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, 0);
245 spin_unlock(&rt_b->rt_runtime_lock);
248 #ifdef CONFIG_RT_GROUP_SCHED
249 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
251 hrtimer_cancel(&rt_b->rt_period_timer);
256 * sched_domains_mutex serializes calls to arch_init_sched_domains,
257 * detach_destroy_domains and partition_sched_domains.
259 static DEFINE_MUTEX(sched_domains_mutex);
261 #ifdef CONFIG_GROUP_SCHED
263 #include <linux/cgroup.h>
267 static LIST_HEAD(task_groups);
269 /* task group related information */
271 #ifdef CONFIG_CGROUP_SCHED
272 struct cgroup_subsys_state css;
275 #ifdef CONFIG_USER_SCHED
279 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* schedulable entities of this group on each cpu */
281 struct sched_entity **se;
282 /* runqueue "owned" by this group on each cpu */
283 struct cfs_rq **cfs_rq;
284 unsigned long shares;
287 #ifdef CONFIG_RT_GROUP_SCHED
288 struct sched_rt_entity **rt_se;
289 struct rt_rq **rt_rq;
291 struct rt_bandwidth rt_bandwidth;
295 struct list_head list;
297 struct task_group *parent;
298 struct list_head siblings;
299 struct list_head children;
302 #ifdef CONFIG_USER_SCHED
304 /* Helper function to pass uid information to create_sched_user() */
305 void set_tg_uid(struct user_struct *user)
307 user->tg->uid = user->uid;
312 * Every UID task group (including init_task_group aka UID-0) will
313 * be a child to this group.
315 struct task_group root_task_group;
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 /* Default task group's sched entity on each cpu */
319 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
320 /* Default task group's cfs_rq on each cpu */
321 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_FAIR_GROUP_SCHED */
324 #ifdef CONFIG_RT_GROUP_SCHED
325 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
326 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
327 #endif /* CONFIG_RT_GROUP_SCHED */
328 #else /* !CONFIG_USER_SCHED */
329 #define root_task_group init_task_group
330 #endif /* CONFIG_USER_SCHED */
332 /* task_group_lock serializes add/remove of task groups and also changes to
333 * a task group's cpu shares.
335 static DEFINE_SPINLOCK(task_group_lock);
338 static int root_task_group_empty(void)
340 return list_empty(&root_task_group.children);
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 #ifdef CONFIG_USER_SCHED
346 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
347 #else /* !CONFIG_USER_SCHED */
348 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
349 #endif /* CONFIG_USER_SCHED */
352 * A weight of 0 or 1 can cause arithmetics problems.
353 * A weight of a cfs_rq is the sum of weights of which entities
354 * are queued on this cfs_rq, so a weight of a entity should not be
355 * too large, so as the shares value of a task group.
356 * (The default weight is 1024 - so there's no practical
357 * limitation from this.)
360 #define MAX_SHARES (1UL << 18)
362 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
365 /* Default task group.
366 * Every task in system belong to this group at bootup.
368 struct task_group init_task_group;
370 /* return group to which a task belongs */
371 static inline struct task_group *task_group(struct task_struct *p)
373 struct task_group *tg;
375 #ifdef CONFIG_USER_SCHED
377 tg = __task_cred(p)->user->tg;
379 #elif defined(CONFIG_CGROUP_SCHED)
380 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
381 struct task_group, css);
383 tg = &init_task_group;
388 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
389 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
393 p->se.parent = task_group(p)->se[cpu];
396 #ifdef CONFIG_RT_GROUP_SCHED
397 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
398 p->rt.parent = task_group(p)->rt_se[cpu];
405 static int root_task_group_empty(void)
411 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
412 static inline struct task_group *task_group(struct task_struct *p)
417 #endif /* CONFIG_GROUP_SCHED */
419 /* CFS-related fields in a runqueue */
421 struct load_weight load;
422 unsigned long nr_running;
427 struct rb_root tasks_timeline;
428 struct rb_node *rb_leftmost;
430 struct list_head tasks;
431 struct list_head *balance_iterator;
434 * 'curr' points to currently running entity on this cfs_rq.
435 * It is set to NULL otherwise (i.e when none are currently running).
437 struct sched_entity *curr, *next, *last;
439 unsigned int nr_spread_over;
441 #ifdef CONFIG_FAIR_GROUP_SCHED
442 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
445 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
446 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447 * (like users, containers etc.)
449 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450 * list is used during load balance.
452 struct list_head leaf_cfs_rq_list;
453 struct task_group *tg; /* group that "owns" this runqueue */
457 * the part of load.weight contributed by tasks
459 unsigned long task_weight;
462 * h_load = weight * f(tg)
464 * Where f(tg) is the recursive weight fraction assigned to
467 unsigned long h_load;
470 * this cpu's part of tg->shares
472 unsigned long shares;
475 * load.weight at the time we set shares
477 unsigned long rq_weight;
482 /* Real-Time classes' related field in a runqueue: */
484 struct rt_prio_array active;
485 unsigned long rt_nr_running;
486 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
488 int curr; /* highest queued rt task prio */
490 int next; /* next highest */
495 unsigned long rt_nr_migratory;
497 struct plist_head pushable_tasks;
502 /* Nests inside the rq lock: */
503 spinlock_t rt_runtime_lock;
505 #ifdef CONFIG_RT_GROUP_SCHED
506 unsigned long rt_nr_boosted;
509 struct list_head leaf_rt_rq_list;
510 struct task_group *tg;
511 struct sched_rt_entity *rt_se;
518 * We add the notion of a root-domain which will be used to define per-domain
519 * variables. Each exclusive cpuset essentially defines an island domain by
520 * fully partitioning the member cpus from any other cpuset. Whenever a new
521 * exclusive cpuset is created, we also create and attach a new root-domain
528 cpumask_var_t online;
531 * The "RT overload" flag: it gets set if a CPU has more than
532 * one runnable RT task.
534 cpumask_var_t rto_mask;
537 struct cpupri cpupri;
539 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
541 * Preferred wake up cpu nominated by sched_mc balance that will be
542 * used when most cpus are idle in the system indicating overall very
543 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
545 unsigned int sched_mc_preferred_wakeup_cpu;
550 * By default the system creates a single root-domain with all cpus as
551 * members (mimicking the global state we have today).
553 static struct root_domain def_root_domain;
558 * This is the main, per-CPU runqueue data structure.
560 * Locking rule: those places that want to lock multiple runqueues
561 * (such as the load balancing or the thread migration code), lock
562 * acquire operations must be ordered by ascending &runqueue.
569 * nr_running and cpu_load should be in the same cacheline because
570 * remote CPUs use both these fields when doing load calculation.
572 unsigned long nr_running;
573 #define CPU_LOAD_IDX_MAX 5
574 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
576 unsigned long last_tick_seen;
577 unsigned char in_nohz_recently;
579 /* capture load from *all* tasks on this cpu: */
580 struct load_weight load;
581 unsigned long nr_load_updates;
583 u64 nr_migrations_in;
588 #ifdef CONFIG_FAIR_GROUP_SCHED
589 /* list of leaf cfs_rq on this cpu: */
590 struct list_head leaf_cfs_rq_list;
592 #ifdef CONFIG_RT_GROUP_SCHED
593 struct list_head leaf_rt_rq_list;
597 * This is part of a global counter where only the total sum
598 * over all CPUs matters. A task can increase this counter on
599 * one CPU and if it got migrated afterwards it may decrease
600 * it on another CPU. Always updated under the runqueue lock:
602 unsigned long nr_uninterruptible;
604 struct task_struct *curr, *idle;
605 unsigned long next_balance;
606 struct mm_struct *prev_mm;
613 struct root_domain *rd;
614 struct sched_domain *sd;
616 unsigned char idle_at_tick;
617 /* For active balancing */
620 /* cpu of this runqueue: */
624 unsigned long avg_load_per_task;
626 struct task_struct *migration_thread;
627 struct list_head migration_queue;
630 /* calc_load related fields */
631 unsigned long calc_load_update;
632 long calc_load_active;
634 #ifdef CONFIG_SCHED_HRTICK
636 int hrtick_csd_pending;
637 struct call_single_data hrtick_csd;
639 struct hrtimer hrtick_timer;
642 #ifdef CONFIG_SCHEDSTATS
644 struct sched_info rq_sched_info;
645 unsigned long long rq_cpu_time;
646 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
648 /* sys_sched_yield() stats */
649 unsigned int yld_count;
651 /* schedule() stats */
652 unsigned int sched_switch;
653 unsigned int sched_count;
654 unsigned int sched_goidle;
656 /* try_to_wake_up() stats */
657 unsigned int ttwu_count;
658 unsigned int ttwu_local;
661 unsigned int bkl_count;
665 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
667 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
669 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
672 static inline int cpu_of(struct rq *rq)
682 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
683 * See detach_destroy_domains: synchronize_sched for details.
685 * The domain tree of any CPU may only be accessed from within
686 * preempt-disabled sections.
688 #define for_each_domain(cpu, __sd) \
689 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
691 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
692 #define this_rq() (&__get_cpu_var(runqueues))
693 #define task_rq(p) cpu_rq(task_cpu(p))
694 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
696 inline void update_rq_clock(struct rq *rq)
698 rq->clock = sched_clock_cpu(cpu_of(rq));
702 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
704 #ifdef CONFIG_SCHED_DEBUG
705 # define const_debug __read_mostly
707 # define const_debug static const
713 * Returns true if the current cpu runqueue is locked.
714 * This interface allows printk to be called with the runqueue lock
715 * held and know whether or not it is OK to wake up the klogd.
717 int runqueue_is_locked(void)
720 struct rq *rq = cpu_rq(cpu);
723 ret = spin_is_locked(&rq->lock);
729 * Debugging: various feature bits
732 #define SCHED_FEAT(name, enabled) \
733 __SCHED_FEAT_##name ,
736 #include "sched_features.h"
741 #define SCHED_FEAT(name, enabled) \
742 (1UL << __SCHED_FEAT_##name) * enabled |
744 const_debug unsigned int sysctl_sched_features =
745 #include "sched_features.h"
750 #ifdef CONFIG_SCHED_DEBUG
751 #define SCHED_FEAT(name, enabled) \
754 static __read_mostly char *sched_feat_names[] = {
755 #include "sched_features.h"
761 static int sched_feat_show(struct seq_file *m, void *v)
765 for (i = 0; sched_feat_names[i]; i++) {
766 if (!(sysctl_sched_features & (1UL << i)))
768 seq_printf(m, "%s ", sched_feat_names[i]);
776 sched_feat_write(struct file *filp, const char __user *ubuf,
777 size_t cnt, loff_t *ppos)
787 if (copy_from_user(&buf, ubuf, cnt))
792 if (strncmp(buf, "NO_", 3) == 0) {
797 for (i = 0; sched_feat_names[i]; i++) {
798 int len = strlen(sched_feat_names[i]);
800 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
802 sysctl_sched_features &= ~(1UL << i);
804 sysctl_sched_features |= (1UL << i);
809 if (!sched_feat_names[i])
817 static int sched_feat_open(struct inode *inode, struct file *filp)
819 return single_open(filp, sched_feat_show, NULL);
822 static struct file_operations sched_feat_fops = {
823 .open = sched_feat_open,
824 .write = sched_feat_write,
827 .release = single_release,
830 static __init int sched_init_debug(void)
832 debugfs_create_file("sched_features", 0644, NULL, NULL,
837 late_initcall(sched_init_debug);
841 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
844 * Number of tasks to iterate in a single balance run.
845 * Limited because this is done with IRQs disabled.
847 const_debug unsigned int sysctl_sched_nr_migrate = 32;
850 * ratelimit for updating the group shares.
853 unsigned int sysctl_sched_shares_ratelimit = 250000;
856 * Inject some fuzzyness into changing the per-cpu group shares
857 * this avoids remote rq-locks at the expense of fairness.
860 unsigned int sysctl_sched_shares_thresh = 4;
863 * period over which we measure -rt task cpu usage in us.
866 unsigned int sysctl_sched_rt_period = 1000000;
868 static __read_mostly int scheduler_running;
871 * part of the period that we allow rt tasks to run in us.
874 int sysctl_sched_rt_runtime = 950000;
876 static inline u64 global_rt_period(void)
878 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
881 static inline u64 global_rt_runtime(void)
883 if (sysctl_sched_rt_runtime < 0)
886 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
889 #ifndef prepare_arch_switch
890 # define prepare_arch_switch(next) do { } while (0)
892 #ifndef finish_arch_switch
893 # define finish_arch_switch(prev) do { } while (0)
896 static inline int task_current(struct rq *rq, struct task_struct *p)
898 return rq->curr == p;
901 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
902 static inline int task_running(struct rq *rq, struct task_struct *p)
904 return task_current(rq, p);
907 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
911 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
913 #ifdef CONFIG_DEBUG_SPINLOCK
914 /* this is a valid case when another task releases the spinlock */
915 rq->lock.owner = current;
918 * If we are tracking spinlock dependencies then we have to
919 * fix up the runqueue lock - which gets 'carried over' from
922 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
924 spin_unlock_irq(&rq->lock);
927 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
928 static inline int task_running(struct rq *rq, struct task_struct *p)
933 return task_current(rq, p);
937 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
941 * We can optimise this out completely for !SMP, because the
942 * SMP rebalancing from interrupt is the only thing that cares
947 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
948 spin_unlock_irq(&rq->lock);
950 spin_unlock(&rq->lock);
954 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
958 * After ->oncpu is cleared, the task can be moved to a different CPU.
959 * We must ensure this doesn't happen until the switch is completely
965 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
969 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
972 * __task_rq_lock - lock the runqueue a given task resides on.
973 * Must be called interrupts disabled.
975 static inline struct rq *__task_rq_lock(struct task_struct *p)
979 struct rq *rq = task_rq(p);
980 spin_lock(&rq->lock);
981 if (likely(rq == task_rq(p)))
983 spin_unlock(&rq->lock);
988 * task_rq_lock - lock the runqueue a given task resides on and disable
989 * interrupts. Note the ordering: we can safely lookup the task_rq without
990 * explicitly disabling preemption.
992 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
998 local_irq_save(*flags);
1000 spin_lock(&rq->lock);
1001 if (likely(rq == task_rq(p)))
1003 spin_unlock_irqrestore(&rq->lock, *flags);
1007 void task_rq_unlock_wait(struct task_struct *p)
1009 struct rq *rq = task_rq(p);
1011 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1012 spin_unlock_wait(&rq->lock);
1015 static void __task_rq_unlock(struct rq *rq)
1016 __releases(rq->lock)
1018 spin_unlock(&rq->lock);
1021 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1022 __releases(rq->lock)
1024 spin_unlock_irqrestore(&rq->lock, *flags);
1028 * this_rq_lock - lock this runqueue and disable interrupts.
1030 static struct rq *this_rq_lock(void)
1031 __acquires(rq->lock)
1035 local_irq_disable();
1037 spin_lock(&rq->lock);
1042 #ifdef CONFIG_SCHED_HRTICK
1044 * Use HR-timers to deliver accurate preemption points.
1046 * Its all a bit involved since we cannot program an hrt while holding the
1047 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1050 * When we get rescheduled we reprogram the hrtick_timer outside of the
1056 * - enabled by features
1057 * - hrtimer is actually high res
1059 static inline int hrtick_enabled(struct rq *rq)
1061 if (!sched_feat(HRTICK))
1063 if (!cpu_active(cpu_of(rq)))
1065 return hrtimer_is_hres_active(&rq->hrtick_timer);
1068 static void hrtick_clear(struct rq *rq)
1070 if (hrtimer_active(&rq->hrtick_timer))
1071 hrtimer_cancel(&rq->hrtick_timer);
1075 * High-resolution timer tick.
1076 * Runs from hardirq context with interrupts disabled.
1078 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1080 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1082 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1084 spin_lock(&rq->lock);
1085 update_rq_clock(rq);
1086 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1087 spin_unlock(&rq->lock);
1089 return HRTIMER_NORESTART;
1094 * called from hardirq (IPI) context
1096 static void __hrtick_start(void *arg)
1098 struct rq *rq = arg;
1100 spin_lock(&rq->lock);
1101 hrtimer_restart(&rq->hrtick_timer);
1102 rq->hrtick_csd_pending = 0;
1103 spin_unlock(&rq->lock);
1107 * Called to set the hrtick timer state.
1109 * called with rq->lock held and irqs disabled
1111 static void hrtick_start(struct rq *rq, u64 delay)
1113 struct hrtimer *timer = &rq->hrtick_timer;
1114 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1116 hrtimer_set_expires(timer, time);
1118 if (rq == this_rq()) {
1119 hrtimer_restart(timer);
1120 } else if (!rq->hrtick_csd_pending) {
1121 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1122 rq->hrtick_csd_pending = 1;
1127 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1129 int cpu = (int)(long)hcpu;
1132 case CPU_UP_CANCELED:
1133 case CPU_UP_CANCELED_FROZEN:
1134 case CPU_DOWN_PREPARE:
1135 case CPU_DOWN_PREPARE_FROZEN:
1137 case CPU_DEAD_FROZEN:
1138 hrtick_clear(cpu_rq(cpu));
1145 static __init void init_hrtick(void)
1147 hotcpu_notifier(hotplug_hrtick, 0);
1151 * Called to set the hrtick timer state.
1153 * called with rq->lock held and irqs disabled
1155 static void hrtick_start(struct rq *rq, u64 delay)
1157 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1158 HRTIMER_MODE_REL, 0);
1161 static inline void init_hrtick(void)
1164 #endif /* CONFIG_SMP */
1166 static void init_rq_hrtick(struct rq *rq)
1169 rq->hrtick_csd_pending = 0;
1171 rq->hrtick_csd.flags = 0;
1172 rq->hrtick_csd.func = __hrtick_start;
1173 rq->hrtick_csd.info = rq;
1176 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1177 rq->hrtick_timer.function = hrtick;
1179 #else /* CONFIG_SCHED_HRTICK */
1180 static inline void hrtick_clear(struct rq *rq)
1184 static inline void init_rq_hrtick(struct rq *rq)
1188 static inline void init_hrtick(void)
1191 #endif /* CONFIG_SCHED_HRTICK */
1194 * resched_task - mark a task 'to be rescheduled now'.
1196 * On UP this means the setting of the need_resched flag, on SMP it
1197 * might also involve a cross-CPU call to trigger the scheduler on
1202 #ifndef tsk_is_polling
1203 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1206 static void resched_task(struct task_struct *p)
1210 assert_spin_locked(&task_rq(p)->lock);
1212 if (test_tsk_need_resched(p))
1215 set_tsk_need_resched(p);
1218 if (cpu == smp_processor_id())
1221 /* NEED_RESCHED must be visible before we test polling */
1223 if (!tsk_is_polling(p))
1224 smp_send_reschedule(cpu);
1227 static void resched_cpu(int cpu)
1229 struct rq *rq = cpu_rq(cpu);
1230 unsigned long flags;
1232 if (!spin_trylock_irqsave(&rq->lock, flags))
1234 resched_task(cpu_curr(cpu));
1235 spin_unlock_irqrestore(&rq->lock, flags);
1240 * When add_timer_on() enqueues a timer into the timer wheel of an
1241 * idle CPU then this timer might expire before the next timer event
1242 * which is scheduled to wake up that CPU. In case of a completely
1243 * idle system the next event might even be infinite time into the
1244 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1245 * leaves the inner idle loop so the newly added timer is taken into
1246 * account when the CPU goes back to idle and evaluates the timer
1247 * wheel for the next timer event.
1249 void wake_up_idle_cpu(int cpu)
1251 struct rq *rq = cpu_rq(cpu);
1253 if (cpu == smp_processor_id())
1257 * This is safe, as this function is called with the timer
1258 * wheel base lock of (cpu) held. When the CPU is on the way
1259 * to idle and has not yet set rq->curr to idle then it will
1260 * be serialized on the timer wheel base lock and take the new
1261 * timer into account automatically.
1263 if (rq->curr != rq->idle)
1267 * We can set TIF_RESCHED on the idle task of the other CPU
1268 * lockless. The worst case is that the other CPU runs the
1269 * idle task through an additional NOOP schedule()
1271 set_tsk_need_resched(rq->idle);
1273 /* NEED_RESCHED must be visible before we test polling */
1275 if (!tsk_is_polling(rq->idle))
1276 smp_send_reschedule(cpu);
1278 #endif /* CONFIG_NO_HZ */
1280 #else /* !CONFIG_SMP */
1281 static void resched_task(struct task_struct *p)
1283 assert_spin_locked(&task_rq(p)->lock);
1284 set_tsk_need_resched(p);
1286 #endif /* CONFIG_SMP */
1288 #if BITS_PER_LONG == 32
1289 # define WMULT_CONST (~0UL)
1291 # define WMULT_CONST (1UL << 32)
1294 #define WMULT_SHIFT 32
1297 * Shift right and round:
1299 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1302 * delta *= weight / lw
1304 static unsigned long
1305 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1306 struct load_weight *lw)
1310 if (!lw->inv_weight) {
1311 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1314 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1318 tmp = (u64)delta_exec * weight;
1320 * Check whether we'd overflow the 64-bit multiplication:
1322 if (unlikely(tmp > WMULT_CONST))
1323 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1326 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1328 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1331 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1337 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1344 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1345 * of tasks with abnormal "nice" values across CPUs the contribution that
1346 * each task makes to its run queue's load is weighted according to its
1347 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1348 * scaled version of the new time slice allocation that they receive on time
1352 #define WEIGHT_IDLEPRIO 3
1353 #define WMULT_IDLEPRIO 1431655765
1356 * Nice levels are multiplicative, with a gentle 10% change for every
1357 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1358 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1359 * that remained on nice 0.
1361 * The "10% effect" is relative and cumulative: from _any_ nice level,
1362 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1363 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1364 * If a task goes up by ~10% and another task goes down by ~10% then
1365 * the relative distance between them is ~25%.)
1367 static const int prio_to_weight[40] = {
1368 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1369 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1370 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1371 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1372 /* 0 */ 1024, 820, 655, 526, 423,
1373 /* 5 */ 335, 272, 215, 172, 137,
1374 /* 10 */ 110, 87, 70, 56, 45,
1375 /* 15 */ 36, 29, 23, 18, 15,
1379 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1381 * In cases where the weight does not change often, we can use the
1382 * precalculated inverse to speed up arithmetics by turning divisions
1383 * into multiplications:
1385 static const u32 prio_to_wmult[40] = {
1386 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1387 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1388 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1389 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1390 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1391 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1392 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1393 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1396 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1399 * runqueue iterator, to support SMP load-balancing between different
1400 * scheduling classes, without having to expose their internal data
1401 * structures to the load-balancing proper:
1403 struct rq_iterator {
1405 struct task_struct *(*start)(void *);
1406 struct task_struct *(*next)(void *);
1410 static unsigned long
1411 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1412 unsigned long max_load_move, struct sched_domain *sd,
1413 enum cpu_idle_type idle, int *all_pinned,
1414 int *this_best_prio, struct rq_iterator *iterator);
1417 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1418 struct sched_domain *sd, enum cpu_idle_type idle,
1419 struct rq_iterator *iterator);
1422 /* Time spent by the tasks of the cpu accounting group executing in ... */
1423 enum cpuacct_stat_index {
1424 CPUACCT_STAT_USER, /* ... user mode */
1425 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1427 CPUACCT_STAT_NSTATS,
1430 #ifdef CONFIG_CGROUP_CPUACCT
1431 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1432 static void cpuacct_update_stats(struct task_struct *tsk,
1433 enum cpuacct_stat_index idx, cputime_t val);
1435 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1436 static inline void cpuacct_update_stats(struct task_struct *tsk,
1437 enum cpuacct_stat_index idx, cputime_t val) {}
1440 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1442 update_load_add(&rq->load, load);
1445 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1447 update_load_sub(&rq->load, load);
1450 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1451 typedef int (*tg_visitor)(struct task_group *, void *);
1454 * Iterate the full tree, calling @down when first entering a node and @up when
1455 * leaving it for the final time.
1457 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1459 struct task_group *parent, *child;
1463 parent = &root_task_group;
1465 ret = (*down)(parent, data);
1468 list_for_each_entry_rcu(child, &parent->children, siblings) {
1475 ret = (*up)(parent, data);
1480 parent = parent->parent;
1489 static int tg_nop(struct task_group *tg, void *data)
1496 static unsigned long source_load(int cpu, int type);
1497 static unsigned long target_load(int cpu, int type);
1498 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1500 static unsigned long cpu_avg_load_per_task(int cpu)
1502 struct rq *rq = cpu_rq(cpu);
1503 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1506 rq->avg_load_per_task = rq->load.weight / nr_running;
1508 rq->avg_load_per_task = 0;
1510 return rq->avg_load_per_task;
1513 #ifdef CONFIG_FAIR_GROUP_SCHED
1515 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1518 * Calculate and set the cpu's group shares.
1521 update_group_shares_cpu(struct task_group *tg, int cpu,
1522 unsigned long sd_shares, unsigned long sd_rq_weight)
1524 unsigned long shares;
1525 unsigned long rq_weight;
1530 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1533 * \Sum shares * rq_weight
1534 * shares = -----------------------
1538 shares = (sd_shares * rq_weight) / sd_rq_weight;
1539 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1541 if (abs(shares - tg->se[cpu]->load.weight) >
1542 sysctl_sched_shares_thresh) {
1543 struct rq *rq = cpu_rq(cpu);
1544 unsigned long flags;
1546 spin_lock_irqsave(&rq->lock, flags);
1547 tg->cfs_rq[cpu]->shares = shares;
1549 __set_se_shares(tg->se[cpu], shares);
1550 spin_unlock_irqrestore(&rq->lock, flags);
1555 * Re-compute the task group their per cpu shares over the given domain.
1556 * This needs to be done in a bottom-up fashion because the rq weight of a
1557 * parent group depends on the shares of its child groups.
1559 static int tg_shares_up(struct task_group *tg, void *data)
1561 unsigned long weight, rq_weight = 0;
1562 unsigned long shares = 0;
1563 struct sched_domain *sd = data;
1566 for_each_cpu(i, sched_domain_span(sd)) {
1568 * If there are currently no tasks on the cpu pretend there
1569 * is one of average load so that when a new task gets to
1570 * run here it will not get delayed by group starvation.
1572 weight = tg->cfs_rq[i]->load.weight;
1574 weight = NICE_0_LOAD;
1576 tg->cfs_rq[i]->rq_weight = weight;
1577 rq_weight += weight;
1578 shares += tg->cfs_rq[i]->shares;
1581 if ((!shares && rq_weight) || shares > tg->shares)
1582 shares = tg->shares;
1584 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1585 shares = tg->shares;
1587 for_each_cpu(i, sched_domain_span(sd))
1588 update_group_shares_cpu(tg, i, shares, rq_weight);
1594 * Compute the cpu's hierarchical load factor for each task group.
1595 * This needs to be done in a top-down fashion because the load of a child
1596 * group is a fraction of its parents load.
