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;
439 struct sched_rt_entity *rt_se;
446 * We add the notion of a root-domain which will be used to define per-domain
447 * variables. Each exclusive cpuset essentially defines an island domain by
448 * fully partitioning the member cpus from any other cpuset. Whenever a new
449 * exclusive cpuset is created, we also create and attach a new root-domain
456 cpumask_var_t online;
459 * The "RT overload" flag: it gets set if a CPU has more than
460 * one runnable RT task.
462 cpumask_var_t rto_mask;
465 struct cpupri cpupri;
470 * By default the system creates a single root-domain with all cpus as
471 * members (mimicking the global state we have today).
473 static struct root_domain def_root_domain;
478 * This is the main, per-CPU runqueue data structure.
480 * Locking rule: those places that want to lock multiple runqueues
481 * (such as the load balancing or the thread migration code), lock
482 * acquire operations must be ordered by ascending &runqueue.
489 * nr_running and cpu_load should be in the same cacheline because
490 * remote CPUs use both these fields when doing load calculation.
492 unsigned long nr_running;
493 #define CPU_LOAD_IDX_MAX 5
494 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
496 unsigned char in_nohz_recently;
498 /* capture load from *all* tasks on this cpu: */
499 struct load_weight load;
500 unsigned long nr_load_updates;
506 #ifdef CONFIG_FAIR_GROUP_SCHED
507 /* list of leaf cfs_rq on this cpu: */
508 struct list_head leaf_cfs_rq_list;
510 #ifdef CONFIG_RT_GROUP_SCHED
511 struct list_head leaf_rt_rq_list;
515 * This is part of a global counter where only the total sum
516 * over all CPUs matters. A task can increase this counter on
517 * one CPU and if it got migrated afterwards it may decrease
518 * it on another CPU. Always updated under the runqueue lock:
520 unsigned long nr_uninterruptible;
522 struct task_struct *curr, *idle;
523 unsigned long next_balance;
524 struct mm_struct *prev_mm;
531 struct root_domain *rd;
532 struct sched_domain *sd;
534 unsigned char idle_at_tick;
535 /* For active balancing */
539 /* cpu of this runqueue: */
543 unsigned long avg_load_per_task;
545 struct task_struct *migration_thread;
546 struct list_head migration_queue;
554 /* calc_load related fields */
555 unsigned long calc_load_update;
556 long calc_load_active;
558 #ifdef CONFIG_SCHED_HRTICK
560 int hrtick_csd_pending;
561 struct call_single_data hrtick_csd;
563 struct hrtimer hrtick_timer;
566 #ifdef CONFIG_SCHEDSTATS
568 struct sched_info rq_sched_info;
569 unsigned long long rq_cpu_time;
570 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
572 /* sys_sched_yield() stats */
573 unsigned int yld_count;
575 /* schedule() stats */
576 unsigned int sched_switch;
577 unsigned int sched_count;
578 unsigned int sched_goidle;
580 /* try_to_wake_up() stats */
581 unsigned int ttwu_count;
582 unsigned int ttwu_local;
585 unsigned int bkl_count;
589 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
592 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
594 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
597 static inline int cpu_of(struct rq *rq)
607 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
608 * See detach_destroy_domains: synchronize_sched for details.
610 * The domain tree of any CPU may only be accessed from within
611 * preempt-disabled sections.
613 #define for_each_domain(cpu, __sd) \
614 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
616 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
617 #define this_rq() (&__get_cpu_var(runqueues))
618 #define task_rq(p) cpu_rq(task_cpu(p))
619 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
620 #define raw_rq() (&__raw_get_cpu_var(runqueues))
622 inline void update_rq_clock(struct rq *rq)
624 rq->clock = sched_clock_cpu(cpu_of(rq));
628 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
630 #ifdef CONFIG_SCHED_DEBUG
631 # define const_debug __read_mostly
633 # define const_debug static const
638 * @cpu: the processor in question.
640 * Returns true if the current cpu runqueue is locked.
641 * This interface allows printk to be called with the runqueue lock
642 * held and know whether or not it is OK to wake up the klogd.
644 int runqueue_is_locked(int cpu)
646 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
650 * Debugging: various feature bits
653 #define SCHED_FEAT(name, enabled) \
654 __SCHED_FEAT_##name ,
657 #include "sched_features.h"
662 #define SCHED_FEAT(name, enabled) \
663 (1UL << __SCHED_FEAT_##name) * enabled |
665 const_debug unsigned int sysctl_sched_features =
666 #include "sched_features.h"
671 #ifdef CONFIG_SCHED_DEBUG
672 #define SCHED_FEAT(name, enabled) \
675 static __read_mostly char *sched_feat_names[] = {
676 #include "sched_features.h"
682 static int sched_feat_show(struct seq_file *m, void *v)
686 for (i = 0; sched_feat_names[i]; i++) {
687 if (!(sysctl_sched_features & (1UL << i)))
689 seq_printf(m, "%s ", sched_feat_names[i]);
697 sched_feat_write(struct file *filp, const char __user *ubuf,
698 size_t cnt, loff_t *ppos)
708 if (copy_from_user(&buf, ubuf, cnt))
713 if (strncmp(buf, "NO_", 3) == 0) {
718 for (i = 0; sched_feat_names[i]; i++) {
719 int len = strlen(sched_feat_names[i]);
721 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
723 sysctl_sched_features &= ~(1UL << i);
725 sysctl_sched_features |= (1UL << i);
730 if (!sched_feat_names[i])
738 static int sched_feat_open(struct inode *inode, struct file *filp)
740 return single_open(filp, sched_feat_show, NULL);
743 static const struct file_operations sched_feat_fops = {
744 .open = sched_feat_open,
745 .write = sched_feat_write,
748 .release = single_release,
751 static __init int sched_init_debug(void)
753 debugfs_create_file("sched_features", 0644, NULL, NULL,
758 late_initcall(sched_init_debug);
762 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
765 * Number of tasks to iterate in a single balance run.
766 * Limited because this is done with IRQs disabled.
768 const_debug unsigned int sysctl_sched_nr_migrate = 32;
771 * ratelimit for updating the group shares.
774 unsigned int sysctl_sched_shares_ratelimit = 250000;
775 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
778 * Inject some fuzzyness into changing the per-cpu group shares
779 * this avoids remote rq-locks at the expense of fairness.
782 unsigned int sysctl_sched_shares_thresh = 4;
785 * period over which we average the RT time consumption, measured
790 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
793 * period over which we measure -rt task cpu usage in us.
796 unsigned int sysctl_sched_rt_period = 1000000;
798 static __read_mostly int scheduler_running;
801 * part of the period that we allow rt tasks to run in us.
804 int sysctl_sched_rt_runtime = 950000;
806 static inline u64 global_rt_period(void)
808 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
811 static inline u64 global_rt_runtime(void)
813 if (sysctl_sched_rt_runtime < 0)
816 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
819 #ifndef prepare_arch_switch
820 # define prepare_arch_switch(next) do { } while (0)
822 #ifndef finish_arch_switch
823 # define finish_arch_switch(prev) do { } while (0)
826 static inline int task_current(struct rq *rq, struct task_struct *p)
828 return rq->curr == p;
831 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
832 static inline int task_running(struct rq *rq, struct task_struct *p)
834 return task_current(rq, p);
837 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
841 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
843 #ifdef CONFIG_DEBUG_SPINLOCK
844 /* this is a valid case when another task releases the spinlock */
845 rq->lock.owner = current;
848 * If we are tracking spinlock dependencies then we have to
849 * fix up the runqueue lock - which gets 'carried over' from
852 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
854 raw_spin_unlock_irq(&rq->lock);
857 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
858 static inline int task_running(struct rq *rq, struct task_struct *p)
863 return task_current(rq, p);
867 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
871 * We can optimise this out completely for !SMP, because the
872 * SMP rebalancing from interrupt is the only thing that cares
877 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
878 raw_spin_unlock_irq(&rq->lock);
880 raw_spin_unlock(&rq->lock);
884 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
888 * After ->oncpu is cleared, the task can be moved to a different CPU.
889 * We must ensure this doesn't happen until the switch is completely
895 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
899 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
902 * __task_rq_lock - lock the runqueue a given task resides on.
903 * Must be called interrupts disabled.
905 static inline struct rq *__task_rq_lock(struct task_struct *p)
909 struct rq *rq = task_rq(p);
910 raw_spin_lock(&rq->lock);
911 if (likely(rq == task_rq(p)))
913 raw_spin_unlock(&rq->lock);
918 * task_rq_lock - lock the runqueue a given task resides on and disable
919 * interrupts. Note the ordering: we can safely lookup the task_rq without
920 * explicitly disabling preemption.
922 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
928 local_irq_save(*flags);
930 raw_spin_lock(&rq->lock);
931 if (likely(rq == task_rq(p)))
933 raw_spin_unlock_irqrestore(&rq->lock, *flags);
937 void task_rq_unlock_wait(struct task_struct *p)
939 struct rq *rq = task_rq(p);
941 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
942 raw_spin_unlock_wait(&rq->lock);
945 static void __task_rq_unlock(struct rq *rq)
948 raw_spin_unlock(&rq->lock);
951 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
954 raw_spin_unlock_irqrestore(&rq->lock, *flags);
958 * this_rq_lock - lock this runqueue and disable interrupts.
960 static struct rq *this_rq_lock(void)
967 raw_spin_lock(&rq->lock);
972 #ifdef CONFIG_SCHED_HRTICK
974 * Use HR-timers to deliver accurate preemption points.
976 * Its all a bit involved since we cannot program an hrt while holding the
977 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
980 * When we get rescheduled we reprogram the hrtick_timer outside of the
986 * - enabled by features
987 * - hrtimer is actually high res
989 static inline int hrtick_enabled(struct rq *rq)
991 if (!sched_feat(HRTICK))
993 if (!cpu_active(cpu_of(rq)))
995 return hrtimer_is_hres_active(&rq->hrtick_timer);
998 static void hrtick_clear(struct rq *rq)
1000 if (hrtimer_active(&rq->hrtick_timer))
1001 hrtimer_cancel(&rq->hrtick_timer);
1005 * High-resolution timer tick.
1006 * Runs from hardirq context with interrupts disabled.
1008 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1010 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1012 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1014 raw_spin_lock(&rq->lock);
1015 update_rq_clock(rq);
1016 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1017 raw_spin_unlock(&rq->lock);
1019 return HRTIMER_NORESTART;
1024 * called from hardirq (IPI) context
1026 static void __hrtick_start(void *arg)
1028 struct rq *rq = arg;
1030 raw_spin_lock(&rq->lock);
1031 hrtimer_restart(&rq->hrtick_timer);
1032 rq->hrtick_csd_pending = 0;
1033 raw_spin_unlock(&rq->lock);
1037 * Called to set the hrtick timer state.
1039 * called with rq->lock held and irqs disabled
1041 static void hrtick_start(struct rq *rq, u64 delay)
1043 struct hrtimer *timer = &rq->hrtick_timer;
1044 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1046 hrtimer_set_expires(timer, time);
1048 if (rq == this_rq()) {
1049 hrtimer_restart(timer);
1050 } else if (!rq->hrtick_csd_pending) {
1051 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1052 rq->hrtick_csd_pending = 1;
1057 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1059 int cpu = (int)(long)hcpu;
1062 case CPU_UP_CANCELED:
1063 case CPU_UP_CANCELED_FROZEN:
1064 case CPU_DOWN_PREPARE:
1065 case CPU_DOWN_PREPARE_FROZEN:
1067 case CPU_DEAD_FROZEN:
1068 hrtick_clear(cpu_rq(cpu));
1075 static __init void init_hrtick(void)
1077 hotcpu_notifier(hotplug_hrtick, 0);
1081 * Called to set the hrtick timer state.
1083 * called with rq->lock held and irqs disabled
1085 static void hrtick_start(struct rq *rq, u64 delay)
1087 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1088 HRTIMER_MODE_REL_PINNED, 0);
1091 static inline void init_hrtick(void)
1094 #endif /* CONFIG_SMP */
1096 static void init_rq_hrtick(struct rq *rq)
1099 rq->hrtick_csd_pending = 0;
1101 rq->hrtick_csd.flags = 0;
1102 rq->hrtick_csd.func = __hrtick_start;
1103 rq->hrtick_csd.info = rq;
1106 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1107 rq->hrtick_timer.function = hrtick;
1109 #else /* CONFIG_SCHED_HRTICK */
1110 static inline void hrtick_clear(struct rq *rq)
1114 static inline void init_rq_hrtick(struct rq *rq)
1118 static inline void init_hrtick(void)
1121 #endif /* CONFIG_SCHED_HRTICK */
1124 * resched_task - mark a task 'to be rescheduled now'.
1126 * On UP this means the setting of the need_resched flag, on SMP it
1127 * might also involve a cross-CPU call to trigger the scheduler on
1132 #ifndef tsk_is_polling
1133 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1136 static void resched_task(struct task_struct *p)
1140 assert_raw_spin_locked(&task_rq(p)->lock);
1142 if (test_tsk_need_resched(p))
1145 set_tsk_need_resched(p);
1148 if (cpu == smp_processor_id())
1151 /* NEED_RESCHED must be visible before we test polling */
1153 if (!tsk_is_polling(p))
1154 smp_send_reschedule(cpu);
1157 static void resched_cpu(int cpu)
1159 struct rq *rq = cpu_rq(cpu);
1160 unsigned long flags;
1162 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1164 resched_task(cpu_curr(cpu));
1165 raw_spin_unlock_irqrestore(&rq->lock, flags);
1170 * When add_timer_on() enqueues a timer into the timer wheel of an
1171 * idle CPU then this timer might expire before the next timer event
1172 * which is scheduled to wake up that CPU. In case of a completely
1173 * idle system the next event might even be infinite time into the
1174 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1175 * leaves the inner idle loop so the newly added timer is taken into
1176 * account when the CPU goes back to idle and evaluates the timer
1177 * wheel for the next timer event.
1179 void wake_up_idle_cpu(int cpu)
1181 struct rq *rq = cpu_rq(cpu);
1183 if (cpu == smp_processor_id())
1187 * This is safe, as this function is called with the timer
1188 * wheel base lock of (cpu) held. When the CPU is on the way
1189 * to idle and has not yet set rq->curr to idle then it will
1190 * be serialized on the timer wheel base lock and take the new
1191 * timer into account automatically.
1193 if (rq->curr != rq->idle)
1197 * We can set TIF_RESCHED on the idle task of the other CPU
1198 * lockless. The worst case is that the other CPU runs the
1199 * idle task through an additional NOOP schedule()
1201 set_tsk_need_resched(rq->idle);
1203 /* NEED_RESCHED must be visible before we test polling */
1205 if (!tsk_is_polling(rq->idle))
1206 smp_send_reschedule(cpu);
1208 #endif /* CONFIG_NO_HZ */
1210 static u64 sched_avg_period(void)
1212 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1215 static void sched_avg_update(struct rq *rq)
1217 s64 period = sched_avg_period();
1219 while ((s64)(rq->clock - rq->age_stamp) > period) {
1220 rq->age_stamp += period;
1225 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1227 rq->rt_avg += rt_delta;
1228 sched_avg_update(rq);
1231 #else /* !CONFIG_SMP */
1232 static void resched_task(struct task_struct *p)
1234 assert_raw_spin_locked(&task_rq(p)->lock);
1235 set_tsk_need_resched(p);
1238 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1241 #endif /* CONFIG_SMP */
1243 #if BITS_PER_LONG == 32
1244 # define WMULT_CONST (~0UL)
1246 # define WMULT_CONST (1UL << 32)
1249 #define WMULT_SHIFT 32
1252 * Shift right and round:
1254 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1257 * delta *= weight / lw
1259 static unsigned long
1260 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1261 struct load_weight *lw)
1265 if (!lw->inv_weight) {
1266 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1269 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1273 tmp = (u64)delta_exec * weight;
1275 * Check whether we'd overflow the 64-bit multiplication:
1277 if (unlikely(tmp > WMULT_CONST))
1278 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1281 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1283 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1286 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1292 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1299 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1300 * of tasks with abnormal "nice" values across CPUs the contribution that
1301 * each task makes to its run queue's load is weighted according to its
1302 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1303 * scaled version of the new time slice allocation that they receive on time
1307 #define WEIGHT_IDLEPRIO 3
1308 #define WMULT_IDLEPRIO 1431655765
1311 * Nice levels are multiplicative, with a gentle 10% change for every
1312 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1313 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1314 * that remained on nice 0.
1316 * The "10% effect" is relative and cumulative: from _any_ nice level,
1317 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1318 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1319 * If a task goes up by ~10% and another task goes down by ~10% then
1320 * the relative distance between them is ~25%.)
1322 static const int prio_to_weight[40] = {
1323 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1324 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1325 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1326 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1327 /* 0 */ 1024, 820, 655, 526, 423,
1328 /* 5 */ 335, 272, 215, 172, 137,
1329 /* 10 */ 110, 87, 70, 56, 45,
1330 /* 15 */ 36, 29, 23, 18, 15,
1334 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1336 * In cases where the weight does not change often, we can use the
1337 * precalculated inverse to speed up arithmetics by turning divisions
1338 * into multiplications:
1340 static const u32 prio_to_wmult[40] = {
1341 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1342 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1343 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1344 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1345 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1346 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1347 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1348 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1351 /* Time spent by the tasks of the cpu accounting group executing in ... */
1352 enum cpuacct_stat_index {
1353 CPUACCT_STAT_USER, /* ... user mode */
1354 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1356 CPUACCT_STAT_NSTATS,
1359 #ifdef CONFIG_CGROUP_CPUACCT
1360 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1361 static void cpuacct_update_stats(struct task_struct *tsk,
1362 enum cpuacct_stat_index idx, cputime_t val);
1364 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1365 static inline void cpuacct_update_stats(struct task_struct *tsk,
1366 enum cpuacct_stat_index idx, cputime_t val) {}
1369 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1371 update_load_add(&rq->load, load);
1374 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1376 update_load_sub(&rq->load, load);
1379 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1380 typedef int (*tg_visitor)(struct task_group *, void *);
1383 * Iterate the full tree, calling @down when first entering a node and @up when
1384 * leaving it for the final time.
1386 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1388 struct task_group *parent, *child;
1392 parent = &root_task_group;
1394 ret = (*down)(parent, data);
1397 list_for_each_entry_rcu(child, &parent->children, siblings) {
1404 ret = (*up)(parent, data);
1409 parent = parent->parent;
1418 static int tg_nop(struct task_group *tg, void *data)
1425 /* Used instead of source_load when we know the type == 0 */
1426 static unsigned long weighted_cpuload(const int cpu)
1428 return cpu_rq(cpu)->load.weight;
1432 * Return a low guess at the load of a migration-source cpu weighted
1433 * according to the scheduling class and "nice" value.
1435 * We want to under-estimate the load of migration sources, to
1436 * balance conservatively.
1438 static unsigned long source_load(int cpu, int type)
1440 struct rq *rq = cpu_rq(cpu);
1441 unsigned long total = weighted_cpuload(cpu);
1443 if (type == 0 || !sched_feat(LB_BIAS))
1446 return min(rq->cpu_load[type-1], total);
1450 * Return a high guess at the load of a migration-target cpu weighted
1451 * according to the scheduling class and "nice" value.
1453 static unsigned long target_load(int cpu, int type)
1455 struct rq *rq = cpu_rq(cpu);
1456 unsigned long total = weighted_cpuload(cpu);
1458 if (type == 0 || !sched_feat(LB_BIAS))
1461 return max(rq->cpu_load[type-1], total);
1464 static struct sched_group *group_of(int cpu)
1466 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1474 static unsigned long power_of(int cpu)
1476 struct sched_group *group = group_of(cpu);
1479 return SCHED_LOAD_SCALE;
1481 return group->cpu_power;
1484 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1486 static unsigned long cpu_avg_load_per_task(int cpu)
1488 struct rq *rq = cpu_rq(cpu);
1489 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1492 rq->avg_load_per_task = rq->load.weight / nr_running;
1494 rq->avg_load_per_task = 0;
1496 return rq->avg_load_per_task;
1499 #ifdef CONFIG_FAIR_GROUP_SCHED
1501 static __read_mostly unsigned long *update_shares_data;
1503 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1506 * Calculate and set the cpu's group shares.
1508 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1509 unsigned long sd_shares,
1510 unsigned long sd_rq_weight,
1511 unsigned long *usd_rq_weight)
1513 unsigned long shares, rq_weight;
1516 rq_weight = usd_rq_weight[cpu];
1519 rq_weight = NICE_0_LOAD;
1523 * \Sum_j shares_j * rq_weight_i
1524 * shares_i = -----------------------------
1525 * \Sum_j rq_weight_j
1527 shares = (sd_shares * rq_weight) / sd_rq_weight;
1528 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1530 if (abs(shares - tg->se[cpu]->load.weight) >
1531 sysctl_sched_shares_thresh) {
1532 struct rq *rq = cpu_rq(cpu);
1533 unsigned long flags;
1535 raw_spin_lock_irqsave(&rq->lock, flags);
1536 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1537 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1538 __set_se_shares(tg->se[cpu], shares);
1539 raw_spin_unlock_irqrestore(&rq->lock, flags);
1544 * Re-compute the task group their per cpu shares over the given domain.
1545 * This needs to be done in a bottom-up fashion because the rq weight of a
1546 * parent group depends on the shares of its child groups.
1548 static int tg_shares_up(struct task_group *tg, void *data)
1550 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1551 unsigned long *usd_rq_weight;
1552 struct sched_domain *sd = data;
1553 unsigned long flags;
1559 local_irq_save(flags);
1560 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1562 for_each_cpu(i, sched_domain_span(sd)) {
1563 weight = tg->cfs_rq[i]->load.weight;
1564 usd_rq_weight[i] = weight;
1566 rq_weight += weight;
1568 * If there are currently no tasks on the cpu pretend there
1569 * is one of average load so that when a new task gets to
1570 * run here it will not get delayed by group starvation.
1573 weight = NICE_0_LOAD;
1575 sum_weight += weight;
1576 shares += tg->cfs_rq[i]->shares;
1580 rq_weight = sum_weight;
1582 if ((!shares && rq_weight) || shares > tg->shares)
1583 shares = tg->shares;
1585 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1586 shares = tg->shares;
1588 for_each_cpu(i, sched_domain_span(sd))
1589 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1591 local_irq_restore(flags);
1597 * Compute the cpu's hierarchical load factor for each task group.
1598 * This needs to be done in a top-down fashion because the load of a child
1599 * group is a fraction of its parents load.
