p, new_cpu);
2022 struct migration_req {
2023 struct list_head list;
2025 struct task_struct *task;
2028 struct completion done;
2032 * The task's runqueue lock must be held.
2033 * Returns true if you have to wait for migration thread.
2036 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2038 struct rq *rq = task_rq(p);
2041 * If the task is not on a runqueue (and not running), then
2042 * the next wake-up will properly place the task.
2044 if (!p->se.on_rq && !task_running(rq, p))
2047 init_completion(&req->done);
2049 req->dest_cpu = dest_cpu;
2050 list_add(&req->list, &rq->migration_queue);
2056 * wait_task_context_switch - wait for a thread to complete at least one
2059 * @p must not be current.
2061 void wait_task_context_switch(struct task_struct *p)
2063 unsigned long nvcsw, nivcsw, flags;
2071 * The runqueue is assigned before the actual context
2072 * switch. We need to take the runqueue lock.
2074 * We could check initially without the lock but it is
2075 * very likely that we need to take the lock in every
2078 rq = task_rq_lock(p, &flags);
2079 running = task_running(rq, p);
2080 task_rq_unlock(rq, &flags);
2082 if (likely(!running))
2085 * The switch count is incremented before the actual
2086 * context switch. We thus wait for two switches to be
2087 * sure at least one completed.
2089 if ((p->nvcsw - nvcsw) > 1)
2091 if ((p->nivcsw - nivcsw) > 1)
2099 * wait_task_inactive - wait for a thread to unschedule.
2101 * If @match_state is nonzero, it's the @p->state value just checked and
2102 * not expected to change. If it changes, i.e. @p might have woken up,
2103 * then return zero. When we succeed in waiting for @p to be off its CPU,
2104 * we return a positive number (its total switch count). If a second call
2105 * a short while later returns the same number, the caller can be sure that
2106 * @p has remained unscheduled the whole time.
2108 * The caller must ensure that the task *will* unschedule sometime soon,
2109 * else this function might spin for a *long* time. This function can't
2110 * be called with interrupts off, or it may introduce deadlock with
2111 * smp_call_function() if an IPI is sent by the same process we are
2112 * waiting to become inactive.
2114 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2116 unsigned long flags;
2123 * We do the initial early heuristics without holding
2124 * any task-queue locks at all. We'll only try to get
2125 * the runqueue lock when things look like they will
2131 * If the task is actively running on another CPU
2132 * still, just relax and busy-wait without holding
2135 * NOTE! Since we don't hold any locks, it's not
2136 * even sure that "rq" stays as the right runqueue!
2137 * But we don't care, since "task_running()" will
2138 * return false if the runqueue has changed and p
2139 * is actually now running somewhere else!
2141 while (task_running(rq, p)) {
2142 if (match_state && unlikely(p->state != match_state))
2148 * Ok, time to look more closely! We need the rq
2149 * lock now, to be *sure*. If we're wrong, we'll
2150 * just go back and repeat.
2152 rq = task_rq_lock(p, &flags);
2153 trace_sched_wait_task(rq, p);
2154 running = task_running(rq, p);
2155 on_rq = p->se.on_rq;
2157 if (!match_state || p->state == match_state)
2158 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2159 task_rq_unlock(rq, &flags);
2162 * If it changed from the expected state, bail out now.
2164 if (unlikely(!ncsw))
2168 * Was it really running after all now that we
2169 * checked with the proper locks actually held?
2171 * Oops. Go back and try again..
2173 if (unlikely(running)) {
2179 * It's not enough that it's not actively running,
2180 * it must be off the runqueue _entirely_, and not
2183 * So if it was still runnable (but just not actively
2184 * running right now), it's preempted, and we should
2185 * yield - it could be a while.
2187 if (unlikely(on_rq)) {
2188 schedule_timeout_uninterruptible(1);
2193 * Ahh, all good. It wasn't running, and it wasn't
2194 * runnable, which means that it will never become
2195 * running in the future either. We're all done!
2204 * kick_process - kick a running thread to enter/exit the kernel
2205 * @p: the to-be-kicked thread
2207 * Cause a process which is running on another CPU to enter
2208 * kernel-mode, without any delay. (to get signals handled.)
2210 * NOTE: this function doesnt have to take the runqueue lock,
2211 * because all it wants to ensure is that the remote task enters
2212 * the kernel. If the IPI races and the task has been migrated
2213 * to another CPU then no harm is done and the purpose has been
2216 void kick_process(struct task_struct *p)
2222 if ((cpu != smp_processor_id()) && task_curr(p))
2223 smp_send_reschedule(cpu);
2226 EXPORT_SYMBOL_GPL(kick_process);
2227 #endif /* CONFIG_SMP */
2230 * task_oncpu_function_call - call a function on the cpu on which a task runs
2231 * @p: the task to evaluate
2232 * @func: the function to be called
2233 * @info: the function call argument
2235 * Calls the function @func when the task is currently running. This might
2236 * be on the current CPU, which just calls the function directly
2238 void task_oncpu_function_call(struct task_struct *p,
2239 void (*func) (void *info), void *info)
2246 smp_call_function_single(cpu, func, info, 1);
2251 static int select_fallback_rq(int cpu, struct task_struct *p)
2254 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2256 /* Look for allowed, online CPU in same node. */
2257 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2258 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2261 /* Any allowed, online CPU? */
2262 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2263 if (dest_cpu < nr_cpu_ids)
2266 /* No more Mr. Nice Guy. */
2267 if (dest_cpu >= nr_cpu_ids) {
2269 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2271 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2274 * Don't tell them about moving exiting tasks or
2275 * kernel threads (both mm NULL), since they never
2278 if (p->mm && printk_ratelimit()) {
2279 printk(KERN_INFO "process %d (%s) no "
2280 "longer affine to cpu%d\n",
2281 task_pid_nr(p), p->comm, cpu);
2289 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2290 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2293 * exec: is unstable, retry loop
2294 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2297 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2299 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2302 * In order not to call set_task_cpu() on a blocking task we need
2303 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2306 * Since this is common to all placement strategies, this lives here.
2308 * [ this allows ->select_task() to simply return task_cpu(p) and
2309 * not worry about this generic constraint ]
2311 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2313 cpu = select_fallback_rq(task_cpu(p), p);
2320 * try_to_wake_up - wake up a thread
2321 * @p: the to-be-woken-up thread
2322 * @state: the mask of task states that can be woken
2323 * @sync: do a synchronous wakeup?
2325 * Put it on the run-queue if it's not already there. The "current"
2326 * thread is always on the run-queue (except when the actual
2327 * re-schedule is in progress), and as such you're allowed to do
2328 * the simpler "current->state = TASK_RUNNING" to mark yourself
2329 * runnable without the overhead of this.
2331 * returns failure only if the task is already active.
2333 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2336 int cpu, orig_cpu, this_cpu, success = 0;
2337 unsigned long flags;
2338 struct rq *rq, *orig_rq;
2340 if (!sched_feat(SYNC_WAKEUPS))
2341 wake_flags &= ~WF_SYNC;
2343 this_cpu = get_cpu();
2346 rq = orig_rq = task_rq_lock(p, &flags);
2347 update_rq_clock(rq);
2348 if (!(p->state & state))
2358 if (unlikely(task_running(rq, p)))
2362 * In order to handle concurrent wakeups and release the rq->lock
2363 * we put the task in TASK_WAKING state.
2365 * First fix up the nr_uninterruptible count:
2367 if (task_contributes_to_load(p))
2368 rq->nr_uninterruptible--;
2369 p->state = TASK_WAKING;
2371 if (p->sched_class->task_waking)
2372 p->sched_class->task_waking(rq, p);
2374 __task_rq_unlock(rq);
2376 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2377 if (cpu != orig_cpu)
2378 set_task_cpu(p, cpu);
2380 rq = __task_rq_lock(p);
2381 update_rq_clock(rq);
2383 WARN_ON(p->state != TASK_WAKING);
2386 #ifdef CONFIG_SCHEDSTATS
2387 schedstat_inc(rq, ttwu_count);
2388 if (cpu == this_cpu)
2389 schedstat_inc(rq, ttwu_local);
2391 struct sched_domain *sd;
2392 for_each_domain(this_cpu, sd) {
2393 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2394 schedstat_inc(sd, ttwu_wake_remote);
2399 #endif /* CONFIG_SCHEDSTATS */
2402 #endif /* CONFIG_SMP */
2403 schedstat_inc(p, se.nr_wakeups);
2404 if (wake_flags & WF_SYNC)
2405 schedstat_inc(p, se.nr_wakeups_sync);
2406 if (orig_cpu != cpu)
2407 schedstat_inc(p, se.nr_wakeups_migrate);
2408 if (cpu == this_cpu)
2409 schedstat_inc(p, se.nr_wakeups_local);
2411 schedstat_inc(p, se.nr_wakeups_remote);
2412 activate_task(rq, p, 1);
2416 * Only attribute actual wakeups done by this task.
2418 if (!in_interrupt()) {
2419 struct sched_entity *se = ¤t->se;
2420 u64 sample = se->sum_exec_runtime;
2422 if (se->last_wakeup)
2423 sample -= se->last_wakeup;
2425 sample -= se->start_runtime;
2426 update_avg(&se->avg_wakeup, sample);
2428 se->last_wakeup = se->sum_exec_runtime;
2432 trace_sched_wakeup(rq, p, success);
2433 check_preempt_curr(rq, p, wake_flags);
2435 p->state = TASK_RUNNING;
2437 if (p->sched_class->task_woken)
2438 p->sched_class->task_woken(rq, p);
2440 if (unlikely(rq->idle_stamp)) {
2441 u64 delta = rq->clock - rq->idle_stamp;
2442 u64 max = 2*sysctl_sched_migration_cost;
2447 update_avg(&rq->avg_idle, delta);
2452 task_rq_unlock(rq, &flags);
2459 * wake_up_process - Wake up a specific process
2460 * @p: The process to be woken up.
2462 * Attempt to wake up the nominated process and move it to the set of runnable
2463 * processes. Returns 1 if the process was woken up, 0 if it was already
2466 * It may be assumed that this function implies a write memory barrier before
2467 * changing the task state if and only if any tasks are woken up.
2469 int wake_up_process(struct task_struct *p)
2471 return try_to_wake_up(p, TASK_ALL, 0);
2473 EXPORT_SYMBOL(wake_up_process);
2475 int wake_up_state(struct task_struct *p, unsigned int state)
2477 return try_to_wake_up(p, state, 0);
2481 * Perform scheduler related setup for a newly forked process p.
2482 * p is forked by current.
2484 * __sched_fork() is basic setup used by init_idle() too:
2486 static void __sched_fork(struct task_struct *p)
2488 p->se.exec_start = 0;
2489 p->se.sum_exec_runtime = 0;
2490 p->se.prev_sum_exec_runtime = 0;
2491 p->se.nr_migrations = 0;
2492 p->se.last_wakeup = 0;
2493 p->se.avg_overlap = 0;
2494 p->se.start_runtime = 0;
2495 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2497 #ifdef CONFIG_SCHEDSTATS
2498 p->se.wait_start = 0;
2500 p->se.wait_count = 0;
2503 p->se.sleep_start = 0;
2504 p->se.sleep_max = 0;
2505 p->se.sum_sleep_runtime = 0;
2507 p->se.block_start = 0;
2508 p->se.block_max = 0;
2510 p->se.slice_max = 0;
2512 p->se.nr_migrations_cold = 0;
2513 p->se.nr_failed_migrations_affine = 0;
2514 p->se.nr_failed_migrations_running = 0;
2515 p->se.nr_failed_migrations_hot = 0;
2516 p->se.nr_forced_migrations = 0;
2518 p->se.nr_wakeups = 0;
2519 p->se.nr_wakeups_sync = 0;
2520 p->se.nr_wakeups_migrate = 0;
2521 p->se.nr_wakeups_local = 0;
2522 p->se.nr_wakeups_remote = 0;
2523 p->se.nr_wakeups_affine = 0;
2524 p->se.nr_wakeups_affine_attempts = 0;
2525 p->se.nr_wakeups_passive = 0;
2526 p->se.nr_wakeups_idle = 0;
2530 INIT_LIST_HEAD(&p->rt.run_list);
2532 INIT_LIST_HEAD(&p->se.group_node);
2534 #ifdef CONFIG_PREEMPT_NOTIFIERS
2535 INIT_HLIST_HEAD(&p->preempt_notifiers);
2540 * fork()/clone()-time setup:
2542 void sched_fork(struct task_struct *p, int clone_flags)
2544 int cpu = get_cpu();
2548 * We mark the process as waking here. This guarantees that
2549 * nobody will actually run it, and a signal or other external
2550 * event cannot wake it up and insert it on the runqueue either.
2552 p->state = TASK_WAKING;
2555 * Revert to default priority/policy on fork if requested.
2557 if (unlikely(p->sched_reset_on_fork)) {
2558 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2559 p->policy = SCHED_NORMAL;
2560 p->normal_prio = p->static_prio;
2563 if (PRIO_TO_NICE(p->static_prio) < 0) {
2564 p->static_prio = NICE_TO_PRIO(0);
2565 p->normal_prio = p->static_prio;
2570 * We don't need the reset flag anymore after the fork. It has
2571 * fulfilled its duty:
2573 p->sched_reset_on_fork = 0;
2577 * Make sure we do not leak PI boosting priority to the child.
2579 p->prio = current->normal_prio;
2581 if (!rt_prio(p->prio))
2582 p->sched_class = &fair_sched_class;
2584 if (p->sched_class->task_fork)
2585 p->sched_class->task_fork(p);
2587 set_task_cpu(p, cpu);
2589 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2590 if (likely(sched_info_on()))
2591 memset(&p->sched_info, 0, sizeof(p->sched_info));
2593 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2596 #ifdef CONFIG_PREEMPT
2597 /* Want to start with kernel preemption disabled. */
2598 task_thread_info(p)->preempt_count = 1;
2600 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2606 * wake_up_new_task - wake up a newly created task for the first time.
2608 * This function will do some initial scheduler statistics housekeeping
2609 * that must be done for every newly created context, then puts the task
2610 * on the runqueue and wakes it.
2612 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2614 unsigned long flags;
2616 int cpu __maybe_unused = get_cpu();
2620 * Fork balancing, do it here and not earlier because:
2621 * - cpus_allowed can change in the fork path
2622 * - any previously selected cpu might disappear through hotplug
2624 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2625 * ->cpus_allowed is stable, we have preemption disabled, meaning
2626 * cpu_online_mask is stable.
2628 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2629 set_task_cpu(p, cpu);
2632 rq = task_rq_lock(p, &flags);
2633 BUG_ON(p->state != TASK_WAKING);
2634 p->state = TASK_RUNNING;
2635 update_rq_clock(rq);
2636 activate_task(rq, p, 0);
2637 trace_sched_wakeup_new(rq, p, 1);
2638 check_preempt_curr(rq, p, WF_FORK);
2640 if (p->sched_class->task_woken)
2641 p->sched_class->task_woken(rq, p);
2643 task_rq_unlock(rq, &flags);
2647 #ifdef CONFIG_PREEMPT_NOTIFIERS
2650 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2651 * @notifier: notifier struct to register
2653 void preempt_notifier_register(struct preempt_notifier *notifier)
2655 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2657 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2660 * preempt_notifier_unregister - no longer interested in preemption notifications
2661 * @notifier: notifier struct to unregister
2663 * This is safe to call from within a preemption notifier.
2665 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2667 hlist_del(¬ifier->link);
2669 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2671 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2673 struct preempt_notifier *notifier;
2674 struct hlist_node *node;
2676 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2677 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2681 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2682 struct task_struct *next)
2684 struct preempt_notifier *notifier;
2685 struct hlist_node *node;
2687 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2688 notifier->ops->sched_out(notifier, next);
2691 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2693 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2698 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2699 struct task_struct *next)
2703 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2706 * prepare_task_switch - prepare to switch tasks
2707 * @rq: the runqueue preparing to switch
2708 * @prev: the current task that is being switched out
2709 * @next: the task we are going to switch to.
2711 * This is called with the rq lock held and interrupts off. It must
2712 * be paired with a subsequent finish_task_switch after the context
2715 * prepare_task_switch sets up locking and calls architecture specific
2719 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2720 struct task_struct *next)
2722 fire_sched_out_preempt_notifiers(prev, next);
2723 prepare_lock_switch(rq, next);
2724 prepare_arch_switch(next);
2728 * finish_task_switch - clean up after a task-switch
2729 * @rq: runqueue associated with task-switch
2730 * @prev: the thread we just switched away from.
2732 * finish_task_switch must be called after the context switch, paired
2733 * with a prepare_task_switch call before the context switch.
2734 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2735 * and do any other architecture-specific cleanup actions.
2737 * Note that we may have delayed dropping an mm in context_switch(). If
2738 * so, we finish that here outside of the runqueue lock. (Doing it
2739 * with the lock held can cause deadlocks; see schedule() for
2742 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2743 __releases(rq->lock)
2745 struct mm_struct *mm = rq->prev_mm;
2751 * A task struct has one reference for the use as "current".
2752 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2753 * schedule one last time. The schedule call will never return, and
2754 * the scheduled task must drop that reference.
2755 * The test for TASK_DEAD must occur while the runqueue locks are
2756 * still held, otherwise prev could be scheduled on another cpu, die
2757 * there before we look at prev->state, and then the reference would
2759 * Manfred Spraul <manfred@colorfullife.com>
2761 prev_state = prev->state;
2762 finish_arch_switch(prev);
2763 perf_event_task_sched_in(current, cpu_of(rq));
2764 finish_lock_switch(rq, prev);
2766 fire_sched_in_preempt_notifiers(current);
2769 if (unlikely(prev_state == TASK_DEAD)) {
2771 * Remove function-return probe instances associated with this
2772 * task and put them back on the free list.
2774 kprobe_flush_task(prev);
2775 put_task_struct(prev);
2781 /* assumes rq->lock is held */
2782 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2784 if (prev->sched_class->pre_schedule)
2785 prev->sched_class->pre_schedule(rq, prev);
2788 /* rq->lock is NOT held, but preemption is disabled */
2789 static inline void post_schedule(struct rq *rq)
2791 if (rq->post_schedule) {
2792 unsigned long flags;
2794 raw_spin_lock_irqsave(&rq->lock, flags);
2795 if (rq->curr->sched_class->post_schedule)
2796 rq->curr->sched_class->post_schedule(rq);
2797 raw_spin_unlock_irqrestore(&rq->lock, flags);
2799 rq->post_schedule = 0;
2805 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2809 static inline void post_schedule(struct rq *rq)
2816 * schedule_tail - first thing a freshly forked thread must call.
2817 * @prev: the thread we just switched away from.
2819 asmlinkage void schedule_tail(struct task_struct *prev)
2820 __releases(rq->lock)
2822 struct rq *rq = this_rq();
2824 finish_task_switch(rq, prev);
2827 * FIXME: do we need to worry about rq being invalidated by the
2832 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2833 /* In this case, finish_task_switch does not reenable preemption */
2836 if (current->set_child_tid)
2837 put_user(task_pid_vnr(current), current->set_child_tid);
2841 * context_switch - switch to the new MM and the new
2842 * thread's register state.
