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;
1860 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1863 p->se.start_runtime = p->se.sum_exec_runtime;
1865 sched_info_queued(p);
1866 p->sched_class->enqueue_task(rq, p, wakeup, head);
1870 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1873 if (p->se.last_wakeup) {
1874 update_avg(&p->se.avg_overlap,
1875 p->se.sum_exec_runtime - p->se.last_wakeup);
1876 p->se.last_wakeup = 0;
1878 update_avg(&p->se.avg_wakeup,
1879 sysctl_sched_wakeup_granularity);
1883 sched_info_dequeued(p);
1884 p->sched_class->dequeue_task(rq, p, sleep);
1889 * activate_task - move a task to the runqueue.
1891 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1893 if (task_contributes_to_load(p))
1894 rq->nr_uninterruptible--;
1896 enqueue_task(rq, p, wakeup, false);
1901 * deactivate_task - remove a task from the runqueue.
1903 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1905 if (task_contributes_to_load(p))
1906 rq->nr_uninterruptible++;
1908 dequeue_task(rq, p, sleep);
1912 #include "sched_idletask.c"
1913 #include "sched_fair.c"
1914 #include "sched_rt.c"
1915 #ifdef CONFIG_SCHED_DEBUG
1916 # include "sched_debug.c"
1920 * __normal_prio - return the priority that is based on the static prio
1922 static inline int __normal_prio(struct task_struct *p)
1924 return p->static_prio;
1928 * Calculate the expected normal priority: i.e. priority
1929 * without taking RT-inheritance into account. Might be
1930 * boosted by interactivity modifiers. Changes upon fork,
1931 * setprio syscalls, and whenever the interactivity
1932 * estimator recalculates.
1934 static inline int normal_prio(struct task_struct *p)
1938 if (task_has_rt_policy(p))
1939 prio = MAX_RT_PRIO-1 - p->rt_priority;
1941 prio = __normal_prio(p);
1946 * Calculate the current priority, i.e. the priority
1947 * taken into account by the scheduler. This value might
1948 * be boosted by RT tasks, or might be boosted by
1949 * interactivity modifiers. Will be RT if the task got
1950 * RT-boosted. If not then it returns p->normal_prio.
1952 static int effective_prio(struct task_struct *p)
1954 p->normal_prio = normal_prio(p);
1956 * If we are RT tasks or we were boosted to RT priority,
1957 * keep the priority unchanged. Otherwise, update priority
1958 * to the normal priority:
1960 if (!rt_prio(p->prio))
1961 return p->normal_prio;
1966 * task_curr - is this task currently executing on a CPU?
1967 * @p: the task in question.
1969 inline int task_curr(const struct task_struct *p)
1971 return cpu_curr(task_cpu(p)) == p;
1974 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1975 const struct sched_class *prev_class,
1976 int oldprio, int running)
1978 if (prev_class != p->sched_class) {
1979 if (prev_class->switched_from)
1980 prev_class->switched_from(rq, p, running);
1981 p->sched_class->switched_to(rq, p, running);
1983 p->sched_class->prio_changed(rq, p, oldprio, running);
1988 * Is this task likely cache-hot:
1991 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1995 if (p->sched_class != &fair_sched_class)
1999 * Buddy candidates are cache hot:
2001 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2002 (&p->se == cfs_rq_of(&p->se)->next ||
2003 &p->se == cfs_rq_of(&p->se)->last))
2006 if (sysctl_sched_migration_cost == -1)
2008 if (sysctl_sched_migration_cost == 0)
2011 delta = now - p->se.exec_start;
2013 return delta < (s64)sysctl_sched_migration_cost;
2016 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2018 #ifdef CONFIG_SCHED_DEBUG
2020 * We should never call set_task_cpu() on a blocked task,
2021 * ttwu() will sort out the placement.
2023 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2024 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2027 trace_sched_migrate_task(p, new_cpu);
2029 if (task_cpu(p) != new_cpu) {
2030 p->se.nr_migrations++;
2031 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2034 __set_task_cpu(p, new_cpu);
2037 struct migration_req {
2038 struct list_head list;
2040 struct task_struct *task;
2043 struct completion done;
2047 * The task's runqueue lock must be held.
2048 * Returns true if you have to wait for migration thread.
2051 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2053 struct rq *rq = task_rq(p);
2056 * If the task is not on a runqueue (and not running), then
2057 * the next wake-up will properly place the task.
2059 if (!p->se.on_rq && !task_running(rq, p))
2062 init_completion(&req->done);
2064 req->dest_cpu = dest_cpu;
2065 list_add(&req->list, &rq->migration_queue);
2071 * wait_task_context_switch - wait for a thread to complete at least one
2074 * @p must not be current.
2076 void wait_task_context_switch(struct task_struct *p)
2078 unsigned long nvcsw, nivcsw, flags;
2086 * The runqueue is assigned before the actual context
2087 * switch. We need to take the runqueue lock.
2089 * We could check initially without the lock but it is
2090 * very likely that we need to take the lock in every
2093 rq = task_rq_lock(p, &flags);
2094 running = task_running(rq, p);
2095 task_rq_unlock(rq, &flags);
2097 if (likely(!running))
2100 * The switch count is incremented before the actual
2101 * context switch. We thus wait for two switches to be
2102 * sure at least one completed.
2104 if ((p->nvcsw - nvcsw) > 1)
2106 if ((p->nivcsw - nivcsw) > 1)
2114 * wait_task_inactive - wait for a thread to unschedule.
2116 * If @match_state is nonzero, it's the @p->state value just checked and
2117 * not expected to change. If it changes, i.e. @p might have woken up,
2118 * then return zero. When we succeed in waiting for @p to be off its CPU,
2119 * we return a positive number (its total switch count). If a second call
2120 * a short while later returns the same number, the caller can be sure that
2121 * @p has remained unscheduled the whole time.
2123 * The caller must ensure that the task *will* unschedule sometime soon,
2124 * else this function might spin for a *long* time. This function can't
2125 * be called with interrupts off, or it may introduce deadlock with
2126 * smp_call_function() if an IPI is sent by the same process we are
2127 * waiting to become inactive.
2129 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2131 unsigned long flags;
2138 * We do the initial early heuristics without holding
2139 * any task-queue locks at all. We'll only try to get
2140 * the runqueue lock when things look like they will
2146 * If the task is actively running on another CPU
2147 * still, just relax and busy-wait without holding
2150 * NOTE! Since we don't hold any locks, it's not
2151 * even sure that "rq" stays as the right runqueue!
2152 * But we don't care, since "task_running()" will
2153 * return false if the runqueue has changed and p
2154 * is actually now running somewhere else!
2156 while (task_running(rq, p)) {
2157 if (match_state && unlikely(p->state != match_state))
2163 * Ok, time to look more closely! We need the rq
2164 * lock now, to be *sure*. If we're wrong, we'll
2165 * just go back and repeat.
2167 rq = task_rq_lock(p, &flags);
2168 trace_sched_wait_task(rq, p);
2169 running = task_running(rq, p);
2170 on_rq = p->se.on_rq;
2172 if (!match_state || p->state == match_state)
2173 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2174 task_rq_unlock(rq, &flags);
2177 * If it changed from the expected state, bail out now.
2179 if (unlikely(!ncsw))
2183 * Was it really running after all now that we
2184 * checked with the proper locks actually held?
2186 * Oops. Go back and try again..
2188 if (unlikely(running)) {
2194 * It's not enough that it's not actively running,
2195 * it must be off the runqueue _entirely_, and not
2198 * So if it was still runnable (but just not actively
2199 * running right now), it's preempted, and we should
2200 * yield - it could be a while.
2202 if (unlikely(on_rq)) {
2203 schedule_timeout_uninterruptible(1);
2208 * Ahh, all good. It wasn't running, and it wasn't
2209 * runnable, which means that it will never become
2210 * running in the future either. We're all done!
2219 * kick_process - kick a running thread to enter/exit the kernel
2220 * @p: the to-be-kicked thread
2222 * Cause a process which is running on another CPU to enter
2223 * kernel-mode, without any delay. (to get signals handled.)
2225 * NOTE: this function doesnt have to take the runqueue lock,
2226 * because all it wants to ensure is that the remote task enters
2227 * the kernel. If the IPI races and the task has been migrated
2228 * to another CPU then no harm is done and the purpose has been
2231 void kick_process(struct task_struct *p)
2237 if ((cpu != smp_processor_id()) && task_curr(p))
2238 smp_send_reschedule(cpu);
2241 EXPORT_SYMBOL_GPL(kick_process);
2242 #endif /* CONFIG_SMP */
2245 * task_oncpu_function_call - call a function on the cpu on which a task runs
2246 * @p: the task to evaluate
2247 * @func: the function to be called
2248 * @info: the function call argument
2250 * Calls the function @func when the task is currently running. This might
2251 * be on the current CPU, which just calls the function directly
2253 void task_oncpu_function_call(struct task_struct *p,
2254 void (*func) (void *info), void *info)
2261 smp_call_function_single(cpu, func, info, 1);
2266 static int select_fallback_rq(int cpu, struct task_struct *p)
2269 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2271 /* Look for allowed, online CPU in same node. */
2272 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2273 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2276 /* Any allowed, online CPU? */
2277 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2278 if (dest_cpu < nr_cpu_ids)
2281 /* No more Mr. Nice Guy. */
2282 if (dest_cpu >= nr_cpu_ids) {
2284 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2286 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2289 * Don't tell them about moving exiting tasks or
2290 * kernel threads (both mm NULL), since they never
2293 if (p->mm && printk_ratelimit()) {
2294 printk(KERN_INFO "process %d (%s) no "
2295 "longer affine to cpu%d\n",
2296 task_pid_nr(p), p->comm, cpu);
2306 * - fork, @p is stable because it isn't on the tasklist yet
2308 * - exec, @p is unstable, retry loop
2310 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so
2311 * we should be good.
2314 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2316 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2319 * In order not to call set_task_cpu() on a blocking task we need
2320 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2323 * Since this is common to all placement strategies, this lives here.
2325 * [ this allows ->select_task() to simply return task_cpu(p) and
2326 * not worry about this generic constraint ]
2328 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2330 cpu = select_fallback_rq(task_cpu(p), p);
2337 * try_to_wake_up - wake up a thread
2338 * @p: the to-be-woken-up thread
2339 * @state: the mask of task states that can be woken
2340 * @sync: do a synchronous wakeup?
2342 * Put it on the run-queue if it's not already there. The "current"
2343 * thread is always on the run-queue (except when the actual
2344 * re-schedule is in progress), and as such you're allowed to do
2345 * the simpler "current->state = TASK_RUNNING" to mark yourself
2346 * runnable without the overhead of this.
2348 * returns failure only if the task is already active.
2350 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2353 int cpu, orig_cpu, this_cpu, success = 0;
2354 unsigned long flags;
2355 struct rq *rq, *orig_rq;
2357 if (!sched_feat(SYNC_WAKEUPS))
2358 wake_flags &= ~WF_SYNC;
2360 this_cpu = get_cpu();
2363 rq = orig_rq = task_rq_lock(p, &flags);
2364 update_rq_clock(rq);
2365 if (!(p->state & state))
2375 if (unlikely(task_running(rq, p)))
2379 * In order to handle concurrent wakeups and release the rq->lock
2380 * we put the task in TASK_WAKING state.
2382 * First fix up the nr_uninterruptible count:
2384 if (task_contributes_to_load(p))
2385 rq->nr_uninterruptible--;
2386 p->state = TASK_WAKING;
2388 if (p->sched_class->task_waking)
2389 p->sched_class->task_waking(rq, p);
2391 __task_rq_unlock(rq);
2393 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2394 if (cpu != orig_cpu)
2395 set_task_cpu(p, cpu);
2397 rq = __task_rq_lock(p);
2398 update_rq_clock(rq);
2400 WARN_ON(p->state != TASK_WAKING);
2403 #ifdef CONFIG_SCHEDSTATS
2404 schedstat_inc(rq, ttwu_count);
2405 if (cpu == this_cpu)
2406 schedstat_inc(rq, ttwu_local);
2408 struct sched_domain *sd;
2409 for_each_domain(this_cpu, sd) {
2410 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2411 schedstat_inc(sd, ttwu_wake_remote);
2416 #endif /* CONFIG_SCHEDSTATS */
2419 #endif /* CONFIG_SMP */
2420 schedstat_inc(p, se.nr_wakeups);
2421 if (wake_flags & WF_SYNC)
2422 schedstat_inc(p, se.nr_wakeups_sync);
2423 if (orig_cpu != cpu)
2424 schedstat_inc(p, se.nr_wakeups_migrate);
2425 if (cpu == this_cpu)
2426 schedstat_inc(p, se.nr_wakeups_local);
2428 schedstat_inc(p, se.nr_wakeups_remote);
2429 activate_task(rq, p, 1);
2433 * Only attribute actual wakeups done by this task.
2435 if (!in_interrupt()) {
2436 struct sched_entity *se = ¤t->se;
2437 u64 sample = se->sum_exec_runtime;
2439 if (se->last_wakeup)
2440 sample -= se->last_wakeup;
2442 sample -= se->start_runtime;
2443 update_avg(&se->avg_wakeup, sample);
2445 se->last_wakeup = se->sum_exec_runtime;
2449 trace_sched_wakeup(rq, p, success);
2450 check_preempt_curr(rq, p, wake_flags);
2452 p->state = TASK_RUNNING;
2454 if (p->sched_class->task_woken)
2455 p->sched_class->task_woken(rq, p);
2457 if (unlikely(rq->idle_stamp)) {
2458 u64 delta = rq->clock - rq->idle_stamp;
2459 u64 max = 2*sysctl_sched_migration_cost;
2464 update_avg(&rq->avg_idle, delta);
2469 task_rq_unlock(rq, &flags);
2476 * wake_up_process - Wake up a specific process
2477 * @p: The process to be woken up.
2479 * Attempt to wake up the nominated process and move it to the set of runnable
2480 * processes. Returns 1 if the process was woken up, 0 if it was already
2483 * It may be assumed that this function implies a write memory barrier before
2484 * changing the task state if and only if any tasks are woken up.
2486 int wake_up_process(struct task_struct *p)
2488 return try_to_wake_up(p, TASK_ALL, 0);
2490 EXPORT_SYMBOL(wake_up_process);
2492 int wake_up_state(struct task_struct *p, unsigned int state)
2494 return try_to_wake_up(p, state, 0);
2498 * Perform scheduler related setup for a newly forked process p.
2499 * p is forked by current.
2501 * __sched_fork() is basic setup used by init_idle() too:
2503 static void __sched_fork(struct task_struct *p)
2505 p->se.exec_start = 0;
2506 p->se.sum_exec_runtime = 0;
2507 p->se.prev_sum_exec_runtime = 0;
2508 p->se.nr_migrations = 0;
2509 p->se.last_wakeup = 0;
2510 p->se.avg_overlap = 0;
2511 p->se.start_runtime = 0;
2512 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2514 #ifdef CONFIG_SCHEDSTATS
2515 p->se.wait_start = 0;
2517 p->se.wait_count = 0;
2520 p->se.sleep_start = 0;
2521 p->se.sleep_max = 0;
2522 p->se.sum_sleep_runtime = 0;
2524 p->se.block_start = 0;
2525 p->se.block_max = 0;
2527 p->se.slice_max = 0;
2529 p->se.nr_migrations_cold = 0;
2530 p->se.nr_failed_migrations_affine = 0;
2531 p->se.nr_failed_migrations_running = 0;
2532 p->se.nr_failed_migrations_hot = 0;
2533 p->se.nr_forced_migrations = 0;
2535 p->se.nr_wakeups = 0;
2536 p->se.nr_wakeups_sync = 0;
2537 p->se.nr_wakeups_migrate = 0;
2538 p->se.nr_wakeups_local = 0;
2539 p->se.nr_wakeups_remote = 0;
2540 p->se.nr_wakeups_affine = 0;
2541 p->se.nr_wakeups_affine_attempts = 0;
2542 p->se.nr_wakeups_passive = 0;
2543 p->se.nr_wakeups_idle = 0;
2547 INIT_LIST_HEAD(&p->rt.run_list);
2549 INIT_LIST_HEAD(&p->se.group_node);
2551 #ifdef CONFIG_PREEMPT_NOTIFIERS
2552 INIT_HLIST_HEAD(&p->preempt_notifiers);
2557 * fork()/clone()-time setup:
2559 void sched_fork(struct task_struct *p, int clone_flags)
2561 int cpu = get_cpu();
2565 * We mark the process as waking here. This guarantees that
2566 * nobody will actually run it, and a signal or other external
2567 * event cannot wake it up and insert it on the runqueue either.
2569 p->state = TASK_WAKING;
2572 * Revert to default priority/policy on fork if requested.
2574 if (unlikely(p->sched_reset_on_fork)) {
2575 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2576 p->policy = SCHED_NORMAL;
2577 p->normal_prio = p->static_prio;
2580 if (PRIO_TO_NICE(p->static_prio) < 0) {
2581 p->static_prio = NICE_TO_PRIO(0);
2582 p->normal_prio = p->static_prio;
2587 * We don't need the reset flag anymore after the fork. It has
2588 * fulfilled its duty:
2590 p->sched_reset_on_fork = 0;
2594 * Make sure we do not leak PI boosting priority to the child.
2596 p->prio = current->normal_prio;
2598 if (!rt_prio(p->prio))
2599 p->sched_class = &fair_sched_class;
2601 if (p->sched_class->task_fork)
2602 p->sched_class->task_fork(p);
2605 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2607 set_task_cpu(p, cpu);
2609 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2610 if (likely(sched_info_on()))
2611 memset(&p->sched_info, 0, sizeof(p->sched_info));
2613 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2616 #ifdef CONFIG_PREEMPT
2617 /* Want to start with kernel preemption disabled. */
2618 task_thread_info(p)->preempt_count = 1;
2620 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2626 * wake_up_new_task - wake up a newly created task for the first time.
2628 * This function will do some initial scheduler statistics housekeeping
2629 * that must be done for every newly created context, then puts the task
2630 * on the runqueue and wakes it.
2632 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2634 unsigned long flags;
2637 rq = task_rq_lock(p, &flags);
2638 BUG_ON(p->state != TASK_WAKING);
2639 p->state = TASK_RUNNING;
2640 update_rq_clock(rq);
2641 activate_task(rq, p, 0);
2642 trace_sched_wakeup_new(rq, p, 1);
2643 check_preempt_curr(rq, p, WF_FORK);
2645 if (p->sched_class->task_woken)
2646 p->sched_class->task_woken(rq, p);
2648 task_rq_unlock(rq, &flags);
2651 #ifdef CONFIG_PREEMPT_NOTIFIERS
2654 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2655 * @notifier: notifier struct to register
2657 void preempt_notifier_register(struct preempt_notifier *notifier)
2659 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2661 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2664 * preempt_notifier_unregister - no longer interested in preemption notifications
2665 * @notifier: notifier struct to unregister
2667 * This is safe to call from within a preemption notifier.
2669 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2671 hlist_del(¬ifier->link);
2673 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2675 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2677 struct preempt_notifier *notifier;
2678 struct hlist_node *node;
2680 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2681 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2685 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2686 struct task_struct *next)
2688 struct preempt_notifier *notifier;
2689 struct hlist_node *node;
2691 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2692 notifier->ops->sched_out(notifier, next);
2695 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2697 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2702 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2703 struct task_struct *next)
2707 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2710 * prepare_task_switch - prepare to switch tasks
2711 * @rq: the runqueue preparing to switch
2712 * @prev: the current task that is being switched out
2713 * @next: the task we are going to switch to.
