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>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
123 static inline int rt_policy(int policy)
125 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
130 static inline int task_has_rt_policy(struct task_struct *p)
132 return rt_policy(p->policy);
136 * This is the priority-queue data structure of the RT scheduling class:
138 struct rt_prio_array {
139 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
140 struct list_head queue[MAX_RT_PRIO];
143 struct rt_bandwidth {
144 /* nests inside the rq lock: */
145 raw_spinlock_t rt_runtime_lock;
148 struct hrtimer rt_period_timer;
151 static struct rt_bandwidth def_rt_bandwidth;
153 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
155 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
157 struct rt_bandwidth *rt_b =
158 container_of(timer, struct rt_bandwidth, rt_period_timer);
164 now = hrtimer_cb_get_time(timer);
165 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
170 idle = do_sched_rt_period_timer(rt_b, overrun);
173 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
177 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
179 rt_b->rt_period = ns_to_ktime(period);
180 rt_b->rt_runtime = runtime;
182 raw_spin_lock_init(&rt_b->rt_runtime_lock);
184 hrtimer_init(&rt_b->rt_period_timer,
185 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
186 rt_b->rt_period_timer.function = sched_rt_period_timer;
189 static inline int rt_bandwidth_enabled(void)
191 return sysctl_sched_rt_runtime >= 0;
194 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
198 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
201 if (hrtimer_active(&rt_b->rt_period_timer))
204 raw_spin_lock(&rt_b->rt_runtime_lock);
209 if (hrtimer_active(&rt_b->rt_period_timer))
212 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
213 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
215 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
216 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
217 delta = ktime_to_ns(ktime_sub(hard, soft));
218 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
219 HRTIMER_MODE_ABS_PINNED, 0);
221 raw_spin_unlock(&rt_b->rt_runtime_lock);
224 #ifdef CONFIG_RT_GROUP_SCHED
225 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
227 hrtimer_cancel(&rt_b->rt_period_timer);
232 * sched_domains_mutex serializes calls to arch_init_sched_domains,
233 * detach_destroy_domains and partition_sched_domains.
235 static DEFINE_MUTEX(sched_domains_mutex);
237 #ifdef CONFIG_CGROUP_SCHED
239 #include <linux/cgroup.h>
243 static LIST_HEAD(task_groups);
245 /* task group related information */
247 struct cgroup_subsys_state css;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 /* schedulable entities of this group on each cpu */
251 struct sched_entity **se;
252 /* runqueue "owned" by this group on each cpu */
253 struct cfs_rq **cfs_rq;
254 unsigned long shares;
257 #ifdef CONFIG_RT_GROUP_SCHED
258 struct sched_rt_entity **rt_se;
259 struct rt_rq **rt_rq;
261 struct rt_bandwidth rt_bandwidth;
265 struct list_head list;
267 struct task_group *parent;
268 struct list_head siblings;
269 struct list_head children;
272 #define root_task_group init_task_group
274 /* task_group_lock serializes add/remove of task groups and also changes to
275 * a task group's cpu shares.
277 static DEFINE_SPINLOCK(task_group_lock);
279 #ifdef CONFIG_FAIR_GROUP_SCHED
282 static int root_task_group_empty(void)
284 return list_empty(&root_task_group.children);
288 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
299 #define MAX_SHARES (1UL << 18)
301 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group init_task_group;
309 /* return group to which a task belongs */
310 static inline struct task_group *task_group(struct task_struct *p)
312 struct task_group *tg;
314 #ifdef CONFIG_CGROUP_SCHED
315 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
316 struct task_group, css);
318 tg = &init_task_group;
323 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
324 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
326 #ifdef CONFIG_FAIR_GROUP_SCHED
327 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
328 p->se.parent = task_group(p)->se[cpu];
331 #ifdef CONFIG_RT_GROUP_SCHED
332 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
333 p->rt.parent = task_group(p)->rt_se[cpu];
339 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
340 static inline struct task_group *task_group(struct task_struct *p)
345 #endif /* CONFIG_CGROUP_SCHED */
347 /* CFS-related fields in a runqueue */
349 struct load_weight load;
350 unsigned long nr_running;
355 struct rb_root tasks_timeline;
356 struct rb_node *rb_leftmost;
358 struct list_head tasks;
359 struct list_head *balance_iterator;
362 * 'curr' points to currently running entity on this cfs_rq.
363 * It is set to NULL otherwise (i.e when none are currently running).
365 struct sched_entity *curr, *next, *last;
367 unsigned int nr_spread_over;
369 #ifdef CONFIG_FAIR_GROUP_SCHED
370 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
373 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
374 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
375 * (like users, containers etc.)
377 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
378 * list is used during load balance.
380 struct list_head leaf_cfs_rq_list;
381 struct task_group *tg; /* group that "owns" this runqueue */
385 * the part of load.weight contributed by tasks
387 unsigned long task_weight;
390 * h_load = weight * f(tg)
392 * Where f(tg) is the recursive weight fraction assigned to
395 unsigned long h_load;
398 * this cpu's part of tg->shares
400 unsigned long shares;
403 * load.weight at the time we set shares
405 unsigned long rq_weight;
410 /* Real-Time classes' related field in a runqueue: */
412 struct rt_prio_array active;
413 unsigned long rt_nr_running;
414 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
416 int curr; /* highest queued rt task prio */
418 int next; /* next highest */
423 unsigned long rt_nr_migratory;
424 unsigned long rt_nr_total;
426 struct plist_head pushable_tasks;
431 /* Nests inside the rq lock: */
432 raw_spinlock_t rt_runtime_lock;
434 #ifdef CONFIG_RT_GROUP_SCHED
435 unsigned long rt_nr_boosted;
438 struct list_head leaf_rt_rq_list;
439 struct task_group *tg;
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];
497 unsigned char in_nohz_recently;
499 unsigned int skip_clock_update;
501 /* capture load from *all* tasks on this cpu: */
502 struct load_weight load;
503 unsigned long nr_load_updates;
509 #ifdef CONFIG_FAIR_GROUP_SCHED
510 /* list of leaf cfs_rq on this cpu: */
511 struct list_head leaf_cfs_rq_list;
513 #ifdef CONFIG_RT_GROUP_SCHED
514 struct list_head leaf_rt_rq_list;
518 * This is part of a global counter where only the total sum
519 * over all CPUs matters. A task can increase this counter on
520 * one CPU and if it got migrated afterwards it may decrease
521 * it on another CPU. Always updated under the runqueue lock:
523 unsigned long nr_uninterruptible;
525 struct task_struct *curr, *idle;
526 unsigned long next_balance;
527 struct mm_struct *prev_mm;
534 struct root_domain *rd;
535 struct sched_domain *sd;
537 unsigned char idle_at_tick;
538 /* For active balancing */
542 /* cpu of this runqueue: */
546 unsigned long avg_load_per_task;
548 struct task_struct *migration_thread;
549 struct list_head migration_queue;
557 /* calc_load related fields */
558 unsigned long calc_load_update;
559 long calc_load_active;
561 #ifdef CONFIG_SCHED_HRTICK
563 int hrtick_csd_pending;
564 struct call_single_data hrtick_csd;
566 struct hrtimer hrtick_timer;
569 #ifdef CONFIG_SCHEDSTATS
571 struct sched_info rq_sched_info;
572 unsigned long long rq_cpu_time;
573 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
575 /* sys_sched_yield() stats */
576 unsigned int yld_count;
578 /* schedule() stats */
579 unsigned int sched_switch;
580 unsigned int sched_count;
581 unsigned int sched_goidle;
583 /* try_to_wake_up() stats */
584 unsigned int ttwu_count;
585 unsigned int ttwu_local;
588 unsigned int bkl_count;
592 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
595 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
597 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
600 * A queue event has occurred, and we're going to schedule. In
601 * this case, we can save a useless back to back clock update.
603 if (test_tsk_need_resched(p))
604 rq->skip_clock_update = 1;
607 static inline int cpu_of(struct rq *rq)
616 #define rcu_dereference_check_sched_domain(p) \
617 rcu_dereference_check((p), \
618 rcu_read_lock_sched_held() || \
619 lockdep_is_held(&sched_domains_mutex))
622 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
623 * See detach_destroy_domains: synchronize_sched for details.
625 * The domain tree of any CPU may only be accessed from within
626 * preempt-disabled sections.
628 #define for_each_domain(cpu, __sd) \
629 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
631 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
632 #define this_rq() (&__get_cpu_var(runqueues))
633 #define task_rq(p) cpu_rq(task_cpu(p))
634 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
635 #define raw_rq() (&__raw_get_cpu_var(runqueues))
637 inline void update_rq_clock(struct rq *rq)
639 if (!rq->skip_clock_update)
640 rq->clock = sched_clock_cpu(cpu_of(rq));
644 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
646 #ifdef CONFIG_SCHED_DEBUG
647 # define const_debug __read_mostly
649 # define const_debug static const
654 * @cpu: the processor in question.
656 * Returns true if the current cpu runqueue is locked.
657 * This interface allows printk to be called with the runqueue lock
658 * held and know whether or not it is OK to wake up the klogd.
660 int runqueue_is_locked(int cpu)
662 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
666 * Debugging: various feature bits
669 #define SCHED_FEAT(name, enabled) \
670 __SCHED_FEAT_##name ,
673 #include "sched_features.h"
678 #define SCHED_FEAT(name, enabled) \
679 (1UL << __SCHED_FEAT_##name) * enabled |
681 const_debug unsigned int sysctl_sched_features =
682 #include "sched_features.h"
687 #ifdef CONFIG_SCHED_DEBUG
688 #define SCHED_FEAT(name, enabled) \
691 static __read_mostly char *sched_feat_names[] = {
692 #include "sched_features.h"
698 static int sched_feat_show(struct seq_file *m, void *v)
702 for (i = 0; sched_feat_names[i]; i++) {
703 if (!(sysctl_sched_features & (1UL << i)))
705 seq_printf(m, "%s ", sched_feat_names[i]);
713 sched_feat_write(struct file *filp, const char __user *ubuf,
714 size_t cnt, loff_t *ppos)
724 if (copy_from_user(&buf, ubuf, cnt))
729 if (strncmp(buf, "NO_", 3) == 0) {
734 for (i = 0; sched_feat_names[i]; i++) {
735 int len = strlen(sched_feat_names[i]);
737 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
739 sysctl_sched_features &= ~(1UL << i);
741 sysctl_sched_features |= (1UL << i);
746 if (!sched_feat_names[i])
754 static int sched_feat_open(struct inode *inode, struct file *filp)
756 return single_open(filp, sched_feat_show, NULL);
759 static const struct file_operations sched_feat_fops = {
760 .open = sched_feat_open,
761 .write = sched_feat_write,
764 .release = single_release,
767 static __init int sched_init_debug(void)
769 debugfs_create_file("sched_features", 0644, NULL, NULL,
774 late_initcall(sched_init_debug);
778 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
781 * Number of tasks to iterate in a single balance run.
782 * Limited because this is done with IRQs disabled.
784 const_debug unsigned int sysctl_sched_nr_migrate = 32;
787 * ratelimit for updating the group shares.
790 unsigned int sysctl_sched_shares_ratelimit = 250000;
791 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
794 * Inject some fuzzyness into changing the per-cpu group shares
795 * this avoids remote rq-locks at the expense of fairness.
798 unsigned int sysctl_sched_shares_thresh = 4;
801 * period over which we average the RT time consumption, measured
806 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
809 * period over which we measure -rt task cpu usage in us.
812 unsigned int sysctl_sched_rt_period = 1000000;
814 static __read_mostly int scheduler_running;
817 * part of the period that we allow rt tasks to run in us.
820 int sysctl_sched_rt_runtime = 950000;
822 static inline u64 global_rt_period(void)
824 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
827 static inline u64 global_rt_runtime(void)
829 if (sysctl_sched_rt_runtime < 0)
832 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
835 #ifndef prepare_arch_switch
836 # define prepare_arch_switch(next) do { } while (0)
838 #ifndef finish_arch_switch
839 # define finish_arch_switch(prev) do { } while (0)
842 static inline int task_current(struct rq *rq, struct task_struct *p)
844 return rq->curr == p;
847 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
848 static inline int task_running(struct rq *rq, struct task_struct *p)
850 return task_current(rq, p);
853 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
857 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
859 #ifdef CONFIG_DEBUG_SPINLOCK
860 /* this is a valid case when another task releases the spinlock */
861 rq->lock.owner = current;
864 * If we are tracking spinlock dependencies then we have to
865 * fix up the runqueue lock - which gets 'carried over' from
868 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
870 raw_spin_unlock_irq(&rq->lock);
873 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
874 static inline int task_running(struct rq *rq, struct task_struct *p)
879 return task_current(rq, p);
883 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
887 * We can optimise this out completely for !SMP, because the
888 * SMP rebalancing from interrupt is the only thing that cares
893 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
894 raw_spin_unlock_irq(&rq->lock);
896 raw_spin_unlock(&rq->lock);
900 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
904 * After ->oncpu is cleared, the task can be moved to a different CPU.
905 * We must ensure this doesn't happen until the switch is completely
911 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
915 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
918 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
921 static inline int task_is_waking(struct task_struct *p)
923 return unlikely(p->state == TASK_WAKING);
927 * __task_rq_lock - lock the runqueue a given task resides on.
928 * Must be called interrupts disabled.
930 static inline struct rq *__task_rq_lock(struct task_struct *p)
937 raw_spin_lock(&rq->lock);
938 if (likely(rq == task_rq(p)))
940 raw_spin_unlock(&rq->lock);
945 * task_rq_lock - lock the runqueue a given task resides on and disable
946 * interrupts. Note the ordering: we can safely lookup the task_rq without
947 * explicitly disabling preemption.
949 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
955 local_irq_save(*flags);
957 raw_spin_lock(&rq->lock);
958 if (likely(rq == task_rq(p)))
960 raw_spin_unlock_irqrestore(&rq->lock, *flags);
964 void task_rq_unlock_wait(struct task_struct *p)
966 struct rq *rq = task_rq(p);
968 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
969 raw_spin_unlock_wait(&rq->lock);
972 static void __task_rq_unlock(struct rq *rq)
975 raw_spin_unlock(&rq->lock);
978 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
981 raw_spin_unlock_irqrestore(&rq->lock, *flags);
985 * this_rq_lock - lock this runqueue and disable interrupts.
987 static struct rq *this_rq_lock(void)
994 raw_spin_lock(&rq->lock);
999 #ifdef CONFIG_SCHED_HRTICK
1001 * Use HR-timers to deliver accurate preemption points.
1003 * Its all a bit involved since we cannot program an hrt while holding the
1004 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1007 * When we get rescheduled we reprogram the hrtick_timer outside of the
1013 * - enabled by features
1014 * - hrtimer is actually high res
1016 static inline int hrtick_enabled(struct rq *rq)
1018 if (!sched_feat(HRTICK))
1020 if (!cpu_active(cpu_of(rq)))
1022 return hrtimer_is_hres_active(&rq->hrtick_timer);
1025 static void hrtick_clear(struct rq *rq)
1027 if (hrtimer_active(&rq->hrtick_timer))
1028 hrtimer_cancel(&rq->hrtick_timer);
1032 * High-resolution timer tick.
1033 * Runs from hardirq context with interrupts disabled.
1035 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1037 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1039 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1041 raw_spin_lock(&rq->lock);
1042 update_rq_clock(rq);
1043 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1044 raw_spin_unlock(&rq->lock);
1046 return HRTIMER_NORESTART;
1051 * called from hardirq (IPI) context
1053 static void __hrtick_start(void *arg)
1055 struct rq *rq = arg;
1057 raw_spin_lock(&rq->lock);
1058 hrtimer_restart(&rq->hrtick_timer);
1059 rq->hrtick_csd_pending = 0;
1060 raw_spin_unlock(&rq->lock);
1064 * Called to set the hrtick timer state.
1066 * called with rq->lock held and irqs disabled
1068 static void hrtick_start(struct rq *rq, u64 delay)
1070 struct hrtimer *timer = &rq->hrtick_timer;
1071 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1073 hrtimer_set_expires(timer, time);
1075 if (rq == this_rq()) {
1076 hrtimer_restart(timer);
1077 } else if (!rq->hrtick_csd_pending) {
1078 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1079 rq->hrtick_csd_pending = 1;
1084 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1086 int cpu = (int)(long)hcpu;
1089 case CPU_UP_CANCELED:
1090 case CPU_UP_CANCELED_FROZEN:
1091 case CPU_DOWN_PREPARE:
1092 case CPU_DOWN_PREPARE_FROZEN:
1094 case CPU_DEAD_FROZEN:
1095 hrtick_clear(cpu_rq(cpu));
1102 static __init void init_hrtick(void)
1104 hotcpu_notifier(hotplug_hrtick, 0);
1108 * Called to set the hrtick timer state.
1110 * called with rq->lock held and irqs disabled
1112 static void hrtick_start(struct rq *rq, u64 delay)
1114 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1115 HRTIMER_MODE_REL_PINNED, 0);
1118 static inline void init_hrtick(void)
1121 #endif /* CONFIG_SMP */
1123 static void init_rq_hrtick(struct rq *rq)
1126 rq->hrtick_csd_pending = 0;
1128 rq->hrtick_csd.flags = 0;
1129 rq->hrtick_csd.func = __hrtick_start;
1130 rq->hrtick_csd.info = rq;
1133 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1134 rq->hrtick_timer.function = hrtick;
1136 #else /* CONFIG_SCHED_HRTICK */
1137 static inline void hrtick_clear(struct rq *rq)
1141 static inline void init_rq_hrtick(struct rq *rq)
1145 static inline void init_hrtick(void)
1148 #endif /* CONFIG_SCHED_HRTICK */
1151 * resched_task - mark a task 'to be rescheduled now'.
1153 * On UP this means the setting of the need_resched flag, on SMP it
1154 * might also involve a cross-CPU call to trigger the scheduler on
1159 #ifndef tsk_is_polling
1160 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1163 static void resched_task(struct task_struct *p)
1167 assert_raw_spin_locked(&task_rq(p)->lock);
1169 if (test_tsk_need_resched(p))
1172 set_tsk_need_resched(p);
1175 if (cpu == smp_processor_id())
1178 /* NEED_RESCHED must be visible before we test polling */
1180 if (!tsk_is_polling(p))
1181 smp_send_reschedule(cpu);
1184 static void resched_cpu(int cpu)
1186 struct rq *rq = cpu_rq(cpu);
1187 unsigned long flags;
1189 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1191 resched_task(cpu_curr(cpu));
1192 raw_spin_unlock_irqrestore(&rq->lock, flags);
1197 * When add_timer_on() enqueues a timer into the timer wheel of an
1198 * idle CPU then this timer might expire before the next timer event
1199 * which is scheduled to wake up that CPU. In case of a completely
1200 * idle system the next event might even be infinite time into the
1201 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1202 * leaves the inner idle loop so the newly added timer is taken into
1203 * account when the CPU goes back to idle and evaluates the timer
1204 * wheel for the next timer event.
1206 void wake_up_idle_cpu(int cpu)
1208 struct rq *rq = cpu_rq(cpu);
1210 if (cpu == smp_processor_id())
1214 * This is safe, as this function is called with the timer
1215 * wheel base lock of (cpu) held. When the CPU is on the way
1216 * to idle and has not yet set rq->curr to idle then it will
1217 * be serialized on the timer wheel base lock and take the new
1218 * timer into account automatically.
1220 if (rq->curr != rq->idle)
1224 * We can set TIF_RESCHED on the idle task of the other CPU
1225 * lockless. The worst case is that the other CPU runs the
1226 * idle task through an additional NOOP schedule()
1228 set_tsk_need_resched(rq->idle);
1230 /* NEED_RESCHED must be visible before we test polling */
1232 if (!tsk_is_polling(rq->idle))
1233 smp_send_reschedule(cpu);
1236 int nohz_ratelimit(int cpu)
1238 struct rq *rq = cpu_rq(cpu);
1239 u64 diff = rq->clock - rq->nohz_stamp;
1241 rq->nohz_stamp = rq->clock;
1243 return diff < (NSEC_PER_SEC / HZ) >> 1;
1246 #endif /* CONFIG_NO_HZ */
1248 static u64 sched_avg_period(void)
1250 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1253 static void sched_avg_update(struct rq *rq)
1255 s64 period = sched_avg_period();
1257 while ((s64)(rq->clock - rq->age_stamp) > period) {
1258 rq->age_stamp += period;
1263 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1265 rq->rt_avg += rt_delta;
1266 sched_avg_update(rq);
1269 #else /* !CONFIG_SMP */
1270 static void resched_task(struct task_struct *p)
1272 assert_raw_spin_locked(&task_rq(p)->lock);
1273 set_tsk_need_resched(p);
1276 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1279 #endif /* CONFIG_SMP */
1281 #if BITS_PER_LONG == 32
1282 # define WMULT_CONST (~0UL)
1284 # define WMULT_CONST (1UL << 32)
1287 #define WMULT_SHIFT 32
1290 * Shift right and round:
1292 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1295 * delta *= weight / lw
1297 static unsigned long
1298 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1299 struct load_weight *lw)
1303 if (!lw->inv_weight) {
1304 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1307 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1311 tmp = (u64)delta_exec * weight;
1313 * Check whether we'd overflow the 64-bit multiplication:
1315 if (unlikely(tmp > WMULT_CONST))
1316 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1319 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1321 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1324 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1330 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1337 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1338 * of tasks with abnormal "nice" values across CPUs the contribution that
1339 * each task makes to its run queue's load is weighted according to its
1340 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1341 * scaled version of the new time slice allocation that they receive on time
1345 #define WEIGHT_IDLEPRIO 3
1346 #define WMULT_IDLEPRIO 1431655765
1349 * Nice levels are multiplicative, with a gentle 10% change for every
1350 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1351 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1352 * that remained on nice 0.
1354 * The "10% effect" is relative and cumulative: from _any_ nice level,
1355 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1356 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1357 * If a task goes up by ~10% and another task goes down by ~10% then
1358 * the relative distance between them is ~25%.)
1360 static const int prio_to_weight[40] = {
1361 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1362 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1363 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1364 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1365 /* 0 */ 1024, 820, 655, 526, 423,
1366 /* 5 */ 335, 272, 215, 172, 137,
1367 /* 10 */ 110, 87, 70, 56, 45,
1368 /* 15 */ 36, 29, 23, 18, 15,
1372 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1374 * In cases where the weight does not change often, we can use the
1375 * precalculated inverse to speed up arithmetics by turning divisions
1376 * into multiplications:
1378 static const u32 prio_to_wmult[40] = {
1379 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1380 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1381 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1382 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1383 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1384 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1385 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1386 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1389 /* Time spent by the tasks of the cpu accounting group executing in ... */
1390 enum cpuacct_stat_index {
1391 CPUACCT_STAT_USER, /* ... user mode */
1392 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1394 CPUACCT_STAT_NSTATS,
1397 #ifdef CONFIG_CGROUP_CPUACCT
1398 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1399 static void cpuacct_update_stats(struct task_struct *tsk,
1400 enum cpuacct_stat_index idx, cputime_t val);
1402 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1403 static inline void cpuacct_update_stats(struct task_struct *tsk,
1404 enum cpuacct_stat_index idx, cputime_t val) {}
1407 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1409 update_load_add(&rq->load, load);
1412 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1414 update_load_sub(&rq->load, load);
1417 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1418 typedef int (*tg_visitor)(struct task_group *, void *);
1421 * Iterate the full tree, calling @down when first entering a node and @up when
1422 * leaving it for the final time.