2845 context_switch(struct rq *rq, struct task_struct *prev,
2846 struct task_struct *next)
2848 struct mm_struct *mm, *oldmm;
2850 prepare_task_switch(rq, prev, next);
2851 trace_sched_switch(rq, prev, next);
2853 oldmm = prev->active_mm;
2855 * For paravirt, this is coupled with an exit in switch_to to
2856 * combine the page table reload and the switch backend into
2859 arch_start_context_switch(prev);
2862 next->active_mm = oldmm;
2863 atomic_inc(&oldmm->mm_count);
2864 enter_lazy_tlb(oldmm, next);
2866 switch_mm(oldmm, mm, next);
2868 if (likely(!prev->mm)) {
2869 prev->active_mm = NULL;
2870 rq->prev_mm = oldmm;
2873 * Since the runqueue lock will be released by the next
2874 * task (which is an invalid locking op but in the case
2875 * of the scheduler it's an obvious special-case), so we
2876 * do an early lockdep release here:
2878 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2879 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2882 /* Here we just switch the register state and the stack. */
2883 switch_to(prev, next, prev);
2887 * this_rq must be evaluated again because prev may have moved
2888 * CPUs since it called schedule(), thus the 'rq' on its stack
2889 * frame will be invalid.
2891 finish_task_switch(this_rq(), prev);
2895 * nr_running, nr_uninterruptible and nr_context_switches:
2897 * externally visible scheduler statistics: current number of runnable
2898 * threads, current number of uninterruptible-sleeping threads, total
2899 * number of context switches performed since bootup.
2901 unsigned long nr_running(void)
2903 unsigned long i, sum = 0;
2905 for_each_online_cpu(i)
2906 sum += cpu_rq(i)->nr_running;
2911 unsigned long nr_uninterruptible(void)
2913 unsigned long i, sum = 0;
2915 for_each_possible_cpu(i)
2916 sum += cpu_rq(i)->nr_uninterruptible;
2919 * Since we read the counters lockless, it might be slightly
2920 * inaccurate. Do not allow it to go below zero though:
2922 if (unlikely((long)sum < 0))
2928 unsigned long long nr_context_switches(void)
2931 unsigned long long sum = 0;
2933 for_each_possible_cpu(i)
2934 sum += cpu_rq(i)->nr_switches;
2939 unsigned long nr_iowait(void)
2941 unsigned long i, sum = 0;
2943 for_each_possible_cpu(i)
2944 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2949 unsigned long nr_iowait_cpu(void)
2951 struct rq *this = this_rq();
2952 return atomic_read(&this->nr_iowait);
2955 unsigned long this_cpu_load(void)
2957 struct rq *this = this_rq();
2958 return this->cpu_load[0];
2962 /* Variables and functions for calc_load */
2963 static atomic_long_t calc_load_tasks;
2964 static unsigned long calc_load_update;
2965 unsigned long avenrun[3];
2966 EXPORT_SYMBOL(avenrun);
2969 * get_avenrun - get the load average array
2970 * @loads: pointer to dest load array
2971 * @offset: offset to add
2972 * @shift: shift count to shift the result left
2974 * These values are estimates at best, so no need for locking.
2976 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2978 loads[0] = (avenrun[0] + offset) << shift;
2979 loads[1] = (avenrun[1] + offset) << shift;
2980 loads[2] = (avenrun[2] + offset) << shift;
2983 static unsigned long
2984 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2987 load += active * (FIXED_1 - exp);
2988 return load >> FSHIFT;
2992 * calc_load - update the avenrun load estimates 10 ticks after the
2993 * CPUs have updated calc_load_tasks.
2995 void calc_global_load(void)
2997 unsigned long upd = calc_load_update + 10;
3000 if (time_before(jiffies, upd))
3003 active = atomic_long_read(&calc_load_tasks);
3004 active = active > 0 ? active * FIXED_1 : 0;
3006 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3007 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3008 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3010 calc_load_update += LOAD_FREQ;
3014 * Either called from update_cpu_load() or from a cpu going idle
3016 static void calc_load_account_active(struct rq *this_rq)
3018 long nr_active, delta;
3020 nr_active = this_rq->nr_running;
3021 nr_active += (long) this_rq->nr_uninterruptible;
3023 if (nr_active != this_rq->calc_load_active) {
3024 delta = nr_active - this_rq->calc_load_active;
3025 this_rq->calc_load_active = nr_active;
3026 atomic_long_add(delta, &calc_load_tasks);
3031 * Update rq->cpu_load[] statistics. This function is usually called every
3032 * scheduler tick (TICK_NSEC).
3034 static void update_cpu_load(struct rq *this_rq)
3036 unsigned long this_load = this_rq->load.weight;
3039 this_rq->nr_load_updates++;
3041 /* Update our load: */
3042 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3043 unsigned long old_load, new_load;
3045 /* scale is effectively 1 << i now, and >> i divides by scale */
3047 old_load = this_rq->cpu_load[i];
3048 new_load = this_load;
3050 * Round up the averaging division if load is increasing. This
3051 * prevents us from getting stuck on 9 if the load is 10, for
3054 if (new_load > old_load)
3055 new_load += scale-1;
3056 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3059 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3060 this_rq->calc_load_update += LOAD_FREQ;
3061 calc_load_account_active(this_rq);
3068 * sched_exec - execve() is a valuable balancing opportunity, because at
3069 * this point the task has the smallest effective memory and cache footprint.
3071 void sched_exec(void)
3073 struct task_struct *p = current;
3074 struct migration_req req;
3075 int dest_cpu, this_cpu;
3076 unsigned long flags;
3080 this_cpu = get_cpu();
3081 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3082 if (dest_cpu == this_cpu) {
3087 rq = task_rq_lock(p, &flags);
3091 * select_task_rq() can race against ->cpus_allowed
3093 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3094 || unlikely(!cpu_active(dest_cpu))) {
3095 task_rq_unlock(rq, &flags);
3099 /* force the process onto the specified CPU */
3100 if (migrate_task(p, dest_cpu, &req)) {
3101 /* Need to wait for migration thread (might exit: take ref). */
3102 struct task_struct *mt = rq->migration_thread;
3104 get_task_struct(mt);
3105 task_rq_unlock(rq, &flags);
3106 wake_up_process(mt);
3107 put_task_struct(mt);
3108 wait_for_completion(&req.done);
3112 task_rq_unlock(rq, &flags);
3117 DEFINE_PER_CPU(struct kernel_stat, kstat);
3119 EXPORT_PER_CPU_SYMBOL(kstat);
3122 * Return any ns on the sched_clock that have not yet been accounted in
3123 * @p in case that task is currently running.
3125 * Called with task_rq_lock() held on @rq.
3127 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3131 if (task_current(rq, p)) {
3132 update_rq_clock(rq);
3133 ns = rq->clock - p->se.exec_start;
3141 unsigned long long task_delta_exec(struct task_struct *p)
3143 unsigned long flags;
3147 rq = task_rq_lock(p, &flags);
3148 ns = do_task_delta_exec(p, rq);
3149 task_rq_unlock(rq, &flags);
3155 * Return accounted runtime for the task.
3156 * In case the task is currently running, return the runtime plus current's
3157 * pending runtime that have not been accounted yet.
3159 unsigned long long task_sched_runtime(struct task_struct *p)
3161 unsigned long flags;
3165 rq = task_rq_lock(p, &flags);
3166 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3167 task_rq_unlock(rq, &flags);
3173 * Return sum_exec_runtime for the thread group.
3174 * In case the task is currently running, return the sum plus current's
3175 * pending runtime that have not been accounted yet.
3177 * Note that the thread group might have other running tasks as well,
3178 * so the return value not includes other pending runtime that other
3179 * running tasks might have.
3181 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3183 struct task_cputime totals;
3184 unsigned long flags;
3188 rq = task_rq_lock(p, &flags);
3189 thread_group_cputime(p, &totals);
3190 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3191 task_rq_unlock(rq, &flags);
3197 * Account user cpu time to a process.
3198 * @p: the process that the cpu time gets accounted to
3199 * @cputime: the cpu time spent in user space since the last update
3200 * @cputime_scaled: cputime scaled by cpu frequency
3202 void account_user_time(struct task_struct *p, cputime_t cputime,
3203 cputime_t cputime_scaled)
3205 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3208 /* Add user time to process. */
3209 p->utime = cputime_add(p->utime, cputime);
3210 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3211 account_group_user_time(p, cputime);
3213 /* Add user time to cpustat. */
3214 tmp = cputime_to_cputime64(cputime);
3215 if (TASK_NICE(p) > 0)
3216 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3218 cpustat->user = cputime64_add(cpustat->user, tmp);
3220 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3221 /* Account for user time used */
3222 acct_update_integrals(p);
3226 * Account guest cpu time to a process.
3227 * @p: the process that the cpu time gets accounted to
3228 * @cputime: the cpu time spent in virtual machine since the last update
3229 * @cputime_scaled: cputime scaled by cpu frequency
3231 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3232 cputime_t cputime_scaled)
3235 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3237 tmp = cputime_to_cputime64(cputime);
3239 /* Add guest time to process. */
3240 p->utime = cputime_add(p->utime, cputime);
3241 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3242 account_group_user_time(p, cputime);
3243 p->gtime = cputime_add(p->gtime, cputime);
3245 /* Add guest time to cpustat. */
3246 if (TASK_NICE(p) > 0) {
3247 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3248 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3250 cpustat->user = cputime64_add(cpustat->user, tmp);
3251 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3256 * Account system cpu time to a process.
3257 * @p: the process that the cpu time gets accounted to
3258 * @hardirq_offset: the offset to subtract from hardirq_count()
3259 * @cputime: the cpu time spent in kernel space since the last update
3260 * @cputime_scaled: cputime scaled by cpu frequency
3262 void account_system_time(struct task_struct *p, int hardirq_offset,
3263 cputime_t cputime, cputime_t cputime_scaled)
3265 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3268 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3269 account_guest_time(p, cputime, cputime_scaled);
3273 /* Add system time to process. */
3274 p->stime = cputime_add(p->stime, cputime);
3275 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3276 account_group_system_time(p, cputime);
3278 /* Add system time to cpustat. */
3279 tmp = cputime_to_cputime64(cputime);
3280 if (hardirq_count() - hardirq_offset)
3281 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3282 else if (softirq_count())
3283 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3285 cpustat->system = cputime64_add(cpustat->system, tmp);
3287 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3289 /* Account for system time used */
3290 acct_update_integrals(p);
3294 * Account for involuntary wait time.
3295 * @steal: the cpu time spent in involuntary wait
3297 void account_steal_time(cputime_t cputime)
3299 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3300 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3302 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3306 * Account for idle time.
3307 * @cputime: the cpu time spent in idle wait
3309 void account_idle_time(cputime_t cputime)
3311 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3312 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3313 struct rq *rq = this_rq();
3315 if (atomic_read(&rq->nr_iowait) > 0)
3316 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3318 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3321 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3324 * Account a single tick of cpu time.
3325 * @p: the process that the cpu time gets accounted to
3326 * @user_tick: indicates if the tick is a user or a system tick
3328 void account_process_tick(struct task_struct *p, int user_tick)
3330 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3331 struct rq *rq = this_rq();
3334 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3335 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3336 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3339 account_idle_time(cputime_one_jiffy);
3343 * Account multiple ticks of steal time.
3344 * @p: the process from which the cpu time has been stolen
3345 * @ticks: number of stolen ticks
3347 void account_steal_ticks(unsigned long ticks)
3349 account_steal_time(jiffies_to_cputime(ticks));
3353 * Account multiple ticks of idle time.
3354 * @ticks: number of stolen ticks
3356 void account_idle_ticks(unsigned long ticks)
3358 account_idle_time(jiffies_to_cputime(ticks));
3364 * Use precise platform statistics if available:
3366 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3367 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3373 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3375 struct task_cputime cputime;
3377 thread_group_cputime(p, &cputime);
3379 *ut = cputime.utime;
3380 *st = cputime.stime;
3384 #ifndef nsecs_to_cputime
3385 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3388 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3390 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3393 * Use CFS's precise accounting:
3395 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3400 temp = (u64)(rtime * utime);
3401 do_div(temp, total);
3402 utime = (cputime_t)temp;
3407 * Compare with previous values, to keep monotonicity:
3409 p->prev_utime = max(p->prev_utime, utime);
3410 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3412 *ut = p->prev_utime;
3413 *st = p->prev_stime;
3417 * Must be called with siglock held.
3419 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3421 struct signal_struct *sig = p->signal;
3422 struct task_cputime cputime;
3423 cputime_t rtime, utime, total;
3425 thread_group_cputime(p, &cputime);
3427 total = cputime_add(cputime.utime, cputime.stime);
3428 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3433 temp = (u64)(rtime * cputime.utime);
3434 do_div(temp, total);
3435 utime = (cputime_t)temp;
3439 sig->prev_utime = max(sig->prev_utime, utime);
3440 sig->prev_stime = max(sig->prev_stime,
3441 cputime_sub(rtime, sig->prev_utime));
3443 *ut = sig->prev_utime;
3444 *st = sig->prev_stime;
3449 * This function gets called by the timer code, with HZ frequency.
3450 * We call it with interrupts disabled.
3452 * It also gets called by the fork code, when changing the parent's
3455 void scheduler_tick(void)
3457 int cpu = smp_processor_id();
3458 struct rq *rq = cpu_rq(cpu);
3459 struct task_struct *curr = rq->curr;
3463 raw_spin_lock(&rq->lock);
3464 update_rq_clock(rq);
3465 update_cpu_load(rq);
3466 curr->sched_class->task_tick(rq, curr, 0);
3467 raw_spin_unlock(&rq->lock);
3469 perf_event_task_tick(curr, cpu);
3472 rq->idle_at_tick = idle_cpu(cpu);
3473 trigger_load_balance(rq, cpu);
3477 notrace unsigned long get_parent_ip(unsigned long addr)
3479 if (in_lock_functions(addr)) {
3480 addr = CALLER_ADDR2;
3481 if (in_lock_functions(addr))
3482 addr = CALLER_ADDR3;
3487 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3488 defined(CONFIG_PREEMPT_TRACER))
3490 void __kprobes add_preempt_count(int val)
3492 #ifdef CONFIG_DEBUG_PREEMPT
3496 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3499 preempt_count() += val;
3500 #ifdef CONFIG_DEBUG_PREEMPT
3502 * Spinlock count overflowing soon?
3504 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3507 if (preempt_count() == val)
3508 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3510 EXPORT_SYMBOL(add_preempt_count);
3512 void __kprobes sub_preempt_count(int val)
3514 #ifdef CONFIG_DEBUG_PREEMPT
3518 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3521 * Is the spinlock portion underflowing?
3523 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3524 !(preempt_count() & PREEMPT_MASK)))
3528 if (preempt_count() == val)
3529 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3530 preempt_count() -= val;
3532 EXPORT_SYMBOL(sub_preempt_count);
3537 * Print scheduling while atomic bug:
3539 static noinline void __schedule_bug(struct task_struct *prev)
3541 struct pt_regs *regs = get_irq_regs();
3543 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3544 prev->comm, prev->pid, preempt_count());
3546 debug_show_held_locks(prev);
3548 if (irqs_disabled())
3549 print_irqtrace_events(prev);
3558 * Various schedule()-time debugging checks and statistics:
3560 static inline void schedule_debug(struct task_struct *prev)
3563 * Test if we are atomic. Since do_exit() needs to call into
3564 * schedule() atomically, we ignore that path for now.
3565 * Otherwise, whine if we are scheduling when we should not be.
3567 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3568 __schedule_bug(prev);
3570 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3572 schedstat_inc(this_rq(), sched_count);
3573 #ifdef CONFIG_SCHEDSTATS
3574 if (unlikely(prev->lock_depth >= 0)) {
3575 schedstat_inc(this_rq(), bkl_count);
3576 schedstat_inc(prev, sched_info.bkl_count);
3581 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3583 if (prev->state == TASK_RUNNING) {
3584 u64 runtime = prev->se.sum_exec_runtime;
3586 runtime -= prev->se.prev_sum_exec_runtime;
3587 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
3590 * In order to avoid avg_overlap growing stale when we are
3591 * indeed overlapping and hence not getting put to sleep, grow
3592 * the avg_overlap on preemption.
3594 * We use the average preemption runtime because that
3595 * correlates to the amount of cache footprint a task can
3598 update_avg(&prev->se.avg_overlap, runtime);
3600 prev->sched_class->put_prev_task(rq, prev);
3604 * Pick up the highest-prio task:
3606 static inline struct task_struct *
3607 pick_next_task(struct rq *rq)
3609 const struct sched_class *class;
3610 struct task_struct *p;
3613 * Optimization: we know that if all tasks are in
3614 * the fair class we can call that function directly:
3616 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3617 p = fair_sched_class.pick_next_task(rq);
3622 class = sched_class_highest;
3624 p = class->pick_next_task(rq);
3628 * Will never be NULL as the idle class always
3629 * returns a non-NULL p:
3631 class = class->next;
3636 * schedule() is the main scheduler function.
3638 asmlinkage void __sched schedule(void)
3640 struct task_struct *prev, *next;
3641 unsigned long *switch_count;
3647 cpu = smp_processor_id();
3651 switch_count = &prev->nivcsw;
3653 release_kernel_lock(prev);
3654 need_resched_nonpreemptible:
3656 schedule_debug(prev);
3658 if (sched_feat(HRTICK))
3661 raw_spin_lock_irq(&rq->lock);
3662 update_rq_clock(rq);
3663 clear_tsk_need_resched(prev);
3665 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3666 if (unlikely(signal_pending_state(prev->state, prev)))
3667 prev->state = TASK_RUNNING;
3669 deactivate_task(rq, prev, 1);
3670 switch_count = &prev->nvcsw;
3673 pre_schedule(rq, prev);
3675 if (unlikely(!rq->nr_running))
3676 idle_balance(cpu, rq);
3678 put_prev_task(rq, prev);
3679 next = pick_next_task(rq);
3681 if (likely(prev != next)) {
3682 sched_info_switch(prev, next);
3683 perf_event_task_sched_out(prev, next, cpu);
3689 context_switch(rq, prev, next); /* unlocks the rq */
3691 * the context switch might have flipped the stack from under
3692 * us, hence refresh the local variables.
3694 cpu = smp_processor_id();
3697 raw_spin_unlock_irq(&rq->lock);
3701 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3703 switch_count = &prev->nivcsw;
3704 goto need_resched_nonpreemptible;
3707 preempt_enable_no_resched();
3711 EXPORT_SYMBOL(schedule);
3713 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3715 * Look out! "owner" is an entirely speculative pointer
3716 * access and not reliable.
