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;
496 unsigned long rt_nr_total;
498 struct plist_head pushable_tasks;
503 /* Nests inside the rq lock: */
504 spinlock_t rt_runtime_lock;
506 #ifdef CONFIG_RT_GROUP_SCHED
507 unsigned long rt_nr_boosted;
510 struct list_head leaf_rt_rq_list;
511 struct task_group *tg;
512 struct sched_rt_entity *rt_se;
519 * We add the notion of a root-domain which will be used to define per-domain
520 * variables. Each exclusive cpuset essentially defines an island domain by
521 * fully partitioning the member cpus from any other cpuset. Whenever a new
522 * exclusive cpuset is created, we also create and attach a new root-domain
529 cpumask_var_t online;
532 * The "RT overload" flag: it gets set if a CPU has more than
533 * one runnable RT task.
535 cpumask_var_t rto_mask;
538 struct cpupri cpupri;
540 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
542 * Preferred wake up cpu nominated by sched_mc balance that will be
543 * used when most cpus are idle in the system indicating overall very
544 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
546 unsigned int sched_mc_preferred_wakeup_cpu;
551 * By default the system creates a single root-domain with all cpus as
552 * members (mimicking the global state we have today).
554 static struct root_domain def_root_domain;
559 * This is the main, per-CPU runqueue data structure.
561 * Locking rule: those places that want to lock multiple runqueues
562 * (such as the load balancing or the thread migration code), lock
563 * acquire operations must be ordered by ascending &runqueue.
570 * nr_running and cpu_load should be in the same cacheline because
571 * remote CPUs use both these fields when doing load calculation.
573 unsigned long nr_running;
574 #define CPU_LOAD_IDX_MAX 5
575 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
577 unsigned long last_tick_seen;
578 unsigned char in_nohz_recently;
580 /* capture load from *all* tasks on this cpu: */
581 struct load_weight load;
582 unsigned long nr_load_updates;
584 u64 nr_migrations_in;
589 #ifdef CONFIG_FAIR_GROUP_SCHED
590 /* list of leaf cfs_rq on this cpu: */
591 struct list_head leaf_cfs_rq_list;
593 #ifdef CONFIG_RT_GROUP_SCHED
594 struct list_head leaf_rt_rq_list;
598 * This is part of a global counter where only the total sum
599 * over all CPUs matters. A task can increase this counter on
600 * one CPU and if it got migrated afterwards it may decrease
601 * it on another CPU. Always updated under the runqueue lock:
603 unsigned long nr_uninterruptible;
605 struct task_struct *curr, *idle;
606 unsigned long next_balance;
607 struct mm_struct *prev_mm;
614 struct root_domain *rd;
615 struct sched_domain *sd;
617 unsigned char idle_at_tick;
618 /* For active balancing */
621 /* cpu of this runqueue: */
625 unsigned long avg_load_per_task;
627 struct task_struct *migration_thread;
628 struct list_head migration_queue;
631 /* calc_load related fields */
632 unsigned long calc_load_update;
633 long calc_load_active;
635 #ifdef CONFIG_SCHED_HRTICK
637 int hrtick_csd_pending;
638 struct call_single_data hrtick_csd;
640 struct hrtimer hrtick_timer;
643 #ifdef CONFIG_SCHEDSTATS
645 struct sched_info rq_sched_info;
646 unsigned long long rq_cpu_time;
647 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
649 /* sys_sched_yield() stats */
650 unsigned int yld_count;
652 /* schedule() stats */
653 unsigned int sched_switch;
654 unsigned int sched_count;
655 unsigned int sched_goidle;
657 /* try_to_wake_up() stats */
658 unsigned int ttwu_count;
659 unsigned int ttwu_local;
662 unsigned int bkl_count;
666 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
668 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
670 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
673 static inline int cpu_of(struct rq *rq)
683 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
684 * See detach_destroy_domains: synchronize_sched for details.
686 * The domain tree of any CPU may only be accessed from within
687 * preempt-disabled sections.
689 #define for_each_domain(cpu, __sd) \
690 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
692 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
693 #define this_rq() (&__get_cpu_var(runqueues))
694 #define task_rq(p) cpu_rq(task_cpu(p))
695 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
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;
2577 p->se.wait_count = 0;
2580 p->se.sleep_start = 0;
2581 p->se.sleep_max = 0;
2582 p->se.sum_sleep_runtime = 0;
2584 p->se.block_start = 0;
2585 p->se.block_max = 0;
2587 p->se.slice_max = 0;
2589 p->se.nr_migrations_cold = 0;
2590 p->se.nr_failed_migrations_affine = 0;
2591 p->se.nr_failed_migrations_running = 0;
2592 p->se.nr_failed_migrations_hot = 0;
2593 p->se.nr_forced_migrations = 0;
2594 p->se.nr_forced2_migrations = 0;
2596 p->se.nr_wakeups = 0;
2597 p->se.nr_wakeups_sync = 0;
2598 p->se.nr_wakeups_migrate = 0;
2599 p->se.nr_wakeups_local = 0;
2600 p->se.nr_wakeups_remote = 0;
2601 p->se.nr_wakeups_affine = 0;
2602 p->se.nr_wakeups_affine_attempts = 0;
2603 p->se.nr_wakeups_passive = 0;
2604 p->se.nr_wakeups_idle = 0;
2608 INIT_LIST_HEAD(&p->rt.run_list);
2610 INIT_LIST_HEAD(&p->se.group_node);
2612 #ifdef CONFIG_PREEMPT_NOTIFIERS
2613 INIT_HLIST_HEAD(&p->preempt_notifiers);
2617 * We mark the process as running here, but have not actually
2618 * inserted it onto the runqueue yet. This guarantees that
2619 * nobody will actually run it, and a signal or other external
2620 * event cannot wake it up and insert it on the runqueue either.
2622 p->state = TASK_RUNNING;
2626 * fork()/clone()-time setup:
2628 void sched_fork(struct task_struct *p, int clone_flags)
2630 int cpu = get_cpu();
2635 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2637 set_task_cpu(p, cpu);
2640 * Make sure we do not leak PI boosting priority to the child:
2642 p->prio = current->normal_prio;
2643 if (!rt_prio(p->prio))
2644 p->sched_class = &fair_sched_class;
2646 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2647 if (likely(sched_info_on()))
2648 memset(&p->sched_info, 0, sizeof(p->sched_info));
2650 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2653 #ifdef CONFIG_PREEMPT
2654 /* Want to start with kernel preemption disabled. */
2655 task_thread_info(p)->preempt_count = 1;
2657 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2663 * wake_up_new_task - wake up a newly created task for the first time.
2665 * This function will do some initial scheduler statistics housekeeping
2666 * that must be done for every newly created context, then puts the task
2667 * on the runqueue and wakes it.
2669 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2671 unsigned long flags;
2674 rq = task_rq_lock(p, &flags);
2675 BUG_ON(p->state != TASK_RUNNING);
2676 update_rq_clock(rq);
2678 p->prio = effective_prio(p);
2680 if (!p->sched_class->task_new || !current->se.on_rq) {
2681 activate_task(rq, p, 0);
2684 * Let the scheduling class do new task startup
2685 * management (if any):
2687 p->sched_class->task_new(rq, p);
2690 trace_sched_wakeup_new(rq, p, 1);
2691 check_preempt_curr(rq, p, 0);
2693 if (p->sched_class->task_wake_up)
2694 p->sched_class->task_wake_up(rq, p);
2696 task_rq_unlock(rq, &flags);
2699 #ifdef CONFIG_PREEMPT_NOTIFIERS
2702 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2703 * @notifier: notifier struct to register
2705 void preempt_notifier_register(struct preempt_notifier *notifier)
2707 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2709 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2712 * preempt_notifier_unregister - no longer interested in preemption notifications
2713 * @notifier: notifier struct to unregister
2715 * This is safe to call from within a preemption notifier.
2717 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2719 hlist_del(¬ifier->link);
2721 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2723 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2725 struct preempt_notifier *notifier;
2726 struct hlist_node *node;
2728 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2729 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2733 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2734 struct task_struct *next)
2736 struct preempt_notifier *notifier;
2737 struct hlist_node *node;
2739 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2740 notifier->ops->sched_out(notifier, next);
2743 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2745 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2750 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2751 struct task_struct *next)
2755 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2758 * prepare_task_switch - prepare to switch tasks
2759 * @rq: the runqueue preparing to switch
2760 * @prev: the current task that is being switched out
2761 * @next: the task we are going to switch to.
2763 * This is called with the rq lock held and interrupts off. It must
2764 * be paired with a subsequent finish_task_switch after the context
2767 * prepare_task_switch sets up locking and calls architecture specific
2771 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2772 struct task_struct *next)
2774 fire_sched_out_preempt_notifiers(prev, next);
2775 prepare_lock_switch(rq, next);
2776 prepare_arch_switch(next);
2780 * finish_task_switch - clean up after a task-switch
2781 * @rq: runqueue associated with task-switch
2782 * @prev: the thread we just switched away from.
2784 * finish_task_switch must be called after the context switch, paired
2785 * with a prepare_task_switch call before the context switch.
2786 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2787 * and do any other architecture-specific cleanup actions.
2789 * Note that we may have delayed dropping an mm in context_switch(). If
2790 * so, we finish that here outside of the runqueue lock. (Doing it
2791 * with the lock held can cause deadlocks; see schedule() for
2794 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2795 __releases(rq->lock)
2797 struct mm_struct *mm = rq->prev_mm;
2800 int post_schedule = 0;
2802 if (current->sched_class->needs_post_schedule)
2803 post_schedule = current->sched_class->needs_post_schedule(rq);
2809 * A task struct has one reference for the use as "current".
2810 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2811 * schedule one last time. The schedule call will never return, and
2812 * the scheduled task must drop that reference.
2813 * The test for TASK_DEAD must occur while the runqueue locks are
2814 * still held, otherwise prev could be scheduled on another cpu, die
2815 * there before we look at prev->state, and then the reference would
2817 * Manfred Spraul <manfred@colorfullife.com>
2819 prev_state = prev->state;
2820 finish_arch_switch(prev);
2821 perf_counter_task_sched_in(current, cpu_of(rq));
2822 finish_lock_switch(rq, prev);
2825 current->sched_class->post_schedule(rq);
2828 fire_sched_in_preempt_notifiers(current);
2831 if (unlikely(prev_state == TASK_DEAD)) {
2833 * Remove function-return probe instances associated with this
2834 * task and put them back on the free list.
2836 kprobe_flush_task(prev);
2837 put_task_struct(prev);
2842 * schedule_tail - first thing a freshly forked thread must call.
2843 * @prev: the thread we just switched away from.
2845 asmlinkage void schedule_tail(struct task_struct *prev)
2846 __releases(rq->lock)
2848 struct rq *rq = this_rq();
2850 finish_task_switch(rq, prev);
2851 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2852 /* In this case, finish_task_switch does not reenable preemption */
2855 if (current->set_child_tid)
2856 put_user(task_pid_vnr(current), current->set_child_tid);
2860 * context_switch - switch to the new MM and the new
2861 * thread's register state.
2864 context_switch(struct rq *rq, struct task_struct *prev,
2865 struct task_struct *next)
2867 struct mm_struct *mm, *oldmm;
2869 prepare_task_switch(rq, prev, next);
2870 trace_sched_switch(rq, prev, next);
2872 oldmm = prev->active_mm;
2874 * For paravirt, this is coupled with an exit in switch_to to
2875 * combine the page table reload and the switch backend into
2878 arch_start_context_switch(prev);
2880 if (unlikely(!mm)) {
2881 next->active_mm = oldmm;
2882 atomic_inc(&oldmm->mm_count);
2883 enter_lazy_tlb(oldmm, next);
2885 switch_mm(oldmm, mm, next);
2887 if (unlikely(!prev->mm)) {
2888 prev->active_mm = NULL;
2889 rq->prev_mm = oldmm;
2892 * Since the runqueue lock will be released by the next
2893 * task (which is an invalid locking op but in the case
2894 * of the scheduler it's an obvious special-case), so we
2895 * do an early lockdep release here:
2897 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2898 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2901 /* Here we just switch the register state and the stack. */
2902 switch_to(prev, next, prev);
2906 * this_rq must be evaluated again because prev may have moved
2907 * CPUs since it called schedule(), thus the 'rq' on its stack
2908 * frame will be invalid.
2910 finish_task_switch(this_rq(), prev);
2914 * nr_running, nr_uninterruptible and nr_context_switches:
2916 * externally visible scheduler statistics: current number of runnable
2917 * threads, current number of uninterruptible-sleeping threads, total
2918 * number of context switches performed since bootup.
2920 unsigned long nr_running(void)
2922 unsigned long i, sum = 0;
2924 for_each_online_cpu(i)
2925 sum += cpu_rq(i)->nr_running;
2930 unsigned long nr_uninterruptible(void)
2932 unsigned long i, sum = 0;
2934 for_each_possible_cpu(i)
2935 sum += cpu_rq(i)->nr_uninterruptible;
2938 * Since we read the counters lockless, it might be slightly
2939 * inaccurate. Do not allow it to go below zero though:
2941 if (unlikely((long)sum < 0))
2947 unsigned long long nr_context_switches(void)
2950 unsigned long long sum = 0;
2952 for_each_possible_cpu(i)
2953 sum += cpu_rq(i)->nr_switches;
2958 unsigned long nr_iowait(void)
2960 unsigned long i, sum = 0;
2962 for_each_possible_cpu(i)
2963 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2968 /* Variables and functions for calc_load */
2969 static atomic_long_t calc_load_tasks;
2970 static unsigned long calc_load_update;
2971 unsigned long avenrun[3];
2972 EXPORT_SYMBOL(avenrun);
2975 * get_avenrun - get the load average array
2976 * @loads: pointer to dest load array
2977 * @offset: offset to add
2978 * @shift: shift count to shift the result left
2980 * These values are estimates at best, so no need for locking.
2982 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2984 loads[0] = (avenrun[0] + offset) << shift;
2985 loads[1] = (avenrun[1] + offset) << shift;
2986 loads[2] = (avenrun[2] + offset) << shift;
2989 static unsigned long
2990 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2993 load += active * (FIXED_1 - exp);
2994 return load >> FSHIFT;
2998 * calc_load - update the avenrun load estimates 10 ticks after the
2999 * CPUs have updated calc_load_tasks.
3001 void calc_global_load(void)
3003 unsigned long upd = calc_load_update + 10;
3006 if (time_before(jiffies, upd))
3009 active = atomic_long_read(&calc_load_tasks);
3010 active = active > 0 ? active * FIXED_1 : 0;
3012 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3013 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3014 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3016 calc_load_update += LOAD_FREQ;
3020 * Either called from update_cpu_load() or from a cpu going idle
3022 static void calc_load_account_active(struct rq *this_rq)
3024 long nr_active, delta;
3026 nr_active = this_rq->nr_running;
3027 nr_active += (long) this_rq->nr_uninterruptible;
3029 if (nr_active != this_rq->calc_load_active) {
3030 delta = nr_active - this_rq->calc_load_active;
3031 this_rq->calc_load_active = nr_active;
3032 atomic_long_add(delta, &calc_load_tasks);
3037 * Externally visible per-cpu scheduler statistics:
3038 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3040 u64 cpu_nr_migrations(int cpu)
3042 return cpu_rq(cpu)->nr_migrations_in;
3046 * Update rq->cpu_load[] statistics. This function is usually called every
3047 * scheduler tick (TICK_NSEC).
3049 static void update_cpu_load(struct rq *this_rq)
3051 unsigned long this_load = this_rq->load.weight;
3054 this_rq->nr_load_updates++;
3056 /* Update our load: */
3057 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3058 unsigned long old_load, new_load;
3060 /* scale is effectively 1 << i now, and >> i divides by scale */
3062 old_load = this_rq->cpu_load[i];
3063 new_load = this_load;
3065 * Round up the averaging division if load is increasing. This
3066 * prevents us from getting stuck on 9 if the load is 10, for
3069 if (new_load > old_load)
3070 new_load += scale-1;
3071 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3074 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3075 this_rq->calc_load_update += LOAD_FREQ;
3076 calc_load_account_active(this_rq);
3083 * double_rq_lock - safely lock two runqueues
3085 * Note this does not disable interrupts like task_rq_lock,
3086 * you need to do so manually before calling.
3088 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3089 __acquires(rq1->lock)
3090 __acquires(rq2->lock)
3092 BUG_ON(!irqs_disabled());
3094 spin_lock(&rq1->lock);
3095 __acquire(rq2->lock); /* Fake it out ;) */
3098 spin_lock(&rq1->lock);
3099 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3101 spin_lock(&rq2->lock);
3102 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3105 update_rq_clock(rq1);
3106 update_rq_clock(rq2);
3110 * double_rq_unlock - safely unlock two runqueues
3112 * Note this does not restore interrupts like task_rq_unlock,
3113 * you need to do so manually after calling.
3115 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3116 __releases(rq1->lock)
3117 __releases(rq2->lock)
3119 spin_unlock(&rq1->lock);
3121 spin_unlock(&rq2->lock);
3123 __release(rq2->lock);
3127 * If dest_cpu is allowed for this process, migrate the task to it.
3128 * This is accomplished by forcing the cpu_allowed mask to only
3129 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3130 * the cpu_allowed mask is restored.
3132 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3134 struct migration_req req;
3135 unsigned long flags;
3138 rq = task_rq_lock(p, &flags);
3139 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3140 || unlikely(!cpu_active(dest_cpu)))
3143 /* force the process onto the specified CPU */
3144 if (migrate_task(p, dest_cpu, &req)) {
3145 /* Need to wait for migration thread (might exit: take ref). */
3146 struct task_struct *mt = rq->migration_thread;
3148 get_task_struct(mt);
3149 task_rq_unlock(rq, &flags);
3150 wake_up_process(mt);
3151 put_task_struct(mt);
3152 wait_for_completion(&req.done);
3157 task_rq_unlock(rq, &flags);
3161 * sched_exec - execve() is a valuable balancing opportunity, because at
3162 * this point the task has the smallest effective memory and cache footprint.
3164 void sched_exec(void)
3166 int new_cpu, this_cpu = get_cpu();
3167 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3169 if (new_cpu != this_cpu)
3170 sched_migrate_task(current, new_cpu);
3174 * pull_task - move a task from a remote runqueue to the local runqueue.
3175 * Both runqueues must be locked.
3177 static void pull_task(struct rq *src_rq, struct task_struct *p,
3178 struct rq *this_rq, int this_cpu)
3180 deactivate_task(src_rq, p, 0);
3181 set_task_cpu(p, this_cpu);
3182 activate_task(this_rq, p, 0);
3184 * Note that idle threads have a prio of MAX_PRIO, for this test
3185 * to be always true for them.
3187 check_preempt_curr(this_rq, p, 0);
3191 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3194 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3195 struct sched_domain *sd, enum cpu_idle_type idle,
3198 int tsk_cache_hot = 0;
3200 * We do not migrate tasks that are:
3201 * 1) running (obviously), or
3202 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3203 * 3) are cache-hot on their current CPU.
3205 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3206 schedstat_inc(p, se.nr_failed_migrations_affine);
3211 if (task_running(rq, p)) {
3212 schedstat_inc(p, se.nr_failed_migrations_running);
3217 * Aggressive migration if:
3218 * 1) task is cache cold, or
3219 * 2) too many balance attempts have failed.
3222 tsk_cache_hot = task_hot(p, rq->clock, sd);
3223 if (!tsk_cache_hot ||
3224 sd->nr_balance_failed > sd->cache_nice_tries) {
3225 #ifdef CONFIG_SCHEDSTATS
3226 if (tsk_cache_hot) {
3227 schedstat_inc(sd, lb_hot_gained[idle]);
3228 schedstat_inc(p, se.nr_forced_migrations);
3234 if (tsk_cache_hot) {
3235 schedstat_inc(p, se.nr_failed_migrations_hot);
3241 static unsigned long
3242 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3243 unsigned long max_load_move, struct sched_domain *sd,
3244 enum cpu_idle_type idle, int *all_pinned,
3245 int *this_best_prio, struct rq_iterator *iterator)
3247 int loops = 0, pulled = 0, pinned = 0;
3248 struct task_struct *p;
3249 long rem_load_move = max_load_move;
3251 if (max_load_move == 0)
3257 * Start the load-balancing iterator:
3259 p = iterator->start(iterator->arg);
3261 if (!p || loops++ > sysctl_sched_nr_migrate)
3264 if ((p->se.load.weight >> 1) > rem_load_move ||
3265 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3266 p = iterator->next(iterator->arg);
3270 pull_task(busiest, p, this_rq, this_cpu);
3272 rem_load_move -= p->se.load.weight;
3274 #ifdef CONFIG_PREEMPT
3276 * NEWIDLE balancing is a source of latency, so preemptible kernels
3277 * will stop after the first task is pulled to minimize the critical
3280 if (idle == CPU_NEWLY_IDLE)
3285 * We only want to steal up to the prescribed amount of weighted load.
3287 if (rem_load_move > 0) {
3288 if (p->prio < *this_best_prio)
3289 *this_best_prio = p->prio;
3290 p = iterator->next(iterator->arg);
3295 * Right now, this is one of only two places pull_task() is called,
3296 * so we can safely collect pull_task() stats here rather than
3297 * inside pull_task().
3299 schedstat_add(sd, lb_gained[idle], pulled);
3302 *all_pinned = pinned;
3304 return max_load_move - rem_load_move;
3308 * move_tasks tries to move up to max_load_move weighted load from busiest to
3309 * this_rq, as part of a balancing operation within domain "sd".
3310 * Returns 1 if successful and 0 otherwise.
3312 * Called with both runqueues locked.
3314 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3315 unsigned long max_load_move,
3316 struct sched_domain *sd, enum cpu_idle_type idle,
3319 const struct sched_class *class = sched_class_highest;
3320 unsigned long total_load_moved = 0;
3321 int this_best_prio = this_rq->curr->prio;
3325 class->load_balance(this_rq, this_cpu, busiest,
3326 max_load_move - total_load_moved,
3327 sd, idle, all_pinned, &this_best_prio);
3328 class = class->next;
3330 #ifdef CONFIG_PREEMPT
3332 * NEWIDLE balancing is a source of latency, so preemptible
3333 * kernels will stop after the first task is pulled to minimize
3334 * the critical section.
3336 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3339 } while (class && max_load_move > total_load_moved);
3341 return total_load_moved > 0;
3345 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3346 struct sched_domain *sd, enum cpu_idle_type idle,
3347 struct rq_iterator *iterator)
3349 struct task_struct *p = iterator->start(iterator->arg);
3353 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3354 pull_task(busiest, p, this_rq, this_cpu);
3356 * Right now, this is only the second place pull_task()
3357 * is called, so we can safely collect pull_task()
3358 * stats here rather than inside pull_task().