1598 static int tg_load_down(struct task_group *tg, void *data)
1601 long cpu = (long)data;
1604 load = cpu_rq(cpu)->load.weight;
1606 load = tg->parent->cfs_rq[cpu]->h_load;
1607 load *= tg->cfs_rq[cpu]->shares;
1608 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1611 tg->cfs_rq[cpu]->h_load = load;
1616 static void update_shares(struct sched_domain *sd)
1618 u64 now = cpu_clock(raw_smp_processor_id());
1619 s64 elapsed = now - sd->last_update;
1621 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1622 sd->last_update = now;
1623 walk_tg_tree(tg_nop, tg_shares_up, sd);
1627 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1629 spin_unlock(&rq->lock);
1631 spin_lock(&rq->lock);
1634 static void update_h_load(long cpu)
1636 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1641 static inline void update_shares(struct sched_domain *sd)
1645 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1651 #ifdef CONFIG_PREEMPT
1654 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1655 * way at the expense of forcing extra atomic operations in all
1656 * invocations. This assures that the double_lock is acquired using the
1657 * same underlying policy as the spinlock_t on this architecture, which
1658 * reduces latency compared to the unfair variant below. However, it
1659 * also adds more overhead and therefore may reduce throughput.
1661 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1662 __releases(this_rq->lock)
1663 __acquires(busiest->lock)
1664 __acquires(this_rq->lock)
1666 spin_unlock(&this_rq->lock);
1667 double_rq_lock(this_rq, busiest);
1674 * Unfair double_lock_balance: Optimizes throughput at the expense of
1675 * latency by eliminating extra atomic operations when the locks are
1676 * already in proper order on entry. This favors lower cpu-ids and will
1677 * grant the double lock to lower cpus over higher ids under contention,
1678 * regardless of entry order into the function.
1680 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1681 __releases(this_rq->lock)
1682 __acquires(busiest->lock)
1683 __acquires(this_rq->lock)
1687 if (unlikely(!spin_trylock(&busiest->lock))) {
1688 if (busiest < this_rq) {
1689 spin_unlock(&this_rq->lock);
1690 spin_lock(&busiest->lock);
1691 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1694 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1699 #endif /* CONFIG_PREEMPT */
1702 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1704 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1706 if (unlikely(!irqs_disabled())) {
1707 /* printk() doesn't work good under rq->lock */
1708 spin_unlock(&this_rq->lock);
1712 return _double_lock_balance(this_rq, busiest);
1715 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1716 __releases(busiest->lock)
1718 spin_unlock(&busiest->lock);
1719 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1723 #ifdef CONFIG_FAIR_GROUP_SCHED
1724 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1727 cfs_rq->shares = shares;
1732 static void calc_load_account_active(struct rq *this_rq);
1734 #include "sched_stats.h"
1735 #include "sched_idletask.c"
1736 #include "sched_fair.c"
1737 #include "sched_rt.c"
1738 #ifdef CONFIG_SCHED_DEBUG
1739 # include "sched_debug.c"
1742 #define sched_class_highest (&rt_sched_class)
1743 #define for_each_class(class) \
1744 for (class = sched_class_highest; class; class = class->next)
1746 static void inc_nr_running(struct rq *rq)
1751 static void dec_nr_running(struct rq *rq)
1756 static void set_load_weight(struct task_struct *p)
1758 if (task_has_rt_policy(p)) {
1759 p->se.load.weight = prio_to_weight[0] * 2;
1760 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1765 * SCHED_IDLE tasks get minimal weight:
1767 if (p->policy == SCHED_IDLE) {
1768 p->se.load.weight = WEIGHT_IDLEPRIO;
1769 p->se.load.inv_weight = WMULT_IDLEPRIO;
1773 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1774 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1777 static void update_avg(u64 *avg, u64 sample)
1779 s64 diff = sample - *avg;
1783 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1786 p->se.start_runtime = p->se.sum_exec_runtime;
1788 sched_info_queued(p);
1789 p->sched_class->enqueue_task(rq, p, wakeup);
1793 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1796 if (p->se.last_wakeup) {
1797 update_avg(&p->se.avg_overlap,
1798 p->se.sum_exec_runtime - p->se.last_wakeup);
1799 p->se.last_wakeup = 0;
1801 update_avg(&p->se.avg_wakeup,
1802 sysctl_sched_wakeup_granularity);
1806 sched_info_dequeued(p);
1807 p->sched_class->dequeue_task(rq, p, sleep);
1812 * __normal_prio - return the priority that is based on the static prio
1814 static inline int __normal_prio(struct task_struct *p)
1816 return p->static_prio;
1820 * Calculate the expected normal priority: i.e. priority
1821 * without taking RT-inheritance into account. Might be
1822 * boosted by interactivity modifiers. Changes upon fork,
1823 * setprio syscalls, and whenever the interactivity
1824 * estimator recalculates.
1826 static inline int normal_prio(struct task_struct *p)
1830 if (task_has_rt_policy(p))
1831 prio = MAX_RT_PRIO-1 - p->rt_priority;
1833 prio = __normal_prio(p);
1838 * Calculate the current priority, i.e. the priority
1839 * taken into account by the scheduler. This value might
1840 * be boosted by RT tasks, or might be boosted by
1841 * interactivity modifiers. Will be RT if the task got
1842 * RT-boosted. If not then it returns p->normal_prio.
1844 static int effective_prio(struct task_struct *p)
1846 p->normal_prio = normal_prio(p);
1848 * If we are RT tasks or we were boosted to RT priority,
1849 * keep the priority unchanged. Otherwise, update priority
1850 * to the normal priority:
1852 if (!rt_prio(p->prio))
1853 return p->normal_prio;
1858 * activate_task - move a task to the runqueue.
1860 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1862 if (task_contributes_to_load(p))
1863 rq->nr_uninterruptible--;
1865 enqueue_task(rq, p, wakeup);
1870 * deactivate_task - remove a task from the runqueue.
1872 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1874 if (task_contributes_to_load(p))
1875 rq->nr_uninterruptible++;
1877 dequeue_task(rq, p, sleep);
1882 * task_curr - is this task currently executing on a CPU?
1883 * @p: the task in question.
1885 inline int task_curr(const struct task_struct *p)
1887 return cpu_curr(task_cpu(p)) == p;
1890 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1892 set_task_rq(p, cpu);
1895 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1896 * successfuly executed on another CPU. We must ensure that updates of
1897 * per-task data have been completed by this moment.
1900 task_thread_info(p)->cpu = cpu;
1904 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1905 const struct sched_class *prev_class,
1906 int oldprio, int running)
1908 if (prev_class != p->sched_class) {
1909 if (prev_class->switched_from)
1910 prev_class->switched_from(rq, p, running);
1911 p->sched_class->switched_to(rq, p, running);
1913 p->sched_class->prio_changed(rq, p, oldprio, running);
1918 /* Used instead of source_load when we know the type == 0 */
1919 static unsigned long weighted_cpuload(const int cpu)
1921 return cpu_rq(cpu)->load.weight;
1925 * Is this task likely cache-hot:
1928 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1933 * Buddy candidates are cache hot:
1935 if (sched_feat(CACHE_HOT_BUDDY) &&
1936 (&p->se == cfs_rq_of(&p->se)->next ||
1937 &p->se == cfs_rq_of(&p->se)->last))
1940 if (p->sched_class != &fair_sched_class)
1943 if (sysctl_sched_migration_cost == -1)
1945 if (sysctl_sched_migration_cost == 0)
1948 delta = now - p->se.exec_start;
1950 return delta < (s64)sysctl_sched_migration_cost;
1954 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1956 int old_cpu = task_cpu(p);
1957 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1958 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1959 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1962 clock_offset = old_rq->clock - new_rq->clock;
1964 trace_sched_migrate_task(p, new_cpu);
1966 #ifdef CONFIG_SCHEDSTATS
1967 if (p->se.wait_start)
1968 p->se.wait_start -= clock_offset;
1969 if (p->se.sleep_start)
1970 p->se.sleep_start -= clock_offset;
1971 if (p->se.block_start)
1972 p->se.block_start -= clock_offset;
1974 if (old_cpu != new_cpu) {
1975 p->se.nr_migrations++;
1976 new_rq->nr_migrations_in++;
1977 #ifdef CONFIG_SCHEDSTATS
1978 if (task_hot(p, old_rq->clock, NULL))
1979 schedstat_inc(p, se.nr_forced2_migrations);
1981 perf_counter_task_migration(p, new_cpu);
1983 p->se.vruntime -= old_cfsrq->min_vruntime -
1984 new_cfsrq->min_vruntime;
1986 __set_task_cpu(p, new_cpu);
1989 struct migration_req {
1990 struct list_head list;
1992 struct task_struct *task;
1995 struct completion done;
1999 * The task's runqueue lock must be held.
2000 * Returns true if you have to wait for migration thread.
2003 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2005 struct rq *rq = task_rq(p);
2008 * If the task is not on a runqueue (and not running), then
2009 * it is sufficient to simply update the task's cpu field.
2011 if (!p->se.on_rq && !task_running(rq, p)) {
2012 set_task_cpu(p, dest_cpu);
2016 init_completion(&req->done);
2018 req->dest_cpu = dest_cpu;
2019 list_add(&req->list, &rq->migration_queue);
2025 * wait_task_context_switch - wait for a thread to complete at least one
2028 * @p must not be current.
2030 void wait_task_context_switch(struct task_struct *p)
2032 unsigned long nvcsw, nivcsw, flags;
2040 * The runqueue is assigned before the actual context
2041 * switch. We need to take the runqueue lock.
2043 * We could check initially without the lock but it is
2044 * very likely that we need to take the lock in every
2047 rq = task_rq_lock(p, &flags);
2048 running = task_running(rq, p);
2049 task_rq_unlock(rq, &flags);
2051 if (likely(!running))
2054 * The switch count is incremented before the actual
2055 * context switch. We thus wait for two switches to be
2056 * sure at least one completed.
2058 if ((p->nvcsw - nvcsw) > 1)
2060 if ((p->nivcsw - nivcsw) > 1)
2068 * wait_task_inactive - wait for a thread to unschedule.
2070 * If @match_state is nonzero, it's the @p->state value just checked and
2071 * not expected to change. If it changes, i.e. @p might have woken up,
2072 * then return zero. When we succeed in waiting for @p to be off its CPU,
2073 * we return a positive number (its total switch count). If a second call
2074 * a short while later returns the same number, the caller can be sure that
2075 * @p has remained unscheduled the whole time.
2077 * The caller must ensure that the task *will* unschedule sometime soon,
2078 * else this function might spin for a *long* time. This function can't
2079 * be called with interrupts off, or it may introduce deadlock with
2080 * smp_call_function() if an IPI is sent by the same process we are
2081 * waiting to become inactive.
2083 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2085 unsigned long flags;
2092 * We do the initial early heuristics without holding
2093 * any task-queue locks at all. We'll only try to get
2094 * the runqueue lock when things look like they will
2100 * If the task is actively running on another CPU
2101 * still, just relax and busy-wait without holding
2104 * NOTE! Since we don't hold any locks, it's not
2105 * even sure that "rq" stays as the right runqueue!
2106 * But we don't care, since "task_running()" will
2107 * return false if the runqueue has changed and p
2108 * is actually now running somewhere else!
2110 while (task_running(rq, p)) {
2111 if (match_state && unlikely(p->state != match_state))
2117 * Ok, time to look more closely! We need the rq
2118 * lock now, to be *sure*. If we're wrong, we'll
2119 * just go back and repeat.
2121 rq = task_rq_lock(p, &flags);
2122 trace_sched_wait_task(rq, p);
2123 running = task_running(rq, p);
2124 on_rq = p->se.on_rq;
2126 if (!match_state || p->state == match_state)
2127 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2128 task_rq_unlock(rq, &flags);
2131 * If it changed from the expected state, bail out now.
2133 if (unlikely(!ncsw))
2137 * Was it really running after all now that we
2138 * checked with the proper locks actually held?
2140 * Oops. Go back and try again..
2142 if (unlikely(running)) {
2148 * It's not enough that it's not actively running,
2149 * it must be off the runqueue _entirely_, and not
2152 * So if it was still runnable (but just not actively
2153 * running right now), it's preempted, and we should
2154 * yield - it could be a while.
2156 if (unlikely(on_rq)) {
2157 schedule_timeout_uninterruptible(1);
2162 * Ahh, all good. It wasn't running, and it wasn't
2163 * runnable, which means that it will never become
2164 * running in the future either. We're all done!
2173 * kick_process - kick a running thread to enter/exit the kernel
2174 * @p: the to-be-kicked thread
2176 * Cause a process which is running on another CPU to enter
2177 * kernel-mode, without any delay. (to get signals handled.)
2179 * NOTE: this function doesnt have to take the runqueue lock,
2180 * because all it wants to ensure is that the remote task enters
2181 * the kernel. If the IPI races and the task has been migrated
2182 * to another CPU then no harm is done and the purpose has been
2185 void kick_process(struct task_struct *p)
2191 if ((cpu != smp_processor_id()) && task_curr(p))
2192 smp_send_reschedule(cpu);
2195 EXPORT_SYMBOL_GPL(kick_process);
2198 * Return a low guess at the load of a migration-source cpu weighted
2199 * according to the scheduling class and "nice" value.
2201 * We want to under-estimate the load of migration sources, to
2202 * balance conservatively.
2204 static unsigned long source_load(int cpu, int type)
2206 struct rq *rq = cpu_rq(cpu);
2207 unsigned long total = weighted_cpuload(cpu);
2209 if (type == 0 || !sched_feat(LB_BIAS))
2212 return min(rq->cpu_load[type-1], total);
2216 * Return a high guess at the load of a migration-target cpu weighted
2217 * according to the scheduling class and "nice" value.
2219 static unsigned long target_load(int cpu, int type)
2221 struct rq *rq = cpu_rq(cpu);
2222 unsigned long total = weighted_cpuload(cpu);
2224 if (type == 0 || !sched_feat(LB_BIAS))
2227 return max(rq->cpu_load[type-1], total);
2231 * find_idlest_group finds and returns the least busy CPU group within the
2234 static struct sched_group *
2235 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2237 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2238 unsigned long min_load = ULONG_MAX, this_load = 0;
2239 int load_idx = sd->forkexec_idx;
2240 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2243 unsigned long load, avg_load;
2247 /* Skip over this group if it has no CPUs allowed */
2248 if (!cpumask_intersects(sched_group_cpus(group),
2252 local_group = cpumask_test_cpu(this_cpu,
2253 sched_group_cpus(group));
2255 /* Tally up the load of all CPUs in the group */
2258 for_each_cpu(i, sched_group_cpus(group)) {
2259 /* Bias balancing toward cpus of our domain */
2261 load = source_load(i, load_idx);
2263 load = target_load(i, load_idx);
2268 /* Adjust by relative CPU power of the group */
2269 avg_load = sg_div_cpu_power(group,
2270 avg_load * SCHED_LOAD_SCALE);
2273 this_load = avg_load;
2275 } else if (avg_load < min_load) {
2276 min_load = avg_load;
2279 } while (group = group->next, group != sd->groups);
2281 if (!idlest || 100*this_load < imbalance*min_load)
2287 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2290 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2292 unsigned long load, min_load = ULONG_MAX;
2296 /* Traverse only the allowed CPUs */
2297 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2298 load = weighted_cpuload(i);
2300 if (load < min_load || (load == min_load && i == this_cpu)) {
2310 * sched_balance_self: balance the current task (running on cpu) in domains
2311 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2314 * Balance, ie. select the least loaded group.
2316 * Returns the target CPU number, or the same CPU if no balancing is needed.
2318 * preempt must be disabled.
2320 static int sched_balance_self(int cpu, int flag)
2322 struct task_struct *t = current;
2323 struct sched_domain *tmp, *sd = NULL;
2325 for_each_domain(cpu, tmp) {
2327 * If power savings logic is enabled for a domain, stop there.
2329 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2331 if (tmp->flags & flag)
2339 struct sched_group *group;
2340 int new_cpu, weight;
2342 if (!(sd->flags & flag)) {
2347 group = find_idlest_group(sd, t, cpu);
2353 new_cpu = find_idlest_cpu(group, t, cpu);
2354 if (new_cpu == -1 || new_cpu == cpu) {
2355 /* Now try balancing at a lower domain level of cpu */
2360 /* Now try balancing at a lower domain level of new_cpu */
2362 weight = cpumask_weight(sched_domain_span(sd));
2364 for_each_domain(cpu, tmp) {
2365 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2367 if (tmp->flags & flag)
2370 /* while loop will break here if sd == NULL */
2376 #endif /* CONFIG_SMP */
2379 * task_oncpu_function_call - call a function on the cpu on which a task runs
2380 * @p: the task to evaluate
2381 * @func: the function to be called
2382 * @info: the function call argument
2384 * Calls the function @func when the task is currently running. This might
2385 * be on the current CPU, which just calls the function directly
2387 void task_oncpu_function_call(struct task_struct *p,
2388 void (*func) (void *info), void *info)
2395 smp_call_function_single(cpu, func, info, 1);
2400 * try_to_wake_up - wake up a thread
2401 * @p: the to-be-woken-up thread
2402 * @state: the mask of task states that can be woken
2403 * @sync: do a synchronous wakeup?
2405 * Put it on the run-queue if it's not already there. The "current"
2406 * thread is always on the run-queue (except when the actual
2407 * re-schedule is in progress), and as such you're allowed to do
2408 * the simpler "current->state = TASK_RUNNING" to mark yourself
2409 * runnable without the overhead of this.
2411 * returns failure only if the task is already active.
2413 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2415 int cpu, orig_cpu, this_cpu, success = 0;
2416 unsigned long flags;
2420 if (!sched_feat(SYNC_WAKEUPS))
2424 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2425 struct sched_domain *sd;
2427 this_cpu = raw_smp_processor_id();
2430 for_each_domain(this_cpu, sd) {
2431 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2440 rq = task_rq_lock(p, &flags);
2441 update_rq_clock(rq);
2442 old_state = p->state;
2443 if (!(old_state & state))
2451 this_cpu = smp_processor_id();
2454 if (unlikely(task_running(rq, p)))
2457 cpu = p->sched_class->select_task_rq(p, sync);
2458 if (cpu != orig_cpu) {
2459 set_task_cpu(p, cpu);
2460 task_rq_unlock(rq, &flags);
2461 /* might preempt at this point */
2462 rq = task_rq_lock(p, &flags);
2463 old_state = p->state;
2464 if (!(old_state & state))
2469 this_cpu = smp_processor_id();
2473 #ifdef CONFIG_SCHEDSTATS
2474 schedstat_inc(rq, ttwu_count);
2475 if (cpu == this_cpu)
2476 schedstat_inc(rq, ttwu_local);
2478 struct sched_domain *sd;
2479 for_each_domain(this_cpu, sd) {
2480 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2481 schedstat_inc(sd, ttwu_wake_remote);
2486 #endif /* CONFIG_SCHEDSTATS */
2489 #endif /* CONFIG_SMP */
2490 schedstat_inc(p, se.nr_wakeups);
2492 schedstat_inc(p, se.nr_wakeups_sync);
2493 if (orig_cpu != cpu)
2494 schedstat_inc(p, se.nr_wakeups_migrate);
2495 if (cpu == this_cpu)
2496 schedstat_inc(p, se.nr_wakeups_local);
2498 schedstat_inc(p, se.nr_wakeups_remote);
2499 activate_task(rq, p, 1);
2503 * Only attribute actual wakeups done by this task.
2505 if (!in_interrupt()) {
2506 struct sched_entity *se = ¤t->se;
2507 u64 sample = se->sum_exec_runtime;
2509 if (se->last_wakeup)
2510 sample -= se->last_wakeup;
2512 sample -= se->start_runtime;
2513 update_avg(&se->avg_wakeup, sample);
2515 se->last_wakeup = se->sum_exec_runtime;
2519 trace_sched_wakeup(rq, p, success);
2520 check_preempt_curr(rq, p, sync);
2522 p->state = TASK_RUNNING;
2524 if (p->sched_class->task_wake_up)
2525 p->sched_class->task_wake_up(rq, p);
2528 task_rq_unlock(rq, &flags);
2534 * wake_up_process - Wake up a specific process
2535 * @p: The process to be woken up.
2537 * Attempt to wake up the nominated process and move it to the set of runnable
2538 * processes. Returns 1 if the process was woken up, 0 if it was already
2541 * It may be assumed that this function implies a write memory barrier before
2542 * changing the task state if and only if any tasks are woken up.
2544 int wake_up_process(struct task_struct *p)
2546 return try_to_wake_up(p, TASK_ALL, 0);
2548 EXPORT_SYMBOL(wake_up_process);
2550 int wake_up_state(struct task_struct *p, unsigned int state)
2552 return try_to_wake_up(p, state, 0);
2556 * Perform scheduler related setup for a newly forked process p.
2557 * p is forked by current.
2559 * __sched_fork() is basic setup used by init_idle() too:
2561 static void __sched_fork(struct task_struct *p)
2563 p->se.exec_start = 0;
2564 p->se.sum_exec_runtime = 0;
2565 p->se.prev_sum_exec_runtime = 0;
2566 p->se.nr_migrations = 0;
2567 p->se.last_wakeup = 0;
2568 p->se.avg_overlap = 0;
2569 p->se.start_runtime = 0;
2570 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2572 #ifdef CONFIG_SCHEDSTATS
2573 p->se.wait_start = 0;
2574 p->se.sum_sleep_runtime = 0;
2575 p->se.sleep_start = 0;
2576 p->se.block_start = 0;
2577 p->se.sleep_max = 0;
2578 p->se.block_max = 0;
2580 p->se.slice_max = 0;
2584 INIT_LIST_HEAD(&p->rt.run_list);
2586 INIT_LIST_HEAD(&p->se.group_node);
2588 #ifdef CONFIG_PREEMPT_NOTIFIERS
2589 INIT_HLIST_HEAD(&p->preempt_notifiers);
2593 * We mark the process as running here, but have not actually
2594 * inserted it onto the runqueue yet. This guarantees that
2595 * nobody will actually run it, and a signal or other external
2596 * event cannot wake it up and insert it on the runqueue either.
2598 p->state = TASK_RUNNING;
2602 * fork()/clone()-time setup:
2604 void sched_fork(struct task_struct *p, int clone_flags)
2606 int cpu = get_cpu();
2611 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2613 set_task_cpu(p, cpu);
2616 * Revert to default priority/policy on fork if requested. Make sure we
2617 * do not leak PI boosting priority to the child.
2619 if (current->sched_reset_on_fork &&
2620 (p->policy == SCHED_FIFO || p->policy == SCHED_RR))
2621 p->policy = SCHED_NORMAL;
2623 if (current->sched_reset_on_fork &&
2624 (current->normal_prio < DEFAULT_PRIO))
2625 p->prio = DEFAULT_PRIO;
2627 p->prio = current->normal_prio;
2629 if (!rt_prio(p->prio))
2630 p->sched_class = &fair_sched_class;
2633 * We don't need the reset flag anymore after the fork. It has
2634 * fulfilled its duty:
2636 p->sched_reset_on_fork = 0;
2638 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2639 if (likely(sched_info_on()))
2640 memset(&p->sched_info, 0, sizeof(p->sched_info));
2642 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2645 #ifdef CONFIG_PREEMPT
2646 /* Want to start with kernel preemption disabled. */
2647 task_thread_info(p)->preempt_count = 1;
2649 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2655 * wake_up_new_task - wake up a newly created task for the first time.
2657 * This function will do some initial scheduler statistics housekeeping
2658 * that must be done for every newly created context, then puts the task
2659 * on the runqueue and wakes it.
2661 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2663 unsigned long flags;
2666 rq = task_rq_lock(p, &flags);
2667 BUG_ON(p->state != TASK_RUNNING);
2668 update_rq_clock(rq);
2670 p->prio = effective_prio(p);
2672 if (!p->sched_class->task_new || !current->se.on_rq) {
2673 activate_task(rq, p, 0);
2676 * Let the scheduling class do new task startup
2677 * management (if any):
2679 p->sched_class->task_new(rq, p);
2682 trace_sched_wakeup_new(rq, p, 1);
2683 check_preempt_curr(rq, p, 0);
2685 if (p->sched_class->task_wake_up)
2686 p->sched_class->task_wake_up(rq, p);
2688 task_rq_unlock(rq, &flags);
2691 #ifdef CONFIG_PREEMPT_NOTIFIERS
2694 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2695 * @notifier: notifier struct to register
2697 void preempt_notifier_register(struct preempt_notifier *notifier)
2699 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2701 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2704 * preempt_notifier_unregister - no longer interested in preemption notifications
2705 * @notifier: notifier struct to unregister
2707 * This is safe to call from within a preemption notifier.
2709 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2711 hlist_del(¬ifier->link);
2713 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2715 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2717 struct preempt_notifier *notifier;
2718 struct hlist_node *node;
2720 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2721 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2725 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2726 struct task_struct *next)
2728 struct preempt_notifier *notifier;
2729 struct hlist_node *node;
2731 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2732 notifier->ops->sched_out(notifier, next);
2735 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2737 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2742 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2743 struct task_struct *next)
2747 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2750 * prepare_task_switch - prepare to switch tasks
2751 * @rq: the runqueue preparing to switch
2752 * @prev: the current task that is being switched out
2753 * @next: the task we are going to switch to.
2755 * This is called with the rq lock held and interrupts off. It must
2756 * be paired with a subsequent finish_task_switch after the context
2759 * prepare_task_switch sets up locking and calls architecture specific
2763 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2764 struct task_struct *next)
2766 fire_sched_out_preempt_notifiers(prev, next);
2767 prepare_lock_switch(rq, next);
2768 prepare_arch_switch(next);
2772 * finish_task_switch - clean up after a task-switch
2773 * @rq: runqueue associated with task-switch
2774 * @prev: the thread we just switched away from.
2776 * finish_task_switch must be called after the context switch, paired
2777 * with a prepare_task_switch call before the context switch.
2778 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2779 * and do any other architecture-specific cleanup actions.
2781 * Note that we may have delayed dropping an mm in context_switch(). If
2782 * so, we finish that here outside of the runqueue lock. (Doing it
2783 * with the lock held can cause deadlocks; see schedule() for
2786 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2787 __releases(rq->lock)
2789 struct mm_struct *mm = rq->prev_mm;
2792 int post_schedule = 0;
2794 if (current->sched_class->needs_post_schedule)
2795 post_schedule = current->sched_class->needs_post_schedule(rq);
2801 * A task struct has one reference for the use as "current".
2802 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2803 * schedule one last time. The schedule call will never return, and
2804 * the scheduled task must drop that reference.
2805 * The test for TASK_DEAD must occur while the runqueue locks are
2806 * still held, otherwise prev could be scheduled on another cpu, die
2807 * there before we look at prev->state, and then the reference would
2809 * Manfred Spraul <manfred@colorfullife.com>
2811 prev_state = prev->state;
2812 finish_arch_switch(prev);
2813 perf_counter_task_sched_in(current, cpu_of(rq));
2814 finish_lock_switch(rq, prev);
2817 current->sched_class->post_schedule(rq);
2820 fire_sched_in_preempt_notifiers(current);
2823 if (unlikely(prev_state == TASK_DEAD)) {
2825 * Remove function-return probe instances associated with this
2826 * task and put them back on the free list.
2828 kprobe_flush_task(prev);
2829 put_task_struct(prev);
2834 * schedule_tail - first thing a freshly forked thread must call.
2835 * @prev: the thread we just switched away from.