1601 static int tg_load_down(struct task_group *tg, void *data)
1604 long cpu = (long)data;
1607 load = cpu_rq(cpu)->load.weight;
1609 load = tg->parent->cfs_rq[cpu]->h_load;
1610 load *= tg->cfs_rq[cpu]->shares;
1611 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1614 tg->cfs_rq[cpu]->h_load = load;
1619 static void update_shares(struct sched_domain *sd)
1624 if (root_task_group_empty())
1627 now = cpu_clock(raw_smp_processor_id());
1628 elapsed = now - sd->last_update;
1630 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1631 sd->last_update = now;
1632 walk_tg_tree(tg_nop, tg_shares_up, sd);
1636 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1638 if (root_task_group_empty())
1641 raw_spin_unlock(&rq->lock);
1643 raw_spin_lock(&rq->lock);
1646 static void update_h_load(long cpu)
1648 if (root_task_group_empty())
1651 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1656 static inline void update_shares(struct sched_domain *sd)
1660 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1666 #ifdef CONFIG_PREEMPT
1668 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1671 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1672 * way at the expense of forcing extra atomic operations in all
1673 * invocations. This assures that the double_lock is acquired using the
1674 * same underlying policy as the spinlock_t on this architecture, which
1675 * reduces latency compared to the unfair variant below. However, it
1676 * also adds more overhead and therefore may reduce throughput.
1678 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1679 __releases(this_rq->lock)
1680 __acquires(busiest->lock)
1681 __acquires(this_rq->lock)
1683 raw_spin_unlock(&this_rq->lock);
1684 double_rq_lock(this_rq, busiest);
1691 * Unfair double_lock_balance: Optimizes throughput at the expense of
1692 * latency by eliminating extra atomic operations when the locks are
1693 * already in proper order on entry. This favors lower cpu-ids and will
1694 * grant the double lock to lower cpus over higher ids under contention,
1695 * regardless of entry order into the function.
1697 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1698 __releases(this_rq->lock)
1699 __acquires(busiest->lock)
1700 __acquires(this_rq->lock)
1704 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1705 if (busiest < this_rq) {
1706 raw_spin_unlock(&this_rq->lock);
1707 raw_spin_lock(&busiest->lock);
1708 raw_spin_lock_nested(&this_rq->lock,
1709 SINGLE_DEPTH_NESTING);
1712 raw_spin_lock_nested(&busiest->lock,
1713 SINGLE_DEPTH_NESTING);
1718 #endif /* CONFIG_PREEMPT */
1721 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1723 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1725 if (unlikely(!irqs_disabled())) {
1726 /* printk() doesn't work good under rq->lock */
1727 raw_spin_unlock(&this_rq->lock);
1731 return _double_lock_balance(this_rq, busiest);
1734 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1735 __releases(busiest->lock)
1737 raw_spin_unlock(&busiest->lock);
1738 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1742 * double_rq_lock - safely lock two runqueues
1744 * Note this does not disable interrupts like task_rq_lock,
1745 * you need to do so manually before calling.
1747 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1748 __acquires(rq1->lock)
1749 __acquires(rq2->lock)
1751 BUG_ON(!irqs_disabled());
1753 raw_spin_lock(&rq1->lock);
1754 __acquire(rq2->lock); /* Fake it out ;) */
1757 raw_spin_lock(&rq1->lock);
1758 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1760 raw_spin_lock(&rq2->lock);
1761 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1764 update_rq_clock(rq1);
1765 update_rq_clock(rq2);
1769 * double_rq_unlock - safely unlock two runqueues
1771 * Note this does not restore interrupts like task_rq_unlock,
1772 * you need to do so manually after calling.
1774 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1775 __releases(rq1->lock)
1776 __releases(rq2->lock)
1778 raw_spin_unlock(&rq1->lock);
1780 raw_spin_unlock(&rq2->lock);
1782 __release(rq2->lock);
1787 #ifdef CONFIG_FAIR_GROUP_SCHED
1788 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1791 cfs_rq->shares = shares;
1796 static void calc_load_account_active(struct rq *this_rq);
1797 static void update_sysctl(void);
1798 static int get_update_sysctl_factor(void);
1800 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1802 set_task_rq(p, cpu);
1805 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1806 * successfuly executed on another CPU. We must ensure that updates of
1807 * per-task data have been completed by this moment.
1810 task_thread_info(p)->cpu = cpu;
1814 static const struct sched_class rt_sched_class;
1816 #define sched_class_highest (&rt_sched_class)
1817 #define for_each_class(class) \
1818 for (class = sched_class_highest; class; class = class->next)
1820 #include "sched_stats.h"
1822 static void inc_nr_running(struct rq *rq)
1827 static void dec_nr_running(struct rq *rq)
1832 static void set_load_weight(struct task_struct *p)
1834 if (task_has_rt_policy(p)) {
1835 p->se.load.weight = prio_to_weight[0] * 2;
1836 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1841 * SCHED_IDLE tasks get minimal weight:
1843 if (p->policy == SCHED_IDLE) {
1844 p->se.load.weight = WEIGHT_IDLEPRIO;
1845 p->se.load.inv_weight = WMULT_IDLEPRIO;
1849 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1850 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1853 static void update_avg(u64 *avg, u64 sample)
1855 s64 diff = sample - *avg;
1859 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1862 p->se.start_runtime = p->se.sum_exec_runtime;
1864 sched_info_queued(p);
1865 p->sched_class->enqueue_task(rq, p, wakeup);
1869 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1872 if (p->se.last_wakeup) {
1873 update_avg(&p->se.avg_overlap,
1874 p->se.sum_exec_runtime - p->se.last_wakeup);
1875 p->se.last_wakeup = 0;
1877 update_avg(&p->se.avg_wakeup,
1878 sysctl_sched_wakeup_granularity);
1882 sched_info_dequeued(p);
1883 p->sched_class->dequeue_task(rq, p, sleep);
1888 * activate_task - move a task to the runqueue.
1890 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1892 if (task_contributes_to_load(p))
1893 rq->nr_uninterruptible--;
1895 enqueue_task(rq, p, wakeup);
1900 * deactivate_task - remove a task from the runqueue.
1902 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1904 if (task_contributes_to_load(p))
1905 rq->nr_uninterruptible++;
1907 dequeue_task(rq, p, sleep);
1911 #include "sched_idletask.c"
1912 #include "sched_fair.c"
1913 #include "sched_rt.c"
1914 #ifdef CONFIG_SCHED_DEBUG
1915 # include "sched_debug.c"
1919 * __normal_prio - return the priority that is based on the static prio
1921 static inline int __normal_prio(struct task_struct *p)
1923 return p->static_prio;
1927 * Calculate the expected normal priority: i.e. priority
1928 * without taking RT-inheritance into account. Might be
1929 * boosted by interactivity modifiers. Changes upon fork,
1930 * setprio syscalls, and whenever the interactivity
1931 * estimator recalculates.
1933 static inline int normal_prio(struct task_struct *p)
1937 if (task_has_rt_policy(p))
1938 prio = MAX_RT_PRIO-1 - p->rt_priority;
1940 prio = __normal_prio(p);
1945 * Calculate the current priority, i.e. the priority
1946 * taken into account by the scheduler. This value might
1947 * be boosted by RT tasks, or might be boosted by
1948 * interactivity modifiers. Will be RT if the task got
1949 * RT-boosted. If not then it returns p->normal_prio.
1951 static int effective_prio(struct task_struct *p)
1953 p->normal_prio = normal_prio(p);
1955 * If we are RT tasks or we were boosted to RT priority,
1956 * keep the priority unchanged. Otherwise, update priority
1957 * to the normal priority:
1959 if (!rt_prio(p->prio))
1960 return p->normal_prio;
1965 * task_curr - is this task currently executing on a CPU?
1966 * @p: the task in question.
1968 inline int task_curr(const struct task_struct *p)
1970 return cpu_curr(task_cpu(p)) == p;
1973 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1974 const struct sched_class *prev_class,
1975 int oldprio, int running)
1977 if (prev_class != p->sched_class) {
1978 if (prev_class->switched_from)
1979 prev_class->switched_from(rq, p, running);
1980 p->sched_class->switched_to(rq, p, running);
1982 p->sched_class->prio_changed(rq, p, oldprio, running);
1987 * Is this task likely cache-hot:
1990 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1994 if (p->sched_class != &fair_sched_class)
1998 * Buddy candidates are cache hot:
2000 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2001 (&p->se == cfs_rq_of(&p->se)->next ||
2002 &p->se == cfs_rq_of(&p->se)->last))
2005 if (sysctl_sched_migration_cost == -1)
2007 if (sysctl_sched_migration_cost == 0)
2010 delta = now - p->se.exec_start;
2012 return delta < (s64)sysctl_sched_migration_cost;
2015 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2017 #ifdef CONFIG_SCHED_DEBUG
2019 * We should never call set_task_cpu() on a blocked task,
2020 * ttwu() will sort out the placement.
2022 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2023 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2026 trace_sched_migrate_task(p, new_cpu);
2028 if (task_cpu(p) != new_cpu) {
2029 p->se.nr_migrations++;
2030 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2033 __set_task_cpu(p, new_cpu);
2036 struct migration_req {
2037 struct list_head list;
2039 struct task_struct *task;
2042 struct completion done;
2046 * The task's runqueue lock must be held.
2047 * Returns true if you have to wait for migration thread.
2050 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2052 struct rq *rq = task_rq(p);
2055 * If the task is not on a runqueue (and not running), then
2056 * the next wake-up will properly place the task.
2058 if (!p->se.on_rq && !task_running(rq, p))
2061 init_completion(&req->done);
2063 req->dest_cpu = dest_cpu;
2064 list_add(&req->list, &rq->migration_queue);
2070 * wait_task_context_switch - wait for a thread to complete at least one
2073 * @p must not be current.
2075 void wait_task_context_switch(struct task_struct *p)
2077 unsigned long nvcsw, nivcsw, flags;
2085 * The runqueue is assigned before the actual context
2086 * switch. We need to take the runqueue lock.
2088 * We could check initially without the lock but it is
2089 * very likely that we need to take the lock in every
2092 rq = task_rq_lock(p, &flags);
2093 running = task_running(rq, p);
2094 task_rq_unlock(rq, &flags);
2096 if (likely(!running))
2099 * The switch count is incremented before the actual
2100 * context switch. We thus wait for two switches to be
2101 * sure at least one completed.
2103 if ((p->nvcsw - nvcsw) > 1)
2105 if ((p->nivcsw - nivcsw) > 1)
2113 * wait_task_inactive - wait for a thread to unschedule.
2115 * If @match_state is nonzero, it's the @p->state value just checked and
2116 * not expected to change. If it changes, i.e. @p might have woken up,
2117 * then return zero. When we succeed in waiting for @p to be off its CPU,
2118 * we return a positive number (its total switch count). If a second call
2119 * a short while later returns the same number, the caller can be sure that
2120 * @p has remained unscheduled the whole time.
2122 * The caller must ensure that the task *will* unschedule sometime soon,
2123 * else this function might spin for a *long* time. This function can't
2124 * be called with interrupts off, or it may introduce deadlock with
2125 * smp_call_function() if an IPI is sent by the same process we are
2126 * waiting to become inactive.
2128 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2130 unsigned long flags;
2137 * We do the initial early heuristics without holding
2138 * any task-queue locks at all. We'll only try to get
2139 * the runqueue lock when things look like they will
2145 * If the task is actively running on another CPU
2146 * still, just relax and busy-wait without holding
2149 * NOTE! Since we don't hold any locks, it's not
2150 * even sure that "rq" stays as the right runqueue!
2151 * But we don't care, since "task_running()" will
2152 * return false if the runqueue has changed and p
2153 * is actually now running somewhere else!
2155 while (task_running(rq, p)) {
2156 if (match_state && unlikely(p->state != match_state))
2162 * Ok, time to look more closely! We need the rq
2163 * lock now, to be *sure*. If we're wrong, we'll
2164 * just go back and repeat.
2166 rq = task_rq_lock(p, &flags);
2167 trace_sched_wait_task(rq, p);
2168 running = task_running(rq, p);
2169 on_rq = p->se.on_rq;
2171 if (!match_state || p->state == match_state)
2172 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2173 task_rq_unlock(rq, &flags);
2176 * If it changed from the expected state, bail out now.
2178 if (unlikely(!ncsw))
2182 * Was it really running after all now that we
2183 * checked with the proper locks actually held?
2185 * Oops. Go back and try again..
2187 if (unlikely(running)) {
2193 * It's not enough that it's not actively running,
2194 * it must be off the runqueue _entirely_, and not
2197 * So if it was still runnable (but just not actively
2198 * running right now), it's preempted, and we should
2199 * yield - it could be a while.
2201 if (unlikely(on_rq)) {
2202 schedule_timeout_uninterruptible(1);
2207 * Ahh, all good. It wasn't running, and it wasn't
2208 * runnable, which means that it will never become
2209 * running in the future either. We're all done!
2218 * kick_process - kick a running thread to enter/exit the kernel
2219 * @p: the to-be-kicked thread
2221 * Cause a process which is running on another CPU to enter
2222 * kernel-mode, without any delay. (to get signals handled.)
2224 * NOTE: this function doesnt have to take the runqueue lock,
2225 * because all it wants to ensure is that the remote task enters
2226 * the kernel. If the IPI races and the task has been migrated
2227 * to another CPU then no harm is done and the purpose has been
2230 void kick_process(struct task_struct *p)
2236 if ((cpu != smp_processor_id()) && task_curr(p))
2237 smp_send_reschedule(cpu);
2240 EXPORT_SYMBOL_GPL(kick_process);
2241 #endif /* CONFIG_SMP */
2244 * task_oncpu_function_call - call a function on the cpu on which a task runs
2245 * @p: the task to evaluate
2246 * @func: the function to be called
2247 * @info: the function call argument
2249 * Calls the function @func when the task is currently running. This might
2250 * be on the current CPU, which just calls the function directly
2252 void task_oncpu_function_call(struct task_struct *p,
2253 void (*func) (void *info), void *info)
2260 smp_call_function_single(cpu, func, info, 1);
2265 static int select_fallback_rq(int cpu, struct task_struct *p)
2268 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2270 /* Look for allowed, online CPU in same node. */
2271 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2272 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2275 /* Any allowed, online CPU? */
2276 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2277 if (dest_cpu < nr_cpu_ids)
2280 /* No more Mr. Nice Guy. */
2281 if (dest_cpu >= nr_cpu_ids) {
2283 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2285 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2288 * Don't tell them about moving exiting tasks or
2289 * kernel threads (both mm NULL), since they never
2292 if (p->mm && printk_ratelimit()) {
2293 printk(KERN_INFO "process %d (%s) no "
2294 "longer affine to cpu%d\n",
2295 task_pid_nr(p), p->comm, cpu);
2305 * - fork, @p is stable because it isn't on the tasklist yet
2307 * - exec, @p is unstable, retry loop
2309 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so
2310 * we should be good.
2313 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2315 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2318 * In order not to call set_task_cpu() on a blocking task we need
2319 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2322 * Since this is common to all placement strategies, this lives here.
2324 * [ this allows ->select_task() to simply return task_cpu(p) and
2325 * not worry about this generic constraint ]
2327 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2329 cpu = select_fallback_rq(task_cpu(p), p);
2336 * try_to_wake_up - wake up a thread
2337 * @p: the to-be-woken-up thread
2338 * @state: the mask of task states that can be woken
2339 * @sync: do a synchronous wakeup?
2341 * Put it on the run-queue if it's not already there. The "current"
2342 * thread is always on the run-queue (except when the actual
2343 * re-schedule is in progress), and as such you're allowed to do
2344 * the simpler "current->state = TASK_RUNNING" to mark yourself
2345 * runnable without the overhead of this.
2347 * returns failure only if the task is already active.
2349 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2352 int cpu, orig_cpu, this_cpu, success = 0;
2353 unsigned long flags;
2354 struct rq *rq, *orig_rq;
2356 if (!sched_feat(SYNC_WAKEUPS))
2357 wake_flags &= ~WF_SYNC;
2359 this_cpu = get_cpu();
2362 rq = orig_rq = task_rq_lock(p, &flags);
2363 update_rq_clock(rq);
2364 if (!(p->state & state))
2374 if (unlikely(task_running(rq, p)))
2378 * In order to handle concurrent wakeups and release the rq->lock
2379 * we put the task in TASK_WAKING state.
2381 * First fix up the nr_uninterruptible count:
2383 if (task_contributes_to_load(p))
2384 rq->nr_uninterruptible--;
2385 p->state = TASK_WAKING;
2387 if (p->sched_class->task_waking)
2388 p->sched_class->task_waking(rq, p);
2390 __task_rq_unlock(rq);
2392 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2393 if (cpu != orig_cpu)
2394 set_task_cpu(p, cpu);
2396 rq = __task_rq_lock(p);
2397 update_rq_clock(rq);
2399 WARN_ON(p->state != TASK_WAKING);
2402 #ifdef CONFIG_SCHEDSTATS
2403 schedstat_inc(rq, ttwu_count);
2404 if (cpu == this_cpu)
2405 schedstat_inc(rq, ttwu_local);
2407 struct sched_domain *sd;
2408 for_each_domain(this_cpu, sd) {
2409 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2410 schedstat_inc(sd, ttwu_wake_remote);
2415 #endif /* CONFIG_SCHEDSTATS */
2418 #endif /* CONFIG_SMP */
2419 schedstat_inc(p, se.nr_wakeups);
2420 if (wake_flags & WF_SYNC)
2421 schedstat_inc(p, se.nr_wakeups_sync);
2422 if (orig_cpu != cpu)
2423 schedstat_inc(p, se.nr_wakeups_migrate);
2424 if (cpu == this_cpu)
2425 schedstat_inc(p, se.nr_wakeups_local);
2427 schedstat_inc(p, se.nr_wakeups_remote);
2428 activate_task(rq, p, 1);
2432 * Only attribute actual wakeups done by this task.
2434 if (!in_interrupt()) {
2435 struct sched_entity *se = ¤t->se;
2436 u64 sample = se->sum_exec_runtime;
2438 if (se->last_wakeup)
2439 sample -= se->last_wakeup;
2441 sample -= se->start_runtime;
2442 update_avg(&se->avg_wakeup, sample);
2444 se->last_wakeup = se->sum_exec_runtime;
2448 trace_sched_wakeup(rq, p, success);
2449 check_preempt_curr(rq, p, wake_flags);
2451 p->state = TASK_RUNNING;
2453 if (p->sched_class->task_woken)
2454 p->sched_class->task_woken(rq, p);
2456 if (unlikely(rq->idle_stamp)) {
2457 u64 delta = rq->clock - rq->idle_stamp;
2458 u64 max = 2*sysctl_sched_migration_cost;
2463 update_avg(&rq->avg_idle, delta);
2468 task_rq_unlock(rq, &flags);
2475 * wake_up_process - Wake up a specific process
2476 * @p: The process to be woken up.
2478 * Attempt to wake up the nominated process and move it to the set of runnable
2479 * processes. Returns 1 if the process was woken up, 0 if it was already
2482 * It may be assumed that this function implies a write memory barrier before
2483 * changing the task state if and only if any tasks are woken up.
2485 int wake_up_process(struct task_struct *p)
2487 return try_to_wake_up(p, TASK_ALL, 0);
2489 EXPORT_SYMBOL(wake_up_process);
2491 int wake_up_state(struct task_struct *p, unsigned int state)
2493 return try_to_wake_up(p, state, 0);
2497 * Perform scheduler related setup for a newly forked process p.
2498 * p is forked by current.
2500 * __sched_fork() is basic setup used by init_idle() too:
2502 static void __sched_fork(struct task_struct *p)
2504 p->se.exec_start = 0;
2505 p->se.sum_exec_runtime = 0;
2506 p->se.prev_sum_exec_runtime = 0;
2507 p->se.nr_migrations = 0;
2508 p->se.last_wakeup = 0;
2509 p->se.avg_overlap = 0;
2510 p->se.start_runtime = 0;
2511 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2513 #ifdef CONFIG_SCHEDSTATS
2514 p->se.wait_start = 0;
2516 p->se.wait_count = 0;
2519 p->se.sleep_start = 0;
2520 p->se.sleep_max = 0;
2521 p->se.sum_sleep_runtime = 0;
2523 p->se.block_start = 0;
2524 p->se.block_max = 0;
2526 p->se.slice_max = 0;
2528 p->se.nr_migrations_cold = 0;
2529 p->se.nr_failed_migrations_affine = 0;
2530 p->se.nr_failed_migrations_running = 0;
2531 p->se.nr_failed_migrations_hot = 0;
2532 p->se.nr_forced_migrations = 0;
2534 p->se.nr_wakeups = 0;
2535 p->se.nr_wakeups_sync = 0;
2536 p->se.nr_wakeups_migrate = 0;
2537 p->se.nr_wakeups_local = 0;
2538 p->se.nr_wakeups_remote = 0;
2539 p->se.nr_wakeups_affine = 0;
2540 p->se.nr_wakeups_affine_attempts = 0;
2541 p->se.nr_wakeups_passive = 0;
2542 p->se.nr_wakeups_idle = 0;
2546 INIT_LIST_HEAD(&p->rt.run_list);
2548 INIT_LIST_HEAD(&p->se.group_node);
2550 #ifdef CONFIG_PREEMPT_NOTIFIERS
2551 INIT_HLIST_HEAD(&p->preempt_notifiers);
2556 * fork()/clone()-time setup:
2558 void sched_fork(struct task_struct *p, int clone_flags)
2560 int cpu = get_cpu();
2564 * We mark the process as waking here. This guarantees that
2565 * nobody will actually run it, and a signal or other external
2566 * event cannot wake it up and insert it on the runqueue either.
2568 p->state = TASK_WAKING;
2571 * Revert to default priority/policy on fork if requested.
2573 if (unlikely(p->sched_reset_on_fork)) {
2574 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2575 p->policy = SCHED_NORMAL;
2576 p->normal_prio = p->static_prio;
2579 if (PRIO_TO_NICE(p->static_prio) < 0) {
2580 p->static_prio = NICE_TO_PRIO(0);
2581 p->normal_prio = p->static_prio;
2586 * We don't need the reset flag anymore after the fork. It has
2587 * fulfilled its duty:
2589 p->sched_reset_on_fork = 0;
2593 * Make sure we do not leak PI boosting priority to the child.
2595 p->prio = current->normal_prio;
2597 if (!rt_prio(p->prio))
2598 p->sched_class = &fair_sched_class;
2600 if (p->sched_class->task_fork)
2601 p->sched_class->task_fork(p);
2604 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2606 set_task_cpu(p, cpu);
2608 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2609 if (likely(sched_info_on()))
2610 memset(&p->sched_info, 0, sizeof(p->sched_info));
2612 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2615 #ifdef CONFIG_PREEMPT
2616 /* Want to start with kernel preemption disabled. */
2617 task_thread_info(p)->preempt_count = 1;
2619 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2625 * wake_up_new_task - wake up a newly created task for the first time.
