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_counter.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/reciprocal_div.h>
68 #include <linux/unistd.h>
69 #include <linux/pagemap.h>
70 #include <linux/hrtimer.h>
71 #include <linux/tick.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.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)
125 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
128 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129 * Since cpu_power is a 'constant', we can use a reciprocal divide.
131 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
133 return reciprocal_divide(load, sg->reciprocal_cpu_power);
137 * Each time a sched group cpu_power is changed,
138 * we must compute its reciprocal value
140 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
142 sg->__cpu_power += val;
143 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
147 static inline int rt_policy(int policy)
149 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
154 static inline int task_has_rt_policy(struct task_struct *p)
156 return rt_policy(p->policy);
160 * This is the priority-queue data structure of the RT scheduling class:
162 struct rt_prio_array {
163 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
164 struct list_head queue[MAX_RT_PRIO];
167 struct rt_bandwidth {
168 /* nests inside the rq lock: */
169 spinlock_t rt_runtime_lock;
172 struct hrtimer rt_period_timer;
175 static struct rt_bandwidth def_rt_bandwidth;
177 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
179 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
181 struct rt_bandwidth *rt_b =
182 container_of(timer, struct rt_bandwidth, rt_period_timer);
188 now = hrtimer_cb_get_time(timer);
189 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
194 idle = do_sched_rt_period_timer(rt_b, overrun);
197 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
201 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
203 rt_b->rt_period = ns_to_ktime(period);
204 rt_b->rt_runtime = runtime;
206 spin_lock_init(&rt_b->rt_runtime_lock);
208 hrtimer_init(&rt_b->rt_period_timer,
209 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
210 rt_b->rt_period_timer.function = sched_rt_period_timer;
213 static inline int rt_bandwidth_enabled(void)
215 return sysctl_sched_rt_runtime >= 0;
218 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 spin_lock(&rt_b->rt_runtime_lock);
233 if (hrtimer_active(&rt_b->rt_period_timer))
236 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
237 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
239 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
240 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
241 delta = ktime_to_ns(ktime_sub(hard, soft));
242 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
243 HRTIMER_MODE_ABS_PINNED, 0);
245 spin_unlock(&rt_b->rt_runtime_lock);
248 #ifdef CONFIG_RT_GROUP_SCHED
249 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
251 hrtimer_cancel(&rt_b->rt_period_timer);
256 * sched_domains_mutex serializes calls to arch_init_sched_domains,
257 * detach_destroy_domains and partition_sched_domains.
259 static DEFINE_MUTEX(sched_domains_mutex);
261 #ifdef CONFIG_GROUP_SCHED
263 #include <linux/cgroup.h>
267 static LIST_HEAD(task_groups);
269 /* task group related information */
271 #ifdef CONFIG_CGROUP_SCHED
272 struct cgroup_subsys_state css;
275 #ifdef CONFIG_USER_SCHED
279 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* schedulable entities of this group on each cpu */
281 struct sched_entity **se;
282 /* runqueue "owned" by this group on each cpu */
283 struct cfs_rq **cfs_rq;
284 unsigned long shares;
287 #ifdef CONFIG_RT_GROUP_SCHED
288 struct sched_rt_entity **rt_se;
289 struct rt_rq **rt_rq;
291 struct rt_bandwidth rt_bandwidth;
295 struct list_head list;
297 struct task_group *parent;
298 struct list_head siblings;
299 struct list_head children;
302 #ifdef CONFIG_USER_SCHED
304 /* Helper function to pass uid information to create_sched_user() */
305 void set_tg_uid(struct user_struct *user)
307 user->tg->uid = user->uid;
312 * Every UID task group (including init_task_group aka UID-0) will
313 * be a child to this group.
315 struct task_group root_task_group;
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 /* Default task group's sched entity on each cpu */
319 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
320 /* Default task group's cfs_rq on each cpu */
321 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_FAIR_GROUP_SCHED */
324 #ifdef CONFIG_RT_GROUP_SCHED
325 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
326 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
327 #endif /* CONFIG_RT_GROUP_SCHED */
328 #else /* !CONFIG_USER_SCHED */
329 #define root_task_group init_task_group
330 #endif /* CONFIG_USER_SCHED */
332 /* task_group_lock serializes add/remove of task groups and also changes to
333 * a task group's cpu shares.
335 static DEFINE_SPINLOCK(task_group_lock);
338 static int root_task_group_empty(void)
340 return list_empty(&root_task_group.children);
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 #ifdef CONFIG_USER_SCHED
346 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
347 #else /* !CONFIG_USER_SCHED */
348 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
349 #endif /* CONFIG_USER_SCHED */
352 * A weight of 0 or 1 can cause arithmetics problems.
353 * A weight of a cfs_rq is the sum of weights of which entities
354 * are queued on this cfs_rq, so a weight of a entity should not be
355 * too large, so as the shares value of a task group.
356 * (The default weight is 1024 - so there's no practical
357 * limitation from this.)
360 #define MAX_SHARES (1UL << 18)
362 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
365 /* Default task group.
366 * Every task in system belong to this group at bootup.
368 struct task_group init_task_group;
370 /* return group to which a task belongs */
371 static inline struct task_group *task_group(struct task_struct *p)
373 struct task_group *tg;
375 #ifdef CONFIG_USER_SCHED
377 tg = __task_cred(p)->user->tg;
379 #elif defined(CONFIG_CGROUP_SCHED)
380 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
381 struct task_group, css);
383 tg = &init_task_group;
388 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
389 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
393 p->se.parent = task_group(p)->se[cpu];
396 #ifdef CONFIG_RT_GROUP_SCHED
397 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
398 p->rt.parent = task_group(p)->rt_se[cpu];
405 static int root_task_group_empty(void)
411 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
412 static inline struct task_group *task_group(struct task_struct *p)
417 #endif /* CONFIG_GROUP_SCHED */
419 /* CFS-related fields in a runqueue */
421 struct load_weight load;
422 unsigned long nr_running;
427 struct rb_root tasks_timeline;
428 struct rb_node *rb_leftmost;
430 struct list_head tasks;
431 struct list_head *balance_iterator;
434 * 'curr' points to currently running entity on this cfs_rq.
435 * It is set to NULL otherwise (i.e when none are currently running).
437 struct sched_entity *curr, *next, *last;
439 unsigned int nr_spread_over;
441 #ifdef CONFIG_FAIR_GROUP_SCHED
442 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
445 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
446 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447 * (like users, containers etc.)
449 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450 * list is used during load balance.
452 struct list_head leaf_cfs_rq_list;
453 struct task_group *tg; /* group that "owns" this runqueue */
457 * the part of load.weight contributed by tasks
459 unsigned long task_weight;
462 * h_load = weight * f(tg)
464 * Where f(tg) is the recursive weight fraction assigned to
467 unsigned long h_load;
470 * this cpu's part of tg->shares
472 unsigned long shares;
475 * load.weight at the time we set shares
477 unsigned long rq_weight;
482 /* Real-Time classes' related field in a runqueue: */
484 struct rt_prio_array active;
485 unsigned long rt_nr_running;
486 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
488 int curr; /* highest queued rt task prio */
490 int next; /* next highest */
495 unsigned long rt_nr_migratory;
497 struct plist_head pushable_tasks;
502 /* Nests inside the rq lock: */
503 spinlock_t rt_runtime_lock;
505 #ifdef CONFIG_RT_GROUP_SCHED
506 unsigned long rt_nr_boosted;
509 struct list_head leaf_rt_rq_list;
510 struct task_group *tg;
511 struct sched_rt_entity *rt_se;
518 * We add the notion of a root-domain which will be used to define per-domain
519 * variables. Each exclusive cpuset essentially defines an island domain by
520 * fully partitioning the member cpus from any other cpuset. Whenever a new
521 * exclusive cpuset is created, we also create and attach a new root-domain
528 cpumask_var_t online;
531 * The "RT overload" flag: it gets set if a CPU has more than
532 * one runnable RT task.
534 cpumask_var_t rto_mask;
537 struct cpupri cpupri;
539 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
541 * Preferred wake up cpu nominated by sched_mc balance that will be
542 * used when most cpus are idle in the system indicating overall very
543 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
545 unsigned int sched_mc_preferred_wakeup_cpu;
550 * By default the system creates a single root-domain with all cpus as
551 * members (mimicking the global state we have today).
553 static struct root_domain def_root_domain;
558 * This is the main, per-CPU runqueue data structure.
560 * Locking rule: those places that want to lock multiple runqueues
561 * (such as the load balancing or the thread migration code), lock
562 * acquire operations must be ordered by ascending &runqueue.
569 * nr_running and cpu_load should be in the same cacheline because
570 * remote CPUs use both these fields when doing load calculation.
572 unsigned long nr_running;
573 #define CPU_LOAD_IDX_MAX 5
574 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
576 unsigned long last_tick_seen;
577 unsigned char in_nohz_recently;
579 /* capture load from *all* tasks on this cpu: */
580 struct load_weight load;
581 unsigned long nr_load_updates;
583 u64 nr_migrations_in;
588 #ifdef CONFIG_FAIR_GROUP_SCHED
589 /* list of leaf cfs_rq on this cpu: */
590 struct list_head leaf_cfs_rq_list;
592 #ifdef CONFIG_RT_GROUP_SCHED
593 struct list_head leaf_rt_rq_list;
597 * This is part of a global counter where only the total sum
598 * over all CPUs matters. A task can increase this counter on
599 * one CPU and if it got migrated afterwards it may decrease
600 * it on another CPU. Always updated under the runqueue lock:
602 unsigned long nr_uninterruptible;
604 struct task_struct *curr, *idle;
605 unsigned long next_balance;
606 struct mm_struct *prev_mm;
613 struct root_domain *rd;
614 struct sched_domain *sd;
616 unsigned char idle_at_tick;
617 /* For active balancing */
620 /* cpu of this runqueue: */
624 unsigned long avg_load_per_task;
626 struct task_struct *migration_thread;
627 struct list_head migration_queue;
630 /* calc_load related fields */
631 unsigned long calc_load_update;
632 long calc_load_active;
634 #ifdef CONFIG_SCHED_HRTICK
636 int hrtick_csd_pending;
637 struct call_single_data hrtick_csd;
639 struct hrtimer hrtick_timer;
642 #ifdef CONFIG_SCHEDSTATS
644 struct sched_info rq_sched_info;
645 unsigned long long rq_cpu_time;
646 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
648 /* sys_sched_yield() stats */
649 unsigned int yld_count;
651 /* schedule() stats */
652 unsigned int sched_switch;
653 unsigned int sched_count;
654 unsigned int sched_goidle;
656 /* try_to_wake_up() stats */
657 unsigned int ttwu_count;
658 unsigned int ttwu_local;
661 unsigned int bkl_count;
665 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
667 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
669 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
672 static inline int cpu_of(struct rq *rq)
682 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
683 * See detach_destroy_domains: synchronize_sched for details.
685 * The domain tree of any CPU may only be accessed from within
686 * preempt-disabled sections.
688 #define for_each_domain(cpu, __sd) \
689 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
691 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
692 #define this_rq() (&__get_cpu_var(runqueues))
693 #define task_rq(p) cpu_rq(task_cpu(p))
694 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
695 #define raw_rq() (&__raw_get_cpu_var(runqueues))
697 inline void update_rq_clock(struct rq *rq)
699 rq->clock = sched_clock_cpu(cpu_of(rq));
703 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
705 #ifdef CONFIG_SCHED_DEBUG
706 # define const_debug __read_mostly
708 # define const_debug static const
714 * Returns true if the current cpu runqueue is locked.
715 * This interface allows printk to be called with the runqueue lock
716 * held and know whether or not it is OK to wake up the klogd.
718 int runqueue_is_locked(void)
721 struct rq *rq = cpu_rq(cpu);
724 ret = spin_is_locked(&rq->lock);
730 * Debugging: various feature bits
733 #define SCHED_FEAT(name, enabled) \
734 __SCHED_FEAT_##name ,
737 #include "sched_features.h"
742 #define SCHED_FEAT(name, enabled) \
743 (1UL << __SCHED_FEAT_##name) * enabled |
745 const_debug unsigned int sysctl_sched_features =
746 #include "sched_features.h"
751 #ifdef CONFIG_SCHED_DEBUG
752 #define SCHED_FEAT(name, enabled) \
755 static __read_mostly char *sched_feat_names[] = {
756 #include "sched_features.h"
762 static int sched_feat_show(struct seq_file *m, void *v)
766 for (i = 0; sched_feat_names[i]; i++) {
767 if (!(sysctl_sched_features & (1UL << i)))
769 seq_printf(m, "%s ", sched_feat_names[i]);
777 sched_feat_write(struct file *filp, const char __user *ubuf,
778 size_t cnt, loff_t *ppos)
788 if (copy_from_user(&buf, ubuf, cnt))
793 if (strncmp(buf, "NO_", 3) == 0) {
798 for (i = 0; sched_feat_names[i]; i++) {
799 int len = strlen(sched_feat_names[i]);
801 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
803 sysctl_sched_features &= ~(1UL << i);
805 sysctl_sched_features |= (1UL << i);
810 if (!sched_feat_names[i])
818 static int sched_feat_open(struct inode *inode, struct file *filp)
820 return single_open(filp, sched_feat_show, NULL);
823 static struct file_operations sched_feat_fops = {
824 .open = sched_feat_open,
825 .write = sched_feat_write,
828 .release = single_release,
831 static __init int sched_init_debug(void)
833 debugfs_create_file("sched_features", 0644, NULL, NULL,
838 late_initcall(sched_init_debug);
842 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
845 * Number of tasks to iterate in a single balance run.
846 * Limited because this is done with IRQs disabled.
848 const_debug unsigned int sysctl_sched_nr_migrate = 32;
851 * ratelimit for updating the group shares.
854 unsigned int sysctl_sched_shares_ratelimit = 250000;
857 * Inject some fuzzyness into changing the per-cpu group shares
858 * this avoids remote rq-locks at the expense of fairness.
861 unsigned int sysctl_sched_shares_thresh = 4;
864 * period over which we measure -rt task cpu usage in us.
867 unsigned int sysctl_sched_rt_period = 1000000;
869 static __read_mostly int scheduler_running;
872 * part of the period that we allow rt tasks to run in us.
875 int sysctl_sched_rt_runtime = 950000;
877 static inline u64 global_rt_period(void)
879 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
882 static inline u64 global_rt_runtime(void)
884 if (sysctl_sched_rt_runtime < 0)
887 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
890 #ifndef prepare_arch_switch
891 # define prepare_arch_switch(next) do { } while (0)
893 #ifndef finish_arch_switch
894 # define finish_arch_switch(prev) do { } while (0)
897 static inline int task_current(struct rq *rq, struct task_struct *p)
899 return rq->curr == p;
902 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
903 static inline int task_running(struct rq *rq, struct task_struct *p)
905 return task_current(rq, p);
908 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
912 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
914 #ifdef CONFIG_DEBUG_SPINLOCK
915 /* this is a valid case when another task releases the spinlock */
916 rq->lock.owner = current;
919 * If we are tracking spinlock dependencies then we have to
920 * fix up the runqueue lock - which gets 'carried over' from
923 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
925 spin_unlock_irq(&rq->lock);
928 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
929 static inline int task_running(struct rq *rq, struct task_struct *p)
934 return task_current(rq, p);
938 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
942 * We can optimise this out completely for !SMP, because the
943 * SMP rebalancing from interrupt is the only thing that cares
948 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
949 spin_unlock_irq(&rq->lock);
951 spin_unlock(&rq->lock);
955 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
959 * After ->oncpu is cleared, the task can be moved to a different CPU.
960 * We must ensure this doesn't happen until the switch is completely
966 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
970 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
973 * __task_rq_lock - lock the runqueue a given task resides on.
974 * Must be called interrupts disabled.
976 static inline struct rq *__task_rq_lock(struct task_struct *p)
980 struct rq *rq = task_rq(p);
981 spin_lock(&rq->lock);
982 if (likely(rq == task_rq(p)))
984 spin_unlock(&rq->lock);
989 * task_rq_lock - lock the runqueue a given task resides on and disable
990 * interrupts. Note the ordering: we can safely lookup the task_rq without
991 * explicitly disabling preemption.
993 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
999 local_irq_save(*flags);
1001 spin_lock(&rq->lock);
1002 if (likely(rq == task_rq(p)))
1004 spin_unlock_irqrestore(&rq->lock, *flags);
1008 void task_rq_unlock_wait(struct task_struct *p)
1010 struct rq *rq = task_rq(p);
1012 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1013 spin_unlock_wait(&rq->lock);
1016 static void __task_rq_unlock(struct rq *rq)
1017 __releases(rq->lock)
1019 spin_unlock(&rq->lock);
1022 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1023 __releases(rq->lock)
1025 spin_unlock_irqrestore(&rq->lock, *flags);
1029 * this_rq_lock - lock this runqueue and disable interrupts.
1031 static struct rq *this_rq_lock(void)
1032 __acquires(rq->lock)
1036 local_irq_disable();
1038 spin_lock(&rq->lock);
1043 #ifdef CONFIG_SCHED_HRTICK
1045 * Use HR-timers to deliver accurate preemption points.
1047 * Its all a bit involved since we cannot program an hrt while holding the
1048 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1051 * When we get rescheduled we reprogram the hrtick_timer outside of the
1057 * - enabled by features
1058 * - hrtimer is actually high res
1060 static inline int hrtick_enabled(struct rq *rq)
1062 if (!sched_feat(HRTICK))
1064 if (!cpu_active(cpu_of(rq)))
1066 return hrtimer_is_hres_active(&rq->hrtick_timer);
1069 static void hrtick_clear(struct rq *rq)
1071 if (hrtimer_active(&rq->hrtick_timer))
1072 hrtimer_cancel(&rq->hrtick_timer);
1076 * High-resolution timer tick.
1077 * Runs from hardirq context with interrupts disabled.
1079 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1081 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1083 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1085 spin_lock(&rq->lock);
1086 update_rq_clock(rq);
1087 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1088 spin_unlock(&rq->lock);
1090 return HRTIMER_NORESTART;
1095 * called from hardirq (IPI) context
1097 static void __hrtick_start(void *arg)
1099 struct rq *rq = arg;
1101 spin_lock(&rq->lock);
1102 hrtimer_restart(&rq->hrtick_timer);
1103 rq->hrtick_csd_pending = 0;
1104 spin_unlock(&rq->lock);
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 struct hrtimer *timer = &rq->hrtick_timer;
1115 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1117 hrtimer_set_expires(timer, time);
1119 if (rq == this_rq()) {
1120 hrtimer_restart(timer);
1121 } else if (!rq->hrtick_csd_pending) {
1122 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1123 rq->hrtick_csd_pending = 1;
1128 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1130 int cpu = (int)(long)hcpu;
1133 case CPU_UP_CANCELED:
1134 case CPU_UP_CANCELED_FROZEN:
1135 case CPU_DOWN_PREPARE:
1136 case CPU_DOWN_PREPARE_FROZEN:
1138 case CPU_DEAD_FROZEN:
1139 hrtick_clear(cpu_rq(cpu));
1146 static __init void init_hrtick(void)
1148 hotcpu_notifier(hotplug_hrtick, 0);
1152 * Called to set the hrtick timer state.
1154 * called with rq->lock held and irqs disabled
1156 static void hrtick_start(struct rq *rq, u64 delay)
1158 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1159 HRTIMER_MODE_REL_PINNED, 0);
1162 static inline void init_hrtick(void)
1165 #endif /* CONFIG_SMP */
1167 static void init_rq_hrtick(struct rq *rq)
1170 rq->hrtick_csd_pending = 0;
1172 rq->hrtick_csd.flags = 0;
1173 rq->hrtick_csd.func = __hrtick_start;
1174 rq->hrtick_csd.info = rq;
1177 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1178 rq->hrtick_timer.function = hrtick;
1180 #else /* CONFIG_SCHED_HRTICK */
1181 static inline void hrtick_clear(struct rq *rq)
1185 static inline void init_rq_hrtick(struct rq *rq)
1189 static inline void init_hrtick(void)
1192 #endif /* CONFIG_SCHED_HRTICK */
1195 * resched_task - mark a task 'to be rescheduled now'.
1197 * On UP this means the setting of the need_resched flag, on SMP it
1198 * might also involve a cross-CPU call to trigger the scheduler on
1203 #ifndef tsk_is_polling
1204 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1207 static void resched_task(struct task_struct *p)
1211 assert_spin_locked(&task_rq(p)->lock);
1213 if (test_tsk_need_resched(p))
1216 set_tsk_need_resched(p);
1219 if (cpu == smp_processor_id())
1222 /* NEED_RESCHED must be visible before we test polling */
1224 if (!tsk_is_polling(p))
1225 smp_send_reschedule(cpu);
1228 static void resched_cpu(int cpu)
1230 struct rq *rq = cpu_rq(cpu);
1231 unsigned long flags;
1233 if (!spin_trylock_irqsave(&rq->lock, flags))
1235 resched_task(cpu_curr(cpu));
1236 spin_unlock_irqrestore(&rq->lock, flags);
1241 * When add_timer_on() enqueues a timer into the timer wheel of an
1242 * idle CPU then this timer might expire before the next timer event
1243 * which is scheduled to wake up that CPU. In case of a completely
1244 * idle system the next event might even be infinite time into the
1245 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1246 * leaves the inner idle loop so the newly added timer is taken into
1247 * account when the CPU goes back to idle and evaluates the timer
1248 * wheel for the next timer event.
1250 void wake_up_idle_cpu(int cpu)
1252 struct rq *rq = cpu_rq(cpu);
1254 if (cpu == smp_processor_id())
1258 * This is safe, as this function is called with the timer
1259 * wheel base lock of (cpu) held. When the CPU is on the way
1260 * to idle and has not yet set rq->curr to idle then it will
1261 * be serialized on the timer wheel base lock and take the new
1262 * timer into account automatically.
1264 if (rq->curr != rq->idle)
1268 * We can set TIF_RESCHED on the idle task of the other CPU
1269 * lockless. The worst case is that the other CPU runs the
1270 * idle task through an additional NOOP schedule()
1272 set_tsk_need_resched(rq->idle);
1274 /* NEED_RESCHED must be visible before we test polling */
1276 if (!tsk_is_polling(rq->idle))
1277 smp_send_reschedule(cpu);
1279 #endif /* CONFIG_NO_HZ */
1281 #else /* !CONFIG_SMP */
1282 static void resched_task(struct task_struct *p)
1284 assert_spin_locked(&task_rq(p)->lock);
1285 set_tsk_need_resched(p);
1287 #endif /* CONFIG_SMP */
1289 #if BITS_PER_LONG == 32
1290 # define WMULT_CONST (~0UL)
1292 # define WMULT_CONST (1UL << 32)
1295 #define WMULT_SHIFT 32
1298 * Shift right and round:
1300 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1303 * delta *= weight / lw
1305 static unsigned long
1306 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1307 struct load_weight *lw)
1311 if (!lw->inv_weight) {
1312 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1315 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1319 tmp = (u64)delta_exec * weight;
1321 * Check whether we'd overflow the 64-bit multiplication:
1323 if (unlikely(tmp > WMULT_CONST))
1324 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1327 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1329 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1332 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1338 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1345 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1346 * of tasks with abnormal "nice" values across CPUs the contribution that
1347 * each task makes to its run queue's load is weighted according to its
1348 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1349 * scaled version of the new time slice allocation that they receive on time
1353 #define WEIGHT_IDLEPRIO 3
1354 #define WMULT_IDLEPRIO 1431655765
1357 * Nice levels are multiplicative, with a gentle 10% change for every
1358 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1359 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1360 * that remained on nice 0.
1362 * The "10% effect" is relative and cumulative: from _any_ nice level,
1363 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1364 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1365 * If a task goes up by ~10% and another task goes down by ~10% then
1366 * the relative distance between them is ~25%.)
1368 static const int prio_to_weight[40] = {
1369 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1370 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1371 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1372 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1373 /* 0 */ 1024, 820, 655, 526, 423,
1374 /* 5 */ 335, 272, 215, 172, 137,
1375 /* 10 */ 110, 87, 70, 56, 45,
1376 /* 15 */ 36, 29, 23, 18, 15,
1380 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1382 * In cases where the weight does not change often, we can use the
1383 * precalculated inverse to speed up arithmetics by turning divisions
1384 * into multiplications:
1386 static const u32 prio_to_wmult[40] = {
1387 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1388 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1389 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1390 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1391 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1392 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1393 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1394 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1397 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1400 * runqueue iterator, to support SMP load-balancing between different
1401 * scheduling classes, without having to expose their internal data
1402 * structures to the load-balancing proper:
1404 struct rq_iterator {
1406 struct task_struct *(*start)(void *);
1407 struct task_struct *(*next)(void *);
1411 static unsigned long
1412 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1413 unsigned long max_load_move, struct sched_domain *sd,
1414 enum cpu_idle_type idle, int *all_pinned,
1415 int *this_best_prio, struct rq_iterator *iterator);
1418 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1419 struct sched_domain *sd, enum cpu_idle_type idle,
1420 struct rq_iterator *iterator);
1423 /* Time spent by the tasks of the cpu accounting group executing in ... */
1424 enum cpuacct_stat_index {
1425 CPUACCT_STAT_USER, /* ... user mode */
1426 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1428 CPUACCT_STAT_NSTATS,
1431 #ifdef CONFIG_CGROUP_CPUACCT
1432 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1433 static void cpuacct_update_stats(struct task_struct *tsk,
1434 enum cpuacct_stat_index idx, cputime_t val);
1436 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1437 static inline void cpuacct_update_stats(struct task_struct *tsk,
1438 enum cpuacct_stat_index idx, cputime_t val) {}
1441 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1443 update_load_add(&rq->load, load);
1446 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1448 update_load_sub(&rq->load, load);
1451 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1452 typedef int (*tg_visitor)(struct task_group *, void *);
1455 * Iterate the full tree, calling @down when first entering a node and @up when
1456 * leaving it for the final time.
1458 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1460 struct task_group *parent, *child;
1464 parent = &root_task_group;
1466 ret = (*down)(parent, data);
1469 list_for_each_entry_rcu(child, &parent->children, siblings) {
1476 ret = (*up)(parent, data);
1481 parent = parent->parent;
1490 static int tg_nop(struct task_group *tg, void *data)
1497 static unsigned long source_load(int cpu, int type);
1498 static unsigned long target_load(int cpu, int type);
1499 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1501 static unsigned long cpu_avg_load_per_task(int cpu)
1503 struct rq *rq = cpu_rq(cpu);
1504 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1507 rq->avg_load_per_task = rq->load.weight / nr_running;
1509 rq->avg_load_per_task = 0;
1511 return rq->avg_load_per_task;
1514 #ifdef CONFIG_FAIR_GROUP_SCHED
1516 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1519 * Calculate and set the cpu's group shares.
1522 update_group_shares_cpu(struct task_group *tg, int cpu,
1523 unsigned long sd_shares, unsigned long sd_rq_weight)
1525 unsigned long shares;
1526 unsigned long rq_weight;
1531 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1534 * \Sum shares * rq_weight
1535 * shares = -----------------------
1539 shares = (sd_shares * rq_weight) / sd_rq_weight;
1540 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1542 if (abs(shares - tg->se[cpu]->load.weight) >
1543 sysctl_sched_shares_thresh) {
1544 struct rq *rq = cpu_rq(cpu);
1545 unsigned long flags;
1547 spin_lock_irqsave(&rq->lock, flags);
1548 tg->cfs_rq[cpu]->shares = shares;
1550 __set_se_shares(tg->se[cpu], shares);
1551 spin_unlock_irqrestore(&rq->lock, flags);
1556 * Re-compute the task group their per cpu shares over the given domain.
1557 * This needs to be done in a bottom-up fashion because the rq weight of a
1558 * parent group depends on the shares of its child groups.
1560 static int tg_shares_up(struct task_group *tg, void *data)
1562 unsigned long weight, rq_weight = 0;
1563 unsigned long shares = 0;
1564 struct sched_domain *sd = data;
1567 for_each_cpu(i, sched_domain_span(sd)) {
1569 * If there are currently no tasks on the cpu pretend there
1570 * is one of average load so that when a new task gets to
1571 * run here it will not get delayed by group starvation.