2715 * This is called with the rq lock held and interrupts off. It must
2716 * be paired with a subsequent finish_task_switch after the context
2719 * prepare_task_switch sets up locking and calls architecture specific
2723 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2724 struct task_struct *next)
2726 fire_sched_out_preempt_notifiers(prev, next);
2727 prepare_lock_switch(rq, next);
2728 prepare_arch_switch(next);
2732 * finish_task_switch - clean up after a task-switch
2733 * @rq: runqueue associated with task-switch
2734 * @prev: the thread we just switched away from.
2736 * finish_task_switch must be called after the context switch, paired
2737 * with a prepare_task_switch call before the context switch.
2738 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2739 * and do any other architecture-specific cleanup actions.
2741 * Note that we may have delayed dropping an mm in context_switch(). If
2742 * so, we finish that here outside of the runqueue lock. (Doing it
2743 * with the lock held can cause deadlocks; see schedule() for
2746 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2747 __releases(rq->lock)
2749 struct mm_struct *mm = rq->prev_mm;
2755 * A task struct has one reference for the use as "current".
2756 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2757 * schedule one last time. The schedule call will never return, and
2758 * the scheduled task must drop that reference.
2759 * The test for TASK_DEAD must occur while the runqueue locks are
2760 * still held, otherwise prev could be scheduled on another cpu, die
2761 * there before we look at prev->state, and then the reference would
2763 * Manfred Spraul <manfred@colorfullife.com>
2765 prev_state = prev->state;
2766 finish_arch_switch(prev);
2767 perf_event_task_sched_in(current, cpu_of(rq));
2768 finish_lock_switch(rq, prev);
2770 fire_sched_in_preempt_notifiers(current);
2773 if (unlikely(prev_state == TASK_DEAD)) {
2775 * Remove function-return probe instances associated with this
2776 * task and put them back on the free list.
2778 kprobe_flush_task(prev);
2779 put_task_struct(prev);
2785 /* assumes rq->lock is held */
2786 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2788 if (prev->sched_class->pre_schedule)
2789 prev->sched_class->pre_schedule(rq, prev);
2792 /* rq->lock is NOT held, but preemption is disabled */
2793 static inline void post_schedule(struct rq *rq)
2795 if (rq->post_schedule) {
2796 unsigned long flags;
2798 raw_spin_lock_irqsave(&rq->lock, flags);
2799 if (rq->curr->sched_class->post_schedule)
2800 rq->curr->sched_class->post_schedule(rq);
2801 raw_spin_unlock_irqrestore(&rq->lock, flags);
2803 rq->post_schedule = 0;
2809 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2813 static inline void post_schedule(struct rq *rq)
2820 * schedule_tail - first thing a freshly forked thread must call.
2821 * @prev: the thread we just switched away from.
2823 asmlinkage void schedule_tail(struct task_struct *prev)
2824 __releases(rq->lock)
2826 struct rq *rq = this_rq();
2828 finish_task_switch(rq, prev);
2831 * FIXME: do we need to worry about rq being invalidated by the
2836 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2837 /* In this case, finish_task_switch does not reenable preemption */
2840 if (current->set_child_tid)
2841 put_user(task_pid_vnr(current), current->set_child_tid);
2845 * context_switch - switch to the new MM and the new
2846 * thread's register state.
2849 context_switch(struct rq *rq, struct task_struct *prev,
2850 struct task_struct *next)
2852 struct mm_struct *mm, *oldmm;
2854 prepare_task_switch(rq, prev, next);
2855 trace_sched_switch(rq, prev, next);
2857 oldmm = prev->active_mm;
2859 * For paravirt, this is coupled with an exit in switch_to to
2860 * combine the page table reload and the switch backend into
2863 arch_start_context_switch(prev);
2866 next->active_mm = oldmm;
2867 atomic_inc(&oldmm->mm_count);
2868 enter_lazy_tlb(oldmm, next);
2870 switch_mm(oldmm, mm, next);
2872 if (likely(!prev->mm)) {
2873 prev->active_mm = NULL;
2874 rq->prev_mm = oldmm;
2877 * Since the runqueue lock will be released by the next
2878 * task (which is an invalid locking op but in the case
2879 * of the scheduler it's an obvious special-case), so we
2880 * do an early lockdep release here:
2882 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2883 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2886 /* Here we just switch the register state and the stack. */
2887 switch_to(prev, next, prev);
2891 * this_rq must be evaluated again because prev may have moved
2892 * CPUs since it called schedule(), thus the 'rq' on its stack
2893 * frame will be invalid.
2895 finish_task_switch(this_rq(), prev);
2899 * nr_running, nr_uninterruptible and nr_context_switches:
2901 * externally visible scheduler statistics: current number of runnable
2902 * threads, current number of uninterruptible-sleeping threads, total
2903 * number of context switches performed since bootup.
2905 unsigned long nr_running(void)
2907 unsigned long i, sum = 0;
2909 for_each_online_cpu(i)
2910 sum += cpu_rq(i)->nr_running;
2915 unsigned long nr_uninterruptible(void)
2917 unsigned long i, sum = 0;
2919 for_each_possible_cpu(i)
2920 sum += cpu_rq(i)->nr_uninterruptible;
2923 * Since we read the counters lockless, it might be slightly
2924 * inaccurate. Do not allow it to go below zero though:
2926 if (unlikely((long)sum < 0))
2932 unsigned long long nr_context_switches(void)
2935 unsigned long long sum = 0;
2937 for_each_possible_cpu(i)
2938 sum += cpu_rq(i)->nr_switches;
2943 unsigned long nr_iowait(void)
2945 unsigned long i, sum = 0;
2947 for_each_possible_cpu(i)
2948 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2953 unsigned long nr_iowait_cpu(void)
2955 struct rq *this = this_rq();
2956 return atomic_read(&this->nr_iowait);
2959 unsigned long this_cpu_load(void)
2961 struct rq *this = this_rq();
2962 return this->cpu_load[0];
2966 /* Variables and functions for calc_load */
2967 static atomic_long_t calc_load_tasks;
2968 static unsigned long calc_load_update;
2969 unsigned long avenrun[3];
2970 EXPORT_SYMBOL(avenrun);
2973 * get_avenrun - get the load average array
2974 * @loads: pointer to dest load array
2975 * @offset: offset to add
2976 * @shift: shift count to shift the result left
2978 * These values are estimates at best, so no need for locking.
2980 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2982 loads[0] = (avenrun[0] + offset) << shift;
2983 loads[1] = (avenrun[1] + offset) << shift;
2984 loads[2] = (avenrun[2] + offset) << shift;
2987 static unsigned long
2988 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2991 load += active * (FIXED_1 - exp);
2992 return load >> FSHIFT;
2996 * calc_load - update the avenrun load estimates 10 ticks after the
2997 * CPUs have updated calc_load_tasks.
2999 void calc_global_load(void)
3001 unsigned long upd = calc_load_update + 10;
3004 if (time_before(jiffies, upd))
3007 active = atomic_long_read(&calc_load_tasks);
3008 active = active > 0 ? active * FIXED_1 : 0;
3010 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3011 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3012 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3014 calc_load_update += LOAD_FREQ;
3018 * Either called from update_cpu_load() or from a cpu going idle
3020 static void calc_load_account_active(struct rq *this_rq)
3022 long nr_active, delta;
3024 nr_active = this_rq->nr_running;
3025 nr_active += (long) this_rq->nr_uninterruptible;
3027 if (nr_active != this_rq->calc_load_active) {
3028 delta = nr_active - this_rq->calc_load_active;
3029 this_rq->calc_load_active = nr_active;
3030 atomic_long_add(delta, &calc_load_tasks);
3035 * Update rq->cpu_load[] statistics. This function is usually called every
3036 * scheduler tick (TICK_NSEC).
3038 static void update_cpu_load(struct rq *this_rq)
3040 unsigned long this_load = this_rq->load.weight;
3043 this_rq->nr_load_updates++;
3045 /* Update our load: */
3046 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3047 unsigned long old_load, new_load;
3049 /* scale is effectively 1 << i now, and >> i divides by scale */
3051 old_load = this_rq->cpu_load[i];
3052 new_load = this_load;
3054 * Round up the averaging division if load is increasing. This
3055 * prevents us from getting stuck on 9 if the load is 10, for
3058 if (new_load > old_load)
3059 new_load += scale-1;
3060 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3063 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3064 this_rq->calc_load_update += LOAD_FREQ;
3065 calc_load_account_active(this_rq);
3072 * sched_exec - execve() is a valuable balancing opportunity, because at
3073 * this point the task has the smallest effective memory and cache footprint.
3075 void sched_exec(void)
3077 struct task_struct *p = current;
3078 struct migration_req req;
3079 int dest_cpu, this_cpu;
3080 unsigned long flags;
3084 this_cpu = get_cpu();
3085 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3086 if (dest_cpu == this_cpu) {
3091 rq = task_rq_lock(p, &flags);
3095 * select_task_rq() can race against ->cpus_allowed
3097 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3098 || unlikely(!cpu_active(dest_cpu))) {
3099 task_rq_unlock(rq, &flags);
3103 /* force the process onto the specified CPU */
3104 if (migrate_task(p, dest_cpu, &req)) {
3105 /* Need to wait for migration thread (might exit: take ref). */
3106 struct task_struct *mt = rq->migration_thread;
3108 get_task_struct(mt);
3109 task_rq_unlock(rq, &flags);
3110 wake_up_process(mt);
3111 put_task_struct(mt);
3112 wait_for_completion(&req.done);
3116 task_rq_unlock(rq, &flags);
3121 DEFINE_PER_CPU(struct kernel_stat, kstat);
3123 EXPORT_PER_CPU_SYMBOL(kstat);
3126 * Return any ns on the sched_clock that have not yet been accounted in
3127 * @p in case that task is currently running.
3129 * Called with task_rq_lock() held on @rq.
3131 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3135 if (task_current(rq, p)) {
3136 update_rq_clock(rq);
3137 ns = rq->clock - p->se.exec_start;
3145 unsigned long long task_delta_exec(struct task_struct *p)
3147 unsigned long flags;
3151 rq = task_rq_lock(p, &flags);
3152 ns = do_task_delta_exec(p, rq);
3153 task_rq_unlock(rq, &flags);
3159 * Return accounted runtime for the task.
3160 * In case the task is currently running, return the runtime plus current's
3161 * pending runtime that have not been accounted yet.
3163 unsigned long long task_sched_runtime(struct task_struct *p)
3165 unsigned long flags;
3169 rq = task_rq_lock(p, &flags);
3170 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3171 task_rq_unlock(rq, &flags);
3177 * Return sum_exec_runtime for the thread group.
3178 * In case the task is currently running, return the sum plus current's
3179 * pending runtime that have not been accounted yet.
3181 * Note that the thread group might have other running tasks as well,
3182 * so the return value not includes other pending runtime that other
3183 * running tasks might have.
3185 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3187 struct task_cputime totals;
3188 unsigned long flags;
3192 rq = task_rq_lock(p, &flags);
3193 thread_group_cputime(p, &totals);
3194 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3195 task_rq_unlock(rq, &flags);
3201 * Account user cpu time to a process.
3202 * @p: the process that the cpu time gets accounted to
3203 * @cputime: the cpu time spent in user space since the last update
3204 * @cputime_scaled: cputime scaled by cpu frequency
3206 void account_user_time(struct task_struct *p, cputime_t cputime,
3207 cputime_t cputime_scaled)
3209 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3212 /* Add user time to process. */
3213 p->utime = cputime_add(p->utime, cputime);
3214 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3215 account_group_user_time(p, cputime);
3217 /* Add user time to cpustat. */
3218 tmp = cputime_to_cputime64(cputime);
3219 if (TASK_NICE(p) > 0)
3220 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3222 cpustat->user = cputime64_add(cpustat->user, tmp);
3224 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3225 /* Account for user time used */
3226 acct_update_integrals(p);
3230 * Account guest cpu time to a process.
3231 * @p: the process that the cpu time gets accounted to
3232 * @cputime: the cpu time spent in virtual machine since the last update
3233 * @cputime_scaled: cputime scaled by cpu frequency
3235 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3236 cputime_t cputime_scaled)
3239 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3241 tmp = cputime_to_cputime64(cputime);
3243 /* Add guest time to process. */
3244 p->utime = cputime_add(p->utime, cputime);
3245 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3246 account_group_user_time(p, cputime);
3247 p->gtime = cputime_add(p->gtime, cputime);
3249 /* Add guest time to cpustat. */
3250 if (TASK_NICE(p) > 0) {
3251 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3252 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3254 cpustat->user = cputime64_add(cpustat->user, tmp);
3255 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3260 * Account system cpu time to a process.
3261 * @p: the process that the cpu time gets accounted to
3262 * @hardirq_offset: the offset to subtract from hardirq_count()
3263 * @cputime: the cpu time spent in kernel space since the last update
3264 * @cputime_scaled: cputime scaled by cpu frequency
3266 void account_system_time(struct task_struct *p, int hardirq_offset,
3267 cputime_t cputime, cputime_t cputime_scaled)
3269 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3272 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3273 account_guest_time(p, cputime, cputime_scaled);
3277 /* Add system time to process. */
3278 p->stime = cputime_add(p->stime, cputime);
3279 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3280 account_group_system_time(p, cputime);
3282 /* Add system time to cpustat. */
3283 tmp = cputime_to_cputime64(cputime);
3284 if (hardirq_count() - hardirq_offset)
3285 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3286 else if (softirq_count())
3287 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3289 cpustat->system = cputime64_add(cpustat->system, tmp);
3291 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3293 /* Account for system time used */
3294 acct_update_integrals(p);
3298 * Account for involuntary wait time.
3299 * @steal: the cpu time spent in involuntary wait
3301 void account_steal_time(cputime_t cputime)
3303 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3304 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3306 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3310 * Account for idle time.
3311 * @cputime: the cpu time spent in idle wait
3313 void account_idle_time(cputime_t cputime)
3315 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3316 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3317 struct rq *rq = this_rq();
3319 if (atomic_read(&rq->nr_iowait) > 0)
3320 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3322 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3325 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3328 * Account a single tick of cpu time.
3329 * @p: the process that the cpu time gets accounted to
3330 * @user_tick: indicates if the tick is a user or a system tick
3332 void account_process_tick(struct task_struct *p, int user_tick)
3334 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3335 struct rq *rq = this_rq();
3338 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3339 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3340 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3343 account_idle_time(cputime_one_jiffy);
3347 * Account multiple ticks of steal time.
3348 * @p: the process from which the cpu time has been stolen
3349 * @ticks: number of stolen ticks
3351 void account_steal_ticks(unsigned long ticks)
3353 account_steal_time(jiffies_to_cputime(ticks));
3357 * Account multiple ticks of idle time.
3358 * @ticks: number of stolen ticks
3360 void account_idle_ticks(unsigned long ticks)
3362 account_idle_time(jiffies_to_cputime(ticks));
3368 * Use precise platform statistics if available:
3370 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3371 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3377 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3379 struct task_cputime cputime;
3381 thread_group_cputime(p, &cputime);
3383 *ut = cputime.utime;
3384 *st = cputime.stime;
3388 #ifndef nsecs_to_cputime
3389 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3392 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3394 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3397 * Use CFS's precise accounting:
3399 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3404 temp = (u64)(rtime * utime);
3405 do_div(temp, total);
3406 utime = (cputime_t)temp;
3411 * Compare with previous values, to keep monotonicity:
3413 p->prev_utime = max(p->prev_utime, utime);
3414 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3416 *ut = p->prev_utime;
3417 *st = p->prev_stime;
3421 * Must be called with siglock held.
3423 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3425 struct signal_struct *sig = p->signal;
3426 struct task_cputime cputime;
3427 cputime_t rtime, utime, total;
3429 thread_group_cputime(p, &cputime);
3431 total = cputime_add(cputime.utime, cputime.stime);
3432 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3437 temp = (u64)(rtime * cputime.utime);
3438 do_div(temp, total);
3439 utime = (cputime_t)temp;
3443 sig->prev_utime = max(sig->prev_utime, utime);
3444 sig->prev_stime = max(sig->prev_stime,
3445 cputime_sub(rtime, sig->prev_utime));
3447 *ut = sig->prev_utime;
3448 *st = sig->prev_stime;
3453 * This function gets called by the timer code, with HZ frequency.
3454 * We call it with interrupts disabled.
3456 * It also gets called by the fork code, when changing the parent's
3459 void scheduler_tick(void)
3461 int cpu = smp_processor_id();
3462 struct rq *rq = cpu_rq(cpu);
3463 struct task_struct *curr = rq->curr;
3467 raw_spin_lock(&rq->lock);
3468 update_rq_clock(rq);
3469 update_cpu_load(rq);
3470 curr->sched_class->task_tick(rq, curr, 0);
3471 raw_spin_unlock(&rq->lock);
3473 perf_event_task_tick(curr, cpu);
3476 rq->idle_at_tick = idle_cpu(cpu);
3477 trigger_load_balance(rq, cpu);
3481 notrace unsigned long get_parent_ip(unsigned long addr)
3483 if (in_lock_functions(addr)) {
3484 addr = CALLER_ADDR2;
3485 if (in_lock_functions(addr))
3486 addr = CALLER_ADDR3;
3491 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3492 defined(CONFIG_PREEMPT_TRACER))
3494 void __kprobes add_preempt_count(int val)
3496 #ifdef CONFIG_DEBUG_PREEMPT
3500 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3503 preempt_count() += val;
3504 #ifdef CONFIG_DEBUG_PREEMPT
3506 * Spinlock count overflowing soon?
3508 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3511 if (preempt_count() == val)
3512 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3514 EXPORT_SYMBOL(add_preempt_count);
3516 void __kprobes sub_preempt_count(int val)
3518 #ifdef CONFIG_DEBUG_PREEMPT
3522 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3525 * Is the spinlock portion underflowing?
3527 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3528 !(preempt_count() & PREEMPT_MASK)))
3532 if (preempt_count() == val)
3533 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3534 preempt_count() -= val;
3536 EXPORT_SYMBOL(sub_preempt_count);
3541 * Print scheduling while atomic bug:
3543 static noinline void __schedule_bug(struct task_struct *prev)
3545 struct pt_regs *regs = get_irq_regs();
3547 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3548 prev->comm, prev->pid, preempt_count());
3550 debug_show_held_locks(prev);
3552 if (irqs_disabled())
3553 print_irqtrace_events(prev);
3562 * Various schedule()-time debugging checks and statistics:
3564 static inline void schedule_debug(struct task_struct *prev)
3567 * Test if we are atomic. Since do_exit() needs to call into
3568 * schedule() atomically, we ignore that path for now.
3569 * Otherwise, whine if we are scheduling when we should not be.
3571 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3572 __schedule_bug(prev);
3574 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3576 schedstat_inc(this_rq(), sched_count);
3577 #ifdef CONFIG_SCHEDSTATS
3578 if (unlikely(prev->lock_depth >= 0)) {
3579 schedstat_inc(this_rq(), bkl_count);
3580 schedstat_inc(prev, sched_info.bkl_count);
3585 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3587 if (prev->state == TASK_RUNNING) {
3588 u64 runtime = prev->se.sum_exec_runtime;
3590 runtime -= prev->se.prev_sum_exec_runtime;
3591 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
3594 * In order to avoid avg_overlap growing stale when we are
3595 * indeed overlapping and hence not getting put to sleep, grow
3596 * the avg_overlap on preemption.