2837 asmlinkage void schedule_tail(struct task_struct *prev)
2838 __releases(rq->lock)
2840 struct rq *rq = this_rq();
2842 finish_task_switch(rq, prev);
2843 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2844 /* In this case, finish_task_switch does not reenable preemption */
2847 if (current->set_child_tid)
2848 put_user(task_pid_vnr(current), current->set_child_tid);
2852 * context_switch - switch to the new MM and the new
2853 * thread's register state.
2856 context_switch(struct rq *rq, struct task_struct *prev,
2857 struct task_struct *next)
2859 struct mm_struct *mm, *oldmm;
2861 prepare_task_switch(rq, prev, next);
2862 trace_sched_switch(rq, prev, next);
2864 oldmm = prev->active_mm;
2866 * For paravirt, this is coupled with an exit in switch_to to
2867 * combine the page table reload and the switch backend into
2870 arch_start_context_switch(prev);
2872 if (unlikely(!mm)) {
2873 next->active_mm = oldmm;
2874 atomic_inc(&oldmm->mm_count);
2875 enter_lazy_tlb(oldmm, next);
2877 switch_mm(oldmm, mm, next);
2879 if (unlikely(!prev->mm)) {
2880 prev->active_mm = NULL;
2881 rq->prev_mm = oldmm;
2884 * Since the runqueue lock will be released by the next
2885 * task (which is an invalid locking op but in the case
2886 * of the scheduler it's an obvious special-case), so we
2887 * do an early lockdep release here:
2889 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2890 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2893 /* Here we just switch the register state and the stack. */
2894 switch_to(prev, next, prev);
2898 * this_rq must be evaluated again because prev may have moved
2899 * CPUs since it called schedule(), thus the 'rq' on its stack
2900 * frame will be invalid.
2902 finish_task_switch(this_rq(), prev);
2906 * nr_running, nr_uninterruptible and nr_context_switches:
2908 * externally visible scheduler statistics: current number of runnable
2909 * threads, current number of uninterruptible-sleeping threads, total
2910 * number of context switches performed since bootup.
2912 unsigned long nr_running(void)
2914 unsigned long i, sum = 0;
2916 for_each_online_cpu(i)
2917 sum += cpu_rq(i)->nr_running;
2922 unsigned long nr_uninterruptible(void)
2924 unsigned long i, sum = 0;
2926 for_each_possible_cpu(i)
2927 sum += cpu_rq(i)->nr_uninterruptible;
2930 * Since we read the counters lockless, it might be slightly
2931 * inaccurate. Do not allow it to go below zero though:
2933 if (unlikely((long)sum < 0))
2939 unsigned long long nr_context_switches(void)
2942 unsigned long long sum = 0;
2944 for_each_possible_cpu(i)
2945 sum += cpu_rq(i)->nr_switches;
2950 unsigned long nr_iowait(void)
2952 unsigned long i, sum = 0;
2954 for_each_possible_cpu(i)
2955 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2960 /* Variables and functions for calc_load */
2961 static atomic_long_t calc_load_tasks;
2962 static unsigned long calc_load_update;
2963 unsigned long avenrun[3];
2964 EXPORT_SYMBOL(avenrun);
2967 * get_avenrun - get the load average array
2968 * @loads: pointer to dest load array
2969 * @offset: offset to add
2970 * @shift: shift count to shift the result left
2972 * These values are estimates at best, so no need for locking.
2974 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2976 loads[0] = (avenrun[0] + offset) << shift;
2977 loads[1] = (avenrun[1] + offset) << shift;
2978 loads[2] = (avenrun[2] + offset) << shift;
2981 static unsigned long
2982 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2985 load += active * (FIXED_1 - exp);
2986 return load >> FSHIFT;
2990 * calc_load - update the avenrun load estimates 10 ticks after the
2991 * CPUs have updated calc_load_tasks.
2993 void calc_global_load(void)
2995 unsigned long upd = calc_load_update + 10;
2998 if (time_before(jiffies, upd))
3001 active = atomic_long_read(&calc_load_tasks);
3002 active = active > 0 ? active * FIXED_1 : 0;
3004 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3005 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3006 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3008 calc_load_update += LOAD_FREQ;
3012 * Either called from update_cpu_load() or from a cpu going idle
3014 static void calc_load_account_active(struct rq *this_rq)
3016 long nr_active, delta;
3018 nr_active = this_rq->nr_running;
3019 nr_active += (long) this_rq->nr_uninterruptible;
3021 if (nr_active != this_rq->calc_load_active) {
3022 delta = nr_active - this_rq->calc_load_active;
3023 this_rq->calc_load_active = nr_active;
3024 atomic_long_add(delta, &calc_load_tasks);
3029 * Externally visible per-cpu scheduler statistics:
3030 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3032 u64 cpu_nr_migrations(int cpu)
3034 return cpu_rq(cpu)->nr_migrations_in;
3038 * Update rq->cpu_load[] statistics. This function is usually called every
3039 * scheduler tick (TICK_NSEC).
3041 static void update_cpu_load(struct rq *this_rq)
3043 unsigned long this_load = this_rq->load.weight;
3046 this_rq->nr_load_updates++;
3048 /* Update our load: */
3049 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3050 unsigned long old_load, new_load;
3052 /* scale is effectively 1 << i now, and >> i divides by scale */
3054 old_load = this_rq->cpu_load[i];
3055 new_load = this_load;
3057 * Round up the averaging division if load is increasing. This
3058 * prevents us from getting stuck on 9 if the load is 10, for
3061 if (new_load > old_load)
3062 new_load += scale-1;
3063 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3066 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3067 this_rq->calc_load_update += LOAD_FREQ;
3068 calc_load_account_active(this_rq);
3075 * double_rq_lock - safely lock two runqueues
3077 * Note this does not disable interrupts like task_rq_lock,
3078 * you need to do so manually before calling.
3080 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3081 __acquires(rq1->lock)
3082 __acquires(rq2->lock)
3084 BUG_ON(!irqs_disabled());
3086 spin_lock(&rq1->lock);
3087 __acquire(rq2->lock); /* Fake it out ;) */
3090 spin_lock(&rq1->lock);
3091 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3093 spin_lock(&rq2->lock);
3094 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3097 update_rq_clock(rq1);
3098 update_rq_clock(rq2);
3102 * double_rq_unlock - safely unlock two runqueues
3104 * Note this does not restore interrupts like task_rq_unlock,
3105 * you need to do so manually after calling.
3107 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3108 __releases(rq1->lock)
3109 __releases(rq2->lock)
3111 spin_unlock(&rq1->lock);
3113 spin_unlock(&rq2->lock);
3115 __release(rq2->lock);
3119 * If dest_cpu is allowed for this process, migrate the task to it.
3120 * This is accomplished by forcing the cpu_allowed mask to only
3121 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3122 * the cpu_allowed mask is restored.
3124 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3126 struct migration_req req;
3127 unsigned long flags;
3130 rq = task_rq_lock(p, &flags);
3131 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3132 || unlikely(!cpu_active(dest_cpu)))
3135 /* force the process onto the specified CPU */
3136 if (migrate_task(p, dest_cpu, &req)) {
3137 /* Need to wait for migration thread (might exit: take ref). */
3138 struct task_struct *mt = rq->migration_thread;
3140 get_task_struct(mt);
3141 task_rq_unlock(rq, &flags);
3142 wake_up_process(mt);
3143 put_task_struct(mt);
3144 wait_for_completion(&req.done);
3149 task_rq_unlock(rq, &flags);
3153 * sched_exec - execve() is a valuable balancing opportunity, because at
3154 * this point the task has the smallest effective memory and cache footprint.
3156 void sched_exec(void)
3158 int new_cpu, this_cpu = get_cpu();
3159 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3161 if (new_cpu != this_cpu)
3162 sched_migrate_task(current, new_cpu);
3166 * pull_task - move a task from a remote runqueue to the local runqueue.
3167 * Both runqueues must be locked.
3169 static void pull_task(struct rq *src_rq, struct task_struct *p,
3170 struct rq *this_rq, int this_cpu)
3172 deactivate_task(src_rq, p, 0);
3173 set_task_cpu(p, this_cpu);
3174 activate_task(this_rq, p, 0);
3176 * Note that idle threads have a prio of MAX_PRIO, for this test
3177 * to be always true for them.
3179 check_preempt_curr(this_rq, p, 0);
3183 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3186 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3187 struct sched_domain *sd, enum cpu_idle_type idle,
3190 int tsk_cache_hot = 0;
3192 * We do not migrate tasks that are:
3193 * 1) running (obviously), or
3194 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3195 * 3) are cache-hot on their current CPU.
3197 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3198 schedstat_inc(p, se.nr_failed_migrations_affine);
3203 if (task_running(rq, p)) {
3204 schedstat_inc(p, se.nr_failed_migrations_running);
3209 * Aggressive migration if:
3210 * 1) task is cache cold, or
3211 * 2) too many balance attempts have failed.
3214 tsk_cache_hot = task_hot(p, rq->clock, sd);
3215 if (!tsk_cache_hot ||
3216 sd->nr_balance_failed > sd->cache_nice_tries) {
3217 #ifdef CONFIG_SCHEDSTATS
3218 if (tsk_cache_hot) {
3219 schedstat_inc(sd, lb_hot_gained[idle]);
3220 schedstat_inc(p, se.nr_forced_migrations);
3226 if (tsk_cache_hot) {
3227 schedstat_inc(p, se.nr_failed_migrations_hot);
3233 static unsigned long
3234 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3235 unsigned long max_load_move, struct sched_domain *sd,
3236 enum cpu_idle_type idle, int *all_pinned,
3237 int *this_best_prio, struct rq_iterator *iterator)
3239 int loops = 0, pulled = 0, pinned = 0;
3240 struct task_struct *p;
3241 long rem_load_move = max_load_move;
3243 if (max_load_move == 0)
3249 * Start the load-balancing iterator:
3251 p = iterator->start(iterator->arg);
3253 if (!p || loops++ > sysctl_sched_nr_migrate)
3256 if ((p->se.load.weight >> 1) > rem_load_move ||
3257 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3258 p = iterator->next(iterator->arg);
3262 pull_task(busiest, p, this_rq, this_cpu);
3264 rem_load_move -= p->se.load.weight;
3266 #ifdef CONFIG_PREEMPT
3268 * NEWIDLE balancing is a source of latency, so preemptible kernels
3269 * will stop after the first task is pulled to minimize the critical
3272 if (idle == CPU_NEWLY_IDLE)
3277 * We only want to steal up to the prescribed amount of weighted load.
3279 if (rem_load_move > 0) {
3280 if (p->prio < *this_best_prio)
3281 *this_best_prio = p->prio;
3282 p = iterator->next(iterator->arg);
3287 * Right now, this is one of only two places pull_task() is called,
3288 * so we can safely collect pull_task() stats here rather than
3289 * inside pull_task().
3291 schedstat_add(sd, lb_gained[idle], pulled);
3294 *all_pinned = pinned;
3296 return max_load_move - rem_load_move;
3300 * move_tasks tries to move up to max_load_move weighted load from busiest to
3301 * this_rq, as part of a balancing operation within domain "sd".
3302 * Returns 1 if successful and 0 otherwise.
3304 * Called with both runqueues locked.
3306 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3307 unsigned long max_load_move,
3308 struct sched_domain *sd, enum cpu_idle_type idle,
3311 const struct sched_class *class = sched_class_highest;
3312 unsigned long total_load_moved = 0;
3313 int this_best_prio = this_rq->curr->prio;
3317 class->load_balance(this_rq, this_cpu, busiest,
3318 max_load_move - total_load_moved,
3319 sd, idle, all_pinned, &this_best_prio);
3320 class = class->next;
3322 #ifdef CONFIG_PREEMPT
3324 * NEWIDLE balancing is a source of latency, so preemptible
3325 * kernels will stop after the first task is pulled to minimize
3326 * the critical section.
3328 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3331 } while (class && max_load_move > total_load_moved);
3333 return total_load_moved > 0;
3337 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3338 struct sched_domain *sd, enum cpu_idle_type idle,
3339 struct rq_iterator *iterator)
3341 struct task_struct *p = iterator->start(iterator->arg);
3345 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3346 pull_task(busiest, p, this_rq, this_cpu);
3348 * Right now, this is only the second place pull_task()
3349 * is called, so we can safely collect pull_task()
3350 * stats here rather than inside pull_task().
3352 schedstat_inc(sd, lb_gained[idle]);
3356 p = iterator->next(iterator->arg);
3363 * move_one_task tries to move exactly one task from busiest to this_rq, as
3364 * part of active balancing operations within "domain".
3365 * Returns 1 if successful and 0 otherwise.
3367 * Called with both runqueues locked.
3369 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3370 struct sched_domain *sd, enum cpu_idle_type idle)
3372 const struct sched_class *class;
3374 for (class = sched_class_highest; class; class = class->next)
3375 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3380 /********** Helpers for find_busiest_group ************************/
3382 * sd_lb_stats - Structure to store the statistics of a sched_domain
3383 * during load balancing.
3385 struct sd_lb_stats {
3386 struct sched_group *busiest; /* Busiest group in this sd */
3387 struct sched_group *this; /* Local group in this sd */
3388 unsigned long total_load; /* Total load of all groups in sd */
3389 unsigned long total_pwr; /* Total power of all groups in sd */
3390 unsigned long avg_load; /* Average load across all groups in sd */
3392 /** Statistics of this group */
3393 unsigned long this_load;
3394 unsigned long this_load_per_task;
3395 unsigned long this_nr_running;
3397 /* Statistics of the busiest group */
3398 unsigned long max_load;
3399 unsigned long busiest_load_per_task;
3400 unsigned long busiest_nr_running;
3402 int group_imb; /* Is there imbalance in this sd */
3403 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3404 int power_savings_balance; /* Is powersave balance needed for this sd */
3405 struct sched_group *group_min; /* Least loaded group in sd */
3406 struct sched_group *group_leader; /* Group which relieves group_min */
3407 unsigned long min_load_per_task; /* load_per_task in group_min */
3408 unsigned long leader_nr_running; /* Nr running of group_leader */
3409 unsigned long min_nr_running; /* Nr running of group_min */
3414 * sg_lb_stats - stats of a sched_group required for load_balancing
3416 struct sg_lb_stats {
3417 unsigned long avg_load; /*Avg load across the CPUs of the group */
3418 unsigned long group_load; /* Total load over the CPUs of the group */
3419 unsigned long sum_nr_running; /* Nr tasks running in the group */
3420 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3421 unsigned long group_capacity;
3422 int group_imb; /* Is there an imbalance in the group ? */
3426 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3427 * @group: The group whose first cpu is to be returned.
3429 static inline unsigned int group_first_cpu(struct sched_group *group)
3431 return cpumask_first(sched_group_cpus(group));
3435 * get_sd_load_idx - Obtain the load index for a given sched domain.
3436 * @sd: The sched_domain whose load_idx is to be obtained.
3437 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3439 static inline int get_sd_load_idx(struct sched_domain *sd,
3440 enum cpu_idle_type idle)
3446 load_idx = sd->busy_idx;
3449 case CPU_NEWLY_IDLE:
3450 load_idx = sd->newidle_idx;
3453 load_idx = sd->idle_idx;
3461 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3463 * init_sd_power_savings_stats - Initialize power savings statistics for
3464 * the given sched_domain, during load balancing.
3466 * @sd: Sched domain whose power-savings statistics are to be initialized.
3467 * @sds: Variable containing the statistics for sd.
3468 * @idle: Idle status of the CPU at which we're performing load-balancing.
3470 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3471 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3474 * Busy processors will not participate in power savings
3477 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3478 sds->power_savings_balance = 0;
3480 sds->power_savings_balance = 1;
3481 sds->min_nr_running = ULONG_MAX;
3482 sds->leader_nr_running = 0;
3487 * update_sd_power_savings_stats - Update the power saving stats for a
3488 * sched_domain while performing load balancing.
3490 * @group: sched_group belonging to the sched_domain under consideration.
3491 * @sds: Variable containing the statistics of the sched_domain
3492 * @local_group: Does group contain the CPU for which we're performing
3494 * @sgs: Variable containing the statistics of the group.
3496 static inline void update_sd_power_savings_stats(struct sched_group *group,
3497 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3500 if (!sds->power_savings_balance)
3504 * If the local group is idle or completely loaded
3505 * no need to do power savings balance at this domain
3507 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3508 !sds->this_nr_running))
3509 sds->power_savings_balance = 0;
3512 * If a group is already running at full capacity or idle,
3513 * don't include that group in power savings calculations
3515 if (!sds->power_savings_balance ||
3516 sgs->sum_nr_running >= sgs->group_capacity ||
3517 !sgs->sum_nr_running)
3521 * Calculate the group which has the least non-idle load.
3522 * This is the group from where we need to pick up the load
3525 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3526 (sgs->sum_nr_running == sds->min_nr_running &&
3527 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3528 sds->group_min = group;
3529 sds->min_nr_running = sgs->sum_nr_running;
3530 sds->min_load_per_task = sgs->sum_weighted_load /
3531 sgs->sum_nr_running;
3535 * Calculate the group which is almost near its
3536 * capacity but still has some space to pick up some load
3537 * from other group and save more power
3539 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3542 if (sgs->sum_nr_running > sds->leader_nr_running ||
3543 (sgs->sum_nr_running == sds->leader_nr_running &&
3544 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3545 sds->group_leader = group;
3546 sds->leader_nr_running = sgs->sum_nr_running;
3551 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3552 * @sds: Variable containing the statistics of the sched_domain
3553 * under consideration.
3554 * @this_cpu: Cpu at which we're currently performing load-balancing.
3555 * @imbalance: Variable to store the imbalance.
3558 * Check if we have potential to perform some power-savings balance.
3559 * If yes, set the busiest group to be the least loaded group in the
3560 * sched_domain, so that it's CPUs can be put to idle.
3562 * Returns 1 if there is potential to perform power-savings balance.
3565 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3566 int this_cpu, unsigned long *imbalance)
3568 if (!sds->power_savings_balance)
3571 if (sds->this != sds->group_leader ||
3572 sds->group_leader == sds->group_min)
3575 *imbalance = sds->min_load_per_task;
3576 sds->busiest = sds->group_min;
3578 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3579 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3580 group_first_cpu(sds->group_leader);
3586 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3587 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3588 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3593 static inline void update_sd_power_savings_stats(struct sched_group *group,
3594 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3599 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3600 int this_cpu, unsigned long *imbalance)
3604 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3608 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3609 * @group: sched_group whose statistics are to be updated.
3610 * @this_cpu: Cpu for which load balance is currently performed.
3611 * @idle: Idle status of this_cpu
3612 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3613 * @sd_idle: Idle status of the sched_domain containing group.
3614 * @local_group: Does group contain this_cpu.
3615 * @cpus: Set of cpus considered for load balancing.
3616 * @balance: Should we balance.
3617 * @sgs: variable to hold the statistics for this group.
3619 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3620 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3621 int local_group, const struct cpumask *cpus,
3622 int *balance, struct sg_lb_stats *sgs)
3624 unsigned long load, max_cpu_load, min_cpu_load;
3626 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3627 unsigned long sum_avg_load_per_task;
3628 unsigned long avg_load_per_task;
3631 balance_cpu = group_first_cpu(group);
3633 /* Tally up the load of all CPUs in the group */
3634 sum_avg_load_per_task = avg_load_per_task = 0;
3636 min_cpu_load = ~0UL;
3638 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3639 struct rq *rq = cpu_rq(i);
3641 if (*sd_idle && rq->nr_running)
3644 /* Bias balancing toward cpus of our domain */
3646 if (idle_cpu(i) && !first_idle_cpu) {
3651 load = target_load(i, load_idx);
3653 load = source_load(i, load_idx);
3654 if (load > max_cpu_load)
3655 max_cpu_load = load;
3656 if (min_cpu_load > load)
3657 min_cpu_load = load;
3660 sgs->group_load += load;
3661 sgs->sum_nr_running += rq->nr_running;
3662 sgs->sum_weighted_load += weighted_cpuload(i);
3664 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3668 * First idle cpu or the first cpu(busiest) in this sched group
3669 * is eligible for doing load balancing at this and above
3670 * domains. In the newly idle case, we will allow all the cpu's
3671 * to do the newly idle load balance.
3673 if (idle != CPU_NEWLY_IDLE && local_group &&
3674 balance_cpu != this_cpu && balance) {
3679 /* Adjust by relative CPU power of the group */
3680 sgs->avg_load = sg_div_cpu_power(group,
3681 sgs->group_load * SCHED_LOAD_SCALE);
3685 * Consider the group unbalanced when the imbalance is larger
3686 * than the average weight of two tasks.
3688 * APZ: with cgroup the avg task weight can vary wildly and
3689 * might not be a suitable number - should we keep a
3690 * normalized nr_running number somewhere that negates
3693 avg_load_per_task = sg_div_cpu_power(group,
3694 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3696 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3699 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3704 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3705 * @sd: sched_domain whose statistics are to be updated.
3706 * @this_cpu: Cpu for which load balance is currently performed.
3707 * @idle: Idle status of this_cpu
3708 * @sd_idle: Idle status of the sched_domain containing group.
3709 * @cpus: Set of cpus considered for load balancing.
3710 * @balance: Should we balance.
3711 * @sds: variable to hold the statistics for this sched_domain.
3713 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3714 enum cpu_idle_type idle, int *sd_idle,
3715 const struct cpumask *cpus, int *balance,
3716 struct sd_lb_stats *sds)
3718 struct sched_group *group = sd->groups;
3719 struct sg_lb_stats sgs;
3722 init_sd_power_savings_stats(sd, sds, idle);
3723 load_idx = get_sd_load_idx(sd, idle);
3728 local_group = cpumask_test_cpu(this_cpu,
3729 sched_group_cpus(group));
3730 memset(&sgs, 0, sizeof(sgs));
3731 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3732 local_group, cpus, balance, &sgs);
3734 if (local_group && balance && !(*balance))
3737 sds->total_load += sgs.group_load;
3738 sds->total_pwr += group->__cpu_power;
3741 sds->this_load = sgs.avg_load;
3743 sds->this_nr_running = sgs.sum_nr_running;
3744 sds->this_load_per_task = sgs.sum_weighted_load;
3745 } else if (sgs.avg_load > sds->max_load &&
3746 (sgs.sum_nr_running > sgs.group_capacity ||
3748 sds->max_load = sgs.avg_load;
3749 sds->busiest = group;
3750 sds->busiest_nr_running = sgs.sum_nr_running;
3751 sds->busiest_load_per_task = sgs.sum_weighted_load;
3752 sds->group_imb = sgs.group_imb;
3755 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3756 group = group->next;
3757 } while (group != sd->groups);
3762 * fix_small_imbalance - Calculate the minor imbalance that exists
3763 * amongst the groups of a sched_domain, during
3765 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3766 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3767 * @imbalance: Variable to store the imbalance.
3769 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3770 int this_cpu, unsigned long *imbalance)
3772 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3773 unsigned int imbn = 2;
3775 if (sds->this_nr_running) {
3776 sds->this_load_per_task /= sds->this_nr_running;
3777 if (sds->busiest_load_per_task >
3778 sds->this_load_per_task)
3781 sds->this_load_per_task =
3782 cpu_avg_load_per_task(this_cpu);
3784 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3785 sds->busiest_load_per_task * imbn) {
3786 *imbalance = sds->busiest_load_per_task;
3791 * OK, we don't have enough imbalance to justify moving tasks,
3792 * however we may be able to increase total CPU power used by
3796 pwr_now += sds->busiest->__cpu_power *
3797 min(sds->busiest_load_per_task, sds->max_load);
3798 pwr_now += sds->this->__cpu_power *
3799 min(sds->this_load_per_task, sds->this_load);
3800 pwr_now /= SCHED_LOAD_SCALE;
3802 /* Amount of load we'd subtract */
3803 tmp = sg_div_cpu_power(sds->busiest,
3804 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3805 if (sds->max_load > tmp)
3806 pwr_move += sds->busiest->__cpu_power *
3807 min(sds->busiest_load_per_task, sds->max_load - tmp);
3809 /* Amount of load we'd add */
3810 if (sds->max_load * sds->busiest->__cpu_power <
3811 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3812 tmp = sg_div_cpu_power(sds->this,
3813 sds->max_load * sds->busiest->__cpu_power);
3815 tmp = sg_div_cpu_power(sds->this,
3816 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3817 pwr_move += sds->this->__cpu_power *
3818 min(sds->this_load_per_task, sds->this_load + tmp);
3819 pwr_move /= SCHED_LOAD_SCALE;
3821 /* Move if we gain throughput */
3822 if (pwr_move > pwr_now)
3823 *imbalance = sds->busiest_load_per_task;
3827 * calculate_imbalance - Calculate the amount of imbalance present within the
3828 * groups of a given sched_domain during load balance.
3829 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3830 * @this_cpu: Cpu for which currently load balance is being performed.
3831 * @imbalance: The variable to store the imbalance.
3833 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3834 unsigned long *imbalance)
3836 unsigned long max_pull;
3838 * In the presence of smp nice balancing, certain scenarios can have
3839 * max load less than avg load(as we skip the groups at or below
3840 * its cpu_power, while calculating max_load..)
3842 if (sds->max_load < sds->avg_load) {
3844 return fix_small_imbalance(sds, this_cpu, imbalance);
3847 /* Don't want to pull so many tasks that a group would go idle */
3848 max_pull = min(sds->max_load - sds->avg_load,
3849 sds->max_load - sds->busiest_load_per_task);
3851 /* How much load to actually move to equalise the imbalance */
3852 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3853 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3857 * if *imbalance is less than the average load per runnable task
3858 * there is no gaurantee that any tasks will be moved so we'll have
3859 * a think about bumping its value to force at least one task to be
3862 if (*imbalance < sds->busiest_load_per_task)
3863 return fix_small_imbalance(sds, this_cpu, imbalance);
3866 /******* find_busiest_group() helpers end here *********************/
3869 * find_busiest_group - Returns the busiest group within the sched_domain
3870 * if there is an imbalance. If there isn't an imbalance, and
3871 * the user has opted for power-savings, it returns a group whose
3872 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3873 * such a group exists.
3875 * Also calculates the amount of weighted load which should be moved
3876 * to restore balance.
3878 * @sd: The sched_domain whose busiest group is to be returned.
3879 * @this_cpu: The cpu for which load balancing is currently being performed.
3880 * @imbalance: Variable which stores amount of weighted load which should
3881 * be moved to restore balance/put a group to idle.
3882 * @idle: The idle status of this_cpu.
3883 * @sd_idle: The idleness of sd
3884 * @cpus: The set of CPUs under consideration for load-balancing.
3885 * @balance: Pointer to a variable indicating if this_cpu
3886 * is the appropriate cpu to perform load balancing at this_level.
3888 * Returns: - the busiest group if imbalance exists.
3889 * - If no imbalance and user has opted for power-savings balance,
3890 * return the least loaded group whose CPUs can be
3891 * put to idle by rebalancing its tasks onto our group.
3893 static struct sched_group *
3894 find_busiest_group(struct sched_domain *sd, int this_cpu,
3895 unsigned long *imbalance, enum cpu_idle_type idle,
3896 int *sd_idle, const struct cpumask *cpus, int *balance)
3898 struct sd_lb_stats sds;
3900 memset(&sds, 0, sizeof(sds));
3903 * Compute the various statistics relavent for load balancing at
3906 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3909 /* Cases where imbalance does not exist from POV of this_cpu */
3910 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3912 * 2) There is no busy sibling group to pull from.