3360 schedstat_inc(sd, lb_gained[idle]);
3364 p = iterator->next(iterator->arg);
3371 * move_one_task tries to move exactly one task from busiest to this_rq, as
3372 * part of active balancing operations within "domain".
3373 * Returns 1 if successful and 0 otherwise.
3375 * Called with both runqueues locked.
3377 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3378 struct sched_domain *sd, enum cpu_idle_type idle)
3380 const struct sched_class *class;
3382 for (class = sched_class_highest; class; class = class->next)
3383 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3388 /********** Helpers for find_busiest_group ************************/
3390 * sd_lb_stats - Structure to store the statistics of a sched_domain
3391 * during load balancing.
3393 struct sd_lb_stats {
3394 struct sched_group *busiest; /* Busiest group in this sd */
3395 struct sched_group *this; /* Local group in this sd */
3396 unsigned long total_load; /* Total load of all groups in sd */
3397 unsigned long total_pwr; /* Total power of all groups in sd */
3398 unsigned long avg_load; /* Average load across all groups in sd */
3400 /** Statistics of this group */
3401 unsigned long this_load;
3402 unsigned long this_load_per_task;
3403 unsigned long this_nr_running;
3405 /* Statistics of the busiest group */
3406 unsigned long max_load;
3407 unsigned long busiest_load_per_task;
3408 unsigned long busiest_nr_running;
3410 int group_imb; /* Is there imbalance in this sd */
3411 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3412 int power_savings_balance; /* Is powersave balance needed for this sd */
3413 struct sched_group *group_min; /* Least loaded group in sd */
3414 struct sched_group *group_leader; /* Group which relieves group_min */
3415 unsigned long min_load_per_task; /* load_per_task in group_min */
3416 unsigned long leader_nr_running; /* Nr running of group_leader */
3417 unsigned long min_nr_running; /* Nr running of group_min */
3422 * sg_lb_stats - stats of a sched_group required for load_balancing
3424 struct sg_lb_stats {
3425 unsigned long avg_load; /*Avg load across the CPUs of the group */
3426 unsigned long group_load; /* Total load over the CPUs of the group */
3427 unsigned long sum_nr_running; /* Nr tasks running in the group */
3428 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3429 unsigned long group_capacity;
3430 int group_imb; /* Is there an imbalance in the group ? */
3434 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3435 * @group: The group whose first cpu is to be returned.
3437 static inline unsigned int group_first_cpu(struct sched_group *group)
3439 return cpumask_first(sched_group_cpus(group));
3443 * get_sd_load_idx - Obtain the load index for a given sched domain.
3444 * @sd: The sched_domain whose load_idx is to be obtained.
3445 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3447 static inline int get_sd_load_idx(struct sched_domain *sd,
3448 enum cpu_idle_type idle)
3454 load_idx = sd->busy_idx;
3457 case CPU_NEWLY_IDLE:
3458 load_idx = sd->newidle_idx;
3461 load_idx = sd->idle_idx;
3469 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3471 * init_sd_power_savings_stats - Initialize power savings statistics for
3472 * the given sched_domain, during load balancing.
3474 * @sd: Sched domain whose power-savings statistics are to be initialized.
3475 * @sds: Variable containing the statistics for sd.
3476 * @idle: Idle status of the CPU at which we're performing load-balancing.
3478 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3479 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3482 * Busy processors will not participate in power savings
3485 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3486 sds->power_savings_balance = 0;
3488 sds->power_savings_balance = 1;
3489 sds->min_nr_running = ULONG_MAX;
3490 sds->leader_nr_running = 0;
3495 * update_sd_power_savings_stats - Update the power saving stats for a
3496 * sched_domain while performing load balancing.
3498 * @group: sched_group belonging to the sched_domain under consideration.
3499 * @sds: Variable containing the statistics of the sched_domain
3500 * @local_group: Does group contain the CPU for which we're performing
3502 * @sgs: Variable containing the statistics of the group.
3504 static inline void update_sd_power_savings_stats(struct sched_group *group,
3505 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3508 if (!sds->power_savings_balance)
3512 * If the local group is idle or completely loaded
3513 * no need to do power savings balance at this domain
3515 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3516 !sds->this_nr_running))
3517 sds->power_savings_balance = 0;
3520 * If a group is already running at full capacity or idle,
3521 * don't include that group in power savings calculations
3523 if (!sds->power_savings_balance ||
3524 sgs->sum_nr_running >= sgs->group_capacity ||
3525 !sgs->sum_nr_running)
3529 * Calculate the group which has the least non-idle load.
3530 * This is the group from where we need to pick up the load
3533 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3534 (sgs->sum_nr_running == sds->min_nr_running &&
3535 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3536 sds->group_min = group;
3537 sds->min_nr_running = sgs->sum_nr_running;
3538 sds->min_load_per_task = sgs->sum_weighted_load /
3539 sgs->sum_nr_running;
3543 * Calculate the group which is almost near its
3544 * capacity but still has some space to pick up some load
3545 * from other group and save more power
3547 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3550 if (sgs->sum_nr_running > sds->leader_nr_running ||
3551 (sgs->sum_nr_running == sds->leader_nr_running &&
3552 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3553 sds->group_leader = group;
3554 sds->leader_nr_running = sgs->sum_nr_running;
3559 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3560 * @sds: Variable containing the statistics of the sched_domain
3561 * under consideration.
3562 * @this_cpu: Cpu at which we're currently performing load-balancing.
3563 * @imbalance: Variable to store the imbalance.
3566 * Check if we have potential to perform some power-savings balance.
3567 * If yes, set the busiest group to be the least loaded group in the
3568 * sched_domain, so that it's CPUs can be put to idle.
3570 * Returns 1 if there is potential to perform power-savings balance.
3573 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3574 int this_cpu, unsigned long *imbalance)
3576 if (!sds->power_savings_balance)
3579 if (sds->this != sds->group_leader ||
3580 sds->group_leader == sds->group_min)
3583 *imbalance = sds->min_load_per_task;
3584 sds->busiest = sds->group_min;
3586 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3587 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3588 group_first_cpu(sds->group_leader);
3594 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3595 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3596 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3601 static inline void update_sd_power_savings_stats(struct sched_group *group,
3602 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3607 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3608 int this_cpu, unsigned long *imbalance)
3612 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3616 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3617 * @group: sched_group whose statistics are to be updated.
3618 * @this_cpu: Cpu for which load balance is currently performed.
3619 * @idle: Idle status of this_cpu
3620 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3621 * @sd_idle: Idle status of the sched_domain containing group.
3622 * @local_group: Does group contain this_cpu.
3623 * @cpus: Set of cpus considered for load balancing.
3624 * @balance: Should we balance.
3625 * @sgs: variable to hold the statistics for this group.
3627 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3628 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3629 int local_group, const struct cpumask *cpus,
3630 int *balance, struct sg_lb_stats *sgs)
3632 unsigned long load, max_cpu_load, min_cpu_load;
3634 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3635 unsigned long sum_avg_load_per_task;
3636 unsigned long avg_load_per_task;
3639 balance_cpu = group_first_cpu(group);
3641 /* Tally up the load of all CPUs in the group */
3642 sum_avg_load_per_task = avg_load_per_task = 0;
3644 min_cpu_load = ~0UL;
3646 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3647 struct rq *rq = cpu_rq(i);
3649 if (*sd_idle && rq->nr_running)
3652 /* Bias balancing toward cpus of our domain */
3654 if (idle_cpu(i) && !first_idle_cpu) {
3659 load = target_load(i, load_idx);
3661 load = source_load(i, load_idx);
3662 if (load > max_cpu_load)
3663 max_cpu_load = load;
3664 if (min_cpu_load > load)
3665 min_cpu_load = load;
3668 sgs->group_load += load;
3669 sgs->sum_nr_running += rq->nr_running;
3670 sgs->sum_weighted_load += weighted_cpuload(i);
3672 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3676 * First idle cpu or the first cpu(busiest) in this sched group
3677 * is eligible for doing load balancing at this and above
3678 * domains. In the newly idle case, we will allow all the cpu's
3679 * to do the newly idle load balance.
3681 if (idle != CPU_NEWLY_IDLE && local_group &&
3682 balance_cpu != this_cpu && balance) {
3687 /* Adjust by relative CPU power of the group */
3688 sgs->avg_load = sg_div_cpu_power(group,
3689 sgs->group_load * SCHED_LOAD_SCALE);
3693 * Consider the group unbalanced when the imbalance is larger
3694 * than the average weight of two tasks.
3696 * APZ: with cgroup the avg task weight can vary wildly and
3697 * might not be a suitable number - should we keep a
3698 * normalized nr_running number somewhere that negates
3701 avg_load_per_task = sg_div_cpu_power(group,
3702 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3704 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3707 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3712 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3713 * @sd: sched_domain whose statistics are to be updated.
3714 * @this_cpu: Cpu for which load balance is currently performed.
3715 * @idle: Idle status of this_cpu
3716 * @sd_idle: Idle status of the sched_domain containing group.
3717 * @cpus: Set of cpus considered for load balancing.
3718 * @balance: Should we balance.
3719 * @sds: variable to hold the statistics for this sched_domain.
3721 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3722 enum cpu_idle_type idle, int *sd_idle,
3723 const struct cpumask *cpus, int *balance,
3724 struct sd_lb_stats *sds)
3726 struct sched_group *group = sd->groups;
3727 struct sg_lb_stats sgs;
3730 init_sd_power_savings_stats(sd, sds, idle);
3731 load_idx = get_sd_load_idx(sd, idle);
3736 local_group = cpumask_test_cpu(this_cpu,
3737 sched_group_cpus(group));
3738 memset(&sgs, 0, sizeof(sgs));
3739 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3740 local_group, cpus, balance, &sgs);
3742 if (local_group && balance && !(*balance))
3745 sds->total_load += sgs.group_load;
3746 sds->total_pwr += group->__cpu_power;
3749 sds->this_load = sgs.avg_load;
3751 sds->this_nr_running = sgs.sum_nr_running;
3752 sds->this_load_per_task = sgs.sum_weighted_load;
3753 } else if (sgs.avg_load > sds->max_load &&
3754 (sgs.sum_nr_running > sgs.group_capacity ||
3756 sds->max_load = sgs.avg_load;
3757 sds->busiest = group;
3758 sds->busiest_nr_running = sgs.sum_nr_running;
3759 sds->busiest_load_per_task = sgs.sum_weighted_load;
3760 sds->group_imb = sgs.group_imb;
3763 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3764 group = group->next;
3765 } while (group != sd->groups);
3770 * fix_small_imbalance - Calculate the minor imbalance that exists
3771 * amongst the groups of a sched_domain, during
3773 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3774 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3775 * @imbalance: Variable to store the imbalance.
3777 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3778 int this_cpu, unsigned long *imbalance)
3780 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3781 unsigned int imbn = 2;
3783 if (sds->this_nr_running) {
3784 sds->this_load_per_task /= sds->this_nr_running;
3785 if (sds->busiest_load_per_task >
3786 sds->this_load_per_task)
3789 sds->this_load_per_task =
3790 cpu_avg_load_per_task(this_cpu);
3792 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3793 sds->busiest_load_per_task * imbn) {
3794 *imbalance = sds->busiest_load_per_task;
3799 * OK, we don't have enough imbalance to justify moving tasks,
3800 * however we may be able to increase total CPU power used by
3804 pwr_now += sds->busiest->__cpu_power *
3805 min(sds->busiest_load_per_task, sds->max_load);
3806 pwr_now += sds->this->__cpu_power *
3807 min(sds->this_load_per_task, sds->this_load);
3808 pwr_now /= SCHED_LOAD_SCALE;
3810 /* Amount of load we'd subtract */
3811 tmp = sg_div_cpu_power(sds->busiest,
3812 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3813 if (sds->max_load > tmp)
3814 pwr_move += sds->busiest->__cpu_power *
3815 min(sds->busiest_load_per_task, sds->max_load - tmp);
3817 /* Amount of load we'd add */
3818 if (sds->max_load * sds->busiest->__cpu_power <
3819 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3820 tmp = sg_div_cpu_power(sds->this,
3821 sds->max_load * sds->busiest->__cpu_power);
3823 tmp = sg_div_cpu_power(sds->this,
3824 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3825 pwr_move += sds->this->__cpu_power *
3826 min(sds->this_load_per_task, sds->this_load + tmp);
3827 pwr_move /= SCHED_LOAD_SCALE;
3829 /* Move if we gain throughput */
3830 if (pwr_move > pwr_now)
3831 *imbalance = sds->busiest_load_per_task;
3835 * calculate_imbalance - Calculate the amount of imbalance present within the
3836 * groups of a given sched_domain during load balance.
3837 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3838 * @this_cpu: Cpu for which currently load balance is being performed.
3839 * @imbalance: The variable to store the imbalance.
3841 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3842 unsigned long *imbalance)
3844 unsigned long max_pull;
3846 * In the presence of smp nice balancing, certain scenarios can have
3847 * max load less than avg load(as we skip the groups at or below
3848 * its cpu_power, while calculating max_load..)
3850 if (sds->max_load < sds->avg_load) {
3852 return fix_small_imbalance(sds, this_cpu, imbalance);
3855 /* Don't want to pull so many tasks that a group would go idle */
3856 max_pull = min(sds->max_load - sds->avg_load,
3857 sds->max_load - sds->busiest_load_per_task);
3859 /* How much load to actually move to equalise the imbalance */
3860 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3861 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3865 * if *imbalance is less than the average load per runnable task
3866 * there is no gaurantee that any tasks will be moved so we'll have
3867 * a think about bumping its value to force at least one task to be
3870 if (*imbalance < sds->busiest_load_per_task)
3871 return fix_small_imbalance(sds, this_cpu, imbalance);
3874 /******* find_busiest_group() helpers end here *********************/
3877 * find_busiest_group - Returns the busiest group within the sched_domain
3878 * if there is an imbalance. If there isn't an imbalance, and
3879 * the user has opted for power-savings, it returns a group whose
3880 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3881 * such a group exists.
3883 * Also calculates the amount of weighted load which should be moved
3884 * to restore balance.
3886 * @sd: The sched_domain whose busiest group is to be returned.
3887 * @this_cpu: The cpu for which load balancing is currently being performed.
3888 * @imbalance: Variable which stores amount of weighted load which should
3889 * be moved to restore balance/put a group to idle.
3890 * @idle: The idle status of this_cpu.
3891 * @sd_idle: The idleness of sd
3892 * @cpus: The set of CPUs under consideration for load-balancing.
3893 * @balance: Pointer to a variable indicating if this_cpu
3894 * is the appropriate cpu to perform load balancing at this_level.
3896 * Returns: - the busiest group if imbalance exists.
3897 * - If no imbalance and user has opted for power-savings balance,
3898 * return the least loaded group whose CPUs can be
3899 * put to idle by rebalancing its tasks onto our group.
3901 static struct sched_group *
3902 find_busiest_group(struct sched_domain *sd, int this_cpu,
3903 unsigned long *imbalance, enum cpu_idle_type idle,
3904 int *sd_idle, const struct cpumask *cpus, int *balance)
3906 struct sd_lb_stats sds;
3908 memset(&sds, 0, sizeof(sds));
3911 * Compute the various statistics relavent for load balancing at
3914 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3917 /* Cases where imbalance does not exist from POV of this_cpu */
3918 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3920 * 2) There is no busy sibling group to pull from.
3921 * 3) This group is the busiest group.
3922 * 4) This group is more busy than the avg busieness at this
3924 * 5) The imbalance is within the specified limit.
3925 * 6) Any rebalance would lead to ping-pong
3927 if (balance && !(*balance))
3930 if (!sds.busiest || sds.busiest_nr_running == 0)
3933 if (sds.this_load >= sds.max_load)
3936 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3938 if (sds.this_load >= sds.avg_load)
3941 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3944 sds.busiest_load_per_task /= sds.busiest_nr_running;
3946 sds.busiest_load_per_task =
3947 min(sds.busiest_load_per_task, sds.avg_load);
3950 * We're trying to get all the cpus to the average_load, so we don't
3951 * want to push ourselves above the average load, nor do we wish to
3952 * reduce the max loaded cpu below the average load, as either of these
3953 * actions would just result in more rebalancing later, and ping-pong
3954 * tasks around. Thus we look for the minimum possible imbalance.
3955 * Negative imbalances (*we* are more loaded than anyone else) will
3956 * be counted as no imbalance for these purposes -- we can't fix that
3957 * by pulling tasks to us. Be careful of negative numbers as they'll
3958 * appear as very large values with unsigned longs.
3960 if (sds.max_load <= sds.busiest_load_per_task)
3963 /* Looks like there is an imbalance. Compute it */
3964 calculate_imbalance(&sds, this_cpu, imbalance);
3969 * There is no obvious imbalance. But check if we can do some balancing
3972 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3980 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3983 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3984 unsigned long imbalance, const struct cpumask *cpus)
3986 struct rq *busiest = NULL, *rq;
3987 unsigned long max_load = 0;
3990 for_each_cpu(i, sched_group_cpus(group)) {
3993 if (!cpumask_test_cpu(i, cpus))
3997 wl = weighted_cpuload(i);
3999 if (rq->nr_running == 1 && wl > imbalance)
4002 if (wl > max_load) {
4012 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4013 * so long as it is large enough.
4015 #define MAX_PINNED_INTERVAL 512
4017 /* Working cpumask for load_balance and load_balance_newidle. */
4018 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4021 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4022 * tasks if there is an imbalance.
4024 static int load_balance(int this_cpu, struct rq *this_rq,
4025 struct sched_domain *sd, enum cpu_idle_type idle,
4028 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4029 struct sched_group *group;
4030 unsigned long imbalance;
4032 unsigned long flags;
4033 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4035 cpumask_setall(cpus);
4038 * When power savings policy is enabled for the parent domain, idle
4039 * sibling can pick up load irrespective of busy siblings. In this case,
4040 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4041 * portraying it as CPU_NOT_IDLE.
4043 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4044 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4047 schedstat_inc(sd, lb_count[idle]);
4051 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4058 schedstat_inc(sd, lb_nobusyg[idle]);
4062 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4064 schedstat_inc(sd, lb_nobusyq[idle]);
4068 BUG_ON(busiest == this_rq);
4070 schedstat_add(sd, lb_imbalance[idle], imbalance);
4073 if (busiest->nr_running > 1) {
4075 * Attempt to move tasks. If find_busiest_group has found
4076 * an imbalance but busiest->nr_running <= 1, the group is
4077 * still unbalanced. ld_moved simply stays zero, so it is
4078 * correctly treated as an imbalance.
4080 local_irq_save(flags);
4081 double_rq_lock(this_rq, busiest);
4082 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4083 imbalance, sd, idle, &all_pinned);
4084 double_rq_unlock(this_rq, busiest);
4085 local_irq_restore(flags);
4088 * some other cpu did the load balance for us.
4090 if (ld_moved && this_cpu != smp_processor_id())
4091 resched_cpu(this_cpu);
4093 /* All tasks on this runqueue were pinned by CPU affinity */
4094 if (unlikely(all_pinned)) {
4095 cpumask_clear_cpu(cpu_of(busiest), cpus);
4096 if (!cpumask_empty(cpus))
4103 schedstat_inc(sd, lb_failed[idle]);
4104 sd->nr_balance_failed++;
4106 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4108 spin_lock_irqsave(&busiest->lock, flags);
4110 /* don't kick the migration_thread, if the curr
4111 * task on busiest cpu can't be moved to this_cpu
4113 if (!cpumask_test_cpu(this_cpu,
4114 &busiest->curr->cpus_allowed)) {
4115 spin_unlock_irqrestore(&busiest->lock, flags);
4117 goto out_one_pinned;
4120 if (!busiest->active_balance) {
4121 busiest->active_balance = 1;
4122 busiest->push_cpu = this_cpu;
4125 spin_unlock_irqrestore(&busiest->lock, flags);
4127 wake_up_process(busiest->migration_thread);
4130 * We've kicked active balancing, reset the failure
4133 sd->nr_balance_failed = sd->cache_nice_tries+1;
4136 sd->nr_balance_failed = 0;
4138 if (likely(!active_balance)) {
4139 /* We were unbalanced, so reset the balancing interval */
4140 sd->balance_interval = sd->min_interval;
4143 * If we've begun active balancing, start to back off. This
4144 * case may not be covered by the all_pinned logic if there
4145 * is only 1 task on the busy runqueue (because we don't call
4148 if (sd->balance_interval < sd->max_interval)
4149 sd->balance_interval *= 2;
4152 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4153 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4159 schedstat_inc(sd, lb_balanced[idle]);
4161 sd->nr_balance_failed = 0;
4164 /* tune up the balancing interval */
4165 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4166 (sd->balance_interval < sd->max_interval))
4167 sd->balance_interval *= 2;
4169 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4170 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4181 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4182 * tasks if there is an imbalance.
4184 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4185 * this_rq is locked.
4188 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4190 struct sched_group *group;
4191 struct rq *busiest = NULL;
4192 unsigned long imbalance;
4196 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4198 cpumask_setall(cpus);
4201 * When power savings policy is enabled for the parent domain, idle
4202 * sibling can pick up load irrespective of busy siblings. In this case,
4203 * let the state of idle sibling percolate up as IDLE, instead of
4204 * portraying it as CPU_NOT_IDLE.
4206 if (sd->flags & SD_SHARE_CPUPOWER &&
4207 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4210 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4212 update_shares_locked(this_rq, sd);
4213 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4214 &sd_idle, cpus, NULL);
4216 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4220 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4222 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4226 BUG_ON(busiest == this_rq);
4228 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4231 if (busiest->nr_running > 1) {
4232 /* Attempt to move tasks */
4233 double_lock_balance(this_rq, busiest);
4234 /* this_rq->clock is already updated */
4235 update_rq_clock(busiest);
4236 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4237 imbalance, sd, CPU_NEWLY_IDLE,
4239 double_unlock_balance(this_rq, busiest);
4241 if (unlikely(all_pinned)) {
4242 cpumask_clear_cpu(cpu_of(busiest), cpus);
4243 if (!cpumask_empty(cpus))
4249 int active_balance = 0;
4251 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4252 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4253 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4256 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4259 if (sd->nr_balance_failed++ < 2)
4263 * The only task running in a non-idle cpu can be moved to this
4264 * cpu in an attempt to completely freeup the other CPU
4265 * package. The same method used to move task in load_balance()
4266 * have been extended for load_balance_newidle() to speedup
4267 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4269 * The package power saving logic comes from
4270 * find_busiest_group(). If there are no imbalance, then
4271 * f_b_g() will return NULL. However when sched_mc={1,2} then
4272 * f_b_g() will select a group from which a running task may be
4273 * pulled to this cpu in order to make the other package idle.