1573 weight = tg->cfs_rq[i]->load.weight;
1575 weight = NICE_0_LOAD;
1577 tg->cfs_rq[i]->rq_weight = weight;
1578 rq_weight += weight;
1579 shares += tg->cfs_rq[i]->shares;
1582 if ((!shares && rq_weight) || shares > tg->shares)
1583 shares = tg->shares;
1585 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1586 shares = tg->shares;
1588 for_each_cpu(i, sched_domain_span(sd))
1589 update_group_shares_cpu(tg, i, shares, rq_weight);
1595 * Compute the cpu's hierarchical load factor for each task group.
1596 * This needs to be done in a top-down fashion because the load of a child
1597 * group is a fraction of its parents load.
1599 static int tg_load_down(struct task_group *tg, void *data)
1602 long cpu = (long)data;
1605 load = cpu_rq(cpu)->load.weight;
1607 load = tg->parent->cfs_rq[cpu]->h_load;
1608 load *= tg->cfs_rq[cpu]->shares;
1609 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1612 tg->cfs_rq[cpu]->h_load = load;
1617 static void update_shares(struct sched_domain *sd)
1619 u64 now = cpu_clock(raw_smp_processor_id());
1620 s64 elapsed = now - sd->last_update;
1622 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1623 sd->last_update = now;
1624 walk_tg_tree(tg_nop, tg_shares_up, sd);
1628 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1630 spin_unlock(&rq->lock);
1632 spin_lock(&rq->lock);
1635 static void update_h_load(long cpu)
1637 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1642 static inline void update_shares(struct sched_domain *sd)
1646 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1652 #ifdef CONFIG_PREEMPT
1655 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1656 * way at the expense of forcing extra atomic operations in all
1657 * invocations. This assures that the double_lock is acquired using the
1658 * same underlying policy as the spinlock_t on this architecture, which
1659 * reduces latency compared to the unfair variant below. However, it
1660 * also adds more overhead and therefore may reduce throughput.
1662 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1663 __releases(this_rq->lock)
1664 __acquires(busiest->lock)
1665 __acquires(this_rq->lock)
1667 spin_unlock(&this_rq->lock);
1668 double_rq_lock(this_rq, busiest);
1675 * Unfair double_lock_balance: Optimizes throughput at the expense of
1676 * latency by eliminating extra atomic operations when the locks are
1677 * already in proper order on entry. This favors lower cpu-ids and will
1678 * grant the double lock to lower cpus over higher ids under contention,
1679 * regardless of entry order into the function.
1681 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1682 __releases(this_rq->lock)
1683 __acquires(busiest->lock)
1684 __acquires(this_rq->lock)
1688 if (unlikely(!spin_trylock(&busiest->lock))) {
1689 if (busiest < this_rq) {
1690 spin_unlock(&this_rq->lock);
1691 spin_lock(&busiest->lock);
1692 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1695 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1700 #endif /* CONFIG_PREEMPT */
1703 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1705 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1707 if (unlikely(!irqs_disabled())) {
1708 /* printk() doesn't work good under rq->lock */
1709 spin_unlock(&this_rq->lock);
1713 return _double_lock_balance(this_rq, busiest);
1716 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1717 __releases(busiest->lock)
1719 spin_unlock(&busiest->lock);
1720 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1724 #ifdef CONFIG_FAIR_GROUP_SCHED
1725 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1728 cfs_rq->shares = shares;
1733 static void calc_load_account_active(struct rq *this_rq);
1735 #include "sched_stats.h"
1736 #include "sched_idletask.c"
1737 #include "sched_fair.c"
1738 #include "sched_rt.c"
1739 #ifdef CONFIG_SCHED_DEBUG
1740 # include "sched_debug.c"
1743 #define sched_class_highest (&rt_sched_class)
1744 #define for_each_class(class) \
1745 for (class = sched_class_highest; class; class = class->next)
1747 static void inc_nr_running(struct rq *rq)
1752 static void dec_nr_running(struct rq *rq)
1757 static void set_load_weight(struct task_struct *p)
1759 if (task_has_rt_policy(p)) {
1760 p->se.load.weight = prio_to_weight[0] * 2;
1761 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1766 * SCHED_IDLE tasks get minimal weight:
1768 if (p->policy == SCHED_IDLE) {
1769 p->se.load.weight = WEIGHT_IDLEPRIO;
1770 p->se.load.inv_weight = WMULT_IDLEPRIO;
1774 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1775 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1778 static void update_avg(u64 *avg, u64 sample)
1780 s64 diff = sample - *avg;
1784 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1787 p->se.start_runtime = p->se.sum_exec_runtime;
1789 sched_info_queued(p);
1790 p->sched_class->enqueue_task(rq, p, wakeup);
1794 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1797 if (p->se.last_wakeup) {
1798 update_avg(&p->se.avg_overlap,
1799 p->se.sum_exec_runtime - p->se.last_wakeup);
1800 p->se.last_wakeup = 0;
1802 update_avg(&p->se.avg_wakeup,
1803 sysctl_sched_wakeup_granularity);
1807 sched_info_dequeued(p);
1808 p->sched_class->dequeue_task(rq, p, sleep);
1813 * __normal_prio - return the priority that is based on the static prio
1815 static inline int __normal_prio(struct task_struct *p)
1817 return p->static_prio;
1821 * Calculate the expected normal priority: i.e. priority
1822 * without taking RT-inheritance into account. Might be
1823 * boosted by interactivity modifiers. Changes upon fork,
1824 * setprio syscalls, and whenever the interactivity
1825 * estimator recalculates.
1827 static inline int normal_prio(struct task_struct *p)
1831 if (task_has_rt_policy(p))
1832 prio = MAX_RT_PRIO-1 - p->rt_priority;
1834 prio = __normal_prio(p);
1839 * Calculate the current priority, i.e. the priority
1840 * taken into account by the scheduler. This value might
1841 * be boosted by RT tasks, or might be boosted by
1842 * interactivity modifiers. Will be RT if the task got
1843 * RT-boosted. If not then it returns p->normal_prio.
1845 static int effective_prio(struct task_struct *p)
1847 p->normal_prio = normal_prio(p);
1849 * If we are RT tasks or we were boosted to RT priority,
1850 * keep the priority unchanged. Otherwise, update priority
1851 * to the normal priority:
1853 if (!rt_prio(p->prio))
1854 return p->normal_prio;
1859 * activate_task - move a task to the runqueue.
1861 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1863 if (task_contributes_to_load(p))
1864 rq->nr_uninterruptible--;
1866 enqueue_task(rq, p, wakeup);
1871 * deactivate_task - remove a task from the runqueue.
1873 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1875 if (task_contributes_to_load(p))
1876 rq->nr_uninterruptible++;
1878 dequeue_task(rq, p, sleep);
1883 * task_curr - is this task currently executing on a CPU?
1884 * @p: the task in question.
1886 inline int task_curr(const struct task_struct *p)
1888 return cpu_curr(task_cpu(p)) == p;
1891 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1893 set_task_rq(p, cpu);
1896 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1897 * successfuly executed on another CPU. We must ensure that updates of
1898 * per-task data have been completed by this moment.
1901 task_thread_info(p)->cpu = cpu;
1905 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1906 const struct sched_class *prev_class,
1907 int oldprio, int running)
1909 if (prev_class != p->sched_class) {
1910 if (prev_class->switched_from)
1911 prev_class->switched_from(rq, p, running);
1912 p->sched_class->switched_to(rq, p, running);
1914 p->sched_class->prio_changed(rq, p, oldprio, running);
1919 /* Used instead of source_load when we know the type == 0 */
1920 static unsigned long weighted_cpuload(const int cpu)
1922 return cpu_rq(cpu)->load.weight;
1926 * Is this task likely cache-hot:
1929 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1934 * Buddy candidates are cache hot:
1936 if (sched_feat(CACHE_HOT_BUDDY) &&
1937 (&p->se == cfs_rq_of(&p->se)->next ||
1938 &p->se == cfs_rq_of(&p->se)->last))
1941 if (p->sched_class != &fair_sched_class)
1944 if (sysctl_sched_migration_cost == -1)
1946 if (sysctl_sched_migration_cost == 0)
1949 delta = now - p->se.exec_start;
1951 return delta < (s64)sysctl_sched_migration_cost;
1955 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1957 int old_cpu = task_cpu(p);
1958 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1959 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1960 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1963 clock_offset = old_rq->clock - new_rq->clock;
1965 trace_sched_migrate_task(p, new_cpu);
1967 #ifdef CONFIG_SCHEDSTATS
1968 if (p->se.wait_start)
1969 p->se.wait_start -= clock_offset;
1970 if (p->se.sleep_start)
1971 p->se.sleep_start -= clock_offset;
1972 if (p->se.block_start)
1973 p->se.block_start -= clock_offset;
1975 if (old_cpu != new_cpu) {
1976 p->se.nr_migrations++;
1977 new_rq->nr_migrations_in++;
1978 #ifdef CONFIG_SCHEDSTATS
1979 if (task_hot(p, old_rq->clock, NULL))
1980 schedstat_inc(p, se.nr_forced2_migrations);
1982 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
1985 p->se.vruntime -= old_cfsrq->min_vruntime -
1986 new_cfsrq->min_vruntime;
1988 __set_task_cpu(p, new_cpu);
1991 struct migration_req {
1992 struct list_head list;
1994 struct task_struct *task;
1997 struct completion done;
2001 * The task's runqueue lock must be held.
2002 * Returns true if you have to wait for migration thread.
2005 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2007 struct rq *rq = task_rq(p);
2010 * If the task is not on a runqueue (and not running), then
2011 * it is sufficient to simply update the task's cpu field.
2013 if (!p->se.on_rq && !task_running(rq, p)) {
2014 set_task_cpu(p, dest_cpu);
2018 init_completion(&req->done);
2020 req->dest_cpu = dest_cpu;
2021 list_add(&req->list, &rq->migration_queue);
2027 * wait_task_context_switch - wait for a thread to complete at least one
2030 * @p must not be current.
2032 void wait_task_context_switch(struct task_struct *p)
2034 unsigned long nvcsw, nivcsw, flags;
2042 * The runqueue is assigned before the actual context
2043 * switch. We need to take the runqueue lock.
2045 * We could check initially without the lock but it is
2046 * very likely that we need to take the lock in every
2049 rq = task_rq_lock(p, &flags);
2050 running = task_running(rq, p);
2051 task_rq_unlock(rq, &flags);
2053 if (likely(!running))
2056 * The switch count is incremented before the actual
2057 * context switch. We thus wait for two switches to be
2058 * sure at least one completed.
2060 if ((p->nvcsw - nvcsw) > 1)
2062 if ((p->nivcsw - nivcsw) > 1)
2070 * wait_task_inactive - wait for a thread to unschedule.
2072 * If @match_state is nonzero, it's the @p->state value just checked and
2073 * not expected to change. If it changes, i.e. @p might have woken up,
2074 * then return zero. When we succeed in waiting for @p to be off its CPU,
2075 * we return a positive number (its total switch count). If a second call
2076 * a short while later returns the same number, the caller can be sure that
2077 * @p has remained unscheduled the whole time.
2079 * The caller must ensure that the task *will* unschedule sometime soon,
2080 * else this function might spin for a *long* time. This function can't
2081 * be called with interrupts off, or it may introduce deadlock with
2082 * smp_call_function() if an IPI is sent by the same process we are
2083 * waiting to become inactive.
2085 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2087 unsigned long flags;
2094 * We do the initial early heuristics without holding
2095 * any task-queue locks at all. We'll only try to get
2096 * the runqueue lock when things look like they will
2102 * If the task is actively running on another CPU
2103 * still, just relax and busy-wait without holding
2106 * NOTE! Since we don't hold any locks, it's not
2107 * even sure that "rq" stays as the right runqueue!
2108 * But we don't care, since "task_running()" will
2109 * return false if the runqueue has changed and p
2110 * is actually now running somewhere else!
2112 while (task_running(rq, p)) {
2113 if (match_state && unlikely(p->state != match_state))
2119 * Ok, time to look more closely! We need the rq
2120 * lock now, to be *sure*. If we're wrong, we'll
2121 * just go back and repeat.
2123 rq = task_rq_lock(p, &flags);
2124 trace_sched_wait_task(rq, p);
2125 running = task_running(rq, p);
2126 on_rq = p->se.on_rq;
2128 if (!match_state || p->state == match_state)
2129 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2130 task_rq_unlock(rq, &flags);
2133 * If it changed from the expected state, bail out now.
2135 if (unlikely(!ncsw))
2139 * Was it really running after all now that we
2140 * checked with the proper locks actually held?
2142 * Oops. Go back and try again..
2144 if (unlikely(running)) {
2150 * It's not enough that it's not actively running,
2151 * it must be off the runqueue _entirely_, and not
2154 * So if it was still runnable (but just not actively
2155 * running right now), it's preempted, and we should
2156 * yield - it could be a while.
2158 if (unlikely(on_rq)) {
2159 schedule_timeout_uninterruptible(1);
2164 * Ahh, all good. It wasn't running, and it wasn't
2165 * runnable, which means that it will never become
2166 * running in the future either. We're all done!
2175 * kick_process - kick a running thread to enter/exit the kernel
2176 * @p: the to-be-kicked thread
2178 * Cause a process which is running on another CPU to enter
2179 * kernel-mode, without any delay. (to get signals handled.)
2181 * NOTE: this function doesnt have to take the runqueue lock,
2182 * because all it wants to ensure is that the remote task enters
2183 * the kernel. If the IPI races and the task has been migrated
2184 * to another CPU then no harm is done and the purpose has been
2187 void kick_process(struct task_struct *p)
2193 if ((cpu != smp_processor_id()) && task_curr(p))
2194 smp_send_reschedule(cpu);
2197 EXPORT_SYMBOL_GPL(kick_process);
2200 * Return a low guess at the load of a migration-source cpu weighted
2201 * according to the scheduling class and "nice" value.
2203 * We want to under-estimate the load of migration sources, to
2204 * balance conservatively.
2206 static unsigned long source_load(int cpu, int type)
2208 struct rq *rq = cpu_rq(cpu);
2209 unsigned long total = weighted_cpuload(cpu);
2211 if (type == 0 || !sched_feat(LB_BIAS))
2214 return min(rq->cpu_load[type-1], total);
2218 * Return a high guess at the load of a migration-target cpu weighted
2219 * according to the scheduling class and "nice" value.
2221 static unsigned long target_load(int cpu, int type)
2223 struct rq *rq = cpu_rq(cpu);
2224 unsigned long total = weighted_cpuload(cpu);
2226 if (type == 0 || !sched_feat(LB_BIAS))
2229 return max(rq->cpu_load[type-1], total);
2233 * find_idlest_group finds and returns the least busy CPU group within the
2236 static struct sched_group *
2237 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2239 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2240 unsigned long min_load = ULONG_MAX, this_load = 0;
2241 int load_idx = sd->forkexec_idx;
2242 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2245 unsigned long load, avg_load;
2249 /* Skip over this group if it has no CPUs allowed */
2250 if (!cpumask_intersects(sched_group_cpus(group),
2254 local_group = cpumask_test_cpu(this_cpu,
2255 sched_group_cpus(group));
2257 /* Tally up the load of all CPUs in the group */
2260 for_each_cpu(i, sched_group_cpus(group)) {
2261 /* Bias balancing toward cpus of our domain */
2263 load = source_load(i, load_idx);
2265 load = target_load(i, load_idx);
2270 /* Adjust by relative CPU power of the group */
2271 avg_load = sg_div_cpu_power(group,
2272 avg_load * SCHED_LOAD_SCALE);
2275 this_load = avg_load;
2277 } else if (avg_load < min_load) {
2278 min_load = avg_load;
2281 } while (group = group->next, group != sd->groups);
2283 if (!idlest || 100*this_load < imbalance*min_load)
2289 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2292 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2294 unsigned long load, min_load = ULONG_MAX;
2298 /* Traverse only the allowed CPUs */
2299 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2300 load = weighted_cpuload(i);
2302 if (load < min_load || (load == min_load && i == this_cpu)) {
2312 * sched_balance_self: balance the current task (running on cpu) in domains
2313 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2316 * Balance, ie. select the least loaded group.
2318 * Returns the target CPU number, or the same CPU if no balancing is needed.
2320 * preempt must be disabled.
2322 static int sched_balance_self(int cpu, int flag)
2324 struct task_struct *t = current;
2325 struct sched_domain *tmp, *sd = NULL;
2327 for_each_domain(cpu, tmp) {
2329 * If power savings logic is enabled for a domain, stop there.
2331 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2333 if (tmp->flags & flag)
2341 struct sched_group *group;
2342 int new_cpu, weight;
2344 if (!(sd->flags & flag)) {
2349 group = find_idlest_group(sd, t, cpu);
2355 new_cpu = find_idlest_cpu(group, t, cpu);
2356 if (new_cpu == -1 || new_cpu == cpu) {
2357 /* Now try balancing at a lower domain level of cpu */
2362 /* Now try balancing at a lower domain level of new_cpu */
2364 weight = cpumask_weight(sched_domain_span(sd));
2366 for_each_domain(cpu, tmp) {
2367 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2369 if (tmp->flags & flag)
2372 /* while loop will break here if sd == NULL */
2378 #endif /* CONFIG_SMP */
2381 * task_oncpu_function_call - call a function on the cpu on which a task runs
2382 * @p: the task to evaluate
2383 * @func: the function to be called
2384 * @info: the function call argument
2386 * Calls the function @func when the task is currently running. This might
2387 * be on the current CPU, which just calls the function directly
2389 void task_oncpu_function_call(struct task_struct *p,
2390 void (*func) (void *info), void *info)
2397 smp_call_function_single(cpu, func, info, 1);
2402 * try_to_wake_up - wake up a thread
2403 * @p: the to-be-woken-up thread
2404 * @state: the mask of task states that can be woken
2405 * @sync: do a synchronous wakeup?
2407 * Put it on the run-queue if it's not already there. The "current"
2408 * thread is always on the run-queue (except when the actual
2409 * re-schedule is in progress), and as such you're allowed to do
2410 * the simpler "current->state = TASK_RUNNING" to mark yourself
2411 * runnable without the overhead of this.
2413 * returns failure only if the task is already active.
2415 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2417 int cpu, orig_cpu, this_cpu, success = 0;
2418 unsigned long flags;
2422 if (!sched_feat(SYNC_WAKEUPS))
2426 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2427 struct sched_domain *sd;
2429 this_cpu = raw_smp_processor_id();
2432 for_each_domain(this_cpu, sd) {
2433 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2442 rq = task_rq_lock(p, &flags);
2443 update_rq_clock(rq);
2444 old_state = p->state;
2445 if (!(old_state & state))
2453 this_cpu = smp_processor_id();
2456 if (unlikely(task_running(rq, p)))
2459 cpu = p->sched_class->select_task_rq(p, sync);
2460 if (cpu != orig_cpu) {
2461 set_task_cpu(p, cpu);
2462 task_rq_unlock(rq, &flags);
2463 /* might preempt at this point */
2464 rq = task_rq_lock(p, &flags);
2465 old_state = p->state;
2466 if (!(old_state & state))
2471 this_cpu = smp_processor_id();
2475 #ifdef CONFIG_SCHEDSTATS
2476 schedstat_inc(rq, ttwu_count);
2477 if (cpu == this_cpu)
2478 schedstat_inc(rq, ttwu_local);
2480 struct sched_domain *sd;
2481 for_each_domain(this_cpu, sd) {
2482 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2483 schedstat_inc(sd, ttwu_wake_remote);
2488 #endif /* CONFIG_SCHEDSTATS */
2491 #endif /* CONFIG_SMP */
2492 schedstat_inc(p, se.nr_wakeups);
2494 schedstat_inc(p, se.nr_wakeups_sync);
2495 if (orig_cpu != cpu)
2496 schedstat_inc(p, se.nr_wakeups_migrate);
2497 if (cpu == this_cpu)
2498 schedstat_inc(p, se.nr_wakeups_local);
2500 schedstat_inc(p, se.nr_wakeups_remote);
2501 activate_task(rq, p, 1);
2505 * Only attribute actual wakeups done by this task.
2507 if (!in_interrupt()) {
2508 struct sched_entity *se = ¤t->se;
2509 u64 sample = se->sum_exec_runtime;
2511 if (se->last_wakeup)
2512 sample -= se->last_wakeup;
2514 sample -= se->start_runtime;
2515 update_avg(&se->avg_wakeup, sample);
2517 se->last_wakeup = se->sum_exec_runtime;
2521 trace_sched_wakeup(rq, p, success);
2522 check_preempt_curr(rq, p, sync);
2524 p->state = TASK_RUNNING;
2526 if (p->sched_class->task_wake_up)
2527 p->sched_class->task_wake_up(rq, p);
2530 task_rq_unlock(rq, &flags);
2536 * wake_up_process - Wake up a specific process
2537 * @p: The process to be woken up.
2539 * Attempt to wake up the nominated process and move it to the set of runnable
2540 * processes. Returns 1 if the process was woken up, 0 if it was already
2543 * It may be assumed that this function implies a write memory barrier before
2544 * changing the task state if and only if any tasks are woken up.
2546 int wake_up_process(struct task_struct *p)
2548 return try_to_wake_up(p, TASK_ALL, 0);
2550 EXPORT_SYMBOL(wake_up_process);
2552 int wake_up_state(struct task_struct *p, unsigned int state)
2554 return try_to_wake_up(p, state, 0);
2558 * Perform scheduler related setup for a newly forked process p.
2559 * p is forked by current.
2561 * __sched_fork() is basic setup used by init_idle() too:
2563 static void __sched_fork(struct task_struct *p)
2565 p->se.exec_start = 0;
2566 p->se.sum_exec_runtime = 0;
2567 p->se.prev_sum_exec_runtime = 0;
2568 p->se.nr_migrations = 0;
2569 p->se.last_wakeup = 0;
2570 p->se.avg_overlap = 0;
2571 p->se.start_runtime = 0;
2572 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2574 #ifdef CONFIG_SCHEDSTATS
2575 p->se.wait_start = 0;
2576 p->se.sum_sleep_runtime = 0;
2577 p->se.sleep_start = 0;
2578 p->se.block_start = 0;
2579 p->se.sleep_max = 0;
2580 p->se.block_max = 0;
2582 p->se.slice_max = 0;
2586 INIT_LIST_HEAD(&p->rt.run_list);
2588 INIT_LIST_HEAD(&p->se.group_node);
2590 #ifdef CONFIG_PREEMPT_NOTIFIERS
2591 INIT_HLIST_HEAD(&p->preempt_notifiers);
2595 * We mark the process as running here, but have not actually
2596 * inserted it onto the runqueue yet. This guarantees that
2597 * nobody will actually run it, and a signal or other external
2598 * event cannot wake it up and insert it on the runqueue either.
2600 p->state = TASK_RUNNING;
2604 * fork()/clone()-time setup:
2606 void sched_fork(struct task_struct *p, int clone_flags)
2608 int cpu = get_cpu();
2613 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2615 set_task_cpu(p, cpu);
2618 * Make sure we do not leak PI boosting priority to the child.
2620 p->prio = current->normal_prio;
2623 * Revert to default priority/policy on fork if requested.
2625 if (unlikely(p->sched_reset_on_fork)) {
2626 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2627 p->policy = SCHED_NORMAL;
2629 if (p->normal_prio < DEFAULT_PRIO)
2630 p->prio = DEFAULT_PRIO;
2632 if (PRIO_TO_NICE(p->static_prio) < 0) {
2633 p->static_prio = NICE_TO_PRIO(0);
2638 * We don't need the reset flag anymore after the fork. It has
2639 * fulfilled its duty:
2641 p->sched_reset_on_fork = 0;
2644 if (!rt_prio(p->prio))
2645 p->sched_class = &fair_sched_class;
2647 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2648 if (likely(sched_info_on()))
2649 memset(&p->sched_info, 0, sizeof(p->sched_info));
2651 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2654 #ifdef CONFIG_PREEMPT
2655 /* Want to start with kernel preemption disabled. */
2656 task_thread_info(p)->preempt_count = 1;
2658 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2664 * wake_up_new_task - wake up a newly created task for the first time.
2666 * This function will do some initial scheduler statistics housekeeping
2667 * that must be done for every newly created context, then puts the task
2668 * on the runqueue and wakes it.
2670 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2672 unsigned long flags;
2675 rq = task_rq_lock(p, &flags);
2676 BUG_ON(p->state != TASK_RUNNING);
2677 update_rq_clock(rq);
2679 p->prio = effective_prio(p);
2681 if (!p->sched_class->task_new || !current->se.on_rq) {
2682 activate_task(rq, p, 0);
2685 * Let the scheduling class do new task startup
2686 * management (if any):
2688 p->sched_class->task_new(rq, p);
2691 trace_sched_wakeup_new(rq, p, 1);
2692 check_preempt_curr(rq, p, 0);
2694 if (p->sched_class->task_wake_up)
2695 p->sched_class->task_wake_up(rq, p);
2697 task_rq_unlock(rq, &flags);
2700 #ifdef CONFIG_PREEMPT_NOTIFIERS
2703 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2704 * @notifier: notifier struct to register
2706 void preempt_notifier_register(struct preempt_notifier *notifier)
2708 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2710 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2713 * preempt_notifier_unregister - no longer interested in preemption notifications
2714 * @notifier: notifier struct to unregister
2716 * This is safe to call from within a preemption notifier.
2718 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2720 hlist_del(¬ifier->link);
2722 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2724 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2726 struct preempt_notifier *notifier;
2727 struct hlist_node *node;
2729 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2730 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2734 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2735 struct task_struct *next)
2737 struct preempt_notifier *notifier;
2738 struct hlist_node *node;
2740 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2741 notifier->ops->sched_out(notifier, next);
2744 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2746 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2751 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2752 struct task_struct *next)
2756 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2759 * prepare_task_switch - prepare to switch tasks
2760 * @rq: the runqueue preparing to switch
2761 * @prev: the current task that is being switched out
2762 * @next: the task we are going to switch to.
2764 * This is called with the rq lock held and interrupts off. It must
2765 * be paired with a subsequent finish_task_switch after the context
2768 * prepare_task_switch sets up locking and calls architecture specific
2772 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2773 struct task_struct *next)
2775 fire_sched_out_preempt_notifiers(prev, next);
2776 prepare_lock_switch(rq, next);
2777 prepare_arch_switch(next);
2781 * finish_task_switch - clean up after a task-switch
2782 * @rq: runqueue associated with task-switch
2783 * @prev: the thread we just switched away from.
2785 * finish_task_switch must be called after the context switch, paired
2786 * with a prepare_task_switch call before the context switch.
2787 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2788 * and do any other architecture-specific cleanup actions.
2790 * Note that we may have delayed dropping an mm in context_switch(). If
2791 * so, we finish that here outside of the runqueue lock. (Doing it
2792 * with the lock held can cause deadlocks; see schedule() for
2795 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2796 __releases(rq->lock)
2798 struct mm_struct *mm = rq->prev_mm;
2801 int post_schedule = 0;
2803 if (current->sched_class->needs_post_schedule)
2804 post_schedule = current->sched_class->needs_post_schedule(rq);
2810 * A task struct has one reference for the use as "current".
2811 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2812 * schedule one last time. The schedule call will never return, and
2813 * the scheduled task must drop that reference.
2814 * The test for TASK_DEAD must occur while the runqueue locks are
2815 * still held, otherwise prev could be scheduled on another cpu, die
2816 * there before we look at prev->state, and then the reference would
2818 * Manfred Spraul <manfred@colorfullife.com>
2820 prev_state = prev->state;
2821 finish_arch_switch(prev);
2822 perf_counter_task_sched_in(current, cpu_of(rq));
2823 finish_lock_switch(rq, prev);
2826 current->sched_class->post_schedule(rq);
2829 fire_sched_in_preempt_notifiers(current);
2832 if (unlikely(prev_state == TASK_DEAD)) {
2834 * Remove function-return probe instances associated with this
2835 * task and put them back on the free list.
2837 kprobe_flush_task(prev);
2838 put_task_struct(prev);
2843 * schedule_tail - first thing a freshly forked thread must call.
2844 * @prev: the thread we just switched away from.
2846 asmlinkage void schedule_tail(struct task_struct *prev)
2847 __releases(rq->lock)
2849 struct rq *rq = this_rq();
2851 finish_task_switch(rq, prev);
2852 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2853 /* In this case, finish_task_switch does not reenable preemption */
2856 if (current->set_child_tid)
2857 put_user(task_pid_vnr(current), current->set_child_tid);
2861 * context_switch - switch to the new MM and the new
2862 * thread's register state.