3718 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3723 if (!sched_feat(OWNER_SPIN))
3726 #ifdef CONFIG_DEBUG_PAGEALLOC
3728 * Need to access the cpu field knowing that
3729 * DEBUG_PAGEALLOC could have unmapped it if
3730 * the mutex owner just released it and exited.
3732 if (probe_kernel_address(&owner->cpu, cpu))
3739 * Even if the access succeeded (likely case),
3740 * the cpu field may no longer be valid.
3742 if (cpu >= nr_cpumask_bits)
3746 * We need to validate that we can do a
3747 * get_cpu() and that we have the percpu area.
3749 if (!cpu_online(cpu))
3756 * Owner changed, break to re-assess state.
3758 if (lock->owner != owner)
3762 * Is that owner really running on that cpu?
3764 if (task_thread_info(rq->curr) != owner || need_resched())
3774 #ifdef CONFIG_PREEMPT
3776 * this is the entry point to schedule() from in-kernel preemption
3777 * off of preempt_enable. Kernel preemptions off return from interrupt
3778 * occur there and call schedule directly.
3780 asmlinkage void __sched preempt_schedule(void)
3782 struct thread_info *ti = current_thread_info();
3785 * If there is a non-zero preempt_count or interrupts are disabled,
3786 * we do not want to preempt the current task. Just return..
3788 if (likely(ti->preempt_count || irqs_disabled()))
3792 add_preempt_count(PREEMPT_ACTIVE);
3794 sub_preempt_count(PREEMPT_ACTIVE);
3797 * Check again in case we missed a preemption opportunity
3798 * between schedule and now.
3801 } while (need_resched());
3803 EXPORT_SYMBOL(preempt_schedule);
3806 * this is the entry point to schedule() from kernel preemption
3807 * off of irq context.
3808 * Note, that this is called and return with irqs disabled. This will
3809 * protect us against recursive calling from irq.
3811 asmlinkage void __sched preempt_schedule_irq(void)
3813 struct thread_info *ti = current_thread_info();
3815 /* Catch callers which need to be fixed */
3816 BUG_ON(ti->preempt_count || !irqs_disabled());
3819 add_preempt_count(PREEMPT_ACTIVE);
3822 local_irq_disable();
3823 sub_preempt_count(PREEMPT_ACTIVE);
3826 * Check again in case we missed a preemption opportunity
3827 * between schedule and now.
3830 } while (need_resched());
3833 #endif /* CONFIG_PREEMPT */
3835 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3838 return try_to_wake_up(curr->private, mode, wake_flags);
3840 EXPORT_SYMBOL(default_wake_function);
3843 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3844 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3845 * number) then we wake all the non-exclusive tasks and one exclusive task.
3847 * There are circumstances in which we can try to wake a task which has already
3848 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3849 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3851 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3852 int nr_exclusive, int wake_flags, void *key)
3854 wait_queue_t *curr, *next;
3856 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3857 unsigned flags = curr->flags;
3859 if (curr->func(curr, mode, wake_flags, key) &&
3860 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3866 * __wake_up - wake up threads blocked on a waitqueue.
3868 * @mode: which threads
3869 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3870 * @key: is directly passed to the wakeup function
3872 * It may be assumed that this function implies a write memory barrier before
3873 * changing the task state if and only if any tasks are woken up.
3875 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3876 int nr_exclusive, void *key)
3878 unsigned long flags;
3880 spin_lock_irqsave(&q->lock, flags);
3881 __wake_up_common(q, mode, nr_exclusive, 0, key);
3882 spin_unlock_irqrestore(&q->lock, flags);
3884 EXPORT_SYMBOL(__wake_up);
3887 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3889 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3891 __wake_up_common(q, mode, 1, 0, NULL);
3894 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3896 __wake_up_common(q, mode, 1, 0, key);
3900 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3902 * @mode: which threads
3903 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3904 * @key: opaque value to be passed to wakeup targets
3906 * The sync wakeup differs that the waker knows that it will schedule
3907 * away soon, so while the target thread will be woken up, it will not
3908 * be migrated to another CPU - ie. the two threads are 'synchronized'
3909 * with each other. This can prevent needless bouncing between CPUs.
3911 * On UP it can prevent extra preemption.
3913 * It may be assumed that this function implies a write memory barrier before
3914 * changing the task state if and only if any tasks are woken up.
3916 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3917 int nr_exclusive, void *key)
3919 unsigned long flags;
3920 int wake_flags = WF_SYNC;
3925 if (unlikely(!nr_exclusive))
3928 spin_lock_irqsave(&q->lock, flags);
3929 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3930 spin_unlock_irqrestore(&q->lock, flags);
3932 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3935 * __wake_up_sync - see __wake_up_sync_key()
3937 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3939 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3941 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3944 * complete: - signals a single thread waiting on this completion
3945 * @x: holds the state of this particular completion
3947 * This will wake up a single thread waiting on this completion. Threads will be
3948 * awakened in the same order in which they were queued.
3950 * See also complete_all(), wait_for_completion() and related routines.
3952 * It may be assumed that this function implies a write memory barrier before
3953 * changing the task state if and only if any tasks are woken up.
3955 void complete(struct completion *x)
3957 unsigned long flags;
3959 spin_lock_irqsave(&x->wait.lock, flags);
3961 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3962 spin_unlock_irqrestore(&x->wait.lock, flags);
3964 EXPORT_SYMBOL(complete);
3967 * complete_all: - signals all threads waiting on this completion
3968 * @x: holds the state of this particular completion
3970 * This will wake up all threads waiting on this particular completion event.
3972 * It may be assumed that this function implies a write memory barrier before
3973 * changing the task state if and only if any tasks are woken up.
3975 void complete_all(struct completion *x)
3977 unsigned long flags;
3979 spin_lock_irqsave(&x->wait.lock, flags);
3980 x->done += UINT_MAX/2;
3981 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3982 spin_unlock_irqrestore(&x->wait.lock, flags);
3984 EXPORT_SYMBOL(complete_all);
3986 static inline long __sched
3987 do_wait_for_common(struct completion *x, long timeout, int state)
3990 DECLARE_WAITQUEUE(wait, current);
3992 wait.flags |= WQ_FLAG_EXCLUSIVE;
3993 __add_wait_queue_tail(&x->wait, &wait);
3995 if (signal_pending_state(state, current)) {
3996 timeout = -ERESTARTSYS;
3999 __set_current_state(state);
4000 spin_unlock_irq(&x->wait.lock);
4001 timeout = schedule_timeout(timeout);
4002 spin_lock_irq(&x->wait.lock);
4003 } while (!x->done && timeout);
4004 __remove_wait_queue(&x->wait, &wait);
4009 return timeout ?: 1;
4013 wait_for_common(struct completion *x, long timeout, int state)
4017 spin_lock_irq(&x->wait.lock);
4018 timeout = do_wait_for_common(x, timeout, state);
4019 spin_unlock_irq(&x->wait.lock);
4024 * wait_for_completion: - waits for completion of a task
4025 * @x: holds the state of this particular completion
4027 * This waits to be signaled for completion of a specific task. It is NOT
4028 * interruptible and there is no timeout.
4030 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4031 * and interrupt capability. Also see complete().
4033 void __sched wait_for_completion(struct completion *x)
4035 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4037 EXPORT_SYMBOL(wait_for_completion);
4040 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4041 * @x: holds the state of this particular completion
4042 * @timeout: timeout value in jiffies
4044 * This waits for either a completion of a specific task to be signaled or for a
4045 * specified timeout to expire. The timeout is in jiffies. It is not
4048 unsigned long __sched
4049 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4051 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4053 EXPORT_SYMBOL(wait_for_completion_timeout);
4056 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4057 * @x: holds the state of this particular completion
4059 * This waits for completion of a specific task to be signaled. It is
4062 int __sched wait_for_completion_interruptible(struct completion *x)
4064 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4065 if (t == -ERESTARTSYS)
4069 EXPORT_SYMBOL(wait_for_completion_interruptible);
4072 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4073 * @x: holds the state of this particular completion
4074 * @timeout: timeout value in jiffies
4076 * This waits for either a completion of a specific task to be signaled or for a
4077 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4079 unsigned long __sched
4080 wait_for_completion_interruptible_timeout(struct completion *x,
4081 unsigned long timeout)
4083 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4085 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4088 * wait_for_completion_killable: - waits for completion of a task (killable)
4089 * @x: holds the state of this particular completion
4091 * This waits to be signaled for completion of a specific task. It can be
4092 * interrupted by a kill signal.
4094 int __sched wait_for_completion_killable(struct completion *x)
4096 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4097 if (t == -ERESTARTSYS)
4101 EXPORT_SYMBOL(wait_for_completion_killable);
4104 * try_wait_for_completion - try to decrement a completion without blocking
4105 * @x: completion structure
4107 * Returns: 0 if a decrement cannot be done without blocking
4108 * 1 if a decrement succeeded.
4110 * If a completion is being used as a counting completion,
4111 * attempt to decrement the counter without blocking. This
4112 * enables us to avoid waiting if the resource the completion
4113 * is protecting is not available.
4115 bool try_wait_for_completion(struct completion *x)
4117 unsigned long flags;
4120 spin_lock_irqsave(&x->wait.lock, flags);
4125 spin_unlock_irqrestore(&x->wait.lock, flags);
4128 EXPORT_SYMBOL(try_wait_for_completion);
4131 * completion_done - Test to see if a completion has any waiters
4132 * @x: completion structure
4134 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4135 * 1 if there are no waiters.
4138 bool completion_done(struct completion *x)
4140 unsigned long flags;
4143 spin_lock_irqsave(&x->wait.lock, flags);
4146 spin_unlock_irqrestore(&x->wait.lock, flags);
4149 EXPORT_SYMBOL(completion_done);
4152 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4154 unsigned long flags;
4157 init_waitqueue_entry(&wait, current);
4159 __set_current_state(state);
4161 spin_lock_irqsave(&q->lock, flags);
4162 __add_wait_queue(q, &wait);
4163 spin_unlock(&q->lock);
4164 timeout = schedule_timeout(timeout);
4165 spin_lock_irq(&q->lock);
4166 __remove_wait_queue(q, &wait);
4167 spin_unlock_irqrestore(&q->lock, flags);
4172 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4174 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4176 EXPORT_SYMBOL(interruptible_sleep_on);
4179 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4181 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4183 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4185 void __sched sleep_on(wait_queue_head_t *q)
4187 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4189 EXPORT_SYMBOL(sleep_on);
4191 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4193 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4195 EXPORT_SYMBOL(sleep_on_timeout);
4197 #ifdef CONFIG_RT_MUTEXES
4200 * rt_mutex_setprio - set the current priority of a task
4202 * @prio: prio value (kernel-internal form)
4204 * This function changes the 'effective' priority of a task. It does
4205 * not touch ->normal_prio like __setscheduler().
4207 * Used by the rt_mutex code to implement priority inheritance logic.
4209 void rt_mutex_setprio(struct task_struct *p, int prio)
4211 unsigned long flags;
4212 int oldprio, on_rq, running;
4214 const struct sched_class *prev_class = p->sched_class;
4216 BUG_ON(prio < 0 || prio > MAX_PRIO);
4218 rq = task_rq_lock(p, &flags);
4219 update_rq_clock(rq);
4222 on_rq = p->se.on_rq;
4223 running = task_current(rq, p);
4225 dequeue_task(rq, p, 0);
4227 p->sched_class->put_prev_task(rq, p);
4230 p->sched_class = &rt_sched_class;
4232 p->sched_class = &fair_sched_class;
4237 p->sched_class->set_curr_task(rq);
4239 enqueue_task(rq, p, 0, oldprio < prio);
4241 check_class_changed(rq, p, prev_class, oldprio, running);
4243 task_rq_unlock(rq, &flags);
4248 void set_user_nice(struct task_struct *p, long nice)
4250 int old_prio, delta, on_rq;
4251 unsigned long flags;
4254 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4257 * We have to be careful, if called from sys_setpriority(),
4258 * the task might be in the middle of scheduling on another CPU.
4260 rq = task_rq_lock(p, &flags);
4261 update_rq_clock(rq);
4263 * The RT priorities are set via sched_setscheduler(), but we still
4264 * allow the 'normal' nice value to be set - but as expected
4265 * it wont have any effect on scheduling until the task is
4266 * SCHED_FIFO/SCHED_RR:
4268 if (task_has_rt_policy(p)) {
4269 p->static_prio = NICE_TO_PRIO(nice);
4272 on_rq = p->se.on_rq;
4274 dequeue_task(rq, p, 0);
4276 p->static_prio = NICE_TO_PRIO(nice);
4279 p->prio = effective_prio(p);
4280 delta = p->prio - old_prio;
4283 enqueue_task(rq, p, 0, false);
4285 * If the task increased its priority or is running and
4286 * lowered its priority, then reschedule its CPU:
4288 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4289 resched_task(rq->curr);
4292 task_rq_unlock(rq, &flags);
4294 EXPORT_SYMBOL(set_user_nice);
4297 * can_nice - check if a task can reduce its nice value
4301 int can_nice(const struct task_struct *p, const int nice)
4303 /* convert nice value [19,-20] to rlimit style value [1,40] */
4304 int nice_rlim = 20 - nice;
4306 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4307 capable(CAP_SYS_NICE));
4310 #ifdef __ARCH_WANT_SYS_NICE
4313 * sys_nice - change the priority of the current process.
4314 * @increment: priority increment
4316 * sys_setpriority is a more generic, but much slower function that
4317 * does similar things.
4319 SYSCALL_DEFINE1(nice, int, increment)
4324 * Setpriority might change our priority at the same moment.
4325 * We don't have to worry. Conceptually one call occurs first
4326 * and we have a single winner.
4328 if (increment < -40)
4333 nice = TASK_NICE(current) + increment;
4339 if (increment < 0 && !can_nice(current, nice))
4342 retval = security_task_setnice(current, nice);
4346 set_user_nice(current, nice);
4353 * task_prio - return the priority value of a given task.
4354 * @p: the task in question.
4356 * This is the priority value as seen by users in /proc.
4357 * RT tasks are offset by -200. Normal tasks are centered
4358 * around 0, value goes from -16 to +15.
4360 int task_prio(const struct task_struct *p)
4362 return p->prio - MAX_RT_PRIO;
4366 * task_nice - return the nice value of a given task.
4367 * @p: the task in question.
4369 int task_nice(const struct task_struct *p)
4371 return TASK_NICE(p);
4373 EXPORT_SYMBOL(task_nice);
4376 * idle_cpu - is a given cpu idle currently?
4377 * @cpu: the processor in question.
4379 int idle_cpu(int cpu)
4381 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4385 * idle_task - return the idle task for a given cpu.
4386 * @cpu: the processor in question.
4388 struct task_struct *idle_task(int cpu)
4390 return cpu_rq(cpu)->idle;
4394 * find_process_by_pid - find a process with a matching PID value.
4395 * @pid: the pid in question.
4397 static struct task_struct *find_process_by_pid(pid_t pid)
4399 return pid ? find_task_by_vpid(pid) : current;
4402 /* Actually do priority change: must hold rq lock. */
4404 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4406 BUG_ON(p->se.on_rq);
4409 p->rt_priority = prio;
4410 p->normal_prio = normal_prio(p);
4411 /* we are holding p->pi_lock already */
4412 p->prio = rt_mutex_getprio(p);
4413 if (rt_prio(p->prio))
4414 p->sched_class = &rt_sched_class;
4416 p->sched_class = &fair_sched_class;
4421 * check the target process has a UID that matches the current process's
4423 static bool check_same_owner(struct task_struct *p)
4425 const struct cred *cred = current_cred(), *pcred;
4429 pcred = __task_cred(p);
4430 match = (cred->euid == pcred->euid ||
4431 cred->euid == pcred->uid);
4436 static int __sched_setscheduler(struct task_struct *p, int policy,
4437 struct sched_param *param, bool user)
4439 int retval, oldprio, oldpolicy = -1, on_rq, running;
4440 unsigned long flags;
4441 const struct sched_class *prev_class = p->sched_class;
4445 /* may grab non-irq protected spin_locks */
4446 BUG_ON(in_interrupt());
4448 /* double check policy once rq lock held */
4450 reset_on_fork = p->sched_reset_on_fork;
4451 policy = oldpolicy = p->policy;
4453 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4454 policy &= ~SCHED_RESET_ON_FORK;
4456 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4457 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4458 policy != SCHED_IDLE)
4463 * Valid priorities for SCHED_FIFO and SCHED_RR are
4464 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4465 * SCHED_BATCH and SCHED_IDLE is 0.
4467 if (param->sched_priority < 0 ||
4468 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4469 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4471 if (rt_policy(policy) != (param->sched_priority != 0))
4475 * Allow unprivileged RT tasks to decrease priority:
4477 if (user && !capable(CAP_SYS_NICE)) {
4478 if (rt_policy(policy)) {
4479 unsigned long rlim_rtprio;
4481 if (!lock_task_sighand(p, &flags))
4483 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4484 unlock_task_sighand(p, &flags);
4486 /* can't set/change the rt policy */
4487 if (policy != p->policy && !rlim_rtprio)
4490 /* can't increase priority */
4491 if (param->sched_priority > p->rt_priority &&
4492 param->sched_priority > rlim_rtprio)
4496 * Like positive nice levels, dont allow tasks to
4497 * move out of SCHED_IDLE either:
4499 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4502 /* can't change other user's priorities */
4503 if (!check_same_owner(p))
4506 /* Normal users shall not reset the sched_reset_on_fork flag */
4507 if (p->sched_reset_on_fork && !reset_on_fork)
4512 #ifdef CONFIG_RT_GROUP_SCHED
4514 * Do not allow realtime tasks into groups that have no runtime
4517 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4518 task_group(p)->rt_bandwidth.rt_runtime == 0)
4522 retval = security_task_setscheduler(p, policy, param);
4528 * make sure no PI-waiters arrive (or leave) while we are
4529 * changing the priority of the task:
4531 raw_spin_lock_irqsave(&p->pi_lock, flags);
4533 * To be able to change p->policy safely, the apropriate
4534 * runqueue lock must be held.
4536 rq = __task_rq_lock(p);
4537 /* recheck policy now with rq lock held */
4538 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4539 policy = oldpolicy = -1;
4540 __task_rq_unlock(rq);
4541 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4544 update_rq_clock(rq);
4545 on_rq = p->se.on_rq;
4546 running = task_current(rq, p);
4548 deactivate_task(rq, p, 0);
4550 p->sched_class->put_prev_task(rq, p);
4552 p->sched_reset_on_fork = reset_on_fork;
4555 __setscheduler(rq, p, policy, param->sched_priority);
4558 p->sched_class->set_curr_task(rq);
4560 activate_task(rq, p, 0);
4562 check_class_changed(rq, p, prev_class, oldprio, running);
4564 __task_rq_unlock(rq);
4565 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4567 rt_mutex_adjust_pi(p);
4573 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4574 * @p: the task in question.
4575 * @policy: new policy.
4576 * @param: structure containing the new RT priority.
4578 * NOTE that the task may be already dead.