3913 * 3) This group is the busiest group.
3914 * 4) This group is more busy than the avg busieness at this
3916 * 5) The imbalance is within the specified limit.
3917 * 6) Any rebalance would lead to ping-pong
3919 if (balance && !(*balance))
3922 if (!sds.busiest || sds.busiest_nr_running == 0)
3925 if (sds.this_load >= sds.max_load)
3928 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3930 if (sds.this_load >= sds.avg_load)
3933 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3936 sds.busiest_load_per_task /= sds.busiest_nr_running;
3938 sds.busiest_load_per_task =
3939 min(sds.busiest_load_per_task, sds.avg_load);
3942 * We're trying to get all the cpus to the average_load, so we don't
3943 * want to push ourselves above the average load, nor do we wish to
3944 * reduce the max loaded cpu below the average load, as either of these
3945 * actions would just result in more rebalancing later, and ping-pong
3946 * tasks around. Thus we look for the minimum possible imbalance.
3947 * Negative imbalances (*we* are more loaded than anyone else) will
3948 * be counted as no imbalance for these purposes -- we can't fix that
3949 * by pulling tasks to us. Be careful of negative numbers as they'll
3950 * appear as very large values with unsigned longs.
3952 if (sds.max_load <= sds.busiest_load_per_task)
3955 /* Looks like there is an imbalance. Compute it */
3956 calculate_imbalance(&sds, this_cpu, imbalance);
3961 * There is no obvious imbalance. But check if we can do some balancing
3964 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3972 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3975 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3976 unsigned long imbalance, const struct cpumask *cpus)
3978 struct rq *busiest = NULL, *rq;
3979 unsigned long max_load = 0;
3982 for_each_cpu(i, sched_group_cpus(group)) {
3985 if (!cpumask_test_cpu(i, cpus))
3989 wl = weighted_cpuload(i);
3991 if (rq->nr_running == 1 && wl > imbalance)
3994 if (wl > max_load) {
4004 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4005 * so long as it is large enough.
4007 #define MAX_PINNED_INTERVAL 512
4009 /* Working cpumask for load_balance and load_balance_newidle. */
4010 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4013 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4014 * tasks if there is an imbalance.
4016 static int load_balance(int this_cpu, struct rq *this_rq,
4017 struct sched_domain *sd, enum cpu_idle_type idle,
4020 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4021 struct sched_group *group;
4022 unsigned long imbalance;
4024 unsigned long flags;
4025 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4027 cpumask_setall(cpus);
4030 * When power savings policy is enabled for the parent domain, idle
4031 * sibling can pick up load irrespective of busy siblings. In this case,
4032 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4033 * portraying it as CPU_NOT_IDLE.
4035 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4036 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4039 schedstat_inc(sd, lb_count[idle]);
4043 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4050 schedstat_inc(sd, lb_nobusyg[idle]);
4054 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4056 schedstat_inc(sd, lb_nobusyq[idle]);
4060 BUG_ON(busiest == this_rq);
4062 schedstat_add(sd, lb_imbalance[idle], imbalance);
4065 if (busiest->nr_running > 1) {
4067 * Attempt to move tasks. If find_busiest_group has found
4068 * an imbalance but busiest->nr_running <= 1, the group is
4069 * still unbalanced. ld_moved simply stays zero, so it is
4070 * correctly treated as an imbalance.
4072 local_irq_save(flags);
4073 double_rq_lock(this_rq, busiest);
4074 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4075 imbalance, sd, idle, &all_pinned);
4076 double_rq_unlock(this_rq, busiest);
4077 local_irq_restore(flags);
4080 * some other cpu did the load balance for us.
4082 if (ld_moved && this_cpu != smp_processor_id())
4083 resched_cpu(this_cpu);
4085 /* All tasks on this runqueue were pinned by CPU affinity */
4086 if (unlikely(all_pinned)) {
4087 cpumask_clear_cpu(cpu_of(busiest), cpus);
4088 if (!cpumask_empty(cpus))
4095 schedstat_inc(sd, lb_failed[idle]);
4096 sd->nr_balance_failed++;
4098 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4100 spin_lock_irqsave(&busiest->lock, flags);
4102 /* don't kick the migration_thread, if the curr
4103 * task on busiest cpu can't be moved to this_cpu
4105 if (!cpumask_test_cpu(this_cpu,
4106 &busiest->curr->cpus_allowed)) {
4107 spin_unlock_irqrestore(&busiest->lock, flags);
4109 goto out_one_pinned;
4112 if (!busiest->active_balance) {
4113 busiest->active_balance = 1;
4114 busiest->push_cpu = this_cpu;
4117 spin_unlock_irqrestore(&busiest->lock, flags);
4119 wake_up_process(busiest->migration_thread);
4122 * We've kicked active balancing, reset the failure
4125 sd->nr_balance_failed = sd->cache_nice_tries+1;
4128 sd->nr_balance_failed = 0;
4130 if (likely(!active_balance)) {
4131 /* We were unbalanced, so reset the balancing interval */
4132 sd->balance_interval = sd->min_interval;
4135 * If we've begun active balancing, start to back off. This
4136 * case may not be covered by the all_pinned logic if there
4137 * is only 1 task on the busy runqueue (because we don't call
4140 if (sd->balance_interval < sd->max_interval)
4141 sd->balance_interval *= 2;
4144 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4145 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4151 schedstat_inc(sd, lb_balanced[idle]);
4153 sd->nr_balance_failed = 0;
4156 /* tune up the balancing interval */
4157 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4158 (sd->balance_interval < sd->max_interval))
4159 sd->balance_interval *= 2;
4161 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4162 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4173 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4174 * tasks if there is an imbalance.
4176 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4177 * this_rq is locked.
4180 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4182 struct sched_group *group;
4183 struct rq *busiest = NULL;
4184 unsigned long imbalance;
4188 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4190 cpumask_setall(cpus);
4193 * When power savings policy is enabled for the parent domain, idle
4194 * sibling can pick up load irrespective of busy siblings. In this case,
4195 * let the state of idle sibling percolate up as IDLE, instead of
4196 * portraying it as CPU_NOT_IDLE.
4198 if (sd->flags & SD_SHARE_CPUPOWER &&
4199 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4202 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4204 update_shares_locked(this_rq, sd);
4205 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4206 &sd_idle, cpus, NULL);
4208 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4212 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4214 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4218 BUG_ON(busiest == this_rq);
4220 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4223 if (busiest->nr_running > 1) {
4224 /* Attempt to move tasks */
4225 double_lock_balance(this_rq, busiest);
4226 /* this_rq->clock is already updated */
4227 update_rq_clock(busiest);
4228 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4229 imbalance, sd, CPU_NEWLY_IDLE,
4231 double_unlock_balance(this_rq, busiest);
4233 if (unlikely(all_pinned)) {
4234 cpumask_clear_cpu(cpu_of(busiest), cpus);
4235 if (!cpumask_empty(cpus))
4241 int active_balance = 0;
4243 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4244 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4245 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4248 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4251 if (sd->nr_balance_failed++ < 2)
4255 * The only task running in a non-idle cpu can be moved to this
4256 * cpu in an attempt to completely freeup the other CPU
4257 * package. The same method used to move task in load_balance()
4258 * have been extended for load_balance_newidle() to speedup
4259 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4261 * The package power saving logic comes from
4262 * find_busiest_group(). If there are no imbalance, then
4263 * f_b_g() will return NULL. However when sched_mc={1,2} then
4264 * f_b_g() will select a group from which a running task may be
4265 * pulled to this cpu in order to make the other package idle.
4266 * If there is no opportunity to make a package idle and if
4267 * there are no imbalance, then f_b_g() will return NULL and no
4268 * action will be taken in load_balance_newidle().
4270 * Under normal task pull operation due to imbalance, there
4271 * will be more than one task in the source run queue and
4272 * move_tasks() will succeed. ld_moved will be true and this
4273 * active balance code will not be triggered.
4276 /* Lock busiest in correct order while this_rq is held */
4277 double_lock_balance(this_rq, busiest);
4280 * don't kick the migration_thread, if the curr
4281 * task on busiest cpu can't be moved to this_cpu
4283 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4284 double_unlock_balance(this_rq, busiest);
4289 if (!busiest->active_balance) {
4290 busiest->active_balance = 1;
4291 busiest->push_cpu = this_cpu;
4295 double_unlock_balance(this_rq, busiest);
4297 * Should not call ttwu while holding a rq->lock
4299 spin_unlock(&this_rq->lock);
4301 wake_up_process(busiest->migration_thread);
4302 spin_lock(&this_rq->lock);
4305 sd->nr_balance_failed = 0;
4307 update_shares_locked(this_rq, sd);
4311 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4312 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4313 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4315 sd->nr_balance_failed = 0;
4321 * idle_balance is called by schedule() if this_cpu is about to become
4322 * idle. Attempts to pull tasks from other CPUs.
4324 static void idle_balance(int this_cpu, struct rq *this_rq)
4326 struct sched_domain *sd;
4327 int pulled_task = 0;
4328 unsigned long next_balance = jiffies + HZ;
4330 for_each_domain(this_cpu, sd) {
4331 unsigned long interval;
4333 if (!(sd->flags & SD_LOAD_BALANCE))
4336 if (sd->flags & SD_BALANCE_NEWIDLE)
4337 /* If we've pulled tasks over stop searching: */
4338 pulled_task = load_balance_newidle(this_cpu, this_rq,
4341 interval = msecs_to_jiffies(sd->balance_interval);
4342 if (time_after(next_balance, sd->last_balance + interval))
4343 next_balance = sd->last_balance + interval;
4347 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4349 * We are going idle. next_balance may be set based on
4350 * a busy processor. So reset next_balance.
4352 this_rq->next_balance = next_balance;
4357 * active_load_balance is run by migration threads. It pushes running tasks
4358 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4359 * running on each physical CPU where possible, and avoids physical /
4360 * logical imbalances.
4362 * Called with busiest_rq locked.
4364 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4366 int target_cpu = busiest_rq->push_cpu;
4367 struct sched_domain *sd;
4368 struct rq *target_rq;
4370 /* Is there any task to move? */
4371 if (busiest_rq->nr_running <= 1)
4374 target_rq = cpu_rq(target_cpu);
4377 * This condition is "impossible", if it occurs
4378 * we need to fix it. Originally reported by
4379 * Bjorn Helgaas on a 128-cpu setup.
4381 BUG_ON(busiest_rq == target_rq);
4383 /* move a task from busiest_rq to target_rq */
4384 double_lock_balance(busiest_rq, target_rq);
4385 update_rq_clock(busiest_rq);
4386 update_rq_clock(target_rq);
4388 /* Search for an sd spanning us and the target CPU. */
4389 for_each_domain(target_cpu, sd) {
4390 if ((sd->flags & SD_LOAD_BALANCE) &&
4391 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4396 schedstat_inc(sd, alb_count);
4398 if (move_one_task(target_rq, target_cpu, busiest_rq,
4400 schedstat_inc(sd, alb_pushed);
4402 schedstat_inc(sd, alb_failed);
4404 double_unlock_balance(busiest_rq, target_rq);
4409 atomic_t load_balancer;
4410 cpumask_var_t cpu_mask;
4411 cpumask_var_t ilb_grp_nohz_mask;
4412 } nohz ____cacheline_aligned = {
4413 .load_balancer = ATOMIC_INIT(-1),
4416 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4418 * lowest_flag_domain - Return lowest sched_domain containing flag.
4419 * @cpu: The cpu whose lowest level of sched domain is to
4421 * @flag: The flag to check for the lowest sched_domain
4422 * for the given cpu.
4424 * Returns the lowest sched_domain of a cpu which contains the given flag.
4426 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4428 struct sched_domain *sd;
4430 for_each_domain(cpu, sd)
4431 if (sd && (sd->flags & flag))
4438 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4439 * @cpu: The cpu whose domains we're iterating over.
4440 * @sd: variable holding the value of the power_savings_sd
4442 * @flag: The flag to filter the sched_domains to be iterated.
4444 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4445 * set, starting from the lowest sched_domain to the highest.
4447 #define for_each_flag_domain(cpu, sd, flag) \
4448 for (sd = lowest_flag_domain(cpu, flag); \
4449 (sd && (sd->flags & flag)); sd = sd->parent)
4452 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4453 * @ilb_group: group to be checked for semi-idleness
4455 * Returns: 1 if the group is semi-idle. 0 otherwise.
4457 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4458 * and atleast one non-idle CPU. This helper function checks if the given
4459 * sched_group is semi-idle or not.
4461 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4463 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4464 sched_group_cpus(ilb_group));
4467 * A sched_group is semi-idle when it has atleast one busy cpu
4468 * and atleast one idle cpu.
4470 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4473 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4479 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4480 * @cpu: The cpu which is nominating a new idle_load_balancer.
4482 * Returns: Returns the id of the idle load balancer if it exists,
4483 * Else, returns >= nr_cpu_ids.
4485 * This algorithm picks the idle load balancer such that it belongs to a
4486 * semi-idle powersavings sched_domain. The idea is to try and avoid
4487 * completely idle packages/cores just for the purpose of idle load balancing
4488 * when there are other idle cpu's which are better suited for that job.
4490 static int find_new_ilb(int cpu)
4492 struct sched_domain *sd;
4493 struct sched_group *ilb_group;
4496 * Have idle load balancer selection from semi-idle packages only
4497 * when power-aware load balancing is enabled
4499 if (!(sched_smt_power_savings || sched_mc_power_savings))
4503 * Optimize for the case when we have no idle CPUs or only one
4504 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4506 if (cpumask_weight(nohz.cpu_mask) < 2)
4509 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4510 ilb_group = sd->groups;
4513 if (is_semi_idle_group(ilb_group))
4514 return cpumask_first(nohz.ilb_grp_nohz_mask);
4516 ilb_group = ilb_group->next;
4518 } while (ilb_group != sd->groups);
4522 return cpumask_first(nohz.cpu_mask);
4524 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4525 static inline int find_new_ilb(int call_cpu)
4527 return cpumask_first(nohz.cpu_mask);
4532 * This routine will try to nominate the ilb (idle load balancing)
4533 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4534 * load balancing on behalf of all those cpus. If all the cpus in the system
4535 * go into this tickless mode, then there will be no ilb owner (as there is
4536 * no need for one) and all the cpus will sleep till the next wakeup event
4539 * For the ilb owner, tick is not stopped. And this tick will be used
4540 * for idle load balancing. ilb owner will still be part of
4543 * While stopping the tick, this cpu will become the ilb owner if there
4544 * is no other owner. And will be the owner till that cpu becomes busy
4545 * or if all cpus in the system stop their ticks at which point
4546 * there is no need for ilb owner.
4548 * When the ilb owner becomes busy, it nominates another owner, during the
4549 * next busy scheduler_tick()
4551 int select_nohz_load_balancer(int stop_tick)
4553 int cpu = smp_processor_id();
4556 cpu_rq(cpu)->in_nohz_recently = 1;
4558 if (!cpu_active(cpu)) {
4559 if (atomic_read(&nohz.load_balancer) != cpu)
4563 * If we are going offline and still the leader,
4566 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4572 cpumask_set_cpu(cpu, nohz.cpu_mask);
4574 /* time for ilb owner also to sleep */
4575 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4576 if (atomic_read(&nohz.load_balancer) == cpu)
4577 atomic_set(&nohz.load_balancer, -1);
4581 if (atomic_read(&nohz.load_balancer) == -1) {
4582 /* make me the ilb owner */
4583 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4585 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4588 if (!(sched_smt_power_savings ||
4589 sched_mc_power_savings))
4592 * Check to see if there is a more power-efficient
4595 new_ilb = find_new_ilb(cpu);
4596 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4597 atomic_set(&nohz.load_balancer, -1);
4598 resched_cpu(new_ilb);
4604 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4607 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4609 if (atomic_read(&nohz.load_balancer) == cpu)
4610 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4617 static DEFINE_SPINLOCK(balancing);
4620 * It checks each scheduling domain to see if it is due to be balanced,
4621 * and initiates a balancing operation if so.
4623 * Balancing parameters are set up in arch_init_sched_domains.
4625 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4628 struct rq *rq = cpu_rq(cpu);
4629 unsigned long interval;
4630 struct sched_domain *sd;
4631 /* Earliest time when we have to do rebalance again */
4632 unsigned long next_balance = jiffies + 60*HZ;
4633 int update_next_balance = 0;
4636 for_each_domain(cpu, sd) {
4637 if (!(sd->flags & SD_LOAD_BALANCE))
4640 interval = sd->balance_interval;
4641 if (idle != CPU_IDLE)
4642 interval *= sd->busy_factor;
4644 /* scale ms to jiffies */
4645 interval = msecs_to_jiffies(interval);
4646 if (unlikely(!interval))
4648 if (interval > HZ*NR_CPUS/10)
4649 interval = HZ*NR_CPUS/10;
4651 need_serialize = sd->flags & SD_SERIALIZE;
4653 if (need_serialize) {
4654 if (!spin_trylock(&balancing))
4658 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4659 if (load_balance(cpu, rq, sd, idle, &balance)) {
4661 * We've pulled tasks over so either we're no
4662 * longer idle, or one of our SMT siblings is
4665 idle = CPU_NOT_IDLE;
4667 sd->last_balance = jiffies;
4670 spin_unlock(&balancing);
4672 if (time_after(next_balance, sd->last_balance + interval)) {
4673 next_balance = sd->last_balance + interval;
4674 update_next_balance = 1;
4678 * Stop the load balance at this level. There is another
4679 * CPU in our sched group which is doing load balancing more
4687 * next_balance will be updated only when there is a need.
4688 * When the cpu is attached to null domain for ex, it will not be
4691 if (likely(update_next_balance))
4692 rq->next_balance = next_balance;
4696 * run_rebalance_domains is triggered when needed from the scheduler tick.
4697 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4698 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4700 static void run_rebalance_domains(struct softirq_action *h)
4702 int this_cpu = smp_processor_id();
4703 struct rq *this_rq = cpu_rq(this_cpu);
4704 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4705 CPU_IDLE : CPU_NOT_IDLE;
4707 rebalance_domains(this_cpu, idle);
4711 * If this cpu is the owner for idle load balancing, then do the
4712 * balancing on behalf of the other idle cpus whose ticks are
4715 if (this_rq->idle_at_tick &&
4716 atomic_read(&nohz.load_balancer) == this_cpu) {
4720 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4721 if (balance_cpu == this_cpu)
4725 * If this cpu gets work to do, stop the load balancing
4726 * work being done for other cpus. Next load
4727 * balancing owner will pick it up.
4732 rebalance_domains(balance_cpu, CPU_IDLE);
4734 rq = cpu_rq(balance_cpu);
4735 if (time_after(this_rq->next_balance, rq->next_balance))
4736 this_rq->next_balance = rq->next_balance;
4742 static inline int on_null_domain(int cpu)
4744 return !rcu_dereference(cpu_rq(cpu)->sd);
4748 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4750 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4751 * idle load balancing owner or decide to stop the periodic load balancing,
4752 * if the whole system is idle.
4754 static inline void trigger_load_balance(struct rq *rq, int cpu)
4758 * If we were in the nohz mode recently and busy at the current
4759 * scheduler tick, then check if we need to nominate new idle
4762 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4763 rq->in_nohz_recently = 0;
4765 if (atomic_read(&nohz.load_balancer) == cpu) {
4766 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4767 atomic_set(&nohz.load_balancer, -1);
4770 if (atomic_read(&nohz.load_balancer) == -1) {
4771 int ilb = find_new_ilb(cpu);
4773 if (ilb < nr_cpu_ids)
4779 * If this cpu is idle and doing idle load balancing for all the
4780 * cpus with ticks stopped, is it time for that to stop?
4782 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4783 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4789 * If this cpu is idle and the idle load balancing is done by
4790 * someone else, then no need raise the SCHED_SOFTIRQ
4792 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4793 cpumask_test_cpu(cpu, nohz.cpu_mask))
4796 /* Don't need to rebalance while attached to NULL domain */
4797 if (time_after_eq(jiffies, rq->next_balance) &&
4798 likely(!on_null_domain(cpu)))
4799 raise_softirq(SCHED_SOFTIRQ);
4802 #else /* CONFIG_SMP */
4805 * on UP we do not need to balance between CPUs:
4807 static inline void idle_balance(int cpu, struct rq *rq)
4813 DEFINE_PER_CPU(struct kernel_stat, kstat);
4815 EXPORT_PER_CPU_SYMBOL(kstat);
4818 * Return any ns on the sched_clock that have not yet been accounted in
4819 * @p in case that task is currently running.
4821 * Called with task_rq_lock() held on @rq.
4823 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4827 if (task_current(rq, p)) {
4828 update_rq_clock(rq);
4829 ns = rq->clock - p->se.exec_start;
4837 unsigned long long task_delta_exec(struct task_struct *p)
4839 unsigned long flags;
4843 rq = task_rq_lock(p, &flags);
4844 ns = do_task_delta_exec(p, rq);
4845 task_rq_unlock(rq, &flags);
4851 * Return accounted runtime for the task.
4852 * In case the task is currently running, return the runtime plus current's
4853 * pending runtime that have not been accounted yet.
4855 unsigned long long task_sched_runtime(struct task_struct *p)
4857 unsigned long flags;
4861 rq = task_rq_lock(p, &flags);
4862 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4863 task_rq_unlock(rq, &flags);
4869 * Return sum_exec_runtime for the thread group.
4870 * In case the task is currently running, return the sum plus current's
4871 * pending runtime that have not been accounted yet.
4873 * Note that the thread group might have other running tasks as well,
4874 * so the return value not includes other pending runtime that other
4875 * running tasks might have.
4877 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4879 struct task_cputime totals;
4880 unsigned long flags;
4884 rq = task_rq_lock(p, &flags);
4885 thread_group_cputime(p, &totals);
4886 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4887 task_rq_unlock(rq, &flags);
4893 * Account user cpu time to a process.
4894 * @p: the process that the cpu time gets accounted to
4895 * @cputime: the cpu time spent in user space since the last update
4896 * @cputime_scaled: cputime scaled by cpu frequency
4898 void account_user_time(struct task_struct *p, cputime_t cputime,
4899 cputime_t cputime_scaled)
4901 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4904 /* Add user time to process. */
4905 p->utime = cputime_add(p->utime, cputime);
4906 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4907 account_group_user_time(p, cputime);
4909 /* Add user time to cpustat. */
4910 tmp = cputime_to_cputime64(cputime);
4911 if (TASK_NICE(p) > 0)
4912 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4914 cpustat->user = cputime64_add(cpustat->user, tmp);
4916 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4917 /* Account for user time used */
4918 acct_update_integrals(p);
4922 * Account guest cpu time to a process.
4923 * @p: the process that the cpu time gets accounted to
4924 * @cputime: the cpu time spent in virtual machine since the last update
4925 * @cputime_scaled: cputime scaled by cpu frequency
4927 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4928 cputime_t cputime_scaled)
4931 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4933 tmp = cputime_to_cputime64(cputime);
4935 /* Add guest time to process. */
4936 p->utime = cputime_add(p->utime, cputime);
4937 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4938 account_group_user_time(p, cputime);
4939 p->gtime = cputime_add(p->gtime, cputime);
4941 /* Add guest time to cpustat. */
4942 cpustat->user = cputime64_add(cpustat->user, tmp);
4943 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4947 * Account system cpu time to a process.
4948 * @p: the process that the cpu time gets accounted to
4949 * @hardirq_offset: the offset to subtract from hardirq_count()
4950 * @cputime: the cpu time spent in kernel space since the last update
4951 * @cputime_scaled: cputime scaled by cpu frequency
4953 void account_system_time(struct task_struct *p, int hardirq_offset,
4954 cputime_t cputime, cputime_t cputime_scaled)
4956 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4959 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4960 account_guest_time(p, cputime, cputime_scaled);
4964 /* Add system time to process. */
4965 p->stime = cputime_add(p->stime, cputime);
4966 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4967 account_group_system_time(p, cputime);
4969 /* Add system time to cpustat. */
4970 tmp = cputime_to_cputime64(cputime);
4971 if (hardirq_count() - hardirq_offset)
4972 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4973 else if (softirq_count())
4974 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4976 cpustat->system = cputime64_add(cpustat->system, tmp);
4978 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4980 /* Account for system time used */
4981 acct_update_integrals(p);
4985 * Account for involuntary wait time.
4986 * @steal: the cpu time spent in involuntary wait
4988 void account_steal_time(cputime_t cputime)
4990 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4991 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4993 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4997 * Account for idle time.
4998 * @cputime: the cpu time spent in idle wait
5000 void account_idle_time(cputime_t cputime)
5002 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5003 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5004 struct rq *rq = this_rq();
5006 if (atomic_read(&rq->nr_iowait) > 0)
5007 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5009 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5012 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5015 * Account a single tick of cpu time.
5016 * @p: the process that the cpu time gets accounted to
5017 * @user_tick: indicates if the tick is a user or a system tick
5019 void account_process_tick(struct task_struct *p, int user_tick)
5021 cputime_t one_jiffy = jiffies_to_cputime(1);
5022 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5023 struct rq *rq = this_rq();
5026 account_user_time(p, one_jiffy, one_jiffy_scaled);
5027 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5028 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5031 account_idle_time(one_jiffy);
5035 * Account multiple ticks of steal time.
5036 * @p: the process from which the cpu time has been stolen
5037 * @ticks: number of stolen ticks
5039 void account_steal_ticks(unsigned long ticks)
5041 account_steal_time(jiffies_to_cputime(ticks));
5045 * Account multiple ticks of idle time.
5046 * @ticks: number of stolen ticks
5048 void account_idle_ticks(unsigned long ticks)
5050 account_idle_time(jiffies_to_cputime(ticks));
5056 * Use precise platform statistics if available:
5058 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5059 cputime_t task_utime(struct task_struct *p)
5064 cputime_t task_stime(struct task_struct *p)
5069 cputime_t task_utime(struct task_struct *p)
5071 clock_t utime = cputime_to_clock_t(p->utime),
5072 total = utime + cputime_to_clock_t(p->stime);
5076 * Use CFS's precise accounting:
5078 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5082 do_div(temp, total);
5084 utime = (clock_t)temp;
5086 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5087 return p->prev_utime;
5090 cputime_t task_stime(struct task_struct *p)
5095 * Use CFS's precise accounting. (we subtract utime from
5096 * the total, to make sure the total observed by userspace
5097 * grows monotonically - apps rely on that):
5099 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5100 cputime_to_clock_t(task_utime(p));
5103 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5105 return p->prev_stime;
5109 inline cputime_t task_gtime(struct task_struct *p)
5115 * This function gets called by the timer code, with HZ frequency.
5116 * We call it with interrupts disabled.