4274 * If there is no opportunity to make a package idle and if
4275 * there are no imbalance, then f_b_g() will return NULL and no
4276 * action will be taken in load_balance_newidle().
4278 * Under normal task pull operation due to imbalance, there
4279 * will be more than one task in the source run queue and
4280 * move_tasks() will succeed. ld_moved will be true and this
4281 * active balance code will not be triggered.
4284 /* Lock busiest in correct order while this_rq is held */
4285 double_lock_balance(this_rq, busiest);
4288 * don't kick the migration_thread, if the curr
4289 * task on busiest cpu can't be moved to this_cpu
4291 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4292 double_unlock_balance(this_rq, busiest);
4297 if (!busiest->active_balance) {
4298 busiest->active_balance = 1;
4299 busiest->push_cpu = this_cpu;
4303 double_unlock_balance(this_rq, busiest);
4305 * Should not call ttwu while holding a rq->lock
4307 spin_unlock(&this_rq->lock);
4309 wake_up_process(busiest->migration_thread);
4310 spin_lock(&this_rq->lock);
4313 sd->nr_balance_failed = 0;
4315 update_shares_locked(this_rq, sd);
4319 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4320 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4321 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4323 sd->nr_balance_failed = 0;
4329 * idle_balance is called by schedule() if this_cpu is about to become
4330 * idle. Attempts to pull tasks from other CPUs.
4332 static void idle_balance(int this_cpu, struct rq *this_rq)
4334 struct sched_domain *sd;
4335 int pulled_task = 0;
4336 unsigned long next_balance = jiffies + HZ;
4338 for_each_domain(this_cpu, sd) {
4339 unsigned long interval;
4341 if (!(sd->flags & SD_LOAD_BALANCE))
4344 if (sd->flags & SD_BALANCE_NEWIDLE)
4345 /* If we've pulled tasks over stop searching: */
4346 pulled_task = load_balance_newidle(this_cpu, this_rq,
4349 interval = msecs_to_jiffies(sd->balance_interval);
4350 if (time_after(next_balance, sd->last_balance + interval))
4351 next_balance = sd->last_balance + interval;
4355 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4357 * We are going idle. next_balance may be set based on
4358 * a busy processor. So reset next_balance.
4360 this_rq->next_balance = next_balance;
4365 * active_load_balance is run by migration threads. It pushes running tasks
4366 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4367 * running on each physical CPU where possible, and avoids physical /
4368 * logical imbalances.
4370 * Called with busiest_rq locked.
4372 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4374 int target_cpu = busiest_rq->push_cpu;
4375 struct sched_domain *sd;
4376 struct rq *target_rq;
4378 /* Is there any task to move? */
4379 if (busiest_rq->nr_running <= 1)
4382 target_rq = cpu_rq(target_cpu);
4385 * This condition is "impossible", if it occurs
4386 * we need to fix it. Originally reported by
4387 * Bjorn Helgaas on a 128-cpu setup.
4389 BUG_ON(busiest_rq == target_rq);
4391 /* move a task from busiest_rq to target_rq */
4392 double_lock_balance(busiest_rq, target_rq);
4393 update_rq_clock(busiest_rq);
4394 update_rq_clock(target_rq);
4396 /* Search for an sd spanning us and the target CPU. */
4397 for_each_domain(target_cpu, sd) {
4398 if ((sd->flags & SD_LOAD_BALANCE) &&
4399 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4404 schedstat_inc(sd, alb_count);
4406 if (move_one_task(target_rq, target_cpu, busiest_rq,
4408 schedstat_inc(sd, alb_pushed);
4410 schedstat_inc(sd, alb_failed);
4412 double_unlock_balance(busiest_rq, target_rq);
4417 atomic_t load_balancer;
4418 cpumask_var_t cpu_mask;
4419 cpumask_var_t ilb_grp_nohz_mask;
4420 } nohz ____cacheline_aligned = {
4421 .load_balancer = ATOMIC_INIT(-1),
4424 int get_nohz_load_balancer(void)
4426 return atomic_read(&nohz.load_balancer);
4429 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4431 * lowest_flag_domain - Return lowest sched_domain containing flag.
4432 * @cpu: The cpu whose lowest level of sched domain is to
4434 * @flag: The flag to check for the lowest sched_domain
4435 * for the given cpu.
4437 * Returns the lowest sched_domain of a cpu which contains the given flag.
4439 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4441 struct sched_domain *sd;
4443 for_each_domain(cpu, sd)
4444 if (sd && (sd->flags & flag))
4451 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4452 * @cpu: The cpu whose domains we're iterating over.
4453 * @sd: variable holding the value of the power_savings_sd
4455 * @flag: The flag to filter the sched_domains to be iterated.
4457 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4458 * set, starting from the lowest sched_domain to the highest.
4460 #define for_each_flag_domain(cpu, sd, flag) \
4461 for (sd = lowest_flag_domain(cpu, flag); \
4462 (sd && (sd->flags & flag)); sd = sd->parent)
4465 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4466 * @ilb_group: group to be checked for semi-idleness
4468 * Returns: 1 if the group is semi-idle. 0 otherwise.
4470 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4471 * and atleast one non-idle CPU. This helper function checks if the given
4472 * sched_group is semi-idle or not.
4474 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4476 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4477 sched_group_cpus(ilb_group));
4480 * A sched_group is semi-idle when it has atleast one busy cpu
4481 * and atleast one idle cpu.
4483 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4486 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4492 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4493 * @cpu: The cpu which is nominating a new idle_load_balancer.
4495 * Returns: Returns the id of the idle load balancer if it exists,
4496 * Else, returns >= nr_cpu_ids.
4498 * This algorithm picks the idle load balancer such that it belongs to a
4499 * semi-idle powersavings sched_domain. The idea is to try and avoid
4500 * completely idle packages/cores just for the purpose of idle load balancing
4501 * when there are other idle cpu's which are better suited for that job.
4503 static int find_new_ilb(int cpu)
4505 struct sched_domain *sd;
4506 struct sched_group *ilb_group;
4509 * Have idle load balancer selection from semi-idle packages only
4510 * when power-aware load balancing is enabled
4512 if (!(sched_smt_power_savings || sched_mc_power_savings))
4516 * Optimize for the case when we have no idle CPUs or only one
4517 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4519 if (cpumask_weight(nohz.cpu_mask) < 2)
4522 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4523 ilb_group = sd->groups;
4526 if (is_semi_idle_group(ilb_group))
4527 return cpumask_first(nohz.ilb_grp_nohz_mask);
4529 ilb_group = ilb_group->next;
4531 } while (ilb_group != sd->groups);
4535 return cpumask_first(nohz.cpu_mask);
4537 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4538 static inline int find_new_ilb(int call_cpu)
4540 return cpumask_first(nohz.cpu_mask);
4545 * This routine will try to nominate the ilb (idle load balancing)
4546 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4547 * load balancing on behalf of all those cpus. If all the cpus in the system
4548 * go into this tickless mode, then there will be no ilb owner (as there is
4549 * no need for one) and all the cpus will sleep till the next wakeup event
4552 * For the ilb owner, tick is not stopped. And this tick will be used
4553 * for idle load balancing. ilb owner will still be part of
4556 * While stopping the tick, this cpu will become the ilb owner if there
4557 * is no other owner. And will be the owner till that cpu becomes busy
4558 * or if all cpus in the system stop their ticks at which point
4559 * there is no need for ilb owner.
4561 * When the ilb owner becomes busy, it nominates another owner, during the
4562 * next busy scheduler_tick()
4564 int select_nohz_load_balancer(int stop_tick)
4566 int cpu = smp_processor_id();
4569 cpu_rq(cpu)->in_nohz_recently = 1;
4571 if (!cpu_active(cpu)) {
4572 if (atomic_read(&nohz.load_balancer) != cpu)
4576 * If we are going offline and still the leader,
4579 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4585 cpumask_set_cpu(cpu, nohz.cpu_mask);
4587 /* time for ilb owner also to sleep */
4588 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4589 if (atomic_read(&nohz.load_balancer) == cpu)
4590 atomic_set(&nohz.load_balancer, -1);
4594 if (atomic_read(&nohz.load_balancer) == -1) {
4595 /* make me the ilb owner */
4596 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4598 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4601 if (!(sched_smt_power_savings ||
4602 sched_mc_power_savings))
4605 * Check to see if there is a more power-efficient
4608 new_ilb = find_new_ilb(cpu);
4609 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4610 atomic_set(&nohz.load_balancer, -1);
4611 resched_cpu(new_ilb);
4617 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4620 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4622 if (atomic_read(&nohz.load_balancer) == cpu)
4623 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4630 static DEFINE_SPINLOCK(balancing);
4633 * It checks each scheduling domain to see if it is due to be balanced,
4634 * and initiates a balancing operation if so.
4636 * Balancing parameters are set up in arch_init_sched_domains.
4638 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4641 struct rq *rq = cpu_rq(cpu);
4642 unsigned long interval;
4643 struct sched_domain *sd;
4644 /* Earliest time when we have to do rebalance again */
4645 unsigned long next_balance = jiffies + 60*HZ;
4646 int update_next_balance = 0;
4649 for_each_domain(cpu, sd) {
4650 if (!(sd->flags & SD_LOAD_BALANCE))
4653 interval = sd->balance_interval;
4654 if (idle != CPU_IDLE)
4655 interval *= sd->busy_factor;
4657 /* scale ms to jiffies */
4658 interval = msecs_to_jiffies(interval);
4659 if (unlikely(!interval))
4661 if (interval > HZ*NR_CPUS/10)
4662 interval = HZ*NR_CPUS/10;
4664 need_serialize = sd->flags & SD_SERIALIZE;
4666 if (need_serialize) {
4667 if (!spin_trylock(&balancing))
4671 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4672 if (load_balance(cpu, rq, sd, idle, &balance)) {
4674 * We've pulled tasks over so either we're no
4675 * longer idle, or one of our SMT siblings is
4678 idle = CPU_NOT_IDLE;
4680 sd->last_balance = jiffies;
4683 spin_unlock(&balancing);
4685 if (time_after(next_balance, sd->last_balance + interval)) {
4686 next_balance = sd->last_balance + interval;
4687 update_next_balance = 1;
4691 * Stop the load balance at this level. There is another
4692 * CPU in our sched group which is doing load balancing more
4700 * next_balance will be updated only when there is a need.
4701 * When the cpu is attached to null domain for ex, it will not be
4704 if (likely(update_next_balance))
4705 rq->next_balance = next_balance;
4709 * run_rebalance_domains is triggered when needed from the scheduler tick.
4710 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4711 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4713 static void run_rebalance_domains(struct softirq_action *h)
4715 int this_cpu = smp_processor_id();
4716 struct rq *this_rq = cpu_rq(this_cpu);
4717 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4718 CPU_IDLE : CPU_NOT_IDLE;
4720 rebalance_domains(this_cpu, idle);
4724 * If this cpu is the owner for idle load balancing, then do the
4725 * balancing on behalf of the other idle cpus whose ticks are
4728 if (this_rq->idle_at_tick &&
4729 atomic_read(&nohz.load_balancer) == this_cpu) {
4733 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4734 if (balance_cpu == this_cpu)
4738 * If this cpu gets work to do, stop the load balancing
4739 * work being done for other cpus. Next load
4740 * balancing owner will pick it up.
4745 rebalance_domains(balance_cpu, CPU_IDLE);
4747 rq = cpu_rq(balance_cpu);
4748 if (time_after(this_rq->next_balance, rq->next_balance))
4749 this_rq->next_balance = rq->next_balance;
4755 static inline int on_null_domain(int cpu)
4757 return !rcu_dereference(cpu_rq(cpu)->sd);
4761 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4763 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4764 * idle load balancing owner or decide to stop the periodic load balancing,
4765 * if the whole system is idle.
4767 static inline void trigger_load_balance(struct rq *rq, int cpu)
4771 * If we were in the nohz mode recently and busy at the current
4772 * scheduler tick, then check if we need to nominate new idle
4775 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4776 rq->in_nohz_recently = 0;
4778 if (atomic_read(&nohz.load_balancer) == cpu) {
4779 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4780 atomic_set(&nohz.load_balancer, -1);
4783 if (atomic_read(&nohz.load_balancer) == -1) {
4784 int ilb = find_new_ilb(cpu);
4786 if (ilb < nr_cpu_ids)
4792 * If this cpu is idle and doing idle load balancing for all the
4793 * cpus with ticks stopped, is it time for that to stop?
4795 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4796 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4802 * If this cpu is idle and the idle load balancing is done by
4803 * someone else, then no need raise the SCHED_SOFTIRQ
4805 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4806 cpumask_test_cpu(cpu, nohz.cpu_mask))
4809 /* Don't need to rebalance while attached to NULL domain */
4810 if (time_after_eq(jiffies, rq->next_balance) &&
4811 likely(!on_null_domain(cpu)))
4812 raise_softirq(SCHED_SOFTIRQ);
4815 #else /* CONFIG_SMP */
4818 * on UP we do not need to balance between CPUs:
4820 static inline void idle_balance(int cpu, struct rq *rq)
4826 DEFINE_PER_CPU(struct kernel_stat, kstat);
4828 EXPORT_PER_CPU_SYMBOL(kstat);
4831 * Return any ns on the sched_clock that have not yet been accounted in
4832 * @p in case that task is currently running.
4834 * Called with task_rq_lock() held on @rq.
4836 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4840 if (task_current(rq, p)) {
4841 update_rq_clock(rq);
4842 ns = rq->clock - p->se.exec_start;
4850 unsigned long long task_delta_exec(struct task_struct *p)
4852 unsigned long flags;
4856 rq = task_rq_lock(p, &flags);
4857 ns = do_task_delta_exec(p, rq);
4858 task_rq_unlock(rq, &flags);
4864 * Return accounted runtime for the task.
4865 * In case the task is currently running, return the runtime plus current's
4866 * pending runtime that have not been accounted yet.
4868 unsigned long long task_sched_runtime(struct task_struct *p)
4870 unsigned long flags;
4874 rq = task_rq_lock(p, &flags);
4875 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4876 task_rq_unlock(rq, &flags);
4882 * Return sum_exec_runtime for the thread group.
4883 * In case the task is currently running, return the sum plus current's
4884 * pending runtime that have not been accounted yet.
4886 * Note that the thread group might have other running tasks as well,
4887 * so the return value not includes other pending runtime that other
4888 * running tasks might have.
4890 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4892 struct task_cputime totals;
4893 unsigned long flags;
4897 rq = task_rq_lock(p, &flags);
4898 thread_group_cputime(p, &totals);
4899 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4900 task_rq_unlock(rq, &flags);
4906 * Account user cpu time to a process.
4907 * @p: the process that the cpu time gets accounted to
4908 * @cputime: the cpu time spent in user space since the last update
4909 * @cputime_scaled: cputime scaled by cpu frequency
4911 void account_user_time(struct task_struct *p, cputime_t cputime,
4912 cputime_t cputime_scaled)
4914 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4917 /* Add user time to process. */
4918 p->utime = cputime_add(p->utime, cputime);
4919 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4920 account_group_user_time(p, cputime);
4922 /* Add user time to cpustat. */
4923 tmp = cputime_to_cputime64(cputime);
4924 if (TASK_NICE(p) > 0)
4925 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4927 cpustat->user = cputime64_add(cpustat->user, tmp);
4929 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4930 /* Account for user time used */
4931 acct_update_integrals(p);
4935 * Account guest cpu time to a process.
4936 * @p: the process that the cpu time gets accounted to
4937 * @cputime: the cpu time spent in virtual machine since the last update
4938 * @cputime_scaled: cputime scaled by cpu frequency
4940 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4941 cputime_t cputime_scaled)
4944 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4946 tmp = cputime_to_cputime64(cputime);
4948 /* Add guest time to process. */
4949 p->utime = cputime_add(p->utime, cputime);
4950 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4951 account_group_user_time(p, cputime);
4952 p->gtime = cputime_add(p->gtime, cputime);
4954 /* Add guest time to cpustat. */
4955 cpustat->user = cputime64_add(cpustat->user, tmp);
4956 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4960 * Account system cpu time to a process.
4961 * @p: the process that the cpu time gets accounted to
4962 * @hardirq_offset: the offset to subtract from hardirq_count()
4963 * @cputime: the cpu time spent in kernel space since the last update
4964 * @cputime_scaled: cputime scaled by cpu frequency
4966 void account_system_time(struct task_struct *p, int hardirq_offset,
4967 cputime_t cputime, cputime_t cputime_scaled)
4969 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4972 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4973 account_guest_time(p, cputime, cputime_scaled);
4977 /* Add system time to process. */
4978 p->stime = cputime_add(p->stime, cputime);
4979 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4980 account_group_system_time(p, cputime);
4982 /* Add system time to cpustat. */
4983 tmp = cputime_to_cputime64(cputime);
4984 if (hardirq_count() - hardirq_offset)
4985 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4986 else if (softirq_count())
4987 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4989 cpustat->system = cputime64_add(cpustat->system, tmp);
4991 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4993 /* Account for system time used */
4994 acct_update_integrals(p);
4998 * Account for involuntary wait time.
4999 * @steal: the cpu time spent in involuntary wait
5001 void account_steal_time(cputime_t cputime)
5003 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5004 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5006 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5010 * Account for idle time.
5011 * @cputime: the cpu time spent in idle wait
5013 void account_idle_time(cputime_t cputime)
5015 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5016 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5017 struct rq *rq = this_rq();
5019 if (atomic_read(&rq->nr_iowait) > 0)
5020 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5022 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5025 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5028 * Account a single tick of cpu time.
5029 * @p: the process that the cpu time gets accounted to
5030 * @user_tick: indicates if the tick is a user or a system tick
5032 void account_process_tick(struct task_struct *p, int user_tick)
5034 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5035 struct rq *rq = this_rq();
5038 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5039 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5040 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5043 account_idle_time(cputime_one_jiffy);
5047 * Account multiple ticks of steal time.
5048 * @p: the process from which the cpu time has been stolen
5049 * @ticks: number of stolen ticks
5051 void account_steal_ticks(unsigned long ticks)
5053 account_steal_time(jiffies_to_cputime(ticks));
5057 * Account multiple ticks of idle time.
5058 * @ticks: number of stolen ticks
5060 void account_idle_ticks(unsigned long ticks)
5062 account_idle_time(jiffies_to_cputime(ticks));
5068 * Use precise platform statistics if available:
5070 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5071 cputime_t task_utime(struct task_struct *p)
5076 cputime_t task_stime(struct task_struct *p)
5081 cputime_t task_utime(struct task_struct *p)
5083 clock_t utime = cputime_to_clock_t(p->utime),
5084 total = utime + cputime_to_clock_t(p->stime);
5088 * Use CFS's precise accounting:
5090 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5094 do_div(temp, total);
5096 utime = (clock_t)temp;
5098 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5099 return p->prev_utime;
5102 cputime_t task_stime(struct task_struct *p)
5107 * Use CFS's precise accounting. (we subtract utime from
5108 * the total, to make sure the total observed by userspace
5109 * grows monotonically - apps rely on that):
5111 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5112 cputime_to_clock_t(task_utime(p));
5115 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5117 return p->prev_stime;
5121 inline cputime_t task_gtime(struct task_struct *p)
5127 * This function gets called by the timer code, with HZ frequency.
5128 * We call it with interrupts disabled.
5130 * It also gets called by the fork code, when changing the parent's
5133 void scheduler_tick(void)
5135 int cpu = smp_processor_id();
5136 struct rq *rq = cpu_rq(cpu);
5137 struct task_struct *curr = rq->curr;
5141 spin_lock(&rq->lock);
5142 update_rq_clock(rq);
5143 update_cpu_load(rq);
5144 curr->sched_class->task_tick(rq, curr, 0);
5145 spin_unlock(&rq->lock);
5147 perf_counter_task_tick(curr, cpu);
5150 rq->idle_at_tick = idle_cpu(cpu);
5151 trigger_load_balance(rq, cpu);
5155 notrace unsigned long get_parent_ip(unsigned long addr)
5157 if (in_lock_functions(addr)) {
5158 addr = CALLER_ADDR2;
5159 if (in_lock_functions(addr))
5160 addr = CALLER_ADDR3;
5165 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5166 defined(CONFIG_PREEMPT_TRACER))
5168 void __kprobes add_preempt_count(int val)
5170 #ifdef CONFIG_DEBUG_PREEMPT
5174 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5177 preempt_count() += val;
5178 #ifdef CONFIG_DEBUG_PREEMPT
5180 * Spinlock count overflowing soon?
5182 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5185 if (preempt_count() == val)
5186 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5188 EXPORT_SYMBOL(add_preempt_count);
5190 void __kprobes sub_preempt_count(int val)
5192 #ifdef CONFIG_DEBUG_PREEMPT
5196 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5199 * Is the spinlock portion underflowing?
5201 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5202 !(preempt_count() & PREEMPT_MASK)))
5206 if (preempt_count() == val)
5207 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5208 preempt_count() -= val;
5210 EXPORT_SYMBOL(sub_preempt_count);
5215 * Print scheduling while atomic bug:
5217 static noinline void __schedule_bug(struct task_struct *prev)
5219 struct pt_regs *regs = get_irq_regs();
5221 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5222 prev->comm, prev->pid, preempt_count());
5224 debug_show_held_locks(prev);
5226 if (irqs_disabled())
5227 print_irqtrace_events(prev);
5236 * Various schedule()-time debugging checks and statistics:
5238 static inline void schedule_debug(struct task_struct *prev)
5241 * Test if we are atomic. Since do_exit() needs to call into
5242 * schedule() atomically, we ignore that path for now.
5243 * Otherwise, whine if we are scheduling when we should not be.
5245 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5246 __schedule_bug(prev);
5248 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5250 schedstat_inc(this_rq(), sched_count);
5251 #ifdef CONFIG_SCHEDSTATS
5252 if (unlikely(prev->lock_depth >= 0)) {
5253 schedstat_inc(this_rq(), bkl_count);
5254 schedstat_inc(prev, sched_info.bkl_count);
5259 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5261 if (prev->state == TASK_RUNNING) {
5262 u64 runtime = prev->se.sum_exec_runtime;
5264 runtime -= prev->se.prev_sum_exec_runtime;
5265 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5268 * In order to avoid avg_overlap growing stale when we are
5269 * indeed overlapping and hence not getting put to sleep, grow
5270 * the avg_overlap on preemption.
