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);
2291 * - fork, @p is stable because it isn't on the tasklist yet
2293 * - exec, @p is unstable, retry loop
2295 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so
2296 * we should be good.
2299 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2301 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2304 * In order not to call set_task_cpu() on a blocking task we need
2305 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2308 * Since this is common to all placement strategies, this lives here.
2310 * [ this allows ->select_task() to simply return task_cpu(p) and
2311 * not worry about this generic constraint ]
2313 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2315 cpu = select_fallback_rq(task_cpu(p), p);
2322 * try_to_wake_up - wake up a thread
2323 * @p: the to-be-woken-up thread
2324 * @state: the mask of task states that can be woken
2325 * @sync: do a synchronous wakeup?
2327 * Put it on the run-queue if it's not already there. The "current"
2328 * thread is always on the run-queue (except when the actual
2329 * re-schedule is in progress), and as such you're allowed to do
2330 * the simpler "current->state = TASK_RUNNING" to mark yourself
2331 * runnable without the overhead of this.
2333 * returns failure only if the task is already active.
2335 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2338 int cpu, orig_cpu, this_cpu, success = 0;
2339 unsigned long flags;
2340 struct rq *rq, *orig_rq;
2342 if (!sched_feat(SYNC_WAKEUPS))
2343 wake_flags &= ~WF_SYNC;
2345 this_cpu = get_cpu();
2348 rq = orig_rq = task_rq_lock(p, &flags);
2349 update_rq_clock(rq);
2350 if (!(p->state & state))
2360 if (unlikely(task_running(rq, p)))
2364 * In order to handle concurrent wakeups and release the rq->lock
2365 * we put the task in TASK_WAKING state.
2367 * First fix up the nr_uninterruptible count:
2369 if (task_contributes_to_load(p))
2370 rq->nr_uninterruptible--;
2371 p->state = TASK_WAKING;
2373 if (p->sched_class->task_waking)
2374 p->sched_class->task_waking(rq, p);
2376 __task_rq_unlock(rq);
2378 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2379 if (cpu != orig_cpu)
2380 set_task_cpu(p, cpu);
2382 rq = __task_rq_lock(p);
2383 update_rq_clock(rq);
2385 WARN_ON(p->state != TASK_WAKING);
2388 #ifdef CONFIG_SCHEDSTATS
2389 schedstat_inc(rq, ttwu_count);
2390 if (cpu == this_cpu)
2391 schedstat_inc(rq, ttwu_local);
2393 struct sched_domain *sd;
2394 for_each_domain(this_cpu, sd) {
2395 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2396 schedstat_inc(sd, ttwu_wake_remote);
2401 #endif /* CONFIG_SCHEDSTATS */
2404 #endif /* CONFIG_SMP */
2405 schedstat_inc(p, se.nr_wakeups);
2406 if (wake_flags & WF_SYNC)
2407 schedstat_inc(p, se.nr_wakeups_sync);
2408 if (orig_cpu != cpu)
2409 schedstat_inc(p, se.nr_wakeups_migrate);
2410 if (cpu == this_cpu)
2411 schedstat_inc(p, se.nr_wakeups_local);
2413 schedstat_inc(p, se.nr_wakeups_remote);
2414 activate_task(rq, p, 1);
2418 * Only attribute actual wakeups done by this task.
2420 if (!in_interrupt()) {
2421 struct sched_entity *se = ¤t->se;
2422 u64 sample = se->sum_exec_runtime;
2424 if (se->last_wakeup)
2425 sample -= se->last_wakeup;
2427 sample -= se->start_runtime;
2428 update_avg(&se->avg_wakeup, sample);
2430 se->last_wakeup = se->sum_exec_runtime;
2434 trace_sched_wakeup(rq, p, success);
2435 check_preempt_curr(rq, p, wake_flags);
2437 p->state = TASK_RUNNING;
2439 if (p->sched_class->task_woken)
2440 p->sched_class->task_woken(rq, p);
2442 if (unlikely(rq->idle_stamp)) {
2443 u64 delta = rq->clock - rq->idle_stamp;
2444 u64 max = 2*sysctl_sched_migration_cost;
2449 update_avg(&rq->avg_idle, delta);
2454 task_rq_unlock(rq, &flags);
2461 * wake_up_process - Wake up a specific process
2462 * @p: The process to be woken up.
2464 * Attempt to wake up the nominated process and move it to the set of runnable
2465 * processes. Returns 1 if the process was woken up, 0 if it was already
2468 * It may be assumed that this function implies a write memory barrier before
2469 * changing the task state if and only if any tasks are woken up.
2471 int wake_up_process(struct task_struct *p)
2473 return try_to_wake_up(p, TASK_ALL, 0);
2475 EXPORT_SYMBOL(wake_up_process);
2477 int wake_up_state(struct task_struct *p, unsigned int state)
2479 return try_to_wake_up(p, state, 0);
2483 * Perform scheduler related setup for a newly forked process p.
2484 * p is forked by current.
2486 * __sched_fork() is basic setup used by init_idle() too:
2488 static void __sched_fork(struct task_struct *p)
2490 p->se.exec_start = 0;
2491 p->se.sum_exec_runtime = 0;
2492 p->se.prev_sum_exec_runtime = 0;
2493 p->se.nr_migrations = 0;
2494 p->se.last_wakeup = 0;
2495 p->se.avg_overlap = 0;
2496 p->se.start_runtime = 0;
2497 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2499 #ifdef CONFIG_SCHEDSTATS
2500 p->se.wait_start = 0;
2502 p->se.wait_count = 0;
2505 p->se.sleep_start = 0;
2506 p->se.sleep_max = 0;
2507 p->se.sum_sleep_runtime = 0;
2509 p->se.block_start = 0;
2510 p->se.block_max = 0;
2512 p->se.slice_max = 0;
2514 p->se.nr_migrations_cold = 0;
2515 p->se.nr_failed_migrations_affine = 0;
2516 p->se.nr_failed_migrations_running = 0;
2517 p->se.nr_failed_migrations_hot = 0;
2518 p->se.nr_forced_migrations = 0;
2520 p->se.nr_wakeups = 0;
2521 p->se.nr_wakeups_sync = 0;
2522 p->se.nr_wakeups_migrate = 0;
2523 p->se.nr_wakeups_local = 0;
2524 p->se.nr_wakeups_remote = 0;
2525 p->se.nr_wakeups_affine = 0;
2526 p->se.nr_wakeups_affine_attempts = 0;
2527 p->se.nr_wakeups_passive = 0;
2528 p->se.nr_wakeups_idle = 0;
2532 INIT_LIST_HEAD(&p->rt.run_list);
2534 INIT_LIST_HEAD(&p->se.group_node);
2536 #ifdef CONFIG_PREEMPT_NOTIFIERS
2537 INIT_HLIST_HEAD(&p->preempt_notifiers);
2542 * fork()/clone()-time setup:
2544 void sched_fork(struct task_struct *p, int clone_flags)
2546 int cpu = get_cpu();
2550 * We mark the process as waking here. This guarantees that
2551 * nobody will actually run it, and a signal or other external
2552 * event cannot wake it up and insert it on the runqueue either.
2554 p->state = TASK_WAKING;
2557 * Revert to default priority/policy on fork if requested.
2559 if (unlikely(p->sched_reset_on_fork)) {
2560 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2561 p->policy = SCHED_NORMAL;
2562 p->normal_prio = p->static_prio;
2565 if (PRIO_TO_NICE(p->static_prio) < 0) {
2566 p->static_prio = NICE_TO_PRIO(0);
2567 p->normal_prio = p->static_prio;
2572 * We don't need the reset flag anymore after the fork. It has
2573 * fulfilled its duty:
2575 p->sched_reset_on_fork = 0;
2579 * Make sure we do not leak PI boosting priority to the child.
2581 p->prio = current->normal_prio;
2583 if (!rt_prio(p->prio))
2584 p->sched_class = &fair_sched_class;
2586 if (p->sched_class->task_fork)
2587 p->sched_class->task_fork(p);
2590 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2592 set_task_cpu(p, cpu);
2594 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2595 if (likely(sched_info_on()))
2596 memset(&p->sched_info, 0, sizeof(p->sched_info));
2598 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2601 #ifdef CONFIG_PREEMPT
2602 /* Want to start with kernel preemption disabled. */
2603 task_thread_info(p)->preempt_count = 1;
2605 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2611 * wake_up_new_task - wake up a newly created task for the first time.
2613 * This function will do some initial scheduler statistics housekeeping
2614 * that must be done for every newly created context, then puts the task
2615 * on the runqueue and wakes it.
2617 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2619 unsigned long flags;
2622 rq = task_rq_lock(p, &flags);
2623 BUG_ON(p->state != TASK_WAKING);
2624 p->state = TASK_RUNNING;
2625 update_rq_clock(rq);
2626 activate_task(rq, p, 0);
2627 trace_sched_wakeup_new(rq, p, 1);
2628 check_preempt_curr(rq, p, WF_FORK);
2630 if (p->sched_class->task_woken)
2631 p->sched_class->task_woken(rq, p);
2633 task_rq_unlock(rq, &flags);
2636 #ifdef CONFIG_PREEMPT_NOTIFIERS
2639 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2640 * @notifier: notifier struct to register
2642 void preempt_notifier_register(struct preempt_notifier *notifier)
2644 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2646 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2649 * preempt_notifier_unregister - no longer interested in preemption notifications
2650 * @notifier: notifier struct to unregister
2652 * This is safe to call from within a preemption notifier.
2654 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2656 hlist_del(¬ifier->link);
2658 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2660 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2662 struct preempt_notifier *notifier;
2663 struct hlist_node *node;
2665 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2666 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2670 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2671 struct task_struct *next)
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_out(notifier, next);
2680 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2682 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2687 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2688 struct task_struct *next)
2692 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2695 * prepare_task_switch - prepare to switch tasks
2696 * @rq: the runqueue preparing to switch
2697 * @prev: the current task that is being switched out
2698 * @next: the task we are going to switch to.
2700 * This is called with the rq lock held and interrupts off. It must
2701 * be paired with a subsequent finish_task_switch after the context
2704 * prepare_task_switch sets up locking and calls architecture specific
2708 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2709 struct task_struct *next)
2711 fire_sched_out_preempt_notifiers(prev, next);
2712 prepare_lock_switch(rq, next);
2713 prepare_arch_switch(next);
2717 * finish_task_switch - clean up after a task-switch
2718 * @rq: runqueue associated with task-switch
2719 * @prev: the thread we just switched away from.
2721 * finish_task_switch must be called after the context switch, paired
2722 * with a prepare_task_switch call before the context switch.
2723 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2724 * and do any other architecture-specific cleanup actions.
2726 * Note that we may have delayed dropping an mm in context_switch(). If
2727 * so, we finish that here outside of the runqueue lock. (Doing it
2728 * with the lock held can cause deadlocks; see schedule() for
2731 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2732 __releases(rq->lock)
2734 struct mm_struct *mm = rq->prev_mm;
2740 * A task struct has one reference for the use as "current".
2741 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2742 * schedule one last time. The schedule call will never return, and
2743 * the scheduled task must drop that reference.
2744 * The test for TASK_DEAD must occur while the runqueue locks are
2745 * still held, otherwise prev could be scheduled on another cpu, die
2746 * there before we look at prev->state, and then the reference would
2748 * Manfred Spraul <manfred@colorfullife.com>
2750 prev_state = prev->state;
2751 finish_arch_switch(prev);
2752 perf_event_task_sched_in(current, cpu_of(rq));
2753 finish_lock_switch(rq, prev);
2755 fire_sched_in_preempt_notifiers(current);
2758 if (unlikely(prev_state == TASK_DEAD)) {
2760 * Remove function-return probe instances associated with this
2761 * task and put them back on the free list.
2763 kprobe_flush_task(prev);
2764 put_task_struct(prev);
2770 /* assumes rq->lock is held */
2771 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2773 if (prev->sched_class->pre_schedule)
2774 prev->sched_class->pre_schedule(rq, prev);
2777 /* rq->lock is NOT held, but preemption is disabled */
2778 static inline void post_schedule(struct rq *rq)
2780 if (rq->post_schedule) {
2781 unsigned long flags;
2783 raw_spin_lock_irqsave(&rq->lock, flags);
2784 if (rq->curr->sched_class->post_schedule)
2785 rq->curr->sched_class->post_schedule(rq);
2786 raw_spin_unlock_irqrestore(&rq->lock, flags);
2788 rq->post_schedule = 0;
2794 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2798 static inline void post_schedule(struct rq *rq)
2805 * schedule_tail - first thing a freshly forked thread must call.
2806 * @prev: the thread we just switched away from.
2808 asmlinkage void schedule_tail(struct task_struct *prev)
2809 __releases(rq->lock)
2811 struct rq *rq = this_rq();
2813 finish_task_switch(rq, prev);
2816 * FIXME: do we need to worry about rq being invalidated by the
2821 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2822 /* In this case, finish_task_switch does not reenable preemption */
2825 if (current->set_child_tid)
2826 put_user(task_pid_vnr(current), current->set_child_tid);
2830 * context_switch - switch to the new MM and the new
2831 * thread's register state.
2834 context_switch(struct rq *rq, struct task_struct *prev,
2835 struct task_struct *next)
2837 struct mm_struct *mm, *oldmm;
2839 prepare_task_switch(rq, prev, next);
2840 trace_sched_switch(rq, prev, next);
2842 oldmm = prev->active_mm;
2844 * For paravirt, this is coupled with an exit in switch_to to
2845 * combine the page table reload and the switch backend into
2848 arch_start_context_switch(prev);
2851 next->active_mm = oldmm;
2852 atomic_inc(&oldmm->mm_count);
2853 enter_lazy_tlb(oldmm, next);
2855 switch_mm(oldmm, mm, next);
2857 if (likely(!prev->mm)) {
2858 prev->active_mm = NULL;
2859 rq->prev_mm = oldmm;
2862 * Since the runqueue lock will be released by the next
2863 * task (which is an invalid locking op but in the case
2864 * of the scheduler it's an obvious special-case), so we
2865 * do an early lockdep release here:
2867 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2868 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2871 /* Here we just switch the register state and the stack. */
2872 switch_to(prev, next, prev);
2876 * this_rq must be evaluated again because prev may have moved
2877 * CPUs since it called schedule(), thus the 'rq' on its stack
2878 * frame will be invalid.
2880 finish_task_switch(this_rq(), prev);
2884 * nr_running, nr_uninterruptible and nr_context_switches:
2886 * externally visible scheduler statistics: current number of runnable
2887 * threads, current number of uninterruptible-sleeping threads, total
2888 * number of context switches performed since bootup.
2890 unsigned long nr_running(void)
2892 unsigned long i, sum = 0;
2894 for_each_online_cpu(i)
2895 sum += cpu_rq(i)->nr_running;
2900 unsigned long nr_uninterruptible(void)
2902 unsigned long i, sum = 0;
2904 for_each_possible_cpu(i)
2905 sum += cpu_rq(i)->nr_uninterruptible;
2908 * Since we read the counters lockless, it might be slightly
2909 * inaccurate. Do not allow it to go below zero though:
2911 if (unlikely((long)sum < 0))
2917 unsigned long long nr_context_switches(void)
2920 unsigned long long sum = 0;
2922 for_each_possible_cpu(i)
2923 sum += cpu_rq(i)->nr_switches;
2928 unsigned long nr_iowait(void)
2930 unsigned long i, sum = 0;
2932 for_each_possible_cpu(i)
2933 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2938 unsigned long nr_iowait_cpu(void)
2940 struct rq *this = this_rq();
2941 return atomic_read(&this->nr_iowait);
2944 unsigned long this_cpu_load(void)
2946 struct rq *this = this_rq();
2947 return this->cpu_load[0];
2951 /* Variables and functions for calc_load */
2952 static atomic_long_t calc_load_tasks;
2953 static unsigned long calc_load_update;
2954 unsigned long avenrun[3];
2955 EXPORT_SYMBOL(avenrun);
2958 * get_avenrun - get the load average array
2959 * @loads: pointer to dest load array
2960 * @offset: offset to add
2961 * @shift: shift count to shift the result left
2963 * These values are estimates at best, so no need for locking.
2965 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2967 loads[0] = (avenrun[0] + offset) << shift;
2968 loads[1] = (avenrun[1] + offset) << shift;
2969 loads[2] = (avenrun[2] + offset) << shift;
2972 static unsigned long
2973 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2976 load += active * (FIXED_1 - exp);
2977 return load >> FSHIFT;
2981 * calc_load - update the avenrun load estimates 10 ticks after the
2982 * CPUs have updated calc_load_tasks.
2984 void calc_global_load(void)
2986 unsigned long upd = calc_load_update + 10;
2989 if (time_before(jiffies, upd))
2992 active = atomic_long_read(&calc_load_tasks);
2993 active = active > 0 ? active * FIXED_1 : 0;
2995 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2996 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2997 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2999 calc_load_update += LOAD_FREQ;
3003 * Either called from update_cpu_load() or from a cpu going idle
3005 static void calc_load_account_active(struct rq *this_rq)
3007 long nr_active, delta;
3009 nr_active = this_rq->nr_running;
3010 nr_active += (long) this_rq->nr_uninterruptible;
3012 if (nr_active != this_rq->calc_load_active) {
3013 delta = nr_active - this_rq->calc_load_active;
3014 this_rq->calc_load_active = nr_active;
3015 atomic_long_add(delta, &calc_load_tasks);
3020 * Update rq->cpu_load[] statistics. This function is usually called every
3021 * scheduler tick (TICK_NSEC).
3023 static void update_cpu_load(struct rq *this_rq)
3025 unsigned long this_load = this_rq->load.weight;
3028 this_rq->nr_load_updates++;
3030 /* Update our load: */
3031 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3032 unsigned long old_load, new_load;
3034 /* scale is effectively 1 << i now, and >> i divides by scale */
3036 old_load = this_rq->cpu_load[i];
3037 new_load = this_load;
3039 * Round up the averaging division if load is increasing. This
3040 * prevents us from getting stuck on 9 if the load is 10, for
3043 if (new_load > old_load)
3044 new_load += scale-1;
3045 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3048 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3049 this_rq->calc_load_update += LOAD_FREQ;
3050 calc_load_account_active(this_rq);
3057 * sched_exec - execve() is a valuable balancing opportunity, because at
3058 * this point the task has the smallest effective memory and cache footprint.
