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_h_load(long cpu)
1638 if (root_task_group_empty())
1641 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1646 static inline void update_shares(struct sched_domain *sd)
1652 #ifdef CONFIG_PREEMPT
1654 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1657 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1658 * way at the expense of forcing extra atomic operations in all
1659 * invocations. This assures that the double_lock is acquired using the
1660 * same underlying policy as the spinlock_t on this architecture, which
1661 * reduces latency compared to the unfair variant below. However, it
1662 * also adds more overhead and therefore may reduce throughput.
1664 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1665 __releases(this_rq->lock)
1666 __acquires(busiest->lock)
1667 __acquires(this_rq->lock)
1669 raw_spin_unlock(&this_rq->lock);
1670 double_rq_lock(this_rq, busiest);
1677 * Unfair double_lock_balance: Optimizes throughput at the expense of
1678 * latency by eliminating extra atomic operations when the locks are
1679 * already in proper order on entry. This favors lower cpu-ids and will
1680 * grant the double lock to lower cpus over higher ids under contention,
1681 * regardless of entry order into the function.
1683 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1684 __releases(this_rq->lock)
1685 __acquires(busiest->lock)
1686 __acquires(this_rq->lock)
1690 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1691 if (busiest < this_rq) {
1692 raw_spin_unlock(&this_rq->lock);
1693 raw_spin_lock(&busiest->lock);
1694 raw_spin_lock_nested(&this_rq->lock,
1695 SINGLE_DEPTH_NESTING);
1698 raw_spin_lock_nested(&busiest->lock,
1699 SINGLE_DEPTH_NESTING);
1704 #endif /* CONFIG_PREEMPT */
1707 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1709 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1711 if (unlikely(!irqs_disabled())) {
1712 /* printk() doesn't work good under rq->lock */
1713 raw_spin_unlock(&this_rq->lock);
1717 return _double_lock_balance(this_rq, busiest);
1720 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1721 __releases(busiest->lock)
1723 raw_spin_unlock(&busiest->lock);
1724 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1728 * double_rq_lock - safely lock two runqueues
1730 * Note this does not disable interrupts like task_rq_lock,
1731 * you need to do so manually before calling.
1733 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1734 __acquires(rq1->lock)
1735 __acquires(rq2->lock)
1737 BUG_ON(!irqs_disabled());
1739 raw_spin_lock(&rq1->lock);
1740 __acquire(rq2->lock); /* Fake it out ;) */
1743 raw_spin_lock(&rq1->lock);
1744 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1746 raw_spin_lock(&rq2->lock);
1747 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1750 update_rq_clock(rq1);
1751 update_rq_clock(rq2);
1755 * double_rq_unlock - safely unlock two runqueues
1757 * Note this does not restore interrupts like task_rq_unlock,
1758 * you need to do so manually after calling.
1760 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1761 __releases(rq1->lock)
1762 __releases(rq2->lock)
1764 raw_spin_unlock(&rq1->lock);
1766 raw_spin_unlock(&rq2->lock);
1768 __release(rq2->lock);
1773 #ifdef CONFIG_FAIR_GROUP_SCHED
1774 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1777 cfs_rq->shares = shares;
1782 static void calc_load_account_active(struct rq *this_rq);
1783 static void update_sysctl(void);
1784 static int get_update_sysctl_factor(void);
1786 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1788 set_task_rq(p, cpu);
1791 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1792 * successfuly executed on another CPU. We must ensure that updates of
1793 * per-task data have been completed by this moment.
1796 task_thread_info(p)->cpu = cpu;
1800 static const struct sched_class rt_sched_class;
1802 #define sched_class_highest (&rt_sched_class)
1803 #define for_each_class(class) \
1804 for (class = sched_class_highest; class; class = class->next)
1806 #include "sched_stats.h"
1808 static void inc_nr_running(struct rq *rq)
1813 static void dec_nr_running(struct rq *rq)
1818 static void set_load_weight(struct task_struct *p)
1820 if (task_has_rt_policy(p)) {
1821 p->se.load.weight = prio_to_weight[0] * 2;
1822 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1827 * SCHED_IDLE tasks get minimal weight:
1829 if (p->policy == SCHED_IDLE) {
1830 p->se.load.weight = WEIGHT_IDLEPRIO;
1831 p->se.load.inv_weight = WMULT_IDLEPRIO;
1835 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1836 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1839 static void update_avg(u64 *avg, u64 sample)
1841 s64 diff = sample - *avg;
1846 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1849 p->se.start_runtime = p->se.sum_exec_runtime;
1851 sched_info_queued(p);
1852 p->sched_class->enqueue_task(rq, p, wakeup, head);
1856 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1859 if (p->se.last_wakeup) {
1860 update_avg(&p->se.avg_overlap,
1861 p->se.sum_exec_runtime - p->se.last_wakeup);
1862 p->se.last_wakeup = 0;
1864 update_avg(&p->se.avg_wakeup,
1865 sysctl_sched_wakeup_granularity);
1869 sched_info_dequeued(p);
1870 p->sched_class->dequeue_task(rq, p, sleep);
1875 * activate_task - move a task to the runqueue.
1877 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1879 if (task_contributes_to_load(p))
1880 rq->nr_uninterruptible--;
1882 enqueue_task(rq, p, wakeup, false);
1887 * deactivate_task - remove a task from the runqueue.
1889 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1891 if (task_contributes_to_load(p))
1892 rq->nr_uninterruptible++;
1894 dequeue_task(rq, p, sleep);
1898 #include "sched_idletask.c"
1899 #include "sched_fair.c"
1900 #include "sched_rt.c"
1901 #ifdef CONFIG_SCHED_DEBUG
1902 # include "sched_debug.c"
1906 * __normal_prio - return the priority that is based on the static prio
1908 static inline int __normal_prio(struct task_struct *p)
1910 return p->static_prio;
1914 * Calculate the expected normal priority: i.e. priority
1915 * without taking RT-inheritance into account. Might be
1916 * boosted by interactivity modifiers. Changes upon fork,
1917 * setprio syscalls, and whenever the interactivity
1918 * estimator recalculates.
1920 static inline int normal_prio(struct task_struct *p)
1924 if (task_has_rt_policy(p))
1925 prio = MAX_RT_PRIO-1 - p->rt_priority;
1927 prio = __normal_prio(p);
1932 * Calculate the current priority, i.e. the priority
1933 * taken into account by the scheduler. This value might
1934 * be boosted by RT tasks, or might be boosted by
1935 * interactivity modifiers. Will be RT if the task got
1936 * RT-boosted. If not then it returns p->normal_prio.
1938 static int effective_prio(struct task_struct *p)
1940 p->normal_prio = normal_prio(p);
1942 * If we are RT tasks or we were boosted to RT priority,
1943 * keep the priority unchanged. Otherwise, update priority
1944 * to the normal priority:
1946 if (!rt_prio(p->prio))
1947 return p->normal_prio;
1952 * task_curr - is this task currently executing on a CPU?
1953 * @p: the task in question.
1955 inline int task_curr(const struct task_struct *p)
1957 return cpu_curr(task_cpu(p)) == p;
1960 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1961 const struct sched_class *prev_class,
1962 int oldprio, int running)
1964 if (prev_class != p->sched_class) {
1965 if (prev_class->switched_from)
1966 prev_class->switched_from(rq, p, running);
1967 p->sched_class->switched_to(rq, p, running);
1969 p->sched_class->prio_changed(rq, p, oldprio, running);
1974 * Is this task likely cache-hot:
1977 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1981 if (p->sched_class != &fair_sched_class)
1985 * Buddy candidates are cache hot:
1987 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
1988 (&p->se == cfs_rq_of(&p->se)->next ||
1989 &p->se == cfs_rq_of(&p->se)->last))
1992 if (sysctl_sched_migration_cost == -1)
1994 if (sysctl_sched_migration_cost == 0)
1997 delta = now - p->se.exec_start;
1999 return delta < (s64)sysctl_sched_migration_cost;
2002 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2004 #ifdef CONFIG_SCHED_DEBUG
2006 * We should never call set_task_cpu() on a blocked task,
2007 * ttwu() will sort out the placement.
2009 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2010 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2013 trace_sched_migrate_task(p, new_cpu);
2015 if (task_cpu(p) != new_cpu) {
2016 p->se.nr_migrations++;
2017 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2020 __set_task_cpu(p, new_cpu);
2023 struct migration_req {
2024 struct list_head list;
2026 struct task_struct *task;
2029 struct completion done;
2033 * The task's runqueue lock must be held.
2034 * Returns true if you have to wait for migration thread.
2037 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2039 struct rq *rq = task_rq(p);
2042 * If the task is not on a runqueue (and not running), then
2043 * the next wake-up will properly place the task.
2045 if (!p->se.on_rq && !task_running(rq, p))
2048 init_completion(&req->done);
2050 req->dest_cpu = dest_cpu;
2051 list_add(&req->list, &rq->migration_queue);
2057 * wait_task_context_switch - wait for a thread to complete at least one
2060 * @p must not be current.
2062 void wait_task_context_switch(struct task_struct *p)
2064 unsigned long nvcsw, nivcsw, flags;
2072 * The runqueue is assigned before the actual context
2073 * switch. We need to take the runqueue lock.
2075 * We could check initially without the lock but it is
2076 * very likely that we need to take the lock in every
2079 rq = task_rq_lock(p, &flags);
2080 running = task_running(rq, p);
2081 task_rq_unlock(rq, &flags);
2083 if (likely(!running))
2086 * The switch count is incremented before the actual
2087 * context switch. We thus wait for two switches to be
2088 * sure at least one completed.
2090 if ((p->nvcsw - nvcsw) > 1)
2092 if ((p->nivcsw - nivcsw) > 1)
2100 * wait_task_inactive - wait for a thread to unschedule.
2102 * If @match_state is nonzero, it's the @p->state value just checked and
2103 * not expected to change. If it changes, i.e. @p might have woken up,
2104 * then return zero. When we succeed in waiting for @p to be off its CPU,
2105 * we return a positive number (its total switch count). If a second call
2106 * a short while later returns the same number, the caller can be sure that
2107 * @p has remained unscheduled the whole time.
2109 * The caller must ensure that the task *will* unschedule sometime soon,
2110 * else this function might spin for a *long* time. This function can't
2111 * be called with interrupts off, or it may introduce deadlock with
2112 * smp_call_function() if an IPI is sent by the same process we are
2113 * waiting to become inactive.
2115 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2117 unsigned long flags;
2124 * We do the initial early heuristics without holding
2125 * any task-queue locks at all. We'll only try to get
2126 * the runqueue lock when things look like they will
2132 * If the task is actively running on another CPU
2133 * still, just relax and busy-wait without holding
2136 * NOTE! Since we don't hold any locks, it's not
2137 * even sure that "rq" stays as the right runqueue!
2138 * But we don't care, since "task_running()" will
2139 * return false if the runqueue has changed and p
2140 * is actually now running somewhere else!
2142 while (task_running(rq, p)) {
2143 if (match_state && unlikely(p->state != match_state))
2149 * Ok, time to look more closely! We need the rq
2150 * lock now, to be *sure*. If we're wrong, we'll
2151 * just go back and repeat.
2153 rq = task_rq_lock(p, &flags);
2154 trace_sched_wait_task(rq, p);
2155 running = task_running(rq, p);
2156 on_rq = p->se.on_rq;
2158 if (!match_state || p->state == match_state)
2159 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2160 task_rq_unlock(rq, &flags);
2163 * If it changed from the expected state, bail out now.
2165 if (unlikely(!ncsw))
2169 * Was it really running after all now that we
2170 * checked with the proper locks actually held?
2172 * Oops. Go back and try again..
2174 if (unlikely(running)) {
2180 * It's not enough that it's not actively running,
2181 * it must be off the runqueue _entirely_, and not
2184 * So if it was still runnable (but just not actively
2185 * running right now), it's preempted, and we should
2186 * yield - it could be a while.
2188 if (unlikely(on_rq)) {
2189 schedule_timeout_uninterruptible(1);
2194 * Ahh, all good. It wasn't running, and it wasn't
2195 * runnable, which means that it will never become
2196 * running in the future either. We're all done!
2205 * kick_process - kick a running thread to enter/exit the kernel
2206 * @p: the to-be-kicked thread
2208 * Cause a process which is running on another CPU to enter
2209 * kernel-mode, without any delay. (to get signals handled.)
2211 * NOTE: this function doesnt have to take the runqueue lock,
2212 * because all it wants to ensure is that the remote task enters
2213 * the kernel. If the IPI races and the task has been migrated
2214 * to another CPU then no harm is done and the purpose has been
2217 void kick_process(struct task_struct *p)
2223 if ((cpu != smp_processor_id()) && task_curr(p))
2224 smp_send_reschedule(cpu);
2227 EXPORT_SYMBOL_GPL(kick_process);
2228 #endif /* CONFIG_SMP */
2231 * task_oncpu_function_call - call a function on the cpu on which a task runs
2232 * @p: the task to evaluate
2233 * @func: the function to be called
2234 * @info: the function call argument
2236 * Calls the function @func when the task is currently running. This might
2237 * be on the current CPU, which just calls the function directly
2239 void task_oncpu_function_call(struct task_struct *p,
2240 void (*func) (void *info), void *info)
2247 smp_call_function_single(cpu, func, info, 1);
2252 static int select_fallback_rq(int cpu, struct task_struct *p)
2255 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2257 /* Look for allowed, online CPU in same node. */
2258 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2259 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2262 /* Any allowed, online CPU? */
2263 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2264 if (dest_cpu < nr_cpu_ids)
2267 /* No more Mr. Nice Guy. */
2268 if (dest_cpu >= nr_cpu_ids) {
2270 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2272 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2275 * Don't tell them about moving exiting tasks or
2276 * kernel threads (both mm NULL), since they never
2279 if (p->mm && printk_ratelimit()) {
2280 printk(KERN_INFO "process %d (%s) no "
2281 "longer affine to cpu%d\n",
2282 task_pid_nr(p), p->comm, cpu);
2292 * - fork, @p is stable because it isn't on the tasklist yet
2294 * - exec, @p is unstable, retry loop
2296 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so
2297 * we should be good.
2300 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2302 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2305 * In order not to call set_task_cpu() on a blocking task we need
2306 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2309 * Since this is common to all placement strategies, this lives here.
2311 * [ this allows ->select_task() to simply return task_cpu(p) and
2312 * not worry about this generic constraint ]
2314 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2316 cpu = select_fallback_rq(task_cpu(p), p);
2323 * try_to_wake_up - wake up a thread
2324 * @p: the to-be-woken-up thread
2325 * @state: the mask of task states that can be woken
2326 * @sync: do a synchronous wakeup?
2328 * Put it on the run-queue if it's not already there. The "current"
2329 * thread is always on the run-queue (except when the actual
2330 * re-schedule is in progress), and as such you're allowed to do
2331 * the simpler "current->state = TASK_RUNNING" to mark yourself
2332 * runnable without the overhead of this.
2334 * returns failure only if the task is already active.
2336 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2339 int cpu, orig_cpu, this_cpu, success = 0;
2340 unsigned long flags;
2341 struct rq *rq, *orig_rq;
2343 if (!sched_feat(SYNC_WAKEUPS))
2344 wake_flags &= ~WF_SYNC;
2346 this_cpu = get_cpu();
2349 rq = orig_rq = task_rq_lock(p, &flags);
2350 update_rq_clock(rq);
2351 if (!(p->state & state))
2361 if (unlikely(task_running(rq, p)))
2365 * In order to handle concurrent wakeups and release the rq->lock
2366 * we put the task in TASK_WAKING state.
2368 * First fix up the nr_uninterruptible count:
2370 if (task_contributes_to_load(p))
2371 rq->nr_uninterruptible--;
2372 p->state = TASK_WAKING;
2374 if (p->sched_class->task_waking)
2375 p->sched_class->task_waking(rq, p);
2377 __task_rq_unlock(rq);
2379 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2380 if (cpu != orig_cpu)
2381 set_task_cpu(p, cpu);
2383 rq = __task_rq_lock(p);
2384 update_rq_clock(rq);
2386 WARN_ON(p->state != TASK_WAKING);
2389 #ifdef CONFIG_SCHEDSTATS
2390 schedstat_inc(rq, ttwu_count);
2391 if (cpu == this_cpu)
2392 schedstat_inc(rq, ttwu_local);
2394 struct sched_domain *sd;
2395 for_each_domain(this_cpu, sd) {
2396 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2397 schedstat_inc(sd, ttwu_wake_remote);
2402 #endif /* CONFIG_SCHEDSTATS */
2405 #endif /* CONFIG_SMP */
2406 schedstat_inc(p, se.nr_wakeups);
2407 if (wake_flags & WF_SYNC)
2408 schedstat_inc(p, se.nr_wakeups_sync);
2409 if (orig_cpu != cpu)
2410 schedstat_inc(p, se.nr_wakeups_migrate);
2411 if (cpu == this_cpu)
2412 schedstat_inc(p, se.nr_wakeups_local);
2414 schedstat_inc(p, se.nr_wakeups_remote);
2415 activate_task(rq, p, 1);
2419 * Only attribute actual wakeups done by this task.
2421 if (!in_interrupt()) {
2422 struct sched_entity *se = ¤t->se;
2423 u64 sample = se->sum_exec_runtime;
2425 if (se->last_wakeup)
2426 sample -= se->last_wakeup;
2428 sample -= se->start_runtime;
2429 update_avg(&se->avg_wakeup, sample);
2431 se->last_wakeup = se->sum_exec_runtime;
2435 trace_sched_wakeup(rq, p, success);
2436 check_preempt_curr(rq, p, wake_flags);
2438 p->state = TASK_RUNNING;
2440 if (p->sched_class->task_woken)
2441 p->sched_class->task_woken(rq, p);
2443 if (unlikely(rq->idle_stamp)) {
2444 u64 delta = rq->clock - rq->idle_stamp;
2445 u64 max = 2*sysctl_sched_migration_cost;
2450 update_avg(&rq->avg_idle, delta);
2455 task_rq_unlock(rq, &flags);
2462 * wake_up_process - Wake up a specific process
2463 * @p: The process to be woken up.
2465 * Attempt to wake up the nominated process and move it to the set of runnable
2466 * processes. Returns 1 if the process was woken up, 0 if it was already
2469 * It may be assumed that this function implies a write memory barrier before
2470 * changing the task state if and only if any tasks are woken up.
2472 int wake_up_process(struct task_struct *p)
2474 return try_to_wake_up(p, TASK_ALL, 0);
2476 EXPORT_SYMBOL(wake_up_process);
2478 int wake_up_state(struct task_struct *p, unsigned int state)
2480 return try_to_wake_up(p, state, 0);
2484 * Perform scheduler related setup for a newly forked process p.
2485 * p is forked by current.
2487 * __sched_fork() is basic setup used by init_idle() too:
2489 static void __sched_fork(struct task_struct *p)
2491 p->se.exec_start = 0;
2492 p->se.sum_exec_runtime = 0;
2493 p->se.prev_sum_exec_runtime = 0;
2494 p->se.nr_migrations = 0;
2495 p->se.last_wakeup = 0;
2496 p->se.avg_overlap = 0;
2497 p->se.start_runtime = 0;
2498 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2500 #ifdef CONFIG_SCHEDSTATS
2501 p->se.wait_start = 0;
2503 p->se.wait_count = 0;
2506 p->se.sleep_start = 0;
2507 p->se.sleep_max = 0;
2508 p->se.sum_sleep_runtime = 0;
2510 p->se.block_start = 0;
2511 p->se.block_max = 0;
2513 p->se.slice_max = 0;
2515 p->se.nr_migrations_cold = 0;
2516 p->se.nr_failed_migrations_affine = 0;
2517 p->se.nr_failed_migrations_running = 0;
2518 p->se.nr_failed_migrations_hot = 0;
2519 p->se.nr_forced_migrations = 0;
2521 p->se.nr_wakeups = 0;
2522 p->se.nr_wakeups_sync = 0;
2523 p->se.nr_wakeups_migrate = 0;
2524 p->se.nr_wakeups_local = 0;
2525 p->se.nr_wakeups_remote = 0;
2526 p->se.nr_wakeups_affine = 0;
2527 p->se.nr_wakeups_affine_attempts = 0;
2528 p->se.nr_wakeups_passive = 0;
2529 p->se.nr_wakeups_idle = 0;
2533 INIT_LIST_HEAD(&p->rt.run_list);
2535 INIT_LIST_HEAD(&p->se.group_node);
2537 #ifdef CONFIG_PREEMPT_NOTIFIERS
2538 INIT_HLIST_HEAD(&p->preempt_notifiers);
2543 * fork()/clone()-time setup:
2545 void sched_fork(struct task_struct *p, int clone_flags)
2547 int cpu = get_cpu();
2551 * We mark the process as waking here. This guarantees that
2552 * nobody will actually run it, and a signal or other external
2553 * event cannot wake it up and insert it on the runqueue either.