5272 * We use the average preemption runtime because that
5273 * correlates to the amount of cache footprint a task can
5276 update_avg(&prev->se.avg_overlap, runtime);
5278 prev->sched_class->put_prev_task(rq, prev);
5282 * Pick up the highest-prio task:
5284 static inline struct task_struct *
5285 pick_next_task(struct rq *rq)
5287 const struct sched_class *class;
5288 struct task_struct *p;
5291 * Optimization: we know that if all tasks are in
5292 * the fair class we can call that function directly:
5294 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5295 p = fair_sched_class.pick_next_task(rq);
5300 class = sched_class_highest;
5302 p = class->pick_next_task(rq);
5306 * Will never be NULL as the idle class always
5307 * returns a non-NULL p:
5309 class = class->next;
5314 * schedule() is the main scheduler function.
5316 asmlinkage void __sched schedule(void)
5318 struct task_struct *prev, *next;
5319 unsigned long *switch_count;
5325 cpu = smp_processor_id();
5329 switch_count = &prev->nivcsw;
5331 release_kernel_lock(prev);
5332 need_resched_nonpreemptible:
5334 schedule_debug(prev);
5336 if (sched_feat(HRTICK))
5339 spin_lock_irq(&rq->lock);
5340 update_rq_clock(rq);
5341 clear_tsk_need_resched(prev);
5343 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5344 if (unlikely(signal_pending_state(prev->state, prev)))
5345 prev->state = TASK_RUNNING;
5347 deactivate_task(rq, prev, 1);
5348 switch_count = &prev->nvcsw;
5352 if (prev->sched_class->pre_schedule)
5353 prev->sched_class->pre_schedule(rq, prev);
5356 if (unlikely(!rq->nr_running))
5357 idle_balance(cpu, rq);
5359 put_prev_task(rq, prev);
5360 next = pick_next_task(rq);
5362 if (likely(prev != next)) {
5363 sched_info_switch(prev, next);
5364 perf_counter_task_sched_out(prev, next, cpu);
5370 context_switch(rq, prev, next); /* unlocks the rq */
5372 * the context switch might have flipped the stack from under
5373 * us, hence refresh the local variables.
5375 cpu = smp_processor_id();
5378 spin_unlock_irq(&rq->lock);
5380 if (unlikely(reacquire_kernel_lock(current) < 0))
5381 goto need_resched_nonpreemptible;
5383 preempt_enable_no_resched();
5387 EXPORT_SYMBOL(schedule);
5391 * Look out! "owner" is an entirely speculative pointer
5392 * access and not reliable.
5394 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5399 if (!sched_feat(OWNER_SPIN))
5402 #ifdef CONFIG_DEBUG_PAGEALLOC
5404 * Need to access the cpu field knowing that
5405 * DEBUG_PAGEALLOC could have unmapped it if
5406 * the mutex owner just released it and exited.
5408 if (probe_kernel_address(&owner->cpu, cpu))
5415 * Even if the access succeeded (likely case),
5416 * the cpu field may no longer be valid.
5418 if (cpu >= nr_cpumask_bits)
5422 * We need to validate that we can do a
5423 * get_cpu() and that we have the percpu area.
5425 if (!cpu_online(cpu))
5432 * Owner changed, break to re-assess state.
5434 if (lock->owner != owner)
5438 * Is that owner really running on that cpu?
5440 if (task_thread_info(rq->curr) != owner || need_resched())
5450 #ifdef CONFIG_PREEMPT
5452 * this is the entry point to schedule() from in-kernel preemption
5453 * off of preempt_enable. Kernel preemptions off return from interrupt
5454 * occur there and call schedule directly.
5456 asmlinkage void __sched preempt_schedule(void)
5458 struct thread_info *ti = current_thread_info();
5461 * If there is a non-zero preempt_count or interrupts are disabled,
5462 * we do not want to preempt the current task. Just return..
5464 if (likely(ti->preempt_count || irqs_disabled()))
5468 add_preempt_count(PREEMPT_ACTIVE);
5470 sub_preempt_count(PREEMPT_ACTIVE);
5473 * Check again in case we missed a preemption opportunity
5474 * between schedule and now.
5477 } while (need_resched());
5479 EXPORT_SYMBOL(preempt_schedule);
5482 * this is the entry point to schedule() from kernel preemption
5483 * off of irq context.
5484 * Note, that this is called and return with irqs disabled. This will
5485 * protect us against recursive calling from irq.
5487 asmlinkage void __sched preempt_schedule_irq(void)
5489 struct thread_info *ti = current_thread_info();
5491 /* Catch callers which need to be fixed */
5492 BUG_ON(ti->preempt_count || !irqs_disabled());
5495 add_preempt_count(PREEMPT_ACTIVE);
5498 local_irq_disable();
5499 sub_preempt_count(PREEMPT_ACTIVE);
5502 * Check again in case we missed a preemption opportunity
5503 * between schedule and now.
5506 } while (need_resched());
5509 #endif /* CONFIG_PREEMPT */
5511 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5514 return try_to_wake_up(curr->private, mode, sync);
5516 EXPORT_SYMBOL(default_wake_function);
5519 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5520 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5521 * number) then we wake all the non-exclusive tasks and one exclusive task.
5523 * There are circumstances in which we can try to wake a task which has already
5524 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5525 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5527 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5528 int nr_exclusive, int sync, void *key)
5530 wait_queue_t *curr, *next;
5532 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5533 unsigned flags = curr->flags;
5535 if (curr->func(curr, mode, sync, key) &&
5536 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5542 * __wake_up - wake up threads blocked on a waitqueue.
5544 * @mode: which threads
5545 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5546 * @key: is directly passed to the wakeup function
5548 * It may be assumed that this function implies a write memory barrier before
5549 * changing the task state if and only if any tasks are woken up.
5551 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5552 int nr_exclusive, void *key)
5554 unsigned long flags;
5556 spin_lock_irqsave(&q->lock, flags);
5557 __wake_up_common(q, mode, nr_exclusive, 0, key);
5558 spin_unlock_irqrestore(&q->lock, flags);
5560 EXPORT_SYMBOL(__wake_up);
5563 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5565 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5567 __wake_up_common(q, mode, 1, 0, NULL);
5570 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5572 __wake_up_common(q, mode, 1, 0, key);
5576 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5578 * @mode: which threads
5579 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5580 * @key: opaque value to be passed to wakeup targets
5582 * The sync wakeup differs that the waker knows that it will schedule
5583 * away soon, so while the target thread will be woken up, it will not
5584 * be migrated to another CPU - ie. the two threads are 'synchronized'
5585 * with each other. This can prevent needless bouncing between CPUs.
5587 * On UP it can prevent extra preemption.
5589 * It may be assumed that this function implies a write memory barrier before
5590 * changing the task state if and only if any tasks are woken up.
5592 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5593 int nr_exclusive, void *key)
5595 unsigned long flags;
5601 if (unlikely(!nr_exclusive))
5604 spin_lock_irqsave(&q->lock, flags);
5605 __wake_up_common(q, mode, nr_exclusive, sync, key);
5606 spin_unlock_irqrestore(&q->lock, flags);
5608 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5611 * __wake_up_sync - see __wake_up_sync_key()
5613 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5615 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5617 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5620 * complete: - signals a single thread waiting on this completion
5621 * @x: holds the state of this particular completion
5623 * This will wake up a single thread waiting on this completion. Threads will be
5624 * awakened in the same order in which they were queued.
5626 * See also complete_all(), wait_for_completion() and related routines.
5628 * It may be assumed that this function implies a write memory barrier before
5629 * changing the task state if and only if any tasks are woken up.
5631 void complete(struct completion *x)
5633 unsigned long flags;
5635 spin_lock_irqsave(&x->wait.lock, flags);
5637 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5638 spin_unlock_irqrestore(&x->wait.lock, flags);
5640 EXPORT_SYMBOL(complete);
5643 * complete_all: - signals all threads waiting on this completion
5644 * @x: holds the state of this particular completion
5646 * This will wake up all threads waiting on this particular completion event.
5648 * It may be assumed that this function implies a write memory barrier before
5649 * changing the task state if and only if any tasks are woken up.
5651 void complete_all(struct completion *x)
5653 unsigned long flags;
5655 spin_lock_irqsave(&x->wait.lock, flags);
5656 x->done += UINT_MAX/2;
5657 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5658 spin_unlock_irqrestore(&x->wait.lock, flags);
5660 EXPORT_SYMBOL(complete_all);
5662 static inline long __sched
5663 do_wait_for_common(struct completion *x, long timeout, int state)
5666 DECLARE_WAITQUEUE(wait, current);
5668 wait.flags |= WQ_FLAG_EXCLUSIVE;
5669 __add_wait_queue_tail(&x->wait, &wait);
5671 if (signal_pending_state(state, current)) {
5672 timeout = -ERESTARTSYS;
5675 __set_current_state(state);
5676 spin_unlock_irq(&x->wait.lock);
5677 timeout = schedule_timeout(timeout);
5678 spin_lock_irq(&x->wait.lock);
5679 } while (!x->done && timeout);
5680 __remove_wait_queue(&x->wait, &wait);
5685 return timeout ?: 1;
5689 wait_for_common(struct completion *x, long timeout, int state)
5693 spin_lock_irq(&x->wait.lock);
5694 timeout = do_wait_for_common(x, timeout, state);
5695 spin_unlock_irq(&x->wait.lock);
5700 * wait_for_completion: - waits for completion of a task
5701 * @x: holds the state of this particular completion
5703 * This waits to be signaled for completion of a specific task. It is NOT
5704 * interruptible and there is no timeout.
5706 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5707 * and interrupt capability. Also see complete().
5709 void __sched wait_for_completion(struct completion *x)
5711 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5713 EXPORT_SYMBOL(wait_for_completion);
5716 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5717 * @x: holds the state of this particular completion
5718 * @timeout: timeout value in jiffies
5720 * This waits for either a completion of a specific task to be signaled or for a
5721 * specified timeout to expire. The timeout is in jiffies. It is not
5724 unsigned long __sched
5725 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5727 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5729 EXPORT_SYMBOL(wait_for_completion_timeout);
5732 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5733 * @x: holds the state of this particular completion
5735 * This waits for completion of a specific task to be signaled. It is
5738 int __sched wait_for_completion_interruptible(struct completion *x)
5740 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5741 if (t == -ERESTARTSYS)
5745 EXPORT_SYMBOL(wait_for_completion_interruptible);
5748 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5749 * @x: holds the state of this particular completion
5750 * @timeout: timeout value in jiffies
5752 * This waits for either a completion of a specific task to be signaled or for a
5753 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5755 unsigned long __sched
5756 wait_for_completion_interruptible_timeout(struct completion *x,
5757 unsigned long timeout)
5759 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5761 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5764 * wait_for_completion_killable: - waits for completion of a task (killable)
5765 * @x: holds the state of this particular completion
5767 * This waits to be signaled for completion of a specific task. It can be
5768 * interrupted by a kill signal.
5770 int __sched wait_for_completion_killable(struct completion *x)
5772 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5773 if (t == -ERESTARTSYS)
5777 EXPORT_SYMBOL(wait_for_completion_killable);
5780 * try_wait_for_completion - try to decrement a completion without blocking
5781 * @x: completion structure
5783 * Returns: 0 if a decrement cannot be done without blocking
5784 * 1 if a decrement succeeded.
5786 * If a completion is being used as a counting completion,
5787 * attempt to decrement the counter without blocking. This
5788 * enables us to avoid waiting if the resource the completion
5789 * is protecting is not available.
5791 bool try_wait_for_completion(struct completion *x)
5795 spin_lock_irq(&x->wait.lock);
5800 spin_unlock_irq(&x->wait.lock);
5803 EXPORT_SYMBOL(try_wait_for_completion);
5806 * completion_done - Test to see if a completion has any waiters
5807 * @x: completion structure
5809 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5810 * 1 if there are no waiters.
5813 bool completion_done(struct completion *x)
5817 spin_lock_irq(&x->wait.lock);
5820 spin_unlock_irq(&x->wait.lock);
5823 EXPORT_SYMBOL(completion_done);
5826 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5828 unsigned long flags;
5831 init_waitqueue_entry(&wait, current);
5833 __set_current_state(state);
5835 spin_lock_irqsave(&q->lock, flags);
5836 __add_wait_queue(q, &wait);
5837 spin_unlock(&q->lock);
5838 timeout = schedule_timeout(timeout);
5839 spin_lock_irq(&q->lock);
5840 __remove_wait_queue(q, &wait);
5841 spin_unlock_irqrestore(&q->lock, flags);
5846 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5848 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5850 EXPORT_SYMBOL(interruptible_sleep_on);
5853 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5855 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5857 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5859 void __sched sleep_on(wait_queue_head_t *q)
5861 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5863 EXPORT_SYMBOL(sleep_on);
5865 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5867 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5869 EXPORT_SYMBOL(sleep_on_timeout);
5871 #ifdef CONFIG_RT_MUTEXES
5874 * rt_mutex_setprio - set the current priority of a task
5876 * @prio: prio value (kernel-internal form)
5878 * This function changes the 'effective' priority of a task. It does
5879 * not touch ->normal_prio like __setscheduler().
5881 * Used by the rt_mutex code to implement priority inheritance logic.
5883 void rt_mutex_setprio(struct task_struct *p, int prio)
5885 unsigned long flags;
5886 int oldprio, on_rq, running;
5888 const struct sched_class *prev_class = p->sched_class;
5890 BUG_ON(prio < 0 || prio > MAX_PRIO);
5892 rq = task_rq_lock(p, &flags);
5893 update_rq_clock(rq);
5896 on_rq = p->se.on_rq;
5897 running = task_current(rq, p);
5899 dequeue_task(rq, p, 0);
5901 p->sched_class->put_prev_task(rq, p);
5904 p->sched_class = &rt_sched_class;
5906 p->sched_class = &fair_sched_class;
5911 p->sched_class->set_curr_task(rq);
5913 enqueue_task(rq, p, 0);
5915 check_class_changed(rq, p, prev_class, oldprio, running);
5917 task_rq_unlock(rq, &flags);
5922 void set_user_nice(struct task_struct *p, long nice)
5924 int old_prio, delta, on_rq;
5925 unsigned long flags;
5928 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5931 * We have to be careful, if called from sys_setpriority(),
5932 * the task might be in the middle of scheduling on another CPU.
5934 rq = task_rq_lock(p, &flags);
5935 update_rq_clock(rq);
5937 * The RT priorities are set via sched_setscheduler(), but we still
5938 * allow the 'normal' nice value to be set - but as expected
5939 * it wont have any effect on scheduling until the task is
5940 * SCHED_FIFO/SCHED_RR:
5942 if (task_has_rt_policy(p)) {
5943 p->static_prio = NICE_TO_PRIO(nice);
5946 on_rq = p->se.on_rq;
5948 dequeue_task(rq, p, 0);
5950 p->static_prio = NICE_TO_PRIO(nice);
5953 p->prio = effective_prio(p);
5954 delta = p->prio - old_prio;
5957 enqueue_task(rq, p, 0);
5959 * If the task increased its priority or is running and
5960 * lowered its priority, then reschedule its CPU:
5962 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5963 resched_task(rq->curr);
5966 task_rq_unlock(rq, &flags);
5968 EXPORT_SYMBOL(set_user_nice);
5971 * can_nice - check if a task can reduce its nice value
5975 int can_nice(const struct task_struct *p, const int nice)
5977 /* convert nice value [19,-20] to rlimit style value [1,40] */
5978 int nice_rlim = 20 - nice;
5980 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5981 capable(CAP_SYS_NICE));
5984 #ifdef __ARCH_WANT_SYS_NICE
5987 * sys_nice - change the priority of the current process.
5988 * @increment: priority increment
5990 * sys_setpriority is a more generic, but much slower function that
5991 * does similar things.
5993 SYSCALL_DEFINE1(nice, int, increment)
5998 * Setpriority might change our priority at the same moment.
5999 * We don't have to worry. Conceptually one call occurs first
6000 * and we have a single winner.
6002 if (increment < -40)
6007 nice = TASK_NICE(current) + increment;
6013 if (increment < 0 && !can_nice(current, nice))
6016 retval = security_task_setnice(current, nice);
6020 set_user_nice(current, nice);
6027 * task_prio - return the priority value of a given task.
6028 * @p: the task in question.
6030 * This is the priority value as seen by users in /proc.
6031 * RT tasks are offset by -200. Normal tasks are centered
6032 * around 0, value goes from -16 to +15.
6034 int task_prio(const struct task_struct *p)
6036 return p->prio - MAX_RT_PRIO;
6040 * task_nice - return the nice value of a given task.
6041 * @p: the task in question.
6043 int task_nice(const struct task_struct *p)
6045 return TASK_NICE(p);
6047 EXPORT_SYMBOL(task_nice);
6050 * idle_cpu - is a given cpu idle currently?
6051 * @cpu: the processor in question.
6053 int idle_cpu(int cpu)
6055 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6059 * idle_task - return the idle task for a given cpu.
6060 * @cpu: the processor in question.
6062 struct task_struct *idle_task(int cpu)
6064 return cpu_rq(cpu)->idle;
6068 * find_process_by_pid - find a process with a matching PID value.
6069 * @pid: the pid in question.
6071 static struct task_struct *find_process_by_pid(pid_t pid)
6073 return pid ? find_task_by_vpid(pid) : current;
6076 /* Actually do priority change: must hold rq lock. */
6078 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6080 BUG_ON(p->se.on_rq);
6083 switch (p->policy) {
6087 p->sched_class = &fair_sched_class;
6091 p->sched_class = &rt_sched_class;
6095 p->rt_priority = prio;
6096 p->normal_prio = normal_prio(p);
6097 /* we are holding p->pi_lock already */
6098 p->prio = rt_mutex_getprio(p);
6103 * check the target process has a UID that matches the current process's
6105 static bool check_same_owner(struct task_struct *p)
6107 const struct cred *cred = current_cred(), *pcred;
6111 pcred = __task_cred(p);
6112 match = (cred->euid == pcred->euid ||
6113 cred->euid == pcred->uid);
6118 static int __sched_setscheduler(struct task_struct *p, int policy,
6119 struct sched_param *param, bool user)
6121 int retval, oldprio, oldpolicy = -1, on_rq, running;
6122 unsigned long flags;
6123 const struct sched_class *prev_class = p->sched_class;
6126 /* may grab non-irq protected spin_locks */
6127 BUG_ON(in_interrupt());
6129 /* double check policy once rq lock held */
6131 policy = oldpolicy = p->policy;
6132 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
6133 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6134 policy != SCHED_IDLE)
6137 * Valid priorities for SCHED_FIFO and SCHED_RR are
6138 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6139 * SCHED_BATCH and SCHED_IDLE is 0.
6141 if (param->sched_priority < 0 ||
6142 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6143 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6145 if (rt_policy(policy) != (param->sched_priority != 0))
6149 * Allow unprivileged RT tasks to decrease priority:
6151 if (user && !capable(CAP_SYS_NICE)) {
6152 if (rt_policy(policy)) {
6153 unsigned long rlim_rtprio;
6155 if (!lock_task_sighand(p, &flags))
6157 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6158 unlock_task_sighand(p, &flags);
6160 /* can't set/change the rt policy */
6161 if (policy != p->policy && !rlim_rtprio)
6164 /* can't increase priority */
6165 if (param->sched_priority > p->rt_priority &&
6166 param->sched_priority > rlim_rtprio)
6170 * Like positive nice levels, dont allow tasks to
6171 * move out of SCHED_IDLE either:
6173 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6176 /* can't change other user's priorities */
6177 if (!check_same_owner(p))
6182 #ifdef CONFIG_RT_GROUP_SCHED
6184 * Do not allow realtime tasks into groups that have no runtime
6187 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6188 task_group(p)->rt_bandwidth.rt_runtime == 0)
6192 retval = security_task_setscheduler(p, policy, param);
6198 * make sure no PI-waiters arrive (or leave) while we are
6199 * changing the priority of the task:
6201 spin_lock_irqsave(&p->pi_lock, flags);
6203 * To be able to change p->policy safely, the apropriate
6204 * runqueue lock must be held.
6206 rq = __task_rq_lock(p);
6207 /* recheck policy now with rq lock held */
6208 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6209 policy = oldpolicy = -1;
6210 __task_rq_unlock(rq);
6211 spin_unlock_irqrestore(&p->pi_lock, flags);
6214 update_rq_clock(rq);
6215 on_rq = p->se.on_rq;
6216 running = task_current(rq, p);
6218 deactivate_task(rq, p, 0);
6220 p->sched_class->put_prev_task(rq, p);
6223 __setscheduler(rq, p, policy, param->sched_priority);
6226 p->sched_class->set_curr_task(rq);
6228 activate_task(rq, p, 0);
6230 check_class_changed(rq, p, prev_class, oldprio, running);
6232 __task_rq_unlock(rq);
6233 spin_unlock_irqrestore(&p->pi_lock, flags);
6235 rt_mutex_adjust_pi(p);
6241 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6242 * @p: the task in question.
6243 * @policy: new policy.
6244 * @param: structure containing the new RT priority.
6246 * NOTE that the task may be already dead.
6248 int sched_setscheduler(struct task_struct *p, int policy,
6249 struct sched_param *param)
6251 return __sched_setscheduler(p, policy, param, true);
6253 EXPORT_SYMBOL_GPL(sched_setscheduler);
6256 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6257 * @p: the task in question.
6258 * @policy: new policy.
6259 * @param: structure containing the new RT priority.
6261 * Just like sched_setscheduler, only don't bother checking if the
6262 * current context has permission. For example, this is needed in
6263 * stop_machine(): we create temporary high priority worker threads,
6264 * but our caller might not have that capability.
6266 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6267 struct sched_param *param)
6269 return __sched_setscheduler(p, policy, param, false);
6273 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6275 struct sched_param lparam;
6276 struct task_struct *p;
6279 if (!param || pid < 0)
6281 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6286 p = find_process_by_pid(pid);
6288 retval = sched_setscheduler(p, policy, &lparam);
6295 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6296 * @pid: the pid in question.
6297 * @policy: new policy.
6298 * @param: structure containing the new RT priority.
6300 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6301 struct sched_param __user *, param)
6303 /* negative values for policy are not valid */
6307 return do_sched_setscheduler(pid, policy, param);
6311 * sys_sched_setparam - set/change the RT priority of a thread
6312 * @pid: the pid in question.
6313 * @param: structure containing the new RT priority.
6315 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6317 return do_sched_setscheduler(pid, -1, param);
6321 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6322 * @pid: the pid in question.
6324 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6326 struct task_struct *p;
6333 read_lock(&tasklist_lock);
6334 p = find_process_by_pid(pid);
6336 retval = security_task_getscheduler(p);
6340 read_unlock(&tasklist_lock);
6345 * sys_sched_getscheduler - get the RT priority of a thread
6346 * @pid: the pid in question.