3060 void sched_exec(void)
3062 struct task_struct *p = current;
3063 struct migration_req req;
3064 int dest_cpu, this_cpu;
3065 unsigned long flags;
3069 this_cpu = get_cpu();
3070 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3071 if (dest_cpu == this_cpu) {
3076 rq = task_rq_lock(p, &flags);
3080 * select_task_rq() can race against ->cpus_allowed
3082 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3083 || unlikely(!cpu_active(dest_cpu))) {
3084 task_rq_unlock(rq, &flags);
3088 /* force the process onto the specified CPU */
3089 if (migrate_task(p, dest_cpu, &req)) {
3090 /* Need to wait for migration thread (might exit: take ref). */
3091 struct task_struct *mt = rq->migration_thread;
3093 get_task_struct(mt);
3094 task_rq_unlock(rq, &flags);
3095 wake_up_process(mt);
3096 put_task_struct(mt);
3097 wait_for_completion(&req.done);
3101 task_rq_unlock(rq, &flags);
3106 DEFINE_PER_CPU(struct kernel_stat, kstat);
3108 EXPORT_PER_CPU_SYMBOL(kstat);
3111 * Return any ns on the sched_clock that have not yet been accounted in
3112 * @p in case that task is currently running.
3114 * Called with task_rq_lock() held on @rq.
3116 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3120 if (task_current(rq, p)) {
3121 update_rq_clock(rq);
3122 ns = rq->clock - p->se.exec_start;
3130 unsigned long long task_delta_exec(struct task_struct *p)
3132 unsigned long flags;
3136 rq = task_rq_lock(p, &flags);
3137 ns = do_task_delta_exec(p, rq);
3138 task_rq_unlock(rq, &flags);
3144 * Return accounted runtime for the task.
3145 * In case the task is currently running, return the runtime plus current's
3146 * pending runtime that have not been accounted yet.
3148 unsigned long long task_sched_runtime(struct task_struct *p)
3150 unsigned long flags;
3154 rq = task_rq_lock(p, &flags);
3155 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3156 task_rq_unlock(rq, &flags);
3162 * Return sum_exec_runtime for the thread group.
3163 * In case the task is currently running, return the sum plus current's
3164 * pending runtime that have not been accounted yet.
3166 * Note that the thread group might have other running tasks as well,
3167 * so the return value not includes other pending runtime that other
3168 * running tasks might have.
3170 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3172 struct task_cputime totals;
3173 unsigned long flags;
3177 rq = task_rq_lock(p, &flags);
3178 thread_group_cputime(p, &totals);
3179 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3180 task_rq_unlock(rq, &flags);
3186 * Account user cpu time to a process.
3187 * @p: the process that the cpu time gets accounted to
3188 * @cputime: the cpu time spent in user space since the last update
3189 * @cputime_scaled: cputime scaled by cpu frequency
3191 void account_user_time(struct task_struct *p, cputime_t cputime,
3192 cputime_t cputime_scaled)
3194 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3197 /* Add user time to process. */
3198 p->utime = cputime_add(p->utime, cputime);
3199 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3200 account_group_user_time(p, cputime);
3202 /* Add user time to cpustat. */
3203 tmp = cputime_to_cputime64(cputime);
3204 if (TASK_NICE(p) > 0)
3205 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3207 cpustat->user = cputime64_add(cpustat->user, tmp);
3209 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3210 /* Account for user time used */
3211 acct_update_integrals(p);
3215 * Account guest cpu time to a process.
3216 * @p: the process that the cpu time gets accounted to
3217 * @cputime: the cpu time spent in virtual machine since the last update
3218 * @cputime_scaled: cputime scaled by cpu frequency
3220 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3221 cputime_t cputime_scaled)
3224 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3226 tmp = cputime_to_cputime64(cputime);
3228 /* Add guest time to process. */
3229 p->utime = cputime_add(p->utime, cputime);
3230 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3231 account_group_user_time(p, cputime);
3232 p->gtime = cputime_add(p->gtime, cputime);
3234 /* Add guest time to cpustat. */
3235 if (TASK_NICE(p) > 0) {
3236 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3237 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3239 cpustat->user = cputime64_add(cpustat->user, tmp);
3240 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3245 * Account system cpu time to a process.
3246 * @p: the process that the cpu time gets accounted to
3247 * @hardirq_offset: the offset to subtract from hardirq_count()
3248 * @cputime: the cpu time spent in kernel space since the last update
3249 * @cputime_scaled: cputime scaled by cpu frequency
3251 void account_system_time(struct task_struct *p, int hardirq_offset,
3252 cputime_t cputime, cputime_t cputime_scaled)
3254 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3257 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3258 account_guest_time(p, cputime, cputime_scaled);
3262 /* Add system time to process. */
3263 p->stime = cputime_add(p->stime, cputime);
3264 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3265 account_group_system_time(p, cputime);
3267 /* Add system time to cpustat. */
3268 tmp = cputime_to_cputime64(cputime);
3269 if (hardirq_count() - hardirq_offset)
3270 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3271 else if (softirq_count())
3272 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3274 cpustat->system = cputime64_add(cpustat->system, tmp);
3276 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3278 /* Account for system time used */
3279 acct_update_integrals(p);
3283 * Account for involuntary wait time.
3284 * @steal: the cpu time spent in involuntary wait
3286 void account_steal_time(cputime_t cputime)
3288 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3289 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3291 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3295 * Account for idle time.
3296 * @cputime: the cpu time spent in idle wait
3298 void account_idle_time(cputime_t cputime)
3300 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3301 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3302 struct rq *rq = this_rq();
3304 if (atomic_read(&rq->nr_iowait) > 0)
3305 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3307 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3310 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3313 * Account a single tick of cpu time.
3314 * @p: the process that the cpu time gets accounted to
3315 * @user_tick: indicates if the tick is a user or a system tick
3317 void account_process_tick(struct task_struct *p, int user_tick)
3319 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3320 struct rq *rq = this_rq();
3323 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3324 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3325 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3328 account_idle_time(cputime_one_jiffy);
3332 * Account multiple ticks of steal time.
3333 * @p: the process from which the cpu time has been stolen
3334 * @ticks: number of stolen ticks
3336 void account_steal_ticks(unsigned long ticks)
3338 account_steal_time(jiffies_to_cputime(ticks));
3342 * Account multiple ticks of idle time.
3343 * @ticks: number of stolen ticks
3345 void account_idle_ticks(unsigned long ticks)
3347 account_idle_time(jiffies_to_cputime(ticks));
3353 * Use precise platform statistics if available:
3355 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3356 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3362 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3364 struct task_cputime cputime;
3366 thread_group_cputime(p, &cputime);
3368 *ut = cputime.utime;
3369 *st = cputime.stime;
3373 #ifndef nsecs_to_cputime
3374 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3377 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3379 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3382 * Use CFS's precise accounting:
3384 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3389 temp = (u64)(rtime * utime);
3390 do_div(temp, total);
3391 utime = (cputime_t)temp;
3396 * Compare with previous values, to keep monotonicity:
3398 p->prev_utime = max(p->prev_utime, utime);
3399 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3401 *ut = p->prev_utime;
3402 *st = p->prev_stime;
3406 * Must be called with siglock held.
3408 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3410 struct signal_struct *sig = p->signal;
3411 struct task_cputime cputime;
3412 cputime_t rtime, utime, total;
3414 thread_group_cputime(p, &cputime);
3416 total = cputime_add(cputime.utime, cputime.stime);
3417 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3422 temp = (u64)(rtime * cputime.utime);
3423 do_div(temp, total);
3424 utime = (cputime_t)temp;
3428 sig->prev_utime = max(sig->prev_utime, utime);
3429 sig->prev_stime = max(sig->prev_stime,
3430 cputime_sub(rtime, sig->prev_utime));
3432 *ut = sig->prev_utime;
3433 *st = sig->prev_stime;
3438 * This function gets called by the timer code, with HZ frequency.
3439 * We call it with interrupts disabled.
3441 * It also gets called by the fork code, when changing the parent's
3444 void scheduler_tick(void)
3446 int cpu = smp_processor_id();
3447 struct rq *rq = cpu_rq(cpu);
3448 struct task_struct *curr = rq->curr;
3452 raw_spin_lock(&rq->lock);
3453 update_rq_clock(rq);
3454 update_cpu_load(rq);
3455 curr->sched_class->task_tick(rq, curr, 0);
3456 raw_spin_unlock(&rq->lock);
3458 perf_event_task_tick(curr, cpu);
3461 rq->idle_at_tick = idle_cpu(cpu);
3462 trigger_load_balance(rq, cpu);
3466 notrace unsigned long get_parent_ip(unsigned long addr)
3468 if (in_lock_functions(addr)) {
3469 addr = CALLER_ADDR2;
3470 if (in_lock_functions(addr))
3471 addr = CALLER_ADDR3;
3476 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3477 defined(CONFIG_PREEMPT_TRACER))
3479 void __kprobes add_preempt_count(int val)
3481 #ifdef CONFIG_DEBUG_PREEMPT
3485 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3488 preempt_count() += val;
3489 #ifdef CONFIG_DEBUG_PREEMPT
3491 * Spinlock count overflowing soon?
3493 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3496 if (preempt_count() == val)
3497 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3499 EXPORT_SYMBOL(add_preempt_count);
3501 void __kprobes sub_preempt_count(int val)
3503 #ifdef CONFIG_DEBUG_PREEMPT
3507 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3510 * Is the spinlock portion underflowing?
3512 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3513 !(preempt_count() & PREEMPT_MASK)))
3517 if (preempt_count() == val)
3518 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3519 preempt_count() -= val;
3521 EXPORT_SYMBOL(sub_preempt_count);
3526 * Print scheduling while atomic bug:
3528 static noinline void __schedule_bug(struct task_struct *prev)
3530 struct pt_regs *regs = get_irq_regs();
3532 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3533 prev->comm, prev->pid, preempt_count());
3535 debug_show_held_locks(prev);
3537 if (irqs_disabled())
3538 print_irqtrace_events(prev);
3547 * Various schedule()-time debugging checks and statistics:
3549 static inline void schedule_debug(struct task_struct *prev)
3552 * Test if we are atomic. Since do_exit() needs to call into
3553 * schedule() atomically, we ignore that path for now.
3554 * Otherwise, whine if we are scheduling when we should not be.
3556 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3557 __schedule_bug(prev);
3559 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3561 schedstat_inc(this_rq(), sched_count);
3562 #ifdef CONFIG_SCHEDSTATS
3563 if (unlikely(prev->lock_depth >= 0)) {
3564 schedstat_inc(this_rq(), bkl_count);
3565 schedstat_inc(prev, sched_info.bkl_count);
3570 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3572 if (prev->state == TASK_RUNNING) {
3573 u64 runtime = prev->se.sum_exec_runtime;
3575 runtime -= prev->se.prev_sum_exec_runtime;
3576 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
3579 * In order to avoid avg_overlap growing stale when we are
3580 * indeed overlapping and hence not getting put to sleep, grow
3581 * the avg_overlap on preemption.
3583 * We use the average preemption runtime because that
3584 * correlates to the amount of cache footprint a task can
3587 update_avg(&prev->se.avg_overlap, runtime);
3589 prev->sched_class->put_prev_task(rq, prev);
3593 * Pick up the highest-prio task:
3595 static inline struct task_struct *
3596 pick_next_task(struct rq *rq)
3598 const struct sched_class *class;
3599 struct task_struct *p;
3602 * Optimization: we know that if all tasks are in
3603 * the fair class we can call that function directly:
3605 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3606 p = fair_sched_class.pick_next_task(rq);
3611 class = sched_class_highest;
3613 p = class->pick_next_task(rq);
3617 * Will never be NULL as the idle class always
3618 * returns a non-NULL p:
3620 class = class->next;
3625 * schedule() is the main scheduler function.
3627 asmlinkage void __sched schedule(void)
3629 struct task_struct *prev, *next;
3630 unsigned long *switch_count;
3636 cpu = smp_processor_id();
3640 switch_count = &prev->nivcsw;
3642 release_kernel_lock(prev);
3643 need_resched_nonpreemptible:
3645 schedule_debug(prev);
3647 if (sched_feat(HRTICK))
3650 raw_spin_lock_irq(&rq->lock);
3651 update_rq_clock(rq);
3652 clear_tsk_need_resched(prev);
3654 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3655 if (unlikely(signal_pending_state(prev->state, prev)))
3656 prev->state = TASK_RUNNING;
3658 deactivate_task(rq, prev, 1);
3659 switch_count = &prev->nvcsw;
3662 pre_schedule(rq, prev);
3664 if (unlikely(!rq->nr_running))
3665 idle_balance(cpu, rq);
3667 put_prev_task(rq, prev);
3668 next = pick_next_task(rq);
3670 if (likely(prev != next)) {
3671 sched_info_switch(prev, next);
3672 perf_event_task_sched_out(prev, next, cpu);
3678 context_switch(rq, prev, next); /* unlocks the rq */
3680 * the context switch might have flipped the stack from under
3681 * us, hence refresh the local variables.
3683 cpu = smp_processor_id();
3686 raw_spin_unlock_irq(&rq->lock);
3690 if (unlikely(reacquire_kernel_lock(current) < 0))
3691 goto need_resched_nonpreemptible;
3693 preempt_enable_no_resched();
3697 EXPORT_SYMBOL(schedule);
3699 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3701 * Look out! "owner" is an entirely speculative pointer
3702 * access and not reliable.
3704 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3709 if (!sched_feat(OWNER_SPIN))
3712 #ifdef CONFIG_DEBUG_PAGEALLOC
3714 * Need to access the cpu field knowing that
3715 * DEBUG_PAGEALLOC could have unmapped it if
3716 * the mutex owner just released it and exited.
3718 if (probe_kernel_address(&owner->cpu, cpu))
3725 * Even if the access succeeded (likely case),
3726 * the cpu field may no longer be valid.
3728 if (cpu >= nr_cpumask_bits)
3732 * We need to validate that we can do a
3733 * get_cpu() and that we have the percpu area.
3735 if (!cpu_online(cpu))
3742 * Owner changed, break to re-assess state.
3744 if (lock->owner != owner)
3748 * Is that owner really running on that cpu?
3750 if (task_thread_info(rq->curr) != owner || need_resched())
3760 #ifdef CONFIG_PREEMPT
3762 * this is the entry point to schedule() from in-kernel preemption
3763 * off of preempt_enable. Kernel preemptions off return from interrupt
3764 * occur there and call schedule directly.
3766 asmlinkage void __sched preempt_schedule(void)
3768 struct thread_info *ti = current_thread_info();
3771 * If there is a non-zero preempt_count or interrupts are disabled,
3772 * we do not want to preempt the current task. Just return..
3774 if (likely(ti->preempt_count || irqs_disabled()))
3778 add_preempt_count(PREEMPT_ACTIVE);
3780 sub_preempt_count(PREEMPT_ACTIVE);
3783 * Check again in case we missed a preemption opportunity
3784 * between schedule and now.
3787 } while (need_resched());
3789 EXPORT_SYMBOL(preempt_schedule);
3792 * this is the entry point to schedule() from kernel preemption
3793 * off of irq context.
3794 * Note, that this is called and return with irqs disabled. This will
3795 * protect us against recursive calling from irq.
3797 asmlinkage void __sched preempt_schedule_irq(void)
3799 struct thread_info *ti = current_thread_info();
3801 /* Catch callers which need to be fixed */
3802 BUG_ON(ti->preempt_count || !irqs_disabled());
3805 add_preempt_count(PREEMPT_ACTIVE);
3808 local_irq_disable();
3809 sub_preempt_count(PREEMPT_ACTIVE);
3812 * Check again in case we missed a preemption opportunity
3813 * between schedule and now.
3816 } while (need_resched());
3819 #endif /* CONFIG_PREEMPT */
3821 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3824 return try_to_wake_up(curr->private, mode, wake_flags);
3826 EXPORT_SYMBOL(default_wake_function);
3829 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3830 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3831 * number) then we wake all the non-exclusive tasks and one exclusive task.
3833 * There are circumstances in which we can try to wake a task which has already
3834 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3835 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3837 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3838 int nr_exclusive, int wake_flags, void *key)
3840 wait_queue_t *curr, *next;
3842 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3843 unsigned flags = curr->flags;
3845 if (curr->func(curr, mode, wake_flags, key) &&
3846 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3852 * __wake_up - wake up threads blocked on a waitqueue.
3854 * @mode: which threads
3855 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3856 * @key: is directly passed to the wakeup function
3858 * It may be assumed that this function implies a write memory barrier before
3859 * changing the task state if and only if any tasks are woken up.
3861 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3862 int nr_exclusive, void *key)
3864 unsigned long flags;
3866 spin_lock_irqsave(&q->lock, flags);
3867 __wake_up_common(q, mode, nr_exclusive, 0, key);
3868 spin_unlock_irqrestore(&q->lock, flags);
3870 EXPORT_SYMBOL(__wake_up);
3873 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3875 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3877 __wake_up_common(q, mode, 1, 0, NULL);
3880 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3882 __wake_up_common(q, mode, 1, 0, key);
3886 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3888 * @mode: which threads
3889 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3890 * @key: opaque value to be passed to wakeup targets
3892 * The sync wakeup differs that the waker knows that it will schedule
3893 * away soon, so while the target thread will be woken up, it will not
3894 * be migrated to another CPU - ie. the two threads are 'synchronized'
3895 * with each other. This can prevent needless bouncing between CPUs.
3897 * On UP it can prevent extra preemption.
3899 * It may be assumed that this function implies a write memory barrier before
3900 * changing the task state if and only if any tasks are woken up.
3902 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3903 int nr_exclusive, void *key)
3905 unsigned long flags;
3906 int wake_flags = WF_SYNC;
3911 if (unlikely(!nr_exclusive))
3914 spin_lock_irqsave(&q->lock, flags);
3915 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3916 spin_unlock_irqrestore(&q->lock, flags);
3918 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3921 * __wake_up_sync - see __wake_up_sync_key()
3923 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3925 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3927 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3930 * complete: - signals a single thread waiting on this completion
3931 * @x: holds the state of this particular completion
3933 * This will wake up a single thread waiting on this completion. Threads will be
3934 * awakened in the same order in which they were queued.