3598 * We use the average preemption runtime because that
3599 * correlates to the amount of cache footprint a task can
3602 update_avg(&prev->se.avg_overlap, runtime);
3604 prev->sched_class->put_prev_task(rq, prev);
3608 * Pick up the highest-prio task:
3610 static inline struct task_struct *
3611 pick_next_task(struct rq *rq)
3613 const struct sched_class *class;
3614 struct task_struct *p;
3617 * Optimization: we know that if all tasks are in
3618 * the fair class we can call that function directly:
3620 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3621 p = fair_sched_class.pick_next_task(rq);
3626 class = sched_class_highest;
3628 p = class->pick_next_task(rq);
3632 * Will never be NULL as the idle class always
3633 * returns a non-NULL p:
3635 class = class->next;
3640 * schedule() is the main scheduler function.
3642 asmlinkage void __sched schedule(void)
3644 struct task_struct *prev, *next;
3645 unsigned long *switch_count;
3651 cpu = smp_processor_id();
3655 switch_count = &prev->nivcsw;
3657 release_kernel_lock(prev);
3658 need_resched_nonpreemptible:
3660 schedule_debug(prev);
3662 if (sched_feat(HRTICK))
3665 raw_spin_lock_irq(&rq->lock);
3666 update_rq_clock(rq);
3667 clear_tsk_need_resched(prev);
3669 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3670 if (unlikely(signal_pending_state(prev->state, prev)))
3671 prev->state = TASK_RUNNING;
3673 deactivate_task(rq, prev, 1);
3674 switch_count = &prev->nvcsw;
3677 pre_schedule(rq, prev);
3679 if (unlikely(!rq->nr_running))
3680 idle_balance(cpu, rq);
3682 put_prev_task(rq, prev);
3683 next = pick_next_task(rq);
3685 if (likely(prev != next)) {
3686 sched_info_switch(prev, next);
3687 perf_event_task_sched_out(prev, next, cpu);
3693 context_switch(rq, prev, next); /* unlocks the rq */
3695 * the context switch might have flipped the stack from under
3696 * us, hence refresh the local variables.
3698 cpu = smp_processor_id();
3701 raw_spin_unlock_irq(&rq->lock);
3705 if (unlikely(reacquire_kernel_lock(current) < 0))
3706 goto need_resched_nonpreemptible;
3708 preempt_enable_no_resched();
3712 EXPORT_SYMBOL(schedule);
3714 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3716 * Look out! "owner" is an entirely speculative pointer
3717 * access and not reliable.
3719 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3724 if (!sched_feat(OWNER_SPIN))
3727 #ifdef CONFIG_DEBUG_PAGEALLOC
3729 * Need to access the cpu field knowing that
3730 * DEBUG_PAGEALLOC could have unmapped it if
3731 * the mutex owner just released it and exited.
3733 if (probe_kernel_address(&owner->cpu, cpu))
3740 * Even if the access succeeded (likely case),
3741 * the cpu field may no longer be valid.
3743 if (cpu >= nr_cpumask_bits)
3747 * We need to validate that we can do a
3748 * get_cpu() and that we have the percpu area.
3750 if (!cpu_online(cpu))
3757 * Owner changed, break to re-assess state.
3759 if (lock->owner != owner)
3763 * Is that owner really running on that cpu?
3765 if (task_thread_info(rq->curr) != owner || need_resched())
3775 #ifdef CONFIG_PREEMPT
3777 * this is the entry point to schedule() from in-kernel preemption
3778 * off of preempt_enable. Kernel preemptions off return from interrupt
3779 * occur there and call schedule directly.
3781 asmlinkage void __sched preempt_schedule(void)
3783 struct thread_info *ti = current_thread_info();
3786 * If there is a non-zero preempt_count or interrupts are disabled,
3787 * we do not want to preempt the current task. Just return..
3789 if (likely(ti->preempt_count || irqs_disabled()))
3793 add_preempt_count(PREEMPT_ACTIVE);
3795 sub_preempt_count(PREEMPT_ACTIVE);
3798 * Check again in case we missed a preemption opportunity
3799 * between schedule and now.
3802 } while (need_resched());
3804 EXPORT_SYMBOL(preempt_schedule);
3807 * this is the entry point to schedule() from kernel preemption
3808 * off of irq context.
3809 * Note, that this is called and return with irqs disabled. This will
3810 * protect us against recursive calling from irq.
3812 asmlinkage void __sched preempt_schedule_irq(void)
3814 struct thread_info *ti = current_thread_info();
3816 /* Catch callers which need to be fixed */
3817 BUG_ON(ti->preempt_count || !irqs_disabled());
3820 add_preempt_count(PREEMPT_ACTIVE);
3823 local_irq_disable();
3824 sub_preempt_count(PREEMPT_ACTIVE);
3827 * Check again in case we missed a preemption opportunity
3828 * between schedule and now.
3831 } while (need_resched());
3834 #endif /* CONFIG_PREEMPT */
3836 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3839 return try_to_wake_up(curr->private, mode, wake_flags);
3841 EXPORT_SYMBOL(default_wake_function);
3844 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3845 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3846 * number) then we wake all the non-exclusive tasks and one exclusive task.
3848 * There are circumstances in which we can try to wake a task which has already
3849 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3850 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3852 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3853 int nr_exclusive, int wake_flags, void *key)
3855 wait_queue_t *curr, *next;
3857 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3858 unsigned flags = curr->flags;
3860 if (curr->func(curr, mode, wake_flags, key) &&
3861 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3867 * __wake_up - wake up threads blocked on a waitqueue.
3869 * @mode: which threads
3870 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3871 * @key: is directly passed to the wakeup function
3873 * It may be assumed that this function implies a write memory barrier before
3874 * changing the task state if and only if any tasks are woken up.
3876 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3877 int nr_exclusive, void *key)
3879 unsigned long flags;
3881 spin_lock_irqsave(&q->lock, flags);
3882 __wake_up_common(q, mode, nr_exclusive, 0, key);
3883 spin_unlock_irqrestore(&q->lock, flags);
3885 EXPORT_SYMBOL(__wake_up);
3888 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3890 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3892 __wake_up_common(q, mode, 1, 0, NULL);
3895 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3897 __wake_up_common(q, mode, 1, 0, key);
3901 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3903 * @mode: which threads
3904 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3905 * @key: opaque value to be passed to wakeup targets
3907 * The sync wakeup differs that the waker knows that it will schedule
3908 * away soon, so while the target thread will be woken up, it will not
3909 * be migrated to another CPU - ie. the two threads are 'synchronized'
3910 * with each other. This can prevent needless bouncing between CPUs.
3912 * On UP it can prevent extra preemption.
3914 * It may be assumed that this function implies a write memory barrier before
3915 * changing the task state if and only if any tasks are woken up.
3917 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3918 int nr_exclusive, void *key)
3920 unsigned long flags;
3921 int wake_flags = WF_SYNC;
3926 if (unlikely(!nr_exclusive))
3929 spin_lock_irqsave(&q->lock, flags);
3930 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3931 spin_unlock_irqrestore(&q->lock, flags);
3933 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3936 * __wake_up_sync - see __wake_up_sync_key()
3938 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3940 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3942 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3945 * complete: - signals a single thread waiting on this completion
3946 * @x: holds the state of this particular completion
3948 * This will wake up a single thread waiting on this completion. Threads will be
3949 * awakened in the same order in which they were queued.
3951 * See also complete_all(), wait_for_completion() and related routines.
3953 * It may be assumed that this function implies a write memory barrier before
3954 * changing the task state if and only if any tasks are woken up.
3956 void complete(struct completion *x)
3958 unsigned long flags;
3960 spin_lock_irqsave(&x->wait.lock, flags);
3962 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3963 spin_unlock_irqrestore(&x->wait.lock, flags);
3965 EXPORT_SYMBOL(complete);
3968 * complete_all: - signals all threads waiting on this completion
3969 * @x: holds the state of this particular completion
3971 * This will wake up all threads waiting on this particular completion event.
3973 * It may be assumed that this function implies a write memory barrier before
3974 * changing the task state if and only if any tasks are woken up.
3976 void complete_all(struct completion *x)
3978 unsigned long flags;
3980 spin_lock_irqsave(&x->wait.lock, flags);
3981 x->done += UINT_MAX/2;
3982 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3983 spin_unlock_irqrestore(&x->wait.lock, flags);
3985 EXPORT_SYMBOL(complete_all);
3987 static inline long __sched
3988 do_wait_for_common(struct completion *x, long timeout, int state)
3991 DECLARE_WAITQUEUE(wait, current);
3993 wait.flags |= WQ_FLAG_EXCLUSIVE;
3994 __add_wait_queue_tail(&x->wait, &wait);
3996 if (signal_pending_state(state, current)) {
3997 timeout = -ERESTARTSYS;
4000 __set_current_state(state);
4001 spin_unlock_irq(&x->wait.lock);
4002 timeout = schedule_timeout(timeout);
4003 spin_lock_irq(&x->wait.lock);
4004 } while (!x->done && timeout);
4005 __remove_wait_queue(&x->wait, &wait);
4010 return timeout ?: 1;
4014 wait_for_common(struct completion *x, long timeout, int state)
4018 spin_lock_irq(&x->wait.lock);
4019 timeout = do_wait_for_common(x, timeout, state);
4020 spin_unlock_irq(&x->wait.lock);
4025 * wait_for_completion: - waits for completion of a task
4026 * @x: holds the state of this particular completion
4028 * This waits to be signaled for completion of a specific task. It is NOT
4029 * interruptible and there is no timeout.
4031 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4032 * and interrupt capability. Also see complete().
4034 void __sched wait_for_completion(struct completion *x)
4036 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4038 EXPORT_SYMBOL(wait_for_completion);
4041 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4042 * @x: holds the state of this particular completion
4043 * @timeout: timeout value in jiffies
4045 * This waits for either a completion of a specific task to be signaled or for a
4046 * specified timeout to expire. The timeout is in jiffies. It is not
4049 unsigned long __sched
4050 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4052 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4054 EXPORT_SYMBOL(wait_for_completion_timeout);
4057 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4058 * @x: holds the state of this particular completion
4060 * This waits for completion of a specific task to be signaled. It is
4063 int __sched wait_for_completion_interruptible(struct completion *x)
4065 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4066 if (t == -ERESTARTSYS)
4070 EXPORT_SYMBOL(wait_for_completion_interruptible);
4073 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4074 * @x: holds the state of this particular completion
4075 * @timeout: timeout value in jiffies
4077 * This waits for either a completion of a specific task to be signaled or for a
4078 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4080 unsigned long __sched
4081 wait_for_completion_interruptible_timeout(struct completion *x,
4082 unsigned long timeout)
4084 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4086 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4089 * wait_for_completion_killable: - waits for completion of a task (killable)
4090 * @x: holds the state of this particular completion
4092 * This waits to be signaled for completion of a specific task. It can be
4093 * interrupted by a kill signal.
4095 int __sched wait_for_completion_killable(struct completion *x)
4097 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4098 if (t == -ERESTARTSYS)
4102 EXPORT_SYMBOL(wait_for_completion_killable);
4105 * try_wait_for_completion - try to decrement a completion without blocking
4106 * @x: completion structure
4108 * Returns: 0 if a decrement cannot be done without blocking
4109 * 1 if a decrement succeeded.
4111 * If a completion is being used as a counting completion,
4112 * attempt to decrement the counter without blocking. This
4113 * enables us to avoid waiting if the resource the completion
4114 * is protecting is not available.
4116 bool try_wait_for_completion(struct completion *x)
4118 unsigned long flags;
4121 spin_lock_irqsave(&x->wait.lock, flags);
4126 spin_unlock_irqrestore(&x->wait.lock, flags);
4129 EXPORT_SYMBOL(try_wait_for_completion);
4132 * completion_done - Test to see if a completion has any waiters
4133 * @x: completion structure
4135 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4136 * 1 if there are no waiters.
4139 bool completion_done(struct completion *x)
4141 unsigned long flags;
4144 spin_lock_irqsave(&x->wait.lock, flags);
4147 spin_unlock_irqrestore(&x->wait.lock, flags);
4150 EXPORT_SYMBOL(completion_done);
4153 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4155 unsigned long flags;
4158 init_waitqueue_entry(&wait, current);
4160 __set_current_state(state);
4162 spin_lock_irqsave(&q->lock, flags);
4163 __add_wait_queue(q, &wait);
4164 spin_unlock(&q->lock);
4165 timeout = schedule_timeout(timeout);
4166 spin_lock_irq(&q->lock);
4167 __remove_wait_queue(q, &wait);
4168 spin_unlock_irqrestore(&q->lock, flags);
4173 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4175 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4177 EXPORT_SYMBOL(interruptible_sleep_on);
4180 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4182 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4184 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4186 void __sched sleep_on(wait_queue_head_t *q)
4188 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4190 EXPORT_SYMBOL(sleep_on);
4192 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4194 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4196 EXPORT_SYMBOL(sleep_on_timeout);
4198 #ifdef CONFIG_RT_MUTEXES
4201 * rt_mutex_setprio - set the current priority of a task
4203 * @prio: prio value (kernel-internal form)
4205 * This function changes the 'effective' priority of a task. It does
4206 * not touch ->normal_prio like __setscheduler().
4208 * Used by the rt_mutex code to implement priority inheritance logic.
4210 void rt_mutex_setprio(struct task_struct *p, int prio)
4212 unsigned long flags;
4213 int oldprio, on_rq, running;
4215 const struct sched_class *prev_class = p->sched_class;
4217 BUG_ON(prio < 0 || prio > MAX_PRIO);
4219 rq = task_rq_lock(p, &flags);
4220 update_rq_clock(rq);
4223 on_rq = p->se.on_rq;
4224 running = task_current(rq, p);
4226 dequeue_task(rq, p, 0);
4228 p->sched_class->put_prev_task(rq, p);
4231 p->sched_class = &rt_sched_class;
4233 p->sched_class = &fair_sched_class;
4238 p->sched_class->set_curr_task(rq);
4240 enqueue_task(rq, p, 0, false);
4242 check_class_changed(rq, p, prev_class, oldprio, running);
4244 task_rq_unlock(rq, &flags);
4249 void set_user_nice(struct task_struct *p, long nice)
4251 int old_prio, delta, on_rq;
4252 unsigned long flags;
4255 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4258 * We have to be careful, if called from sys_setpriority(),
4259 * the task might be in the middle of scheduling on another CPU.
4261 rq = task_rq_lock(p, &flags);
4262 update_rq_clock(rq);
4264 * The RT priorities are set via sched_setscheduler(), but we still
4265 * allow the 'normal' nice value to be set - but as expected
4266 * it wont have any effect on scheduling until the task is
4267 * SCHED_FIFO/SCHED_RR:
4269 if (task_has_rt_policy(p)) {
4270 p->static_prio = NICE_TO_PRIO(nice);
4273 on_rq = p->se.on_rq;
4275 dequeue_task(rq, p, 0);
4277 p->static_prio = NICE_TO_PRIO(nice);
4280 p->prio = effective_prio(p);
4281 delta = p->prio - old_prio;
4284 enqueue_task(rq, p, 0, false);
4286 * If the task increased its priority or is running and
4287 * lowered its priority, then reschedule its CPU:
4289 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4290 resched_task(rq->curr);
4293 task_rq_unlock(rq, &flags);
4295 EXPORT_SYMBOL(set_user_nice);
4298 * can_nice - check if a task can reduce its nice value
4302 int can_nice(const struct task_struct *p, const int nice)
4304 /* convert nice value [19,-20] to rlimit style value [1,40] */
4305 int nice_rlim = 20 - nice;
4307 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4308 capable(CAP_SYS_NICE));
4311 #ifdef __ARCH_WANT_SYS_NICE
4314 * sys_nice - change the priority of the current process.
4315 * @increment: priority increment
4317 * sys_setpriority is a more generic, but much slower function that
4318 * does similar things.
4320 SYSCALL_DEFINE1(nice, int, increment)
4325 * Setpriority might change our priority at the same moment.
4326 * We don't have to worry. Conceptually one call occurs first
4327 * and we have a single winner.
4329 if (increment < -40)
4334 nice = TASK_NICE(current) + increment;
4340 if (increment < 0 && !can_nice(current, nice))
4343 retval = security_task_setnice(current, nice);
4347 set_user_nice(current, nice);
4354 * task_prio - return the priority value of a given task.
4355 * @p: the task in question.
4357 * This is the priority value as seen by users in /proc.
4358 * RT tasks are offset by -200. Normal tasks are centered
4359 * around 0, value goes from -16 to +15.
4361 int task_prio(const struct task_struct *p)
4363 return p->prio - MAX_RT_PRIO;
4367 * task_nice - return the nice value of a given task.
4368 * @p: the task in question.
4370 int task_nice(const struct task_struct *p)
4372 return TASK_NICE(p);
4374 EXPORT_SYMBOL(task_nice);
4377 * idle_cpu - is a given cpu idle currently?
4378 * @cpu: the processor in question.
4380 int idle_cpu(int cpu)
4382 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4386 * idle_task - return the idle task for a given cpu.
4387 * @cpu: the processor in question.
4389 struct task_struct *idle_task(int cpu)
4391 return cpu_rq(cpu)->idle;
4395 * find_process_by_pid - find a process with a matching PID value.
4396 * @pid: the pid in question.
4398 static struct task_struct *find_process_by_pid(pid_t pid)
4400 return pid ? find_task_by_vpid(pid) : current;
4403 /* Actually do priority change: must hold rq lock. */
4405 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4407 BUG_ON(p->se.on_rq);
4410 p->rt_priority = prio;
4411 p->normal_prio = normal_prio(p);
4412 /* we are holding p->pi_lock already */
4413 p->prio = rt_mutex_getprio(p);
4414 if (rt_prio(p->prio))
4415 p->sched_class = &rt_sched_class;
4417 p->sched_class = &fair_sched_class;
4422 * check the target process has a UID that matches the current process's
4424 static bool check_same_owner(struct task_struct *p)
4426 const struct cred *cred = current_cred(), *pcred;
4430 pcred = __task_cred(p);
4431 match = (cred->euid == pcred->euid ||
4432 cred->euid == pcred->uid);
4437 static int __sched_setscheduler(struct task_struct *p, int policy,
4438 struct sched_param *param, bool user)
4440 int retval, oldprio, oldpolicy = -1, on_rq, running;
4441 unsigned long flags;
4442 const struct sched_class *prev_class = p->sched_class;
4446 /* may grab non-irq protected spin_locks */
4447 BUG_ON(in_interrupt());
4449 /* double check policy once rq lock held */
4451 reset_on_fork = p->sched_reset_on_fork;
4452 policy = oldpolicy = p->policy;
4454 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4455 policy &= ~SCHED_RESET_ON_FORK;
4457 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4458 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4459 policy != SCHED_IDLE)
4464 * Valid priorities for SCHED_FIFO and SCHED_RR are
4465 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4466 * SCHED_BATCH and SCHED_IDLE is 0.