5118 * It also gets called by the fork code, when changing the parent's
5121 void scheduler_tick(void)
5123 int cpu = smp_processor_id();
5124 struct rq *rq = cpu_rq(cpu);
5125 struct task_struct *curr = rq->curr;
5129 spin_lock(&rq->lock);
5130 update_rq_clock(rq);
5131 update_cpu_load(rq);
5132 curr->sched_class->task_tick(rq, curr, 0);
5133 spin_unlock(&rq->lock);
5135 perf_counter_task_tick(curr, cpu);
5138 rq->idle_at_tick = idle_cpu(cpu);
5139 trigger_load_balance(rq, cpu);
5143 notrace unsigned long get_parent_ip(unsigned long addr)
5145 if (in_lock_functions(addr)) {
5146 addr = CALLER_ADDR2;
5147 if (in_lock_functions(addr))
5148 addr = CALLER_ADDR3;
5153 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5154 defined(CONFIG_PREEMPT_TRACER))
5156 void __kprobes add_preempt_count(int val)
5158 #ifdef CONFIG_DEBUG_PREEMPT
5162 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5165 preempt_count() += val;
5166 #ifdef CONFIG_DEBUG_PREEMPT
5168 * Spinlock count overflowing soon?
5170 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5173 if (preempt_count() == val)
5174 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5176 EXPORT_SYMBOL(add_preempt_count);
5178 void __kprobes sub_preempt_count(int val)
5180 #ifdef CONFIG_DEBUG_PREEMPT
5184 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5187 * Is the spinlock portion underflowing?
5189 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5190 !(preempt_count() & PREEMPT_MASK)))
5194 if (preempt_count() == val)
5195 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5196 preempt_count() -= val;
5198 EXPORT_SYMBOL(sub_preempt_count);
5203 * Print scheduling while atomic bug:
5205 static noinline void __schedule_bug(struct task_struct *prev)
5207 struct pt_regs *regs = get_irq_regs();
5209 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5210 prev->comm, prev->pid, preempt_count());
5212 debug_show_held_locks(prev);
5214 if (irqs_disabled())
5215 print_irqtrace_events(prev);
5224 * Various schedule()-time debugging checks and statistics:
5226 static inline void schedule_debug(struct task_struct *prev)
5229 * Test if we are atomic. Since do_exit() needs to call into
5230 * schedule() atomically, we ignore that path for now.
5231 * Otherwise, whine if we are scheduling when we should not be.
5233 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5234 __schedule_bug(prev);
5236 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5238 schedstat_inc(this_rq(), sched_count);
5239 #ifdef CONFIG_SCHEDSTATS
5240 if (unlikely(prev->lock_depth >= 0)) {
5241 schedstat_inc(this_rq(), bkl_count);
5242 schedstat_inc(prev, sched_info.bkl_count);
5247 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5249 if (prev->state == TASK_RUNNING) {
5250 u64 runtime = prev->se.sum_exec_runtime;
5252 runtime -= prev->se.prev_sum_exec_runtime;
5253 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5256 * In order to avoid avg_overlap growing stale when we are
5257 * indeed overlapping and hence not getting put to sleep, grow
5258 * the avg_overlap on preemption.
5260 * We use the average preemption runtime because that
5261 * correlates to the amount of cache footprint a task can
5264 update_avg(&prev->se.avg_overlap, runtime);
5266 prev->sched_class->put_prev_task(rq, prev);
5270 * Pick up the highest-prio task:
5272 static inline struct task_struct *
5273 pick_next_task(struct rq *rq)
5275 const struct sched_class *class;
5276 struct task_struct *p;
5279 * Optimization: we know that if all tasks are in
5280 * the fair class we can call that function directly:
5282 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5283 p = fair_sched_class.pick_next_task(rq);
5288 class = sched_class_highest;
5290 p = class->pick_next_task(rq);
5294 * Will never be NULL as the idle class always
5295 * returns a non-NULL p:
5297 class = class->next;
5302 * schedule() is the main scheduler function.
5304 asmlinkage void __sched schedule(void)
5306 struct task_struct *prev, *next;
5307 unsigned long *switch_count;
5313 cpu = smp_processor_id();
5317 switch_count = &prev->nivcsw;
5319 release_kernel_lock(prev);
5320 need_resched_nonpreemptible:
5322 schedule_debug(prev);
5324 if (sched_feat(HRTICK))
5327 spin_lock_irq(&rq->lock);
5328 update_rq_clock(rq);
5329 clear_tsk_need_resched(prev);
5331 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5332 if (unlikely(signal_pending_state(prev->state, prev)))
5333 prev->state = TASK_RUNNING;
5335 deactivate_task(rq, prev, 1);
5336 switch_count = &prev->nvcsw;
5340 if (prev->sched_class->pre_schedule)
5341 prev->sched_class->pre_schedule(rq, prev);
5344 if (unlikely(!rq->nr_running))
5345 idle_balance(cpu, rq);
5347 put_prev_task(rq, prev);
5348 next = pick_next_task(rq);
5350 if (likely(prev != next)) {
5351 sched_info_switch(prev, next);
5352 perf_counter_task_sched_out(prev, next, cpu);
5358 context_switch(rq, prev, next); /* unlocks the rq */
5360 * the context switch might have flipped the stack from under
5361 * us, hence refresh the local variables.
5363 cpu = smp_processor_id();
5366 spin_unlock_irq(&rq->lock);
5368 if (unlikely(reacquire_kernel_lock(current) < 0))
5369 goto need_resched_nonpreemptible;
5371 preempt_enable_no_resched();
5375 EXPORT_SYMBOL(schedule);
5379 * Look out! "owner" is an entirely speculative pointer
5380 * access and not reliable.
5382 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5387 if (!sched_feat(OWNER_SPIN))
5390 #ifdef CONFIG_DEBUG_PAGEALLOC
5392 * Need to access the cpu field knowing that
5393 * DEBUG_PAGEALLOC could have unmapped it if
5394 * the mutex owner just released it and exited.
5396 if (probe_kernel_address(&owner->cpu, cpu))
5403 * Even if the access succeeded (likely case),
5404 * the cpu field may no longer be valid.
5406 if (cpu >= nr_cpumask_bits)
5410 * We need to validate that we can do a
5411 * get_cpu() and that we have the percpu area.
5413 if (!cpu_online(cpu))
5420 * Owner changed, break to re-assess state.
5422 if (lock->owner != owner)
5426 * Is that owner really running on that cpu?
5428 if (task_thread_info(rq->curr) != owner || need_resched())
5438 #ifdef CONFIG_PREEMPT
5440 * this is the entry point to schedule() from in-kernel preemption
5441 * off of preempt_enable. Kernel preemptions off return from interrupt
5442 * occur there and call schedule directly.
5444 asmlinkage void __sched preempt_schedule(void)
5446 struct thread_info *ti = current_thread_info();
5449 * If there is a non-zero preempt_count or interrupts are disabled,
5450 * we do not want to preempt the current task. Just return..
5452 if (likely(ti->preempt_count || irqs_disabled()))
5456 add_preempt_count(PREEMPT_ACTIVE);
5458 sub_preempt_count(PREEMPT_ACTIVE);
5461 * Check again in case we missed a preemption opportunity
5462 * between schedule and now.
5465 } while (need_resched());
5467 EXPORT_SYMBOL(preempt_schedule);
5470 * this is the entry point to schedule() from kernel preemption
5471 * off of irq context.
5472 * Note, that this is called and return with irqs disabled. This will
5473 * protect us against recursive calling from irq.
5475 asmlinkage void __sched preempt_schedule_irq(void)
5477 struct thread_info *ti = current_thread_info();
5479 /* Catch callers which need to be fixed */
5480 BUG_ON(ti->preempt_count || !irqs_disabled());
5483 add_preempt_count(PREEMPT_ACTIVE);
5486 local_irq_disable();
5487 sub_preempt_count(PREEMPT_ACTIVE);
5490 * Check again in case we missed a preemption opportunity
5491 * between schedule and now.
5494 } while (need_resched());
5497 #endif /* CONFIG_PREEMPT */
5499 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5502 return try_to_wake_up(curr->private, mode, sync);
5504 EXPORT_SYMBOL(default_wake_function);
5507 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5508 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5509 * number) then we wake all the non-exclusive tasks and one exclusive task.
5511 * There are circumstances in which we can try to wake a task which has already
5512 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5513 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5515 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5516 int nr_exclusive, int sync, void *key)
5518 wait_queue_t *curr, *next;
5520 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5521 unsigned flags = curr->flags;
5523 if (curr->func(curr, mode, sync, key) &&
5524 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5530 * __wake_up - wake up threads blocked on a waitqueue.
5532 * @mode: which threads
5533 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5534 * @key: is directly passed to the wakeup function
5536 * It may be assumed that this function implies a write memory barrier before
5537 * changing the task state if and only if any tasks are woken up.
5539 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5540 int nr_exclusive, void *key)
5542 unsigned long flags;
5544 spin_lock_irqsave(&q->lock, flags);
5545 __wake_up_common(q, mode, nr_exclusive, 0, key);
5546 spin_unlock_irqrestore(&q->lock, flags);
5548 EXPORT_SYMBOL(__wake_up);
5551 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5553 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5555 __wake_up_common(q, mode, 1, 0, NULL);
5558 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5560 __wake_up_common(q, mode, 1, 0, key);
5564 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5566 * @mode: which threads
5567 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5568 * @key: opaque value to be passed to wakeup targets
5570 * The sync wakeup differs that the waker knows that it will schedule
5571 * away soon, so while the target thread will be woken up, it will not
5572 * be migrated to another CPU - ie. the two threads are 'synchronized'
5573 * with each other. This can prevent needless bouncing between CPUs.
5575 * On UP it can prevent extra preemption.
5577 * It may be assumed that this function implies a write memory barrier before
5578 * changing the task state if and only if any tasks are woken up.
5580 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5581 int nr_exclusive, void *key)
5583 unsigned long flags;
5589 if (unlikely(!nr_exclusive))
5592 spin_lock_irqsave(&q->lock, flags);
5593 __wake_up_common(q, mode, nr_exclusive, sync, key);
5594 spin_unlock_irqrestore(&q->lock, flags);
5596 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5599 * __wake_up_sync - see __wake_up_sync_key()
5601 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5603 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5605 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5608 * complete: - signals a single thread waiting on this completion
5609 * @x: holds the state of this particular completion
5611 * This will wake up a single thread waiting on this completion. Threads will be
5612 * awakened in the same order in which they were queued.
5614 * See also complete_all(), wait_for_completion() and related routines.
5616 * It may be assumed that this function implies a write memory barrier before
5617 * changing the task state if and only if any tasks are woken up.
5619 void complete(struct completion *x)
5621 unsigned long flags;
5623 spin_lock_irqsave(&x->wait.lock, flags);
5625 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5626 spin_unlock_irqrestore(&x->wait.lock, flags);
5628 EXPORT_SYMBOL(complete);
5631 * complete_all: - signals all threads waiting on this completion
5632 * @x: holds the state of this particular completion
5634 * This will wake up all threads waiting on this particular completion event.
5636 * It may be assumed that this function implies a write memory barrier before
5637 * changing the task state if and only if any tasks are woken up.
5639 void complete_all(struct completion *x)
5641 unsigned long flags;
5643 spin_lock_irqsave(&x->wait.lock, flags);
5644 x->done += UINT_MAX/2;
5645 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5646 spin_unlock_irqrestore(&x->wait.lock, flags);
5648 EXPORT_SYMBOL(complete_all);
5650 static inline long __sched
5651 do_wait_for_common(struct completion *x, long timeout, int state)
5654 DECLARE_WAITQUEUE(wait, current);
5656 wait.flags |= WQ_FLAG_EXCLUSIVE;
5657 __add_wait_queue_tail(&x->wait, &wait);
5659 if (signal_pending_state(state, current)) {
5660 timeout = -ERESTARTSYS;
5663 __set_current_state(state);
5664 spin_unlock_irq(&x->wait.lock);
5665 timeout = schedule_timeout(timeout);
5666 spin_lock_irq(&x->wait.lock);
5667 } while (!x->done && timeout);
5668 __remove_wait_queue(&x->wait, &wait);
5673 return timeout ?: 1;
5677 wait_for_common(struct completion *x, long timeout, int state)
5681 spin_lock_irq(&x->wait.lock);
5682 timeout = do_wait_for_common(x, timeout, state);
5683 spin_unlock_irq(&x->wait.lock);
5688 * wait_for_completion: - waits for completion of a task
5689 * @x: holds the state of this particular completion
5691 * This waits to be signaled for completion of a specific task. It is NOT
5692 * interruptible and there is no timeout.
5694 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5695 * and interrupt capability. Also see complete().
5697 void __sched wait_for_completion(struct completion *x)
5699 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5701 EXPORT_SYMBOL(wait_for_completion);
5704 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5705 * @x: holds the state of this particular completion
5706 * @timeout: timeout value in jiffies
5708 * This waits for either a completion of a specific task to be signaled or for a
5709 * specified timeout to expire. The timeout is in jiffies. It is not
5712 unsigned long __sched
5713 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5715 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5717 EXPORT_SYMBOL(wait_for_completion_timeout);
5720 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5721 * @x: holds the state of this particular completion
5723 * This waits for completion of a specific task to be signaled. It is
5726 int __sched wait_for_completion_interruptible(struct completion *x)
5728 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5729 if (t == -ERESTARTSYS)
5733 EXPORT_SYMBOL(wait_for_completion_interruptible);
5736 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5737 * @x: holds the state of this particular completion
5738 * @timeout: timeout value in jiffies
5740 * This waits for either a completion of a specific task to be signaled or for a
5741 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5743 unsigned long __sched
5744 wait_for_completion_interruptible_timeout(struct completion *x,
5745 unsigned long timeout)
5747 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5749 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5752 * wait_for_completion_killable: - waits for completion of a task (killable)
5753 * @x: holds the state of this particular completion
5755 * This waits to be signaled for completion of a specific task. It can be
5756 * interrupted by a kill signal.
5758 int __sched wait_for_completion_killable(struct completion *x)
5760 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5761 if (t == -ERESTARTSYS)
5765 EXPORT_SYMBOL(wait_for_completion_killable);
5768 * try_wait_for_completion - try to decrement a completion without blocking
5769 * @x: completion structure
5771 * Returns: 0 if a decrement cannot be done without blocking
5772 * 1 if a decrement succeeded.
5774 * If a completion is being used as a counting completion,
5775 * attempt to decrement the counter without blocking. This
5776 * enables us to avoid waiting if the resource the completion
5777 * is protecting is not available.
5779 bool try_wait_for_completion(struct completion *x)
5783 spin_lock_irq(&x->wait.lock);
5788 spin_unlock_irq(&x->wait.lock);
5791 EXPORT_SYMBOL(try_wait_for_completion);
5794 * completion_done - Test to see if a completion has any waiters
5795 * @x: completion structure
5797 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5798 * 1 if there are no waiters.
5801 bool completion_done(struct completion *x)
5805 spin_lock_irq(&x->wait.lock);
5808 spin_unlock_irq(&x->wait.lock);
5811 EXPORT_SYMBOL(completion_done);
5814 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5816 unsigned long flags;
5819 init_waitqueue_entry(&wait, current);
5821 __set_current_state(state);
5823 spin_lock_irqsave(&q->lock, flags);
5824 __add_wait_queue(q, &wait);
5825 spin_unlock(&q->lock);
5826 timeout = schedule_timeout(timeout);
5827 spin_lock_irq(&q->lock);
5828 __remove_wait_queue(q, &wait);
5829 spin_unlock_irqrestore(&q->lock, flags);
5834 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5836 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5838 EXPORT_SYMBOL(interruptible_sleep_on);
5841 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5843 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5845 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5847 void __sched sleep_on(wait_queue_head_t *q)
5849 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5851 EXPORT_SYMBOL(sleep_on);
5853 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5855 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5857 EXPORT_SYMBOL(sleep_on_timeout);
5859 #ifdef CONFIG_RT_MUTEXES
5862 * rt_mutex_setprio - set the current priority of a task
5864 * @prio: prio value (kernel-internal form)
5866 * This function changes the 'effective' priority of a task. It does
5867 * not touch ->normal_prio like __setscheduler().
5869 * Used by the rt_mutex code to implement priority inheritance logic.
5871 void rt_mutex_setprio(struct task_struct *p, int prio)
5873 unsigned long flags;
5874 int oldprio, on_rq, running;
5876 const struct sched_class *prev_class = p->sched_class;
5878 BUG_ON(prio < 0 || prio > MAX_PRIO);
5880 rq = task_rq_lock(p, &flags);
5881 update_rq_clock(rq);
5884 on_rq = p->se.on_rq;
5885 running = task_current(rq, p);
5887 dequeue_task(rq, p, 0);
5889 p->sched_class->put_prev_task(rq, p);
5892 p->sched_class = &rt_sched_class;
5894 p->sched_class = &fair_sched_class;
5899 p->sched_class->set_curr_task(rq);
5901 enqueue_task(rq, p, 0);
5903 check_class_changed(rq, p, prev_class, oldprio, running);
5905 task_rq_unlock(rq, &flags);
5910 void set_user_nice(struct task_struct *p, long nice)
5912 int old_prio, delta, on_rq;
5913 unsigned long flags;
5916 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5919 * We have to be careful, if called from sys_setpriority(),
5920 * the task might be in the middle of scheduling on another CPU.
5922 rq = task_rq_lock(p, &flags);
5923 update_rq_clock(rq);
5925 * The RT priorities are set via sched_setscheduler(), but we still
5926 * allow the 'normal' nice value to be set - but as expected
5927 * it wont have any effect on scheduling until the task is
5928 * SCHED_FIFO/SCHED_RR:
5930 if (task_has_rt_policy(p)) {
5931 p->static_prio = NICE_TO_PRIO(nice);
5934 on_rq = p->se.on_rq;
5936 dequeue_task(rq, p, 0);
5938 p->static_prio = NICE_TO_PRIO(nice);
5941 p->prio = effective_prio(p);
5942 delta = p->prio - old_prio;
5945 enqueue_task(rq, p, 0);
5947 * If the task increased its priority or is running and
5948 * lowered its priority, then reschedule its CPU:
5950 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5951 resched_task(rq->curr);
5954 task_rq_unlock(rq, &flags);
5956 EXPORT_SYMBOL(set_user_nice);
5959 * can_nice - check if a task can reduce its nice value
5963 int can_nice(const struct task_struct *p, const int nice)
5965 /* convert nice value [19,-20] to rlimit style value [1,40] */
5966 int nice_rlim = 20 - nice;
5968 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5969 capable(CAP_SYS_NICE));
5972 #ifdef __ARCH_WANT_SYS_NICE
5975 * sys_nice - change the priority of the current process.
5976 * @increment: priority increment
5978 * sys_setpriority is a more generic, but much slower function that
5979 * does similar things.
5981 SYSCALL_DEFINE1(nice, int, increment)
5986 * Setpriority might change our priority at the same moment.
5987 * We don't have to worry. Conceptually one call occurs first
5988 * and we have a single winner.
5990 if (increment < -40)
5995 nice = TASK_NICE(current) + increment;
6001 if (increment < 0 && !can_nice(current, nice))
6004 retval = security_task_setnice(current, nice);
6008 set_user_nice(current, nice);
6015 * task_prio - return the priority value of a given task.
6016 * @p: the task in question.
6018 * This is the priority value as seen by users in /proc.
6019 * RT tasks are offset by -200. Normal tasks are centered
6020 * around 0, value goes from -16 to +15.
6022 int task_prio(const struct task_struct *p)
6024 return p->prio - MAX_RT_PRIO;
6028 * task_nice - return the nice value of a given task.
6029 * @p: the task in question.
6031 int task_nice(const struct task_struct *p)
6033 return TASK_NICE(p);
6035 EXPORT_SYMBOL(task_nice);
6038 * idle_cpu - is a given cpu idle currently?
6039 * @cpu: the processor in question.
6041 int idle_cpu(int cpu)
6043 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6047 * idle_task - return the idle task for a given cpu.
6048 * @cpu: the processor in question.
6050 struct task_struct *idle_task(int cpu)
6052 return cpu_rq(cpu)->idle;
6056 * find_process_by_pid - find a process with a matching PID value.
6057 * @pid: the pid in question.
6059 static struct task_struct *find_process_by_pid(pid_t pid)
6061 return pid ? find_task_by_vpid(pid) : current;
6064 /* Actually do priority change: must hold rq lock. */
6066 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6068 BUG_ON(p->se.on_rq);
6071 switch (p->policy) {
6075 p->sched_class = &fair_sched_class;
6079 p->sched_class = &rt_sched_class;
6083 p->rt_priority = prio;
6084 p->normal_prio = normal_prio(p);
6085 /* we are holding p->pi_lock already */
6086 p->prio = rt_mutex_getprio(p);
6091 * check the target process has a UID that matches the current process's
6093 static bool check_same_owner(struct task_struct *p)
6095 const struct cred *cred = current_cred(), *pcred;
6099 pcred = __task_cred(p);
6100 match = (cred->euid == pcred->euid ||
6101 cred->euid == pcred->uid);
6106 static int __sched_setscheduler(struct task_struct *p, int policy,
6107 struct sched_param *param, bool user)
6109 int retval, oldprio, oldpolicy = -1, on_rq, running;
6110 unsigned long flags;
6111 const struct sched_class *prev_class = p->sched_class;
6115 /* may grab non-irq protected spin_locks */
6116 BUG_ON(in_interrupt());
6118 /* double check policy once rq lock held */
6120 reset_on_fork = p->sched_reset_on_fork;
6121 policy = oldpolicy = p->policy;
6123 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6124 policy &= ~SCHED_RESET_ON_FORK;
6126 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6127 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6128 policy != SCHED_IDLE)
6133 * Valid priorities for SCHED_FIFO and SCHED_RR are
6134 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6135 * SCHED_BATCH and SCHED_IDLE is 0.
6137 if (param->sched_priority < 0 ||
6138 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6139 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6141 if (rt_policy(policy) != (param->sched_priority != 0))
6145 * Allow unprivileged RT tasks to decrease priority:
6147 if (user && !capable(CAP_SYS_NICE)) {
6148 if (rt_policy(policy)) {
6149 unsigned long rlim_rtprio;
6151 if (!lock_task_sighand(p, &flags))
6153 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6154 unlock_task_sighand(p, &flags);
6156 /* can't set/change the rt policy */
6157 if (policy != p->policy && !rlim_rtprio)
6160 /* can't increase priority */
6161 if (param->sched_priority > p->rt_priority &&
6162 param->sched_priority > rlim_rtprio)
6166 * Like positive nice levels, dont allow tasks to
6167 * move out of SCHED_IDLE either:
6169 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6172 /* can't change other user's priorities */
6173 if (!check_same_owner(p))
6176 /* Normal users shall not reset the sched_reset_on_fork flag */
6177 if (p->sched_reset_on_fork && !reset_on_fork)
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);
6222 p->sched_reset_on_fork = reset_on_fork;
6225 __setscheduler(rq, p, policy, param->sched_priority);
6228 p->sched_class->set_curr_task(rq);
6230 activate_task(rq, p, 0);
6232 check_class_changed(rq, p, prev_class, oldprio, running);
6234 __task_rq_unlock(rq);
6235 spin_unlock_irqrestore(&p->pi_lock, flags);
6237 rt_mutex_adjust_pi(p);
6243 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6244 * @p: the task in question.
6245 * @policy: new policy.
6246 * @param: structure containing the new RT priority.
6248 * NOTE that the task may be already dead.
6250 int sched_setscheduler(struct task_struct *p, int policy,
6251 struct sched_param *param)
6253 return __sched_setscheduler(p, policy, param, true);
6255 EXPORT_SYMBOL_GPL(sched_setscheduler);
6258 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6259 * @p: the task in question.
6260 * @policy: new policy.
6261 * @param: structure containing the new RT priority.
6263 * Just like sched_setscheduler, only don't bother checking if the
6264 * current context has permission. For example, this is needed in
6265 * stop_machine(): we create temporary high priority worker threads,
6266 * but our caller might not have that capability.
6268 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6269 struct sched_param *param)
6271 return __sched_setscheduler(p, policy, param, false);
6275 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6277 struct sched_param lparam;
6278 struct task_struct *p;
6281 if (!param || pid < 0)
6283 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6288 p = find_process_by_pid(pid);
6290 retval = sched_setscheduler(p, policy, &lparam);
6297 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6298 * @pid: the pid in question.
6299 * @policy: new policy.
6300 * @param: structure containing the new RT priority.
6302 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6303 struct sched_param __user *, param)
6305 /* negative values for policy are not valid */
6309 return do_sched_setscheduler(pid, policy, param);
6313 * sys_sched_setparam - set/change the RT priority of a thread
6314 * @pid: the pid in question.
6315 * @param: structure containing the new RT priority.
6317 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6319 return do_sched_setscheduler(pid, -1, param);
6323 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6324 * @pid: the pid in question.
6326 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6328 struct task_struct *p;
6335 read_lock(&tasklist_lock);
6336 p = find_process_by_pid(pid);
6338 retval = security_task_getscheduler(p);
6341 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6343 read_unlock(&tasklist_lock);
6348 * sys_sched_getparam - get the RT priority of a thread
6349 * @pid: the pid in question.
6350 * @param: structure containing the RT priority.
6352 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6354 struct sched_param lp;
6355 struct task_struct *p;
6358 if (!param || pid < 0)
6361 read_lock(&tasklist_lock);
6362 p = find_process_by_pid(pid);
6367 retval = security_task_getscheduler(p);
6371 lp.sched_priority = p->rt_priority;
6372 read_unlock(&tasklist_lock);
6375 * This one might sleep, we cannot do it with a spinlock held ...
6377 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6382 read_unlock(&tasklist_lock);
6386 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6388 cpumask_var_t cpus_allowed, new_mask;
6389 struct task_struct *p;
6393 read_lock(&tasklist_lock);
6395 p = find_process_by_pid(pid);
6397 read_unlock(&tasklist_lock);
6403 * It is not safe to call set_cpus_allowed with the
6404 * tasklist_lock held. We will bump the task_struct's
6405 * usage count and then drop tasklist_lock.
6408 read_unlock(&tasklist_lock);
6410 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6414 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6416 goto out_free_cpus_allowed;
6419 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6422 retval = security_task_setscheduler(p, 0, NULL);
6426 cpuset_cpus_allowed(p, cpus_allowed);
6427 cpumask_and(new_mask, in_mask, cpus_allowed);
6429 retval = set_cpus_allowed_ptr(p, new_mask);
6432 cpuset_cpus_allowed(p, cpus_allowed);
6433 if (!cpumask_subset(new_mask, cpus_allowed)) {
6435 * We must have raced with a concurrent cpuset
6436 * update. Just reset the cpus_allowed to the
6437 * cpuset's cpus_allowed
6439 cpumask_copy(new_mask, cpus_allowed);
6444 free_cpumask_var(new_mask);
6445 out_free_cpus_allowed:
6446 free_cpumask_var(cpus_allowed);
6453 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6454 struct cpumask *new_mask)
6456 if (len < cpumask_size())
6457 cpumask_clear(new_mask);
6458 else if (len > cpumask_size())
6459 len = cpumask_size();
6461 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6465 * sys_sched_setaffinity - set the cpu affinity of a process
6466 * @pid: pid of the process
6467 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6468 * @user_mask_ptr: user-space pointer to the new cpu mask
6470 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6471 unsigned long __user *, user_mask_ptr)
6473 cpumask_var_t new_mask;
6476 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6479 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6481 retval = sched_setaffinity(pid, new_mask);
6482 free_cpumask_var(new_mask);
6486 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6488 struct task_struct *p;
6492 read_lock(&tasklist_lock);
6495 p = find_process_by_pid(pid);
6499 retval = security_task_getscheduler(p);
6503 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6506 read_unlock(&tasklist_lock);
6513 * sys_sched_getaffinity - get the cpu affinity of a process
6514 * @pid: pid of the process
6515 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6516 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6518 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6519 unsigned long __user *, user_mask_ptr)
6524 if (len < cpumask_size())
6527 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6530 ret = sched_getaffinity(pid, mask);
6532 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6535 ret = cpumask_size();
6537 free_cpumask_var(mask);
6543 * sys_sched_yield - yield the current processor to other threads.