3936 * See also complete_all(), wait_for_completion() and related routines.
3938 * It may be assumed that this function implies a write memory barrier before
3939 * changing the task state if and only if any tasks are woken up.
3941 void complete(struct completion *x)
3943 unsigned long flags;
3945 spin_lock_irqsave(&x->wait.lock, flags);
3947 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3948 spin_unlock_irqrestore(&x->wait.lock, flags);
3950 EXPORT_SYMBOL(complete);
3953 * complete_all: - signals all threads waiting on this completion
3954 * @x: holds the state of this particular completion
3956 * This will wake up all threads waiting on this particular completion event.
3958 * It may be assumed that this function implies a write memory barrier before
3959 * changing the task state if and only if any tasks are woken up.
3961 void complete_all(struct completion *x)
3963 unsigned long flags;
3965 spin_lock_irqsave(&x->wait.lock, flags);
3966 x->done += UINT_MAX/2;
3967 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3968 spin_unlock_irqrestore(&x->wait.lock, flags);
3970 EXPORT_SYMBOL(complete_all);
3972 static inline long __sched
3973 do_wait_for_common(struct completion *x, long timeout, int state)
3976 DECLARE_WAITQUEUE(wait, current);
3978 wait.flags |= WQ_FLAG_EXCLUSIVE;
3979 __add_wait_queue_tail(&x->wait, &wait);
3981 if (signal_pending_state(state, current)) {
3982 timeout = -ERESTARTSYS;
3985 __set_current_state(state);
3986 spin_unlock_irq(&x->wait.lock);
3987 timeout = schedule_timeout(timeout);
3988 spin_lock_irq(&x->wait.lock);
3989 } while (!x->done && timeout);
3990 __remove_wait_queue(&x->wait, &wait);
3995 return timeout ?: 1;
3999 wait_for_common(struct completion *x, long timeout, int state)
4003 spin_lock_irq(&x->wait.lock);
4004 timeout = do_wait_for_common(x, timeout, state);
4005 spin_unlock_irq(&x->wait.lock);
4010 * wait_for_completion: - waits for completion of a task
4011 * @x: holds the state of this particular completion
4013 * This waits to be signaled for completion of a specific task. It is NOT
4014 * interruptible and there is no timeout.
4016 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4017 * and interrupt capability. Also see complete().
4019 void __sched wait_for_completion(struct completion *x)
4021 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4023 EXPORT_SYMBOL(wait_for_completion);
4026 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4027 * @x: holds the state of this particular completion
4028 * @timeout: timeout value in jiffies
4030 * This waits for either a completion of a specific task to be signaled or for a
4031 * specified timeout to expire. The timeout is in jiffies. It is not
4034 unsigned long __sched
4035 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4037 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4039 EXPORT_SYMBOL(wait_for_completion_timeout);
4042 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4043 * @x: holds the state of this particular completion
4045 * This waits for completion of a specific task to be signaled. It is
4048 int __sched wait_for_completion_interruptible(struct completion *x)
4050 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4051 if (t == -ERESTARTSYS)
4055 EXPORT_SYMBOL(wait_for_completion_interruptible);
4058 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4059 * @x: holds the state of this particular completion
4060 * @timeout: timeout value in jiffies
4062 * This waits for either a completion of a specific task to be signaled or for a
4063 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4065 unsigned long __sched
4066 wait_for_completion_interruptible_timeout(struct completion *x,
4067 unsigned long timeout)
4069 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4071 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4074 * wait_for_completion_killable: - waits for completion of a task (killable)
4075 * @x: holds the state of this particular completion
4077 * This waits to be signaled for completion of a specific task. It can be
4078 * interrupted by a kill signal.
4080 int __sched wait_for_completion_killable(struct completion *x)
4082 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4083 if (t == -ERESTARTSYS)
4087 EXPORT_SYMBOL(wait_for_completion_killable);
4090 * try_wait_for_completion - try to decrement a completion without blocking
4091 * @x: completion structure
4093 * Returns: 0 if a decrement cannot be done without blocking
4094 * 1 if a decrement succeeded.
4096 * If a completion is being used as a counting completion,
4097 * attempt to decrement the counter without blocking. This
4098 * enables us to avoid waiting if the resource the completion
4099 * is protecting is not available.
4101 bool try_wait_for_completion(struct completion *x)
4103 unsigned long flags;
4106 spin_lock_irqsave(&x->wait.lock, flags);
4111 spin_unlock_irqrestore(&x->wait.lock, flags);
4114 EXPORT_SYMBOL(try_wait_for_completion);
4117 * completion_done - Test to see if a completion has any waiters
4118 * @x: completion structure
4120 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4121 * 1 if there are no waiters.
4124 bool completion_done(struct completion *x)
4126 unsigned long flags;
4129 spin_lock_irqsave(&x->wait.lock, flags);
4132 spin_unlock_irqrestore(&x->wait.lock, flags);
4135 EXPORT_SYMBOL(completion_done);
4138 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4140 unsigned long flags;
4143 init_waitqueue_entry(&wait, current);
4145 __set_current_state(state);
4147 spin_lock_irqsave(&q->lock, flags);
4148 __add_wait_queue(q, &wait);
4149 spin_unlock(&q->lock);
4150 timeout = schedule_timeout(timeout);
4151 spin_lock_irq(&q->lock);
4152 __remove_wait_queue(q, &wait);
4153 spin_unlock_irqrestore(&q->lock, flags);
4158 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4160 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4162 EXPORT_SYMBOL(interruptible_sleep_on);
4165 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4167 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4169 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4171 void __sched sleep_on(wait_queue_head_t *q)
4173 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4175 EXPORT_SYMBOL(sleep_on);
4177 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4179 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4181 EXPORT_SYMBOL(sleep_on_timeout);
4183 #ifdef CONFIG_RT_MUTEXES
4186 * rt_mutex_setprio - set the current priority of a task
4188 * @prio: prio value (kernel-internal form)
4190 * This function changes the 'effective' priority of a task. It does
4191 * not touch ->normal_prio like __setscheduler().
4193 * Used by the rt_mutex code to implement priority inheritance logic.
4195 void rt_mutex_setprio(struct task_struct *p, int prio)
4197 unsigned long flags;
4198 int oldprio, on_rq, running;
4200 const struct sched_class *prev_class = p->sched_class;
4202 BUG_ON(prio < 0 || prio > MAX_PRIO);
4204 rq = task_rq_lock(p, &flags);
4205 update_rq_clock(rq);
4208 on_rq = p->se.on_rq;
4209 running = task_current(rq, p);
4211 dequeue_task(rq, p, 0);
4213 p->sched_class->put_prev_task(rq, p);
4216 p->sched_class = &rt_sched_class;
4218 p->sched_class = &fair_sched_class;
4223 p->sched_class->set_curr_task(rq);
4225 enqueue_task(rq, p, 0, oldprio < prio);
4227 check_class_changed(rq, p, prev_class, oldprio, running);
4229 task_rq_unlock(rq, &flags);
4234 void set_user_nice(struct task_struct *p, long nice)
4236 int old_prio, delta, on_rq;
4237 unsigned long flags;
4240 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4243 * We have to be careful, if called from sys_setpriority(),
4244 * the task might be in the middle of scheduling on another CPU.
4246 rq = task_rq_lock(p, &flags);
4247 update_rq_clock(rq);
4249 * The RT priorities are set via sched_setscheduler(), but we still
4250 * allow the 'normal' nice value to be set - but as expected
4251 * it wont have any effect on scheduling until the task is
4252 * SCHED_FIFO/SCHED_RR:
4254 if (task_has_rt_policy(p)) {
4255 p->static_prio = NICE_TO_PRIO(nice);
4258 on_rq = p->se.on_rq;
4260 dequeue_task(rq, p, 0);
4262 p->static_prio = NICE_TO_PRIO(nice);
4265 p->prio = effective_prio(p);
4266 delta = p->prio - old_prio;
4269 enqueue_task(rq, p, 0, false);
4271 * If the task increased its priority or is running and
4272 * lowered its priority, then reschedule its CPU:
4274 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4275 resched_task(rq->curr);
4278 task_rq_unlock(rq, &flags);
4280 EXPORT_SYMBOL(set_user_nice);
4283 * can_nice - check if a task can reduce its nice value
4287 int can_nice(const struct task_struct *p, const int nice)
4289 /* convert nice value [19,-20] to rlimit style value [1,40] */
4290 int nice_rlim = 20 - nice;
4292 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4293 capable(CAP_SYS_NICE));
4296 #ifdef __ARCH_WANT_SYS_NICE
4299 * sys_nice - change the priority of the current process.
4300 * @increment: priority increment
4302 * sys_setpriority is a more generic, but much slower function that
4303 * does similar things.
4305 SYSCALL_DEFINE1(nice, int, increment)
4310 * Setpriority might change our priority at the same moment.
4311 * We don't have to worry. Conceptually one call occurs first
4312 * and we have a single winner.
4314 if (increment < -40)
4319 nice = TASK_NICE(current) + increment;
4325 if (increment < 0 && !can_nice(current, nice))
4328 retval = security_task_setnice(current, nice);
4332 set_user_nice(current, nice);
4339 * task_prio - return the priority value of a given task.
4340 * @p: the task in question.
4342 * This is the priority value as seen by users in /proc.
4343 * RT tasks are offset by -200. Normal tasks are centered
4344 * around 0, value goes from -16 to +15.
4346 int task_prio(const struct task_struct *p)
4348 return p->prio - MAX_RT_PRIO;
4352 * task_nice - return the nice value of a given task.
4353 * @p: the task in question.
4355 int task_nice(const struct task_struct *p)
4357 return TASK_NICE(p);
4359 EXPORT_SYMBOL(task_nice);
4362 * idle_cpu - is a given cpu idle currently?
4363 * @cpu: the processor in question.
4365 int idle_cpu(int cpu)
4367 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4371 * idle_task - return the idle task for a given cpu.
4372 * @cpu: the processor in question.
4374 struct task_struct *idle_task(int cpu)
4376 return cpu_rq(cpu)->idle;
4380 * find_process_by_pid - find a process with a matching PID value.
4381 * @pid: the pid in question.
4383 static struct task_struct *find_process_by_pid(pid_t pid)
4385 return pid ? find_task_by_vpid(pid) : current;
4388 /* Actually do priority change: must hold rq lock. */
4390 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4392 BUG_ON(p->se.on_rq);
4395 p->rt_priority = prio;
4396 p->normal_prio = normal_prio(p);
4397 /* we are holding p->pi_lock already */
4398 p->prio = rt_mutex_getprio(p);
4399 if (rt_prio(p->prio))
4400 p->sched_class = &rt_sched_class;
4402 p->sched_class = &fair_sched_class;
4407 * check the target process has a UID that matches the current process's
4409 static bool check_same_owner(struct task_struct *p)
4411 const struct cred *cred = current_cred(), *pcred;
4415 pcred = __task_cred(p);
4416 match = (cred->euid == pcred->euid ||
4417 cred->euid == pcred->uid);
4422 static int __sched_setscheduler(struct task_struct *p, int policy,
4423 struct sched_param *param, bool user)
4425 int retval, oldprio, oldpolicy = -1, on_rq, running;
4426 unsigned long flags;
4427 const struct sched_class *prev_class = p->sched_class;
4431 /* may grab non-irq protected spin_locks */
4432 BUG_ON(in_interrupt());
4434 /* double check policy once rq lock held */
4436 reset_on_fork = p->sched_reset_on_fork;
4437 policy = oldpolicy = p->policy;
4439 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4440 policy &= ~SCHED_RESET_ON_FORK;
4442 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4443 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4444 policy != SCHED_IDLE)
4449 * Valid priorities for SCHED_FIFO and SCHED_RR are
4450 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4451 * SCHED_BATCH and SCHED_IDLE is 0.
4453 if (param->sched_priority < 0 ||
4454 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4455 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4457 if (rt_policy(policy) != (param->sched_priority != 0))
4461 * Allow unprivileged RT tasks to decrease priority:
4463 if (user && !capable(CAP_SYS_NICE)) {
4464 if (rt_policy(policy)) {
4465 unsigned long rlim_rtprio;
4467 if (!lock_task_sighand(p, &flags))
4469 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4470 unlock_task_sighand(p, &flags);
4472 /* can't set/change the rt policy */
4473 if (policy != p->policy && !rlim_rtprio)
4476 /* can't increase priority */
4477 if (param->sched_priority > p->rt_priority &&
4478 param->sched_priority > rlim_rtprio)
4482 * Like positive nice levels, dont allow tasks to
4483 * move out of SCHED_IDLE either:
4485 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4488 /* can't change other user's priorities */
4489 if (!check_same_owner(p))
4492 /* Normal users shall not reset the sched_reset_on_fork flag */
4493 if (p->sched_reset_on_fork && !reset_on_fork)
4498 #ifdef CONFIG_RT_GROUP_SCHED
4500 * Do not allow realtime tasks into groups that have no runtime
4503 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4504 task_group(p)->rt_bandwidth.rt_runtime == 0)
4508 retval = security_task_setscheduler(p, policy, param);
4514 * make sure no PI-waiters arrive (or leave) while we are
4515 * changing the priority of the task:
4517 raw_spin_lock_irqsave(&p->pi_lock, flags);
4519 * To be able to change p->policy safely, the apropriate
4520 * runqueue lock must be held.
4522 rq = __task_rq_lock(p);
4523 /* recheck policy now with rq lock held */
4524 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4525 policy = oldpolicy = -1;
4526 __task_rq_unlock(rq);
4527 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4530 update_rq_clock(rq);
4531 on_rq = p->se.on_rq;
4532 running = task_current(rq, p);
4534 deactivate_task(rq, p, 0);
4536 p->sched_class->put_prev_task(rq, p);
4538 p->sched_reset_on_fork = reset_on_fork;
4541 __setscheduler(rq, p, policy, param->sched_priority);
4544 p->sched_class->set_curr_task(rq);
4546 activate_task(rq, p, 0);
4548 check_class_changed(rq, p, prev_class, oldprio, running);
4550 __task_rq_unlock(rq);
4551 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4553 rt_mutex_adjust_pi(p);
4559 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4560 * @p: the task in question.
4561 * @policy: new policy.
4562 * @param: structure containing the new RT priority.
4564 * NOTE that the task may be already dead.
4566 int sched_setscheduler(struct task_struct *p, int policy,
4567 struct sched_param *param)
4569 return __sched_setscheduler(p, policy, param, true);
4571 EXPORT_SYMBOL_GPL(sched_setscheduler);
4574 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4575 * @p: the task in question.
4576 * @policy: new policy.
4577 * @param: structure containing the new RT priority.
4579 * Just like sched_setscheduler, only don't bother checking if the
4580 * current context has permission. For example, this is needed in
4581 * stop_machine(): we create temporary high priority worker threads,
4582 * but our caller might not have that capability.
4584 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4585 struct sched_param *param)
4587 return __sched_setscheduler(p, policy, param, false);
4591 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4593 struct sched_param lparam;
4594 struct task_struct *p;
4597 if (!param || pid < 0)
4599 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4604 p = find_process_by_pid(pid);
4606 retval = sched_setscheduler(p, policy, &lparam);
4613 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4614 * @pid: the pid in question.
4615 * @policy: new policy.
4616 * @param: structure containing the new RT priority.
4618 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4619 struct sched_param __user *, param)
4621 /* negative values for policy are not valid */
4625 return do_sched_setscheduler(pid, policy, param);
4629 * sys_sched_setparam - set/change the RT priority of a thread
4630 * @pid: the pid in question.
4631 * @param: structure containing the new RT priority.
4633 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4635 return do_sched_setscheduler(pid, -1, param);
4639 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4640 * @pid: the pid in question.
4642 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4644 struct task_struct *p;
4652 p = find_process_by_pid(pid);
4654 retval = security_task_getscheduler(p);
4657 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4664 * sys_sched_getparam - get the RT priority of a thread
4665 * @pid: the pid in question.
4666 * @param: structure containing the RT priority.
4668 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4670 struct sched_param lp;
4671 struct task_struct *p;
4674 if (!param || pid < 0)
4678 p = find_process_by_pid(pid);
4683 retval = security_task_getscheduler(p);
4687 lp.sched_priority = p->rt_priority;
4691 * This one might sleep, we cannot do it with a spinlock held ...
4693 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4702 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4704 cpumask_var_t cpus_allowed, new_mask;
4705 struct task_struct *p;
4711 p = find_process_by_pid(pid);
4718 /* Prevent p going away */
4722 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4726 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4728 goto out_free_cpus_allowed;
4731 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4734 retval = security_task_setscheduler(p, 0, NULL);
4738 cpuset_cpus_allowed(p, cpus_allowed);
4739 cpumask_and(new_mask, in_mask, cpus_allowed);
4741 retval = set_cpus_allowed_ptr(p, new_mask);
4744 cpuset_cpus_allowed(p, cpus_allowed);
4745 if (!cpumask_subset(new_mask, cpus_allowed)) {
4747 * We must have raced with a concurrent cpuset
4748 * update. Just reset the cpus_allowed to the
4749 * cpuset's cpus_allowed
4751 cpumask_copy(new_mask, cpus_allowed);
4756 free_cpumask_var(new_mask);
4757 out_free_cpus_allowed:
4758 free_cpumask_var(cpus_allowed);
4765 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4766 struct cpumask *new_mask)
4768 if (len < cpumask_size())
4769 cpumask_clear(new_mask);
4770 else if (len > cpumask_size())
4771 len = cpumask_size();
4773 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4777 * sys_sched_setaffinity - set the cpu affinity of a process
4778 * @pid: pid of the process
4779 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4780 * @user_mask_ptr: user-space pointer to the new cpu mask
4782 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4783 unsigned long __user *, user_mask_ptr)
4785 cpumask_var_t new_mask;
4788 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4791 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4793 retval = sched_setaffinity(pid, new_mask);
4794 free_cpumask_var(new_mask);
4798 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4800 struct task_struct *p;
4801 unsigned long flags;
4809 p = find_process_by_pid(pid);
4813 retval = security_task_getscheduler(p);
4817 rq = task_rq_lock(p, &flags);
4818 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4819 task_rq_unlock(rq, &flags);
4829 * sys_sched_getaffinity - get the cpu affinity of a process
4830 * @pid: pid of the process
4831 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4832 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4834 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4835 unsigned long __user *, user_mask_ptr)
4840 if (len < cpumask_size())
4843 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4846 ret = sched_getaffinity(pid, mask);
4848 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
4851 ret = cpumask_size();
4853 free_cpumask_var(mask);
4859 * sys_sched_yield - yield the current processor to other threads.