4468 if (param->sched_priority < 0 ||
4469 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4470 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4472 if (rt_policy(policy) != (param->sched_priority != 0))
4476 * Allow unprivileged RT tasks to decrease priority:
4478 if (user && !capable(CAP_SYS_NICE)) {
4479 if (rt_policy(policy)) {
4480 unsigned long rlim_rtprio;
4482 if (!lock_task_sighand(p, &flags))
4484 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4485 unlock_task_sighand(p, &flags);
4487 /* can't set/change the rt policy */
4488 if (policy != p->policy && !rlim_rtprio)
4491 /* can't increase priority */
4492 if (param->sched_priority > p->rt_priority &&
4493 param->sched_priority > rlim_rtprio)
4497 * Like positive nice levels, dont allow tasks to
4498 * move out of SCHED_IDLE either:
4500 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4503 /* can't change other user's priorities */
4504 if (!check_same_owner(p))
4507 /* Normal users shall not reset the sched_reset_on_fork flag */
4508 if (p->sched_reset_on_fork && !reset_on_fork)
4513 #ifdef CONFIG_RT_GROUP_SCHED
4515 * Do not allow realtime tasks into groups that have no runtime
4518 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4519 task_group(p)->rt_bandwidth.rt_runtime == 0)
4523 retval = security_task_setscheduler(p, policy, param);
4529 * make sure no PI-waiters arrive (or leave) while we are
4530 * changing the priority of the task:
4532 raw_spin_lock_irqsave(&p->pi_lock, flags);
4534 * To be able to change p->policy safely, the apropriate
4535 * runqueue lock must be held.
4537 rq = __task_rq_lock(p);
4538 /* recheck policy now with rq lock held */
4539 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4540 policy = oldpolicy = -1;
4541 __task_rq_unlock(rq);
4542 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4545 update_rq_clock(rq);
4546 on_rq = p->se.on_rq;
4547 running = task_current(rq, p);
4549 deactivate_task(rq, p, 0);
4551 p->sched_class->put_prev_task(rq, p);
4553 p->sched_reset_on_fork = reset_on_fork;
4556 __setscheduler(rq, p, policy, param->sched_priority);
4559 p->sched_class->set_curr_task(rq);
4561 activate_task(rq, p, 0);
4563 check_class_changed(rq, p, prev_class, oldprio, running);
4565 __task_rq_unlock(rq);
4566 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4568 rt_mutex_adjust_pi(p);
4574 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4575 * @p: the task in question.
4576 * @policy: new policy.
4577 * @param: structure containing the new RT priority.
4579 * NOTE that the task may be already dead.
4581 int sched_setscheduler(struct task_struct *p, int policy,
4582 struct sched_param *param)
4584 return __sched_setscheduler(p, policy, param, true);
4586 EXPORT_SYMBOL_GPL(sched_setscheduler);
4589 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4590 * @p: the task in question.
4591 * @policy: new policy.
4592 * @param: structure containing the new RT priority.
4594 * Just like sched_setscheduler, only don't bother checking if the
4595 * current context has permission. For example, this is needed in
4596 * stop_machine(): we create temporary high priority worker threads,
4597 * but our caller might not have that capability.
4599 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4600 struct sched_param *param)
4602 return __sched_setscheduler(p, policy, param, false);
4606 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4608 struct sched_param lparam;
4609 struct task_struct *p;
4612 if (!param || pid < 0)
4614 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4619 p = find_process_by_pid(pid);
4621 retval = sched_setscheduler(p, policy, &lparam);
4628 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4629 * @pid: the pid in question.
4630 * @policy: new policy.
4631 * @param: structure containing the new RT priority.
4633 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4634 struct sched_param __user *, param)
4636 /* negative values for policy are not valid */
4640 return do_sched_setscheduler(pid, policy, param);
4644 * sys_sched_setparam - set/change the RT priority of a thread
4645 * @pid: the pid in question.
4646 * @param: structure containing the new RT priority.
4648 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4650 return do_sched_setscheduler(pid, -1, param);
4654 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4655 * @pid: the pid in question.
4657 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4659 struct task_struct *p;
4667 p = find_process_by_pid(pid);
4669 retval = security_task_getscheduler(p);
4672 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4679 * sys_sched_getparam - get the RT priority of a thread
4680 * @pid: the pid in question.
4681 * @param: structure containing the RT priority.
4683 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4685 struct sched_param lp;
4686 struct task_struct *p;
4689 if (!param || pid < 0)
4693 p = find_process_by_pid(pid);
4698 retval = security_task_getscheduler(p);
4702 lp.sched_priority = p->rt_priority;
4706 * This one might sleep, we cannot do it with a spinlock held ...
4708 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4717 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4719 cpumask_var_t cpus_allowed, new_mask;
4720 struct task_struct *p;
4726 p = find_process_by_pid(pid);
4733 /* Prevent p going away */
4737 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4741 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4743 goto out_free_cpus_allowed;
4746 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4749 retval = security_task_setscheduler(p, 0, NULL);
4753 cpuset_cpus_allowed(p, cpus_allowed);
4754 cpumask_and(new_mask, in_mask, cpus_allowed);
4756 retval = set_cpus_allowed_ptr(p, new_mask);
4759 cpuset_cpus_allowed(p, cpus_allowed);
4760 if (!cpumask_subset(new_mask, cpus_allowed)) {
4762 * We must have raced with a concurrent cpuset
4763 * update. Just reset the cpus_allowed to the
4764 * cpuset's cpus_allowed
4766 cpumask_copy(new_mask, cpus_allowed);
4771 free_cpumask_var(new_mask);
4772 out_free_cpus_allowed:
4773 free_cpumask_var(cpus_allowed);
4780 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4781 struct cpumask *new_mask)
4783 if (len < cpumask_size())
4784 cpumask_clear(new_mask);
4785 else if (len > cpumask_size())
4786 len = cpumask_size();
4788 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4792 * sys_sched_setaffinity - set the cpu affinity of a process
4793 * @pid: pid of the process
4794 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4795 * @user_mask_ptr: user-space pointer to the new cpu mask
4797 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4798 unsigned long __user *, user_mask_ptr)
4800 cpumask_var_t new_mask;
4803 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4806 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4808 retval = sched_setaffinity(pid, new_mask);
4809 free_cpumask_var(new_mask);
4813 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4815 struct task_struct *p;
4816 unsigned long flags;
4824 p = find_process_by_pid(pid);
4828 retval = security_task_getscheduler(p);
4832 rq = task_rq_lock(p, &flags);
4833 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4834 task_rq_unlock(rq, &flags);
4844 * sys_sched_getaffinity - get the cpu affinity of a process
4845 * @pid: pid of the process
4846 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4847 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4849 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4850 unsigned long __user *, user_mask_ptr)
4855 if (len < cpumask_size())
4858 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4861 ret = sched_getaffinity(pid, mask);
4863 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
4866 ret = cpumask_size();
4868 free_cpumask_var(mask);
4874 * sys_sched_yield - yield the current processor to other threads.
4876 * This function yields the current CPU to other tasks. If there are no
4877 * other threads running on this CPU then this function will return.
4879 SYSCALL_DEFINE0(sched_yield)
4881 struct rq *rq = this_rq_lock();
4883 schedstat_inc(rq, yld_count);
4884 current->sched_class->yield_task(rq);
4887 * Since we are going to call schedule() anyway, there's
4888 * no need to preempt or enable interrupts:
4890 __release(rq->lock);
4891 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4892 do_raw_spin_unlock(&rq->lock);
4893 preempt_enable_no_resched();
4900 static inline int should_resched(void)
4902 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4905 static void __cond_resched(void)
4907 add_preempt_count(PREEMPT_ACTIVE);
4909 sub_preempt_count(PREEMPT_ACTIVE);
4912 int __sched _cond_resched(void)
4914 if (should_resched()) {
4920 EXPORT_SYMBOL(_cond_resched);
4923 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4924 * call schedule, and on return reacquire the lock.
4926 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4927 * operations here to prevent schedule() from being called twice (once via
4928 * spin_unlock(), once by hand).
4930 int __cond_resched_lock(spinlock_t *lock)
4932 int resched = should_resched();
4935 lockdep_assert_held(lock);
4937 if (spin_needbreak(lock) || resched) {
4948 EXPORT_SYMBOL(__cond_resched_lock);
4950 int __sched __cond_resched_softirq(void)
4952 BUG_ON(!in_softirq());
4954 if (should_resched()) {
4962 EXPORT_SYMBOL(__cond_resched_softirq);
4965 * yield - yield the current processor to other threads.
4967 * This is a shortcut for kernel-space yielding - it marks the
4968 * thread runnable and calls sys_sched_yield().
4970 void __sched yield(void)
4972 set_current_state(TASK_RUNNING);
4975 EXPORT_SYMBOL(yield);
4978 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4979 * that process accounting knows that this is a task in IO wait state.
4981 void __sched io_schedule(void)
4983 struct rq *rq = raw_rq();
4985 delayacct_blkio_start();
4986 atomic_inc(&rq->nr_iowait);
4987 current->in_iowait = 1;
4989 current->in_iowait = 0;
4990 atomic_dec(&rq->nr_iowait);
4991 delayacct_blkio_end();
4993 EXPORT_SYMBOL(io_schedule);
4995 long __sched io_schedule_timeout(long timeout)
4997 struct rq *rq = raw_rq();
5000 delayacct_blkio_start();
5001 atomic_inc(&rq->nr_iowait);
5002 current->in_iowait = 1;
5003 ret = schedule_timeout(timeout);
5004 current->in_iowait = 0;
5005 atomic_dec(&rq->nr_iowait);
5006 delayacct_blkio_end();
5011 * sys_sched_get_priority_max - return maximum RT priority.
5012 * @policy: scheduling class.
5014 * this syscall returns the maximum rt_priority that can be used
5015 * by a given scheduling class.
5017 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5024 ret = MAX_USER_RT_PRIO-1;
5036 * sys_sched_get_priority_min - return minimum RT priority.
5037 * @policy: scheduling class.
5039 * this syscall returns the minimum rt_priority that can be used
5040 * by a given scheduling class.
5042 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5060 * sys_sched_rr_get_interval - return the default timeslice of a process.
5061 * @pid: pid of the process.
5062 * @interval: userspace pointer to the timeslice value.
5064 * this syscall writes the default timeslice value of a given process
5065 * into the user-space timespec buffer. A value of '0' means infinity.
5067 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5068 struct timespec __user *, interval)
5070 struct task_struct *p;
5071 unsigned int time_slice;
5072 unsigned long flags;
5082 p = find_process_by_pid(pid);
5086 retval = security_task_getscheduler(p);
5090 rq = task_rq_lock(p, &flags);
5091 time_slice = p->sched_class->get_rr_interval(rq, p);
5092 task_rq_unlock(rq, &flags);
5095 jiffies_to_timespec(time_slice, &t);
5096 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5104 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5106 void sched_show_task(struct task_struct *p)
5108 unsigned long free = 0;
5111 state = p->state ? __ffs(p->state) + 1 : 0;
5112 printk(KERN_INFO "%-13.13s %c", p->comm,
5113 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5114 #if BITS_PER_LONG == 32
5115 if (state == TASK_RUNNING)
5116 printk(KERN_CONT " running ");
5118 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5120 if (state == TASK_RUNNING)
5121 printk(KERN_CONT " running task ");
5123 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5125 #ifdef CONFIG_DEBUG_STACK_USAGE
5126 free = stack_not_used(p);
5128 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5129 task_pid_nr(p), task_pid_nr(p->real_parent),
5130 (unsigned long)task_thread_info(p)->flags);
5132 show_stack(p, NULL);
5135 void show_state_filter(unsigned long state_filter)
5137 struct task_struct *g, *p;
5139 #if BITS_PER_LONG == 32
5141 " task PC stack pid father\n");
5144 " task PC stack pid father\n");
5146 read_lock(&tasklist_lock);
5147 do_each_thread(g, p) {
5149 * reset the NMI-timeout, listing all files on a slow
5150 * console might take alot of time:
5152 touch_nmi_watchdog();
5153 if (!state_filter || (p->state & state_filter))
5155 } while_each_thread(g, p);
5157 touch_all_softlockup_watchdogs();
5159 #ifdef CONFIG_SCHED_DEBUG
5160 sysrq_sched_debug_show();
5162 read_unlock(&tasklist_lock);
5164 * Only show locks if all tasks are dumped:
5167 debug_show_all_locks();
5170 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5172 idle->sched_class = &idle_sched_class;
5176 * init_idle - set up an idle thread for a given CPU
5177 * @idle: task in question
5178 * @cpu: cpu the idle task belongs to
5180 * NOTE: this function does not set the idle thread's NEED_RESCHED
5181 * flag, to make booting more robust.
5183 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5185 struct rq *rq = cpu_rq(cpu);
5186 unsigned long flags;
5188 raw_spin_lock_irqsave(&rq->lock, flags);
5191 idle->state = TASK_RUNNING;
5192 idle->se.exec_start = sched_clock();
5194 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5195 __set_task_cpu(idle, cpu);
5197 rq->curr = rq->idle = idle;
5198 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5201 raw_spin_unlock_irqrestore(&rq->lock, flags);
5203 /* Set the preempt count _outside_ the spinlocks! */
5204 #if defined(CONFIG_PREEMPT)
5205 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5207 task_thread_info(idle)->preempt_count = 0;
5210 * The idle tasks have their own, simple scheduling class:
5212 idle->sched_class = &idle_sched_class;
5213 ftrace_graph_init_task(idle);
5217 * In a system that switches off the HZ timer nohz_cpu_mask
5218 * indicates which cpus entered this state. This is used
5219 * in the rcu update to wait only for active cpus. For system
5220 * which do not switch off the HZ timer nohz_cpu_mask should
5221 * always be CPU_BITS_NONE.
5223 cpumask_var_t nohz_cpu_mask;
5226 * Increase the granularity value when there are more CPUs,
5227 * because with more CPUs the 'effective latency' as visible
5228 * to users decreases. But the relationship is not linear,
5229 * so pick a second-best guess by going with the log2 of the
5232 * This idea comes from the SD scheduler of Con Kolivas:
5234 static int get_update_sysctl_factor(void)
5236 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5237 unsigned int factor;
5239 switch (sysctl_sched_tunable_scaling) {
5240 case SCHED_TUNABLESCALING_NONE:
5243 case SCHED_TUNABLESCALING_LINEAR:
5246 case SCHED_TUNABLESCALING_LOG:
5248 factor = 1 + ilog2(cpus);
5255 static void update_sysctl(void)
5257 unsigned int factor = get_update_sysctl_factor();
5259 #define SET_SYSCTL(name) \
5260 (sysctl_##name = (factor) * normalized_sysctl_##name)
5261 SET_SYSCTL(sched_min_granularity);
5262 SET_SYSCTL(sched_latency);
5263 SET_SYSCTL(sched_wakeup_granularity);
5264 SET_SYSCTL(sched_shares_ratelimit);
5268 static inline void sched_init_granularity(void)
5275 * This is how migration works:
5277 * 1) we queue a struct migration_req structure in the source CPU's
5278 * runqueue and wake up that CPU's migration thread.
5279 * 2) we down() the locked semaphore => thread blocks.
5280 * 3) migration thread wakes up (implicitly it forces the migrated
5281 * thread off the CPU)
5282 * 4) it gets the migration request and checks whether the migrated
5283 * task is still in the wrong runqueue.
5284 * 5) if it's in the wrong runqueue then the migration thread removes
5285 * it and puts it into the right queue.
5286 * 6) migration thread up()s the semaphore.
5287 * 7) we wake up and the migration is done.
5291 * Change a given task's CPU affinity. Migrate the thread to a
5292 * proper CPU and schedule it away if the CPU it's executing on
5293 * is removed from the allowed bitmask.
5295 * NOTE: the caller must have a valid reference to the task, the
5296 * task must not exit() & deallocate itself prematurely. The
5297 * call is not atomic; no spinlocks may be held.
5299 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5301 struct migration_req req;
5302 unsigned long flags;
5307 * Since we rely on wake-ups to migrate sleeping tasks, don't change
5308 * the ->cpus_allowed mask from under waking tasks, which would be
5309 * possible when we change rq->lock in ttwu(), so synchronize against
5310 * TASK_WAKING to avoid that.
5313 while (p->state == TASK_WAKING)
5316 rq = task_rq_lock(p, &flags);
5318 if (p->state == TASK_WAKING) {
5319 task_rq_unlock(rq, &flags);
5323 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5328 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5329 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5334 if (p->sched_class->set_cpus_allowed)
5335 p->sched_class->set_cpus_allowed(p, new_mask);
5337 cpumask_copy(&p->cpus_allowed, new_mask);
5338 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5341 /* Can the task run on the task's current CPU? If so, we're done */
5342 if (cpumask_test_cpu(task_cpu(p), new_mask))
5345 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5346 /* Need help from migration thread: drop lock and wait. */
5347 struct task_struct *mt = rq->migration_thread;
5349 get_task_struct(mt);
5350 task_rq_unlock(rq, &flags);
5351 wake_up_process(rq->migration_thread);
5352 put_task_struct(mt);
5353 wait_for_completion(&req.done);
5354 tlb_migrate_finish(p->mm);
5358 task_rq_unlock(rq, &flags);
5362 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5365 * Move (not current) task off this cpu, onto dest cpu. We're doing
5366 * this because either it can't run here any more (set_cpus_allowed()
5367 * away from this CPU, or CPU going down), or because we're
5368 * attempting to rebalance this task on exec (sched_exec).
5370 * So we race with normal scheduler movements, but that's OK, as long
5371 * as the task is no longer on this CPU.
5373 * Returns non-zero if task was successfully migrated.
5375 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5377 struct rq *rq_dest, *rq_src;
5380 if (unlikely(!cpu_active(dest_cpu)))
5383 rq_src = cpu_rq(src_cpu);
5384 rq_dest = cpu_rq(dest_cpu);
5386 double_rq_lock(rq_src, rq_dest);
5387 /* Already moved. */
5388 if (task_cpu(p) != src_cpu)
5390 /* Affinity changed (again). */
5391 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5395 * If we're not on a rq, the next wake-up will ensure we're
5399 deactivate_task(rq_src, p, 0);
5400 set_task_cpu(p, dest_cpu);
5401 activate_task(rq_dest, p, 0);
5402 check_preempt_curr(rq_dest, p, 0);
5407 double_rq_unlock(rq_src, rq_dest);
5411 #define RCU_MIGRATION_IDLE 0
5412 #define RCU_MIGRATION_NEED_QS 1
5413 #define RCU_MIGRATION_GOT_QS 2
5414 #define RCU_MIGRATION_MUST_SYNC 3
5417 * migration_thread - this is a highprio system thread that performs
5418 * thread migration by bumping thread off CPU then 'pushing' onto
5421 static int migration_thread(void *data)
5424 int cpu = (long)data;
5428 BUG_ON(rq->migration_thread != current);
5430 set_current_state(TASK_INTERRUPTIBLE);
5431 while (!kthread_should_stop()) {
5432 struct migration_req *req;
5433 struct list_head *head;
5435 raw_spin_lock_irq(&rq->lock);
5437 if (cpu_is_offline(cpu)) {
5438 raw_spin_unlock_irq(&rq->lock);
5442 if (rq->active_balance) {
5443 active_load_balance(rq, cpu);
5444 rq->active_balance = 0;
5447 head = &rq->migration_queue;
5449 if (list_empty(head)) {
5450 raw_spin_unlock_irq(&rq->lock);
5452 set_current_state(TASK_INTERRUPTIBLE);
5455 req = list_entry(head->next, struct migration_req, list);
5456 list_del_init(head->next);
5458 if (req->task != NULL) {
5459 raw_spin_unlock(&rq->lock);
5460 __migrate_task(req->task, cpu, req->dest_cpu);
5461 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5462 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5463 raw_spin_unlock(&rq->lock);
5465 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5466 raw_spin_unlock(&rq->lock);
5467 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5471 complete(&req->done);
5473 __set_current_state(TASK_RUNNING);
5478 #ifdef CONFIG_HOTPLUG_CPU
5480 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5484 local_irq_disable();
5485 ret = __migrate_task(p, src_cpu, dest_cpu);
5491 * Figure out where task on dead CPU should go, use force if necessary.