2555 p->state = TASK_WAKING;
2558 * Revert to default priority/policy on fork if requested.
2560 if (unlikely(p->sched_reset_on_fork)) {
2561 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2562 p->policy = SCHED_NORMAL;
2563 p->normal_prio = p->static_prio;
2566 if (PRIO_TO_NICE(p->static_prio) < 0) {
2567 p->static_prio = NICE_TO_PRIO(0);
2568 p->normal_prio = p->static_prio;
2573 * We don't need the reset flag anymore after the fork. It has
2574 * fulfilled its duty:
2576 p->sched_reset_on_fork = 0;
2580 * Make sure we do not leak PI boosting priority to the child.
2582 p->prio = current->normal_prio;
2584 if (!rt_prio(p->prio))
2585 p->sched_class = &fair_sched_class;
2587 if (p->sched_class->task_fork)
2588 p->sched_class->task_fork(p);
2591 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2593 set_task_cpu(p, cpu);
2595 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2596 if (likely(sched_info_on()))
2597 memset(&p->sched_info, 0, sizeof(p->sched_info));
2599 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2602 #ifdef CONFIG_PREEMPT
2603 /* Want to start with kernel preemption disabled. */
2604 task_thread_info(p)->preempt_count = 1;
2606 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2612 * wake_up_new_task - wake up a newly created task for the first time.
2614 * This function will do some initial scheduler statistics housekeeping
2615 * that must be done for every newly created context, then puts the task
2616 * on the runqueue and wakes it.
2618 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2620 unsigned long flags;
2623 rq = task_rq_lock(p, &flags);
2624 BUG_ON(p->state != TASK_WAKING);
2625 p->state = TASK_RUNNING;
2626 update_rq_clock(rq);
2627 activate_task(rq, p, 0);
2628 trace_sched_wakeup_new(rq, p, 1);
2629 check_preempt_curr(rq, p, WF_FORK);
2631 if (p->sched_class->task_woken)
2632 p->sched_class->task_woken(rq, p);
2634 task_rq_unlock(rq, &flags);
2637 #ifdef CONFIG_PREEMPT_NOTIFIERS
2640 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2641 * @notifier: notifier struct to register
2643 void preempt_notifier_register(struct preempt_notifier *notifier)
2645 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2647 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2650 * preempt_notifier_unregister - no longer interested in preemption notifications
2651 * @notifier: notifier struct to unregister
2653 * This is safe to call from within a preemption notifier.
2655 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2657 hlist_del(¬ifier->link);
2659 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2661 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2663 struct preempt_notifier *notifier;
2664 struct hlist_node *node;
2666 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2667 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2671 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2672 struct task_struct *next)
2674 struct preempt_notifier *notifier;
2675 struct hlist_node *node;
2677 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2678 notifier->ops->sched_out(notifier, next);
2681 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2683 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2688 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2689 struct task_struct *next)
2693 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2696 * prepare_task_switch - prepare to switch tasks
2697 * @rq: the runqueue preparing to switch
2698 * @prev: the current task that is being switched out
2699 * @next: the task we are going to switch to.
2701 * This is called with the rq lock held and interrupts off. It must
2702 * be paired with a subsequent finish_task_switch after the context
2705 * prepare_task_switch sets up locking and calls architecture specific
2709 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2710 struct task_struct *next)
2712 fire_sched_out_preempt_notifiers(prev, next);
2713 prepare_lock_switch(rq, next);
2714 prepare_arch_switch(next);
2718 * finish_task_switch - clean up after a task-switch
2719 * @rq: runqueue associated with task-switch
2720 * @prev: the thread we just switched away from.
2722 * finish_task_switch must be called after the context switch, paired
2723 * with a prepare_task_switch call before the context switch.
2724 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2725 * and do any other architecture-specific cleanup actions.
2727 * Note that we may have delayed dropping an mm in context_switch(). If
2728 * so, we finish that here outside of the runqueue lock. (Doing it
2729 * with the lock held can cause deadlocks; see schedule() for
2732 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2733 __releases(rq->lock)
2735 struct mm_struct *mm = rq->prev_mm;
2741 * A task struct has one reference for the use as "current".
2742 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2743 * schedule one last time. The schedule call will never return, and
2744 * the scheduled task must drop that reference.
2745 * The test for TASK_DEAD must occur while the runqueue locks are
2746 * still held, otherwise prev could be scheduled on another cpu, die
2747 * there before we look at prev->state, and then the reference would
2749 * Manfred Spraul <manfred@colorfullife.com>
2751 prev_state = prev->state;
2752 finish_arch_switch(prev);
2753 perf_event_task_sched_in(current, cpu_of(rq));
2754 finish_lock_switch(rq, prev);
2756 fire_sched_in_preempt_notifiers(current);
2759 if (unlikely(prev_state == TASK_DEAD)) {
2761 * Remove function-return probe instances associated with this
2762 * task and put them back on the free list.
2764 kprobe_flush_task(prev);
2765 put_task_struct(prev);
2771 /* assumes rq->lock is held */
2772 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2774 if (prev->sched_class->pre_schedule)
2775 prev->sched_class->pre_schedule(rq, prev);
2778 /* rq->lock is NOT held, but preemption is disabled */
2779 static inline void post_schedule(struct rq *rq)
2781 if (rq->post_schedule) {
2782 unsigned long flags;
2784 raw_spin_lock_irqsave(&rq->lock, flags);
2785 if (rq->curr->sched_class->post_schedule)
2786 rq->curr->sched_class->post_schedule(rq);
2787 raw_spin_unlock_irqrestore(&rq->lock, flags);
2789 rq->post_schedule = 0;
2795 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2799 static inline void post_schedule(struct rq *rq)
2806 * schedule_tail - first thing a freshly forked thread must call.
2807 * @prev: the thread we just switched away from.
2809 asmlinkage void schedule_tail(struct task_struct *prev)
2810 __releases(rq->lock)
2812 struct rq *rq = this_rq();
2814 finish_task_switch(rq, prev);
2817 * FIXME: do we need to worry about rq being invalidated by the
2822 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2823 /* In this case, finish_task_switch does not reenable preemption */
2826 if (current->set_child_tid)
2827 put_user(task_pid_vnr(current), current->set_child_tid);
2831 * context_switch - switch to the new MM and the new
2832 * thread's register state.
2835 context_switch(struct rq *rq, struct task_struct *prev,
2836 struct task_struct *next)
2838 struct mm_struct *mm, *oldmm;
2840 prepare_task_switch(rq, prev, next);
2841 trace_sched_switch(rq, prev, next);
2843 oldmm = prev->active_mm;
2845 * For paravirt, this is coupled with an exit in switch_to to
2846 * combine the page table reload and the switch backend into
2849 arch_start_context_switch(prev);
2852 next->active_mm = oldmm;
2853 atomic_inc(&oldmm->mm_count);
2854 enter_lazy_tlb(oldmm, next);
2856 switch_mm(oldmm, mm, next);
2858 if (likely(!prev->mm)) {
2859 prev->active_mm = NULL;
2860 rq->prev_mm = oldmm;
2863 * Since the runqueue lock will be released by the next
2864 * task (which is an invalid locking op but in the case
2865 * of the scheduler it's an obvious special-case), so we
2866 * do an early lockdep release here:
2868 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2869 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2872 /* Here we just switch the register state and the stack. */
2873 switch_to(prev, next, prev);
2877 * this_rq must be evaluated again because prev may have moved
2878 * CPUs since it called schedule(), thus the 'rq' on its stack
2879 * frame will be invalid.
2881 finish_task_switch(this_rq(), prev);
2885 * nr_running, nr_uninterruptible and nr_context_switches:
2887 * externally visible scheduler statistics: current number of runnable
2888 * threads, current number of uninterruptible-sleeping threads, total
2889 * number of context switches performed since bootup.
2891 unsigned long nr_running(void)
2893 unsigned long i, sum = 0;
2895 for_each_online_cpu(i)
2896 sum += cpu_rq(i)->nr_running;
2901 unsigned long nr_uninterruptible(void)
2903 unsigned long i, sum = 0;
2905 for_each_possible_cpu(i)
2906 sum += cpu_rq(i)->nr_uninterruptible;
2909 * Since we read the counters lockless, it might be slightly
2910 * inaccurate. Do not allow it to go below zero though:
2912 if (unlikely((long)sum < 0))
2918 unsigned long long nr_context_switches(void)
2921 unsigned long long sum = 0;
2923 for_each_possible_cpu(i)
2924 sum += cpu_rq(i)->nr_switches;
2929 unsigned long nr_iowait(void)
2931 unsigned long i, sum = 0;
2933 for_each_possible_cpu(i)
2934 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2939 unsigned long nr_iowait_cpu(void)
2941 struct rq *this = this_rq();
2942 return atomic_read(&this->nr_iowait);
2945 unsigned long this_cpu_load(void)
2947 struct rq *this = this_rq();
2948 return this->cpu_load[0];
2952 /* Variables and functions for calc_load */
2953 static atomic_long_t calc_load_tasks;
2954 static unsigned long calc_load_update;
2955 unsigned long avenrun[3];
2956 EXPORT_SYMBOL(avenrun);
2959 * get_avenrun - get the load average array
2960 * @loads: pointer to dest load array
2961 * @offset: offset to add
2962 * @shift: shift count to shift the result left
2964 * These values are estimates at best, so no need for locking.
2966 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2968 loads[0] = (avenrun[0] + offset) << shift;
2969 loads[1] = (avenrun[1] + offset) << shift;
2970 loads[2] = (avenrun[2] + offset) << shift;
2973 static unsigned long
2974 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2977 load += active * (FIXED_1 - exp);
2978 return load >> FSHIFT;
2982 * calc_load - update the avenrun load estimates 10 ticks after the
2983 * CPUs have updated calc_load_tasks.
2985 void calc_global_load(void)
2987 unsigned long upd = calc_load_update + 10;
2990 if (time_before(jiffies, upd))
2993 active = atomic_long_read(&calc_load_tasks);
2994 active = active > 0 ? active * FIXED_1 : 0;
2996 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2997 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2998 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3000 calc_load_update += LOAD_FREQ;
3004 * Either called from update_cpu_load() or from a cpu going idle
3006 static void calc_load_account_active(struct rq *this_rq)
3008 long nr_active, delta;
3010 nr_active = this_rq->nr_running;
3011 nr_active += (long) this_rq->nr_uninterruptible;
3013 if (nr_active != this_rq->calc_load_active) {
3014 delta = nr_active - this_rq->calc_load_active;
3015 this_rq->calc_load_active = nr_active;
3016 atomic_long_add(delta, &calc_load_tasks);
3021 * Update rq->cpu_load[] statistics. This function is usually called every
3022 * scheduler tick (TICK_NSEC).
3024 static void update_cpu_load(struct rq *this_rq)
3026 unsigned long this_load = this_rq->load.weight;
3029 this_rq->nr_load_updates++;
3031 /* Update our load: */
3032 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3033 unsigned long old_load, new_load;
3035 /* scale is effectively 1 << i now, and >> i divides by scale */
3037 old_load = this_rq->cpu_load[i];
3038 new_load = this_load;
3040 * Round up the averaging division if load is increasing. This
3041 * prevents us from getting stuck on 9 if the load is 10, for
3044 if (new_load > old_load)
3045 new_load += scale-1;
3046 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3049 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3050 this_rq->calc_load_update += LOAD_FREQ;
3051 calc_load_account_active(this_rq);
3058 * sched_exec - execve() is a valuable balancing opportunity, because at
3059 * this point the task has the smallest effective memory and cache footprint.
3061 void sched_exec(void)
3063 struct task_struct *p = current;
3064 struct migration_req req;
3065 int dest_cpu, this_cpu;
3066 unsigned long flags;
3070 this_cpu = get_cpu();
3071 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3072 if (dest_cpu == this_cpu) {
3077 rq = task_rq_lock(p, &flags);
3081 * select_task_rq() can race against ->cpus_allowed
3083 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3084 || unlikely(!cpu_active(dest_cpu))) {
3085 task_rq_unlock(rq, &flags);
3089 /* force the process onto the specified CPU */
3090 if (migrate_task(p, dest_cpu, &req)) {
3091 /* Need to wait for migration thread (might exit: take ref). */
3092 struct task_struct *mt = rq->migration_thread;
3094 get_task_struct(mt);
3095 task_rq_unlock(rq, &flags);
3096 wake_up_process(mt);
3097 put_task_struct(mt);
3098 wait_for_completion(&req.done);
3102 task_rq_unlock(rq, &flags);
3107 DEFINE_PER_CPU(struct kernel_stat, kstat);
3109 EXPORT_PER_CPU_SYMBOL(kstat);
3112 * Return any ns on the sched_clock that have not yet been accounted in
3113 * @p in case that task is currently running.
3115 * Called with task_rq_lock() held on @rq.
3117 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3121 if (task_current(rq, p)) {
3122 update_rq_clock(rq);
3123 ns = rq->clock - p->se.exec_start;
3131 unsigned long long task_delta_exec(struct task_struct *p)
3133 unsigned long flags;
3137 rq = task_rq_lock(p, &flags);
3138 ns = do_task_delta_exec(p, rq);
3139 task_rq_unlock(rq, &flags);
3145 * Return accounted runtime for the task.
3146 * In case the task is currently running, return the runtime plus current's
3147 * pending runtime that have not been accounted yet.
3149 unsigned long long task_sched_runtime(struct task_struct *p)
3151 unsigned long flags;
3155 rq = task_rq_lock(p, &flags);
3156 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3157 task_rq_unlock(rq, &flags);
3163 * Return sum_exec_runtime for the thread group.
3164 * In case the task is currently running, return the sum plus current's
3165 * pending runtime that have not been accounted yet.
3167 * Note that the thread group might have other running tasks as well,
3168 * so the return value not includes other pending runtime that other
3169 * running tasks might have.
3171 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3173 struct task_cputime totals;
3174 unsigned long flags;
3178 rq = task_rq_lock(p, &flags);
3179 thread_group_cputime(p, &totals);
3180 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3181 task_rq_unlock(rq, &flags);
3187 * Account user cpu time to a process.
3188 * @p: the process that the cpu time gets accounted to
3189 * @cputime: the cpu time spent in user space since the last update
3190 * @cputime_scaled: cputime scaled by cpu frequency
3192 void account_user_time(struct task_struct *p, cputime_t cputime,
3193 cputime_t cputime_scaled)
3195 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3198 /* Add user time to process. */
3199 p->utime = cputime_add(p->utime, cputime);
3200 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3201 account_group_user_time(p, cputime);
3203 /* Add user time to cpustat. */
3204 tmp = cputime_to_cputime64(cputime);
3205 if (TASK_NICE(p) > 0)
3206 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3208 cpustat->user = cputime64_add(cpustat->user, tmp);
3210 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3211 /* Account for user time used */
3212 acct_update_integrals(p);
3216 * Account guest cpu time to a process.
3217 * @p: the process that the cpu time gets accounted to
3218 * @cputime: the cpu time spent in virtual machine since the last update
3219 * @cputime_scaled: cputime scaled by cpu frequency
3221 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3222 cputime_t cputime_scaled)
3225 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3227 tmp = cputime_to_cputime64(cputime);
3229 /* Add guest time to process. */
3230 p->utime = cputime_add(p->utime, cputime);
3231 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3232 account_group_user_time(p, cputime);
3233 p->gtime = cputime_add(p->gtime, cputime);
3235 /* Add guest time to cpustat. */
3236 if (TASK_NICE(p) > 0) {
3237 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3238 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3240 cpustat->user = cputime64_add(cpustat->user, tmp);
3241 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3246 * Account system cpu time to a process.
3247 * @p: the process that the cpu time gets accounted to
3248 * @hardirq_offset: the offset to subtract from hardirq_count()
3249 * @cputime: the cpu time spent in kernel space since the last update
3250 * @cputime_scaled: cputime scaled by cpu frequency
3252 void account_system_time(struct task_struct *p, int hardirq_offset,
3253 cputime_t cputime, cputime_t cputime_scaled)
3255 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3258 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3259 account_guest_time(p, cputime, cputime_scaled);
3263 /* Add system time to process. */
3264 p->stime = cputime_add(p->stime, cputime);
3265 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3266 account_group_system_time(p, cputime);
3268 /* Add system time to cpustat. */
3269 tmp = cputime_to_cputime64(cputime);
3270 if (hardirq_count() - hardirq_offset)
3271 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3272 else if (softirq_count())
3273 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3275 cpustat->system = cputime64_add(cpustat->system, tmp);
3277 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3279 /* Account for system time used */
3280 acct_update_integrals(p);
3284 * Account for involuntary wait time.
3285 * @steal: the cpu time spent in involuntary wait
3287 void account_steal_time(cputime_t cputime)
3289 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3290 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3292 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3296 * Account for idle time.
3297 * @cputime: the cpu time spent in idle wait
3299 void account_idle_time(cputime_t cputime)
3301 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3302 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3303 struct rq *rq = this_rq();
3305 if (atomic_read(&rq->nr_iowait) > 0)
3306 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3308 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3311 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3314 * Account a single tick of cpu time.
3315 * @p: the process that the cpu time gets accounted to
3316 * @user_tick: indicates if the tick is a user or a system tick
3318 void account_process_tick(struct task_struct *p, int user_tick)
3320 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3321 struct rq *rq = this_rq();
3324 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3325 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3326 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3329 account_idle_time(cputime_one_jiffy);
3333 * Account multiple ticks of steal time.
3334 * @p: the process from which the cpu time has been stolen
3335 * @ticks: number of stolen ticks
3337 void account_steal_ticks(unsigned long ticks)
3339 account_steal_time(jiffies_to_cputime(ticks));
3343 * Account multiple ticks of idle time.
3344 * @ticks: number of stolen ticks
3346 void account_idle_ticks(unsigned long ticks)
3348 account_idle_time(jiffies_to_cputime(ticks));
3354 * Use precise platform statistics if available:
3356 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3357 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3363 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3365 struct task_cputime cputime;
3367 thread_group_cputime(p, &cputime);
3369 *ut = cputime.utime;
3370 *st = cputime.stime;
3374 #ifndef nsecs_to_cputime
3375 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3378 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3380 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3383 * Use CFS's precise accounting:
3385 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3390 temp = (u64)(rtime * utime);
3391 do_div(temp, total);
3392 utime = (cputime_t)temp;
3397 * Compare with previous values, to keep monotonicity:
3399 p->prev_utime = max(p->prev_utime, utime);
3400 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3402 *ut = p->prev_utime;
3403 *st = p->prev_stime;
3407 * Must be called with siglock held.
3409 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3411 struct signal_struct *sig = p->signal;
3412 struct task_cputime cputime;
3413 cputime_t rtime, utime, total;
3415 thread_group_cputime(p, &cputime);
3417 total = cputime_add(cputime.utime, cputime.stime);
3418 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3423 temp = (u64)(rtime * cputime.utime);
3424 do_div(temp, total);
3425 utime = (cputime_t)temp;
3429 sig->prev_utime = max(sig->prev_utime, utime);
3430 sig->prev_stime = max(sig->prev_stime,
3431 cputime_sub(rtime, sig->prev_utime));
3433 *ut = sig->prev_utime;
3434 *st = sig->prev_stime;
3439 * This function gets called by the timer code, with HZ frequency.