4861 * This function yields the current CPU to other tasks. If there are no
4862 * other threads running on this CPU then this function will return.
4864 SYSCALL_DEFINE0(sched_yield)
4866 struct rq *rq = this_rq_lock();
4868 schedstat_inc(rq, yld_count);
4869 current->sched_class->yield_task(rq);
4872 * Since we are going to call schedule() anyway, there's
4873 * no need to preempt or enable interrupts:
4875 __release(rq->lock);
4876 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4877 do_raw_spin_unlock(&rq->lock);
4878 preempt_enable_no_resched();
4885 static inline int should_resched(void)
4887 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4890 static void __cond_resched(void)
4892 add_preempt_count(PREEMPT_ACTIVE);
4894 sub_preempt_count(PREEMPT_ACTIVE);
4897 int __sched _cond_resched(void)
4899 if (should_resched()) {
4905 EXPORT_SYMBOL(_cond_resched);
4908 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4909 * call schedule, and on return reacquire the lock.
4911 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4912 * operations here to prevent schedule() from being called twice (once via
4913 * spin_unlock(), once by hand).
4915 int __cond_resched_lock(spinlock_t *lock)
4917 int resched = should_resched();
4920 lockdep_assert_held(lock);
4922 if (spin_needbreak(lock) || resched) {
4933 EXPORT_SYMBOL(__cond_resched_lock);
4935 int __sched __cond_resched_softirq(void)
4937 BUG_ON(!in_softirq());
4939 if (should_resched()) {
4947 EXPORT_SYMBOL(__cond_resched_softirq);
4950 * yield - yield the current processor to other threads.
4952 * This is a shortcut for kernel-space yielding - it marks the
4953 * thread runnable and calls sys_sched_yield().
4955 void __sched yield(void)
4957 set_current_state(TASK_RUNNING);
4960 EXPORT_SYMBOL(yield);
4963 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4964 * that process accounting knows that this is a task in IO wait state.
4966 void __sched io_schedule(void)
4968 struct rq *rq = raw_rq();
4970 delayacct_blkio_start();
4971 atomic_inc(&rq->nr_iowait);
4972 current->in_iowait = 1;
4974 current->in_iowait = 0;
4975 atomic_dec(&rq->nr_iowait);
4976 delayacct_blkio_end();
4978 EXPORT_SYMBOL(io_schedule);
4980 long __sched io_schedule_timeout(long timeout)
4982 struct rq *rq = raw_rq();
4985 delayacct_blkio_start();
4986 atomic_inc(&rq->nr_iowait);
4987 current->in_iowait = 1;
4988 ret = schedule_timeout(timeout);
4989 current->in_iowait = 0;
4990 atomic_dec(&rq->nr_iowait);
4991 delayacct_blkio_end();
4996 * sys_sched_get_priority_max - return maximum RT priority.
4997 * @policy: scheduling class.
4999 * this syscall returns the maximum rt_priority that can be used
5000 * by a given scheduling class.
5002 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5009 ret = MAX_USER_RT_PRIO-1;
5021 * sys_sched_get_priority_min - return minimum RT priority.
5022 * @policy: scheduling class.
5024 * this syscall returns the minimum rt_priority that can be used
5025 * by a given scheduling class.
5027 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5045 * sys_sched_rr_get_interval - return the default timeslice of a process.
5046 * @pid: pid of the process.
5047 * @interval: userspace pointer to the timeslice value.
5049 * this syscall writes the default timeslice value of a given process
5050 * into the user-space timespec buffer. A value of '0' means infinity.
5052 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5053 struct timespec __user *, interval)
5055 struct task_struct *p;
5056 unsigned int time_slice;
5057 unsigned long flags;
5067 p = find_process_by_pid(pid);
5071 retval = security_task_getscheduler(p);
5075 rq = task_rq_lock(p, &flags);
5076 time_slice = p->sched_class->get_rr_interval(rq, p);
5077 task_rq_unlock(rq, &flags);
5080 jiffies_to_timespec(time_slice, &t);
5081 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5089 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5091 void sched_show_task(struct task_struct *p)
5093 unsigned long free = 0;
5096 state = p->state ? __ffs(p->state) + 1 : 0;
5097 printk(KERN_INFO "%-13.13s %c", p->comm,
5098 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5099 #if BITS_PER_LONG == 32
5100 if (state == TASK_RUNNING)
5101 printk(KERN_CONT " running ");
5103 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5105 if (state == TASK_RUNNING)
5106 printk(KERN_CONT " running task ");
5108 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5110 #ifdef CONFIG_DEBUG_STACK_USAGE
5111 free = stack_not_used(p);
5113 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5114 task_pid_nr(p), task_pid_nr(p->real_parent),
5115 (unsigned long)task_thread_info(p)->flags);
5117 show_stack(p, NULL);
5120 void show_state_filter(unsigned long state_filter)
5122 struct task_struct *g, *p;
5124 #if BITS_PER_LONG == 32
5126 " task PC stack pid father\n");
5129 " task PC stack pid father\n");
5131 read_lock(&tasklist_lock);
5132 do_each_thread(g, p) {
5134 * reset the NMI-timeout, listing all files on a slow
5135 * console might take alot of time:
5137 touch_nmi_watchdog();
5138 if (!state_filter || (p->state & state_filter))
5140 } while_each_thread(g, p);
5142 touch_all_softlockup_watchdogs();
5144 #ifdef CONFIG_SCHED_DEBUG
5145 sysrq_sched_debug_show();
5147 read_unlock(&tasklist_lock);
5149 * Only show locks if all tasks are dumped:
5152 debug_show_all_locks();
5155 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5157 idle->sched_class = &idle_sched_class;
5161 * init_idle - set up an idle thread for a given CPU
5162 * @idle: task in question
5163 * @cpu: cpu the idle task belongs to
5165 * NOTE: this function does not set the idle thread's NEED_RESCHED
5166 * flag, to make booting more robust.
5168 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5170 struct rq *rq = cpu_rq(cpu);
5171 unsigned long flags;
5173 raw_spin_lock_irqsave(&rq->lock, flags);
5176 idle->state = TASK_RUNNING;
5177 idle->se.exec_start = sched_clock();
5179 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5180 __set_task_cpu(idle, cpu);
5182 rq->curr = rq->idle = idle;
5183 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5186 raw_spin_unlock_irqrestore(&rq->lock, flags);
5188 /* Set the preempt count _outside_ the spinlocks! */
5189 #if defined(CONFIG_PREEMPT)
5190 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5192 task_thread_info(idle)->preempt_count = 0;
5195 * The idle tasks have their own, simple scheduling class:
5197 idle->sched_class = &idle_sched_class;
5198 ftrace_graph_init_task(idle);
5202 * In a system that switches off the HZ timer nohz_cpu_mask
5203 * indicates which cpus entered this state. This is used
5204 * in the rcu update to wait only for active cpus. For system
5205 * which do not switch off the HZ timer nohz_cpu_mask should
5206 * always be CPU_BITS_NONE.
5208 cpumask_var_t nohz_cpu_mask;
5211 * Increase the granularity value when there are more CPUs,
5212 * because with more CPUs the 'effective latency' as visible
5213 * to users decreases. But the relationship is not linear,
5214 * so pick a second-best guess by going with the log2 of the
5217 * This idea comes from the SD scheduler of Con Kolivas:
5219 static int get_update_sysctl_factor(void)
5221 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5222 unsigned int factor;
5224 switch (sysctl_sched_tunable_scaling) {
5225 case SCHED_TUNABLESCALING_NONE:
5228 case SCHED_TUNABLESCALING_LINEAR:
5231 case SCHED_TUNABLESCALING_LOG:
5233 factor = 1 + ilog2(cpus);
5240 static void update_sysctl(void)
5242 unsigned int factor = get_update_sysctl_factor();
5244 #define SET_SYSCTL(name) \
5245 (sysctl_##name = (factor) * normalized_sysctl_##name)
5246 SET_SYSCTL(sched_min_granularity);
5247 SET_SYSCTL(sched_latency);
5248 SET_SYSCTL(sched_wakeup_granularity);
5249 SET_SYSCTL(sched_shares_ratelimit);
5253 static inline void sched_init_granularity(void)
5260 * This is how migration works:
5262 * 1) we queue a struct migration_req structure in the source CPU's
5263 * runqueue and wake up that CPU's migration thread.
5264 * 2) we down() the locked semaphore => thread blocks.
5265 * 3) migration thread wakes up (implicitly it forces the migrated
5266 * thread off the CPU)
5267 * 4) it gets the migration request and checks whether the migrated
5268 * task is still in the wrong runqueue.
5269 * 5) if it's in the wrong runqueue then the migration thread removes
5270 * it and puts it into the right queue.
5271 * 6) migration thread up()s the semaphore.
5272 * 7) we wake up and the migration is done.
5276 * Change a given task's CPU affinity. Migrate the thread to a
5277 * proper CPU and schedule it away if the CPU it's executing on
5278 * is removed from the allowed bitmask.
5280 * NOTE: the caller must have a valid reference to the task, the
5281 * task must not exit() & deallocate itself prematurely. The
5282 * call is not atomic; no spinlocks may be held.
5284 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5286 struct migration_req req;
5287 unsigned long flags;
5292 * Since we rely on wake-ups to migrate sleeping tasks, don't change
5293 * the ->cpus_allowed mask from under waking tasks, which would be
5294 * possible when we change rq->lock in ttwu(), so synchronize against
5295 * TASK_WAKING to avoid that.
5298 while (p->state == TASK_WAKING)
5301 rq = task_rq_lock(p, &flags);
5303 if (p->state == TASK_WAKING) {
5304 task_rq_unlock(rq, &flags);
5308 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5313 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5314 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5319 if (p->sched_class->set_cpus_allowed)
5320 p->sched_class->set_cpus_allowed(p, new_mask);
5322 cpumask_copy(&p->cpus_allowed, new_mask);
5323 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5326 /* Can the task run on the task's current CPU? If so, we're done */
5327 if (cpumask_test_cpu(task_cpu(p), new_mask))
5330 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5331 /* Need help from migration thread: drop lock and wait. */
5332 struct task_struct *mt = rq->migration_thread;
5334 get_task_struct(mt);
5335 task_rq_unlock(rq, &flags);
5336 wake_up_process(rq->migration_thread);
5337 put_task_struct(mt);
5338 wait_for_completion(&req.done);
5339 tlb_migrate_finish(p->mm);
5343 task_rq_unlock(rq, &flags);
5347 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5350 * Move (not current) task off this cpu, onto dest cpu. We're doing
5351 * this because either it can't run here any more (set_cpus_allowed()
5352 * away from this CPU, or CPU going down), or because we're
5353 * attempting to rebalance this task on exec (sched_exec).
5355 * So we race with normal scheduler movements, but that's OK, as long
5356 * as the task is no longer on this CPU.
5358 * Returns non-zero if task was successfully migrated.
5360 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5362 struct rq *rq_dest, *rq_src;
5365 if (unlikely(!cpu_active(dest_cpu)))
5368 rq_src = cpu_rq(src_cpu);
5369 rq_dest = cpu_rq(dest_cpu);
5371 double_rq_lock(rq_src, rq_dest);
5372 /* Already moved. */
5373 if (task_cpu(p) != src_cpu)
5375 /* Affinity changed (again). */
5376 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5380 * If we're not on a rq, the next wake-up will ensure we're
5384 deactivate_task(rq_src, p, 0);
5385 set_task_cpu(p, dest_cpu);
5386 activate_task(rq_dest, p, 0);
5387 check_preempt_curr(rq_dest, p, 0);
5392 double_rq_unlock(rq_src, rq_dest);
5396 #define RCU_MIGRATION_IDLE 0
5397 #define RCU_MIGRATION_NEED_QS 1
5398 #define RCU_MIGRATION_GOT_QS 2
5399 #define RCU_MIGRATION_MUST_SYNC 3
5402 * migration_thread - this is a highprio system thread that performs
5403 * thread migration by bumping thread off CPU then 'pushing' onto
5406 static int migration_thread(void *data)
5409 int cpu = (long)data;
5413 BUG_ON(rq->migration_thread != current);
5415 set_current_state(TASK_INTERRUPTIBLE);
5416 while (!kthread_should_stop()) {
5417 struct migration_req *req;
5418 struct list_head *head;
5420 raw_spin_lock_irq(&rq->lock);
5422 if (cpu_is_offline(cpu)) {
5423 raw_spin_unlock_irq(&rq->lock);
5427 if (rq->active_balance) {
5428 active_load_balance(rq, cpu);
5429 rq->active_balance = 0;
5432 head = &rq->migration_queue;
5434 if (list_empty(head)) {
5435 raw_spin_unlock_irq(&rq->lock);
5437 set_current_state(TASK_INTERRUPTIBLE);
5440 req = list_entry(head->next, struct migration_req, list);
5441 list_del_init(head->next);
5443 if (req->task != NULL) {
5444 raw_spin_unlock(&rq->lock);
5445 __migrate_task(req->task, cpu, req->dest_cpu);
5446 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5447 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5448 raw_spin_unlock(&rq->lock);
5450 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5451 raw_spin_unlock(&rq->lock);
5452 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5456 complete(&req->done);
5458 __set_current_state(TASK_RUNNING);
5463 #ifdef CONFIG_HOTPLUG_CPU
5465 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5469 local_irq_disable();
5470 ret = __migrate_task(p, src_cpu, dest_cpu);
5476 * Figure out where task on dead CPU should go, use force if necessary.
5478 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5483 dest_cpu = select_fallback_rq(dead_cpu, p);
5485 /* It can have affinity changed while we were choosing. */
5486 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
5491 * While a dead CPU has no uninterruptible tasks queued at this point,
5492 * it might still have a nonzero ->nr_uninterruptible counter, because
5493 * for performance reasons the counter is not stricly tracking tasks to
5494 * their home CPUs. So we just add the counter to another CPU's counter,
5495 * to keep the global sum constant after CPU-down:
5497 static void migrate_nr_uninterruptible(struct rq *rq_src)
5499 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5500 unsigned long flags;
5502 local_irq_save(flags);
5503 double_rq_lock(rq_src, rq_dest);
5504 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5505 rq_src->nr_uninterruptible = 0;
5506 double_rq_unlock(rq_src, rq_dest);
5507 local_irq_restore(flags);
5510 /* Run through task list and migrate tasks from the dead cpu. */
5511 static void migrate_live_tasks(int src_cpu)
5513 struct task_struct *p, *t;
5515 read_lock(&tasklist_lock);
5517 do_each_thread(t, p) {
5521 if (task_cpu(p) == src_cpu)
5522 move_task_off_dead_cpu(src_cpu, p);
5523 } while_each_thread(t, p);
5525 read_unlock(&tasklist_lock);
5529 * Schedules idle task to be the next runnable task on current CPU.
5530 * It does so by boosting its priority to highest possible.
5531 * Used by CPU offline code.
5533 void sched_idle_next(void)
5535 int this_cpu = smp_processor_id();
5536 struct rq *rq = cpu_rq(this_cpu);
5537 struct task_struct *p = rq->idle;
5538 unsigned long flags;
5540 /* cpu has to be offline */
5541 BUG_ON(cpu_online(this_cpu));
5544 * Strictly not necessary since rest of the CPUs are stopped by now
5545 * and interrupts disabled on the current cpu.
5547 raw_spin_lock_irqsave(&rq->lock, flags);
5549 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5551 update_rq_clock(rq);
5552 activate_task(rq, p, 0);
5554 raw_spin_unlock_irqrestore(&rq->lock, flags);
5558 * Ensures that the idle task is using init_mm right before its cpu goes
5561 void idle_task_exit(void)
5563 struct mm_struct *mm = current->active_mm;
5565 BUG_ON(cpu_online(smp_processor_id()));
5568 switch_mm(mm, &init_mm, current);
5572 /* called under rq->lock with disabled interrupts */
5573 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5575 struct rq *rq = cpu_rq(dead_cpu);
5577 /* Must be exiting, otherwise would be on tasklist. */
5578 BUG_ON(!p->exit_state);
5580 /* Cannot have done final schedule yet: would have vanished. */
5581 BUG_ON(p->state == TASK_DEAD);
5586 * Drop lock around migration; if someone else moves it,
5587 * that's OK. No task can be added to this CPU, so iteration is
5590 raw_spin_unlock_irq(&rq->lock);
5591 move_task_off_dead_cpu(dead_cpu, p);
5592 raw_spin_lock_irq(&rq->lock);
5597 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5598 static void migrate_dead_tasks(unsigned int dead_cpu)
5600 struct rq *rq = cpu_rq(dead_cpu);
5601 struct task_struct *next;
5604 if (!rq->nr_running)
5606 update_rq_clock(rq);
5607 next = pick_next_task(rq);
5610 next->sched_class->put_prev_task(rq, next);
5611 migrate_dead(dead_cpu, next);
5617 * remove the tasks which were accounted by rq from calc_load_tasks.
5619 static void calc_global_load_remove(struct rq *rq)
5621 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5622 rq->calc_load_active = 0;
5624 #endif /* CONFIG_HOTPLUG_CPU */
5626 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5628 static struct ctl_table sd_ctl_dir[] = {
5630 .procname = "sched_domain",
5636 static struct ctl_table sd_ctl_root[] = {
5638 .procname = "kernel",
5640 .child = sd_ctl_dir,
5645 static struct ctl_table *sd_alloc_ctl_entry(int n)
5647 struct ctl_table *entry =
5648 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5653 static void sd_free_ctl_entry(struct ctl_table **tablep)
5655 struct ctl_table *entry;
5658 * In the intermediate directories, both the child directory and
5659 * procname are dynamically allocated and could fail but the mode
5660 * will always be set. In the lowest directory the names are
5661 * static strings and all have proc handlers.