5493 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5498 dest_cpu = select_fallback_rq(dead_cpu, p);
5500 /* It can have affinity changed while we were choosing. */
5501 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
5506 * While a dead CPU has no uninterruptible tasks queued at this point,
5507 * it might still have a nonzero ->nr_uninterruptible counter, because
5508 * for performance reasons the counter is not stricly tracking tasks to
5509 * their home CPUs. So we just add the counter to another CPU's counter,
5510 * to keep the global sum constant after CPU-down:
5512 static void migrate_nr_uninterruptible(struct rq *rq_src)
5514 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5515 unsigned long flags;
5517 local_irq_save(flags);
5518 double_rq_lock(rq_src, rq_dest);
5519 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5520 rq_src->nr_uninterruptible = 0;
5521 double_rq_unlock(rq_src, rq_dest);
5522 local_irq_restore(flags);
5525 /* Run through task list and migrate tasks from the dead cpu. */
5526 static void migrate_live_tasks(int src_cpu)
5528 struct task_struct *p, *t;
5530 read_lock(&tasklist_lock);
5532 do_each_thread(t, p) {
5536 if (task_cpu(p) == src_cpu)
5537 move_task_off_dead_cpu(src_cpu, p);
5538 } while_each_thread(t, p);
5540 read_unlock(&tasklist_lock);
5544 * Schedules idle task to be the next runnable task on current CPU.
5545 * It does so by boosting its priority to highest possible.
5546 * Used by CPU offline code.
5548 void sched_idle_next(void)
5550 int this_cpu = smp_processor_id();
5551 struct rq *rq = cpu_rq(this_cpu);
5552 struct task_struct *p = rq->idle;
5553 unsigned long flags;
5555 /* cpu has to be offline */
5556 BUG_ON(cpu_online(this_cpu));
5559 * Strictly not necessary since rest of the CPUs are stopped by now
5560 * and interrupts disabled on the current cpu.
5562 raw_spin_lock_irqsave(&rq->lock, flags);
5564 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5566 update_rq_clock(rq);
5567 activate_task(rq, p, 0);
5569 raw_spin_unlock_irqrestore(&rq->lock, flags);
5573 * Ensures that the idle task is using init_mm right before its cpu goes
5576 void idle_task_exit(void)
5578 struct mm_struct *mm = current->active_mm;
5580 BUG_ON(cpu_online(smp_processor_id()));
5583 switch_mm(mm, &init_mm, current);
5587 /* called under rq->lock with disabled interrupts */
5588 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5590 struct rq *rq = cpu_rq(dead_cpu);
5592 /* Must be exiting, otherwise would be on tasklist. */
5593 BUG_ON(!p->exit_state);
5595 /* Cannot have done final schedule yet: would have vanished. */
5596 BUG_ON(p->state == TASK_DEAD);
5601 * Drop lock around migration; if someone else moves it,
5602 * that's OK. No task can be added to this CPU, so iteration is
5605 raw_spin_unlock_irq(&rq->lock);
5606 move_task_off_dead_cpu(dead_cpu, p);
5607 raw_spin_lock_irq(&rq->lock);
5612 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5613 static void migrate_dead_tasks(unsigned int dead_cpu)
5615 struct rq *rq = cpu_rq(dead_cpu);
5616 struct task_struct *next;
5619 if (!rq->nr_running)
5621 update_rq_clock(rq);
5622 next = pick_next_task(rq);
5625 next->sched_class->put_prev_task(rq, next);
5626 migrate_dead(dead_cpu, next);
5632 * remove the tasks which were accounted by rq from calc_load_tasks.
5634 static void calc_global_load_remove(struct rq *rq)
5636 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5637 rq->calc_load_active = 0;
5639 #endif /* CONFIG_HOTPLUG_CPU */
5641 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5643 static struct ctl_table sd_ctl_dir[] = {
5645 .procname = "sched_domain",
5651 static struct ctl_table sd_ctl_root[] = {
5653 .procname = "kernel",
5655 .child = sd_ctl_dir,
5660 static struct ctl_table *sd_alloc_ctl_entry(int n)
5662 struct ctl_table *entry =
5663 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5668 static void sd_free_ctl_entry(struct ctl_table **tablep)
5670 struct ctl_table *entry;
5673 * In the intermediate directories, both the child directory and
5674 * procname are dynamically allocated and could fail but the mode
5675 * will always be set. In the lowest directory the names are
5676 * static strings and all have proc handlers.
5678 for (entry = *tablep; entry->mode; entry++) {
5680 sd_free_ctl_entry(&entry->child);
5681 if (entry->proc_handler == NULL)
5682 kfree(entry->procname);
5690 set_table_entry(struct ctl_table *entry,
5691 const char *procname, void *data, int maxlen,
5692 mode_t mode, proc_handler *proc_handler)
5694 entry->procname = procname;
5696 entry->maxlen = maxlen;
5698 entry->proc_handler = proc_handler;
5701 static struct ctl_table *
5702 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5704 struct ctl_table *table = sd_alloc_ctl_entry(13);
5709 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5710 sizeof(long), 0644, proc_doulongvec_minmax);
5711 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5712 sizeof(long), 0644, proc_doulongvec_minmax);
5713 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5714 sizeof(int), 0644, proc_dointvec_minmax);
5715 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5716 sizeof(int), 0644, proc_dointvec_minmax);
5717 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5718 sizeof(int), 0644, proc_dointvec_minmax);
5719 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5720 sizeof(int), 0644, proc_dointvec_minmax);
5721 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5722 sizeof(int), 0644, proc_dointvec_minmax);
5723 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5724 sizeof(int), 0644, proc_dointvec_minmax);
5725 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5726 sizeof(int), 0644, proc_dointvec_minmax);
5727 set_table_entry(&table[9], "cache_nice_tries",
5728 &sd->cache_nice_tries,
5729 sizeof(int), 0644, proc_dointvec_minmax);
5730 set_table_entry(&table[10], "flags", &sd->flags,
5731 sizeof(int), 0644, proc_dointvec_minmax);
5732 set_table_entry(&table[11], "name", sd->name,
5733 CORENAME_MAX_SIZE, 0444, proc_dostring);
5734 /* &table[12] is terminator */
5739 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5741 struct ctl_table *entry, *table;
5742 struct sched_domain *sd;
5743 int domain_num = 0, i;
5746 for_each_domain(cpu, sd)
5748 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5753 for_each_domain(cpu, sd) {
5754 snprintf(buf, 32, "domain%d", i);
5755 entry->procname = kstrdup(buf, GFP_KERNEL);
5757 entry->child = sd_alloc_ctl_domain_table(sd);
5764 static struct ctl_table_header *sd_sysctl_header;
5765 static void register_sched_domain_sysctl(void)
5767 int i, cpu_num = num_possible_cpus();
5768 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5771 WARN_ON(sd_ctl_dir[0].child);
5772 sd_ctl_dir[0].child = entry;
5777 for_each_possible_cpu(i) {
5778 snprintf(buf, 32, "cpu%d", i);
5779 entry->procname = kstrdup(buf, GFP_KERNEL);
5781 entry->child = sd_alloc_ctl_cpu_table(i);
5785 WARN_ON(sd_sysctl_header);
5786 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5789 /* may be called multiple times per register */
5790 static void unregister_sched_domain_sysctl(void)
5792 if (sd_sysctl_header)
5793 unregister_sysctl_table(sd_sysctl_header);
5794 sd_sysctl_header = NULL;
5795 if (sd_ctl_dir[0].child)
5796 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5799 static void register_sched_domain_sysctl(void)
5802 static void unregister_sched_domain_sysctl(void)
5807 static void set_rq_online(struct rq *rq)
5810 const struct sched_class *class;
5812 cpumask_set_cpu(rq->cpu, rq->rd->online);
5815 for_each_class(class) {
5816 if (class->rq_online)
5817 class->rq_online(rq);
5822 static void set_rq_offline(struct rq *rq)
5825 const struct sched_class *class;
5827 for_each_class(class) {
5828 if (class->rq_offline)
5829 class->rq_offline(rq);
5832 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5838 * migration_call - callback that gets triggered when a CPU is added.
5839 * Here we can start up the necessary migration thread for the new CPU.
5841 static int __cpuinit
5842 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5844 struct task_struct *p;
5845 int cpu = (long)hcpu;
5846 unsigned long flags;
5851 case CPU_UP_PREPARE:
5852 case CPU_UP_PREPARE_FROZEN:
5853 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5856 kthread_bind(p, cpu);
5857 /* Must be high prio: stop_machine expects to yield to it. */
5858 rq = task_rq_lock(p, &flags);
5859 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5860 task_rq_unlock(rq, &flags);
5862 cpu_rq(cpu)->migration_thread = p;
5863 rq->calc_load_update = calc_load_update;
5867 case CPU_ONLINE_FROZEN:
5868 /* Strictly unnecessary, as first user will wake it. */
5869 wake_up_process(cpu_rq(cpu)->migration_thread);
5871 /* Update our root-domain */
5873 raw_spin_lock_irqsave(&rq->lock, flags);
5875 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5879 raw_spin_unlock_irqrestore(&rq->lock, flags);
5882 #ifdef CONFIG_HOTPLUG_CPU
5883 case CPU_UP_CANCELED:
5884 case CPU_UP_CANCELED_FROZEN:
5885 if (!cpu_rq(cpu)->migration_thread)
5887 /* Unbind it from offline cpu so it can run. Fall thru. */
5888 kthread_bind(cpu_rq(cpu)->migration_thread,
5889 cpumask_any(cpu_online_mask));
5890 kthread_stop(cpu_rq(cpu)->migration_thread);
5891 put_task_struct(cpu_rq(cpu)->migration_thread);
5892 cpu_rq(cpu)->migration_thread = NULL;
5896 case CPU_DEAD_FROZEN:
5897 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5898 migrate_live_tasks(cpu);
5900 kthread_stop(rq->migration_thread);
5901 put_task_struct(rq->migration_thread);
5902 rq->migration_thread = NULL;
5903 /* Idle task back to normal (off runqueue, low prio) */
5904 raw_spin_lock_irq(&rq->lock);
5905 update_rq_clock(rq);
5906 deactivate_task(rq, rq->idle, 0);
5907 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5908 rq->idle->sched_class = &idle_sched_class;
5909 migrate_dead_tasks(cpu);
5910 raw_spin_unlock_irq(&rq->lock);
5912 migrate_nr_uninterruptible(rq);
5913 BUG_ON(rq->nr_running != 0);
5914 calc_global_load_remove(rq);
5916 * No need to migrate the tasks: it was best-effort if
5917 * they didn't take sched_hotcpu_mutex. Just wake up
5920 raw_spin_lock_irq(&rq->lock);
5921 while (!list_empty(&rq->migration_queue)) {
5922 struct migration_req *req;
5924 req = list_entry(rq->migration_queue.next,
5925 struct migration_req, list);
5926 list_del_init(&req->list);
5927 raw_spin_unlock_irq(&rq->lock);
5928 complete(&req->done);
5929 raw_spin_lock_irq(&rq->lock);
5931 raw_spin_unlock_irq(&rq->lock);
5935 case CPU_DYING_FROZEN:
5936 /* Update our root-domain */
5938 raw_spin_lock_irqsave(&rq->lock, flags);
5940 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5943 raw_spin_unlock_irqrestore(&rq->lock, flags);
5951 * Register at high priority so that task migration (migrate_all_tasks)
5952 * happens before everything else. This has to be lower priority than
5953 * the notifier in the perf_event subsystem, though.
5955 static struct notifier_block __cpuinitdata migration_notifier = {
5956 .notifier_call = migration_call,
5960 static int __init migration_init(void)
5962 void *cpu = (void *)(long)smp_processor_id();
5965 /* Start one for the boot CPU: */
5966 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5967 BUG_ON(err == NOTIFY_BAD);
5968 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5969 register_cpu_notifier(&migration_notifier);
5973 early_initcall(migration_init);
5978 #ifdef CONFIG_SCHED_DEBUG
5980 static __read_mostly int sched_domain_debug_enabled;
5982 static int __init sched_domain_debug_setup(char *str)
5984 sched_domain_debug_enabled = 1;
5988 early_param("sched_debug", sched_domain_debug_setup);
5990 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5991 struct cpumask *groupmask)
5993 struct sched_group *group = sd->groups;
5996 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5997 cpumask_clear(groupmask);
5999 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6001 if (!(sd->flags & SD_LOAD_BALANCE)) {
6002 printk("does not load-balance\n");
6004 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6009 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6011 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6012 printk(KERN_ERR "ERROR: domain->span does not contain "
6015 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6016 printk(KERN_ERR "ERROR: domain->groups does not contain"
6020 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6024 printk(KERN_ERR "ERROR: group is NULL\n");
6028 if (!group->cpu_power) {
6029 printk(KERN_CONT "\n");
6030 printk(KERN_ERR "ERROR: domain->cpu_power not "
6035 if (!cpumask_weight(sched_group_cpus(group))) {
6036 printk(KERN_CONT "\n");
6037 printk(KERN_ERR "ERROR: empty group\n");
6041 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6042 printk(KERN_CONT "\n");
6043 printk(KERN_ERR "ERROR: repeated CPUs\n");
6047 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6049 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6051 printk(KERN_CONT " %s", str);
6052 if (group->cpu_power != SCHED_LOAD_SCALE) {
6053 printk(KERN_CONT " (cpu_power = %d)",
6057 group = group->next;
6058 } while (group != sd->groups);
6059 printk(KERN_CONT "\n");
6061 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6062 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6065 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6066 printk(KERN_ERR "ERROR: parent span is not a superset "
6067 "of domain->span\n");
6071 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6073 cpumask_var_t groupmask;
6076 if (!sched_domain_debug_enabled)
6080 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6084 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6086 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6087 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6092 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6099 free_cpumask_var(groupmask);
6101 #else /* !CONFIG_SCHED_DEBUG */
6102 # define sched_domain_debug(sd, cpu) do { } while (0)
6103 #endif /* CONFIG_SCHED_DEBUG */
6105 static int sd_degenerate(struct sched_domain *sd)
6107 if (cpumask_weight(sched_domain_span(sd)) == 1)
6110 /* Following flags need at least 2 groups */
6111 if (sd->flags & (SD_LOAD_BALANCE |
6112 SD_BALANCE_NEWIDLE |
6116 SD_SHARE_PKG_RESOURCES)) {
6117 if (sd->groups != sd->groups->next)
6121 /* Following flags don't use groups */
6122 if (sd->flags & (SD_WAKE_AFFINE))
6129 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6131 unsigned long cflags = sd->flags, pflags = parent->flags;
6133 if (sd_degenerate(parent))
6136 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6139 /* Flags needing groups don't count if only 1 group in parent */
6140 if (parent->groups == parent->groups->next) {
6141 pflags &= ~(SD_LOAD_BALANCE |
6142 SD_BALANCE_NEWIDLE |
6146 SD_SHARE_PKG_RESOURCES);
6147 if (nr_node_ids == 1)
6148 pflags &= ~SD_SERIALIZE;
6150 if (~cflags & pflags)
6156 static void free_rootdomain(struct root_domain *rd)
6158 synchronize_sched();
6160 cpupri_cleanup(&rd->cpupri);
6162 free_cpumask_var(rd->rto_mask);
6163 free_cpumask_var(rd->online);
6164 free_cpumask_var(rd->span);
6168 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6170 struct root_domain *old_rd = NULL;
6171 unsigned long flags;
6173 raw_spin_lock_irqsave(&rq->lock, flags);
6178 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6181 cpumask_clear_cpu(rq->cpu, old_rd->span);
6184 * If we dont want to free the old_rt yet then
6185 * set old_rd to NULL to skip the freeing later
6188 if (!atomic_dec_and_test(&old_rd->refcount))
6192 atomic_inc(&rd->refcount);
6195 cpumask_set_cpu(rq->cpu, rd->span);
6196 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6199 raw_spin_unlock_irqrestore(&rq->lock, flags);
6202 free_rootdomain(old_rd);
6205 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6207 gfp_t gfp = GFP_KERNEL;
6209 memset(rd, 0, sizeof(*rd));
6214 if (!alloc_cpumask_var(&rd->span, gfp))
6216 if (!alloc_cpumask_var(&rd->online, gfp))
6218 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6221 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6226 free_cpumask_var(rd->rto_mask);
6228 free_cpumask_var(rd->online);
6230 free_cpumask_var(rd->span);
6235 static void init_defrootdomain(void)
6237 init_rootdomain(&def_root_domain, true);
6239 atomic_set(&def_root_domain.refcount, 1);
6242 static struct root_domain *alloc_rootdomain(void)
6244 struct root_domain *rd;
6246 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6250 if (init_rootdomain(rd, false) != 0) {
6259 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6260 * hold the hotplug lock.
6263 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6265 struct rq *rq = cpu_rq(cpu);
6266 struct sched_domain *tmp;
6268 /* Remove the sched domains which do not contribute to scheduling. */
6269 for (tmp = sd; tmp; ) {
6270 struct sched_domain *parent = tmp->parent;
6274 if (sd_parent_degenerate(tmp, parent)) {
6275 tmp->parent = parent->parent;
6277 parent->parent->child = tmp;
6282 if (sd && sd_degenerate(sd)) {
6288 sched_domain_debug(sd, cpu);
6290 rq_attach_root(rq, rd);
6291 rcu_assign_pointer(rq->sd, sd);
6294 /* cpus with isolated domains */
6295 static cpumask_var_t cpu_isolated_map;
6297 /* Setup the mask of cpus configured for isolated domains */
6298 static int __init isolated_cpu_setup(char *str)
6300 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6301 cpulist_parse(str, cpu_isolated_map);
6305 __setup("isolcpus=", isolated_cpu_setup);
6308 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6309 * to a function which identifies what group(along with sched group) a CPU
6310 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6311 * (due to the fact that we keep track of groups covered with a struct cpumask).
6313 * init_sched_build_groups will build a circular linked list of the groups
6314 * covered by the given span, and will set each group's ->cpumask correctly,
6315 * and ->cpu_power to 0.
6318 init_sched_build_groups(const struct cpumask *span,
6319 const struct cpumask *cpu_map,
6320 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6321 struct sched_group **sg,
6322 struct cpumask *tmpmask),
6323 struct cpumask *covered, struct cpumask *tmpmask)
6325 struct sched_group *first = NULL, *last = NULL;
6328 cpumask_clear(covered);
6330 for_each_cpu(i, span) {
6331 struct sched_group *sg;
6332 int group = group_fn(i, cpu_map, &sg, tmpmask);
6335 if (cpumask_test_cpu(i, covered))
6338 cpumask_clear(sched_group_cpus(sg));
6341 for_each_cpu(j, span) {
6342 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6345 cpumask_set_cpu(j, covered);
6346 cpumask_set_cpu(j, sched_group_cpus(sg));
6357 #define SD_NODES_PER_DOMAIN 16
6362 * find_next_best_node - find the next node to include in a sched_domain
6363 * @node: node whose sched_domain we're building
6364 * @used_nodes: nodes already in the sched_domain
6366 * Find the next node to include in a given scheduling domain. Simply
6367 * finds the closest node not already in the @used_nodes map.