1424 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1426 struct task_group *parent, *child;
1430 parent = &root_task_group;
1432 ret = (*down)(parent, data);
1435 list_for_each_entry_rcu(child, &parent->children, siblings) {
1442 ret = (*up)(parent, data);
1447 parent = parent->parent;
1456 static int tg_nop(struct task_group *tg, void *data)
1463 /* Used instead of source_load when we know the type == 0 */
1464 static unsigned long weighted_cpuload(const int cpu)
1466 return cpu_rq(cpu)->load.weight;
1470 * Return a low guess at the load of a migration-source cpu weighted
1471 * according to the scheduling class and "nice" value.
1473 * We want to under-estimate the load of migration sources, to
1474 * balance conservatively.
1476 static unsigned long source_load(int cpu, int type)
1478 struct rq *rq = cpu_rq(cpu);
1479 unsigned long total = weighted_cpuload(cpu);
1481 if (type == 0 || !sched_feat(LB_BIAS))
1484 return min(rq->cpu_load[type-1], total);
1488 * Return a high guess at the load of a migration-target cpu weighted
1489 * according to the scheduling class and "nice" value.
1491 static unsigned long target_load(int cpu, int type)
1493 struct rq *rq = cpu_rq(cpu);
1494 unsigned long total = weighted_cpuload(cpu);
1496 if (type == 0 || !sched_feat(LB_BIAS))
1499 return max(rq->cpu_load[type-1], total);
1502 static struct sched_group *group_of(int cpu)
1504 struct sched_domain *sd = rcu_dereference_sched(cpu_rq(cpu)->sd);
1512 static unsigned long power_of(int cpu)
1514 struct sched_group *group = group_of(cpu);
1517 return SCHED_LOAD_SCALE;
1519 return group->cpu_power;
1522 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1524 static unsigned long cpu_avg_load_per_task(int cpu)
1526 struct rq *rq = cpu_rq(cpu);
1527 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1530 rq->avg_load_per_task = rq->load.weight / nr_running;
1532 rq->avg_load_per_task = 0;
1534 return rq->avg_load_per_task;
1537 #ifdef CONFIG_FAIR_GROUP_SCHED
1539 static __read_mostly unsigned long __percpu *update_shares_data;
1541 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1544 * Calculate and set the cpu's group shares.
1546 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1547 unsigned long sd_shares,
1548 unsigned long sd_rq_weight,
1549 unsigned long *usd_rq_weight)
1551 unsigned long shares, rq_weight;
1554 rq_weight = usd_rq_weight[cpu];
1557 rq_weight = NICE_0_LOAD;
1561 * \Sum_j shares_j * rq_weight_i
1562 * shares_i = -----------------------------
1563 * \Sum_j rq_weight_j
1565 shares = (sd_shares * rq_weight) / sd_rq_weight;
1566 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1568 if (abs(shares - tg->se[cpu]->load.weight) >
1569 sysctl_sched_shares_thresh) {
1570 struct rq *rq = cpu_rq(cpu);
1571 unsigned long flags;
1573 raw_spin_lock_irqsave(&rq->lock, flags);
1574 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1575 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1576 __set_se_shares(tg->se[cpu], shares);
1577 raw_spin_unlock_irqrestore(&rq->lock, flags);
1582 * Re-compute the task group their per cpu shares over the given domain.
1583 * This needs to be done in a bottom-up fashion because the rq weight of a
1584 * parent group depends on the shares of its child groups.
1586 static int tg_shares_up(struct task_group *tg, void *data)
1588 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1589 unsigned long *usd_rq_weight;
1590 struct sched_domain *sd = data;
1591 unsigned long flags;
1597 local_irq_save(flags);
1598 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1600 for_each_cpu(i, sched_domain_span(sd)) {
1601 weight = tg->cfs_rq[i]->load.weight;
1602 usd_rq_weight[i] = weight;
1604 rq_weight += weight;
1606 * If there are currently no tasks on the cpu pretend there
1607 * is one of average load so that when a new task gets to
1608 * run here it will not get delayed by group starvation.
1611 weight = NICE_0_LOAD;
1613 sum_weight += weight;
1614 shares += tg->cfs_rq[i]->shares;
1618 rq_weight = sum_weight;
1620 if ((!shares && rq_weight) || shares > tg->shares)
1621 shares = tg->shares;
1623 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1624 shares = tg->shares;
1626 for_each_cpu(i, sched_domain_span(sd))
1627 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1629 local_irq_restore(flags);
1635 * Compute the cpu's hierarchical load factor for each task group.
1636 * This needs to be done in a top-down fashion because the load of a child
1637 * group is a fraction of its parents load.
1639 static int tg_load_down(struct task_group *tg, void *data)
1642 long cpu = (long)data;
1645 load = cpu_rq(cpu)->load.weight;
1647 load = tg->parent->cfs_rq[cpu]->h_load;
1648 load *= tg->cfs_rq[cpu]->shares;
1649 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1652 tg->cfs_rq[cpu]->h_load = load;
1657 static void update_shares(struct sched_domain *sd)
1662 if (root_task_group_empty())
1665 now = cpu_clock(raw_smp_processor_id());
1666 elapsed = now - sd->last_update;
1668 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1669 sd->last_update = now;
1670 walk_tg_tree(tg_nop, tg_shares_up, sd);
1674 static void update_h_load(long cpu)
1676 if (root_task_group_empty())
1679 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1684 static inline void update_shares(struct sched_domain *sd)
1690 #ifdef CONFIG_PREEMPT
1692 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1695 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1696 * way at the expense of forcing extra atomic operations in all
1697 * invocations. This assures that the double_lock is acquired using the
1698 * same underlying policy as the spinlock_t on this architecture, which
1699 * reduces latency compared to the unfair variant below. However, it
1700 * also adds more overhead and therefore may reduce throughput.
1702 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1703 __releases(this_rq->lock)
1704 __acquires(busiest->lock)
1705 __acquires(this_rq->lock)
1707 raw_spin_unlock(&this_rq->lock);
1708 double_rq_lock(this_rq, busiest);
1715 * Unfair double_lock_balance: Optimizes throughput at the expense of
1716 * latency by eliminating extra atomic operations when the locks are
1717 * already in proper order on entry. This favors lower cpu-ids and will
1718 * grant the double lock to lower cpus over higher ids under contention,
1719 * regardless of entry order into the function.
1721 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1722 __releases(this_rq->lock)
1723 __acquires(busiest->lock)
1724 __acquires(this_rq->lock)
1728 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1729 if (busiest < this_rq) {
1730 raw_spin_unlock(&this_rq->lock);
1731 raw_spin_lock(&busiest->lock);
1732 raw_spin_lock_nested(&this_rq->lock,
1733 SINGLE_DEPTH_NESTING);
1736 raw_spin_lock_nested(&busiest->lock,
1737 SINGLE_DEPTH_NESTING);
1742 #endif /* CONFIG_PREEMPT */
1745 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1747 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1749 if (unlikely(!irqs_disabled())) {
1750 /* printk() doesn't work good under rq->lock */
1751 raw_spin_unlock(&this_rq->lock);
1755 return _double_lock_balance(this_rq, busiest);
1758 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1759 __releases(busiest->lock)
1761 raw_spin_unlock(&busiest->lock);
1762 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1766 * double_rq_lock - safely lock two runqueues
1768 * Note this does not disable interrupts like task_rq_lock,
1769 * you need to do so manually before calling.
1771 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1772 __acquires(rq1->lock)
1773 __acquires(rq2->lock)
1775 BUG_ON(!irqs_disabled());
1777 raw_spin_lock(&rq1->lock);
1778 __acquire(rq2->lock); /* Fake it out ;) */
1781 raw_spin_lock(&rq1->lock);
1782 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1784 raw_spin_lock(&rq2->lock);
1785 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1791 * double_rq_unlock - safely unlock two runqueues
1793 * Note this does not restore interrupts like task_rq_unlock,
1794 * you need to do so manually after calling.
1796 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1797 __releases(rq1->lock)
1798 __releases(rq2->lock)
1800 raw_spin_unlock(&rq1->lock);
1802 raw_spin_unlock(&rq2->lock);
1804 __release(rq2->lock);
1809 #ifdef CONFIG_FAIR_GROUP_SCHED
1810 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1813 cfs_rq->shares = shares;
1818 static void calc_load_account_active(struct rq *this_rq);
1819 static void update_sysctl(void);
1820 static int get_update_sysctl_factor(void);
1822 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1824 set_task_rq(p, cpu);
1827 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1828 * successfuly executed on another CPU. We must ensure that updates of
1829 * per-task data have been completed by this moment.
1832 task_thread_info(p)->cpu = cpu;
1836 static const struct sched_class rt_sched_class;
1838 #define sched_class_highest (&rt_sched_class)
1839 #define for_each_class(class) \
1840 for (class = sched_class_highest; class; class = class->next)
1842 #include "sched_stats.h"
1844 static void inc_nr_running(struct rq *rq)
1849 static void dec_nr_running(struct rq *rq)
1854 static void set_load_weight(struct task_struct *p)
1856 if (task_has_rt_policy(p)) {
1857 p->se.load.weight = prio_to_weight[0] * 2;
1858 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1863 * SCHED_IDLE tasks get minimal weight:
1865 if (p->policy == SCHED_IDLE) {
1866 p->se.load.weight = WEIGHT_IDLEPRIO;
1867 p->se.load.inv_weight = WMULT_IDLEPRIO;
1871 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1872 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1875 static void update_avg(u64 *avg, u64 sample)
1877 s64 diff = sample - *avg;
1881 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1883 update_rq_clock(rq);
1884 sched_info_queued(p);
1885 p->sched_class->enqueue_task(rq, p, flags);
1889 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1891 update_rq_clock(rq);
1892 sched_info_dequeued(p);
1893 p->sched_class->dequeue_task(rq, p, flags);
1898 * activate_task - move a task to the runqueue.
1900 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1902 if (task_contributes_to_load(p))
1903 rq->nr_uninterruptible--;
1905 enqueue_task(rq, p, flags);
1910 * deactivate_task - remove a task from the runqueue.
1912 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1914 if (task_contributes_to_load(p))
1915 rq->nr_uninterruptible++;
1917 dequeue_task(rq, p, flags);
1921 #include "sched_idletask.c"
1922 #include "sched_fair.c"
1923 #include "sched_rt.c"
1924 #ifdef CONFIG_SCHED_DEBUG
1925 # include "sched_debug.c"
1929 * __normal_prio - return the priority that is based on the static prio
1931 static inline int __normal_prio(struct task_struct *p)
1933 return p->static_prio;
1937 * Calculate the expected normal priority: i.e. priority
1938 * without taking RT-inheritance into account. Might be
1939 * boosted by interactivity modifiers. Changes upon fork,
1940 * setprio syscalls, and whenever the interactivity
1941 * estimator recalculates.
1943 static inline int normal_prio(struct task_struct *p)
1947 if (task_has_rt_policy(p))
1948 prio = MAX_RT_PRIO-1 - p->rt_priority;
1950 prio = __normal_prio(p);
1955 * Calculate the current priority, i.e. the priority
1956 * taken into account by the scheduler. This value might
1957 * be boosted by RT tasks, or might be boosted by
1958 * interactivity modifiers. Will be RT if the task got
1959 * RT-boosted. If not then it returns p->normal_prio.
1961 static int effective_prio(struct task_struct *p)
1963 p->normal_prio = normal_prio(p);
1965 * If we are RT tasks or we were boosted to RT priority,
1966 * keep the priority unchanged. Otherwise, update priority
1967 * to the normal priority:
1969 if (!rt_prio(p->prio))
1970 return p->normal_prio;
1975 * task_curr - is this task currently executing on a CPU?
1976 * @p: the task in question.
1978 inline int task_curr(const struct task_struct *p)
1980 return cpu_curr(task_cpu(p)) == p;
1983 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1984 const struct sched_class *prev_class,
1985 int oldprio, int running)
1987 if (prev_class != p->sched_class) {
1988 if (prev_class->switched_from)
1989 prev_class->switched_from(rq, p, running);
1990 p->sched_class->switched_to(rq, p, running);
1992 p->sched_class->prio_changed(rq, p, oldprio, running);
1997 * Is this task likely cache-hot:
2000 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2004 if (p->sched_class != &fair_sched_class)
2008 * Buddy candidates are cache hot:
2010 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2011 (&p->se == cfs_rq_of(&p->se)->next ||
2012 &p->se == cfs_rq_of(&p->se)->last))
2015 if (sysctl_sched_migration_cost == -1)
2017 if (sysctl_sched_migration_cost == 0)
2020 delta = now - p->se.exec_start;
2022 return delta < (s64)sysctl_sched_migration_cost;
2025 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2027 #ifdef CONFIG_SCHED_DEBUG
2029 * We should never call set_task_cpu() on a blocked task,
2030 * ttwu() will sort out the placement.
2032 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2033 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2036 trace_sched_migrate_task(p, new_cpu);
2038 if (task_cpu(p) != new_cpu) {
2039 p->se.nr_migrations++;
2040 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2043 __set_task_cpu(p, new_cpu);
2046 struct migration_req {
2047 struct list_head list;
2049 struct task_struct *task;
2052 struct completion done;
2056 * The task's runqueue lock must be held.
2057 * Returns true if you have to wait for migration thread.
2060 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2062 struct rq *rq = task_rq(p);
2065 * If the task is not on a runqueue (and not running), then
2066 * the next wake-up will properly place the task.
2068 if (!p->se.on_rq && !task_running(rq, p))
2071 init_completion(&req->done);
2073 req->dest_cpu = dest_cpu;
2074 list_add(&req->list, &rq->migration_queue);
2080 * wait_task_context_switch - wait for a thread to complete at least one
2083 * @p must not be current.
2085 void wait_task_context_switch(struct task_struct *p)
2087 unsigned long nvcsw, nivcsw, flags;
2095 * The runqueue is assigned before the actual context
2096 * switch. We need to take the runqueue lock.
2098 * We could check initially without the lock but it is
2099 * very likely that we need to take the lock in every
2102 rq = task_rq_lock(p, &flags);
2103 running = task_running(rq, p);
2104 task_rq_unlock(rq, &flags);
2106 if (likely(!running))
2109 * The switch count is incremented before the actual
2110 * context switch. We thus wait for two switches to be
2111 * sure at least one completed.
2113 if ((p->nvcsw - nvcsw) > 1)
2115 if ((p->nivcsw - nivcsw) > 1)
2123 * wait_task_inactive - wait for a thread to unschedule.
2125 * If @match_state is nonzero, it's the @p->state value just checked and
2126 * not expected to change. If it changes, i.e. @p might have woken up,
2127 * then return zero. When we succeed in waiting for @p to be off its CPU,
2128 * we return a positive number (its total switch count). If a second call
2129 * a short while later returns the same number, the caller can be sure that
2130 * @p has remained unscheduled the whole time.
2132 * The caller must ensure that the task *will* unschedule sometime soon,
2133 * else this function might spin for a *long* time. This function can't
2134 * be called with interrupts off, or it may introduce deadlock with
2135 * smp_call_function() if an IPI is sent by the same process we are
2136 * waiting to become inactive.
2138 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2140 unsigned long flags;
2147 * We do the initial early heuristics without holding
2148 * any task-queue locks at all. We'll only try to get
2149 * the runqueue lock when things look like they will
2155 * If the task is actively running on another CPU
2156 * still, just relax and busy-wait without holding
2159 * NOTE! Since we don't hold any locks, it's not
2160 * even sure that "rq" stays as the right runqueue!
2161 * But we don't care, since "task_running()" will
2162 * return false if the runqueue has changed and p
2163 * is actually now running somewhere else!
2165 while (task_running(rq, p)) {
2166 if (match_state && unlikely(p->state != match_state))
2172 * Ok, time to look more closely! We need the rq
2173 * lock now, to be *sure*. If we're wrong, we'll
2174 * just go back and repeat.
2176 rq = task_rq_lock(p, &flags);
2177 trace_sched_wait_task(rq, p);
2178 running = task_running(rq, p);
2179 on_rq = p->se.on_rq;
2181 if (!match_state || p->state == match_state)
2182 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2183 task_rq_unlock(rq, &flags);
2186 * If it changed from the expected state, bail out now.
2188 if (unlikely(!ncsw))
2192 * Was it really running after all now that we
2193 * checked with the proper locks actually held?
2195 * Oops. Go back and try again..
2197 if (unlikely(running)) {
2203 * It's not enough that it's not actively running,
2204 * it must be off the runqueue _entirely_, and not
2207 * So if it was still runnable (but just not actively
2208 * running right now), it's preempted, and we should
2209 * yield - it could be a while.
2211 if (unlikely(on_rq)) {
2212 schedule_timeout_uninterruptible(1);
2217 * Ahh, all good. It wasn't running, and it wasn't
2218 * runnable, which means that it will never become
2219 * running in the future either. We're all done!
2228 * kick_process - kick a running thread to enter/exit the kernel
2229 * @p: the to-be-kicked thread
2231 * Cause a process which is running on another CPU to enter
2232 * kernel-mode, without any delay. (to get signals handled.)
2234 * NOTE: this function doesnt have to take the runqueue lock,
2235 * because all it wants to ensure is that the remote task enters
2236 * the kernel. If the IPI races and the task has been migrated
2237 * to another CPU then no harm is done and the purpose has been
2240 void kick_process(struct task_struct *p)
2246 if ((cpu != smp_processor_id()) && task_curr(p))
2247 smp_send_reschedule(cpu);
2250 EXPORT_SYMBOL_GPL(kick_process);
2251 #endif /* CONFIG_SMP */
2254 * task_oncpu_function_call - call a function on the cpu on which a task runs
2255 * @p: the task to evaluate
2256 * @func: the function to be called
2257 * @info: the function call argument
2259 * Calls the function @func when the task is currently running. This might
2260 * be on the current CPU, which just calls the function directly
2262 void task_oncpu_function_call(struct task_struct *p,
2263 void (*func) (void *info), void *info)
2270 smp_call_function_single(cpu, func, info, 1);
2276 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2278 static int select_fallback_rq(int cpu, struct task_struct *p)
2281 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2283 /* Look for allowed, online CPU in same node. */
2284 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2285 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2288 /* Any allowed, online CPU? */
2289 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2290 if (dest_cpu < nr_cpu_ids)
2293 /* No more Mr. Nice Guy. */
2294 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2295 dest_cpu = cpuset_cpus_allowed_fallback(p);
2297 * Don't tell them about moving exiting tasks or
2298 * kernel threads (both mm NULL), since they never
2301 if (p->mm && printk_ratelimit()) {
2302 printk(KERN_INFO "process %d (%s) no "
2303 "longer affine to cpu%d\n",
2304 task_pid_nr(p), p->comm, cpu);
2312 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2315 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2317 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2320 * In order not to call set_task_cpu() on a blocking task we need
2321 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2324 * Since this is common to all placement strategies, this lives here.
2326 * [ this allows ->select_task() to simply return task_cpu(p) and
2327 * not worry about this generic constraint ]
2329 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2331 cpu = select_fallback_rq(task_cpu(p), p);
2338 * try_to_wake_up - wake up a thread
2339 * @p: the to-be-woken-up thread
2340 * @state: the mask of task states that can be woken
2341 * @sync: do a synchronous wakeup?
2343 * Put it on the run-queue if it's not already there. The "current"
2344 * thread is always on the run-queue (except when the actual
2345 * re-schedule is in progress), and as such you're allowed to do
2346 * the simpler "current->state = TASK_RUNNING" to mark yourself
2347 * runnable without the overhead of this.
2349 * returns failure only if the task is already active.
2351 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2354 int cpu, orig_cpu, this_cpu, success = 0;
2355 unsigned long flags;
2356 unsigned long en_flags = ENQUEUE_WAKEUP;
2359 this_cpu = get_cpu();
2362 rq = task_rq_lock(p, &flags);
2363 if (!(p->state & state))
2373 if (unlikely(task_running(rq, p)))
2377 * In order to handle concurrent wakeups and release the rq->lock
2378 * we put the task in TASK_WAKING state.
2380 * First fix up the nr_uninterruptible count:
2382 if (task_contributes_to_load(p)) {
2383 if (likely(cpu_online(orig_cpu)))
2384 rq->nr_uninterruptible--;
2386 this_rq()->nr_uninterruptible--;
2388 p->state = TASK_WAKING;
2390 if (p->sched_class->task_waking) {
2391 p->sched_class->task_waking(rq, p);
2392 en_flags |= ENQUEUE_WAKING;
2395 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2396 if (cpu != orig_cpu)
2397 set_task_cpu(p, cpu);
2398 __task_rq_unlock(rq);
2401 raw_spin_lock(&rq->lock);
2404 * We migrated the task without holding either rq->lock, however
2405 * since the task is not on the task list itself, nobody else
2406 * will try and migrate the task, hence the rq should match the
2407 * cpu we just moved it to.
2409 WARN_ON(task_cpu(p) != cpu);
2410 WARN_ON(p->state != TASK_WAKING);
2412 #ifdef CONFIG_SCHEDSTATS
2413 schedstat_inc(rq, ttwu_count);
2414 if (cpu == this_cpu)
2415 schedstat_inc(rq, ttwu_local);
2417 struct sched_domain *sd;
2418 for_each_domain(this_cpu, sd) {
2419 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2420 schedstat_inc(sd, ttwu_wake_remote);
2425 #endif /* CONFIG_SCHEDSTATS */
2428 #endif /* CONFIG_SMP */
2429 schedstat_inc(p, se.statistics.nr_wakeups);
2430 if (wake_flags & WF_SYNC)
2431 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2432 if (orig_cpu != cpu)
2433 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2434 if (cpu == this_cpu)
2435 schedstat_inc(p, se.statistics.nr_wakeups_local);
2437 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2438 activate_task(rq, p, en_flags);
2442 trace_sched_wakeup(rq, p, success);
2443 check_preempt_curr(rq, p, wake_flags);
2445 p->state = TASK_RUNNING;
2447 if (p->sched_class->task_woken)
2448 p->sched_class->task_woken(rq, p);
2450 if (unlikely(rq->idle_stamp)) {
2451 u64 delta = rq->clock - rq->idle_stamp;
2452 u64 max = 2*sysctl_sched_migration_cost;
2457 update_avg(&rq->avg_idle, delta);
2462 task_rq_unlock(rq, &flags);
2469 * wake_up_process - Wake up a specific process
2470 * @p: The process to be woken up.
2472 * Attempt to wake up the nominated process and move it to the set of runnable
2473 * processes. Returns 1 if the process was woken up, 0 if it was already
2476 * It may be assumed that this function implies a write memory barrier before
2477 * changing the task state if and only if any tasks are woken up.
2479 int wake_up_process(struct task_struct *p)
2481 return try_to_wake_up(p, TASK_ALL, 0);
2483 EXPORT_SYMBOL(wake_up_process);
2485 int wake_up_state(struct task_struct *p, unsigned int state)
2487 return try_to_wake_up(p, state, 0);
2491 * Perform scheduler related setup for a newly forked process p.