5663 for (entry = *tablep; entry->mode; entry++) {
5665 sd_free_ctl_entry(&entry->child);
5666 if (entry->proc_handler == NULL)
5667 kfree(entry->procname);
5675 set_table_entry(struct ctl_table *entry,
5676 const char *procname, void *data, int maxlen,
5677 mode_t mode, proc_handler *proc_handler)
5679 entry->procname = procname;
5681 entry->maxlen = maxlen;
5683 entry->proc_handler = proc_handler;
5686 static struct ctl_table *
5687 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5689 struct ctl_table *table = sd_alloc_ctl_entry(13);
5694 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5695 sizeof(long), 0644, proc_doulongvec_minmax);
5696 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5697 sizeof(long), 0644, proc_doulongvec_minmax);
5698 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5699 sizeof(int), 0644, proc_dointvec_minmax);
5700 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5701 sizeof(int), 0644, proc_dointvec_minmax);
5702 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5703 sizeof(int), 0644, proc_dointvec_minmax);
5704 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5705 sizeof(int), 0644, proc_dointvec_minmax);
5706 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5707 sizeof(int), 0644, proc_dointvec_minmax);
5708 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5709 sizeof(int), 0644, proc_dointvec_minmax);
5710 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5711 sizeof(int), 0644, proc_dointvec_minmax);
5712 set_table_entry(&table[9], "cache_nice_tries",
5713 &sd->cache_nice_tries,
5714 sizeof(int), 0644, proc_dointvec_minmax);
5715 set_table_entry(&table[10], "flags", &sd->flags,
5716 sizeof(int), 0644, proc_dointvec_minmax);
5717 set_table_entry(&table[11], "name", sd->name,
5718 CORENAME_MAX_SIZE, 0444, proc_dostring);
5719 /* &table[12] is terminator */
5724 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5726 struct ctl_table *entry, *table;
5727 struct sched_domain *sd;
5728 int domain_num = 0, i;
5731 for_each_domain(cpu, sd)
5733 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5738 for_each_domain(cpu, sd) {
5739 snprintf(buf, 32, "domain%d", i);
5740 entry->procname = kstrdup(buf, GFP_KERNEL);
5742 entry->child = sd_alloc_ctl_domain_table(sd);
5749 static struct ctl_table_header *sd_sysctl_header;
5750 static void register_sched_domain_sysctl(void)
5752 int i, cpu_num = num_possible_cpus();
5753 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5756 WARN_ON(sd_ctl_dir[0].child);
5757 sd_ctl_dir[0].child = entry;
5762 for_each_possible_cpu(i) {
5763 snprintf(buf, 32, "cpu%d", i);
5764 entry->procname = kstrdup(buf, GFP_KERNEL);
5766 entry->child = sd_alloc_ctl_cpu_table(i);
5770 WARN_ON(sd_sysctl_header);
5771 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5774 /* may be called multiple times per register */
5775 static void unregister_sched_domain_sysctl(void)
5777 if (sd_sysctl_header)
5778 unregister_sysctl_table(sd_sysctl_header);
5779 sd_sysctl_header = NULL;
5780 if (sd_ctl_dir[0].child)
5781 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5784 static void register_sched_domain_sysctl(void)
5787 static void unregister_sched_domain_sysctl(void)
5792 static void set_rq_online(struct rq *rq)
5795 const struct sched_class *class;
5797 cpumask_set_cpu(rq->cpu, rq->rd->online);
5800 for_each_class(class) {
5801 if (class->rq_online)
5802 class->rq_online(rq);
5807 static void set_rq_offline(struct rq *rq)
5810 const struct sched_class *class;
5812 for_each_class(class) {
5813 if (class->rq_offline)
5814 class->rq_offline(rq);
5817 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5823 * migration_call - callback that gets triggered when a CPU is added.
5824 * Here we can start up the necessary migration thread for the new CPU.
5826 static int __cpuinit
5827 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5829 struct task_struct *p;
5830 int cpu = (long)hcpu;
5831 unsigned long flags;
5836 case CPU_UP_PREPARE:
5837 case CPU_UP_PREPARE_FROZEN:
5838 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5841 kthread_bind(p, cpu);
5842 /* Must be high prio: stop_machine expects to yield to it. */
5843 rq = task_rq_lock(p, &flags);
5844 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5845 task_rq_unlock(rq, &flags);
5847 cpu_rq(cpu)->migration_thread = p;
5848 rq->calc_load_update = calc_load_update;
5852 case CPU_ONLINE_FROZEN:
5853 /* Strictly unnecessary, as first user will wake it. */
5854 wake_up_process(cpu_rq(cpu)->migration_thread);
5856 /* Update our root-domain */
5858 raw_spin_lock_irqsave(&rq->lock, flags);
5860 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5864 raw_spin_unlock_irqrestore(&rq->lock, flags);
5867 #ifdef CONFIG_HOTPLUG_CPU
5868 case CPU_UP_CANCELED:
5869 case CPU_UP_CANCELED_FROZEN:
5870 if (!cpu_rq(cpu)->migration_thread)
5872 /* Unbind it from offline cpu so it can run. Fall thru. */
5873 kthread_bind(cpu_rq(cpu)->migration_thread,
5874 cpumask_any(cpu_online_mask));
5875 kthread_stop(cpu_rq(cpu)->migration_thread);
5876 put_task_struct(cpu_rq(cpu)->migration_thread);
5877 cpu_rq(cpu)->migration_thread = NULL;
5881 case CPU_DEAD_FROZEN:
5882 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5883 migrate_live_tasks(cpu);
5885 kthread_stop(rq->migration_thread);
5886 put_task_struct(rq->migration_thread);
5887 rq->migration_thread = NULL;
5888 /* Idle task back to normal (off runqueue, low prio) */
5889 raw_spin_lock_irq(&rq->lock);
5890 update_rq_clock(rq);
5891 deactivate_task(rq, rq->idle, 0);
5892 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5893 rq->idle->sched_class = &idle_sched_class;
5894 migrate_dead_tasks(cpu);
5895 raw_spin_unlock_irq(&rq->lock);
5897 migrate_nr_uninterruptible(rq);
5898 BUG_ON(rq->nr_running != 0);
5899 calc_global_load_remove(rq);
5901 * No need to migrate the tasks: it was best-effort if
5902 * they didn't take sched_hotcpu_mutex. Just wake up
5905 raw_spin_lock_irq(&rq->lock);
5906 while (!list_empty(&rq->migration_queue)) {
5907 struct migration_req *req;
5909 req = list_entry(rq->migration_queue.next,
5910 struct migration_req, list);
5911 list_del_init(&req->list);
5912 raw_spin_unlock_irq(&rq->lock);
5913 complete(&req->done);
5914 raw_spin_lock_irq(&rq->lock);
5916 raw_spin_unlock_irq(&rq->lock);
5920 case CPU_DYING_FROZEN:
5921 /* Update our root-domain */
5923 raw_spin_lock_irqsave(&rq->lock, flags);
5925 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5928 raw_spin_unlock_irqrestore(&rq->lock, flags);
5936 * Register at high priority so that task migration (migrate_all_tasks)
5937 * happens before everything else. This has to be lower priority than
5938 * the notifier in the perf_event subsystem, though.
5940 static struct notifier_block __cpuinitdata migration_notifier = {
5941 .notifier_call = migration_call,
5945 static int __init migration_init(void)
5947 void *cpu = (void *)(long)smp_processor_id();
5950 /* Start one for the boot CPU: */
5951 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5952 BUG_ON(err == NOTIFY_BAD);
5953 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5954 register_cpu_notifier(&migration_notifier);
5958 early_initcall(migration_init);
5963 #ifdef CONFIG_SCHED_DEBUG
5965 static __read_mostly int sched_domain_debug_enabled;
5967 static int __init sched_domain_debug_setup(char *str)
5969 sched_domain_debug_enabled = 1;
5973 early_param("sched_debug", sched_domain_debug_setup);
5975 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5976 struct cpumask *groupmask)
5978 struct sched_group *group = sd->groups;
5981 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5982 cpumask_clear(groupmask);
5984 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5986 if (!(sd->flags & SD_LOAD_BALANCE)) {
5987 printk("does not load-balance\n");
5989 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5994 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5996 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5997 printk(KERN_ERR "ERROR: domain->span does not contain "
6000 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6001 printk(KERN_ERR "ERROR: domain->groups does not contain"
6005 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6009 printk(KERN_ERR "ERROR: group is NULL\n");
6013 if (!group->cpu_power) {
6014 printk(KERN_CONT "\n");
6015 printk(KERN_ERR "ERROR: domain->cpu_power not "
6020 if (!cpumask_weight(sched_group_cpus(group))) {
6021 printk(KERN_CONT "\n");
6022 printk(KERN_ERR "ERROR: empty group\n");
6026 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6027 printk(KERN_CONT "\n");
6028 printk(KERN_ERR "ERROR: repeated CPUs\n");
6032 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6034 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6036 printk(KERN_CONT " %s", str);
6037 if (group->cpu_power != SCHED_LOAD_SCALE) {
6038 printk(KERN_CONT " (cpu_power = %d)",
6042 group = group->next;
6043 } while (group != sd->groups);
6044 printk(KERN_CONT "\n");
6046 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6047 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6050 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6051 printk(KERN_ERR "ERROR: parent span is not a superset "
6052 "of domain->span\n");
6056 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6058 cpumask_var_t groupmask;
6061 if (!sched_domain_debug_enabled)
6065 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6069 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6071 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6072 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6077 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6084 free_cpumask_var(groupmask);
6086 #else /* !CONFIG_SCHED_DEBUG */
6087 # define sched_domain_debug(sd, cpu) do { } while (0)
6088 #endif /* CONFIG_SCHED_DEBUG */
6090 static int sd_degenerate(struct sched_domain *sd)
6092 if (cpumask_weight(sched_domain_span(sd)) == 1)
6095 /* Following flags need at least 2 groups */
6096 if (sd->flags & (SD_LOAD_BALANCE |
6097 SD_BALANCE_NEWIDLE |
6101 SD_SHARE_PKG_RESOURCES)) {
6102 if (sd->groups != sd->groups->next)
6106 /* Following flags don't use groups */
6107 if (sd->flags & (SD_WAKE_AFFINE))
6114 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6116 unsigned long cflags = sd->flags, pflags = parent->flags;
6118 if (sd_degenerate(parent))
6121 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6124 /* Flags needing groups don't count if only 1 group in parent */
6125 if (parent->groups == parent->groups->next) {
6126 pflags &= ~(SD_LOAD_BALANCE |
6127 SD_BALANCE_NEWIDLE |
6131 SD_SHARE_PKG_RESOURCES);
6132 if (nr_node_ids == 1)
6133 pflags &= ~SD_SERIALIZE;
6135 if (~cflags & pflags)
6141 static void free_rootdomain(struct root_domain *rd)
6143 synchronize_sched();
6145 cpupri_cleanup(&rd->cpupri);
6147 free_cpumask_var(rd->rto_mask);
6148 free_cpumask_var(rd->online);
6149 free_cpumask_var(rd->span);
6153 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6155 struct root_domain *old_rd = NULL;
6156 unsigned long flags;
6158 raw_spin_lock_irqsave(&rq->lock, flags);
6163 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6166 cpumask_clear_cpu(rq->cpu, old_rd->span);
6169 * If we dont want to free the old_rt yet then
6170 * set old_rd to NULL to skip the freeing later
6173 if (!atomic_dec_and_test(&old_rd->refcount))
6177 atomic_inc(&rd->refcount);
6180 cpumask_set_cpu(rq->cpu, rd->span);
6181 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6184 raw_spin_unlock_irqrestore(&rq->lock, flags);
6187 free_rootdomain(old_rd);
6190 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6192 gfp_t gfp = GFP_KERNEL;
6194 memset(rd, 0, sizeof(*rd));
6199 if (!alloc_cpumask_var(&rd->span, gfp))
6201 if (!alloc_cpumask_var(&rd->online, gfp))
6203 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6206 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6211 free_cpumask_var(rd->rto_mask);
6213 free_cpumask_var(rd->online);
6215 free_cpumask_var(rd->span);
6220 static void init_defrootdomain(void)
6222 init_rootdomain(&def_root_domain, true);
6224 atomic_set(&def_root_domain.refcount, 1);
6227 static struct root_domain *alloc_rootdomain(void)
6229 struct root_domain *rd;
6231 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6235 if (init_rootdomain(rd, false) != 0) {
6244 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6245 * hold the hotplug lock.
6248 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6250 struct rq *rq = cpu_rq(cpu);
6251 struct sched_domain *tmp;
6253 /* Remove the sched domains which do not contribute to scheduling. */
6254 for (tmp = sd; tmp; ) {
6255 struct sched_domain *parent = tmp->parent;
6259 if (sd_parent_degenerate(tmp, parent)) {
6260 tmp->parent = parent->parent;
6262 parent->parent->child = tmp;
6267 if (sd && sd_degenerate(sd)) {
6273 sched_domain_debug(sd, cpu);
6275 rq_attach_root(rq, rd);
6276 rcu_assign_pointer(rq->sd, sd);
6279 /* cpus with isolated domains */
6280 static cpumask_var_t cpu_isolated_map;
6282 /* Setup the mask of cpus configured for isolated domains */
6283 static int __init isolated_cpu_setup(char *str)
6285 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6286 cpulist_parse(str, cpu_isolated_map);
6290 __setup("isolcpus=", isolated_cpu_setup);
6293 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6294 * to a function which identifies what group(along with sched group) a CPU
6295 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6296 * (due to the fact that we keep track of groups covered with a struct cpumask).
6298 * init_sched_build_groups will build a circular linked list of the groups
6299 * covered by the given span, and will set each group's ->cpumask correctly,
6300 * and ->cpu_power to 0.
6303 init_sched_build_groups(const struct cpumask *span,
6304 const struct cpumask *cpu_map,
6305 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6306 struct sched_group **sg,
6307 struct cpumask *tmpmask),
6308 struct cpumask *covered, struct cpumask *tmpmask)
6310 struct sched_group *first = NULL, *last = NULL;
6313 cpumask_clear(covered);
6315 for_each_cpu(i, span) {
6316 struct sched_group *sg;
6317 int group = group_fn(i, cpu_map, &sg, tmpmask);
6320 if (cpumask_test_cpu(i, covered))
6323 cpumask_clear(sched_group_cpus(sg));
6326 for_each_cpu(j, span) {
6327 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6330 cpumask_set_cpu(j, covered);
6331 cpumask_set_cpu(j, sched_group_cpus(sg));
6342 #define SD_NODES_PER_DOMAIN 16
6347 * find_next_best_node - find the next node to include in a sched_domain
6348 * @node: node whose sched_domain we're building
6349 * @used_nodes: nodes already in the sched_domain
6351 * Find the next node to include in a given scheduling domain. Simply
6352 * finds the closest node not already in the @used_nodes map.
6354 * Should use nodemask_t.
6356 static int find_next_best_node(int node, nodemask_t *used_nodes)
6358 int i, n, val, min_val, best_node = 0;
6362 for (i = 0; i < nr_node_ids; i++) {
6363 /* Start at @node */
6364 n = (node + i) % nr_node_ids;
6366 if (!nr_cpus_node(n))
6369 /* Skip already used nodes */
6370 if (node_isset(n, *used_nodes))
6373 /* Simple min distance search */
6374 val = node_distance(node, n);
6376 if (val < min_val) {
6382 node_set(best_node, *used_nodes);
6387 * sched_domain_node_span - get a cpumask for a node's sched_domain
6388 * @node: node whose cpumask we're constructing
6389 * @span: resulting cpumask
6391 * Given a node, construct a good cpumask for its sched_domain to span. It
6392 * should be one that prevents unnecessary balancing, but also spreads tasks
6395 static void sched_domain_node_span(int node, struct cpumask *span)
6397 nodemask_t used_nodes;
6400 cpumask_clear(span);
6401 nodes_clear(used_nodes);
6403 cpumask_or(span, span, cpumask_of_node(node));
6404 node_set(node, used_nodes);
6406 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6407 int next_node = find_next_best_node(node, &used_nodes);
6409 cpumask_or(span, span, cpumask_of_node(next_node));
6412 #endif /* CONFIG_NUMA */
6414 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6417 * The cpus mask in sched_group and sched_domain hangs off the end.
6419 * ( See the the comments in include/linux/sched.h:struct sched_group
6420 * and struct sched_domain. )
6422 struct static_sched_group {
6423 struct sched_group sg;
6424 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6427 struct static_sched_domain {
6428 struct sched_domain sd;
6429 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6435 cpumask_var_t domainspan;
6436 cpumask_var_t covered;
6437 cpumask_var_t notcovered;
6439 cpumask_var_t nodemask;
6440 cpumask_var_t this_sibling_map;
6441 cpumask_var_t this_core_map;
6442 cpumask_var_t send_covered;
6443 cpumask_var_t tmpmask;
6444 struct sched_group **sched_group_nodes;
6445 struct root_domain *rd;
6449 sa_sched_groups = 0,
6454 sa_this_sibling_map,
6456 sa_sched_group_nodes,
6466 * SMT sched-domains:
6468 #ifdef CONFIG_SCHED_SMT
6469 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6470 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6473 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6474 struct sched_group **sg, struct cpumask *unused)
6477 *sg = &per_cpu(sched_groups, cpu).sg;
6480 #endif /* CONFIG_SCHED_SMT */
6483 * multi-core sched-domains:
6485 #ifdef CONFIG_SCHED_MC
6486 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6487 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6488 #endif /* CONFIG_SCHED_MC */
6490 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6492 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6493 struct sched_group **sg, struct cpumask *mask)
6497 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6498 group = cpumask_first(mask);
6500 *sg = &per_cpu(sched_group_core, group).sg;
6503 #elif defined(CONFIG_SCHED_MC)
6505 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6506 struct sched_group **sg, struct cpumask *unused)
6509 *sg = &per_cpu(sched_group_core, cpu).sg;
6514 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6515 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6518 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6519 struct sched_group **sg, struct cpumask *mask)
6522 #ifdef CONFIG_SCHED_MC
6523 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6524 group = cpumask_first(mask);
6525 #elif defined(CONFIG_SCHED_SMT)
6526 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6527 group = cpumask_first(mask);
6532 *sg = &per_cpu(sched_group_phys, group).sg;
6538 * The init_sched_build_groups can't handle what we want to do with node
6539 * groups, so roll our own. Now each node has its own list of groups which
6540 * gets dynamically allocated.