6369 * Should use nodemask_t.
6371 static int find_next_best_node(int node, nodemask_t *used_nodes)
6373 int i, n, val, min_val, best_node = 0;
6377 for (i = 0; i < nr_node_ids; i++) {
6378 /* Start at @node */
6379 n = (node + i) % nr_node_ids;
6381 if (!nr_cpus_node(n))
6384 /* Skip already used nodes */
6385 if (node_isset(n, *used_nodes))
6388 /* Simple min distance search */
6389 val = node_distance(node, n);
6391 if (val < min_val) {
6397 node_set(best_node, *used_nodes);
6402 * sched_domain_node_span - get a cpumask for a node's sched_domain
6403 * @node: node whose cpumask we're constructing
6404 * @span: resulting cpumask
6406 * Given a node, construct a good cpumask for its sched_domain to span. It
6407 * should be one that prevents unnecessary balancing, but also spreads tasks
6410 static void sched_domain_node_span(int node, struct cpumask *span)
6412 nodemask_t used_nodes;
6415 cpumask_clear(span);
6416 nodes_clear(used_nodes);
6418 cpumask_or(span, span, cpumask_of_node(node));
6419 node_set(node, used_nodes);
6421 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6422 int next_node = find_next_best_node(node, &used_nodes);
6424 cpumask_or(span, span, cpumask_of_node(next_node));
6427 #endif /* CONFIG_NUMA */
6429 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6432 * The cpus mask in sched_group and sched_domain hangs off the end.
6434 * ( See the the comments in include/linux/sched.h:struct sched_group
6435 * and struct sched_domain. )
6437 struct static_sched_group {
6438 struct sched_group sg;
6439 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6442 struct static_sched_domain {
6443 struct sched_domain sd;
6444 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6450 cpumask_var_t domainspan;
6451 cpumask_var_t covered;
6452 cpumask_var_t notcovered;
6454 cpumask_var_t nodemask;
6455 cpumask_var_t this_sibling_map;
6456 cpumask_var_t this_core_map;
6457 cpumask_var_t send_covered;
6458 cpumask_var_t tmpmask;
6459 struct sched_group **sched_group_nodes;
6460 struct root_domain *rd;
6464 sa_sched_groups = 0,
6469 sa_this_sibling_map,
6471 sa_sched_group_nodes,
6481 * SMT sched-domains:
6483 #ifdef CONFIG_SCHED_SMT
6484 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6485 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6488 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6489 struct sched_group **sg, struct cpumask *unused)
6492 *sg = &per_cpu(sched_groups, cpu).sg;
6495 #endif /* CONFIG_SCHED_SMT */
6498 * multi-core sched-domains:
6500 #ifdef CONFIG_SCHED_MC
6501 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6502 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6503 #endif /* CONFIG_SCHED_MC */
6505 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6507 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6508 struct sched_group **sg, struct cpumask *mask)
6512 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6513 group = cpumask_first(mask);
6515 *sg = &per_cpu(sched_group_core, group).sg;
6518 #elif defined(CONFIG_SCHED_MC)
6520 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6521 struct sched_group **sg, struct cpumask *unused)
6524 *sg = &per_cpu(sched_group_core, cpu).sg;
6529 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6530 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6533 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6534 struct sched_group **sg, struct cpumask *mask)
6537 #ifdef CONFIG_SCHED_MC
6538 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6539 group = cpumask_first(mask);
6540 #elif defined(CONFIG_SCHED_SMT)
6541 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6542 group = cpumask_first(mask);
6547 *sg = &per_cpu(sched_group_phys, group).sg;
6553 * The init_sched_build_groups can't handle what we want to do with node
6554 * groups, so roll our own. Now each node has its own list of groups which
6555 * gets dynamically allocated.
6557 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6558 static struct sched_group ***sched_group_nodes_bycpu;
6560 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6561 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6563 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6564 struct sched_group **sg,
6565 struct cpumask *nodemask)
6569 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6570 group = cpumask_first(nodemask);
6573 *sg = &per_cpu(sched_group_allnodes, group).sg;
6577 static void init_numa_sched_groups_power(struct sched_group *group_head)
6579 struct sched_group *sg = group_head;
6585 for_each_cpu(j, sched_group_cpus(sg)) {
6586 struct sched_domain *sd;
6588 sd = &per_cpu(phys_domains, j).sd;
6589 if (j != group_first_cpu(sd->groups)) {
6591 * Only add "power" once for each
6597 sg->cpu_power += sd->groups->cpu_power;
6600 } while (sg != group_head);
6603 static int build_numa_sched_groups(struct s_data *d,
6604 const struct cpumask *cpu_map, int num)
6606 struct sched_domain *sd;
6607 struct sched_group *sg, *prev;
6610 cpumask_clear(d->covered);
6611 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6612 if (cpumask_empty(d->nodemask)) {
6613 d->sched_group_nodes[num] = NULL;
6617 sched_domain_node_span(num, d->domainspan);
6618 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6620 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6623 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6627 d->sched_group_nodes[num] = sg;
6629 for_each_cpu(j, d->nodemask) {
6630 sd = &per_cpu(node_domains, j).sd;
6635 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6637 cpumask_or(d->covered, d->covered, d->nodemask);
6640 for (j = 0; j < nr_node_ids; j++) {
6641 n = (num + j) % nr_node_ids;
6642 cpumask_complement(d->notcovered, d->covered);
6643 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6644 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6645 if (cpumask_empty(d->tmpmask))
6647 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6648 if (cpumask_empty(d->tmpmask))
6650 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6654 "Can not alloc domain group for node %d\n", j);
6658 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6659 sg->next = prev->next;
6660 cpumask_or(d->covered, d->covered, d->tmpmask);
6667 #endif /* CONFIG_NUMA */
6670 /* Free memory allocated for various sched_group structures */
6671 static void free_sched_groups(const struct cpumask *cpu_map,
6672 struct cpumask *nodemask)
6676 for_each_cpu(cpu, cpu_map) {
6677 struct sched_group **sched_group_nodes
6678 = sched_group_nodes_bycpu[cpu];
6680 if (!sched_group_nodes)
6683 for (i = 0; i < nr_node_ids; i++) {
6684 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6686 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6687 if (cpumask_empty(nodemask))
6697 if (oldsg != sched_group_nodes[i])
6700 kfree(sched_group_nodes);
6701 sched_group_nodes_bycpu[cpu] = NULL;
6704 #else /* !CONFIG_NUMA */
6705 static void free_sched_groups(const struct cpumask *cpu_map,
6706 struct cpumask *nodemask)
6709 #endif /* CONFIG_NUMA */
6712 * Initialize sched groups cpu_power.
6714 * cpu_power indicates the capacity of sched group, which is used while
6715 * distributing the load between different sched groups in a sched domain.
6716 * Typically cpu_power for all the groups in a sched domain will be same unless
6717 * there are asymmetries in the topology. If there are asymmetries, group
6718 * having more cpu_power will pickup more load compared to the group having
6721 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6723 struct sched_domain *child;
6724 struct sched_group *group;
6728 WARN_ON(!sd || !sd->groups);
6730 if (cpu != group_first_cpu(sd->groups))
6735 sd->groups->cpu_power = 0;
6738 power = SCHED_LOAD_SCALE;
6739 weight = cpumask_weight(sched_domain_span(sd));
6741 * SMT siblings share the power of a single core.
6742 * Usually multiple threads get a better yield out of
6743 * that one core than a single thread would have,
6744 * reflect that in sd->smt_gain.
6746 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6747 power *= sd->smt_gain;
6749 power >>= SCHED_LOAD_SHIFT;
6751 sd->groups->cpu_power += power;
6756 * Add cpu_power of each child group to this groups cpu_power.
6758 group = child->groups;
6760 sd->groups->cpu_power += group->cpu_power;
6761 group = group->next;
6762 } while (group != child->groups);
6766 * Initializers for schedule domains
6767 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6770 #ifdef CONFIG_SCHED_DEBUG
6771 # define SD_INIT_NAME(sd, type) sd->name = #type
6773 # define SD_INIT_NAME(sd, type) do { } while (0)
6776 #define SD_INIT(sd, type) sd_init_##type(sd)
6778 #define SD_INIT_FUNC(type) \
6779 static noinline void sd_init_##type(struct sched_domain *sd) \
6781 memset(sd, 0, sizeof(*sd)); \
6782 *sd = SD_##type##_INIT; \
6783 sd->level = SD_LV_##type; \
6784 SD_INIT_NAME(sd, type); \
6789 SD_INIT_FUNC(ALLNODES)
6792 #ifdef CONFIG_SCHED_SMT
6793 SD_INIT_FUNC(SIBLING)
6795 #ifdef CONFIG_SCHED_MC
6799 static int default_relax_domain_level = -1;
6801 static int __init setup_relax_domain_level(char *str)
6805 val = simple_strtoul(str, NULL, 0);
6806 if (val < SD_LV_MAX)
6807 default_relax_domain_level = val;
6811 __setup("relax_domain_level=", setup_relax_domain_level);
6813 static void set_domain_attribute(struct sched_domain *sd,
6814 struct sched_domain_attr *attr)
6818 if (!attr || attr->relax_domain_level < 0) {
6819 if (default_relax_domain_level < 0)
6822 request = default_relax_domain_level;
6824 request = attr->relax_domain_level;
6825 if (request < sd->level) {
6826 /* turn off idle balance on this domain */
6827 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6829 /* turn on idle balance on this domain */
6830 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6834 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6835 const struct cpumask *cpu_map)
6838 case sa_sched_groups:
6839 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6840 d->sched_group_nodes = NULL;
6842 free_rootdomain(d->rd); /* fall through */
6844 free_cpumask_var(d->tmpmask); /* fall through */
6845 case sa_send_covered:
6846 free_cpumask_var(d->send_covered); /* fall through */
6847 case sa_this_core_map:
6848 free_cpumask_var(d->this_core_map); /* fall through */
6849 case sa_this_sibling_map:
6850 free_cpumask_var(d->this_sibling_map); /* fall through */
6852 free_cpumask_var(d->nodemask); /* fall through */
6853 case sa_sched_group_nodes:
6855 kfree(d->sched_group_nodes); /* fall through */
6857 free_cpumask_var(d->notcovered); /* fall through */
6859 free_cpumask_var(d->covered); /* fall through */
6861 free_cpumask_var(d->domainspan); /* fall through */
6868 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6869 const struct cpumask *cpu_map)
6872 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6874 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6875 return sa_domainspan;
6876 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6878 /* Allocate the per-node list of sched groups */
6879 d->sched_group_nodes = kcalloc(nr_node_ids,
6880 sizeof(struct sched_group *), GFP_KERNEL);
6881 if (!d->sched_group_nodes) {
6882 printk(KERN_WARNING "Can not alloc sched group node list\n");
6883 return sa_notcovered;
6885 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6887 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6888 return sa_sched_group_nodes;
6889 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6891 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6892 return sa_this_sibling_map;
6893 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6894 return sa_this_core_map;
6895 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6896 return sa_send_covered;
6897 d->rd = alloc_rootdomain();
6899 printk(KERN_WARNING "Cannot alloc root domain\n");
6902 return sa_rootdomain;
6905 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6906 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6908 struct sched_domain *sd = NULL;
6910 struct sched_domain *parent;
6913 if (cpumask_weight(cpu_map) >
6914 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6915 sd = &per_cpu(allnodes_domains, i).sd;
6916 SD_INIT(sd, ALLNODES);
6917 set_domain_attribute(sd, attr);
6918 cpumask_copy(sched_domain_span(sd), cpu_map);
6919 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6924 sd = &per_cpu(node_domains, i).sd;
6926 set_domain_attribute(sd, attr);
6927 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6928 sd->parent = parent;
6931 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6936 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6937 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6938 struct sched_domain *parent, int i)
6940 struct sched_domain *sd;
6941 sd = &per_cpu(phys_domains, i).sd;
6943 set_domain_attribute(sd, attr);
6944 cpumask_copy(sched_domain_span(sd), d->nodemask);
6945 sd->parent = parent;
6948 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6952 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6953 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6954 struct sched_domain *parent, int i)
6956 struct sched_domain *sd = parent;
6957 #ifdef CONFIG_SCHED_MC
6958 sd = &per_cpu(core_domains, i).sd;
6960 set_domain_attribute(sd, attr);
6961 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6962 sd->parent = parent;
6964 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6969 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6970 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6971 struct sched_domain *parent, int i)
6973 struct sched_domain *sd = parent;
6974 #ifdef CONFIG_SCHED_SMT
6975 sd = &per_cpu(cpu_domains, i).sd;
6976 SD_INIT(sd, SIBLING);
6977 set_domain_attribute(sd, attr);
6978 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6979 sd->parent = parent;
6981 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6986 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6987 const struct cpumask *cpu_map, int cpu)
6990 #ifdef CONFIG_SCHED_SMT
6991 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6992 cpumask_and(d->this_sibling_map, cpu_map,
6993 topology_thread_cpumask(cpu));
6994 if (cpu == cpumask_first(d->this_sibling_map))
6995 init_sched_build_groups(d->this_sibling_map, cpu_map,
6997 d->send_covered, d->tmpmask);
7000 #ifdef CONFIG_SCHED_MC
7001 case SD_LV_MC: /* set up multi-core groups */
7002 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7003 if (cpu == cpumask_first(d->this_core_map))
7004 init_sched_build_groups(d->this_core_map, cpu_map,
7006 d->send_covered, d->tmpmask);
7009 case SD_LV_CPU: /* set up physical groups */
7010 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7011 if (!cpumask_empty(d->nodemask))
7012 init_sched_build_groups(d->nodemask, cpu_map,
7014 d->send_covered, d->tmpmask);
7017 case SD_LV_ALLNODES:
7018 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7019 d->send_covered, d->tmpmask);
7028 * Build sched domains for a given set of cpus and attach the sched domains
7029 * to the individual cpus
7031 static int __build_sched_domains(const struct cpumask *cpu_map,
7032 struct sched_domain_attr *attr)
7034 enum s_alloc alloc_state = sa_none;
7036 struct sched_domain *sd;
7042 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7043 if (alloc_state != sa_rootdomain)
7045 alloc_state = sa_sched_groups;
7048 * Set up domains for cpus specified by the cpu_map.
7050 for_each_cpu(i, cpu_map) {
7051 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7054 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7055 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7056 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7057 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7060 for_each_cpu(i, cpu_map) {
7061 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7062 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7065 /* Set up physical groups */
7066 for (i = 0; i < nr_node_ids; i++)
7067 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7070 /* Set up node groups */
7072 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7074 for (i = 0; i < nr_node_ids; i++)
7075 if (build_numa_sched_groups(&d, cpu_map, i))
7079 /* Calculate CPU power for physical packages and nodes */
7080 #ifdef CONFIG_SCHED_SMT
7081 for_each_cpu(i, cpu_map) {
7082 sd = &per_cpu(cpu_domains, i).sd;
7083 init_sched_groups_power(i, sd);
7086 #ifdef CONFIG_SCHED_MC
7087 for_each_cpu(i, cpu_map) {
7088 sd = &per_cpu(core_domains, i).sd;
7089 init_sched_groups_power(i, sd);
7093 for_each_cpu(i, cpu_map) {
7094 sd = &per_cpu(phys_domains, i).sd;
7095 init_sched_groups_power(i, sd);
7099 for (i = 0; i < nr_node_ids; i++)
7100 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7102 if (d.sd_allnodes) {
7103 struct sched_group *sg;
7105 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7107 init_numa_sched_groups_power(sg);
7111 /* Attach the domains */
7112 for_each_cpu(i, cpu_map) {
7113 #ifdef CONFIG_SCHED_SMT
7114 sd = &per_cpu(cpu_domains, i).sd;
7115 #elif defined(CONFIG_SCHED_MC)
7116 sd = &per_cpu(core_domains, i).sd;
7118 sd = &per_cpu(phys_domains, i).sd;
7120 cpu_attach_domain(sd, d.rd, i);
7123 d.sched_group_nodes = NULL; /* don't free this we still need it */
7124 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7128 __free_domain_allocs(&d, alloc_state, cpu_map);
7132 static int build_sched_domains(const struct cpumask *cpu_map)
7134 return __build_sched_domains(cpu_map, NULL);
7137 static cpumask_var_t *doms_cur; /* current sched domains */
7138 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7139 static struct sched_domain_attr *dattr_cur;
7140 /* attribues of custom domains in 'doms_cur' */
7143 * Special case: If a kmalloc of a doms_cur partition (array of
7144 * cpumask) fails, then fallback to a single sched domain,
7145 * as determined by the single cpumask fallback_doms.
7147 static cpumask_var_t fallback_doms;
7150 * arch_update_cpu_topology lets virtualized architectures update the
7151 * cpu core maps. It is supposed to return 1 if the topology changed
7152 * or 0 if it stayed the same.
7154 int __attribute__((weak)) arch_update_cpu_topology(void)
7159 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7162 cpumask_var_t *doms;
7164 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7167 for (i = 0; i < ndoms; i++) {
7168 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7169 free_sched_domains(doms, i);
7176 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7179 for (i = 0; i < ndoms; i++)
7180 free_cpumask_var(doms[i]);
7185 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7186 * For now this just excludes isolated cpus, but could be used to
7187 * exclude other special cases in the future.
7189 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7193 arch_update_cpu_topology();
7195 doms_cur = alloc_sched_domains(ndoms_cur);
7197 doms_cur = &fallback_doms;
7198 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7200 err = build_sched_domains(doms_cur[0]);
7201 register_sched_domain_sysctl();
7206 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7207 struct cpumask *tmpmask)
7209 free_sched_groups(cpu_map, tmpmask);
7213 * Detach sched domains from a group of cpus specified in cpu_map
7214 * These cpus will now be attached to the NULL domain
7216 static void detach_destroy_domains(const struct cpumask *cpu_map)
7218 /* Save because hotplug lock held. */
7219 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7222 for_each_cpu(i, cpu_map)
7223 cpu_attach_domain(NULL, &def_root_domain, i);
7224 synchronize_sched();
7225 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7228 /* handle null as "default" */
7229 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7230 struct sched_domain_attr *new, int idx_new)
7232 struct sched_domain_attr tmp;
7239 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7240 new ? (new + idx_new) : &tmp,
7241 sizeof(struct sched_domain_attr));
7245 * Partition sched domains as specified by the 'ndoms_new'
7246 * cpumasks in the array doms_new[] of cpumasks. This compares
7247 * doms_new[] to the current sched domain partitioning, doms_cur[].
7248 * It destroys each deleted domain and builds each new domain.
7250 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7251 * The masks don't intersect (don't overlap.) We should setup one
7252 * sched domain for each mask. CPUs not in any of the cpumasks will
7253 * not be load balanced. If the same cpumask appears both in the
7254 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7257 * The passed in 'doms_new' should be allocated using
7258 * alloc_sched_domains. This routine takes ownership of it and will
7259 * free_sched_domains it when done with it. If the caller failed the
7260 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7261 * and partition_sched_domains() will fallback to the single partition
7262 * 'fallback_doms', it also forces the domains to be rebuilt.