2865 context_switch(struct rq *rq, struct task_struct *prev,
2866 struct task_struct *next)
2868 struct mm_struct *mm, *oldmm;
2870 prepare_task_switch(rq, prev, next);
2871 trace_sched_switch(rq, prev, next);
2873 oldmm = prev->active_mm;
2875 * For paravirt, this is coupled with an exit in switch_to to
2876 * combine the page table reload and the switch backend into
2879 arch_start_context_switch(prev);
2881 if (unlikely(!mm)) {
2882 next->active_mm = oldmm;
2883 atomic_inc(&oldmm->mm_count);
2884 enter_lazy_tlb(oldmm, next);
2886 switch_mm(oldmm, mm, next);
2888 if (unlikely(!prev->mm)) {
2889 prev->active_mm = NULL;
2890 rq->prev_mm = oldmm;
2893 * Since the runqueue lock will be released by the next
2894 * task (which is an invalid locking op but in the case
2895 * of the scheduler it's an obvious special-case), so we
2896 * do an early lockdep release here:
2898 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2899 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2902 /* Here we just switch the register state and the stack. */
2903 switch_to(prev, next, prev);
2907 * this_rq must be evaluated again because prev may have moved
2908 * CPUs since it called schedule(), thus the 'rq' on its stack
2909 * frame will be invalid.
2911 finish_task_switch(this_rq(), prev);
2915 * nr_running, nr_uninterruptible and nr_context_switches:
2917 * externally visible scheduler statistics: current number of runnable
2918 * threads, current number of uninterruptible-sleeping threads, total
2919 * number of context switches performed since bootup.
2921 unsigned long nr_running(void)
2923 unsigned long i, sum = 0;
2925 for_each_online_cpu(i)
2926 sum += cpu_rq(i)->nr_running;
2931 unsigned long nr_uninterruptible(void)
2933 unsigned long i, sum = 0;
2935 for_each_possible_cpu(i)
2936 sum += cpu_rq(i)->nr_uninterruptible;
2939 * Since we read the counters lockless, it might be slightly
2940 * inaccurate. Do not allow it to go below zero though:
2942 if (unlikely((long)sum < 0))
2948 unsigned long long nr_context_switches(void)
2951 unsigned long long sum = 0;
2953 for_each_possible_cpu(i)
2954 sum += cpu_rq(i)->nr_switches;
2959 unsigned long nr_iowait(void)
2961 unsigned long i, sum = 0;
2963 for_each_possible_cpu(i)
2964 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2969 /* Variables and functions for calc_load */
2970 static atomic_long_t calc_load_tasks;
2971 static unsigned long calc_load_update;
2972 unsigned long avenrun[3];
2973 EXPORT_SYMBOL(avenrun);
2976 * get_avenrun - get the load average array
2977 * @loads: pointer to dest load array
2978 * @offset: offset to add
2979 * @shift: shift count to shift the result left
2981 * These values are estimates at best, so no need for locking.
2983 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2985 loads[0] = (avenrun[0] + offset) << shift;
2986 loads[1] = (avenrun[1] + offset) << shift;
2987 loads[2] = (avenrun[2] + offset) << shift;
2990 static unsigned long
2991 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2994 load += active * (FIXED_1 - exp);
2995 return load >> FSHIFT;
2999 * calc_load - update the avenrun load estimates 10 ticks after the
3000 * CPUs have updated calc_load_tasks.
3002 void calc_global_load(void)
3004 unsigned long upd = calc_load_update + 10;
3007 if (time_before(jiffies, upd))
3010 active = atomic_long_read(&calc_load_tasks);
3011 active = active > 0 ? active * FIXED_1 : 0;
3013 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3014 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3015 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3017 calc_load_update += LOAD_FREQ;
3021 * Either called from update_cpu_load() or from a cpu going idle
3023 static void calc_load_account_active(struct rq *this_rq)
3025 long nr_active, delta;
3027 nr_active = this_rq->nr_running;
3028 nr_active += (long) this_rq->nr_uninterruptible;
3030 if (nr_active != this_rq->calc_load_active) {
3031 delta = nr_active - this_rq->calc_load_active;
3032 this_rq->calc_load_active = nr_active;
3033 atomic_long_add(delta, &calc_load_tasks);
3038 * Externally visible per-cpu scheduler statistics:
3039 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3041 u64 cpu_nr_migrations(int cpu)
3043 return cpu_rq(cpu)->nr_migrations_in;
3047 * Update rq->cpu_load[] statistics. This function is usually called every
3048 * scheduler tick (TICK_NSEC).
3050 static void update_cpu_load(struct rq *this_rq)
3052 unsigned long this_load = this_rq->load.weight;
3055 this_rq->nr_load_updates++;
3057 /* Update our load: */
3058 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3059 unsigned long old_load, new_load;
3061 /* scale is effectively 1 << i now, and >> i divides by scale */
3063 old_load = this_rq->cpu_load[i];
3064 new_load = this_load;
3066 * Round up the averaging division if load is increasing. This
3067 * prevents us from getting stuck on 9 if the load is 10, for
3070 if (new_load > old_load)
3071 new_load += scale-1;
3072 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3075 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3076 this_rq->calc_load_update += LOAD_FREQ;
3077 calc_load_account_active(this_rq);
3084 * double_rq_lock - safely lock two runqueues
3086 * Note this does not disable interrupts like task_rq_lock,
3087 * you need to do so manually before calling.
3089 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3090 __acquires(rq1->lock)
3091 __acquires(rq2->lock)
3093 BUG_ON(!irqs_disabled());
3095 spin_lock(&rq1->lock);
3096 __acquire(rq2->lock); /* Fake it out ;) */
3099 spin_lock(&rq1->lock);
3100 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3102 spin_lock(&rq2->lock);
3103 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3106 update_rq_clock(rq1);
3107 update_rq_clock(rq2);
3111 * double_rq_unlock - safely unlock two runqueues
3113 * Note this does not restore interrupts like task_rq_unlock,
3114 * you need to do so manually after calling.
3116 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3117 __releases(rq1->lock)
3118 __releases(rq2->lock)
3120 spin_unlock(&rq1->lock);
3122 spin_unlock(&rq2->lock);
3124 __release(rq2->lock);
3128 * If dest_cpu is allowed for this process, migrate the task to it.
3129 * This is accomplished by forcing the cpu_allowed mask to only
3130 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3131 * the cpu_allowed mask is restored.
3133 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3135 struct migration_req req;
3136 unsigned long flags;
3139 rq = task_rq_lock(p, &flags);
3140 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3141 || unlikely(!cpu_active(dest_cpu)))
3144 /* force the process onto the specified CPU */
3145 if (migrate_task(p, dest_cpu, &req)) {
3146 /* Need to wait for migration thread (might exit: take ref). */
3147 struct task_struct *mt = rq->migration_thread;
3149 get_task_struct(mt);
3150 task_rq_unlock(rq, &flags);
3151 wake_up_process(mt);
3152 put_task_struct(mt);
3153 wait_for_completion(&req.done);
3158 task_rq_unlock(rq, &flags);
3162 * sched_exec - execve() is a valuable balancing opportunity, because at
3163 * this point the task has the smallest effective memory and cache footprint.
3165 void sched_exec(void)
3167 int new_cpu, this_cpu = get_cpu();
3168 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3170 if (new_cpu != this_cpu)
3171 sched_migrate_task(current, new_cpu);
3175 * pull_task - move a task from a remote runqueue to the local runqueue.
3176 * Both runqueues must be locked.
3178 static void pull_task(struct rq *src_rq, struct task_struct *p,
3179 struct rq *this_rq, int this_cpu)
3181 deactivate_task(src_rq, p, 0);
3182 set_task_cpu(p, this_cpu);
3183 activate_task(this_rq, p, 0);
3185 * Note that idle threads have a prio of MAX_PRIO, for this test
3186 * to be always true for them.
3188 check_preempt_curr(this_rq, p, 0);
3192 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3195 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3196 struct sched_domain *sd, enum cpu_idle_type idle,
3199 int tsk_cache_hot = 0;
3201 * We do not migrate tasks that are:
3202 * 1) running (obviously), or
3203 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3204 * 3) are cache-hot on their current CPU.
3206 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3207 schedstat_inc(p, se.nr_failed_migrations_affine);
3212 if (task_running(rq, p)) {
3213 schedstat_inc(p, se.nr_failed_migrations_running);
3218 * Aggressive migration if:
3219 * 1) task is cache cold, or
3220 * 2) too many balance attempts have failed.
3223 tsk_cache_hot = task_hot(p, rq->clock, sd);
3224 if (!tsk_cache_hot ||
3225 sd->nr_balance_failed > sd->cache_nice_tries) {
3226 #ifdef CONFIG_SCHEDSTATS
3227 if (tsk_cache_hot) {
3228 schedstat_inc(sd, lb_hot_gained[idle]);
3229 schedstat_inc(p, se.nr_forced_migrations);
3235 if (tsk_cache_hot) {
3236 schedstat_inc(p, se.nr_failed_migrations_hot);
3242 static unsigned long
3243 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3244 unsigned long max_load_move, struct sched_domain *sd,
3245 enum cpu_idle_type idle, int *all_pinned,
3246 int *this_best_prio, struct rq_iterator *iterator)
3248 int loops = 0, pulled = 0, pinned = 0;
3249 struct task_struct *p;
3250 long rem_load_move = max_load_move;
3252 if (max_load_move == 0)
3258 * Start the load-balancing iterator:
3260 p = iterator->start(iterator->arg);
3262 if (!p || loops++ > sysctl_sched_nr_migrate)
3265 if ((p->se.load.weight >> 1) > rem_load_move ||
3266 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3267 p = iterator->next(iterator->arg);
3271 pull_task(busiest, p, this_rq, this_cpu);
3273 rem_load_move -= p->se.load.weight;
3275 #ifdef CONFIG_PREEMPT
3277 * NEWIDLE balancing is a source of latency, so preemptible kernels
3278 * will stop after the first task is pulled to minimize the critical
3281 if (idle == CPU_NEWLY_IDLE)
3286 * We only want to steal up to the prescribed amount of weighted load.
3288 if (rem_load_move > 0) {
3289 if (p->prio < *this_best_prio)
3290 *this_best_prio = p->prio;
3291 p = iterator->next(iterator->arg);
3296 * Right now, this is one of only two places pull_task() is called,
3297 * so we can safely collect pull_task() stats here rather than
3298 * inside pull_task().
3300 schedstat_add(sd, lb_gained[idle], pulled);
3303 *all_pinned = pinned;
3305 return max_load_move - rem_load_move;
3309 * move_tasks tries to move up to max_load_move weighted load from busiest to
3310 * this_rq, as part of a balancing operation within domain "sd".
3311 * Returns 1 if successful and 0 otherwise.
3313 * Called with both runqueues locked.
3315 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3316 unsigned long max_load_move,
3317 struct sched_domain *sd, enum cpu_idle_type idle,
3320 const struct sched_class *class = sched_class_highest;
3321 unsigned long total_load_moved = 0;
3322 int this_best_prio = this_rq->curr->prio;
3326 class->load_balance(this_rq, this_cpu, busiest,
3327 max_load_move - total_load_moved,
3328 sd, idle, all_pinned, &this_best_prio);
3329 class = class->next;
3331 #ifdef CONFIG_PREEMPT
3333 * NEWIDLE balancing is a source of latency, so preemptible
3334 * kernels will stop after the first task is pulled to minimize
3335 * the critical section.
3337 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3340 } while (class && max_load_move > total_load_moved);
3342 return total_load_moved > 0;
3346 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3347 struct sched_domain *sd, enum cpu_idle_type idle,
3348 struct rq_iterator *iterator)
3350 struct task_struct *p = iterator->start(iterator->arg);
3354 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3355 pull_task(busiest, p, this_rq, this_cpu);
3357 * Right now, this is only the second place pull_task()
3358 * is called, so we can safely collect pull_task()
3359 * stats here rather than inside pull_task().
3361 schedstat_inc(sd, lb_gained[idle]);
3365 p = iterator->next(iterator->arg);
3372 * move_one_task tries to move exactly one task from busiest to this_rq, as
3373 * part of active balancing operations within "domain".
3374 * Returns 1 if successful and 0 otherwise.
3376 * Called with both runqueues locked.
3378 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3379 struct sched_domain *sd, enum cpu_idle_type idle)
3381 const struct sched_class *class;
3383 for (class = sched_class_highest; class; class = class->next)
3384 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3389 /********** Helpers for find_busiest_group ************************/
3391 * sd_lb_stats - Structure to store the statistics of a sched_domain
3392 * during load balancing.
3394 struct sd_lb_stats {
3395 struct sched_group *busiest; /* Busiest group in this sd */
3396 struct sched_group *this; /* Local group in this sd */
3397 unsigned long total_load; /* Total load of all groups in sd */
3398 unsigned long total_pwr; /* Total power of all groups in sd */
3399 unsigned long avg_load; /* Average load across all groups in sd */
3401 /** Statistics of this group */
3402 unsigned long this_load;
3403 unsigned long this_load_per_task;
3404 unsigned long this_nr_running;
3406 /* Statistics of the busiest group */
3407 unsigned long max_load;
3408 unsigned long busiest_load_per_task;
3409 unsigned long busiest_nr_running;
3411 int group_imb; /* Is there imbalance in this sd */
3412 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3413 int power_savings_balance; /* Is powersave balance needed for this sd */
3414 struct sched_group *group_min; /* Least loaded group in sd */
3415 struct sched_group *group_leader; /* Group which relieves group_min */
3416 unsigned long min_load_per_task; /* load_per_task in group_min */
3417 unsigned long leader_nr_running; /* Nr running of group_leader */
3418 unsigned long min_nr_running; /* Nr running of group_min */
3423 * sg_lb_stats - stats of a sched_group required for load_balancing
3425 struct sg_lb_stats {
3426 unsigned long avg_load; /*Avg load across the CPUs of the group */
3427 unsigned long group_load; /* Total load over the CPUs of the group */
3428 unsigned long sum_nr_running; /* Nr tasks running in the group */
3429 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3430 unsigned long group_capacity;
3431 int group_imb; /* Is there an imbalance in the group ? */
3435 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3436 * @group: The group whose first cpu is to be returned.
3438 static inline unsigned int group_first_cpu(struct sched_group *group)
3440 return cpumask_first(sched_group_cpus(group));
3444 * get_sd_load_idx - Obtain the load index for a given sched domain.
3445 * @sd: The sched_domain whose load_idx is to be obtained.
3446 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3448 static inline int get_sd_load_idx(struct sched_domain *sd,
3449 enum cpu_idle_type idle)
3455 load_idx = sd->busy_idx;
3458 case CPU_NEWLY_IDLE:
3459 load_idx = sd->newidle_idx;
3462 load_idx = sd->idle_idx;
3470 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3472 * init_sd_power_savings_stats - Initialize power savings statistics for
3473 * the given sched_domain, during load balancing.
3475 * @sd: Sched domain whose power-savings statistics are to be initialized.
3476 * @sds: Variable containing the statistics for sd.
3477 * @idle: Idle status of the CPU at which we're performing load-balancing.
3479 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3480 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3483 * Busy processors will not participate in power savings
3486 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3487 sds->power_savings_balance = 0;
3489 sds->power_savings_balance = 1;
3490 sds->min_nr_running = ULONG_MAX;
3491 sds->leader_nr_running = 0;
3496 * update_sd_power_savings_stats - Update the power saving stats for a
3497 * sched_domain while performing load balancing.
3499 * @group: sched_group belonging to the sched_domain under consideration.
3500 * @sds: Variable containing the statistics of the sched_domain
3501 * @local_group: Does group contain the CPU for which we're performing
3503 * @sgs: Variable containing the statistics of the group.
3505 static inline void update_sd_power_savings_stats(struct sched_group *group,
3506 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3509 if (!sds->power_savings_balance)
3513 * If the local group is idle or completely loaded
3514 * no need to do power savings balance at this domain
3516 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3517 !sds->this_nr_running))
3518 sds->power_savings_balance = 0;
3521 * If a group is already running at full capacity or idle,
3522 * don't include that group in power savings calculations
3524 if (!sds->power_savings_balance ||
3525 sgs->sum_nr_running >= sgs->group_capacity ||
3526 !sgs->sum_nr_running)
3530 * Calculate the group which has the least non-idle load.
3531 * This is the group from where we need to pick up the load
3534 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3535 (sgs->sum_nr_running == sds->min_nr_running &&
3536 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3537 sds->group_min = group;
3538 sds->min_nr_running = sgs->sum_nr_running;
3539 sds->min_load_per_task = sgs->sum_weighted_load /
3540 sgs->sum_nr_running;
3544 * Calculate the group which is almost near its
3545 * capacity but still has some space to pick up some load
3546 * from other group and save more power
3548 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3551 if (sgs->sum_nr_running > sds->leader_nr_running ||
3552 (sgs->sum_nr_running == sds->leader_nr_running &&
3553 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3554 sds->group_leader = group;
3555 sds->leader_nr_running = sgs->sum_nr_running;
3560 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3561 * @sds: Variable containing the statistics of the sched_domain
3562 * under consideration.
3563 * @this_cpu: Cpu at which we're currently performing load-balancing.
3564 * @imbalance: Variable to store the imbalance.
3567 * Check if we have potential to perform some power-savings balance.
3568 * If yes, set the busiest group to be the least loaded group in the
3569 * sched_domain, so that it's CPUs can be put to idle.
3571 * Returns 1 if there is potential to perform power-savings balance.
3574 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3575 int this_cpu, unsigned long *imbalance)
3577 if (!sds->power_savings_balance)
3580 if (sds->this != sds->group_leader ||
3581 sds->group_leader == sds->group_min)
3584 *imbalance = sds->min_load_per_task;
3585 sds->busiest = sds->group_min;
3587 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3588 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3589 group_first_cpu(sds->group_leader);
3595 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3596 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3597 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3602 static inline void update_sd_power_savings_stats(struct sched_group *group,
3603 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3608 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3609 int this_cpu, unsigned long *imbalance)
3613 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3617 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3618 * @group: sched_group whose statistics are to be updated.
3619 * @this_cpu: Cpu for which load balance is currently performed.
3620 * @idle: Idle status of this_cpu
3621 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3622 * @sd_idle: Idle status of the sched_domain containing group.
3623 * @local_group: Does group contain this_cpu.
3624 * @cpus: Set of cpus considered for load balancing.
3625 * @balance: Should we balance.
3626 * @sgs: variable to hold the statistics for this group.
3628 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3629 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3630 int local_group, const struct cpumask *cpus,
3631 int *balance, struct sg_lb_stats *sgs)
3633 unsigned long load, max_cpu_load, min_cpu_load;
3635 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3636 unsigned long sum_avg_load_per_task;
3637 unsigned long avg_load_per_task;
3640 balance_cpu = group_first_cpu(group);
3642 /* Tally up the load of all CPUs in the group */
3643 sum_avg_load_per_task = avg_load_per_task = 0;
3645 min_cpu_load = ~0UL;
3647 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3648 struct rq *rq = cpu_rq(i);
3650 if (*sd_idle && rq->nr_running)
3653 /* Bias balancing toward cpus of our domain */
3655 if (idle_cpu(i) && !first_idle_cpu) {
3660 load = target_load(i, load_idx);
3662 load = source_load(i, load_idx);
3663 if (load > max_cpu_load)
3664 max_cpu_load = load;
3665 if (min_cpu_load > load)
3666 min_cpu_load = load;
3669 sgs->group_load += load;
3670 sgs->sum_nr_running += rq->nr_running;
3671 sgs->sum_weighted_load += weighted_cpuload(i);
3673 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3677 * First idle cpu or the first cpu(busiest) in this sched group
3678 * is eligible for doing load balancing at this and above
3679 * domains. In the newly idle case, we will allow all the cpu's
3680 * to do the newly idle load balance.
3682 if (idle != CPU_NEWLY_IDLE && local_group &&
3683 balance_cpu != this_cpu && balance) {
3688 /* Adjust by relative CPU power of the group */
3689 sgs->avg_load = sg_div_cpu_power(group,
3690 sgs->group_load * SCHED_LOAD_SCALE);
3694 * Consider the group unbalanced when the imbalance is larger
3695 * than the average weight of two tasks.
3697 * APZ: with cgroup the avg task weight can vary wildly and
3698 * might not be a suitable number - should we keep a
3699 * normalized nr_running number somewhere that negates
3702 avg_load_per_task = sg_div_cpu_power(group,
3703 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3705 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3708 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3713 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3714 * @sd: sched_domain whose statistics are to be updated.
3715 * @this_cpu: Cpu for which load balance is currently performed.
3716 * @idle: Idle status of this_cpu
3717 * @sd_idle: Idle status of the sched_domain containing group.
3718 * @cpus: Set of cpus considered for load balancing.
3719 * @balance: Should we balance.
3720 * @sds: variable to hold the statistics for this sched_domain.
3722 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3723 enum cpu_idle_type idle, int *sd_idle,
3724 const struct cpumask *cpus, int *balance,
3725 struct sd_lb_stats *sds)
3727 struct sched_group *group = sd->groups;
3728 struct sg_lb_stats sgs;
3731 init_sd_power_savings_stats(sd, sds, idle);
3732 load_idx = get_sd_load_idx(sd, idle);
3737 local_group = cpumask_test_cpu(this_cpu,
3738 sched_group_cpus(group));
3739 memset(&sgs, 0, sizeof(sgs));
3740 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3741 local_group, cpus, balance, &sgs);
3743 if (local_group && balance && !(*balance))
3746 sds->total_load += sgs.group_load;
3747 sds->total_pwr += group->__cpu_power;
3750 sds->this_load = sgs.avg_load;
3752 sds->this_nr_running = sgs.sum_nr_running;
3753 sds->this_load_per_task = sgs.sum_weighted_load;
3754 } else if (sgs.avg_load > sds->max_load &&
3755 (sgs.sum_nr_running > sgs.group_capacity ||
3757 sds->max_load = sgs.avg_load;
3758 sds->busiest = group;
3759 sds->busiest_nr_running = sgs.sum_nr_running;
3760 sds->busiest_load_per_task = sgs.sum_weighted_load;
3761 sds->group_imb = sgs.group_imb;
3764 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3765 group = group->next;
3766 } while (group != sd->groups);
3771 * fix_small_imbalance - Calculate the minor imbalance that exists
3772 * amongst the groups of a sched_domain, during
3774 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3775 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3776 * @imbalance: Variable to store the imbalance.
3778 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3779 int this_cpu, unsigned long *imbalance)
3781 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3782 unsigned int imbn = 2;
3784 if (sds->this_nr_running) {
3785 sds->this_load_per_task /= sds->this_nr_running;
3786 if (sds->busiest_load_per_task >
3787 sds->this_load_per_task)
3790 sds->this_load_per_task =
3791 cpu_avg_load_per_task(this_cpu);
3793 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3794 sds->busiest_load_per_task * imbn) {
3795 *imbalance = sds->busiest_load_per_task;
3800 * OK, we don't have enough imbalance to justify moving tasks,
3801 * however we may be able to increase total CPU power used by
3805 pwr_now += sds->busiest->__cpu_power *
3806 min(sds->busiest_load_per_task, sds->max_load);
3807 pwr_now += sds->this->__cpu_power *
3808 min(sds->this_load_per_task, sds->this_load);
3809 pwr_now /= SCHED_LOAD_SCALE;
3811 /* Amount of load we'd subtract */
3812 tmp = sg_div_cpu_power(sds->busiest,
3813 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3814 if (sds->max_load > tmp)
3815 pwr_move += sds->busiest->__cpu_power *
3816 min(sds->busiest_load_per_task, sds->max_load - tmp);
3818 /* Amount of load we'd add */
3819 if (sds->max_load * sds->busiest->__cpu_power <
3820 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3821 tmp = sg_div_cpu_power(sds->this,
3822 sds->max_load * sds->busiest->__cpu_power);
3824 tmp = sg_div_cpu_power(sds->this,
3825 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3826 pwr_move += sds->this->__cpu_power *
3827 min(sds->this_load_per_task, sds->this_load + tmp);
3828 pwr_move /= SCHED_LOAD_SCALE;
3830 /* Move if we gain throughput */
3831 if (pwr_move > pwr_now)
3832 *imbalance = sds->busiest_load_per_task;
3836 * calculate_imbalance - Calculate the amount of imbalance present within the
3837 * groups of a given sched_domain during load balance.
3838 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3839 * @this_cpu: Cpu for which currently load balance is being performed.
3840 * @imbalance: The variable to store the imbalance.
3842 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3843 unsigned long *imbalance)
3845 unsigned long max_pull;
3847 * In the presence of smp nice balancing, certain scenarios can have
3848 * max load less than avg load(as we skip the groups at or below
3849 * its cpu_power, while calculating max_load..)
3851 if (sds->max_load < sds->avg_load) {
3853 return fix_small_imbalance(sds, this_cpu, imbalance);
3856 /* Don't want to pull so many tasks that a group would go idle */
3857 max_pull = min(sds->max_load - sds->avg_load,
3858 sds->max_load - sds->busiest_load_per_task);
3860 /* How much load to actually move to equalise the imbalance */
3861 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3862 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3866 * if *imbalance is less than the average load per runnable task
3867 * there is no gaurantee that any tasks will be moved so we'll have
3868 * a think about bumping its value to force at least one task to be
3871 if (*imbalance < sds->busiest_load_per_task)
3872 return fix_small_imbalance(sds, this_cpu, imbalance);
3875 /******* find_busiest_group() helpers end here *********************/
3878 * find_busiest_group - Returns the busiest group within the sched_domain
3879 * if there is an imbalance. If there isn't an imbalance, and
3880 * the user has opted for power-savings, it returns a group whose
3881 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3882 * such a group exists.
3884 * Also calculates the amount of weighted load which should be moved
3885 * to restore balance.
3887 * @sd: The sched_domain whose busiest group is to be returned.
3888 * @this_cpu: The cpu for which load balancing is currently being performed.
3889 * @imbalance: Variable which stores amount of weighted load which should
3890 * be moved to restore balance/put a group to idle.
3891 * @idle: The idle status of this_cpu.
3892 * @sd_idle: The idleness of sd
3893 * @cpus: The set of CPUs under consideration for load-balancing.
3894 * @balance: Pointer to a variable indicating if this_cpu
3895 * is the appropriate cpu to perform load balancing at this_level.
3897 * Returns: - the busiest group if imbalance exists.
3898 * - If no imbalance and user has opted for power-savings balance,
3899 * return the least loaded group whose CPUs can be
3900 * put to idle by rebalancing its tasks onto our group.
3902 static struct sched_group *
3903 find_busiest_group(struct sched_domain *sd, int this_cpu,
3904 unsigned long *imbalance, enum cpu_idle_type idle,
3905 int *sd_idle, const struct cpumask *cpus, int *balance)
3907 struct sd_lb_stats sds;
3909 memset(&sds, 0, sizeof(sds));
3912 * Compute the various statistics relavent for load balancing at
3915 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3918 /* Cases where imbalance does not exist from POV of this_cpu */
3919 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3921 * 2) There is no busy sibling group to pull from.
3922 * 3) This group is the busiest group.
3923 * 4) This group is more busy than the avg busieness at this
3925 * 5) The imbalance is within the specified limit.
3926 * 6) Any rebalance would lead to ping-pong
3928 if (balance && !(*balance))
3931 if (!sds.busiest || sds.busiest_nr_running == 0)
3934 if (sds.this_load >= sds.max_load)
3937 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3939 if (sds.this_load >= sds.avg_load)
3942 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3945 sds.busiest_load_per_task /= sds.busiest_nr_running;
3947 sds.busiest_load_per_task =
3948 min(sds.busiest_load_per_task, sds.avg_load);
3951 * We're trying to get all the cpus to the average_load, so we don't
3952 * want to push ourselves above the average load, nor do we wish to
3953 * reduce the max loaded cpu below the average load, as either of these
3954 * actions would just result in more rebalancing later, and ping-pong
3955 * tasks around. Thus we look for the minimum possible imbalance.
3956 * Negative imbalances (*we* are more loaded than anyone else) will
3957 * be counted as no imbalance for these purposes -- we can't fix that
3958 * by pulling tasks to us. Be careful of negative numbers as they'll
3959 * appear as very large values with unsigned longs.