2492 * p is forked by current.
2494 * __sched_fork() is basic setup used by init_idle() too:
2496 static void __sched_fork(struct task_struct *p)
2498 p->se.exec_start = 0;
2499 p->se.sum_exec_runtime = 0;
2500 p->se.prev_sum_exec_runtime = 0;
2501 p->se.nr_migrations = 0;
2503 #ifdef CONFIG_SCHEDSTATS
2504 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2507 INIT_LIST_HEAD(&p->rt.run_list);
2509 INIT_LIST_HEAD(&p->se.group_node);
2511 #ifdef CONFIG_PREEMPT_NOTIFIERS
2512 INIT_HLIST_HEAD(&p->preempt_notifiers);
2517 * fork()/clone()-time setup:
2519 void sched_fork(struct task_struct *p, int clone_flags)
2521 int cpu = get_cpu();
2525 * We mark the process as running here. This guarantees that
2526 * nobody will actually run it, and a signal or other external
2527 * event cannot wake it up and insert it on the runqueue either.
2529 p->state = TASK_RUNNING;
2532 * Revert to default priority/policy on fork if requested.
2534 if (unlikely(p->sched_reset_on_fork)) {
2535 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2536 p->policy = SCHED_NORMAL;
2537 p->normal_prio = p->static_prio;
2540 if (PRIO_TO_NICE(p->static_prio) < 0) {
2541 p->static_prio = NICE_TO_PRIO(0);
2542 p->normal_prio = p->static_prio;
2547 * We don't need the reset flag anymore after the fork. It has
2548 * fulfilled its duty:
2550 p->sched_reset_on_fork = 0;
2554 * Make sure we do not leak PI boosting priority to the child.
2556 p->prio = current->normal_prio;
2558 if (!rt_prio(p->prio))
2559 p->sched_class = &fair_sched_class;
2561 if (p->sched_class->task_fork)
2562 p->sched_class->task_fork(p);
2564 set_task_cpu(p, cpu);
2566 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2567 if (likely(sched_info_on()))
2568 memset(&p->sched_info, 0, sizeof(p->sched_info));
2570 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2573 #ifdef CONFIG_PREEMPT
2574 /* Want to start with kernel preemption disabled. */
2575 task_thread_info(p)->preempt_count = 1;
2577 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2583 * wake_up_new_task - wake up a newly created task for the first time.
2585 * This function will do some initial scheduler statistics housekeeping
2586 * that must be done for every newly created context, then puts the task
2587 * on the runqueue and wakes it.
2589 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2591 unsigned long flags;
2593 int cpu __maybe_unused = get_cpu();
2596 rq = task_rq_lock(p, &flags);
2597 p->state = TASK_WAKING;
2600 * Fork balancing, do it here and not earlier because:
2601 * - cpus_allowed can change in the fork path
2602 * - any previously selected cpu might disappear through hotplug
2604 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2605 * without people poking at ->cpus_allowed.
2607 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2608 set_task_cpu(p, cpu);
2610 p->state = TASK_RUNNING;
2611 task_rq_unlock(rq, &flags);
2614 rq = task_rq_lock(p, &flags);
2615 activate_task(rq, p, 0);
2616 trace_sched_wakeup_new(rq, p, 1);
2617 check_preempt_curr(rq, p, WF_FORK);
2619 if (p->sched_class->task_woken)
2620 p->sched_class->task_woken(rq, p);
2622 task_rq_unlock(rq, &flags);
2626 #ifdef CONFIG_PREEMPT_NOTIFIERS
2629 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2630 * @notifier: notifier struct to register
2632 void preempt_notifier_register(struct preempt_notifier *notifier)
2634 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2636 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2639 * preempt_notifier_unregister - no longer interested in preemption notifications
2640 * @notifier: notifier struct to unregister
2642 * This is safe to call from within a preemption notifier.
2644 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2646 hlist_del(¬ifier->link);
2648 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2650 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2652 struct preempt_notifier *notifier;
2653 struct hlist_node *node;
2655 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2656 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2660 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2661 struct task_struct *next)
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_out(notifier, next);
2670 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2672 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2677 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2678 struct task_struct *next)
2682 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2685 * prepare_task_switch - prepare to switch tasks
2686 * @rq: the runqueue preparing to switch
2687 * @prev: the current task that is being switched out
2688 * @next: the task we are going to switch to.
2690 * This is called with the rq lock held and interrupts off. It must
2691 * be paired with a subsequent finish_task_switch after the context
2694 * prepare_task_switch sets up locking and calls architecture specific
2698 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2699 struct task_struct *next)
2701 fire_sched_out_preempt_notifiers(prev, next);
2702 prepare_lock_switch(rq, next);
2703 prepare_arch_switch(next);
2707 * finish_task_switch - clean up after a task-switch
2708 * @rq: runqueue associated with task-switch
2709 * @prev: the thread we just switched away from.
2711 * finish_task_switch must be called after the context switch, paired
2712 * with a prepare_task_switch call before the context switch.
2713 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2714 * and do any other architecture-specific cleanup actions.
2716 * Note that we may have delayed dropping an mm in context_switch(). If
2717 * so, we finish that here outside of the runqueue lock. (Doing it
2718 * with the lock held can cause deadlocks; see schedule() for
2721 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2722 __releases(rq->lock)
2724 struct mm_struct *mm = rq->prev_mm;
2730 * A task struct has one reference for the use as "current".
2731 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2732 * schedule one last time. The schedule call will never return, and
2733 * the scheduled task must drop that reference.
2734 * The test for TASK_DEAD must occur while the runqueue locks are
2735 * still held, otherwise prev could be scheduled on another cpu, die
2736 * there before we look at prev->state, and then the reference would
2738 * Manfred Spraul <manfred@colorfullife.com>
2740 prev_state = prev->state;
2741 finish_arch_switch(prev);
2742 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2743 local_irq_disable();
2744 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2745 perf_event_task_sched_in(current);
2746 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2748 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2749 finish_lock_switch(rq, prev);
2751 fire_sched_in_preempt_notifiers(current);
2754 if (unlikely(prev_state == TASK_DEAD)) {
2756 * Remove function-return probe instances associated with this
2757 * task and put them back on the free list.
2759 kprobe_flush_task(prev);
2760 put_task_struct(prev);
2766 /* assumes rq->lock is held */
2767 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2769 if (prev->sched_class->pre_schedule)
2770 prev->sched_class->pre_schedule(rq, prev);
2773 /* rq->lock is NOT held, but preemption is disabled */
2774 static inline void post_schedule(struct rq *rq)
2776 if (rq->post_schedule) {
2777 unsigned long flags;
2779 raw_spin_lock_irqsave(&rq->lock, flags);
2780 if (rq->curr->sched_class->post_schedule)
2781 rq->curr->sched_class->post_schedule(rq);
2782 raw_spin_unlock_irqrestore(&rq->lock, flags);
2784 rq->post_schedule = 0;
2790 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2794 static inline void post_schedule(struct rq *rq)
2801 * schedule_tail - first thing a freshly forked thread must call.
2802 * @prev: the thread we just switched away from.
2804 asmlinkage void schedule_tail(struct task_struct *prev)
2805 __releases(rq->lock)
2807 struct rq *rq = this_rq();
2809 finish_task_switch(rq, prev);
2812 * FIXME: do we need to worry about rq being invalidated by the
2817 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2818 /* In this case, finish_task_switch does not reenable preemption */
2821 if (current->set_child_tid)
2822 put_user(task_pid_vnr(current), current->set_child_tid);
2826 * context_switch - switch to the new MM and the new
2827 * thread's register state.
2830 context_switch(struct rq *rq, struct task_struct *prev,
2831 struct task_struct *next)
2833 struct mm_struct *mm, *oldmm;
2835 prepare_task_switch(rq, prev, next);
2836 trace_sched_switch(rq, prev, next);
2838 oldmm = prev->active_mm;
2840 * For paravirt, this is coupled with an exit in switch_to to
2841 * combine the page table reload and the switch backend into
2844 arch_start_context_switch(prev);
2847 next->active_mm = oldmm;
2848 atomic_inc(&oldmm->mm_count);
2849 enter_lazy_tlb(oldmm, next);
2851 switch_mm(oldmm, mm, next);
2853 if (likely(!prev->mm)) {
2854 prev->active_mm = NULL;
2855 rq->prev_mm = oldmm;
2858 * Since the runqueue lock will be released by the next
2859 * task (which is an invalid locking op but in the case
2860 * of the scheduler it's an obvious special-case), so we
2861 * do an early lockdep release here:
2863 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2864 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2867 /* Here we just switch the register state and the stack. */
2868 switch_to(prev, next, prev);
2872 * this_rq must be evaluated again because prev may have moved
2873 * CPUs since it called schedule(), thus the 'rq' on its stack
2874 * frame will be invalid.
2876 finish_task_switch(this_rq(), prev);
2880 * nr_running, nr_uninterruptible and nr_context_switches:
2882 * externally visible scheduler statistics: current number of runnable
2883 * threads, current number of uninterruptible-sleeping threads, total
2884 * number of context switches performed since bootup.
2886 unsigned long nr_running(void)
2888 unsigned long i, sum = 0;
2890 for_each_online_cpu(i)
2891 sum += cpu_rq(i)->nr_running;
2896 unsigned long nr_uninterruptible(void)
2898 unsigned long i, sum = 0;
2900 for_each_possible_cpu(i)
2901 sum += cpu_rq(i)->nr_uninterruptible;
2904 * Since we read the counters lockless, it might be slightly
2905 * inaccurate. Do not allow it to go below zero though:
2907 if (unlikely((long)sum < 0))
2913 unsigned long long nr_context_switches(void)
2916 unsigned long long sum = 0;
2918 for_each_possible_cpu(i)
2919 sum += cpu_rq(i)->nr_switches;
2924 unsigned long nr_iowait(void)
2926 unsigned long i, sum = 0;
2928 for_each_possible_cpu(i)
2929 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2934 unsigned long nr_iowait_cpu(void)
2936 struct rq *this = this_rq();
2937 return atomic_read(&this->nr_iowait);
2940 unsigned long this_cpu_load(void)
2942 struct rq *this = this_rq();
2943 return this->cpu_load[0];
2947 /* Variables and functions for calc_load */
2948 static atomic_long_t calc_load_tasks;
2949 static unsigned long calc_load_update;
2950 unsigned long avenrun[3];
2951 EXPORT_SYMBOL(avenrun);
2954 * get_avenrun - get the load average array
2955 * @loads: pointer to dest load array
2956 * @offset: offset to add
2957 * @shift: shift count to shift the result left
2959 * These values are estimates at best, so no need for locking.
2961 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2963 loads[0] = (avenrun[0] + offset) << shift;
2964 loads[1] = (avenrun[1] + offset) << shift;
2965 loads[2] = (avenrun[2] + offset) << shift;
2968 static unsigned long
2969 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2972 load += active * (FIXED_1 - exp);
2973 return load >> FSHIFT;
2977 * calc_load - update the avenrun load estimates 10 ticks after the
2978 * CPUs have updated calc_load_tasks.
2980 void calc_global_load(void)
2982 unsigned long upd = calc_load_update + 10;
2985 if (time_before(jiffies, upd))
2988 active = atomic_long_read(&calc_load_tasks);
2989 active = active > 0 ? active * FIXED_1 : 0;
2991 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2992 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2993 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2995 calc_load_update += LOAD_FREQ;
2999 * Either called from update_cpu_load() or from a cpu going idle
3001 static void calc_load_account_active(struct rq *this_rq)
3003 long nr_active, delta;
3005 nr_active = this_rq->nr_running;
3006 nr_active += (long) this_rq->nr_uninterruptible;
3008 if (nr_active != this_rq->calc_load_active) {
3009 delta = nr_active - this_rq->calc_load_active;
3010 this_rq->calc_load_active = nr_active;
3011 atomic_long_add(delta, &calc_load_tasks);
3016 * Update rq->cpu_load[] statistics. This function is usually called every
3017 * scheduler tick (TICK_NSEC).
3019 static void update_cpu_load(struct rq *this_rq)
3021 unsigned long this_load = this_rq->load.weight;
3024 this_rq->nr_load_updates++;
3026 /* Update our load: */
3027 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3028 unsigned long old_load, new_load;
3030 /* scale is effectively 1 << i now, and >> i divides by scale */
3032 old_load = this_rq->cpu_load[i];
3033 new_load = this_load;
3035 * Round up the averaging division if load is increasing. This
3036 * prevents us from getting stuck on 9 if the load is 10, for
3039 if (new_load > old_load)
3040 new_load += scale-1;
3041 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3044 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3045 this_rq->calc_load_update += LOAD_FREQ;
3046 calc_load_account_active(this_rq);
3053 * sched_exec - execve() is a valuable balancing opportunity, because at
3054 * this point the task has the smallest effective memory and cache footprint.
3056 void sched_exec(void)
3058 struct task_struct *p = current;
3059 struct migration_req req;
3060 unsigned long flags;
3064 rq = task_rq_lock(p, &flags);
3065 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3066 if (dest_cpu == smp_processor_id())
3070 * select_task_rq() can race against ->cpus_allowed
3072 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3073 likely(cpu_active(dest_cpu)) &&
3074 migrate_task(p, dest_cpu, &req)) {
3075 /* Need to wait for migration thread (might exit: take ref). */
3076 struct task_struct *mt = rq->migration_thread;
3078 get_task_struct(mt);
3079 task_rq_unlock(rq, &flags);
3080 wake_up_process(mt);
3081 put_task_struct(mt);
3082 wait_for_completion(&req.done);
3087 task_rq_unlock(rq, &flags);
3092 DEFINE_PER_CPU(struct kernel_stat, kstat);
3094 EXPORT_PER_CPU_SYMBOL(kstat);
3097 * Return any ns on the sched_clock that have not yet been accounted in
3098 * @p in case that task is currently running.
3100 * Called with task_rq_lock() held on @rq.
3102 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3106 if (task_current(rq, p)) {
3107 update_rq_clock(rq);
3108 ns = rq->clock - p->se.exec_start;
3116 unsigned long long task_delta_exec(struct task_struct *p)
3118 unsigned long flags;
3122 rq = task_rq_lock(p, &flags);
3123 ns = do_task_delta_exec(p, rq);
3124 task_rq_unlock(rq, &flags);
3130 * Return accounted runtime for the task.
3131 * In case the task is currently running, return the runtime plus current's
3132 * pending runtime that have not been accounted yet.
3134 unsigned long long task_sched_runtime(struct task_struct *p)
3136 unsigned long flags;
3140 rq = task_rq_lock(p, &flags);
3141 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3142 task_rq_unlock(rq, &flags);
3148 * Return sum_exec_runtime for the thread group.
3149 * In case the task is currently running, return the sum plus current's
3150 * pending runtime that have not been accounted yet.
3152 * Note that the thread group might have other running tasks as well,
3153 * so the return value not includes other pending runtime that other
3154 * running tasks might have.
3156 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3158 struct task_cputime totals;
3159 unsigned long flags;
3163 rq = task_rq_lock(p, &flags);
3164 thread_group_cputime(p, &totals);
3165 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3166 task_rq_unlock(rq, &flags);
3172 * Account user cpu time to a process.
3173 * @p: the process that the cpu time gets accounted to
3174 * @cputime: the cpu time spent in user space since the last update
3175 * @cputime_scaled: cputime scaled by cpu frequency
3177 void account_user_time(struct task_struct *p, cputime_t cputime,
3178 cputime_t cputime_scaled)
3180 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3183 /* Add user time to process. */
3184 p->utime = cputime_add(p->utime, cputime);
3185 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3186 account_group_user_time(p, cputime);
3188 /* Add user time to cpustat. */
3189 tmp = cputime_to_cputime64(cputime);
3190 if (TASK_NICE(p) > 0)
3191 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3193 cpustat->user = cputime64_add(cpustat->user, tmp);
3195 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3196 /* Account for user time used */
3197 acct_update_integrals(p);
3201 * Account guest cpu time to a process.
3202 * @p: the process that the cpu time gets accounted to
3203 * @cputime: the cpu time spent in virtual machine since the last update
3204 * @cputime_scaled: cputime scaled by cpu frequency
3206 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3207 cputime_t cputime_scaled)
3210 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3212 tmp = cputime_to_cputime64(cputime);
3214 /* Add guest time to process. */
3215 p->utime = cputime_add(p->utime, cputime);
3216 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3217 account_group_user_time(p, cputime);
3218 p->gtime = cputime_add(p->gtime, cputime);
3220 /* Add guest time to cpustat. */
3221 if (TASK_NICE(p) > 0) {
3222 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3223 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3225 cpustat->user = cputime64_add(cpustat->user, tmp);
3226 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3231 * Account system cpu time to a process.
3232 * @p: the process that the cpu time gets accounted to
3233 * @hardirq_offset: the offset to subtract from hardirq_count()
3234 * @cputime: the cpu time spent in kernel space since the last update
3235 * @cputime_scaled: cputime scaled by cpu frequency
3237 void account_system_time(struct task_struct *p, int hardirq_offset,
3238 cputime_t cputime, cputime_t cputime_scaled)
3240 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3243 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3244 account_guest_time(p, cputime, cputime_scaled);
3248 /* Add system time to process. */
3249 p->stime = cputime_add(p->stime, cputime);
3250 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3251 account_group_system_time(p, cputime);
3253 /* Add system time to cpustat. */
3254 tmp = cputime_to_cputime64(cputime);
3255 if (hardirq_count() - hardirq_offset)
3256 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3257 else if (softirq_count())
3258 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3260 cpustat->system = cputime64_add(cpustat->system, tmp);
3262 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3264 /* Account for system time used */
3265 acct_update_integrals(p);
3269 * Account for involuntary wait time.
3270 * @steal: the cpu time spent in involuntary wait
3272 void account_steal_time(cputime_t cputime)
3274 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3275 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3277 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3281 * Account for idle time.
3282 * @cputime: the cpu time spent in idle wait
3284 void account_idle_time(cputime_t cputime)
3286 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3287 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3288 struct rq *rq = this_rq();
3290 if (atomic_read(&rq->nr_iowait) > 0)
3291 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3293 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3296 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3299 * Account a single tick of cpu time.
3300 * @p: the process that the cpu time gets accounted to
3301 * @user_tick: indicates if the tick is a user or a system tick
3303 void account_process_tick(struct task_struct *p, int user_tick)
3305 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3306 struct rq *rq = this_rq();
3309 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3310 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3311 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3314 account_idle_time(cputime_one_jiffy);
3318 * Account multiple ticks of steal time.
3319 * @p: the process from which the cpu time has been stolen
3320 * @ticks: number of stolen ticks
3322 void account_steal_ticks(unsigned long ticks)
3324 account_steal_time(jiffies_to_cputime(ticks));
3328 * Account multiple ticks of idle time.
3329 * @ticks: number of stolen ticks
3331 void account_idle_ticks(unsigned long ticks)
3333 account_idle_time(jiffies_to_cputime(ticks));
3339 * Use precise platform statistics if available:
3341 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3342 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3348 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3350 struct task_cputime cputime;
3352 thread_group_cputime(p, &cputime);
3354 *ut = cputime.utime;
3355 *st = cputime.stime;
3359 #ifndef nsecs_to_cputime
3360 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3363 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3365 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3368 * Use CFS's precise accounting:
3370 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3375 temp = (u64)(rtime * utime);
3376 do_div(temp, total);
3377 utime = (cputime_t)temp;
3382 * Compare with previous values, to keep monotonicity:
3384 p->prev_utime = max(p->prev_utime, utime);
3385 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3387 *ut = p->prev_utime;
3388 *st = p->prev_stime;
3392 * Must be called with siglock held.
3394 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3396 struct signal_struct *sig = p->signal;
3397 struct task_cputime cputime;
3398 cputime_t rtime, utime, total;
3400 thread_group_cputime(p, &cputime);
3402 total = cputime_add(cputime.utime, cputime.stime);
3403 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3408 temp = (u64)(rtime * cputime.utime);
3409 do_div(temp, total);
3410 utime = (cputime_t)temp;
3414 sig->prev_utime = max(sig->prev_utime, utime);
3415 sig->prev_stime = max(sig->prev_stime,
3416 cputime_sub(rtime, sig->prev_utime));
3418 *ut = sig->prev_utime;
3419 *st = sig->prev_stime;
3424 * This function gets called by the timer code, with HZ frequency.
3425 * We call it with interrupts disabled.
3427 * It also gets called by the fork code, when changing the parent's
3430 void scheduler_tick(void)
3432 int cpu = smp_processor_id();
3433 struct rq *rq = cpu_rq(cpu);
3434 struct task_struct *curr = rq->curr;
3438 raw_spin_lock(&rq->lock);
3439 update_rq_clock(rq);
3440 update_cpu_load(rq);
3441 curr->sched_class->task_tick(rq, curr, 0);
3442 raw_spin_unlock(&rq->lock);
3444 perf_event_task_tick(curr);
3447 rq->idle_at_tick = idle_cpu(cpu);
3448 trigger_load_balance(rq, cpu);
3452 notrace unsigned long get_parent_ip(unsigned long addr)
3454 if (in_lock_functions(addr)) {
3455 addr = CALLER_ADDR2;
3456 if (in_lock_functions(addr))
3457 addr = CALLER_ADDR3;
3462 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3463 defined(CONFIG_PREEMPT_TRACER))
3465 void __kprobes add_preempt_count(int val)
3467 #ifdef CONFIG_DEBUG_PREEMPT
3471 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3474 preempt_count() += val;
3475 #ifdef CONFIG_DEBUG_PREEMPT
3477 * Spinlock count overflowing soon?
3479 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3482 if (preempt_count() == val)
3483 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3485 EXPORT_SYMBOL(add_preempt_count);
3487 void __kprobes sub_preempt_count(int val)
3489 #ifdef CONFIG_DEBUG_PREEMPT
3493 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3496 * Is the spinlock portion underflowing?
3498 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3499 !(preempt_count() & PREEMPT_MASK)))
3503 if (preempt_count() == val)
3504 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3505 preempt_count() -= val;
3507 EXPORT_SYMBOL(sub_preempt_count);
3512 * Print scheduling while atomic bug:
3514 static noinline void __schedule_bug(struct task_struct *prev)
3516 struct pt_regs *regs = get_irq_regs();
3518 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3519 prev->comm, prev->pid, preempt_count());
3521 debug_show_held_locks(prev);
3523 if (irqs_disabled())
3524 print_irqtrace_events(prev);
3533 * Various schedule()-time debugging checks and statistics:
3535 static inline void schedule_debug(struct task_struct *prev)
3538 * Test if we are atomic. Since do_exit() needs to call into
3539 * schedule() atomically, we ignore that path for now.
3540 * Otherwise, whine if we are scheduling when we should not be.