7264 * If doms_new == NULL it will be replaced with cpu_online_mask.
7265 * ndoms_new == 0 is a special case for destroying existing domains,
7266 * and it will not create the default domain.
7268 * Call with hotplug lock held
7270 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7271 struct sched_domain_attr *dattr_new)
7276 mutex_lock(&sched_domains_mutex);
7278 /* always unregister in case we don't destroy any domains */
7279 unregister_sched_domain_sysctl();
7281 /* Let architecture update cpu core mappings. */
7282 new_topology = arch_update_cpu_topology();
7284 n = doms_new ? ndoms_new : 0;
7286 /* Destroy deleted domains */
7287 for (i = 0; i < ndoms_cur; i++) {
7288 for (j = 0; j < n && !new_topology; j++) {
7289 if (cpumask_equal(doms_cur[i], doms_new[j])
7290 && dattrs_equal(dattr_cur, i, dattr_new, j))
7293 /* no match - a current sched domain not in new doms_new[] */
7294 detach_destroy_domains(doms_cur[i]);
7299 if (doms_new == NULL) {
7301 doms_new = &fallback_doms;
7302 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7303 WARN_ON_ONCE(dattr_new);
7306 /* Build new domains */
7307 for (i = 0; i < ndoms_new; i++) {
7308 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7309 if (cpumask_equal(doms_new[i], doms_cur[j])
7310 && dattrs_equal(dattr_new, i, dattr_cur, j))
7313 /* no match - add a new doms_new */
7314 __build_sched_domains(doms_new[i],
7315 dattr_new ? dattr_new + i : NULL);
7320 /* Remember the new sched domains */
7321 if (doms_cur != &fallback_doms)
7322 free_sched_domains(doms_cur, ndoms_cur);
7323 kfree(dattr_cur); /* kfree(NULL) is safe */
7324 doms_cur = doms_new;
7325 dattr_cur = dattr_new;
7326 ndoms_cur = ndoms_new;
7328 register_sched_domain_sysctl();
7330 mutex_unlock(&sched_domains_mutex);
7333 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7334 static void arch_reinit_sched_domains(void)
7338 /* Destroy domains first to force the rebuild */
7339 partition_sched_domains(0, NULL, NULL);
7341 rebuild_sched_domains();
7345 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7347 unsigned int level = 0;
7349 if (sscanf(buf, "%u", &level) != 1)
7353 * level is always be positive so don't check for
7354 * level < POWERSAVINGS_BALANCE_NONE which is 0
7355 * What happens on 0 or 1 byte write,
7356 * need to check for count as well?
7359 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7363 sched_smt_power_savings = level;
7365 sched_mc_power_savings = level;
7367 arch_reinit_sched_domains();
7372 #ifdef CONFIG_SCHED_MC
7373 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7376 return sprintf(page, "%u\n", sched_mc_power_savings);
7378 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7379 const char *buf, size_t count)
7381 return sched_power_savings_store(buf, count, 0);
7383 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7384 sched_mc_power_savings_show,
7385 sched_mc_power_savings_store);
7388 #ifdef CONFIG_SCHED_SMT
7389 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7392 return sprintf(page, "%u\n", sched_smt_power_savings);
7394 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7395 const char *buf, size_t count)
7397 return sched_power_savings_store(buf, count, 1);
7399 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7400 sched_smt_power_savings_show,
7401 sched_smt_power_savings_store);
7404 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7408 #ifdef CONFIG_SCHED_SMT
7410 err = sysfs_create_file(&cls->kset.kobj,
7411 &attr_sched_smt_power_savings.attr);
7413 #ifdef CONFIG_SCHED_MC
7414 if (!err && mc_capable())
7415 err = sysfs_create_file(&cls->kset.kobj,
7416 &attr_sched_mc_power_savings.attr);
7420 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7422 #ifndef CONFIG_CPUSETS
7424 * Add online and remove offline CPUs from the scheduler domains.
7425 * When cpusets are enabled they take over this function.
7427 static int update_sched_domains(struct notifier_block *nfb,
7428 unsigned long action, void *hcpu)
7432 case CPU_ONLINE_FROZEN:
7433 case CPU_DOWN_PREPARE:
7434 case CPU_DOWN_PREPARE_FROZEN:
7435 case CPU_DOWN_FAILED:
7436 case CPU_DOWN_FAILED_FROZEN:
7437 partition_sched_domains(1, NULL, NULL);
7446 static int update_runtime(struct notifier_block *nfb,
7447 unsigned long action, void *hcpu)
7449 int cpu = (int)(long)hcpu;
7452 case CPU_DOWN_PREPARE:
7453 case CPU_DOWN_PREPARE_FROZEN:
7454 disable_runtime(cpu_rq(cpu));
7457 case CPU_DOWN_FAILED:
7458 case CPU_DOWN_FAILED_FROZEN:
7460 case CPU_ONLINE_FROZEN:
7461 enable_runtime(cpu_rq(cpu));
7469 void __init sched_init_smp(void)
7471 cpumask_var_t non_isolated_cpus;
7473 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7474 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7476 #if defined(CONFIG_NUMA)
7477 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7479 BUG_ON(sched_group_nodes_bycpu == NULL);
7482 mutex_lock(&sched_domains_mutex);
7483 arch_init_sched_domains(cpu_active_mask);
7484 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7485 if (cpumask_empty(non_isolated_cpus))
7486 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7487 mutex_unlock(&sched_domains_mutex);
7490 #ifndef CONFIG_CPUSETS
7491 /* XXX: Theoretical race here - CPU may be hotplugged now */
7492 hotcpu_notifier(update_sched_domains, 0);
7495 /* RT runtime code needs to handle some hotplug events */
7496 hotcpu_notifier(update_runtime, 0);
7500 /* Move init over to a non-isolated CPU */
7501 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7503 sched_init_granularity();
7504 free_cpumask_var(non_isolated_cpus);
7506 init_sched_rt_class();
7509 void __init sched_init_smp(void)
7511 sched_init_granularity();
7513 #endif /* CONFIG_SMP */
7515 const_debug unsigned int sysctl_timer_migration = 1;
7517 int in_sched_functions(unsigned long addr)
7519 return in_lock_functions(addr) ||
7520 (addr >= (unsigned long)__sched_text_start
7521 && addr < (unsigned long)__sched_text_end);
7524 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7526 cfs_rq->tasks_timeline = RB_ROOT;
7527 INIT_LIST_HEAD(&cfs_rq->tasks);
7528 #ifdef CONFIG_FAIR_GROUP_SCHED
7531 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7534 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7536 struct rt_prio_array *array;
7539 array = &rt_rq->active;
7540 for (i = 0; i < MAX_RT_PRIO; i++) {
7541 INIT_LIST_HEAD(array->queue + i);
7542 __clear_bit(i, array->bitmap);
7544 /* delimiter for bitsearch: */
7545 __set_bit(MAX_RT_PRIO, array->bitmap);
7547 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7548 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7550 rt_rq->highest_prio.next = MAX_RT_PRIO;
7554 rt_rq->rt_nr_migratory = 0;
7555 rt_rq->overloaded = 0;
7556 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7560 rt_rq->rt_throttled = 0;
7561 rt_rq->rt_runtime = 0;
7562 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7564 #ifdef CONFIG_RT_GROUP_SCHED
7565 rt_rq->rt_nr_boosted = 0;
7570 #ifdef CONFIG_FAIR_GROUP_SCHED
7571 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7572 struct sched_entity *se, int cpu, int add,
7573 struct sched_entity *parent)
7575 struct rq *rq = cpu_rq(cpu);
7576 tg->cfs_rq[cpu] = cfs_rq;
7577 init_cfs_rq(cfs_rq, rq);
7580 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7583 /* se could be NULL for init_task_group */
7588 se->cfs_rq = &rq->cfs;
7590 se->cfs_rq = parent->my_q;
7593 se->load.weight = tg->shares;
7594 se->load.inv_weight = 0;
7595 se->parent = parent;
7599 #ifdef CONFIG_RT_GROUP_SCHED
7600 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7601 struct sched_rt_entity *rt_se, int cpu, int add,
7602 struct sched_rt_entity *parent)
7604 struct rq *rq = cpu_rq(cpu);
7606 tg->rt_rq[cpu] = rt_rq;
7607 init_rt_rq(rt_rq, rq);
7609 rt_rq->rt_se = rt_se;
7610 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7612 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7614 tg->rt_se[cpu] = rt_se;
7619 rt_se->rt_rq = &rq->rt;
7621 rt_se->rt_rq = parent->my_q;
7623 rt_se->my_q = rt_rq;
7624 rt_se->parent = parent;
7625 INIT_LIST_HEAD(&rt_se->run_list);
7629 void __init sched_init(void)
7632 unsigned long alloc_size = 0, ptr;
7634 #ifdef CONFIG_FAIR_GROUP_SCHED
7635 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7637 #ifdef CONFIG_RT_GROUP_SCHED
7638 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7640 #ifdef CONFIG_CPUMASK_OFFSTACK
7641 alloc_size += num_possible_cpus() * cpumask_size();
7644 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7646 #ifdef CONFIG_FAIR_GROUP_SCHED
7647 init_task_group.se = (struct sched_entity **)ptr;
7648 ptr += nr_cpu_ids * sizeof(void **);
7650 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7651 ptr += nr_cpu_ids * sizeof(void **);
7653 #endif /* CONFIG_FAIR_GROUP_SCHED */
7654 #ifdef CONFIG_RT_GROUP_SCHED
7655 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7656 ptr += nr_cpu_ids * sizeof(void **);
7658 init_task_group.rt_rq = (struct rt_rq **)ptr;
7659 ptr += nr_cpu_ids * sizeof(void **);
7661 #endif /* CONFIG_RT_GROUP_SCHED */
7662 #ifdef CONFIG_CPUMASK_OFFSTACK
7663 for_each_possible_cpu(i) {
7664 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7665 ptr += cpumask_size();
7667 #endif /* CONFIG_CPUMASK_OFFSTACK */
7671 init_defrootdomain();
7674 init_rt_bandwidth(&def_rt_bandwidth,
7675 global_rt_period(), global_rt_runtime());
7677 #ifdef CONFIG_RT_GROUP_SCHED
7678 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7679 global_rt_period(), global_rt_runtime());
7680 #endif /* CONFIG_RT_GROUP_SCHED */
7682 #ifdef CONFIG_CGROUP_SCHED
7683 list_add(&init_task_group.list, &task_groups);
7684 INIT_LIST_HEAD(&init_task_group.children);
7686 #endif /* CONFIG_CGROUP_SCHED */
7688 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7689 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7690 __alignof__(unsigned long));
7692 for_each_possible_cpu(i) {
7696 raw_spin_lock_init(&rq->lock);
7698 rq->calc_load_active = 0;
7699 rq->calc_load_update = jiffies + LOAD_FREQ;
7700 init_cfs_rq(&rq->cfs, rq);
7701 init_rt_rq(&rq->rt, rq);
7702 #ifdef CONFIG_FAIR_GROUP_SCHED
7703 init_task_group.shares = init_task_group_load;
7704 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7705 #ifdef CONFIG_CGROUP_SCHED
7707 * How much cpu bandwidth does init_task_group get?
7709 * In case of task-groups formed thr' the cgroup filesystem, it
7710 * gets 100% of the cpu resources in the system. This overall
7711 * system cpu resource is divided among the tasks of
7712 * init_task_group and its child task-groups in a fair manner,
7713 * based on each entity's (task or task-group's) weight
7714 * (se->load.weight).
7716 * In other words, if init_task_group has 10 tasks of weight
7717 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7718 * then A0's share of the cpu resource is:
7720 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7722 * We achieve this by letting init_task_group's tasks sit
7723 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7725 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7727 #endif /* CONFIG_FAIR_GROUP_SCHED */
7729 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7730 #ifdef CONFIG_RT_GROUP_SCHED
7731 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7732 #ifdef CONFIG_CGROUP_SCHED
7733 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7737 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7738 rq->cpu_load[j] = 0;
7742 rq->post_schedule = 0;
7743 rq->active_balance = 0;
7744 rq->next_balance = jiffies;
7748 rq->migration_thread = NULL;
7750 rq->avg_idle = 2*sysctl_sched_migration_cost;
7751 INIT_LIST_HEAD(&rq->migration_queue);
7752 rq_attach_root(rq, &def_root_domain);
7755 atomic_set(&rq->nr_iowait, 0);
7758 set_load_weight(&init_task);
7760 #ifdef CONFIG_PREEMPT_NOTIFIERS
7761 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7765 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7768 #ifdef CONFIG_RT_MUTEXES
7769 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7773 * The boot idle thread does lazy MMU switching as well:
7775 atomic_inc(&init_mm.mm_count);
7776 enter_lazy_tlb(&init_mm, current);
7779 * Make us the idle thread. Technically, schedule() should not be
7780 * called from this thread, however somewhere below it might be,
7781 * but because we are the idle thread, we just pick up running again
7782 * when this runqueue becomes "idle".
7784 init_idle(current, smp_processor_id());
7786 calc_load_update = jiffies + LOAD_FREQ;
7789 * During early bootup we pretend to be a normal task:
7791 current->sched_class = &fair_sched_class;
7793 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7794 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7797 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7798 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7800 /* May be allocated at isolcpus cmdline parse time */
7801 if (cpu_isolated_map == NULL)
7802 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7807 scheduler_running = 1;
7810 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7811 static inline int preempt_count_equals(int preempt_offset)
7813 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7815 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7818 void __might_sleep(const char *file, int line, int preempt_offset)
7821 static unsigned long prev_jiffy; /* ratelimiting */
7823 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7824 system_state != SYSTEM_RUNNING || oops_in_progress)
7826 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7828 prev_jiffy = jiffies;
7831 "BUG: sleeping function called from invalid context at %s:%d\n",
7834 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7835 in_atomic(), irqs_disabled(),
7836 current->pid, current->comm);
7838 debug_show_held_locks(current);
7839 if (irqs_disabled())
7840 print_irqtrace_events(current);
7844 EXPORT_SYMBOL(__might_sleep);
7847 #ifdef CONFIG_MAGIC_SYSRQ
7848 static void normalize_task(struct rq *rq, struct task_struct *p)
7852 update_rq_clock(rq);
7853 on_rq = p->se.on_rq;
7855 deactivate_task(rq, p, 0);
7856 __setscheduler(rq, p, SCHED_NORMAL, 0);
7858 activate_task(rq, p, 0);
7859 resched_task(rq->curr);
7863 void normalize_rt_tasks(void)
7865 struct task_struct *g, *p;
7866 unsigned long flags;
7869 read_lock_irqsave(&tasklist_lock, flags);
7870 do_each_thread(g, p) {
7872 * Only normalize user tasks:
7877 p->se.exec_start = 0;
7878 #ifdef CONFIG_SCHEDSTATS
7879 p->se.wait_start = 0;
7880 p->se.sleep_start = 0;
7881 p->se.block_start = 0;
7886 * Renice negative nice level userspace
7889 if (TASK_NICE(p) < 0 && p->mm)
7890 set_user_nice(p, 0);
7894 raw_spin_lock(&p->pi_lock);
7895 rq = __task_rq_lock(p);
7897 normalize_task(rq, p);
7899 __task_rq_unlock(rq);
7900 raw_spin_unlock(&p->pi_lock);
7901 } while_each_thread(g, p);
7903 read_unlock_irqrestore(&tasklist_lock, flags);
7906 #endif /* CONFIG_MAGIC_SYSRQ */
7910 * These functions are only useful for the IA64 MCA handling.
7912 * They can only be called when the whole system has been
7913 * stopped - every CPU needs to be quiescent, and no scheduling
7914 * activity can take place. Using them for anything else would
7915 * be a serious bug, and as a result, they aren't even visible
7916 * under any other configuration.
7920 * curr_task - return the current task for a given cpu.
7921 * @cpu: the processor in question.
7923 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7925 struct task_struct *curr_task(int cpu)
7927 return cpu_curr(cpu);
7931 * set_curr_task - set the current task for a given cpu.
7932 * @cpu: the processor in question.
7933 * @p: the task pointer to set.
7935 * Description: This function must only be used when non-maskable interrupts
7936 * are serviced on a separate stack. It allows the architecture to switch the
7937 * notion of the current task on a cpu in a non-blocking manner. This function
7938 * must be called with all CPU's synchronized, and interrupts disabled, the
7939 * and caller must save the original value of the current task (see
7940 * curr_task() above) and restore that value before reenabling interrupts and
7941 * re-starting the system.
7943 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7945 void set_curr_task(int cpu, struct task_struct *p)
7952 #ifdef CONFIG_FAIR_GROUP_SCHED
7953 static void free_fair_sched_group(struct task_group *tg)
7957 for_each_possible_cpu(i) {
7959 kfree(tg->cfs_rq[i]);
7969 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7971 struct cfs_rq *cfs_rq;
7972 struct sched_entity *se;
7976 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7979 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7983 tg->shares = NICE_0_LOAD;
7985 for_each_possible_cpu(i) {
7988 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7989 GFP_KERNEL, cpu_to_node(i));
7993 se = kzalloc_node(sizeof(struct sched_entity),
7994 GFP_KERNEL, cpu_to_node(i));
7998 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8009 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8011 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8012 &cpu_rq(cpu)->leaf_cfs_rq_list);
8015 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8017 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8019 #else /* !CONFG_FAIR_GROUP_SCHED */
8020 static inline void free_fair_sched_group(struct task_group *tg)
8025 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8030 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8034 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8037 #endif /* CONFIG_FAIR_GROUP_SCHED */
8039 #ifdef CONFIG_RT_GROUP_SCHED
8040 static void free_rt_sched_group(struct task_group *tg)
8044 destroy_rt_bandwidth(&tg->rt_bandwidth);
8046 for_each_possible_cpu(i) {
8048 kfree(tg->rt_rq[i]);
8050 kfree(tg->rt_se[i]);
8058 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8060 struct rt_rq *rt_rq;
8061 struct sched_rt_entity *rt_se;
8065 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8068 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8072 init_rt_bandwidth(&tg->rt_bandwidth,
8073 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8075 for_each_possible_cpu(i) {
8078 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8079 GFP_KERNEL, cpu_to_node(i));
8083 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8084 GFP_KERNEL, cpu_to_node(i));
8088 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8099 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8101 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8102 &cpu_rq(cpu)->leaf_rt_rq_list);
8105 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8107 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8109 #else /* !CONFIG_RT_GROUP_SCHED */
8110 static inline void free_rt_sched_group(struct task_group *tg)
8115 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8120 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8124 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8127 #endif /* CONFIG_RT_GROUP_SCHED */
8129 #ifdef CONFIG_CGROUP_SCHED
8130 static void free_sched_group(struct task_group *tg)
8132 free_fair_sched_group(tg);
8133 free_rt_sched_group(tg);
8137 /* allocate runqueue etc for a new task group */
8138 struct task_group *sched_create_group(struct task_group *parent)
8140 struct task_group *tg;
8141 unsigned long flags;
8144 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8146 return ERR_PTR(-ENOMEM);
8148 if (!alloc_fair_sched_group(tg, parent))
8151 if (!alloc_rt_sched_group(tg, parent))
8154 spin_lock_irqsave(&task_group_lock, flags);
8155 for_each_possible_cpu(i) {
8156 register_fair_sched_group(tg, i);
8157 register_rt_sched_group(tg, i);
8159 list_add_rcu(&tg->list, &task_groups);
8161 WARN_ON(!parent); /* root should already exist */
8163 tg->parent = parent;
8164 INIT_LIST_HEAD(&tg->children);
8165 list_add_rcu(&tg->siblings, &parent->children);
8166 spin_unlock_irqrestore(&task_group_lock, flags);
8171 free_sched_group(tg);
8172 return ERR_PTR(-ENOMEM);
8175 /* rcu callback to free various structures associated with a task group */
8176 static void free_sched_group_rcu(struct rcu_head *rhp)
8178 /* now it should be safe to free those cfs_rqs */
8179 free_sched_group(container_of(rhp, struct task_group, rcu));
8182 /* Destroy runqueue etc associated with a task group */
8183 void sched_destroy_group(struct task_group *tg)
8185 unsigned long flags;
8188 spin_lock_irqsave(&task_group_lock, flags);
8189 for_each_possible_cpu(i) {
8190 unregister_fair_sched_group(tg, i);
8191 unregister_rt_sched_group(tg, i);
8193 list_del_rcu(&tg->list);
8194 list_del_rcu(&tg->siblings);
8195 spin_unlock_irqrestore(&task_group_lock, flags);
8197 /* wait for possible concurrent references to cfs_rqs complete */
8198 call_rcu(&tg->rcu, free_sched_group_rcu);
8201 /* change task's runqueue when it moves between groups.