6545 * This function yields the current CPU to other tasks. If there are no
6546 * other threads running on this CPU then this function will return.
6548 SYSCALL_DEFINE0(sched_yield)
6550 struct rq *rq = this_rq_lock();
6552 schedstat_inc(rq, yld_count);
6553 current->sched_class->yield_task(rq);
6556 * Since we are going to call schedule() anyway, there's
6557 * no need to preempt or enable interrupts:
6559 __release(rq->lock);
6560 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6561 _raw_spin_unlock(&rq->lock);
6562 preempt_enable_no_resched();
6569 static void __cond_resched(void)
6571 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6572 __might_sleep(__FILE__, __LINE__);
6575 * The BKS might be reacquired before we have dropped
6576 * PREEMPT_ACTIVE, which could trigger a second
6577 * cond_resched() call.
6580 add_preempt_count(PREEMPT_ACTIVE);
6582 sub_preempt_count(PREEMPT_ACTIVE);
6583 } while (need_resched());
6586 int __sched _cond_resched(void)
6588 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
6589 system_state == SYSTEM_RUNNING) {
6595 EXPORT_SYMBOL(_cond_resched);
6598 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6599 * call schedule, and on return reacquire the lock.
6601 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6602 * operations here to prevent schedule() from being called twice (once via
6603 * spin_unlock(), once by hand).
6605 int cond_resched_lock(spinlock_t *lock)
6607 int resched = need_resched() && system_state == SYSTEM_RUNNING;
6610 if (spin_needbreak(lock) || resched) {
6612 if (resched && need_resched())
6621 EXPORT_SYMBOL(cond_resched_lock);
6623 int __sched cond_resched_softirq(void)
6625 BUG_ON(!in_softirq());
6627 if (need_resched() && system_state == SYSTEM_RUNNING) {
6635 EXPORT_SYMBOL(cond_resched_softirq);
6638 * yield - yield the current processor to other threads.
6640 * This is a shortcut for kernel-space yielding - it marks the
6641 * thread runnable and calls sys_sched_yield().
6643 void __sched yield(void)
6645 set_current_state(TASK_RUNNING);
6648 EXPORT_SYMBOL(yield);
6651 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6652 * that process accounting knows that this is a task in IO wait state.
6654 * But don't do that if it is a deliberate, throttling IO wait (this task
6655 * has set its backing_dev_info: the queue against which it should throttle)
6657 void __sched io_schedule(void)
6659 struct rq *rq = &__raw_get_cpu_var(runqueues);
6661 delayacct_blkio_start();
6662 atomic_inc(&rq->nr_iowait);
6664 atomic_dec(&rq->nr_iowait);
6665 delayacct_blkio_end();
6667 EXPORT_SYMBOL(io_schedule);
6669 long __sched io_schedule_timeout(long timeout)
6671 struct rq *rq = &__raw_get_cpu_var(runqueues);
6674 delayacct_blkio_start();
6675 atomic_inc(&rq->nr_iowait);
6676 ret = schedule_timeout(timeout);
6677 atomic_dec(&rq->nr_iowait);
6678 delayacct_blkio_end();
6683 * sys_sched_get_priority_max - return maximum RT priority.
6684 * @policy: scheduling class.
6686 * this syscall returns the maximum rt_priority that can be used
6687 * by a given scheduling class.
6689 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6696 ret = MAX_USER_RT_PRIO-1;
6708 * sys_sched_get_priority_min - return minimum RT priority.
6709 * @policy: scheduling class.
6711 * this syscall returns the minimum rt_priority that can be used
6712 * by a given scheduling class.
6714 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6732 * sys_sched_rr_get_interval - return the default timeslice of a process.
6733 * @pid: pid of the process.
6734 * @interval: userspace pointer to the timeslice value.
6736 * this syscall writes the default timeslice value of a given process
6737 * into the user-space timespec buffer. A value of '0' means infinity.
6739 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6740 struct timespec __user *, interval)
6742 struct task_struct *p;
6743 unsigned int time_slice;
6751 read_lock(&tasklist_lock);
6752 p = find_process_by_pid(pid);
6756 retval = security_task_getscheduler(p);
6761 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6762 * tasks that are on an otherwise idle runqueue:
6765 if (p->policy == SCHED_RR) {
6766 time_slice = DEF_TIMESLICE;
6767 } else if (p->policy != SCHED_FIFO) {
6768 struct sched_entity *se = &p->se;
6769 unsigned long flags;
6772 rq = task_rq_lock(p, &flags);
6773 if (rq->cfs.load.weight)
6774 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6775 task_rq_unlock(rq, &flags);
6777 read_unlock(&tasklist_lock);
6778 jiffies_to_timespec(time_slice, &t);
6779 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6783 read_unlock(&tasklist_lock);
6787 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6789 void sched_show_task(struct task_struct *p)
6791 unsigned long free = 0;
6794 state = p->state ? __ffs(p->state) + 1 : 0;
6795 printk(KERN_INFO "%-13.13s %c", p->comm,
6796 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6797 #if BITS_PER_LONG == 32
6798 if (state == TASK_RUNNING)
6799 printk(KERN_CONT " running ");
6801 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6803 if (state == TASK_RUNNING)
6804 printk(KERN_CONT " running task ");
6806 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6808 #ifdef CONFIG_DEBUG_STACK_USAGE
6809 free = stack_not_used(p);
6811 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6812 task_pid_nr(p), task_pid_nr(p->real_parent),
6813 (unsigned long)task_thread_info(p)->flags);
6815 show_stack(p, NULL);
6818 void show_state_filter(unsigned long state_filter)
6820 struct task_struct *g, *p;
6822 #if BITS_PER_LONG == 32
6824 " task PC stack pid father\n");
6827 " task PC stack pid father\n");
6829 read_lock(&tasklist_lock);
6830 do_each_thread(g, p) {
6832 * reset the NMI-timeout, listing all files on a slow
6833 * console might take alot of time:
6835 touch_nmi_watchdog();
6836 if (!state_filter || (p->state & state_filter))
6838 } while_each_thread(g, p);
6840 touch_all_softlockup_watchdogs();
6842 #ifdef CONFIG_SCHED_DEBUG
6843 sysrq_sched_debug_show();
6845 read_unlock(&tasklist_lock);
6847 * Only show locks if all tasks are dumped:
6849 if (state_filter == -1)
6850 debug_show_all_locks();
6853 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6855 idle->sched_class = &idle_sched_class;
6859 * init_idle - set up an idle thread for a given CPU
6860 * @idle: task in question
6861 * @cpu: cpu the idle task belongs to
6863 * NOTE: this function does not set the idle thread's NEED_RESCHED
6864 * flag, to make booting more robust.
6866 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6868 struct rq *rq = cpu_rq(cpu);
6869 unsigned long flags;
6871 spin_lock_irqsave(&rq->lock, flags);
6874 idle->se.exec_start = sched_clock();
6876 idle->prio = idle->normal_prio = MAX_PRIO;
6877 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6878 __set_task_cpu(idle, cpu);
6880 rq->curr = rq->idle = idle;
6881 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6884 spin_unlock_irqrestore(&rq->lock, flags);
6886 /* Set the preempt count _outside_ the spinlocks! */
6887 #if defined(CONFIG_PREEMPT)
6888 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6890 task_thread_info(idle)->preempt_count = 0;
6893 * The idle tasks have their own, simple scheduling class:
6895 idle->sched_class = &idle_sched_class;
6896 ftrace_graph_init_task(idle);
6900 * In a system that switches off the HZ timer nohz_cpu_mask
6901 * indicates which cpus entered this state. This is used
6902 * in the rcu update to wait only for active cpus. For system
6903 * which do not switch off the HZ timer nohz_cpu_mask should
6904 * always be CPU_BITS_NONE.
6906 cpumask_var_t nohz_cpu_mask;
6909 * Increase the granularity value when there are more CPUs,
6910 * because with more CPUs the 'effective latency' as visible
6911 * to users decreases. But the relationship is not linear,
6912 * so pick a second-best guess by going with the log2 of the
6915 * This idea comes from the SD scheduler of Con Kolivas:
6917 static inline void sched_init_granularity(void)
6919 unsigned int factor = 1 + ilog2(num_online_cpus());
6920 const unsigned long limit = 200000000;
6922 sysctl_sched_min_granularity *= factor;
6923 if (sysctl_sched_min_granularity > limit)
6924 sysctl_sched_min_granularity = limit;
6926 sysctl_sched_latency *= factor;
6927 if (sysctl_sched_latency > limit)
6928 sysctl_sched_latency = limit;
6930 sysctl_sched_wakeup_granularity *= factor;
6932 sysctl_sched_shares_ratelimit *= factor;
6937 * This is how migration works:
6939 * 1) we queue a struct migration_req structure in the source CPU's
6940 * runqueue and wake up that CPU's migration thread.
6941 * 2) we down() the locked semaphore => thread blocks.
6942 * 3) migration thread wakes up (implicitly it forces the migrated
6943 * thread off the CPU)
6944 * 4) it gets the migration request and checks whether the migrated
6945 * task is still in the wrong runqueue.
6946 * 5) if it's in the wrong runqueue then the migration thread removes
6947 * it and puts it into the right queue.
6948 * 6) migration thread up()s the semaphore.
6949 * 7) we wake up and the migration is done.
6953 * Change a given task's CPU affinity. Migrate the thread to a
6954 * proper CPU and schedule it away if the CPU it's executing on
6955 * is removed from the allowed bitmask.
6957 * NOTE: the caller must have a valid reference to the task, the
6958 * task must not exit() & deallocate itself prematurely. The
6959 * call is not atomic; no spinlocks may be held.
6961 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6963 struct migration_req req;
6964 unsigned long flags;
6968 rq = task_rq_lock(p, &flags);
6969 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6974 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6975 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6980 if (p->sched_class->set_cpus_allowed)
6981 p->sched_class->set_cpus_allowed(p, new_mask);
6983 cpumask_copy(&p->cpus_allowed, new_mask);
6984 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6987 /* Can the task run on the task's current CPU? If so, we're done */
6988 if (cpumask_test_cpu(task_cpu(p), new_mask))
6991 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6992 /* Need help from migration thread: drop lock and wait. */
6993 task_rq_unlock(rq, &flags);
6994 wake_up_process(rq->migration_thread);
6995 wait_for_completion(&req.done);
6996 tlb_migrate_finish(p->mm);
7000 task_rq_unlock(rq, &flags);
7004 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7007 * Move (not current) task off this cpu, onto dest cpu. We're doing
7008 * this because either it can't run here any more (set_cpus_allowed()
7009 * away from this CPU, or CPU going down), or because we're
7010 * attempting to rebalance this task on exec (sched_exec).
7012 * So we race with normal scheduler movements, but that's OK, as long
7013 * as the task is no longer on this CPU.
7015 * Returns non-zero if task was successfully migrated.
7017 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7019 struct rq *rq_dest, *rq_src;
7022 if (unlikely(!cpu_active(dest_cpu)))
7025 rq_src = cpu_rq(src_cpu);
7026 rq_dest = cpu_rq(dest_cpu);
7028 double_rq_lock(rq_src, rq_dest);
7029 /* Already moved. */
7030 if (task_cpu(p) != src_cpu)
7032 /* Affinity changed (again). */
7033 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7036 on_rq = p->se.on_rq;
7038 deactivate_task(rq_src, p, 0);
7040 set_task_cpu(p, dest_cpu);
7042 activate_task(rq_dest, p, 0);
7043 check_preempt_curr(rq_dest, p, 0);
7048 double_rq_unlock(rq_src, rq_dest);
7053 * migration_thread - this is a highprio system thread that performs
7054 * thread migration by bumping thread off CPU then 'pushing' onto
7057 static int migration_thread(void *data)
7059 int cpu = (long)data;
7063 BUG_ON(rq->migration_thread != current);
7065 set_current_state(TASK_INTERRUPTIBLE);
7066 while (!kthread_should_stop()) {
7067 struct migration_req *req;
7068 struct list_head *head;
7070 spin_lock_irq(&rq->lock);
7072 if (cpu_is_offline(cpu)) {
7073 spin_unlock_irq(&rq->lock);
7077 if (rq->active_balance) {
7078 active_load_balance(rq, cpu);
7079 rq->active_balance = 0;
7082 head = &rq->migration_queue;
7084 if (list_empty(head)) {
7085 spin_unlock_irq(&rq->lock);
7087 set_current_state(TASK_INTERRUPTIBLE);
7090 req = list_entry(head->next, struct migration_req, list);
7091 list_del_init(head->next);
7093 spin_unlock(&rq->lock);
7094 __migrate_task(req->task, cpu, req->dest_cpu);
7097 complete(&req->done);
7099 __set_current_state(TASK_RUNNING);
7103 /* Wait for kthread_stop */
7104 set_current_state(TASK_INTERRUPTIBLE);
7105 while (!kthread_should_stop()) {
7107 set_current_state(TASK_INTERRUPTIBLE);
7109 __set_current_state(TASK_RUNNING);
7113 #ifdef CONFIG_HOTPLUG_CPU
7115 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7119 local_irq_disable();
7120 ret = __migrate_task(p, src_cpu, dest_cpu);
7126 * Figure out where task on dead CPU should go, use force if necessary.
7128 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7131 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7134 /* Look for allowed, online CPU in same node. */
7135 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7136 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7139 /* Any allowed, online CPU? */
7140 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7141 if (dest_cpu < nr_cpu_ids)
7144 /* No more Mr. Nice Guy. */
7145 if (dest_cpu >= nr_cpu_ids) {
7146 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7147 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7150 * Don't tell them about moving exiting tasks or
7151 * kernel threads (both mm NULL), since they never
7154 if (p->mm && printk_ratelimit()) {
7155 printk(KERN_INFO "process %d (%s) no "
7156 "longer affine to cpu%d\n",
7157 task_pid_nr(p), p->comm, dead_cpu);
7162 /* It can have affinity changed while we were choosing. */
7163 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7168 * While a dead CPU has no uninterruptible tasks queued at this point,
7169 * it might still have a nonzero ->nr_uninterruptible counter, because
7170 * for performance reasons the counter is not stricly tracking tasks to
7171 * their home CPUs. So we just add the counter to another CPU's counter,
7172 * to keep the global sum constant after CPU-down:
7174 static void migrate_nr_uninterruptible(struct rq *rq_src)
7176 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7177 unsigned long flags;
7179 local_irq_save(flags);
7180 double_rq_lock(rq_src, rq_dest);
7181 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7182 rq_src->nr_uninterruptible = 0;
7183 double_rq_unlock(rq_src, rq_dest);
7184 local_irq_restore(flags);
7187 /* Run through task list and migrate tasks from the dead cpu. */
7188 static void migrate_live_tasks(int src_cpu)
7190 struct task_struct *p, *t;
7192 read_lock(&tasklist_lock);
7194 do_each_thread(t, p) {
7198 if (task_cpu(p) == src_cpu)
7199 move_task_off_dead_cpu(src_cpu, p);
7200 } while_each_thread(t, p);
7202 read_unlock(&tasklist_lock);
7206 * Schedules idle task to be the next runnable task on current CPU.
7207 * It does so by boosting its priority to highest possible.
7208 * Used by CPU offline code.
7210 void sched_idle_next(void)
7212 int this_cpu = smp_processor_id();
7213 struct rq *rq = cpu_rq(this_cpu);
7214 struct task_struct *p = rq->idle;
7215 unsigned long flags;
7217 /* cpu has to be offline */
7218 BUG_ON(cpu_online(this_cpu));
7221 * Strictly not necessary since rest of the CPUs are stopped by now
7222 * and interrupts disabled on the current cpu.
7224 spin_lock_irqsave(&rq->lock, flags);
7226 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7228 update_rq_clock(rq);
7229 activate_task(rq, p, 0);
7231 spin_unlock_irqrestore(&rq->lock, flags);
7235 * Ensures that the idle task is using init_mm right before its cpu goes
7238 void idle_task_exit(void)
7240 struct mm_struct *mm = current->active_mm;
7242 BUG_ON(cpu_online(smp_processor_id()));
7245 switch_mm(mm, &init_mm, current);
7249 /* called under rq->lock with disabled interrupts */
7250 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7252 struct rq *rq = cpu_rq(dead_cpu);
7254 /* Must be exiting, otherwise would be on tasklist. */
7255 BUG_ON(!p->exit_state);
7257 /* Cannot have done final schedule yet: would have vanished. */
7258 BUG_ON(p->state == TASK_DEAD);
7263 * Drop lock around migration; if someone else moves it,
7264 * that's OK. No task can be added to this CPU, so iteration is
7267 spin_unlock_irq(&rq->lock);
7268 move_task_off_dead_cpu(dead_cpu, p);
7269 spin_lock_irq(&rq->lock);
7274 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7275 static void migrate_dead_tasks(unsigned int dead_cpu)
7277 struct rq *rq = cpu_rq(dead_cpu);
7278 struct task_struct *next;
7281 if (!rq->nr_running)
7283 update_rq_clock(rq);
7284 next = pick_next_task(rq);
7287 next->sched_class->put_prev_task(rq, next);
7288 migrate_dead(dead_cpu, next);
7294 * remove the tasks which were accounted by rq from calc_load_tasks.
7296 static void calc_global_load_remove(struct rq *rq)
7298 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7300 #endif /* CONFIG_HOTPLUG_CPU */
7302 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7304 static struct ctl_table sd_ctl_dir[] = {
7306 .procname = "sched_domain",
7312 static struct ctl_table sd_ctl_root[] = {
7314 .ctl_name = CTL_KERN,
7315 .procname = "kernel",
7317 .child = sd_ctl_dir,
7322 static struct ctl_table *sd_alloc_ctl_entry(int n)
7324 struct ctl_table *entry =
7325 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7330 static void sd_free_ctl_entry(struct ctl_table **tablep)
7332 struct ctl_table *entry;
7335 * In the intermediate directories, both the child directory and
7336 * procname are dynamically allocated and could fail but the mode
7337 * will always be set. In the lowest directory the names are
7338 * static strings and all have proc handlers.
7340 for (entry = *tablep; entry->mode; entry++) {
7342 sd_free_ctl_entry(&entry->child);
7343 if (entry->proc_handler == NULL)
7344 kfree(entry->procname);
7352 set_table_entry(struct ctl_table *entry,
7353 const char *procname, void *data, int maxlen,
7354 mode_t mode, proc_handler *proc_handler)
7356 entry->procname = procname;
7358 entry->maxlen = maxlen;
7360 entry->proc_handler = proc_handler;
7363 static struct ctl_table *
7364 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7366 struct ctl_table *table = sd_alloc_ctl_entry(13);
7371 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7372 sizeof(long), 0644, proc_doulongvec_minmax);
7373 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7374 sizeof(long), 0644, proc_doulongvec_minmax);
7375 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7376 sizeof(int), 0644, proc_dointvec_minmax);
7377 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7378 sizeof(int), 0644, proc_dointvec_minmax);
7379 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7380 sizeof(int), 0644, proc_dointvec_minmax);
7381 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7382 sizeof(int), 0644, proc_dointvec_minmax);
7383 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7384 sizeof(int), 0644, proc_dointvec_minmax);
7385 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7386 sizeof(int), 0644, proc_dointvec_minmax);
7387 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7388 sizeof(int), 0644, proc_dointvec_minmax);
7389 set_table_entry(&table[9], "cache_nice_tries",
7390 &sd->cache_nice_tries,
7391 sizeof(int), 0644, proc_dointvec_minmax);
7392 set_table_entry(&table[10], "flags", &sd->flags,
7393 sizeof(int), 0644, proc_dointvec_minmax);
7394 set_table_entry(&table[11], "name", sd->name,
7395 CORENAME_MAX_SIZE, 0444, proc_dostring);
7396 /* &table[12] is terminator */
7401 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7403 struct ctl_table *entry, *table;
7404 struct sched_domain *sd;
7405 int domain_num = 0, i;
7408 for_each_domain(cpu, sd)
7410 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7415 for_each_domain(cpu, sd) {
7416 snprintf(buf, 32, "domain%d", i);
7417 entry->procname = kstrdup(buf, GFP_KERNEL);
7419 entry->child = sd_alloc_ctl_domain_table(sd);
7426 static struct ctl_table_header *sd_sysctl_header;
7427 static void register_sched_domain_sysctl(void)
7429 int i, cpu_num = num_online_cpus();
7430 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7433 WARN_ON(sd_ctl_dir[0].child);
7434 sd_ctl_dir[0].child = entry;
7439 for_each_online_cpu(i) {
7440 snprintf(buf, 32, "cpu%d", i);
7441 entry->procname = kstrdup(buf, GFP_KERNEL);
7443 entry->child = sd_alloc_ctl_cpu_table(i);
7447 WARN_ON(sd_sysctl_header);
7448 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7451 /* may be called multiple times per register */
7452 static void unregister_sched_domain_sysctl(void)
7454 if (sd_sysctl_header)
7455 unregister_sysctl_table(sd_sysctl_header);
7456 sd_sysctl_header = NULL;
7457 if (sd_ctl_dir[0].child)
7458 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7461 static void register_sched_domain_sysctl(void)
7464 static void unregister_sched_domain_sysctl(void)
7469 static void set_rq_online(struct rq *rq)
7472 const struct sched_class *class;
7474 cpumask_set_cpu(rq->cpu, rq->rd->online);
7477 for_each_class(class) {
7478 if (class->rq_online)
7479 class->rq_online(rq);
7484 static void set_rq_offline(struct rq *rq)
7487 const struct sched_class *class;
7489 for_each_class(class) {
7490 if (class->rq_offline)
7491 class->rq_offline(rq);
7494 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7500 * migration_call - callback that gets triggered when a CPU is added.
7501 * Here we can start up the necessary migration thread for the new CPU.
7503 static int __cpuinit
7504 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7506 struct task_struct *p;
7507 int cpu = (long)hcpu;
7508 unsigned long flags;
7513 case CPU_UP_PREPARE:
7514 case CPU_UP_PREPARE_FROZEN:
7515 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7518 kthread_bind(p, cpu);
7519 /* Must be high prio: stop_machine expects to yield to it. */
7520 rq = task_rq_lock(p, &flags);
7521 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7522 task_rq_unlock(rq, &flags);
7523 cpu_rq(cpu)->migration_thread = p;
7527 case CPU_ONLINE_FROZEN:
7528 /* Strictly unnecessary, as first user will wake it. */
7529 wake_up_process(cpu_rq(cpu)->migration_thread);
7531 /* Update our root-domain */
7533 spin_lock_irqsave(&rq->lock, flags);
7534 rq->calc_load_update = calc_load_update;
7535 rq->calc_load_active = 0;
7537 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7541 spin_unlock_irqrestore(&rq->lock, flags);
7544 #ifdef CONFIG_HOTPLUG_CPU
7545 case CPU_UP_CANCELED:
7546 case CPU_UP_CANCELED_FROZEN:
7547 if (!cpu_rq(cpu)->migration_thread)
7549 /* Unbind it from offline cpu so it can run. Fall thru. */
7550 kthread_bind(cpu_rq(cpu)->migration_thread,
7551 cpumask_any(cpu_online_mask));
7552 kthread_stop(cpu_rq(cpu)->migration_thread);
7553 cpu_rq(cpu)->migration_thread = NULL;
7557 case CPU_DEAD_FROZEN:
7558 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7559 migrate_live_tasks(cpu);
7561 kthread_stop(rq->migration_thread);
7562 rq->migration_thread = NULL;
7563 /* Idle task back to normal (off runqueue, low prio) */
7564 spin_lock_irq(&rq->lock);
7565 update_rq_clock(rq);
7566 deactivate_task(rq, rq->idle, 0);
7567 rq->idle->static_prio = MAX_PRIO;
7568 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7569 rq->idle->sched_class = &idle_sched_class;
7570 migrate_dead_tasks(cpu);
7571 spin_unlock_irq(&rq->lock);
7573 migrate_nr_uninterruptible(rq);
7574 BUG_ON(rq->nr_running != 0);
7575 calc_global_load_remove(rq);
7577 * No need to migrate the tasks: it was best-effort if
7578 * they didn't take sched_hotcpu_mutex. Just wake up
7581 spin_lock_irq(&rq->lock);
7582 while (!list_empty(&rq->migration_queue)) {
7583 struct migration_req *req;
7585 req = list_entry(rq->migration_queue.next,
7586 struct migration_req, list);
7587 list_del_init(&req->list);
7588 spin_unlock_irq(&rq->lock);
7589 complete(&req->done);
7590 spin_lock_irq(&rq->lock);
7592 spin_unlock_irq(&rq->lock);
7596 case CPU_DYING_FROZEN:
7597 /* Update our root-domain */
7599 spin_lock_irqsave(&rq->lock, flags);
7601 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7604 spin_unlock_irqrestore(&rq->lock, flags);
7612 * Register at high priority so that task migration (migrate_all_tasks)
7613 * happens before everything else. This has to be lower priority than
7614 * the notifier in the perf_counter subsystem, though.