2627 * This function will do some initial scheduler statistics housekeeping
2628 * that must be done for every newly created context, then puts the task
2629 * on the runqueue and wakes it.
2631 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2633 unsigned long flags;
2636 rq = task_rq_lock(p, &flags);
2637 BUG_ON(p->state != TASK_WAKING);
2638 p->state = TASK_RUNNING;
2639 update_rq_clock(rq);
2640 activate_task(rq, p, 0);
2641 trace_sched_wakeup_new(rq, p, 1);
2642 check_preempt_curr(rq, p, WF_FORK);
2644 if (p->sched_class->task_woken)
2645 p->sched_class->task_woken(rq, p);
2647 task_rq_unlock(rq, &flags);
2650 #ifdef CONFIG_PREEMPT_NOTIFIERS
2653 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2654 * @notifier: notifier struct to register
2656 void preempt_notifier_register(struct preempt_notifier *notifier)
2658 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2660 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2663 * preempt_notifier_unregister - no longer interested in preemption notifications
2664 * @notifier: notifier struct to unregister
2666 * This is safe to call from within a preemption notifier.
2668 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2670 hlist_del(¬ifier->link);
2672 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2674 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2676 struct preempt_notifier *notifier;
2677 struct hlist_node *node;
2679 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2680 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2684 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2685 struct task_struct *next)
2687 struct preempt_notifier *notifier;
2688 struct hlist_node *node;
2690 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2691 notifier->ops->sched_out(notifier, next);
2694 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2696 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2701 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2702 struct task_struct *next)
2706 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2709 * prepare_task_switch - prepare to switch tasks
2710 * @rq: the runqueue preparing to switch
2711 * @prev: the current task that is being switched out
2712 * @next: the task we are going to switch to.
2714 * This is called with the rq lock held and interrupts off. It must
2715 * be paired with a subsequent finish_task_switch after the context
2718 * prepare_task_switch sets up locking and calls architecture specific
2722 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2723 struct task_struct *next)
2725 fire_sched_out_preempt_notifiers(prev, next);
2726 prepare_lock_switch(rq, next);
2727 prepare_arch_switch(next);
2731 * finish_task_switch - clean up after a task-switch
2732 * @rq: runqueue associated with task-switch
2733 * @prev: the thread we just switched away from.
2735 * finish_task_switch must be called after the context switch, paired
2736 * with a prepare_task_switch call before the context switch.
2737 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2738 * and do any other architecture-specific cleanup actions.
2740 * Note that we may have delayed dropping an mm in context_switch(). If
2741 * so, we finish that here outside of the runqueue lock. (Doing it
2742 * with the lock held can cause deadlocks; see schedule() for
2745 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2746 __releases(rq->lock)
2748 struct mm_struct *mm = rq->prev_mm;
2754 * A task struct has one reference for the use as "current".
2755 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2756 * schedule one last time. The schedule call will never return, and
2757 * the scheduled task must drop that reference.
2758 * The test for TASK_DEAD must occur while the runqueue locks are
2759 * still held, otherwise prev could be scheduled on another cpu, die
2760 * there before we look at prev->state, and then the reference would
2762 * Manfred Spraul <manfred@colorfullife.com>
2764 prev_state = prev->state;
2765 finish_arch_switch(prev);
2766 perf_event_task_sched_in(current, cpu_of(rq));
2767 finish_lock_switch(rq, prev);
2769 fire_sched_in_preempt_notifiers(current);
2772 if (unlikely(prev_state == TASK_DEAD)) {
2774 * Remove function-return probe instances associated with this
2775 * task and put them back on the free list.
2777 kprobe_flush_task(prev);
2778 put_task_struct(prev);
2784 /* assumes rq->lock is held */
2785 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2787 if (prev->sched_class->pre_schedule)
2788 prev->sched_class->pre_schedule(rq, prev);
2791 /* rq->lock is NOT held, but preemption is disabled */
2792 static inline void post_schedule(struct rq *rq)
2794 if (rq->post_schedule) {
2795 unsigned long flags;
2797 raw_spin_lock_irqsave(&rq->lock, flags);
2798 if (rq->curr->sched_class->post_schedule)
2799 rq->curr->sched_class->post_schedule(rq);
2800 raw_spin_unlock_irqrestore(&rq->lock, flags);
2802 rq->post_schedule = 0;
2808 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2812 static inline void post_schedule(struct rq *rq)
2819 * schedule_tail - first thing a freshly forked thread must call.
2820 * @prev: the thread we just switched away from.
2822 asmlinkage void schedule_tail(struct task_struct *prev)
2823 __releases(rq->lock)
2825 struct rq *rq = this_rq();
2827 finish_task_switch(rq, prev);
2830 * FIXME: do we need to worry about rq being invalidated by the
2835 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2836 /* In this case, finish_task_switch does not reenable preemption */
2839 if (current->set_child_tid)
2840 put_user(task_pid_vnr(current), current->set_child_tid);
2844 * context_switch - switch to the new MM and the new
2845 * thread's register state.
2848 context_switch(struct rq *rq, struct task_struct *prev,
2849 struct task_struct *next)
2851 struct mm_struct *mm, *oldmm;
2853 prepare_task_switch(rq, prev, next);
2854 trace_sched_switch(rq, prev, next);
2856 oldmm = prev->active_mm;
2858 * For paravirt, this is coupled with an exit in switch_to to
2859 * combine the page table reload and the switch backend into
2862 arch_start_context_switch(prev);
2865 next->active_mm = oldmm;
2866 atomic_inc(&oldmm->mm_count);
2867 enter_lazy_tlb(oldmm, next);
2869 switch_mm(oldmm, mm, next);
2871 if (likely(!prev->mm)) {
2872 prev->active_mm = NULL;
2873 rq->prev_mm = oldmm;
2876 * Since the runqueue lock will be released by the next
2877 * task (which is an invalid locking op but in the case
2878 * of the scheduler it's an obvious special-case), so we
2879 * do an early lockdep release here:
2881 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2882 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2885 /* Here we just switch the register state and the stack. */
2886 switch_to(prev, next, prev);
2890 * this_rq must be evaluated again because prev may have moved
2891 * CPUs since it called schedule(), thus the 'rq' on its stack
2892 * frame will be invalid.
2894 finish_task_switch(this_rq(), prev);
2898 * nr_running, nr_uninterruptible and nr_context_switches:
2900 * externally visible scheduler statistics: current number of runnable
2901 * threads, current number of uninterruptible-sleeping threads, total
2902 * number of context switches performed since bootup.
2904 unsigned long nr_running(void)
2906 unsigned long i, sum = 0;
2908 for_each_online_cpu(i)
2909 sum += cpu_rq(i)->nr_running;
2914 unsigned long nr_uninterruptible(void)
2916 unsigned long i, sum = 0;
2918 for_each_possible_cpu(i)
2919 sum += cpu_rq(i)->nr_uninterruptible;
2922 * Since we read the counters lockless, it might be slightly
2923 * inaccurate. Do not allow it to go below zero though:
2925 if (unlikely((long)sum < 0))
2931 unsigned long long nr_context_switches(void)
2934 unsigned long long sum = 0;
2936 for_each_possible_cpu(i)
2937 sum += cpu_rq(i)->nr_switches;
2942 unsigned long nr_iowait(void)
2944 unsigned long i, sum = 0;
2946 for_each_possible_cpu(i)
2947 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2952 unsigned long nr_iowait_cpu(void)
2954 struct rq *this = this_rq();
2955 return atomic_read(&this->nr_iowait);
2958 unsigned long this_cpu_load(void)
2960 struct rq *this = this_rq();
2961 return this->cpu_load[0];
2965 /* Variables and functions for calc_load */
2966 static atomic_long_t calc_load_tasks;
2967 static unsigned long calc_load_update;
2968 unsigned long avenrun[3];
2969 EXPORT_SYMBOL(avenrun);
2972 * get_avenrun - get the load average array
2973 * @loads: pointer to dest load array
2974 * @offset: offset to add
2975 * @shift: shift count to shift the result left
2977 * These values are estimates at best, so no need for locking.
2979 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2981 loads[0] = (avenrun[0] + offset) << shift;
2982 loads[1] = (avenrun[1] + offset) << shift;
2983 loads[2] = (avenrun[2] + offset) << shift;
2986 static unsigned long
2987 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2990 load += active * (FIXED_1 - exp);
2991 return load >> FSHIFT;
2995 * calc_load - update the avenrun load estimates 10 ticks after the
2996 * CPUs have updated calc_load_tasks.
2998 void calc_global_load(void)
3000 unsigned long upd = calc_load_update + 10;
3003 if (time_before(jiffies, upd))
3006 active = atomic_long_read(&calc_load_tasks);
3007 active = active > 0 ? active * FIXED_1 : 0;
3009 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3010 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3011 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3013 calc_load_update += LOAD_FREQ;
3017 * Either called from update_cpu_load() or from a cpu going idle
3019 static void calc_load_account_active(struct rq *this_rq)
3021 long nr_active, delta;
3023 nr_active = this_rq->nr_running;
3024 nr_active += (long) this_rq->nr_uninterruptible;
3026 if (nr_active != this_rq->calc_load_active) {
3027 delta = nr_active - this_rq->calc_load_active;
3028 this_rq->calc_load_active = nr_active;
3029 atomic_long_add(delta, &calc_load_tasks);
3034 * Update rq->cpu_load[] statistics. This function is usually called every
3035 * scheduler tick (TICK_NSEC).
3037 static void update_cpu_load(struct rq *this_rq)
3039 unsigned long this_load = this_rq->load.weight;
3042 this_rq->nr_load_updates++;
3044 /* Update our load: */
3045 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3046 unsigned long old_load, new_load;
3048 /* scale is effectively 1 << i now, and >> i divides by scale */
3050 old_load = this_rq->cpu_load[i];
3051 new_load = this_load;
3053 * Round up the averaging division if load is increasing. This
3054 * prevents us from getting stuck on 9 if the load is 10, for
3057 if (new_load > old_load)
3058 new_load += scale-1;
3059 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3062 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3063 this_rq->calc_load_update += LOAD_FREQ;
3064 calc_load_account_active(this_rq);
3071 * sched_exec - execve() is a valuable balancing opportunity, because at
3072 * this point the task has the smallest effective memory and cache footprint.
3074 void sched_exec(void)
3076 struct task_struct *p = current;
3077 struct migration_req req;
3078 int dest_cpu, this_cpu;
3079 unsigned long flags;
3083 this_cpu = get_cpu();
3084 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3085 if (dest_cpu == this_cpu) {
3090 rq = task_rq_lock(p, &flags);
3094 * select_task_rq() can race against ->cpus_allowed
3096 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3097 || unlikely(!cpu_active(dest_cpu))) {
3098 task_rq_unlock(rq, &flags);
3102 /* force the process onto the specified CPU */
3103 if (migrate_task(p, dest_cpu, &req)) {
3104 /* Need to wait for migration thread (might exit: take ref). */
3105 struct task_struct *mt = rq->migration_thread;
3107 get_task_struct(mt);
3108 task_rq_unlock(rq, &flags);
3109 wake_up_process(mt);
3110 put_task_struct(mt);
3111 wait_for_completion(&req.done);
3115 task_rq_unlock(rq, &flags);
3120 DEFINE_PER_CPU(struct kernel_stat, kstat);
3122 EXPORT_PER_CPU_SYMBOL(kstat);
3125 * Return any ns on the sched_clock that have not yet been accounted in
3126 * @p in case that task is currently running.
3128 * Called with task_rq_lock() held on @rq.
3130 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3134 if (task_current(rq, p)) {
3135 update_rq_clock(rq);
3136 ns = rq->clock - p->se.exec_start;
3144 unsigned long long task_delta_exec(struct task_struct *p)
3146 unsigned long flags;
3150 rq = task_rq_lock(p, &flags);
3151 ns = do_task_delta_exec(p, rq);
3152 task_rq_unlock(rq, &flags);
3158 * Return accounted runtime for the task.
3159 * In case the task is currently running, return the runtime plus current's
3160 * pending runtime that have not been accounted yet.
3162 unsigned long long task_sched_runtime(struct task_struct *p)
3164 unsigned long flags;
3168 rq = task_rq_lock(p, &flags);
3169 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3170 task_rq_unlock(rq, &flags);
3176 * Return sum_exec_runtime for the thread group.
3177 * In case the task is currently running, return the sum plus current's
3178 * pending runtime that have not been accounted yet.
3180 * Note that the thread group might have other running tasks as well,
3181 * so the return value not includes other pending runtime that other
3182 * running tasks might have.
3184 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3186 struct task_cputime totals;
3187 unsigned long flags;
3191 rq = task_rq_lock(p, &flags);
3192 thread_group_cputime(p, &totals);
3193 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3194 task_rq_unlock(rq, &flags);
3200 * Account user cpu time to a process.
3201 * @p: the process that the cpu time gets accounted to
3202 * @cputime: the cpu time spent in user space since the last update
3203 * @cputime_scaled: cputime scaled by cpu frequency
3205 void account_user_time(struct task_struct *p, cputime_t cputime,
3206 cputime_t cputime_scaled)
3208 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3211 /* Add user time to process. */
3212 p->utime = cputime_add(p->utime, cputime);
3213 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3214 account_group_user_time(p, cputime);
3216 /* Add user time to cpustat. */
3217 tmp = cputime_to_cputime64(cputime);
3218 if (TASK_NICE(p) > 0)
3219 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3221 cpustat->user = cputime64_add(cpustat->user, tmp);
3223 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3224 /* Account for user time used */
3225 acct_update_integrals(p);
3229 * Account guest cpu time to a process.
3230 * @p: the process that the cpu time gets accounted to
3231 * @cputime: the cpu time spent in virtual machine since the last update
3232 * @cputime_scaled: cputime scaled by cpu frequency
3234 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3235 cputime_t cputime_scaled)
3238 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3240 tmp = cputime_to_cputime64(cputime);
3242 /* Add guest time to process. */
3243 p->utime = cputime_add(p->utime, cputime);
3244 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3245 account_group_user_time(p, cputime);
3246 p->gtime = cputime_add(p->gtime, cputime);
3248 /* Add guest time to cpustat. */
3249 if (TASK_NICE(p) > 0) {
3250 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3251 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3253 cpustat->user = cputime64_add(cpustat->user, tmp);
3254 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3259 * Account system cpu time to a process.
3260 * @p: the process that the cpu time gets accounted to
3261 * @hardirq_offset: the offset to subtract from hardirq_count()
3262 * @cputime: the cpu time spent in kernel space since the last update
3263 * @cputime_scaled: cputime scaled by cpu frequency
3265 void account_system_time(struct task_struct *p, int hardirq_offset,
3266 cputime_t cputime, cputime_t cputime_scaled)
3268 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3271 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3272 account_guest_time(p, cputime, cputime_scaled);
3276 /* Add system time to process. */
3277 p->stime = cputime_add(p->stime, cputime);
3278 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3279 account_group_system_time(p, cputime);
3281 /* Add system time to cpustat. */
3282 tmp = cputime_to_cputime64(cputime);
3283 if (hardirq_count() - hardirq_offset)
3284 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3285 else if (softirq_count())
3286 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3288 cpustat->system = cputime64_add(cpustat->system, tmp);
3290 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3292 /* Account for system time used */
3293 acct_update_integrals(p);
3297 * Account for involuntary wait time.
3298 * @steal: the cpu time spent in involuntary wait
3300 void account_steal_time(cputime_t cputime)
3302 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3303 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3305 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3309 * Account for idle time.
3310 * @cputime: the cpu time spent in idle wait
3312 void account_idle_time(cputime_t cputime)
3314 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3315 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3316 struct rq *rq = this_rq();
3318 if (atomic_read(&rq->nr_iowait) > 0)
3319 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3321 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3324 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3327 * Account a single tick of cpu time.
3328 * @p: the process that the cpu time gets accounted to
3329 * @user_tick: indicates if the tick is a user or a system tick
3331 void account_process_tick(struct task_struct *p, int user_tick)
3333 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3334 struct rq *rq = this_rq();
3337 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3338 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3339 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3342 account_idle_time(cputime_one_jiffy);
3346 * Account multiple ticks of steal time.
3347 * @p: the process from which the cpu time has been stolen
3348 * @ticks: number of stolen ticks
3350 void account_steal_ticks(unsigned long ticks)
3352 account_steal_time(jiffies_to_cputime(ticks));
3356 * Account multiple ticks of idle time.
3357 * @ticks: number of stolen ticks
3359 void account_idle_ticks(unsigned long ticks)
3361 account_idle_time(jiffies_to_cputime(ticks));
3367 * Use precise platform statistics if available:
3369 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3370 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3376 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3378 struct task_cputime cputime;
3380 thread_group_cputime(p, &cputime);
3382 *ut = cputime.utime;
3383 *st = cputime.stime;
3387 #ifndef nsecs_to_cputime
3388 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3391 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3393 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3396 * Use CFS's precise accounting:
3398 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3403 temp = (u64)(rtime * utime);
3404 do_div(temp, total);
3405 utime = (cputime_t)temp;
3410 * Compare with previous values, to keep monotonicity:
3412 p->prev_utime = max(p->prev_utime, utime);
3413 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3415 *ut = p->prev_utime;
3416 *st = p->prev_stime;
3420 * Must be called with siglock held.
3422 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3424 struct signal_struct *sig = p->signal;
3425 struct task_cputime cputime;
3426 cputime_t rtime, utime, total;
3428 thread_group_cputime(p, &cputime);
3430 total = cputime_add(cputime.utime, cputime.stime);
3431 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3436 temp = (u64)(rtime * cputime.utime);
3437 do_div(temp, total);
3438 utime = (cputime_t)temp;
3442 sig->prev_utime = max(sig->prev_utime, utime);
3443 sig->prev_stime = max(sig->prev_stime,
3444 cputime_sub(rtime, sig->prev_utime));
3446 *ut = sig->prev_utime;
3447 *st = sig->prev_stime;
3452 * This function gets called by the timer code, with HZ frequency.
3453 * We call it with interrupts disabled.
3455 * It also gets called by the fork code, when changing the parent's
3458 void scheduler_tick(void)
3460 int cpu = smp_processor_id();
3461 struct rq *rq = cpu_rq(cpu);
3462 struct task_struct *curr = rq->curr;
3466 raw_spin_lock(&rq->lock);
3467 update_rq_clock(rq);
3468 update_cpu_load(rq);
3469 curr->sched_class->task_tick(rq, curr, 0);
3470 raw_spin_unlock(&rq->lock);
3472 perf_event_task_tick(curr, cpu);
3475 rq->idle_at_tick = idle_cpu(cpu);
3476 trigger_load_balance(rq, cpu);
3480 notrace unsigned long get_parent_ip(unsigned long addr)
3482 if (in_lock_functions(addr)) {
3483 addr = CALLER_ADDR2;
3484 if (in_lock_functions(addr))
3485 addr = CALLER_ADDR3;
3490 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3491 defined(CONFIG_PREEMPT_TRACER))
3493 void __kprobes add_preempt_count(int val)
3495 #ifdef CONFIG_DEBUG_PREEMPT
3499 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3502 preempt_count() += val;
3503 #ifdef CONFIG_DEBUG_PREEMPT
3505 * Spinlock count overflowing soon?
3507 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3510 if (preempt_count() == val)
3511 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3513 EXPORT_SYMBOL(add_preempt_count);
3515 void __kprobes sub_preempt_count(int val)
3517 #ifdef CONFIG_DEBUG_PREEMPT
3521 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3524 * Is the spinlock portion underflowing?
3526 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3527 !(preempt_count() & PREEMPT_MASK)))
3531 if (preempt_count() == val)
3532 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3533 preempt_count() -= val;
3535 EXPORT_SYMBOL(sub_preempt_count);
3540 * Print scheduling while atomic bug:
3542 static noinline void __schedule_bug(struct task_struct *prev)
3544 struct pt_regs *regs = get_irq_regs();
3546 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3547 prev->comm, prev->pid, preempt_count());
3549 debug_show_held_locks(prev);
3551 if (irqs_disabled())
3552 print_irqtrace_events(prev);
3561 * Various schedule()-time debugging checks and statistics:
3563 static inline void schedule_debug(struct task_struct *prev)
3566 * Test if we are atomic. Since do_exit() needs to call into
3567 * schedule() atomically, we ignore that path for now.
3568 * Otherwise, whine if we are scheduling when we should not be.
3570 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3571 __schedule_bug(prev);
3573 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3575 schedstat_inc(this_rq(), sched_count);
3576 #ifdef CONFIG_SCHEDSTATS
3577 if (unlikely(prev->lock_depth >= 0)) {
3578 schedstat_inc(this_rq(), bkl_count);
3579 schedstat_inc(prev, sched_info.bkl_count);
3584 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3586 if (prev->state == TASK_RUNNING) {
3587 u64 runtime = prev->se.sum_exec_runtime;
3589 runtime -= prev->se.prev_sum_exec_runtime;
3590 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
3593 * In order to avoid avg_overlap growing stale when we are
3594 * indeed overlapping and hence not getting put to sleep, grow
3595 * the avg_overlap on preemption.
3597 * We use the average preemption runtime because that
3598 * correlates to the amount of cache footprint a task can
3601 update_avg(&prev->se.avg_overlap, runtime);
3603 prev->sched_class->put_prev_task(rq, prev);
3607 * Pick up the highest-prio task:
3609 static inline struct task_struct *
3610 pick_next_task(struct rq *rq)
3612 const struct sched_class *class;
3613 struct task_struct *p;
3616 * Optimization: we know that if all tasks are in
3617 * the fair class we can call that function directly:
3619 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3620 p = fair_sched_class.pick_next_task(rq);
3625 class = sched_class_highest;
3627 p = class->pick_next_task(rq);
3631 * Will never be NULL as the idle class always
3632 * returns a non-NULL p:
3634 class = class->next;
3639 * schedule() is the main scheduler function.
3641 asmlinkage void __sched schedule(void)
3643 struct task_struct *prev, *next;
3644 unsigned long *switch_count;
3650 cpu = smp_processor_id();
3654 switch_count = &prev->nivcsw;
3656 release_kernel_lock(prev);
3657 need_resched_nonpreemptible:
3659 schedule_debug(prev);
3661 if (sched_feat(HRTICK))
3664 raw_spin_lock_irq(&rq->lock);
3665 update_rq_clock(rq);
3666 clear_tsk_need_resched(prev);
3668 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3669 if (unlikely(signal_pending_state(prev->state, prev)))
3670 prev->state = TASK_RUNNING;
3672 deactivate_task(rq, prev, 1);
3673 switch_count = &prev->nvcsw;
3676 pre_schedule(rq, prev);
3678 if (unlikely(!rq->nr_running))
3679 idle_balance(cpu, rq);
3681 put_prev_task(rq, prev);
3682 next = pick_next_task(rq);
3684 if (likely(prev != next)) {
3685 sched_info_switch(prev, next);
3686 perf_event_task_sched_out(prev, next, cpu);
3692 context_switch(rq, prev, next); /* unlocks the rq */
3694 * the context switch might have flipped the stack from under
3695 * us, hence refresh the local variables.