6542 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6543 static struct sched_group ***sched_group_nodes_bycpu;
6545 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6546 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6548 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6549 struct sched_group **sg,
6550 struct cpumask *nodemask)
6554 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6555 group = cpumask_first(nodemask);
6558 *sg = &per_cpu(sched_group_allnodes, group).sg;
6562 static void init_numa_sched_groups_power(struct sched_group *group_head)
6564 struct sched_group *sg = group_head;
6570 for_each_cpu(j, sched_group_cpus(sg)) {
6571 struct sched_domain *sd;
6573 sd = &per_cpu(phys_domains, j).sd;
6574 if (j != group_first_cpu(sd->groups)) {
6576 * Only add "power" once for each
6582 sg->cpu_power += sd->groups->cpu_power;
6585 } while (sg != group_head);
6588 static int build_numa_sched_groups(struct s_data *d,
6589 const struct cpumask *cpu_map, int num)
6591 struct sched_domain *sd;
6592 struct sched_group *sg, *prev;
6595 cpumask_clear(d->covered);
6596 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6597 if (cpumask_empty(d->nodemask)) {
6598 d->sched_group_nodes[num] = NULL;
6602 sched_domain_node_span(num, d->domainspan);
6603 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6605 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6608 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6612 d->sched_group_nodes[num] = sg;
6614 for_each_cpu(j, d->nodemask) {
6615 sd = &per_cpu(node_domains, j).sd;
6620 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6622 cpumask_or(d->covered, d->covered, d->nodemask);
6625 for (j = 0; j < nr_node_ids; j++) {
6626 n = (num + j) % nr_node_ids;
6627 cpumask_complement(d->notcovered, d->covered);
6628 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6629 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6630 if (cpumask_empty(d->tmpmask))
6632 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6633 if (cpumask_empty(d->tmpmask))
6635 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6639 "Can not alloc domain group for node %d\n", j);
6643 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6644 sg->next = prev->next;
6645 cpumask_or(d->covered, d->covered, d->tmpmask);
6652 #endif /* CONFIG_NUMA */
6655 /* Free memory allocated for various sched_group structures */
6656 static void free_sched_groups(const struct cpumask *cpu_map,
6657 struct cpumask *nodemask)
6661 for_each_cpu(cpu, cpu_map) {
6662 struct sched_group **sched_group_nodes
6663 = sched_group_nodes_bycpu[cpu];
6665 if (!sched_group_nodes)
6668 for (i = 0; i < nr_node_ids; i++) {
6669 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6671 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6672 if (cpumask_empty(nodemask))
6682 if (oldsg != sched_group_nodes[i])
6685 kfree(sched_group_nodes);
6686 sched_group_nodes_bycpu[cpu] = NULL;
6689 #else /* !CONFIG_NUMA */
6690 static void free_sched_groups(const struct cpumask *cpu_map,
6691 struct cpumask *nodemask)
6694 #endif /* CONFIG_NUMA */
6697 * Initialize sched groups cpu_power.
6699 * cpu_power indicates the capacity of sched group, which is used while
6700 * distributing the load between different sched groups in a sched domain.
6701 * Typically cpu_power for all the groups in a sched domain will be same unless
6702 * there are asymmetries in the topology. If there are asymmetries, group
6703 * having more cpu_power will pickup more load compared to the group having
6706 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6708 struct sched_domain *child;
6709 struct sched_group *group;
6713 WARN_ON(!sd || !sd->groups);
6715 if (cpu != group_first_cpu(sd->groups))
6720 sd->groups->cpu_power = 0;
6723 power = SCHED_LOAD_SCALE;
6724 weight = cpumask_weight(sched_domain_span(sd));
6726 * SMT siblings share the power of a single core.
6727 * Usually multiple threads get a better yield out of
6728 * that one core than a single thread would have,
6729 * reflect that in sd->smt_gain.
6731 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6732 power *= sd->smt_gain;
6734 power >>= SCHED_LOAD_SHIFT;
6736 sd->groups->cpu_power += power;
6741 * Add cpu_power of each child group to this groups cpu_power.
6743 group = child->groups;
6745 sd->groups->cpu_power += group->cpu_power;
6746 group = group->next;
6747 } while (group != child->groups);
6751 * Initializers for schedule domains
6752 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6755 #ifdef CONFIG_SCHED_DEBUG
6756 # define SD_INIT_NAME(sd, type) sd->name = #type
6758 # define SD_INIT_NAME(sd, type) do { } while (0)
6761 #define SD_INIT(sd, type) sd_init_##type(sd)
6763 #define SD_INIT_FUNC(type) \
6764 static noinline void sd_init_##type(struct sched_domain *sd) \
6766 memset(sd, 0, sizeof(*sd)); \
6767 *sd = SD_##type##_INIT; \
6768 sd->level = SD_LV_##type; \
6769 SD_INIT_NAME(sd, type); \
6774 SD_INIT_FUNC(ALLNODES)
6777 #ifdef CONFIG_SCHED_SMT
6778 SD_INIT_FUNC(SIBLING)
6780 #ifdef CONFIG_SCHED_MC
6784 static int default_relax_domain_level = -1;
6786 static int __init setup_relax_domain_level(char *str)
6790 val = simple_strtoul(str, NULL, 0);
6791 if (val < SD_LV_MAX)
6792 default_relax_domain_level = val;
6796 __setup("relax_domain_level=", setup_relax_domain_level);
6798 static void set_domain_attribute(struct sched_domain *sd,
6799 struct sched_domain_attr *attr)
6803 if (!attr || attr->relax_domain_level < 0) {
6804 if (default_relax_domain_level < 0)
6807 request = default_relax_domain_level;
6809 request = attr->relax_domain_level;
6810 if (request < sd->level) {
6811 /* turn off idle balance on this domain */
6812 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6814 /* turn on idle balance on this domain */
6815 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6819 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6820 const struct cpumask *cpu_map)
6823 case sa_sched_groups:
6824 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6825 d->sched_group_nodes = NULL;
6827 free_rootdomain(d->rd); /* fall through */
6829 free_cpumask_var(d->tmpmask); /* fall through */
6830 case sa_send_covered:
6831 free_cpumask_var(d->send_covered); /* fall through */
6832 case sa_this_core_map:
6833 free_cpumask_var(d->this_core_map); /* fall through */
6834 case sa_this_sibling_map:
6835 free_cpumask_var(d->this_sibling_map); /* fall through */
6837 free_cpumask_var(d->nodemask); /* fall through */
6838 case sa_sched_group_nodes:
6840 kfree(d->sched_group_nodes); /* fall through */
6842 free_cpumask_var(d->notcovered); /* fall through */
6844 free_cpumask_var(d->covered); /* fall through */
6846 free_cpumask_var(d->domainspan); /* fall through */
6853 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6854 const struct cpumask *cpu_map)
6857 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6859 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6860 return sa_domainspan;
6861 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6863 /* Allocate the per-node list of sched groups */
6864 d->sched_group_nodes = kcalloc(nr_node_ids,
6865 sizeof(struct sched_group *), GFP_KERNEL);
6866 if (!d->sched_group_nodes) {
6867 printk(KERN_WARNING "Can not alloc sched group node list\n");
6868 return sa_notcovered;
6870 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6872 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6873 return sa_sched_group_nodes;
6874 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6876 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6877 return sa_this_sibling_map;
6878 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6879 return sa_this_core_map;
6880 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6881 return sa_send_covered;
6882 d->rd = alloc_rootdomain();
6884 printk(KERN_WARNING "Cannot alloc root domain\n");
6887 return sa_rootdomain;
6890 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6891 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6893 struct sched_domain *sd = NULL;
6895 struct sched_domain *parent;
6898 if (cpumask_weight(cpu_map) >
6899 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6900 sd = &per_cpu(allnodes_domains, i).sd;
6901 SD_INIT(sd, ALLNODES);
6902 set_domain_attribute(sd, attr);
6903 cpumask_copy(sched_domain_span(sd), cpu_map);
6904 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6909 sd = &per_cpu(node_domains, i).sd;
6911 set_domain_attribute(sd, attr);
6912 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6913 sd->parent = parent;
6916 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6921 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6922 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6923 struct sched_domain *parent, int i)
6925 struct sched_domain *sd;
6926 sd = &per_cpu(phys_domains, i).sd;
6928 set_domain_attribute(sd, attr);
6929 cpumask_copy(sched_domain_span(sd), d->nodemask);
6930 sd->parent = parent;
6933 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6937 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6938 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6939 struct sched_domain *parent, int i)
6941 struct sched_domain *sd = parent;
6942 #ifdef CONFIG_SCHED_MC
6943 sd = &per_cpu(core_domains, i).sd;
6945 set_domain_attribute(sd, attr);
6946 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6947 sd->parent = parent;
6949 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6954 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6955 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6956 struct sched_domain *parent, int i)
6958 struct sched_domain *sd = parent;
6959 #ifdef CONFIG_SCHED_SMT
6960 sd = &per_cpu(cpu_domains, i).sd;
6961 SD_INIT(sd, SIBLING);
6962 set_domain_attribute(sd, attr);
6963 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6964 sd->parent = parent;
6966 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6971 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6972 const struct cpumask *cpu_map, int cpu)
6975 #ifdef CONFIG_SCHED_SMT
6976 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6977 cpumask_and(d->this_sibling_map, cpu_map,
6978 topology_thread_cpumask(cpu));
6979 if (cpu == cpumask_first(d->this_sibling_map))
6980 init_sched_build_groups(d->this_sibling_map, cpu_map,
6982 d->send_covered, d->tmpmask);
6985 #ifdef CONFIG_SCHED_MC
6986 case SD_LV_MC: /* set up multi-core groups */
6987 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6988 if (cpu == cpumask_first(d->this_core_map))
6989 init_sched_build_groups(d->this_core_map, cpu_map,
6991 d->send_covered, d->tmpmask);
6994 case SD_LV_CPU: /* set up physical groups */
6995 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
6996 if (!cpumask_empty(d->nodemask))
6997 init_sched_build_groups(d->nodemask, cpu_map,
6999 d->send_covered, d->tmpmask);
7002 case SD_LV_ALLNODES:
7003 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7004 d->send_covered, d->tmpmask);
7013 * Build sched domains for a given set of cpus and attach the sched domains
7014 * to the individual cpus
7016 static int __build_sched_domains(const struct cpumask *cpu_map,
7017 struct sched_domain_attr *attr)
7019 enum s_alloc alloc_state = sa_none;
7021 struct sched_domain *sd;
7027 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7028 if (alloc_state != sa_rootdomain)
7030 alloc_state = sa_sched_groups;
7033 * Set up domains for cpus specified by the cpu_map.
7035 for_each_cpu(i, cpu_map) {
7036 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7039 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7040 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7041 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7042 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7045 for_each_cpu(i, cpu_map) {
7046 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7047 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7050 /* Set up physical groups */
7051 for (i = 0; i < nr_node_ids; i++)
7052 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7055 /* Set up node groups */
7057 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7059 for (i = 0; i < nr_node_ids; i++)
7060 if (build_numa_sched_groups(&d, cpu_map, i))
7064 /* Calculate CPU power for physical packages and nodes */
7065 #ifdef CONFIG_SCHED_SMT
7066 for_each_cpu(i, cpu_map) {
7067 sd = &per_cpu(cpu_domains, i).sd;
7068 init_sched_groups_power(i, sd);
7071 #ifdef CONFIG_SCHED_MC
7072 for_each_cpu(i, cpu_map) {
7073 sd = &per_cpu(core_domains, i).sd;
7074 init_sched_groups_power(i, sd);
7078 for_each_cpu(i, cpu_map) {
7079 sd = &per_cpu(phys_domains, i).sd;
7080 init_sched_groups_power(i, sd);
7084 for (i = 0; i < nr_node_ids; i++)
7085 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7087 if (d.sd_allnodes) {
7088 struct sched_group *sg;
7090 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7092 init_numa_sched_groups_power(sg);
7096 /* Attach the domains */
7097 for_each_cpu(i, cpu_map) {
7098 #ifdef CONFIG_SCHED_SMT
7099 sd = &per_cpu(cpu_domains, i).sd;
7100 #elif defined(CONFIG_SCHED_MC)
7101 sd = &per_cpu(core_domains, i).sd;
7103 sd = &per_cpu(phys_domains, i).sd;
7105 cpu_attach_domain(sd, d.rd, i);
7108 d.sched_group_nodes = NULL; /* don't free this we still need it */
7109 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7113 __free_domain_allocs(&d, alloc_state, cpu_map);
7117 static int build_sched_domains(const struct cpumask *cpu_map)
7119 return __build_sched_domains(cpu_map, NULL);
7122 static cpumask_var_t *doms_cur; /* current sched domains */
7123 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7124 static struct sched_domain_attr *dattr_cur;
7125 /* attribues of custom domains in 'doms_cur' */
7128 * Special case: If a kmalloc of a doms_cur partition (array of
7129 * cpumask) fails, then fallback to a single sched domain,
7130 * as determined by the single cpumask fallback_doms.
7132 static cpumask_var_t fallback_doms;
7135 * arch_update_cpu_topology lets virtualized architectures update the
7136 * cpu core maps. It is supposed to return 1 if the topology changed
7137 * or 0 if it stayed the same.
7139 int __attribute__((weak)) arch_update_cpu_topology(void)
7144 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7147 cpumask_var_t *doms;
7149 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7152 for (i = 0; i < ndoms; i++) {
7153 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7154 free_sched_domains(doms, i);
7161 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7164 for (i = 0; i < ndoms; i++)
7165 free_cpumask_var(doms[i]);
7170 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7171 * For now this just excludes isolated cpus, but could be used to
7172 * exclude other special cases in the future.
7174 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7178 arch_update_cpu_topology();
7180 doms_cur = alloc_sched_domains(ndoms_cur);
7182 doms_cur = &fallback_doms;
7183 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7185 err = build_sched_domains(doms_cur[0]);
7186 register_sched_domain_sysctl();
7191 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7192 struct cpumask *tmpmask)
7194 free_sched_groups(cpu_map, tmpmask);
7198 * Detach sched domains from a group of cpus specified in cpu_map
7199 * These cpus will now be attached to the NULL domain
7201 static void detach_destroy_domains(const struct cpumask *cpu_map)
7203 /* Save because hotplug lock held. */
7204 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7207 for_each_cpu(i, cpu_map)
7208 cpu_attach_domain(NULL, &def_root_domain, i);
7209 synchronize_sched();
7210 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7213 /* handle null as "default" */
7214 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7215 struct sched_domain_attr *new, int idx_new)
7217 struct sched_domain_attr tmp;
7224 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7225 new ? (new + idx_new) : &tmp,
7226 sizeof(struct sched_domain_attr));
7230 * Partition sched domains as specified by the 'ndoms_new'
7231 * cpumasks in the array doms_new[] of cpumasks. This compares
7232 * doms_new[] to the current sched domain partitioning, doms_cur[].
7233 * It destroys each deleted domain and builds each new domain.
7235 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7236 * The masks don't intersect (don't overlap.) We should setup one
7237 * sched domain for each mask. CPUs not in any of the cpumasks will
7238 * not be load balanced. If the same cpumask appears both in the
7239 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7242 * The passed in 'doms_new' should be allocated using
7243 * alloc_sched_domains. This routine takes ownership of it and will
7244 * free_sched_domains it when done with it. If the caller failed the
7245 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7246 * and partition_sched_domains() will fallback to the single partition
7247 * 'fallback_doms', it also forces the domains to be rebuilt.
7249 * If doms_new == NULL it will be replaced with cpu_online_mask.
7250 * ndoms_new == 0 is a special case for destroying existing domains,
7251 * and it will not create the default domain.
7253 * Call with hotplug lock held
7255 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7256 struct sched_domain_attr *dattr_new)
7261 mutex_lock(&sched_domains_mutex);
7263 /* always unregister in case we don't destroy any domains */
7264 unregister_sched_domain_sysctl();
7266 /* Let architecture update cpu core mappings. */
7267 new_topology = arch_update_cpu_topology();
7269 n = doms_new ? ndoms_new : 0;
7271 /* Destroy deleted domains */
7272 for (i = 0; i < ndoms_cur; i++) {
7273 for (j = 0; j < n && !new_topology; j++) {
7274 if (cpumask_equal(doms_cur[i], doms_new[j])
7275 && dattrs_equal(dattr_cur, i, dattr_new, j))
7278 /* no match - a current sched domain not in new doms_new[] */
7279 detach_destroy_domains(doms_cur[i]);
7284 if (doms_new == NULL) {
7286 doms_new = &fallback_doms;
7287 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7288 WARN_ON_ONCE(dattr_new);
7291 /* Build new domains */
7292 for (i = 0; i < ndoms_new; i++) {
7293 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7294 if (cpumask_equal(doms_new[i], doms_cur[j])
7295 && dattrs_equal(dattr_new, i, dattr_cur, j))
7298 /* no match - add a new doms_new */
7299 __build_sched_domains(doms_new[i],
7300 dattr_new ? dattr_new + i : NULL);
7305 /* Remember the new sched domains */
7306 if (doms_cur != &fallback_doms)
7307 free_sched_domains(doms_cur, ndoms_cur);
7308 kfree(dattr_cur); /* kfree(NULL) is safe */
7309 doms_cur = doms_new;
7310 dattr_cur = dattr_new;
7311 ndoms_cur = ndoms_new;
7313 register_sched_domain_sysctl();
7315 mutex_unlock(&sched_domains_mutex);
7318 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7319 static void arch_reinit_sched_domains(void)
7323 /* Destroy domains first to force the rebuild */
7324 partition_sched_domains(0, NULL, NULL);
7326 rebuild_sched_domains();
7330 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7332 unsigned int level = 0;
7334 if (sscanf(buf, "%u", &level) != 1)
7338 * level is always be positive so don't check for
7339 * level < POWERSAVINGS_BALANCE_NONE which is 0
7340 * What happens on 0 or 1 byte write,
7341 * need to check for count as well?