7616 static struct notifier_block __cpuinitdata migration_notifier = {
7617 .notifier_call = migration_call,
7621 static int __init migration_init(void)
7623 void *cpu = (void *)(long)smp_processor_id();
7626 /* Start one for the boot CPU: */
7627 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7628 BUG_ON(err == NOTIFY_BAD);
7629 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7630 register_cpu_notifier(&migration_notifier);
7634 early_initcall(migration_init);
7639 #ifdef CONFIG_SCHED_DEBUG
7641 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7642 struct cpumask *groupmask)
7644 struct sched_group *group = sd->groups;
7647 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7648 cpumask_clear(groupmask);
7650 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7652 if (!(sd->flags & SD_LOAD_BALANCE)) {
7653 printk("does not load-balance\n");
7655 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7660 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7662 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7663 printk(KERN_ERR "ERROR: domain->span does not contain "
7666 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7667 printk(KERN_ERR "ERROR: domain->groups does not contain"
7671 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7675 printk(KERN_ERR "ERROR: group is NULL\n");
7679 if (!group->__cpu_power) {
7680 printk(KERN_CONT "\n");
7681 printk(KERN_ERR "ERROR: domain->cpu_power not "
7686 if (!cpumask_weight(sched_group_cpus(group))) {
7687 printk(KERN_CONT "\n");
7688 printk(KERN_ERR "ERROR: empty group\n");
7692 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7693 printk(KERN_CONT "\n");
7694 printk(KERN_ERR "ERROR: repeated CPUs\n");
7698 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7700 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7702 printk(KERN_CONT " %s", str);
7703 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7704 printk(KERN_CONT " (__cpu_power = %d)",
7705 group->__cpu_power);
7708 group = group->next;
7709 } while (group != sd->groups);
7710 printk(KERN_CONT "\n");
7712 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7713 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7716 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7717 printk(KERN_ERR "ERROR: parent span is not a superset "
7718 "of domain->span\n");
7722 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7724 cpumask_var_t groupmask;
7728 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7732 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7734 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7735 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7740 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7747 free_cpumask_var(groupmask);
7749 #else /* !CONFIG_SCHED_DEBUG */
7750 # define sched_domain_debug(sd, cpu) do { } while (0)
7751 #endif /* CONFIG_SCHED_DEBUG */
7753 static int sd_degenerate(struct sched_domain *sd)
7755 if (cpumask_weight(sched_domain_span(sd)) == 1)
7758 /* Following flags need at least 2 groups */
7759 if (sd->flags & (SD_LOAD_BALANCE |
7760 SD_BALANCE_NEWIDLE |
7764 SD_SHARE_PKG_RESOURCES)) {
7765 if (sd->groups != sd->groups->next)
7769 /* Following flags don't use groups */
7770 if (sd->flags & (SD_WAKE_IDLE |
7779 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7781 unsigned long cflags = sd->flags, pflags = parent->flags;
7783 if (sd_degenerate(parent))
7786 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7789 /* Does parent contain flags not in child? */
7790 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7791 if (cflags & SD_WAKE_AFFINE)
7792 pflags &= ~SD_WAKE_BALANCE;
7793 /* Flags needing groups don't count if only 1 group in parent */
7794 if (parent->groups == parent->groups->next) {
7795 pflags &= ~(SD_LOAD_BALANCE |
7796 SD_BALANCE_NEWIDLE |
7800 SD_SHARE_PKG_RESOURCES);
7801 if (nr_node_ids == 1)
7802 pflags &= ~SD_SERIALIZE;
7804 if (~cflags & pflags)
7810 static void free_rootdomain(struct root_domain *rd)
7812 cpupri_cleanup(&rd->cpupri);
7814 free_cpumask_var(rd->rto_mask);
7815 free_cpumask_var(rd->online);
7816 free_cpumask_var(rd->span);
7820 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7822 struct root_domain *old_rd = NULL;
7823 unsigned long flags;
7825 spin_lock_irqsave(&rq->lock, flags);
7830 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7833 cpumask_clear_cpu(rq->cpu, old_rd->span);
7836 * If we dont want to free the old_rt yet then
7837 * set old_rd to NULL to skip the freeing later
7840 if (!atomic_dec_and_test(&old_rd->refcount))
7844 atomic_inc(&rd->refcount);
7847 cpumask_set_cpu(rq->cpu, rd->span);
7848 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7851 spin_unlock_irqrestore(&rq->lock, flags);
7854 free_rootdomain(old_rd);
7857 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7859 gfp_t gfp = GFP_KERNEL;
7861 memset(rd, 0, sizeof(*rd));
7866 if (!alloc_cpumask_var(&rd->span, gfp))
7868 if (!alloc_cpumask_var(&rd->online, gfp))
7870 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7873 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7878 free_cpumask_var(rd->rto_mask);
7880 free_cpumask_var(rd->online);
7882 free_cpumask_var(rd->span);
7887 static void init_defrootdomain(void)
7889 init_rootdomain(&def_root_domain, true);
7891 atomic_set(&def_root_domain.refcount, 1);
7894 static struct root_domain *alloc_rootdomain(void)
7896 struct root_domain *rd;
7898 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7902 if (init_rootdomain(rd, false) != 0) {
7911 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7912 * hold the hotplug lock.
7915 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7917 struct rq *rq = cpu_rq(cpu);
7918 struct sched_domain *tmp;
7920 /* Remove the sched domains which do not contribute to scheduling. */
7921 for (tmp = sd; tmp; ) {
7922 struct sched_domain *parent = tmp->parent;
7926 if (sd_parent_degenerate(tmp, parent)) {
7927 tmp->parent = parent->parent;
7929 parent->parent->child = tmp;
7934 if (sd && sd_degenerate(sd)) {
7940 sched_domain_debug(sd, cpu);
7942 rq_attach_root(rq, rd);
7943 rcu_assign_pointer(rq->sd, sd);
7946 /* cpus with isolated domains */
7947 static cpumask_var_t cpu_isolated_map;
7949 /* Setup the mask of cpus configured for isolated domains */
7950 static int __init isolated_cpu_setup(char *str)
7952 cpulist_parse(str, cpu_isolated_map);
7956 __setup("isolcpus=", isolated_cpu_setup);
7959 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7960 * to a function which identifies what group(along with sched group) a CPU
7961 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7962 * (due to the fact that we keep track of groups covered with a struct cpumask).
7964 * init_sched_build_groups will build a circular linked list of the groups
7965 * covered by the given span, and will set each group's ->cpumask correctly,
7966 * and ->cpu_power to 0.
7969 init_sched_build_groups(const struct cpumask *span,
7970 const struct cpumask *cpu_map,
7971 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7972 struct sched_group **sg,
7973 struct cpumask *tmpmask),
7974 struct cpumask *covered, struct cpumask *tmpmask)
7976 struct sched_group *first = NULL, *last = NULL;
7979 cpumask_clear(covered);
7981 for_each_cpu(i, span) {
7982 struct sched_group *sg;
7983 int group = group_fn(i, cpu_map, &sg, tmpmask);
7986 if (cpumask_test_cpu(i, covered))
7989 cpumask_clear(sched_group_cpus(sg));
7990 sg->__cpu_power = 0;
7992 for_each_cpu(j, span) {
7993 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7996 cpumask_set_cpu(j, covered);
7997 cpumask_set_cpu(j, sched_group_cpus(sg));
8008 #define SD_NODES_PER_DOMAIN 16
8013 * find_next_best_node - find the next node to include in a sched_domain
8014 * @node: node whose sched_domain we're building
8015 * @used_nodes: nodes already in the sched_domain
8017 * Find the next node to include in a given scheduling domain. Simply
8018 * finds the closest node not already in the @used_nodes map.
8020 * Should use nodemask_t.
8022 static int find_next_best_node(int node, nodemask_t *used_nodes)
8024 int i, n, val, min_val, best_node = 0;
8028 for (i = 0; i < nr_node_ids; i++) {
8029 /* Start at @node */
8030 n = (node + i) % nr_node_ids;
8032 if (!nr_cpus_node(n))
8035 /* Skip already used nodes */
8036 if (node_isset(n, *used_nodes))
8039 /* Simple min distance search */
8040 val = node_distance(node, n);
8042 if (val < min_val) {
8048 node_set(best_node, *used_nodes);
8053 * sched_domain_node_span - get a cpumask for a node's sched_domain
8054 * @node: node whose cpumask we're constructing
8055 * @span: resulting cpumask
8057 * Given a node, construct a good cpumask for its sched_domain to span. It
8058 * should be one that prevents unnecessary balancing, but also spreads tasks
8061 static void sched_domain_node_span(int node, struct cpumask *span)
8063 nodemask_t used_nodes;
8066 cpumask_clear(span);
8067 nodes_clear(used_nodes);
8069 cpumask_or(span, span, cpumask_of_node(node));
8070 node_set(node, used_nodes);
8072 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8073 int next_node = find_next_best_node(node, &used_nodes);
8075 cpumask_or(span, span, cpumask_of_node(next_node));
8078 #endif /* CONFIG_NUMA */
8080 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8083 * The cpus mask in sched_group and sched_domain hangs off the end.
8085 * ( See the the comments in include/linux/sched.h:struct sched_group
8086 * and struct sched_domain. )
8088 struct static_sched_group {
8089 struct sched_group sg;
8090 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8093 struct static_sched_domain {
8094 struct sched_domain sd;
8095 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8099 * SMT sched-domains:
8101 #ifdef CONFIG_SCHED_SMT
8102 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8103 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8106 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8107 struct sched_group **sg, struct cpumask *unused)
8110 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8113 #endif /* CONFIG_SCHED_SMT */
8116 * multi-core sched-domains:
8118 #ifdef CONFIG_SCHED_MC
8119 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8120 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8121 #endif /* CONFIG_SCHED_MC */
8123 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8125 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8126 struct sched_group **sg, struct cpumask *mask)
8130 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8131 group = cpumask_first(mask);
8133 *sg = &per_cpu(sched_group_core, group).sg;
8136 #elif defined(CONFIG_SCHED_MC)
8138 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8139 struct sched_group **sg, struct cpumask *unused)
8142 *sg = &per_cpu(sched_group_core, cpu).sg;
8147 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8148 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8151 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8152 struct sched_group **sg, struct cpumask *mask)
8155 #ifdef CONFIG_SCHED_MC
8156 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8157 group = cpumask_first(mask);
8158 #elif defined(CONFIG_SCHED_SMT)
8159 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8160 group = cpumask_first(mask);
8165 *sg = &per_cpu(sched_group_phys, group).sg;
8171 * The init_sched_build_groups can't handle what we want to do with node
8172 * groups, so roll our own. Now each node has its own list of groups which
8173 * gets dynamically allocated.
8175 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8176 static struct sched_group ***sched_group_nodes_bycpu;
8178 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8179 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8181 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8182 struct sched_group **sg,
8183 struct cpumask *nodemask)
8187 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8188 group = cpumask_first(nodemask);
8191 *sg = &per_cpu(sched_group_allnodes, group).sg;
8195 static void init_numa_sched_groups_power(struct sched_group *group_head)
8197 struct sched_group *sg = group_head;
8203 for_each_cpu(j, sched_group_cpus(sg)) {
8204 struct sched_domain *sd;
8206 sd = &per_cpu(phys_domains, j).sd;
8207 if (j != group_first_cpu(sd->groups)) {
8209 * Only add "power" once for each
8215 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8218 } while (sg != group_head);
8220 #endif /* CONFIG_NUMA */
8223 /* Free memory allocated for various sched_group structures */
8224 static void free_sched_groups(const struct cpumask *cpu_map,
8225 struct cpumask *nodemask)
8229 for_each_cpu(cpu, cpu_map) {
8230 struct sched_group **sched_group_nodes
8231 = sched_group_nodes_bycpu[cpu];
8233 if (!sched_group_nodes)
8236 for (i = 0; i < nr_node_ids; i++) {
8237 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8239 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8240 if (cpumask_empty(nodemask))
8250 if (oldsg != sched_group_nodes[i])
8253 kfree(sched_group_nodes);
8254 sched_group_nodes_bycpu[cpu] = NULL;
8257 #else /* !CONFIG_NUMA */
8258 static void free_sched_groups(const struct cpumask *cpu_map,
8259 struct cpumask *nodemask)
8262 #endif /* CONFIG_NUMA */
8265 * Initialize sched groups cpu_power.
8267 * cpu_power indicates the capacity of sched group, which is used while
8268 * distributing the load between different sched groups in a sched domain.
8269 * Typically cpu_power for all the groups in a sched domain will be same unless
8270 * there are asymmetries in the topology. If there are asymmetries, group
8271 * having more cpu_power will pickup more load compared to the group having
8274 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8275 * the maximum number of tasks a group can handle in the presence of other idle
8276 * or lightly loaded groups in the same sched domain.
8278 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8280 struct sched_domain *child;
8281 struct sched_group *group;
8283 WARN_ON(!sd || !sd->groups);
8285 if (cpu != group_first_cpu(sd->groups))
8290 sd->groups->__cpu_power = 0;
8293 * For perf policy, if the groups in child domain share resources
8294 * (for example cores sharing some portions of the cache hierarchy
8295 * or SMT), then set this domain groups cpu_power such that each group
8296 * can handle only one task, when there are other idle groups in the
8297 * same sched domain.
8299 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8301 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8302 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8307 * add cpu_power of each child group to this groups cpu_power
8309 group = child->groups;
8311 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8312 group = group->next;
8313 } while (group != child->groups);
8317 * Initializers for schedule domains
8318 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8321 #ifdef CONFIG_SCHED_DEBUG
8322 # define SD_INIT_NAME(sd, type) sd->name = #type
8324 # define SD_INIT_NAME(sd, type) do { } while (0)
8327 #define SD_INIT(sd, type) sd_init_##type(sd)
8329 #define SD_INIT_FUNC(type) \
8330 static noinline void sd_init_##type(struct sched_domain *sd) \
8332 memset(sd, 0, sizeof(*sd)); \
8333 *sd = SD_##type##_INIT; \
8334 sd->level = SD_LV_##type; \
8335 SD_INIT_NAME(sd, type); \
8340 SD_INIT_FUNC(ALLNODES)
8343 #ifdef CONFIG_SCHED_SMT
8344 SD_INIT_FUNC(SIBLING)
8346 #ifdef CONFIG_SCHED_MC
8350 static int default_relax_domain_level = -1;
8352 static int __init setup_relax_domain_level(char *str)
8356 val = simple_strtoul(str, NULL, 0);
8357 if (val < SD_LV_MAX)
8358 default_relax_domain_level = val;
8362 __setup("relax_domain_level=", setup_relax_domain_level);
8364 static void set_domain_attribute(struct sched_domain *sd,
8365 struct sched_domain_attr *attr)
8369 if (!attr || attr->relax_domain_level < 0) {
8370 if (default_relax_domain_level < 0)
8373 request = default_relax_domain_level;
8375 request = attr->relax_domain_level;
8376 if (request < sd->level) {
8377 /* turn off idle balance on this domain */
8378 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8380 /* turn on idle balance on this domain */
8381 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8386 * Build sched domains for a given set of cpus and attach the sched domains
8387 * to the individual cpus
8389 static int __build_sched_domains(const struct cpumask *cpu_map,
8390 struct sched_domain_attr *attr)
8392 int i, err = -ENOMEM;
8393 struct root_domain *rd;
8394 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8397 cpumask_var_t domainspan, covered, notcovered;
8398 struct sched_group **sched_group_nodes = NULL;
8399 int sd_allnodes = 0;
8401 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8403 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8404 goto free_domainspan;
8405 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
8409 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8410 goto free_notcovered;
8411 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8413 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8414 goto free_this_sibling_map;
8415 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8416 goto free_this_core_map;
8417 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8418 goto free_send_covered;
8422 * Allocate the per-node list of sched groups
8424 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8426 if (!sched_group_nodes) {
8427 printk(KERN_WARNING "Can not alloc sched group node list\n");
8432 rd = alloc_rootdomain();
8434 printk(KERN_WARNING "Cannot alloc root domain\n");
8435 goto free_sched_groups;
8439 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8443 * Set up domains for cpus specified by the cpu_map.
8445 for_each_cpu(i, cpu_map) {
8446 struct sched_domain *sd = NULL, *p;
8448 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8451 if (cpumask_weight(cpu_map) >
8452 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8453 sd = &per_cpu(allnodes_domains, i).sd;
8454 SD_INIT(sd, ALLNODES);
8455 set_domain_attribute(sd, attr);
8456 cpumask_copy(sched_domain_span(sd), cpu_map);
8457 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8463 sd = &per_cpu(node_domains, i).sd;
8465 set_domain_attribute(sd, attr);
8466 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8470 cpumask_and(sched_domain_span(sd),
8471 sched_domain_span(sd), cpu_map);
8475 sd = &per_cpu(phys_domains, i).sd;
8477 set_domain_attribute(sd, attr);
8478 cpumask_copy(sched_domain_span(sd), nodemask);
8482 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8484 #ifdef CONFIG_SCHED_MC
8486 sd = &per_cpu(core_domains, i).sd;
8488 set_domain_attribute(sd, attr);
8489 cpumask_and(sched_domain_span(sd), cpu_map,
8490 cpu_coregroup_mask(i));
8493 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8496 #ifdef CONFIG_SCHED_SMT
8498 sd = &per_cpu(cpu_domains, i).sd;
8499 SD_INIT(sd, SIBLING);
8500 set_domain_attribute(sd, attr);
8501 cpumask_and(sched_domain_span(sd),
8502 topology_thread_cpumask(i), cpu_map);
8505 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8509 #ifdef CONFIG_SCHED_SMT
8510 /* Set up CPU (sibling) groups */
8511 for_each_cpu(i, cpu_map) {
8512 cpumask_and(this_sibling_map,
8513 topology_thread_cpumask(i), cpu_map);
8514 if (i != cpumask_first(this_sibling_map))
8517 init_sched_build_groups(this_sibling_map, cpu_map,
8519 send_covered, tmpmask);
8523 #ifdef CONFIG_SCHED_MC
8524 /* Set up multi-core groups */
8525 for_each_cpu(i, cpu_map) {
8526 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8527 if (i != cpumask_first(this_core_map))
8530 init_sched_build_groups(this_core_map, cpu_map,
8532 send_covered, tmpmask);
8536 /* Set up physical groups */
8537 for (i = 0; i < nr_node_ids; i++) {
8538 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8539 if (cpumask_empty(nodemask))
8542 init_sched_build_groups(nodemask, cpu_map,
8544 send_covered, tmpmask);
8548 /* Set up node groups */
8550 init_sched_build_groups(cpu_map, cpu_map,
8551 &cpu_to_allnodes_group,
8552 send_covered, tmpmask);
8555 for (i = 0; i < nr_node_ids; i++) {
8556 /* Set up node groups */
8557 struct sched_group *sg, *prev;
8560 cpumask_clear(covered);
8561 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8562 if (cpumask_empty(nodemask)) {
8563 sched_group_nodes[i] = NULL;
8567 sched_domain_node_span(i, domainspan);
8568 cpumask_and(domainspan, domainspan, cpu_map);
8570 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8573 printk(KERN_WARNING "Can not alloc domain group for "
8577 sched_group_nodes[i] = sg;
8578 for_each_cpu(j, nodemask) {
8579 struct sched_domain *sd;
8581 sd = &per_cpu(node_domains, j).sd;
8584 sg->__cpu_power = 0;
8585 cpumask_copy(sched_group_cpus(sg), nodemask);
8587 cpumask_or(covered, covered, nodemask);
8590 for (j = 0; j < nr_node_ids; j++) {
8591 int n = (i + j) % nr_node_ids;
8593 cpumask_complement(notcovered, covered);
8594 cpumask_and(tmpmask, notcovered, cpu_map);
8595 cpumask_and(tmpmask, tmpmask, domainspan);
8596 if (cpumask_empty(tmpmask))
8599 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8600 if (cpumask_empty(tmpmask))
8603 sg = kmalloc_node(sizeof(struct sched_group) +
8608 "Can not alloc domain group for node %d\n", j);
8611 sg->__cpu_power = 0;
8612 cpumask_copy(sched_group_cpus(sg), tmpmask);
8613 sg->next = prev->next;
8614 cpumask_or(covered, covered, tmpmask);
8621 /* Calculate CPU power for physical packages and nodes */
8622 #ifdef CONFIG_SCHED_SMT
8623 for_each_cpu(i, cpu_map) {
8624 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8626 init_sched_groups_power(i, sd);
8629 #ifdef CONFIG_SCHED_MC
8630 for_each_cpu(i, cpu_map) {
8631 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8633 init_sched_groups_power(i, sd);
8637 for_each_cpu(i, cpu_map) {
8638 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8640 init_sched_groups_power(i, sd);
8644 for (i = 0; i < nr_node_ids; i++)
8645 init_numa_sched_groups_power(sched_group_nodes[i]);
8648 struct sched_group *sg;
8650 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8652 init_numa_sched_groups_power(sg);
8656 /* Attach the domains */
8657 for_each_cpu(i, cpu_map) {
8658 struct sched_domain *sd;
8659 #ifdef CONFIG_SCHED_SMT
8660 sd = &per_cpu(cpu_domains, i).sd;
8661 #elif defined(CONFIG_SCHED_MC)
8662 sd = &per_cpu(core_domains, i).sd;
8664 sd = &per_cpu(phys_domains, i).sd;
8666 cpu_attach_domain(sd, rd, i);
8672 free_cpumask_var(tmpmask);
8674 free_cpumask_var(send_covered);
8676 free_cpumask_var(this_core_map);
8677 free_this_sibling_map:
8678 free_cpumask_var(this_sibling_map);
8680 free_cpumask_var(nodemask);
8683 free_cpumask_var(notcovered);
8685 free_cpumask_var(covered);
8687 free_cpumask_var(domainspan);
8694 kfree(sched_group_nodes);
8700 free_sched_groups(cpu_map, tmpmask);
8701 free_rootdomain(rd);
8706 static int build_sched_domains(const struct cpumask *cpu_map)
8708 return __build_sched_domains(cpu_map, NULL);
8711 static struct cpumask *doms_cur; /* current sched domains */
8712 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8713 static struct sched_domain_attr *dattr_cur;
8714 /* attribues of custom domains in 'doms_cur' */
8717 * Special case: If a kmalloc of a doms_cur partition (array of
8718 * cpumask) fails, then fallback to a single sched domain,
8719 * as determined by the single cpumask fallback_doms.
8721 static cpumask_var_t fallback_doms;
8724 * arch_update_cpu_topology lets virtualized architectures update the
8725 * cpu core maps. It is supposed to return 1 if the topology changed
8726 * or 0 if it stayed the same.
8728 int __attribute__((weak)) arch_update_cpu_topology(void)
8734 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8735 * For now this just excludes isolated cpus, but could be used to
8736 * exclude other special cases in the future.
8738 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8742 arch_update_cpu_topology();
8744 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8746 doms_cur = fallback_doms;
8747 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8749 err = build_sched_domains(doms_cur);
8750 register_sched_domain_sysctl();
8755 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8756 struct cpumask *tmpmask)
8758 free_sched_groups(cpu_map, tmpmask);
8762 * Detach sched domains from a group of cpus specified in cpu_map
8763 * These cpus will now be attached to the NULL domain
8765 static void detach_destroy_domains(const struct cpumask *cpu_map)
8767 /* Save because hotplug lock held. */
8768 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8771 for_each_cpu(i, cpu_map)
8772 cpu_attach_domain(NULL, &def_root_domain, i);
8773 synchronize_sched();
8774 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8777 /* handle null as "default" */
8778 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8779 struct sched_domain_attr *new, int idx_new)
8781 struct sched_domain_attr tmp;
8788 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8789 new ? (new + idx_new) : &tmp,
8790 sizeof(struct sched_domain_attr));
8794 * Partition sched domains as specified by the 'ndoms_new'
8795 * cpumasks in the array doms_new[] of cpumasks. This compares
8796 * doms_new[] to the current sched domain partitioning, doms_cur[].
8797 * It destroys each deleted domain and builds each new domain.
8799 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8800 * The masks don't intersect (don't overlap.) We should setup one
8801 * sched domain for each mask. CPUs not in any of the cpumasks will
8802 * not be load balanced. If the same cpumask appears both in the
8803 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8806 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8807 * ownership of it and will kfree it when done with it. If the caller
8808 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8809 * ndoms_new == 1, and partition_sched_domains() will fallback to
8810 * the single partition 'fallback_doms', it also forces the domains
8813 * If doms_new == NULL it will be replaced with cpu_online_mask.
8814 * ndoms_new == 0 is a special case for destroying existing domains,
8815 * and it will not create the default domain.
8817 * Call with hotplug lock held
8819 /* FIXME: Change to struct cpumask *doms_new[] */
8820 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8821 struct sched_domain_attr *dattr_new)
8826 mutex_lock(&sched_domains_mutex);
8828 /* always unregister in case we don't destroy any domains */
8829 unregister_sched_domain_sysctl();
8831 /* Let architecture update cpu core mappings. */
8832 new_topology = arch_update_cpu_topology();
8834 n = doms_new ? ndoms_new : 0;
8836 /* Destroy deleted domains */
8837 for (i = 0; i < ndoms_cur; i++) {
8838 for (j = 0; j < n && !new_topology; j++) {
8839 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8840 && dattrs_equal(dattr_cur, i, dattr_new, j))
8843 /* no match - a current sched domain not in new doms_new[] */
8844 detach_destroy_domains(doms_cur + i);
8849 if (doms_new == NULL) {
8851 doms_new = fallback_doms;
8852 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8853 WARN_ON_ONCE(dattr_new);
8856 /* Build new domains */
8857 for (i = 0; i < ndoms_new; i++) {
8858 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8859 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8860 && dattrs_equal(dattr_new, i, dattr_cur, j))
8863 /* no match - add a new doms_new */
8864 __build_sched_domains(doms_new + i,
8865 dattr_new ? dattr_new + i : NULL);
8870 /* Remember the new sched domains */
8871 if (doms_cur != fallback_doms)
8873 kfree(dattr_cur); /* kfree(NULL) is safe */
8874 doms_cur = doms_new;
8875 dattr_cur = dattr_new;
8876 ndoms_cur = ndoms_new;
8878 register_sched_domain_sysctl();
8880 mutex_unlock(&sched_domains_mutex);
8883 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8884 static void arch_reinit_sched_domains(void)
8888 /* Destroy domains first to force the rebuild */
8889 partition_sched_domains(0, NULL, NULL);
8891 rebuild_sched_domains();
8895 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8897 unsigned int level = 0;
8899 if (sscanf(buf, "%u", &level) != 1)
8903 * level is always be positive so don't check for
8904 * level < POWERSAVINGS_BALANCE_NONE which is 0
8905 * What happens on 0 or 1 byte write,
8906 * need to check for count as well?
8909 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8913 sched_smt_power_savings = level;
8915 sched_mc_power_savings = level;
8917 arch_reinit_sched_domains();
8922 #ifdef CONFIG_SCHED_MC
8923 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8926 return sprintf(page, "%u\n", sched_mc_power_savings);
8928 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8929 const char *buf, size_t count)
8931 return sched_power_savings_store(buf, count, 0);
8933 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8934 sched_mc_power_savings_show,
8935 sched_mc_power_savings_store);
8938 #ifdef CONFIG_SCHED_SMT
8939 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8942 return sprintf(page, "%u\n", sched_smt_power_savings);
8944 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8945 const char *buf, size_t count)
8947 return sched_power_savings_store(buf, count, 1);
8949 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8950 sched_smt_power_savings_show,
8951 sched_smt_power_savings_store);
8954 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8958 #ifdef CONFIG_SCHED_SMT
8960 err = sysfs_create_file(&cls->kset.kobj,
8961 &attr_sched_smt_power_savings.attr);
8963 #ifdef CONFIG_SCHED_MC
8964 if (!err && mc_capable())
8965 err = sysfs_create_file(&cls->kset.kobj,
8966 &attr_sched_mc_power_savings.attr);
8970 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8972 #ifndef CONFIG_CPUSETS
8974 * Add online and remove offline CPUs from the scheduler domains.
8975 * When cpusets are enabled they take over this function.