3697 cpu = smp_processor_id();
3700 raw_spin_unlock_irq(&rq->lock);
3704 if (unlikely(reacquire_kernel_lock(current) < 0))
3705 goto need_resched_nonpreemptible;
3707 preempt_enable_no_resched();
3711 EXPORT_SYMBOL(schedule);
3713 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3715 * Look out! "owner" is an entirely speculative pointer
3716 * access and not reliable.
3718 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3723 if (!sched_feat(OWNER_SPIN))
3726 #ifdef CONFIG_DEBUG_PAGEALLOC
3728 * Need to access the cpu field knowing that
3729 * DEBUG_PAGEALLOC could have unmapped it if
3730 * the mutex owner just released it and exited.
3732 if (probe_kernel_address(&owner->cpu, cpu))
3739 * Even if the access succeeded (likely case),
3740 * the cpu field may no longer be valid.
3742 if (cpu >= nr_cpumask_bits)
3746 * We need to validate that we can do a
3747 * get_cpu() and that we have the percpu area.
3749 if (!cpu_online(cpu))
3756 * Owner changed, break to re-assess state.
3758 if (lock->owner != owner)
3762 * Is that owner really running on that cpu?
3764 if (task_thread_info(rq->curr) != owner || need_resched())
3774 #ifdef CONFIG_PREEMPT
3776 * this is the entry point to schedule() from in-kernel preemption
3777 * off of preempt_enable. Kernel preemptions off return from interrupt
3778 * occur there and call schedule directly.
3780 asmlinkage void __sched preempt_schedule(void)
3782 struct thread_info *ti = current_thread_info();
3785 * If there is a non-zero preempt_count or interrupts are disabled,
3786 * we do not want to preempt the current task. Just return..
3788 if (likely(ti->preempt_count || irqs_disabled()))
3792 add_preempt_count(PREEMPT_ACTIVE);
3794 sub_preempt_count(PREEMPT_ACTIVE);
3797 * Check again in case we missed a preemption opportunity
3798 * between schedule and now.
3801 } while (need_resched());
3803 EXPORT_SYMBOL(preempt_schedule);
3806 * this is the entry point to schedule() from kernel preemption
3807 * off of irq context.
3808 * Note, that this is called and return with irqs disabled. This will
3809 * protect us against recursive calling from irq.
3811 asmlinkage void __sched preempt_schedule_irq(void)
3813 struct thread_info *ti = current_thread_info();
3815 /* Catch callers which need to be fixed */
3816 BUG_ON(ti->preempt_count || !irqs_disabled());
3819 add_preempt_count(PREEMPT_ACTIVE);
3822 local_irq_disable();
3823 sub_preempt_count(PREEMPT_ACTIVE);
3826 * Check again in case we missed a preemption opportunity
3827 * between schedule and now.
3830 } while (need_resched());
3833 #endif /* CONFIG_PREEMPT */
3835 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3838 return try_to_wake_up(curr->private, mode, wake_flags);
3840 EXPORT_SYMBOL(default_wake_function);
3843 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3844 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3845 * number) then we wake all the non-exclusive tasks and one exclusive task.
3847 * There are circumstances in which we can try to wake a task which has already
3848 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3849 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3851 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3852 int nr_exclusive, int wake_flags, void *key)
3854 wait_queue_t *curr, *next;
3856 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3857 unsigned flags = curr->flags;
3859 if (curr->func(curr, mode, wake_flags, key) &&
3860 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3866 * __wake_up - wake up threads blocked on a waitqueue.
3868 * @mode: which threads
3869 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3870 * @key: is directly passed to the wakeup function
3872 * It may be assumed that this function implies a write memory barrier before
3873 * changing the task state if and only if any tasks are woken up.
3875 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3876 int nr_exclusive, void *key)
3878 unsigned long flags;
3880 spin_lock_irqsave(&q->lock, flags);
3881 __wake_up_common(q, mode, nr_exclusive, 0, key);
3882 spin_unlock_irqrestore(&q->lock, flags);
3884 EXPORT_SYMBOL(__wake_up);
3887 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3889 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3891 __wake_up_common(q, mode, 1, 0, NULL);
3894 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3896 __wake_up_common(q, mode, 1, 0, key);
3900 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3902 * @mode: which threads
3903 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3904 * @key: opaque value to be passed to wakeup targets
3906 * The sync wakeup differs that the waker knows that it will schedule
3907 * away soon, so while the target thread will be woken up, it will not
3908 * be migrated to another CPU - ie. the two threads are 'synchronized'
3909 * with each other. This can prevent needless bouncing between CPUs.
3911 * On UP it can prevent extra preemption.
3913 * It may be assumed that this function implies a write memory barrier before
3914 * changing the task state if and only if any tasks are woken up.
3916 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3917 int nr_exclusive, void *key)
3919 unsigned long flags;
3920 int wake_flags = WF_SYNC;
3925 if (unlikely(!nr_exclusive))
3928 spin_lock_irqsave(&q->lock, flags);
3929 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3930 spin_unlock_irqrestore(&q->lock, flags);
3932 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3935 * __wake_up_sync - see __wake_up_sync_key()
3937 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3939 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3941 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3944 * complete: - signals a single thread waiting on this completion
3945 * @x: holds the state of this particular completion
3947 * This will wake up a single thread waiting on this completion. Threads will be
3948 * awakened in the same order in which they were queued.
3950 * See also complete_all(), wait_for_completion() and related routines.
3952 * It may be assumed that this function implies a write memory barrier before
3953 * changing the task state if and only if any tasks are woken up.
3955 void complete(struct completion *x)
3957 unsigned long flags;
3959 spin_lock_irqsave(&x->wait.lock, flags);
3961 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3962 spin_unlock_irqrestore(&x->wait.lock, flags);
3964 EXPORT_SYMBOL(complete);
3967 * complete_all: - signals all threads waiting on this completion
3968 * @x: holds the state of this particular completion
3970 * This will wake up all threads waiting on this particular completion event.
3972 * It may be assumed that this function implies a write memory barrier before
3973 * changing the task state if and only if any tasks are woken up.
3975 void complete_all(struct completion *x)
3977 unsigned long flags;
3979 spin_lock_irqsave(&x->wait.lock, flags);
3980 x->done += UINT_MAX/2;
3981 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3982 spin_unlock_irqrestore(&x->wait.lock, flags);
3984 EXPORT_SYMBOL(complete_all);
3986 static inline long __sched
3987 do_wait_for_common(struct completion *x, long timeout, int state)
3990 DECLARE_WAITQUEUE(wait, current);
3992 wait.flags |= WQ_FLAG_EXCLUSIVE;
3993 __add_wait_queue_tail(&x->wait, &wait);
3995 if (signal_pending_state(state, current)) {
3996 timeout = -ERESTARTSYS;
3999 __set_current_state(state);
4000 spin_unlock_irq(&x->wait.lock);
4001 timeout = schedule_timeout(timeout);
4002 spin_lock_irq(&x->wait.lock);
4003 } while (!x->done && timeout);
4004 __remove_wait_queue(&x->wait, &wait);
4009 return timeout ?: 1;
4013 wait_for_common(struct completion *x, long timeout, int state)
4017 spin_lock_irq(&x->wait.lock);
4018 timeout = do_wait_for_common(x, timeout, state);
4019 spin_unlock_irq(&x->wait.lock);
4024 * wait_for_completion: - waits for completion of a task
4025 * @x: holds the state of this particular completion
4027 * This waits to be signaled for completion of a specific task. It is NOT
4028 * interruptible and there is no timeout.
4030 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4031 * and interrupt capability. Also see complete().
4033 void __sched wait_for_completion(struct completion *x)
4035 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4037 EXPORT_SYMBOL(wait_for_completion);
4040 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4041 * @x: holds the state of this particular completion
4042 * @timeout: timeout value in jiffies
4044 * This waits for either a completion of a specific task to be signaled or for a
4045 * specified timeout to expire. The timeout is in jiffies. It is not
4048 unsigned long __sched
4049 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4051 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4053 EXPORT_SYMBOL(wait_for_completion_timeout);
4056 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4057 * @x: holds the state of this particular completion
4059 * This waits for completion of a specific task to be signaled. It is
4062 int __sched wait_for_completion_interruptible(struct completion *x)
4064 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4065 if (t == -ERESTARTSYS)
4069 EXPORT_SYMBOL(wait_for_completion_interruptible);
4072 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4073 * @x: holds the state of this particular completion
4074 * @timeout: timeout value in jiffies
4076 * This waits for either a completion of a specific task to be signaled or for a
4077 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4079 unsigned long __sched
4080 wait_for_completion_interruptible_timeout(struct completion *x,
4081 unsigned long timeout)
4083 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4085 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4088 * wait_for_completion_killable: - waits for completion of a task (killable)
4089 * @x: holds the state of this particular completion
4091 * This waits to be signaled for completion of a specific task. It can be
4092 * interrupted by a kill signal.
4094 int __sched wait_for_completion_killable(struct completion *x)
4096 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4097 if (t == -ERESTARTSYS)
4101 EXPORT_SYMBOL(wait_for_completion_killable);
4104 * try_wait_for_completion - try to decrement a completion without blocking
4105 * @x: completion structure
4107 * Returns: 0 if a decrement cannot be done without blocking
4108 * 1 if a decrement succeeded.
4110 * If a completion is being used as a counting completion,
4111 * attempt to decrement the counter without blocking. This
4112 * enables us to avoid waiting if the resource the completion
4113 * is protecting is not available.
4115 bool try_wait_for_completion(struct completion *x)
4117 unsigned long flags;
4120 spin_lock_irqsave(&x->wait.lock, flags);
4125 spin_unlock_irqrestore(&x->wait.lock, flags);
4128 EXPORT_SYMBOL(try_wait_for_completion);
4131 * completion_done - Test to see if a completion has any waiters
4132 * @x: completion structure
4134 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4135 * 1 if there are no waiters.
4138 bool completion_done(struct completion *x)
4140 unsigned long flags;
4143 spin_lock_irqsave(&x->wait.lock, flags);
4146 spin_unlock_irqrestore(&x->wait.lock, flags);
4149 EXPORT_SYMBOL(completion_done);
4152 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4154 unsigned long flags;
4157 init_waitqueue_entry(&wait, current);
4159 __set_current_state(state);
4161 spin_lock_irqsave(&q->lock, flags);
4162 __add_wait_queue(q, &wait);
4163 spin_unlock(&q->lock);
4164 timeout = schedule_timeout(timeout);
4165 spin_lock_irq(&q->lock);
4166 __remove_wait_queue(q, &wait);
4167 spin_unlock_irqrestore(&q->lock, flags);
4172 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4174 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4176 EXPORT_SYMBOL(interruptible_sleep_on);
4179 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4181 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4183 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4185 void __sched sleep_on(wait_queue_head_t *q)
4187 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4189 EXPORT_SYMBOL(sleep_on);
4191 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4193 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4195 EXPORT_SYMBOL(sleep_on_timeout);
4197 #ifdef CONFIG_RT_MUTEXES
4200 * rt_mutex_setprio - set the current priority of a task
4202 * @prio: prio value (kernel-internal form)
4204 * This function changes the 'effective' priority of a task. It does
4205 * not touch ->normal_prio like __setscheduler().
4207 * Used by the rt_mutex code to implement priority inheritance logic.
4209 void rt_mutex_setprio(struct task_struct *p, int prio)
4211 unsigned long flags;
4212 int oldprio, on_rq, running;
4214 const struct sched_class *prev_class = p->sched_class;
4216 BUG_ON(prio < 0 || prio > MAX_PRIO);
4218 rq = task_rq_lock(p, &flags);
4219 update_rq_clock(rq);
4222 on_rq = p->se.on_rq;
4223 running = task_current(rq, p);
4225 dequeue_task(rq, p, 0);
4227 p->sched_class->put_prev_task(rq, p);
4230 p->sched_class = &rt_sched_class;
4232 p->sched_class = &fair_sched_class;
4237 p->sched_class->set_curr_task(rq);
4239 enqueue_task(rq, p, 0);
4241 check_class_changed(rq, p, prev_class, oldprio, running);
4243 task_rq_unlock(rq, &flags);
4248 void set_user_nice(struct task_struct *p, long nice)
4250 int old_prio, delta, on_rq;
4251 unsigned long flags;
4254 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4257 * We have to be careful, if called from sys_setpriority(),
4258 * the task might be in the middle of scheduling on another CPU.
4260 rq = task_rq_lock(p, &flags);
4261 update_rq_clock(rq);
4263 * The RT priorities are set via sched_setscheduler(), but we still
4264 * allow the 'normal' nice value to be set - but as expected
4265 * it wont have any effect on scheduling until the task is
4266 * SCHED_FIFO/SCHED_RR:
4268 if (task_has_rt_policy(p)) {
4269 p->static_prio = NICE_TO_PRIO(nice);
4272 on_rq = p->se.on_rq;
4274 dequeue_task(rq, p, 0);
4276 p->static_prio = NICE_TO_PRIO(nice);
4279 p->prio = effective_prio(p);
4280 delta = p->prio - old_prio;
4283 enqueue_task(rq, p, 0);
4285 * If the task increased its priority or is running and
4286 * lowered its priority, then reschedule its CPU:
4288 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4289 resched_task(rq->curr);
4292 task_rq_unlock(rq, &flags);
4294 EXPORT_SYMBOL(set_user_nice);
4297 * can_nice - check if a task can reduce its nice value
4301 int can_nice(const struct task_struct *p, const int nice)
4303 /* convert nice value [19,-20] to rlimit style value [1,40] */
4304 int nice_rlim = 20 - nice;
4306 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4307 capable(CAP_SYS_NICE));
4310 #ifdef __ARCH_WANT_SYS_NICE
4313 * sys_nice - change the priority of the current process.
4314 * @increment: priority increment
4316 * sys_setpriority is a more generic, but much slower function that
4317 * does similar things.
4319 SYSCALL_DEFINE1(nice, int, increment)
4324 * Setpriority might change our priority at the same moment.
4325 * We don't have to worry. Conceptually one call occurs first
4326 * and we have a single winner.
4328 if (increment < -40)
4333 nice = TASK_NICE(current) + increment;
4339 if (increment < 0 && !can_nice(current, nice))
4342 retval = security_task_setnice(current, nice);
4346 set_user_nice(current, nice);
4353 * task_prio - return the priority value of a given task.
4354 * @p: the task in question.
4356 * This is the priority value as seen by users in /proc.
4357 * RT tasks are offset by -200. Normal tasks are centered
4358 * around 0, value goes from -16 to +15.
4360 int task_prio(const struct task_struct *p)
4362 return p->prio - MAX_RT_PRIO;
4366 * task_nice - return the nice value of a given task.
4367 * @p: the task in question.
4369 int task_nice(const struct task_struct *p)
4371 return TASK_NICE(p);
4373 EXPORT_SYMBOL(task_nice);
4376 * idle_cpu - is a given cpu idle currently?
4377 * @cpu: the processor in question.
4379 int idle_cpu(int cpu)
4381 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4385 * idle_task - return the idle task for a given cpu.
4386 * @cpu: the processor in question.
4388 struct task_struct *idle_task(int cpu)
4390 return cpu_rq(cpu)->idle;
4394 * find_process_by_pid - find a process with a matching PID value.
4395 * @pid: the pid in question.
4397 static struct task_struct *find_process_by_pid(pid_t pid)
4399 return pid ? find_task_by_vpid(pid) : current;
4402 /* Actually do priority change: must hold rq lock. */
4404 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4406 BUG_ON(p->se.on_rq);
4409 p->rt_priority = prio;
4410 p->normal_prio = normal_prio(p);
4411 /* we are holding p->pi_lock already */
4412 p->prio = rt_mutex_getprio(p);
4413 if (rt_prio(p->prio))
4414 p->sched_class = &rt_sched_class;
4416 p->sched_class = &fair_sched_class;
4421 * check the target process has a UID that matches the current process's
4423 static bool check_same_owner(struct task_struct *p)
4425 const struct cred *cred = current_cred(), *pcred;
4429 pcred = __task_cred(p);
4430 match = (cred->euid == pcred->euid ||
4431 cred->euid == pcred->uid);
4436 static int __sched_setscheduler(struct task_struct *p, int policy,
4437 struct sched_param *param, bool user)
4439 int retval, oldprio, oldpolicy = -1, on_rq, running;
4440 unsigned long flags;
4441 const struct sched_class *prev_class = p->sched_class;
4445 /* may grab non-irq protected spin_locks */
4446 BUG_ON(in_interrupt());
4448 /* double check policy once rq lock held */
4450 reset_on_fork = p->sched_reset_on_fork;
4451 policy = oldpolicy = p->policy;
4453 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4454 policy &= ~SCHED_RESET_ON_FORK;
4456 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4457 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4458 policy != SCHED_IDLE)
4463 * Valid priorities for SCHED_FIFO and SCHED_RR are
4464 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4465 * SCHED_BATCH and SCHED_IDLE is 0.
4467 if (param->sched_priority < 0 ||
4468 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4469 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4471 if (rt_policy(policy) != (param->sched_priority != 0))
4475 * Allow unprivileged RT tasks to decrease priority:
4477 if (user && !capable(CAP_SYS_NICE)) {
4478 if (rt_policy(policy)) {
4479 unsigned long rlim_rtprio;
4481 if (!lock_task_sighand(p, &flags))
4483 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4484 unlock_task_sighand(p, &flags);
4486 /* can't set/change the rt policy */
4487 if (policy != p->policy && !rlim_rtprio)
4490 /* can't increase priority */
4491 if (param->sched_priority > p->rt_priority &&
4492 param->sched_priority > rlim_rtprio)
4496 * Like positive nice levels, dont allow tasks to
4497 * move out of SCHED_IDLE either:
4499 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4502 /* can't change other user's priorities */
4503 if (!check_same_owner(p))
4506 /* Normal users shall not reset the sched_reset_on_fork flag */
4507 if (p->sched_reset_on_fork && !reset_on_fork)
4512 #ifdef CONFIG_RT_GROUP_SCHED
4514 * Do not allow realtime tasks into groups that have no runtime
4517 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4518 task_group(p)->rt_bandwidth.rt_runtime == 0)
4522 retval = security_task_setscheduler(p, policy, param);
4528 * make sure no PI-waiters arrive (or leave) while we are
4529 * changing the priority of the task:
4531 raw_spin_lock_irqsave(&p->pi_lock, flags);
4533 * To be able to change p->policy safely, the apropriate
4534 * runqueue lock must be held.
4536 rq = __task_rq_lock(p);
4537 /* recheck policy now with rq lock held */
4538 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4539 policy = oldpolicy = -1;
4540 __task_rq_unlock(rq);
4541 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4544 update_rq_clock(rq);
4545 on_rq = p->se.on_rq;
4546 running = task_current(rq, p);
4548 deactivate_task(rq, p, 0);
4550 p->sched_class->put_prev_task(rq, p);
4552 p->sched_reset_on_fork = reset_on_fork;
4555 __setscheduler(rq, p, policy, param->sched_priority);
4558 p->sched_class->set_curr_task(rq);
4560 activate_task(rq, p, 0);
4562 check_class_changed(rq, p, prev_class, oldprio, running);
4564 __task_rq_unlock(rq);
4565 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4567 rt_mutex_adjust_pi(p);
4573 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4574 * @p: the task in question.
4575 * @policy: new policy.
4576 * @param: structure containing the new RT priority.
4578 * NOTE that the task may be already dead.
4580 int sched_setscheduler(struct task_struct *p, int policy,
4581 struct sched_param *param)
4583 return __sched_setscheduler(p, policy, param, true);
4585 EXPORT_SYMBOL_GPL(sched_setscheduler);
4588 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4589 * @p: the task in question.
4590 * @policy: new policy.
4591 * @param: structure containing the new RT priority.
4593 * Just like sched_setscheduler, only don't bother checking if the
4594 * current context has permission. For example, this is needed in
4595 * stop_machine(): we create temporary high priority worker threads,
4596 * but our caller might not have that capability.
4598 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4599 struct sched_param *param)
4601 return __sched_setscheduler(p, policy, param, false);
4605 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4607 struct sched_param lparam;
4608 struct task_struct *p;
4611 if (!param || pid < 0)
4613 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4618 p = find_process_by_pid(pid);
4620 retval = sched_setscheduler(p, policy, &lparam);
4627 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4628 * @pid: the pid in question.
4629 * @policy: new policy.
4630 * @param: structure containing the new RT priority.
4632 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4633 struct sched_param __user *, param)
4635 /* negative values for policy are not valid */
4639 return do_sched_setscheduler(pid, policy, param);
4643 * sys_sched_setparam - set/change the RT priority of a thread
4644 * @pid: the pid in question.
4645 * @param: structure containing the new RT priority.
4647 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4649 return do_sched_setscheduler(pid, -1, param);
4653 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4654 * @pid: the pid in question.
4656 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4658 struct task_struct *p;
4666 p = find_process_by_pid(pid);
4668 retval = security_task_getscheduler(p);
4671 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4678 * sys_sched_getparam - get the RT priority of a thread
4679 * @pid: the pid in question.
4680 * @param: structure containing the RT priority.
4682 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4684 struct sched_param lp;
4685 struct task_struct *p;
4688 if (!param || pid < 0)
4692 p = find_process_by_pid(pid);
4697 retval = security_task_getscheduler(p);
4701 lp.sched_priority = p->rt_priority;
4705 * This one might sleep, we cannot do it with a spinlock held ...