4580 int sched_setscheduler(struct task_struct *p, int policy,
4581 struct sched_param *param)
4583 return __sched_setscheduler(p, policy, param, true);
4585 EXPORT_SYMBOL_GPL(sched_setscheduler);
4588 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4589 * @p: the task in question.
4590 * @policy: new policy.
4591 * @param: structure containing the new RT priority.
4593 * Just like sched_setscheduler, only don't bother checking if the
4594 * current context has permission. For example, this is needed in
4595 * stop_machine(): we create temporary high priority worker threads,
4596 * but our caller might not have that capability.
4598 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4599 struct sched_param *param)
4601 return __sched_setscheduler(p, policy, param, false);
4605 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4607 struct sched_param lparam;
4608 struct task_struct *p;
4611 if (!param || pid < 0)
4613 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4618 p = find_process_by_pid(pid);
4620 retval = sched_setscheduler(p, policy, &lparam);
4627 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4628 * @pid: the pid in question.
4629 * @policy: new policy.
4630 * @param: structure containing the new RT priority.
4632 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4633 struct sched_param __user *, param)
4635 /* negative values for policy are not valid */
4639 return do_sched_setscheduler(pid, policy, param);
4643 * sys_sched_setparam - set/change the RT priority of a thread
4644 * @pid: the pid in question.
4645 * @param: structure containing the new RT priority.
4647 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4649 return do_sched_setscheduler(pid, -1, param);
4653 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4654 * @pid: the pid in question.
4656 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4658 struct task_struct *p;
4666 p = find_process_by_pid(pid);
4668 retval = security_task_getscheduler(p);
4671 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4678 * sys_sched_getparam - get the RT priority of a thread
4679 * @pid: the pid in question.
4680 * @param: structure containing the RT priority.
4682 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4684 struct sched_param lp;
4685 struct task_struct *p;
4688 if (!param || pid < 0)
4692 p = find_process_by_pid(pid);
4697 retval = security_task_getscheduler(p);
4701 lp.sched_priority = p->rt_priority;
4705 * This one might sleep, we cannot do it with a spinlock held ...
4707 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4716 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4718 cpumask_var_t cpus_allowed, new_mask;
4719 struct task_struct *p;
4725 p = find_process_by_pid(pid);
4732 /* Prevent p going away */
4736 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4740 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4742 goto out_free_cpus_allowed;
4745 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4748 retval = security_task_setscheduler(p, 0, NULL);
4752 cpuset_cpus_allowed(p, cpus_allowed);
4753 cpumask_and(new_mask, in_mask, cpus_allowed);
4755 retval = set_cpus_allowed_ptr(p, new_mask);
4758 cpuset_cpus_allowed(p, cpus_allowed);
4759 if (!cpumask_subset(new_mask, cpus_allowed)) {
4761 * We must have raced with a concurrent cpuset
4762 * update. Just reset the cpus_allowed to the
4763 * cpuset's cpus_allowed
4765 cpumask_copy(new_mask, cpus_allowed);
4770 free_cpumask_var(new_mask);
4771 out_free_cpus_allowed:
4772 free_cpumask_var(cpus_allowed);
4779 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4780 struct cpumask *new_mask)
4782 if (len < cpumask_size())
4783 cpumask_clear(new_mask);
4784 else if (len > cpumask_size())
4785 len = cpumask_size();
4787 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4791 * sys_sched_setaffinity - set the cpu affinity of a process
4792 * @pid: pid of the process
4793 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4794 * @user_mask_ptr: user-space pointer to the new cpu mask
4796 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4797 unsigned long __user *, user_mask_ptr)
4799 cpumask_var_t new_mask;
4802 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4805 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4807 retval = sched_setaffinity(pid, new_mask);
4808 free_cpumask_var(new_mask);
4812 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4814 struct task_struct *p;
4815 unsigned long flags;
4823 p = find_process_by_pid(pid);
4827 retval = security_task_getscheduler(p);
4831 rq = task_rq_lock(p, &flags);
4832 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4833 task_rq_unlock(rq, &flags);
4843 * sys_sched_getaffinity - get the cpu affinity of a process
4844 * @pid: pid of the process
4845 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4846 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4848 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4849 unsigned long __user *, user_mask_ptr)
4854 if (len < cpumask_size())
4857 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4860 ret = sched_getaffinity(pid, mask);
4862 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
4865 ret = cpumask_size();
4867 free_cpumask_var(mask);
4873 * sys_sched_yield - yield the current processor to other threads.
4875 * This function yields the current CPU to other tasks. If there are no
4876 * other threads running on this CPU then this function will return.
4878 SYSCALL_DEFINE0(sched_yield)
4880 struct rq *rq = this_rq_lock();
4882 schedstat_inc(rq, yld_count);
4883 current->sched_class->yield_task(rq);
4886 * Since we are going to call schedule() anyway, there's
4887 * no need to preempt or enable interrupts:
4889 __release(rq->lock);
4890 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4891 do_raw_spin_unlock(&rq->lock);
4892 preempt_enable_no_resched();
4899 static inline int should_resched(void)
4901 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4904 static void __cond_resched(void)
4906 add_preempt_count(PREEMPT_ACTIVE);
4908 sub_preempt_count(PREEMPT_ACTIVE);
4911 int __sched _cond_resched(void)
4913 if (should_resched()) {
4919 EXPORT_SYMBOL(_cond_resched);
4922 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4923 * call schedule, and on return reacquire the lock.
4925 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4926 * operations here to prevent schedule() from being called twice (once via
4927 * spin_unlock(), once by hand).
4929 int __cond_resched_lock(spinlock_t *lock)
4931 int resched = should_resched();
4934 lockdep_assert_held(lock);
4936 if (spin_needbreak(lock) || resched) {
4947 EXPORT_SYMBOL(__cond_resched_lock);
4949 int __sched __cond_resched_softirq(void)
4951 BUG_ON(!in_softirq());
4953 if (should_resched()) {
4961 EXPORT_SYMBOL(__cond_resched_softirq);
4964 * yield - yield the current processor to other threads.
4966 * This is a shortcut for kernel-space yielding - it marks the
4967 * thread runnable and calls sys_sched_yield().
4969 void __sched yield(void)
4971 set_current_state(TASK_RUNNING);
4974 EXPORT_SYMBOL(yield);
4977 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4978 * that process accounting knows that this is a task in IO wait state.
4980 void __sched io_schedule(void)
4982 struct rq *rq = raw_rq();
4984 delayacct_blkio_start();
4985 atomic_inc(&rq->nr_iowait);
4986 current->in_iowait = 1;
4988 current->in_iowait = 0;
4989 atomic_dec(&rq->nr_iowait);
4990 delayacct_blkio_end();
4992 EXPORT_SYMBOL(io_schedule);
4994 long __sched io_schedule_timeout(long timeout)
4996 struct rq *rq = raw_rq();
4999 delayacct_blkio_start();
5000 atomic_inc(&rq->nr_iowait);
5001 current->in_iowait = 1;
5002 ret = schedule_timeout(timeout);
5003 current->in_iowait = 0;
5004 atomic_dec(&rq->nr_iowait);
5005 delayacct_blkio_end();
5010 * sys_sched_get_priority_max - return maximum RT priority.
5011 * @policy: scheduling class.
5013 * this syscall returns the maximum rt_priority that can be used
5014 * by a given scheduling class.
5016 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5023 ret = MAX_USER_RT_PRIO-1;
5035 * sys_sched_get_priority_min - return minimum RT priority.
5036 * @policy: scheduling class.
5038 * this syscall returns the minimum rt_priority that can be used
5039 * by a given scheduling class.
5041 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5059 * sys_sched_rr_get_interval - return the default timeslice of a process.
5060 * @pid: pid of the process.
5061 * @interval: userspace pointer to the timeslice value.
5063 * this syscall writes the default timeslice value of a given process
5064 * into the user-space timespec buffer. A value of '0' means infinity.
5066 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5067 struct timespec __user *, interval)
5069 struct task_struct *p;
5070 unsigned int time_slice;
5071 unsigned long flags;
5081 p = find_process_by_pid(pid);
5085 retval = security_task_getscheduler(p);
5089 rq = task_rq_lock(p, &flags);
5090 time_slice = p->sched_class->get_rr_interval(rq, p);
5091 task_rq_unlock(rq, &flags);
5094 jiffies_to_timespec(time_slice, &t);
5095 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5103 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5105 void sched_show_task(struct task_struct *p)
5107 unsigned long free = 0;
5110 state = p->state ? __ffs(p->state) + 1 : 0;
5111 printk(KERN_INFO "%-13.13s %c", p->comm,
5112 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5113 #if BITS_PER_LONG == 32
5114 if (state == TASK_RUNNING)
5115 printk(KERN_CONT " running ");
5117 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5119 if (state == TASK_RUNNING)
5120 printk(KERN_CONT " running task ");
5122 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5124 #ifdef CONFIG_DEBUG_STACK_USAGE
5125 free = stack_not_used(p);
5127 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5128 task_pid_nr(p), task_pid_nr(p->real_parent),
5129 (unsigned long)task_thread_info(p)->flags);
5131 show_stack(p, NULL);
5134 void show_state_filter(unsigned long state_filter)
5136 struct task_struct *g, *p;
5138 #if BITS_PER_LONG == 32
5140 " task PC stack pid father\n");
5143 " task PC stack pid father\n");
5145 read_lock(&tasklist_lock);
5146 do_each_thread(g, p) {
5148 * reset the NMI-timeout, listing all files on a slow
5149 * console might take alot of time:
5151 touch_nmi_watchdog();
5152 if (!state_filter || (p->state & state_filter))
5154 } while_each_thread(g, p);
5156 touch_all_softlockup_watchdogs();
5158 #ifdef CONFIG_SCHED_DEBUG
5159 sysrq_sched_debug_show();
5161 read_unlock(&tasklist_lock);
5163 * Only show locks if all tasks are dumped:
5166 debug_show_all_locks();
5169 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5171 idle->sched_class = &idle_sched_class;
5175 * init_idle - set up an idle thread for a given CPU
5176 * @idle: task in question
5177 * @cpu: cpu the idle task belongs to
5179 * NOTE: this function does not set the idle thread's NEED_RESCHED
5180 * flag, to make booting more robust.
5182 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5184 struct rq *rq = cpu_rq(cpu);
5185 unsigned long flags;
5187 raw_spin_lock_irqsave(&rq->lock, flags);
5190 idle->state = TASK_RUNNING;
5191 idle->se.exec_start = sched_clock();
5193 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5194 __set_task_cpu(idle, cpu);
5196 rq->curr = rq->idle = idle;
5197 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5200 raw_spin_unlock_irqrestore(&rq->lock, flags);
5202 /* Set the preempt count _outside_ the spinlocks! */
5203 #if defined(CONFIG_PREEMPT)
5204 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5206 task_thread_info(idle)->preempt_count = 0;
5209 * The idle tasks have their own, simple scheduling class:
5211 idle->sched_class = &idle_sched_class;
5212 ftrace_graph_init_task(idle);
5216 * In a system that switches off the HZ timer nohz_cpu_mask
5217 * indicates which cpus entered this state. This is used
5218 * in the rcu update to wait only for active cpus. For system
5219 * which do not switch off the HZ timer nohz_cpu_mask should
5220 * always be CPU_BITS_NONE.
5222 cpumask_var_t nohz_cpu_mask;
5225 * Increase the granularity value when there are more CPUs,
5226 * because with more CPUs the 'effective latency' as visible
5227 * to users decreases. But the relationship is not linear,
5228 * so pick a second-best guess by going with the log2 of the
5231 * This idea comes from the SD scheduler of Con Kolivas:
5233 static int get_update_sysctl_factor(void)
5235 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5236 unsigned int factor;
5238 switch (sysctl_sched_tunable_scaling) {
5239 case SCHED_TUNABLESCALING_NONE:
5242 case SCHED_TUNABLESCALING_LINEAR:
5245 case SCHED_TUNABLESCALING_LOG:
5247 factor = 1 + ilog2(cpus);
5254 static void update_sysctl(void)
5256 unsigned int factor = get_update_sysctl_factor();
5258 #define SET_SYSCTL(name) \
5259 (sysctl_##name = (factor) * normalized_sysctl_##name)
5260 SET_SYSCTL(sched_min_granularity);
5261 SET_SYSCTL(sched_latency);
5262 SET_SYSCTL(sched_wakeup_granularity);
5263 SET_SYSCTL(sched_shares_ratelimit);
5267 static inline void sched_init_granularity(void)
5274 * This is how migration works:
5276 * 1) we queue a struct migration_req structure in the source CPU's
5277 * runqueue and wake up that CPU's migration thread.
5278 * 2) we down() the locked semaphore => thread blocks.
5279 * 3) migration thread wakes up (implicitly it forces the migrated
5280 * thread off the CPU)
5281 * 4) it gets the migration request and checks whether the migrated
5282 * task is still in the wrong runqueue.
5283 * 5) if it's in the wrong runqueue then the migration thread removes
5284 * it and puts it into the right queue.
5285 * 6) migration thread up()s the semaphore.
5286 * 7) we wake up and the migration is done.
5290 * Change a given task's CPU affinity. Migrate the thread to a
5291 * proper CPU and schedule it away if the CPU it's executing on
5292 * is removed from the allowed bitmask.
5294 * NOTE: the caller must have a valid reference to the task, the
5295 * task must not exit() & deallocate itself prematurely. The
5296 * call is not atomic; no spinlocks may be held.
5298 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5300 struct migration_req req;
5301 unsigned long flags;
5306 * Since we rely on wake-ups to migrate sleeping tasks, don't change
5307 * the ->cpus_allowed mask from under waking tasks, which would be
5308 * possible when we change rq->lock in ttwu(), so synchronize against
5309 * TASK_WAKING to avoid that.
5311 * Make an exception for freshly cloned tasks, since cpuset namespaces
5312 * might move the task about, we have to validate the target in
5313 * wake_up_new_task() anyway since the cpu might have gone away.
5316 while (p->state == TASK_WAKING && !(p->flags & PF_STARTING))
5319 rq = task_rq_lock(p, &flags);
5321 if (p->state == TASK_WAKING && !(p->flags & PF_STARTING)) {
5322 task_rq_unlock(rq, &flags);
5326 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5331 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5332 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5337 if (p->sched_class->set_cpus_allowed)
5338 p->sched_class->set_cpus_allowed(p, new_mask);
5340 cpumask_copy(&p->cpus_allowed, new_mask);
5341 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5344 /* Can the task run on the task's current CPU? If so, we're done */
5345 if (cpumask_test_cpu(task_cpu(p), new_mask))
5348 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5349 /* Need help from migration thread: drop lock and wait. */
5350 struct task_struct *mt = rq->migration_thread;
5352 get_task_struct(mt);
5353 task_rq_unlock(rq, &flags);
5354 wake_up_process(rq->migration_thread);
5355 put_task_struct(mt);
5356 wait_for_completion(&req.done);
5357 tlb_migrate_finish(p->mm);
5361 task_rq_unlock(rq, &flags);
5365 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5368 * Move (not current) task off this cpu, onto dest cpu. We're doing
5369 * this because either it can't run here any more (set_cpus_allowed()
5370 * away from this CPU, or CPU going down), or because we're
5371 * attempting to rebalance this task on exec (sched_exec).
5373 * So we race with normal scheduler movements, but that's OK, as long
5374 * as the task is no longer on this CPU.
5376 * Returns non-zero if task was successfully migrated.
5378 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5380 struct rq *rq_dest, *rq_src;
5383 if (unlikely(!cpu_active(dest_cpu)))
5386 rq_src = cpu_rq(src_cpu);
5387 rq_dest = cpu_rq(dest_cpu);
5389 double_rq_lock(rq_src, rq_dest);
5390 /* Already moved. */
5391 if (task_cpu(p) != src_cpu)
5393 /* Affinity changed (again). */
5394 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5398 * If we're not on a rq, the next wake-up will ensure we're
5402 deactivate_task(rq_src, p, 0);
5403 set_task_cpu(p, dest_cpu);
5404 activate_task(rq_dest, p, 0);
5405 check_preempt_curr(rq_dest, p, 0);
5410 double_rq_unlock(rq_src, rq_dest);
5414 #define RCU_MIGRATION_IDLE 0
5415 #define RCU_MIGRATION_NEED_QS 1
5416 #define RCU_MIGRATION_GOT_QS 2
5417 #define RCU_MIGRATION_MUST_SYNC 3
5420 * migration_thread - this is a highprio system thread that performs
5421 * thread migration by bumping thread off CPU then 'pushing' onto
5424 static int migration_thread(void *data)
5427 int cpu = (long)data;
5431 BUG_ON(rq->migration_thread != current);
5433 set_current_state(TASK_INTERRUPTIBLE);
5434 while (!kthread_should_stop()) {
5435 struct migration_req *req;
5436 struct list_head *head;
5438 raw_spin_lock_irq(&rq->lock);
5440 if (cpu_is_offline(cpu)) {
5441 raw_spin_unlock_irq(&rq->lock);
5445 if (rq->active_balance) {
5446 active_load_balance(rq, cpu);
5447 rq->active_balance = 0;
5450 head = &rq->migration_queue;
5452 if (list_empty(head)) {
5453 raw_spin_unlock_irq(&rq->lock);
5455 set_current_state(TASK_INTERRUPTIBLE);
5458 req = list_entry(head->next, struct migration_req, list);
5459 list_del_init(head->next);
5461 if (req->task != NULL) {
5462 raw_spin_unlock(&rq->lock);
5463 __migrate_task(req->task, cpu, req->dest_cpu);
5464 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5465 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5466 raw_spin_unlock(&rq->lock);
5468 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5469 raw_spin_unlock(&rq->lock);
5470 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5474 complete(&req->done);
5476 __set_current_state(TASK_RUNNING);
5481 #ifdef CONFIG_HOTPLUG_CPU
5483 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5487 local_irq_disable();
5488 ret = __migrate_task(p, src_cpu, dest_cpu);
5494 * Figure out where task on dead CPU should go, use force if necessary.
5496 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5501 dest_cpu = select_fallback_rq(dead_cpu, p);
5503 /* It can have affinity changed while we were choosing. */
5504 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
5509 * While a dead CPU has no uninterruptible tasks queued at this point,
5510 * it might still have a nonzero ->nr_uninterruptible counter, because
5511 * for performance reasons the counter is not stricly tracking tasks to
5512 * their home CPUs. So we just add the counter to another CPU's counter,
5513 * to keep the global sum constant after CPU-down:
5515 static void migrate_nr_uninterruptible(struct rq *rq_src)
5517 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5518 unsigned long flags;
5520 local_irq_save(flags);
5521 double_rq_lock(rq_src, rq_dest);
5522 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5523 rq_src->nr_uninterruptible = 0;
5524 double_rq_unlock(rq_src, rq_dest);
5525 local_irq_restore(flags);
5528 /* Run through task list and migrate tasks from the dead cpu. */
5529 static void migrate_live_tasks(int src_cpu)
5531 struct task_struct *p, *t;
5533 read_lock(&tasklist_lock);
5535 do_each_thread(t, p) {
5539 if (task_cpu(p) == src_cpu)
5540 move_task_off_dead_cpu(src_cpu, p);
5541 } while_each_thread(t, p);
5543 read_unlock(&tasklist_lock);
5547 * Schedules idle task to be the next runnable task on current CPU.