3440 * We call it with interrupts disabled.
3442 * It also gets called by the fork code, when changing the parent's
3445 void scheduler_tick(void)
3447 int cpu = smp_processor_id();
3448 struct rq *rq = cpu_rq(cpu);
3449 struct task_struct *curr = rq->curr;
3453 raw_spin_lock(&rq->lock);
3454 update_rq_clock(rq);
3455 update_cpu_load(rq);
3456 curr->sched_class->task_tick(rq, curr, 0);
3457 raw_spin_unlock(&rq->lock);
3459 perf_event_task_tick(curr, cpu);
3462 rq->idle_at_tick = idle_cpu(cpu);
3463 trigger_load_balance(rq, cpu);
3467 notrace unsigned long get_parent_ip(unsigned long addr)
3469 if (in_lock_functions(addr)) {
3470 addr = CALLER_ADDR2;
3471 if (in_lock_functions(addr))
3472 addr = CALLER_ADDR3;
3477 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3478 defined(CONFIG_PREEMPT_TRACER))
3480 void __kprobes add_preempt_count(int val)
3482 #ifdef CONFIG_DEBUG_PREEMPT
3486 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3489 preempt_count() += val;
3490 #ifdef CONFIG_DEBUG_PREEMPT
3492 * Spinlock count overflowing soon?
3494 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3497 if (preempt_count() == val)
3498 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3500 EXPORT_SYMBOL(add_preempt_count);
3502 void __kprobes sub_preempt_count(int val)
3504 #ifdef CONFIG_DEBUG_PREEMPT
3508 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3511 * Is the spinlock portion underflowing?
3513 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3514 !(preempt_count() & PREEMPT_MASK)))
3518 if (preempt_count() == val)
3519 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3520 preempt_count() -= val;
3522 EXPORT_SYMBOL(sub_preempt_count);
3527 * Print scheduling while atomic bug:
3529 static noinline void __schedule_bug(struct task_struct *prev)
3531 struct pt_regs *regs = get_irq_regs();
3533 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3534 prev->comm, prev->pid, preempt_count());
3536 debug_show_held_locks(prev);
3538 if (irqs_disabled())
3539 print_irqtrace_events(prev);
3548 * Various schedule()-time debugging checks and statistics:
3550 static inline void schedule_debug(struct task_struct *prev)
3553 * Test if we are atomic. Since do_exit() needs to call into
3554 * schedule() atomically, we ignore that path for now.
3555 * Otherwise, whine if we are scheduling when we should not be.
3557 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3558 __schedule_bug(prev);
3560 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3562 schedstat_inc(this_rq(), sched_count);
3563 #ifdef CONFIG_SCHEDSTATS
3564 if (unlikely(prev->lock_depth >= 0)) {
3565 schedstat_inc(this_rq(), bkl_count);
3566 schedstat_inc(prev, sched_info.bkl_count);
3571 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3573 if (prev->state == TASK_RUNNING) {
3574 u64 runtime = prev->se.sum_exec_runtime;
3576 runtime -= prev->se.prev_sum_exec_runtime;
3577 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
3580 * In order to avoid avg_overlap growing stale when we are
3581 * indeed overlapping and hence not getting put to sleep, grow
3582 * the avg_overlap on preemption.
3584 * We use the average preemption runtime because that
3585 * correlates to the amount of cache footprint a task can
3588 update_avg(&prev->se.avg_overlap, runtime);
3590 prev->sched_class->put_prev_task(rq, prev);
3594 * Pick up the highest-prio task:
3596 static inline struct task_struct *
3597 pick_next_task(struct rq *rq)
3599 const struct sched_class *class;
3600 struct task_struct *p;
3603 * Optimization: we know that if all tasks are in
3604 * the fair class we can call that function directly:
3606 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3607 p = fair_sched_class.pick_next_task(rq);
3612 class = sched_class_highest;
3614 p = class->pick_next_task(rq);
3618 * Will never be NULL as the idle class always
3619 * returns a non-NULL p:
3621 class = class->next;
3626 * schedule() is the main scheduler function.
3628 asmlinkage void __sched schedule(void)
3630 struct task_struct *prev, *next;
3631 unsigned long *switch_count;
3637 cpu = smp_processor_id();
3641 switch_count = &prev->nivcsw;
3643 release_kernel_lock(prev);
3644 need_resched_nonpreemptible:
3646 schedule_debug(prev);
3648 if (sched_feat(HRTICK))
3651 raw_spin_lock_irq(&rq->lock);
3652 update_rq_clock(rq);
3653 clear_tsk_need_resched(prev);
3655 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3656 if (unlikely(signal_pending_state(prev->state, prev)))
3657 prev->state = TASK_RUNNING;
3659 deactivate_task(rq, prev, 1);
3660 switch_count = &prev->nvcsw;
3663 pre_schedule(rq, prev);
3665 if (unlikely(!rq->nr_running))
3666 idle_balance(cpu, rq);
3668 put_prev_task(rq, prev);
3669 next = pick_next_task(rq);
3671 if (likely(prev != next)) {
3672 sched_info_switch(prev, next);
3673 perf_event_task_sched_out(prev, next, cpu);
3679 context_switch(rq, prev, next); /* unlocks the rq */
3681 * the context switch might have flipped the stack from under
3682 * us, hence refresh the local variables.
3684 cpu = smp_processor_id();
3687 raw_spin_unlock_irq(&rq->lock);
3691 if (unlikely(reacquire_kernel_lock(current) < 0))
3692 goto need_resched_nonpreemptible;
3694 preempt_enable_no_resched();
3698 EXPORT_SYMBOL(schedule);
3700 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3702 * Look out! "owner" is an entirely speculative pointer
3703 * access and not reliable.
3705 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3710 if (!sched_feat(OWNER_SPIN))
3713 #ifdef CONFIG_DEBUG_PAGEALLOC
3715 * Need to access the cpu field knowing that
3716 * DEBUG_PAGEALLOC could have unmapped it if
3717 * the mutex owner just released it and exited.
3719 if (probe_kernel_address(&owner->cpu, cpu))
3726 * Even if the access succeeded (likely case),
3727 * the cpu field may no longer be valid.
3729 if (cpu >= nr_cpumask_bits)
3733 * We need to validate that we can do a
3734 * get_cpu() and that we have the percpu area.
3736 if (!cpu_online(cpu))
3743 * Owner changed, break to re-assess state.
3745 if (lock->owner != owner)
3749 * Is that owner really running on that cpu?
3751 if (task_thread_info(rq->curr) != owner || need_resched())
3761 #ifdef CONFIG_PREEMPT
3763 * this is the entry point to schedule() from in-kernel preemption
3764 * off of preempt_enable. Kernel preemptions off return from interrupt
3765 * occur there and call schedule directly.
3767 asmlinkage void __sched preempt_schedule(void)
3769 struct thread_info *ti = current_thread_info();
3772 * If there is a non-zero preempt_count or interrupts are disabled,
3773 * we do not want to preempt the current task. Just return..
3775 if (likely(ti->preempt_count || irqs_disabled()))
3779 add_preempt_count(PREEMPT_ACTIVE);
3781 sub_preempt_count(PREEMPT_ACTIVE);
3784 * Check again in case we missed a preemption opportunity
3785 * between schedule and now.
3788 } while (need_resched());
3790 EXPORT_SYMBOL(preempt_schedule);
3793 * this is the entry point to schedule() from kernel preemption
3794 * off of irq context.
3795 * Note, that this is called and return with irqs disabled. This will
3796 * protect us against recursive calling from irq.
3798 asmlinkage void __sched preempt_schedule_irq(void)
3800 struct thread_info *ti = current_thread_info();
3802 /* Catch callers which need to be fixed */
3803 BUG_ON(ti->preempt_count || !irqs_disabled());
3806 add_preempt_count(PREEMPT_ACTIVE);
3809 local_irq_disable();
3810 sub_preempt_count(PREEMPT_ACTIVE);
3813 * Check again in case we missed a preemption opportunity
3814 * between schedule and now.
3817 } while (need_resched());
3820 #endif /* CONFIG_PREEMPT */
3822 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3825 return try_to_wake_up(curr->private, mode, wake_flags);
3827 EXPORT_SYMBOL(default_wake_function);
3830 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3831 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3832 * number) then we wake all the non-exclusive tasks and one exclusive task.
3834 * There are circumstances in which we can try to wake a task which has already
3835 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3836 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3838 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3839 int nr_exclusive, int wake_flags, void *key)
3841 wait_queue_t *curr, *next;
3843 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3844 unsigned flags = curr->flags;
3846 if (curr->func(curr, mode, wake_flags, key) &&
3847 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3853 * __wake_up - wake up threads blocked on a waitqueue.
3855 * @mode: which threads
3856 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3857 * @key: is directly passed to the wakeup function
3859 * It may be assumed that this function implies a write memory barrier before
3860 * changing the task state if and only if any tasks are woken up.
3862 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3863 int nr_exclusive, void *key)
3865 unsigned long flags;
3867 spin_lock_irqsave(&q->lock, flags);
3868 __wake_up_common(q, mode, nr_exclusive, 0, key);
3869 spin_unlock_irqrestore(&q->lock, flags);
3871 EXPORT_SYMBOL(__wake_up);
3874 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3876 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3878 __wake_up_common(q, mode, 1, 0, NULL);
3881 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3883 __wake_up_common(q, mode, 1, 0, key);
3887 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3889 * @mode: which threads
3890 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3891 * @key: opaque value to be passed to wakeup targets
3893 * The sync wakeup differs that the waker knows that it will schedule
3894 * away soon, so while the target thread will be woken up, it will not
3895 * be migrated to another CPU - ie. the two threads are 'synchronized'
3896 * with each other. This can prevent needless bouncing between CPUs.
3898 * On UP it can prevent extra preemption.
3900 * It may be assumed that this function implies a write memory barrier before
3901 * changing the task state if and only if any tasks are woken up.
3903 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3904 int nr_exclusive, void *key)
3906 unsigned long flags;
3907 int wake_flags = WF_SYNC;
3912 if (unlikely(!nr_exclusive))
3915 spin_lock_irqsave(&q->lock, flags);
3916 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3917 spin_unlock_irqrestore(&q->lock, flags);
3919 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3922 * __wake_up_sync - see __wake_up_sync_key()
3924 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3926 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3928 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3931 * complete: - signals a single thread waiting on this completion
3932 * @x: holds the state of this particular completion
3934 * This will wake up a single thread waiting on this completion. Threads will be
3935 * awakened in the same order in which they were queued.
3937 * See also complete_all(), wait_for_completion() and related routines.
3939 * It may be assumed that this function implies a write memory barrier before
3940 * changing the task state if and only if any tasks are woken up.
3942 void complete(struct completion *x)
3944 unsigned long flags;
3946 spin_lock_irqsave(&x->wait.lock, flags);
3948 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3949 spin_unlock_irqrestore(&x->wait.lock, flags);
3951 EXPORT_SYMBOL(complete);
3954 * complete_all: - signals all threads waiting on this completion
3955 * @x: holds the state of this particular completion
3957 * This will wake up all threads waiting on this particular completion event.
3959 * It may be assumed that this function implies a write memory barrier before
3960 * changing the task state if and only if any tasks are woken up.
3962 void complete_all(struct completion *x)
3964 unsigned long flags;
3966 spin_lock_irqsave(&x->wait.lock, flags);
3967 x->done += UINT_MAX/2;
3968 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3969 spin_unlock_irqrestore(&x->wait.lock, flags);
3971 EXPORT_SYMBOL(complete_all);
3973 static inline long __sched
3974 do_wait_for_common(struct completion *x, long timeout, int state)
3977 DECLARE_WAITQUEUE(wait, current);
3979 wait.flags |= WQ_FLAG_EXCLUSIVE;
3980 __add_wait_queue_tail(&x->wait, &wait);
3982 if (signal_pending_state(state, current)) {
3983 timeout = -ERESTARTSYS;
3986 __set_current_state(state);
3987 spin_unlock_irq(&x->wait.lock);
3988 timeout = schedule_timeout(timeout);
3989 spin_lock_irq(&x->wait.lock);
3990 } while (!x->done && timeout);
3991 __remove_wait_queue(&x->wait, &wait);
3996 return timeout ?: 1;
4000 wait_for_common(struct completion *x, long timeout, int state)
4004 spin_lock_irq(&x->wait.lock);
4005 timeout = do_wait_for_common(x, timeout, state);
4006 spin_unlock_irq(&x->wait.lock);
4011 * wait_for_completion: - waits for completion of a task
4012 * @x: holds the state of this particular completion
4014 * This waits to be signaled for completion of a specific task. It is NOT
4015 * interruptible and there is no timeout.
4017 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4018 * and interrupt capability. Also see complete().
4020 void __sched wait_for_completion(struct completion *x)
4022 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4024 EXPORT_SYMBOL(wait_for_completion);
4027 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4028 * @x: holds the state of this particular completion
4029 * @timeout: timeout value in jiffies
4031 * This waits for either a completion of a specific task to be signaled or for a
4032 * specified timeout to expire. The timeout is in jiffies. It is not
4035 unsigned long __sched
4036 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4038 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4040 EXPORT_SYMBOL(wait_for_completion_timeout);
4043 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4044 * @x: holds the state of this particular completion
4046 * This waits for completion of a specific task to be signaled. It is
4049 int __sched wait_for_completion_interruptible(struct completion *x)
4051 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4052 if (t == -ERESTARTSYS)
4056 EXPORT_SYMBOL(wait_for_completion_interruptible);
4059 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4060 * @x: holds the state of this particular completion
4061 * @timeout: timeout value in jiffies
4063 * This waits for either a completion of a specific task to be signaled or for a
4064 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4066 unsigned long __sched
4067 wait_for_completion_interruptible_timeout(struct completion *x,
4068 unsigned long timeout)
4070 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4072 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4075 * wait_for_completion_killable: - waits for completion of a task (killable)
4076 * @x: holds the state of this particular completion
4078 * This waits to be signaled for completion of a specific task. It can be
4079 * interrupted by a kill signal.
4081 int __sched wait_for_completion_killable(struct completion *x)
4083 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4084 if (t == -ERESTARTSYS)
4088 EXPORT_SYMBOL(wait_for_completion_killable);
4091 * try_wait_for_completion - try to decrement a completion without blocking
4092 * @x: completion structure
4094 * Returns: 0 if a decrement cannot be done without blocking
4095 * 1 if a decrement succeeded.
4097 * If a completion is being used as a counting completion,
4098 * attempt to decrement the counter without blocking. This
4099 * enables us to avoid waiting if the resource the completion
4100 * is protecting is not available.
4102 bool try_wait_for_completion(struct completion *x)
4104 unsigned long flags;
4107 spin_lock_irqsave(&x->wait.lock, flags);
4112 spin_unlock_irqrestore(&x->wait.lock, flags);
4115 EXPORT_SYMBOL(try_wait_for_completion);
4118 * completion_done - Test to see if a completion has any waiters
4119 * @x: completion structure
4121 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4122 * 1 if there are no waiters.
4125 bool completion_done(struct completion *x)
4127 unsigned long flags;
4130 spin_lock_irqsave(&x->wait.lock, flags);
4133 spin_unlock_irqrestore(&x->wait.lock, flags);
4136 EXPORT_SYMBOL(completion_done);
4139 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4141 unsigned long flags;
4144 init_waitqueue_entry(&wait, current);
4146 __set_current_state(state);
4148 spin_lock_irqsave(&q->lock, flags);
4149 __add_wait_queue(q, &wait);
4150 spin_unlock(&q->lock);
4151 timeout = schedule_timeout(timeout);
4152 spin_lock_irq(&q->lock);
4153 __remove_wait_queue(q, &wait);
4154 spin_unlock_irqrestore(&q->lock, flags);
4159 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4161 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4163 EXPORT_SYMBOL(interruptible_sleep_on);
4166 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4168 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4170 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4172 void __sched sleep_on(wait_queue_head_t *q)
4174 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4176 EXPORT_SYMBOL(sleep_on);
4178 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4180 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4182 EXPORT_SYMBOL(sleep_on_timeout);
4184 #ifdef CONFIG_RT_MUTEXES
4187 * rt_mutex_setprio - set the current priority of a task
4189 * @prio: prio value (kernel-internal form)
4191 * This function changes the 'effective' priority of a task. It does
4192 * not touch ->normal_prio like __setscheduler().
4194 * Used by the rt_mutex code to implement priority inheritance logic.
4196 void rt_mutex_setprio(struct task_struct *p, int prio)
4198 unsigned long flags;
4199 int oldprio, on_rq, running;
4201 const struct sched_class *prev_class = p->sched_class;
4203 BUG_ON(prio < 0 || prio > MAX_PRIO);
4205 rq = task_rq_lock(p, &flags);
4206 update_rq_clock(rq);
4209 on_rq = p->se.on_rq;
4210 running = task_current(rq, p);
4212 dequeue_task(rq, p, 0);
4214 p->sched_class->put_prev_task(rq, p);
4217 p->sched_class = &rt_sched_class;
4219 p->sched_class = &fair_sched_class;
4224 p->sched_class->set_curr_task(rq);
4226 enqueue_task(rq, p, 0, oldprio < prio);
4228 check_class_changed(rq, p, prev_class, oldprio, running);
4230 task_rq_unlock(rq, &flags);
4235 void set_user_nice(struct task_struct *p, long nice)
4237 int old_prio, delta, on_rq;
4238 unsigned long flags;
4241 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4244 * We have to be careful, if called from sys_setpriority(),
4245 * the task might be in the middle of scheduling on another CPU.
4247 rq = task_rq_lock(p, &flags);
4248 update_rq_clock(rq);
4250 * The RT priorities are set via sched_setscheduler(), but we still
4251 * allow the 'normal' nice value to be set - but as expected
4252 * it wont have any effect on scheduling until the task is
4253 * SCHED_FIFO/SCHED_RR:
4255 if (task_has_rt_policy(p)) {
4256 p->static_prio = NICE_TO_PRIO(nice);
4259 on_rq = p->se.on_rq;
4261 dequeue_task(rq, p, 0);
4263 p->static_prio = NICE_TO_PRIO(nice);
4266 p->prio = effective_prio(p);
4267 delta = p->prio - old_prio;
4270 enqueue_task(rq, p, 0, false);
4272 * If the task increased its priority or is running and
4273 * lowered its priority, then reschedule its CPU:
4275 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4276 resched_task(rq->curr);
4279 task_rq_unlock(rq, &flags);
4281 EXPORT_SYMBOL(set_user_nice);
4284 * can_nice - check if a task can reduce its nice value
4288 int can_nice(const struct task_struct *p, const int nice)
4290 /* convert nice value [19,-20] to rlimit style value [1,40] */
4291 int nice_rlim = 20 - nice;
4293 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4294 capable(CAP_SYS_NICE));
4297 #ifdef __ARCH_WANT_SYS_NICE
4300 * sys_nice - change the priority of the current process.
4301 * @increment: priority increment
4303 * sys_setpriority is a more generic, but much slower function that
4304 * does similar things.
4306 SYSCALL_DEFINE1(nice, int, increment)
4311 * Setpriority might change our priority at the same moment.
4312 * We don't have to worry. Conceptually one call occurs first
4313 * and we have a single winner.
4315 if (increment < -40)
4320 nice = TASK_NICE(current) + increment;
4326 if (increment < 0 && !can_nice(current, nice))
4329 retval = security_task_setnice(current, nice);
4333 set_user_nice(current, nice);
4340 * task_prio - return the priority value of a given task.
4341 * @p: the task in question.
4343 * This is the priority value as seen by users in /proc.