3542 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3543 __schedule_bug(prev);
3545 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3547 schedstat_inc(this_rq(), sched_count);
3548 #ifdef CONFIG_SCHEDSTATS
3549 if (unlikely(prev->lock_depth >= 0)) {
3550 schedstat_inc(this_rq(), bkl_count);
3551 schedstat_inc(prev, sched_info.bkl_count);
3556 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3559 update_rq_clock(rq);
3560 rq->skip_clock_update = 0;
3561 prev->sched_class->put_prev_task(rq, prev);
3565 * Pick up the highest-prio task:
3567 static inline struct task_struct *
3568 pick_next_task(struct rq *rq)
3570 const struct sched_class *class;
3571 struct task_struct *p;
3574 * Optimization: we know that if all tasks are in
3575 * the fair class we can call that function directly:
3577 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3578 p = fair_sched_class.pick_next_task(rq);
3583 class = sched_class_highest;
3585 p = class->pick_next_task(rq);
3589 * Will never be NULL as the idle class always
3590 * returns a non-NULL p:
3592 class = class->next;
3597 * schedule() is the main scheduler function.
3599 asmlinkage void __sched schedule(void)
3601 struct task_struct *prev, *next;
3602 unsigned long *switch_count;
3608 cpu = smp_processor_id();
3612 switch_count = &prev->nivcsw;
3614 release_kernel_lock(prev);
3615 need_resched_nonpreemptible:
3617 schedule_debug(prev);
3619 if (sched_feat(HRTICK))
3622 raw_spin_lock_irq(&rq->lock);
3623 clear_tsk_need_resched(prev);
3625 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3626 if (unlikely(signal_pending_state(prev->state, prev)))
3627 prev->state = TASK_RUNNING;
3629 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3630 switch_count = &prev->nvcsw;
3633 pre_schedule(rq, prev);
3635 if (unlikely(!rq->nr_running))
3636 idle_balance(cpu, rq);
3638 put_prev_task(rq, prev);
3639 next = pick_next_task(rq);
3641 if (likely(prev != next)) {
3642 sched_info_switch(prev, next);
3643 perf_event_task_sched_out(prev, next);
3649 context_switch(rq, prev, next); /* unlocks the rq */
3651 * the context switch might have flipped the stack from under
3652 * us, hence refresh the local variables.
3654 cpu = smp_processor_id();
3657 raw_spin_unlock_irq(&rq->lock);
3661 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3663 switch_count = &prev->nivcsw;
3664 goto need_resched_nonpreemptible;
3667 preempt_enable_no_resched();
3671 EXPORT_SYMBOL(schedule);
3673 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3675 * Look out! "owner" is an entirely speculative pointer
3676 * access and not reliable.
3678 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3683 if (!sched_feat(OWNER_SPIN))
3686 #ifdef CONFIG_DEBUG_PAGEALLOC
3688 * Need to access the cpu field knowing that
3689 * DEBUG_PAGEALLOC could have unmapped it if
3690 * the mutex owner just released it and exited.
3692 if (probe_kernel_address(&owner->cpu, cpu))
3699 * Even if the access succeeded (likely case),
3700 * the cpu field may no longer be valid.
3702 if (cpu >= nr_cpumask_bits)
3706 * We need to validate that we can do a
3707 * get_cpu() and that we have the percpu area.
3709 if (!cpu_online(cpu))
3716 * Owner changed, break to re-assess state.
3718 if (lock->owner != owner)
3722 * Is that owner really running on that cpu?
3724 if (task_thread_info(rq->curr) != owner || need_resched())
3734 #ifdef CONFIG_PREEMPT
3736 * this is the entry point to schedule() from in-kernel preemption
3737 * off of preempt_enable. Kernel preemptions off return from interrupt
3738 * occur there and call schedule directly.
3740 asmlinkage void __sched preempt_schedule(void)
3742 struct thread_info *ti = current_thread_info();
3745 * If there is a non-zero preempt_count or interrupts are disabled,
3746 * we do not want to preempt the current task. Just return..
3748 if (likely(ti->preempt_count || irqs_disabled()))
3752 add_preempt_count(PREEMPT_ACTIVE);
3754 sub_preempt_count(PREEMPT_ACTIVE);
3757 * Check again in case we missed a preemption opportunity
3758 * between schedule and now.
3761 } while (need_resched());
3763 EXPORT_SYMBOL(preempt_schedule);
3766 * this is the entry point to schedule() from kernel preemption
3767 * off of irq context.
3768 * Note, that this is called and return with irqs disabled. This will
3769 * protect us against recursive calling from irq.
3771 asmlinkage void __sched preempt_schedule_irq(void)
3773 struct thread_info *ti = current_thread_info();
3775 /* Catch callers which need to be fixed */
3776 BUG_ON(ti->preempt_count || !irqs_disabled());
3779 add_preempt_count(PREEMPT_ACTIVE);
3782 local_irq_disable();
3783 sub_preempt_count(PREEMPT_ACTIVE);
3786 * Check again in case we missed a preemption opportunity
3787 * between schedule and now.
3790 } while (need_resched());
3793 #endif /* CONFIG_PREEMPT */
3795 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3798 return try_to_wake_up(curr->private, mode, wake_flags);
3800 EXPORT_SYMBOL(default_wake_function);
3803 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3804 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3805 * number) then we wake all the non-exclusive tasks and one exclusive task.
3807 * There are circumstances in which we can try to wake a task which has already
3808 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3809 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3811 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3812 int nr_exclusive, int wake_flags, void *key)
3814 wait_queue_t *curr, *next;
3816 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3817 unsigned flags = curr->flags;
3819 if (curr->func(curr, mode, wake_flags, key) &&
3820 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3826 * __wake_up - wake up threads blocked on a waitqueue.
3828 * @mode: which threads
3829 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3830 * @key: is directly passed to the wakeup function
3832 * It may be assumed that this function implies a write memory barrier before
3833 * changing the task state if and only if any tasks are woken up.
3835 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3836 int nr_exclusive, void *key)
3838 unsigned long flags;
3840 spin_lock_irqsave(&q->lock, flags);
3841 __wake_up_common(q, mode, nr_exclusive, 0, key);
3842 spin_unlock_irqrestore(&q->lock, flags);
3844 EXPORT_SYMBOL(__wake_up);
3847 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3849 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3851 __wake_up_common(q, mode, 1, 0, NULL);
3854 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3856 __wake_up_common(q, mode, 1, 0, key);
3860 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3862 * @mode: which threads
3863 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3864 * @key: opaque value to be passed to wakeup targets
3866 * The sync wakeup differs that the waker knows that it will schedule
3867 * away soon, so while the target thread will be woken up, it will not
3868 * be migrated to another CPU - ie. the two threads are 'synchronized'
3869 * with each other. This can prevent needless bouncing between CPUs.
3871 * On UP it can prevent extra preemption.
3873 * It may be assumed that this function implies a write memory barrier before
3874 * changing the task state if and only if any tasks are woken up.
3876 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3877 int nr_exclusive, void *key)
3879 unsigned long flags;
3880 int wake_flags = WF_SYNC;
3885 if (unlikely(!nr_exclusive))
3888 spin_lock_irqsave(&q->lock, flags);
3889 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3890 spin_unlock_irqrestore(&q->lock, flags);
3892 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3895 * __wake_up_sync - see __wake_up_sync_key()
3897 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3899 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3901 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3904 * complete: - signals a single thread waiting on this completion
3905 * @x: holds the state of this particular completion
3907 * This will wake up a single thread waiting on this completion. Threads will be
3908 * awakened in the same order in which they were queued.
3910 * See also complete_all(), wait_for_completion() and related routines.
3912 * It may be assumed that this function implies a write memory barrier before
3913 * changing the task state if and only if any tasks are woken up.
3915 void complete(struct completion *x)
3917 unsigned long flags;
3919 spin_lock_irqsave(&x->wait.lock, flags);
3921 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3922 spin_unlock_irqrestore(&x->wait.lock, flags);
3924 EXPORT_SYMBOL(complete);
3927 * complete_all: - signals all threads waiting on this completion
3928 * @x: holds the state of this particular completion
3930 * This will wake up all threads waiting on this particular completion event.
3932 * It may be assumed that this function implies a write memory barrier before
3933 * changing the task state if and only if any tasks are woken up.
3935 void complete_all(struct completion *x)
3937 unsigned long flags;
3939 spin_lock_irqsave(&x->wait.lock, flags);
3940 x->done += UINT_MAX/2;
3941 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3942 spin_unlock_irqrestore(&x->wait.lock, flags);
3944 EXPORT_SYMBOL(complete_all);
3946 static inline long __sched
3947 do_wait_for_common(struct completion *x, long timeout, int state)
3950 DECLARE_WAITQUEUE(wait, current);
3952 wait.flags |= WQ_FLAG_EXCLUSIVE;
3953 __add_wait_queue_tail(&x->wait, &wait);
3955 if (signal_pending_state(state, current)) {
3956 timeout = -ERESTARTSYS;
3959 __set_current_state(state);
3960 spin_unlock_irq(&x->wait.lock);
3961 timeout = schedule_timeout(timeout);
3962 spin_lock_irq(&x->wait.lock);
3963 } while (!x->done && timeout);
3964 __remove_wait_queue(&x->wait, &wait);
3969 return timeout ?: 1;
3973 wait_for_common(struct completion *x, long timeout, int state)
3977 spin_lock_irq(&x->wait.lock);
3978 timeout = do_wait_for_common(x, timeout, state);
3979 spin_unlock_irq(&x->wait.lock);
3984 * wait_for_completion: - waits for completion of a task
3985 * @x: holds the state of this particular completion
3987 * This waits to be signaled for completion of a specific task. It is NOT
3988 * interruptible and there is no timeout.
3990 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3991 * and interrupt capability. Also see complete().
3993 void __sched wait_for_completion(struct completion *x)
3995 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3997 EXPORT_SYMBOL(wait_for_completion);
4000 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4001 * @x: holds the state of this particular completion
4002 * @timeout: timeout value in jiffies
4004 * This waits for either a completion of a specific task to be signaled or for a
4005 * specified timeout to expire. The timeout is in jiffies. It is not
4008 unsigned long __sched
4009 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4011 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4013 EXPORT_SYMBOL(wait_for_completion_timeout);
4016 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4017 * @x: holds the state of this particular completion
4019 * This waits for completion of a specific task to be signaled. It is
4022 int __sched wait_for_completion_interruptible(struct completion *x)
4024 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4025 if (t == -ERESTARTSYS)
4029 EXPORT_SYMBOL(wait_for_completion_interruptible);
4032 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4033 * @x: holds the state of this particular completion
4034 * @timeout: timeout value in jiffies
4036 * This waits for either a completion of a specific task to be signaled or for a
4037 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4039 unsigned long __sched
4040 wait_for_completion_interruptible_timeout(struct completion *x,
4041 unsigned long timeout)
4043 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4045 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4048 * wait_for_completion_killable: - waits for completion of a task (killable)
4049 * @x: holds the state of this particular completion
4051 * This waits to be signaled for completion of a specific task. It can be
4052 * interrupted by a kill signal.
4054 int __sched wait_for_completion_killable(struct completion *x)
4056 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4057 if (t == -ERESTARTSYS)
4061 EXPORT_SYMBOL(wait_for_completion_killable);
4064 * try_wait_for_completion - try to decrement a completion without blocking
4065 * @x: completion structure
4067 * Returns: 0 if a decrement cannot be done without blocking
4068 * 1 if a decrement succeeded.
4070 * If a completion is being used as a counting completion,
4071 * attempt to decrement the counter without blocking. This
4072 * enables us to avoid waiting if the resource the completion
4073 * is protecting is not available.
4075 bool try_wait_for_completion(struct completion *x)
4077 unsigned long flags;
4080 spin_lock_irqsave(&x->wait.lock, flags);
4085 spin_unlock_irqrestore(&x->wait.lock, flags);
4088 EXPORT_SYMBOL(try_wait_for_completion);
4091 * completion_done - Test to see if a completion has any waiters
4092 * @x: completion structure
4094 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4095 * 1 if there are no waiters.
4098 bool completion_done(struct completion *x)
4100 unsigned long flags;
4103 spin_lock_irqsave(&x->wait.lock, flags);
4106 spin_unlock_irqrestore(&x->wait.lock, flags);
4109 EXPORT_SYMBOL(completion_done);
4112 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4114 unsigned long flags;
4117 init_waitqueue_entry(&wait, current);
4119 __set_current_state(state);
4121 spin_lock_irqsave(&q->lock, flags);
4122 __add_wait_queue(q, &wait);
4123 spin_unlock(&q->lock);
4124 timeout = schedule_timeout(timeout);
4125 spin_lock_irq(&q->lock);
4126 __remove_wait_queue(q, &wait);
4127 spin_unlock_irqrestore(&q->lock, flags);
4132 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4134 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4136 EXPORT_SYMBOL(interruptible_sleep_on);
4139 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4141 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4143 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4145 void __sched sleep_on(wait_queue_head_t *q)
4147 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4149 EXPORT_SYMBOL(sleep_on);
4151 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4153 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4155 EXPORT_SYMBOL(sleep_on_timeout);
4157 #ifdef CONFIG_RT_MUTEXES
4160 * rt_mutex_setprio - set the current priority of a task
4162 * @prio: prio value (kernel-internal form)
4164 * This function changes the 'effective' priority of a task. It does
4165 * not touch ->normal_prio like __setscheduler().
4167 * Used by the rt_mutex code to implement priority inheritance logic.
4169 void rt_mutex_setprio(struct task_struct *p, int prio)
4171 unsigned long flags;
4172 int oldprio, on_rq, running;
4174 const struct sched_class *prev_class;
4176 BUG_ON(prio < 0 || prio > MAX_PRIO);
4178 rq = task_rq_lock(p, &flags);
4181 prev_class = p->sched_class;
4182 on_rq = p->se.on_rq;
4183 running = task_current(rq, p);
4185 dequeue_task(rq, p, 0);
4187 p->sched_class->put_prev_task(rq, p);
4190 p->sched_class = &rt_sched_class;
4192 p->sched_class = &fair_sched_class;
4197 p->sched_class->set_curr_task(rq);
4199 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4201 check_class_changed(rq, p, prev_class, oldprio, running);
4203 task_rq_unlock(rq, &flags);
4208 void set_user_nice(struct task_struct *p, long nice)
4210 int old_prio, delta, on_rq;
4211 unsigned long flags;
4214 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4217 * We have to be careful, if called from sys_setpriority(),
4218 * the task might be in the middle of scheduling on another CPU.
4220 rq = task_rq_lock(p, &flags);
4222 * The RT priorities are set via sched_setscheduler(), but we still
4223 * allow the 'normal' nice value to be set - but as expected
4224 * it wont have any effect on scheduling until the task is
4225 * SCHED_FIFO/SCHED_RR:
4227 if (task_has_rt_policy(p)) {
4228 p->static_prio = NICE_TO_PRIO(nice);
4231 on_rq = p->se.on_rq;
4233 dequeue_task(rq, p, 0);
4235 p->static_prio = NICE_TO_PRIO(nice);
4238 p->prio = effective_prio(p);
4239 delta = p->prio - old_prio;
4242 enqueue_task(rq, p, 0);
4244 * If the task increased its priority or is running and
4245 * lowered its priority, then reschedule its CPU:
4247 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4248 resched_task(rq->curr);
4251 task_rq_unlock(rq, &flags);
4253 EXPORT_SYMBOL(set_user_nice);
4256 * can_nice - check if a task can reduce its nice value
4260 int can_nice(const struct task_struct *p, const int nice)
4262 /* convert nice value [19,-20] to rlimit style value [1,40] */
4263 int nice_rlim = 20 - nice;
4265 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4266 capable(CAP_SYS_NICE));
4269 #ifdef __ARCH_WANT_SYS_NICE
4272 * sys_nice - change the priority of the current process.
4273 * @increment: priority increment
4275 * sys_setpriority is a more generic, but much slower function that
4276 * does similar things.
4278 SYSCALL_DEFINE1(nice, int, increment)
4283 * Setpriority might change our priority at the same moment.
4284 * We don't have to worry. Conceptually one call occurs first
4285 * and we have a single winner.
4287 if (increment < -40)
4292 nice = TASK_NICE(current) + increment;
4298 if (increment < 0 && !can_nice(current, nice))
4301 retval = security_task_setnice(current, nice);
4305 set_user_nice(current, nice);
4312 * task_prio - return the priority value of a given task.
4313 * @p: the task in question.
4315 * This is the priority value as seen by users in /proc.
4316 * RT tasks are offset by -200. Normal tasks are centered
4317 * around 0, value goes from -16 to +15.
4319 int task_prio(const struct task_struct *p)
4321 return p->prio - MAX_RT_PRIO;
4325 * task_nice - return the nice value of a given task.
4326 * @p: the task in question.
4328 int task_nice(const struct task_struct *p)
4330 return TASK_NICE(p);
4332 EXPORT_SYMBOL(task_nice);
4335 * idle_cpu - is a given cpu idle currently?
4336 * @cpu: the processor in question.
4338 int idle_cpu(int cpu)
4340 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4344 * idle_task - return the idle task for a given cpu.
4345 * @cpu: the processor in question.
4347 struct task_struct *idle_task(int cpu)
4349 return cpu_rq(cpu)->idle;
4353 * find_process_by_pid - find a process with a matching PID value.
4354 * @pid: the pid in question.
4356 static struct task_struct *find_process_by_pid(pid_t pid)
4358 return pid ? find_task_by_vpid(pid) : current;
4361 /* Actually do priority change: must hold rq lock. */
4363 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4365 BUG_ON(p->se.on_rq);
4368 p->rt_priority = prio;
4369 p->normal_prio = normal_prio(p);
4370 /* we are holding p->pi_lock already */
4371 p->prio = rt_mutex_getprio(p);
4372 if (rt_prio(p->prio))
4373 p->sched_class = &rt_sched_class;
4375 p->sched_class = &fair_sched_class;
4380 * check the target process has a UID that matches the current process's
4382 static bool check_same_owner(struct task_struct *p)
4384 const struct cred *cred = current_cred(), *pcred;
4388 pcred = __task_cred(p);
4389 match = (cred->euid == pcred->euid ||
4390 cred->euid == pcred->uid);
4395 static int __sched_setscheduler(struct task_struct *p, int policy,
4396 struct sched_param *param, bool user)
4398 int retval, oldprio, oldpolicy = -1, on_rq, running;
4399 unsigned long flags;
4400 const struct sched_class *prev_class;
4404 /* may grab non-irq protected spin_locks */
4405 BUG_ON(in_interrupt());
4407 /* double check policy once rq lock held */
4409 reset_on_fork = p->sched_reset_on_fork;
4410 policy = oldpolicy = p->policy;
4412 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4413 policy &= ~SCHED_RESET_ON_FORK;
4415 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4416 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4417 policy != SCHED_IDLE)
4422 * Valid priorities for SCHED_FIFO and SCHED_RR are
4423 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4424 * SCHED_BATCH and SCHED_IDLE is 0.
4426 if (param->sched_priority < 0 ||
4427 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4428 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4430 if (rt_policy(policy) != (param->sched_priority != 0))
4434 * Allow unprivileged RT tasks to decrease priority:
4436 if (user && !capable(CAP_SYS_NICE)) {
4437 if (rt_policy(policy)) {
4438 unsigned long rlim_rtprio;
4440 if (!lock_task_sighand(p, &flags))
4442 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4443 unlock_task_sighand(p, &flags);
4445 /* can't set/change the rt policy */
4446 if (policy != p->policy && !rlim_rtprio)
4449 /* can't increase priority */
4450 if (param->sched_priority > p->rt_priority &&
4451 param->sched_priority > rlim_rtprio)
4455 * Like positive nice levels, dont allow tasks to
4456 * move out of SCHED_IDLE either:
4458 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4461 /* can't change other user's priorities */
4462 if (!check_same_owner(p))
4465 /* Normal users shall not reset the sched_reset_on_fork flag */
4466 if (p->sched_reset_on_fork && !reset_on_fork)
4471 #ifdef CONFIG_RT_GROUP_SCHED
4473 * Do not allow realtime tasks into groups that have no runtime
4476 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4477 task_group(p)->rt_bandwidth.rt_runtime == 0)
4481 retval = security_task_setscheduler(p, policy, param);
4487 * make sure no PI-waiters arrive (or leave) while we are
4488 * changing the priority of the task:
4490 raw_spin_lock_irqsave(&p->pi_lock, flags);
4492 * To be able to change p->policy safely, the apropriate
4493 * runqueue lock must be held.
4495 rq = __task_rq_lock(p);
4496 /* recheck policy now with rq lock held */
4497 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4498 policy = oldpolicy = -1;
4499 __task_rq_unlock(rq);
4500 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4503 on_rq = p->se.on_rq;
4504 running = task_current(rq, p);
4506 deactivate_task(rq, p, 0);
4508 p->sched_class->put_prev_task(rq, p);
4510 p->sched_reset_on_fork = reset_on_fork;
4513 prev_class = p->sched_class;
4514 __setscheduler(rq, p, policy, param->sched_priority);
4517 p->sched_class->set_curr_task(rq);
4519 activate_task(rq, p, 0);
4521 check_class_changed(rq, p, prev_class, oldprio, running);
4523 __task_rq_unlock(rq);
4524 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4526 rt_mutex_adjust_pi(p);
4532 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4533 * @p: the task in question.
4534 * @policy: new policy.
4535 * @param: structure containing the new RT priority.
4537 * NOTE that the task may be already dead.
4539 int sched_setscheduler(struct task_struct *p, int policy,
4540 struct sched_param *param)
4542 return __sched_setscheduler(p, policy, param, true);
4544 EXPORT_SYMBOL_GPL(sched_setscheduler);
4547 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4548 * @p: the task in question.
4549 * @policy: new policy.
4550 * @param: structure containing the new RT priority.
4552 * Just like sched_setscheduler, only don't bother checking if the
4553 * current context has permission. For example, this is needed in
4554 * stop_machine(): we create temporary high priority worker threads,
4555 * but our caller might not have that capability.
4557 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4558 struct sched_param *param)
4560 return __sched_setscheduler(p, policy, param, false);
4564 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4566 struct sched_param lparam;
4567 struct task_struct *p;
4570 if (!param || pid < 0)
4572 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4577 p = find_process_by_pid(pid);
4579 retval = sched_setscheduler(p, policy, &lparam);
4586 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4587 * @pid: the pid in question.
4588 * @policy: new policy.
4589 * @param: structure containing the new RT priority.
4591 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4592 struct sched_param __user *, param)
4594 /* negative values for policy are not valid */
4598 return do_sched_setscheduler(pid, policy, param);
4602 * sys_sched_setparam - set/change the RT priority of a thread
4603 * @pid: the pid in question.
4604 * @param: structure containing the new RT priority.
4606 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4608 return do_sched_setscheduler(pid, -1, param);
4612 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4613 * @pid: the pid in question.
4615 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4617 struct task_struct *p;
4625 p = find_process_by_pid(pid);
4627 retval = security_task_getscheduler(p);
4630 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4637 * sys_sched_getparam - get the RT priority of a thread
4638 * @pid: the pid in question.
4639 * @param: structure containing the RT priority.
4641 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4643 struct sched_param lp;
4644 struct task_struct *p;
4647 if (!param || pid < 0)
4651 p = find_process_by_pid(pid);
4656 retval = security_task_getscheduler(p);
4660 lp.sched_priority = p->rt_priority;
4664 * This one might sleep, we cannot do it with a spinlock held ...