5548 * It does so by boosting its priority to highest possible.
5549 * Used by CPU offline code.
5551 void sched_idle_next(void)
5553 int this_cpu = smp_processor_id();
5554 struct rq *rq = cpu_rq(this_cpu);
5555 struct task_struct *p = rq->idle;
5556 unsigned long flags;
5558 /* cpu has to be offline */
5559 BUG_ON(cpu_online(this_cpu));
5562 * Strictly not necessary since rest of the CPUs are stopped by now
5563 * and interrupts disabled on the current cpu.
5565 raw_spin_lock_irqsave(&rq->lock, flags);
5567 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5569 update_rq_clock(rq);
5570 activate_task(rq, p, 0);
5572 raw_spin_unlock_irqrestore(&rq->lock, flags);
5576 * Ensures that the idle task is using init_mm right before its cpu goes
5579 void idle_task_exit(void)
5581 struct mm_struct *mm = current->active_mm;
5583 BUG_ON(cpu_online(smp_processor_id()));
5586 switch_mm(mm, &init_mm, current);
5590 /* called under rq->lock with disabled interrupts */
5591 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5593 struct rq *rq = cpu_rq(dead_cpu);
5595 /* Must be exiting, otherwise would be on tasklist. */
5596 BUG_ON(!p->exit_state);
5598 /* Cannot have done final schedule yet: would have vanished. */
5599 BUG_ON(p->state == TASK_DEAD);
5604 * Drop lock around migration; if someone else moves it,
5605 * that's OK. No task can be added to this CPU, so iteration is
5608 raw_spin_unlock_irq(&rq->lock);
5609 move_task_off_dead_cpu(dead_cpu, p);
5610 raw_spin_lock_irq(&rq->lock);
5615 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5616 static void migrate_dead_tasks(unsigned int dead_cpu)
5618 struct rq *rq = cpu_rq(dead_cpu);
5619 struct task_struct *next;
5622 if (!rq->nr_running)
5624 update_rq_clock(rq);
5625 next = pick_next_task(rq);
5628 next->sched_class->put_prev_task(rq, next);
5629 migrate_dead(dead_cpu, next);
5635 * remove the tasks which were accounted by rq from calc_load_tasks.
5637 static void calc_global_load_remove(struct rq *rq)
5639 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5640 rq->calc_load_active = 0;
5642 #endif /* CONFIG_HOTPLUG_CPU */
5644 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5646 static struct ctl_table sd_ctl_dir[] = {
5648 .procname = "sched_domain",
5654 static struct ctl_table sd_ctl_root[] = {
5656 .procname = "kernel",
5658 .child = sd_ctl_dir,
5663 static struct ctl_table *sd_alloc_ctl_entry(int n)
5665 struct ctl_table *entry =
5666 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5671 static void sd_free_ctl_entry(struct ctl_table **tablep)
5673 struct ctl_table *entry;
5676 * In the intermediate directories, both the child directory and
5677 * procname are dynamically allocated and could fail but the mode
5678 * will always be set. In the lowest directory the names are
5679 * static strings and all have proc handlers.
5681 for (entry = *tablep; entry->mode; entry++) {
5683 sd_free_ctl_entry(&entry->child);
5684 if (entry->proc_handler == NULL)
5685 kfree(entry->procname);
5693 set_table_entry(struct ctl_table *entry,
5694 const char *procname, void *data, int maxlen,
5695 mode_t mode, proc_handler *proc_handler)
5697 entry->procname = procname;
5699 entry->maxlen = maxlen;
5701 entry->proc_handler = proc_handler;
5704 static struct ctl_table *
5705 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5707 struct ctl_table *table = sd_alloc_ctl_entry(13);
5712 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5713 sizeof(long), 0644, proc_doulongvec_minmax);
5714 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5715 sizeof(long), 0644, proc_doulongvec_minmax);
5716 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5717 sizeof(int), 0644, proc_dointvec_minmax);
5718 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5719 sizeof(int), 0644, proc_dointvec_minmax);
5720 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5721 sizeof(int), 0644, proc_dointvec_minmax);
5722 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5723 sizeof(int), 0644, proc_dointvec_minmax);
5724 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5725 sizeof(int), 0644, proc_dointvec_minmax);
5726 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5727 sizeof(int), 0644, proc_dointvec_minmax);
5728 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5729 sizeof(int), 0644, proc_dointvec_minmax);
5730 set_table_entry(&table[9], "cache_nice_tries",
5731 &sd->cache_nice_tries,
5732 sizeof(int), 0644, proc_dointvec_minmax);
5733 set_table_entry(&table[10], "flags", &sd->flags,
5734 sizeof(int), 0644, proc_dointvec_minmax);
5735 set_table_entry(&table[11], "name", sd->name,
5736 CORENAME_MAX_SIZE, 0444, proc_dostring);
5737 /* &table[12] is terminator */
5742 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5744 struct ctl_table *entry, *table;
5745 struct sched_domain *sd;
5746 int domain_num = 0, i;
5749 for_each_domain(cpu, sd)
5751 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5756 for_each_domain(cpu, sd) {
5757 snprintf(buf, 32, "domain%d", i);
5758 entry->procname = kstrdup(buf, GFP_KERNEL);
5760 entry->child = sd_alloc_ctl_domain_table(sd);
5767 static struct ctl_table_header *sd_sysctl_header;
5768 static void register_sched_domain_sysctl(void)
5770 int i, cpu_num = num_possible_cpus();
5771 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5774 WARN_ON(sd_ctl_dir[0].child);
5775 sd_ctl_dir[0].child = entry;
5780 for_each_possible_cpu(i) {
5781 snprintf(buf, 32, "cpu%d", i);
5782 entry->procname = kstrdup(buf, GFP_KERNEL);
5784 entry->child = sd_alloc_ctl_cpu_table(i);
5788 WARN_ON(sd_sysctl_header);
5789 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5792 /* may be called multiple times per register */
5793 static void unregister_sched_domain_sysctl(void)
5795 if (sd_sysctl_header)
5796 unregister_sysctl_table(sd_sysctl_header);
5797 sd_sysctl_header = NULL;
5798 if (sd_ctl_dir[0].child)
5799 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5802 static void register_sched_domain_sysctl(void)
5805 static void unregister_sched_domain_sysctl(void)
5810 static void set_rq_online(struct rq *rq)
5813 const struct sched_class *class;
5815 cpumask_set_cpu(rq->cpu, rq->rd->online);
5818 for_each_class(class) {
5819 if (class->rq_online)
5820 class->rq_online(rq);
5825 static void set_rq_offline(struct rq *rq)
5828 const struct sched_class *class;
5830 for_each_class(class) {
5831 if (class->rq_offline)
5832 class->rq_offline(rq);
5835 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5841 * migration_call - callback that gets triggered when a CPU is added.
5842 * Here we can start up the necessary migration thread for the new CPU.
5844 static int __cpuinit
5845 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5847 struct task_struct *p;
5848 int cpu = (long)hcpu;
5849 unsigned long flags;
5854 case CPU_UP_PREPARE:
5855 case CPU_UP_PREPARE_FROZEN:
5856 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5859 kthread_bind(p, cpu);
5860 /* Must be high prio: stop_machine expects to yield to it. */
5861 rq = task_rq_lock(p, &flags);
5862 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5863 task_rq_unlock(rq, &flags);
5865 cpu_rq(cpu)->migration_thread = p;
5866 rq->calc_load_update = calc_load_update;
5870 case CPU_ONLINE_FROZEN:
5871 /* Strictly unnecessary, as first user will wake it. */
5872 wake_up_process(cpu_rq(cpu)->migration_thread);
5874 /* Update our root-domain */
5876 raw_spin_lock_irqsave(&rq->lock, flags);
5878 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5882 raw_spin_unlock_irqrestore(&rq->lock, flags);
5885 #ifdef CONFIG_HOTPLUG_CPU
5886 case CPU_UP_CANCELED:
5887 case CPU_UP_CANCELED_FROZEN:
5888 if (!cpu_rq(cpu)->migration_thread)
5890 /* Unbind it from offline cpu so it can run. Fall thru. */
5891 kthread_bind(cpu_rq(cpu)->migration_thread,
5892 cpumask_any(cpu_online_mask));
5893 kthread_stop(cpu_rq(cpu)->migration_thread);
5894 put_task_struct(cpu_rq(cpu)->migration_thread);
5895 cpu_rq(cpu)->migration_thread = NULL;
5899 case CPU_DEAD_FROZEN:
5900 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5901 migrate_live_tasks(cpu);
5903 kthread_stop(rq->migration_thread);
5904 put_task_struct(rq->migration_thread);
5905 rq->migration_thread = NULL;
5906 /* Idle task back to normal (off runqueue, low prio) */
5907 raw_spin_lock_irq(&rq->lock);
5908 update_rq_clock(rq);
5909 deactivate_task(rq, rq->idle, 0);
5910 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5911 rq->idle->sched_class = &idle_sched_class;
5912 migrate_dead_tasks(cpu);
5913 raw_spin_unlock_irq(&rq->lock);
5915 migrate_nr_uninterruptible(rq);
5916 BUG_ON(rq->nr_running != 0);
5917 calc_global_load_remove(rq);
5919 * No need to migrate the tasks: it was best-effort if
5920 * they didn't take sched_hotcpu_mutex. Just wake up
5923 raw_spin_lock_irq(&rq->lock);
5924 while (!list_empty(&rq->migration_queue)) {
5925 struct migration_req *req;
5927 req = list_entry(rq->migration_queue.next,
5928 struct migration_req, list);
5929 list_del_init(&req->list);
5930 raw_spin_unlock_irq(&rq->lock);
5931 complete(&req->done);
5932 raw_spin_lock_irq(&rq->lock);
5934 raw_spin_unlock_irq(&rq->lock);
5938 case CPU_DYING_FROZEN:
5939 /* Update our root-domain */
5941 raw_spin_lock_irqsave(&rq->lock, flags);
5943 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5946 raw_spin_unlock_irqrestore(&rq->lock, flags);
5954 * Register at high priority so that task migration (migrate_all_tasks)
5955 * happens before everything else. This has to be lower priority than
5956 * the notifier in the perf_event subsystem, though.
5958 static struct notifier_block __cpuinitdata migration_notifier = {
5959 .notifier_call = migration_call,
5963 static int __init migration_init(void)
5965 void *cpu = (void *)(long)smp_processor_id();
5968 /* Start one for the boot CPU: */
5969 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5970 BUG_ON(err == NOTIFY_BAD);
5971 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5972 register_cpu_notifier(&migration_notifier);
5976 early_initcall(migration_init);
5981 #ifdef CONFIG_SCHED_DEBUG
5983 static __read_mostly int sched_domain_debug_enabled;
5985 static int __init sched_domain_debug_setup(char *str)
5987 sched_domain_debug_enabled = 1;
5991 early_param("sched_debug", sched_domain_debug_setup);
5993 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5994 struct cpumask *groupmask)
5996 struct sched_group *group = sd->groups;
5999 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6000 cpumask_clear(groupmask);
6002 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6004 if (!(sd->flags & SD_LOAD_BALANCE)) {
6005 printk("does not load-balance\n");
6007 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6012 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6014 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6015 printk(KERN_ERR "ERROR: domain->span does not contain "
6018 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6019 printk(KERN_ERR "ERROR: domain->groups does not contain"
6023 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6027 printk(KERN_ERR "ERROR: group is NULL\n");
6031 if (!group->cpu_power) {
6032 printk(KERN_CONT "\n");
6033 printk(KERN_ERR "ERROR: domain->cpu_power not "
6038 if (!cpumask_weight(sched_group_cpus(group))) {
6039 printk(KERN_CONT "\n");
6040 printk(KERN_ERR "ERROR: empty group\n");
6044 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6045 printk(KERN_CONT "\n");
6046 printk(KERN_ERR "ERROR: repeated CPUs\n");
6050 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6052 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6054 printk(KERN_CONT " %s", str);
6055 if (group->cpu_power != SCHED_LOAD_SCALE) {
6056 printk(KERN_CONT " (cpu_power = %d)",
6060 group = group->next;
6061 } while (group != sd->groups);
6062 printk(KERN_CONT "\n");
6064 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6065 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6068 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6069 printk(KERN_ERR "ERROR: parent span is not a superset "
6070 "of domain->span\n");
6074 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6076 cpumask_var_t groupmask;
6079 if (!sched_domain_debug_enabled)
6083 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6087 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6089 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6090 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6095 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6102 free_cpumask_var(groupmask);
6104 #else /* !CONFIG_SCHED_DEBUG */
6105 # define sched_domain_debug(sd, cpu) do { } while (0)
6106 #endif /* CONFIG_SCHED_DEBUG */
6108 static int sd_degenerate(struct sched_domain *sd)
6110 if (cpumask_weight(sched_domain_span(sd)) == 1)
6113 /* Following flags need at least 2 groups */
6114 if (sd->flags & (SD_LOAD_BALANCE |
6115 SD_BALANCE_NEWIDLE |
6119 SD_SHARE_PKG_RESOURCES)) {
6120 if (sd->groups != sd->groups->next)
6124 /* Following flags don't use groups */
6125 if (sd->flags & (SD_WAKE_AFFINE))
6132 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6134 unsigned long cflags = sd->flags, pflags = parent->flags;
6136 if (sd_degenerate(parent))
6139 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6142 /* Flags needing groups don't count if only 1 group in parent */
6143 if (parent->groups == parent->groups->next) {
6144 pflags &= ~(SD_LOAD_BALANCE |
6145 SD_BALANCE_NEWIDLE |
6149 SD_SHARE_PKG_RESOURCES);
6150 if (nr_node_ids == 1)
6151 pflags &= ~SD_SERIALIZE;
6153 if (~cflags & pflags)
6159 static void free_rootdomain(struct root_domain *rd)
6161 synchronize_sched();
6163 cpupri_cleanup(&rd->cpupri);
6165 free_cpumask_var(rd->rto_mask);
6166 free_cpumask_var(rd->online);
6167 free_cpumask_var(rd->span);
6171 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6173 struct root_domain *old_rd = NULL;
6174 unsigned long flags;
6176 raw_spin_lock_irqsave(&rq->lock, flags);
6181 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6184 cpumask_clear_cpu(rq->cpu, old_rd->span);
6187 * If we dont want to free the old_rt yet then
6188 * set old_rd to NULL to skip the freeing later
6191 if (!atomic_dec_and_test(&old_rd->refcount))
6195 atomic_inc(&rd->refcount);
6198 cpumask_set_cpu(rq->cpu, rd->span);
6199 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6202 raw_spin_unlock_irqrestore(&rq->lock, flags);
6205 free_rootdomain(old_rd);
6208 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6210 gfp_t gfp = GFP_KERNEL;
6212 memset(rd, 0, sizeof(*rd));
6217 if (!alloc_cpumask_var(&rd->span, gfp))
6219 if (!alloc_cpumask_var(&rd->online, gfp))
6221 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6224 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6229 free_cpumask_var(rd->rto_mask);
6231 free_cpumask_var(rd->online);
6233 free_cpumask_var(rd->span);
6238 static void init_defrootdomain(void)
6240 init_rootdomain(&def_root_domain, true);
6242 atomic_set(&def_root_domain.refcount, 1);
6245 static struct root_domain *alloc_rootdomain(void)
6247 struct root_domain *rd;
6249 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6253 if (init_rootdomain(rd, false) != 0) {
6262 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6263 * hold the hotplug lock.
6266 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6268 struct rq *rq = cpu_rq(cpu);
6269 struct sched_domain *tmp;
6271 /* Remove the sched domains which do not contribute to scheduling. */
6272 for (tmp = sd; tmp; ) {
6273 struct sched_domain *parent = tmp->parent;
6277 if (sd_parent_degenerate(tmp, parent)) {
6278 tmp->parent = parent->parent;
6280 parent->parent->child = tmp;
6285 if (sd && sd_degenerate(sd)) {
6291 sched_domain_debug(sd, cpu);
6293 rq_attach_root(rq, rd);
6294 rcu_assign_pointer(rq->sd, sd);
6297 /* cpus with isolated domains */
6298 static cpumask_var_t cpu_isolated_map;
6300 /* Setup the mask of cpus configured for isolated domains */
6301 static int __init isolated_cpu_setup(char *str)
6303 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6304 cpulist_parse(str, cpu_isolated_map);
6308 __setup("isolcpus=", isolated_cpu_setup);
6311 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6312 * to a function which identifies what group(along with sched group) a CPU
6313 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6314 * (due to the fact that we keep track of groups covered with a struct cpumask).
6316 * init_sched_build_groups will build a circular linked list of the groups
6317 * covered by the given span, and will set each group's ->cpumask correctly,
6318 * and ->cpu_power to 0.
6321 init_sched_build_groups(const struct cpumask *span,
6322 const struct cpumask *cpu_map,
6323 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6324 struct sched_group **sg,
6325 struct cpumask *tmpmask),
6326 struct cpumask *covered, struct cpumask *tmpmask)
6328 struct sched_group *first = NULL, *last = NULL;
6331 cpumask_clear(covered);
6333 for_each_cpu(i, span) {
6334 struct sched_group *sg;
6335 int group = group_fn(i, cpu_map, &sg, tmpmask);
6338 if (cpumask_test_cpu(i, covered))
6341 cpumask_clear(sched_group_cpus(sg));
6344 for_each_cpu(j, span) {
6345 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6348 cpumask_set_cpu(j, covered);
6349 cpumask_set_cpu(j, sched_group_cpus(sg));
6360 #define SD_NODES_PER_DOMAIN 16
6365 * find_next_best_node - find the next node to include in a sched_domain
6366 * @node: node whose sched_domain we're building
6367 * @used_nodes: nodes already in the sched_domain
6369 * Find the next node to include in a given scheduling domain. Simply
6370 * finds the closest node not already in the @used_nodes map.
6372 * Should use nodemask_t.
6374 static int find_next_best_node(int node, nodemask_t *used_nodes)
6376 int i, n, val, min_val, best_node = 0;
6380 for (i = 0; i < nr_node_ids; i++) {
6381 /* Start at @node */
6382 n = (node + i) % nr_node_ids;
6384 if (!nr_cpus_node(n))
6387 /* Skip already used nodes */
6388 if (node_isset(n, *used_nodes))
6391 /* Simple min distance search */
6392 val = node_distance(node, n);
6394 if (val < min_val) {
6400 node_set(best_node, *used_nodes);
6405 * sched_domain_node_span - get a cpumask for a node's sched_domain
6406 * @node: node whose cpumask we're constructing
6407 * @span: resulting cpumask
6409 * Given a node, construct a good cpumask for its sched_domain to span. It
6410 * should be one that prevents unnecessary balancing, but also spreads tasks
6413 static void sched_domain_node_span(int node, struct cpumask *span)
6415 nodemask_t used_nodes;
6418 cpumask_clear(span);
6419 nodes_clear(used_nodes);
6421 cpumask_or(span, span, cpumask_of_node(node));
6422 node_set(node, used_nodes);
6424 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6425 int next_node = find_next_best_node(node, &used_nodes);
6427 cpumask_or(span, span, cpumask_of_node(next_node));
6430 #endif /* CONFIG_NUMA */
6432 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6435 * The cpus mask in sched_group and sched_domain hangs off the end.