7344 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7348 sched_smt_power_savings = level;
7350 sched_mc_power_savings = level;
7352 arch_reinit_sched_domains();
7357 #ifdef CONFIG_SCHED_MC
7358 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7361 return sprintf(page, "%u\n", sched_mc_power_savings);
7363 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7364 const char *buf, size_t count)
7366 return sched_power_savings_store(buf, count, 0);
7368 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7369 sched_mc_power_savings_show,
7370 sched_mc_power_savings_store);
7373 #ifdef CONFIG_SCHED_SMT
7374 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7377 return sprintf(page, "%u\n", sched_smt_power_savings);
7379 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7380 const char *buf, size_t count)
7382 return sched_power_savings_store(buf, count, 1);
7384 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7385 sched_smt_power_savings_show,
7386 sched_smt_power_savings_store);
7389 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7393 #ifdef CONFIG_SCHED_SMT
7395 err = sysfs_create_file(&cls->kset.kobj,
7396 &attr_sched_smt_power_savings.attr);
7398 #ifdef CONFIG_SCHED_MC
7399 if (!err && mc_capable())
7400 err = sysfs_create_file(&cls->kset.kobj,
7401 &attr_sched_mc_power_savings.attr);
7405 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7407 #ifndef CONFIG_CPUSETS
7409 * Add online and remove offline CPUs from the scheduler domains.
7410 * When cpusets are enabled they take over this function.
7412 static int update_sched_domains(struct notifier_block *nfb,
7413 unsigned long action, void *hcpu)
7417 case CPU_ONLINE_FROZEN:
7418 case CPU_DOWN_PREPARE:
7419 case CPU_DOWN_PREPARE_FROZEN:
7420 case CPU_DOWN_FAILED:
7421 case CPU_DOWN_FAILED_FROZEN:
7422 partition_sched_domains(1, NULL, NULL);
7431 static int update_runtime(struct notifier_block *nfb,
7432 unsigned long action, void *hcpu)
7434 int cpu = (int)(long)hcpu;
7437 case CPU_DOWN_PREPARE:
7438 case CPU_DOWN_PREPARE_FROZEN:
7439 disable_runtime(cpu_rq(cpu));
7442 case CPU_DOWN_FAILED:
7443 case CPU_DOWN_FAILED_FROZEN:
7445 case CPU_ONLINE_FROZEN:
7446 enable_runtime(cpu_rq(cpu));
7454 void __init sched_init_smp(void)
7456 cpumask_var_t non_isolated_cpus;
7458 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7459 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7461 #if defined(CONFIG_NUMA)
7462 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7464 BUG_ON(sched_group_nodes_bycpu == NULL);
7467 mutex_lock(&sched_domains_mutex);
7468 arch_init_sched_domains(cpu_active_mask);
7469 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7470 if (cpumask_empty(non_isolated_cpus))
7471 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7472 mutex_unlock(&sched_domains_mutex);
7475 #ifndef CONFIG_CPUSETS
7476 /* XXX: Theoretical race here - CPU may be hotplugged now */
7477 hotcpu_notifier(update_sched_domains, 0);
7480 /* RT runtime code needs to handle some hotplug events */
7481 hotcpu_notifier(update_runtime, 0);
7485 /* Move init over to a non-isolated CPU */
7486 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7488 sched_init_granularity();
7489 free_cpumask_var(non_isolated_cpus);
7491 init_sched_rt_class();
7494 void __init sched_init_smp(void)
7496 sched_init_granularity();
7498 #endif /* CONFIG_SMP */
7500 const_debug unsigned int sysctl_timer_migration = 1;
7502 int in_sched_functions(unsigned long addr)
7504 return in_lock_functions(addr) ||
7505 (addr >= (unsigned long)__sched_text_start
7506 && addr < (unsigned long)__sched_text_end);
7509 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7511 cfs_rq->tasks_timeline = RB_ROOT;
7512 INIT_LIST_HEAD(&cfs_rq->tasks);
7513 #ifdef CONFIG_FAIR_GROUP_SCHED
7516 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7519 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7521 struct rt_prio_array *array;
7524 array = &rt_rq->active;
7525 for (i = 0; i < MAX_RT_PRIO; i++) {
7526 INIT_LIST_HEAD(array->queue + i);
7527 __clear_bit(i, array->bitmap);
7529 /* delimiter for bitsearch: */
7530 __set_bit(MAX_RT_PRIO, array->bitmap);
7532 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7533 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7535 rt_rq->highest_prio.next = MAX_RT_PRIO;
7539 rt_rq->rt_nr_migratory = 0;
7540 rt_rq->overloaded = 0;
7541 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7545 rt_rq->rt_throttled = 0;
7546 rt_rq->rt_runtime = 0;
7547 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7549 #ifdef CONFIG_RT_GROUP_SCHED
7550 rt_rq->rt_nr_boosted = 0;
7555 #ifdef CONFIG_FAIR_GROUP_SCHED
7556 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7557 struct sched_entity *se, int cpu, int add,
7558 struct sched_entity *parent)
7560 struct rq *rq = cpu_rq(cpu);
7561 tg->cfs_rq[cpu] = cfs_rq;
7562 init_cfs_rq(cfs_rq, rq);
7565 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7568 /* se could be NULL for init_task_group */
7573 se->cfs_rq = &rq->cfs;
7575 se->cfs_rq = parent->my_q;
7578 se->load.weight = tg->shares;
7579 se->load.inv_weight = 0;
7580 se->parent = parent;
7584 #ifdef CONFIG_RT_GROUP_SCHED
7585 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7586 struct sched_rt_entity *rt_se, int cpu, int add,
7587 struct sched_rt_entity *parent)
7589 struct rq *rq = cpu_rq(cpu);
7591 tg->rt_rq[cpu] = rt_rq;
7592 init_rt_rq(rt_rq, rq);
7594 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7596 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7598 tg->rt_se[cpu] = rt_se;
7603 rt_se->rt_rq = &rq->rt;
7605 rt_se->rt_rq = parent->my_q;
7607 rt_se->my_q = rt_rq;
7608 rt_se->parent = parent;
7609 INIT_LIST_HEAD(&rt_se->run_list);
7613 void __init sched_init(void)
7616 unsigned long alloc_size = 0, ptr;
7618 #ifdef CONFIG_FAIR_GROUP_SCHED
7619 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7621 #ifdef CONFIG_RT_GROUP_SCHED
7622 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7624 #ifdef CONFIG_CPUMASK_OFFSTACK
7625 alloc_size += num_possible_cpus() * cpumask_size();
7628 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7630 #ifdef CONFIG_FAIR_GROUP_SCHED
7631 init_task_group.se = (struct sched_entity **)ptr;
7632 ptr += nr_cpu_ids * sizeof(void **);
7634 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7635 ptr += nr_cpu_ids * sizeof(void **);
7637 #endif /* CONFIG_FAIR_GROUP_SCHED */
7638 #ifdef CONFIG_RT_GROUP_SCHED
7639 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7640 ptr += nr_cpu_ids * sizeof(void **);
7642 init_task_group.rt_rq = (struct rt_rq **)ptr;
7643 ptr += nr_cpu_ids * sizeof(void **);
7645 #endif /* CONFIG_RT_GROUP_SCHED */
7646 #ifdef CONFIG_CPUMASK_OFFSTACK
7647 for_each_possible_cpu(i) {
7648 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7649 ptr += cpumask_size();
7651 #endif /* CONFIG_CPUMASK_OFFSTACK */
7655 init_defrootdomain();
7658 init_rt_bandwidth(&def_rt_bandwidth,
7659 global_rt_period(), global_rt_runtime());
7661 #ifdef CONFIG_RT_GROUP_SCHED
7662 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7663 global_rt_period(), global_rt_runtime());
7664 #endif /* CONFIG_RT_GROUP_SCHED */
7666 #ifdef CONFIG_CGROUP_SCHED
7667 list_add(&init_task_group.list, &task_groups);
7668 INIT_LIST_HEAD(&init_task_group.children);
7670 #endif /* CONFIG_CGROUP_SCHED */
7672 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7673 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7674 __alignof__(unsigned long));
7676 for_each_possible_cpu(i) {
7680 raw_spin_lock_init(&rq->lock);
7682 rq->calc_load_active = 0;
7683 rq->calc_load_update = jiffies + LOAD_FREQ;
7684 init_cfs_rq(&rq->cfs, rq);
7685 init_rt_rq(&rq->rt, rq);
7686 #ifdef CONFIG_FAIR_GROUP_SCHED
7687 init_task_group.shares = init_task_group_load;
7688 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7689 #ifdef CONFIG_CGROUP_SCHED
7691 * How much cpu bandwidth does init_task_group get?
7693 * In case of task-groups formed thr' the cgroup filesystem, it
7694 * gets 100% of the cpu resources in the system. This overall
7695 * system cpu resource is divided among the tasks of
7696 * init_task_group and its child task-groups in a fair manner,
7697 * based on each entity's (task or task-group's) weight
7698 * (se->load.weight).
7700 * In other words, if init_task_group has 10 tasks of weight
7701 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7702 * then A0's share of the cpu resource is:
7704 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7706 * We achieve this by letting init_task_group's tasks sit
7707 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7709 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7711 #endif /* CONFIG_FAIR_GROUP_SCHED */
7713 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7714 #ifdef CONFIG_RT_GROUP_SCHED
7715 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7716 #ifdef CONFIG_CGROUP_SCHED
7717 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7721 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7722 rq->cpu_load[j] = 0;
7726 rq->post_schedule = 0;
7727 rq->active_balance = 0;
7728 rq->next_balance = jiffies;
7732 rq->migration_thread = NULL;
7734 rq->avg_idle = 2*sysctl_sched_migration_cost;
7735 INIT_LIST_HEAD(&rq->migration_queue);
7736 rq_attach_root(rq, &def_root_domain);
7739 atomic_set(&rq->nr_iowait, 0);
7742 set_load_weight(&init_task);
7744 #ifdef CONFIG_PREEMPT_NOTIFIERS
7745 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7749 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7752 #ifdef CONFIG_RT_MUTEXES
7753 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7757 * The boot idle thread does lazy MMU switching as well:
7759 atomic_inc(&init_mm.mm_count);
7760 enter_lazy_tlb(&init_mm, current);
7763 * Make us the idle thread. Technically, schedule() should not be
7764 * called from this thread, however somewhere below it might be,
7765 * but because we are the idle thread, we just pick up running again
7766 * when this runqueue becomes "idle".
7768 init_idle(current, smp_processor_id());
7770 calc_load_update = jiffies + LOAD_FREQ;
7773 * During early bootup we pretend to be a normal task:
7775 current->sched_class = &fair_sched_class;
7777 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7778 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7781 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7782 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7784 /* May be allocated at isolcpus cmdline parse time */
7785 if (cpu_isolated_map == NULL)
7786 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7791 scheduler_running = 1;
7794 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7795 static inline int preempt_count_equals(int preempt_offset)
7797 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7799 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7802 void __might_sleep(const char *file, int line, int preempt_offset)
7805 static unsigned long prev_jiffy; /* ratelimiting */
7807 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7808 system_state != SYSTEM_RUNNING || oops_in_progress)
7810 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7812 prev_jiffy = jiffies;
7815 "BUG: sleeping function called from invalid context at %s:%d\n",
7818 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7819 in_atomic(), irqs_disabled(),
7820 current->pid, current->comm);
7822 debug_show_held_locks(current);
7823 if (irqs_disabled())
7824 print_irqtrace_events(current);
7828 EXPORT_SYMBOL(__might_sleep);
7831 #ifdef CONFIG_MAGIC_SYSRQ
7832 static void normalize_task(struct rq *rq, struct task_struct *p)
7836 update_rq_clock(rq);
7837 on_rq = p->se.on_rq;
7839 deactivate_task(rq, p, 0);
7840 __setscheduler(rq, p, SCHED_NORMAL, 0);
7842 activate_task(rq, p, 0);
7843 resched_task(rq->curr);
7847 void normalize_rt_tasks(void)
7849 struct task_struct *g, *p;
7850 unsigned long flags;
7853 read_lock_irqsave(&tasklist_lock, flags);
7854 do_each_thread(g, p) {
7856 * Only normalize user tasks:
7861 p->se.exec_start = 0;
7862 #ifdef CONFIG_SCHEDSTATS
7863 p->se.wait_start = 0;
7864 p->se.sleep_start = 0;
7865 p->se.block_start = 0;
7870 * Renice negative nice level userspace
7873 if (TASK_NICE(p) < 0 && p->mm)
7874 set_user_nice(p, 0);
7878 raw_spin_lock(&p->pi_lock);
7879 rq = __task_rq_lock(p);
7881 normalize_task(rq, p);
7883 __task_rq_unlock(rq);
7884 raw_spin_unlock(&p->pi_lock);
7885 } while_each_thread(g, p);
7887 read_unlock_irqrestore(&tasklist_lock, flags);
7890 #endif /* CONFIG_MAGIC_SYSRQ */
7894 * These functions are only useful for the IA64 MCA handling.
7896 * They can only be called when the whole system has been
7897 * stopped - every CPU needs to be quiescent, and no scheduling
7898 * activity can take place. Using them for anything else would
7899 * be a serious bug, and as a result, they aren't even visible
7900 * under any other configuration.
7904 * curr_task - return the current task for a given cpu.
7905 * @cpu: the processor in question.
7907 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7909 struct task_struct *curr_task(int cpu)
7911 return cpu_curr(cpu);
7915 * set_curr_task - set the current task for a given cpu.
7916 * @cpu: the processor in question.
7917 * @p: the task pointer to set.
7919 * Description: This function must only be used when non-maskable interrupts
7920 * are serviced on a separate stack. It allows the architecture to switch the
7921 * notion of the current task on a cpu in a non-blocking manner. This function
7922 * must be called with all CPU's synchronized, and interrupts disabled, the
7923 * and caller must save the original value of the current task (see
7924 * curr_task() above) and restore that value before reenabling interrupts and
7925 * re-starting the system.
7927 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7929 void set_curr_task(int cpu, struct task_struct *p)
7936 #ifdef CONFIG_FAIR_GROUP_SCHED
7937 static void free_fair_sched_group(struct task_group *tg)
7941 for_each_possible_cpu(i) {
7943 kfree(tg->cfs_rq[i]);
7953 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7955 struct cfs_rq *cfs_rq;
7956 struct sched_entity *se;
7960 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7963 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7967 tg->shares = NICE_0_LOAD;
7969 for_each_possible_cpu(i) {
7972 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7973 GFP_KERNEL, cpu_to_node(i));
7977 se = kzalloc_node(sizeof(struct sched_entity),
7978 GFP_KERNEL, cpu_to_node(i));
7982 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
7993 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7995 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7996 &cpu_rq(cpu)->leaf_cfs_rq_list);
7999 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8001 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8003 #else /* !CONFG_FAIR_GROUP_SCHED */
8004 static inline void free_fair_sched_group(struct task_group *tg)
8009 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8014 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8018 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8021 #endif /* CONFIG_FAIR_GROUP_SCHED */
8023 #ifdef CONFIG_RT_GROUP_SCHED
8024 static void free_rt_sched_group(struct task_group *tg)
8028 destroy_rt_bandwidth(&tg->rt_bandwidth);
8030 for_each_possible_cpu(i) {
8032 kfree(tg->rt_rq[i]);
8034 kfree(tg->rt_se[i]);
8042 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8044 struct rt_rq *rt_rq;
8045 struct sched_rt_entity *rt_se;
8049 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8052 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8056 init_rt_bandwidth(&tg->rt_bandwidth,
8057 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8059 for_each_possible_cpu(i) {
8062 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8063 GFP_KERNEL, cpu_to_node(i));
8067 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8068 GFP_KERNEL, cpu_to_node(i));
8072 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8083 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8085 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8086 &cpu_rq(cpu)->leaf_rt_rq_list);
8089 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8091 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8093 #else /* !CONFIG_RT_GROUP_SCHED */
8094 static inline void free_rt_sched_group(struct task_group *tg)
8099 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8104 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8108 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8111 #endif /* CONFIG_RT_GROUP_SCHED */
8113 #ifdef CONFIG_CGROUP_SCHED
8114 static void free_sched_group(struct task_group *tg)
8116 free_fair_sched_group(tg);
8117 free_rt_sched_group(tg);
8121 /* allocate runqueue etc for a new task group */
8122 struct task_group *sched_create_group(struct task_group *parent)
8124 struct task_group *tg;
8125 unsigned long flags;
8128 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8130 return ERR_PTR(-ENOMEM);
8132 if (!alloc_fair_sched_group(tg, parent))
8135 if (!alloc_rt_sched_group(tg, parent))
8138 spin_lock_irqsave(&task_group_lock, flags);
8139 for_each_possible_cpu(i) {
8140 register_fair_sched_group(tg, i);
8141 register_rt_sched_group(tg, i);
8143 list_add_rcu(&tg->list, &task_groups);
8145 WARN_ON(!parent); /* root should already exist */
8147 tg->parent = parent;
8148 INIT_LIST_HEAD(&tg->children);
8149 list_add_rcu(&tg->siblings, &parent->children);
8150 spin_unlock_irqrestore(&task_group_lock, flags);
8155 free_sched_group(tg);
8156 return ERR_PTR(-ENOMEM);
8159 /* rcu callback to free various structures associated with a task group */
8160 static void free_sched_group_rcu(struct rcu_head *rhp)
8162 /* now it should be safe to free those cfs_rqs */
8163 free_sched_group(container_of(rhp, struct task_group, rcu));
8166 /* Destroy runqueue etc associated with a task group */
8167 void sched_destroy_group(struct task_group *tg)
8169 unsigned long flags;
8172 spin_lock_irqsave(&task_group_lock, flags);
8173 for_each_possible_cpu(i) {
8174 unregister_fair_sched_group(tg, i);
8175 unregister_rt_sched_group(tg, i);
8177 list_del_rcu(&tg->list);
8178 list_del_rcu(&tg->siblings);
8179 spin_unlock_irqrestore(&task_group_lock, flags);
8181 /* wait for possible concurrent references to cfs_rqs complete */
8182 call_rcu(&tg->rcu, free_sched_group_rcu);
8185 /* change task's runqueue when it moves between groups.