4344 * RT tasks are offset by -200. Normal tasks are centered
4345 * around 0, value goes from -16 to +15.
4347 int task_prio(const struct task_struct *p)
4349 return p->prio - MAX_RT_PRIO;
4353 * task_nice - return the nice value of a given task.
4354 * @p: the task in question.
4356 int task_nice(const struct task_struct *p)
4358 return TASK_NICE(p);
4360 EXPORT_SYMBOL(task_nice);
4363 * idle_cpu - is a given cpu idle currently?
4364 * @cpu: the processor in question.
4366 int idle_cpu(int cpu)
4368 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4372 * idle_task - return the idle task for a given cpu.
4373 * @cpu: the processor in question.
4375 struct task_struct *idle_task(int cpu)
4377 return cpu_rq(cpu)->idle;
4381 * find_process_by_pid - find a process with a matching PID value.
4382 * @pid: the pid in question.
4384 static struct task_struct *find_process_by_pid(pid_t pid)
4386 return pid ? find_task_by_vpid(pid) : current;
4389 /* Actually do priority change: must hold rq lock. */
4391 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4393 BUG_ON(p->se.on_rq);
4396 p->rt_priority = prio;
4397 p->normal_prio = normal_prio(p);
4398 /* we are holding p->pi_lock already */
4399 p->prio = rt_mutex_getprio(p);
4400 if (rt_prio(p->prio))
4401 p->sched_class = &rt_sched_class;
4403 p->sched_class = &fair_sched_class;
4408 * check the target process has a UID that matches the current process's
4410 static bool check_same_owner(struct task_struct *p)
4412 const struct cred *cred = current_cred(), *pcred;
4416 pcred = __task_cred(p);
4417 match = (cred->euid == pcred->euid ||
4418 cred->euid == pcred->uid);
4423 static int __sched_setscheduler(struct task_struct *p, int policy,
4424 struct sched_param *param, bool user)
4426 int retval, oldprio, oldpolicy = -1, on_rq, running;
4427 unsigned long flags;
4428 const struct sched_class *prev_class = p->sched_class;
4432 /* may grab non-irq protected spin_locks */
4433 BUG_ON(in_interrupt());
4435 /* double check policy once rq lock held */
4437 reset_on_fork = p->sched_reset_on_fork;
4438 policy = oldpolicy = p->policy;
4440 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4441 policy &= ~SCHED_RESET_ON_FORK;
4443 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4444 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4445 policy != SCHED_IDLE)
4450 * Valid priorities for SCHED_FIFO and SCHED_RR are
4451 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4452 * SCHED_BATCH and SCHED_IDLE is 0.
4454 if (param->sched_priority < 0 ||
4455 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4456 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4458 if (rt_policy(policy) != (param->sched_priority != 0))
4462 * Allow unprivileged RT tasks to decrease priority:
4464 if (user && !capable(CAP_SYS_NICE)) {
4465 if (rt_policy(policy)) {
4466 unsigned long rlim_rtprio;
4468 if (!lock_task_sighand(p, &flags))
4470 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4471 unlock_task_sighand(p, &flags);
4473 /* can't set/change the rt policy */
4474 if (policy != p->policy && !rlim_rtprio)
4477 /* can't increase priority */
4478 if (param->sched_priority > p->rt_priority &&
4479 param->sched_priority > rlim_rtprio)
4483 * Like positive nice levels, dont allow tasks to
4484 * move out of SCHED_IDLE either:
4486 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4489 /* can't change other user's priorities */
4490 if (!check_same_owner(p))
4493 /* Normal users shall not reset the sched_reset_on_fork flag */
4494 if (p->sched_reset_on_fork && !reset_on_fork)
4499 #ifdef CONFIG_RT_GROUP_SCHED
4501 * Do not allow realtime tasks into groups that have no runtime
4504 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4505 task_group(p)->rt_bandwidth.rt_runtime == 0)
4509 retval = security_task_setscheduler(p, policy, param);
4515 * make sure no PI-waiters arrive (or leave) while we are
4516 * changing the priority of the task:
4518 raw_spin_lock_irqsave(&p->pi_lock, flags);
4520 * To be able to change p->policy safely, the apropriate
4521 * runqueue lock must be held.
4523 rq = __task_rq_lock(p);
4524 /* recheck policy now with rq lock held */
4525 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4526 policy = oldpolicy = -1;
4527 __task_rq_unlock(rq);
4528 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4531 update_rq_clock(rq);
4532 on_rq = p->se.on_rq;
4533 running = task_current(rq, p);
4535 deactivate_task(rq, p, 0);
4537 p->sched_class->put_prev_task(rq, p);
4539 p->sched_reset_on_fork = reset_on_fork;
4542 __setscheduler(rq, p, policy, param->sched_priority);
4545 p->sched_class->set_curr_task(rq);
4547 activate_task(rq, p, 0);
4549 check_class_changed(rq, p, prev_class, oldprio, running);
4551 __task_rq_unlock(rq);
4552 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4554 rt_mutex_adjust_pi(p);
4560 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4561 * @p: the task in question.
4562 * @policy: new policy.
4563 * @param: structure containing the new RT priority.
4565 * NOTE that the task may be already dead.
4567 int sched_setscheduler(struct task_struct *p, int policy,
4568 struct sched_param *param)
4570 return __sched_setscheduler(p, policy, param, true);
4572 EXPORT_SYMBOL_GPL(sched_setscheduler);
4575 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4576 * @p: the task in question.
4577 * @policy: new policy.
4578 * @param: structure containing the new RT priority.
4580 * Just like sched_setscheduler, only don't bother checking if the
4581 * current context has permission. For example, this is needed in
4582 * stop_machine(): we create temporary high priority worker threads,
4583 * but our caller might not have that capability.
4585 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4586 struct sched_param *param)
4588 return __sched_setscheduler(p, policy, param, false);
4592 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4594 struct sched_param lparam;
4595 struct task_struct *p;
4598 if (!param || pid < 0)
4600 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4605 p = find_process_by_pid(pid);
4607 retval = sched_setscheduler(p, policy, &lparam);
4614 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4615 * @pid: the pid in question.
4616 * @policy: new policy.
4617 * @param: structure containing the new RT priority.
4619 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4620 struct sched_param __user *, param)
4622 /* negative values for policy are not valid */
4626 return do_sched_setscheduler(pid, policy, param);
4630 * sys_sched_setparam - set/change the RT priority of a thread
4631 * @pid: the pid in question.
4632 * @param: structure containing the new RT priority.
4634 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4636 return do_sched_setscheduler(pid, -1, param);
4640 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4641 * @pid: the pid in question.
4643 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4645 struct task_struct *p;
4653 p = find_process_by_pid(pid);
4655 retval = security_task_getscheduler(p);
4658 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4665 * sys_sched_getparam - get the RT priority of a thread
4666 * @pid: the pid in question.
4667 * @param: structure containing the RT priority.
4669 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4671 struct sched_param lp;
4672 struct task_struct *p;
4675 if (!param || pid < 0)
4679 p = find_process_by_pid(pid);
4684 retval = security_task_getscheduler(p);
4688 lp.sched_priority = p->rt_priority;
4692 * This one might sleep, we cannot do it with a spinlock held ...
4694 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4703 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4705 cpumask_var_t cpus_allowed, new_mask;
4706 struct task_struct *p;
4712 p = find_process_by_pid(pid);
4719 /* Prevent p going away */
4723 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4727 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4729 goto out_free_cpus_allowed;
4732 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4735 retval = security_task_setscheduler(p, 0, NULL);
4739 cpuset_cpus_allowed(p, cpus_allowed);
4740 cpumask_and(new_mask, in_mask, cpus_allowed);
4742 retval = set_cpus_allowed_ptr(p, new_mask);
4745 cpuset_cpus_allowed(p, cpus_allowed);
4746 if (!cpumask_subset(new_mask, cpus_allowed)) {
4748 * We must have raced with a concurrent cpuset
4749 * update. Just reset the cpus_allowed to the
4750 * cpuset's cpus_allowed
4752 cpumask_copy(new_mask, cpus_allowed);
4757 free_cpumask_var(new_mask);
4758 out_free_cpus_allowed:
4759 free_cpumask_var(cpus_allowed);
4766 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4767 struct cpumask *new_mask)
4769 if (len < cpumask_size())
4770 cpumask_clear(new_mask);
4771 else if (len > cpumask_size())
4772 len = cpumask_size();
4774 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4778 * sys_sched_setaffinity - set the cpu affinity of a process
4779 * @pid: pid of the process
4780 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4781 * @user_mask_ptr: user-space pointer to the new cpu mask
4783 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4784 unsigned long __user *, user_mask_ptr)
4786 cpumask_var_t new_mask;
4789 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4792 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4794 retval = sched_setaffinity(pid, new_mask);
4795 free_cpumask_var(new_mask);
4799 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4801 struct task_struct *p;
4802 unsigned long flags;
4810 p = find_process_by_pid(pid);
4814 retval = security_task_getscheduler(p);
4818 rq = task_rq_lock(p, &flags);
4819 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4820 task_rq_unlock(rq, &flags);
4830 * sys_sched_getaffinity - get the cpu affinity of a process
4831 * @pid: pid of the process
4832 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4833 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4835 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4836 unsigned long __user *, user_mask_ptr)
4841 if (len < cpumask_size())
4844 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4847 ret = sched_getaffinity(pid, mask);
4849 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
4852 ret = cpumask_size();
4854 free_cpumask_var(mask);
4860 * sys_sched_yield - yield the current processor to other threads.
4862 * This function yields the current CPU to other tasks. If there are no
4863 * other threads running on this CPU then this function will return.
4865 SYSCALL_DEFINE0(sched_yield)
4867 struct rq *rq = this_rq_lock();
4869 schedstat_inc(rq, yld_count);
4870 current->sched_class->yield_task(rq);
4873 * Since we are going to call schedule() anyway, there's
4874 * no need to preempt or enable interrupts:
4876 __release(rq->lock);
4877 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4878 do_raw_spin_unlock(&rq->lock);
4879 preempt_enable_no_resched();
4886 static inline int should_resched(void)
4888 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4891 static void __cond_resched(void)
4893 add_preempt_count(PREEMPT_ACTIVE);
4895 sub_preempt_count(PREEMPT_ACTIVE);
4898 int __sched _cond_resched(void)
4900 if (should_resched()) {
4906 EXPORT_SYMBOL(_cond_resched);
4909 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4910 * call schedule, and on return reacquire the lock.
4912 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4913 * operations here to prevent schedule() from being called twice (once via
4914 * spin_unlock(), once by hand).
4916 int __cond_resched_lock(spinlock_t *lock)
4918 int resched = should_resched();
4921 lockdep_assert_held(lock);
4923 if (spin_needbreak(lock) || resched) {
4934 EXPORT_SYMBOL(__cond_resched_lock);
4936 int __sched __cond_resched_softirq(void)
4938 BUG_ON(!in_softirq());
4940 if (should_resched()) {
4948 EXPORT_SYMBOL(__cond_resched_softirq);
4951 * yield - yield the current processor to other threads.
4953 * This is a shortcut for kernel-space yielding - it marks the
4954 * thread runnable and calls sys_sched_yield().
4956 void __sched yield(void)
4958 set_current_state(TASK_RUNNING);
4961 EXPORT_SYMBOL(yield);
4964 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4965 * that process accounting knows that this is a task in IO wait state.
4967 void __sched io_schedule(void)
4969 struct rq *rq = raw_rq();
4971 delayacct_blkio_start();
4972 atomic_inc(&rq->nr_iowait);
4973 current->in_iowait = 1;
4975 current->in_iowait = 0;
4976 atomic_dec(&rq->nr_iowait);
4977 delayacct_blkio_end();
4979 EXPORT_SYMBOL(io_schedule);
4981 long __sched io_schedule_timeout(long timeout)
4983 struct rq *rq = raw_rq();
4986 delayacct_blkio_start();
4987 atomic_inc(&rq->nr_iowait);
4988 current->in_iowait = 1;
4989 ret = schedule_timeout(timeout);
4990 current->in_iowait = 0;
4991 atomic_dec(&rq->nr_iowait);
4992 delayacct_blkio_end();
4997 * sys_sched_get_priority_max - return maximum RT priority.
4998 * @policy: scheduling class.
5000 * this syscall returns the maximum rt_priority that can be used
5001 * by a given scheduling class.
5003 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5010 ret = MAX_USER_RT_PRIO-1;
5022 * sys_sched_get_priority_min - return minimum RT priority.
5023 * @policy: scheduling class.
5025 * this syscall returns the minimum rt_priority that can be used
5026 * by a given scheduling class.
5028 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5046 * sys_sched_rr_get_interval - return the default timeslice of a process.
5047 * @pid: pid of the process.
5048 * @interval: userspace pointer to the timeslice value.
5050 * this syscall writes the default timeslice value of a given process
5051 * into the user-space timespec buffer. A value of '0' means infinity.
5053 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5054 struct timespec __user *, interval)
5056 struct task_struct *p;
5057 unsigned int time_slice;
5058 unsigned long flags;
5068 p = find_process_by_pid(pid);
5072 retval = security_task_getscheduler(p);
5076 rq = task_rq_lock(p, &flags);
5077 time_slice = p->sched_class->get_rr_interval(rq, p);
5078 task_rq_unlock(rq, &flags);
5081 jiffies_to_timespec(time_slice, &t);
5082 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5090 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5092 void sched_show_task(struct task_struct *p)
5094 unsigned long free = 0;
5097 state = p->state ? __ffs(p->state) + 1 : 0;
5098 printk(KERN_INFO "%-13.13s %c", p->comm,
5099 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5100 #if BITS_PER_LONG == 32
5101 if (state == TASK_RUNNING)
5102 printk(KERN_CONT " running ");
5104 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5106 if (state == TASK_RUNNING)
5107 printk(KERN_CONT " running task ");
5109 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5111 #ifdef CONFIG_DEBUG_STACK_USAGE
5112 free = stack_not_used(p);
5114 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5115 task_pid_nr(p), task_pid_nr(p->real_parent),
5116 (unsigned long)task_thread_info(p)->flags);
5118 show_stack(p, NULL);
5121 void show_state_filter(unsigned long state_filter)
5123 struct task_struct *g, *p;
5125 #if BITS_PER_LONG == 32
5127 " task PC stack pid father\n");
5130 " task PC stack pid father\n");
5132 read_lock(&tasklist_lock);
5133 do_each_thread(g, p) {
5135 * reset the NMI-timeout, listing all files on a slow
5136 * console might take alot of time:
5138 touch_nmi_watchdog();
5139 if (!state_filter || (p->state & state_filter))
5141 } while_each_thread(g, p);
5143 touch_all_softlockup_watchdogs();
5145 #ifdef CONFIG_SCHED_DEBUG
5146 sysrq_sched_debug_show();
5148 read_unlock(&tasklist_lock);
5150 * Only show locks if all tasks are dumped:
5153 debug_show_all_locks();
5156 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5158 idle->sched_class = &idle_sched_class;
5162 * init_idle - set up an idle thread for a given CPU
5163 * @idle: task in question
5164 * @cpu: cpu the idle task belongs to
5166 * NOTE: this function does not set the idle thread's NEED_RESCHED
5167 * flag, to make booting more robust.
5169 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5171 struct rq *rq = cpu_rq(cpu);
5172 unsigned long flags;
5174 raw_spin_lock_irqsave(&rq->lock, flags);
5177 idle->state = TASK_RUNNING;
5178 idle->se.exec_start = sched_clock();
5180 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5181 __set_task_cpu(idle, cpu);
5183 rq->curr = rq->idle = idle;
5184 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5187 raw_spin_unlock_irqrestore(&rq->lock, flags);
5189 /* Set the preempt count _outside_ the spinlocks! */
5190 #if defined(CONFIG_PREEMPT)
5191 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5193 task_thread_info(idle)->preempt_count = 0;
5196 * The idle tasks have their own, simple scheduling class:
5198 idle->sched_class = &idle_sched_class;
5199 ftrace_graph_init_task(idle);
5203 * In a system that switches off the HZ timer nohz_cpu_mask
5204 * indicates which cpus entered this state. This is used
5205 * in the rcu update to wait only for active cpus. For system
5206 * which do not switch off the HZ timer nohz_cpu_mask should
5207 * always be CPU_BITS_NONE.
5209 cpumask_var_t nohz_cpu_mask;
5212 * Increase the granularity value when there are more CPUs,
5213 * because with more CPUs the 'effective latency' as visible
5214 * to users decreases. But the relationship is not linear,
5215 * so pick a second-best guess by going with the log2 of the
5218 * This idea comes from the SD scheduler of Con Kolivas:
5220 static int get_update_sysctl_factor(void)
5222 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5223 unsigned int factor;
5225 switch (sysctl_sched_tunable_scaling) {
5226 case SCHED_TUNABLESCALING_NONE:
5229 case SCHED_TUNABLESCALING_LINEAR:
5232 case SCHED_TUNABLESCALING_LOG:
5234 factor = 1 + ilog2(cpus);
5241 static void update_sysctl(void)
5243 unsigned int factor = get_update_sysctl_factor();
5245 #define SET_SYSCTL(name) \
5246 (sysctl_##name = (factor) * normalized_sysctl_##name)
5247 SET_SYSCTL(sched_min_granularity);
5248 SET_SYSCTL(sched_latency);
5249 SET_SYSCTL(sched_wakeup_granularity);
5250 SET_SYSCTL(sched_shares_ratelimit);
5254 static inline void sched_init_granularity(void)
5261 * This is how migration works:
5263 * 1) we queue a struct migration_req structure in the source CPU's
5264 * runqueue and wake up that CPU's migration thread.
5265 * 2) we down() the locked semaphore => thread blocks.
5266 * 3) migration thread wakes up (implicitly it forces the migrated
5267 * thread off the CPU)
5268 * 4) it gets the migration request and checks whether the migrated
5269 * task is still in the wrong runqueue.
5270 * 5) if it's in the wrong runqueue then the migration thread removes
5271 * it and puts it into the right queue.
5272 * 6) migration thread up()s the semaphore.
5273 * 7) we wake up and the migration is done.
5277 * Change a given task's CPU affinity. Migrate the thread to a
5278 * proper CPU and schedule it away if the CPU it's executing on
5279 * is removed from the allowed bitmask.
5281 * NOTE: the caller must have a valid reference to the task, the
5282 * task must not exit() & deallocate itself prematurely. The
5283 * call is not atomic; no spinlocks may be held.
5285 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5287 struct migration_req req;
5288 unsigned long flags;
5293 * Since we rely on wake-ups to migrate sleeping tasks, don't change
5294 * the ->cpus_allowed mask from under waking tasks, which would be
5295 * possible when we change rq->lock in ttwu(), so synchronize against
5296 * TASK_WAKING to avoid that.
5299 while (p->state == TASK_WAKING)
5302 rq = task_rq_lock(p, &flags);
5304 if (p->state == TASK_WAKING) {
5305 task_rq_unlock(rq, &flags);
5309 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5314 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5315 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5320 if (p->sched_class->set_cpus_allowed)
5321 p->sched_class->set_cpus_allowed(p, new_mask);
5323 cpumask_copy(&p->cpus_allowed, new_mask);
5324 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5327 /* Can the task run on the task's current CPU? If so, we're done */
5328 if (cpumask_test_cpu(task_cpu(p), new_mask))
5331 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5332 /* Need help from migration thread: drop lock and wait. */
5333 struct task_struct *mt = rq->migration_thread;
5335 get_task_struct(mt);
5336 task_rq_unlock(rq, &flags);
5337 wake_up_process(rq->migration_thread);
5338 put_task_struct(mt);
5339 wait_for_completion(&req.done);
5340 tlb_migrate_finish(p->mm);
5344 task_rq_unlock(rq, &flags);
5348 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5351 * Move (not current) task off this cpu, onto dest cpu. We're doing
5352 * this because either it can't run here any more (set_cpus_allowed()
5353 * away from this CPU, or CPU going down), or because we're
5354 * attempting to rebalance this task on exec (sched_exec).