3961 if (sds.max_load <= sds.busiest_load_per_task)
3964 /* Looks like there is an imbalance. Compute it */
3965 calculate_imbalance(&sds, this_cpu, imbalance);
3970 * There is no obvious imbalance. But check if we can do some balancing
3973 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3981 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3984 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3985 unsigned long imbalance, const struct cpumask *cpus)
3987 struct rq *busiest = NULL, *rq;
3988 unsigned long max_load = 0;
3991 for_each_cpu(i, sched_group_cpus(group)) {
3994 if (!cpumask_test_cpu(i, cpus))
3998 wl = weighted_cpuload(i);
4000 if (rq->nr_running == 1 && wl > imbalance)
4003 if (wl > max_load) {
4013 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4014 * so long as it is large enough.
4016 #define MAX_PINNED_INTERVAL 512
4018 /* Working cpumask for load_balance and load_balance_newidle. */
4019 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4022 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4023 * tasks if there is an imbalance.
4025 static int load_balance(int this_cpu, struct rq *this_rq,
4026 struct sched_domain *sd, enum cpu_idle_type idle,
4029 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4030 struct sched_group *group;
4031 unsigned long imbalance;
4033 unsigned long flags;
4034 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4036 cpumask_setall(cpus);
4039 * When power savings policy is enabled for the parent domain, idle
4040 * sibling can pick up load irrespective of busy siblings. In this case,
4041 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4042 * portraying it as CPU_NOT_IDLE.
4044 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4045 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4048 schedstat_inc(sd, lb_count[idle]);
4052 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4059 schedstat_inc(sd, lb_nobusyg[idle]);
4063 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4065 schedstat_inc(sd, lb_nobusyq[idle]);
4069 BUG_ON(busiest == this_rq);
4071 schedstat_add(sd, lb_imbalance[idle], imbalance);
4074 if (busiest->nr_running > 1) {
4076 * Attempt to move tasks. If find_busiest_group has found
4077 * an imbalance but busiest->nr_running <= 1, the group is
4078 * still unbalanced. ld_moved simply stays zero, so it is
4079 * correctly treated as an imbalance.
4081 local_irq_save(flags);
4082 double_rq_lock(this_rq, busiest);
4083 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4084 imbalance, sd, idle, &all_pinned);
4085 double_rq_unlock(this_rq, busiest);
4086 local_irq_restore(flags);
4089 * some other cpu did the load balance for us.
4091 if (ld_moved && this_cpu != smp_processor_id())
4092 resched_cpu(this_cpu);
4094 /* All tasks on this runqueue were pinned by CPU affinity */
4095 if (unlikely(all_pinned)) {
4096 cpumask_clear_cpu(cpu_of(busiest), cpus);
4097 if (!cpumask_empty(cpus))
4104 schedstat_inc(sd, lb_failed[idle]);
4105 sd->nr_balance_failed++;
4107 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4109 spin_lock_irqsave(&busiest->lock, flags);
4111 /* don't kick the migration_thread, if the curr
4112 * task on busiest cpu can't be moved to this_cpu
4114 if (!cpumask_test_cpu(this_cpu,
4115 &busiest->curr->cpus_allowed)) {
4116 spin_unlock_irqrestore(&busiest->lock, flags);
4118 goto out_one_pinned;
4121 if (!busiest->active_balance) {
4122 busiest->active_balance = 1;
4123 busiest->push_cpu = this_cpu;
4126 spin_unlock_irqrestore(&busiest->lock, flags);
4128 wake_up_process(busiest->migration_thread);
4131 * We've kicked active balancing, reset the failure
4134 sd->nr_balance_failed = sd->cache_nice_tries+1;
4137 sd->nr_balance_failed = 0;
4139 if (likely(!active_balance)) {
4140 /* We were unbalanced, so reset the balancing interval */
4141 sd->balance_interval = sd->min_interval;
4144 * If we've begun active balancing, start to back off. This
4145 * case may not be covered by the all_pinned logic if there
4146 * is only 1 task on the busy runqueue (because we don't call
4149 if (sd->balance_interval < sd->max_interval)
4150 sd->balance_interval *= 2;
4153 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4154 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4160 schedstat_inc(sd, lb_balanced[idle]);
4162 sd->nr_balance_failed = 0;
4165 /* tune up the balancing interval */
4166 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4167 (sd->balance_interval < sd->max_interval))
4168 sd->balance_interval *= 2;
4170 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4171 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4182 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4183 * tasks if there is an imbalance.
4185 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4186 * this_rq is locked.
4189 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4191 struct sched_group *group;
4192 struct rq *busiest = NULL;
4193 unsigned long imbalance;
4197 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4199 cpumask_setall(cpus);
4202 * When power savings policy is enabled for the parent domain, idle
4203 * sibling can pick up load irrespective of busy siblings. In this case,
4204 * let the state of idle sibling percolate up as IDLE, instead of
4205 * portraying it as CPU_NOT_IDLE.
4207 if (sd->flags & SD_SHARE_CPUPOWER &&
4208 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4211 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4213 update_shares_locked(this_rq, sd);
4214 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4215 &sd_idle, cpus, NULL);
4217 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4221 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4223 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4227 BUG_ON(busiest == this_rq);
4229 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4232 if (busiest->nr_running > 1) {
4233 /* Attempt to move tasks */
4234 double_lock_balance(this_rq, busiest);
4235 /* this_rq->clock is already updated */
4236 update_rq_clock(busiest);
4237 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4238 imbalance, sd, CPU_NEWLY_IDLE,
4240 double_unlock_balance(this_rq, busiest);
4242 if (unlikely(all_pinned)) {
4243 cpumask_clear_cpu(cpu_of(busiest), cpus);
4244 if (!cpumask_empty(cpus))
4250 int active_balance = 0;
4252 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4253 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4254 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4257 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4260 if (sd->nr_balance_failed++ < 2)
4264 * The only task running in a non-idle cpu can be moved to this
4265 * cpu in an attempt to completely freeup the other CPU
4266 * package. The same method used to move task in load_balance()
4267 * have been extended for load_balance_newidle() to speedup
4268 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4270 * The package power saving logic comes from
4271 * find_busiest_group(). If there are no imbalance, then
4272 * f_b_g() will return NULL. However when sched_mc={1,2} then
4273 * f_b_g() will select a group from which a running task may be
4274 * pulled to this cpu in order to make the other package idle.
4275 * If there is no opportunity to make a package idle and if
4276 * there are no imbalance, then f_b_g() will return NULL and no
4277 * action will be taken in load_balance_newidle().
4279 * Under normal task pull operation due to imbalance, there
4280 * will be more than one task in the source run queue and
4281 * move_tasks() will succeed. ld_moved will be true and this
4282 * active balance code will not be triggered.
4285 /* Lock busiest in correct order while this_rq is held */
4286 double_lock_balance(this_rq, busiest);
4289 * don't kick the migration_thread, if the curr
4290 * task on busiest cpu can't be moved to this_cpu
4292 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4293 double_unlock_balance(this_rq, busiest);
4298 if (!busiest->active_balance) {
4299 busiest->active_balance = 1;
4300 busiest->push_cpu = this_cpu;
4304 double_unlock_balance(this_rq, busiest);
4306 * Should not call ttwu while holding a rq->lock
4308 spin_unlock(&this_rq->lock);
4310 wake_up_process(busiest->migration_thread);
4311 spin_lock(&this_rq->lock);
4314 sd->nr_balance_failed = 0;
4316 update_shares_locked(this_rq, sd);
4320 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4321 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4322 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4324 sd->nr_balance_failed = 0;
4330 * idle_balance is called by schedule() if this_cpu is about to become
4331 * idle. Attempts to pull tasks from other CPUs.
4333 static void idle_balance(int this_cpu, struct rq *this_rq)
4335 struct sched_domain *sd;
4336 int pulled_task = 0;
4337 unsigned long next_balance = jiffies + HZ;
4339 for_each_domain(this_cpu, sd) {
4340 unsigned long interval;
4342 if (!(sd->flags & SD_LOAD_BALANCE))
4345 if (sd->flags & SD_BALANCE_NEWIDLE)
4346 /* If we've pulled tasks over stop searching: */
4347 pulled_task = load_balance_newidle(this_cpu, this_rq,
4350 interval = msecs_to_jiffies(sd->balance_interval);
4351 if (time_after(next_balance, sd->last_balance + interval))
4352 next_balance = sd->last_balance + interval;
4356 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4358 * We are going idle. next_balance may be set based on
4359 * a busy processor. So reset next_balance.
4361 this_rq->next_balance = next_balance;
4366 * active_load_balance is run by migration threads. It pushes running tasks
4367 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4368 * running on each physical CPU where possible, and avoids physical /
4369 * logical imbalances.
4371 * Called with busiest_rq locked.
4373 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4375 int target_cpu = busiest_rq->push_cpu;
4376 struct sched_domain *sd;
4377 struct rq *target_rq;
4379 /* Is there any task to move? */
4380 if (busiest_rq->nr_running <= 1)
4383 target_rq = cpu_rq(target_cpu);
4386 * This condition is "impossible", if it occurs
4387 * we need to fix it. Originally reported by
4388 * Bjorn Helgaas on a 128-cpu setup.
4390 BUG_ON(busiest_rq == target_rq);
4392 /* move a task from busiest_rq to target_rq */
4393 double_lock_balance(busiest_rq, target_rq);
4394 update_rq_clock(busiest_rq);
4395 update_rq_clock(target_rq);
4397 /* Search for an sd spanning us and the target CPU. */
4398 for_each_domain(target_cpu, sd) {
4399 if ((sd->flags & SD_LOAD_BALANCE) &&
4400 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4405 schedstat_inc(sd, alb_count);
4407 if (move_one_task(target_rq, target_cpu, busiest_rq,
4409 schedstat_inc(sd, alb_pushed);
4411 schedstat_inc(sd, alb_failed);
4413 double_unlock_balance(busiest_rq, target_rq);
4418 atomic_t load_balancer;
4419 cpumask_var_t cpu_mask;
4420 cpumask_var_t ilb_grp_nohz_mask;
4421 } nohz ____cacheline_aligned = {
4422 .load_balancer = ATOMIC_INIT(-1),
4425 int get_nohz_load_balancer(void)
4427 return atomic_read(&nohz.load_balancer);
4430 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4432 * lowest_flag_domain - Return lowest sched_domain containing flag.
4433 * @cpu: The cpu whose lowest level of sched domain is to
4435 * @flag: The flag to check for the lowest sched_domain
4436 * for the given cpu.
4438 * Returns the lowest sched_domain of a cpu which contains the given flag.
4440 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4442 struct sched_domain *sd;
4444 for_each_domain(cpu, sd)
4445 if (sd && (sd->flags & flag))
4452 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4453 * @cpu: The cpu whose domains we're iterating over.
4454 * @sd: variable holding the value of the power_savings_sd
4456 * @flag: The flag to filter the sched_domains to be iterated.
4458 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4459 * set, starting from the lowest sched_domain to the highest.
4461 #define for_each_flag_domain(cpu, sd, flag) \
4462 for (sd = lowest_flag_domain(cpu, flag); \
4463 (sd && (sd->flags & flag)); sd = sd->parent)
4466 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4467 * @ilb_group: group to be checked for semi-idleness
4469 * Returns: 1 if the group is semi-idle. 0 otherwise.
4471 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4472 * and atleast one non-idle CPU. This helper function checks if the given
4473 * sched_group is semi-idle or not.
4475 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4477 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4478 sched_group_cpus(ilb_group));
4481 * A sched_group is semi-idle when it has atleast one busy cpu
4482 * and atleast one idle cpu.
4484 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4487 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4493 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4494 * @cpu: The cpu which is nominating a new idle_load_balancer.
4496 * Returns: Returns the id of the idle load balancer if it exists,
4497 * Else, returns >= nr_cpu_ids.
4499 * This algorithm picks the idle load balancer such that it belongs to a
4500 * semi-idle powersavings sched_domain. The idea is to try and avoid
4501 * completely idle packages/cores just for the purpose of idle load balancing
4502 * when there are other idle cpu's which are better suited for that job.
4504 static int find_new_ilb(int cpu)
4506 struct sched_domain *sd;
4507 struct sched_group *ilb_group;
4510 * Have idle load balancer selection from semi-idle packages only
4511 * when power-aware load balancing is enabled
4513 if (!(sched_smt_power_savings || sched_mc_power_savings))
4517 * Optimize for the case when we have no idle CPUs or only one
4518 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4520 if (cpumask_weight(nohz.cpu_mask) < 2)
4523 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4524 ilb_group = sd->groups;
4527 if (is_semi_idle_group(ilb_group))
4528 return cpumask_first(nohz.ilb_grp_nohz_mask);
4530 ilb_group = ilb_group->next;
4532 } while (ilb_group != sd->groups);
4536 return cpumask_first(nohz.cpu_mask);
4538 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4539 static inline int find_new_ilb(int call_cpu)
4541 return cpumask_first(nohz.cpu_mask);
4546 * This routine will try to nominate the ilb (idle load balancing)
4547 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4548 * load balancing on behalf of all those cpus. If all the cpus in the system
4549 * go into this tickless mode, then there will be no ilb owner (as there is
4550 * no need for one) and all the cpus will sleep till the next wakeup event
4553 * For the ilb owner, tick is not stopped. And this tick will be used
4554 * for idle load balancing. ilb owner will still be part of
4557 * While stopping the tick, this cpu will become the ilb owner if there
4558 * is no other owner. And will be the owner till that cpu becomes busy
4559 * or if all cpus in the system stop their ticks at which point
4560 * there is no need for ilb owner.
4562 * When the ilb owner becomes busy, it nominates another owner, during the
4563 * next busy scheduler_tick()
4565 int select_nohz_load_balancer(int stop_tick)
4567 int cpu = smp_processor_id();
4570 cpu_rq(cpu)->in_nohz_recently = 1;
4572 if (!cpu_active(cpu)) {
4573 if (atomic_read(&nohz.load_balancer) != cpu)
4577 * If we are going offline and still the leader,
4580 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4586 cpumask_set_cpu(cpu, nohz.cpu_mask);
4588 /* time for ilb owner also to sleep */
4589 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4590 if (atomic_read(&nohz.load_balancer) == cpu)
4591 atomic_set(&nohz.load_balancer, -1);
4595 if (atomic_read(&nohz.load_balancer) == -1) {
4596 /* make me the ilb owner */
4597 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4599 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4602 if (!(sched_smt_power_savings ||
4603 sched_mc_power_savings))
4606 * Check to see if there is a more power-efficient
4609 new_ilb = find_new_ilb(cpu);
4610 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4611 atomic_set(&nohz.load_balancer, -1);
4612 resched_cpu(new_ilb);
4618 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4621 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4623 if (atomic_read(&nohz.load_balancer) == cpu)
4624 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4631 static DEFINE_SPINLOCK(balancing);
4634 * It checks each scheduling domain to see if it is due to be balanced,
4635 * and initiates a balancing operation if so.
4637 * Balancing parameters are set up in arch_init_sched_domains.
4639 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4642 struct rq *rq = cpu_rq(cpu);
4643 unsigned long interval;
4644 struct sched_domain *sd;
4645 /* Earliest time when we have to do rebalance again */
4646 unsigned long next_balance = jiffies + 60*HZ;
4647 int update_next_balance = 0;
4650 for_each_domain(cpu, sd) {
4651 if (!(sd->flags & SD_LOAD_BALANCE))
4654 interval = sd->balance_interval;
4655 if (idle != CPU_IDLE)
4656 interval *= sd->busy_factor;
4658 /* scale ms to jiffies */
4659 interval = msecs_to_jiffies(interval);
4660 if (unlikely(!interval))
4662 if (interval > HZ*NR_CPUS/10)
4663 interval = HZ*NR_CPUS/10;
4665 need_serialize = sd->flags & SD_SERIALIZE;
4667 if (need_serialize) {
4668 if (!spin_trylock(&balancing))
4672 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4673 if (load_balance(cpu, rq, sd, idle, &balance)) {
4675 * We've pulled tasks over so either we're no
4676 * longer idle, or one of our SMT siblings is
4679 idle = CPU_NOT_IDLE;
4681 sd->last_balance = jiffies;
4684 spin_unlock(&balancing);
4686 if (time_after(next_balance, sd->last_balance + interval)) {
4687 next_balance = sd->last_balance + interval;
4688 update_next_balance = 1;
4692 * Stop the load balance at this level. There is another
4693 * CPU in our sched group which is doing load balancing more
4701 * next_balance will be updated only when there is a need.
4702 * When the cpu is attached to null domain for ex, it will not be
4705 if (likely(update_next_balance))
4706 rq->next_balance = next_balance;
4710 * run_rebalance_domains is triggered when needed from the scheduler tick.
4711 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4712 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4714 static void run_rebalance_domains(struct softirq_action *h)
4716 int this_cpu = smp_processor_id();
4717 struct rq *this_rq = cpu_rq(this_cpu);
4718 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4719 CPU_IDLE : CPU_NOT_IDLE;
4721 rebalance_domains(this_cpu, idle);
4725 * If this cpu is the owner for idle load balancing, then do the
4726 * balancing on behalf of the other idle cpus whose ticks are
4729 if (this_rq->idle_at_tick &&
4730 atomic_read(&nohz.load_balancer) == this_cpu) {
4734 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4735 if (balance_cpu == this_cpu)
4739 * If this cpu gets work to do, stop the load balancing
4740 * work being done for other cpus. Next load
4741 * balancing owner will pick it up.
4746 rebalance_domains(balance_cpu, CPU_IDLE);
4748 rq = cpu_rq(balance_cpu);
4749 if (time_after(this_rq->next_balance, rq->next_balance))
4750 this_rq->next_balance = rq->next_balance;
4756 static inline int on_null_domain(int cpu)
4758 return !rcu_dereference(cpu_rq(cpu)->sd);
4762 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4764 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4765 * idle load balancing owner or decide to stop the periodic load balancing,
4766 * if the whole system is idle.
4768 static inline void trigger_load_balance(struct rq *rq, int cpu)
4772 * If we were in the nohz mode recently and busy at the current
4773 * scheduler tick, then check if we need to nominate new idle
4776 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4777 rq->in_nohz_recently = 0;
4779 if (atomic_read(&nohz.load_balancer) == cpu) {
4780 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4781 atomic_set(&nohz.load_balancer, -1);
4784 if (atomic_read(&nohz.load_balancer) == -1) {
4785 int ilb = find_new_ilb(cpu);
4787 if (ilb < nr_cpu_ids)
4793 * If this cpu is idle and doing idle load balancing for all the
4794 * cpus with ticks stopped, is it time for that to stop?
4796 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4797 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4803 * If this cpu is idle and the idle load balancing is done by
4804 * someone else, then no need raise the SCHED_SOFTIRQ
4806 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4807 cpumask_test_cpu(cpu, nohz.cpu_mask))
4810 /* Don't need to rebalance while attached to NULL domain */
4811 if (time_after_eq(jiffies, rq->next_balance) &&
4812 likely(!on_null_domain(cpu)))
4813 raise_softirq(SCHED_SOFTIRQ);
4816 #else /* CONFIG_SMP */
4819 * on UP we do not need to balance between CPUs:
4821 static inline void idle_balance(int cpu, struct rq *rq)
4827 DEFINE_PER_CPU(struct kernel_stat, kstat);
4829 EXPORT_PER_CPU_SYMBOL(kstat);
4832 * Return any ns on the sched_clock that have not yet been accounted in
4833 * @p in case that task is currently running.
4835 * Called with task_rq_lock() held on @rq.
4837 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4841 if (task_current(rq, p)) {
4842 update_rq_clock(rq);
4843 ns = rq->clock - p->se.exec_start;
4851 unsigned long long task_delta_exec(struct task_struct *p)
4853 unsigned long flags;
4857 rq = task_rq_lock(p, &flags);
4858 ns = do_task_delta_exec(p, rq);
4859 task_rq_unlock(rq, &flags);
4865 * Return accounted runtime for the task.
4866 * In case the task is currently running, return the runtime plus current's
4867 * pending runtime that have not been accounted yet.
4869 unsigned long long task_sched_runtime(struct task_struct *p)
4871 unsigned long flags;
4875 rq = task_rq_lock(p, &flags);
4876 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4877 task_rq_unlock(rq, &flags);
4883 * Return sum_exec_runtime for the thread group.
4884 * In case the task is currently running, return the sum plus current's
4885 * pending runtime that have not been accounted yet.
4887 * Note that the thread group might have other running tasks as well,
4888 * so the return value not includes other pending runtime that other
4889 * running tasks might have.
4891 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4893 struct task_cputime totals;
4894 unsigned long flags;
4898 rq = task_rq_lock(p, &flags);
4899 thread_group_cputime(p, &totals);
4900 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4901 task_rq_unlock(rq, &flags);
4907 * Account user cpu time to a process.
4908 * @p: the process that the cpu time gets accounted to
4909 * @cputime: the cpu time spent in user space since the last update
4910 * @cputime_scaled: cputime scaled by cpu frequency
4912 void account_user_time(struct task_struct *p, cputime_t cputime,
4913 cputime_t cputime_scaled)
4915 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4918 /* Add user time to process. */
4919 p->utime = cputime_add(p->utime, cputime);
4920 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4921 account_group_user_time(p, cputime);
4923 /* Add user time to cpustat. */
4924 tmp = cputime_to_cputime64(cputime);
4925 if (TASK_NICE(p) > 0)
4926 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4928 cpustat->user = cputime64_add(cpustat->user, tmp);
4930 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4931 /* Account for user time used */
4932 acct_update_integrals(p);
4936 * Account guest cpu time to a process.
4937 * @p: the process that the cpu time gets accounted to
4938 * @cputime: the cpu time spent in virtual machine since the last update
4939 * @cputime_scaled: cputime scaled by cpu frequency
4941 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4942 cputime_t cputime_scaled)
4945 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4947 tmp = cputime_to_cputime64(cputime);
4949 /* Add guest time to process. */
4950 p->utime = cputime_add(p->utime, cputime);
4951 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4952 account_group_user_time(p, cputime);
4953 p->gtime = cputime_add(p->gtime, cputime);
4955 /* Add guest time to cpustat. */
4956 cpustat->user = cputime64_add(cpustat->user, tmp);
4957 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4961 * Account system cpu time to a process.
4962 * @p: the process that the cpu time gets accounted to
4963 * @hardirq_offset: the offset to subtract from hardirq_count()
4964 * @cputime: the cpu time spent in kernel space since the last update
4965 * @cputime_scaled: cputime scaled by cpu frequency
4967 void account_system_time(struct task_struct *p, int hardirq_offset,
4968 cputime_t cputime, cputime_t cputime_scaled)
4970 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4973 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4974 account_guest_time(p, cputime, cputime_scaled);
4978 /* Add system time to process. */
4979 p->stime = cputime_add(p->stime, cputime);
4980 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4981 account_group_system_time(p, cputime);
4983 /* Add system time to cpustat. */
4984 tmp = cputime_to_cputime64(cputime);
4985 if (hardirq_count() - hardirq_offset)
4986 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4987 else if (softirq_count())
4988 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4990 cpustat->system = cputime64_add(cpustat->system, tmp);
4992 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4994 /* Account for system time used */
4995 acct_update_integrals(p);
4999 * Account for involuntary wait time.
5000 * @steal: the cpu time spent in involuntary wait
5002 void account_steal_time(cputime_t cputime)
5004 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5005 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5007 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5011 * Account for idle time.
5012 * @cputime: the cpu time spent in idle wait
5014 void account_idle_time(cputime_t cputime)
5016 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5017 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5018 struct rq *rq = this_rq();
5020 if (atomic_read(&rq->nr_iowait) > 0)
5021 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5023 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5026 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5029 * Account a single tick of cpu time.
5030 * @p: the process that the cpu time gets accounted to
5031 * @user_tick: indicates if the tick is a user or a system tick
5033 void account_process_tick(struct task_struct *p, int user_tick)
5035 cputime_t one_jiffy = jiffies_to_cputime(1);
5036 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5037 struct rq *rq = this_rq();
5040 account_user_time(p, one_jiffy, one_jiffy_scaled);
5041 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5042 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5045 account_idle_time(one_jiffy);
5049 * Account multiple ticks of steal time.
5050 * @p: the process from which the cpu time has been stolen
5051 * @ticks: number of stolen ticks
5053 void account_steal_ticks(unsigned long ticks)
5055 account_steal_time(jiffies_to_cputime(ticks));
5059 * Account multiple ticks of idle time.
5060 * @ticks: number of stolen ticks
5062 void account_idle_ticks(unsigned long ticks)
5064 account_idle_time(jiffies_to_cputime(ticks));
5070 * Use precise platform statistics if available:
5072 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5073 cputime_t task_utime(struct task_struct *p)
5078 cputime_t task_stime(struct task_struct *p)
5083 cputime_t task_utime(struct task_struct *p)
5085 clock_t utime = cputime_to_clock_t(p->utime),
5086 total = utime + cputime_to_clock_t(p->stime);
5090 * Use CFS's precise accounting:
5092 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5096 do_div(temp, total);
5098 utime = (clock_t)temp;
5100 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5101 return p->prev_utime;
5104 cputime_t task_stime(struct task_struct *p)
5109 * Use CFS's precise accounting. (we subtract utime from
5110 * the total, to make sure the total observed by userspace
5111 * grows monotonically - apps rely on that):
5113 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5114 cputime_to_clock_t(task_utime(p));
5117 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5119 return p->prev_stime;
5123 inline cputime_t task_gtime(struct task_struct *p)
5129 * This function gets called by the timer code, with HZ frequency.
5130 * We call it with interrupts disabled.
5132 * It also gets called by the fork code, when changing the parent's
5135 void scheduler_tick(void)
5137 int cpu = smp_processor_id();
5138 struct rq *rq = cpu_rq(cpu);
5139 struct task_struct *curr = rq->curr;
5143 spin_lock(&rq->lock);
5144 update_rq_clock(rq);
5145 update_cpu_load(rq);
5146 curr->sched_class->task_tick(rq, curr, 0);
5147 spin_unlock(&rq->lock);
5149 perf_counter_task_tick(curr, cpu);
5152 rq->idle_at_tick = idle_cpu(cpu);
5153 trigger_load_balance(rq, cpu);
5157 notrace unsigned long get_parent_ip(unsigned long addr)
5159 if (in_lock_functions(addr)) {
5160 addr = CALLER_ADDR2;
5161 if (in_lock_functions(addr))
5162 addr = CALLER_ADDR3;
5167 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5168 defined(CONFIG_PREEMPT_TRACER))
5170 void __kprobes add_preempt_count(int val)
5172 #ifdef CONFIG_DEBUG_PREEMPT
5176 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5179 preempt_count() += val;
5180 #ifdef CONFIG_DEBUG_PREEMPT
5182 * Spinlock count overflowing soon?
5184 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5187 if (preempt_count() == val)
5188 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5190 EXPORT_SYMBOL(add_preempt_count);
5192 void __kprobes sub_preempt_count(int val)
5194 #ifdef CONFIG_DEBUG_PREEMPT
5198 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5201 * Is the spinlock portion underflowing?
5203 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5204 !(preempt_count() & PREEMPT_MASK)))
5208 if (preempt_count() == val)
5209 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5210 preempt_count() -= val;
5212 EXPORT_SYMBOL(sub_preempt_count);
5217 * Print scheduling while atomic bug:
5219 static noinline void __schedule_bug(struct task_struct *prev)
5221 struct pt_regs *regs = get_irq_regs();
5223 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5224 prev->comm, prev->pid, preempt_count());
5226 debug_show_held_locks(prev);
5228 if (irqs_disabled())
5229 print_irqtrace_events(prev);
5238 * Various schedule()-time debugging checks and statistics:
5240 static inline void schedule_debug(struct task_struct *prev)
5243 * Test if we are atomic. Since do_exit() needs to call into
5244 * schedule() atomically, we ignore that path for now.
5245 * Otherwise, whine if we are scheduling when we should not be.
5247 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5248 __schedule_bug(prev);
5250 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5252 schedstat_inc(this_rq(), sched_count);
5253 #ifdef CONFIG_SCHEDSTATS
5254 if (unlikely(prev->lock_depth >= 0)) {
5255 schedstat_inc(this_rq(), bkl_count);
5256 schedstat_inc(prev, sched_info.bkl_count);
5261 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5263 if (prev->state == TASK_RUNNING) {
5264 u64 runtime = prev->se.sum_exec_runtime;
5266 runtime -= prev->se.prev_sum_exec_runtime;
5267 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5270 * In order to avoid avg_overlap growing stale when we are
5271 * indeed overlapping and hence not getting put to sleep, grow
5272 * the avg_overlap on preemption.