4707 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4716 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4718 cpumask_var_t cpus_allowed, new_mask;
4719 struct task_struct *p;
4725 p = find_process_by_pid(pid);
4732 /* Prevent p going away */
4736 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4740 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4742 goto out_free_cpus_allowed;
4745 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4748 retval = security_task_setscheduler(p, 0, NULL);
4752 cpuset_cpus_allowed(p, cpus_allowed);
4753 cpumask_and(new_mask, in_mask, cpus_allowed);
4755 retval = set_cpus_allowed_ptr(p, new_mask);
4758 cpuset_cpus_allowed(p, cpus_allowed);
4759 if (!cpumask_subset(new_mask, cpus_allowed)) {
4761 * We must have raced with a concurrent cpuset
4762 * update. Just reset the cpus_allowed to the
4763 * cpuset's cpus_allowed
4765 cpumask_copy(new_mask, cpus_allowed);
4770 free_cpumask_var(new_mask);
4771 out_free_cpus_allowed:
4772 free_cpumask_var(cpus_allowed);
4779 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4780 struct cpumask *new_mask)
4782 if (len < cpumask_size())
4783 cpumask_clear(new_mask);
4784 else if (len > cpumask_size())
4785 len = cpumask_size();
4787 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4791 * sys_sched_setaffinity - set the cpu affinity of a process
4792 * @pid: pid of the process
4793 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4794 * @user_mask_ptr: user-space pointer to the new cpu mask
4796 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4797 unsigned long __user *, user_mask_ptr)
4799 cpumask_var_t new_mask;
4802 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4805 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4807 retval = sched_setaffinity(pid, new_mask);
4808 free_cpumask_var(new_mask);
4812 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4814 struct task_struct *p;
4815 unsigned long flags;
4823 p = find_process_by_pid(pid);
4827 retval = security_task_getscheduler(p);
4831 rq = task_rq_lock(p, &flags);
4832 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4833 task_rq_unlock(rq, &flags);
4843 * sys_sched_getaffinity - get the cpu affinity of a process
4844 * @pid: pid of the process
4845 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4846 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4848 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4849 unsigned long __user *, user_mask_ptr)
4854 if (len < cpumask_size())
4857 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4860 ret = sched_getaffinity(pid, mask);
4862 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
4865 ret = cpumask_size();
4867 free_cpumask_var(mask);
4873 * sys_sched_yield - yield the current processor to other threads.
4875 * This function yields the current CPU to other tasks. If there are no
4876 * other threads running on this CPU then this function will return.
4878 SYSCALL_DEFINE0(sched_yield)
4880 struct rq *rq = this_rq_lock();
4882 schedstat_inc(rq, yld_count);
4883 current->sched_class->yield_task(rq);
4886 * Since we are going to call schedule() anyway, there's
4887 * no need to preempt or enable interrupts:
4889 __release(rq->lock);
4890 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4891 do_raw_spin_unlock(&rq->lock);
4892 preempt_enable_no_resched();
4899 static inline int should_resched(void)
4901 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4904 static void __cond_resched(void)
4906 add_preempt_count(PREEMPT_ACTIVE);
4908 sub_preempt_count(PREEMPT_ACTIVE);
4911 int __sched _cond_resched(void)
4913 if (should_resched()) {
4919 EXPORT_SYMBOL(_cond_resched);
4922 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4923 * call schedule, and on return reacquire the lock.
4925 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4926 * operations here to prevent schedule() from being called twice (once via
4927 * spin_unlock(), once by hand).
4929 int __cond_resched_lock(spinlock_t *lock)
4931 int resched = should_resched();
4934 lockdep_assert_held(lock);
4936 if (spin_needbreak(lock) || resched) {
4947 EXPORT_SYMBOL(__cond_resched_lock);
4949 int __sched __cond_resched_softirq(void)
4951 BUG_ON(!in_softirq());
4953 if (should_resched()) {
4961 EXPORT_SYMBOL(__cond_resched_softirq);
4964 * yield - yield the current processor to other threads.
4966 * This is a shortcut for kernel-space yielding - it marks the
4967 * thread runnable and calls sys_sched_yield().
4969 void __sched yield(void)
4971 set_current_state(TASK_RUNNING);
4974 EXPORT_SYMBOL(yield);
4977 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4978 * that process accounting knows that this is a task in IO wait state.
4980 void __sched io_schedule(void)
4982 struct rq *rq = raw_rq();
4984 delayacct_blkio_start();
4985 atomic_inc(&rq->nr_iowait);
4986 current->in_iowait = 1;
4988 current->in_iowait = 0;
4989 atomic_dec(&rq->nr_iowait);
4990 delayacct_blkio_end();
4992 EXPORT_SYMBOL(io_schedule);
4994 long __sched io_schedule_timeout(long timeout)
4996 struct rq *rq = raw_rq();
4999 delayacct_blkio_start();
5000 atomic_inc(&rq->nr_iowait);
5001 current->in_iowait = 1;
5002 ret = schedule_timeout(timeout);
5003 current->in_iowait = 0;
5004 atomic_dec(&rq->nr_iowait);
5005 delayacct_blkio_end();
5010 * sys_sched_get_priority_max - return maximum RT priority.
5011 * @policy: scheduling class.
5013 * this syscall returns the maximum rt_priority that can be used
5014 * by a given scheduling class.
5016 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5023 ret = MAX_USER_RT_PRIO-1;
5035 * sys_sched_get_priority_min - return minimum RT priority.
5036 * @policy: scheduling class.
5038 * this syscall returns the minimum rt_priority that can be used
5039 * by a given scheduling class.
5041 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5059 * sys_sched_rr_get_interval - return the default timeslice of a process.
5060 * @pid: pid of the process.
5061 * @interval: userspace pointer to the timeslice value.
5063 * this syscall writes the default timeslice value of a given process
5064 * into the user-space timespec buffer. A value of '0' means infinity.
5066 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5067 struct timespec __user *, interval)
5069 struct task_struct *p;
5070 unsigned int time_slice;
5071 unsigned long flags;
5081 p = find_process_by_pid(pid);
5085 retval = security_task_getscheduler(p);
5089 rq = task_rq_lock(p, &flags);
5090 time_slice = p->sched_class->get_rr_interval(rq, p);
5091 task_rq_unlock(rq, &flags);
5094 jiffies_to_timespec(time_slice, &t);
5095 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5103 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5105 void sched_show_task(struct task_struct *p)
5107 unsigned long free = 0;
5110 state = p->state ? __ffs(p->state) + 1 : 0;
5111 printk(KERN_INFO "%-13.13s %c", p->comm,
5112 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5113 #if BITS_PER_LONG == 32
5114 if (state == TASK_RUNNING)
5115 printk(KERN_CONT " running ");
5117 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5119 if (state == TASK_RUNNING)
5120 printk(KERN_CONT " running task ");
5122 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5124 #ifdef CONFIG_DEBUG_STACK_USAGE
5125 free = stack_not_used(p);
5127 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5128 task_pid_nr(p), task_pid_nr(p->real_parent),
5129 (unsigned long)task_thread_info(p)->flags);
5131 show_stack(p, NULL);
5134 void show_state_filter(unsigned long state_filter)
5136 struct task_struct *g, *p;
5138 #if BITS_PER_LONG == 32
5140 " task PC stack pid father\n");
5143 " task PC stack pid father\n");
5145 read_lock(&tasklist_lock);
5146 do_each_thread(g, p) {
5148 * reset the NMI-timeout, listing all files on a slow
5149 * console might take alot of time:
5151 touch_nmi_watchdog();
5152 if (!state_filter || (p->state & state_filter))
5154 } while_each_thread(g, p);
5156 touch_all_softlockup_watchdogs();
5158 #ifdef CONFIG_SCHED_DEBUG
5159 sysrq_sched_debug_show();
5161 read_unlock(&tasklist_lock);
5163 * Only show locks if all tasks are dumped:
5166 debug_show_all_locks();
5169 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5171 idle->sched_class = &idle_sched_class;
5175 * init_idle - set up an idle thread for a given CPU
5176 * @idle: task in question
5177 * @cpu: cpu the idle task belongs to
5179 * NOTE: this function does not set the idle thread's NEED_RESCHED
5180 * flag, to make booting more robust.
5182 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5184 struct rq *rq = cpu_rq(cpu);
5185 unsigned long flags;
5187 raw_spin_lock_irqsave(&rq->lock, flags);
5190 idle->state = TASK_RUNNING;
5191 idle->se.exec_start = sched_clock();
5193 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5194 __set_task_cpu(idle, cpu);
5196 rq->curr = rq->idle = idle;
5197 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5200 raw_spin_unlock_irqrestore(&rq->lock, flags);
5202 /* Set the preempt count _outside_ the spinlocks! */
5203 #if defined(CONFIG_PREEMPT)
5204 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5206 task_thread_info(idle)->preempt_count = 0;
5209 * The idle tasks have their own, simple scheduling class:
5211 idle->sched_class = &idle_sched_class;
5212 ftrace_graph_init_task(idle);
5216 * In a system that switches off the HZ timer nohz_cpu_mask
5217 * indicates which cpus entered this state. This is used
5218 * in the rcu update to wait only for active cpus. For system
5219 * which do not switch off the HZ timer nohz_cpu_mask should
5220 * always be CPU_BITS_NONE.
5222 cpumask_var_t nohz_cpu_mask;
5225 * Increase the granularity value when there are more CPUs,
5226 * because with more CPUs the 'effective latency' as visible
5227 * to users decreases. But the relationship is not linear,
5228 * so pick a second-best guess by going with the log2 of the
5231 * This idea comes from the SD scheduler of Con Kolivas:
5233 static int get_update_sysctl_factor(void)
5235 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5236 unsigned int factor;
5238 switch (sysctl_sched_tunable_scaling) {
5239 case SCHED_TUNABLESCALING_NONE:
5242 case SCHED_TUNABLESCALING_LINEAR:
5245 case SCHED_TUNABLESCALING_LOG:
5247 factor = 1 + ilog2(cpus);
5254 static void update_sysctl(void)
5256 unsigned int factor = get_update_sysctl_factor();
5258 #define SET_SYSCTL(name) \
5259 (sysctl_##name = (factor) * normalized_sysctl_##name)
5260 SET_SYSCTL(sched_min_granularity);
5261 SET_SYSCTL(sched_latency);
5262 SET_SYSCTL(sched_wakeup_granularity);
5263 SET_SYSCTL(sched_shares_ratelimit);
5267 static inline void sched_init_granularity(void)
5274 * This is how migration works:
5276 * 1) we queue a struct migration_req structure in the source CPU's
5277 * runqueue and wake up that CPU's migration thread.
5278 * 2) we down() the locked semaphore => thread blocks.
5279 * 3) migration thread wakes up (implicitly it forces the migrated
5280 * thread off the CPU)
5281 * 4) it gets the migration request and checks whether the migrated
5282 * task is still in the wrong runqueue.
5283 * 5) if it's in the wrong runqueue then the migration thread removes
5284 * it and puts it into the right queue.
5285 * 6) migration thread up()s the semaphore.
5286 * 7) we wake up and the migration is done.
5290 * Change a given task's CPU affinity. Migrate the thread to a
5291 * proper CPU and schedule it away if the CPU it's executing on
5292 * is removed from the allowed bitmask.
5294 * NOTE: the caller must have a valid reference to the task, the
5295 * task must not exit() & deallocate itself prematurely. The
5296 * call is not atomic; no spinlocks may be held.
5298 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5300 struct migration_req req;
5301 unsigned long flags;
5306 * Since we rely on wake-ups to migrate sleeping tasks, don't change
5307 * the ->cpus_allowed mask from under waking tasks, which would be
5308 * possible when we change rq->lock in ttwu(), so synchronize against
5309 * TASK_WAKING to avoid that.
5312 while (p->state == TASK_WAKING)
5315 rq = task_rq_lock(p, &flags);
5317 if (p->state == TASK_WAKING) {
5318 task_rq_unlock(rq, &flags);
5322 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5327 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5328 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5333 if (p->sched_class->set_cpus_allowed)
5334 p->sched_class->set_cpus_allowed(p, new_mask);
5336 cpumask_copy(&p->cpus_allowed, new_mask);
5337 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5340 /* Can the task run on the task's current CPU? If so, we're done */
5341 if (cpumask_test_cpu(task_cpu(p), new_mask))
5344 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5345 /* Need help from migration thread: drop lock and wait. */
5346 struct task_struct *mt = rq->migration_thread;
5348 get_task_struct(mt);
5349 task_rq_unlock(rq, &flags);
5350 wake_up_process(rq->migration_thread);
5351 put_task_struct(mt);
5352 wait_for_completion(&req.done);
5353 tlb_migrate_finish(p->mm);
5357 task_rq_unlock(rq, &flags);
5361 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5364 * Move (not current) task off this cpu, onto dest cpu. We're doing
5365 * this because either it can't run here any more (set_cpus_allowed()
5366 * away from this CPU, or CPU going down), or because we're
5367 * attempting to rebalance this task on exec (sched_exec).
5369 * So we race with normal scheduler movements, but that's OK, as long
5370 * as the task is no longer on this CPU.
5372 * Returns non-zero if task was successfully migrated.
5374 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5376 struct rq *rq_dest, *rq_src;
5379 if (unlikely(!cpu_active(dest_cpu)))
5382 rq_src = cpu_rq(src_cpu);
5383 rq_dest = cpu_rq(dest_cpu);
5385 double_rq_lock(rq_src, rq_dest);
5386 /* Already moved. */
5387 if (task_cpu(p) != src_cpu)
5389 /* Affinity changed (again). */
5390 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5394 * If we're not on a rq, the next wake-up will ensure we're
5398 deactivate_task(rq_src, p, 0);
5399 set_task_cpu(p, dest_cpu);
5400 activate_task(rq_dest, p, 0);
5401 check_preempt_curr(rq_dest, p, 0);
5406 double_rq_unlock(rq_src, rq_dest);
5410 #define RCU_MIGRATION_IDLE 0
5411 #define RCU_MIGRATION_NEED_QS 1
5412 #define RCU_MIGRATION_GOT_QS 2
5413 #define RCU_MIGRATION_MUST_SYNC 3
5416 * migration_thread - this is a highprio system thread that performs
5417 * thread migration by bumping thread off CPU then 'pushing' onto
5420 static int migration_thread(void *data)
5423 int cpu = (long)data;
5427 BUG_ON(rq->migration_thread != current);
5429 set_current_state(TASK_INTERRUPTIBLE);
5430 while (!kthread_should_stop()) {
5431 struct migration_req *req;
5432 struct list_head *head;
5434 raw_spin_lock_irq(&rq->lock);
5436 if (cpu_is_offline(cpu)) {
5437 raw_spin_unlock_irq(&rq->lock);
5441 if (rq->active_balance) {
5442 active_load_balance(rq, cpu);
5443 rq->active_balance = 0;
5446 head = &rq->migration_queue;
5448 if (list_empty(head)) {
5449 raw_spin_unlock_irq(&rq->lock);
5451 set_current_state(TASK_INTERRUPTIBLE);
5454 req = list_entry(head->next, struct migration_req, list);
5455 list_del_init(head->next);
5457 if (req->task != NULL) {
5458 raw_spin_unlock(&rq->lock);
5459 __migrate_task(req->task, cpu, req->dest_cpu);
5460 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5461 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5462 raw_spin_unlock(&rq->lock);
5464 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5465 raw_spin_unlock(&rq->lock);
5466 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5470 complete(&req->done);
5472 __set_current_state(TASK_RUNNING);
5477 #ifdef CONFIG_HOTPLUG_CPU
5479 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5483 local_irq_disable();
5484 ret = __migrate_task(p, src_cpu, dest_cpu);
5490 * Figure out where task on dead CPU should go, use force if necessary.
5492 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5497 dest_cpu = select_fallback_rq(dead_cpu, p);
5499 /* It can have affinity changed while we were choosing. */
5500 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
5505 * While a dead CPU has no uninterruptible tasks queued at this point,
5506 * it might still have a nonzero ->nr_uninterruptible counter, because
5507 * for performance reasons the counter is not stricly tracking tasks to
5508 * their home CPUs. So we just add the counter to another CPU's counter,
5509 * to keep the global sum constant after CPU-down:
5511 static void migrate_nr_uninterruptible(struct rq *rq_src)
5513 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5514 unsigned long flags;
5516 local_irq_save(flags);
5517 double_rq_lock(rq_src, rq_dest);
5518 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5519 rq_src->nr_uninterruptible = 0;
5520 double_rq_unlock(rq_src, rq_dest);
5521 local_irq_restore(flags);
5524 /* Run through task list and migrate tasks from the dead cpu. */
5525 static void migrate_live_tasks(int src_cpu)
5527 struct task_struct *p, *t;
5529 read_lock(&tasklist_lock);
5531 do_each_thread(t, p) {
5535 if (task_cpu(p) == src_cpu)
5536 move_task_off_dead_cpu(src_cpu, p);
5537 } while_each_thread(t, p);
5539 read_unlock(&tasklist_lock);
5543 * Schedules idle task to be the next runnable task on current CPU.
5544 * It does so by boosting its priority to highest possible.
5545 * Used by CPU offline code.
5547 void sched_idle_next(void)
5549 int this_cpu = smp_processor_id();
5550 struct rq *rq = cpu_rq(this_cpu);
5551 struct task_struct *p = rq->idle;
5552 unsigned long flags;
5554 /* cpu has to be offline */
5555 BUG_ON(cpu_online(this_cpu));
5558 * Strictly not necessary since rest of the CPUs are stopped by now
5559 * and interrupts disabled on the current cpu.
5561 raw_spin_lock_irqsave(&rq->lock, flags);
5563 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5565 update_rq_clock(rq);
5566 activate_task(rq, p, 0);
5568 raw_spin_unlock_irqrestore(&rq->lock, flags);
5572 * Ensures that the idle task is using init_mm right before its cpu goes
5575 void idle_task_exit(void)
5577 struct mm_struct *mm = current->active_mm;
5579 BUG_ON(cpu_online(smp_processor_id()));
5582 switch_mm(mm, &init_mm, current);
5586 /* called under rq->lock with disabled interrupts */
5587 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5589 struct rq *rq = cpu_rq(dead_cpu);
5591 /* Must be exiting, otherwise would be on tasklist. */
5592 BUG_ON(!p->exit_state);
5594 /* Cannot have done final schedule yet: would have vanished. */
5595 BUG_ON(p->state == TASK_DEAD);
5600 * Drop lock around migration; if someone else moves it,
5601 * that's OK. No task can be added to this CPU, so iteration is
5604 raw_spin_unlock_irq(&rq->lock);
5605 move_task_off_dead_cpu(dead_cpu, p);
5606 raw_spin_lock_irq(&rq->lock);
5611 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5612 static void migrate_dead_tasks(unsigned int dead_cpu)
5614 struct rq *rq = cpu_rq(dead_cpu);
5615 struct task_struct *next;
5618 if (!rq->nr_running)
5620 update_rq_clock(rq);
5621 next = pick_next_task(rq);
5624 next->sched_class->put_prev_task(rq, next);
5625 migrate_dead(dead_cpu, next);
5631 * remove the tasks which were accounted by rq from calc_load_tasks.
5633 static void calc_global_load_remove(struct rq *rq)
5635 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5636 rq->calc_load_active = 0;
5638 #endif /* CONFIG_HOTPLUG_CPU */
5640 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5642 static struct ctl_table sd_ctl_dir[] = {
5644 .procname = "sched_domain",
5650 static struct ctl_table sd_ctl_root[] = {
5652 .procname = "kernel",
5654 .child = sd_ctl_dir,
5659 static struct ctl_table *sd_alloc_ctl_entry(int n)
5661 struct ctl_table *entry =
5662 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5667 static void sd_free_ctl_entry(struct ctl_table **tablep)
5669 struct ctl_table *entry;
5672 * In the intermediate directories, both the child directory and
5673 * procname are dynamically allocated and could fail but the mode
5674 * will always be set. In the lowest directory the names are
5675 * static strings and all have proc handlers.
5677 for (entry = *tablep; entry->mode; entry++) {
5679 sd_free_ctl_entry(&entry->child);
5680 if (entry->proc_handler == NULL)
5681 kfree(entry->procname);
5689 set_table_entry(struct ctl_table *entry,
5690 const char *procname, void *data, int maxlen,
5691 mode_t mode, proc_handler *proc_handler)
5693 entry->procname = procname;
5695 entry->maxlen = maxlen;
5697 entry->proc_handler = proc_handler;
5700 static struct ctl_table *
5701 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5703 struct ctl_table *table = sd_alloc_ctl_entry(13);
5708 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5709 sizeof(long), 0644, proc_doulongvec_minmax);
5710 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5711 sizeof(long), 0644, proc_doulongvec_minmax);
5712 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5713 sizeof(int), 0644, proc_dointvec_minmax);
5714 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5715 sizeof(int), 0644, proc_dointvec_minmax);
5716 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5717 sizeof(int), 0644, proc_dointvec_minmax);
5718 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5719 sizeof(int), 0644, proc_dointvec_minmax);
5720 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5721 sizeof(int), 0644, proc_dointvec_minmax);
5722 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5723 sizeof(int), 0644, proc_dointvec_minmax);
5724 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5725 sizeof(int), 0644, proc_dointvec_minmax);
5726 set_table_entry(&table[9], "cache_nice_tries",
5727 &sd->cache_nice_tries,
5728 sizeof(int), 0644, proc_dointvec_minmax);
5729 set_table_entry(&table[10], "flags", &sd->flags,
5730 sizeof(int), 0644, proc_dointvec_minmax);
5731 set_table_entry(&table[11], "name", sd->name,
5732 CORENAME_MAX_SIZE, 0444, proc_dostring);
5733 /* &table[12] is terminator */
5738 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5740 struct ctl_table *entry, *table;
5741 struct sched_domain *sd;
5742 int domain_num = 0, i;
5745 for_each_domain(cpu, sd)
5747 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5752 for_each_domain(cpu, sd) {
5753 snprintf(buf, 32, "domain%d", i);
5754 entry->procname = kstrdup(buf, GFP_KERNEL);
5756 entry->child = sd_alloc_ctl_domain_table(sd);
5763 static struct ctl_table_header *sd_sysctl_header;
5764 static void register_sched_domain_sysctl(void)
5766 int i, cpu_num = num_possible_cpus();
5767 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5770 WARN_ON(sd_ctl_dir[0].child);
5771 sd_ctl_dir[0].child = entry;
5776 for_each_possible_cpu(i) {
5777 snprintf(buf, 32, "cpu%d", i);
5778 entry->procname = kstrdup(buf, GFP_KERNEL);
5780 entry->child = sd_alloc_ctl_cpu_table(i);
5784 WARN_ON(sd_sysctl_header);
5785 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5788 /* may be called multiple times per register */
5789 static void unregister_sched_domain_sysctl(void)
5791 if (sd_sysctl_header)
5792 unregister_sysctl_table(sd_sysctl_header);
5793 sd_sysctl_header = NULL;
5794 if (sd_ctl_dir[0].child)
5795 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5798 static void register_sched_domain_sysctl(void)
5801 static void unregister_sched_domain_sysctl(void)
5806 static void set_rq_online(struct rq *rq)
5809 const struct sched_class *class;
5811 cpumask_set_cpu(rq->cpu, rq->rd->online);
5814 for_each_class(class) {
5815 if (class->rq_online)
5816 class->rq_online(rq);
5821 static void set_rq_offline(struct rq *rq)
5824 const struct sched_class *class;
5826 for_each_class(class) {
5827 if (class->rq_offline)
5828 class->rq_offline(rq);
5831 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5837 * migration_call - callback that gets triggered when a CPU is added.
5838 * Here we can start up the necessary migration thread for the new CPU.