6437 * ( See the the comments in include/linux/sched.h:struct sched_group
6438 * and struct sched_domain. )
6440 struct static_sched_group {
6441 struct sched_group sg;
6442 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6445 struct static_sched_domain {
6446 struct sched_domain sd;
6447 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6453 cpumask_var_t domainspan;
6454 cpumask_var_t covered;
6455 cpumask_var_t notcovered;
6457 cpumask_var_t nodemask;
6458 cpumask_var_t this_sibling_map;
6459 cpumask_var_t this_core_map;
6460 cpumask_var_t send_covered;
6461 cpumask_var_t tmpmask;
6462 struct sched_group **sched_group_nodes;
6463 struct root_domain *rd;
6467 sa_sched_groups = 0,
6472 sa_this_sibling_map,
6474 sa_sched_group_nodes,
6484 * SMT sched-domains:
6486 #ifdef CONFIG_SCHED_SMT
6487 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6488 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6491 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6492 struct sched_group **sg, struct cpumask *unused)
6495 *sg = &per_cpu(sched_groups, cpu).sg;
6498 #endif /* CONFIG_SCHED_SMT */
6501 * multi-core sched-domains:
6503 #ifdef CONFIG_SCHED_MC
6504 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6505 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6506 #endif /* CONFIG_SCHED_MC */
6508 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6510 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6511 struct sched_group **sg, struct cpumask *mask)
6515 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6516 group = cpumask_first(mask);
6518 *sg = &per_cpu(sched_group_core, group).sg;
6521 #elif defined(CONFIG_SCHED_MC)
6523 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6524 struct sched_group **sg, struct cpumask *unused)
6527 *sg = &per_cpu(sched_group_core, cpu).sg;
6532 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6533 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6536 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6537 struct sched_group **sg, struct cpumask *mask)
6540 #ifdef CONFIG_SCHED_MC
6541 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6542 group = cpumask_first(mask);
6543 #elif defined(CONFIG_SCHED_SMT)
6544 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6545 group = cpumask_first(mask);
6550 *sg = &per_cpu(sched_group_phys, group).sg;
6556 * The init_sched_build_groups can't handle what we want to do with node
6557 * groups, so roll our own. Now each node has its own list of groups which
6558 * gets dynamically allocated.
6560 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6561 static struct sched_group ***sched_group_nodes_bycpu;
6563 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6564 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6566 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6567 struct sched_group **sg,
6568 struct cpumask *nodemask)
6572 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6573 group = cpumask_first(nodemask);
6576 *sg = &per_cpu(sched_group_allnodes, group).sg;
6580 static void init_numa_sched_groups_power(struct sched_group *group_head)
6582 struct sched_group *sg = group_head;
6588 for_each_cpu(j, sched_group_cpus(sg)) {
6589 struct sched_domain *sd;
6591 sd = &per_cpu(phys_domains, j).sd;
6592 if (j != group_first_cpu(sd->groups)) {
6594 * Only add "power" once for each
6600 sg->cpu_power += sd->groups->cpu_power;
6603 } while (sg != group_head);
6606 static int build_numa_sched_groups(struct s_data *d,
6607 const struct cpumask *cpu_map, int num)
6609 struct sched_domain *sd;
6610 struct sched_group *sg, *prev;
6613 cpumask_clear(d->covered);
6614 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6615 if (cpumask_empty(d->nodemask)) {
6616 d->sched_group_nodes[num] = NULL;
6620 sched_domain_node_span(num, d->domainspan);
6621 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6623 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6626 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6630 d->sched_group_nodes[num] = sg;
6632 for_each_cpu(j, d->nodemask) {
6633 sd = &per_cpu(node_domains, j).sd;
6638 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6640 cpumask_or(d->covered, d->covered, d->nodemask);
6643 for (j = 0; j < nr_node_ids; j++) {
6644 n = (num + j) % nr_node_ids;
6645 cpumask_complement(d->notcovered, d->covered);
6646 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6647 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6648 if (cpumask_empty(d->tmpmask))
6650 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6651 if (cpumask_empty(d->tmpmask))
6653 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6657 "Can not alloc domain group for node %d\n", j);
6661 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6662 sg->next = prev->next;
6663 cpumask_or(d->covered, d->covered, d->tmpmask);
6670 #endif /* CONFIG_NUMA */
6673 /* Free memory allocated for various sched_group structures */
6674 static void free_sched_groups(const struct cpumask *cpu_map,
6675 struct cpumask *nodemask)
6679 for_each_cpu(cpu, cpu_map) {
6680 struct sched_group **sched_group_nodes
6681 = sched_group_nodes_bycpu[cpu];
6683 if (!sched_group_nodes)
6686 for (i = 0; i < nr_node_ids; i++) {
6687 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6689 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6690 if (cpumask_empty(nodemask))
6700 if (oldsg != sched_group_nodes[i])
6703 kfree(sched_group_nodes);
6704 sched_group_nodes_bycpu[cpu] = NULL;
6707 #else /* !CONFIG_NUMA */
6708 static void free_sched_groups(const struct cpumask *cpu_map,
6709 struct cpumask *nodemask)
6712 #endif /* CONFIG_NUMA */
6715 * Initialize sched groups cpu_power.
6717 * cpu_power indicates the capacity of sched group, which is used while
6718 * distributing the load between different sched groups in a sched domain.
6719 * Typically cpu_power for all the groups in a sched domain will be same unless
6720 * there are asymmetries in the topology. If there are asymmetries, group
6721 * having more cpu_power will pickup more load compared to the group having
6724 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6726 struct sched_domain *child;
6727 struct sched_group *group;
6731 WARN_ON(!sd || !sd->groups);
6733 if (cpu != group_first_cpu(sd->groups))
6738 sd->groups->cpu_power = 0;
6741 power = SCHED_LOAD_SCALE;
6742 weight = cpumask_weight(sched_domain_span(sd));
6744 * SMT siblings share the power of a single core.
6745 * Usually multiple threads get a better yield out of
6746 * that one core than a single thread would have,
6747 * reflect that in sd->smt_gain.
6749 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6750 power *= sd->smt_gain;
6752 power >>= SCHED_LOAD_SHIFT;
6754 sd->groups->cpu_power += power;
6759 * Add cpu_power of each child group to this groups cpu_power.
6761 group = child->groups;
6763 sd->groups->cpu_power += group->cpu_power;
6764 group = group->next;
6765 } while (group != child->groups);
6769 * Initializers for schedule domains
6770 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6773 #ifdef CONFIG_SCHED_DEBUG
6774 # define SD_INIT_NAME(sd, type) sd->name = #type
6776 # define SD_INIT_NAME(sd, type) do { } while (0)
6779 #define SD_INIT(sd, type) sd_init_##type(sd)
6781 #define SD_INIT_FUNC(type) \
6782 static noinline void sd_init_##type(struct sched_domain *sd) \
6784 memset(sd, 0, sizeof(*sd)); \
6785 *sd = SD_##type##_INIT; \
6786 sd->level = SD_LV_##type; \
6787 SD_INIT_NAME(sd, type); \
6792 SD_INIT_FUNC(ALLNODES)
6795 #ifdef CONFIG_SCHED_SMT
6796 SD_INIT_FUNC(SIBLING)
6798 #ifdef CONFIG_SCHED_MC
6802 static int default_relax_domain_level = -1;
6804 static int __init setup_relax_domain_level(char *str)
6808 val = simple_strtoul(str, NULL, 0);
6809 if (val < SD_LV_MAX)
6810 default_relax_domain_level = val;
6814 __setup("relax_domain_level=", setup_relax_domain_level);
6816 static void set_domain_attribute(struct sched_domain *sd,
6817 struct sched_domain_attr *attr)
6821 if (!attr || attr->relax_domain_level < 0) {
6822 if (default_relax_domain_level < 0)
6825 request = default_relax_domain_level;
6827 request = attr->relax_domain_level;
6828 if (request < sd->level) {
6829 /* turn off idle balance on this domain */
6830 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6832 /* turn on idle balance on this domain */
6833 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6837 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6838 const struct cpumask *cpu_map)
6841 case sa_sched_groups:
6842 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6843 d->sched_group_nodes = NULL;
6845 free_rootdomain(d->rd); /* fall through */
6847 free_cpumask_var(d->tmpmask); /* fall through */
6848 case sa_send_covered:
6849 free_cpumask_var(d->send_covered); /* fall through */
6850 case sa_this_core_map:
6851 free_cpumask_var(d->this_core_map); /* fall through */
6852 case sa_this_sibling_map:
6853 free_cpumask_var(d->this_sibling_map); /* fall through */
6855 free_cpumask_var(d->nodemask); /* fall through */
6856 case sa_sched_group_nodes:
6858 kfree(d->sched_group_nodes); /* fall through */
6860 free_cpumask_var(d->notcovered); /* fall through */
6862 free_cpumask_var(d->covered); /* fall through */
6864 free_cpumask_var(d->domainspan); /* fall through */
6871 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6872 const struct cpumask *cpu_map)
6875 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6877 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6878 return sa_domainspan;
6879 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6881 /* Allocate the per-node list of sched groups */
6882 d->sched_group_nodes = kcalloc(nr_node_ids,
6883 sizeof(struct sched_group *), GFP_KERNEL);
6884 if (!d->sched_group_nodes) {
6885 printk(KERN_WARNING "Can not alloc sched group node list\n");
6886 return sa_notcovered;
6888 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6890 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6891 return sa_sched_group_nodes;
6892 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6894 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6895 return sa_this_sibling_map;
6896 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6897 return sa_this_core_map;
6898 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6899 return sa_send_covered;
6900 d->rd = alloc_rootdomain();
6902 printk(KERN_WARNING "Cannot alloc root domain\n");
6905 return sa_rootdomain;
6908 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6909 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6911 struct sched_domain *sd = NULL;
6913 struct sched_domain *parent;
6916 if (cpumask_weight(cpu_map) >
6917 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6918 sd = &per_cpu(allnodes_domains, i).sd;
6919 SD_INIT(sd, ALLNODES);
6920 set_domain_attribute(sd, attr);
6921 cpumask_copy(sched_domain_span(sd), cpu_map);
6922 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6927 sd = &per_cpu(node_domains, i).sd;
6929 set_domain_attribute(sd, attr);
6930 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6931 sd->parent = parent;
6934 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6939 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6940 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6941 struct sched_domain *parent, int i)
6943 struct sched_domain *sd;
6944 sd = &per_cpu(phys_domains, i).sd;
6946 set_domain_attribute(sd, attr);
6947 cpumask_copy(sched_domain_span(sd), d->nodemask);
6948 sd->parent = parent;
6951 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6955 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6956 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6957 struct sched_domain *parent, int i)
6959 struct sched_domain *sd = parent;
6960 #ifdef CONFIG_SCHED_MC
6961 sd = &per_cpu(core_domains, i).sd;
6963 set_domain_attribute(sd, attr);
6964 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6965 sd->parent = parent;
6967 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6972 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6973 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6974 struct sched_domain *parent, int i)
6976 struct sched_domain *sd = parent;
6977 #ifdef CONFIG_SCHED_SMT
6978 sd = &per_cpu(cpu_domains, i).sd;
6979 SD_INIT(sd, SIBLING);
6980 set_domain_attribute(sd, attr);
6981 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6982 sd->parent = parent;
6984 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6989 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6990 const struct cpumask *cpu_map, int cpu)
6993 #ifdef CONFIG_SCHED_SMT
6994 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6995 cpumask_and(d->this_sibling_map, cpu_map,
6996 topology_thread_cpumask(cpu));
6997 if (cpu == cpumask_first(d->this_sibling_map))
6998 init_sched_build_groups(d->this_sibling_map, cpu_map,
7000 d->send_covered, d->tmpmask);
7003 #ifdef CONFIG_SCHED_MC
7004 case SD_LV_MC: /* set up multi-core groups */
7005 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7006 if (cpu == cpumask_first(d->this_core_map))
7007 init_sched_build_groups(d->this_core_map, cpu_map,
7009 d->send_covered, d->tmpmask);
7012 case SD_LV_CPU: /* set up physical groups */
7013 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7014 if (!cpumask_empty(d->nodemask))
7015 init_sched_build_groups(d->nodemask, cpu_map,
7017 d->send_covered, d->tmpmask);
7020 case SD_LV_ALLNODES:
7021 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7022 d->send_covered, d->tmpmask);
7031 * Build sched domains for a given set of cpus and attach the sched domains
7032 * to the individual cpus
7034 static int __build_sched_domains(const struct cpumask *cpu_map,
7035 struct sched_domain_attr *attr)
7037 enum s_alloc alloc_state = sa_none;
7039 struct sched_domain *sd;
7045 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7046 if (alloc_state != sa_rootdomain)
7048 alloc_state = sa_sched_groups;
7051 * Set up domains for cpus specified by the cpu_map.
7053 for_each_cpu(i, cpu_map) {
7054 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7057 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7058 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7059 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7060 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7063 for_each_cpu(i, cpu_map) {
7064 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7065 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7068 /* Set up physical groups */
7069 for (i = 0; i < nr_node_ids; i++)
7070 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7073 /* Set up node groups */
7075 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7077 for (i = 0; i < nr_node_ids; i++)
7078 if (build_numa_sched_groups(&d, cpu_map, i))
7082 /* Calculate CPU power for physical packages and nodes */
7083 #ifdef CONFIG_SCHED_SMT
7084 for_each_cpu(i, cpu_map) {
7085 sd = &per_cpu(cpu_domains, i).sd;
7086 init_sched_groups_power(i, sd);
7089 #ifdef CONFIG_SCHED_MC
7090 for_each_cpu(i, cpu_map) {
7091 sd = &per_cpu(core_domains, i).sd;
7092 init_sched_groups_power(i, sd);
7096 for_each_cpu(i, cpu_map) {
7097 sd = &per_cpu(phys_domains, i).sd;
7098 init_sched_groups_power(i, sd);
7102 for (i = 0; i < nr_node_ids; i++)
7103 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7105 if (d.sd_allnodes) {
7106 struct sched_group *sg;
7108 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7110 init_numa_sched_groups_power(sg);
7114 /* Attach the domains */
7115 for_each_cpu(i, cpu_map) {
7116 #ifdef CONFIG_SCHED_SMT
7117 sd = &per_cpu(cpu_domains, i).sd;
7118 #elif defined(CONFIG_SCHED_MC)
7119 sd = &per_cpu(core_domains, i).sd;
7121 sd = &per_cpu(phys_domains, i).sd;
7123 cpu_attach_domain(sd, d.rd, i);
7126 d.sched_group_nodes = NULL; /* don't free this we still need it */
7127 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7131 __free_domain_allocs(&d, alloc_state, cpu_map);
7135 static int build_sched_domains(const struct cpumask *cpu_map)
7137 return __build_sched_domains(cpu_map, NULL);
7140 static cpumask_var_t *doms_cur; /* current sched domains */
7141 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7142 static struct sched_domain_attr *dattr_cur;
7143 /* attribues of custom domains in 'doms_cur' */
7146 * Special case: If a kmalloc of a doms_cur partition (array of
7147 * cpumask) fails, then fallback to a single sched domain,
7148 * as determined by the single cpumask fallback_doms.
7150 static cpumask_var_t fallback_doms;
7153 * arch_update_cpu_topology lets virtualized architectures update the
7154 * cpu core maps. It is supposed to return 1 if the topology changed
7155 * or 0 if it stayed the same.
7157 int __attribute__((weak)) arch_update_cpu_topology(void)
7162 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7165 cpumask_var_t *doms;
7167 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7170 for (i = 0; i < ndoms; i++) {
7171 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7172 free_sched_domains(doms, i);
7179 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7182 for (i = 0; i < ndoms; i++)
7183 free_cpumask_var(doms[i]);
7188 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7189 * For now this just excludes isolated cpus, but could be used to
7190 * exclude other special cases in the future.
7192 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7196 arch_update_cpu_topology();
7198 doms_cur = alloc_sched_domains(ndoms_cur);
7200 doms_cur = &fallback_doms;
7201 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7203 err = build_sched_domains(doms_cur[0]);
7204 register_sched_domain_sysctl();
7209 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7210 struct cpumask *tmpmask)
7212 free_sched_groups(cpu_map, tmpmask);
7216 * Detach sched domains from a group of cpus specified in cpu_map
7217 * These cpus will now be attached to the NULL domain
7219 static void detach_destroy_domains(const struct cpumask *cpu_map)
7221 /* Save because hotplug lock held. */
7222 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7225 for_each_cpu(i, cpu_map)
7226 cpu_attach_domain(NULL, &def_root_domain, i);
7227 synchronize_sched();
7228 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7231 /* handle null as "default" */
7232 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7233 struct sched_domain_attr *new, int idx_new)
7235 struct sched_domain_attr tmp;
7242 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7243 new ? (new + idx_new) : &tmp,
7244 sizeof(struct sched_domain_attr));
7248 * Partition sched domains as specified by the 'ndoms_new'
7249 * cpumasks in the array doms_new[] of cpumasks. This compares
7250 * doms_new[] to the current sched domain partitioning, doms_cur[].
7251 * It destroys each deleted domain and builds each new domain.
7253 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7254 * The masks don't intersect (don't overlap.) We should setup one
7255 * sched domain for each mask. CPUs not in any of the cpumasks will
7256 * not be load balanced. If the same cpumask appears both in the
7257 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7260 * The passed in 'doms_new' should be allocated using
7261 * alloc_sched_domains. This routine takes ownership of it and will
7262 * free_sched_domains it when done with it. If the caller failed the
7263 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7264 * and partition_sched_domains() will fallback to the single partition
7265 * 'fallback_doms', it also forces the domains to be rebuilt.
7267 * If doms_new == NULL it will be replaced with cpu_online_mask.
7268 * ndoms_new == 0 is a special case for destroying existing domains,
7269 * and it will not create the default domain.