8186 * The caller of this function should have put the task in its new group
8187 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8188 * reflect its new group.
8190 void sched_move_task(struct task_struct *tsk)
8193 unsigned long flags;
8196 rq = task_rq_lock(tsk, &flags);
8198 update_rq_clock(rq);
8200 running = task_current(rq, tsk);
8201 on_rq = tsk->se.on_rq;
8204 dequeue_task(rq, tsk, 0);
8205 if (unlikely(running))
8206 tsk->sched_class->put_prev_task(rq, tsk);
8208 set_task_rq(tsk, task_cpu(tsk));
8210 #ifdef CONFIG_FAIR_GROUP_SCHED
8211 if (tsk->sched_class->moved_group)
8212 tsk->sched_class->moved_group(tsk, on_rq);
8215 if (unlikely(running))
8216 tsk->sched_class->set_curr_task(rq);
8218 enqueue_task(rq, tsk, 0, false);
8220 task_rq_unlock(rq, &flags);
8222 #endif /* CONFIG_CGROUP_SCHED */
8224 #ifdef CONFIG_FAIR_GROUP_SCHED
8225 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8227 struct cfs_rq *cfs_rq = se->cfs_rq;
8232 dequeue_entity(cfs_rq, se, 0);
8234 se->load.weight = shares;
8235 se->load.inv_weight = 0;
8238 enqueue_entity(cfs_rq, se, 0);
8241 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8243 struct cfs_rq *cfs_rq = se->cfs_rq;
8244 struct rq *rq = cfs_rq->rq;
8245 unsigned long flags;
8247 raw_spin_lock_irqsave(&rq->lock, flags);
8248 __set_se_shares(se, shares);
8249 raw_spin_unlock_irqrestore(&rq->lock, flags);
8252 static DEFINE_MUTEX(shares_mutex);
8254 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8257 unsigned long flags;
8260 * We can't change the weight of the root cgroup.
8265 if (shares < MIN_SHARES)
8266 shares = MIN_SHARES;
8267 else if (shares > MAX_SHARES)
8268 shares = MAX_SHARES;
8270 mutex_lock(&shares_mutex);
8271 if (tg->shares == shares)
8274 spin_lock_irqsave(&task_group_lock, flags);
8275 for_each_possible_cpu(i)
8276 unregister_fair_sched_group(tg, i);
8277 list_del_rcu(&tg->siblings);
8278 spin_unlock_irqrestore(&task_group_lock, flags);
8280 /* wait for any ongoing reference to this group to finish */
8281 synchronize_sched();
8284 * Now we are free to modify the group's share on each cpu
8285 * w/o tripping rebalance_share or load_balance_fair.
8287 tg->shares = shares;
8288 for_each_possible_cpu(i) {
8292 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8293 set_se_shares(tg->se[i], shares);
8297 * Enable load balance activity on this group, by inserting it back on
8298 * each cpu's rq->leaf_cfs_rq_list.
8300 spin_lock_irqsave(&task_group_lock, flags);
8301 for_each_possible_cpu(i)
8302 register_fair_sched_group(tg, i);
8303 list_add_rcu(&tg->siblings, &tg->parent->children);
8304 spin_unlock_irqrestore(&task_group_lock, flags);
8306 mutex_unlock(&shares_mutex);
8310 unsigned long sched_group_shares(struct task_group *tg)
8316 #ifdef CONFIG_RT_GROUP_SCHED
8318 * Ensure that the real time constraints are schedulable.
8320 static DEFINE_MUTEX(rt_constraints_mutex);
8322 static unsigned long to_ratio(u64 period, u64 runtime)
8324 if (runtime == RUNTIME_INF)
8327 return div64_u64(runtime << 20, period);
8330 /* Must be called with tasklist_lock held */
8331 static inline int tg_has_rt_tasks(struct task_group *tg)
8333 struct task_struct *g, *p;
8335 do_each_thread(g, p) {
8336 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8338 } while_each_thread(g, p);
8343 struct rt_schedulable_data {
8344 struct task_group *tg;
8349 static int tg_schedulable(struct task_group *tg, void *data)
8351 struct rt_schedulable_data *d = data;
8352 struct task_group *child;
8353 unsigned long total, sum = 0;
8354 u64 period, runtime;
8356 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8357 runtime = tg->rt_bandwidth.rt_runtime;
8360 period = d->rt_period;
8361 runtime = d->rt_runtime;
8365 * Cannot have more runtime than the period.
8367 if (runtime > period && runtime != RUNTIME_INF)
8371 * Ensure we don't starve existing RT tasks.
8373 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8376 total = to_ratio(period, runtime);
8379 * Nobody can have more than the global setting allows.
8381 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8385 * The sum of our children's runtime should not exceed our own.
8387 list_for_each_entry_rcu(child, &tg->children, siblings) {
8388 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8389 runtime = child->rt_bandwidth.rt_runtime;
8391 if (child == d->tg) {
8392 period = d->rt_period;
8393 runtime = d->rt_runtime;
8396 sum += to_ratio(period, runtime);
8405 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8407 struct rt_schedulable_data data = {
8409 .rt_period = period,
8410 .rt_runtime = runtime,
8413 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8416 static int tg_set_bandwidth(struct task_group *tg,
8417 u64 rt_period, u64 rt_runtime)
8421 mutex_lock(&rt_constraints_mutex);
8422 read_lock(&tasklist_lock);
8423 err = __rt_schedulable(tg, rt_period, rt_runtime);
8427 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8428 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8429 tg->rt_bandwidth.rt_runtime = rt_runtime;
8431 for_each_possible_cpu(i) {
8432 struct rt_rq *rt_rq = tg->rt_rq[i];
8434 raw_spin_lock(&rt_rq->rt_runtime_lock);
8435 rt_rq->rt_runtime = rt_runtime;
8436 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8438 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8440 read_unlock(&tasklist_lock);
8441 mutex_unlock(&rt_constraints_mutex);
8446 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8448 u64 rt_runtime, rt_period;
8450 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8451 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8452 if (rt_runtime_us < 0)
8453 rt_runtime = RUNTIME_INF;
8455 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8458 long sched_group_rt_runtime(struct task_group *tg)
8462 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8465 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8466 do_div(rt_runtime_us, NSEC_PER_USEC);
8467 return rt_runtime_us;
8470 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8472 u64 rt_runtime, rt_period;
8474 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8475 rt_runtime = tg->rt_bandwidth.rt_runtime;
8480 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8483 long sched_group_rt_period(struct task_group *tg)
8487 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8488 do_div(rt_period_us, NSEC_PER_USEC);
8489 return rt_period_us;
8492 static int sched_rt_global_constraints(void)
8494 u64 runtime, period;
8497 if (sysctl_sched_rt_period <= 0)
8500 runtime = global_rt_runtime();
8501 period = global_rt_period();
8504 * Sanity check on the sysctl variables.
8506 if (runtime > period && runtime != RUNTIME_INF)
8509 mutex_lock(&rt_constraints_mutex);
8510 read_lock(&tasklist_lock);
8511 ret = __rt_schedulable(NULL, 0, 0);
8512 read_unlock(&tasklist_lock);
8513 mutex_unlock(&rt_constraints_mutex);
8518 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8520 /* Don't accept realtime tasks when there is no way for them to run */
8521 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8527 #else /* !CONFIG_RT_GROUP_SCHED */
8528 static int sched_rt_global_constraints(void)
8530 unsigned long flags;
8533 if (sysctl_sched_rt_period <= 0)
8537 * There's always some RT tasks in the root group
8538 * -- migration, kstopmachine etc..
8540 if (sysctl_sched_rt_runtime == 0)
8543 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8544 for_each_possible_cpu(i) {
8545 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8547 raw_spin_lock(&rt_rq->rt_runtime_lock);
8548 rt_rq->rt_runtime = global_rt_runtime();
8549 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8551 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8555 #endif /* CONFIG_RT_GROUP_SCHED */
8557 int sched_rt_handler(struct ctl_table *table, int write,
8558 void __user *buffer, size_t *lenp,
8562 int old_period, old_runtime;
8563 static DEFINE_MUTEX(mutex);
8566 old_period = sysctl_sched_rt_period;
8567 old_runtime = sysctl_sched_rt_runtime;
8569 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8571 if (!ret && write) {
8572 ret = sched_rt_global_constraints();
8574 sysctl_sched_rt_period = old_period;
8575 sysctl_sched_rt_runtime = old_runtime;
8577 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8578 def_rt_bandwidth.rt_period =
8579 ns_to_ktime(global_rt_period());
8582 mutex_unlock(&mutex);
8587 #ifdef CONFIG_CGROUP_SCHED
8589 /* return corresponding task_group object of a cgroup */
8590 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8592 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8593 struct task_group, css);
8596 static struct cgroup_subsys_state *
8597 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8599 struct task_group *tg, *parent;
8601 if (!cgrp->parent) {
8602 /* This is early initialization for the top cgroup */
8603 return &init_task_group.css;
8606 parent = cgroup_tg(cgrp->parent);
8607 tg = sched_create_group(parent);
8609 return ERR_PTR(-ENOMEM);
8615 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8617 struct task_group *tg = cgroup_tg(cgrp);
8619 sched_destroy_group(tg);
8623 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8625 #ifdef CONFIG_RT_GROUP_SCHED
8626 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8629 /* We don't support RT-tasks being in separate groups */
8630 if (tsk->sched_class != &fair_sched_class)
8637 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8638 struct task_struct *tsk, bool threadgroup)
8640 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8644 struct task_struct *c;
8646 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8647 retval = cpu_cgroup_can_attach_task(cgrp, c);
8659 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8660 struct cgroup *old_cont, struct task_struct *tsk,
8663 sched_move_task(tsk);
8665 struct task_struct *c;
8667 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8674 #ifdef CONFIG_FAIR_GROUP_SCHED
8675 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8678 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8681 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8683 struct task_group *tg = cgroup_tg(cgrp);
8685 return (u64) tg->shares;
8687 #endif /* CONFIG_FAIR_GROUP_SCHED */
8689 #ifdef CONFIG_RT_GROUP_SCHED
8690 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8693 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8696 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8698 return sched_group_rt_runtime(cgroup_tg(cgrp));
8701 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8704 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8707 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8709 return sched_group_rt_period(cgroup_tg(cgrp));
8711 #endif /* CONFIG_RT_GROUP_SCHED */
8713 static struct cftype cpu_files[] = {
8714 #ifdef CONFIG_FAIR_GROUP_SCHED
8717 .read_u64 = cpu_shares_read_u64,
8718 .write_u64 = cpu_shares_write_u64,
8721 #ifdef CONFIG_RT_GROUP_SCHED
8723 .name = "rt_runtime_us",
8724 .read_s64 = cpu_rt_runtime_read,
8725 .write_s64 = cpu_rt_runtime_write,
8728 .name = "rt_period_us",
8729 .read_u64 = cpu_rt_period_read_uint,
8730 .write_u64 = cpu_rt_period_write_uint,
8735 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8737 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8740 struct cgroup_subsys cpu_cgroup_subsys = {
8742 .create = cpu_cgroup_create,
8743 .destroy = cpu_cgroup_destroy,
8744 .can_attach = cpu_cgroup_can_attach,
8745 .attach = cpu_cgroup_attach,
8746 .populate = cpu_cgroup_populate,
8747 .subsys_id = cpu_cgroup_subsys_id,
8751 #endif /* CONFIG_CGROUP_SCHED */
8753 #ifdef CONFIG_CGROUP_CPUACCT
8756 * CPU accounting code for task groups.
8758 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8759 * (balbir@in.ibm.com).
8762 /* track cpu usage of a group of tasks and its child groups */
8764 struct cgroup_subsys_state css;
8765 /* cpuusage holds pointer to a u64-type object on every cpu */
8767 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8768 struct cpuacct *parent;
8771 struct cgroup_subsys cpuacct_subsys;
8773 /* return cpu accounting group corresponding to this container */
8774 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8776 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8777 struct cpuacct, css);
8780 /* return cpu accounting group to which this task belongs */
8781 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8783 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8784 struct cpuacct, css);
8787 /* create a new cpu accounting group */
8788 static struct cgroup_subsys_state *cpuacct_create(
8789 struct cgroup_subsys *ss, struct cgroup *cgrp)
8791 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8797 ca->cpuusage = alloc_percpu(u64);
8801 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8802 if (percpu_counter_init(&ca->cpustat[i], 0))
8803 goto out_free_counters;
8806 ca->parent = cgroup_ca(cgrp->parent);
8812 percpu_counter_destroy(&ca->cpustat[i]);
8813 free_percpu(ca->cpuusage);
8817 return ERR_PTR(-ENOMEM);
8820 /* destroy an existing cpu accounting group */
8822 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8824 struct cpuacct *ca = cgroup_ca(cgrp);
8827 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8828 percpu_counter_destroy(&ca->cpustat[i]);
8829 free_percpu(ca->cpuusage);
8833 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8835 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8838 #ifndef CONFIG_64BIT
8840 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8842 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8844 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8852 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8854 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8856 #ifndef CONFIG_64BIT
8858 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8860 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8862 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8868 /* return total cpu usage (in nanoseconds) of a group */
8869 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8871 struct cpuacct *ca = cgroup_ca(cgrp);
8872 u64 totalcpuusage = 0;
8875 for_each_present_cpu(i)
8876 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8878 return totalcpuusage;
8881 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8884 struct cpuacct *ca = cgroup_ca(cgrp);
8893 for_each_present_cpu(i)
8894 cpuacct_cpuusage_write(ca, i, 0);
8900 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8903 struct cpuacct *ca = cgroup_ca(cgroup);
8907 for_each_present_cpu(i) {
8908 percpu = cpuacct_cpuusage_read(ca, i);
8909 seq_printf(m, "%llu ", (unsigned long long) percpu);
8911 seq_printf(m, "\n");
8915 static const char *cpuacct_stat_desc[] = {
8916 [CPUACCT_STAT_USER] = "user",
8917 [CPUACCT_STAT_SYSTEM] = "system",
8920 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8921 struct cgroup_map_cb *cb)
8923 struct cpuacct *ca = cgroup_ca(cgrp);
8926 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8927 s64 val = percpu_counter_read(&ca->cpustat[i]);
8928 val = cputime64_to_clock_t(val);
8929 cb->fill(cb, cpuacct_stat_desc[i], val);
8934 static struct cftype files[] = {
8937 .read_u64 = cpuusage_read,
8938 .write_u64 = cpuusage_write,
8941 .name = "usage_percpu",
8942 .read_seq_string = cpuacct_percpu_seq_read,
8946 .read_map = cpuacct_stats_show,
8950 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8952 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8956 * charge this task's execution time to its accounting group.
8958 * called with rq->lock held.
8960 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8965 if (unlikely(!cpuacct_subsys.active))
8968 cpu = task_cpu(tsk);
8974 for (; ca; ca = ca->parent) {
8975 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8976 *cpuusage += cputime;
8983 * Charge the system/user time to the task's accounting group.
8985 static void cpuacct_update_stats(struct task_struct *tsk,
8986 enum cpuacct_stat_index idx, cputime_t val)
8990 if (unlikely(!cpuacct_subsys.active))
8997 percpu_counter_add(&ca->cpustat[idx], val);
9003 struct cgroup_subsys cpuacct_subsys = {
9005 .create = cpuacct_create,
9006 .destroy = cpuacct_destroy,
9007 .populate = cpuacct_populate,
9008 .subsys_id = cpuacct_subsys_id,
9010 #endif /* CONFIG_CGROUP_CPUACCT */
9014 int rcu_expedited_torture_stats(char *page)
9018 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9020 void synchronize_sched_expedited(void)
9023 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9025 #else /* #ifndef CONFIG_SMP */
9027 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9028 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9030 #define RCU_EXPEDITED_STATE_POST -2
9031 #define RCU_EXPEDITED_STATE_IDLE -1
9033 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9035 int rcu_expedited_torture_stats(char *page)
9040 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9041 for_each_online_cpu(cpu) {
9042 cnt += sprintf(&page[cnt], " %d:%d",
9043 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9045 cnt += sprintf(&page[cnt], "\n");
9048 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9050 static long synchronize_sched_expedited_count;
9053 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9054 * approach to force grace period to end quickly. This consumes
9055 * significant time on all CPUs, and is thus not recommended for
9056 * any sort of common-case code.
9058 * Note that it is illegal to call this function while holding any
9059 * lock that is acquired by a CPU-hotplug notifier. Failing to
9060 * observe this restriction will result in deadlock.
9062 void synchronize_sched_expedited(void)
9065 unsigned long flags;
9066 bool need_full_sync = 0;
9068 struct migration_req *req;
9072 smp_mb(); /* ensure prior mod happens before capturing snap. */
9073 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9075 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9077 if (trycount++ < 10)
9078 udelay(trycount * num_online_cpus());
9080 synchronize_sched();
9083 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9084 smp_mb(); /* ensure test happens before caller kfree */
9089 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9090 for_each_online_cpu(cpu) {
9092 req = &per_cpu(rcu_migration_req, cpu);
9093 init_completion(&req->done);
9095 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9096 raw_spin_lock_irqsave(&rq->lock, flags);
9097 list_add(&req->list, &rq->migration_queue);
9098 raw_spin_unlock_irqrestore(&rq->lock, flags);
9099 wake_up_process(rq->migration_thread);
9101 for_each_online_cpu(cpu) {
9102 rcu_expedited_state = cpu;
9103 req = &per_cpu(rcu_migration_req, cpu);
9105 wait_for_completion(&req->done);
9106 raw_spin_lock_irqsave(&rq->lock, flags);
9107 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9109 req->dest_cpu = RCU_MIGRATION_IDLE;
9110 raw_spin_unlock_irqrestore(&rq->lock, flags);
9112 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9113 synchronize_sched_expedited_count++;
9114 mutex_unlock(&rcu_sched_expedited_mutex);
9117 synchronize_sched();
9119 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9121 #endif /* #else #ifndef CONFIG_SMP */