5274 * We use the average preemption runtime because that
5275 * correlates to the amount of cache footprint a task can
5278 update_avg(&prev->se.avg_overlap, runtime);
5280 prev->sched_class->put_prev_task(rq, prev);
5284 * Pick up the highest-prio task:
5286 static inline struct task_struct *
5287 pick_next_task(struct rq *rq)
5289 const struct sched_class *class;
5290 struct task_struct *p;
5293 * Optimization: we know that if all tasks are in
5294 * the fair class we can call that function directly:
5296 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5297 p = fair_sched_class.pick_next_task(rq);
5302 class = sched_class_highest;
5304 p = class->pick_next_task(rq);
5308 * Will never be NULL as the idle class always
5309 * returns a non-NULL p:
5311 class = class->next;
5316 * schedule() is the main scheduler function.
5318 asmlinkage void __sched schedule(void)
5320 struct task_struct *prev, *next;
5321 unsigned long *switch_count;
5327 cpu = smp_processor_id();
5331 switch_count = &prev->nivcsw;
5333 release_kernel_lock(prev);
5334 need_resched_nonpreemptible:
5336 schedule_debug(prev);
5338 if (sched_feat(HRTICK))
5341 spin_lock_irq(&rq->lock);
5342 update_rq_clock(rq);
5343 clear_tsk_need_resched(prev);
5345 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5346 if (unlikely(signal_pending_state(prev->state, prev)))
5347 prev->state = TASK_RUNNING;
5349 deactivate_task(rq, prev, 1);
5350 switch_count = &prev->nvcsw;
5354 if (prev->sched_class->pre_schedule)
5355 prev->sched_class->pre_schedule(rq, prev);
5358 if (unlikely(!rq->nr_running))
5359 idle_balance(cpu, rq);
5361 put_prev_task(rq, prev);
5362 next = pick_next_task(rq);
5364 if (likely(prev != next)) {
5365 sched_info_switch(prev, next);
5366 perf_counter_task_sched_out(prev, next, cpu);
5372 context_switch(rq, prev, next); /* unlocks the rq */
5374 * the context switch might have flipped the stack from under
5375 * us, hence refresh the local variables.
5377 cpu = smp_processor_id();
5380 spin_unlock_irq(&rq->lock);
5382 if (unlikely(reacquire_kernel_lock(current) < 0))
5383 goto need_resched_nonpreemptible;
5385 preempt_enable_no_resched();
5389 EXPORT_SYMBOL(schedule);
5393 * Look out! "owner" is an entirely speculative pointer
5394 * access and not reliable.
5396 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5401 if (!sched_feat(OWNER_SPIN))
5404 #ifdef CONFIG_DEBUG_PAGEALLOC
5406 * Need to access the cpu field knowing that
5407 * DEBUG_PAGEALLOC could have unmapped it if
5408 * the mutex owner just released it and exited.
5410 if (probe_kernel_address(&owner->cpu, cpu))
5417 * Even if the access succeeded (likely case),
5418 * the cpu field may no longer be valid.
5420 if (cpu >= nr_cpumask_bits)
5424 * We need to validate that we can do a
5425 * get_cpu() and that we have the percpu area.
5427 if (!cpu_online(cpu))
5434 * Owner changed, break to re-assess state.
5436 if (lock->owner != owner)
5440 * Is that owner really running on that cpu?
5442 if (task_thread_info(rq->curr) != owner || need_resched())
5452 #ifdef CONFIG_PREEMPT
5454 * this is the entry point to schedule() from in-kernel preemption
5455 * off of preempt_enable. Kernel preemptions off return from interrupt
5456 * occur there and call schedule directly.
5458 asmlinkage void __sched preempt_schedule(void)
5460 struct thread_info *ti = current_thread_info();
5463 * If there is a non-zero preempt_count or interrupts are disabled,
5464 * we do not want to preempt the current task. Just return..
5466 if (likely(ti->preempt_count || irqs_disabled()))
5470 add_preempt_count(PREEMPT_ACTIVE);
5472 sub_preempt_count(PREEMPT_ACTIVE);
5475 * Check again in case we missed a preemption opportunity
5476 * between schedule and now.
5479 } while (need_resched());
5481 EXPORT_SYMBOL(preempt_schedule);
5484 * this is the entry point to schedule() from kernel preemption
5485 * off of irq context.
5486 * Note, that this is called and return with irqs disabled. This will
5487 * protect us against recursive calling from irq.
5489 asmlinkage void __sched preempt_schedule_irq(void)
5491 struct thread_info *ti = current_thread_info();
5493 /* Catch callers which need to be fixed */
5494 BUG_ON(ti->preempt_count || !irqs_disabled());
5497 add_preempt_count(PREEMPT_ACTIVE);
5500 local_irq_disable();
5501 sub_preempt_count(PREEMPT_ACTIVE);
5504 * Check again in case we missed a preemption opportunity
5505 * between schedule and now.
5508 } while (need_resched());
5511 #endif /* CONFIG_PREEMPT */
5513 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5516 return try_to_wake_up(curr->private, mode, sync);
5518 EXPORT_SYMBOL(default_wake_function);
5521 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5522 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5523 * number) then we wake all the non-exclusive tasks and one exclusive task.
5525 * There are circumstances in which we can try to wake a task which has already
5526 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5527 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5529 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5530 int nr_exclusive, int sync, void *key)
5532 wait_queue_t *curr, *next;
5534 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5535 unsigned flags = curr->flags;
5537 if (curr->func(curr, mode, sync, key) &&
5538 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5544 * __wake_up - wake up threads blocked on a waitqueue.
5546 * @mode: which threads
5547 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5548 * @key: is directly passed to the wakeup function
5550 * It may be assumed that this function implies a write memory barrier before
5551 * changing the task state if and only if any tasks are woken up.
5553 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5554 int nr_exclusive, void *key)
5556 unsigned long flags;
5558 spin_lock_irqsave(&q->lock, flags);
5559 __wake_up_common(q, mode, nr_exclusive, 0, key);
5560 spin_unlock_irqrestore(&q->lock, flags);
5562 EXPORT_SYMBOL(__wake_up);
5565 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5567 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5569 __wake_up_common(q, mode, 1, 0, NULL);
5572 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5574 __wake_up_common(q, mode, 1, 0, key);
5578 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5580 * @mode: which threads
5581 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5582 * @key: opaque value to be passed to wakeup targets
5584 * The sync wakeup differs that the waker knows that it will schedule
5585 * away soon, so while the target thread will be woken up, it will not
5586 * be migrated to another CPU - ie. the two threads are 'synchronized'
5587 * with each other. This can prevent needless bouncing between CPUs.
5589 * On UP it can prevent extra preemption.
5591 * It may be assumed that this function implies a write memory barrier before
5592 * changing the task state if and only if any tasks are woken up.
5594 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5595 int nr_exclusive, void *key)
5597 unsigned long flags;
5603 if (unlikely(!nr_exclusive))
5606 spin_lock_irqsave(&q->lock, flags);
5607 __wake_up_common(q, mode, nr_exclusive, sync, key);
5608 spin_unlock_irqrestore(&q->lock, flags);
5610 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5613 * __wake_up_sync - see __wake_up_sync_key()
5615 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5617 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5619 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5622 * complete: - signals a single thread waiting on this completion
5623 * @x: holds the state of this particular completion
5625 * This will wake up a single thread waiting on this completion. Threads will be
5626 * awakened in the same order in which they were queued.
5628 * See also complete_all(), wait_for_completion() and related routines.
5630 * It may be assumed that this function implies a write memory barrier before
5631 * changing the task state if and only if any tasks are woken up.
5633 void complete(struct completion *x)
5635 unsigned long flags;
5637 spin_lock_irqsave(&x->wait.lock, flags);
5639 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5640 spin_unlock_irqrestore(&x->wait.lock, flags);
5642 EXPORT_SYMBOL(complete);
5645 * complete_all: - signals all threads waiting on this completion
5646 * @x: holds the state of this particular completion
5648 * This will wake up all threads waiting on this particular completion event.
5650 * It may be assumed that this function implies a write memory barrier before
5651 * changing the task state if and only if any tasks are woken up.
5653 void complete_all(struct completion *x)
5655 unsigned long flags;
5657 spin_lock_irqsave(&x->wait.lock, flags);
5658 x->done += UINT_MAX/2;
5659 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5660 spin_unlock_irqrestore(&x->wait.lock, flags);
5662 EXPORT_SYMBOL(complete_all);
5664 static inline long __sched
5665 do_wait_for_common(struct completion *x, long timeout, int state)
5668 DECLARE_WAITQUEUE(wait, current);
5670 wait.flags |= WQ_FLAG_EXCLUSIVE;
5671 __add_wait_queue_tail(&x->wait, &wait);
5673 if (signal_pending_state(state, current)) {
5674 timeout = -ERESTARTSYS;
5677 __set_current_state(state);
5678 spin_unlock_irq(&x->wait.lock);
5679 timeout = schedule_timeout(timeout);
5680 spin_lock_irq(&x->wait.lock);
5681 } while (!x->done && timeout);
5682 __remove_wait_queue(&x->wait, &wait);
5687 return timeout ?: 1;
5691 wait_for_common(struct completion *x, long timeout, int state)
5695 spin_lock_irq(&x->wait.lock);
5696 timeout = do_wait_for_common(x, timeout, state);
5697 spin_unlock_irq(&x->wait.lock);
5702 * wait_for_completion: - waits for completion of a task
5703 * @x: holds the state of this particular completion
5705 * This waits to be signaled for completion of a specific task. It is NOT
5706 * interruptible and there is no timeout.
5708 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5709 * and interrupt capability. Also see complete().
5711 void __sched wait_for_completion(struct completion *x)
5713 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5715 EXPORT_SYMBOL(wait_for_completion);
5718 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5719 * @x: holds the state of this particular completion
5720 * @timeout: timeout value in jiffies
5722 * This waits for either a completion of a specific task to be signaled or for a
5723 * specified timeout to expire. The timeout is in jiffies. It is not
5726 unsigned long __sched
5727 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5729 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5731 EXPORT_SYMBOL(wait_for_completion_timeout);
5734 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5735 * @x: holds the state of this particular completion
5737 * This waits for completion of a specific task to be signaled. It is
5740 int __sched wait_for_completion_interruptible(struct completion *x)
5742 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5743 if (t == -ERESTARTSYS)
5747 EXPORT_SYMBOL(wait_for_completion_interruptible);
5750 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5751 * @x: holds the state of this particular completion
5752 * @timeout: timeout value in jiffies
5754 * This waits for either a completion of a specific task to be signaled or for a
5755 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5757 unsigned long __sched
5758 wait_for_completion_interruptible_timeout(struct completion *x,
5759 unsigned long timeout)
5761 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5763 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5766 * wait_for_completion_killable: - waits for completion of a task (killable)
5767 * @x: holds the state of this particular completion
5769 * This waits to be signaled for completion of a specific task. It can be
5770 * interrupted by a kill signal.
5772 int __sched wait_for_completion_killable(struct completion *x)
5774 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5775 if (t == -ERESTARTSYS)
5779 EXPORT_SYMBOL(wait_for_completion_killable);
5782 * try_wait_for_completion - try to decrement a completion without blocking
5783 * @x: completion structure
5785 * Returns: 0 if a decrement cannot be done without blocking
5786 * 1 if a decrement succeeded.
5788 * If a completion is being used as a counting completion,
5789 * attempt to decrement the counter without blocking. This
5790 * enables us to avoid waiting if the resource the completion
5791 * is protecting is not available.
5793 bool try_wait_for_completion(struct completion *x)
5797 spin_lock_irq(&x->wait.lock);
5802 spin_unlock_irq(&x->wait.lock);
5805 EXPORT_SYMBOL(try_wait_for_completion);
5808 * completion_done - Test to see if a completion has any waiters
5809 * @x: completion structure
5811 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5812 * 1 if there are no waiters.
5815 bool completion_done(struct completion *x)
5819 spin_lock_irq(&x->wait.lock);
5822 spin_unlock_irq(&x->wait.lock);
5825 EXPORT_SYMBOL(completion_done);
5828 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5830 unsigned long flags;
5833 init_waitqueue_entry(&wait, current);
5835 __set_current_state(state);
5837 spin_lock_irqsave(&q->lock, flags);
5838 __add_wait_queue(q, &wait);
5839 spin_unlock(&q->lock);
5840 timeout = schedule_timeout(timeout);
5841 spin_lock_irq(&q->lock);
5842 __remove_wait_queue(q, &wait);
5843 spin_unlock_irqrestore(&q->lock, flags);
5848 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5850 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5852 EXPORT_SYMBOL(interruptible_sleep_on);
5855 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5857 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5859 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5861 void __sched sleep_on(wait_queue_head_t *q)
5863 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5865 EXPORT_SYMBOL(sleep_on);
5867 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5869 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5871 EXPORT_SYMBOL(sleep_on_timeout);
5873 #ifdef CONFIG_RT_MUTEXES
5876 * rt_mutex_setprio - set the current priority of a task
5878 * @prio: prio value (kernel-internal form)
5880 * This function changes the 'effective' priority of a task. It does
5881 * not touch ->normal_prio like __setscheduler().
5883 * Used by the rt_mutex code to implement priority inheritance logic.
5885 void rt_mutex_setprio(struct task_struct *p, int prio)
5887 unsigned long flags;
5888 int oldprio, on_rq, running;
5890 const struct sched_class *prev_class = p->sched_class;
5892 BUG_ON(prio < 0 || prio > MAX_PRIO);
5894 rq = task_rq_lock(p, &flags);
5895 update_rq_clock(rq);
5898 on_rq = p->se.on_rq;
5899 running = task_current(rq, p);
5901 dequeue_task(rq, p, 0);
5903 p->sched_class->put_prev_task(rq, p);
5906 p->sched_class = &rt_sched_class;
5908 p->sched_class = &fair_sched_class;
5913 p->sched_class->set_curr_task(rq);
5915 enqueue_task(rq, p, 0);
5917 check_class_changed(rq, p, prev_class, oldprio, running);
5919 task_rq_unlock(rq, &flags);
5924 void set_user_nice(struct task_struct *p, long nice)
5926 int old_prio, delta, on_rq;
5927 unsigned long flags;
5930 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5933 * We have to be careful, if called from sys_setpriority(),
5934 * the task might be in the middle of scheduling on another CPU.
5936 rq = task_rq_lock(p, &flags);
5937 update_rq_clock(rq);
5939 * The RT priorities are set via sched_setscheduler(), but we still
5940 * allow the 'normal' nice value to be set - but as expected
5941 * it wont have any effect on scheduling until the task is
5942 * SCHED_FIFO/SCHED_RR:
5944 if (task_has_rt_policy(p)) {
5945 p->static_prio = NICE_TO_PRIO(nice);
5948 on_rq = p->se.on_rq;
5950 dequeue_task(rq, p, 0);
5952 p->static_prio = NICE_TO_PRIO(nice);
5955 p->prio = effective_prio(p);
5956 delta = p->prio - old_prio;
5959 enqueue_task(rq, p, 0);
5961 * If the task increased its priority or is running and
5962 * lowered its priority, then reschedule its CPU:
5964 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5965 resched_task(rq->curr);
5968 task_rq_unlock(rq, &flags);
5970 EXPORT_SYMBOL(set_user_nice);
5973 * can_nice - check if a task can reduce its nice value
5977 int can_nice(const struct task_struct *p, const int nice)
5979 /* convert nice value [19,-20] to rlimit style value [1,40] */
5980 int nice_rlim = 20 - nice;
5982 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5983 capable(CAP_SYS_NICE));
5986 #ifdef __ARCH_WANT_SYS_NICE
5989 * sys_nice - change the priority of the current process.
5990 * @increment: priority increment
5992 * sys_setpriority is a more generic, but much slower function that
5993 * does similar things.
5995 SYSCALL_DEFINE1(nice, int, increment)
6000 * Setpriority might change our priority at the same moment.
6001 * We don't have to worry. Conceptually one call occurs first
6002 * and we have a single winner.
6004 if (increment < -40)
6009 nice = TASK_NICE(current) + increment;
6015 if (increment < 0 && !can_nice(current, nice))
6018 retval = security_task_setnice(current, nice);
6022 set_user_nice(current, nice);
6029 * task_prio - return the priority value of a given task.
6030 * @p: the task in question.
6032 * This is the priority value as seen by users in /proc.
6033 * RT tasks are offset by -200. Normal tasks are centered
6034 * around 0, value goes from -16 to +15.
6036 int task_prio(const struct task_struct *p)
6038 return p->prio - MAX_RT_PRIO;
6042 * task_nice - return the nice value of a given task.
6043 * @p: the task in question.
6045 int task_nice(const struct task_struct *p)
6047 return TASK_NICE(p);
6049 EXPORT_SYMBOL(task_nice);
6052 * idle_cpu - is a given cpu idle currently?
6053 * @cpu: the processor in question.
6055 int idle_cpu(int cpu)
6057 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6061 * idle_task - return the idle task for a given cpu.
6062 * @cpu: the processor in question.
6064 struct task_struct *idle_task(int cpu)
6066 return cpu_rq(cpu)->idle;
6070 * find_process_by_pid - find a process with a matching PID value.
6071 * @pid: the pid in question.
6073 static struct task_struct *find_process_by_pid(pid_t pid)
6075 return pid ? find_task_by_vpid(pid) : current;
6078 /* Actually do priority change: must hold rq lock. */
6080 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6082 BUG_ON(p->se.on_rq);
6085 switch (p->policy) {
6089 p->sched_class = &fair_sched_class;
6093 p->sched_class = &rt_sched_class;
6097 p->rt_priority = prio;
6098 p->normal_prio = normal_prio(p);
6099 /* we are holding p->pi_lock already */
6100 p->prio = rt_mutex_getprio(p);
6105 * check the target process has a UID that matches the current process's
6107 static bool check_same_owner(struct task_struct *p)
6109 const struct cred *cred = current_cred(), *pcred;
6113 pcred = __task_cred(p);
6114 match = (cred->euid == pcred->euid ||
6115 cred->euid == pcred->uid);
6120 static int __sched_setscheduler(struct task_struct *p, int policy,
6121 struct sched_param *param, bool user)
6123 int retval, oldprio, oldpolicy = -1, on_rq, running;
6124 unsigned long flags;
6125 const struct sched_class *prev_class = p->sched_class;
6129 /* may grab non-irq protected spin_locks */
6130 BUG_ON(in_interrupt());
6132 /* double check policy once rq lock held */
6134 reset_on_fork = p->sched_reset_on_fork;
6135 policy = oldpolicy = p->policy;
6137 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6138 policy &= ~SCHED_RESET_ON_FORK;
6140 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6141 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6142 policy != SCHED_IDLE)
6147 * Valid priorities for SCHED_FIFO and SCHED_RR are
6148 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6149 * SCHED_BATCH and SCHED_IDLE is 0.
6151 if (param->sched_priority < 0 ||
6152 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6153 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6155 if (rt_policy(policy) != (param->sched_priority != 0))
6159 * Allow unprivileged RT tasks to decrease priority:
6161 if (user && !capable(CAP_SYS_NICE)) {
6162 if (rt_policy(policy)) {
6163 unsigned long rlim_rtprio;
6165 if (!lock_task_sighand(p, &flags))
6167 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6168 unlock_task_sighand(p, &flags);
6170 /* can't set/change the rt policy */
6171 if (policy != p->policy && !rlim_rtprio)
6174 /* can't increase priority */
6175 if (param->sched_priority > p->rt_priority &&
6176 param->sched_priority > rlim_rtprio)
6180 * Like positive nice levels, dont allow tasks to
6181 * move out of SCHED_IDLE either:
6183 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6186 /* can't change other user's priorities */
6187 if (!check_same_owner(p))
6190 /* Normal users shall not reset the sched_reset_on_fork flag */
6191 if (p->sched_reset_on_fork && !reset_on_fork)
6196 #ifdef CONFIG_RT_GROUP_SCHED
6198 * Do not allow realtime tasks into groups that have no runtime
6201 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6202 task_group(p)->rt_bandwidth.rt_runtime == 0)
6206 retval = security_task_setscheduler(p, policy, param);
6212 * make sure no PI-waiters arrive (or leave) while we are
6213 * changing the priority of the task:
6215 spin_lock_irqsave(&p->pi_lock, flags);
6217 * To be able to change p->policy safely, the apropriate
6218 * runqueue lock must be held.
6220 rq = __task_rq_lock(p);
6221 /* recheck policy now with rq lock held */
6222 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6223 policy = oldpolicy = -1;
6224 __task_rq_unlock(rq);
6225 spin_unlock_irqrestore(&p->pi_lock, flags);
6228 update_rq_clock(rq);
6229 on_rq = p->se.on_rq;
6230 running = task_current(rq, p);
6232 deactivate_task(rq, p, 0);
6234 p->sched_class->put_prev_task(rq, p);
6236 p->sched_reset_on_fork = reset_on_fork;
6239 __setscheduler(rq, p, policy, param->sched_priority);
6242 p->sched_class->set_curr_task(rq);
6244 activate_task(rq, p, 0);
6246 check_class_changed(rq, p, prev_class, oldprio, running);
6248 __task_rq_unlock(rq);
6249 spin_unlock_irqrestore(&p->pi_lock, flags);
6251 rt_mutex_adjust_pi(p);
6257 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6258 * @p: the task in question.
6259 * @policy: new policy.
6260 * @param: structure containing the new RT priority.
6262 * NOTE that the task may be already dead.
6264 int sched_setscheduler(struct task_struct *p, int policy,
6265 struct sched_param *param)
6267 return __sched_setscheduler(p, policy, param, true);
6269 EXPORT_SYMBOL_GPL(sched_setscheduler);
6272 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6273 * @p: the task in question.
6274 * @policy: new policy.
6275 * @param: structure containing the new RT priority.
6277 * Just like sched_setscheduler, only don't bother checking if the
6278 * current context has permission. For example, this is needed in
6279 * stop_machine(): we create temporary high priority worker threads,
6280 * but our caller might not have that capability.
6282 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6283 struct sched_param *param)
6285 return __sched_setscheduler(p, policy, param, false);
6289 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6291 struct sched_param lparam;
6292 struct task_struct *p;
6295 if (!param || pid < 0)
6297 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6302 p = find_process_by_pid(pid);
6304 retval = sched_setscheduler(p, policy, &lparam);
6311 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6312 * @pid: the pid in question.
6313 * @policy: new policy.
6314 * @param: structure containing the new RT priority.
6316 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6317 struct sched_param __user *, param)
6319 /* negative values for policy are not valid */
6323 return do_sched_setscheduler(pid, policy, param);
6327 * sys_sched_setparam - set/change the RT priority of a thread
6328 * @pid: the pid in question.
6329 * @param: structure containing the new RT priority.
6331 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6333 return do_sched_setscheduler(pid, -1, param);
6337 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6338 * @pid: the pid in question.
6340 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6342 struct task_struct *p;
6349 read_lock(&tasklist_lock);
6350 p = find_process_by_pid(pid);
6352 retval = security_task_getscheduler(p);
6355 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6357 read_unlock(&tasklist_lock);
6362 * sys_sched_getparam - get the RT priority of a thread
6363 * @pid: the pid in question.
6364 * @param: structure containing the RT priority.
6366 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6368 struct sched_param lp;
6369 struct task_struct *p;
6372 if (!param || pid < 0)
6375 read_lock(&tasklist_lock);
6376 p = find_process_by_pid(pid);
6381 retval = security_task_getscheduler(p);
6385 lp.sched_priority = p->rt_priority;
6386 read_unlock(&tasklist_lock);
6389 * This one might sleep, we cannot do it with a spinlock held ...
6391 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6396 read_unlock(&tasklist_lock);
6400 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6402 cpumask_var_t cpus_allowed, new_mask;
6403 struct task_struct *p;
6407 read_lock(&tasklist_lock);
6409 p = find_process_by_pid(pid);
6411 read_unlock(&tasklist_lock);
6417 * It is not safe to call set_cpus_allowed with the
6418 * tasklist_lock held. We will bump the task_struct's
6419 * usage count and then drop tasklist_lock.
6422 read_unlock(&tasklist_lock);
6424 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6428 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6430 goto out_free_cpus_allowed;
6433 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6436 retval = security_task_setscheduler(p, 0, NULL);
6440 cpuset_cpus_allowed(p, cpus_allowed);
6441 cpumask_and(new_mask, in_mask, cpus_allowed);
6443 retval = set_cpus_allowed_ptr(p, new_mask);
6446 cpuset_cpus_allowed(p, cpus_allowed);
6447 if (!cpumask_subset(new_mask, cpus_allowed)) {
6449 * We must have raced with a concurrent cpuset
6450 * update. Just reset the cpus_allowed to the
6451 * cpuset's cpus_allowed
6453 cpumask_copy(new_mask, cpus_allowed);
6458 free_cpumask_var(new_mask);
6459 out_free_cpus_allowed:
6460 free_cpumask_var(cpus_allowed);
6467 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6468 struct cpumask *new_mask)
6470 if (len < cpumask_size())
6471 cpumask_clear(new_mask);
6472 else if (len > cpumask_size())
6473 len = cpumask_size();
6475 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6479 * sys_sched_setaffinity - set the cpu affinity of a process
6480 * @pid: pid of the process
6481 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6482 * @user_mask_ptr: user-space pointer to the new cpu mask
6484 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6485 unsigned long __user *, user_mask_ptr)
6487 cpumask_var_t new_mask;
6490 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6493 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6495 retval = sched_setaffinity(pid, new_mask);
6496 free_cpumask_var(new_mask);
6500 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6502 struct task_struct *p;
6506 read_lock(&tasklist_lock);
6509 p = find_process_by_pid(pid);
6513 retval = security_task_getscheduler(p);
6517 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6520 read_unlock(&tasklist_lock);
6527 * sys_sched_getaffinity - get the cpu affinity of a process
6528 * @pid: pid of the process
6529 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6530 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6532 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6533 unsigned long __user *, user_mask_ptr)
6538 if (len < cpumask_size())
6541 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6544 ret = sched_getaffinity(pid, mask);
6546 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6549 ret = cpumask_size();
6551 free_cpumask_var(mask);
6557 * sys_sched_yield - yield the current processor to other threads.
6559 * This function yields the current CPU to other tasks. If there are no
6560 * other threads running on this CPU then this function will return.
6562 SYSCALL_DEFINE0(sched_yield)
6564 struct rq *rq = this_rq_lock();
6566 schedstat_inc(rq, yld_count);
6567 current->sched_class->yield_task(rq);
6570 * Since we are going to call schedule() anyway, there's
6571 * no need to preempt or enable interrupts:
6573 __release(rq->lock);
6574 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6575 _raw_spin_unlock(&rq->lock);
6576 preempt_enable_no_resched();
6583 static void __cond_resched(void)
6585 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6586 __might_sleep(__FILE__, __LINE__);
6589 * The BKS might be reacquired before we have dropped
6590 * PREEMPT_ACTIVE, which could trigger a second
6591 * cond_resched() call.
6594 add_preempt_count(PREEMPT_ACTIVE);
6596 sub_preempt_count(PREEMPT_ACTIVE);
6597 } while (need_resched());
6600 int __sched _cond_resched(void)
6602 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
6603 system_state == SYSTEM_RUNNING) {
6609 EXPORT_SYMBOL(_cond_resched);
6612 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6613 * call schedule, and on return reacquire the lock.
6615 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6616 * operations here to prevent schedule() from being called twice (once via
6617 * spin_unlock(), once by hand).
6619 int cond_resched_lock(spinlock_t *lock)
6621 int resched = need_resched() && system_state == SYSTEM_RUNNING;
6624 if (spin_needbreak(lock) || resched) {
6626 if (resched && need_resched())
6635 EXPORT_SYMBOL(cond_resched_lock);
6637 int __sched cond_resched_softirq(void)
6639 BUG_ON(!in_softirq());
6641 if (need_resched() && system_state == SYSTEM_RUNNING) {
6649 EXPORT_SYMBOL(cond_resched_softirq);
6652 * yield - yield the current processor to other threads.
6654 * This is a shortcut for kernel-space yielding - it marks the
6655 * thread runnable and calls sys_sched_yield().
6657 void __sched yield(void)
6659 set_current_state(TASK_RUNNING);
6662 EXPORT_SYMBOL(yield);
6665 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6666 * that process accounting knows that this is a task in IO wait state.