6347 * @param: structure containing the RT priority.
6349 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6351 struct sched_param lp;
6352 struct task_struct *p;
6355 if (!param || pid < 0)
6358 read_lock(&tasklist_lock);
6359 p = find_process_by_pid(pid);
6364 retval = security_task_getscheduler(p);
6368 lp.sched_priority = p->rt_priority;
6369 read_unlock(&tasklist_lock);
6372 * This one might sleep, we cannot do it with a spinlock held ...
6374 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6379 read_unlock(&tasklist_lock);
6383 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6385 cpumask_var_t cpus_allowed, new_mask;
6386 struct task_struct *p;
6390 read_lock(&tasklist_lock);
6392 p = find_process_by_pid(pid);
6394 read_unlock(&tasklist_lock);
6400 * It is not safe to call set_cpus_allowed with the
6401 * tasklist_lock held. We will bump the task_struct's
6402 * usage count and then drop tasklist_lock.
6405 read_unlock(&tasklist_lock);
6407 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6411 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6413 goto out_free_cpus_allowed;
6416 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6419 retval = security_task_setscheduler(p, 0, NULL);
6423 cpuset_cpus_allowed(p, cpus_allowed);
6424 cpumask_and(new_mask, in_mask, cpus_allowed);
6426 retval = set_cpus_allowed_ptr(p, new_mask);
6429 cpuset_cpus_allowed(p, cpus_allowed);
6430 if (!cpumask_subset(new_mask, cpus_allowed)) {
6432 * We must have raced with a concurrent cpuset
6433 * update. Just reset the cpus_allowed to the
6434 * cpuset's cpus_allowed
6436 cpumask_copy(new_mask, cpus_allowed);
6441 free_cpumask_var(new_mask);
6442 out_free_cpus_allowed:
6443 free_cpumask_var(cpus_allowed);
6450 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6451 struct cpumask *new_mask)
6453 if (len < cpumask_size())
6454 cpumask_clear(new_mask);
6455 else if (len > cpumask_size())
6456 len = cpumask_size();
6458 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6462 * sys_sched_setaffinity - set the cpu affinity of a process
6463 * @pid: pid of the process
6464 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6465 * @user_mask_ptr: user-space pointer to the new cpu mask
6467 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6468 unsigned long __user *, user_mask_ptr)
6470 cpumask_var_t new_mask;
6473 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6476 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6478 retval = sched_setaffinity(pid, new_mask);
6479 free_cpumask_var(new_mask);
6483 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6485 struct task_struct *p;
6489 read_lock(&tasklist_lock);
6492 p = find_process_by_pid(pid);
6496 retval = security_task_getscheduler(p);
6500 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6503 read_unlock(&tasklist_lock);
6510 * sys_sched_getaffinity - get the cpu affinity of a process
6511 * @pid: pid of the process
6512 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6513 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6515 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6516 unsigned long __user *, user_mask_ptr)
6521 if (len < cpumask_size())
6524 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6527 ret = sched_getaffinity(pid, mask);
6529 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6532 ret = cpumask_size();
6534 free_cpumask_var(mask);
6540 * sys_sched_yield - yield the current processor to other threads.
6542 * This function yields the current CPU to other tasks. If there are no
6543 * other threads running on this CPU then this function will return.
6545 SYSCALL_DEFINE0(sched_yield)
6547 struct rq *rq = this_rq_lock();
6549 schedstat_inc(rq, yld_count);
6550 current->sched_class->yield_task(rq);
6553 * Since we are going to call schedule() anyway, there's
6554 * no need to preempt or enable interrupts:
6556 __release(rq->lock);
6557 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6558 _raw_spin_unlock(&rq->lock);
6559 preempt_enable_no_resched();
6566 static inline int should_resched(void)
6568 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6571 static void __cond_resched(void)
6573 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6574 __might_sleep(__FILE__, __LINE__);
6577 * The BKS might be reacquired before we have dropped
6578 * PREEMPT_ACTIVE, which could trigger a second
6579 * cond_resched() call.
6582 add_preempt_count(PREEMPT_ACTIVE);
6584 sub_preempt_count(PREEMPT_ACTIVE);
6585 } while (need_resched());
6588 int __sched _cond_resched(void)
6590 if (should_resched()) {
6596 EXPORT_SYMBOL(_cond_resched);
6599 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6600 * call schedule, and on return reacquire the lock.
6602 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6603 * operations here to prevent schedule() from being called twice (once via
6604 * spin_unlock(), once by hand).
6606 int cond_resched_lock(spinlock_t *lock)
6608 int resched = should_resched();
6611 if (spin_needbreak(lock) || resched) {
6622 EXPORT_SYMBOL(cond_resched_lock);
6624 int __sched cond_resched_softirq(void)
6626 BUG_ON(!in_softirq());
6628 if (should_resched()) {
6636 EXPORT_SYMBOL(cond_resched_softirq);
6639 * yield - yield the current processor to other threads.
6641 * This is a shortcut for kernel-space yielding - it marks the
6642 * thread runnable and calls sys_sched_yield().
6644 void __sched yield(void)
6646 set_current_state(TASK_RUNNING);
6649 EXPORT_SYMBOL(yield);
6652 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6653 * that process accounting knows that this is a task in IO wait state.
6655 * But don't do that if it is a deliberate, throttling IO wait (this task
6656 * has set its backing_dev_info: the queue against which it should throttle)
6658 void __sched io_schedule(void)
6660 struct rq *rq = &__raw_get_cpu_var(runqueues);
6662 delayacct_blkio_start();
6663 atomic_inc(&rq->nr_iowait);
6665 atomic_dec(&rq->nr_iowait);
6666 delayacct_blkio_end();
6668 EXPORT_SYMBOL(io_schedule);
6670 long __sched io_schedule_timeout(long timeout)
6672 struct rq *rq = &__raw_get_cpu_var(runqueues);
6675 delayacct_blkio_start();
6676 atomic_inc(&rq->nr_iowait);
6677 ret = schedule_timeout(timeout);
6678 atomic_dec(&rq->nr_iowait);
6679 delayacct_blkio_end();
6684 * sys_sched_get_priority_max - return maximum RT priority.
6685 * @policy: scheduling class.
6687 * this syscall returns the maximum rt_priority that can be used
6688 * by a given scheduling class.
6690 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6697 ret = MAX_USER_RT_PRIO-1;
6709 * sys_sched_get_priority_min - return minimum RT priority.
6710 * @policy: scheduling class.
6712 * this syscall returns the minimum rt_priority that can be used
6713 * by a given scheduling class.
6715 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6733 * sys_sched_rr_get_interval - return the default timeslice of a process.
6734 * @pid: pid of the process.
6735 * @interval: userspace pointer to the timeslice value.
6737 * this syscall writes the default timeslice value of a given process
6738 * into the user-space timespec buffer. A value of '0' means infinity.
6740 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6741 struct timespec __user *, interval)
6743 struct task_struct *p;
6744 unsigned int time_slice;
6752 read_lock(&tasklist_lock);
6753 p = find_process_by_pid(pid);
6757 retval = security_task_getscheduler(p);
6762 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6763 * tasks that are on an otherwise idle runqueue:
6766 if (p->policy == SCHED_RR) {
6767 time_slice = DEF_TIMESLICE;
6768 } else if (p->policy != SCHED_FIFO) {
6769 struct sched_entity *se = &p->se;
6770 unsigned long flags;
6773 rq = task_rq_lock(p, &flags);
6774 if (rq->cfs.load.weight)
6775 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6776 task_rq_unlock(rq, &flags);
6778 read_unlock(&tasklist_lock);
6779 jiffies_to_timespec(time_slice, &t);
6780 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6784 read_unlock(&tasklist_lock);
6788 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6790 void sched_show_task(struct task_struct *p)
6792 unsigned long free = 0;
6795 state = p->state ? __ffs(p->state) + 1 : 0;
6796 printk(KERN_INFO "%-13.13s %c", p->comm,
6797 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6798 #if BITS_PER_LONG == 32
6799 if (state == TASK_RUNNING)
6800 printk(KERN_CONT " running ");
6802 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6804 if (state == TASK_RUNNING)
6805 printk(KERN_CONT " running task ");
6807 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6809 #ifdef CONFIG_DEBUG_STACK_USAGE
6810 free = stack_not_used(p);
6812 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6813 task_pid_nr(p), task_pid_nr(p->real_parent),
6814 (unsigned long)task_thread_info(p)->flags);
6816 show_stack(p, NULL);
6819 void show_state_filter(unsigned long state_filter)
6821 struct task_struct *g, *p;
6823 #if BITS_PER_LONG == 32
6825 " task PC stack pid father\n");
6828 " task PC stack pid father\n");
6830 read_lock(&tasklist_lock);
6831 do_each_thread(g, p) {
6833 * reset the NMI-timeout, listing all files on a slow
6834 * console might take alot of time:
6836 touch_nmi_watchdog();
6837 if (!state_filter || (p->state & state_filter))
6839 } while_each_thread(g, p);
6841 touch_all_softlockup_watchdogs();
6843 #ifdef CONFIG_SCHED_DEBUG
6844 sysrq_sched_debug_show();
6846 read_unlock(&tasklist_lock);
6848 * Only show locks if all tasks are dumped:
6850 if (state_filter == -1)
6851 debug_show_all_locks();
6854 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6856 idle->sched_class = &idle_sched_class;
6860 * init_idle - set up an idle thread for a given CPU
6861 * @idle: task in question
6862 * @cpu: cpu the idle task belongs to
6864 * NOTE: this function does not set the idle thread's NEED_RESCHED
6865 * flag, to make booting more robust.
6867 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6869 struct rq *rq = cpu_rq(cpu);
6870 unsigned long flags;
6872 spin_lock_irqsave(&rq->lock, flags);
6875 idle->se.exec_start = sched_clock();
6877 idle->prio = idle->normal_prio = MAX_PRIO;
6878 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6879 __set_task_cpu(idle, cpu);
6881 rq->curr = rq->idle = idle;
6882 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6885 spin_unlock_irqrestore(&rq->lock, flags);
6887 /* Set the preempt count _outside_ the spinlocks! */
6888 #if defined(CONFIG_PREEMPT)
6889 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6891 task_thread_info(idle)->preempt_count = 0;
6894 * The idle tasks have their own, simple scheduling class:
6896 idle->sched_class = &idle_sched_class;
6897 ftrace_graph_init_task(idle);
6901 * In a system that switches off the HZ timer nohz_cpu_mask
6902 * indicates which cpus entered this state. This is used
6903 * in the rcu update to wait only for active cpus. For system
6904 * which do not switch off the HZ timer nohz_cpu_mask should
6905 * always be CPU_BITS_NONE.
6907 cpumask_var_t nohz_cpu_mask;
6910 * Increase the granularity value when there are more CPUs,
6911 * because with more CPUs the 'effective latency' as visible
6912 * to users decreases. But the relationship is not linear,
6913 * so pick a second-best guess by going with the log2 of the
6916 * This idea comes from the SD scheduler of Con Kolivas:
6918 static inline void sched_init_granularity(void)
6920 unsigned int factor = 1 + ilog2(num_online_cpus());
6921 const unsigned long limit = 200000000;
6923 sysctl_sched_min_granularity *= factor;
6924 if (sysctl_sched_min_granularity > limit)
6925 sysctl_sched_min_granularity = limit;
6927 sysctl_sched_latency *= factor;
6928 if (sysctl_sched_latency > limit)
6929 sysctl_sched_latency = limit;
6931 sysctl_sched_wakeup_granularity *= factor;
6933 sysctl_sched_shares_ratelimit *= factor;
6938 * This is how migration works:
6940 * 1) we queue a struct migration_req structure in the source CPU's
6941 * runqueue and wake up that CPU's migration thread.
6942 * 2) we down() the locked semaphore => thread blocks.
6943 * 3) migration thread wakes up (implicitly it forces the migrated
6944 * thread off the CPU)
6945 * 4) it gets the migration request and checks whether the migrated
6946 * task is still in the wrong runqueue.
6947 * 5) if it's in the wrong runqueue then the migration thread removes
6948 * it and puts it into the right queue.
6949 * 6) migration thread up()s the semaphore.
6950 * 7) we wake up and the migration is done.
6954 * Change a given task's CPU affinity. Migrate the thread to a
6955 * proper CPU and schedule it away if the CPU it's executing on
6956 * is removed from the allowed bitmask.
6958 * NOTE: the caller must have a valid reference to the task, the
6959 * task must not exit() & deallocate itself prematurely. The
6960 * call is not atomic; no spinlocks may be held.
6962 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6964 struct migration_req req;
6965 unsigned long flags;
6969 rq = task_rq_lock(p, &flags);
6970 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6975 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6976 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6981 if (p->sched_class->set_cpus_allowed)
6982 p->sched_class->set_cpus_allowed(p, new_mask);
6984 cpumask_copy(&p->cpus_allowed, new_mask);
6985 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6988 /* Can the task run on the task's current CPU? If so, we're done */
6989 if (cpumask_test_cpu(task_cpu(p), new_mask))
6992 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6993 /* Need help from migration thread: drop lock and wait. */
6994 task_rq_unlock(rq, &flags);
6995 wake_up_process(rq->migration_thread);
6996 wait_for_completion(&req.done);
6997 tlb_migrate_finish(p->mm);
7001 task_rq_unlock(rq, &flags);
7005 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7008 * Move (not current) task off this cpu, onto dest cpu. We're doing
7009 * this because either it can't run here any more (set_cpus_allowed()
7010 * away from this CPU, or CPU going down), or because we're
7011 * attempting to rebalance this task on exec (sched_exec).
7013 * So we race with normal scheduler movements, but that's OK, as long
7014 * as the task is no longer on this CPU.
7016 * Returns non-zero if task was successfully migrated.
7018 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7020 struct rq *rq_dest, *rq_src;
7023 if (unlikely(!cpu_active(dest_cpu)))
7026 rq_src = cpu_rq(src_cpu);
7027 rq_dest = cpu_rq(dest_cpu);
7029 double_rq_lock(rq_src, rq_dest);
7030 /* Already moved. */
7031 if (task_cpu(p) != src_cpu)
7033 /* Affinity changed (again). */
7034 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7037 on_rq = p->se.on_rq;
7039 deactivate_task(rq_src, p, 0);
7041 set_task_cpu(p, dest_cpu);
7043 activate_task(rq_dest, p, 0);
7044 check_preempt_curr(rq_dest, p, 0);
7049 double_rq_unlock(rq_src, rq_dest);
7054 * migration_thread - this is a highprio system thread that performs
7055 * thread migration by bumping thread off CPU then 'pushing' onto
7058 static int migration_thread(void *data)
7060 int cpu = (long)data;
7064 BUG_ON(rq->migration_thread != current);
7066 set_current_state(TASK_INTERRUPTIBLE);
7067 while (!kthread_should_stop()) {
7068 struct migration_req *req;
7069 struct list_head *head;
7071 spin_lock_irq(&rq->lock);
7073 if (cpu_is_offline(cpu)) {
7074 spin_unlock_irq(&rq->lock);
7078 if (rq->active_balance) {
7079 active_load_balance(rq, cpu);
7080 rq->active_balance = 0;
7083 head = &rq->migration_queue;
7085 if (list_empty(head)) {
7086 spin_unlock_irq(&rq->lock);
7088 set_current_state(TASK_INTERRUPTIBLE);
7091 req = list_entry(head->next, struct migration_req, list);
7092 list_del_init(head->next);
7094 spin_unlock(&rq->lock);
7095 __migrate_task(req->task, cpu, req->dest_cpu);
7098 complete(&req->done);
7100 __set_current_state(TASK_RUNNING);
7105 #ifdef CONFIG_HOTPLUG_CPU
7107 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7111 local_irq_disable();
7112 ret = __migrate_task(p, src_cpu, dest_cpu);
7118 * Figure out where task on dead CPU should go, use force if necessary.
7120 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7123 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7126 /* Look for allowed, online CPU in same node. */
7127 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7128 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7131 /* Any allowed, online CPU? */
7132 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7133 if (dest_cpu < nr_cpu_ids)
7136 /* No more Mr. Nice Guy. */
7137 if (dest_cpu >= nr_cpu_ids) {
7138 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7139 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7142 * Don't tell them about moving exiting tasks or
7143 * kernel threads (both mm NULL), since they never
7146 if (p->mm && printk_ratelimit()) {
7147 printk(KERN_INFO "process %d (%s) no "
7148 "longer affine to cpu%d\n",
7149 task_pid_nr(p), p->comm, dead_cpu);
7154 /* It can have affinity changed while we were choosing. */
7155 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7160 * While a dead CPU has no uninterruptible tasks queued at this point,
7161 * it might still have a nonzero ->nr_uninterruptible counter, because
7162 * for performance reasons the counter is not stricly tracking tasks to
7163 * their home CPUs. So we just add the counter to another CPU's counter,
7164 * to keep the global sum constant after CPU-down:
7166 static void migrate_nr_uninterruptible(struct rq *rq_src)
7168 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7169 unsigned long flags;
7171 local_irq_save(flags);
7172 double_rq_lock(rq_src, rq_dest);
7173 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7174 rq_src->nr_uninterruptible = 0;
7175 double_rq_unlock(rq_src, rq_dest);
7176 local_irq_restore(flags);
7179 /* Run through task list and migrate tasks from the dead cpu. */
7180 static void migrate_live_tasks(int src_cpu)
7182 struct task_struct *p, *t;
7184 read_lock(&tasklist_lock);
7186 do_each_thread(t, p) {
7190 if (task_cpu(p) == src_cpu)
7191 move_task_off_dead_cpu(src_cpu, p);
7192 } while_each_thread(t, p);
7194 read_unlock(&tasklist_lock);
7198 * Schedules idle task to be the next runnable task on current CPU.
7199 * It does so by boosting its priority to highest possible.
7200 * Used by CPU offline code.
7202 void sched_idle_next(void)
7204 int this_cpu = smp_processor_id();
7205 struct rq *rq = cpu_rq(this_cpu);
7206 struct task_struct *p = rq->idle;
7207 unsigned long flags;
7209 /* cpu has to be offline */
7210 BUG_ON(cpu_online(this_cpu));
7213 * Strictly not necessary since rest of the CPUs are stopped by now
7214 * and interrupts disabled on the current cpu.
7216 spin_lock_irqsave(&rq->lock, flags);
7218 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7220 update_rq_clock(rq);
7221 activate_task(rq, p, 0);
7223 spin_unlock_irqrestore(&rq->lock, flags);
7227 * Ensures that the idle task is using init_mm right before its cpu goes
7230 void idle_task_exit(void)
7232 struct mm_struct *mm = current->active_mm;
7234 BUG_ON(cpu_online(smp_processor_id()));
7237 switch_mm(mm, &init_mm, current);
7241 /* called under rq->lock with disabled interrupts */
7242 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7244 struct rq *rq = cpu_rq(dead_cpu);
7246 /* Must be exiting, otherwise would be on tasklist. */
7247 BUG_ON(!p->exit_state);
7249 /* Cannot have done final schedule yet: would have vanished. */
7250 BUG_ON(p->state == TASK_DEAD);
7255 * Drop lock around migration; if someone else moves it,
7256 * that's OK. No task can be added to this CPU, so iteration is
7259 spin_unlock_irq(&rq->lock);
7260 move_task_off_dead_cpu(dead_cpu, p);
7261 spin_lock_irq(&rq->lock);
7266 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7267 static void migrate_dead_tasks(unsigned int dead_cpu)
7269 struct rq *rq = cpu_rq(dead_cpu);
7270 struct task_struct *next;
7273 if (!rq->nr_running)
7275 update_rq_clock(rq);
7276 next = pick_next_task(rq);
7279 next->sched_class->put_prev_task(rq, next);
7280 migrate_dead(dead_cpu, next);
7286 * remove the tasks which were accounted by rq from calc_load_tasks.
7288 static void calc_global_load_remove(struct rq *rq)
7290 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7291 rq->calc_load_active = 0;
7293 #endif /* CONFIG_HOTPLUG_CPU */
7295 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7297 static struct ctl_table sd_ctl_dir[] = {
7299 .procname = "sched_domain",
7305 static struct ctl_table sd_ctl_root[] = {
7307 .ctl_name = CTL_KERN,
7308 .procname = "kernel",
7310 .child = sd_ctl_dir,
7315 static struct ctl_table *sd_alloc_ctl_entry(int n)
7317 struct ctl_table *entry =
7318 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7323 static void sd_free_ctl_entry(struct ctl_table **tablep)
7325 struct ctl_table *entry;
7328 * In the intermediate directories, both the child directory and
7329 * procname are dynamically allocated and could fail but the mode
7330 * will always be set. In the lowest directory the names are
7331 * static strings and all have proc handlers.
7333 for (entry = *tablep; entry->mode; entry++) {
7335 sd_free_ctl_entry(&entry->child);
7336 if (entry->proc_handler == NULL)
7337 kfree(entry->procname);
7345 set_table_entry(struct ctl_table *entry,
7346 const char *procname, void *data, int maxlen,
7347 mode_t mode, proc_handler *proc_handler)
7349 entry->procname = procname;
7351 entry->maxlen = maxlen;
7353 entry->proc_handler = proc_handler;
7356 static struct ctl_table *
7357 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7359 struct ctl_table *table = sd_alloc_ctl_entry(13);
7364 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7365 sizeof(long), 0644, proc_doulongvec_minmax);
7366 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7367 sizeof(long), 0644, proc_doulongvec_minmax);
7368 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7369 sizeof(int), 0644, proc_dointvec_minmax);
7370 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7371 sizeof(int), 0644, proc_dointvec_minmax);
7372 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7373 sizeof(int), 0644, proc_dointvec_minmax);
7374 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7375 sizeof(int), 0644, proc_dointvec_minmax);
7376 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7377 sizeof(int), 0644, proc_dointvec_minmax);
7378 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7379 sizeof(int), 0644, proc_dointvec_minmax);
7380 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7381 sizeof(int), 0644, proc_dointvec_minmax);
7382 set_table_entry(&table[9], "cache_nice_tries",
7383 &sd->cache_nice_tries,
7384 sizeof(int), 0644, proc_dointvec_minmax);
7385 set_table_entry(&table[10], "flags", &sd->flags,
7386 sizeof(int), 0644, proc_dointvec_minmax);
7387 set_table_entry(&table[11], "name", sd->name,
7388 CORENAME_MAX_SIZE, 0444, proc_dostring);
7389 /* &table[12] is terminator */
7394 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7396 struct ctl_table *entry, *table;
7397 struct sched_domain *sd;
7398 int domain_num = 0, i;
7401 for_each_domain(cpu, sd)
7403 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7408 for_each_domain(cpu, sd) {
7409 snprintf(buf, 32, "domain%d", i);
7410 entry->procname = kstrdup(buf, GFP_KERNEL);
7412 entry->child = sd_alloc_ctl_domain_table(sd);
7419 static struct ctl_table_header *sd_sysctl_header;
7420 static void register_sched_domain_sysctl(void)
7422 int i, cpu_num = num_online_cpus();
7423 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7426 WARN_ON(sd_ctl_dir[0].child);
7427 sd_ctl_dir[0].child = entry;
7432 for_each_online_cpu(i) {
7433 snprintf(buf, 32, "cpu%d", i);
7434 entry->procname = kstrdup(buf, GFP_KERNEL);
7436 entry->child = sd_alloc_ctl_cpu_table(i);
7440 WARN_ON(sd_sysctl_header);
7441 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7444 /* may be called multiple times per register */
7445 static void unregister_sched_domain_sysctl(void)
7447 if (sd_sysctl_header)
7448 unregister_sysctl_table(sd_sysctl_header);
7449 sd_sysctl_header = NULL;
7450 if (sd_ctl_dir[0].child)
7451 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7454 static void register_sched_domain_sysctl(void)
7457 static void unregister_sched_domain_sysctl(void)
7462 static void set_rq_online(struct rq *rq)
7465 const struct sched_class *class;
7467 cpumask_set_cpu(rq->cpu, rq->rd->online);
7470 for_each_class(class) {
7471 if (class->rq_online)
7472 class->rq_online(rq);
7477 static void set_rq_offline(struct rq *rq)
7480 const struct sched_class *class;
7482 for_each_class(class) {
7483 if (class->rq_offline)
7484 class->rq_offline(rq);
7487 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7493 * migration_call - callback that gets triggered when a CPU is added.