5840 static int __cpuinit
5841 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5843 struct task_struct *p;
5844 int cpu = (long)hcpu;
5845 unsigned long flags;
5850 case CPU_UP_PREPARE:
5851 case CPU_UP_PREPARE_FROZEN:
5852 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5855 kthread_bind(p, cpu);
5856 /* Must be high prio: stop_machine expects to yield to it. */
5857 rq = task_rq_lock(p, &flags);
5858 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5859 task_rq_unlock(rq, &flags);
5861 cpu_rq(cpu)->migration_thread = p;
5862 rq->calc_load_update = calc_load_update;
5866 case CPU_ONLINE_FROZEN:
5867 /* Strictly unnecessary, as first user will wake it. */
5868 wake_up_process(cpu_rq(cpu)->migration_thread);
5870 /* Update our root-domain */
5872 raw_spin_lock_irqsave(&rq->lock, flags);
5874 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5878 raw_spin_unlock_irqrestore(&rq->lock, flags);
5881 #ifdef CONFIG_HOTPLUG_CPU
5882 case CPU_UP_CANCELED:
5883 case CPU_UP_CANCELED_FROZEN:
5884 if (!cpu_rq(cpu)->migration_thread)
5886 /* Unbind it from offline cpu so it can run. Fall thru. */
5887 kthread_bind(cpu_rq(cpu)->migration_thread,
5888 cpumask_any(cpu_online_mask));
5889 kthread_stop(cpu_rq(cpu)->migration_thread);
5890 put_task_struct(cpu_rq(cpu)->migration_thread);
5891 cpu_rq(cpu)->migration_thread = NULL;
5895 case CPU_DEAD_FROZEN:
5896 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5897 migrate_live_tasks(cpu);
5899 kthread_stop(rq->migration_thread);
5900 put_task_struct(rq->migration_thread);
5901 rq->migration_thread = NULL;
5902 /* Idle task back to normal (off runqueue, low prio) */
5903 raw_spin_lock_irq(&rq->lock);
5904 update_rq_clock(rq);
5905 deactivate_task(rq, rq->idle, 0);
5906 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5907 rq->idle->sched_class = &idle_sched_class;
5908 migrate_dead_tasks(cpu);
5909 raw_spin_unlock_irq(&rq->lock);
5911 migrate_nr_uninterruptible(rq);
5912 BUG_ON(rq->nr_running != 0);
5913 calc_global_load_remove(rq);
5915 * No need to migrate the tasks: it was best-effort if
5916 * they didn't take sched_hotcpu_mutex. Just wake up
5919 raw_spin_lock_irq(&rq->lock);
5920 while (!list_empty(&rq->migration_queue)) {
5921 struct migration_req *req;
5923 req = list_entry(rq->migration_queue.next,
5924 struct migration_req, list);
5925 list_del_init(&req->list);
5926 raw_spin_unlock_irq(&rq->lock);
5927 complete(&req->done);
5928 raw_spin_lock_irq(&rq->lock);
5930 raw_spin_unlock_irq(&rq->lock);
5934 case CPU_DYING_FROZEN:
5935 /* Update our root-domain */
5937 raw_spin_lock_irqsave(&rq->lock, flags);
5939 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5942 raw_spin_unlock_irqrestore(&rq->lock, flags);
5950 * Register at high priority so that task migration (migrate_all_tasks)
5951 * happens before everything else. This has to be lower priority than
5952 * the notifier in the perf_event subsystem, though.
5954 static struct notifier_block __cpuinitdata migration_notifier = {
5955 .notifier_call = migration_call,
5959 static int __init migration_init(void)
5961 void *cpu = (void *)(long)smp_processor_id();
5964 /* Start one for the boot CPU: */
5965 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5966 BUG_ON(err == NOTIFY_BAD);
5967 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5968 register_cpu_notifier(&migration_notifier);
5972 early_initcall(migration_init);
5977 #ifdef CONFIG_SCHED_DEBUG
5979 static __read_mostly int sched_domain_debug_enabled;
5981 static int __init sched_domain_debug_setup(char *str)
5983 sched_domain_debug_enabled = 1;
5987 early_param("sched_debug", sched_domain_debug_setup);
5989 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5990 struct cpumask *groupmask)
5992 struct sched_group *group = sd->groups;
5995 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5996 cpumask_clear(groupmask);
5998 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6000 if (!(sd->flags & SD_LOAD_BALANCE)) {
6001 printk("does not load-balance\n");
6003 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6008 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6010 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6011 printk(KERN_ERR "ERROR: domain->span does not contain "
6014 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6015 printk(KERN_ERR "ERROR: domain->groups does not contain"
6019 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6023 printk(KERN_ERR "ERROR: group is NULL\n");
6027 if (!group->cpu_power) {
6028 printk(KERN_CONT "\n");
6029 printk(KERN_ERR "ERROR: domain->cpu_power not "
6034 if (!cpumask_weight(sched_group_cpus(group))) {
6035 printk(KERN_CONT "\n");
6036 printk(KERN_ERR "ERROR: empty group\n");
6040 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6041 printk(KERN_CONT "\n");
6042 printk(KERN_ERR "ERROR: repeated CPUs\n");
6046 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6048 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6050 printk(KERN_CONT " %s", str);
6051 if (group->cpu_power != SCHED_LOAD_SCALE) {
6052 printk(KERN_CONT " (cpu_power = %d)",
6056 group = group->next;
6057 } while (group != sd->groups);
6058 printk(KERN_CONT "\n");
6060 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6061 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6064 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6065 printk(KERN_ERR "ERROR: parent span is not a superset "
6066 "of domain->span\n");
6070 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6072 cpumask_var_t groupmask;
6075 if (!sched_domain_debug_enabled)
6079 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6083 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6085 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6086 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6091 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6098 free_cpumask_var(groupmask);
6100 #else /* !CONFIG_SCHED_DEBUG */
6101 # define sched_domain_debug(sd, cpu) do { } while (0)
6102 #endif /* CONFIG_SCHED_DEBUG */
6104 static int sd_degenerate(struct sched_domain *sd)
6106 if (cpumask_weight(sched_domain_span(sd)) == 1)
6109 /* Following flags need at least 2 groups */
6110 if (sd->flags & (SD_LOAD_BALANCE |
6111 SD_BALANCE_NEWIDLE |
6115 SD_SHARE_PKG_RESOURCES)) {
6116 if (sd->groups != sd->groups->next)
6120 /* Following flags don't use groups */
6121 if (sd->flags & (SD_WAKE_AFFINE))
6128 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6130 unsigned long cflags = sd->flags, pflags = parent->flags;
6132 if (sd_degenerate(parent))
6135 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6138 /* Flags needing groups don't count if only 1 group in parent */
6139 if (parent->groups == parent->groups->next) {
6140 pflags &= ~(SD_LOAD_BALANCE |
6141 SD_BALANCE_NEWIDLE |
6145 SD_SHARE_PKG_RESOURCES);
6146 if (nr_node_ids == 1)
6147 pflags &= ~SD_SERIALIZE;
6149 if (~cflags & pflags)
6155 static void free_rootdomain(struct root_domain *rd)
6157 synchronize_sched();
6159 cpupri_cleanup(&rd->cpupri);
6161 free_cpumask_var(rd->rto_mask);
6162 free_cpumask_var(rd->online);
6163 free_cpumask_var(rd->span);
6167 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6169 struct root_domain *old_rd = NULL;
6170 unsigned long flags;
6172 raw_spin_lock_irqsave(&rq->lock, flags);
6177 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6180 cpumask_clear_cpu(rq->cpu, old_rd->span);
6183 * If we dont want to free the old_rt yet then
6184 * set old_rd to NULL to skip the freeing later
6187 if (!atomic_dec_and_test(&old_rd->refcount))
6191 atomic_inc(&rd->refcount);
6194 cpumask_set_cpu(rq->cpu, rd->span);
6195 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6198 raw_spin_unlock_irqrestore(&rq->lock, flags);
6201 free_rootdomain(old_rd);
6204 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6206 gfp_t gfp = GFP_KERNEL;
6208 memset(rd, 0, sizeof(*rd));
6213 if (!alloc_cpumask_var(&rd->span, gfp))
6215 if (!alloc_cpumask_var(&rd->online, gfp))
6217 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6220 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6225 free_cpumask_var(rd->rto_mask);
6227 free_cpumask_var(rd->online);
6229 free_cpumask_var(rd->span);
6234 static void init_defrootdomain(void)
6236 init_rootdomain(&def_root_domain, true);
6238 atomic_set(&def_root_domain.refcount, 1);
6241 static struct root_domain *alloc_rootdomain(void)
6243 struct root_domain *rd;
6245 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6249 if (init_rootdomain(rd, false) != 0) {
6258 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6259 * hold the hotplug lock.
6262 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6264 struct rq *rq = cpu_rq(cpu);
6265 struct sched_domain *tmp;
6267 /* Remove the sched domains which do not contribute to scheduling. */
6268 for (tmp = sd; tmp; ) {
6269 struct sched_domain *parent = tmp->parent;
6273 if (sd_parent_degenerate(tmp, parent)) {
6274 tmp->parent = parent->parent;
6276 parent->parent->child = tmp;
6281 if (sd && sd_degenerate(sd)) {
6287 sched_domain_debug(sd, cpu);
6289 rq_attach_root(rq, rd);
6290 rcu_assign_pointer(rq->sd, sd);
6293 /* cpus with isolated domains */
6294 static cpumask_var_t cpu_isolated_map;
6296 /* Setup the mask of cpus configured for isolated domains */
6297 static int __init isolated_cpu_setup(char *str)
6299 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6300 cpulist_parse(str, cpu_isolated_map);
6304 __setup("isolcpus=", isolated_cpu_setup);
6307 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6308 * to a function which identifies what group(along with sched group) a CPU
6309 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6310 * (due to the fact that we keep track of groups covered with a struct cpumask).
6312 * init_sched_build_groups will build a circular linked list of the groups
6313 * covered by the given span, and will set each group's ->cpumask correctly,
6314 * and ->cpu_power to 0.
6317 init_sched_build_groups(const struct cpumask *span,
6318 const struct cpumask *cpu_map,
6319 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6320 struct sched_group **sg,
6321 struct cpumask *tmpmask),
6322 struct cpumask *covered, struct cpumask *tmpmask)
6324 struct sched_group *first = NULL, *last = NULL;
6327 cpumask_clear(covered);
6329 for_each_cpu(i, span) {
6330 struct sched_group *sg;
6331 int group = group_fn(i, cpu_map, &sg, tmpmask);
6334 if (cpumask_test_cpu(i, covered))
6337 cpumask_clear(sched_group_cpus(sg));
6340 for_each_cpu(j, span) {
6341 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6344 cpumask_set_cpu(j, covered);
6345 cpumask_set_cpu(j, sched_group_cpus(sg));
6356 #define SD_NODES_PER_DOMAIN 16
6361 * find_next_best_node - find the next node to include in a sched_domain
6362 * @node: node whose sched_domain we're building
6363 * @used_nodes: nodes already in the sched_domain
6365 * Find the next node to include in a given scheduling domain. Simply
6366 * finds the closest node not already in the @used_nodes map.
6368 * Should use nodemask_t.
6370 static int find_next_best_node(int node, nodemask_t *used_nodes)
6372 int i, n, val, min_val, best_node = 0;
6376 for (i = 0; i < nr_node_ids; i++) {
6377 /* Start at @node */
6378 n = (node + i) % nr_node_ids;
6380 if (!nr_cpus_node(n))
6383 /* Skip already used nodes */
6384 if (node_isset(n, *used_nodes))
6387 /* Simple min distance search */
6388 val = node_distance(node, n);
6390 if (val < min_val) {
6396 node_set(best_node, *used_nodes);
6401 * sched_domain_node_span - get a cpumask for a node's sched_domain
6402 * @node: node whose cpumask we're constructing
6403 * @span: resulting cpumask
6405 * Given a node, construct a good cpumask for its sched_domain to span. It
6406 * should be one that prevents unnecessary balancing, but also spreads tasks
6409 static void sched_domain_node_span(int node, struct cpumask *span)
6411 nodemask_t used_nodes;
6414 cpumask_clear(span);
6415 nodes_clear(used_nodes);
6417 cpumask_or(span, span, cpumask_of_node(node));
6418 node_set(node, used_nodes);
6420 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6421 int next_node = find_next_best_node(node, &used_nodes);
6423 cpumask_or(span, span, cpumask_of_node(next_node));
6426 #endif /* CONFIG_NUMA */
6428 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6431 * The cpus mask in sched_group and sched_domain hangs off the end.
6433 * ( See the the comments in include/linux/sched.h:struct sched_group
6434 * and struct sched_domain. )
6436 struct static_sched_group {
6437 struct sched_group sg;
6438 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6441 struct static_sched_domain {
6442 struct sched_domain sd;
6443 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6449 cpumask_var_t domainspan;
6450 cpumask_var_t covered;
6451 cpumask_var_t notcovered;
6453 cpumask_var_t nodemask;
6454 cpumask_var_t this_sibling_map;
6455 cpumask_var_t this_core_map;
6456 cpumask_var_t send_covered;
6457 cpumask_var_t tmpmask;
6458 struct sched_group **sched_group_nodes;
6459 struct root_domain *rd;
6463 sa_sched_groups = 0,
6468 sa_this_sibling_map,
6470 sa_sched_group_nodes,
6480 * SMT sched-domains:
6482 #ifdef CONFIG_SCHED_SMT
6483 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6484 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6487 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6488 struct sched_group **sg, struct cpumask *unused)
6491 *sg = &per_cpu(sched_groups, cpu).sg;
6494 #endif /* CONFIG_SCHED_SMT */
6497 * multi-core sched-domains:
6499 #ifdef CONFIG_SCHED_MC
6500 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6501 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6502 #endif /* CONFIG_SCHED_MC */
6504 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6506 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6507 struct sched_group **sg, struct cpumask *mask)
6511 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6512 group = cpumask_first(mask);
6514 *sg = &per_cpu(sched_group_core, group).sg;
6517 #elif defined(CONFIG_SCHED_MC)
6519 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6520 struct sched_group **sg, struct cpumask *unused)
6523 *sg = &per_cpu(sched_group_core, cpu).sg;
6528 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6529 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6532 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6533 struct sched_group **sg, struct cpumask *mask)
6536 #ifdef CONFIG_SCHED_MC
6537 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6538 group = cpumask_first(mask);
6539 #elif defined(CONFIG_SCHED_SMT)
6540 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6541 group = cpumask_first(mask);
6546 *sg = &per_cpu(sched_group_phys, group).sg;
6552 * The init_sched_build_groups can't handle what we want to do with node
6553 * groups, so roll our own. Now each node has its own list of groups which
6554 * gets dynamically allocated.
6556 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6557 static struct sched_group ***sched_group_nodes_bycpu;
6559 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6560 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6562 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6563 struct sched_group **sg,
6564 struct cpumask *nodemask)
6568 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6569 group = cpumask_first(nodemask);
6572 *sg = &per_cpu(sched_group_allnodes, group).sg;
6576 static void init_numa_sched_groups_power(struct sched_group *group_head)
6578 struct sched_group *sg = group_head;
6584 for_each_cpu(j, sched_group_cpus(sg)) {
6585 struct sched_domain *sd;
6587 sd = &per_cpu(phys_domains, j).sd;
6588 if (j != group_first_cpu(sd->groups)) {
6590 * Only add "power" once for each
6596 sg->cpu_power += sd->groups->cpu_power;
6599 } while (sg != group_head);
6602 static int build_numa_sched_groups(struct s_data *d,
6603 const struct cpumask *cpu_map, int num)
6605 struct sched_domain *sd;
6606 struct sched_group *sg, *prev;
6609 cpumask_clear(d->covered);
6610 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6611 if (cpumask_empty(d->nodemask)) {
6612 d->sched_group_nodes[num] = NULL;
6616 sched_domain_node_span(num, d->domainspan);
6617 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6619 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6622 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6626 d->sched_group_nodes[num] = sg;
6628 for_each_cpu(j, d->nodemask) {
6629 sd = &per_cpu(node_domains, j).sd;
6634 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6636 cpumask_or(d->covered, d->covered, d->nodemask);
6639 for (j = 0; j < nr_node_ids; j++) {
6640 n = (num + j) % nr_node_ids;
6641 cpumask_complement(d->notcovered, d->covered);
6642 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6643 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6644 if (cpumask_empty(d->tmpmask))
6646 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6647 if (cpumask_empty(d->tmpmask))
6649 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6653 "Can not alloc domain group for node %d\n", j);
6657 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6658 sg->next = prev->next;
6659 cpumask_or(d->covered, d->covered, d->tmpmask);
6666 #endif /* CONFIG_NUMA */
6669 /* Free memory allocated for various sched_group structures */
6670 static void free_sched_groups(const struct cpumask *cpu_map,
6671 struct cpumask *nodemask)
6675 for_each_cpu(cpu, cpu_map) {
6676 struct sched_group **sched_group_nodes
6677 = sched_group_nodes_bycpu[cpu];
6679 if (!sched_group_nodes)
6682 for (i = 0; i < nr_node_ids; i++) {
6683 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6685 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6686 if (cpumask_empty(nodemask))
6696 if (oldsg != sched_group_nodes[i])
6699 kfree(sched_group_nodes);
6700 sched_group_nodes_bycpu[cpu] = NULL;
6703 #else /* !CONFIG_NUMA */
6704 static void free_sched_groups(const struct cpumask *cpu_map,
6705 struct cpumask *nodemask)
6708 #endif /* CONFIG_NUMA */
6711 * Initialize sched groups cpu_power.
6713 * cpu_power indicates the capacity of sched group, which is used while
6714 * distributing the load between different sched groups in a sched domain.
6715 * Typically cpu_power for all the groups in a sched domain will be same unless
6716 * there are asymmetries in the topology. If there are asymmetries, group
6717 * having more cpu_power will pickup more load compared to the group having
6720 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6722 struct sched_domain *child;
6723 struct sched_group *group;
6727 WARN_ON(!sd || !sd->groups);
6729 if (cpu != group_first_cpu(sd->groups))
6734 sd->groups->cpu_power = 0;
6737 power = SCHED_LOAD_SCALE;
6738 weight = cpumask_weight(sched_domain_span(sd));
6740 * SMT siblings share the power of a single core.
6741 * Usually multiple threads get a better yield out of
6742 * that one core than a single thread would have,
6743 * reflect that in sd->smt_gain.
6745 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6746 power *= sd->smt_gain;
6748 power >>= SCHED_LOAD_SHIFT;
6750 sd->groups->cpu_power += power;
6755 * Add cpu_power of each child group to this groups cpu_power.
6757 group = child->groups;
6759 sd->groups->cpu_power += group->cpu_power;
6760 group = group->next;
6761 } while (group != child->groups);
6765 * Initializers for schedule domains
6766 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6769 #ifdef CONFIG_SCHED_DEBUG
6770 # define SD_INIT_NAME(sd, type) sd->name = #type
6772 # define SD_INIT_NAME(sd, type) do { } while (0)
6775 #define SD_INIT(sd, type) sd_init_##type(sd)
6777 #define SD_INIT_FUNC(type) \
6778 static noinline void sd_init_##type(struct sched_domain *sd) \
6780 memset(sd, 0, sizeof(*sd)); \
6781 *sd = SD_##type##_INIT; \
6782 sd->level = SD_LV_##type; \
6783 SD_INIT_NAME(sd, type); \
6788 SD_INIT_FUNC(ALLNODES)
6791 #ifdef CONFIG_SCHED_SMT
6792 SD_INIT_FUNC(SIBLING)
6794 #ifdef CONFIG_SCHED_MC
6798 static int default_relax_domain_level = -1;
6800 static int __init setup_relax_domain_level(char *str)
6804 val = simple_strtoul(str, NULL, 0);
6805 if (val < SD_LV_MAX)
6806 default_relax_domain_level = val;
6810 __setup("relax_domain_level=", setup_relax_domain_level);
6812 static void set_domain_attribute(struct sched_domain *sd,
6813 struct sched_domain_attr *attr)
6817 if (!attr || attr->relax_domain_level < 0) {
6818 if (default_relax_domain_level < 0)
6821 request = default_relax_domain_level;
6823 request = attr->relax_domain_level;
6824 if (request < sd->level) {
6825 /* turn off idle balance on this domain */
6826 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6828 /* turn on idle balance on this domain */
6829 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6833 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6834 const struct cpumask *cpu_map)
6837 case sa_sched_groups:
6838 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6839 d->sched_group_nodes = NULL;
6841 free_rootdomain(d->rd); /* fall through */
6843 free_cpumask_var(d->tmpmask); /* fall through */
6844 case sa_send_covered:
6845 free_cpumask_var(d->send_covered); /* fall through */
6846 case sa_this_core_map:
6847 free_cpumask_var(d->this_core_map); /* fall through */
6848 case sa_this_sibling_map:
6849 free_cpumask_var(d->this_sibling_map); /* fall through */
6851 free_cpumask_var(d->nodemask); /* fall through */
6852 case sa_sched_group_nodes:
6854 kfree(d->sched_group_nodes); /* fall through */
6856 free_cpumask_var(d->notcovered); /* fall through */
6858 free_cpumask_var(d->covered); /* fall through */
6860 free_cpumask_var(d->domainspan); /* fall through */
6867 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6868 const struct cpumask *cpu_map)
6871 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6873 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6874 return sa_domainspan;
6875 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6877 /* Allocate the per-node list of sched groups */
6878 d->sched_group_nodes = kcalloc(nr_node_ids,
6879 sizeof(struct sched_group *), GFP_KERNEL);
6880 if (!d->sched_group_nodes) {
6881 printk(KERN_WARNING "Can not alloc sched group node list\n");
6882 return sa_notcovered;
6884 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6886 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6887 return sa_sched_group_nodes;
6888 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6890 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6891 return sa_this_sibling_map;
6892 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6893 return sa_this_core_map;
6894 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6895 return sa_send_covered;
6896 d->rd = alloc_rootdomain();
6898 printk(KERN_WARNING "Cannot alloc root domain\n");
6901 return sa_rootdomain;
6904 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6905 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6907 struct sched_domain *sd = NULL;
6909 struct sched_domain *parent;
6912 if (cpumask_weight(cpu_map) >
6913 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6914 sd = &per_cpu(allnodes_domains, i).sd;
6915 SD_INIT(sd, ALLNODES);
6916 set_domain_attribute(sd, attr);
6917 cpumask_copy(sched_domain_span(sd), cpu_map);
6918 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6923 sd = &per_cpu(node_domains, i).sd;
6925 set_domain_attribute(sd, attr);
6926 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6927 sd->parent = parent;
6930 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6935 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6936 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6937 struct sched_domain *parent, int i)
6939 struct sched_domain *sd;
6940 sd = &per_cpu(phys_domains, i).sd;
6942 set_domain_attribute(sd, attr);
6943 cpumask_copy(sched_domain_span(sd), d->nodemask);
6944 sd->parent = parent;
6947 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6951 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6952 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6953 struct sched_domain *parent, int i)
6955 struct sched_domain *sd = parent;
6956 #ifdef CONFIG_SCHED_MC
6957 sd = &per_cpu(core_domains, i).sd;
6959 set_domain_attribute(sd, attr);
6960 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6961 sd->parent = parent;
6963 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6968 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6969 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6970 struct sched_domain *parent, int i)
6972 struct sched_domain *sd = parent;
6973 #ifdef CONFIG_SCHED_SMT
6974 sd = &per_cpu(cpu_domains, i).sd;
6975 SD_INIT(sd, SIBLING);
6976 set_domain_attribute(sd, attr);
6977 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6978 sd->parent = parent;
6980 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6985 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6986 const struct cpumask *cpu_map, int cpu)
6989 #ifdef CONFIG_SCHED_SMT
6990 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6991 cpumask_and(d->this_sibling_map, cpu_map,
6992 topology_thread_cpumask(cpu));
6993 if (cpu == cpumask_first(d->this_sibling_map))
6994 init_sched_build_groups(d->this_sibling_map, cpu_map,
6996 d->send_covered, d->tmpmask);
6999 #ifdef CONFIG_SCHED_MC
7000 case SD_LV_MC: /* set up multi-core groups */
7001 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7002 if (cpu == cpumask_first(d->this_core_map))
7003 init_sched_build_groups(d->this_core_map, cpu_map,
7005 d->send_covered, d->tmpmask);
7008 case SD_LV_CPU: /* set up physical groups */
7009 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7010 if (!cpumask_empty(d->nodemask))
7011 init_sched_build_groups(d->nodemask, cpu_map,
7013 d->send_covered, d->tmpmask);
7016 case SD_LV_ALLNODES:
7017 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7018 d->send_covered, d->tmpmask);
7027 * Build sched domains for a given set of cpus and attach the sched domains
7028 * to the individual cpus
7030 static int __build_sched_domains(const struct cpumask *cpu_map,
7031 struct sched_domain_attr *attr)
7033 enum s_alloc alloc_state = sa_none;
7035 struct sched_domain *sd;
7041 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7042 if (alloc_state != sa_rootdomain)
7044 alloc_state = sa_sched_groups;
7047 * Set up domains for cpus specified by the cpu_map.