4666 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4675 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4677 cpumask_var_t cpus_allowed, new_mask;
4678 struct task_struct *p;
4684 p = find_process_by_pid(pid);
4691 /* Prevent p going away */
4695 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4699 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4701 goto out_free_cpus_allowed;
4704 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4707 retval = security_task_setscheduler(p, 0, NULL);
4711 cpuset_cpus_allowed(p, cpus_allowed);
4712 cpumask_and(new_mask, in_mask, cpus_allowed);
4714 retval = set_cpus_allowed_ptr(p, new_mask);
4717 cpuset_cpus_allowed(p, cpus_allowed);
4718 if (!cpumask_subset(new_mask, cpus_allowed)) {
4720 * We must have raced with a concurrent cpuset
4721 * update. Just reset the cpus_allowed to the
4722 * cpuset's cpus_allowed
4724 cpumask_copy(new_mask, cpus_allowed);
4729 free_cpumask_var(new_mask);
4730 out_free_cpus_allowed:
4731 free_cpumask_var(cpus_allowed);
4738 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4739 struct cpumask *new_mask)
4741 if (len < cpumask_size())
4742 cpumask_clear(new_mask);
4743 else if (len > cpumask_size())
4744 len = cpumask_size();
4746 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4750 * sys_sched_setaffinity - set the cpu affinity of a process
4751 * @pid: pid of the process
4752 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4753 * @user_mask_ptr: user-space pointer to the new cpu mask
4755 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4756 unsigned long __user *, user_mask_ptr)
4758 cpumask_var_t new_mask;
4761 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4764 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4766 retval = sched_setaffinity(pid, new_mask);
4767 free_cpumask_var(new_mask);
4771 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4773 struct task_struct *p;
4774 unsigned long flags;
4782 p = find_process_by_pid(pid);
4786 retval = security_task_getscheduler(p);
4790 rq = task_rq_lock(p, &flags);
4791 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4792 task_rq_unlock(rq, &flags);
4802 * sys_sched_getaffinity - get the cpu affinity of a process
4803 * @pid: pid of the process
4804 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4805 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4807 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4808 unsigned long __user *, user_mask_ptr)
4813 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4815 if (len & (sizeof(unsigned long)-1))
4818 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4821 ret = sched_getaffinity(pid, mask);
4823 size_t retlen = min_t(size_t, len, cpumask_size());
4825 if (copy_to_user(user_mask_ptr, mask, retlen))
4830 free_cpumask_var(mask);
4836 * sys_sched_yield - yield the current processor to other threads.
4838 * This function yields the current CPU to other tasks. If there are no
4839 * other threads running on this CPU then this function will return.
4841 SYSCALL_DEFINE0(sched_yield)
4843 struct rq *rq = this_rq_lock();
4845 schedstat_inc(rq, yld_count);
4846 current->sched_class->yield_task(rq);
4849 * Since we are going to call schedule() anyway, there's
4850 * no need to preempt or enable interrupts:
4852 __release(rq->lock);
4853 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4854 do_raw_spin_unlock(&rq->lock);
4855 preempt_enable_no_resched();
4862 static inline int should_resched(void)
4864 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4867 static void __cond_resched(void)
4869 add_preempt_count(PREEMPT_ACTIVE);
4871 sub_preempt_count(PREEMPT_ACTIVE);
4874 int __sched _cond_resched(void)
4876 if (should_resched()) {
4882 EXPORT_SYMBOL(_cond_resched);
4885 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4886 * call schedule, and on return reacquire the lock.
4888 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4889 * operations here to prevent schedule() from being called twice (once via
4890 * spin_unlock(), once by hand).
4892 int __cond_resched_lock(spinlock_t *lock)
4894 int resched = should_resched();
4897 lockdep_assert_held(lock);
4899 if (spin_needbreak(lock) || resched) {
4910 EXPORT_SYMBOL(__cond_resched_lock);
4912 int __sched __cond_resched_softirq(void)
4914 BUG_ON(!in_softirq());
4916 if (should_resched()) {
4924 EXPORT_SYMBOL(__cond_resched_softirq);
4927 * yield - yield the current processor to other threads.
4929 * This is a shortcut for kernel-space yielding - it marks the
4930 * thread runnable and calls sys_sched_yield().
4932 void __sched yield(void)
4934 set_current_state(TASK_RUNNING);
4937 EXPORT_SYMBOL(yield);
4940 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4941 * that process accounting knows that this is a task in IO wait state.
4943 void __sched io_schedule(void)
4945 struct rq *rq = raw_rq();
4947 delayacct_blkio_start();
4948 atomic_inc(&rq->nr_iowait);
4949 current->in_iowait = 1;
4951 current->in_iowait = 0;
4952 atomic_dec(&rq->nr_iowait);
4953 delayacct_blkio_end();
4955 EXPORT_SYMBOL(io_schedule);
4957 long __sched io_schedule_timeout(long timeout)
4959 struct rq *rq = raw_rq();
4962 delayacct_blkio_start();
4963 atomic_inc(&rq->nr_iowait);
4964 current->in_iowait = 1;
4965 ret = schedule_timeout(timeout);
4966 current->in_iowait = 0;
4967 atomic_dec(&rq->nr_iowait);
4968 delayacct_blkio_end();
4973 * sys_sched_get_priority_max - return maximum RT priority.
4974 * @policy: scheduling class.
4976 * this syscall returns the maximum rt_priority that can be used
4977 * by a given scheduling class.
4979 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4986 ret = MAX_USER_RT_PRIO-1;
4998 * sys_sched_get_priority_min - return minimum RT priority.
4999 * @policy: scheduling class.
5001 * this syscall returns the minimum rt_priority that can be used
5002 * by a given scheduling class.
5004 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5022 * sys_sched_rr_get_interval - return the default timeslice of a process.
5023 * @pid: pid of the process.
5024 * @interval: userspace pointer to the timeslice value.
5026 * this syscall writes the default timeslice value of a given process
5027 * into the user-space timespec buffer. A value of '0' means infinity.
5029 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5030 struct timespec __user *, interval)
5032 struct task_struct *p;
5033 unsigned int time_slice;
5034 unsigned long flags;
5044 p = find_process_by_pid(pid);
5048 retval = security_task_getscheduler(p);
5052 rq = task_rq_lock(p, &flags);
5053 time_slice = p->sched_class->get_rr_interval(rq, p);
5054 task_rq_unlock(rq, &flags);
5057 jiffies_to_timespec(time_slice, &t);
5058 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5066 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5068 void sched_show_task(struct task_struct *p)
5070 unsigned long free = 0;
5073 state = p->state ? __ffs(p->state) + 1 : 0;
5074 printk(KERN_INFO "%-13.13s %c", p->comm,
5075 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5076 #if BITS_PER_LONG == 32
5077 if (state == TASK_RUNNING)
5078 printk(KERN_CONT " running ");
5080 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5082 if (state == TASK_RUNNING)
5083 printk(KERN_CONT " running task ");
5085 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5087 #ifdef CONFIG_DEBUG_STACK_USAGE
5088 free = stack_not_used(p);
5090 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5091 task_pid_nr(p), task_pid_nr(p->real_parent),
5092 (unsigned long)task_thread_info(p)->flags);
5094 show_stack(p, NULL);
5097 void show_state_filter(unsigned long state_filter)
5099 struct task_struct *g, *p;
5101 #if BITS_PER_LONG == 32
5103 " task PC stack pid father\n");
5106 " task PC stack pid father\n");
5108 read_lock(&tasklist_lock);
5109 do_each_thread(g, p) {
5111 * reset the NMI-timeout, listing all files on a slow
5112 * console might take alot of time:
5114 touch_nmi_watchdog();
5115 if (!state_filter || (p->state & state_filter))
5117 } while_each_thread(g, p);
5119 touch_all_softlockup_watchdogs();
5121 #ifdef CONFIG_SCHED_DEBUG
5122 sysrq_sched_debug_show();
5124 read_unlock(&tasklist_lock);
5126 * Only show locks if all tasks are dumped:
5129 debug_show_all_locks();
5132 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5134 idle->sched_class = &idle_sched_class;
5138 * init_idle - set up an idle thread for a given CPU
5139 * @idle: task in question
5140 * @cpu: cpu the idle task belongs to
5142 * NOTE: this function does not set the idle thread's NEED_RESCHED
5143 * flag, to make booting more robust.
5145 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5147 struct rq *rq = cpu_rq(cpu);
5148 unsigned long flags;
5150 raw_spin_lock_irqsave(&rq->lock, flags);
5153 idle->state = TASK_RUNNING;
5154 idle->se.exec_start = sched_clock();
5156 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5157 __set_task_cpu(idle, cpu);
5159 rq->curr = rq->idle = idle;
5160 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5163 raw_spin_unlock_irqrestore(&rq->lock, flags);
5165 /* Set the preempt count _outside_ the spinlocks! */
5166 #if defined(CONFIG_PREEMPT)
5167 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5169 task_thread_info(idle)->preempt_count = 0;
5172 * The idle tasks have their own, simple scheduling class:
5174 idle->sched_class = &idle_sched_class;
5175 ftrace_graph_init_task(idle);
5179 * In a system that switches off the HZ timer nohz_cpu_mask
5180 * indicates which cpus entered this state. This is used
5181 * in the rcu update to wait only for active cpus. For system
5182 * which do not switch off the HZ timer nohz_cpu_mask should
5183 * always be CPU_BITS_NONE.
5185 cpumask_var_t nohz_cpu_mask;
5188 * Increase the granularity value when there are more CPUs,
5189 * because with more CPUs the 'effective latency' as visible
5190 * to users decreases. But the relationship is not linear,
5191 * so pick a second-best guess by going with the log2 of the
5194 * This idea comes from the SD scheduler of Con Kolivas:
5196 static int get_update_sysctl_factor(void)
5198 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5199 unsigned int factor;
5201 switch (sysctl_sched_tunable_scaling) {
5202 case SCHED_TUNABLESCALING_NONE:
5205 case SCHED_TUNABLESCALING_LINEAR:
5208 case SCHED_TUNABLESCALING_LOG:
5210 factor = 1 + ilog2(cpus);
5217 static void update_sysctl(void)
5219 unsigned int factor = get_update_sysctl_factor();
5221 #define SET_SYSCTL(name) \
5222 (sysctl_##name = (factor) * normalized_sysctl_##name)
5223 SET_SYSCTL(sched_min_granularity);
5224 SET_SYSCTL(sched_latency);
5225 SET_SYSCTL(sched_wakeup_granularity);
5226 SET_SYSCTL(sched_shares_ratelimit);
5230 static inline void sched_init_granularity(void)
5237 * This is how migration works:
5239 * 1) we queue a struct migration_req structure in the source CPU's
5240 * runqueue and wake up that CPU's migration thread.
5241 * 2) we down() the locked semaphore => thread blocks.
5242 * 3) migration thread wakes up (implicitly it forces the migrated
5243 * thread off the CPU)
5244 * 4) it gets the migration request and checks whether the migrated
5245 * task is still in the wrong runqueue.
5246 * 5) if it's in the wrong runqueue then the migration thread removes
5247 * it and puts it into the right queue.
5248 * 6) migration thread up()s the semaphore.
5249 * 7) we wake up and the migration is done.
5253 * Change a given task's CPU affinity. Migrate the thread to a
5254 * proper CPU and schedule it away if the CPU it's executing on
5255 * is removed from the allowed bitmask.
5257 * NOTE: the caller must have a valid reference to the task, the
5258 * task must not exit() & deallocate itself prematurely. The
5259 * call is not atomic; no spinlocks may be held.
5261 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5263 struct migration_req req;
5264 unsigned long flags;
5269 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5270 * drop the rq->lock and still rely on ->cpus_allowed.
5273 while (task_is_waking(p))
5275 rq = task_rq_lock(p, &flags);
5276 if (task_is_waking(p)) {
5277 task_rq_unlock(rq, &flags);
5281 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5286 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5287 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5292 if (p->sched_class->set_cpus_allowed)
5293 p->sched_class->set_cpus_allowed(p, new_mask);
5295 cpumask_copy(&p->cpus_allowed, new_mask);
5296 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5299 /* Can the task run on the task's current CPU? If so, we're done */
5300 if (cpumask_test_cpu(task_cpu(p), new_mask))
5303 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5304 /* Need help from migration thread: drop lock and wait. */
5305 struct task_struct *mt = rq->migration_thread;
5307 get_task_struct(mt);
5308 task_rq_unlock(rq, &flags);
5309 wake_up_process(mt);
5310 put_task_struct(mt);
5311 wait_for_completion(&req.done);
5312 tlb_migrate_finish(p->mm);
5316 task_rq_unlock(rq, &flags);
5320 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5323 * Move (not current) task off this cpu, onto dest cpu. We're doing
5324 * this because either it can't run here any more (set_cpus_allowed()
5325 * away from this CPU, or CPU going down), or because we're
5326 * attempting to rebalance this task on exec (sched_exec).
5328 * So we race with normal scheduler movements, but that's OK, as long
5329 * as the task is no longer on this CPU.
5331 * Returns non-zero if task was successfully migrated.
5333 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5335 struct rq *rq_dest, *rq_src;
5338 if (unlikely(!cpu_active(dest_cpu)))
5341 rq_src = cpu_rq(src_cpu);
5342 rq_dest = cpu_rq(dest_cpu);
5344 double_rq_lock(rq_src, rq_dest);
5345 /* Already moved. */
5346 if (task_cpu(p) != src_cpu)
5348 /* Affinity changed (again). */
5349 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5353 * If we're not on a rq, the next wake-up will ensure we're
5357 deactivate_task(rq_src, p, 0);
5358 set_task_cpu(p, dest_cpu);
5359 activate_task(rq_dest, p, 0);
5360 check_preempt_curr(rq_dest, p, 0);
5365 double_rq_unlock(rq_src, rq_dest);
5369 #define RCU_MIGRATION_IDLE 0
5370 #define RCU_MIGRATION_NEED_QS 1
5371 #define RCU_MIGRATION_GOT_QS 2
5372 #define RCU_MIGRATION_MUST_SYNC 3
5375 * migration_thread - this is a highprio system thread that performs
5376 * thread migration by bumping thread off CPU then 'pushing' onto
5379 static int migration_thread(void *data)
5382 int cpu = (long)data;
5386 BUG_ON(rq->migration_thread != current);
5388 set_current_state(TASK_INTERRUPTIBLE);
5389 while (!kthread_should_stop()) {
5390 struct migration_req *req;
5391 struct list_head *head;
5393 raw_spin_lock_irq(&rq->lock);
5395 if (cpu_is_offline(cpu)) {
5396 raw_spin_unlock_irq(&rq->lock);
5400 if (rq->active_balance) {
5401 active_load_balance(rq, cpu);
5402 rq->active_balance = 0;
5405 head = &rq->migration_queue;
5407 if (list_empty(head)) {
5408 raw_spin_unlock_irq(&rq->lock);
5410 set_current_state(TASK_INTERRUPTIBLE);
5413 req = list_entry(head->next, struct migration_req, list);
5414 list_del_init(head->next);
5416 if (req->task != NULL) {
5417 raw_spin_unlock(&rq->lock);
5418 __migrate_task(req->task, cpu, req->dest_cpu);
5419 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5420 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5421 raw_spin_unlock(&rq->lock);
5423 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5424 raw_spin_unlock(&rq->lock);
5425 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5429 complete(&req->done);
5431 __set_current_state(TASK_RUNNING);
5436 #ifdef CONFIG_HOTPLUG_CPU
5438 * Figure out where task on dead CPU should go, use force if necessary.
5440 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5442 struct rq *rq = cpu_rq(dead_cpu);
5443 int needs_cpu, uninitialized_var(dest_cpu);
5444 unsigned long flags;
5446 local_irq_save(flags);
5448 raw_spin_lock(&rq->lock);
5449 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5451 dest_cpu = select_fallback_rq(dead_cpu, p);
5452 raw_spin_unlock(&rq->lock);
5454 * It can only fail if we race with set_cpus_allowed(),
5455 * in the racer should migrate the task anyway.
5458 __migrate_task(p, dead_cpu, dest_cpu);
5459 local_irq_restore(flags);
5463 * While a dead CPU has no uninterruptible tasks queued at this point,
5464 * it might still have a nonzero ->nr_uninterruptible counter, because
5465 * for performance reasons the counter is not stricly tracking tasks to
5466 * their home CPUs. So we just add the counter to another CPU's counter,
5467 * to keep the global sum constant after CPU-down:
5469 static void migrate_nr_uninterruptible(struct rq *rq_src)
5471 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5472 unsigned long flags;
5474 local_irq_save(flags);
5475 double_rq_lock(rq_src, rq_dest);
5476 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5477 rq_src->nr_uninterruptible = 0;
5478 double_rq_unlock(rq_src, rq_dest);
5479 local_irq_restore(flags);
5482 /* Run through task list and migrate tasks from the dead cpu. */
5483 static void migrate_live_tasks(int src_cpu)
5485 struct task_struct *p, *t;
5487 read_lock(&tasklist_lock);
5489 do_each_thread(t, p) {
5493 if (task_cpu(p) == src_cpu)
5494 move_task_off_dead_cpu(src_cpu, p);
5495 } while_each_thread(t, p);
5497 read_unlock(&tasklist_lock);
5501 * Schedules idle task to be the next runnable task on current CPU.
5502 * It does so by boosting its priority to highest possible.
5503 * Used by CPU offline code.
5505 void sched_idle_next(void)
5507 int this_cpu = smp_processor_id();
5508 struct rq *rq = cpu_rq(this_cpu);
5509 struct task_struct *p = rq->idle;
5510 unsigned long flags;
5512 /* cpu has to be offline */
5513 BUG_ON(cpu_online(this_cpu));
5516 * Strictly not necessary since rest of the CPUs are stopped by now
5517 * and interrupts disabled on the current cpu.
5519 raw_spin_lock_irqsave(&rq->lock, flags);
5521 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5523 activate_task(rq, p, 0);
5525 raw_spin_unlock_irqrestore(&rq->lock, flags);
5529 * Ensures that the idle task is using init_mm right before its cpu goes
5532 void idle_task_exit(void)
5534 struct mm_struct *mm = current->active_mm;
5536 BUG_ON(cpu_online(smp_processor_id()));
5539 switch_mm(mm, &init_mm, current);
5543 /* called under rq->lock with disabled interrupts */
5544 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5546 struct rq *rq = cpu_rq(dead_cpu);
5548 /* Must be exiting, otherwise would be on tasklist. */
5549 BUG_ON(!p->exit_state);
5551 /* Cannot have done final schedule yet: would have vanished. */
5552 BUG_ON(p->state == TASK_DEAD);
5557 * Drop lock around migration; if someone else moves it,
5558 * that's OK. No task can be added to this CPU, so iteration is
5561 raw_spin_unlock_irq(&rq->lock);
5562 move_task_off_dead_cpu(dead_cpu, p);
5563 raw_spin_lock_irq(&rq->lock);
5568 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5569 static void migrate_dead_tasks(unsigned int dead_cpu)
5571 struct rq *rq = cpu_rq(dead_cpu);
5572 struct task_struct *next;
5575 if (!rq->nr_running)
5577 next = pick_next_task(rq);
5580 next->sched_class->put_prev_task(rq, next);
5581 migrate_dead(dead_cpu, next);
5587 * remove the tasks which were accounted by rq from calc_load_tasks.
5589 static void calc_global_load_remove(struct rq *rq)
5591 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5592 rq->calc_load_active = 0;
5594 #endif /* CONFIG_HOTPLUG_CPU */
5596 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5598 static struct ctl_table sd_ctl_dir[] = {
5600 .procname = "sched_domain",
5606 static struct ctl_table sd_ctl_root[] = {
5608 .procname = "kernel",
5610 .child = sd_ctl_dir,
5615 static struct ctl_table *sd_alloc_ctl_entry(int n)
5617 struct ctl_table *entry =
5618 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5623 static void sd_free_ctl_entry(struct ctl_table **tablep)
5625 struct ctl_table *entry;
5628 * In the intermediate directories, both the child directory and
5629 * procname are dynamically allocated and could fail but the mode
5630 * will always be set. In the lowest directory the names are
5631 * static strings and all have proc handlers.
5633 for (entry = *tablep; entry->mode; entry++) {
5635 sd_free_ctl_entry(&entry->child);
5636 if (entry->proc_handler == NULL)
5637 kfree(entry->procname);
5645 set_table_entry(struct ctl_table *entry,
5646 const char *procname, void *data, int maxlen,
5647 mode_t mode, proc_handler *proc_handler)
5649 entry->procname = procname;
5651 entry->maxlen = maxlen;
5653 entry->proc_handler = proc_handler;
5656 static struct ctl_table *
5657 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5659 struct ctl_table *table = sd_alloc_ctl_entry(13);
5664 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5665 sizeof(long), 0644, proc_doulongvec_minmax);
5666 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5667 sizeof(long), 0644, proc_doulongvec_minmax);
5668 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5669 sizeof(int), 0644, proc_dointvec_minmax);
5670 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5671 sizeof(int), 0644, proc_dointvec_minmax);
5672 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5673 sizeof(int), 0644, proc_dointvec_minmax);
5674 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5675 sizeof(int), 0644, proc_dointvec_minmax);
5676 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5677 sizeof(int), 0644, proc_dointvec_minmax);
5678 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5679 sizeof(int), 0644, proc_dointvec_minmax);
5680 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5681 sizeof(int), 0644, proc_dointvec_minmax);
5682 set_table_entry(&table[9], "cache_nice_tries",
5683 &sd->cache_nice_tries,
5684 sizeof(int), 0644, proc_dointvec_minmax);
5685 set_table_entry(&table[10], "flags", &sd->flags,
5686 sizeof(int), 0644, proc_dointvec_minmax);
5687 set_table_entry(&table[11], "name", sd->name,
5688 CORENAME_MAX_SIZE, 0444, proc_dostring);
5689 /* &table[12] is terminator */
5694 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5696 struct ctl_table *entry, *table;
5697 struct sched_domain *sd;
5698 int domain_num = 0, i;
5701 for_each_domain(cpu, sd)
5703 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5708 for_each_domain(cpu, sd) {
5709 snprintf(buf, 32, "domain%d", i);
5710 entry->procname = kstrdup(buf, GFP_KERNEL);
5712 entry->child = sd_alloc_ctl_domain_table(sd);
5719 static struct ctl_table_header *sd_sysctl_header;
5720 static void register_sched_domain_sysctl(void)
5722 int i, cpu_num = num_possible_cpus();
5723 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5726 WARN_ON(sd_ctl_dir[0].child);
5727 sd_ctl_dir[0].child = entry;
5732 for_each_possible_cpu(i) {
5733 snprintf(buf, 32, "cpu%d", i);
5734 entry->procname = kstrdup(buf, GFP_KERNEL);
5736 entry->child = sd_alloc_ctl_cpu_table(i);
5740 WARN_ON(sd_sysctl_header);
5741 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5744 /* may be called multiple times per register */
5745 static void unregister_sched_domain_sysctl(void)
5747 if (sd_sysctl_header)
5748 unregister_sysctl_table(sd_sysctl_header);
5749 sd_sysctl_header = NULL;
5750 if (sd_ctl_dir[0].child)
5751 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5754 static void register_sched_domain_sysctl(void)
5757 static void unregister_sched_domain_sysctl(void)
5762 static void set_rq_online(struct rq *rq)
5765 const struct sched_class *class;
5767 cpumask_set_cpu(rq->cpu, rq->rd->online);
5770 for_each_class(class) {
5771 if (class->rq_online)
5772 class->rq_online(rq);
5777 static void set_rq_offline(struct rq *rq)
5780 const struct sched_class *class;
5782 for_each_class(class) {
5783 if (class->rq_offline)
5784 class->rq_offline(rq);
5787 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5793 * migration_call - callback that gets triggered when a CPU is added.