5356 * So we race with normal scheduler movements, but that's OK, as long
5357 * as the task is no longer on this CPU.
5359 * Returns non-zero if task was successfully migrated.
5361 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5363 struct rq *rq_dest, *rq_src;
5366 if (unlikely(!cpu_active(dest_cpu)))
5369 rq_src = cpu_rq(src_cpu);
5370 rq_dest = cpu_rq(dest_cpu);
5372 double_rq_lock(rq_src, rq_dest);
5373 /* Already moved. */
5374 if (task_cpu(p) != src_cpu)
5376 /* Affinity changed (again). */
5377 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5381 * If we're not on a rq, the next wake-up will ensure we're
5385 deactivate_task(rq_src, p, 0);
5386 set_task_cpu(p, dest_cpu);
5387 activate_task(rq_dest, p, 0);
5388 check_preempt_curr(rq_dest, p, 0);
5393 double_rq_unlock(rq_src, rq_dest);
5397 #define RCU_MIGRATION_IDLE 0
5398 #define RCU_MIGRATION_NEED_QS 1
5399 #define RCU_MIGRATION_GOT_QS 2
5400 #define RCU_MIGRATION_MUST_SYNC 3
5403 * migration_thread - this is a highprio system thread that performs
5404 * thread migration by bumping thread off CPU then 'pushing' onto
5407 static int migration_thread(void *data)
5410 int cpu = (long)data;
5414 BUG_ON(rq->migration_thread != current);
5416 set_current_state(TASK_INTERRUPTIBLE);
5417 while (!kthread_should_stop()) {
5418 struct migration_req *req;
5419 struct list_head *head;
5421 raw_spin_lock_irq(&rq->lock);
5423 if (cpu_is_offline(cpu)) {
5424 raw_spin_unlock_irq(&rq->lock);
5428 if (rq->active_balance) {
5429 active_load_balance(rq, cpu);
5430 rq->active_balance = 0;
5433 head = &rq->migration_queue;
5435 if (list_empty(head)) {
5436 raw_spin_unlock_irq(&rq->lock);
5438 set_current_state(TASK_INTERRUPTIBLE);
5441 req = list_entry(head->next, struct migration_req, list);
5442 list_del_init(head->next);
5444 if (req->task != NULL) {
5445 raw_spin_unlock(&rq->lock);
5446 __migrate_task(req->task, cpu, req->dest_cpu);
5447 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5448 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5449 raw_spin_unlock(&rq->lock);
5451 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5452 raw_spin_unlock(&rq->lock);
5453 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5457 complete(&req->done);
5459 __set_current_state(TASK_RUNNING);
5464 #ifdef CONFIG_HOTPLUG_CPU
5466 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5470 local_irq_disable();
5471 ret = __migrate_task(p, src_cpu, dest_cpu);
5477 * Figure out where task on dead CPU should go, use force if necessary.
5479 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5484 dest_cpu = select_fallback_rq(dead_cpu, p);
5486 /* It can have affinity changed while we were choosing. */
5487 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
5492 * While a dead CPU has no uninterruptible tasks queued at this point,
5493 * it might still have a nonzero ->nr_uninterruptible counter, because
5494 * for performance reasons the counter is not stricly tracking tasks to
5495 * their home CPUs. So we just add the counter to another CPU's counter,
5496 * to keep the global sum constant after CPU-down:
5498 static void migrate_nr_uninterruptible(struct rq *rq_src)
5500 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5501 unsigned long flags;
5503 local_irq_save(flags);
5504 double_rq_lock(rq_src, rq_dest);
5505 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5506 rq_src->nr_uninterruptible = 0;
5507 double_rq_unlock(rq_src, rq_dest);
5508 local_irq_restore(flags);
5511 /* Run through task list and migrate tasks from the dead cpu. */
5512 static void migrate_live_tasks(int src_cpu)
5514 struct task_struct *p, *t;
5516 read_lock(&tasklist_lock);
5518 do_each_thread(t, p) {
5522 if (task_cpu(p) == src_cpu)
5523 move_task_off_dead_cpu(src_cpu, p);
5524 } while_each_thread(t, p);
5526 read_unlock(&tasklist_lock);
5530 * Schedules idle task to be the next runnable task on current CPU.
5531 * It does so by boosting its priority to highest possible.
5532 * Used by CPU offline code.
5534 void sched_idle_next(void)
5536 int this_cpu = smp_processor_id();
5537 struct rq *rq = cpu_rq(this_cpu);
5538 struct task_struct *p = rq->idle;
5539 unsigned long flags;
5541 /* cpu has to be offline */
5542 BUG_ON(cpu_online(this_cpu));
5545 * Strictly not necessary since rest of the CPUs are stopped by now
5546 * and interrupts disabled on the current cpu.
5548 raw_spin_lock_irqsave(&rq->lock, flags);
5550 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5552 update_rq_clock(rq);
5553 activate_task(rq, p, 0);
5555 raw_spin_unlock_irqrestore(&rq->lock, flags);
5559 * Ensures that the idle task is using init_mm right before its cpu goes
5562 void idle_task_exit(void)
5564 struct mm_struct *mm = current->active_mm;
5566 BUG_ON(cpu_online(smp_processor_id()));
5569 switch_mm(mm, &init_mm, current);
5573 /* called under rq->lock with disabled interrupts */
5574 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5576 struct rq *rq = cpu_rq(dead_cpu);
5578 /* Must be exiting, otherwise would be on tasklist. */
5579 BUG_ON(!p->exit_state);
5581 /* Cannot have done final schedule yet: would have vanished. */
5582 BUG_ON(p->state == TASK_DEAD);
5587 * Drop lock around migration; if someone else moves it,
5588 * that's OK. No task can be added to this CPU, so iteration is
5591 raw_spin_unlock_irq(&rq->lock);
5592 move_task_off_dead_cpu(dead_cpu, p);
5593 raw_spin_lock_irq(&rq->lock);
5598 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5599 static void migrate_dead_tasks(unsigned int dead_cpu)
5601 struct rq *rq = cpu_rq(dead_cpu);
5602 struct task_struct *next;
5605 if (!rq->nr_running)
5607 update_rq_clock(rq);
5608 next = pick_next_task(rq);
5611 next->sched_class->put_prev_task(rq, next);
5612 migrate_dead(dead_cpu, next);
5618 * remove the tasks which were accounted by rq from calc_load_tasks.
5620 static void calc_global_load_remove(struct rq *rq)
5622 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5623 rq->calc_load_active = 0;
5625 #endif /* CONFIG_HOTPLUG_CPU */
5627 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5629 static struct ctl_table sd_ctl_dir[] = {
5631 .procname = "sched_domain",
5637 static struct ctl_table sd_ctl_root[] = {
5639 .procname = "kernel",
5641 .child = sd_ctl_dir,
5646 static struct ctl_table *sd_alloc_ctl_entry(int n)
5648 struct ctl_table *entry =
5649 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5654 static void sd_free_ctl_entry(struct ctl_table **tablep)
5656 struct ctl_table *entry;
5659 * In the intermediate directories, both the child directory and
5660 * procname are dynamically allocated and could fail but the mode
5661 * will always be set. In the lowest directory the names are
5662 * static strings and all have proc handlers.
5664 for (entry = *tablep; entry->mode; entry++) {
5666 sd_free_ctl_entry(&entry->child);
5667 if (entry->proc_handler == NULL)
5668 kfree(entry->procname);
5676 set_table_entry(struct ctl_table *entry,
5677 const char *procname, void *data, int maxlen,
5678 mode_t mode, proc_handler *proc_handler)
5680 entry->procname = procname;
5682 entry->maxlen = maxlen;
5684 entry->proc_handler = proc_handler;
5687 static struct ctl_table *
5688 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5690 struct ctl_table *table = sd_alloc_ctl_entry(13);
5695 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5696 sizeof(long), 0644, proc_doulongvec_minmax);
5697 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5698 sizeof(long), 0644, proc_doulongvec_minmax);
5699 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5700 sizeof(int), 0644, proc_dointvec_minmax);
5701 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5702 sizeof(int), 0644, proc_dointvec_minmax);
5703 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5704 sizeof(int), 0644, proc_dointvec_minmax);
5705 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5706 sizeof(int), 0644, proc_dointvec_minmax);
5707 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5708 sizeof(int), 0644, proc_dointvec_minmax);
5709 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5710 sizeof(int), 0644, proc_dointvec_minmax);
5711 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5712 sizeof(int), 0644, proc_dointvec_minmax);
5713 set_table_entry(&table[9], "cache_nice_tries",
5714 &sd->cache_nice_tries,
5715 sizeof(int), 0644, proc_dointvec_minmax);
5716 set_table_entry(&table[10], "flags", &sd->flags,
5717 sizeof(int), 0644, proc_dointvec_minmax);
5718 set_table_entry(&table[11], "name", sd->name,
5719 CORENAME_MAX_SIZE, 0444, proc_dostring);
5720 /* &table[12] is terminator */
5725 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5727 struct ctl_table *entry, *table;
5728 struct sched_domain *sd;
5729 int domain_num = 0, i;
5732 for_each_domain(cpu, sd)
5734 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5739 for_each_domain(cpu, sd) {
5740 snprintf(buf, 32, "domain%d", i);
5741 entry->procname = kstrdup(buf, GFP_KERNEL);
5743 entry->child = sd_alloc_ctl_domain_table(sd);
5750 static struct ctl_table_header *sd_sysctl_header;
5751 static void register_sched_domain_sysctl(void)
5753 int i, cpu_num = num_possible_cpus();
5754 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5757 WARN_ON(sd_ctl_dir[0].child);
5758 sd_ctl_dir[0].child = entry;
5763 for_each_possible_cpu(i) {
5764 snprintf(buf, 32, "cpu%d", i);
5765 entry->procname = kstrdup(buf, GFP_KERNEL);
5767 entry->child = sd_alloc_ctl_cpu_table(i);
5771 WARN_ON(sd_sysctl_header);
5772 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5775 /* may be called multiple times per register */
5776 static void unregister_sched_domain_sysctl(void)
5778 if (sd_sysctl_header)
5779 unregister_sysctl_table(sd_sysctl_header);
5780 sd_sysctl_header = NULL;
5781 if (sd_ctl_dir[0].child)
5782 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5785 static void register_sched_domain_sysctl(void)
5788 static void unregister_sched_domain_sysctl(void)
5793 static void set_rq_online(struct rq *rq)
5796 const struct sched_class *class;
5798 cpumask_set_cpu(rq->cpu, rq->rd->online);
5801 for_each_class(class) {
5802 if (class->rq_online)
5803 class->rq_online(rq);
5808 static void set_rq_offline(struct rq *rq)
5811 const struct sched_class *class;
5813 for_each_class(class) {
5814 if (class->rq_offline)
5815 class->rq_offline(rq);
5818 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5824 * migration_call - callback that gets triggered when a CPU is added.
5825 * Here we can start up the necessary migration thread for the new CPU.
5827 static int __cpuinit
5828 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5830 struct task_struct *p;
5831 int cpu = (long)hcpu;
5832 unsigned long flags;
5837 case CPU_UP_PREPARE:
5838 case CPU_UP_PREPARE_FROZEN:
5839 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5842 kthread_bind(p, cpu);
5843 /* Must be high prio: stop_machine expects to yield to it. */
5844 rq = task_rq_lock(p, &flags);
5845 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5846 task_rq_unlock(rq, &flags);
5848 cpu_rq(cpu)->migration_thread = p;
5849 rq->calc_load_update = calc_load_update;
5853 case CPU_ONLINE_FROZEN:
5854 /* Strictly unnecessary, as first user will wake it. */
5855 wake_up_process(cpu_rq(cpu)->migration_thread);
5857 /* Update our root-domain */
5859 raw_spin_lock_irqsave(&rq->lock, flags);
5861 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5865 raw_spin_unlock_irqrestore(&rq->lock, flags);
5868 #ifdef CONFIG_HOTPLUG_CPU
5869 case CPU_UP_CANCELED:
5870 case CPU_UP_CANCELED_FROZEN:
5871 if (!cpu_rq(cpu)->migration_thread)
5873 /* Unbind it from offline cpu so it can run. Fall thru. */
5874 kthread_bind(cpu_rq(cpu)->migration_thread,
5875 cpumask_any(cpu_online_mask));
5876 kthread_stop(cpu_rq(cpu)->migration_thread);
5877 put_task_struct(cpu_rq(cpu)->migration_thread);
5878 cpu_rq(cpu)->migration_thread = NULL;
5882 case CPU_DEAD_FROZEN:
5883 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5884 migrate_live_tasks(cpu);
5886 kthread_stop(rq->migration_thread);
5887 put_task_struct(rq->migration_thread);
5888 rq->migration_thread = NULL;
5889 /* Idle task back to normal (off runqueue, low prio) */
5890 raw_spin_lock_irq(&rq->lock);
5891 update_rq_clock(rq);
5892 deactivate_task(rq, rq->idle, 0);
5893 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5894 rq->idle->sched_class = &idle_sched_class;
5895 migrate_dead_tasks(cpu);
5896 raw_spin_unlock_irq(&rq->lock);
5898 migrate_nr_uninterruptible(rq);
5899 BUG_ON(rq->nr_running != 0);
5900 calc_global_load_remove(rq);
5902 * No need to migrate the tasks: it was best-effort if
5903 * they didn't take sched_hotcpu_mutex. Just wake up
5906 raw_spin_lock_irq(&rq->lock);
5907 while (!list_empty(&rq->migration_queue)) {
5908 struct migration_req *req;
5910 req = list_entry(rq->migration_queue.next,
5911 struct migration_req, list);
5912 list_del_init(&req->list);
5913 raw_spin_unlock_irq(&rq->lock);
5914 complete(&req->done);
5915 raw_spin_lock_irq(&rq->lock);
5917 raw_spin_unlock_irq(&rq->lock);
5921 case CPU_DYING_FROZEN:
5922 /* Update our root-domain */
5924 raw_spin_lock_irqsave(&rq->lock, flags);
5926 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5929 raw_spin_unlock_irqrestore(&rq->lock, flags);
5937 * Register at high priority so that task migration (migrate_all_tasks)
5938 * happens before everything else. This has to be lower priority than
5939 * the notifier in the perf_event subsystem, though.
5941 static struct notifier_block __cpuinitdata migration_notifier = {
5942 .notifier_call = migration_call,
5946 static int __init migration_init(void)
5948 void *cpu = (void *)(long)smp_processor_id();
5951 /* Start one for the boot CPU: */
5952 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5953 BUG_ON(err == NOTIFY_BAD);
5954 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5955 register_cpu_notifier(&migration_notifier);
5959 early_initcall(migration_init);
5964 #ifdef CONFIG_SCHED_DEBUG
5966 static __read_mostly int sched_domain_debug_enabled;
5968 static int __init sched_domain_debug_setup(char *str)
5970 sched_domain_debug_enabled = 1;
5974 early_param("sched_debug", sched_domain_debug_setup);
5976 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5977 struct cpumask *groupmask)
5979 struct sched_group *group = sd->groups;
5982 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5983 cpumask_clear(groupmask);
5985 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5987 if (!(sd->flags & SD_LOAD_BALANCE)) {
5988 printk("does not load-balance\n");
5990 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5995 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5997 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5998 printk(KERN_ERR "ERROR: domain->span does not contain "
6001 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6002 printk(KERN_ERR "ERROR: domain->groups does not contain"
6006 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6010 printk(KERN_ERR "ERROR: group is NULL\n");
6014 if (!group->cpu_power) {
6015 printk(KERN_CONT "\n");
6016 printk(KERN_ERR "ERROR: domain->cpu_power not "
6021 if (!cpumask_weight(sched_group_cpus(group))) {
6022 printk(KERN_CONT "\n");
6023 printk(KERN_ERR "ERROR: empty group\n");
6027 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6028 printk(KERN_CONT "\n");
6029 printk(KERN_ERR "ERROR: repeated CPUs\n");
6033 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6035 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6037 printk(KERN_CONT " %s", str);
6038 if (group->cpu_power != SCHED_LOAD_SCALE) {
6039 printk(KERN_CONT " (cpu_power = %d)",
6043 group = group->next;
6044 } while (group != sd->groups);
6045 printk(KERN_CONT "\n");
6047 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6048 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6051 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6052 printk(KERN_ERR "ERROR: parent span is not a superset "
6053 "of domain->span\n");
6057 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6059 cpumask_var_t groupmask;
6062 if (!sched_domain_debug_enabled)
6066 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6070 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6072 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6073 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6078 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6085 free_cpumask_var(groupmask);
6087 #else /* !CONFIG_SCHED_DEBUG */
6088 # define sched_domain_debug(sd, cpu) do { } while (0)
6089 #endif /* CONFIG_SCHED_DEBUG */
6091 static int sd_degenerate(struct sched_domain *sd)
6093 if (cpumask_weight(sched_domain_span(sd)) == 1)
6096 /* Following flags need at least 2 groups */
6097 if (sd->flags & (SD_LOAD_BALANCE |
6098 SD_BALANCE_NEWIDLE |
6102 SD_SHARE_PKG_RESOURCES)) {
6103 if (sd->groups != sd->groups->next)
6107 /* Following flags don't use groups */
6108 if (sd->flags & (SD_WAKE_AFFINE))
6115 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6117 unsigned long cflags = sd->flags, pflags = parent->flags;
6119 if (sd_degenerate(parent))
6122 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6125 /* Flags needing groups don't count if only 1 group in parent */
6126 if (parent->groups == parent->groups->next) {
6127 pflags &= ~(SD_LOAD_BALANCE |
6128 SD_BALANCE_NEWIDLE |
6132 SD_SHARE_PKG_RESOURCES);
6133 if (nr_node_ids == 1)
6134 pflags &= ~SD_SERIALIZE;
6136 if (~cflags & pflags)
6142 static void free_rootdomain(struct root_domain *rd)
6144 synchronize_sched();
6146 cpupri_cleanup(&rd->cpupri);
6148 free_cpumask_var(rd->rto_mask);
6149 free_cpumask_var(rd->online);
6150 free_cpumask_var(rd->span);
6154 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6156 struct root_domain *old_rd = NULL;
6157 unsigned long flags;
6159 raw_spin_lock_irqsave(&rq->lock, flags);
6164 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6167 cpumask_clear_cpu(rq->cpu, old_rd->span);
6170 * If we dont want to free the old_rt yet then
6171 * set old_rd to NULL to skip the freeing later
6174 if (!atomic_dec_and_test(&old_rd->refcount))
6178 atomic_inc(&rd->refcount);
6181 cpumask_set_cpu(rq->cpu, rd->span);
6182 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6185 raw_spin_unlock_irqrestore(&rq->lock, flags);
6188 free_rootdomain(old_rd);
6191 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6193 gfp_t gfp = GFP_KERNEL;
6195 memset(rd, 0, sizeof(*rd));
6200 if (!alloc_cpumask_var(&rd->span, gfp))
6202 if (!alloc_cpumask_var(&rd->online, gfp))
6204 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6207 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6212 free_cpumask_var(rd->rto_mask);
6214 free_cpumask_var(rd->online);
6216 free_cpumask_var(rd->span);
6221 static void init_defrootdomain(void)
6223 init_rootdomain(&def_root_domain, true);
6225 atomic_set(&def_root_domain.refcount, 1);
6228 static struct root_domain *alloc_rootdomain(void)
6230 struct root_domain *rd;
6232 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6236 if (init_rootdomain(rd, false) != 0) {
6245 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6246 * hold the hotplug lock.