7049 for_each_cpu(i, cpu_map) {
7050 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7053 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7054 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7055 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7056 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7059 for_each_cpu(i, cpu_map) {
7060 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7061 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7064 /* Set up physical groups */
7065 for (i = 0; i < nr_node_ids; i++)
7066 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7069 /* Set up node groups */
7071 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7073 for (i = 0; i < nr_node_ids; i++)
7074 if (build_numa_sched_groups(&d, cpu_map, i))
7078 /* Calculate CPU power for physical packages and nodes */
7079 #ifdef CONFIG_SCHED_SMT
7080 for_each_cpu(i, cpu_map) {
7081 sd = &per_cpu(cpu_domains, i).sd;
7082 init_sched_groups_power(i, sd);
7085 #ifdef CONFIG_SCHED_MC
7086 for_each_cpu(i, cpu_map) {
7087 sd = &per_cpu(core_domains, i).sd;
7088 init_sched_groups_power(i, sd);
7092 for_each_cpu(i, cpu_map) {
7093 sd = &per_cpu(phys_domains, i).sd;
7094 init_sched_groups_power(i, sd);
7098 for (i = 0; i < nr_node_ids; i++)
7099 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7101 if (d.sd_allnodes) {
7102 struct sched_group *sg;
7104 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7106 init_numa_sched_groups_power(sg);
7110 /* Attach the domains */
7111 for_each_cpu(i, cpu_map) {
7112 #ifdef CONFIG_SCHED_SMT
7113 sd = &per_cpu(cpu_domains, i).sd;
7114 #elif defined(CONFIG_SCHED_MC)
7115 sd = &per_cpu(core_domains, i).sd;
7117 sd = &per_cpu(phys_domains, i).sd;
7119 cpu_attach_domain(sd, d.rd, i);
7122 d.sched_group_nodes = NULL; /* don't free this we still need it */
7123 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7127 __free_domain_allocs(&d, alloc_state, cpu_map);
7131 static int build_sched_domains(const struct cpumask *cpu_map)
7133 return __build_sched_domains(cpu_map, NULL);
7136 static cpumask_var_t *doms_cur; /* current sched domains */
7137 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7138 static struct sched_domain_attr *dattr_cur;
7139 /* attribues of custom domains in 'doms_cur' */
7142 * Special case: If a kmalloc of a doms_cur partition (array of
7143 * cpumask) fails, then fallback to a single sched domain,
7144 * as determined by the single cpumask fallback_doms.
7146 static cpumask_var_t fallback_doms;
7149 * arch_update_cpu_topology lets virtualized architectures update the
7150 * cpu core maps. It is supposed to return 1 if the topology changed
7151 * or 0 if it stayed the same.
7153 int __attribute__((weak)) arch_update_cpu_topology(void)
7158 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7161 cpumask_var_t *doms;
7163 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7166 for (i = 0; i < ndoms; i++) {
7167 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7168 free_sched_domains(doms, i);
7175 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7178 for (i = 0; i < ndoms; i++)
7179 free_cpumask_var(doms[i]);
7184 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7185 * For now this just excludes isolated cpus, but could be used to
7186 * exclude other special cases in the future.
7188 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7192 arch_update_cpu_topology();
7194 doms_cur = alloc_sched_domains(ndoms_cur);
7196 doms_cur = &fallback_doms;
7197 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7199 err = build_sched_domains(doms_cur[0]);
7200 register_sched_domain_sysctl();
7205 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7206 struct cpumask *tmpmask)
7208 free_sched_groups(cpu_map, tmpmask);
7212 * Detach sched domains from a group of cpus specified in cpu_map
7213 * These cpus will now be attached to the NULL domain
7215 static void detach_destroy_domains(const struct cpumask *cpu_map)
7217 /* Save because hotplug lock held. */
7218 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7221 for_each_cpu(i, cpu_map)
7222 cpu_attach_domain(NULL, &def_root_domain, i);
7223 synchronize_sched();
7224 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7227 /* handle null as "default" */
7228 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7229 struct sched_domain_attr *new, int idx_new)
7231 struct sched_domain_attr tmp;
7238 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7239 new ? (new + idx_new) : &tmp,
7240 sizeof(struct sched_domain_attr));
7244 * Partition sched domains as specified by the 'ndoms_new'
7245 * cpumasks in the array doms_new[] of cpumasks. This compares
7246 * doms_new[] to the current sched domain partitioning, doms_cur[].
7247 * It destroys each deleted domain and builds each new domain.
7249 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7250 * The masks don't intersect (don't overlap.) We should setup one
7251 * sched domain for each mask. CPUs not in any of the cpumasks will
7252 * not be load balanced. If the same cpumask appears both in the
7253 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7256 * The passed in 'doms_new' should be allocated using
7257 * alloc_sched_domains. This routine takes ownership of it and will
7258 * free_sched_domains it when done with it. If the caller failed the
7259 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7260 * and partition_sched_domains() will fallback to the single partition
7261 * 'fallback_doms', it also forces the domains to be rebuilt.
7263 * If doms_new == NULL it will be replaced with cpu_online_mask.
7264 * ndoms_new == 0 is a special case for destroying existing domains,
7265 * and it will not create the default domain.
7267 * Call with hotplug lock held
7269 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7270 struct sched_domain_attr *dattr_new)
7275 mutex_lock(&sched_domains_mutex);
7277 /* always unregister in case we don't destroy any domains */
7278 unregister_sched_domain_sysctl();
7280 /* Let architecture update cpu core mappings. */
7281 new_topology = arch_update_cpu_topology();
7283 n = doms_new ? ndoms_new : 0;
7285 /* Destroy deleted domains */
7286 for (i = 0; i < ndoms_cur; i++) {
7287 for (j = 0; j < n && !new_topology; j++) {
7288 if (cpumask_equal(doms_cur[i], doms_new[j])
7289 && dattrs_equal(dattr_cur, i, dattr_new, j))
7292 /* no match - a current sched domain not in new doms_new[] */
7293 detach_destroy_domains(doms_cur[i]);
7298 if (doms_new == NULL) {
7300 doms_new = &fallback_doms;
7301 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7302 WARN_ON_ONCE(dattr_new);
7305 /* Build new domains */
7306 for (i = 0; i < ndoms_new; i++) {
7307 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7308 if (cpumask_equal(doms_new[i], doms_cur[j])
7309 && dattrs_equal(dattr_new, i, dattr_cur, j))
7312 /* no match - add a new doms_new */
7313 __build_sched_domains(doms_new[i],
7314 dattr_new ? dattr_new + i : NULL);
7319 /* Remember the new sched domains */
7320 if (doms_cur != &fallback_doms)
7321 free_sched_domains(doms_cur, ndoms_cur);
7322 kfree(dattr_cur); /* kfree(NULL) is safe */
7323 doms_cur = doms_new;
7324 dattr_cur = dattr_new;
7325 ndoms_cur = ndoms_new;
7327 register_sched_domain_sysctl();
7329 mutex_unlock(&sched_domains_mutex);
7332 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7333 static void arch_reinit_sched_domains(void)
7337 /* Destroy domains first to force the rebuild */
7338 partition_sched_domains(0, NULL, NULL);
7340 rebuild_sched_domains();
7344 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7346 unsigned int level = 0;
7348 if (sscanf(buf, "%u", &level) != 1)
7352 * level is always be positive so don't check for
7353 * level < POWERSAVINGS_BALANCE_NONE which is 0
7354 * What happens on 0 or 1 byte write,
7355 * need to check for count as well?
7358 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7362 sched_smt_power_savings = level;
7364 sched_mc_power_savings = level;
7366 arch_reinit_sched_domains();
7371 #ifdef CONFIG_SCHED_MC
7372 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7375 return sprintf(page, "%u\n", sched_mc_power_savings);
7377 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7378 const char *buf, size_t count)
7380 return sched_power_savings_store(buf, count, 0);
7382 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7383 sched_mc_power_savings_show,
7384 sched_mc_power_savings_store);
7387 #ifdef CONFIG_SCHED_SMT
7388 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7391 return sprintf(page, "%u\n", sched_smt_power_savings);
7393 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7394 const char *buf, size_t count)
7396 return sched_power_savings_store(buf, count, 1);
7398 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7399 sched_smt_power_savings_show,
7400 sched_smt_power_savings_store);
7403 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7407 #ifdef CONFIG_SCHED_SMT
7409 err = sysfs_create_file(&cls->kset.kobj,
7410 &attr_sched_smt_power_savings.attr);
7412 #ifdef CONFIG_SCHED_MC
7413 if (!err && mc_capable())
7414 err = sysfs_create_file(&cls->kset.kobj,
7415 &attr_sched_mc_power_savings.attr);
7419 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7421 #ifndef CONFIG_CPUSETS
7423 * Add online and remove offline CPUs from the scheduler domains.
7424 * When cpusets are enabled they take over this function.
7426 static int update_sched_domains(struct notifier_block *nfb,
7427 unsigned long action, void *hcpu)
7431 case CPU_ONLINE_FROZEN:
7432 case CPU_DOWN_PREPARE:
7433 case CPU_DOWN_PREPARE_FROZEN:
7434 case CPU_DOWN_FAILED:
7435 case CPU_DOWN_FAILED_FROZEN:
7436 partition_sched_domains(1, NULL, NULL);
7445 static int update_runtime(struct notifier_block *nfb,
7446 unsigned long action, void *hcpu)
7448 int cpu = (int)(long)hcpu;
7451 case CPU_DOWN_PREPARE:
7452 case CPU_DOWN_PREPARE_FROZEN:
7453 disable_runtime(cpu_rq(cpu));
7456 case CPU_DOWN_FAILED:
7457 case CPU_DOWN_FAILED_FROZEN:
7459 case CPU_ONLINE_FROZEN:
7460 enable_runtime(cpu_rq(cpu));
7468 void __init sched_init_smp(void)
7470 cpumask_var_t non_isolated_cpus;
7472 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7473 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7475 #if defined(CONFIG_NUMA)
7476 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7478 BUG_ON(sched_group_nodes_bycpu == NULL);
7481 mutex_lock(&sched_domains_mutex);
7482 arch_init_sched_domains(cpu_active_mask);
7483 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7484 if (cpumask_empty(non_isolated_cpus))
7485 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7486 mutex_unlock(&sched_domains_mutex);
7489 #ifndef CONFIG_CPUSETS
7490 /* XXX: Theoretical race here - CPU may be hotplugged now */
7491 hotcpu_notifier(update_sched_domains, 0);
7494 /* RT runtime code needs to handle some hotplug events */
7495 hotcpu_notifier(update_runtime, 0);
7499 /* Move init over to a non-isolated CPU */
7500 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7502 sched_init_granularity();
7503 free_cpumask_var(non_isolated_cpus);
7505 init_sched_rt_class();
7508 void __init sched_init_smp(void)
7510 sched_init_granularity();
7512 #endif /* CONFIG_SMP */
7514 const_debug unsigned int sysctl_timer_migration = 1;
7516 int in_sched_functions(unsigned long addr)
7518 return in_lock_functions(addr) ||
7519 (addr >= (unsigned long)__sched_text_start
7520 && addr < (unsigned long)__sched_text_end);
7523 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7525 cfs_rq->tasks_timeline = RB_ROOT;
7526 INIT_LIST_HEAD(&cfs_rq->tasks);
7527 #ifdef CONFIG_FAIR_GROUP_SCHED
7530 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7533 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7535 struct rt_prio_array *array;
7538 array = &rt_rq->active;
7539 for (i = 0; i < MAX_RT_PRIO; i++) {
7540 INIT_LIST_HEAD(array->queue + i);
7541 __clear_bit(i, array->bitmap);
7543 /* delimiter for bitsearch: */
7544 __set_bit(MAX_RT_PRIO, array->bitmap);
7546 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7547 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7549 rt_rq->highest_prio.next = MAX_RT_PRIO;
7553 rt_rq->rt_nr_migratory = 0;
7554 rt_rq->overloaded = 0;
7555 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7559 rt_rq->rt_throttled = 0;
7560 rt_rq->rt_runtime = 0;
7561 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7563 #ifdef CONFIG_RT_GROUP_SCHED
7564 rt_rq->rt_nr_boosted = 0;
7569 #ifdef CONFIG_FAIR_GROUP_SCHED
7570 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7571 struct sched_entity *se, int cpu, int add,
7572 struct sched_entity *parent)
7574 struct rq *rq = cpu_rq(cpu);
7575 tg->cfs_rq[cpu] = cfs_rq;
7576 init_cfs_rq(cfs_rq, rq);
7579 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7582 /* se could be NULL for init_task_group */
7587 se->cfs_rq = &rq->cfs;
7589 se->cfs_rq = parent->my_q;
7592 se->load.weight = tg->shares;
7593 se->load.inv_weight = 0;
7594 se->parent = parent;
7598 #ifdef CONFIG_RT_GROUP_SCHED
7599 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7600 struct sched_rt_entity *rt_se, int cpu, int add,
7601 struct sched_rt_entity *parent)
7603 struct rq *rq = cpu_rq(cpu);
7605 tg->rt_rq[cpu] = rt_rq;
7606 init_rt_rq(rt_rq, rq);
7608 rt_rq->rt_se = rt_se;
7609 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7611 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7613 tg->rt_se[cpu] = rt_se;
7618 rt_se->rt_rq = &rq->rt;
7620 rt_se->rt_rq = parent->my_q;
7622 rt_se->my_q = rt_rq;
7623 rt_se->parent = parent;
7624 INIT_LIST_HEAD(&rt_se->run_list);
7628 void __init sched_init(void)
7631 unsigned long alloc_size = 0, ptr;
7633 #ifdef CONFIG_FAIR_GROUP_SCHED
7634 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7636 #ifdef CONFIG_RT_GROUP_SCHED
7637 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7639 #ifdef CONFIG_CPUMASK_OFFSTACK
7640 alloc_size += num_possible_cpus() * cpumask_size();
7643 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7645 #ifdef CONFIG_FAIR_GROUP_SCHED
7646 init_task_group.se = (struct sched_entity **)ptr;
7647 ptr += nr_cpu_ids * sizeof(void **);
7649 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7650 ptr += nr_cpu_ids * sizeof(void **);
7652 #endif /* CONFIG_FAIR_GROUP_SCHED */
7653 #ifdef CONFIG_RT_GROUP_SCHED
7654 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7655 ptr += nr_cpu_ids * sizeof(void **);
7657 init_task_group.rt_rq = (struct rt_rq **)ptr;
7658 ptr += nr_cpu_ids * sizeof(void **);
7660 #endif /* CONFIG_RT_GROUP_SCHED */
7661 #ifdef CONFIG_CPUMASK_OFFSTACK
7662 for_each_possible_cpu(i) {
7663 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7664 ptr += cpumask_size();
7666 #endif /* CONFIG_CPUMASK_OFFSTACK */
7670 init_defrootdomain();
7673 init_rt_bandwidth(&def_rt_bandwidth,
7674 global_rt_period(), global_rt_runtime());
7676 #ifdef CONFIG_RT_GROUP_SCHED
7677 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7678 global_rt_period(), global_rt_runtime());
7679 #endif /* CONFIG_RT_GROUP_SCHED */
7681 #ifdef CONFIG_CGROUP_SCHED
7682 list_add(&init_task_group.list, &task_groups);
7683 INIT_LIST_HEAD(&init_task_group.children);
7685 #endif /* CONFIG_CGROUP_SCHED */
7687 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7688 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7689 __alignof__(unsigned long));
7691 for_each_possible_cpu(i) {
7695 raw_spin_lock_init(&rq->lock);
7697 rq->calc_load_active = 0;
7698 rq->calc_load_update = jiffies + LOAD_FREQ;
7699 init_cfs_rq(&rq->cfs, rq);
7700 init_rt_rq(&rq->rt, rq);
7701 #ifdef CONFIG_FAIR_GROUP_SCHED
7702 init_task_group.shares = init_task_group_load;
7703 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7704 #ifdef CONFIG_CGROUP_SCHED
7706 * How much cpu bandwidth does init_task_group get?
7708 * In case of task-groups formed thr' the cgroup filesystem, it
7709 * gets 100% of the cpu resources in the system. This overall
7710 * system cpu resource is divided among the tasks of
7711 * init_task_group and its child task-groups in a fair manner,
7712 * based on each entity's (task or task-group's) weight
7713 * (se->load.weight).
7715 * In other words, if init_task_group has 10 tasks of weight
7716 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7717 * then A0's share of the cpu resource is:
7719 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7721 * We achieve this by letting init_task_group's tasks sit
7722 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7724 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7726 #endif /* CONFIG_FAIR_GROUP_SCHED */
7728 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7729 #ifdef CONFIG_RT_GROUP_SCHED
7730 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7731 #ifdef CONFIG_CGROUP_SCHED
7732 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7736 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7737 rq->cpu_load[j] = 0;
7741 rq->post_schedule = 0;
7742 rq->active_balance = 0;
7743 rq->next_balance = jiffies;
7747 rq->migration_thread = NULL;
7749 rq->avg_idle = 2*sysctl_sched_migration_cost;
7750 INIT_LIST_HEAD(&rq->migration_queue);
7751 rq_attach_root(rq, &def_root_domain);
7754 atomic_set(&rq->nr_iowait, 0);
7757 set_load_weight(&init_task);
7759 #ifdef CONFIG_PREEMPT_NOTIFIERS
7760 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7764 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7767 #ifdef CONFIG_RT_MUTEXES
7768 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7772 * The boot idle thread does lazy MMU switching as well:
7774 atomic_inc(&init_mm.mm_count);
7775 enter_lazy_tlb(&init_mm, current);
7778 * Make us the idle thread. Technically, schedule() should not be
7779 * called from this thread, however somewhere below it might be,
7780 * but because we are the idle thread, we just pick up running again
7781 * when this runqueue becomes "idle".
7783 init_idle(current, smp_processor_id());
7785 calc_load_update = jiffies + LOAD_FREQ;
7788 * During early bootup we pretend to be a normal task:
7790 current->sched_class = &fair_sched_class;
7792 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7793 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7796 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7797 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7799 /* May be allocated at isolcpus cmdline parse time */
7800 if (cpu_isolated_map == NULL)
7801 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7806 scheduler_running = 1;
7809 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7810 static inline int preempt_count_equals(int preempt_offset)
7812 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7814 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7817 void __might_sleep(const char *file, int line, int preempt_offset)
7820 static unsigned long prev_jiffy; /* ratelimiting */
7822 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7823 system_state != SYSTEM_RUNNING || oops_in_progress)
7825 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7827 prev_jiffy = jiffies;
7830 "BUG: sleeping function called from invalid context at %s:%d\n",
7833 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7834 in_atomic(), irqs_disabled(),
7835 current->pid, current->comm);
7837 debug_show_held_locks(current);
7838 if (irqs_disabled())
7839 print_irqtrace_events(current);
7843 EXPORT_SYMBOL(__might_sleep);
7846 #ifdef CONFIG_MAGIC_SYSRQ
7847 static void normalize_task(struct rq *rq, struct task_struct *p)
7851 update_rq_clock(rq);
7852 on_rq = p->se.on_rq;
7854 deactivate_task(rq, p, 0);
7855 __setscheduler(rq, p, SCHED_NORMAL, 0);
7857 activate_task(rq, p, 0);
7858 resched_task(rq->curr);
7862 void normalize_rt_tasks(void)
7864 struct task_struct *g, *p;
7865 unsigned long flags;
7868 read_lock_irqsave(&tasklist_lock, flags);
7869 do_each_thread(g, p) {
7871 * Only normalize user tasks:
7876 p->se.exec_start = 0;
7877 #ifdef CONFIG_SCHEDSTATS
7878 p->se.wait_start = 0;
7879 p->se.sleep_start = 0;
7880 p->se.block_start = 0;
7885 * Renice negative nice level userspace
7888 if (TASK_NICE(p) < 0 && p->mm)
7889 set_user_nice(p, 0);
7893 raw_spin_lock(&p->pi_lock);
7894 rq = __task_rq_lock(p);
7896 normalize_task(rq, p);
7898 __task_rq_unlock(rq);
7899 raw_spin_unlock(&p->pi_lock);
7900 } while_each_thread(g, p);
7902 read_unlock_irqrestore(&tasklist_lock, flags);
7905 #endif /* CONFIG_MAGIC_SYSRQ */
7909 * These functions are only useful for the IA64 MCA handling.
7911 * They can only be called when the whole system has been
7912 * stopped - every CPU needs to be quiescent, and no scheduling
7913 * activity can take place. Using them for anything else would
7914 * be a serious bug, and as a result, they aren't even visible
7915 * under any other configuration.
7919 * curr_task - return the current task for a given cpu.