5794 * Here we can start up the necessary migration thread for the new CPU.
5796 static int __cpuinit
5797 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5799 struct task_struct *p;
5800 int cpu = (long)hcpu;
5801 unsigned long flags;
5806 case CPU_UP_PREPARE:
5807 case CPU_UP_PREPARE_FROZEN:
5808 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5811 kthread_bind(p, cpu);
5812 /* Must be high prio: stop_machine expects to yield to it. */
5813 rq = task_rq_lock(p, &flags);
5814 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5815 task_rq_unlock(rq, &flags);
5817 cpu_rq(cpu)->migration_thread = p;
5818 rq->calc_load_update = calc_load_update;
5822 case CPU_ONLINE_FROZEN:
5823 /* Strictly unnecessary, as first user will wake it. */
5824 wake_up_process(cpu_rq(cpu)->migration_thread);
5826 /* Update our root-domain */
5828 raw_spin_lock_irqsave(&rq->lock, flags);
5830 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5834 raw_spin_unlock_irqrestore(&rq->lock, flags);
5837 #ifdef CONFIG_HOTPLUG_CPU
5838 case CPU_UP_CANCELED:
5839 case CPU_UP_CANCELED_FROZEN:
5840 if (!cpu_rq(cpu)->migration_thread)
5842 /* Unbind it from offline cpu so it can run. Fall thru. */
5843 kthread_bind(cpu_rq(cpu)->migration_thread,
5844 cpumask_any(cpu_online_mask));
5845 kthread_stop(cpu_rq(cpu)->migration_thread);
5846 put_task_struct(cpu_rq(cpu)->migration_thread);
5847 cpu_rq(cpu)->migration_thread = NULL;
5851 case CPU_DEAD_FROZEN:
5852 migrate_live_tasks(cpu);
5854 kthread_stop(rq->migration_thread);
5855 put_task_struct(rq->migration_thread);
5856 rq->migration_thread = NULL;
5857 /* Idle task back to normal (off runqueue, low prio) */
5858 raw_spin_lock_irq(&rq->lock);
5859 deactivate_task(rq, rq->idle, 0);
5860 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5861 rq->idle->sched_class = &idle_sched_class;
5862 migrate_dead_tasks(cpu);
5863 raw_spin_unlock_irq(&rq->lock);
5864 migrate_nr_uninterruptible(rq);
5865 BUG_ON(rq->nr_running != 0);
5866 calc_global_load_remove(rq);
5868 * No need to migrate the tasks: it was best-effort if
5869 * they didn't take sched_hotcpu_mutex. Just wake up
5872 raw_spin_lock_irq(&rq->lock);
5873 while (!list_empty(&rq->migration_queue)) {
5874 struct migration_req *req;
5876 req = list_entry(rq->migration_queue.next,
5877 struct migration_req, list);
5878 list_del_init(&req->list);
5879 raw_spin_unlock_irq(&rq->lock);
5880 complete(&req->done);
5881 raw_spin_lock_irq(&rq->lock);
5883 raw_spin_unlock_irq(&rq->lock);
5887 case CPU_DYING_FROZEN:
5888 /* Update our root-domain */
5890 raw_spin_lock_irqsave(&rq->lock, flags);
5892 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5895 raw_spin_unlock_irqrestore(&rq->lock, flags);
5903 * Register at high priority so that task migration (migrate_all_tasks)
5904 * happens before everything else. This has to be lower priority than
5905 * the notifier in the perf_event subsystem, though.
5907 static struct notifier_block __cpuinitdata migration_notifier = {
5908 .notifier_call = migration_call,
5912 static int __init migration_init(void)
5914 void *cpu = (void *)(long)smp_processor_id();
5917 /* Start one for the boot CPU: */
5918 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5919 BUG_ON(err == NOTIFY_BAD);
5920 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5921 register_cpu_notifier(&migration_notifier);
5925 early_initcall(migration_init);
5930 #ifdef CONFIG_SCHED_DEBUG
5932 static __read_mostly int sched_domain_debug_enabled;
5934 static int __init sched_domain_debug_setup(char *str)
5936 sched_domain_debug_enabled = 1;
5940 early_param("sched_debug", sched_domain_debug_setup);
5942 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5943 struct cpumask *groupmask)
5945 struct sched_group *group = sd->groups;
5948 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5949 cpumask_clear(groupmask);
5951 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5953 if (!(sd->flags & SD_LOAD_BALANCE)) {
5954 printk("does not load-balance\n");
5956 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5961 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5963 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5964 printk(KERN_ERR "ERROR: domain->span does not contain "
5967 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5968 printk(KERN_ERR "ERROR: domain->groups does not contain"
5972 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5976 printk(KERN_ERR "ERROR: group is NULL\n");
5980 if (!group->cpu_power) {
5981 printk(KERN_CONT "\n");
5982 printk(KERN_ERR "ERROR: domain->cpu_power not "
5987 if (!cpumask_weight(sched_group_cpus(group))) {
5988 printk(KERN_CONT "\n");
5989 printk(KERN_ERR "ERROR: empty group\n");
5993 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5994 printk(KERN_CONT "\n");
5995 printk(KERN_ERR "ERROR: repeated CPUs\n");
5999 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6001 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6003 printk(KERN_CONT " %s", str);
6004 if (group->cpu_power != SCHED_LOAD_SCALE) {
6005 printk(KERN_CONT " (cpu_power = %d)",
6009 group = group->next;
6010 } while (group != sd->groups);
6011 printk(KERN_CONT "\n");
6013 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6014 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6017 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6018 printk(KERN_ERR "ERROR: parent span is not a superset "
6019 "of domain->span\n");
6023 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6025 cpumask_var_t groupmask;
6028 if (!sched_domain_debug_enabled)
6032 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6036 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6038 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6039 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6044 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6051 free_cpumask_var(groupmask);
6053 #else /* !CONFIG_SCHED_DEBUG */
6054 # define sched_domain_debug(sd, cpu) do { } while (0)
6055 #endif /* CONFIG_SCHED_DEBUG */
6057 static int sd_degenerate(struct sched_domain *sd)
6059 if (cpumask_weight(sched_domain_span(sd)) == 1)
6062 /* Following flags need at least 2 groups */
6063 if (sd->flags & (SD_LOAD_BALANCE |
6064 SD_BALANCE_NEWIDLE |
6068 SD_SHARE_PKG_RESOURCES)) {
6069 if (sd->groups != sd->groups->next)
6073 /* Following flags don't use groups */
6074 if (sd->flags & (SD_WAKE_AFFINE))
6081 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6083 unsigned long cflags = sd->flags, pflags = parent->flags;
6085 if (sd_degenerate(parent))
6088 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6091 /* Flags needing groups don't count if only 1 group in parent */
6092 if (parent->groups == parent->groups->next) {
6093 pflags &= ~(SD_LOAD_BALANCE |
6094 SD_BALANCE_NEWIDLE |
6098 SD_SHARE_PKG_RESOURCES);
6099 if (nr_node_ids == 1)
6100 pflags &= ~SD_SERIALIZE;
6102 if (~cflags & pflags)
6108 static void free_rootdomain(struct root_domain *rd)
6110 synchronize_sched();
6112 cpupri_cleanup(&rd->cpupri);
6114 free_cpumask_var(rd->rto_mask);
6115 free_cpumask_var(rd->online);
6116 free_cpumask_var(rd->span);
6120 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6122 struct root_domain *old_rd = NULL;
6123 unsigned long flags;
6125 raw_spin_lock_irqsave(&rq->lock, flags);
6130 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6133 cpumask_clear_cpu(rq->cpu, old_rd->span);
6136 * If we dont want to free the old_rt yet then
6137 * set old_rd to NULL to skip the freeing later
6140 if (!atomic_dec_and_test(&old_rd->refcount))
6144 atomic_inc(&rd->refcount);
6147 cpumask_set_cpu(rq->cpu, rd->span);
6148 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6151 raw_spin_unlock_irqrestore(&rq->lock, flags);
6154 free_rootdomain(old_rd);
6157 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6159 gfp_t gfp = GFP_KERNEL;
6161 memset(rd, 0, sizeof(*rd));
6166 if (!alloc_cpumask_var(&rd->span, gfp))
6168 if (!alloc_cpumask_var(&rd->online, gfp))
6170 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6173 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6178 free_cpumask_var(rd->rto_mask);
6180 free_cpumask_var(rd->online);
6182 free_cpumask_var(rd->span);
6187 static void init_defrootdomain(void)
6189 init_rootdomain(&def_root_domain, true);
6191 atomic_set(&def_root_domain.refcount, 1);
6194 static struct root_domain *alloc_rootdomain(void)
6196 struct root_domain *rd;
6198 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6202 if (init_rootdomain(rd, false) != 0) {
6211 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6212 * hold the hotplug lock.
6215 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6217 struct rq *rq = cpu_rq(cpu);
6218 struct sched_domain *tmp;
6220 /* Remove the sched domains which do not contribute to scheduling. */
6221 for (tmp = sd; tmp; ) {
6222 struct sched_domain *parent = tmp->parent;
6226 if (sd_parent_degenerate(tmp, parent)) {
6227 tmp->parent = parent->parent;
6229 parent->parent->child = tmp;
6234 if (sd && sd_degenerate(sd)) {
6240 sched_domain_debug(sd, cpu);
6242 rq_attach_root(rq, rd);
6243 rcu_assign_pointer(rq->sd, sd);
6246 /* cpus with isolated domains */
6247 static cpumask_var_t cpu_isolated_map;
6249 /* Setup the mask of cpus configured for isolated domains */
6250 static int __init isolated_cpu_setup(char *str)
6252 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6253 cpulist_parse(str, cpu_isolated_map);
6257 __setup("isolcpus=", isolated_cpu_setup);
6260 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6261 * to a function which identifies what group(along with sched group) a CPU
6262 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6263 * (due to the fact that we keep track of groups covered with a struct cpumask).
6265 * init_sched_build_groups will build a circular linked list of the groups
6266 * covered by the given span, and will set each group's ->cpumask correctly,
6267 * and ->cpu_power to 0.
6270 init_sched_build_groups(const struct cpumask *span,
6271 const struct cpumask *cpu_map,
6272 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6273 struct sched_group **sg,
6274 struct cpumask *tmpmask),
6275 struct cpumask *covered, struct cpumask *tmpmask)
6277 struct sched_group *first = NULL, *last = NULL;
6280 cpumask_clear(covered);
6282 for_each_cpu(i, span) {
6283 struct sched_group *sg;
6284 int group = group_fn(i, cpu_map, &sg, tmpmask);
6287 if (cpumask_test_cpu(i, covered))
6290 cpumask_clear(sched_group_cpus(sg));
6293 for_each_cpu(j, span) {
6294 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6297 cpumask_set_cpu(j, covered);
6298 cpumask_set_cpu(j, sched_group_cpus(sg));
6309 #define SD_NODES_PER_DOMAIN 16
6314 * find_next_best_node - find the next node to include in a sched_domain
6315 * @node: node whose sched_domain we're building
6316 * @used_nodes: nodes already in the sched_domain
6318 * Find the next node to include in a given scheduling domain. Simply
6319 * finds the closest node not already in the @used_nodes map.
6321 * Should use nodemask_t.
6323 static int find_next_best_node(int node, nodemask_t *used_nodes)
6325 int i, n, val, min_val, best_node = 0;
6329 for (i = 0; i < nr_node_ids; i++) {
6330 /* Start at @node */
6331 n = (node + i) % nr_node_ids;
6333 if (!nr_cpus_node(n))
6336 /* Skip already used nodes */
6337 if (node_isset(n, *used_nodes))
6340 /* Simple min distance search */
6341 val = node_distance(node, n);
6343 if (val < min_val) {
6349 node_set(best_node, *used_nodes);
6354 * sched_domain_node_span - get a cpumask for a node's sched_domain
6355 * @node: node whose cpumask we're constructing
6356 * @span: resulting cpumask
6358 * Given a node, construct a good cpumask for its sched_domain to span. It
6359 * should be one that prevents unnecessary balancing, but also spreads tasks
6362 static void sched_domain_node_span(int node, struct cpumask *span)
6364 nodemask_t used_nodes;
6367 cpumask_clear(span);
6368 nodes_clear(used_nodes);
6370 cpumask_or(span, span, cpumask_of_node(node));
6371 node_set(node, used_nodes);
6373 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6374 int next_node = find_next_best_node(node, &used_nodes);
6376 cpumask_or(span, span, cpumask_of_node(next_node));
6379 #endif /* CONFIG_NUMA */
6381 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6384 * The cpus mask in sched_group and sched_domain hangs off the end.
6386 * ( See the the comments in include/linux/sched.h:struct sched_group
6387 * and struct sched_domain. )
6389 struct static_sched_group {
6390 struct sched_group sg;
6391 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6394 struct static_sched_domain {
6395 struct sched_domain sd;
6396 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6402 cpumask_var_t domainspan;
6403 cpumask_var_t covered;
6404 cpumask_var_t notcovered;
6406 cpumask_var_t nodemask;
6407 cpumask_var_t this_sibling_map;
6408 cpumask_var_t this_core_map;
6409 cpumask_var_t send_covered;
6410 cpumask_var_t tmpmask;
6411 struct sched_group **sched_group_nodes;
6412 struct root_domain *rd;
6416 sa_sched_groups = 0,
6421 sa_this_sibling_map,
6423 sa_sched_group_nodes,
6433 * SMT sched-domains:
6435 #ifdef CONFIG_SCHED_SMT
6436 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6437 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6440 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6441 struct sched_group **sg, struct cpumask *unused)
6444 *sg = &per_cpu(sched_groups, cpu).sg;
6447 #endif /* CONFIG_SCHED_SMT */
6450 * multi-core sched-domains:
6452 #ifdef CONFIG_SCHED_MC
6453 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6454 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6455 #endif /* CONFIG_SCHED_MC */
6457 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6459 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6460 struct sched_group **sg, struct cpumask *mask)
6464 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6465 group = cpumask_first(mask);
6467 *sg = &per_cpu(sched_group_core, group).sg;
6470 #elif defined(CONFIG_SCHED_MC)
6472 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6473 struct sched_group **sg, struct cpumask *unused)
6476 *sg = &per_cpu(sched_group_core, cpu).sg;
6481 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6482 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6485 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6486 struct sched_group **sg, struct cpumask *mask)
6489 #ifdef CONFIG_SCHED_MC
6490 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6491 group = cpumask_first(mask);
6492 #elif defined(CONFIG_SCHED_SMT)
6493 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6494 group = cpumask_first(mask);
6499 *sg = &per_cpu(sched_group_phys, group).sg;
6505 * The init_sched_build_groups can't handle what we want to do with node
6506 * groups, so roll our own. Now each node has its own list of groups which
6507 * gets dynamically allocated.
6509 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6510 static struct sched_group ***sched_group_nodes_bycpu;
6512 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6513 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6515 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6516 struct sched_group **sg,
6517 struct cpumask *nodemask)
6521 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6522 group = cpumask_first(nodemask);
6525 *sg = &per_cpu(sched_group_allnodes, group).sg;
6529 static void init_numa_sched_groups_power(struct sched_group *group_head)
6531 struct sched_group *sg = group_head;
6537 for_each_cpu(j, sched_group_cpus(sg)) {
6538 struct sched_domain *sd;
6540 sd = &per_cpu(phys_domains, j).sd;
6541 if (j != group_first_cpu(sd->groups)) {
6543 * Only add "power" once for each
6549 sg->cpu_power += sd->groups->cpu_power;
6552 } while (sg != group_head);
6555 static int build_numa_sched_groups(struct s_data *d,
6556 const struct cpumask *cpu_map, int num)
6558 struct sched_domain *sd;
6559 struct sched_group *sg, *prev;
6562 cpumask_clear(d->covered);
6563 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6564 if (cpumask_empty(d->nodemask)) {
6565 d->sched_group_nodes[num] = NULL;
6569 sched_domain_node_span(num, d->domainspan);
6570 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6572 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6575 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6579 d->sched_group_nodes[num] = sg;
6581 for_each_cpu(j, d->nodemask) {
6582 sd = &per_cpu(node_domains, j).sd;
6587 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6589 cpumask_or(d->covered, d->covered, d->nodemask);
6592 for (j = 0; j < nr_node_ids; j++) {
6593 n = (num + j) % nr_node_ids;
6594 cpumask_complement(d->notcovered, d->covered);
6595 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6596 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6597 if (cpumask_empty(d->tmpmask))
6599 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6600 if (cpumask_empty(d->tmpmask))
6602 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6606 "Can not alloc domain group for node %d\n", j);
6610 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6611 sg->next = prev->next;
6612 cpumask_or(d->covered, d->covered, d->tmpmask);
6619 #endif /* CONFIG_NUMA */
6622 /* Free memory allocated for various sched_group structures */
6623 static void free_sched_groups(const struct cpumask *cpu_map,
6624 struct cpumask *nodemask)
6628 for_each_cpu(cpu, cpu_map) {
6629 struct sched_group **sched_group_nodes
6630 = sched_group_nodes_bycpu[cpu];
6632 if (!sched_group_nodes)
6635 for (i = 0; i < nr_node_ids; i++) {
6636 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6638 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6639 if (cpumask_empty(nodemask))
6649 if (oldsg != sched_group_nodes[i])
6652 kfree(sched_group_nodes);
6653 sched_group_nodes_bycpu[cpu] = NULL;
6656 #else /* !CONFIG_NUMA */
6657 static void free_sched_groups(const struct cpumask *cpu_map,
6658 struct cpumask *nodemask)
6661 #endif /* CONFIG_NUMA */
6664 * Initialize sched groups cpu_power.
6666 * cpu_power indicates the capacity of sched group, which is used while
6667 * distributing the load between different sched groups in a sched domain.
6668 * Typically cpu_power for all the groups in a sched domain will be same unless
6669 * there are asymmetries in the topology. If there are asymmetries, group
6670 * having more cpu_power will pickup more load compared to the group having
6673 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6675 struct sched_domain *child;
6676 struct sched_group *group;
6680 WARN_ON(!sd || !sd->groups);
6682 if (cpu != group_first_cpu(sd->groups))
6687 sd->groups->cpu_power = 0;
6690 power = SCHED_LOAD_SCALE;
6691 weight = cpumask_weight(sched_domain_span(sd));
6693 * SMT siblings share the power of a single core.
6694 * Usually multiple threads get a better yield out of
6695 * that one core than a single thread would have,
6696 * reflect that in sd->smt_gain.
6698 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6699 power *= sd->smt_gain;
6701 power >>= SCHED_LOAD_SHIFT;
6703 sd->groups->cpu_power += power;
6708 * Add cpu_power of each child group to this groups cpu_power.
6710 group = child->groups;
6712 sd->groups->cpu_power += group->cpu_power;
6713 group = group->next;
6714 } while (group != child->groups);
6718 * Initializers for schedule domains
6719 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6722 #ifdef CONFIG_SCHED_DEBUG
6723 # define SD_INIT_NAME(sd, type) sd->name = #type
6725 # define SD_INIT_NAME(sd, type) do { } while (0)
6728 #define SD_INIT(sd, type) sd_init_##type(sd)
6730 #define SD_INIT_FUNC(type) \
6731 static noinline void sd_init_##type(struct sched_domain *sd) \
6733 memset(sd, 0, sizeof(*sd)); \
6734 *sd = SD_##type##_INIT; \
6735 sd->level = SD_LV_##type; \
6736 SD_INIT_NAME(sd, type); \
6741 SD_INIT_FUNC(ALLNODES)
6744 #ifdef CONFIG_SCHED_SMT
6745 SD_INIT_FUNC(SIBLING)
6747 #ifdef CONFIG_SCHED_MC
6751 static int default_relax_domain_level = -1;
6753 static int __init setup_relax_domain_level(char *str)
6757 val = simple_strtoul(str, NULL, 0);
6758 if (val < SD_LV_MAX)
6759 default_relax_domain_level = val;
6763 __setup("relax_domain_level=", setup_relax_domain_level);
6765 static void set_domain_attribute(struct sched_domain *sd,
6766 struct sched_domain_attr *attr)
6770 if (!attr || attr->relax_domain_level < 0) {
6771 if (default_relax_domain_level < 0)
6774 request = default_relax_domain_level;
6776 request = attr->relax_domain_level;
6777 if (request < sd->level) {
6778 /* turn off idle balance on this domain */
6779 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6781 /* turn on idle balance on this domain */
6782 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6786 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6787 const struct cpumask *cpu_map)
6790 case sa_sched_groups:
6791 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6792 d->sched_group_nodes = NULL;
6794 free_rootdomain(d->rd); /* fall through */
6796 free_cpumask_var(d->tmpmask); /* fall through */
6797 case sa_send_covered:
6798 free_cpumask_var(d->send_covered); /* fall through */
6799 case sa_this_core_map:
6800 free_cpumask_var(d->this_core_map); /* fall through */
6801 case sa_this_sibling_map:
6802 free_cpumask_var(d->this_sibling_map); /* fall through */
6804 free_cpumask_var(d->nodemask); /* fall through */
6805 case sa_sched_group_nodes:
6807 kfree(d->sched_group_nodes); /* fall through */
6809 free_cpumask_var(d->notcovered); /* fall through */
6811 free_cpumask_var(d->covered); /* fall through */
6813 free_cpumask_var(d->domainspan); /* fall through */
6820 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6821 const struct cpumask *cpu_map)
6824 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6826 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6827 return sa_domainspan;
6828 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6830 /* Allocate the per-node list of sched groups */
6831 d->sched_group_nodes = kcalloc(nr_node_ids,
6832 sizeof(struct sched_group *), GFP_KERNEL);
6833 if (!d->sched_group_nodes) {
6834 printk(KERN_WARNING "Can not alloc sched group node list\n");
6835 return sa_notcovered;
6837 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6839 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6840 return sa_sched_group_nodes;
6841 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6843 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6844 return sa_this_sibling_map;
6845 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6846 return sa_this_core_map;
6847 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6848 return sa_send_covered;
6849 d->rd = alloc_rootdomain();
6851 printk(KERN_WARNING "Cannot alloc root domain\n");
6854 return sa_rootdomain;
6857 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6858 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6860 struct sched_domain *sd = NULL;
6862 struct sched_domain *parent;
6865 if (cpumask_weight(cpu_map) >
6866 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6867 sd = &per_cpu(allnodes_domains, i).sd;
6868 SD_INIT(sd, ALLNODES);
6869 set_domain_attribute(sd, attr);
6870 cpumask_copy(sched_domain_span(sd), cpu_map);
6871 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6876 sd = &per_cpu(node_domains, i).sd;
6878 set_domain_attribute(sd, attr);
6879 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6880 sd->parent = parent;
6883 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6888 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6889 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6890 struct sched_domain *parent, int i)
6892 struct sched_domain *sd;
6893 sd = &per_cpu(phys_domains, i).sd;
6895 set_domain_attribute(sd, attr);
6896 cpumask_copy(sched_domain_span(sd), d->nodemask);
6897 sd->parent = parent;
6900 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6904 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6905 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6906 struct sched_domain *parent, int i)
6908 struct sched_domain *sd = parent;
6909 #ifdef CONFIG_SCHED_MC
6910 sd = &per_cpu(core_domains, i).sd;
6912 set_domain_attribute(sd, attr);
6913 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6914 sd->parent = parent;
6916 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6921 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6922 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6923 struct sched_domain *parent, int i)
6925 struct sched_domain *sd = parent;
6926 #ifdef CONFIG_SCHED_SMT
6927 sd = &per_cpu(cpu_domains, i).sd;
6928 SD_INIT(sd, SIBLING);
6929 set_domain_attribute(sd, attr);
6930 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6931 sd->parent = parent;
6933 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6938 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6939 const struct cpumask *cpu_map, int cpu)
6942 #ifdef CONFIG_SCHED_SMT
6943 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6944 cpumask_and(d->this_sibling_map, cpu_map,
6945 topology_thread_cpumask(cpu));
6946 if (cpu == cpumask_first(d->this_sibling_map))
6947 init_sched_build_groups(d->this_sibling_map, cpu_map,
6949 d->send_covered, d->tmpmask);
6952 #ifdef CONFIG_SCHED_MC
6953 case SD_LV_MC: /* set up multi-core groups */
6954 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6955 if (cpu == cpumask_first(d->this_core_map))
6956 init_sched_build_groups(d->this_core_map, cpu_map,
6958 d->send_covered, d->tmpmask);
6961 case SD_LV_CPU: /* set up physical groups */
6962 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
6963 if (!cpumask_empty(d->nodemask))
6964 init_sched_build_groups(d->nodemask, cpu_map,
6966 d->send_covered, d->tmpmask);
6969 case SD_LV_ALLNODES:
6970 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
6971 d->send_covered, d->tmpmask);
6980 * Build sched domains for a given set of cpus and attach the sched domains
6981 * to the individual cpus
6983 static int __build_sched_domains(const struct cpumask *cpu_map,
6984 struct sched_domain_attr *attr)
6986 enum s_alloc alloc_state = sa_none;
6988 struct sched_domain *sd;
6994 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6995 if (alloc_state != sa_rootdomain)
6997 alloc_state = sa_sched_groups;
7000 * Set up domains for cpus specified by the cpu_map.