7494 * Here we can start up the necessary migration thread for the new CPU.
7496 static int __cpuinit
7497 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7499 struct task_struct *p;
7500 int cpu = (long)hcpu;
7501 unsigned long flags;
7506 case CPU_UP_PREPARE:
7507 case CPU_UP_PREPARE_FROZEN:
7508 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7511 kthread_bind(p, cpu);
7512 /* Must be high prio: stop_machine expects to yield to it. */
7513 rq = task_rq_lock(p, &flags);
7514 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7515 task_rq_unlock(rq, &flags);
7517 cpu_rq(cpu)->migration_thread = p;
7518 rq->calc_load_update = calc_load_update;
7522 case CPU_ONLINE_FROZEN:
7523 /* Strictly unnecessary, as first user will wake it. */
7524 wake_up_process(cpu_rq(cpu)->migration_thread);
7526 /* Update our root-domain */
7528 spin_lock_irqsave(&rq->lock, flags);
7530 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7534 spin_unlock_irqrestore(&rq->lock, flags);
7537 #ifdef CONFIG_HOTPLUG_CPU
7538 case CPU_UP_CANCELED:
7539 case CPU_UP_CANCELED_FROZEN:
7540 if (!cpu_rq(cpu)->migration_thread)
7542 /* Unbind it from offline cpu so it can run. Fall thru. */
7543 kthread_bind(cpu_rq(cpu)->migration_thread,
7544 cpumask_any(cpu_online_mask));
7545 kthread_stop(cpu_rq(cpu)->migration_thread);
7546 put_task_struct(cpu_rq(cpu)->migration_thread);
7547 cpu_rq(cpu)->migration_thread = NULL;
7551 case CPU_DEAD_FROZEN:
7552 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7553 migrate_live_tasks(cpu);
7555 kthread_stop(rq->migration_thread);
7556 put_task_struct(rq->migration_thread);
7557 rq->migration_thread = NULL;
7558 /* Idle task back to normal (off runqueue, low prio) */
7559 spin_lock_irq(&rq->lock);
7560 update_rq_clock(rq);
7561 deactivate_task(rq, rq->idle, 0);
7562 rq->idle->static_prio = MAX_PRIO;
7563 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7564 rq->idle->sched_class = &idle_sched_class;
7565 migrate_dead_tasks(cpu);
7566 spin_unlock_irq(&rq->lock);
7568 migrate_nr_uninterruptible(rq);
7569 BUG_ON(rq->nr_running != 0);
7570 calc_global_load_remove(rq);
7572 * No need to migrate the tasks: it was best-effort if
7573 * they didn't take sched_hotcpu_mutex. Just wake up
7576 spin_lock_irq(&rq->lock);
7577 while (!list_empty(&rq->migration_queue)) {
7578 struct migration_req *req;
7580 req = list_entry(rq->migration_queue.next,
7581 struct migration_req, list);
7582 list_del_init(&req->list);
7583 spin_unlock_irq(&rq->lock);
7584 complete(&req->done);
7585 spin_lock_irq(&rq->lock);
7587 spin_unlock_irq(&rq->lock);
7591 case CPU_DYING_FROZEN:
7592 /* Update our root-domain */
7594 spin_lock_irqsave(&rq->lock, flags);
7596 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7599 spin_unlock_irqrestore(&rq->lock, flags);
7607 * Register at high priority so that task migration (migrate_all_tasks)
7608 * happens before everything else. This has to be lower priority than
7609 * the notifier in the perf_counter subsystem, though.
7611 static struct notifier_block __cpuinitdata migration_notifier = {
7612 .notifier_call = migration_call,
7616 static int __init migration_init(void)
7618 void *cpu = (void *)(long)smp_processor_id();
7621 /* Start one for the boot CPU: */
7622 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7623 BUG_ON(err == NOTIFY_BAD);
7624 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7625 register_cpu_notifier(&migration_notifier);
7629 early_initcall(migration_init);
7634 #ifdef CONFIG_SCHED_DEBUG
7636 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7637 struct cpumask *groupmask)
7639 struct sched_group *group = sd->groups;
7642 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7643 cpumask_clear(groupmask);
7645 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7647 if (!(sd->flags & SD_LOAD_BALANCE)) {
7648 printk("does not load-balance\n");
7650 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7655 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7657 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7658 printk(KERN_ERR "ERROR: domain->span does not contain "
7661 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7662 printk(KERN_ERR "ERROR: domain->groups does not contain"
7666 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7670 printk(KERN_ERR "ERROR: group is NULL\n");
7674 if (!group->__cpu_power) {
7675 printk(KERN_CONT "\n");
7676 printk(KERN_ERR "ERROR: domain->cpu_power not "
7681 if (!cpumask_weight(sched_group_cpus(group))) {
7682 printk(KERN_CONT "\n");
7683 printk(KERN_ERR "ERROR: empty group\n");
7687 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7688 printk(KERN_CONT "\n");
7689 printk(KERN_ERR "ERROR: repeated CPUs\n");
7693 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7695 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7697 printk(KERN_CONT " %s", str);
7698 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7699 printk(KERN_CONT " (__cpu_power = %d)",
7700 group->__cpu_power);
7703 group = group->next;
7704 } while (group != sd->groups);
7705 printk(KERN_CONT "\n");
7707 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7708 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7711 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7712 printk(KERN_ERR "ERROR: parent span is not a superset "
7713 "of domain->span\n");
7717 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7719 cpumask_var_t groupmask;
7723 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7727 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7729 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7730 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7735 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7742 free_cpumask_var(groupmask);
7744 #else /* !CONFIG_SCHED_DEBUG */
7745 # define sched_domain_debug(sd, cpu) do { } while (0)
7746 #endif /* CONFIG_SCHED_DEBUG */
7748 static int sd_degenerate(struct sched_domain *sd)
7750 if (cpumask_weight(sched_domain_span(sd)) == 1)
7753 /* Following flags need at least 2 groups */
7754 if (sd->flags & (SD_LOAD_BALANCE |
7755 SD_BALANCE_NEWIDLE |
7759 SD_SHARE_PKG_RESOURCES)) {
7760 if (sd->groups != sd->groups->next)
7764 /* Following flags don't use groups */
7765 if (sd->flags & (SD_WAKE_IDLE |
7774 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7776 unsigned long cflags = sd->flags, pflags = parent->flags;
7778 if (sd_degenerate(parent))
7781 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7784 /* Does parent contain flags not in child? */
7785 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7786 if (cflags & SD_WAKE_AFFINE)
7787 pflags &= ~SD_WAKE_BALANCE;
7788 /* Flags needing groups don't count if only 1 group in parent */
7789 if (parent->groups == parent->groups->next) {
7790 pflags &= ~(SD_LOAD_BALANCE |
7791 SD_BALANCE_NEWIDLE |
7795 SD_SHARE_PKG_RESOURCES);
7796 if (nr_node_ids == 1)
7797 pflags &= ~SD_SERIALIZE;
7799 if (~cflags & pflags)
7805 static void free_rootdomain(struct root_domain *rd)
7807 cpupri_cleanup(&rd->cpupri);
7809 free_cpumask_var(rd->rto_mask);
7810 free_cpumask_var(rd->online);
7811 free_cpumask_var(rd->span);
7815 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7817 struct root_domain *old_rd = NULL;
7818 unsigned long flags;
7820 spin_lock_irqsave(&rq->lock, flags);
7825 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7828 cpumask_clear_cpu(rq->cpu, old_rd->span);
7831 * If we dont want to free the old_rt yet then
7832 * set old_rd to NULL to skip the freeing later
7835 if (!atomic_dec_and_test(&old_rd->refcount))
7839 atomic_inc(&rd->refcount);
7842 cpumask_set_cpu(rq->cpu, rd->span);
7843 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7846 spin_unlock_irqrestore(&rq->lock, flags);
7849 free_rootdomain(old_rd);
7852 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7854 gfp_t gfp = GFP_KERNEL;
7856 memset(rd, 0, sizeof(*rd));
7861 if (!alloc_cpumask_var(&rd->span, gfp))
7863 if (!alloc_cpumask_var(&rd->online, gfp))
7865 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7868 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7873 free_cpumask_var(rd->rto_mask);
7875 free_cpumask_var(rd->online);
7877 free_cpumask_var(rd->span);
7882 static void init_defrootdomain(void)
7884 init_rootdomain(&def_root_domain, true);
7886 atomic_set(&def_root_domain.refcount, 1);
7889 static struct root_domain *alloc_rootdomain(void)
7891 struct root_domain *rd;
7893 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7897 if (init_rootdomain(rd, false) != 0) {
7906 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7907 * hold the hotplug lock.
7910 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7912 struct rq *rq = cpu_rq(cpu);
7913 struct sched_domain *tmp;
7915 /* Remove the sched domains which do not contribute to scheduling. */
7916 for (tmp = sd; tmp; ) {
7917 struct sched_domain *parent = tmp->parent;
7921 if (sd_parent_degenerate(tmp, parent)) {
7922 tmp->parent = parent->parent;
7924 parent->parent->child = tmp;
7929 if (sd && sd_degenerate(sd)) {
7935 sched_domain_debug(sd, cpu);
7937 rq_attach_root(rq, rd);
7938 rcu_assign_pointer(rq->sd, sd);
7941 /* cpus with isolated domains */
7942 static cpumask_var_t cpu_isolated_map;
7944 /* Setup the mask of cpus configured for isolated domains */
7945 static int __init isolated_cpu_setup(char *str)
7947 cpulist_parse(str, cpu_isolated_map);
7951 __setup("isolcpus=", isolated_cpu_setup);
7954 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7955 * to a function which identifies what group(along with sched group) a CPU
7956 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7957 * (due to the fact that we keep track of groups covered with a struct cpumask).
7959 * init_sched_build_groups will build a circular linked list of the groups
7960 * covered by the given span, and will set each group's ->cpumask correctly,
7961 * and ->cpu_power to 0.
7964 init_sched_build_groups(const struct cpumask *span,
7965 const struct cpumask *cpu_map,
7966 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7967 struct sched_group **sg,
7968 struct cpumask *tmpmask),
7969 struct cpumask *covered, struct cpumask *tmpmask)
7971 struct sched_group *first = NULL, *last = NULL;
7974 cpumask_clear(covered);
7976 for_each_cpu(i, span) {
7977 struct sched_group *sg;
7978 int group = group_fn(i, cpu_map, &sg, tmpmask);
7981 if (cpumask_test_cpu(i, covered))
7984 cpumask_clear(sched_group_cpus(sg));
7985 sg->__cpu_power = 0;
7987 for_each_cpu(j, span) {
7988 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7991 cpumask_set_cpu(j, covered);
7992 cpumask_set_cpu(j, sched_group_cpus(sg));
8003 #define SD_NODES_PER_DOMAIN 16
8008 * find_next_best_node - find the next node to include in a sched_domain
8009 * @node: node whose sched_domain we're building
8010 * @used_nodes: nodes already in the sched_domain
8012 * Find the next node to include in a given scheduling domain. Simply
8013 * finds the closest node not already in the @used_nodes map.
8015 * Should use nodemask_t.
8017 static int find_next_best_node(int node, nodemask_t *used_nodes)
8019 int i, n, val, min_val, best_node = 0;
8023 for (i = 0; i < nr_node_ids; i++) {
8024 /* Start at @node */
8025 n = (node + i) % nr_node_ids;
8027 if (!nr_cpus_node(n))
8030 /* Skip already used nodes */
8031 if (node_isset(n, *used_nodes))
8034 /* Simple min distance search */
8035 val = node_distance(node, n);
8037 if (val < min_val) {
8043 node_set(best_node, *used_nodes);
8048 * sched_domain_node_span - get a cpumask for a node's sched_domain
8049 * @node: node whose cpumask we're constructing
8050 * @span: resulting cpumask
8052 * Given a node, construct a good cpumask for its sched_domain to span. It
8053 * should be one that prevents unnecessary balancing, but also spreads tasks
8056 static void sched_domain_node_span(int node, struct cpumask *span)
8058 nodemask_t used_nodes;
8061 cpumask_clear(span);
8062 nodes_clear(used_nodes);
8064 cpumask_or(span, span, cpumask_of_node(node));
8065 node_set(node, used_nodes);
8067 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8068 int next_node = find_next_best_node(node, &used_nodes);
8070 cpumask_or(span, span, cpumask_of_node(next_node));
8073 #endif /* CONFIG_NUMA */
8075 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8078 * The cpus mask in sched_group and sched_domain hangs off the end.
8080 * ( See the the comments in include/linux/sched.h:struct sched_group
8081 * and struct sched_domain. )
8083 struct static_sched_group {
8084 struct sched_group sg;
8085 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8088 struct static_sched_domain {
8089 struct sched_domain sd;
8090 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8094 * SMT sched-domains:
8096 #ifdef CONFIG_SCHED_SMT
8097 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8098 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8101 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8102 struct sched_group **sg, struct cpumask *unused)
8105 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8108 #endif /* CONFIG_SCHED_SMT */
8111 * multi-core sched-domains:
8113 #ifdef CONFIG_SCHED_MC
8114 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8115 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8116 #endif /* CONFIG_SCHED_MC */
8118 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8120 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8121 struct sched_group **sg, struct cpumask *mask)
8125 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8126 group = cpumask_first(mask);
8128 *sg = &per_cpu(sched_group_core, group).sg;
8131 #elif defined(CONFIG_SCHED_MC)
8133 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8134 struct sched_group **sg, struct cpumask *unused)
8137 *sg = &per_cpu(sched_group_core, cpu).sg;
8142 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8143 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8146 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8147 struct sched_group **sg, struct cpumask *mask)
8150 #ifdef CONFIG_SCHED_MC
8151 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8152 group = cpumask_first(mask);
8153 #elif defined(CONFIG_SCHED_SMT)
8154 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8155 group = cpumask_first(mask);
8160 *sg = &per_cpu(sched_group_phys, group).sg;
8166 * The init_sched_build_groups can't handle what we want to do with node
8167 * groups, so roll our own. Now each node has its own list of groups which
8168 * gets dynamically allocated.
8170 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8171 static struct sched_group ***sched_group_nodes_bycpu;
8173 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8174 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8176 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8177 struct sched_group **sg,
8178 struct cpumask *nodemask)
8182 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8183 group = cpumask_first(nodemask);
8186 *sg = &per_cpu(sched_group_allnodes, group).sg;
8190 static void init_numa_sched_groups_power(struct sched_group *group_head)
8192 struct sched_group *sg = group_head;
8198 for_each_cpu(j, sched_group_cpus(sg)) {
8199 struct sched_domain *sd;
8201 sd = &per_cpu(phys_domains, j).sd;
8202 if (j != group_first_cpu(sd->groups)) {
8204 * Only add "power" once for each
8210 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8213 } while (sg != group_head);
8215 #endif /* CONFIG_NUMA */
8218 /* Free memory allocated for various sched_group structures */
8219 static void free_sched_groups(const struct cpumask *cpu_map,
8220 struct cpumask *nodemask)
8224 for_each_cpu(cpu, cpu_map) {
8225 struct sched_group **sched_group_nodes
8226 = sched_group_nodes_bycpu[cpu];
8228 if (!sched_group_nodes)
8231 for (i = 0; i < nr_node_ids; i++) {
8232 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8234 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8235 if (cpumask_empty(nodemask))
8245 if (oldsg != sched_group_nodes[i])
8248 kfree(sched_group_nodes);
8249 sched_group_nodes_bycpu[cpu] = NULL;
8252 #else /* !CONFIG_NUMA */
8253 static void free_sched_groups(const struct cpumask *cpu_map,
8254 struct cpumask *nodemask)
8257 #endif /* CONFIG_NUMA */
8260 * Initialize sched groups cpu_power.
8262 * cpu_power indicates the capacity of sched group, which is used while
8263 * distributing the load between different sched groups in a sched domain.
8264 * Typically cpu_power for all the groups in a sched domain will be same unless
8265 * there are asymmetries in the topology. If there are asymmetries, group
8266 * having more cpu_power will pickup more load compared to the group having
8269 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8270 * the maximum number of tasks a group can handle in the presence of other idle
8271 * or lightly loaded groups in the same sched domain.
8273 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8275 struct sched_domain *child;
8276 struct sched_group *group;
8278 WARN_ON(!sd || !sd->groups);
8280 if (cpu != group_first_cpu(sd->groups))
8285 sd->groups->__cpu_power = 0;
8288 * For perf policy, if the groups in child domain share resources
8289 * (for example cores sharing some portions of the cache hierarchy
8290 * or SMT), then set this domain groups cpu_power such that each group
8291 * can handle only one task, when there are other idle groups in the
8292 * same sched domain.
8294 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8296 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8297 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8302 * add cpu_power of each child group to this groups cpu_power
8304 group = child->groups;
8306 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8307 group = group->next;
8308 } while (group != child->groups);
8312 * Initializers for schedule domains
8313 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8316 #ifdef CONFIG_SCHED_DEBUG
8317 # define SD_INIT_NAME(sd, type) sd->name = #type
8319 # define SD_INIT_NAME(sd, type) do { } while (0)
8322 #define SD_INIT(sd, type) sd_init_##type(sd)
8324 #define SD_INIT_FUNC(type) \
8325 static noinline void sd_init_##type(struct sched_domain *sd) \
8327 memset(sd, 0, sizeof(*sd)); \
8328 *sd = SD_##type##_INIT; \
8329 sd->level = SD_LV_##type; \
8330 SD_INIT_NAME(sd, type); \
8335 SD_INIT_FUNC(ALLNODES)
8338 #ifdef CONFIG_SCHED_SMT
8339 SD_INIT_FUNC(SIBLING)
8341 #ifdef CONFIG_SCHED_MC
8345 static int default_relax_domain_level = -1;
8347 static int __init setup_relax_domain_level(char *str)
8351 val = simple_strtoul(str, NULL, 0);
8352 if (val < SD_LV_MAX)
8353 default_relax_domain_level = val;
8357 __setup("relax_domain_level=", setup_relax_domain_level);
8359 static void set_domain_attribute(struct sched_domain *sd,
8360 struct sched_domain_attr *attr)
8364 if (!attr || attr->relax_domain_level < 0) {
8365 if (default_relax_domain_level < 0)
8368 request = default_relax_domain_level;
8370 request = attr->relax_domain_level;
8371 if (request < sd->level) {
8372 /* turn off idle balance on this domain */
8373 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8375 /* turn on idle balance on this domain */
8376 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8381 * Build sched domains for a given set of cpus and attach the sched domains
8382 * to the individual cpus
8384 static int __build_sched_domains(const struct cpumask *cpu_map,
8385 struct sched_domain_attr *attr)
8387 int i, err = -ENOMEM;
8388 struct root_domain *rd;
8389 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8392 cpumask_var_t domainspan, covered, notcovered;
8393 struct sched_group **sched_group_nodes = NULL;
8394 int sd_allnodes = 0;
8396 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8398 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8399 goto free_domainspan;
8400 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
8404 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8405 goto free_notcovered;
8406 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8408 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8409 goto free_this_sibling_map;
8410 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8411 goto free_this_core_map;
8412 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8413 goto free_send_covered;
8417 * Allocate the per-node list of sched groups
8419 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8421 if (!sched_group_nodes) {
8422 printk(KERN_WARNING "Can not alloc sched group node list\n");
8427 rd = alloc_rootdomain();
8429 printk(KERN_WARNING "Cannot alloc root domain\n");
8430 goto free_sched_groups;
8434 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8438 * Set up domains for cpus specified by the cpu_map.