6668 * But don't do that if it is a deliberate, throttling IO wait (this task
6669 * has set its backing_dev_info: the queue against which it should throttle)
6671 void __sched io_schedule(void)
6673 struct rq *rq = raw_rq();
6675 delayacct_blkio_start();
6676 atomic_inc(&rq->nr_iowait);
6678 atomic_dec(&rq->nr_iowait);
6679 delayacct_blkio_end();
6681 EXPORT_SYMBOL(io_schedule);
6683 long __sched io_schedule_timeout(long timeout)
6685 struct rq *rq = raw_rq();
6688 delayacct_blkio_start();
6689 atomic_inc(&rq->nr_iowait);
6690 ret = schedule_timeout(timeout);
6691 atomic_dec(&rq->nr_iowait);
6692 delayacct_blkio_end();
6697 * sys_sched_get_priority_max - return maximum RT priority.
6698 * @policy: scheduling class.
6700 * this syscall returns the maximum rt_priority that can be used
6701 * by a given scheduling class.
6703 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6710 ret = MAX_USER_RT_PRIO-1;
6722 * sys_sched_get_priority_min - return minimum RT priority.
6723 * @policy: scheduling class.
6725 * this syscall returns the minimum rt_priority that can be used
6726 * by a given scheduling class.
6728 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6746 * sys_sched_rr_get_interval - return the default timeslice of a process.
6747 * @pid: pid of the process.
6748 * @interval: userspace pointer to the timeslice value.
6750 * this syscall writes the default timeslice value of a given process
6751 * into the user-space timespec buffer. A value of '0' means infinity.
6753 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6754 struct timespec __user *, interval)
6756 struct task_struct *p;
6757 unsigned int time_slice;
6765 read_lock(&tasklist_lock);
6766 p = find_process_by_pid(pid);
6770 retval = security_task_getscheduler(p);
6775 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6776 * tasks that are on an otherwise idle runqueue:
6779 if (p->policy == SCHED_RR) {
6780 time_slice = DEF_TIMESLICE;
6781 } else if (p->policy != SCHED_FIFO) {
6782 struct sched_entity *se = &p->se;
6783 unsigned long flags;
6786 rq = task_rq_lock(p, &flags);
6787 if (rq->cfs.load.weight)
6788 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6789 task_rq_unlock(rq, &flags);
6791 read_unlock(&tasklist_lock);
6792 jiffies_to_timespec(time_slice, &t);
6793 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6797 read_unlock(&tasklist_lock);
6801 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6803 void sched_show_task(struct task_struct *p)
6805 unsigned long free = 0;
6808 state = p->state ? __ffs(p->state) + 1 : 0;
6809 printk(KERN_INFO "%-13.13s %c", p->comm,
6810 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6811 #if BITS_PER_LONG == 32
6812 if (state == TASK_RUNNING)
6813 printk(KERN_CONT " running ");
6815 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6817 if (state == TASK_RUNNING)
6818 printk(KERN_CONT " running task ");
6820 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6822 #ifdef CONFIG_DEBUG_STACK_USAGE
6823 free = stack_not_used(p);
6825 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6826 task_pid_nr(p), task_pid_nr(p->real_parent),
6827 (unsigned long)task_thread_info(p)->flags);
6829 show_stack(p, NULL);
6832 void show_state_filter(unsigned long state_filter)
6834 struct task_struct *g, *p;
6836 #if BITS_PER_LONG == 32
6838 " task PC stack pid father\n");
6841 " task PC stack pid father\n");
6843 read_lock(&tasklist_lock);
6844 do_each_thread(g, p) {
6846 * reset the NMI-timeout, listing all files on a slow
6847 * console might take alot of time:
6849 touch_nmi_watchdog();
6850 if (!state_filter || (p->state & state_filter))
6852 } while_each_thread(g, p);
6854 touch_all_softlockup_watchdogs();
6856 #ifdef CONFIG_SCHED_DEBUG
6857 sysrq_sched_debug_show();
6859 read_unlock(&tasklist_lock);
6861 * Only show locks if all tasks are dumped:
6863 if (state_filter == -1)
6864 debug_show_all_locks();
6867 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6869 idle->sched_class = &idle_sched_class;
6873 * init_idle - set up an idle thread for a given CPU
6874 * @idle: task in question
6875 * @cpu: cpu the idle task belongs to
6877 * NOTE: this function does not set the idle thread's NEED_RESCHED
6878 * flag, to make booting more robust.
6880 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6882 struct rq *rq = cpu_rq(cpu);
6883 unsigned long flags;
6885 spin_lock_irqsave(&rq->lock, flags);
6888 idle->se.exec_start = sched_clock();
6890 idle->prio = idle->normal_prio = MAX_PRIO;
6891 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6892 __set_task_cpu(idle, cpu);
6894 rq->curr = rq->idle = idle;
6895 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6898 spin_unlock_irqrestore(&rq->lock, flags);
6900 /* Set the preempt count _outside_ the spinlocks! */
6901 #if defined(CONFIG_PREEMPT)
6902 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6904 task_thread_info(idle)->preempt_count = 0;
6907 * The idle tasks have their own, simple scheduling class:
6909 idle->sched_class = &idle_sched_class;
6910 ftrace_graph_init_task(idle);
6914 * In a system that switches off the HZ timer nohz_cpu_mask
6915 * indicates which cpus entered this state. This is used
6916 * in the rcu update to wait only for active cpus. For system
6917 * which do not switch off the HZ timer nohz_cpu_mask should
6918 * always be CPU_BITS_NONE.
6920 cpumask_var_t nohz_cpu_mask;
6923 * Increase the granularity value when there are more CPUs,
6924 * because with more CPUs the 'effective latency' as visible
6925 * to users decreases. But the relationship is not linear,
6926 * so pick a second-best guess by going with the log2 of the
6929 * This idea comes from the SD scheduler of Con Kolivas:
6931 static inline void sched_init_granularity(void)
6933 unsigned int factor = 1 + ilog2(num_online_cpus());
6934 const unsigned long limit = 200000000;
6936 sysctl_sched_min_granularity *= factor;
6937 if (sysctl_sched_min_granularity > limit)
6938 sysctl_sched_min_granularity = limit;
6940 sysctl_sched_latency *= factor;
6941 if (sysctl_sched_latency > limit)
6942 sysctl_sched_latency = limit;
6944 sysctl_sched_wakeup_granularity *= factor;
6946 sysctl_sched_shares_ratelimit *= factor;
6951 * This is how migration works:
6953 * 1) we queue a struct migration_req structure in the source CPU's
6954 * runqueue and wake up that CPU's migration thread.
6955 * 2) we down() the locked semaphore => thread blocks.
6956 * 3) migration thread wakes up (implicitly it forces the migrated
6957 * thread off the CPU)
6958 * 4) it gets the migration request and checks whether the migrated
6959 * task is still in the wrong runqueue.
6960 * 5) if it's in the wrong runqueue then the migration thread removes
6961 * it and puts it into the right queue.
6962 * 6) migration thread up()s the semaphore.
6963 * 7) we wake up and the migration is done.
6967 * Change a given task's CPU affinity. Migrate the thread to a
6968 * proper CPU and schedule it away if the CPU it's executing on
6969 * is removed from the allowed bitmask.
6971 * NOTE: the caller must have a valid reference to the task, the
6972 * task must not exit() & deallocate itself prematurely. The
6973 * call is not atomic; no spinlocks may be held.
6975 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6977 struct migration_req req;
6978 unsigned long flags;
6982 rq = task_rq_lock(p, &flags);
6983 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6988 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6989 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6994 if (p->sched_class->set_cpus_allowed)
6995 p->sched_class->set_cpus_allowed(p, new_mask);
6997 cpumask_copy(&p->cpus_allowed, new_mask);
6998 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7001 /* Can the task run on the task's current CPU? If so, we're done */
7002 if (cpumask_test_cpu(task_cpu(p), new_mask))
7005 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7006 /* Need help from migration thread: drop lock and wait. */
7007 task_rq_unlock(rq, &flags);
7008 wake_up_process(rq->migration_thread);
7009 wait_for_completion(&req.done);
7010 tlb_migrate_finish(p->mm);
7014 task_rq_unlock(rq, &flags);
7018 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7021 * Move (not current) task off this cpu, onto dest cpu. We're doing
7022 * this because either it can't run here any more (set_cpus_allowed()
7023 * away from this CPU, or CPU going down), or because we're
7024 * attempting to rebalance this task on exec (sched_exec).
7026 * So we race with normal scheduler movements, but that's OK, as long
7027 * as the task is no longer on this CPU.
7029 * Returns non-zero if task was successfully migrated.
7031 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7033 struct rq *rq_dest, *rq_src;
7036 if (unlikely(!cpu_active(dest_cpu)))
7039 rq_src = cpu_rq(src_cpu);
7040 rq_dest = cpu_rq(dest_cpu);
7042 double_rq_lock(rq_src, rq_dest);
7043 /* Already moved. */
7044 if (task_cpu(p) != src_cpu)
7046 /* Affinity changed (again). */
7047 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7050 on_rq = p->se.on_rq;
7052 deactivate_task(rq_src, p, 0);
7054 set_task_cpu(p, dest_cpu);
7056 activate_task(rq_dest, p, 0);
7057 check_preempt_curr(rq_dest, p, 0);
7062 double_rq_unlock(rq_src, rq_dest);
7067 * migration_thread - this is a highprio system thread that performs
7068 * thread migration by bumping thread off CPU then 'pushing' onto
7071 static int migration_thread(void *data)
7073 int cpu = (long)data;
7077 BUG_ON(rq->migration_thread != current);
7079 set_current_state(TASK_INTERRUPTIBLE);
7080 while (!kthread_should_stop()) {
7081 struct migration_req *req;
7082 struct list_head *head;
7084 spin_lock_irq(&rq->lock);
7086 if (cpu_is_offline(cpu)) {
7087 spin_unlock_irq(&rq->lock);
7091 if (rq->active_balance) {
7092 active_load_balance(rq, cpu);
7093 rq->active_balance = 0;
7096 head = &rq->migration_queue;
7098 if (list_empty(head)) {
7099 spin_unlock_irq(&rq->lock);
7101 set_current_state(TASK_INTERRUPTIBLE);
7104 req = list_entry(head->next, struct migration_req, list);
7105 list_del_init(head->next);
7107 spin_unlock(&rq->lock);
7108 __migrate_task(req->task, cpu, req->dest_cpu);
7111 complete(&req->done);
7113 __set_current_state(TASK_RUNNING);
7118 #ifdef CONFIG_HOTPLUG_CPU
7120 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7124 local_irq_disable();
7125 ret = __migrate_task(p, src_cpu, dest_cpu);
7131 * Figure out where task on dead CPU should go, use force if necessary.
7133 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7136 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7139 /* Look for allowed, online CPU in same node. */
7140 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7141 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7144 /* Any allowed, online CPU? */
7145 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7146 if (dest_cpu < nr_cpu_ids)
7149 /* No more Mr. Nice Guy. */
7150 if (dest_cpu >= nr_cpu_ids) {
7151 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7152 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7155 * Don't tell them about moving exiting tasks or
7156 * kernel threads (both mm NULL), since they never
7159 if (p->mm && printk_ratelimit()) {
7160 printk(KERN_INFO "process %d (%s) no "
7161 "longer affine to cpu%d\n",
7162 task_pid_nr(p), p->comm, dead_cpu);
7167 /* It can have affinity changed while we were choosing. */
7168 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7173 * While a dead CPU has no uninterruptible tasks queued at this point,
7174 * it might still have a nonzero ->nr_uninterruptible counter, because
7175 * for performance reasons the counter is not stricly tracking tasks to
7176 * their home CPUs. So we just add the counter to another CPU's counter,
7177 * to keep the global sum constant after CPU-down:
7179 static void migrate_nr_uninterruptible(struct rq *rq_src)
7181 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7182 unsigned long flags;
7184 local_irq_save(flags);
7185 double_rq_lock(rq_src, rq_dest);
7186 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7187 rq_src->nr_uninterruptible = 0;
7188 double_rq_unlock(rq_src, rq_dest);
7189 local_irq_restore(flags);
7192 /* Run through task list and migrate tasks from the dead cpu. */
7193 static void migrate_live_tasks(int src_cpu)
7195 struct task_struct *p, *t;
7197 read_lock(&tasklist_lock);
7199 do_each_thread(t, p) {
7203 if (task_cpu(p) == src_cpu)
7204 move_task_off_dead_cpu(src_cpu, p);
7205 } while_each_thread(t, p);
7207 read_unlock(&tasklist_lock);
7211 * Schedules idle task to be the next runnable task on current CPU.
7212 * It does so by boosting its priority to highest possible.
7213 * Used by CPU offline code.
7215 void sched_idle_next(void)
7217 int this_cpu = smp_processor_id();
7218 struct rq *rq = cpu_rq(this_cpu);
7219 struct task_struct *p = rq->idle;
7220 unsigned long flags;
7222 /* cpu has to be offline */
7223 BUG_ON(cpu_online(this_cpu));
7226 * Strictly not necessary since rest of the CPUs are stopped by now
7227 * and interrupts disabled on the current cpu.
7229 spin_lock_irqsave(&rq->lock, flags);
7231 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7233 update_rq_clock(rq);
7234 activate_task(rq, p, 0);
7236 spin_unlock_irqrestore(&rq->lock, flags);
7240 * Ensures that the idle task is using init_mm right before its cpu goes
7243 void idle_task_exit(void)
7245 struct mm_struct *mm = current->active_mm;
7247 BUG_ON(cpu_online(smp_processor_id()));
7250 switch_mm(mm, &init_mm, current);
7254 /* called under rq->lock with disabled interrupts */
7255 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7257 struct rq *rq = cpu_rq(dead_cpu);
7259 /* Must be exiting, otherwise would be on tasklist. */
7260 BUG_ON(!p->exit_state);
7262 /* Cannot have done final schedule yet: would have vanished. */
7263 BUG_ON(p->state == TASK_DEAD);
7268 * Drop lock around migration; if someone else moves it,
7269 * that's OK. No task can be added to this CPU, so iteration is
7272 spin_unlock_irq(&rq->lock);
7273 move_task_off_dead_cpu(dead_cpu, p);
7274 spin_lock_irq(&rq->lock);
7279 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7280 static void migrate_dead_tasks(unsigned int dead_cpu)
7282 struct rq *rq = cpu_rq(dead_cpu);
7283 struct task_struct *next;
7286 if (!rq->nr_running)
7288 update_rq_clock(rq);
7289 next = pick_next_task(rq);
7292 next->sched_class->put_prev_task(rq, next);
7293 migrate_dead(dead_cpu, next);
7299 * remove the tasks which were accounted by rq from calc_load_tasks.
7301 static void calc_global_load_remove(struct rq *rq)
7303 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7305 #endif /* CONFIG_HOTPLUG_CPU */
7307 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7309 static struct ctl_table sd_ctl_dir[] = {
7311 .procname = "sched_domain",
7317 static struct ctl_table sd_ctl_root[] = {
7319 .ctl_name = CTL_KERN,
7320 .procname = "kernel",
7322 .child = sd_ctl_dir,
7327 static struct ctl_table *sd_alloc_ctl_entry(int n)
7329 struct ctl_table *entry =
7330 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7335 static void sd_free_ctl_entry(struct ctl_table **tablep)
7337 struct ctl_table *entry;
7340 * In the intermediate directories, both the child directory and
7341 * procname are dynamically allocated and could fail but the mode
7342 * will always be set. In the lowest directory the names are
7343 * static strings and all have proc handlers.
7345 for (entry = *tablep; entry->mode; entry++) {
7347 sd_free_ctl_entry(&entry->child);
7348 if (entry->proc_handler == NULL)
7349 kfree(entry->procname);
7357 set_table_entry(struct ctl_table *entry,
7358 const char *procname, void *data, int maxlen,
7359 mode_t mode, proc_handler *proc_handler)
7361 entry->procname = procname;
7363 entry->maxlen = maxlen;
7365 entry->proc_handler = proc_handler;
7368 static struct ctl_table *
7369 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7371 struct ctl_table *table = sd_alloc_ctl_entry(13);
7376 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7377 sizeof(long), 0644, proc_doulongvec_minmax);
7378 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7379 sizeof(long), 0644, proc_doulongvec_minmax);
7380 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7381 sizeof(int), 0644, proc_dointvec_minmax);
7382 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7383 sizeof(int), 0644, proc_dointvec_minmax);
7384 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7385 sizeof(int), 0644, proc_dointvec_minmax);
7386 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7387 sizeof(int), 0644, proc_dointvec_minmax);
7388 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7389 sizeof(int), 0644, proc_dointvec_minmax);
7390 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7391 sizeof(int), 0644, proc_dointvec_minmax);
7392 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7393 sizeof(int), 0644, proc_dointvec_minmax);
7394 set_table_entry(&table[9], "cache_nice_tries",
7395 &sd->cache_nice_tries,
7396 sizeof(int), 0644, proc_dointvec_minmax);
7397 set_table_entry(&table[10], "flags", &sd->flags,
7398 sizeof(int), 0644, proc_dointvec_minmax);
7399 set_table_entry(&table[11], "name", sd->name,
7400 CORENAME_MAX_SIZE, 0444, proc_dostring);
7401 /* &table[12] is terminator */
7406 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7408 struct ctl_table *entry, *table;
7409 struct sched_domain *sd;
7410 int domain_num = 0, i;
7413 for_each_domain(cpu, sd)
7415 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7420 for_each_domain(cpu, sd) {
7421 snprintf(buf, 32, "domain%d", i);
7422 entry->procname = kstrdup(buf, GFP_KERNEL);
7424 entry->child = sd_alloc_ctl_domain_table(sd);
7431 static struct ctl_table_header *sd_sysctl_header;
7432 static void register_sched_domain_sysctl(void)
7434 int i, cpu_num = num_online_cpus();
7435 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7438 WARN_ON(sd_ctl_dir[0].child);
7439 sd_ctl_dir[0].child = entry;
7444 for_each_online_cpu(i) {
7445 snprintf(buf, 32, "cpu%d", i);
7446 entry->procname = kstrdup(buf, GFP_KERNEL);
7448 entry->child = sd_alloc_ctl_cpu_table(i);
7452 WARN_ON(sd_sysctl_header);
7453 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7456 /* may be called multiple times per register */
7457 static void unregister_sched_domain_sysctl(void)
7459 if (sd_sysctl_header)
7460 unregister_sysctl_table(sd_sysctl_header);
7461 sd_sysctl_header = NULL;
7462 if (sd_ctl_dir[0].child)
7463 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7466 static void register_sched_domain_sysctl(void)
7469 static void unregister_sched_domain_sysctl(void)
7474 static void set_rq_online(struct rq *rq)
7477 const struct sched_class *class;
7479 cpumask_set_cpu(rq->cpu, rq->rd->online);
7482 for_each_class(class) {
7483 if (class->rq_online)
7484 class->rq_online(rq);
7489 static void set_rq_offline(struct rq *rq)
7492 const struct sched_class *class;
7494 for_each_class(class) {
7495 if (class->rq_offline)
7496 class->rq_offline(rq);
7499 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7505 * migration_call - callback that gets triggered when a CPU is added.
7506 * Here we can start up the necessary migration thread for the new CPU.
7508 static int __cpuinit
7509 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7511 struct task_struct *p;
7512 int cpu = (long)hcpu;
7513 unsigned long flags;
7518 case CPU_UP_PREPARE:
7519 case CPU_UP_PREPARE_FROZEN:
7520 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7523 kthread_bind(p, cpu);
7524 /* Must be high prio: stop_machine expects to yield to it. */
7525 rq = task_rq_lock(p, &flags);
7526 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7527 task_rq_unlock(rq, &flags);
7529 cpu_rq(cpu)->migration_thread = p;
7533 case CPU_ONLINE_FROZEN:
7534 /* Strictly unnecessary, as first user will wake it. */
7535 wake_up_process(cpu_rq(cpu)->migration_thread);
7537 /* Update our root-domain */
7539 spin_lock_irqsave(&rq->lock, flags);
7540 rq->calc_load_update = calc_load_update;
7541 rq->calc_load_active = 0;
7543 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7547 spin_unlock_irqrestore(&rq->lock, flags);
7550 #ifdef CONFIG_HOTPLUG_CPU
7551 case CPU_UP_CANCELED:
7552 case CPU_UP_CANCELED_FROZEN:
7553 if (!cpu_rq(cpu)->migration_thread)
7555 /* Unbind it from offline cpu so it can run. Fall thru. */
7556 kthread_bind(cpu_rq(cpu)->migration_thread,
7557 cpumask_any(cpu_online_mask));
7558 kthread_stop(cpu_rq(cpu)->migration_thread);
7559 put_task_struct(cpu_rq(cpu)->migration_thread);
7560 cpu_rq(cpu)->migration_thread = NULL;
7564 case CPU_DEAD_FROZEN:
7565 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7566 migrate_live_tasks(cpu);
7568 kthread_stop(rq->migration_thread);
7569 put_task_struct(rq->migration_thread);
7570 rq->migration_thread = NULL;
7571 /* Idle task back to normal (off runqueue, low prio) */
7572 spin_lock_irq(&rq->lock);
7573 update_rq_clock(rq);
7574 deactivate_task(rq, rq->idle, 0);
7575 rq->idle->static_prio = MAX_PRIO;
7576 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7577 rq->idle->sched_class = &idle_sched_class;
7578 migrate_dead_tasks(cpu);
7579 spin_unlock_irq(&rq->lock);
7581 migrate_nr_uninterruptible(rq);
7582 BUG_ON(rq->nr_running != 0);
7583 calc_global_load_remove(rq);
7585 * No need to migrate the tasks: it was best-effort if
7586 * they didn't take sched_hotcpu_mutex. Just wake up
7589 spin_lock_irq(&rq->lock);
7590 while (!list_empty(&rq->migration_queue)) {
7591 struct migration_req *req;
7593 req = list_entry(rq->migration_queue.next,
7594 struct migration_req, list);
7595 list_del_init(&req->list);
7596 spin_unlock_irq(&rq->lock);
7597 complete(&req->done);
7598 spin_lock_irq(&rq->lock);
7600 spin_unlock_irq(&rq->lock);
7604 case CPU_DYING_FROZEN:
7605 /* Update our root-domain */
7607 spin_lock_irqsave(&rq->lock, flags);
7609 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7612 spin_unlock_irqrestore(&rq->lock, flags);
7620 * Register at high priority so that task migration (migrate_all_tasks)
7621 * happens before everything else. This has to be lower priority than
7622 * the notifier in the perf_counter subsystem, though.
7624 static struct notifier_block __cpuinitdata migration_notifier = {
7625 .notifier_call = migration_call,
7629 static int __init migration_init(void)
7631 void *cpu = (void *)(long)smp_processor_id();
7634 /* Start one for the boot CPU: */
7635 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7636 BUG_ON(err == NOTIFY_BAD);
7637 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7638 register_cpu_notifier(&migration_notifier);
7642 early_initcall(migration_init);
7647 #ifdef CONFIG_SCHED_DEBUG
7649 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7650 struct cpumask *groupmask)
7652 struct sched_group *group = sd->groups;
7655 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7656 cpumask_clear(groupmask);
7658 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7660 if (!(sd->flags & SD_LOAD_BALANCE)) {
7661 printk("does not load-balance\n");
7663 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7668 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7670 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7671 printk(KERN_ERR "ERROR: domain->span does not contain "
7674 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7675 printk(KERN_ERR "ERROR: domain->groups does not contain"
7679 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7683 printk(KERN_ERR "ERROR: group is NULL\n");
7687 if (!group->__cpu_power) {
7688 printk(KERN_CONT "\n");
7689 printk(KERN_ERR "ERROR: domain->cpu_power not "
7694 if (!cpumask_weight(sched_group_cpus(group))) {
7695 printk(KERN_CONT "\n");
7696 printk(KERN_ERR "ERROR: empty group\n");
7700 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7701 printk(KERN_CONT "\n");
7702 printk(KERN_ERR "ERROR: repeated CPUs\n");
7706 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7708 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7710 printk(KERN_CONT " %s", str);
7711 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7712 printk(KERN_CONT " (__cpu_power = %d)",
7713 group->__cpu_power);
7716 group = group->next;
7717 } while (group != sd->groups);
7718 printk(KERN_CONT "\n");
7720 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7721 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7724 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7725 printk(KERN_ERR "ERROR: parent span is not a superset "
7726 "of domain->span\n");
7730 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7732 cpumask_var_t groupmask;
7736 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7740 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7742 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7743 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7748 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7755 free_cpumask_var(groupmask);
7757 #else /* !CONFIG_SCHED_DEBUG */
7758 # define sched_domain_debug(sd, cpu) do { } while (0)
7759 #endif /* CONFIG_SCHED_DEBUG */
7761 static int sd_degenerate(struct sched_domain *sd)
7763 if (cpumask_weight(sched_domain_span(sd)) == 1)
7766 /* Following flags need at least 2 groups */
7767 if (sd->flags & (SD_LOAD_BALANCE |
7768 SD_BALANCE_NEWIDLE |
7772 SD_SHARE_PKG_RESOURCES)) {
7773 if (sd->groups != sd->groups->next)
7777 /* Following flags don't use groups */
7778 if (sd->flags & (SD_WAKE_IDLE |
7787 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7789 unsigned long cflags = sd->flags, pflags = parent->flags;
7791 if (sd_degenerate(parent))
7794 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7797 /* Does parent contain flags not in child? */
7798 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7799 if (cflags & SD_WAKE_AFFINE)
7800 pflags &= ~SD_WAKE_BALANCE;
7801 /* Flags needing groups don't count if only 1 group in parent */
7802 if (parent->groups == parent->groups->next) {
7803 pflags &= ~(SD_LOAD_BALANCE |
7804 SD_BALANCE_NEWIDLE |
7808 SD_SHARE_PKG_RESOURCES);
7809 if (nr_node_ids == 1)
7810 pflags &= ~SD_SERIALIZE;
7812 if (~cflags & pflags)
7818 static void free_rootdomain(struct root_domain *rd)
7820 cpupri_cleanup(&rd->cpupri);
7822 free_cpumask_var(rd->rto_mask);
7823 free_cpumask_var(rd->online);
7824 free_cpumask_var(rd->span);
7828 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7830 struct root_domain *old_rd = NULL;
7831 unsigned long flags;
7833 spin_lock_irqsave(&rq->lock, flags);
7838 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7841 cpumask_clear_cpu(rq->cpu, old_rd->span);
7844 * If we dont want to free the old_rt yet then
7845 * set old_rd to NULL to skip the freeing later
7848 if (!atomic_dec_and_test(&old_rd->refcount))
7852 atomic_inc(&rd->refcount);
7855 cpumask_set_cpu(rq->cpu, rd->span);
7856 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7859 spin_unlock_irqrestore(&rq->lock, flags);
7862 free_rootdomain(old_rd);
7865 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7867 gfp_t gfp = GFP_KERNEL;
7869 memset(rd, 0, sizeof(*rd));
7874 if (!alloc_cpumask_var(&rd->span, gfp))
7876 if (!alloc_cpumask_var(&rd->online, gfp))
7878 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7881 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7886 free_cpumask_var(rd->rto_mask);
7888 free_cpumask_var(rd->online);
7890 free_cpumask_var(rd->span);
7895 static void init_defrootdomain(void)
7897 init_rootdomain(&def_root_domain, true);
7899 atomic_set(&def_root_domain.refcount, 1);
7902 static struct root_domain *alloc_rootdomain(void)
7904 struct root_domain *rd;
7906 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7910 if (init_rootdomain(rd, false) != 0) {
7919 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7920 * hold the hotplug lock.
7923 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7925 struct rq *rq = cpu_rq(cpu);
7926 struct sched_domain *tmp;
7928 /* Remove the sched domains which do not contribute to scheduling. */
7929 for (tmp = sd; tmp; ) {
7930 struct sched_domain *parent = tmp->parent;
7934 if (sd_parent_degenerate(tmp, parent)) {
7935 tmp->parent = parent->parent;
7937 parent->parent->child = tmp;
7942 if (sd && sd_degenerate(sd)) {
7948 sched_domain_debug(sd, cpu);
7950 rq_attach_root(rq, rd);
7951 rcu_assign_pointer(rq->sd, sd);
7954 /* cpus with isolated domains */
7955 static cpumask_var_t cpu_isolated_map;
7957 /* Setup the mask of cpus configured for isolated domains */
7958 static int __init isolated_cpu_setup(char *str)
7960 cpulist_parse(str, cpu_isolated_map);
7964 __setup("isolcpus=", isolated_cpu_setup);
7967 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7968 * to a function which identifies what group(along with sched group) a CPU
7969 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7970 * (due to the fact that we keep track of groups covered with a struct cpumask).