7002 for_each_cpu(i, cpu_map) {
7003 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7006 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7007 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7008 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7009 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7012 for_each_cpu(i, cpu_map) {
7013 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7014 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7017 /* Set up physical groups */
7018 for (i = 0; i < nr_node_ids; i++)
7019 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7022 /* Set up node groups */
7024 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7026 for (i = 0; i < nr_node_ids; i++)
7027 if (build_numa_sched_groups(&d, cpu_map, i))
7031 /* Calculate CPU power for physical packages and nodes */
7032 #ifdef CONFIG_SCHED_SMT
7033 for_each_cpu(i, cpu_map) {
7034 sd = &per_cpu(cpu_domains, i).sd;
7035 init_sched_groups_power(i, sd);
7038 #ifdef CONFIG_SCHED_MC
7039 for_each_cpu(i, cpu_map) {
7040 sd = &per_cpu(core_domains, i).sd;
7041 init_sched_groups_power(i, sd);
7045 for_each_cpu(i, cpu_map) {
7046 sd = &per_cpu(phys_domains, i).sd;
7047 init_sched_groups_power(i, sd);
7051 for (i = 0; i < nr_node_ids; i++)
7052 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7054 if (d.sd_allnodes) {
7055 struct sched_group *sg;
7057 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7059 init_numa_sched_groups_power(sg);
7063 /* Attach the domains */
7064 for_each_cpu(i, cpu_map) {
7065 #ifdef CONFIG_SCHED_SMT
7066 sd = &per_cpu(cpu_domains, i).sd;
7067 #elif defined(CONFIG_SCHED_MC)
7068 sd = &per_cpu(core_domains, i).sd;
7070 sd = &per_cpu(phys_domains, i).sd;
7072 cpu_attach_domain(sd, d.rd, i);
7075 d.sched_group_nodes = NULL; /* don't free this we still need it */
7076 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7080 __free_domain_allocs(&d, alloc_state, cpu_map);
7084 static int build_sched_domains(const struct cpumask *cpu_map)
7086 return __build_sched_domains(cpu_map, NULL);
7089 static cpumask_var_t *doms_cur; /* current sched domains */
7090 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7091 static struct sched_domain_attr *dattr_cur;
7092 /* attribues of custom domains in 'doms_cur' */
7095 * Special case: If a kmalloc of a doms_cur partition (array of
7096 * cpumask) fails, then fallback to a single sched domain,
7097 * as determined by the single cpumask fallback_doms.
7099 static cpumask_var_t fallback_doms;
7102 * arch_update_cpu_topology lets virtualized architectures update the
7103 * cpu core maps. It is supposed to return 1 if the topology changed
7104 * or 0 if it stayed the same.
7106 int __attribute__((weak)) arch_update_cpu_topology(void)
7111 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7114 cpumask_var_t *doms;
7116 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7119 for (i = 0; i < ndoms; i++) {
7120 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7121 free_sched_domains(doms, i);
7128 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7131 for (i = 0; i < ndoms; i++)
7132 free_cpumask_var(doms[i]);
7137 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7138 * For now this just excludes isolated cpus, but could be used to
7139 * exclude other special cases in the future.
7141 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7145 arch_update_cpu_topology();
7147 doms_cur = alloc_sched_domains(ndoms_cur);
7149 doms_cur = &fallback_doms;
7150 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7152 err = build_sched_domains(doms_cur[0]);
7153 register_sched_domain_sysctl();
7158 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7159 struct cpumask *tmpmask)
7161 free_sched_groups(cpu_map, tmpmask);
7165 * Detach sched domains from a group of cpus specified in cpu_map
7166 * These cpus will now be attached to the NULL domain
7168 static void detach_destroy_domains(const struct cpumask *cpu_map)
7170 /* Save because hotplug lock held. */
7171 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7174 for_each_cpu(i, cpu_map)
7175 cpu_attach_domain(NULL, &def_root_domain, i);
7176 synchronize_sched();
7177 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7180 /* handle null as "default" */
7181 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7182 struct sched_domain_attr *new, int idx_new)
7184 struct sched_domain_attr tmp;
7191 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7192 new ? (new + idx_new) : &tmp,
7193 sizeof(struct sched_domain_attr));
7197 * Partition sched domains as specified by the 'ndoms_new'
7198 * cpumasks in the array doms_new[] of cpumasks. This compares
7199 * doms_new[] to the current sched domain partitioning, doms_cur[].
7200 * It destroys each deleted domain and builds each new domain.
7202 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7203 * The masks don't intersect (don't overlap.) We should setup one
7204 * sched domain for each mask. CPUs not in any of the cpumasks will
7205 * not be load balanced. If the same cpumask appears both in the
7206 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7209 * The passed in 'doms_new' should be allocated using
7210 * alloc_sched_domains. This routine takes ownership of it and will
7211 * free_sched_domains it when done with it. If the caller failed the
7212 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7213 * and partition_sched_domains() will fallback to the single partition
7214 * 'fallback_doms', it also forces the domains to be rebuilt.
7216 * If doms_new == NULL it will be replaced with cpu_online_mask.
7217 * ndoms_new == 0 is a special case for destroying existing domains,
7218 * and it will not create the default domain.
7220 * Call with hotplug lock held
7222 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7223 struct sched_domain_attr *dattr_new)
7228 mutex_lock(&sched_domains_mutex);
7230 /* always unregister in case we don't destroy any domains */
7231 unregister_sched_domain_sysctl();
7233 /* Let architecture update cpu core mappings. */
7234 new_topology = arch_update_cpu_topology();
7236 n = doms_new ? ndoms_new : 0;
7238 /* Destroy deleted domains */
7239 for (i = 0; i < ndoms_cur; i++) {
7240 for (j = 0; j < n && !new_topology; j++) {
7241 if (cpumask_equal(doms_cur[i], doms_new[j])
7242 && dattrs_equal(dattr_cur, i, dattr_new, j))
7245 /* no match - a current sched domain not in new doms_new[] */
7246 detach_destroy_domains(doms_cur[i]);
7251 if (doms_new == NULL) {
7253 doms_new = &fallback_doms;
7254 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7255 WARN_ON_ONCE(dattr_new);
7258 /* Build new domains */
7259 for (i = 0; i < ndoms_new; i++) {
7260 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7261 if (cpumask_equal(doms_new[i], doms_cur[j])
7262 && dattrs_equal(dattr_new, i, dattr_cur, j))
7265 /* no match - add a new doms_new */
7266 __build_sched_domains(doms_new[i],
7267 dattr_new ? dattr_new + i : NULL);
7272 /* Remember the new sched domains */
7273 if (doms_cur != &fallback_doms)
7274 free_sched_domains(doms_cur, ndoms_cur);
7275 kfree(dattr_cur); /* kfree(NULL) is safe */
7276 doms_cur = doms_new;
7277 dattr_cur = dattr_new;
7278 ndoms_cur = ndoms_new;
7280 register_sched_domain_sysctl();
7282 mutex_unlock(&sched_domains_mutex);
7285 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7286 static void arch_reinit_sched_domains(void)
7290 /* Destroy domains first to force the rebuild */
7291 partition_sched_domains(0, NULL, NULL);
7293 rebuild_sched_domains();
7297 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7299 unsigned int level = 0;
7301 if (sscanf(buf, "%u", &level) != 1)
7305 * level is always be positive so don't check for
7306 * level < POWERSAVINGS_BALANCE_NONE which is 0
7307 * What happens on 0 or 1 byte write,
7308 * need to check for count as well?
7311 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7315 sched_smt_power_savings = level;
7317 sched_mc_power_savings = level;
7319 arch_reinit_sched_domains();
7324 #ifdef CONFIG_SCHED_MC
7325 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7326 struct sysdev_class_attribute *attr,
7329 return sprintf(page, "%u\n", sched_mc_power_savings);
7331 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7332 struct sysdev_class_attribute *attr,
7333 const char *buf, size_t count)
7335 return sched_power_savings_store(buf, count, 0);
7337 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7338 sched_mc_power_savings_show,
7339 sched_mc_power_savings_store);
7342 #ifdef CONFIG_SCHED_SMT
7343 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7344 struct sysdev_class_attribute *attr,
7347 return sprintf(page, "%u\n", sched_smt_power_savings);
7349 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7350 struct sysdev_class_attribute *attr,
7351 const char *buf, size_t count)
7353 return sched_power_savings_store(buf, count, 1);
7355 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7356 sched_smt_power_savings_show,
7357 sched_smt_power_savings_store);
7360 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7364 #ifdef CONFIG_SCHED_SMT
7366 err = sysfs_create_file(&cls->kset.kobj,
7367 &attr_sched_smt_power_savings.attr);
7369 #ifdef CONFIG_SCHED_MC
7370 if (!err && mc_capable())
7371 err = sysfs_create_file(&cls->kset.kobj,
7372 &attr_sched_mc_power_savings.attr);
7376 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7378 #ifndef CONFIG_CPUSETS
7380 * Add online and remove offline CPUs from the scheduler domains.
7381 * When cpusets are enabled they take over this function.
7383 static int update_sched_domains(struct notifier_block *nfb,
7384 unsigned long action, void *hcpu)
7388 case CPU_ONLINE_FROZEN:
7389 case CPU_DOWN_PREPARE:
7390 case CPU_DOWN_PREPARE_FROZEN:
7391 case CPU_DOWN_FAILED:
7392 case CPU_DOWN_FAILED_FROZEN:
7393 partition_sched_domains(1, NULL, NULL);
7402 static int update_runtime(struct notifier_block *nfb,
7403 unsigned long action, void *hcpu)
7405 int cpu = (int)(long)hcpu;
7408 case CPU_DOWN_PREPARE:
7409 case CPU_DOWN_PREPARE_FROZEN:
7410 disable_runtime(cpu_rq(cpu));
7413 case CPU_DOWN_FAILED:
7414 case CPU_DOWN_FAILED_FROZEN:
7416 case CPU_ONLINE_FROZEN:
7417 enable_runtime(cpu_rq(cpu));
7425 void __init sched_init_smp(void)
7427 cpumask_var_t non_isolated_cpus;
7429 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7430 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7432 #if defined(CONFIG_NUMA)
7433 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7435 BUG_ON(sched_group_nodes_bycpu == NULL);
7438 mutex_lock(&sched_domains_mutex);
7439 arch_init_sched_domains(cpu_active_mask);
7440 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7441 if (cpumask_empty(non_isolated_cpus))
7442 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7443 mutex_unlock(&sched_domains_mutex);
7446 #ifndef CONFIG_CPUSETS
7447 /* XXX: Theoretical race here - CPU may be hotplugged now */
7448 hotcpu_notifier(update_sched_domains, 0);
7451 /* RT runtime code needs to handle some hotplug events */
7452 hotcpu_notifier(update_runtime, 0);
7456 /* Move init over to a non-isolated CPU */
7457 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7459 sched_init_granularity();
7460 free_cpumask_var(non_isolated_cpus);
7462 init_sched_rt_class();
7465 void __init sched_init_smp(void)
7467 sched_init_granularity();
7469 #endif /* CONFIG_SMP */
7471 const_debug unsigned int sysctl_timer_migration = 1;
7473 int in_sched_functions(unsigned long addr)
7475 return in_lock_functions(addr) ||
7476 (addr >= (unsigned long)__sched_text_start
7477 && addr < (unsigned long)__sched_text_end);
7480 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7482 cfs_rq->tasks_timeline = RB_ROOT;
7483 INIT_LIST_HEAD(&cfs_rq->tasks);
7484 #ifdef CONFIG_FAIR_GROUP_SCHED
7487 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7490 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7492 struct rt_prio_array *array;
7495 array = &rt_rq->active;
7496 for (i = 0; i < MAX_RT_PRIO; i++) {
7497 INIT_LIST_HEAD(array->queue + i);
7498 __clear_bit(i, array->bitmap);
7500 /* delimiter for bitsearch: */
7501 __set_bit(MAX_RT_PRIO, array->bitmap);
7503 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7504 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7506 rt_rq->highest_prio.next = MAX_RT_PRIO;
7510 rt_rq->rt_nr_migratory = 0;
7511 rt_rq->overloaded = 0;
7512 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7516 rt_rq->rt_throttled = 0;
7517 rt_rq->rt_runtime = 0;
7518 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7520 #ifdef CONFIG_RT_GROUP_SCHED
7521 rt_rq->rt_nr_boosted = 0;
7526 #ifdef CONFIG_FAIR_GROUP_SCHED
7527 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7528 struct sched_entity *se, int cpu, int add,
7529 struct sched_entity *parent)
7531 struct rq *rq = cpu_rq(cpu);
7532 tg->cfs_rq[cpu] = cfs_rq;
7533 init_cfs_rq(cfs_rq, rq);
7536 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7539 /* se could be NULL for init_task_group */
7544 se->cfs_rq = &rq->cfs;
7546 se->cfs_rq = parent->my_q;
7549 se->load.weight = tg->shares;
7550 se->load.inv_weight = 0;
7551 se->parent = parent;
7555 #ifdef CONFIG_RT_GROUP_SCHED
7556 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7557 struct sched_rt_entity *rt_se, int cpu, int add,
7558 struct sched_rt_entity *parent)
7560 struct rq *rq = cpu_rq(cpu);
7562 tg->rt_rq[cpu] = rt_rq;
7563 init_rt_rq(rt_rq, rq);
7565 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7567 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7569 tg->rt_se[cpu] = rt_se;
7574 rt_se->rt_rq = &rq->rt;
7576 rt_se->rt_rq = parent->my_q;
7578 rt_se->my_q = rt_rq;
7579 rt_se->parent = parent;
7580 INIT_LIST_HEAD(&rt_se->run_list);
7584 void __init sched_init(void)
7587 unsigned long alloc_size = 0, ptr;
7589 #ifdef CONFIG_FAIR_GROUP_SCHED
7590 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7592 #ifdef CONFIG_RT_GROUP_SCHED
7593 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7595 #ifdef CONFIG_CPUMASK_OFFSTACK
7596 alloc_size += num_possible_cpus() * cpumask_size();
7599 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7601 #ifdef CONFIG_FAIR_GROUP_SCHED
7602 init_task_group.se = (struct sched_entity **)ptr;
7603 ptr += nr_cpu_ids * sizeof(void **);
7605 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7606 ptr += nr_cpu_ids * sizeof(void **);
7608 #endif /* CONFIG_FAIR_GROUP_SCHED */
7609 #ifdef CONFIG_RT_GROUP_SCHED
7610 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7611 ptr += nr_cpu_ids * sizeof(void **);
7613 init_task_group.rt_rq = (struct rt_rq **)ptr;
7614 ptr += nr_cpu_ids * sizeof(void **);
7616 #endif /* CONFIG_RT_GROUP_SCHED */
7617 #ifdef CONFIG_CPUMASK_OFFSTACK
7618 for_each_possible_cpu(i) {
7619 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7620 ptr += cpumask_size();
7622 #endif /* CONFIG_CPUMASK_OFFSTACK */
7626 init_defrootdomain();
7629 init_rt_bandwidth(&def_rt_bandwidth,
7630 global_rt_period(), global_rt_runtime());
7632 #ifdef CONFIG_RT_GROUP_SCHED
7633 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7634 global_rt_period(), global_rt_runtime());
7635 #endif /* CONFIG_RT_GROUP_SCHED */
7637 #ifdef CONFIG_CGROUP_SCHED
7638 list_add(&init_task_group.list, &task_groups);
7639 INIT_LIST_HEAD(&init_task_group.children);
7641 #endif /* CONFIG_CGROUP_SCHED */
7643 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7644 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7645 __alignof__(unsigned long));
7647 for_each_possible_cpu(i) {
7651 raw_spin_lock_init(&rq->lock);
7653 rq->calc_load_active = 0;
7654 rq->calc_load_update = jiffies + LOAD_FREQ;
7655 init_cfs_rq(&rq->cfs, rq);
7656 init_rt_rq(&rq->rt, rq);
7657 #ifdef CONFIG_FAIR_GROUP_SCHED
7658 init_task_group.shares = init_task_group_load;
7659 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7660 #ifdef CONFIG_CGROUP_SCHED
7662 * How much cpu bandwidth does init_task_group get?
7664 * In case of task-groups formed thr' the cgroup filesystem, it
7665 * gets 100% of the cpu resources in the system. This overall
7666 * system cpu resource is divided among the tasks of
7667 * init_task_group and its child task-groups in a fair manner,
7668 * based on each entity's (task or task-group's) weight
7669 * (se->load.weight).
7671 * In other words, if init_task_group has 10 tasks of weight
7672 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7673 * then A0's share of the cpu resource is:
7675 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7677 * We achieve this by letting init_task_group's tasks sit
7678 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7680 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7682 #endif /* CONFIG_FAIR_GROUP_SCHED */
7684 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7685 #ifdef CONFIG_RT_GROUP_SCHED
7686 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7687 #ifdef CONFIG_CGROUP_SCHED
7688 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7692 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7693 rq->cpu_load[j] = 0;
7697 rq->post_schedule = 0;
7698 rq->active_balance = 0;
7699 rq->next_balance = jiffies;
7703 rq->migration_thread = NULL;
7705 rq->avg_idle = 2*sysctl_sched_migration_cost;
7706 INIT_LIST_HEAD(&rq->migration_queue);
7707 rq_attach_root(rq, &def_root_domain);
7710 atomic_set(&rq->nr_iowait, 0);
7713 set_load_weight(&init_task);
7715 #ifdef CONFIG_PREEMPT_NOTIFIERS
7716 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7720 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7723 #ifdef CONFIG_RT_MUTEXES
7724 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7728 * The boot idle thread does lazy MMU switching as well:
7730 atomic_inc(&init_mm.mm_count);
7731 enter_lazy_tlb(&init_mm, current);
7734 * Make us the idle thread. Technically, schedule() should not be
7735 * called from this thread, however somewhere below it might be,
7736 * but because we are the idle thread, we just pick up running again
7737 * when this runqueue becomes "idle".
7739 init_idle(current, smp_processor_id());
7741 calc_load_update = jiffies + LOAD_FREQ;
7744 * During early bootup we pretend to be a normal task:
7746 current->sched_class = &fair_sched_class;
7748 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7749 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7752 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7753 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7755 /* May be allocated at isolcpus cmdline parse time */
7756 if (cpu_isolated_map == NULL)
7757 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7762 scheduler_running = 1;
7765 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7766 static inline int preempt_count_equals(int preempt_offset)
7768 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7770 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7773 void __might_sleep(const char *file, int line, int preempt_offset)
7776 static unsigned long prev_jiffy; /* ratelimiting */
7778 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7779 system_state != SYSTEM_RUNNING || oops_in_progress)
7781 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7783 prev_jiffy = jiffies;
7786 "BUG: sleeping function called from invalid context at %s:%d\n",
7789 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7790 in_atomic(), irqs_disabled(),
7791 current->pid, current->comm);
7793 debug_show_held_locks(current);
7794 if (irqs_disabled())
7795 print_irqtrace_events(current);
7799 EXPORT_SYMBOL(__might_sleep);
7802 #ifdef CONFIG_MAGIC_SYSRQ
7803 static void normalize_task(struct rq *rq, struct task_struct *p)
7807 on_rq = p->se.on_rq;
7809 deactivate_task(rq, p, 0);
7810 __setscheduler(rq, p, SCHED_NORMAL, 0);
7812 activate_task(rq, p, 0);
7813 resched_task(rq->curr);
7817 void normalize_rt_tasks(void)
7819 struct task_struct *g, *p;
7820 unsigned long flags;
7823 read_lock_irqsave(&tasklist_lock, flags);
7824 do_each_thread(g, p) {
7826 * Only normalize user tasks:
7831 p->se.exec_start = 0;
7832 #ifdef CONFIG_SCHEDSTATS
7833 p->se.statistics.wait_start = 0;
7834 p->se.statistics.sleep_start = 0;
7835 p->se.statistics.block_start = 0;
7840 * Renice negative nice level userspace
7843 if (TASK_NICE(p) < 0 && p->mm)
7844 set_user_nice(p, 0);
7848 raw_spin_lock(&p->pi_lock);
7849 rq = __task_rq_lock(p);
7851 normalize_task(rq, p);
7853 __task_rq_unlock(rq);
7854 raw_spin_unlock(&p->pi_lock);
7855 } while_each_thread(g, p);
7857 read_unlock_irqrestore(&tasklist_lock, flags);
7860 #endif /* CONFIG_MAGIC_SYSRQ */
7864 * These functions are only useful for the IA64 MCA handling.
7866 * They can only be called when the whole system has been
7867 * stopped - every CPU needs to be quiescent, and no scheduling
7868 * activity can take place. Using them for anything else would
7869 * be a serious bug, and as a result, they aren't even visible
7870 * under any other configuration.
7874 * curr_task - return the current task for a given cpu.