6249 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6251 struct rq *rq = cpu_rq(cpu);
6252 struct sched_domain *tmp;
6254 /* Remove the sched domains which do not contribute to scheduling. */
6255 for (tmp = sd; tmp; ) {
6256 struct sched_domain *parent = tmp->parent;
6260 if (sd_parent_degenerate(tmp, parent)) {
6261 tmp->parent = parent->parent;
6263 parent->parent->child = tmp;
6268 if (sd && sd_degenerate(sd)) {
6274 sched_domain_debug(sd, cpu);
6276 rq_attach_root(rq, rd);
6277 rcu_assign_pointer(rq->sd, sd);
6280 /* cpus with isolated domains */
6281 static cpumask_var_t cpu_isolated_map;
6283 /* Setup the mask of cpus configured for isolated domains */
6284 static int __init isolated_cpu_setup(char *str)
6286 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6287 cpulist_parse(str, cpu_isolated_map);
6291 __setup("isolcpus=", isolated_cpu_setup);
6294 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6295 * to a function which identifies what group(along with sched group) a CPU
6296 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6297 * (due to the fact that we keep track of groups covered with a struct cpumask).
6299 * init_sched_build_groups will build a circular linked list of the groups
6300 * covered by the given span, and will set each group's ->cpumask correctly,
6301 * and ->cpu_power to 0.
6304 init_sched_build_groups(const struct cpumask *span,
6305 const struct cpumask *cpu_map,
6306 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6307 struct sched_group **sg,
6308 struct cpumask *tmpmask),
6309 struct cpumask *covered, struct cpumask *tmpmask)
6311 struct sched_group *first = NULL, *last = NULL;
6314 cpumask_clear(covered);
6316 for_each_cpu(i, span) {
6317 struct sched_group *sg;
6318 int group = group_fn(i, cpu_map, &sg, tmpmask);
6321 if (cpumask_test_cpu(i, covered))
6324 cpumask_clear(sched_group_cpus(sg));
6327 for_each_cpu(j, span) {
6328 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6331 cpumask_set_cpu(j, covered);
6332 cpumask_set_cpu(j, sched_group_cpus(sg));
6343 #define SD_NODES_PER_DOMAIN 16
6348 * find_next_best_node - find the next node to include in a sched_domain
6349 * @node: node whose sched_domain we're building
6350 * @used_nodes: nodes already in the sched_domain
6352 * Find the next node to include in a given scheduling domain. Simply
6353 * finds the closest node not already in the @used_nodes map.
6355 * Should use nodemask_t.
6357 static int find_next_best_node(int node, nodemask_t *used_nodes)
6359 int i, n, val, min_val, best_node = 0;
6363 for (i = 0; i < nr_node_ids; i++) {
6364 /* Start at @node */
6365 n = (node + i) % nr_node_ids;
6367 if (!nr_cpus_node(n))
6370 /* Skip already used nodes */
6371 if (node_isset(n, *used_nodes))
6374 /* Simple min distance search */
6375 val = node_distance(node, n);
6377 if (val < min_val) {
6383 node_set(best_node, *used_nodes);
6388 * sched_domain_node_span - get a cpumask for a node's sched_domain
6389 * @node: node whose cpumask we're constructing
6390 * @span: resulting cpumask
6392 * Given a node, construct a good cpumask for its sched_domain to span. It
6393 * should be one that prevents unnecessary balancing, but also spreads tasks
6396 static void sched_domain_node_span(int node, struct cpumask *span)
6398 nodemask_t used_nodes;
6401 cpumask_clear(span);
6402 nodes_clear(used_nodes);
6404 cpumask_or(span, span, cpumask_of_node(node));
6405 node_set(node, used_nodes);
6407 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6408 int next_node = find_next_best_node(node, &used_nodes);
6410 cpumask_or(span, span, cpumask_of_node(next_node));
6413 #endif /* CONFIG_NUMA */
6415 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6418 * The cpus mask in sched_group and sched_domain hangs off the end.
6420 * ( See the the comments in include/linux/sched.h:struct sched_group
6421 * and struct sched_domain. )
6423 struct static_sched_group {
6424 struct sched_group sg;
6425 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6428 struct static_sched_domain {
6429 struct sched_domain sd;
6430 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6436 cpumask_var_t domainspan;
6437 cpumask_var_t covered;
6438 cpumask_var_t notcovered;
6440 cpumask_var_t nodemask;
6441 cpumask_var_t this_sibling_map;
6442 cpumask_var_t this_core_map;
6443 cpumask_var_t send_covered;
6444 cpumask_var_t tmpmask;
6445 struct sched_group **sched_group_nodes;
6446 struct root_domain *rd;
6450 sa_sched_groups = 0,
6455 sa_this_sibling_map,
6457 sa_sched_group_nodes,
6467 * SMT sched-domains:
6469 #ifdef CONFIG_SCHED_SMT
6470 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6471 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6474 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6475 struct sched_group **sg, struct cpumask *unused)
6478 *sg = &per_cpu(sched_groups, cpu).sg;
6481 #endif /* CONFIG_SCHED_SMT */
6484 * multi-core sched-domains:
6486 #ifdef CONFIG_SCHED_MC
6487 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6488 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6489 #endif /* CONFIG_SCHED_MC */
6491 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6493 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6494 struct sched_group **sg, struct cpumask *mask)
6498 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6499 group = cpumask_first(mask);
6501 *sg = &per_cpu(sched_group_core, group).sg;
6504 #elif defined(CONFIG_SCHED_MC)
6506 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6507 struct sched_group **sg, struct cpumask *unused)
6510 *sg = &per_cpu(sched_group_core, cpu).sg;
6515 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6516 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6519 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6520 struct sched_group **sg, struct cpumask *mask)
6523 #ifdef CONFIG_SCHED_MC
6524 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6525 group = cpumask_first(mask);
6526 #elif defined(CONFIG_SCHED_SMT)
6527 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6528 group = cpumask_first(mask);
6533 *sg = &per_cpu(sched_group_phys, group).sg;
6539 * The init_sched_build_groups can't handle what we want to do with node
6540 * groups, so roll our own. Now each node has its own list of groups which
6541 * gets dynamically allocated.
6543 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6544 static struct sched_group ***sched_group_nodes_bycpu;
6546 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6547 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6549 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6550 struct sched_group **sg,
6551 struct cpumask *nodemask)
6555 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6556 group = cpumask_first(nodemask);
6559 *sg = &per_cpu(sched_group_allnodes, group).sg;
6563 static void init_numa_sched_groups_power(struct sched_group *group_head)
6565 struct sched_group *sg = group_head;
6571 for_each_cpu(j, sched_group_cpus(sg)) {
6572 struct sched_domain *sd;
6574 sd = &per_cpu(phys_domains, j).sd;
6575 if (j != group_first_cpu(sd->groups)) {
6577 * Only add "power" once for each
6583 sg->cpu_power += sd->groups->cpu_power;
6586 } while (sg != group_head);
6589 static int build_numa_sched_groups(struct s_data *d,
6590 const struct cpumask *cpu_map, int num)
6592 struct sched_domain *sd;
6593 struct sched_group *sg, *prev;
6596 cpumask_clear(d->covered);
6597 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6598 if (cpumask_empty(d->nodemask)) {
6599 d->sched_group_nodes[num] = NULL;
6603 sched_domain_node_span(num, d->domainspan);
6604 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6606 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6609 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6613 d->sched_group_nodes[num] = sg;
6615 for_each_cpu(j, d->nodemask) {
6616 sd = &per_cpu(node_domains, j).sd;
6621 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6623 cpumask_or(d->covered, d->covered, d->nodemask);
6626 for (j = 0; j < nr_node_ids; j++) {
6627 n = (num + j) % nr_node_ids;
6628 cpumask_complement(d->notcovered, d->covered);
6629 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6630 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6631 if (cpumask_empty(d->tmpmask))
6633 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6634 if (cpumask_empty(d->tmpmask))
6636 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6640 "Can not alloc domain group for node %d\n", j);
6644 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6645 sg->next = prev->next;
6646 cpumask_or(d->covered, d->covered, d->tmpmask);
6653 #endif /* CONFIG_NUMA */
6656 /* Free memory allocated for various sched_group structures */
6657 static void free_sched_groups(const struct cpumask *cpu_map,
6658 struct cpumask *nodemask)
6662 for_each_cpu(cpu, cpu_map) {
6663 struct sched_group **sched_group_nodes
6664 = sched_group_nodes_bycpu[cpu];
6666 if (!sched_group_nodes)
6669 for (i = 0; i < nr_node_ids; i++) {
6670 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6672 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6673 if (cpumask_empty(nodemask))
6683 if (oldsg != sched_group_nodes[i])
6686 kfree(sched_group_nodes);
6687 sched_group_nodes_bycpu[cpu] = NULL;
6690 #else /* !CONFIG_NUMA */
6691 static void free_sched_groups(const struct cpumask *cpu_map,
6692 struct cpumask *nodemask)
6695 #endif /* CONFIG_NUMA */
6698 * Initialize sched groups cpu_power.
6700 * cpu_power indicates the capacity of sched group, which is used while
6701 * distributing the load between different sched groups in a sched domain.
6702 * Typically cpu_power for all the groups in a sched domain will be same unless
6703 * there are asymmetries in the topology. If there are asymmetries, group
6704 * having more cpu_power will pickup more load compared to the group having
6707 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6709 struct sched_domain *child;
6710 struct sched_group *group;
6714 WARN_ON(!sd || !sd->groups);
6716 if (cpu != group_first_cpu(sd->groups))
6721 sd->groups->cpu_power = 0;
6724 power = SCHED_LOAD_SCALE;
6725 weight = cpumask_weight(sched_domain_span(sd));
6727 * SMT siblings share the power of a single core.
6728 * Usually multiple threads get a better yield out of
6729 * that one core than a single thread would have,
6730 * reflect that in sd->smt_gain.
6732 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6733 power *= sd->smt_gain;
6735 power >>= SCHED_LOAD_SHIFT;
6737 sd->groups->cpu_power += power;
6742 * Add cpu_power of each child group to this groups cpu_power.
6744 group = child->groups;
6746 sd->groups->cpu_power += group->cpu_power;
6747 group = group->next;
6748 } while (group != child->groups);
6752 * Initializers for schedule domains
6753 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6756 #ifdef CONFIG_SCHED_DEBUG
6757 # define SD_INIT_NAME(sd, type) sd->name = #type
6759 # define SD_INIT_NAME(sd, type) do { } while (0)
6762 #define SD_INIT(sd, type) sd_init_##type(sd)
6764 #define SD_INIT_FUNC(type) \
6765 static noinline void sd_init_##type(struct sched_domain *sd) \
6767 memset(sd, 0, sizeof(*sd)); \
6768 *sd = SD_##type##_INIT; \
6769 sd->level = SD_LV_##type; \
6770 SD_INIT_NAME(sd, type); \
6775 SD_INIT_FUNC(ALLNODES)
6778 #ifdef CONFIG_SCHED_SMT
6779 SD_INIT_FUNC(SIBLING)
6781 #ifdef CONFIG_SCHED_MC
6785 static int default_relax_domain_level = -1;
6787 static int __init setup_relax_domain_level(char *str)
6791 val = simple_strtoul(str, NULL, 0);
6792 if (val < SD_LV_MAX)
6793 default_relax_domain_level = val;
6797 __setup("relax_domain_level=", setup_relax_domain_level);
6799 static void set_domain_attribute(struct sched_domain *sd,
6800 struct sched_domain_attr *attr)
6804 if (!attr || attr->relax_domain_level < 0) {
6805 if (default_relax_domain_level < 0)
6808 request = default_relax_domain_level;
6810 request = attr->relax_domain_level;
6811 if (request < sd->level) {
6812 /* turn off idle balance on this domain */
6813 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6815 /* turn on idle balance on this domain */
6816 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6820 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6821 const struct cpumask *cpu_map)
6824 case sa_sched_groups:
6825 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6826 d->sched_group_nodes = NULL;
6828 free_rootdomain(d->rd); /* fall through */
6830 free_cpumask_var(d->tmpmask); /* fall through */
6831 case sa_send_covered:
6832 free_cpumask_var(d->send_covered); /* fall through */
6833 case sa_this_core_map:
6834 free_cpumask_var(d->this_core_map); /* fall through */
6835 case sa_this_sibling_map:
6836 free_cpumask_var(d->this_sibling_map); /* fall through */
6838 free_cpumask_var(d->nodemask); /* fall through */
6839 case sa_sched_group_nodes:
6841 kfree(d->sched_group_nodes); /* fall through */
6843 free_cpumask_var(d->notcovered); /* fall through */
6845 free_cpumask_var(d->covered); /* fall through */
6847 free_cpumask_var(d->domainspan); /* fall through */
6854 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6855 const struct cpumask *cpu_map)
6858 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6860 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6861 return sa_domainspan;
6862 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6864 /* Allocate the per-node list of sched groups */
6865 d->sched_group_nodes = kcalloc(nr_node_ids,
6866 sizeof(struct sched_group *), GFP_KERNEL);
6867 if (!d->sched_group_nodes) {
6868 printk(KERN_WARNING "Can not alloc sched group node list\n");
6869 return sa_notcovered;
6871 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6873 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6874 return sa_sched_group_nodes;
6875 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6877 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6878 return sa_this_sibling_map;
6879 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6880 return sa_this_core_map;
6881 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6882 return sa_send_covered;
6883 d->rd = alloc_rootdomain();
6885 printk(KERN_WARNING "Cannot alloc root domain\n");
6888 return sa_rootdomain;
6891 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6892 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6894 struct sched_domain *sd = NULL;
6896 struct sched_domain *parent;
6899 if (cpumask_weight(cpu_map) >
6900 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6901 sd = &per_cpu(allnodes_domains, i).sd;
6902 SD_INIT(sd, ALLNODES);
6903 set_domain_attribute(sd, attr);
6904 cpumask_copy(sched_domain_span(sd), cpu_map);
6905 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6910 sd = &per_cpu(node_domains, i).sd;
6912 set_domain_attribute(sd, attr);
6913 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6914 sd->parent = parent;
6917 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6922 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6923 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6924 struct sched_domain *parent, int i)
6926 struct sched_domain *sd;
6927 sd = &per_cpu(phys_domains, i).sd;
6929 set_domain_attribute(sd, attr);
6930 cpumask_copy(sched_domain_span(sd), d->nodemask);
6931 sd->parent = parent;
6934 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6938 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6939 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6940 struct sched_domain *parent, int i)
6942 struct sched_domain *sd = parent;
6943 #ifdef CONFIG_SCHED_MC
6944 sd = &per_cpu(core_domains, i).sd;
6946 set_domain_attribute(sd, attr);
6947 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6948 sd->parent = parent;
6950 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6955 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6956 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6957 struct sched_domain *parent, int i)
6959 struct sched_domain *sd = parent;
6960 #ifdef CONFIG_SCHED_SMT
6961 sd = &per_cpu(cpu_domains, i).sd;
6962 SD_INIT(sd, SIBLING);
6963 set_domain_attribute(sd, attr);
6964 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6965 sd->parent = parent;
6967 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6972 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6973 const struct cpumask *cpu_map, int cpu)
6976 #ifdef CONFIG_SCHED_SMT
6977 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6978 cpumask_and(d->this_sibling_map, cpu_map,
6979 topology_thread_cpumask(cpu));
6980 if (cpu == cpumask_first(d->this_sibling_map))
6981 init_sched_build_groups(d->this_sibling_map, cpu_map,
6983 d->send_covered, d->tmpmask);
6986 #ifdef CONFIG_SCHED_MC
6987 case SD_LV_MC: /* set up multi-core groups */
6988 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6989 if (cpu == cpumask_first(d->this_core_map))
6990 init_sched_build_groups(d->this_core_map, cpu_map,
6992 d->send_covered, d->tmpmask);
6995 case SD_LV_CPU: /* set up physical groups */
6996 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
6997 if (!cpumask_empty(d->nodemask))
6998 init_sched_build_groups(d->nodemask, cpu_map,
7000 d->send_covered, d->tmpmask);
7003 case SD_LV_ALLNODES:
7004 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7005 d->send_covered, d->tmpmask);
7014 * Build sched domains for a given set of cpus and attach the sched domains
7015 * to the individual cpus
7017 static int __build_sched_domains(const struct cpumask *cpu_map,
7018 struct sched_domain_attr *attr)
7020 enum s_alloc alloc_state = sa_none;
7022 struct sched_domain *sd;
7028 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7029 if (alloc_state != sa_rootdomain)
7031 alloc_state = sa_sched_groups;
7034 * Set up domains for cpus specified by the cpu_map.
7036 for_each_cpu(i, cpu_map) {
7037 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7040 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7041 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7042 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7043 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7046 for_each_cpu(i, cpu_map) {
7047 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7048 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7051 /* Set up physical groups */
7052 for (i = 0; i < nr_node_ids; i++)
7053 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7056 /* Set up node groups */
7058 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7060 for (i = 0; i < nr_node_ids; i++)
7061 if (build_numa_sched_groups(&d, cpu_map, i))
7065 /* Calculate CPU power for physical packages and nodes */
7066 #ifdef CONFIG_SCHED_SMT
7067 for_each_cpu(i, cpu_map) {
7068 sd = &per_cpu(cpu_domains, i).sd;
7069 init_sched_groups_power(i, sd);
7072 #ifdef CONFIG_SCHED_MC
7073 for_each_cpu(i, cpu_map) {
7074 sd = &per_cpu(core_domains, i).sd;
7075 init_sched_groups_power(i, sd);
7079 for_each_cpu(i, cpu_map) {
7080 sd = &per_cpu(phys_domains, i).sd;
7081 init_sched_groups_power(i, sd);
7085 for (i = 0; i < nr_node_ids; i++)
7086 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7088 if (d.sd_allnodes) {
7089 struct sched_group *sg;
7091 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7093 init_numa_sched_groups_power(sg);
7097 /* Attach the domains */
7098 for_each_cpu(i, cpu_map) {
7099 #ifdef CONFIG_SCHED_SMT
7100 sd = &per_cpu(cpu_domains, i).sd;
7101 #elif defined(CONFIG_SCHED_MC)
7102 sd = &per_cpu(core_domains, i).sd;
7104 sd = &per_cpu(phys_domains, i).sd;
7106 cpu_attach_domain(sd, d.rd, i);
7109 d.sched_group_nodes = NULL; /* don't free this we still need it */
7110 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7114 __free_domain_allocs(&d, alloc_state, cpu_map);
7118 static int build_sched_domains(const struct cpumask *cpu_map)
7120 return __build_sched_domains(cpu_map, NULL);
7123 static cpumask_var_t *doms_cur; /* current sched domains */
7124 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7125 static struct sched_domain_attr *dattr_cur;
7126 /* attribues of custom domains in 'doms_cur' */
7129 * Special case: If a kmalloc of a doms_cur partition (array of
7130 * cpumask) fails, then fallback to a single sched domain,
7131 * as determined by the single cpumask fallback_doms.
7133 static cpumask_var_t fallback_doms;
7136 * arch_update_cpu_topology lets virtualized architectures update the
7137 * cpu core maps. It is supposed to return 1 if the topology changed
7138 * or 0 if it stayed the same.
7140 int __attribute__((weak)) arch_update_cpu_topology(void)
7145 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7148 cpumask_var_t *doms;
7150 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7153 for (i = 0; i < ndoms; i++) {
7154 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7155 free_sched_domains(doms, i);
7162 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7165 for (i = 0; i < ndoms; i++)
7166 free_cpumask_var(doms[i]);
7171 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7172 * For now this just excludes isolated cpus, but could be used to
7173 * exclude other special cases in the future.