7972 * init_sched_build_groups will build a circular linked list of the groups
7973 * covered by the given span, and will set each group's ->cpumask correctly,
7974 * and ->cpu_power to 0.
7977 init_sched_build_groups(const struct cpumask *span,
7978 const struct cpumask *cpu_map,
7979 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7980 struct sched_group **sg,
7981 struct cpumask *tmpmask),
7982 struct cpumask *covered, struct cpumask *tmpmask)
7984 struct sched_group *first = NULL, *last = NULL;
7987 cpumask_clear(covered);
7989 for_each_cpu(i, span) {
7990 struct sched_group *sg;
7991 int group = group_fn(i, cpu_map, &sg, tmpmask);
7994 if (cpumask_test_cpu(i, covered))
7997 cpumask_clear(sched_group_cpus(sg));
7998 sg->__cpu_power = 0;
8000 for_each_cpu(j, span) {
8001 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8004 cpumask_set_cpu(j, covered);
8005 cpumask_set_cpu(j, sched_group_cpus(sg));
8016 #define SD_NODES_PER_DOMAIN 16
8021 * find_next_best_node - find the next node to include in a sched_domain
8022 * @node: node whose sched_domain we're building
8023 * @used_nodes: nodes already in the sched_domain
8025 * Find the next node to include in a given scheduling domain. Simply
8026 * finds the closest node not already in the @used_nodes map.
8028 * Should use nodemask_t.
8030 static int find_next_best_node(int node, nodemask_t *used_nodes)
8032 int i, n, val, min_val, best_node = 0;
8036 for (i = 0; i < nr_node_ids; i++) {
8037 /* Start at @node */
8038 n = (node + i) % nr_node_ids;
8040 if (!nr_cpus_node(n))
8043 /* Skip already used nodes */
8044 if (node_isset(n, *used_nodes))
8047 /* Simple min distance search */
8048 val = node_distance(node, n);
8050 if (val < min_val) {
8056 node_set(best_node, *used_nodes);
8061 * sched_domain_node_span - get a cpumask for a node's sched_domain
8062 * @node: node whose cpumask we're constructing
8063 * @span: resulting cpumask
8065 * Given a node, construct a good cpumask for its sched_domain to span. It
8066 * should be one that prevents unnecessary balancing, but also spreads tasks
8069 static void sched_domain_node_span(int node, struct cpumask *span)
8071 nodemask_t used_nodes;
8074 cpumask_clear(span);
8075 nodes_clear(used_nodes);
8077 cpumask_or(span, span, cpumask_of_node(node));
8078 node_set(node, used_nodes);
8080 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8081 int next_node = find_next_best_node(node, &used_nodes);
8083 cpumask_or(span, span, cpumask_of_node(next_node));
8086 #endif /* CONFIG_NUMA */
8088 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8091 * The cpus mask in sched_group and sched_domain hangs off the end.
8093 * ( See the the comments in include/linux/sched.h:struct sched_group
8094 * and struct sched_domain. )
8096 struct static_sched_group {
8097 struct sched_group sg;
8098 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8101 struct static_sched_domain {
8102 struct sched_domain sd;
8103 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8107 * SMT sched-domains:
8109 #ifdef CONFIG_SCHED_SMT
8110 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8111 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8114 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8115 struct sched_group **sg, struct cpumask *unused)
8118 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8121 #endif /* CONFIG_SCHED_SMT */
8124 * multi-core sched-domains:
8126 #ifdef CONFIG_SCHED_MC
8127 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8128 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8129 #endif /* CONFIG_SCHED_MC */
8131 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8133 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8134 struct sched_group **sg, struct cpumask *mask)
8138 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8139 group = cpumask_first(mask);
8141 *sg = &per_cpu(sched_group_core, group).sg;
8144 #elif defined(CONFIG_SCHED_MC)
8146 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8147 struct sched_group **sg, struct cpumask *unused)
8150 *sg = &per_cpu(sched_group_core, cpu).sg;
8155 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8156 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8159 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8160 struct sched_group **sg, struct cpumask *mask)
8163 #ifdef CONFIG_SCHED_MC
8164 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8165 group = cpumask_first(mask);
8166 #elif defined(CONFIG_SCHED_SMT)
8167 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8168 group = cpumask_first(mask);
8173 *sg = &per_cpu(sched_group_phys, group).sg;
8179 * The init_sched_build_groups can't handle what we want to do with node
8180 * groups, so roll our own. Now each node has its own list of groups which
8181 * gets dynamically allocated.
8183 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8184 static struct sched_group ***sched_group_nodes_bycpu;
8186 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8187 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8189 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8190 struct sched_group **sg,
8191 struct cpumask *nodemask)
8195 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8196 group = cpumask_first(nodemask);
8199 *sg = &per_cpu(sched_group_allnodes, group).sg;
8203 static void init_numa_sched_groups_power(struct sched_group *group_head)
8205 struct sched_group *sg = group_head;
8211 for_each_cpu(j, sched_group_cpus(sg)) {
8212 struct sched_domain *sd;
8214 sd = &per_cpu(phys_domains, j).sd;
8215 if (j != group_first_cpu(sd->groups)) {
8217 * Only add "power" once for each
8223 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8226 } while (sg != group_head);
8228 #endif /* CONFIG_NUMA */
8231 /* Free memory allocated for various sched_group structures */
8232 static void free_sched_groups(const struct cpumask *cpu_map,
8233 struct cpumask *nodemask)
8237 for_each_cpu(cpu, cpu_map) {
8238 struct sched_group **sched_group_nodes
8239 = sched_group_nodes_bycpu[cpu];
8241 if (!sched_group_nodes)
8244 for (i = 0; i < nr_node_ids; i++) {
8245 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8247 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8248 if (cpumask_empty(nodemask))
8258 if (oldsg != sched_group_nodes[i])
8261 kfree(sched_group_nodes);
8262 sched_group_nodes_bycpu[cpu] = NULL;
8265 #else /* !CONFIG_NUMA */
8266 static void free_sched_groups(const struct cpumask *cpu_map,
8267 struct cpumask *nodemask)
8270 #endif /* CONFIG_NUMA */
8273 * Initialize sched groups cpu_power.
8275 * cpu_power indicates the capacity of sched group, which is used while
8276 * distributing the load between different sched groups in a sched domain.
8277 * Typically cpu_power for all the groups in a sched domain will be same unless
8278 * there are asymmetries in the topology. If there are asymmetries, group
8279 * having more cpu_power will pickup more load compared to the group having
8282 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8283 * the maximum number of tasks a group can handle in the presence of other idle
8284 * or lightly loaded groups in the same sched domain.
8286 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8288 struct sched_domain *child;
8289 struct sched_group *group;
8291 WARN_ON(!sd || !sd->groups);
8293 if (cpu != group_first_cpu(sd->groups))
8298 sd->groups->__cpu_power = 0;
8301 * For perf policy, if the groups in child domain share resources
8302 * (for example cores sharing some portions of the cache hierarchy
8303 * or SMT), then set this domain groups cpu_power such that each group
8304 * can handle only one task, when there are other idle groups in the
8305 * same sched domain.
8307 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8309 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8310 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8315 * add cpu_power of each child group to this groups cpu_power
8317 group = child->groups;
8319 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8320 group = group->next;
8321 } while (group != child->groups);
8325 * Initializers for schedule domains
8326 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8329 #ifdef CONFIG_SCHED_DEBUG
8330 # define SD_INIT_NAME(sd, type) sd->name = #type
8332 # define SD_INIT_NAME(sd, type) do { } while (0)
8335 #define SD_INIT(sd, type) sd_init_##type(sd)
8337 #define SD_INIT_FUNC(type) \
8338 static noinline void sd_init_##type(struct sched_domain *sd) \
8340 memset(sd, 0, sizeof(*sd)); \
8341 *sd = SD_##type##_INIT; \
8342 sd->level = SD_LV_##type; \
8343 SD_INIT_NAME(sd, type); \
8348 SD_INIT_FUNC(ALLNODES)
8351 #ifdef CONFIG_SCHED_SMT
8352 SD_INIT_FUNC(SIBLING)
8354 #ifdef CONFIG_SCHED_MC
8358 static int default_relax_domain_level = -1;
8360 static int __init setup_relax_domain_level(char *str)
8364 val = simple_strtoul(str, NULL, 0);
8365 if (val < SD_LV_MAX)
8366 default_relax_domain_level = val;
8370 __setup("relax_domain_level=", setup_relax_domain_level);
8372 static void set_domain_attribute(struct sched_domain *sd,
8373 struct sched_domain_attr *attr)
8377 if (!attr || attr->relax_domain_level < 0) {
8378 if (default_relax_domain_level < 0)
8381 request = default_relax_domain_level;
8383 request = attr->relax_domain_level;
8384 if (request < sd->level) {
8385 /* turn off idle balance on this domain */
8386 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8388 /* turn on idle balance on this domain */
8389 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8394 * Build sched domains for a given set of cpus and attach the sched domains
8395 * to the individual cpus
8397 static int __build_sched_domains(const struct cpumask *cpu_map,
8398 struct sched_domain_attr *attr)
8400 int i, err = -ENOMEM;
8401 struct root_domain *rd;
8402 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8405 cpumask_var_t domainspan, covered, notcovered;
8406 struct sched_group **sched_group_nodes = NULL;
8407 int sd_allnodes = 0;
8409 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8411 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8412 goto free_domainspan;
8413 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
8417 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8418 goto free_notcovered;
8419 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8421 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8422 goto free_this_sibling_map;
8423 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8424 goto free_this_core_map;
8425 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8426 goto free_send_covered;
8430 * Allocate the per-node list of sched groups
8432 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8434 if (!sched_group_nodes) {
8435 printk(KERN_WARNING "Can not alloc sched group node list\n");
8440 rd = alloc_rootdomain();
8442 printk(KERN_WARNING "Cannot alloc root domain\n");
8443 goto free_sched_groups;
8447 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8451 * Set up domains for cpus specified by the cpu_map.
8453 for_each_cpu(i, cpu_map) {
8454 struct sched_domain *sd = NULL, *p;
8456 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8459 if (cpumask_weight(cpu_map) >
8460 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8461 sd = &per_cpu(allnodes_domains, i).sd;
8462 SD_INIT(sd, ALLNODES);
8463 set_domain_attribute(sd, attr);
8464 cpumask_copy(sched_domain_span(sd), cpu_map);
8465 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8471 sd = &per_cpu(node_domains, i).sd;
8473 set_domain_attribute(sd, attr);
8474 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8478 cpumask_and(sched_domain_span(sd),
8479 sched_domain_span(sd), cpu_map);
8483 sd = &per_cpu(phys_domains, i).sd;
8485 set_domain_attribute(sd, attr);
8486 cpumask_copy(sched_domain_span(sd), nodemask);
8490 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8492 #ifdef CONFIG_SCHED_MC
8494 sd = &per_cpu(core_domains, i).sd;
8496 set_domain_attribute(sd, attr);
8497 cpumask_and(sched_domain_span(sd), cpu_map,
8498 cpu_coregroup_mask(i));
8501 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8504 #ifdef CONFIG_SCHED_SMT
8506 sd = &per_cpu(cpu_domains, i).sd;
8507 SD_INIT(sd, SIBLING);
8508 set_domain_attribute(sd, attr);
8509 cpumask_and(sched_domain_span(sd),
8510 topology_thread_cpumask(i), cpu_map);
8513 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8517 #ifdef CONFIG_SCHED_SMT
8518 /* Set up CPU (sibling) groups */
8519 for_each_cpu(i, cpu_map) {
8520 cpumask_and(this_sibling_map,
8521 topology_thread_cpumask(i), cpu_map);
8522 if (i != cpumask_first(this_sibling_map))
8525 init_sched_build_groups(this_sibling_map, cpu_map,
8527 send_covered, tmpmask);
8531 #ifdef CONFIG_SCHED_MC
8532 /* Set up multi-core groups */
8533 for_each_cpu(i, cpu_map) {
8534 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8535 if (i != cpumask_first(this_core_map))
8538 init_sched_build_groups(this_core_map, cpu_map,
8540 send_covered, tmpmask);
8544 /* Set up physical groups */
8545 for (i = 0; i < nr_node_ids; i++) {
8546 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8547 if (cpumask_empty(nodemask))
8550 init_sched_build_groups(nodemask, cpu_map,
8552 send_covered, tmpmask);
8556 /* Set up node groups */
8558 init_sched_build_groups(cpu_map, cpu_map,
8559 &cpu_to_allnodes_group,
8560 send_covered, tmpmask);
8563 for (i = 0; i < nr_node_ids; i++) {
8564 /* Set up node groups */
8565 struct sched_group *sg, *prev;
8568 cpumask_clear(covered);
8569 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8570 if (cpumask_empty(nodemask)) {
8571 sched_group_nodes[i] = NULL;
8575 sched_domain_node_span(i, domainspan);
8576 cpumask_and(domainspan, domainspan, cpu_map);
8578 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8581 printk(KERN_WARNING "Can not alloc domain group for "
8585 sched_group_nodes[i] = sg;
8586 for_each_cpu(j, nodemask) {
8587 struct sched_domain *sd;
8589 sd = &per_cpu(node_domains, j).sd;
8592 sg->__cpu_power = 0;
8593 cpumask_copy(sched_group_cpus(sg), nodemask);
8595 cpumask_or(covered, covered, nodemask);
8598 for (j = 0; j < nr_node_ids; j++) {
8599 int n = (i + j) % nr_node_ids;
8601 cpumask_complement(notcovered, covered);
8602 cpumask_and(tmpmask, notcovered, cpu_map);
8603 cpumask_and(tmpmask, tmpmask, domainspan);
8604 if (cpumask_empty(tmpmask))
8607 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8608 if (cpumask_empty(tmpmask))
8611 sg = kmalloc_node(sizeof(struct sched_group) +
8616 "Can not alloc domain group for node %d\n", j);
8619 sg->__cpu_power = 0;
8620 cpumask_copy(sched_group_cpus(sg), tmpmask);
8621 sg->next = prev->next;
8622 cpumask_or(covered, covered, tmpmask);
8629 /* Calculate CPU power for physical packages and nodes */
8630 #ifdef CONFIG_SCHED_SMT
8631 for_each_cpu(i, cpu_map) {
8632 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8634 init_sched_groups_power(i, sd);
8637 #ifdef CONFIG_SCHED_MC
8638 for_each_cpu(i, cpu_map) {
8639 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8641 init_sched_groups_power(i, sd);
8645 for_each_cpu(i, cpu_map) {
8646 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8648 init_sched_groups_power(i, sd);
8652 for (i = 0; i < nr_node_ids; i++)
8653 init_numa_sched_groups_power(sched_group_nodes[i]);
8656 struct sched_group *sg;
8658 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8660 init_numa_sched_groups_power(sg);
8664 /* Attach the domains */
8665 for_each_cpu(i, cpu_map) {
8666 struct sched_domain *sd;
8667 #ifdef CONFIG_SCHED_SMT
8668 sd = &per_cpu(cpu_domains, i).sd;
8669 #elif defined(CONFIG_SCHED_MC)
8670 sd = &per_cpu(core_domains, i).sd;
8672 sd = &per_cpu(phys_domains, i).sd;
8674 cpu_attach_domain(sd, rd, i);
8680 free_cpumask_var(tmpmask);
8682 free_cpumask_var(send_covered);
8684 free_cpumask_var(this_core_map);
8685 free_this_sibling_map:
8686 free_cpumask_var(this_sibling_map);
8688 free_cpumask_var(nodemask);
8691 free_cpumask_var(notcovered);
8693 free_cpumask_var(covered);
8695 free_cpumask_var(domainspan);
8702 kfree(sched_group_nodes);
8708 free_sched_groups(cpu_map, tmpmask);
8709 free_rootdomain(rd);
8714 static int build_sched_domains(const struct cpumask *cpu_map)
8716 return __build_sched_domains(cpu_map, NULL);
8719 static struct cpumask *doms_cur; /* current sched domains */
8720 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8721 static struct sched_domain_attr *dattr_cur;
8722 /* attribues of custom domains in 'doms_cur' */
8725 * Special case: If a kmalloc of a doms_cur partition (array of
8726 * cpumask) fails, then fallback to a single sched domain,
8727 * as determined by the single cpumask fallback_doms.
8729 static cpumask_var_t fallback_doms;
8732 * arch_update_cpu_topology lets virtualized architectures update the
8733 * cpu core maps. It is supposed to return 1 if the topology changed
8734 * or 0 if it stayed the same.
8736 int __attribute__((weak)) arch_update_cpu_topology(void)
8742 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8743 * For now this just excludes isolated cpus, but could be used to
8744 * exclude other special cases in the future.
8746 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8750 arch_update_cpu_topology();
8752 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8754 doms_cur = fallback_doms;
8755 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8757 err = build_sched_domains(doms_cur);
8758 register_sched_domain_sysctl();
8763 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8764 struct cpumask *tmpmask)
8766 free_sched_groups(cpu_map, tmpmask);
8770 * Detach sched domains from a group of cpus specified in cpu_map
8771 * These cpus will now be attached to the NULL domain
8773 static void detach_destroy_domains(const struct cpumask *cpu_map)
8775 /* Save because hotplug lock held. */
8776 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8779 for_each_cpu(i, cpu_map)
8780 cpu_attach_domain(NULL, &def_root_domain, i);
8781 synchronize_sched();
8782 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8785 /* handle null as "default" */
8786 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8787 struct sched_domain_attr *new, int idx_new)
8789 struct sched_domain_attr tmp;
8796 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8797 new ? (new + idx_new) : &tmp,
8798 sizeof(struct sched_domain_attr));
8802 * Partition sched domains as specified by the 'ndoms_new'
8803 * cpumasks in the array doms_new[] of cpumasks. This compares
8804 * doms_new[] to the current sched domain partitioning, doms_cur[].
8805 * It destroys each deleted domain and builds each new domain.
8807 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8808 * The masks don't intersect (don't overlap.) We should setup one
8809 * sched domain for each mask. CPUs not in any of the cpumasks will
8810 * not be load balanced. If the same cpumask appears both in the
8811 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8814 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8815 * ownership of it and will kfree it when done with it. If the caller
8816 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8817 * ndoms_new == 1, and partition_sched_domains() will fallback to
8818 * the single partition 'fallback_doms', it also forces the domains
8821 * If doms_new == NULL it will be replaced with cpu_online_mask.
8822 * ndoms_new == 0 is a special case for destroying existing domains,
8823 * and it will not create the default domain.
8825 * Call with hotplug lock held
8827 /* FIXME: Change to struct cpumask *doms_new[] */
8828 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8829 struct sched_domain_attr *dattr_new)
8834 mutex_lock(&sched_domains_mutex);
8836 /* always unregister in case we don't destroy any domains */
8837 unregister_sched_domain_sysctl();
8839 /* Let architecture update cpu core mappings. */
8840 new_topology = arch_update_cpu_topology();
8842 n = doms_new ? ndoms_new : 0;
8844 /* Destroy deleted domains */
8845 for (i = 0; i < ndoms_cur; i++) {
8846 for (j = 0; j < n && !new_topology; j++) {
8847 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8848 && dattrs_equal(dattr_cur, i, dattr_new, j))
8851 /* no match - a current sched domain not in new doms_new[] */
8852 detach_destroy_domains(doms_cur + i);
8857 if (doms_new == NULL) {
8859 doms_new = fallback_doms;
8860 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8861 WARN_ON_ONCE(dattr_new);
8864 /* Build new domains */
8865 for (i = 0; i < ndoms_new; i++) {
8866 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8867 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8868 && dattrs_equal(dattr_new, i, dattr_cur, j))
8871 /* no match - add a new doms_new */
8872 __build_sched_domains(doms_new + i,
8873 dattr_new ? dattr_new + i : NULL);
8878 /* Remember the new sched domains */
8879 if (doms_cur != fallback_doms)
8881 kfree(dattr_cur); /* kfree(NULL) is safe */
8882 doms_cur = doms_new;
8883 dattr_cur = dattr_new;
8884 ndoms_cur = ndoms_new;
8886 register_sched_domain_sysctl();
8888 mutex_unlock(&sched_domains_mutex);
8891 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8892 static void arch_reinit_sched_domains(void)
8896 /* Destroy domains first to force the rebuild */
8897 partition_sched_domains(0, NULL, NULL);
8899 rebuild_sched_domains();
8903 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8905 unsigned int level = 0;
8907 if (sscanf(buf, "%u", &level) != 1)
8911 * level is always be positive so don't check for
8912 * level < POWERSAVINGS_BALANCE_NONE which is 0
8913 * What happens on 0 or 1 byte write,
8914 * need to check for count as well?
8917 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8921 sched_smt_power_savings = level;
8923 sched_mc_power_savings = level;
8925 arch_reinit_sched_domains();
8930 #ifdef CONFIG_SCHED_MC
8931 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8934 return sprintf(page, "%u\n", sched_mc_power_savings);
8936 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8937 const char *buf, size_t count)
8939 return sched_power_savings_store(buf, count, 0);
8941 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8942 sched_mc_power_savings_show,
8943 sched_mc_power_savings_store);
8946 #ifdef CONFIG_SCHED_SMT
8947 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8950 return sprintf(page, "%u\n", sched_smt_power_savings);
8952 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8953 const char *buf, size_t count)
8955 return sched_power_savings_store(buf, count, 1);
8957 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8958 sched_smt_power_savings_show,
8959 sched_smt_power_savings_store);
8962 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8966 #ifdef CONFIG_SCHED_SMT
8968 err = sysfs_create_file(&cls->kset.kobj,
8969 &attr_sched_smt_power_savings.attr);
8971 #ifdef CONFIG_SCHED_MC
8972 if (!err && mc_capable())
8973 err = sysfs_create_file(&cls->kset.kobj,
8974 &attr_sched_mc_power_savings.attr);
8978 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8980 #ifndef CONFIG_CPUSETS
8982 * Add online and remove offline CPUs from the scheduler domains.
8983 * When cpusets are enabled they take over this function.
8985 static int update_sched_domains(struct notifier_block *nfb,
8986 unsigned long action, void *hcpu)
8990 case CPU_ONLINE_FROZEN:
8992 case CPU_DEAD_FROZEN:
8993 partition_sched_domains(1, NULL, NULL);
9002 static int update_runtime(struct notifier_block *nfb,
9003 unsigned long action, void *hcpu)
9005 int cpu = (int)(long)hcpu;
9008 case CPU_DOWN_PREPARE:
9009 case CPU_DOWN_PREPARE_FROZEN:
9010 disable_runtime(cpu_rq(cpu));
9013 case CPU_DOWN_FAILED:
9014 case CPU_DOWN_FAILED_FROZEN:
9016 case CPU_ONLINE_FROZEN:
9017 enable_runtime(cpu_rq(cpu));
9025 void __init sched_init_smp(void)
9027 cpumask_var_t non_isolated_cpus;
9029 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9031 #if defined(CONFIG_NUMA)
9032 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9034 BUG_ON(sched_group_nodes_bycpu == NULL);
9037 mutex_lock(&sched_domains_mutex);
9038 arch_init_sched_domains(cpu_online_mask);
9039 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9040 if (cpumask_empty(non_isolated_cpus))
9041 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9042 mutex_unlock(&sched_domains_mutex);
9045 #ifndef CONFIG_CPUSETS
9046 /* XXX: Theoretical race here - CPU may be hotplugged now */
9047 hotcpu_notifier(update_sched_domains, 0);
9050 /* RT runtime code needs to handle some hotplug events */
9051 hotcpu_notifier(update_runtime, 0);
9055 /* Move init over to a non-isolated CPU */
9056 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9058 sched_init_granularity();
9059 free_cpumask_var(non_isolated_cpus);
9061 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9062 init_sched_rt_class();
9065 void __init sched_init_smp(void)
9067 sched_init_granularity();
9069 #endif /* CONFIG_SMP */
9071 const_debug unsigned int sysctl_timer_migration = 1;
9073 int in_sched_functions(unsigned long addr)
9075 return in_lock_functions(addr) ||
9076 (addr >= (unsigned long)__sched_text_start
9077 && addr < (unsigned long)__sched_text_end);
9080 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9082 cfs_rq->tasks_timeline = RB_ROOT;
9083 INIT_LIST_HEAD(&cfs_rq->tasks);
9084 #ifdef CONFIG_FAIR_GROUP_SCHED
9087 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9090 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9092 struct rt_prio_array *array;
9095 array = &rt_rq->active;
9096 for (i = 0; i < MAX_RT_PRIO; i++) {
9097 INIT_LIST_HEAD(array->queue + i);
9098 __clear_bit(i, array->bitmap);
9100 /* delimiter for bitsearch: */
9101 __set_bit(MAX_RT_PRIO, array->bitmap);
9103 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9104 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9106 rt_rq->highest_prio.next = MAX_RT_PRIO;
9110 rt_rq->rt_nr_migratory = 0;
9111 rt_rq->overloaded = 0;
9112 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
9116 rt_rq->rt_throttled = 0;
9117 rt_rq->rt_runtime = 0;
9118 spin_lock_init(&rt_rq->rt_runtime_lock);
9120 #ifdef CONFIG_RT_GROUP_SCHED
9121 rt_rq->rt_nr_boosted = 0;
9126 #ifdef CONFIG_FAIR_GROUP_SCHED
9127 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9128 struct sched_entity *se, int cpu, int add,
9129 struct sched_entity *parent)
9131 struct rq *rq = cpu_rq(cpu);
9132 tg->cfs_rq[cpu] = cfs_rq;
9133 init_cfs_rq(cfs_rq, rq);
9136 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9139 /* se could be NULL for init_task_group */
9144 se->cfs_rq = &rq->cfs;
9146 se->cfs_rq = parent->my_q;
9149 se->load.weight = tg->shares;
9150 se->load.inv_weight = 0;
9151 se->parent = parent;
9155 #ifdef CONFIG_RT_GROUP_SCHED
9156 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9157 struct sched_rt_entity *rt_se, int cpu, int add,
9158 struct sched_rt_entity *parent)
9160 struct rq *rq = cpu_rq(cpu);
9162 tg->rt_rq[cpu] = rt_rq;
9163 init_rt_rq(rt_rq, rq);
9165 rt_rq->rt_se = rt_se;
9166 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9168 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9170 tg->rt_se[cpu] = rt_se;
9175 rt_se->rt_rq = &rq->rt;
9177 rt_se->rt_rq = parent->my_q;
9179 rt_se->my_q = rt_rq;
9180 rt_se->parent = parent;
9181 INIT_LIST_HEAD(&rt_se->run_list);
9185 void __init sched_init(void)
9188 unsigned long alloc_size = 0, ptr;
9190 #ifdef CONFIG_FAIR_GROUP_SCHED
9191 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9193 #ifdef CONFIG_RT_GROUP_SCHED
9194 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9196 #ifdef CONFIG_USER_SCHED
9199 #ifdef CONFIG_CPUMASK_OFFSTACK
9200 alloc_size += num_possible_cpus() * cpumask_size();
9203 * As sched_init() is called before page_alloc is setup,
9204 * we use alloc_bootmem().