7271 * Call with hotplug lock held
7273 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7274 struct sched_domain_attr *dattr_new)
7279 mutex_lock(&sched_domains_mutex);
7281 /* always unregister in case we don't destroy any domains */
7282 unregister_sched_domain_sysctl();
7284 /* Let architecture update cpu core mappings. */
7285 new_topology = arch_update_cpu_topology();
7287 n = doms_new ? ndoms_new : 0;
7289 /* Destroy deleted domains */
7290 for (i = 0; i < ndoms_cur; i++) {
7291 for (j = 0; j < n && !new_topology; j++) {
7292 if (cpumask_equal(doms_cur[i], doms_new[j])
7293 && dattrs_equal(dattr_cur, i, dattr_new, j))
7296 /* no match - a current sched domain not in new doms_new[] */
7297 detach_destroy_domains(doms_cur[i]);
7302 if (doms_new == NULL) {
7304 doms_new = &fallback_doms;
7305 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7306 WARN_ON_ONCE(dattr_new);
7309 /* Build new domains */
7310 for (i = 0; i < ndoms_new; i++) {
7311 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7312 if (cpumask_equal(doms_new[i], doms_cur[j])
7313 && dattrs_equal(dattr_new, i, dattr_cur, j))
7316 /* no match - add a new doms_new */
7317 __build_sched_domains(doms_new[i],
7318 dattr_new ? dattr_new + i : NULL);
7323 /* Remember the new sched domains */
7324 if (doms_cur != &fallback_doms)
7325 free_sched_domains(doms_cur, ndoms_cur);
7326 kfree(dattr_cur); /* kfree(NULL) is safe */
7327 doms_cur = doms_new;
7328 dattr_cur = dattr_new;
7329 ndoms_cur = ndoms_new;
7331 register_sched_domain_sysctl();
7333 mutex_unlock(&sched_domains_mutex);
7336 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7337 static void arch_reinit_sched_domains(void)
7341 /* Destroy domains first to force the rebuild */
7342 partition_sched_domains(0, NULL, NULL);
7344 rebuild_sched_domains();
7348 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7350 unsigned int level = 0;
7352 if (sscanf(buf, "%u", &level) != 1)
7356 * level is always be positive so don't check for
7357 * level < POWERSAVINGS_BALANCE_NONE which is 0
7358 * What happens on 0 or 1 byte write,
7359 * need to check for count as well?
7362 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7366 sched_smt_power_savings = level;
7368 sched_mc_power_savings = level;
7370 arch_reinit_sched_domains();
7375 #ifdef CONFIG_SCHED_MC
7376 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7379 return sprintf(page, "%u\n", sched_mc_power_savings);
7381 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7382 const char *buf, size_t count)
7384 return sched_power_savings_store(buf, count, 0);
7386 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7387 sched_mc_power_savings_show,
7388 sched_mc_power_savings_store);
7391 #ifdef CONFIG_SCHED_SMT
7392 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7395 return sprintf(page, "%u\n", sched_smt_power_savings);
7397 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7398 const char *buf, size_t count)
7400 return sched_power_savings_store(buf, count, 1);
7402 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7403 sched_smt_power_savings_show,
7404 sched_smt_power_savings_store);
7407 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7411 #ifdef CONFIG_SCHED_SMT
7413 err = sysfs_create_file(&cls->kset.kobj,
7414 &attr_sched_smt_power_savings.attr);
7416 #ifdef CONFIG_SCHED_MC
7417 if (!err && mc_capable())
7418 err = sysfs_create_file(&cls->kset.kobj,
7419 &attr_sched_mc_power_savings.attr);
7423 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7425 #ifndef CONFIG_CPUSETS
7427 * Add online and remove offline CPUs from the scheduler domains.
7428 * When cpusets are enabled they take over this function.
7430 static int update_sched_domains(struct notifier_block *nfb,
7431 unsigned long action, void *hcpu)
7435 case CPU_ONLINE_FROZEN:
7436 case CPU_DOWN_PREPARE:
7437 case CPU_DOWN_PREPARE_FROZEN:
7438 case CPU_DOWN_FAILED:
7439 case CPU_DOWN_FAILED_FROZEN:
7440 partition_sched_domains(1, NULL, NULL);
7449 static int update_runtime(struct notifier_block *nfb,
7450 unsigned long action, void *hcpu)
7452 int cpu = (int)(long)hcpu;
7455 case CPU_DOWN_PREPARE:
7456 case CPU_DOWN_PREPARE_FROZEN:
7457 disable_runtime(cpu_rq(cpu));
7460 case CPU_DOWN_FAILED:
7461 case CPU_DOWN_FAILED_FROZEN:
7463 case CPU_ONLINE_FROZEN:
7464 enable_runtime(cpu_rq(cpu));
7472 void __init sched_init_smp(void)
7474 cpumask_var_t non_isolated_cpus;
7476 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7477 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7479 #if defined(CONFIG_NUMA)
7480 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7482 BUG_ON(sched_group_nodes_bycpu == NULL);
7485 mutex_lock(&sched_domains_mutex);
7486 arch_init_sched_domains(cpu_active_mask);
7487 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7488 if (cpumask_empty(non_isolated_cpus))
7489 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7490 mutex_unlock(&sched_domains_mutex);
7493 #ifndef CONFIG_CPUSETS
7494 /* XXX: Theoretical race here - CPU may be hotplugged now */
7495 hotcpu_notifier(update_sched_domains, 0);
7498 /* RT runtime code needs to handle some hotplug events */
7499 hotcpu_notifier(update_runtime, 0);
7503 /* Move init over to a non-isolated CPU */
7504 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7506 sched_init_granularity();
7507 free_cpumask_var(non_isolated_cpus);
7509 init_sched_rt_class();
7512 void __init sched_init_smp(void)
7514 sched_init_granularity();
7516 #endif /* CONFIG_SMP */
7518 const_debug unsigned int sysctl_timer_migration = 1;
7520 int in_sched_functions(unsigned long addr)
7522 return in_lock_functions(addr) ||
7523 (addr >= (unsigned long)__sched_text_start
7524 && addr < (unsigned long)__sched_text_end);
7527 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7529 cfs_rq->tasks_timeline = RB_ROOT;
7530 INIT_LIST_HEAD(&cfs_rq->tasks);
7531 #ifdef CONFIG_FAIR_GROUP_SCHED
7534 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7537 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7539 struct rt_prio_array *array;
7542 array = &rt_rq->active;
7543 for (i = 0; i < MAX_RT_PRIO; i++) {
7544 INIT_LIST_HEAD(array->queue + i);
7545 __clear_bit(i, array->bitmap);
7547 /* delimiter for bitsearch: */
7548 __set_bit(MAX_RT_PRIO, array->bitmap);
7550 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7551 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7553 rt_rq->highest_prio.next = MAX_RT_PRIO;
7557 rt_rq->rt_nr_migratory = 0;
7558 rt_rq->overloaded = 0;
7559 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7563 rt_rq->rt_throttled = 0;
7564 rt_rq->rt_runtime = 0;
7565 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7567 #ifdef CONFIG_RT_GROUP_SCHED
7568 rt_rq->rt_nr_boosted = 0;
7573 #ifdef CONFIG_FAIR_GROUP_SCHED
7574 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7575 struct sched_entity *se, int cpu, int add,
7576 struct sched_entity *parent)
7578 struct rq *rq = cpu_rq(cpu);
7579 tg->cfs_rq[cpu] = cfs_rq;
7580 init_cfs_rq(cfs_rq, rq);
7583 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7586 /* se could be NULL for init_task_group */
7591 se->cfs_rq = &rq->cfs;
7593 se->cfs_rq = parent->my_q;
7596 se->load.weight = tg->shares;
7597 se->load.inv_weight = 0;
7598 se->parent = parent;
7602 #ifdef CONFIG_RT_GROUP_SCHED
7603 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7604 struct sched_rt_entity *rt_se, int cpu, int add,
7605 struct sched_rt_entity *parent)
7607 struct rq *rq = cpu_rq(cpu);
7609 tg->rt_rq[cpu] = rt_rq;
7610 init_rt_rq(rt_rq, rq);
7612 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7614 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7616 tg->rt_se[cpu] = rt_se;
7621 rt_se->rt_rq = &rq->rt;
7623 rt_se->rt_rq = parent->my_q;
7625 rt_se->my_q = rt_rq;
7626 rt_se->parent = parent;
7627 INIT_LIST_HEAD(&rt_se->run_list);
7631 void __init sched_init(void)
7634 unsigned long alloc_size = 0, ptr;
7636 #ifdef CONFIG_FAIR_GROUP_SCHED
7637 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7639 #ifdef CONFIG_RT_GROUP_SCHED
7640 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7642 #ifdef CONFIG_CPUMASK_OFFSTACK
7643 alloc_size += num_possible_cpus() * cpumask_size();
7646 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7648 #ifdef CONFIG_FAIR_GROUP_SCHED
7649 init_task_group.se = (struct sched_entity **)ptr;
7650 ptr += nr_cpu_ids * sizeof(void **);
7652 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7653 ptr += nr_cpu_ids * sizeof(void **);
7655 #endif /* CONFIG_FAIR_GROUP_SCHED */
7656 #ifdef CONFIG_RT_GROUP_SCHED
7657 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7658 ptr += nr_cpu_ids * sizeof(void **);
7660 init_task_group.rt_rq = (struct rt_rq **)ptr;
7661 ptr += nr_cpu_ids * sizeof(void **);
7663 #endif /* CONFIG_RT_GROUP_SCHED */
7664 #ifdef CONFIG_CPUMASK_OFFSTACK
7665 for_each_possible_cpu(i) {
7666 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7667 ptr += cpumask_size();
7669 #endif /* CONFIG_CPUMASK_OFFSTACK */
7673 init_defrootdomain();
7676 init_rt_bandwidth(&def_rt_bandwidth,
7677 global_rt_period(), global_rt_runtime());
7679 #ifdef CONFIG_RT_GROUP_SCHED
7680 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7681 global_rt_period(), global_rt_runtime());
7682 #endif /* CONFIG_RT_GROUP_SCHED */
7684 #ifdef CONFIG_CGROUP_SCHED
7685 list_add(&init_task_group.list, &task_groups);
7686 INIT_LIST_HEAD(&init_task_group.children);
7688 #endif /* CONFIG_CGROUP_SCHED */
7690 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7691 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7692 __alignof__(unsigned long));
7694 for_each_possible_cpu(i) {
7698 raw_spin_lock_init(&rq->lock);
7700 rq->calc_load_active = 0;
7701 rq->calc_load_update = jiffies + LOAD_FREQ;
7702 init_cfs_rq(&rq->cfs, rq);
7703 init_rt_rq(&rq->rt, rq);
7704 #ifdef CONFIG_FAIR_GROUP_SCHED
7705 init_task_group.shares = init_task_group_load;
7706 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7707 #ifdef CONFIG_CGROUP_SCHED
7709 * How much cpu bandwidth does init_task_group get?
7711 * In case of task-groups formed thr' the cgroup filesystem, it
7712 * gets 100% of the cpu resources in the system. This overall
7713 * system cpu resource is divided among the tasks of
7714 * init_task_group and its child task-groups in a fair manner,
7715 * based on each entity's (task or task-group's) weight
7716 * (se->load.weight).
7718 * In other words, if init_task_group has 10 tasks of weight
7719 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7720 * then A0's share of the cpu resource is:
7722 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7724 * We achieve this by letting init_task_group's tasks sit
7725 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7727 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7729 #endif /* CONFIG_FAIR_GROUP_SCHED */
7731 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7732 #ifdef CONFIG_RT_GROUP_SCHED
7733 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7734 #ifdef CONFIG_CGROUP_SCHED
7735 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7739 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7740 rq->cpu_load[j] = 0;
7744 rq->post_schedule = 0;
7745 rq->active_balance = 0;
7746 rq->next_balance = jiffies;
7750 rq->migration_thread = NULL;
7752 rq->avg_idle = 2*sysctl_sched_migration_cost;
7753 INIT_LIST_HEAD(&rq->migration_queue);
7754 rq_attach_root(rq, &def_root_domain);
7757 atomic_set(&rq->nr_iowait, 0);
7760 set_load_weight(&init_task);
7762 #ifdef CONFIG_PREEMPT_NOTIFIERS
7763 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7767 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7770 #ifdef CONFIG_RT_MUTEXES
7771 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7775 * The boot idle thread does lazy MMU switching as well:
7777 atomic_inc(&init_mm.mm_count);
7778 enter_lazy_tlb(&init_mm, current);
7781 * Make us the idle thread. Technically, schedule() should not be
7782 * called from this thread, however somewhere below it might be,
7783 * but because we are the idle thread, we just pick up running again
7784 * when this runqueue becomes "idle".
7786 init_idle(current, smp_processor_id());
7788 calc_load_update = jiffies + LOAD_FREQ;
7791 * During early bootup we pretend to be a normal task:
7793 current->sched_class = &fair_sched_class;
7795 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7796 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7799 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7800 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7802 /* May be allocated at isolcpus cmdline parse time */
7803 if (cpu_isolated_map == NULL)
7804 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7809 scheduler_running = 1;
7812 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7813 static inline int preempt_count_equals(int preempt_offset)
7815 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7817 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7820 void __might_sleep(const char *file, int line, int preempt_offset)
7823 static unsigned long prev_jiffy; /* ratelimiting */
7825 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7826 system_state != SYSTEM_RUNNING || oops_in_progress)
7828 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7830 prev_jiffy = jiffies;
7833 "BUG: sleeping function called from invalid context at %s:%d\n",
7836 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7837 in_atomic(), irqs_disabled(),
7838 current->pid, current->comm);
7840 debug_show_held_locks(current);
7841 if (irqs_disabled())
7842 print_irqtrace_events(current);
7846 EXPORT_SYMBOL(__might_sleep);
7849 #ifdef CONFIG_MAGIC_SYSRQ
7850 static void normalize_task(struct rq *rq, struct task_struct *p)
7854 update_rq_clock(rq);
7855 on_rq = p->se.on_rq;
7857 deactivate_task(rq, p, 0);
7858 __setscheduler(rq, p, SCHED_NORMAL, 0);
7860 activate_task(rq, p, 0);
7861 resched_task(rq->curr);
7865 void normalize_rt_tasks(void)
7867 struct task_struct *g, *p;
7868 unsigned long flags;
7871 read_lock_irqsave(&tasklist_lock, flags);
7872 do_each_thread(g, p) {
7874 * Only normalize user tasks:
7879 p->se.exec_start = 0;
7880 #ifdef CONFIG_SCHEDSTATS
7881 p->se.wait_start = 0;
7882 p->se.sleep_start = 0;
7883 p->se.block_start = 0;
7888 * Renice negative nice level userspace
7891 if (TASK_NICE(p) < 0 && p->mm)
7892 set_user_nice(p, 0);
7896 raw_spin_lock(&p->pi_lock);
7897 rq = __task_rq_lock(p);
7899 normalize_task(rq, p);
7901 __task_rq_unlock(rq);
7902 raw_spin_unlock(&p->pi_lock);
7903 } while_each_thread(g, p);
7905 read_unlock_irqrestore(&tasklist_lock, flags);
7908 #endif /* CONFIG_MAGIC_SYSRQ */
7912 * These functions are only useful for the IA64 MCA handling.
7914 * They can only be called when the whole system has been
7915 * stopped - every CPU needs to be quiescent, and no scheduling
7916 * activity can take place. Using them for anything else would
7917 * be a serious bug, and as a result, they aren't even visible
7918 * under any other configuration.
7922 * curr_task - return the current task for a given cpu.
7923 * @cpu: the processor in question.
7925 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7927 struct task_struct *curr_task(int cpu)
7929 return cpu_curr(cpu);
7933 * set_curr_task - set the current task for a given cpu.
7934 * @cpu: the processor in question.
7935 * @p: the task pointer to set.
7937 * Description: This function must only be used when non-maskable interrupts
7938 * are serviced on a separate stack. It allows the architecture to switch the
7939 * notion of the current task on a cpu in a non-blocking manner. This function
7940 * must be called with all CPU's synchronized, and interrupts disabled, the
7941 * and caller must save the original value of the current task (see
7942 * curr_task() above) and restore that value before reenabling interrupts and
7943 * re-starting the system.
7945 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7947 void set_curr_task(int cpu, struct task_struct *p)
7954 #ifdef CONFIG_FAIR_GROUP_SCHED
7955 static void free_fair_sched_group(struct task_group *tg)
7959 for_each_possible_cpu(i) {
7961 kfree(tg->cfs_rq[i]);
7971 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7973 struct cfs_rq *cfs_rq;
7974 struct sched_entity *se;
7978 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7981 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7985 tg->shares = NICE_0_LOAD;
7987 for_each_possible_cpu(i) {
7990 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7991 GFP_KERNEL, cpu_to_node(i));
7995 se = kzalloc_node(sizeof(struct sched_entity),
7996 GFP_KERNEL, cpu_to_node(i));
8000 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8011 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8013 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8014 &cpu_rq(cpu)->leaf_cfs_rq_list);
8017 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8019 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8021 #else /* !CONFG_FAIR_GROUP_SCHED */
8022 static inline void free_fair_sched_group(struct task_group *tg)
8027 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8032 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8036 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8039 #endif /* CONFIG_FAIR_GROUP_SCHED */
8041 #ifdef CONFIG_RT_GROUP_SCHED
8042 static void free_rt_sched_group(struct task_group *tg)
8046 destroy_rt_bandwidth(&tg->rt_bandwidth);
8048 for_each_possible_cpu(i) {
8050 kfree(tg->rt_rq[i]);
8052 kfree(tg->rt_se[i]);
8060 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8062 struct rt_rq *rt_rq;
8063 struct sched_rt_entity *rt_se;
8067 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8070 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8074 init_rt_bandwidth(&tg->rt_bandwidth,
8075 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8077 for_each_possible_cpu(i) {
8080 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8081 GFP_KERNEL, cpu_to_node(i));
8085 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8086 GFP_KERNEL, cpu_to_node(i));
8090 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8101 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8103 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8104 &cpu_rq(cpu)->leaf_rt_rq_list);
8107 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8109 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8111 #else /* !CONFIG_RT_GROUP_SCHED */
8112 static inline void free_rt_sched_group(struct task_group *tg)
8117 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8122 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8126 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8129 #endif /* CONFIG_RT_GROUP_SCHED */
8131 #ifdef CONFIG_CGROUP_SCHED
8132 static void free_sched_group(struct task_group *tg)
8134 free_fair_sched_group(tg);
8135 free_rt_sched_group(tg);
8139 /* allocate runqueue etc for a new task group */
8140 struct task_group *sched_create_group(struct task_group *parent)
8142 struct task_group *tg;
8143 unsigned long flags;
8146 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8148 return ERR_PTR(-ENOMEM);
8150 if (!alloc_fair_sched_group(tg, parent))
8153 if (!alloc_rt_sched_group(tg, parent))
8156 spin_lock_irqsave(&task_group_lock, flags);
8157 for_each_possible_cpu(i) {
8158 register_fair_sched_group(tg, i);
8159 register_rt_sched_group(tg, i);
8161 list_add_rcu(&tg->list, &task_groups);
8163 WARN_ON(!parent); /* root should already exist */
8165 tg->parent = parent;
8166 INIT_LIST_HEAD(&tg->children);
8167 list_add_rcu(&tg->siblings, &parent->children);
8168 spin_unlock_irqrestore(&task_group_lock, flags);
8173 free_sched_group(tg);
8174 return ERR_PTR(-ENOMEM);
8177 /* rcu callback to free various structures associated with a task group */
8178 static void free_sched_group_rcu(struct rcu_head *rhp)
8180 /* now it should be safe to free those cfs_rqs */
8181 free_sched_group(container_of(rhp, struct task_group, rcu));
8184 /* Destroy runqueue etc associated with a task group */
8185 void sched_destroy_group(struct task_group *tg)
8187 unsigned long flags;
8190 spin_lock_irqsave(&task_group_lock, flags);
8191 for_each_possible_cpu(i) {
8192 unregister_fair_sched_group(tg, i);
8193 unregister_rt_sched_group(tg, i);
8195 list_del_rcu(&tg->list);
8196 list_del_rcu(&tg->siblings);
8197 spin_unlock_irqrestore(&task_group_lock, flags);
8199 /* wait for possible concurrent references to cfs_rqs complete */
8200 call_rcu(&tg->rcu, free_sched_group_rcu);
8203 /* change task's runqueue when it moves between groups.