7175 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7179 arch_update_cpu_topology();
7181 doms_cur = alloc_sched_domains(ndoms_cur);
7183 doms_cur = &fallback_doms;
7184 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7186 err = build_sched_domains(doms_cur[0]);
7187 register_sched_domain_sysctl();
7192 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7193 struct cpumask *tmpmask)
7195 free_sched_groups(cpu_map, tmpmask);
7199 * Detach sched domains from a group of cpus specified in cpu_map
7200 * These cpus will now be attached to the NULL domain
7202 static void detach_destroy_domains(const struct cpumask *cpu_map)
7204 /* Save because hotplug lock held. */
7205 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7208 for_each_cpu(i, cpu_map)
7209 cpu_attach_domain(NULL, &def_root_domain, i);
7210 synchronize_sched();
7211 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7214 /* handle null as "default" */
7215 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7216 struct sched_domain_attr *new, int idx_new)
7218 struct sched_domain_attr tmp;
7225 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7226 new ? (new + idx_new) : &tmp,
7227 sizeof(struct sched_domain_attr));
7231 * Partition sched domains as specified by the 'ndoms_new'
7232 * cpumasks in the array doms_new[] of cpumasks. This compares
7233 * doms_new[] to the current sched domain partitioning, doms_cur[].
7234 * It destroys each deleted domain and builds each new domain.
7236 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7237 * The masks don't intersect (don't overlap.) We should setup one
7238 * sched domain for each mask. CPUs not in any of the cpumasks will
7239 * not be load balanced. If the same cpumask appears both in the
7240 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7243 * The passed in 'doms_new' should be allocated using
7244 * alloc_sched_domains. This routine takes ownership of it and will
7245 * free_sched_domains it when done with it. If the caller failed the
7246 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7247 * and partition_sched_domains() will fallback to the single partition
7248 * 'fallback_doms', it also forces the domains to be rebuilt.
7250 * If doms_new == NULL it will be replaced with cpu_online_mask.
7251 * ndoms_new == 0 is a special case for destroying existing domains,
7252 * and it will not create the default domain.
7254 * Call with hotplug lock held
7256 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7257 struct sched_domain_attr *dattr_new)
7262 mutex_lock(&sched_domains_mutex);
7264 /* always unregister in case we don't destroy any domains */
7265 unregister_sched_domain_sysctl();
7267 /* Let architecture update cpu core mappings. */
7268 new_topology = arch_update_cpu_topology();
7270 n = doms_new ? ndoms_new : 0;
7272 /* Destroy deleted domains */
7273 for (i = 0; i < ndoms_cur; i++) {
7274 for (j = 0; j < n && !new_topology; j++) {
7275 if (cpumask_equal(doms_cur[i], doms_new[j])
7276 && dattrs_equal(dattr_cur, i, dattr_new, j))
7279 /* no match - a current sched domain not in new doms_new[] */
7280 detach_destroy_domains(doms_cur[i]);
7285 if (doms_new == NULL) {
7287 doms_new = &fallback_doms;
7288 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7289 WARN_ON_ONCE(dattr_new);
7292 /* Build new domains */
7293 for (i = 0; i < ndoms_new; i++) {
7294 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7295 if (cpumask_equal(doms_new[i], doms_cur[j])
7296 && dattrs_equal(dattr_new, i, dattr_cur, j))
7299 /* no match - add a new doms_new */
7300 __build_sched_domains(doms_new[i],
7301 dattr_new ? dattr_new + i : NULL);
7306 /* Remember the new sched domains */
7307 if (doms_cur != &fallback_doms)
7308 free_sched_domains(doms_cur, ndoms_cur);
7309 kfree(dattr_cur); /* kfree(NULL) is safe */
7310 doms_cur = doms_new;
7311 dattr_cur = dattr_new;
7312 ndoms_cur = ndoms_new;
7314 register_sched_domain_sysctl();
7316 mutex_unlock(&sched_domains_mutex);
7319 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7320 static void arch_reinit_sched_domains(void)
7324 /* Destroy domains first to force the rebuild */
7325 partition_sched_domains(0, NULL, NULL);
7327 rebuild_sched_domains();
7331 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7333 unsigned int level = 0;
7335 if (sscanf(buf, "%u", &level) != 1)
7339 * level is always be positive so don't check for
7340 * level < POWERSAVINGS_BALANCE_NONE which is 0
7341 * What happens on 0 or 1 byte write,
7342 * need to check for count as well?
7345 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7349 sched_smt_power_savings = level;
7351 sched_mc_power_savings = level;
7353 arch_reinit_sched_domains();
7358 #ifdef CONFIG_SCHED_MC
7359 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7362 return sprintf(page, "%u\n", sched_mc_power_savings);
7364 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7365 const char *buf, size_t count)
7367 return sched_power_savings_store(buf, count, 0);
7369 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7370 sched_mc_power_savings_show,
7371 sched_mc_power_savings_store);
7374 #ifdef CONFIG_SCHED_SMT
7375 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7378 return sprintf(page, "%u\n", sched_smt_power_savings);
7380 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7381 const char *buf, size_t count)
7383 return sched_power_savings_store(buf, count, 1);
7385 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7386 sched_smt_power_savings_show,
7387 sched_smt_power_savings_store);
7390 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7394 #ifdef CONFIG_SCHED_SMT
7396 err = sysfs_create_file(&cls->kset.kobj,
7397 &attr_sched_smt_power_savings.attr);
7399 #ifdef CONFIG_SCHED_MC
7400 if (!err && mc_capable())
7401 err = sysfs_create_file(&cls->kset.kobj,
7402 &attr_sched_mc_power_savings.attr);
7406 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7408 #ifndef CONFIG_CPUSETS
7410 * Add online and remove offline CPUs from the scheduler domains.
7411 * When cpusets are enabled they take over this function.
7413 static int update_sched_domains(struct notifier_block *nfb,
7414 unsigned long action, void *hcpu)
7418 case CPU_ONLINE_FROZEN:
7419 case CPU_DOWN_PREPARE:
7420 case CPU_DOWN_PREPARE_FROZEN:
7421 case CPU_DOWN_FAILED:
7422 case CPU_DOWN_FAILED_FROZEN:
7423 partition_sched_domains(1, NULL, NULL);
7432 static int update_runtime(struct notifier_block *nfb,
7433 unsigned long action, void *hcpu)
7435 int cpu = (int)(long)hcpu;
7438 case CPU_DOWN_PREPARE:
7439 case CPU_DOWN_PREPARE_FROZEN:
7440 disable_runtime(cpu_rq(cpu));
7443 case CPU_DOWN_FAILED:
7444 case CPU_DOWN_FAILED_FROZEN:
7446 case CPU_ONLINE_FROZEN:
7447 enable_runtime(cpu_rq(cpu));
7455 void __init sched_init_smp(void)
7457 cpumask_var_t non_isolated_cpus;
7459 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7460 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7462 #if defined(CONFIG_NUMA)
7463 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7465 BUG_ON(sched_group_nodes_bycpu == NULL);
7468 mutex_lock(&sched_domains_mutex);
7469 arch_init_sched_domains(cpu_active_mask);
7470 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7471 if (cpumask_empty(non_isolated_cpus))
7472 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7473 mutex_unlock(&sched_domains_mutex);
7476 #ifndef CONFIG_CPUSETS
7477 /* XXX: Theoretical race here - CPU may be hotplugged now */
7478 hotcpu_notifier(update_sched_domains, 0);
7481 /* RT runtime code needs to handle some hotplug events */
7482 hotcpu_notifier(update_runtime, 0);
7486 /* Move init over to a non-isolated CPU */
7487 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7489 sched_init_granularity();
7490 free_cpumask_var(non_isolated_cpus);
7492 init_sched_rt_class();
7495 void __init sched_init_smp(void)
7497 sched_init_granularity();
7499 #endif /* CONFIG_SMP */
7501 const_debug unsigned int sysctl_timer_migration = 1;
7503 int in_sched_functions(unsigned long addr)
7505 return in_lock_functions(addr) ||
7506 (addr >= (unsigned long)__sched_text_start
7507 && addr < (unsigned long)__sched_text_end);
7510 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7512 cfs_rq->tasks_timeline = RB_ROOT;
7513 INIT_LIST_HEAD(&cfs_rq->tasks);
7514 #ifdef CONFIG_FAIR_GROUP_SCHED
7517 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7520 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7522 struct rt_prio_array *array;
7525 array = &rt_rq->active;
7526 for (i = 0; i < MAX_RT_PRIO; i++) {
7527 INIT_LIST_HEAD(array->queue + i);
7528 __clear_bit(i, array->bitmap);
7530 /* delimiter for bitsearch: */
7531 __set_bit(MAX_RT_PRIO, array->bitmap);
7533 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7534 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7536 rt_rq->highest_prio.next = MAX_RT_PRIO;
7540 rt_rq->rt_nr_migratory = 0;
7541 rt_rq->overloaded = 0;
7542 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7546 rt_rq->rt_throttled = 0;
7547 rt_rq->rt_runtime = 0;
7548 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7550 #ifdef CONFIG_RT_GROUP_SCHED
7551 rt_rq->rt_nr_boosted = 0;
7556 #ifdef CONFIG_FAIR_GROUP_SCHED
7557 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7558 struct sched_entity *se, int cpu, int add,
7559 struct sched_entity *parent)
7561 struct rq *rq = cpu_rq(cpu);
7562 tg->cfs_rq[cpu] = cfs_rq;
7563 init_cfs_rq(cfs_rq, rq);
7566 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7569 /* se could be NULL for init_task_group */
7574 se->cfs_rq = &rq->cfs;
7576 se->cfs_rq = parent->my_q;
7579 se->load.weight = tg->shares;
7580 se->load.inv_weight = 0;
7581 se->parent = parent;
7585 #ifdef CONFIG_RT_GROUP_SCHED
7586 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7587 struct sched_rt_entity *rt_se, int cpu, int add,
7588 struct sched_rt_entity *parent)
7590 struct rq *rq = cpu_rq(cpu);
7592 tg->rt_rq[cpu] = rt_rq;
7593 init_rt_rq(rt_rq, rq);
7595 rt_rq->rt_se = rt_se;
7596 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7598 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7600 tg->rt_se[cpu] = rt_se;
7605 rt_se->rt_rq = &rq->rt;
7607 rt_se->rt_rq = parent->my_q;
7609 rt_se->my_q = rt_rq;
7610 rt_se->parent = parent;
7611 INIT_LIST_HEAD(&rt_se->run_list);
7615 void __init sched_init(void)
7618 unsigned long alloc_size = 0, ptr;
7620 #ifdef CONFIG_FAIR_GROUP_SCHED
7621 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7623 #ifdef CONFIG_RT_GROUP_SCHED
7624 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7626 #ifdef CONFIG_CPUMASK_OFFSTACK
7627 alloc_size += num_possible_cpus() * cpumask_size();
7630 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7632 #ifdef CONFIG_FAIR_GROUP_SCHED
7633 init_task_group.se = (struct sched_entity **)ptr;
7634 ptr += nr_cpu_ids * sizeof(void **);
7636 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7637 ptr += nr_cpu_ids * sizeof(void **);
7639 #endif /* CONFIG_FAIR_GROUP_SCHED */
7640 #ifdef CONFIG_RT_GROUP_SCHED
7641 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7642 ptr += nr_cpu_ids * sizeof(void **);
7644 init_task_group.rt_rq = (struct rt_rq **)ptr;
7645 ptr += nr_cpu_ids * sizeof(void **);
7647 #endif /* CONFIG_RT_GROUP_SCHED */
7648 #ifdef CONFIG_CPUMASK_OFFSTACK
7649 for_each_possible_cpu(i) {
7650 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7651 ptr += cpumask_size();
7653 #endif /* CONFIG_CPUMASK_OFFSTACK */
7657 init_defrootdomain();
7660 init_rt_bandwidth(&def_rt_bandwidth,
7661 global_rt_period(), global_rt_runtime());
7663 #ifdef CONFIG_RT_GROUP_SCHED
7664 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7665 global_rt_period(), global_rt_runtime());
7666 #endif /* CONFIG_RT_GROUP_SCHED */
7668 #ifdef CONFIG_CGROUP_SCHED
7669 list_add(&init_task_group.list, &task_groups);
7670 INIT_LIST_HEAD(&init_task_group.children);
7672 #endif /* CONFIG_CGROUP_SCHED */
7674 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7675 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7676 __alignof__(unsigned long));
7678 for_each_possible_cpu(i) {
7682 raw_spin_lock_init(&rq->lock);
7684 rq->calc_load_active = 0;
7685 rq->calc_load_update = jiffies + LOAD_FREQ;
7686 init_cfs_rq(&rq->cfs, rq);
7687 init_rt_rq(&rq->rt, rq);
7688 #ifdef CONFIG_FAIR_GROUP_SCHED
7689 init_task_group.shares = init_task_group_load;
7690 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7691 #ifdef CONFIG_CGROUP_SCHED
7693 * How much cpu bandwidth does init_task_group get?
7695 * In case of task-groups formed thr' the cgroup filesystem, it
7696 * gets 100% of the cpu resources in the system. This overall
7697 * system cpu resource is divided among the tasks of
7698 * init_task_group and its child task-groups in a fair manner,
7699 * based on each entity's (task or task-group's) weight
7700 * (se->load.weight).
7702 * In other words, if init_task_group has 10 tasks of weight
7703 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7704 * then A0's share of the cpu resource is:
7706 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7708 * We achieve this by letting init_task_group's tasks sit
7709 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7711 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7713 #endif /* CONFIG_FAIR_GROUP_SCHED */
7715 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7716 #ifdef CONFIG_RT_GROUP_SCHED
7717 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7718 #ifdef CONFIG_CGROUP_SCHED
7719 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7723 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7724 rq->cpu_load[j] = 0;
7728 rq->post_schedule = 0;
7729 rq->active_balance = 0;
7730 rq->next_balance = jiffies;
7734 rq->migration_thread = NULL;
7736 rq->avg_idle = 2*sysctl_sched_migration_cost;
7737 INIT_LIST_HEAD(&rq->migration_queue);
7738 rq_attach_root(rq, &def_root_domain);
7741 atomic_set(&rq->nr_iowait, 0);
7744 set_load_weight(&init_task);
7746 #ifdef CONFIG_PREEMPT_NOTIFIERS
7747 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7751 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7754 #ifdef CONFIG_RT_MUTEXES
7755 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7759 * The boot idle thread does lazy MMU switching as well:
7761 atomic_inc(&init_mm.mm_count);
7762 enter_lazy_tlb(&init_mm, current);
7765 * Make us the idle thread. Technically, schedule() should not be
7766 * called from this thread, however somewhere below it might be,
7767 * but because we are the idle thread, we just pick up running again
7768 * when this runqueue becomes "idle".
7770 init_idle(current, smp_processor_id());
7772 calc_load_update = jiffies + LOAD_FREQ;
7775 * During early bootup we pretend to be a normal task:
7777 current->sched_class = &fair_sched_class;
7779 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7780 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7783 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7784 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7786 /* May be allocated at isolcpus cmdline parse time */
7787 if (cpu_isolated_map == NULL)
7788 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7793 scheduler_running = 1;
7796 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7797 static inline int preempt_count_equals(int preempt_offset)
7799 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7801 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7804 void __might_sleep(const char *file, int line, int preempt_offset)
7807 static unsigned long prev_jiffy; /* ratelimiting */
7809 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7810 system_state != SYSTEM_RUNNING || oops_in_progress)
7812 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7814 prev_jiffy = jiffies;
7817 "BUG: sleeping function called from invalid context at %s:%d\n",
7820 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7821 in_atomic(), irqs_disabled(),
7822 current->pid, current->comm);
7824 debug_show_held_locks(current);
7825 if (irqs_disabled())
7826 print_irqtrace_events(current);
7830 EXPORT_SYMBOL(__might_sleep);
7833 #ifdef CONFIG_MAGIC_SYSRQ
7834 static void normalize_task(struct rq *rq, struct task_struct *p)
7838 update_rq_clock(rq);
7839 on_rq = p->se.on_rq;
7841 deactivate_task(rq, p, 0);
7842 __setscheduler(rq, p, SCHED_NORMAL, 0);
7844 activate_task(rq, p, 0);
7845 resched_task(rq->curr);
7849 void normalize_rt_tasks(void)
7851 struct task_struct *g, *p;
7852 unsigned long flags;
7855 read_lock_irqsave(&tasklist_lock, flags);
7856 do_each_thread(g, p) {
7858 * Only normalize user tasks:
7863 p->se.exec_start = 0;
7864 #ifdef CONFIG_SCHEDSTATS
7865 p->se.wait_start = 0;
7866 p->se.sleep_start = 0;
7867 p->se.block_start = 0;
7872 * Renice negative nice level userspace
7875 if (TASK_NICE(p) < 0 && p->mm)
7876 set_user_nice(p, 0);
7880 raw_spin_lock(&p->pi_lock);
7881 rq = __task_rq_lock(p);
7883 normalize_task(rq, p);
7885 __task_rq_unlock(rq);
7886 raw_spin_unlock(&p->pi_lock);
7887 } while_each_thread(g, p);
7889 read_unlock_irqrestore(&tasklist_lock, flags);
7892 #endif /* CONFIG_MAGIC_SYSRQ */
7896 * These functions are only useful for the IA64 MCA handling.
7898 * They can only be called when the whole system has been
7899 * stopped - every CPU needs to be quiescent, and no scheduling
7900 * activity can take place. Using them for anything else would
7901 * be a serious bug, and as a result, they aren't even visible
7902 * under any other configuration.
7906 * curr_task - return the current task for a given cpu.
7907 * @cpu: the processor in question.
7909 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7911 struct task_struct *curr_task(int cpu)
7913 return cpu_curr(cpu);
7917 * set_curr_task - set the current task for a given cpu.
7918 * @cpu: the processor in question.
7919 * @p: the task pointer to set.
7921 * Description: This function must only be used when non-maskable interrupts
7922 * are serviced on a separate stack. It allows the architecture to switch the
7923 * notion of the current task on a cpu in a non-blocking manner. This function
7924 * must be called with all CPU's synchronized, and interrupts disabled, the
7925 * and caller must save the original value of the current task (see
7926 * curr_task() above) and restore that value before reenabling interrupts and
7927 * re-starting the system.
7929 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7931 void set_curr_task(int cpu, struct task_struct *p)
7938 #ifdef CONFIG_FAIR_GROUP_SCHED
7939 static void free_fair_sched_group(struct task_group *tg)
7943 for_each_possible_cpu(i) {
7945 kfree(tg->cfs_rq[i]);
7955 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7957 struct cfs_rq *cfs_rq;
7958 struct sched_entity *se;
7962 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7965 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7969 tg->shares = NICE_0_LOAD;
7971 for_each_possible_cpu(i) {
7974 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7975 GFP_KERNEL, cpu_to_node(i));
7979 se = kzalloc_node(sizeof(struct sched_entity),
7980 GFP_KERNEL, cpu_to_node(i));
7984 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
7995 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7997 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7998 &cpu_rq(cpu)->leaf_cfs_rq_list);
8001 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8003 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8005 #else /* !CONFG_FAIR_GROUP_SCHED */
8006 static inline void free_fair_sched_group(struct task_group *tg)
8011 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8016 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8020 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8023 #endif /* CONFIG_FAIR_GROUP_SCHED */
8025 #ifdef CONFIG_RT_GROUP_SCHED
8026 static void free_rt_sched_group(struct task_group *tg)
8030 destroy_rt_bandwidth(&tg->rt_bandwidth);
8032 for_each_possible_cpu(i) {
8034 kfree(tg->rt_rq[i]);
8036 kfree(tg->rt_se[i]);
8044 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8046 struct rt_rq *rt_rq;
8047 struct sched_rt_entity *rt_se;
8051 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8054 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8058 init_rt_bandwidth(&tg->rt_bandwidth,
8059 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8061 for_each_possible_cpu(i) {
8064 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8065 GFP_KERNEL, cpu_to_node(i));
8069 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8070 GFP_KERNEL, cpu_to_node(i));
8074 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8085 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8087 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8088 &cpu_rq(cpu)->leaf_rt_rq_list);
8091 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8093 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8095 #else /* !CONFIG_RT_GROUP_SCHED */
8096 static inline void free_rt_sched_group(struct task_group *tg)
8101 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8106 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8110 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8113 #endif /* CONFIG_RT_GROUP_SCHED */
8115 #ifdef CONFIG_CGROUP_SCHED
8116 static void free_sched_group(struct task_group *tg)
8118 free_fair_sched_group(tg);
8119 free_rt_sched_group(tg);
8123 /* allocate runqueue etc for a new task group */
8124 struct task_group *sched_create_group(struct task_group *parent)
8126 struct task_group *tg;
8127 unsigned long flags;
8130 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8132 return ERR_PTR(-ENOMEM);
8134 if (!alloc_fair_sched_group(tg, parent))
8137 if (!alloc_rt_sched_group(tg, parent))
8140 spin_lock_irqsave(&task_group_lock, flags);
8141 for_each_possible_cpu(i) {
8142 register_fair_sched_group(tg, i);
8143 register_rt_sched_group(tg, i);
8145 list_add_rcu(&tg->list, &task_groups);
8147 WARN_ON(!parent); /* root should already exist */
8149 tg->parent = parent;
8150 INIT_LIST_HEAD(&tg->children);
8151 list_add_rcu(&tg->siblings, &parent->children);
8152 spin_unlock_irqrestore(&task_group_lock, flags);
8157 free_sched_group(tg);
8158 return ERR_PTR(-ENOMEM);
8161 /* rcu callback to free various structures associated with a task group */
8162 static void free_sched_group_rcu(struct rcu_head *rhp)
8164 /* now it should be safe to free those cfs_rqs */
8165 free_sched_group(container_of(rhp, struct task_group, rcu));
8168 /* Destroy runqueue etc associated with a task group */
8169 void sched_destroy_group(struct task_group *tg)
8171 unsigned long flags;
8174 spin_lock_irqsave(&task_group_lock, flags);
8175 for_each_possible_cpu(i) {
8176 unregister_fair_sched_group(tg, i);
8177 unregister_rt_sched_group(tg, i);
8179 list_del_rcu(&tg->list);
8180 list_del_rcu(&tg->siblings);
8181 spin_unlock_irqrestore(&task_group_lock, flags);
8183 /* wait for possible concurrent references to cfs_rqs complete */
8184 call_rcu(&tg->rcu, free_sched_group_rcu);
8187 /* change task's runqueue when it moves between groups.