9207 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9209 #ifdef CONFIG_FAIR_GROUP_SCHED
9210 init_task_group.se = (struct sched_entity **)ptr;
9211 ptr += nr_cpu_ids * sizeof(void **);
9213 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9214 ptr += nr_cpu_ids * sizeof(void **);
9216 #ifdef CONFIG_USER_SCHED
9217 root_task_group.se = (struct sched_entity **)ptr;
9218 ptr += nr_cpu_ids * sizeof(void **);
9220 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9221 ptr += nr_cpu_ids * sizeof(void **);
9222 #endif /* CONFIG_USER_SCHED */
9223 #endif /* CONFIG_FAIR_GROUP_SCHED */
9224 #ifdef CONFIG_RT_GROUP_SCHED
9225 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9226 ptr += nr_cpu_ids * sizeof(void **);
9228 init_task_group.rt_rq = (struct rt_rq **)ptr;
9229 ptr += nr_cpu_ids * sizeof(void **);
9231 #ifdef CONFIG_USER_SCHED
9232 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9233 ptr += nr_cpu_ids * sizeof(void **);
9235 root_task_group.rt_rq = (struct rt_rq **)ptr;
9236 ptr += nr_cpu_ids * sizeof(void **);
9237 #endif /* CONFIG_USER_SCHED */
9238 #endif /* CONFIG_RT_GROUP_SCHED */
9239 #ifdef CONFIG_CPUMASK_OFFSTACK
9240 for_each_possible_cpu(i) {
9241 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9242 ptr += cpumask_size();
9244 #endif /* CONFIG_CPUMASK_OFFSTACK */
9248 init_defrootdomain();
9251 init_rt_bandwidth(&def_rt_bandwidth,
9252 global_rt_period(), global_rt_runtime());
9254 #ifdef CONFIG_RT_GROUP_SCHED
9255 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9256 global_rt_period(), global_rt_runtime());
9257 #ifdef CONFIG_USER_SCHED
9258 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9259 global_rt_period(), RUNTIME_INF);
9260 #endif /* CONFIG_USER_SCHED */
9261 #endif /* CONFIG_RT_GROUP_SCHED */
9263 #ifdef CONFIG_GROUP_SCHED
9264 list_add(&init_task_group.list, &task_groups);
9265 INIT_LIST_HEAD(&init_task_group.children);
9267 #ifdef CONFIG_USER_SCHED
9268 INIT_LIST_HEAD(&root_task_group.children);
9269 init_task_group.parent = &root_task_group;
9270 list_add(&init_task_group.siblings, &root_task_group.children);
9271 #endif /* CONFIG_USER_SCHED */
9272 #endif /* CONFIG_GROUP_SCHED */
9274 for_each_possible_cpu(i) {
9278 spin_lock_init(&rq->lock);
9280 rq->calc_load_active = 0;
9281 rq->calc_load_update = jiffies + LOAD_FREQ;
9282 init_cfs_rq(&rq->cfs, rq);
9283 init_rt_rq(&rq->rt, rq);
9284 #ifdef CONFIG_FAIR_GROUP_SCHED
9285 init_task_group.shares = init_task_group_load;
9286 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9287 #ifdef CONFIG_CGROUP_SCHED
9289 * How much cpu bandwidth does init_task_group get?
9291 * In case of task-groups formed thr' the cgroup filesystem, it
9292 * gets 100% of the cpu resources in the system. This overall
9293 * system cpu resource is divided among the tasks of
9294 * init_task_group and its child task-groups in a fair manner,
9295 * based on each entity's (task or task-group's) weight
9296 * (se->load.weight).
9298 * In other words, if init_task_group has 10 tasks of weight
9299 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9300 * then A0's share of the cpu resource is:
9302 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9304 * We achieve this by letting init_task_group's tasks sit
9305 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9307 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9308 #elif defined CONFIG_USER_SCHED
9309 root_task_group.shares = NICE_0_LOAD;
9310 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9312 * In case of task-groups formed thr' the user id of tasks,
9313 * init_task_group represents tasks belonging to root user.
9314 * Hence it forms a sibling of all subsequent groups formed.
9315 * In this case, init_task_group gets only a fraction of overall
9316 * system cpu resource, based on the weight assigned to root
9317 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9318 * by letting tasks of init_task_group sit in a separate cfs_rq
9319 * (init_cfs_rq) and having one entity represent this group of
9320 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9322 init_tg_cfs_entry(&init_task_group,
9323 &per_cpu(init_cfs_rq, i),
9324 &per_cpu(init_sched_entity, i), i, 1,
9325 root_task_group.se[i]);
9328 #endif /* CONFIG_FAIR_GROUP_SCHED */
9330 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9331 #ifdef CONFIG_RT_GROUP_SCHED
9332 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9333 #ifdef CONFIG_CGROUP_SCHED
9334 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9335 #elif defined CONFIG_USER_SCHED
9336 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9337 init_tg_rt_entry(&init_task_group,
9338 &per_cpu(init_rt_rq, i),
9339 &per_cpu(init_sched_rt_entity, i), i, 1,
9340 root_task_group.rt_se[i]);
9344 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9345 rq->cpu_load[j] = 0;
9349 rq->active_balance = 0;
9350 rq->next_balance = jiffies;
9354 rq->migration_thread = NULL;
9355 INIT_LIST_HEAD(&rq->migration_queue);
9356 rq_attach_root(rq, &def_root_domain);
9359 atomic_set(&rq->nr_iowait, 0);
9362 set_load_weight(&init_task);
9364 #ifdef CONFIG_PREEMPT_NOTIFIERS
9365 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9369 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9372 #ifdef CONFIG_RT_MUTEXES
9373 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9377 * The boot idle thread does lazy MMU switching as well:
9379 atomic_inc(&init_mm.mm_count);
9380 enter_lazy_tlb(&init_mm, current);
9383 * Make us the idle thread. Technically, schedule() should not be
9384 * called from this thread, however somewhere below it might be,
9385 * but because we are the idle thread, we just pick up running again
9386 * when this runqueue becomes "idle".
9388 init_idle(current, smp_processor_id());
9390 calc_load_update = jiffies + LOAD_FREQ;
9393 * During early bootup we pretend to be a normal task:
9395 current->sched_class = &fair_sched_class;
9397 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9398 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9401 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9402 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9404 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9407 perf_counter_init();
9409 scheduler_running = 1;
9412 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9413 void __might_sleep(char *file, int line)
9416 static unsigned long prev_jiffy; /* ratelimiting */
9418 if ((!in_atomic() && !irqs_disabled()) ||
9419 system_state != SYSTEM_RUNNING || oops_in_progress)
9421 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9423 prev_jiffy = jiffies;
9426 "BUG: sleeping function called from invalid context at %s:%d\n",
9429 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9430 in_atomic(), irqs_disabled(),
9431 current->pid, current->comm);
9433 debug_show_held_locks(current);
9434 if (irqs_disabled())
9435 print_irqtrace_events(current);
9439 EXPORT_SYMBOL(__might_sleep);
9442 #ifdef CONFIG_MAGIC_SYSRQ
9443 static void normalize_task(struct rq *rq, struct task_struct *p)
9447 update_rq_clock(rq);
9448 on_rq = p->se.on_rq;
9450 deactivate_task(rq, p, 0);
9451 __setscheduler(rq, p, SCHED_NORMAL, 0);
9453 activate_task(rq, p, 0);
9454 resched_task(rq->curr);
9458 void normalize_rt_tasks(void)
9460 struct task_struct *g, *p;
9461 unsigned long flags;
9464 read_lock_irqsave(&tasklist_lock, flags);
9465 do_each_thread(g, p) {
9467 * Only normalize user tasks:
9472 p->se.exec_start = 0;
9473 #ifdef CONFIG_SCHEDSTATS
9474 p->se.wait_start = 0;
9475 p->se.sleep_start = 0;
9476 p->se.block_start = 0;
9481 * Renice negative nice level userspace
9484 if (TASK_NICE(p) < 0 && p->mm)
9485 set_user_nice(p, 0);
9489 spin_lock(&p->pi_lock);
9490 rq = __task_rq_lock(p);
9492 normalize_task(rq, p);
9494 __task_rq_unlock(rq);
9495 spin_unlock(&p->pi_lock);
9496 } while_each_thread(g, p);
9498 read_unlock_irqrestore(&tasklist_lock, flags);
9501 #endif /* CONFIG_MAGIC_SYSRQ */
9505 * These functions are only useful for the IA64 MCA handling.
9507 * They can only be called when the whole system has been
9508 * stopped - every CPU needs to be quiescent, and no scheduling
9509 * activity can take place. Using them for anything else would
9510 * be a serious bug, and as a result, they aren't even visible
9511 * under any other configuration.
9515 * curr_task - return the current task for a given cpu.
9516 * @cpu: the processor in question.
9518 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9520 struct task_struct *curr_task(int cpu)
9522 return cpu_curr(cpu);
9526 * set_curr_task - set the current task for a given cpu.
9527 * @cpu: the processor in question.
9528 * @p: the task pointer to set.
9530 * Description: This function must only be used when non-maskable interrupts
9531 * are serviced on a separate stack. It allows the architecture to switch the
9532 * notion of the current task on a cpu in a non-blocking manner. This function
9533 * must be called with all CPU's synchronized, and interrupts disabled, the
9534 * and caller must save the original value of the current task (see
9535 * curr_task() above) and restore that value before reenabling interrupts and
9536 * re-starting the system.
9538 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9540 void set_curr_task(int cpu, struct task_struct *p)
9547 #ifdef CONFIG_FAIR_GROUP_SCHED
9548 static void free_fair_sched_group(struct task_group *tg)
9552 for_each_possible_cpu(i) {
9554 kfree(tg->cfs_rq[i]);
9564 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9566 struct cfs_rq *cfs_rq;
9567 struct sched_entity *se;
9571 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9574 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9578 tg->shares = NICE_0_LOAD;
9580 for_each_possible_cpu(i) {
9583 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9584 GFP_KERNEL, cpu_to_node(i));
9588 se = kzalloc_node(sizeof(struct sched_entity),
9589 GFP_KERNEL, cpu_to_node(i));
9593 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9602 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9604 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9605 &cpu_rq(cpu)->leaf_cfs_rq_list);
9608 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9610 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9612 #else /* !CONFG_FAIR_GROUP_SCHED */
9613 static inline void free_fair_sched_group(struct task_group *tg)
9618 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9623 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9627 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9630 #endif /* CONFIG_FAIR_GROUP_SCHED */
9632 #ifdef CONFIG_RT_GROUP_SCHED
9633 static void free_rt_sched_group(struct task_group *tg)
9637 destroy_rt_bandwidth(&tg->rt_bandwidth);
9639 for_each_possible_cpu(i) {
9641 kfree(tg->rt_rq[i]);
9643 kfree(tg->rt_se[i]);
9651 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9653 struct rt_rq *rt_rq;
9654 struct sched_rt_entity *rt_se;
9658 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9661 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9665 init_rt_bandwidth(&tg->rt_bandwidth,
9666 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9668 for_each_possible_cpu(i) {
9671 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9672 GFP_KERNEL, cpu_to_node(i));
9676 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9677 GFP_KERNEL, cpu_to_node(i));
9681 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9690 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9692 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9693 &cpu_rq(cpu)->leaf_rt_rq_list);
9696 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9698 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9700 #else /* !CONFIG_RT_GROUP_SCHED */
9701 static inline void free_rt_sched_group(struct task_group *tg)
9706 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9711 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9715 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9718 #endif /* CONFIG_RT_GROUP_SCHED */
9720 #ifdef CONFIG_GROUP_SCHED
9721 static void free_sched_group(struct task_group *tg)
9723 free_fair_sched_group(tg);
9724 free_rt_sched_group(tg);
9728 /* allocate runqueue etc for a new task group */
9729 struct task_group *sched_create_group(struct task_group *parent)
9731 struct task_group *tg;
9732 unsigned long flags;
9735 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9737 return ERR_PTR(-ENOMEM);
9739 if (!alloc_fair_sched_group(tg, parent))
9742 if (!alloc_rt_sched_group(tg, parent))
9745 spin_lock_irqsave(&task_group_lock, flags);
9746 for_each_possible_cpu(i) {
9747 register_fair_sched_group(tg, i);
9748 register_rt_sched_group(tg, i);
9750 list_add_rcu(&tg->list, &task_groups);
9752 WARN_ON(!parent); /* root should already exist */
9754 tg->parent = parent;
9755 INIT_LIST_HEAD(&tg->children);
9756 list_add_rcu(&tg->siblings, &parent->children);
9757 spin_unlock_irqrestore(&task_group_lock, flags);
9762 free_sched_group(tg);
9763 return ERR_PTR(-ENOMEM);
9766 /* rcu callback to free various structures associated with a task group */
9767 static void free_sched_group_rcu(struct rcu_head *rhp)
9769 /* now it should be safe to free those cfs_rqs */
9770 free_sched_group(container_of(rhp, struct task_group, rcu));
9773 /* Destroy runqueue etc associated with a task group */
9774 void sched_destroy_group(struct task_group *tg)
9776 unsigned long flags;
9779 spin_lock_irqsave(&task_group_lock, flags);
9780 for_each_possible_cpu(i) {
9781 unregister_fair_sched_group(tg, i);
9782 unregister_rt_sched_group(tg, i);
9784 list_del_rcu(&tg->list);
9785 list_del_rcu(&tg->siblings);
9786 spin_unlock_irqrestore(&task_group_lock, flags);
9788 /* wait for possible concurrent references to cfs_rqs complete */
9789 call_rcu(&tg->rcu, free_sched_group_rcu);
9792 /* change task's runqueue when it moves between groups.
9793 * The caller of this function should have put the task in its new group
9794 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9795 * reflect its new group.
9797 void sched_move_task(struct task_struct *tsk)
9800 unsigned long flags;
9803 rq = task_rq_lock(tsk, &flags);
9805 update_rq_clock(rq);
9807 running = task_current(rq, tsk);
9808 on_rq = tsk->se.on_rq;
9811 dequeue_task(rq, tsk, 0);
9812 if (unlikely(running))
9813 tsk->sched_class->put_prev_task(rq, tsk);
9815 set_task_rq(tsk, task_cpu(tsk));
9817 #ifdef CONFIG_FAIR_GROUP_SCHED
9818 if (tsk->sched_class->moved_group)
9819 tsk->sched_class->moved_group(tsk);
9822 if (unlikely(running))
9823 tsk->sched_class->set_curr_task(rq);
9825 enqueue_task(rq, tsk, 0);
9827 task_rq_unlock(rq, &flags);
9829 #endif /* CONFIG_GROUP_SCHED */
9831 #ifdef CONFIG_FAIR_GROUP_SCHED
9832 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9834 struct cfs_rq *cfs_rq = se->cfs_rq;
9839 dequeue_entity(cfs_rq, se, 0);
9841 se->load.weight = shares;
9842 se->load.inv_weight = 0;
9845 enqueue_entity(cfs_rq, se, 0);
9848 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9850 struct cfs_rq *cfs_rq = se->cfs_rq;
9851 struct rq *rq = cfs_rq->rq;
9852 unsigned long flags;
9854 spin_lock_irqsave(&rq->lock, flags);
9855 __set_se_shares(se, shares);
9856 spin_unlock_irqrestore(&rq->lock, flags);
9859 static DEFINE_MUTEX(shares_mutex);
9861 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9864 unsigned long flags;
9867 * We can't change the weight of the root cgroup.
9872 if (shares < MIN_SHARES)
9873 shares = MIN_SHARES;
9874 else if (shares > MAX_SHARES)
9875 shares = MAX_SHARES;
9877 mutex_lock(&shares_mutex);
9878 if (tg->shares == shares)
9881 spin_lock_irqsave(&task_group_lock, flags);
9882 for_each_possible_cpu(i)
9883 unregister_fair_sched_group(tg, i);
9884 list_del_rcu(&tg->siblings);
9885 spin_unlock_irqrestore(&task_group_lock, flags);
9887 /* wait for any ongoing reference to this group to finish */
9888 synchronize_sched();
9891 * Now we are free to modify the group's share on each cpu
9892 * w/o tripping rebalance_share or load_balance_fair.
9894 tg->shares = shares;
9895 for_each_possible_cpu(i) {
9899 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9900 set_se_shares(tg->se[i], shares);
9904 * Enable load balance activity on this group, by inserting it back on
9905 * each cpu's rq->leaf_cfs_rq_list.
9907 spin_lock_irqsave(&task_group_lock, flags);
9908 for_each_possible_cpu(i)
9909 register_fair_sched_group(tg, i);
9910 list_add_rcu(&tg->siblings, &tg->parent->children);
9911 spin_unlock_irqrestore(&task_group_lock, flags);
9913 mutex_unlock(&shares_mutex);
9917 unsigned long sched_group_shares(struct task_group *tg)
9923 #ifdef CONFIG_RT_GROUP_SCHED
9925 * Ensure that the real time constraints are schedulable.
9927 static DEFINE_MUTEX(rt_constraints_mutex);
9929 static unsigned long to_ratio(u64 period, u64 runtime)
9931 if (runtime == RUNTIME_INF)
9934 return div64_u64(runtime << 20, period);
9937 /* Must be called with tasklist_lock held */
9938 static inline int tg_has_rt_tasks(struct task_group *tg)
9940 struct task_struct *g, *p;
9942 do_each_thread(g, p) {
9943 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9945 } while_each_thread(g, p);
9950 struct rt_schedulable_data {
9951 struct task_group *tg;
9956 static int tg_schedulable(struct task_group *tg, void *data)
9958 struct rt_schedulable_data *d = data;
9959 struct task_group *child;
9960 unsigned long total, sum = 0;
9961 u64 period, runtime;
9963 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9964 runtime = tg->rt_bandwidth.rt_runtime;
9967 period = d->rt_period;
9968 runtime = d->rt_runtime;
9971 #ifdef CONFIG_USER_SCHED
9972 if (tg == &root_task_group) {
9973 period = global_rt_period();
9974 runtime = global_rt_runtime();
9979 * Cannot have more runtime than the period.
9981 if (runtime > period && runtime != RUNTIME_INF)
9985 * Ensure we don't starve existing RT tasks.
9987 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9990 total = to_ratio(period, runtime);
9993 * Nobody can have more than the global setting allows.
9995 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9999 * The sum of our children's runtime should not exceed our own.
10001 list_for_each_entry_rcu(child, &tg->children, siblings) {
10002 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10003 runtime = child->rt_bandwidth.rt_runtime;
10005 if (child == d->tg) {
10006 period = d->rt_period;
10007 runtime = d->rt_runtime;
10010 sum += to_ratio(period, runtime);
10019 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10021 struct rt_schedulable_data data = {
10023 .rt_period = period,
10024 .rt_runtime = runtime,
10027 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10030 static int tg_set_bandwidth(struct task_group *tg,
10031 u64 rt_period, u64 rt_runtime)
10035 mutex_lock(&rt_constraints_mutex);
10036 read_lock(&tasklist_lock);
10037 err = __rt_schedulable(tg, rt_period, rt_runtime);
10041 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10042 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10043 tg->rt_bandwidth.rt_runtime = rt_runtime;
10045 for_each_possible_cpu(i) {
10046 struct rt_rq *rt_rq = tg->rt_rq[i];
10048 spin_lock(&rt_rq->rt_runtime_lock);
10049 rt_rq->rt_runtime = rt_runtime;
10050 spin_unlock(&rt_rq->rt_runtime_lock);
10052 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10054 read_unlock(&tasklist_lock);
10055 mutex_unlock(&rt_constraints_mutex);
10060 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10062 u64 rt_runtime, rt_period;
10064 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10065 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10066 if (rt_runtime_us < 0)
10067 rt_runtime = RUNTIME_INF;
10069 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10072 long sched_group_rt_runtime(struct task_group *tg)
10076 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10079 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10080 do_div(rt_runtime_us, NSEC_PER_USEC);
10081 return rt_runtime_us;
10084 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10086 u64 rt_runtime, rt_period;
10088 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10089 rt_runtime = tg->rt_bandwidth.rt_runtime;
10091 if (rt_period == 0)
10094 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10097 long sched_group_rt_period(struct task_group *tg)
10101 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10102 do_div(rt_period_us, NSEC_PER_USEC);
10103 return rt_period_us;
10106 static int sched_rt_global_constraints(void)
10108 u64 runtime, period;
10111 if (sysctl_sched_rt_period <= 0)
10114 runtime = global_rt_runtime();
10115 period = global_rt_period();
10118 * Sanity check on the sysctl variables.
10120 if (runtime > period && runtime != RUNTIME_INF)
10123 mutex_lock(&rt_constraints_mutex);
10124 read_lock(&tasklist_lock);
10125 ret = __rt_schedulable(NULL, 0, 0);
10126 read_unlock(&tasklist_lock);
10127 mutex_unlock(&rt_constraints_mutex);
10132 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10134 /* Don't accept realtime tasks when there is no way for them to run */
10135 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10141 #else /* !CONFIG_RT_GROUP_SCHED */
10142 static int sched_rt_global_constraints(void)
10144 unsigned long flags;
10147 if (sysctl_sched_rt_period <= 0)
10151 * There's always some RT tasks in the root group
10152 * -- migration, kstopmachine etc..
10154 if (sysctl_sched_rt_runtime == 0)
10157 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10158 for_each_possible_cpu(i) {
10159 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10161 spin_lock(&rt_rq->rt_runtime_lock);
10162 rt_rq->rt_runtime = global_rt_runtime();
10163 spin_unlock(&rt_rq->rt_runtime_lock);
10165 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10169 #endif /* CONFIG_RT_GROUP_SCHED */
10171 int sched_rt_handler(struct ctl_table *table, int write,
10172 struct file *filp, void __user *buffer, size_t *lenp,
10176 int old_period, old_runtime;
10177 static DEFINE_MUTEX(mutex);
10179 mutex_lock(&mutex);
10180 old_period = sysctl_sched_rt_period;
10181 old_runtime = sysctl_sched_rt_runtime;
10183 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10185 if (!ret && write) {
10186 ret = sched_rt_global_constraints();
10188 sysctl_sched_rt_period = old_period;
10189 sysctl_sched_rt_runtime = old_runtime;
10191 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10192 def_rt_bandwidth.rt_period =
10193 ns_to_ktime(global_rt_period());
10196 mutex_unlock(&mutex);
10201 #ifdef CONFIG_CGROUP_SCHED
10203 /* return corresponding task_group object of a cgroup */
10204 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10206 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10207 struct task_group, css);
10210 static struct cgroup_subsys_state *
10211 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10213 struct task_group *tg, *parent;
10215 if (!cgrp->parent) {
10216 /* This is early initialization for the top cgroup */
10217 return &init_task_group.css;
10220 parent = cgroup_tg(cgrp->parent);
10221 tg = sched_create_group(parent);
10223 return ERR_PTR(-ENOMEM);
10229 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10231 struct task_group *tg = cgroup_tg(cgrp);
10233 sched_destroy_group(tg);
10237 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10238 struct task_struct *tsk)
10240 #ifdef CONFIG_RT_GROUP_SCHED
10241 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10244 /* We don't support RT-tasks being in separate groups */
10245 if (tsk->sched_class != &fair_sched_class)
10253 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10254 struct cgroup *old_cont, struct task_struct *tsk)
10256 sched_move_task(tsk);
10259 #ifdef CONFIG_FAIR_GROUP_SCHED
10260 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10263 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10266 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10268 struct task_group *tg = cgroup_tg(cgrp);
10270 return (u64) tg->shares;
10272 #endif /* CONFIG_FAIR_GROUP_SCHED */
10274 #ifdef CONFIG_RT_GROUP_SCHED
10275 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10278 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10281 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10283 return sched_group_rt_runtime(cgroup_tg(cgrp));
10286 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10289 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10292 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10294 return sched_group_rt_period(cgroup_tg(cgrp));
10296 #endif /* CONFIG_RT_GROUP_SCHED */
10298 static struct cftype cpu_files[] = {
10299 #ifdef CONFIG_FAIR_GROUP_SCHED
10302 .read_u64 = cpu_shares_read_u64,
10303 .write_u64 = cpu_shares_write_u64,
10306 #ifdef CONFIG_RT_GROUP_SCHED
10308 .name = "rt_runtime_us",
10309 .read_s64 = cpu_rt_runtime_read,
10310 .write_s64 = cpu_rt_runtime_write,
10313 .name = "rt_period_us",
10314 .read_u64 = cpu_rt_period_read_uint,
10315 .write_u64 = cpu_rt_period_write_uint,
10320 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10322 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10325 struct cgroup_subsys cpu_cgroup_subsys = {
10327 .create = cpu_cgroup_create,
10328 .destroy = cpu_cgroup_destroy,
10329 .can_attach = cpu_cgroup_can_attach,
10330 .attach = cpu_cgroup_attach,
10331 .populate = cpu_cgroup_populate,
10332 .subsys_id = cpu_cgroup_subsys_id,
10336 #endif /* CONFIG_CGROUP_SCHED */
10338 #ifdef CONFIG_CGROUP_CPUACCT
10341 * CPU accounting code for task groups.
10343 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10344 * (balbir@in.ibm.com).
10347 /* track cpu usage of a group of tasks and its child groups */
10349 struct cgroup_subsys_state css;
10350 /* cpuusage holds pointer to a u64-type object on every cpu */
10352 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10353 struct cpuacct *parent;
10356 struct cgroup_subsys cpuacct_subsys;
10358 /* return cpu accounting group corresponding to this container */
10359 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10361 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10362 struct cpuacct, css);
10365 /* return cpu accounting group to which this task belongs */
10366 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10368 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10369 struct cpuacct, css);
10372 /* create a new cpu accounting group */
10373 static struct cgroup_subsys_state *cpuacct_create(
10374 struct cgroup_subsys *ss, struct cgroup *cgrp)
10376 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10382 ca->cpuusage = alloc_percpu(u64);
10386 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10387 if (percpu_counter_init(&ca->cpustat[i], 0))
10388 goto out_free_counters;
10391 ca->parent = cgroup_ca(cgrp->parent);
10397 percpu_counter_destroy(&ca->cpustat[i]);
10398 free_percpu(ca->cpuusage);
10402 return ERR_PTR(-ENOMEM);
10405 /* destroy an existing cpu accounting group */
10407 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10409 struct cpuacct *ca = cgroup_ca(cgrp);
10412 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10413 percpu_counter_destroy(&ca->cpustat[i]);
10414 free_percpu(ca->cpuusage);
10418 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10420 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10423 #ifndef CONFIG_64BIT
10425 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10427 spin_lock_irq(&cpu_rq(cpu)->lock);
10429 spin_unlock_irq(&cpu_rq(cpu)->lock);
10437 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10439 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10441 #ifndef CONFIG_64BIT
10443 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10445 spin_lock_irq(&cpu_rq(cpu)->lock);
10447 spin_unlock_irq(&cpu_rq(cpu)->lock);
10453 /* return total cpu usage (in nanoseconds) of a group */
10454 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10456 struct cpuacct *ca = cgroup_ca(cgrp);
10457 u64 totalcpuusage = 0;
10460 for_each_present_cpu(i)
10461 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10463 return totalcpuusage;
10466 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10469 struct cpuacct *ca = cgroup_ca(cgrp);
10478 for_each_present_cpu(i)
10479 cpuacct_cpuusage_write(ca, i, 0);
10485 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10486 struct seq_file *m)
10488 struct cpuacct *ca = cgroup_ca(cgroup);
10492 for_each_present_cpu(i) {
10493 percpu = cpuacct_cpuusage_read(ca, i);
10494 seq_printf(m, "%llu ", (unsigned long long) percpu);
10496 seq_printf(m, "\n");
10500 static const char *cpuacct_stat_desc[] = {
10501 [CPUACCT_STAT_USER] = "user",
10502 [CPUACCT_STAT_SYSTEM] = "system",
10505 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10506 struct cgroup_map_cb *cb)
10508 struct cpuacct *ca = cgroup_ca(cgrp);
10511 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10512 s64 val = percpu_counter_read(&ca->cpustat[i]);
10513 val = cputime64_to_clock_t(val);
10514 cb->fill(cb, cpuacct_stat_desc[i], val);
10519 static struct cftype files[] = {
10522 .read_u64 = cpuusage_read,
10523 .write_u64 = cpuusage_write,
10526 .name = "usage_percpu",
10527 .read_seq_string = cpuacct_percpu_seq_read,
10531 .read_map = cpuacct_stats_show,
10535 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10537 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10541 * charge this task's execution time to its accounting group.
10543 * called with rq->lock held.
10545 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10547 struct cpuacct *ca;
10550 if (unlikely(!cpuacct_subsys.active))
10553 cpu = task_cpu(tsk);
10559 for (; ca; ca = ca->parent) {
10560 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10561 *cpuusage += cputime;
10568 * Charge the system/user time to the task's accounting group.
10570 static void cpuacct_update_stats(struct task_struct *tsk,
10571 enum cpuacct_stat_index idx, cputime_t val)
10573 struct cpuacct *ca;
10575 if (unlikely(!cpuacct_subsys.active))
10582 percpu_counter_add(&ca->cpustat[idx], val);
10588 struct cgroup_subsys cpuacct_subsys = {
10590 .create = cpuacct_create,
10591 .destroy = cpuacct_destroy,
10592 .populate = cpuacct_populate,
10593 .subsys_id = cpuacct_subsys_id,
10595 #endif /* CONFIG_CGROUP_CPUACCT */