7920 * @cpu: the processor in question.
7922 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7924 struct task_struct *curr_task(int cpu)
7926 return cpu_curr(cpu);
7930 * set_curr_task - set the current task for a given cpu.
7931 * @cpu: the processor in question.
7932 * @p: the task pointer to set.
7934 * Description: This function must only be used when non-maskable interrupts
7935 * are serviced on a separate stack. It allows the architecture to switch the
7936 * notion of the current task on a cpu in a non-blocking manner. This function
7937 * must be called with all CPU's synchronized, and interrupts disabled, the
7938 * and caller must save the original value of the current task (see
7939 * curr_task() above) and restore that value before reenabling interrupts and
7940 * re-starting the system.
7942 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7944 void set_curr_task(int cpu, struct task_struct *p)
7951 #ifdef CONFIG_FAIR_GROUP_SCHED
7952 static void free_fair_sched_group(struct task_group *tg)
7956 for_each_possible_cpu(i) {
7958 kfree(tg->cfs_rq[i]);
7968 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7970 struct cfs_rq *cfs_rq;
7971 struct sched_entity *se;
7975 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7978 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7982 tg->shares = NICE_0_LOAD;
7984 for_each_possible_cpu(i) {
7987 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7988 GFP_KERNEL, cpu_to_node(i));
7992 se = kzalloc_node(sizeof(struct sched_entity),
7993 GFP_KERNEL, cpu_to_node(i));
7997 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8008 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8010 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8011 &cpu_rq(cpu)->leaf_cfs_rq_list);
8014 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8016 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8018 #else /* !CONFG_FAIR_GROUP_SCHED */
8019 static inline void free_fair_sched_group(struct task_group *tg)
8024 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8029 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8033 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8036 #endif /* CONFIG_FAIR_GROUP_SCHED */
8038 #ifdef CONFIG_RT_GROUP_SCHED
8039 static void free_rt_sched_group(struct task_group *tg)
8043 destroy_rt_bandwidth(&tg->rt_bandwidth);
8045 for_each_possible_cpu(i) {
8047 kfree(tg->rt_rq[i]);
8049 kfree(tg->rt_se[i]);
8057 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8059 struct rt_rq *rt_rq;
8060 struct sched_rt_entity *rt_se;
8064 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8067 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8071 init_rt_bandwidth(&tg->rt_bandwidth,
8072 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8074 for_each_possible_cpu(i) {
8077 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8078 GFP_KERNEL, cpu_to_node(i));
8082 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8083 GFP_KERNEL, cpu_to_node(i));
8087 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8098 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8100 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8101 &cpu_rq(cpu)->leaf_rt_rq_list);
8104 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8106 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8108 #else /* !CONFIG_RT_GROUP_SCHED */
8109 static inline void free_rt_sched_group(struct task_group *tg)
8114 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8119 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8123 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8126 #endif /* CONFIG_RT_GROUP_SCHED */
8128 #ifdef CONFIG_CGROUP_SCHED
8129 static void free_sched_group(struct task_group *tg)
8131 free_fair_sched_group(tg);
8132 free_rt_sched_group(tg);
8136 /* allocate runqueue etc for a new task group */
8137 struct task_group *sched_create_group(struct task_group *parent)
8139 struct task_group *tg;
8140 unsigned long flags;
8143 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8145 return ERR_PTR(-ENOMEM);
8147 if (!alloc_fair_sched_group(tg, parent))
8150 if (!alloc_rt_sched_group(tg, parent))
8153 spin_lock_irqsave(&task_group_lock, flags);
8154 for_each_possible_cpu(i) {
8155 register_fair_sched_group(tg, i);
8156 register_rt_sched_group(tg, i);
8158 list_add_rcu(&tg->list, &task_groups);
8160 WARN_ON(!parent); /* root should already exist */
8162 tg->parent = parent;
8163 INIT_LIST_HEAD(&tg->children);
8164 list_add_rcu(&tg->siblings, &parent->children);
8165 spin_unlock_irqrestore(&task_group_lock, flags);
8170 free_sched_group(tg);
8171 return ERR_PTR(-ENOMEM);
8174 /* rcu callback to free various structures associated with a task group */
8175 static void free_sched_group_rcu(struct rcu_head *rhp)
8177 /* now it should be safe to free those cfs_rqs */
8178 free_sched_group(container_of(rhp, struct task_group, rcu));
8181 /* Destroy runqueue etc associated with a task group */
8182 void sched_destroy_group(struct task_group *tg)
8184 unsigned long flags;
8187 spin_lock_irqsave(&task_group_lock, flags);
8188 for_each_possible_cpu(i) {
8189 unregister_fair_sched_group(tg, i);
8190 unregister_rt_sched_group(tg, i);
8192 list_del_rcu(&tg->list);
8193 list_del_rcu(&tg->siblings);
8194 spin_unlock_irqrestore(&task_group_lock, flags);
8196 /* wait for possible concurrent references to cfs_rqs complete */
8197 call_rcu(&tg->rcu, free_sched_group_rcu);
8200 /* change task's runqueue when it moves between groups.
8201 * The caller of this function should have put the task in its new group
8202 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8203 * reflect its new group.
8205 void sched_move_task(struct task_struct *tsk)
8208 unsigned long flags;
8211 rq = task_rq_lock(tsk, &flags);
8213 update_rq_clock(rq);
8215 running = task_current(rq, tsk);
8216 on_rq = tsk->se.on_rq;
8219 dequeue_task(rq, tsk, 0);
8220 if (unlikely(running))
8221 tsk->sched_class->put_prev_task(rq, tsk);
8223 set_task_rq(tsk, task_cpu(tsk));
8225 #ifdef CONFIG_FAIR_GROUP_SCHED
8226 if (tsk->sched_class->moved_group)
8227 tsk->sched_class->moved_group(tsk, on_rq);
8230 if (unlikely(running))
8231 tsk->sched_class->set_curr_task(rq);
8233 enqueue_task(rq, tsk, 0);
8235 task_rq_unlock(rq, &flags);
8237 #endif /* CONFIG_CGROUP_SCHED */
8239 #ifdef CONFIG_FAIR_GROUP_SCHED
8240 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8242 struct cfs_rq *cfs_rq = se->cfs_rq;
8247 dequeue_entity(cfs_rq, se, 0);
8249 se->load.weight = shares;
8250 se->load.inv_weight = 0;
8253 enqueue_entity(cfs_rq, se, 0);
8256 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8258 struct cfs_rq *cfs_rq = se->cfs_rq;
8259 struct rq *rq = cfs_rq->rq;
8260 unsigned long flags;
8262 raw_spin_lock_irqsave(&rq->lock, flags);
8263 __set_se_shares(se, shares);
8264 raw_spin_unlock_irqrestore(&rq->lock, flags);
8267 static DEFINE_MUTEX(shares_mutex);
8269 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8272 unsigned long flags;
8275 * We can't change the weight of the root cgroup.
8280 if (shares < MIN_SHARES)
8281 shares = MIN_SHARES;
8282 else if (shares > MAX_SHARES)
8283 shares = MAX_SHARES;
8285 mutex_lock(&shares_mutex);
8286 if (tg->shares == shares)
8289 spin_lock_irqsave(&task_group_lock, flags);
8290 for_each_possible_cpu(i)
8291 unregister_fair_sched_group(tg, i);
8292 list_del_rcu(&tg->siblings);
8293 spin_unlock_irqrestore(&task_group_lock, flags);
8295 /* wait for any ongoing reference to this group to finish */
8296 synchronize_sched();
8299 * Now we are free to modify the group's share on each cpu
8300 * w/o tripping rebalance_share or load_balance_fair.
8302 tg->shares = shares;
8303 for_each_possible_cpu(i) {
8307 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8308 set_se_shares(tg->se[i], shares);
8312 * Enable load balance activity on this group, by inserting it back on
8313 * each cpu's rq->leaf_cfs_rq_list.
8315 spin_lock_irqsave(&task_group_lock, flags);
8316 for_each_possible_cpu(i)
8317 register_fair_sched_group(tg, i);
8318 list_add_rcu(&tg->siblings, &tg->parent->children);
8319 spin_unlock_irqrestore(&task_group_lock, flags);
8321 mutex_unlock(&shares_mutex);
8325 unsigned long sched_group_shares(struct task_group *tg)
8331 #ifdef CONFIG_RT_GROUP_SCHED
8333 * Ensure that the real time constraints are schedulable.
8335 static DEFINE_MUTEX(rt_constraints_mutex);
8337 static unsigned long to_ratio(u64 period, u64 runtime)
8339 if (runtime == RUNTIME_INF)
8342 return div64_u64(runtime << 20, period);
8345 /* Must be called with tasklist_lock held */
8346 static inline int tg_has_rt_tasks(struct task_group *tg)
8348 struct task_struct *g, *p;
8350 do_each_thread(g, p) {
8351 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8353 } while_each_thread(g, p);
8358 struct rt_schedulable_data {
8359 struct task_group *tg;
8364 static int tg_schedulable(struct task_group *tg, void *data)
8366 struct rt_schedulable_data *d = data;
8367 struct task_group *child;
8368 unsigned long total, sum = 0;
8369 u64 period, runtime;
8371 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8372 runtime = tg->rt_bandwidth.rt_runtime;
8375 period = d->rt_period;
8376 runtime = d->rt_runtime;
8380 * Cannot have more runtime than the period.
8382 if (runtime > period && runtime != RUNTIME_INF)
8386 * Ensure we don't starve existing RT tasks.
8388 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8391 total = to_ratio(period, runtime);
8394 * Nobody can have more than the global setting allows.
8396 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8400 * The sum of our children's runtime should not exceed our own.
8402 list_for_each_entry_rcu(child, &tg->children, siblings) {
8403 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8404 runtime = child->rt_bandwidth.rt_runtime;
8406 if (child == d->tg) {
8407 period = d->rt_period;
8408 runtime = d->rt_runtime;
8411 sum += to_ratio(period, runtime);
8420 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8422 struct rt_schedulable_data data = {
8424 .rt_period = period,
8425 .rt_runtime = runtime,
8428 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8431 static int tg_set_bandwidth(struct task_group *tg,
8432 u64 rt_period, u64 rt_runtime)
8436 mutex_lock(&rt_constraints_mutex);
8437 read_lock(&tasklist_lock);
8438 err = __rt_schedulable(tg, rt_period, rt_runtime);
8442 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8443 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8444 tg->rt_bandwidth.rt_runtime = rt_runtime;
8446 for_each_possible_cpu(i) {
8447 struct rt_rq *rt_rq = tg->rt_rq[i];
8449 raw_spin_lock(&rt_rq->rt_runtime_lock);
8450 rt_rq->rt_runtime = rt_runtime;
8451 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8453 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8455 read_unlock(&tasklist_lock);
8456 mutex_unlock(&rt_constraints_mutex);
8461 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8463 u64 rt_runtime, rt_period;
8465 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8466 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8467 if (rt_runtime_us < 0)
8468 rt_runtime = RUNTIME_INF;
8470 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8473 long sched_group_rt_runtime(struct task_group *tg)
8477 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8480 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8481 do_div(rt_runtime_us, NSEC_PER_USEC);
8482 return rt_runtime_us;
8485 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8487 u64 rt_runtime, rt_period;
8489 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8490 rt_runtime = tg->rt_bandwidth.rt_runtime;
8495 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8498 long sched_group_rt_period(struct task_group *tg)
8502 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8503 do_div(rt_period_us, NSEC_PER_USEC);
8504 return rt_period_us;
8507 static int sched_rt_global_constraints(void)
8509 u64 runtime, period;
8512 if (sysctl_sched_rt_period <= 0)
8515 runtime = global_rt_runtime();
8516 period = global_rt_period();
8519 * Sanity check on the sysctl variables.
8521 if (runtime > period && runtime != RUNTIME_INF)
8524 mutex_lock(&rt_constraints_mutex);
8525 read_lock(&tasklist_lock);
8526 ret = __rt_schedulable(NULL, 0, 0);
8527 read_unlock(&tasklist_lock);
8528 mutex_unlock(&rt_constraints_mutex);
8533 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8535 /* Don't accept realtime tasks when there is no way for them to run */
8536 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8542 #else /* !CONFIG_RT_GROUP_SCHED */
8543 static int sched_rt_global_constraints(void)
8545 unsigned long flags;
8548 if (sysctl_sched_rt_period <= 0)
8552 * There's always some RT tasks in the root group
8553 * -- migration, kstopmachine etc..
8555 if (sysctl_sched_rt_runtime == 0)
8558 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8559 for_each_possible_cpu(i) {
8560 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8562 raw_spin_lock(&rt_rq->rt_runtime_lock);
8563 rt_rq->rt_runtime = global_rt_runtime();
8564 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8566 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8570 #endif /* CONFIG_RT_GROUP_SCHED */
8572 int sched_rt_handler(struct ctl_table *table, int write,
8573 void __user *buffer, size_t *lenp,
8577 int old_period, old_runtime;
8578 static DEFINE_MUTEX(mutex);
8581 old_period = sysctl_sched_rt_period;
8582 old_runtime = sysctl_sched_rt_runtime;
8584 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8586 if (!ret && write) {
8587 ret = sched_rt_global_constraints();
8589 sysctl_sched_rt_period = old_period;
8590 sysctl_sched_rt_runtime = old_runtime;
8592 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8593 def_rt_bandwidth.rt_period =
8594 ns_to_ktime(global_rt_period());
8597 mutex_unlock(&mutex);
8602 #ifdef CONFIG_CGROUP_SCHED
8604 /* return corresponding task_group object of a cgroup */
8605 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8607 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8608 struct task_group, css);
8611 static struct cgroup_subsys_state *
8612 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8614 struct task_group *tg, *parent;
8616 if (!cgrp->parent) {
8617 /* This is early initialization for the top cgroup */
8618 return &init_task_group.css;
8621 parent = cgroup_tg(cgrp->parent);
8622 tg = sched_create_group(parent);
8624 return ERR_PTR(-ENOMEM);
8630 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8632 struct task_group *tg = cgroup_tg(cgrp);
8634 sched_destroy_group(tg);
8638 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8640 #ifdef CONFIG_RT_GROUP_SCHED
8641 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8644 /* We don't support RT-tasks being in separate groups */
8645 if (tsk->sched_class != &fair_sched_class)
8652 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8653 struct task_struct *tsk, bool threadgroup)
8655 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8659 struct task_struct *c;
8661 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8662 retval = cpu_cgroup_can_attach_task(cgrp, c);
8674 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8675 struct cgroup *old_cont, struct task_struct *tsk,
8678 sched_move_task(tsk);
8680 struct task_struct *c;
8682 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8689 #ifdef CONFIG_FAIR_GROUP_SCHED
8690 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8693 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8696 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8698 struct task_group *tg = cgroup_tg(cgrp);
8700 return (u64) tg->shares;
8702 #endif /* CONFIG_FAIR_GROUP_SCHED */
8704 #ifdef CONFIG_RT_GROUP_SCHED
8705 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8708 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8711 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8713 return sched_group_rt_runtime(cgroup_tg(cgrp));
8716 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8719 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8722 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8724 return sched_group_rt_period(cgroup_tg(cgrp));
8726 #endif /* CONFIG_RT_GROUP_SCHED */
8728 static struct cftype cpu_files[] = {
8729 #ifdef CONFIG_FAIR_GROUP_SCHED
8732 .read_u64 = cpu_shares_read_u64,
8733 .write_u64 = cpu_shares_write_u64,
8736 #ifdef CONFIG_RT_GROUP_SCHED
8738 .name = "rt_runtime_us",
8739 .read_s64 = cpu_rt_runtime_read,
8740 .write_s64 = cpu_rt_runtime_write,
8743 .name = "rt_period_us",
8744 .read_u64 = cpu_rt_period_read_uint,
8745 .write_u64 = cpu_rt_period_write_uint,
8750 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8752 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8755 struct cgroup_subsys cpu_cgroup_subsys = {
8757 .create = cpu_cgroup_create,
8758 .destroy = cpu_cgroup_destroy,
8759 .can_attach = cpu_cgroup_can_attach,
8760 .attach = cpu_cgroup_attach,
8761 .populate = cpu_cgroup_populate,
8762 .subsys_id = cpu_cgroup_subsys_id,
8766 #endif /* CONFIG_CGROUP_SCHED */
8768 #ifdef CONFIG_CGROUP_CPUACCT
8771 * CPU accounting code for task groups.
8773 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8774 * (balbir@in.ibm.com).
8777 /* track cpu usage of a group of tasks and its child groups */
8779 struct cgroup_subsys_state css;
8780 /* cpuusage holds pointer to a u64-type object on every cpu */
8782 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8783 struct cpuacct *parent;
8786 struct cgroup_subsys cpuacct_subsys;
8788 /* return cpu accounting group corresponding to this container */
8789 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8791 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8792 struct cpuacct, css);
8795 /* return cpu accounting group to which this task belongs */
8796 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8798 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8799 struct cpuacct, css);
8802 /* create a new cpu accounting group */
8803 static struct cgroup_subsys_state *cpuacct_create(
8804 struct cgroup_subsys *ss, struct cgroup *cgrp)
8806 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8812 ca->cpuusage = alloc_percpu(u64);
8816 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8817 if (percpu_counter_init(&ca->cpustat[i], 0))
8818 goto out_free_counters;
8821 ca->parent = cgroup_ca(cgrp->parent);
8827 percpu_counter_destroy(&ca->cpustat[i]);
8828 free_percpu(ca->cpuusage);
8832 return ERR_PTR(-ENOMEM);
8835 /* destroy an existing cpu accounting group */
8837 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8839 struct cpuacct *ca = cgroup_ca(cgrp);
8842 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8843 percpu_counter_destroy(&ca->cpustat[i]);
8844 free_percpu(ca->cpuusage);
8848 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8850 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8853 #ifndef CONFIG_64BIT
8855 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8857 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8859 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8867 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8869 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8871 #ifndef CONFIG_64BIT
8873 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8875 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8877 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8883 /* return total cpu usage (in nanoseconds) of a group */
8884 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8886 struct cpuacct *ca = cgroup_ca(cgrp);
8887 u64 totalcpuusage = 0;
8890 for_each_present_cpu(i)
8891 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8893 return totalcpuusage;
8896 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8899 struct cpuacct *ca = cgroup_ca(cgrp);
8908 for_each_present_cpu(i)
8909 cpuacct_cpuusage_write(ca, i, 0);
8915 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8918 struct cpuacct *ca = cgroup_ca(cgroup);
8922 for_each_present_cpu(i) {
8923 percpu = cpuacct_cpuusage_read(ca, i);
8924 seq_printf(m, "%llu ", (unsigned long long) percpu);
8926 seq_printf(m, "\n");
8930 static const char *cpuacct_stat_desc[] = {
8931 [CPUACCT_STAT_USER] = "user",
8932 [CPUACCT_STAT_SYSTEM] = "system",
8935 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8936 struct cgroup_map_cb *cb)
8938 struct cpuacct *ca = cgroup_ca(cgrp);
8941 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8942 s64 val = percpu_counter_read(&ca->cpustat[i]);
8943 val = cputime64_to_clock_t(val);
8944 cb->fill(cb, cpuacct_stat_desc[i], val);
8949 static struct cftype files[] = {
8952 .read_u64 = cpuusage_read,
8953 .write_u64 = cpuusage_write,
8956 .name = "usage_percpu",
8957 .read_seq_string = cpuacct_percpu_seq_read,
8961 .read_map = cpuacct_stats_show,
8965 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8967 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8971 * charge this task's execution time to its accounting group.
8973 * called with rq->lock held.
8975 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8980 if (unlikely(!cpuacct_subsys.active))
8983 cpu = task_cpu(tsk);
8989 for (; ca; ca = ca->parent) {
8990 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8991 *cpuusage += cputime;
8998 * Charge the system/user time to the task's accounting group.
9000 static void cpuacct_update_stats(struct task_struct *tsk,
9001 enum cpuacct_stat_index idx, cputime_t val)
9005 if (unlikely(!cpuacct_subsys.active))
9012 percpu_counter_add(&ca->cpustat[idx], val);
9018 struct cgroup_subsys cpuacct_subsys = {
9020 .create = cpuacct_create,
9021 .destroy = cpuacct_destroy,
9022 .populate = cpuacct_populate,
9023 .subsys_id = cpuacct_subsys_id,
9025 #endif /* CONFIG_CGROUP_CPUACCT */
9029 int rcu_expedited_torture_stats(char *page)
9033 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9035 void synchronize_sched_expedited(void)
9038 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9040 #else /* #ifndef CONFIG_SMP */
9042 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9043 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9045 #define RCU_EXPEDITED_STATE_POST -2
9046 #define RCU_EXPEDITED_STATE_IDLE -1
9048 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9050 int rcu_expedited_torture_stats(char *page)
9055 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9056 for_each_online_cpu(cpu) {
9057 cnt += sprintf(&page[cnt], " %d:%d",
9058 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9060 cnt += sprintf(&page[cnt], "\n");
9063 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9065 static long synchronize_sched_expedited_count;
9068 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9069 * approach to force grace period to end quickly. This consumes
9070 * significant time on all CPUs, and is thus not recommended for
9071 * any sort of common-case code.
9073 * Note that it is illegal to call this function while holding any
9074 * lock that is acquired by a CPU-hotplug notifier. Failing to
9075 * observe this restriction will result in deadlock.
9077 void synchronize_sched_expedited(void)
9080 unsigned long flags;
9081 bool need_full_sync = 0;
9083 struct migration_req *req;
9087 smp_mb(); /* ensure prior mod happens before capturing snap. */
9088 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9090 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9092 if (trycount++ < 10)
9093 udelay(trycount * num_online_cpus());
9095 synchronize_sched();
9098 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9099 smp_mb(); /* ensure test happens before caller kfree */
9104 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9105 for_each_online_cpu(cpu) {
9107 req = &per_cpu(rcu_migration_req, cpu);
9108 init_completion(&req->done);
9110 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9111 raw_spin_lock_irqsave(&rq->lock, flags);
9112 list_add(&req->list, &rq->migration_queue);
9113 raw_spin_unlock_irqrestore(&rq->lock, flags);
9114 wake_up_process(rq->migration_thread);
9116 for_each_online_cpu(cpu) {
9117 rcu_expedited_state = cpu;
9118 req = &per_cpu(rcu_migration_req, cpu);
9120 wait_for_completion(&req->done);
9121 raw_spin_lock_irqsave(&rq->lock, flags);
9122 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9124 req->dest_cpu = RCU_MIGRATION_IDLE;
9125 raw_spin_unlock_irqrestore(&rq->lock, flags);
9127 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9128 synchronize_sched_expedited_count++;
9129 mutex_unlock(&rcu_sched_expedited_mutex);
9132 synchronize_sched();
9134 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9136 #endif /* #else #ifndef CONFIG_SMP */