8188 * The caller of this function should have put the task in its new group
8189 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8190 * reflect its new group.
8192 void sched_move_task(struct task_struct *tsk)
8195 unsigned long flags;
8198 rq = task_rq_lock(tsk, &flags);
8200 update_rq_clock(rq);
8202 running = task_current(rq, tsk);
8203 on_rq = tsk->se.on_rq;
8206 dequeue_task(rq, tsk, 0);
8207 if (unlikely(running))
8208 tsk->sched_class->put_prev_task(rq, tsk);
8210 set_task_rq(tsk, task_cpu(tsk));
8212 #ifdef CONFIG_FAIR_GROUP_SCHED
8213 if (tsk->sched_class->moved_group)
8214 tsk->sched_class->moved_group(tsk, on_rq);
8217 if (unlikely(running))
8218 tsk->sched_class->set_curr_task(rq);
8220 enqueue_task(rq, tsk, 0, false);
8222 task_rq_unlock(rq, &flags);
8224 #endif /* CONFIG_CGROUP_SCHED */
8226 #ifdef CONFIG_FAIR_GROUP_SCHED
8227 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8229 struct cfs_rq *cfs_rq = se->cfs_rq;
8234 dequeue_entity(cfs_rq, se, 0);
8236 se->load.weight = shares;
8237 se->load.inv_weight = 0;
8240 enqueue_entity(cfs_rq, se, 0);
8243 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8245 struct cfs_rq *cfs_rq = se->cfs_rq;
8246 struct rq *rq = cfs_rq->rq;
8247 unsigned long flags;
8249 raw_spin_lock_irqsave(&rq->lock, flags);
8250 __set_se_shares(se, shares);
8251 raw_spin_unlock_irqrestore(&rq->lock, flags);
8254 static DEFINE_MUTEX(shares_mutex);
8256 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8259 unsigned long flags;
8262 * We can't change the weight of the root cgroup.
8267 if (shares < MIN_SHARES)
8268 shares = MIN_SHARES;
8269 else if (shares > MAX_SHARES)
8270 shares = MAX_SHARES;
8272 mutex_lock(&shares_mutex);
8273 if (tg->shares == shares)
8276 spin_lock_irqsave(&task_group_lock, flags);
8277 for_each_possible_cpu(i)
8278 unregister_fair_sched_group(tg, i);
8279 list_del_rcu(&tg->siblings);
8280 spin_unlock_irqrestore(&task_group_lock, flags);
8282 /* wait for any ongoing reference to this group to finish */
8283 synchronize_sched();
8286 * Now we are free to modify the group's share on each cpu
8287 * w/o tripping rebalance_share or load_balance_fair.
8289 tg->shares = shares;
8290 for_each_possible_cpu(i) {
8294 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8295 set_se_shares(tg->se[i], shares);
8299 * Enable load balance activity on this group, by inserting it back on
8300 * each cpu's rq->leaf_cfs_rq_list.
8302 spin_lock_irqsave(&task_group_lock, flags);
8303 for_each_possible_cpu(i)
8304 register_fair_sched_group(tg, i);
8305 list_add_rcu(&tg->siblings, &tg->parent->children);
8306 spin_unlock_irqrestore(&task_group_lock, flags);
8308 mutex_unlock(&shares_mutex);
8312 unsigned long sched_group_shares(struct task_group *tg)
8318 #ifdef CONFIG_RT_GROUP_SCHED
8320 * Ensure that the real time constraints are schedulable.
8322 static DEFINE_MUTEX(rt_constraints_mutex);
8324 static unsigned long to_ratio(u64 period, u64 runtime)
8326 if (runtime == RUNTIME_INF)
8329 return div64_u64(runtime << 20, period);
8332 /* Must be called with tasklist_lock held */
8333 static inline int tg_has_rt_tasks(struct task_group *tg)
8335 struct task_struct *g, *p;
8337 do_each_thread(g, p) {
8338 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8340 } while_each_thread(g, p);
8345 struct rt_schedulable_data {
8346 struct task_group *tg;
8351 static int tg_schedulable(struct task_group *tg, void *data)
8353 struct rt_schedulable_data *d = data;
8354 struct task_group *child;
8355 unsigned long total, sum = 0;
8356 u64 period, runtime;
8358 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8359 runtime = tg->rt_bandwidth.rt_runtime;
8362 period = d->rt_period;
8363 runtime = d->rt_runtime;
8367 * Cannot have more runtime than the period.
8369 if (runtime > period && runtime != RUNTIME_INF)
8373 * Ensure we don't starve existing RT tasks.
8375 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8378 total = to_ratio(period, runtime);
8381 * Nobody can have more than the global setting allows.
8383 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8387 * The sum of our children's runtime should not exceed our own.
8389 list_for_each_entry_rcu(child, &tg->children, siblings) {
8390 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8391 runtime = child->rt_bandwidth.rt_runtime;
8393 if (child == d->tg) {
8394 period = d->rt_period;
8395 runtime = d->rt_runtime;
8398 sum += to_ratio(period, runtime);
8407 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8409 struct rt_schedulable_data data = {
8411 .rt_period = period,
8412 .rt_runtime = runtime,
8415 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8418 static int tg_set_bandwidth(struct task_group *tg,
8419 u64 rt_period, u64 rt_runtime)
8423 mutex_lock(&rt_constraints_mutex);
8424 read_lock(&tasklist_lock);
8425 err = __rt_schedulable(tg, rt_period, rt_runtime);
8429 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8430 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8431 tg->rt_bandwidth.rt_runtime = rt_runtime;
8433 for_each_possible_cpu(i) {
8434 struct rt_rq *rt_rq = tg->rt_rq[i];
8436 raw_spin_lock(&rt_rq->rt_runtime_lock);
8437 rt_rq->rt_runtime = rt_runtime;
8438 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8440 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8442 read_unlock(&tasklist_lock);
8443 mutex_unlock(&rt_constraints_mutex);
8448 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8450 u64 rt_runtime, rt_period;
8452 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8453 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8454 if (rt_runtime_us < 0)
8455 rt_runtime = RUNTIME_INF;
8457 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8460 long sched_group_rt_runtime(struct task_group *tg)
8464 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8467 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8468 do_div(rt_runtime_us, NSEC_PER_USEC);
8469 return rt_runtime_us;
8472 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8474 u64 rt_runtime, rt_period;
8476 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8477 rt_runtime = tg->rt_bandwidth.rt_runtime;
8482 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8485 long sched_group_rt_period(struct task_group *tg)
8489 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8490 do_div(rt_period_us, NSEC_PER_USEC);
8491 return rt_period_us;
8494 static int sched_rt_global_constraints(void)
8496 u64 runtime, period;
8499 if (sysctl_sched_rt_period <= 0)
8502 runtime = global_rt_runtime();
8503 period = global_rt_period();
8506 * Sanity check on the sysctl variables.
8508 if (runtime > period && runtime != RUNTIME_INF)
8511 mutex_lock(&rt_constraints_mutex);
8512 read_lock(&tasklist_lock);
8513 ret = __rt_schedulable(NULL, 0, 0);
8514 read_unlock(&tasklist_lock);
8515 mutex_unlock(&rt_constraints_mutex);
8520 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8522 /* Don't accept realtime tasks when there is no way for them to run */
8523 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8529 #else /* !CONFIG_RT_GROUP_SCHED */
8530 static int sched_rt_global_constraints(void)
8532 unsigned long flags;
8535 if (sysctl_sched_rt_period <= 0)
8539 * There's always some RT tasks in the root group
8540 * -- migration, kstopmachine etc..
8542 if (sysctl_sched_rt_runtime == 0)
8545 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8546 for_each_possible_cpu(i) {
8547 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8549 raw_spin_lock(&rt_rq->rt_runtime_lock);
8550 rt_rq->rt_runtime = global_rt_runtime();
8551 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8553 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8557 #endif /* CONFIG_RT_GROUP_SCHED */
8559 int sched_rt_handler(struct ctl_table *table, int write,
8560 void __user *buffer, size_t *lenp,
8564 int old_period, old_runtime;
8565 static DEFINE_MUTEX(mutex);
8568 old_period = sysctl_sched_rt_period;
8569 old_runtime = sysctl_sched_rt_runtime;
8571 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8573 if (!ret && write) {
8574 ret = sched_rt_global_constraints();
8576 sysctl_sched_rt_period = old_period;
8577 sysctl_sched_rt_runtime = old_runtime;
8579 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8580 def_rt_bandwidth.rt_period =
8581 ns_to_ktime(global_rt_period());
8584 mutex_unlock(&mutex);
8589 #ifdef CONFIG_CGROUP_SCHED
8591 /* return corresponding task_group object of a cgroup */
8592 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8594 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8595 struct task_group, css);
8598 static struct cgroup_subsys_state *
8599 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8601 struct task_group *tg, *parent;
8603 if (!cgrp->parent) {
8604 /* This is early initialization for the top cgroup */
8605 return &init_task_group.css;
8608 parent = cgroup_tg(cgrp->parent);
8609 tg = sched_create_group(parent);
8611 return ERR_PTR(-ENOMEM);
8617 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8619 struct task_group *tg = cgroup_tg(cgrp);
8621 sched_destroy_group(tg);
8625 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8627 #ifdef CONFIG_RT_GROUP_SCHED
8628 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8631 /* We don't support RT-tasks being in separate groups */
8632 if (tsk->sched_class != &fair_sched_class)
8639 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8640 struct task_struct *tsk, bool threadgroup)
8642 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8646 struct task_struct *c;
8648 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8649 retval = cpu_cgroup_can_attach_task(cgrp, c);
8661 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8662 struct cgroup *old_cont, struct task_struct *tsk,
8665 sched_move_task(tsk);
8667 struct task_struct *c;
8669 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8676 #ifdef CONFIG_FAIR_GROUP_SCHED
8677 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8680 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8683 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8685 struct task_group *tg = cgroup_tg(cgrp);
8687 return (u64) tg->shares;
8689 #endif /* CONFIG_FAIR_GROUP_SCHED */
8691 #ifdef CONFIG_RT_GROUP_SCHED
8692 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8695 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8698 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8700 return sched_group_rt_runtime(cgroup_tg(cgrp));
8703 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8706 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8709 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8711 return sched_group_rt_period(cgroup_tg(cgrp));
8713 #endif /* CONFIG_RT_GROUP_SCHED */
8715 static struct cftype cpu_files[] = {
8716 #ifdef CONFIG_FAIR_GROUP_SCHED
8719 .read_u64 = cpu_shares_read_u64,
8720 .write_u64 = cpu_shares_write_u64,
8723 #ifdef CONFIG_RT_GROUP_SCHED
8725 .name = "rt_runtime_us",
8726 .read_s64 = cpu_rt_runtime_read,
8727 .write_s64 = cpu_rt_runtime_write,
8730 .name = "rt_period_us",
8731 .read_u64 = cpu_rt_period_read_uint,
8732 .write_u64 = cpu_rt_period_write_uint,
8737 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8739 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8742 struct cgroup_subsys cpu_cgroup_subsys = {
8744 .create = cpu_cgroup_create,
8745 .destroy = cpu_cgroup_destroy,
8746 .can_attach = cpu_cgroup_can_attach,
8747 .attach = cpu_cgroup_attach,
8748 .populate = cpu_cgroup_populate,
8749 .subsys_id = cpu_cgroup_subsys_id,
8753 #endif /* CONFIG_CGROUP_SCHED */
8755 #ifdef CONFIG_CGROUP_CPUACCT
8758 * CPU accounting code for task groups.
8760 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8761 * (balbir@in.ibm.com).
8764 /* track cpu usage of a group of tasks and its child groups */
8766 struct cgroup_subsys_state css;
8767 /* cpuusage holds pointer to a u64-type object on every cpu */
8769 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8770 struct cpuacct *parent;
8773 struct cgroup_subsys cpuacct_subsys;
8775 /* return cpu accounting group corresponding to this container */
8776 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8778 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8779 struct cpuacct, css);
8782 /* return cpu accounting group to which this task belongs */
8783 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8785 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8786 struct cpuacct, css);
8789 /* create a new cpu accounting group */
8790 static struct cgroup_subsys_state *cpuacct_create(
8791 struct cgroup_subsys *ss, struct cgroup *cgrp)
8793 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8799 ca->cpuusage = alloc_percpu(u64);
8803 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8804 if (percpu_counter_init(&ca->cpustat[i], 0))
8805 goto out_free_counters;
8808 ca->parent = cgroup_ca(cgrp->parent);
8814 percpu_counter_destroy(&ca->cpustat[i]);
8815 free_percpu(ca->cpuusage);
8819 return ERR_PTR(-ENOMEM);
8822 /* destroy an existing cpu accounting group */
8824 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8826 struct cpuacct *ca = cgroup_ca(cgrp);
8829 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8830 percpu_counter_destroy(&ca->cpustat[i]);
8831 free_percpu(ca->cpuusage);
8835 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8837 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8840 #ifndef CONFIG_64BIT
8842 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8844 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8846 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8854 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8856 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8858 #ifndef CONFIG_64BIT
8860 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8862 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8864 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8870 /* return total cpu usage (in nanoseconds) of a group */
8871 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8873 struct cpuacct *ca = cgroup_ca(cgrp);
8874 u64 totalcpuusage = 0;
8877 for_each_present_cpu(i)
8878 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8880 return totalcpuusage;
8883 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8886 struct cpuacct *ca = cgroup_ca(cgrp);
8895 for_each_present_cpu(i)
8896 cpuacct_cpuusage_write(ca, i, 0);
8902 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8905 struct cpuacct *ca = cgroup_ca(cgroup);
8909 for_each_present_cpu(i) {
8910 percpu = cpuacct_cpuusage_read(ca, i);
8911 seq_printf(m, "%llu ", (unsigned long long) percpu);
8913 seq_printf(m, "\n");
8917 static const char *cpuacct_stat_desc[] = {
8918 [CPUACCT_STAT_USER] = "user",
8919 [CPUACCT_STAT_SYSTEM] = "system",
8922 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8923 struct cgroup_map_cb *cb)
8925 struct cpuacct *ca = cgroup_ca(cgrp);
8928 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8929 s64 val = percpu_counter_read(&ca->cpustat[i]);
8930 val = cputime64_to_clock_t(val);
8931 cb->fill(cb, cpuacct_stat_desc[i], val);
8936 static struct cftype files[] = {
8939 .read_u64 = cpuusage_read,
8940 .write_u64 = cpuusage_write,
8943 .name = "usage_percpu",
8944 .read_seq_string = cpuacct_percpu_seq_read,
8948 .read_map = cpuacct_stats_show,
8952 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8954 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8958 * charge this task's execution time to its accounting group.
8960 * called with rq->lock held.
8962 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8967 if (unlikely(!cpuacct_subsys.active))
8970 cpu = task_cpu(tsk);
8976 for (; ca; ca = ca->parent) {
8977 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8978 *cpuusage += cputime;
8985 * Charge the system/user time to the task's accounting group.
8987 static void cpuacct_update_stats(struct task_struct *tsk,
8988 enum cpuacct_stat_index idx, cputime_t val)
8992 if (unlikely(!cpuacct_subsys.active))
8999 percpu_counter_add(&ca->cpustat[idx], val);
9005 struct cgroup_subsys cpuacct_subsys = {
9007 .create = cpuacct_create,
9008 .destroy = cpuacct_destroy,
9009 .populate = cpuacct_populate,
9010 .subsys_id = cpuacct_subsys_id,
9012 #endif /* CONFIG_CGROUP_CPUACCT */
9016 int rcu_expedited_torture_stats(char *page)
9020 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9022 void synchronize_sched_expedited(void)
9025 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9027 #else /* #ifndef CONFIG_SMP */
9029 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9030 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9032 #define RCU_EXPEDITED_STATE_POST -2
9033 #define RCU_EXPEDITED_STATE_IDLE -1
9035 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9037 int rcu_expedited_torture_stats(char *page)
9042 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9043 for_each_online_cpu(cpu) {
9044 cnt += sprintf(&page[cnt], " %d:%d",
9045 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9047 cnt += sprintf(&page[cnt], "\n");
9050 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9052 static long synchronize_sched_expedited_count;
9055 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9056 * approach to force grace period to end quickly. This consumes
9057 * significant time on all CPUs, and is thus not recommended for
9058 * any sort of common-case code.
9060 * Note that it is illegal to call this function while holding any
9061 * lock that is acquired by a CPU-hotplug notifier. Failing to
9062 * observe this restriction will result in deadlock.
9064 void synchronize_sched_expedited(void)
9067 unsigned long flags;
9068 bool need_full_sync = 0;
9070 struct migration_req *req;
9074 smp_mb(); /* ensure prior mod happens before capturing snap. */
9075 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9077 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9079 if (trycount++ < 10)
9080 udelay(trycount * num_online_cpus());
9082 synchronize_sched();
9085 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9086 smp_mb(); /* ensure test happens before caller kfree */
9091 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9092 for_each_online_cpu(cpu) {
9094 req = &per_cpu(rcu_migration_req, cpu);
9095 init_completion(&req->done);
9097 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9098 raw_spin_lock_irqsave(&rq->lock, flags);
9099 list_add(&req->list, &rq->migration_queue);
9100 raw_spin_unlock_irqrestore(&rq->lock, flags);
9101 wake_up_process(rq->migration_thread);
9103 for_each_online_cpu(cpu) {
9104 rcu_expedited_state = cpu;
9105 req = &per_cpu(rcu_migration_req, cpu);
9107 wait_for_completion(&req->done);
9108 raw_spin_lock_irqsave(&rq->lock, flags);
9109 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9111 req->dest_cpu = RCU_MIGRATION_IDLE;
9112 raw_spin_unlock_irqrestore(&rq->lock, flags);
9114 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9115 synchronize_sched_expedited_count++;
9116 mutex_unlock(&rcu_sched_expedited_mutex);
9119 synchronize_sched();
9121 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9123 #endif /* #else #ifndef CONFIG_SMP */