8204 * The caller of this function should have put the task in its new group
8205 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8206 * reflect its new group.
8208 void sched_move_task(struct task_struct *tsk)
8211 unsigned long flags;
8214 rq = task_rq_lock(tsk, &flags);
8216 update_rq_clock(rq);
8218 running = task_current(rq, tsk);
8219 on_rq = tsk->se.on_rq;
8222 dequeue_task(rq, tsk, 0);
8223 if (unlikely(running))
8224 tsk->sched_class->put_prev_task(rq, tsk);
8226 set_task_rq(tsk, task_cpu(tsk));
8228 #ifdef CONFIG_FAIR_GROUP_SCHED
8229 if (tsk->sched_class->moved_group)
8230 tsk->sched_class->moved_group(tsk, on_rq);
8233 if (unlikely(running))
8234 tsk->sched_class->set_curr_task(rq);
8236 enqueue_task(rq, tsk, 0, false);
8238 task_rq_unlock(rq, &flags);
8240 #endif /* CONFIG_CGROUP_SCHED */
8242 #ifdef CONFIG_FAIR_GROUP_SCHED
8243 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8245 struct cfs_rq *cfs_rq = se->cfs_rq;
8250 dequeue_entity(cfs_rq, se, 0);
8252 se->load.weight = shares;
8253 se->load.inv_weight = 0;
8256 enqueue_entity(cfs_rq, se, 0);
8259 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8261 struct cfs_rq *cfs_rq = se->cfs_rq;
8262 struct rq *rq = cfs_rq->rq;
8263 unsigned long flags;
8265 raw_spin_lock_irqsave(&rq->lock, flags);
8266 __set_se_shares(se, shares);
8267 raw_spin_unlock_irqrestore(&rq->lock, flags);
8270 static DEFINE_MUTEX(shares_mutex);
8272 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8275 unsigned long flags;
8278 * We can't change the weight of the root cgroup.
8283 if (shares < MIN_SHARES)
8284 shares = MIN_SHARES;
8285 else if (shares > MAX_SHARES)
8286 shares = MAX_SHARES;
8288 mutex_lock(&shares_mutex);
8289 if (tg->shares == shares)
8292 spin_lock_irqsave(&task_group_lock, flags);
8293 for_each_possible_cpu(i)
8294 unregister_fair_sched_group(tg, i);
8295 list_del_rcu(&tg->siblings);
8296 spin_unlock_irqrestore(&task_group_lock, flags);
8298 /* wait for any ongoing reference to this group to finish */
8299 synchronize_sched();
8302 * Now we are free to modify the group's share on each cpu
8303 * w/o tripping rebalance_share or load_balance_fair.
8305 tg->shares = shares;
8306 for_each_possible_cpu(i) {
8310 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8311 set_se_shares(tg->se[i], shares);
8315 * Enable load balance activity on this group, by inserting it back on
8316 * each cpu's rq->leaf_cfs_rq_list.
8318 spin_lock_irqsave(&task_group_lock, flags);
8319 for_each_possible_cpu(i)
8320 register_fair_sched_group(tg, i);
8321 list_add_rcu(&tg->siblings, &tg->parent->children);
8322 spin_unlock_irqrestore(&task_group_lock, flags);
8324 mutex_unlock(&shares_mutex);
8328 unsigned long sched_group_shares(struct task_group *tg)
8334 #ifdef CONFIG_RT_GROUP_SCHED
8336 * Ensure that the real time constraints are schedulable.
8338 static DEFINE_MUTEX(rt_constraints_mutex);
8340 static unsigned long to_ratio(u64 period, u64 runtime)
8342 if (runtime == RUNTIME_INF)
8345 return div64_u64(runtime << 20, period);
8348 /* Must be called with tasklist_lock held */
8349 static inline int tg_has_rt_tasks(struct task_group *tg)
8351 struct task_struct *g, *p;
8353 do_each_thread(g, p) {
8354 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8356 } while_each_thread(g, p);
8361 struct rt_schedulable_data {
8362 struct task_group *tg;
8367 static int tg_schedulable(struct task_group *tg, void *data)
8369 struct rt_schedulable_data *d = data;
8370 struct task_group *child;
8371 unsigned long total, sum = 0;
8372 u64 period, runtime;
8374 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8375 runtime = tg->rt_bandwidth.rt_runtime;
8378 period = d->rt_period;
8379 runtime = d->rt_runtime;
8383 * Cannot have more runtime than the period.
8385 if (runtime > period && runtime != RUNTIME_INF)
8389 * Ensure we don't starve existing RT tasks.
8391 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8394 total = to_ratio(period, runtime);
8397 * Nobody can have more than the global setting allows.
8399 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8403 * The sum of our children's runtime should not exceed our own.
8405 list_for_each_entry_rcu(child, &tg->children, siblings) {
8406 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8407 runtime = child->rt_bandwidth.rt_runtime;
8409 if (child == d->tg) {
8410 period = d->rt_period;
8411 runtime = d->rt_runtime;
8414 sum += to_ratio(period, runtime);
8423 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8425 struct rt_schedulable_data data = {
8427 .rt_period = period,
8428 .rt_runtime = runtime,
8431 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8434 static int tg_set_bandwidth(struct task_group *tg,
8435 u64 rt_period, u64 rt_runtime)
8439 mutex_lock(&rt_constraints_mutex);
8440 read_lock(&tasklist_lock);
8441 err = __rt_schedulable(tg, rt_period, rt_runtime);
8445 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8446 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8447 tg->rt_bandwidth.rt_runtime = rt_runtime;
8449 for_each_possible_cpu(i) {
8450 struct rt_rq *rt_rq = tg->rt_rq[i];
8452 raw_spin_lock(&rt_rq->rt_runtime_lock);
8453 rt_rq->rt_runtime = rt_runtime;
8454 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8456 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8458 read_unlock(&tasklist_lock);
8459 mutex_unlock(&rt_constraints_mutex);
8464 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8466 u64 rt_runtime, rt_period;
8468 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8469 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8470 if (rt_runtime_us < 0)
8471 rt_runtime = RUNTIME_INF;
8473 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8476 long sched_group_rt_runtime(struct task_group *tg)
8480 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8483 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8484 do_div(rt_runtime_us, NSEC_PER_USEC);
8485 return rt_runtime_us;
8488 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8490 u64 rt_runtime, rt_period;
8492 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8493 rt_runtime = tg->rt_bandwidth.rt_runtime;
8498 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8501 long sched_group_rt_period(struct task_group *tg)
8505 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8506 do_div(rt_period_us, NSEC_PER_USEC);
8507 return rt_period_us;
8510 static int sched_rt_global_constraints(void)
8512 u64 runtime, period;
8515 if (sysctl_sched_rt_period <= 0)
8518 runtime = global_rt_runtime();
8519 period = global_rt_period();
8522 * Sanity check on the sysctl variables.
8524 if (runtime > period && runtime != RUNTIME_INF)
8527 mutex_lock(&rt_constraints_mutex);
8528 read_lock(&tasklist_lock);
8529 ret = __rt_schedulable(NULL, 0, 0);
8530 read_unlock(&tasklist_lock);
8531 mutex_unlock(&rt_constraints_mutex);
8536 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8538 /* Don't accept realtime tasks when there is no way for them to run */
8539 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8545 #else /* !CONFIG_RT_GROUP_SCHED */
8546 static int sched_rt_global_constraints(void)
8548 unsigned long flags;
8551 if (sysctl_sched_rt_period <= 0)
8555 * There's always some RT tasks in the root group
8556 * -- migration, kstopmachine etc..
8558 if (sysctl_sched_rt_runtime == 0)
8561 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8562 for_each_possible_cpu(i) {
8563 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8565 raw_spin_lock(&rt_rq->rt_runtime_lock);
8566 rt_rq->rt_runtime = global_rt_runtime();
8567 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8569 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8573 #endif /* CONFIG_RT_GROUP_SCHED */
8575 int sched_rt_handler(struct ctl_table *table, int write,
8576 void __user *buffer, size_t *lenp,
8580 int old_period, old_runtime;
8581 static DEFINE_MUTEX(mutex);
8584 old_period = sysctl_sched_rt_period;
8585 old_runtime = sysctl_sched_rt_runtime;
8587 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8589 if (!ret && write) {
8590 ret = sched_rt_global_constraints();
8592 sysctl_sched_rt_period = old_period;
8593 sysctl_sched_rt_runtime = old_runtime;
8595 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8596 def_rt_bandwidth.rt_period =
8597 ns_to_ktime(global_rt_period());
8600 mutex_unlock(&mutex);
8605 #ifdef CONFIG_CGROUP_SCHED
8607 /* return corresponding task_group object of a cgroup */
8608 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8610 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8611 struct task_group, css);
8614 static struct cgroup_subsys_state *
8615 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8617 struct task_group *tg, *parent;
8619 if (!cgrp->parent) {
8620 /* This is early initialization for the top cgroup */
8621 return &init_task_group.css;
8624 parent = cgroup_tg(cgrp->parent);
8625 tg = sched_create_group(parent);
8627 return ERR_PTR(-ENOMEM);
8633 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8635 struct task_group *tg = cgroup_tg(cgrp);
8637 sched_destroy_group(tg);
8641 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8643 #ifdef CONFIG_RT_GROUP_SCHED
8644 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8647 /* We don't support RT-tasks being in separate groups */
8648 if (tsk->sched_class != &fair_sched_class)
8655 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8656 struct task_struct *tsk, bool threadgroup)
8658 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8662 struct task_struct *c;
8664 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8665 retval = cpu_cgroup_can_attach_task(cgrp, c);
8677 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8678 struct cgroup *old_cont, struct task_struct *tsk,
8681 sched_move_task(tsk);
8683 struct task_struct *c;
8685 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8692 #ifdef CONFIG_FAIR_GROUP_SCHED
8693 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8696 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8699 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8701 struct task_group *tg = cgroup_tg(cgrp);
8703 return (u64) tg->shares;
8705 #endif /* CONFIG_FAIR_GROUP_SCHED */
8707 #ifdef CONFIG_RT_GROUP_SCHED
8708 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8711 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8714 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8716 return sched_group_rt_runtime(cgroup_tg(cgrp));
8719 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8722 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8725 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8727 return sched_group_rt_period(cgroup_tg(cgrp));
8729 #endif /* CONFIG_RT_GROUP_SCHED */
8731 static struct cftype cpu_files[] = {
8732 #ifdef CONFIG_FAIR_GROUP_SCHED
8735 .read_u64 = cpu_shares_read_u64,
8736 .write_u64 = cpu_shares_write_u64,
8739 #ifdef CONFIG_RT_GROUP_SCHED
8741 .name = "rt_runtime_us",
8742 .read_s64 = cpu_rt_runtime_read,
8743 .write_s64 = cpu_rt_runtime_write,
8746 .name = "rt_period_us",
8747 .read_u64 = cpu_rt_period_read_uint,
8748 .write_u64 = cpu_rt_period_write_uint,
8753 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8755 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8758 struct cgroup_subsys cpu_cgroup_subsys = {
8760 .create = cpu_cgroup_create,
8761 .destroy = cpu_cgroup_destroy,
8762 .can_attach = cpu_cgroup_can_attach,
8763 .attach = cpu_cgroup_attach,
8764 .populate = cpu_cgroup_populate,
8765 .subsys_id = cpu_cgroup_subsys_id,
8769 #endif /* CONFIG_CGROUP_SCHED */
8771 #ifdef CONFIG_CGROUP_CPUACCT
8774 * CPU accounting code for task groups.
8776 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8777 * (balbir@in.ibm.com).
8780 /* track cpu usage of a group of tasks and its child groups */
8782 struct cgroup_subsys_state css;
8783 /* cpuusage holds pointer to a u64-type object on every cpu */
8785 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8786 struct cpuacct *parent;
8789 struct cgroup_subsys cpuacct_subsys;
8791 /* return cpu accounting group corresponding to this container */
8792 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8794 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8795 struct cpuacct, css);
8798 /* return cpu accounting group to which this task belongs */
8799 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8801 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8802 struct cpuacct, css);
8805 /* create a new cpu accounting group */
8806 static struct cgroup_subsys_state *cpuacct_create(
8807 struct cgroup_subsys *ss, struct cgroup *cgrp)
8809 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8815 ca->cpuusage = alloc_percpu(u64);
8819 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8820 if (percpu_counter_init(&ca->cpustat[i], 0))
8821 goto out_free_counters;
8824 ca->parent = cgroup_ca(cgrp->parent);
8830 percpu_counter_destroy(&ca->cpustat[i]);
8831 free_percpu(ca->cpuusage);
8835 return ERR_PTR(-ENOMEM);
8838 /* destroy an existing cpu accounting group */
8840 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8842 struct cpuacct *ca = cgroup_ca(cgrp);
8845 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8846 percpu_counter_destroy(&ca->cpustat[i]);
8847 free_percpu(ca->cpuusage);
8851 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8853 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8856 #ifndef CONFIG_64BIT
8858 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8860 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8862 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8870 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8872 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8874 #ifndef CONFIG_64BIT
8876 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8878 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8880 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8886 /* return total cpu usage (in nanoseconds) of a group */
8887 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8889 struct cpuacct *ca = cgroup_ca(cgrp);
8890 u64 totalcpuusage = 0;
8893 for_each_present_cpu(i)
8894 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8896 return totalcpuusage;
8899 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8902 struct cpuacct *ca = cgroup_ca(cgrp);
8911 for_each_present_cpu(i)
8912 cpuacct_cpuusage_write(ca, i, 0);
8918 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8921 struct cpuacct *ca = cgroup_ca(cgroup);
8925 for_each_present_cpu(i) {
8926 percpu = cpuacct_cpuusage_read(ca, i);
8927 seq_printf(m, "%llu ", (unsigned long long) percpu);
8929 seq_printf(m, "\n");
8933 static const char *cpuacct_stat_desc[] = {
8934 [CPUACCT_STAT_USER] = "user",
8935 [CPUACCT_STAT_SYSTEM] = "system",
8938 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8939 struct cgroup_map_cb *cb)
8941 struct cpuacct *ca = cgroup_ca(cgrp);
8944 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8945 s64 val = percpu_counter_read(&ca->cpustat[i]);
8946 val = cputime64_to_clock_t(val);
8947 cb->fill(cb, cpuacct_stat_desc[i], val);
8952 static struct cftype files[] = {
8955 .read_u64 = cpuusage_read,
8956 .write_u64 = cpuusage_write,
8959 .name = "usage_percpu",
8960 .read_seq_string = cpuacct_percpu_seq_read,
8964 .read_map = cpuacct_stats_show,
8968 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8970 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8974 * charge this task's execution time to its accounting group.
8976 * called with rq->lock held.
8978 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8983 if (unlikely(!cpuacct_subsys.active))
8986 cpu = task_cpu(tsk);
8992 for (; ca; ca = ca->parent) {
8993 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8994 *cpuusage += cputime;
9001 * Charge the system/user time to the task's accounting group.
9003 static void cpuacct_update_stats(struct task_struct *tsk,
9004 enum cpuacct_stat_index idx, cputime_t val)
9008 if (unlikely(!cpuacct_subsys.active))
9015 percpu_counter_add(&ca->cpustat[idx], val);
9021 struct cgroup_subsys cpuacct_subsys = {
9023 .create = cpuacct_create,
9024 .destroy = cpuacct_destroy,
9025 .populate = cpuacct_populate,
9026 .subsys_id = cpuacct_subsys_id,
9028 #endif /* CONFIG_CGROUP_CPUACCT */
9032 int rcu_expedited_torture_stats(char *page)
9036 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9038 void synchronize_sched_expedited(void)
9041 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9043 #else /* #ifndef CONFIG_SMP */
9045 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9046 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9048 #define RCU_EXPEDITED_STATE_POST -2
9049 #define RCU_EXPEDITED_STATE_IDLE -1
9051 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9053 int rcu_expedited_torture_stats(char *page)
9058 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9059 for_each_online_cpu(cpu) {
9060 cnt += sprintf(&page[cnt], " %d:%d",
9061 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9063 cnt += sprintf(&page[cnt], "\n");
9066 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9068 static long synchronize_sched_expedited_count;
9071 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9072 * approach to force grace period to end quickly. This consumes
9073 * significant time on all CPUs, and is thus not recommended for
9074 * any sort of common-case code.
9076 * Note that it is illegal to call this function while holding any
9077 * lock that is acquired by a CPU-hotplug notifier. Failing to
9078 * observe this restriction will result in deadlock.
9080 void synchronize_sched_expedited(void)
9083 unsigned long flags;
9084 bool need_full_sync = 0;
9086 struct migration_req *req;
9090 smp_mb(); /* ensure prior mod happens before capturing snap. */
9091 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9093 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9095 if (trycount++ < 10)
9096 udelay(trycount * num_online_cpus());
9098 synchronize_sched();
9101 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9102 smp_mb(); /* ensure test happens before caller kfree */
9107 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9108 for_each_online_cpu(cpu) {
9110 req = &per_cpu(rcu_migration_req, cpu);
9111 init_completion(&req->done);
9113 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9114 raw_spin_lock_irqsave(&rq->lock, flags);
9115 list_add(&req->list, &rq->migration_queue);
9116 raw_spin_unlock_irqrestore(&rq->lock, flags);
9117 wake_up_process(rq->migration_thread);
9119 for_each_online_cpu(cpu) {
9120 rcu_expedited_state = cpu;
9121 req = &per_cpu(rcu_migration_req, cpu);
9123 wait_for_completion(&req->done);
9124 raw_spin_lock_irqsave(&rq->lock, flags);
9125 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9127 req->dest_cpu = RCU_MIGRATION_IDLE;
9128 raw_spin_unlock_irqrestore(&rq->lock, flags);
9130 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9131 synchronize_sched_expedited_count++;
9132 mutex_unlock(&rcu_sched_expedited_mutex);
9135 synchronize_sched();
9137 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9139 #endif /* #else #ifndef CONFIG_SMP */