8202 * The caller of this function should have put the task in its new group
8203 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8204 * reflect its new group.
8206 void sched_move_task(struct task_struct *tsk)
8209 unsigned long flags;
8212 rq = task_rq_lock(tsk, &flags);
8214 update_rq_clock(rq);
8216 running = task_current(rq, tsk);
8217 on_rq = tsk->se.on_rq;
8220 dequeue_task(rq, tsk, 0);
8221 if (unlikely(running))
8222 tsk->sched_class->put_prev_task(rq, tsk);
8224 set_task_rq(tsk, task_cpu(tsk));
8226 #ifdef CONFIG_FAIR_GROUP_SCHED
8227 if (tsk->sched_class->moved_group)
8228 tsk->sched_class->moved_group(tsk, on_rq);
8231 if (unlikely(running))
8232 tsk->sched_class->set_curr_task(rq);
8234 enqueue_task(rq, tsk, 0, false);
8236 task_rq_unlock(rq, &flags);
8238 #endif /* CONFIG_CGROUP_SCHED */
8240 #ifdef CONFIG_FAIR_GROUP_SCHED
8241 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8243 struct cfs_rq *cfs_rq = se->cfs_rq;
8248 dequeue_entity(cfs_rq, se, 0);
8250 se->load.weight = shares;
8251 se->load.inv_weight = 0;
8254 enqueue_entity(cfs_rq, se, 0);
8257 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8259 struct cfs_rq *cfs_rq = se->cfs_rq;
8260 struct rq *rq = cfs_rq->rq;
8261 unsigned long flags;
8263 raw_spin_lock_irqsave(&rq->lock, flags);
8264 __set_se_shares(se, shares);
8265 raw_spin_unlock_irqrestore(&rq->lock, flags);
8268 static DEFINE_MUTEX(shares_mutex);
8270 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8273 unsigned long flags;
8276 * We can't change the weight of the root cgroup.
8281 if (shares < MIN_SHARES)
8282 shares = MIN_SHARES;
8283 else if (shares > MAX_SHARES)
8284 shares = MAX_SHARES;
8286 mutex_lock(&shares_mutex);
8287 if (tg->shares == shares)
8290 spin_lock_irqsave(&task_group_lock, flags);
8291 for_each_possible_cpu(i)
8292 unregister_fair_sched_group(tg, i);
8293 list_del_rcu(&tg->siblings);
8294 spin_unlock_irqrestore(&task_group_lock, flags);
8296 /* wait for any ongoing reference to this group to finish */
8297 synchronize_sched();
8300 * Now we are free to modify the group's share on each cpu
8301 * w/o tripping rebalance_share or load_balance_fair.
8303 tg->shares = shares;
8304 for_each_possible_cpu(i) {
8308 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8309 set_se_shares(tg->se[i], shares);
8313 * Enable load balance activity on this group, by inserting it back on
8314 * each cpu's rq->leaf_cfs_rq_list.
8316 spin_lock_irqsave(&task_group_lock, flags);
8317 for_each_possible_cpu(i)
8318 register_fair_sched_group(tg, i);
8319 list_add_rcu(&tg->siblings, &tg->parent->children);
8320 spin_unlock_irqrestore(&task_group_lock, flags);
8322 mutex_unlock(&shares_mutex);
8326 unsigned long sched_group_shares(struct task_group *tg)
8332 #ifdef CONFIG_RT_GROUP_SCHED
8334 * Ensure that the real time constraints are schedulable.
8336 static DEFINE_MUTEX(rt_constraints_mutex);
8338 static unsigned long to_ratio(u64 period, u64 runtime)
8340 if (runtime == RUNTIME_INF)
8343 return div64_u64(runtime << 20, period);
8346 /* Must be called with tasklist_lock held */
8347 static inline int tg_has_rt_tasks(struct task_group *tg)
8349 struct task_struct *g, *p;
8351 do_each_thread(g, p) {
8352 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8354 } while_each_thread(g, p);
8359 struct rt_schedulable_data {
8360 struct task_group *tg;
8365 static int tg_schedulable(struct task_group *tg, void *data)
8367 struct rt_schedulable_data *d = data;
8368 struct task_group *child;
8369 unsigned long total, sum = 0;
8370 u64 period, runtime;
8372 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8373 runtime = tg->rt_bandwidth.rt_runtime;
8376 period = d->rt_period;
8377 runtime = d->rt_runtime;
8381 * Cannot have more runtime than the period.
8383 if (runtime > period && runtime != RUNTIME_INF)
8387 * Ensure we don't starve existing RT tasks.
8389 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8392 total = to_ratio(period, runtime);
8395 * Nobody can have more than the global setting allows.
8397 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8401 * The sum of our children's runtime should not exceed our own.
8403 list_for_each_entry_rcu(child, &tg->children, siblings) {
8404 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8405 runtime = child->rt_bandwidth.rt_runtime;
8407 if (child == d->tg) {
8408 period = d->rt_period;
8409 runtime = d->rt_runtime;
8412 sum += to_ratio(period, runtime);
8421 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8423 struct rt_schedulable_data data = {
8425 .rt_period = period,
8426 .rt_runtime = runtime,
8429 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8432 static int tg_set_bandwidth(struct task_group *tg,
8433 u64 rt_period, u64 rt_runtime)
8437 mutex_lock(&rt_constraints_mutex);
8438 read_lock(&tasklist_lock);
8439 err = __rt_schedulable(tg, rt_period, rt_runtime);
8443 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8444 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8445 tg->rt_bandwidth.rt_runtime = rt_runtime;
8447 for_each_possible_cpu(i) {
8448 struct rt_rq *rt_rq = tg->rt_rq[i];
8450 raw_spin_lock(&rt_rq->rt_runtime_lock);
8451 rt_rq->rt_runtime = rt_runtime;
8452 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8454 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8456 read_unlock(&tasklist_lock);
8457 mutex_unlock(&rt_constraints_mutex);
8462 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8464 u64 rt_runtime, rt_period;
8466 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8467 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8468 if (rt_runtime_us < 0)
8469 rt_runtime = RUNTIME_INF;
8471 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8474 long sched_group_rt_runtime(struct task_group *tg)
8478 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8481 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8482 do_div(rt_runtime_us, NSEC_PER_USEC);
8483 return rt_runtime_us;
8486 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8488 u64 rt_runtime, rt_period;
8490 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8491 rt_runtime = tg->rt_bandwidth.rt_runtime;
8496 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8499 long sched_group_rt_period(struct task_group *tg)
8503 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8504 do_div(rt_period_us, NSEC_PER_USEC);
8505 return rt_period_us;
8508 static int sched_rt_global_constraints(void)
8510 u64 runtime, period;
8513 if (sysctl_sched_rt_period <= 0)
8516 runtime = global_rt_runtime();
8517 period = global_rt_period();
8520 * Sanity check on the sysctl variables.
8522 if (runtime > period && runtime != RUNTIME_INF)
8525 mutex_lock(&rt_constraints_mutex);
8526 read_lock(&tasklist_lock);
8527 ret = __rt_schedulable(NULL, 0, 0);
8528 read_unlock(&tasklist_lock);
8529 mutex_unlock(&rt_constraints_mutex);
8534 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8536 /* Don't accept realtime tasks when there is no way for them to run */
8537 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8543 #else /* !CONFIG_RT_GROUP_SCHED */
8544 static int sched_rt_global_constraints(void)
8546 unsigned long flags;
8549 if (sysctl_sched_rt_period <= 0)
8553 * There's always some RT tasks in the root group
8554 * -- migration, kstopmachine etc..
8556 if (sysctl_sched_rt_runtime == 0)
8559 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8560 for_each_possible_cpu(i) {
8561 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8563 raw_spin_lock(&rt_rq->rt_runtime_lock);
8564 rt_rq->rt_runtime = global_rt_runtime();
8565 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8567 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8571 #endif /* CONFIG_RT_GROUP_SCHED */
8573 int sched_rt_handler(struct ctl_table *table, int write,
8574 void __user *buffer, size_t *lenp,
8578 int old_period, old_runtime;
8579 static DEFINE_MUTEX(mutex);
8582 old_period = sysctl_sched_rt_period;
8583 old_runtime = sysctl_sched_rt_runtime;
8585 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8587 if (!ret && write) {
8588 ret = sched_rt_global_constraints();
8590 sysctl_sched_rt_period = old_period;
8591 sysctl_sched_rt_runtime = old_runtime;
8593 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8594 def_rt_bandwidth.rt_period =
8595 ns_to_ktime(global_rt_period());
8598 mutex_unlock(&mutex);
8603 #ifdef CONFIG_CGROUP_SCHED
8605 /* return corresponding task_group object of a cgroup */
8606 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8608 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8609 struct task_group, css);
8612 static struct cgroup_subsys_state *
8613 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8615 struct task_group *tg, *parent;
8617 if (!cgrp->parent) {
8618 /* This is early initialization for the top cgroup */
8619 return &init_task_group.css;
8622 parent = cgroup_tg(cgrp->parent);
8623 tg = sched_create_group(parent);
8625 return ERR_PTR(-ENOMEM);
8631 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8633 struct task_group *tg = cgroup_tg(cgrp);
8635 sched_destroy_group(tg);
8639 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8641 #ifdef CONFIG_RT_GROUP_SCHED
8642 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8645 /* We don't support RT-tasks being in separate groups */
8646 if (tsk->sched_class != &fair_sched_class)
8653 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8654 struct task_struct *tsk, bool threadgroup)
8656 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8660 struct task_struct *c;
8662 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8663 retval = cpu_cgroup_can_attach_task(cgrp, c);
8675 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8676 struct cgroup *old_cont, struct task_struct *tsk,
8679 sched_move_task(tsk);
8681 struct task_struct *c;
8683 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8690 #ifdef CONFIG_FAIR_GROUP_SCHED
8691 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8694 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8697 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8699 struct task_group *tg = cgroup_tg(cgrp);
8701 return (u64) tg->shares;
8703 #endif /* CONFIG_FAIR_GROUP_SCHED */
8705 #ifdef CONFIG_RT_GROUP_SCHED
8706 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8709 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8712 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8714 return sched_group_rt_runtime(cgroup_tg(cgrp));
8717 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8720 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8723 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8725 return sched_group_rt_period(cgroup_tg(cgrp));
8727 #endif /* CONFIG_RT_GROUP_SCHED */
8729 static struct cftype cpu_files[] = {
8730 #ifdef CONFIG_FAIR_GROUP_SCHED
8733 .read_u64 = cpu_shares_read_u64,
8734 .write_u64 = cpu_shares_write_u64,
8737 #ifdef CONFIG_RT_GROUP_SCHED
8739 .name = "rt_runtime_us",
8740 .read_s64 = cpu_rt_runtime_read,
8741 .write_s64 = cpu_rt_runtime_write,
8744 .name = "rt_period_us",
8745 .read_u64 = cpu_rt_period_read_uint,
8746 .write_u64 = cpu_rt_period_write_uint,
8751 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8753 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8756 struct cgroup_subsys cpu_cgroup_subsys = {
8758 .create = cpu_cgroup_create,
8759 .destroy = cpu_cgroup_destroy,
8760 .can_attach = cpu_cgroup_can_attach,
8761 .attach = cpu_cgroup_attach,
8762 .populate = cpu_cgroup_populate,
8763 .subsys_id = cpu_cgroup_subsys_id,
8767 #endif /* CONFIG_CGROUP_SCHED */
8769 #ifdef CONFIG_CGROUP_CPUACCT
8772 * CPU accounting code for task groups.
8774 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8775 * (balbir@in.ibm.com).
8778 /* track cpu usage of a group of tasks and its child groups */
8780 struct cgroup_subsys_state css;
8781 /* cpuusage holds pointer to a u64-type object on every cpu */
8783 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8784 struct cpuacct *parent;
8787 struct cgroup_subsys cpuacct_subsys;
8789 /* return cpu accounting group corresponding to this container */
8790 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8792 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8793 struct cpuacct, css);
8796 /* return cpu accounting group to which this task belongs */
8797 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8799 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8800 struct cpuacct, css);
8803 /* create a new cpu accounting group */
8804 static struct cgroup_subsys_state *cpuacct_create(
8805 struct cgroup_subsys *ss, struct cgroup *cgrp)
8807 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8813 ca->cpuusage = alloc_percpu(u64);
8817 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8818 if (percpu_counter_init(&ca->cpustat[i], 0))
8819 goto out_free_counters;
8822 ca->parent = cgroup_ca(cgrp->parent);
8828 percpu_counter_destroy(&ca->cpustat[i]);
8829 free_percpu(ca->cpuusage);
8833 return ERR_PTR(-ENOMEM);
8836 /* destroy an existing cpu accounting group */
8838 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8840 struct cpuacct *ca = cgroup_ca(cgrp);
8843 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8844 percpu_counter_destroy(&ca->cpustat[i]);
8845 free_percpu(ca->cpuusage);
8849 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8851 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8854 #ifndef CONFIG_64BIT
8856 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8858 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8860 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8868 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8870 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8872 #ifndef CONFIG_64BIT
8874 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8876 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8878 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8884 /* return total cpu usage (in nanoseconds) of a group */
8885 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8887 struct cpuacct *ca = cgroup_ca(cgrp);
8888 u64 totalcpuusage = 0;
8891 for_each_present_cpu(i)
8892 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8894 return totalcpuusage;
8897 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8900 struct cpuacct *ca = cgroup_ca(cgrp);
8909 for_each_present_cpu(i)
8910 cpuacct_cpuusage_write(ca, i, 0);
8916 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8919 struct cpuacct *ca = cgroup_ca(cgroup);
8923 for_each_present_cpu(i) {
8924 percpu = cpuacct_cpuusage_read(ca, i);
8925 seq_printf(m, "%llu ", (unsigned long long) percpu);
8927 seq_printf(m, "\n");
8931 static const char *cpuacct_stat_desc[] = {
8932 [CPUACCT_STAT_USER] = "user",
8933 [CPUACCT_STAT_SYSTEM] = "system",
8936 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8937 struct cgroup_map_cb *cb)
8939 struct cpuacct *ca = cgroup_ca(cgrp);
8942 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8943 s64 val = percpu_counter_read(&ca->cpustat[i]);
8944 val = cputime64_to_clock_t(val);
8945 cb->fill(cb, cpuacct_stat_desc[i], val);
8950 static struct cftype files[] = {
8953 .read_u64 = cpuusage_read,
8954 .write_u64 = cpuusage_write,
8957 .name = "usage_percpu",
8958 .read_seq_string = cpuacct_percpu_seq_read,
8962 .read_map = cpuacct_stats_show,
8966 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8968 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8972 * charge this task's execution time to its accounting group.
8974 * called with rq->lock held.
8976 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8981 if (unlikely(!cpuacct_subsys.active))
8984 cpu = task_cpu(tsk);
8990 for (; ca; ca = ca->parent) {
8991 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8992 *cpuusage += cputime;
8999 * Charge the system/user time to the task's accounting group.
9001 static void cpuacct_update_stats(struct task_struct *tsk,
9002 enum cpuacct_stat_index idx, cputime_t val)
9006 if (unlikely(!cpuacct_subsys.active))
9013 percpu_counter_add(&ca->cpustat[idx], val);
9019 struct cgroup_subsys cpuacct_subsys = {
9021 .create = cpuacct_create,
9022 .destroy = cpuacct_destroy,
9023 .populate = cpuacct_populate,
9024 .subsys_id = cpuacct_subsys_id,
9026 #endif /* CONFIG_CGROUP_CPUACCT */
9030 int rcu_expedited_torture_stats(char *page)
9034 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9036 void synchronize_sched_expedited(void)
9039 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9041 #else /* #ifndef CONFIG_SMP */
9043 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9044 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9046 #define RCU_EXPEDITED_STATE_POST -2
9047 #define RCU_EXPEDITED_STATE_IDLE -1
9049 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9051 int rcu_expedited_torture_stats(char *page)
9056 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9057 for_each_online_cpu(cpu) {
9058 cnt += sprintf(&page[cnt], " %d:%d",
9059 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9061 cnt += sprintf(&page[cnt], "\n");
9064 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9066 static long synchronize_sched_expedited_count;
9069 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9070 * approach to force grace period to end quickly. This consumes
9071 * significant time on all CPUs, and is thus not recommended for
9072 * any sort of common-case code.
9074 * Note that it is illegal to call this function while holding any
9075 * lock that is acquired by a CPU-hotplug notifier. Failing to
9076 * observe this restriction will result in deadlock.
9078 void synchronize_sched_expedited(void)
9081 unsigned long flags;
9082 bool need_full_sync = 0;
9084 struct migration_req *req;
9088 smp_mb(); /* ensure prior mod happens before capturing snap. */
9089 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9091 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9093 if (trycount++ < 10)
9094 udelay(trycount * num_online_cpus());
9096 synchronize_sched();
9099 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9100 smp_mb(); /* ensure test happens before caller kfree */
9105 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9106 for_each_online_cpu(cpu) {
9108 req = &per_cpu(rcu_migration_req, cpu);
9109 init_completion(&req->done);
9111 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9112 raw_spin_lock_irqsave(&rq->lock, flags);
9113 list_add(&req->list, &rq->migration_queue);
9114 raw_spin_unlock_irqrestore(&rq->lock, flags);
9115 wake_up_process(rq->migration_thread);
9117 for_each_online_cpu(cpu) {
9118 rcu_expedited_state = cpu;
9119 req = &per_cpu(rcu_migration_req, cpu);
9121 wait_for_completion(&req->done);
9122 raw_spin_lock_irqsave(&rq->lock, flags);
9123 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9125 req->dest_cpu = RCU_MIGRATION_IDLE;
9126 raw_spin_unlock_irqrestore(&rq->lock, flags);
9128 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9129 synchronize_sched_expedited_count++;
9130 mutex_unlock(&rcu_sched_expedited_mutex);
9133 synchronize_sched();
9135 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9137 #endif /* #else #ifndef CONFIG_SMP */