8440 for_each_cpu(i, cpu_map) {
8441 struct sched_domain *sd = NULL, *p;
8443 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8446 if (cpumask_weight(cpu_map) >
8447 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8448 sd = &per_cpu(allnodes_domains, i).sd;
8449 SD_INIT(sd, ALLNODES);
8450 set_domain_attribute(sd, attr);
8451 cpumask_copy(sched_domain_span(sd), cpu_map);
8452 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8458 sd = &per_cpu(node_domains, i).sd;
8460 set_domain_attribute(sd, attr);
8461 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8465 cpumask_and(sched_domain_span(sd),
8466 sched_domain_span(sd), cpu_map);
8470 sd = &per_cpu(phys_domains, i).sd;
8472 set_domain_attribute(sd, attr);
8473 cpumask_copy(sched_domain_span(sd), nodemask);
8477 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8479 #ifdef CONFIG_SCHED_MC
8481 sd = &per_cpu(core_domains, i).sd;
8483 set_domain_attribute(sd, attr);
8484 cpumask_and(sched_domain_span(sd), cpu_map,
8485 cpu_coregroup_mask(i));
8488 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8491 #ifdef CONFIG_SCHED_SMT
8493 sd = &per_cpu(cpu_domains, i).sd;
8494 SD_INIT(sd, SIBLING);
8495 set_domain_attribute(sd, attr);
8496 cpumask_and(sched_domain_span(sd),
8497 topology_thread_cpumask(i), cpu_map);
8500 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8504 #ifdef CONFIG_SCHED_SMT
8505 /* Set up CPU (sibling) groups */
8506 for_each_cpu(i, cpu_map) {
8507 cpumask_and(this_sibling_map,
8508 topology_thread_cpumask(i), cpu_map);
8509 if (i != cpumask_first(this_sibling_map))
8512 init_sched_build_groups(this_sibling_map, cpu_map,
8514 send_covered, tmpmask);
8518 #ifdef CONFIG_SCHED_MC
8519 /* Set up multi-core groups */
8520 for_each_cpu(i, cpu_map) {
8521 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8522 if (i != cpumask_first(this_core_map))
8525 init_sched_build_groups(this_core_map, cpu_map,
8527 send_covered, tmpmask);
8531 /* Set up physical groups */
8532 for (i = 0; i < nr_node_ids; i++) {
8533 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8534 if (cpumask_empty(nodemask))
8537 init_sched_build_groups(nodemask, cpu_map,
8539 send_covered, tmpmask);
8543 /* Set up node groups */
8545 init_sched_build_groups(cpu_map, cpu_map,
8546 &cpu_to_allnodes_group,
8547 send_covered, tmpmask);
8550 for (i = 0; i < nr_node_ids; i++) {
8551 /* Set up node groups */
8552 struct sched_group *sg, *prev;
8555 cpumask_clear(covered);
8556 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8557 if (cpumask_empty(nodemask)) {
8558 sched_group_nodes[i] = NULL;
8562 sched_domain_node_span(i, domainspan);
8563 cpumask_and(domainspan, domainspan, cpu_map);
8565 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8568 printk(KERN_WARNING "Can not alloc domain group for "
8572 sched_group_nodes[i] = sg;
8573 for_each_cpu(j, nodemask) {
8574 struct sched_domain *sd;
8576 sd = &per_cpu(node_domains, j).sd;
8579 sg->__cpu_power = 0;
8580 cpumask_copy(sched_group_cpus(sg), nodemask);
8582 cpumask_or(covered, covered, nodemask);
8585 for (j = 0; j < nr_node_ids; j++) {
8586 int n = (i + j) % nr_node_ids;
8588 cpumask_complement(notcovered, covered);
8589 cpumask_and(tmpmask, notcovered, cpu_map);
8590 cpumask_and(tmpmask, tmpmask, domainspan);
8591 if (cpumask_empty(tmpmask))
8594 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8595 if (cpumask_empty(tmpmask))
8598 sg = kmalloc_node(sizeof(struct sched_group) +
8603 "Can not alloc domain group for node %d\n", j);
8606 sg->__cpu_power = 0;
8607 cpumask_copy(sched_group_cpus(sg), tmpmask);
8608 sg->next = prev->next;
8609 cpumask_or(covered, covered, tmpmask);
8616 /* Calculate CPU power for physical packages and nodes */
8617 #ifdef CONFIG_SCHED_SMT
8618 for_each_cpu(i, cpu_map) {
8619 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8621 init_sched_groups_power(i, sd);
8624 #ifdef CONFIG_SCHED_MC
8625 for_each_cpu(i, cpu_map) {
8626 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8628 init_sched_groups_power(i, sd);
8632 for_each_cpu(i, cpu_map) {
8633 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8635 init_sched_groups_power(i, sd);
8639 for (i = 0; i < nr_node_ids; i++)
8640 init_numa_sched_groups_power(sched_group_nodes[i]);
8643 struct sched_group *sg;
8645 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8647 init_numa_sched_groups_power(sg);
8651 /* Attach the domains */
8652 for_each_cpu(i, cpu_map) {
8653 struct sched_domain *sd;
8654 #ifdef CONFIG_SCHED_SMT
8655 sd = &per_cpu(cpu_domains, i).sd;
8656 #elif defined(CONFIG_SCHED_MC)
8657 sd = &per_cpu(core_domains, i).sd;
8659 sd = &per_cpu(phys_domains, i).sd;
8661 cpu_attach_domain(sd, rd, i);
8667 free_cpumask_var(tmpmask);
8669 free_cpumask_var(send_covered);
8671 free_cpumask_var(this_core_map);
8672 free_this_sibling_map:
8673 free_cpumask_var(this_sibling_map);
8675 free_cpumask_var(nodemask);
8678 free_cpumask_var(notcovered);
8680 free_cpumask_var(covered);
8682 free_cpumask_var(domainspan);
8689 kfree(sched_group_nodes);
8695 free_sched_groups(cpu_map, tmpmask);
8696 free_rootdomain(rd);
8701 static int build_sched_domains(const struct cpumask *cpu_map)
8703 return __build_sched_domains(cpu_map, NULL);
8706 static struct cpumask *doms_cur; /* current sched domains */
8707 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8708 static struct sched_domain_attr *dattr_cur;
8709 /* attribues of custom domains in 'doms_cur' */
8712 * Special case: If a kmalloc of a doms_cur partition (array of
8713 * cpumask) fails, then fallback to a single sched domain,
8714 * as determined by the single cpumask fallback_doms.
8716 static cpumask_var_t fallback_doms;
8719 * arch_update_cpu_topology lets virtualized architectures update the
8720 * cpu core maps. It is supposed to return 1 if the topology changed
8721 * or 0 if it stayed the same.
8723 int __attribute__((weak)) arch_update_cpu_topology(void)
8729 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8730 * For now this just excludes isolated cpus, but could be used to
8731 * exclude other special cases in the future.
8733 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8737 arch_update_cpu_topology();
8739 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8741 doms_cur = fallback_doms;
8742 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8744 err = build_sched_domains(doms_cur);
8745 register_sched_domain_sysctl();
8750 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8751 struct cpumask *tmpmask)
8753 free_sched_groups(cpu_map, tmpmask);
8757 * Detach sched domains from a group of cpus specified in cpu_map
8758 * These cpus will now be attached to the NULL domain
8760 static void detach_destroy_domains(const struct cpumask *cpu_map)
8762 /* Save because hotplug lock held. */
8763 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8766 for_each_cpu(i, cpu_map)
8767 cpu_attach_domain(NULL, &def_root_domain, i);
8768 synchronize_sched();
8769 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8772 /* handle null as "default" */
8773 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8774 struct sched_domain_attr *new, int idx_new)
8776 struct sched_domain_attr tmp;
8783 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8784 new ? (new + idx_new) : &tmp,
8785 sizeof(struct sched_domain_attr));
8789 * Partition sched domains as specified by the 'ndoms_new'
8790 * cpumasks in the array doms_new[] of cpumasks. This compares
8791 * doms_new[] to the current sched domain partitioning, doms_cur[].
8792 * It destroys each deleted domain and builds each new domain.
8794 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8795 * The masks don't intersect (don't overlap.) We should setup one
8796 * sched domain for each mask. CPUs not in any of the cpumasks will
8797 * not be load balanced. If the same cpumask appears both in the
8798 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8801 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8802 * ownership of it and will kfree it when done with it. If the caller
8803 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8804 * ndoms_new == 1, and partition_sched_domains() will fallback to
8805 * the single partition 'fallback_doms', it also forces the domains
8808 * If doms_new == NULL it will be replaced with cpu_online_mask.
8809 * ndoms_new == 0 is a special case for destroying existing domains,
8810 * and it will not create the default domain.
8812 * Call with hotplug lock held
8814 /* FIXME: Change to struct cpumask *doms_new[] */
8815 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8816 struct sched_domain_attr *dattr_new)
8821 mutex_lock(&sched_domains_mutex);
8823 /* always unregister in case we don't destroy any domains */
8824 unregister_sched_domain_sysctl();
8826 /* Let architecture update cpu core mappings. */
8827 new_topology = arch_update_cpu_topology();
8829 n = doms_new ? ndoms_new : 0;
8831 /* Destroy deleted domains */
8832 for (i = 0; i < ndoms_cur; i++) {
8833 for (j = 0; j < n && !new_topology; j++) {
8834 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8835 && dattrs_equal(dattr_cur, i, dattr_new, j))
8838 /* no match - a current sched domain not in new doms_new[] */
8839 detach_destroy_domains(doms_cur + i);
8844 if (doms_new == NULL) {
8846 doms_new = fallback_doms;
8847 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8848 WARN_ON_ONCE(dattr_new);
8851 /* Build new domains */
8852 for (i = 0; i < ndoms_new; i++) {
8853 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8854 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8855 && dattrs_equal(dattr_new, i, dattr_cur, j))
8858 /* no match - add a new doms_new */
8859 __build_sched_domains(doms_new + i,
8860 dattr_new ? dattr_new + i : NULL);
8865 /* Remember the new sched domains */
8866 if (doms_cur != fallback_doms)
8868 kfree(dattr_cur); /* kfree(NULL) is safe */
8869 doms_cur = doms_new;
8870 dattr_cur = dattr_new;
8871 ndoms_cur = ndoms_new;
8873 register_sched_domain_sysctl();
8875 mutex_unlock(&sched_domains_mutex);
8878 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8879 static void arch_reinit_sched_domains(void)
8883 /* Destroy domains first to force the rebuild */
8884 partition_sched_domains(0, NULL, NULL);
8886 rebuild_sched_domains();
8890 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8892 unsigned int level = 0;
8894 if (sscanf(buf, "%u", &level) != 1)
8898 * level is always be positive so don't check for
8899 * level < POWERSAVINGS_BALANCE_NONE which is 0
8900 * What happens on 0 or 1 byte write,
8901 * need to check for count as well?
8904 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8908 sched_smt_power_savings = level;
8910 sched_mc_power_savings = level;
8912 arch_reinit_sched_domains();
8917 #ifdef CONFIG_SCHED_MC
8918 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8921 return sprintf(page, "%u\n", sched_mc_power_savings);
8923 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8924 const char *buf, size_t count)
8926 return sched_power_savings_store(buf, count, 0);
8928 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8929 sched_mc_power_savings_show,
8930 sched_mc_power_savings_store);
8933 #ifdef CONFIG_SCHED_SMT
8934 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8937 return sprintf(page, "%u\n", sched_smt_power_savings);
8939 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8940 const char *buf, size_t count)
8942 return sched_power_savings_store(buf, count, 1);
8944 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8945 sched_smt_power_savings_show,
8946 sched_smt_power_savings_store);
8949 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8953 #ifdef CONFIG_SCHED_SMT
8955 err = sysfs_create_file(&cls->kset.kobj,
8956 &attr_sched_smt_power_savings.attr);
8958 #ifdef CONFIG_SCHED_MC
8959 if (!err && mc_capable())
8960 err = sysfs_create_file(&cls->kset.kobj,
8961 &attr_sched_mc_power_savings.attr);
8965 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8967 #ifndef CONFIG_CPUSETS
8969 * Add online and remove offline CPUs from the scheduler domains.
8970 * When cpusets are enabled they take over this function.
8972 static int update_sched_domains(struct notifier_block *nfb,
8973 unsigned long action, void *hcpu)
8977 case CPU_ONLINE_FROZEN:
8979 case CPU_DEAD_FROZEN:
8980 partition_sched_domains(1, NULL, NULL);
8989 static int update_runtime(struct notifier_block *nfb,
8990 unsigned long action, void *hcpu)
8992 int cpu = (int)(long)hcpu;
8995 case CPU_DOWN_PREPARE:
8996 case CPU_DOWN_PREPARE_FROZEN:
8997 disable_runtime(cpu_rq(cpu));
9000 case CPU_DOWN_FAILED:
9001 case CPU_DOWN_FAILED_FROZEN:
9003 case CPU_ONLINE_FROZEN:
9004 enable_runtime(cpu_rq(cpu));
9012 void __init sched_init_smp(void)
9014 cpumask_var_t non_isolated_cpus;
9016 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9018 #if defined(CONFIG_NUMA)
9019 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9021 BUG_ON(sched_group_nodes_bycpu == NULL);
9024 mutex_lock(&sched_domains_mutex);
9025 arch_init_sched_domains(cpu_online_mask);
9026 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9027 if (cpumask_empty(non_isolated_cpus))
9028 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9029 mutex_unlock(&sched_domains_mutex);
9032 #ifndef CONFIG_CPUSETS
9033 /* XXX: Theoretical race here - CPU may be hotplugged now */
9034 hotcpu_notifier(update_sched_domains, 0);
9037 /* RT runtime code needs to handle some hotplug events */
9038 hotcpu_notifier(update_runtime, 0);
9042 /* Move init over to a non-isolated CPU */
9043 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9045 sched_init_granularity();
9046 free_cpumask_var(non_isolated_cpus);
9048 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9049 init_sched_rt_class();
9052 void __init sched_init_smp(void)
9054 sched_init_granularity();
9056 #endif /* CONFIG_SMP */
9058 const_debug unsigned int sysctl_timer_migration = 1;
9060 int in_sched_functions(unsigned long addr)
9062 return in_lock_functions(addr) ||
9063 (addr >= (unsigned long)__sched_text_start
9064 && addr < (unsigned long)__sched_text_end);
9067 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9069 cfs_rq->tasks_timeline = RB_ROOT;
9070 INIT_LIST_HEAD(&cfs_rq->tasks);
9071 #ifdef CONFIG_FAIR_GROUP_SCHED
9074 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9077 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9079 struct rt_prio_array *array;
9082 array = &rt_rq->active;
9083 for (i = 0; i < MAX_RT_PRIO; i++) {
9084 INIT_LIST_HEAD(array->queue + i);
9085 __clear_bit(i, array->bitmap);
9087 /* delimiter for bitsearch: */
9088 __set_bit(MAX_RT_PRIO, array->bitmap);
9090 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9091 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9093 rt_rq->highest_prio.next = MAX_RT_PRIO;
9097 rt_rq->rt_nr_migratory = 0;
9098 rt_rq->overloaded = 0;
9099 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9103 rt_rq->rt_throttled = 0;
9104 rt_rq->rt_runtime = 0;
9105 spin_lock_init(&rt_rq->rt_runtime_lock);
9107 #ifdef CONFIG_RT_GROUP_SCHED
9108 rt_rq->rt_nr_boosted = 0;
9113 #ifdef CONFIG_FAIR_GROUP_SCHED
9114 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9115 struct sched_entity *se, int cpu, int add,
9116 struct sched_entity *parent)
9118 struct rq *rq = cpu_rq(cpu);
9119 tg->cfs_rq[cpu] = cfs_rq;
9120 init_cfs_rq(cfs_rq, rq);
9123 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9126 /* se could be NULL for init_task_group */
9131 se->cfs_rq = &rq->cfs;
9133 se->cfs_rq = parent->my_q;
9136 se->load.weight = tg->shares;
9137 se->load.inv_weight = 0;
9138 se->parent = parent;
9142 #ifdef CONFIG_RT_GROUP_SCHED
9143 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9144 struct sched_rt_entity *rt_se, int cpu, int add,
9145 struct sched_rt_entity *parent)
9147 struct rq *rq = cpu_rq(cpu);
9149 tg->rt_rq[cpu] = rt_rq;
9150 init_rt_rq(rt_rq, rq);
9152 rt_rq->rt_se = rt_se;
9153 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9155 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9157 tg->rt_se[cpu] = rt_se;
9162 rt_se->rt_rq = &rq->rt;
9164 rt_se->rt_rq = parent->my_q;
9166 rt_se->my_q = rt_rq;
9167 rt_se->parent = parent;
9168 INIT_LIST_HEAD(&rt_se->run_list);
9172 void __init sched_init(void)
9175 unsigned long alloc_size = 0, ptr;
9177 #ifdef CONFIG_FAIR_GROUP_SCHED
9178 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9180 #ifdef CONFIG_RT_GROUP_SCHED
9181 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9183 #ifdef CONFIG_USER_SCHED
9186 #ifdef CONFIG_CPUMASK_OFFSTACK
9187 alloc_size += num_possible_cpus() * cpumask_size();
9190 * As sched_init() is called before page_alloc is setup,
9191 * we use alloc_bootmem().
9194 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9196 #ifdef CONFIG_FAIR_GROUP_SCHED
9197 init_task_group.se = (struct sched_entity **)ptr;
9198 ptr += nr_cpu_ids * sizeof(void **);
9200 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9201 ptr += nr_cpu_ids * sizeof(void **);
9203 #ifdef CONFIG_USER_SCHED
9204 root_task_group.se = (struct sched_entity **)ptr;
9205 ptr += nr_cpu_ids * sizeof(void **);
9207 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9208 ptr += nr_cpu_ids * sizeof(void **);
9209 #endif /* CONFIG_USER_SCHED */
9210 #endif /* CONFIG_FAIR_GROUP_SCHED */
9211 #ifdef CONFIG_RT_GROUP_SCHED
9212 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9213 ptr += nr_cpu_ids * sizeof(void **);
9215 init_task_group.rt_rq = (struct rt_rq **)ptr;
9216 ptr += nr_cpu_ids * sizeof(void **);
9218 #ifdef CONFIG_USER_SCHED
9219 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9220 ptr += nr_cpu_ids * sizeof(void **);
9222 root_task_group.rt_rq = (struct rt_rq **)ptr;
9223 ptr += nr_cpu_ids * sizeof(void **);
9224 #endif /* CONFIG_USER_SCHED */
9225 #endif /* CONFIG_RT_GROUP_SCHED */
9226 #ifdef CONFIG_CPUMASK_OFFSTACK
9227 for_each_possible_cpu(i) {
9228 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9229 ptr += cpumask_size();
9231 #endif /* CONFIG_CPUMASK_OFFSTACK */
9235 init_defrootdomain();
9238 init_rt_bandwidth(&def_rt_bandwidth,
9239 global_rt_period(), global_rt_runtime());
9241 #ifdef CONFIG_RT_GROUP_SCHED
9242 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9243 global_rt_period(), global_rt_runtime());
9244 #ifdef CONFIG_USER_SCHED
9245 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9246 global_rt_period(), RUNTIME_INF);
9247 #endif /* CONFIG_USER_SCHED */
9248 #endif /* CONFIG_RT_GROUP_SCHED */
9250 #ifdef CONFIG_GROUP_SCHED
9251 list_add(&init_task_group.list, &task_groups);
9252 INIT_LIST_HEAD(&init_task_group.children);
9254 #ifdef CONFIG_USER_SCHED
9255 INIT_LIST_HEAD(&root_task_group.children);
9256 init_task_group.parent = &root_task_group;
9257 list_add(&init_task_group.siblings, &root_task_group.children);
9258 #endif /* CONFIG_USER_SCHED */
9259 #endif /* CONFIG_GROUP_SCHED */
9261 for_each_possible_cpu(i) {
9265 spin_lock_init(&rq->lock);
9267 rq->calc_load_active = 0;
9268 rq->calc_load_update = jiffies + LOAD_FREQ;
9269 init_cfs_rq(&rq->cfs, rq);
9270 init_rt_rq(&rq->rt, rq);
9271 #ifdef CONFIG_FAIR_GROUP_SCHED
9272 init_task_group.shares = init_task_group_load;
9273 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9274 #ifdef CONFIG_CGROUP_SCHED
9276 * How much cpu bandwidth does init_task_group get?
9278 * In case of task-groups formed thr' the cgroup filesystem, it
9279 * gets 100% of the cpu resources in the system. This overall
9280 * system cpu resource is divided among the tasks of
9281 * init_task_group and its child task-groups in a fair manner,
9282 * based on each entity's (task or task-group's) weight
9283 * (se->load.weight).
9285 * In other words, if init_task_group has 10 tasks of weight
9286 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9287 * then A0's share of the cpu resource is:
9289 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9291 * We achieve this by letting init_task_group's tasks sit
9292 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9294 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9295 #elif defined CONFIG_USER_SCHED
9296 root_task_group.shares = NICE_0_LOAD;
9297 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9299 * In case of task-groups formed thr' the user id of tasks,
9300 * init_task_group represents tasks belonging to root user.
9301 * Hence it forms a sibling of all subsequent groups formed.
9302 * In this case, init_task_group gets only a fraction of overall
9303 * system cpu resource, based on the weight assigned to root
9304 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9305 * by letting tasks of init_task_group sit in a separate cfs_rq
9306 * (init_cfs_rq) and having one entity represent this group of
9307 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9309 init_tg_cfs_entry(&init_task_group,
9310 &per_cpu(init_cfs_rq, i),
9311 &per_cpu(init_sched_entity, i), i, 1,
9312 root_task_group.se[i]);
9315 #endif /* CONFIG_FAIR_GROUP_SCHED */
9317 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9318 #ifdef CONFIG_RT_GROUP_SCHED
9319 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9320 #ifdef CONFIG_CGROUP_SCHED
9321 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9322 #elif defined CONFIG_USER_SCHED
9323 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9324 init_tg_rt_entry(&init_task_group,
9325 &per_cpu(init_rt_rq, i),
9326 &per_cpu(init_sched_rt_entity, i), i, 1,
9327 root_task_group.rt_se[i]);
9331 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9332 rq->cpu_load[j] = 0;
9336 rq->active_balance = 0;
9337 rq->next_balance = jiffies;
9341 rq->migration_thread = NULL;
9342 INIT_LIST_HEAD(&rq->migration_queue);
9343 rq_attach_root(rq, &def_root_domain);
9346 atomic_set(&rq->nr_iowait, 0);
9349 set_load_weight(&init_task);
9351 #ifdef CONFIG_PREEMPT_NOTIFIERS
9352 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9356 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9359 #ifdef CONFIG_RT_MUTEXES
9360 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9364 * The boot idle thread does lazy MMU switching as well:
9366 atomic_inc(&init_mm.mm_count);
9367 enter_lazy_tlb(&init_mm, current);
9370 * Make us the idle thread. Technically, schedule() should not be
9371 * called from this thread, however somewhere below it might be,
9372 * but because we are the idle thread, we just pick up running again
9373 * when this runqueue becomes "idle".