7875 * @cpu: the processor in question.
7877 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7879 struct task_struct *curr_task(int cpu)
7881 return cpu_curr(cpu);
7885 * set_curr_task - set the current task for a given cpu.
7886 * @cpu: the processor in question.
7887 * @p: the task pointer to set.
7889 * Description: This function must only be used when non-maskable interrupts
7890 * are serviced on a separate stack. It allows the architecture to switch the
7891 * notion of the current task on a cpu in a non-blocking manner. This function
7892 * must be called with all CPU's synchronized, and interrupts disabled, the
7893 * and caller must save the original value of the current task (see
7894 * curr_task() above) and restore that value before reenabling interrupts and
7895 * re-starting the system.
7897 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7899 void set_curr_task(int cpu, struct task_struct *p)
7906 #ifdef CONFIG_FAIR_GROUP_SCHED
7907 static void free_fair_sched_group(struct task_group *tg)
7911 for_each_possible_cpu(i) {
7913 kfree(tg->cfs_rq[i]);
7923 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7925 struct cfs_rq *cfs_rq;
7926 struct sched_entity *se;
7930 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7933 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7937 tg->shares = NICE_0_LOAD;
7939 for_each_possible_cpu(i) {
7942 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7943 GFP_KERNEL, cpu_to_node(i));
7947 se = kzalloc_node(sizeof(struct sched_entity),
7948 GFP_KERNEL, cpu_to_node(i));
7952 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
7963 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7965 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7966 &cpu_rq(cpu)->leaf_cfs_rq_list);
7969 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7971 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7973 #else /* !CONFG_FAIR_GROUP_SCHED */
7974 static inline void free_fair_sched_group(struct task_group *tg)
7979 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7984 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7988 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7991 #endif /* CONFIG_FAIR_GROUP_SCHED */
7993 #ifdef CONFIG_RT_GROUP_SCHED
7994 static void free_rt_sched_group(struct task_group *tg)
7998 destroy_rt_bandwidth(&tg->rt_bandwidth);
8000 for_each_possible_cpu(i) {
8002 kfree(tg->rt_rq[i]);
8004 kfree(tg->rt_se[i]);
8012 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8014 struct rt_rq *rt_rq;
8015 struct sched_rt_entity *rt_se;
8019 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8022 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8026 init_rt_bandwidth(&tg->rt_bandwidth,
8027 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8029 for_each_possible_cpu(i) {
8032 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8033 GFP_KERNEL, cpu_to_node(i));
8037 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8038 GFP_KERNEL, cpu_to_node(i));
8042 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8053 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8055 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8056 &cpu_rq(cpu)->leaf_rt_rq_list);
8059 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8061 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8063 #else /* !CONFIG_RT_GROUP_SCHED */
8064 static inline void free_rt_sched_group(struct task_group *tg)
8069 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8074 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8078 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8081 #endif /* CONFIG_RT_GROUP_SCHED */
8083 #ifdef CONFIG_CGROUP_SCHED
8084 static void free_sched_group(struct task_group *tg)
8086 free_fair_sched_group(tg);
8087 free_rt_sched_group(tg);
8091 /* allocate runqueue etc for a new task group */
8092 struct task_group *sched_create_group(struct task_group *parent)
8094 struct task_group *tg;
8095 unsigned long flags;
8098 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8100 return ERR_PTR(-ENOMEM);
8102 if (!alloc_fair_sched_group(tg, parent))
8105 if (!alloc_rt_sched_group(tg, parent))
8108 spin_lock_irqsave(&task_group_lock, flags);
8109 for_each_possible_cpu(i) {
8110 register_fair_sched_group(tg, i);
8111 register_rt_sched_group(tg, i);
8113 list_add_rcu(&tg->list, &task_groups);
8115 WARN_ON(!parent); /* root should already exist */
8117 tg->parent = parent;
8118 INIT_LIST_HEAD(&tg->children);
8119 list_add_rcu(&tg->siblings, &parent->children);
8120 spin_unlock_irqrestore(&task_group_lock, flags);
8125 free_sched_group(tg);
8126 return ERR_PTR(-ENOMEM);
8129 /* rcu callback to free various structures associated with a task group */
8130 static void free_sched_group_rcu(struct rcu_head *rhp)
8132 /* now it should be safe to free those cfs_rqs */
8133 free_sched_group(container_of(rhp, struct task_group, rcu));
8136 /* Destroy runqueue etc associated with a task group */
8137 void sched_destroy_group(struct task_group *tg)
8139 unsigned long flags;
8142 spin_lock_irqsave(&task_group_lock, flags);
8143 for_each_possible_cpu(i) {
8144 unregister_fair_sched_group(tg, i);
8145 unregister_rt_sched_group(tg, i);
8147 list_del_rcu(&tg->list);
8148 list_del_rcu(&tg->siblings);
8149 spin_unlock_irqrestore(&task_group_lock, flags);
8151 /* wait for possible concurrent references to cfs_rqs complete */
8152 call_rcu(&tg->rcu, free_sched_group_rcu);
8155 /* change task's runqueue when it moves between groups.
8156 * The caller of this function should have put the task in its new group
8157 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8158 * reflect its new group.
8160 void sched_move_task(struct task_struct *tsk)
8163 unsigned long flags;
8166 rq = task_rq_lock(tsk, &flags);
8168 running = task_current(rq, tsk);
8169 on_rq = tsk->se.on_rq;
8172 dequeue_task(rq, tsk, 0);
8173 if (unlikely(running))
8174 tsk->sched_class->put_prev_task(rq, tsk);
8176 set_task_rq(tsk, task_cpu(tsk));
8178 #ifdef CONFIG_FAIR_GROUP_SCHED
8179 if (tsk->sched_class->moved_group)
8180 tsk->sched_class->moved_group(tsk, on_rq);
8183 if (unlikely(running))
8184 tsk->sched_class->set_curr_task(rq);
8186 enqueue_task(rq, tsk, 0);
8188 task_rq_unlock(rq, &flags);
8190 #endif /* CONFIG_CGROUP_SCHED */
8192 #ifdef CONFIG_FAIR_GROUP_SCHED
8193 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8195 struct cfs_rq *cfs_rq = se->cfs_rq;
8200 dequeue_entity(cfs_rq, se, 0);
8202 se->load.weight = shares;
8203 se->load.inv_weight = 0;
8206 enqueue_entity(cfs_rq, se, 0);
8209 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8211 struct cfs_rq *cfs_rq = se->cfs_rq;
8212 struct rq *rq = cfs_rq->rq;
8213 unsigned long flags;
8215 raw_spin_lock_irqsave(&rq->lock, flags);
8216 __set_se_shares(se, shares);
8217 raw_spin_unlock_irqrestore(&rq->lock, flags);
8220 static DEFINE_MUTEX(shares_mutex);
8222 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8225 unsigned long flags;
8228 * We can't change the weight of the root cgroup.
8233 if (shares < MIN_SHARES)
8234 shares = MIN_SHARES;
8235 else if (shares > MAX_SHARES)
8236 shares = MAX_SHARES;
8238 mutex_lock(&shares_mutex);
8239 if (tg->shares == shares)
8242 spin_lock_irqsave(&task_group_lock, flags);
8243 for_each_possible_cpu(i)
8244 unregister_fair_sched_group(tg, i);
8245 list_del_rcu(&tg->siblings);
8246 spin_unlock_irqrestore(&task_group_lock, flags);
8248 /* wait for any ongoing reference to this group to finish */
8249 synchronize_sched();
8252 * Now we are free to modify the group's share on each cpu
8253 * w/o tripping rebalance_share or load_balance_fair.
8255 tg->shares = shares;
8256 for_each_possible_cpu(i) {
8260 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8261 set_se_shares(tg->se[i], shares);
8265 * Enable load balance activity on this group, by inserting it back on
8266 * each cpu's rq->leaf_cfs_rq_list.
8268 spin_lock_irqsave(&task_group_lock, flags);
8269 for_each_possible_cpu(i)
8270 register_fair_sched_group(tg, i);
8271 list_add_rcu(&tg->siblings, &tg->parent->children);
8272 spin_unlock_irqrestore(&task_group_lock, flags);
8274 mutex_unlock(&shares_mutex);
8278 unsigned long sched_group_shares(struct task_group *tg)
8284 #ifdef CONFIG_RT_GROUP_SCHED
8286 * Ensure that the real time constraints are schedulable.
8288 static DEFINE_MUTEX(rt_constraints_mutex);
8290 static unsigned long to_ratio(u64 period, u64 runtime)
8292 if (runtime == RUNTIME_INF)
8295 return div64_u64(runtime << 20, period);
8298 /* Must be called with tasklist_lock held */
8299 static inline int tg_has_rt_tasks(struct task_group *tg)
8301 struct task_struct *g, *p;
8303 do_each_thread(g, p) {
8304 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8306 } while_each_thread(g, p);
8311 struct rt_schedulable_data {
8312 struct task_group *tg;
8317 static int tg_schedulable(struct task_group *tg, void *data)
8319 struct rt_schedulable_data *d = data;
8320 struct task_group *child;
8321 unsigned long total, sum = 0;
8322 u64 period, runtime;
8324 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8325 runtime = tg->rt_bandwidth.rt_runtime;
8328 period = d->rt_period;
8329 runtime = d->rt_runtime;
8333 * Cannot have more runtime than the period.
8335 if (runtime > period && runtime != RUNTIME_INF)
8339 * Ensure we don't starve existing RT tasks.
8341 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8344 total = to_ratio(period, runtime);
8347 * Nobody can have more than the global setting allows.
8349 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8353 * The sum of our children's runtime should not exceed our own.
8355 list_for_each_entry_rcu(child, &tg->children, siblings) {
8356 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8357 runtime = child->rt_bandwidth.rt_runtime;
8359 if (child == d->tg) {
8360 period = d->rt_period;
8361 runtime = d->rt_runtime;
8364 sum += to_ratio(period, runtime);
8373 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8375 struct rt_schedulable_data data = {
8377 .rt_period = period,
8378 .rt_runtime = runtime,
8381 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8384 static int tg_set_bandwidth(struct task_group *tg,
8385 u64 rt_period, u64 rt_runtime)
8389 mutex_lock(&rt_constraints_mutex);
8390 read_lock(&tasklist_lock);
8391 err = __rt_schedulable(tg, rt_period, rt_runtime);
8395 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8396 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8397 tg->rt_bandwidth.rt_runtime = rt_runtime;
8399 for_each_possible_cpu(i) {
8400 struct rt_rq *rt_rq = tg->rt_rq[i];
8402 raw_spin_lock(&rt_rq->rt_runtime_lock);
8403 rt_rq->rt_runtime = rt_runtime;
8404 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8406 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8408 read_unlock(&tasklist_lock);
8409 mutex_unlock(&rt_constraints_mutex);
8414 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8416 u64 rt_runtime, rt_period;
8418 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8419 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8420 if (rt_runtime_us < 0)
8421 rt_runtime = RUNTIME_INF;
8423 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8426 long sched_group_rt_runtime(struct task_group *tg)
8430 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8433 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8434 do_div(rt_runtime_us, NSEC_PER_USEC);
8435 return rt_runtime_us;
8438 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8440 u64 rt_runtime, rt_period;
8442 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8443 rt_runtime = tg->rt_bandwidth.rt_runtime;
8448 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8451 long sched_group_rt_period(struct task_group *tg)
8455 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8456 do_div(rt_period_us, NSEC_PER_USEC);
8457 return rt_period_us;
8460 static int sched_rt_global_constraints(void)
8462 u64 runtime, period;
8465 if (sysctl_sched_rt_period <= 0)
8468 runtime = global_rt_runtime();
8469 period = global_rt_period();
8472 * Sanity check on the sysctl variables.
8474 if (runtime > period && runtime != RUNTIME_INF)
8477 mutex_lock(&rt_constraints_mutex);
8478 read_lock(&tasklist_lock);
8479 ret = __rt_schedulable(NULL, 0, 0);
8480 read_unlock(&tasklist_lock);
8481 mutex_unlock(&rt_constraints_mutex);
8486 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8488 /* Don't accept realtime tasks when there is no way for them to run */
8489 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8495 #else /* !CONFIG_RT_GROUP_SCHED */
8496 static int sched_rt_global_constraints(void)
8498 unsigned long flags;
8501 if (sysctl_sched_rt_period <= 0)
8505 * There's always some RT tasks in the root group
8506 * -- migration, kstopmachine etc..
8508 if (sysctl_sched_rt_runtime == 0)
8511 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8512 for_each_possible_cpu(i) {
8513 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8515 raw_spin_lock(&rt_rq->rt_runtime_lock);
8516 rt_rq->rt_runtime = global_rt_runtime();
8517 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8519 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8523 #endif /* CONFIG_RT_GROUP_SCHED */
8525 int sched_rt_handler(struct ctl_table *table, int write,
8526 void __user *buffer, size_t *lenp,
8530 int old_period, old_runtime;
8531 static DEFINE_MUTEX(mutex);
8534 old_period = sysctl_sched_rt_period;
8535 old_runtime = sysctl_sched_rt_runtime;
8537 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8539 if (!ret && write) {
8540 ret = sched_rt_global_constraints();
8542 sysctl_sched_rt_period = old_period;
8543 sysctl_sched_rt_runtime = old_runtime;
8545 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8546 def_rt_bandwidth.rt_period =
8547 ns_to_ktime(global_rt_period());
8550 mutex_unlock(&mutex);
8555 #ifdef CONFIG_CGROUP_SCHED
8557 /* return corresponding task_group object of a cgroup */
8558 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8560 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8561 struct task_group, css);
8564 static struct cgroup_subsys_state *
8565 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8567 struct task_group *tg, *parent;
8569 if (!cgrp->parent) {
8570 /* This is early initialization for the top cgroup */
8571 return &init_task_group.css;
8574 parent = cgroup_tg(cgrp->parent);
8575 tg = sched_create_group(parent);
8577 return ERR_PTR(-ENOMEM);
8583 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8585 struct task_group *tg = cgroup_tg(cgrp);
8587 sched_destroy_group(tg);
8591 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8593 #ifdef CONFIG_RT_GROUP_SCHED
8594 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8597 /* We don't support RT-tasks being in separate groups */
8598 if (tsk->sched_class != &fair_sched_class)
8605 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8606 struct task_struct *tsk, bool threadgroup)
8608 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8612 struct task_struct *c;
8614 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8615 retval = cpu_cgroup_can_attach_task(cgrp, c);
8627 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8628 struct cgroup *old_cont, struct task_struct *tsk,
8631 sched_move_task(tsk);
8633 struct task_struct *c;
8635 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8642 #ifdef CONFIG_FAIR_GROUP_SCHED
8643 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8646 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8649 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8651 struct task_group *tg = cgroup_tg(cgrp);
8653 return (u64) tg->shares;
8655 #endif /* CONFIG_FAIR_GROUP_SCHED */
8657 #ifdef CONFIG_RT_GROUP_SCHED
8658 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8661 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8664 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8666 return sched_group_rt_runtime(cgroup_tg(cgrp));
8669 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8672 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8675 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8677 return sched_group_rt_period(cgroup_tg(cgrp));
8679 #endif /* CONFIG_RT_GROUP_SCHED */
8681 static struct cftype cpu_files[] = {
8682 #ifdef CONFIG_FAIR_GROUP_SCHED
8685 .read_u64 = cpu_shares_read_u64,
8686 .write_u64 = cpu_shares_write_u64,
8689 #ifdef CONFIG_RT_GROUP_SCHED
8691 .name = "rt_runtime_us",
8692 .read_s64 = cpu_rt_runtime_read,
8693 .write_s64 = cpu_rt_runtime_write,
8696 .name = "rt_period_us",
8697 .read_u64 = cpu_rt_period_read_uint,
8698 .write_u64 = cpu_rt_period_write_uint,
8703 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8705 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8708 struct cgroup_subsys cpu_cgroup_subsys = {
8710 .create = cpu_cgroup_create,
8711 .destroy = cpu_cgroup_destroy,
8712 .can_attach = cpu_cgroup_can_attach,
8713 .attach = cpu_cgroup_attach,
8714 .populate = cpu_cgroup_populate,
8715 .subsys_id = cpu_cgroup_subsys_id,
8719 #endif /* CONFIG_CGROUP_SCHED */
8721 #ifdef CONFIG_CGROUP_CPUACCT
8724 * CPU accounting code for task groups.
8726 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8727 * (balbir@in.ibm.com).
8730 /* track cpu usage of a group of tasks and its child groups */
8732 struct cgroup_subsys_state css;
8733 /* cpuusage holds pointer to a u64-type object on every cpu */
8734 u64 __percpu *cpuusage;
8735 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8736 struct cpuacct *parent;
8739 struct cgroup_subsys cpuacct_subsys;
8741 /* return cpu accounting group corresponding to this container */
8742 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8744 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8745 struct cpuacct, css);
8748 /* return cpu accounting group to which this task belongs */
8749 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8751 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8752 struct cpuacct, css);
8755 /* create a new cpu accounting group */
8756 static struct cgroup_subsys_state *cpuacct_create(
8757 struct cgroup_subsys *ss, struct cgroup *cgrp)
8759 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8765 ca->cpuusage = alloc_percpu(u64);
8769 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8770 if (percpu_counter_init(&ca->cpustat[i], 0))
8771 goto out_free_counters;
8774 ca->parent = cgroup_ca(cgrp->parent);
8780 percpu_counter_destroy(&ca->cpustat[i]);
8781 free_percpu(ca->cpuusage);
8785 return ERR_PTR(-ENOMEM);
8788 /* destroy an existing cpu accounting group */
8790 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8792 struct cpuacct *ca = cgroup_ca(cgrp);
8795 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8796 percpu_counter_destroy(&ca->cpustat[i]);
8797 free_percpu(ca->cpuusage);
8801 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8803 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8806 #ifndef CONFIG_64BIT
8808 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8810 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8812 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8820 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8822 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8824 #ifndef CONFIG_64BIT
8826 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8828 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8830 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8836 /* return total cpu usage (in nanoseconds) of a group */
8837 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8839 struct cpuacct *ca = cgroup_ca(cgrp);
8840 u64 totalcpuusage = 0;
8843 for_each_present_cpu(i)
8844 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8846 return totalcpuusage;
8849 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8852 struct cpuacct *ca = cgroup_ca(cgrp);
8861 for_each_present_cpu(i)
8862 cpuacct_cpuusage_write(ca, i, 0);
8868 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8871 struct cpuacct *ca = cgroup_ca(cgroup);
8875 for_each_present_cpu(i) {
8876 percpu = cpuacct_cpuusage_read(ca, i);
8877 seq_printf(m, "%llu ", (unsigned long long) percpu);
8879 seq_printf(m, "\n");
8883 static const char *cpuacct_stat_desc[] = {
8884 [CPUACCT_STAT_USER] = "user",
8885 [CPUACCT_STAT_SYSTEM] = "system",
8888 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8889 struct cgroup_map_cb *cb)
8891 struct cpuacct *ca = cgroup_ca(cgrp);
8894 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8895 s64 val = percpu_counter_read(&ca->cpustat[i]);
8896 val = cputime64_to_clock_t(val);
8897 cb->fill(cb, cpuacct_stat_desc[i], val);
8902 static struct cftype files[] = {
8905 .read_u64 = cpuusage_read,
8906 .write_u64 = cpuusage_write,
8909 .name = "usage_percpu",
8910 .read_seq_string = cpuacct_percpu_seq_read,
8914 .read_map = cpuacct_stats_show,
8918 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8920 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8924 * charge this task's execution time to its accounting group.
8926 * called with rq->lock held.
8928 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8933 if (unlikely(!cpuacct_subsys.active))
8936 cpu = task_cpu(tsk);
8942 for (; ca; ca = ca->parent) {
8943 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8944 *cpuusage += cputime;
8951 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8952 * in cputime_t units. As a result, cpuacct_update_stats calls
8953 * percpu_counter_add with values large enough to always overflow the
8954 * per cpu batch limit causing bad SMP scalability.
8956 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8957 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8958 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8961 #define CPUACCT_BATCH \
8962 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8964 #define CPUACCT_BATCH 0
8968 * Charge the system/user time to the task's accounting group.
8970 static void cpuacct_update_stats(struct task_struct *tsk,
8971 enum cpuacct_stat_index idx, cputime_t val)
8974 int batch = CPUACCT_BATCH;
8976 if (unlikely(!cpuacct_subsys.active))
8983 __percpu_counter_add(&ca->cpustat[idx], val, batch);
8989 struct cgroup_subsys cpuacct_subsys = {
8991 .create = cpuacct_create,
8992 .destroy = cpuacct_destroy,
8993 .populate = cpuacct_populate,
8994 .subsys_id = cpuacct_subsys_id,
8996 #endif /* CONFIG_CGROUP_CPUACCT */
9000 int rcu_expedited_torture_stats(char *page)
9004 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9006 void synchronize_sched_expedited(void)
9009 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9011 #else /* #ifndef CONFIG_SMP */
9013 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9014 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9016 #define RCU_EXPEDITED_STATE_POST -2
9017 #define RCU_EXPEDITED_STATE_IDLE -1
9019 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9021 int rcu_expedited_torture_stats(char *page)
9026 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9027 for_each_online_cpu(cpu) {
9028 cnt += sprintf(&page[cnt], " %d:%d",
9029 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9031 cnt += sprintf(&page[cnt], "\n");
9034 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9036 static long synchronize_sched_expedited_count;
9039 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9040 * approach to force grace period to end quickly. This consumes
9041 * significant time on all CPUs, and is thus not recommended for
9042 * any sort of common-case code.
9044 * Note that it is illegal to call this function while holding any
9045 * lock that is acquired by a CPU-hotplug notifier. Failing to
9046 * observe this restriction will result in deadlock.
9048 void synchronize_sched_expedited(void)
9051 unsigned long flags;
9052 bool need_full_sync = 0;
9054 struct migration_req *req;
9058 smp_mb(); /* ensure prior mod happens before capturing snap. */
9059 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9061 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9063 if (trycount++ < 10)
9064 udelay(trycount * num_online_cpus());
9066 synchronize_sched();
9069 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9070 smp_mb(); /* ensure test happens before caller kfree */
9075 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9076 for_each_online_cpu(cpu) {
9078 req = &per_cpu(rcu_migration_req, cpu);
9079 init_completion(&req->done);
9081 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9082 raw_spin_lock_irqsave(&rq->lock, flags);
9083 list_add(&req->list, &rq->migration_queue);
9084 raw_spin_unlock_irqrestore(&rq->lock, flags);
9085 wake_up_process(rq->migration_thread);
9087 for_each_online_cpu(cpu) {
9088 rcu_expedited_state = cpu;
9089 req = &per_cpu(rcu_migration_req, cpu);
9091 wait_for_completion(&req->done);
9092 raw_spin_lock_irqsave(&rq->lock, flags);
9093 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9095 req->dest_cpu = RCU_MIGRATION_IDLE;
9096 raw_spin_unlock_irqrestore(&rq->lock, flags);
9098 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9099 synchronize_sched_expedited_count++;
9100 mutex_unlock(&rcu_sched_expedited_mutex);
9103 synchronize_sched();
9105 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9107 #endif /* #else #ifndef CONFIG_SMP */