9375 init_idle(current, smp_processor_id());
9377 calc_load_update = jiffies + LOAD_FREQ;
9380 * During early bootup we pretend to be a normal task:
9382 current->sched_class = &fair_sched_class;
9384 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9385 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9388 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9389 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9391 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9394 perf_counter_init();
9396 scheduler_running = 1;
9399 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9400 void __might_sleep(char *file, int line)
9403 static unsigned long prev_jiffy; /* ratelimiting */
9405 if ((!in_atomic() && !irqs_disabled()) ||
9406 system_state != SYSTEM_RUNNING || oops_in_progress)
9408 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9410 prev_jiffy = jiffies;
9413 "BUG: sleeping function called from invalid context at %s:%d\n",
9416 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9417 in_atomic(), irqs_disabled(),
9418 current->pid, current->comm);
9420 debug_show_held_locks(current);
9421 if (irqs_disabled())
9422 print_irqtrace_events(current);
9426 EXPORT_SYMBOL(__might_sleep);
9429 #ifdef CONFIG_MAGIC_SYSRQ
9430 static void normalize_task(struct rq *rq, struct task_struct *p)
9434 update_rq_clock(rq);
9435 on_rq = p->se.on_rq;
9437 deactivate_task(rq, p, 0);
9438 __setscheduler(rq, p, SCHED_NORMAL, 0);
9440 activate_task(rq, p, 0);
9441 resched_task(rq->curr);
9445 void normalize_rt_tasks(void)
9447 struct task_struct *g, *p;
9448 unsigned long flags;
9451 read_lock_irqsave(&tasklist_lock, flags);
9452 do_each_thread(g, p) {
9454 * Only normalize user tasks:
9459 p->se.exec_start = 0;
9460 #ifdef CONFIG_SCHEDSTATS
9461 p->se.wait_start = 0;
9462 p->se.sleep_start = 0;
9463 p->se.block_start = 0;
9468 * Renice negative nice level userspace
9471 if (TASK_NICE(p) < 0 && p->mm)
9472 set_user_nice(p, 0);
9476 spin_lock(&p->pi_lock);
9477 rq = __task_rq_lock(p);
9479 normalize_task(rq, p);
9481 __task_rq_unlock(rq);
9482 spin_unlock(&p->pi_lock);
9483 } while_each_thread(g, p);
9485 read_unlock_irqrestore(&tasklist_lock, flags);
9488 #endif /* CONFIG_MAGIC_SYSRQ */
9492 * These functions are only useful for the IA64 MCA handling.
9494 * They can only be called when the whole system has been
9495 * stopped - every CPU needs to be quiescent, and no scheduling
9496 * activity can take place. Using them for anything else would
9497 * be a serious bug, and as a result, they aren't even visible
9498 * under any other configuration.
9502 * curr_task - return the current task for a given cpu.
9503 * @cpu: the processor in question.
9505 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9507 struct task_struct *curr_task(int cpu)
9509 return cpu_curr(cpu);
9513 * set_curr_task - set the current task for a given cpu.
9514 * @cpu: the processor in question.
9515 * @p: the task pointer to set.
9517 * Description: This function must only be used when non-maskable interrupts
9518 * are serviced on a separate stack. It allows the architecture to switch the
9519 * notion of the current task on a cpu in a non-blocking manner. This function
9520 * must be called with all CPU's synchronized, and interrupts disabled, the
9521 * and caller must save the original value of the current task (see
9522 * curr_task() above) and restore that value before reenabling interrupts and
9523 * re-starting the system.
9525 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9527 void set_curr_task(int cpu, struct task_struct *p)
9534 #ifdef CONFIG_FAIR_GROUP_SCHED
9535 static void free_fair_sched_group(struct task_group *tg)
9539 for_each_possible_cpu(i) {
9541 kfree(tg->cfs_rq[i]);
9551 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9553 struct cfs_rq *cfs_rq;
9554 struct sched_entity *se;
9558 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9561 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9565 tg->shares = NICE_0_LOAD;
9567 for_each_possible_cpu(i) {
9570 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9571 GFP_KERNEL, cpu_to_node(i));
9575 se = kzalloc_node(sizeof(struct sched_entity),
9576 GFP_KERNEL, cpu_to_node(i));
9580 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9589 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9591 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9592 &cpu_rq(cpu)->leaf_cfs_rq_list);
9595 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9597 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9599 #else /* !CONFG_FAIR_GROUP_SCHED */
9600 static inline void free_fair_sched_group(struct task_group *tg)
9605 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9610 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9614 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9617 #endif /* CONFIG_FAIR_GROUP_SCHED */
9619 #ifdef CONFIG_RT_GROUP_SCHED
9620 static void free_rt_sched_group(struct task_group *tg)
9624 destroy_rt_bandwidth(&tg->rt_bandwidth);
9626 for_each_possible_cpu(i) {
9628 kfree(tg->rt_rq[i]);
9630 kfree(tg->rt_se[i]);
9638 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9640 struct rt_rq *rt_rq;
9641 struct sched_rt_entity *rt_se;
9645 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9648 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9652 init_rt_bandwidth(&tg->rt_bandwidth,
9653 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9655 for_each_possible_cpu(i) {
9658 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9659 GFP_KERNEL, cpu_to_node(i));
9663 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9664 GFP_KERNEL, cpu_to_node(i));
9668 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9677 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9679 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9680 &cpu_rq(cpu)->leaf_rt_rq_list);
9683 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9685 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9687 #else /* !CONFIG_RT_GROUP_SCHED */
9688 static inline void free_rt_sched_group(struct task_group *tg)
9693 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9698 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9702 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9705 #endif /* CONFIG_RT_GROUP_SCHED */
9707 #ifdef CONFIG_GROUP_SCHED
9708 static void free_sched_group(struct task_group *tg)
9710 free_fair_sched_group(tg);
9711 free_rt_sched_group(tg);
9715 /* allocate runqueue etc for a new task group */
9716 struct task_group *sched_create_group(struct task_group *parent)
9718 struct task_group *tg;
9719 unsigned long flags;
9722 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9724 return ERR_PTR(-ENOMEM);
9726 if (!alloc_fair_sched_group(tg, parent))
9729 if (!alloc_rt_sched_group(tg, parent))
9732 spin_lock_irqsave(&task_group_lock, flags);
9733 for_each_possible_cpu(i) {
9734 register_fair_sched_group(tg, i);
9735 register_rt_sched_group(tg, i);
9737 list_add_rcu(&tg->list, &task_groups);
9739 WARN_ON(!parent); /* root should already exist */
9741 tg->parent = parent;
9742 INIT_LIST_HEAD(&tg->children);
9743 list_add_rcu(&tg->siblings, &parent->children);
9744 spin_unlock_irqrestore(&task_group_lock, flags);
9749 free_sched_group(tg);
9750 return ERR_PTR(-ENOMEM);
9753 /* rcu callback to free various structures associated with a task group */
9754 static void free_sched_group_rcu(struct rcu_head *rhp)
9756 /* now it should be safe to free those cfs_rqs */
9757 free_sched_group(container_of(rhp, struct task_group, rcu));
9760 /* Destroy runqueue etc associated with a task group */
9761 void sched_destroy_group(struct task_group *tg)
9763 unsigned long flags;
9766 spin_lock_irqsave(&task_group_lock, flags);
9767 for_each_possible_cpu(i) {
9768 unregister_fair_sched_group(tg, i);
9769 unregister_rt_sched_group(tg, i);
9771 list_del_rcu(&tg->list);
9772 list_del_rcu(&tg->siblings);
9773 spin_unlock_irqrestore(&task_group_lock, flags);
9775 /* wait for possible concurrent references to cfs_rqs complete */
9776 call_rcu(&tg->rcu, free_sched_group_rcu);
9779 /* change task's runqueue when it moves between groups.
9780 * The caller of this function should have put the task in its new group
9781 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9782 * reflect its new group.
9784 void sched_move_task(struct task_struct *tsk)
9787 unsigned long flags;
9790 rq = task_rq_lock(tsk, &flags);
9792 update_rq_clock(rq);
9794 running = task_current(rq, tsk);
9795 on_rq = tsk->se.on_rq;
9798 dequeue_task(rq, tsk, 0);
9799 if (unlikely(running))
9800 tsk->sched_class->put_prev_task(rq, tsk);
9802 set_task_rq(tsk, task_cpu(tsk));
9804 #ifdef CONFIG_FAIR_GROUP_SCHED
9805 if (tsk->sched_class->moved_group)
9806 tsk->sched_class->moved_group(tsk);
9809 if (unlikely(running))
9810 tsk->sched_class->set_curr_task(rq);
9812 enqueue_task(rq, tsk, 0);
9814 task_rq_unlock(rq, &flags);
9816 #endif /* CONFIG_GROUP_SCHED */
9818 #ifdef CONFIG_FAIR_GROUP_SCHED
9819 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9821 struct cfs_rq *cfs_rq = se->cfs_rq;
9826 dequeue_entity(cfs_rq, se, 0);
9828 se->load.weight = shares;
9829 se->load.inv_weight = 0;
9832 enqueue_entity(cfs_rq, se, 0);
9835 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9837 struct cfs_rq *cfs_rq = se->cfs_rq;
9838 struct rq *rq = cfs_rq->rq;
9839 unsigned long flags;
9841 spin_lock_irqsave(&rq->lock, flags);
9842 __set_se_shares(se, shares);
9843 spin_unlock_irqrestore(&rq->lock, flags);
9846 static DEFINE_MUTEX(shares_mutex);
9848 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9851 unsigned long flags;
9854 * We can't change the weight of the root cgroup.
9859 if (shares < MIN_SHARES)
9860 shares = MIN_SHARES;
9861 else if (shares > MAX_SHARES)
9862 shares = MAX_SHARES;
9864 mutex_lock(&shares_mutex);
9865 if (tg->shares == shares)
9868 spin_lock_irqsave(&task_group_lock, flags);
9869 for_each_possible_cpu(i)
9870 unregister_fair_sched_group(tg, i);
9871 list_del_rcu(&tg->siblings);
9872 spin_unlock_irqrestore(&task_group_lock, flags);
9874 /* wait for any ongoing reference to this group to finish */
9875 synchronize_sched();
9878 * Now we are free to modify the group's share on each cpu
9879 * w/o tripping rebalance_share or load_balance_fair.
9881 tg->shares = shares;
9882 for_each_possible_cpu(i) {
9886 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9887 set_se_shares(tg->se[i], shares);
9891 * Enable load balance activity on this group, by inserting it back on
9892 * each cpu's rq->leaf_cfs_rq_list.
9894 spin_lock_irqsave(&task_group_lock, flags);
9895 for_each_possible_cpu(i)
9896 register_fair_sched_group(tg, i);
9897 list_add_rcu(&tg->siblings, &tg->parent->children);
9898 spin_unlock_irqrestore(&task_group_lock, flags);
9900 mutex_unlock(&shares_mutex);
9904 unsigned long sched_group_shares(struct task_group *tg)
9910 #ifdef CONFIG_RT_GROUP_SCHED
9912 * Ensure that the real time constraints are schedulable.
9914 static DEFINE_MUTEX(rt_constraints_mutex);
9916 static unsigned long to_ratio(u64 period, u64 runtime)
9918 if (runtime == RUNTIME_INF)
9921 return div64_u64(runtime << 20, period);
9924 /* Must be called with tasklist_lock held */
9925 static inline int tg_has_rt_tasks(struct task_group *tg)
9927 struct task_struct *g, *p;
9929 do_each_thread(g, p) {
9930 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9932 } while_each_thread(g, p);
9937 struct rt_schedulable_data {
9938 struct task_group *tg;
9943 static int tg_schedulable(struct task_group *tg, void *data)
9945 struct rt_schedulable_data *d = data;
9946 struct task_group *child;
9947 unsigned long total, sum = 0;
9948 u64 period, runtime;
9950 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9951 runtime = tg->rt_bandwidth.rt_runtime;
9954 period = d->rt_period;
9955 runtime = d->rt_runtime;
9958 #ifdef CONFIG_USER_SCHED
9959 if (tg == &root_task_group) {
9960 period = global_rt_period();
9961 runtime = global_rt_runtime();
9966 * Cannot have more runtime than the period.
9968 if (runtime > period && runtime != RUNTIME_INF)
9972 * Ensure we don't starve existing RT tasks.
9974 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9977 total = to_ratio(period, runtime);
9980 * Nobody can have more than the global setting allows.
9982 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9986 * The sum of our children's runtime should not exceed our own.
9988 list_for_each_entry_rcu(child, &tg->children, siblings) {
9989 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9990 runtime = child->rt_bandwidth.rt_runtime;
9992 if (child == d->tg) {
9993 period = d->rt_period;
9994 runtime = d->rt_runtime;
9997 sum += to_ratio(period, runtime);
10006 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10008 struct rt_schedulable_data data = {
10010 .rt_period = period,
10011 .rt_runtime = runtime,
10014 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10017 static int tg_set_bandwidth(struct task_group *tg,
10018 u64 rt_period, u64 rt_runtime)
10022 mutex_lock(&rt_constraints_mutex);
10023 read_lock(&tasklist_lock);
10024 err = __rt_schedulable(tg, rt_period, rt_runtime);
10028 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10029 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10030 tg->rt_bandwidth.rt_runtime = rt_runtime;
10032 for_each_possible_cpu(i) {
10033 struct rt_rq *rt_rq = tg->rt_rq[i];
10035 spin_lock(&rt_rq->rt_runtime_lock);
10036 rt_rq->rt_runtime = rt_runtime;
10037 spin_unlock(&rt_rq->rt_runtime_lock);
10039 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10041 read_unlock(&tasklist_lock);
10042 mutex_unlock(&rt_constraints_mutex);
10047 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10049 u64 rt_runtime, rt_period;
10051 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10052 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10053 if (rt_runtime_us < 0)
10054 rt_runtime = RUNTIME_INF;
10056 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10059 long sched_group_rt_runtime(struct task_group *tg)
10063 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10066 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10067 do_div(rt_runtime_us, NSEC_PER_USEC);
10068 return rt_runtime_us;
10071 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10073 u64 rt_runtime, rt_period;
10075 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10076 rt_runtime = tg->rt_bandwidth.rt_runtime;
10078 if (rt_period == 0)
10081 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10084 long sched_group_rt_period(struct task_group *tg)
10088 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10089 do_div(rt_period_us, NSEC_PER_USEC);
10090 return rt_period_us;
10093 static int sched_rt_global_constraints(void)
10095 u64 runtime, period;
10098 if (sysctl_sched_rt_period <= 0)
10101 runtime = global_rt_runtime();
10102 period = global_rt_period();
10105 * Sanity check on the sysctl variables.
10107 if (runtime > period && runtime != RUNTIME_INF)
10110 mutex_lock(&rt_constraints_mutex);
10111 read_lock(&tasklist_lock);
10112 ret = __rt_schedulable(NULL, 0, 0);
10113 read_unlock(&tasklist_lock);
10114 mutex_unlock(&rt_constraints_mutex);
10119 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10121 /* Don't accept realtime tasks when there is no way for them to run */
10122 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10128 #else /* !CONFIG_RT_GROUP_SCHED */
10129 static int sched_rt_global_constraints(void)
10131 unsigned long flags;
10134 if (sysctl_sched_rt_period <= 0)
10138 * There's always some RT tasks in the root group
10139 * -- migration, kstopmachine etc..
10141 if (sysctl_sched_rt_runtime == 0)
10144 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10145 for_each_possible_cpu(i) {
10146 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10148 spin_lock(&rt_rq->rt_runtime_lock);
10149 rt_rq->rt_runtime = global_rt_runtime();
10150 spin_unlock(&rt_rq->rt_runtime_lock);
10152 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10156 #endif /* CONFIG_RT_GROUP_SCHED */
10158 int sched_rt_handler(struct ctl_table *table, int write,
10159 struct file *filp, void __user *buffer, size_t *lenp,
10163 int old_period, old_runtime;
10164 static DEFINE_MUTEX(mutex);
10166 mutex_lock(&mutex);
10167 old_period = sysctl_sched_rt_period;
10168 old_runtime = sysctl_sched_rt_runtime;
10170 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10172 if (!ret && write) {
10173 ret = sched_rt_global_constraints();
10175 sysctl_sched_rt_period = old_period;
10176 sysctl_sched_rt_runtime = old_runtime;
10178 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10179 def_rt_bandwidth.rt_period =
10180 ns_to_ktime(global_rt_period());
10183 mutex_unlock(&mutex);
10188 #ifdef CONFIG_CGROUP_SCHED
10190 /* return corresponding task_group object of a cgroup */
10191 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10193 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10194 struct task_group, css);
10197 static struct cgroup_subsys_state *
10198 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10200 struct task_group *tg, *parent;
10202 if (!cgrp->parent) {
10203 /* This is early initialization for the top cgroup */
10204 return &init_task_group.css;
10207 parent = cgroup_tg(cgrp->parent);
10208 tg = sched_create_group(parent);
10210 return ERR_PTR(-ENOMEM);
10216 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10218 struct task_group *tg = cgroup_tg(cgrp);
10220 sched_destroy_group(tg);
10224 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10225 struct task_struct *tsk)
10227 #ifdef CONFIG_RT_GROUP_SCHED
10228 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10231 /* We don't support RT-tasks being in separate groups */
10232 if (tsk->sched_class != &fair_sched_class)
10240 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10241 struct cgroup *old_cont, struct task_struct *tsk)
10243 sched_move_task(tsk);
10246 #ifdef CONFIG_FAIR_GROUP_SCHED
10247 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10250 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10253 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10255 struct task_group *tg = cgroup_tg(cgrp);
10257 return (u64) tg->shares;
10259 #endif /* CONFIG_FAIR_GROUP_SCHED */
10261 #ifdef CONFIG_RT_GROUP_SCHED
10262 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10265 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10268 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10270 return sched_group_rt_runtime(cgroup_tg(cgrp));
10273 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10276 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10279 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10281 return sched_group_rt_period(cgroup_tg(cgrp));
10283 #endif /* CONFIG_RT_GROUP_SCHED */
10285 static struct cftype cpu_files[] = {
10286 #ifdef CONFIG_FAIR_GROUP_SCHED
10289 .read_u64 = cpu_shares_read_u64,
10290 .write_u64 = cpu_shares_write_u64,
10293 #ifdef CONFIG_RT_GROUP_SCHED
10295 .name = "rt_runtime_us",
10296 .read_s64 = cpu_rt_runtime_read,
10297 .write_s64 = cpu_rt_runtime_write,
10300 .name = "rt_period_us",
10301 .read_u64 = cpu_rt_period_read_uint,
10302 .write_u64 = cpu_rt_period_write_uint,
10307 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10309 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10312 struct cgroup_subsys cpu_cgroup_subsys = {
10314 .create = cpu_cgroup_create,
10315 .destroy = cpu_cgroup_destroy,
10316 .can_attach = cpu_cgroup_can_attach,
10317 .attach = cpu_cgroup_attach,
10318 .populate = cpu_cgroup_populate,
10319 .subsys_id = cpu_cgroup_subsys_id,
10323 #endif /* CONFIG_CGROUP_SCHED */
10325 #ifdef CONFIG_CGROUP_CPUACCT
10328 * CPU accounting code for task groups.
10330 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10331 * (balbir@in.ibm.com).
10334 /* track cpu usage of a group of tasks and its child groups */
10336 struct cgroup_subsys_state css;
10337 /* cpuusage holds pointer to a u64-type object on every cpu */
10339 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10340 struct cpuacct *parent;
10343 struct cgroup_subsys cpuacct_subsys;
10345 /* return cpu accounting group corresponding to this container */
10346 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10348 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10349 struct cpuacct, css);
10352 /* return cpu accounting group to which this task belongs */
10353 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10355 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10356 struct cpuacct, css);
10359 /* create a new cpu accounting group */
10360 static struct cgroup_subsys_state *cpuacct_create(
10361 struct cgroup_subsys *ss, struct cgroup *cgrp)
10363 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10369 ca->cpuusage = alloc_percpu(u64);
10373 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10374 if (percpu_counter_init(&ca->cpustat[i], 0))
10375 goto out_free_counters;
10378 ca->parent = cgroup_ca(cgrp->parent);
10384 percpu_counter_destroy(&ca->cpustat[i]);
10385 free_percpu(ca->cpuusage);
10389 return ERR_PTR(-ENOMEM);
10392 /* destroy an existing cpu accounting group */
10394 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10396 struct cpuacct *ca = cgroup_ca(cgrp);
10399 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10400 percpu_counter_destroy(&ca->cpustat[i]);
10401 free_percpu(ca->cpuusage);
10405 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10407 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10410 #ifndef CONFIG_64BIT
10412 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10414 spin_lock_irq(&cpu_rq(cpu)->lock);
10416 spin_unlock_irq(&cpu_rq(cpu)->lock);
10424 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10426 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10428 #ifndef CONFIG_64BIT
10430 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10432 spin_lock_irq(&cpu_rq(cpu)->lock);
10434 spin_unlock_irq(&cpu_rq(cpu)->lock);
10440 /* return total cpu usage (in nanoseconds) of a group */
10441 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10443 struct cpuacct *ca = cgroup_ca(cgrp);
10444 u64 totalcpuusage = 0;
10447 for_each_present_cpu(i)
10448 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10450 return totalcpuusage;
10453 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10456 struct cpuacct *ca = cgroup_ca(cgrp);
10465 for_each_present_cpu(i)
10466 cpuacct_cpuusage_write(ca, i, 0);
10472 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10473 struct seq_file *m)
10475 struct cpuacct *ca = cgroup_ca(cgroup);
10479 for_each_present_cpu(i) {
10480 percpu = cpuacct_cpuusage_read(ca, i);
10481 seq_printf(m, "%llu ", (unsigned long long) percpu);
10483 seq_printf(m, "\n");
10487 static const char *cpuacct_stat_desc[] = {
10488 [CPUACCT_STAT_USER] = "user",
10489 [CPUACCT_STAT_SYSTEM] = "system",
10492 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10493 struct cgroup_map_cb *cb)
10495 struct cpuacct *ca = cgroup_ca(cgrp);
10498 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10499 s64 val = percpu_counter_read(&ca->cpustat[i]);
10500 val = cputime64_to_clock_t(val);
10501 cb->fill(cb, cpuacct_stat_desc[i], val);
10506 static struct cftype files[] = {
10509 .read_u64 = cpuusage_read,
10510 .write_u64 = cpuusage_write,
10513 .name = "usage_percpu",
10514 .read_seq_string = cpuacct_percpu_seq_read,
10518 .read_map = cpuacct_stats_show,
10522 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10524 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10528 * charge this task's execution time to its accounting group.
10530 * called with rq->lock held.
10532 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10534 struct cpuacct *ca;
10537 if (unlikely(!cpuacct_subsys.active))
10540 cpu = task_cpu(tsk);
10546 for (; ca; ca = ca->parent) {
10547 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10548 *cpuusage += cputime;
10555 * Charge the system/user time to the task's accounting group.
10557 static void cpuacct_update_stats(struct task_struct *tsk,
10558 enum cpuacct_stat_index idx, cputime_t val)
10560 struct cpuacct *ca;
10562 if (unlikely(!cpuacct_subsys.active))
10569 percpu_counter_add(&ca->cpustat[idx], val);
10575 struct cgroup_subsys cpuacct_subsys = {
10577 .create = cpuacct_create,
10578 .destroy = cpuacct_destroy,
10579 .populate = cpuacct_populate,
10580 .subsys_id = cpuacct_subsys_id,
10582 #endif /* CONFIG_CGROUP_CPUACCT */