8977 static int update_sched_domains(struct notifier_block *nfb,
8978 unsigned long action, void *hcpu)
8982 case CPU_ONLINE_FROZEN:
8984 case CPU_DEAD_FROZEN:
8985 partition_sched_domains(1, NULL, NULL);
8994 static int update_runtime(struct notifier_block *nfb,
8995 unsigned long action, void *hcpu)
8997 int cpu = (int)(long)hcpu;
9000 case CPU_DOWN_PREPARE:
9001 case CPU_DOWN_PREPARE_FROZEN:
9002 disable_runtime(cpu_rq(cpu));
9005 case CPU_DOWN_FAILED:
9006 case CPU_DOWN_FAILED_FROZEN:
9008 case CPU_ONLINE_FROZEN:
9009 enable_runtime(cpu_rq(cpu));
9017 void __init sched_init_smp(void)
9019 cpumask_var_t non_isolated_cpus;
9021 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9023 #if defined(CONFIG_NUMA)
9024 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9026 BUG_ON(sched_group_nodes_bycpu == NULL);
9029 mutex_lock(&sched_domains_mutex);
9030 arch_init_sched_domains(cpu_online_mask);
9031 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9032 if (cpumask_empty(non_isolated_cpus))
9033 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9034 mutex_unlock(&sched_domains_mutex);
9037 #ifndef CONFIG_CPUSETS
9038 /* XXX: Theoretical race here - CPU may be hotplugged now */
9039 hotcpu_notifier(update_sched_domains, 0);
9042 /* RT runtime code needs to handle some hotplug events */
9043 hotcpu_notifier(update_runtime, 0);
9047 /* Move init over to a non-isolated CPU */
9048 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9050 sched_init_granularity();
9051 free_cpumask_var(non_isolated_cpus);
9053 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9054 init_sched_rt_class();
9057 void __init sched_init_smp(void)
9059 sched_init_granularity();
9061 #endif /* CONFIG_SMP */
9063 int in_sched_functions(unsigned long addr)
9065 return in_lock_functions(addr) ||
9066 (addr >= (unsigned long)__sched_text_start
9067 && addr < (unsigned long)__sched_text_end);
9070 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9072 cfs_rq->tasks_timeline = RB_ROOT;
9073 INIT_LIST_HEAD(&cfs_rq->tasks);
9074 #ifdef CONFIG_FAIR_GROUP_SCHED
9077 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9080 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9082 struct rt_prio_array *array;
9085 array = &rt_rq->active;
9086 for (i = 0; i < MAX_RT_PRIO; i++) {
9087 INIT_LIST_HEAD(array->queue + i);
9088 __clear_bit(i, array->bitmap);
9090 /* delimiter for bitsearch: */
9091 __set_bit(MAX_RT_PRIO, array->bitmap);
9093 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9094 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9096 rt_rq->highest_prio.next = MAX_RT_PRIO;
9100 rt_rq->rt_nr_migratory = 0;
9101 rt_rq->overloaded = 0;
9102 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
9106 rt_rq->rt_throttled = 0;
9107 rt_rq->rt_runtime = 0;
9108 spin_lock_init(&rt_rq->rt_runtime_lock);
9110 #ifdef CONFIG_RT_GROUP_SCHED
9111 rt_rq->rt_nr_boosted = 0;
9116 #ifdef CONFIG_FAIR_GROUP_SCHED
9117 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9118 struct sched_entity *se, int cpu, int add,
9119 struct sched_entity *parent)
9121 struct rq *rq = cpu_rq(cpu);
9122 tg->cfs_rq[cpu] = cfs_rq;
9123 init_cfs_rq(cfs_rq, rq);
9126 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9129 /* se could be NULL for init_task_group */
9134 se->cfs_rq = &rq->cfs;
9136 se->cfs_rq = parent->my_q;
9139 se->load.weight = tg->shares;
9140 se->load.inv_weight = 0;
9141 se->parent = parent;
9145 #ifdef CONFIG_RT_GROUP_SCHED
9146 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9147 struct sched_rt_entity *rt_se, int cpu, int add,
9148 struct sched_rt_entity *parent)
9150 struct rq *rq = cpu_rq(cpu);
9152 tg->rt_rq[cpu] = rt_rq;
9153 init_rt_rq(rt_rq, rq);
9155 rt_rq->rt_se = rt_se;
9156 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9158 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9160 tg->rt_se[cpu] = rt_se;
9165 rt_se->rt_rq = &rq->rt;
9167 rt_se->rt_rq = parent->my_q;
9169 rt_se->my_q = rt_rq;
9170 rt_se->parent = parent;
9171 INIT_LIST_HEAD(&rt_se->run_list);
9175 void __init sched_init(void)
9178 unsigned long alloc_size = 0, ptr;
9180 #ifdef CONFIG_FAIR_GROUP_SCHED
9181 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9183 #ifdef CONFIG_RT_GROUP_SCHED
9184 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9186 #ifdef CONFIG_USER_SCHED
9189 #ifdef CONFIG_CPUMASK_OFFSTACK
9190 alloc_size += num_possible_cpus() * cpumask_size();
9193 * As sched_init() is called before page_alloc is setup,
9194 * we use alloc_bootmem().
9197 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9199 #ifdef CONFIG_FAIR_GROUP_SCHED
9200 init_task_group.se = (struct sched_entity **)ptr;
9201 ptr += nr_cpu_ids * sizeof(void **);
9203 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9204 ptr += nr_cpu_ids * sizeof(void **);
9206 #ifdef CONFIG_USER_SCHED
9207 root_task_group.se = (struct sched_entity **)ptr;
9208 ptr += nr_cpu_ids * sizeof(void **);
9210 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9211 ptr += nr_cpu_ids * sizeof(void **);
9212 #endif /* CONFIG_USER_SCHED */
9213 #endif /* CONFIG_FAIR_GROUP_SCHED */
9214 #ifdef CONFIG_RT_GROUP_SCHED
9215 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9216 ptr += nr_cpu_ids * sizeof(void **);
9218 init_task_group.rt_rq = (struct rt_rq **)ptr;
9219 ptr += nr_cpu_ids * sizeof(void **);
9221 #ifdef CONFIG_USER_SCHED
9222 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9223 ptr += nr_cpu_ids * sizeof(void **);
9225 root_task_group.rt_rq = (struct rt_rq **)ptr;
9226 ptr += nr_cpu_ids * sizeof(void **);
9227 #endif /* CONFIG_USER_SCHED */
9228 #endif /* CONFIG_RT_GROUP_SCHED */
9229 #ifdef CONFIG_CPUMASK_OFFSTACK
9230 for_each_possible_cpu(i) {
9231 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9232 ptr += cpumask_size();
9234 #endif /* CONFIG_CPUMASK_OFFSTACK */
9238 init_defrootdomain();
9241 init_rt_bandwidth(&def_rt_bandwidth,
9242 global_rt_period(), global_rt_runtime());
9244 #ifdef CONFIG_RT_GROUP_SCHED
9245 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9246 global_rt_period(), global_rt_runtime());
9247 #ifdef CONFIG_USER_SCHED
9248 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9249 global_rt_period(), RUNTIME_INF);
9250 #endif /* CONFIG_USER_SCHED */
9251 #endif /* CONFIG_RT_GROUP_SCHED */
9253 #ifdef CONFIG_GROUP_SCHED
9254 list_add(&init_task_group.list, &task_groups);
9255 INIT_LIST_HEAD(&init_task_group.children);
9257 #ifdef CONFIG_USER_SCHED
9258 INIT_LIST_HEAD(&root_task_group.children);
9259 init_task_group.parent = &root_task_group;
9260 list_add(&init_task_group.siblings, &root_task_group.children);
9261 #endif /* CONFIG_USER_SCHED */
9262 #endif /* CONFIG_GROUP_SCHED */
9264 for_each_possible_cpu(i) {
9268 spin_lock_init(&rq->lock);
9270 rq->calc_load_active = 0;
9271 rq->calc_load_update = jiffies + LOAD_FREQ;
9272 init_cfs_rq(&rq->cfs, rq);
9273 init_rt_rq(&rq->rt, rq);
9274 #ifdef CONFIG_FAIR_GROUP_SCHED
9275 init_task_group.shares = init_task_group_load;
9276 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9277 #ifdef CONFIG_CGROUP_SCHED
9279 * How much cpu bandwidth does init_task_group get?
9281 * In case of task-groups formed thr' the cgroup filesystem, it
9282 * gets 100% of the cpu resources in the system. This overall
9283 * system cpu resource is divided among the tasks of
9284 * init_task_group and its child task-groups in a fair manner,
9285 * based on each entity's (task or task-group's) weight
9286 * (se->load.weight).
9288 * In other words, if init_task_group has 10 tasks of weight
9289 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9290 * then A0's share of the cpu resource is:
9292 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9294 * We achieve this by letting init_task_group's tasks sit
9295 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9297 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9298 #elif defined CONFIG_USER_SCHED
9299 root_task_group.shares = NICE_0_LOAD;
9300 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9302 * In case of task-groups formed thr' the user id of tasks,
9303 * init_task_group represents tasks belonging to root user.
9304 * Hence it forms a sibling of all subsequent groups formed.
9305 * In this case, init_task_group gets only a fraction of overall
9306 * system cpu resource, based on the weight assigned to root
9307 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9308 * by letting tasks of init_task_group sit in a separate cfs_rq
9309 * (init_cfs_rq) and having one entity represent this group of
9310 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9312 init_tg_cfs_entry(&init_task_group,
9313 &per_cpu(init_cfs_rq, i),
9314 &per_cpu(init_sched_entity, i), i, 1,
9315 root_task_group.se[i]);
9318 #endif /* CONFIG_FAIR_GROUP_SCHED */
9320 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9321 #ifdef CONFIG_RT_GROUP_SCHED
9322 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9323 #ifdef CONFIG_CGROUP_SCHED
9324 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9325 #elif defined CONFIG_USER_SCHED
9326 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9327 init_tg_rt_entry(&init_task_group,
9328 &per_cpu(init_rt_rq, i),
9329 &per_cpu(init_sched_rt_entity, i), i, 1,
9330 root_task_group.rt_se[i]);
9334 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9335 rq->cpu_load[j] = 0;
9339 rq->active_balance = 0;
9340 rq->next_balance = jiffies;
9344 rq->migration_thread = NULL;
9345 INIT_LIST_HEAD(&rq->migration_queue);
9346 rq_attach_root(rq, &def_root_domain);
9349 atomic_set(&rq->nr_iowait, 0);
9352 set_load_weight(&init_task);
9354 #ifdef CONFIG_PREEMPT_NOTIFIERS
9355 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9359 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9362 #ifdef CONFIG_RT_MUTEXES
9363 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9367 * The boot idle thread does lazy MMU switching as well:
9369 atomic_inc(&init_mm.mm_count);
9370 enter_lazy_tlb(&init_mm, current);
9373 * Make us the idle thread. Technically, schedule() should not be
9374 * called from this thread, however somewhere below it might be,
9375 * but because we are the idle thread, we just pick up running again
9376 * when this runqueue becomes "idle".
9378 init_idle(current, smp_processor_id());
9380 calc_load_update = jiffies + LOAD_FREQ;
9383 * During early bootup we pretend to be a normal task:
9385 current->sched_class = &fair_sched_class;
9387 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9388 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9391 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9392 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9394 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9397 perf_counter_init();
9399 scheduler_running = 1;
9402 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9403 void __might_sleep(char *file, int line)
9406 static unsigned long prev_jiffy; /* ratelimiting */
9408 if ((!in_atomic() && !irqs_disabled()) ||
9409 system_state != SYSTEM_RUNNING || oops_in_progress)
9411 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9413 prev_jiffy = jiffies;
9416 "BUG: sleeping function called from invalid context at %s:%d\n",
9419 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9420 in_atomic(), irqs_disabled(),
9421 current->pid, current->comm);
9423 debug_show_held_locks(current);
9424 if (irqs_disabled())
9425 print_irqtrace_events(current);
9429 EXPORT_SYMBOL(__might_sleep);
9432 #ifdef CONFIG_MAGIC_SYSRQ
9433 static void normalize_task(struct rq *rq, struct task_struct *p)
9437 update_rq_clock(rq);
9438 on_rq = p->se.on_rq;
9440 deactivate_task(rq, p, 0);
9441 __setscheduler(rq, p, SCHED_NORMAL, 0);
9443 activate_task(rq, p, 0);
9444 resched_task(rq->curr);
9448 void normalize_rt_tasks(void)
9450 struct task_struct *g, *p;
9451 unsigned long flags;
9454 read_lock_irqsave(&tasklist_lock, flags);
9455 do_each_thread(g, p) {
9457 * Only normalize user tasks:
9462 p->se.exec_start = 0;
9463 #ifdef CONFIG_SCHEDSTATS
9464 p->se.wait_start = 0;
9465 p->se.sleep_start = 0;
9466 p->se.block_start = 0;
9471 * Renice negative nice level userspace
9474 if (TASK_NICE(p) < 0 && p->mm)
9475 set_user_nice(p, 0);
9479 spin_lock(&p->pi_lock);
9480 rq = __task_rq_lock(p);
9482 normalize_task(rq, p);
9484 __task_rq_unlock(rq);
9485 spin_unlock(&p->pi_lock);
9486 } while_each_thread(g, p);
9488 read_unlock_irqrestore(&tasklist_lock, flags);
9491 #endif /* CONFIG_MAGIC_SYSRQ */
9495 * These functions are only useful for the IA64 MCA handling.
9497 * They can only be called when the whole system has been
9498 * stopped - every CPU needs to be quiescent, and no scheduling
9499 * activity can take place. Using them for anything else would
9500 * be a serious bug, and as a result, they aren't even visible
9501 * under any other configuration.
9505 * curr_task - return the current task for a given cpu.
9506 * @cpu: the processor in question.
9508 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9510 struct task_struct *curr_task(int cpu)
9512 return cpu_curr(cpu);
9516 * set_curr_task - set the current task for a given cpu.
9517 * @cpu: the processor in question.
9518 * @p: the task pointer to set.
9520 * Description: This function must only be used when non-maskable interrupts
9521 * are serviced on a separate stack. It allows the architecture to switch the
9522 * notion of the current task on a cpu in a non-blocking manner. This function
9523 * must be called with all CPU's synchronized, and interrupts disabled, the
9524 * and caller must save the original value of the current task (see
9525 * curr_task() above) and restore that value before reenabling interrupts and
9526 * re-starting the system.
9528 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9530 void set_curr_task(int cpu, struct task_struct *p)
9537 #ifdef CONFIG_FAIR_GROUP_SCHED
9538 static void free_fair_sched_group(struct task_group *tg)
9542 for_each_possible_cpu(i) {
9544 kfree(tg->cfs_rq[i]);
9554 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9556 struct cfs_rq *cfs_rq;
9557 struct sched_entity *se;
9561 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9564 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9568 tg->shares = NICE_0_LOAD;
9570 for_each_possible_cpu(i) {
9573 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9574 GFP_KERNEL, cpu_to_node(i));
9578 se = kzalloc_node(sizeof(struct sched_entity),
9579 GFP_KERNEL, cpu_to_node(i));
9583 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9592 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9594 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9595 &cpu_rq(cpu)->leaf_cfs_rq_list);
9598 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9600 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9602 #else /* !CONFG_FAIR_GROUP_SCHED */
9603 static inline void free_fair_sched_group(struct task_group *tg)
9608 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9613 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9617 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9620 #endif /* CONFIG_FAIR_GROUP_SCHED */
9622 #ifdef CONFIG_RT_GROUP_SCHED
9623 static void free_rt_sched_group(struct task_group *tg)
9627 destroy_rt_bandwidth(&tg->rt_bandwidth);
9629 for_each_possible_cpu(i) {
9631 kfree(tg->rt_rq[i]);
9633 kfree(tg->rt_se[i]);
9641 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9643 struct rt_rq *rt_rq;
9644 struct sched_rt_entity *rt_se;
9648 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9651 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9655 init_rt_bandwidth(&tg->rt_bandwidth,
9656 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9658 for_each_possible_cpu(i) {
9661 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9662 GFP_KERNEL, cpu_to_node(i));
9666 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9667 GFP_KERNEL, cpu_to_node(i));
9671 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9680 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9682 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9683 &cpu_rq(cpu)->leaf_rt_rq_list);
9686 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9688 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9690 #else /* !CONFIG_RT_GROUP_SCHED */
9691 static inline void free_rt_sched_group(struct task_group *tg)
9696 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9701 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9705 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9708 #endif /* CONFIG_RT_GROUP_SCHED */
9710 #ifdef CONFIG_GROUP_SCHED
9711 static void free_sched_group(struct task_group *tg)
9713 free_fair_sched_group(tg);
9714 free_rt_sched_group(tg);
9718 /* allocate runqueue etc for a new task group */
9719 struct task_group *sched_create_group(struct task_group *parent)
9721 struct task_group *tg;
9722 unsigned long flags;
9725 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9727 return ERR_PTR(-ENOMEM);
9729 if (!alloc_fair_sched_group(tg, parent))
9732 if (!alloc_rt_sched_group(tg, parent))
9735 spin_lock_irqsave(&task_group_lock, flags);
9736 for_each_possible_cpu(i) {
9737 register_fair_sched_group(tg, i);
9738 register_rt_sched_group(tg, i);
9740 list_add_rcu(&tg->list, &task_groups);
9742 WARN_ON(!parent); /* root should already exist */
9744 tg->parent = parent;
9745 INIT_LIST_HEAD(&tg->children);
9746 list_add_rcu(&tg->siblings, &parent->children);
9747 spin_unlock_irqrestore(&task_group_lock, flags);
9752 free_sched_group(tg);
9753 return ERR_PTR(-ENOMEM);
9756 /* rcu callback to free various structures associated with a task group */
9757 static void free_sched_group_rcu(struct rcu_head *rhp)
9759 /* now it should be safe to free those cfs_rqs */
9760 free_sched_group(container_of(rhp, struct task_group, rcu));
9763 /* Destroy runqueue etc associated with a task group */
9764 void sched_destroy_group(struct task_group *tg)
9766 unsigned long flags;
9769 spin_lock_irqsave(&task_group_lock, flags);
9770 for_each_possible_cpu(i) {
9771 unregister_fair_sched_group(tg, i);
9772 unregister_rt_sched_group(tg, i);
9774 list_del_rcu(&tg->list);
9775 list_del_rcu(&tg->siblings);
9776 spin_unlock_irqrestore(&task_group_lock, flags);
9778 /* wait for possible concurrent references to cfs_rqs complete */
9779 call_rcu(&tg->rcu, free_sched_group_rcu);
9782 /* change task's runqueue when it moves between groups.
9783 * The caller of this function should have put the task in its new group
9784 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9785 * reflect its new group.
9787 void sched_move_task(struct task_struct *tsk)
9790 unsigned long flags;
9793 rq = task_rq_lock(tsk, &flags);
9795 update_rq_clock(rq);
9797 running = task_current(rq, tsk);
9798 on_rq = tsk->se.on_rq;
9801 dequeue_task(rq, tsk, 0);
9802 if (unlikely(running))
9803 tsk->sched_class->put_prev_task(rq, tsk);
9805 set_task_rq(tsk, task_cpu(tsk));
9807 #ifdef CONFIG_FAIR_GROUP_SCHED
9808 if (tsk->sched_class->moved_group)
9809 tsk->sched_class->moved_group(tsk);
9812 if (unlikely(running))
9813 tsk->sched_class->set_curr_task(rq);
9815 enqueue_task(rq, tsk, 0);
9817 task_rq_unlock(rq, &flags);
9819 #endif /* CONFIG_GROUP_SCHED */
9821 #ifdef CONFIG_FAIR_GROUP_SCHED
9822 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9824 struct cfs_rq *cfs_rq = se->cfs_rq;
9829 dequeue_entity(cfs_rq, se, 0);
9831 se->load.weight = shares;
9832 se->load.inv_weight = 0;
9835 enqueue_entity(cfs_rq, se, 0);
9838 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9840 struct cfs_rq *cfs_rq = se->cfs_rq;
9841 struct rq *rq = cfs_rq->rq;
9842 unsigned long flags;
9844 spin_lock_irqsave(&rq->lock, flags);
9845 __set_se_shares(se, shares);
9846 spin_unlock_irqrestore(&rq->lock, flags);
9849 static DEFINE_MUTEX(shares_mutex);
9851 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9854 unsigned long flags;
9857 * We can't change the weight of the root cgroup.
9862 if (shares < MIN_SHARES)
9863 shares = MIN_SHARES;
9864 else if (shares > MAX_SHARES)
9865 shares = MAX_SHARES;
9867 mutex_lock(&shares_mutex);
9868 if (tg->shares == shares)
9871 spin_lock_irqsave(&task_group_lock, flags);
9872 for_each_possible_cpu(i)
9873 unregister_fair_sched_group(tg, i);
9874 list_del_rcu(&tg->siblings);
9875 spin_unlock_irqrestore(&task_group_lock, flags);
9877 /* wait for any ongoing reference to this group to finish */
9878 synchronize_sched();
9881 * Now we are free to modify the group's share on each cpu
9882 * w/o tripping rebalance_share or load_balance_fair.
9884 tg->shares = shares;
9885 for_each_possible_cpu(i) {
9889 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9890 set_se_shares(tg->se[i], shares);
9894 * Enable load balance activity on this group, by inserting it back on
9895 * each cpu's rq->leaf_cfs_rq_list.
9897 spin_lock_irqsave(&task_group_lock, flags);
9898 for_each_possible_cpu(i)
9899 register_fair_sched_group(tg, i);
9900 list_add_rcu(&tg->siblings, &tg->parent->children);
9901 spin_unlock_irqrestore(&task_group_lock, flags);
9903 mutex_unlock(&shares_mutex);
9907 unsigned long sched_group_shares(struct task_group *tg)
9913 #ifdef CONFIG_RT_GROUP_SCHED
9915 * Ensure that the real time constraints are schedulable.
9917 static DEFINE_MUTEX(rt_constraints_mutex);
9919 static unsigned long to_ratio(u64 period, u64 runtime)
9921 if (runtime == RUNTIME_INF)
9924 return div64_u64(runtime << 20, period);
9927 /* Must be called with tasklist_lock held */
9928 static inline int tg_has_rt_tasks(struct task_group *tg)
9930 struct task_struct *g, *p;
9932 do_each_thread(g, p) {
9933 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9935 } while_each_thread(g, p);
9940 struct rt_schedulable_data {
9941 struct task_group *tg;
9946 static int tg_schedulable(struct task_group *tg, void *data)
9948 struct rt_schedulable_data *d = data;
9949 struct task_group *child;
9950 unsigned long total, sum = 0;
9951 u64 period, runtime;
9953 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9954 runtime = tg->rt_bandwidth.rt_runtime;
9957 period = d->rt_period;
9958 runtime = d->rt_runtime;
9961 #ifdef CONFIG_USER_SCHED
9962 if (tg == &root_task_group) {
9963 period = global_rt_period();
9964 runtime = global_rt_runtime();
9969 * Cannot have more runtime than the period.
9971 if (runtime > period && runtime != RUNTIME_INF)
9975 * Ensure we don't starve existing RT tasks.
9977 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9980 total = to_ratio(period, runtime);
9983 * Nobody can have more than the global setting allows.
9985 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9989 * The sum of our children's runtime should not exceed our own.
9991 list_for_each_entry_rcu(child, &tg->children, siblings) {
9992 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9993 runtime = child->rt_bandwidth.rt_runtime;
9995 if (child == d->tg) {
9996 period = d->rt_period;
9997 runtime = d->rt_runtime;
10000 sum += to_ratio(period, runtime);
10009 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10011 struct rt_schedulable_data data = {
10013 .rt_period = period,
10014 .rt_runtime = runtime,
10017 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10020 static int tg_set_bandwidth(struct task_group *tg,
10021 u64 rt_period, u64 rt_runtime)
10025 mutex_lock(&rt_constraints_mutex);
10026 read_lock(&tasklist_lock);
10027 err = __rt_schedulable(tg, rt_period, rt_runtime);
10031 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10032 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10033 tg->rt_bandwidth.rt_runtime = rt_runtime;
10035 for_each_possible_cpu(i) {
10036 struct rt_rq *rt_rq = tg->rt_rq[i];
10038 spin_lock(&rt_rq->rt_runtime_lock);
10039 rt_rq->rt_runtime = rt_runtime;
10040 spin_unlock(&rt_rq->rt_runtime_lock);
10042 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10044 read_unlock(&tasklist_lock);
10045 mutex_unlock(&rt_constraints_mutex);
10050 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10052 u64 rt_runtime, rt_period;
10054 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10055 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10056 if (rt_runtime_us < 0)
10057 rt_runtime = RUNTIME_INF;
10059 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10062 long sched_group_rt_runtime(struct task_group *tg)
10066 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10069 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10070 do_div(rt_runtime_us, NSEC_PER_USEC);
10071 return rt_runtime_us;
10074 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10076 u64 rt_runtime, rt_period;
10078 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10079 rt_runtime = tg->rt_bandwidth.rt_runtime;
10081 if (rt_period == 0)
10084 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10087 long sched_group_rt_period(struct task_group *tg)
10091 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10092 do_div(rt_period_us, NSEC_PER_USEC);
10093 return rt_period_us;
10096 static int sched_rt_global_constraints(void)
10098 u64 runtime, period;
10101 if (sysctl_sched_rt_period <= 0)
10104 runtime = global_rt_runtime();
10105 period = global_rt_period();
10108 * Sanity check on the sysctl variables.
10110 if (runtime > period && runtime != RUNTIME_INF)
10113 mutex_lock(&rt_constraints_mutex);
10114 read_lock(&tasklist_lock);
10115 ret = __rt_schedulable(NULL, 0, 0);
10116 read_unlock(&tasklist_lock);
10117 mutex_unlock(&rt_constraints_mutex);
10122 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10124 /* Don't accept realtime tasks when there is no way for them to run */
10125 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10131 #else /* !CONFIG_RT_GROUP_SCHED */
10132 static int sched_rt_global_constraints(void)
10134 unsigned long flags;
10137 if (sysctl_sched_rt_period <= 0)
10141 * There's always some RT tasks in the root group
10142 * -- migration, kstopmachine etc..
10144 if (sysctl_sched_rt_runtime == 0)
10147 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10148 for_each_possible_cpu(i) {
10149 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10151 spin_lock(&rt_rq->rt_runtime_lock);
10152 rt_rq->rt_runtime = global_rt_runtime();
10153 spin_unlock(&rt_rq->rt_runtime_lock);
10155 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10159 #endif /* CONFIG_RT_GROUP_SCHED */
10161 int sched_rt_handler(struct ctl_table *table, int write,
10162 struct file *filp, void __user *buffer, size_t *lenp,
10166 int old_period, old_runtime;
10167 static DEFINE_MUTEX(mutex);
10169 mutex_lock(&mutex);
10170 old_period = sysctl_sched_rt_period;
10171 old_runtime = sysctl_sched_rt_runtime;
10173 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10175 if (!ret && write) {
10176 ret = sched_rt_global_constraints();
10178 sysctl_sched_rt_period = old_period;
10179 sysctl_sched_rt_runtime = old_runtime;
10181 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10182 def_rt_bandwidth.rt_period =
10183 ns_to_ktime(global_rt_period());
10186 mutex_unlock(&mutex);
10191 #ifdef CONFIG_CGROUP_SCHED
10193 /* return corresponding task_group object of a cgroup */
10194 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10196 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10197 struct task_group, css);
10200 static struct cgroup_subsys_state *
10201 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10203 struct task_group *tg, *parent;
10205 if (!cgrp->parent) {
10206 /* This is early initialization for the top cgroup */
10207 return &init_task_group.css;
10210 parent = cgroup_tg(cgrp->parent);
10211 tg = sched_create_group(parent);
10213 return ERR_PTR(-ENOMEM);
10219 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10221 struct task_group *tg = cgroup_tg(cgrp);
10223 sched_destroy_group(tg);
10227 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10228 struct task_struct *tsk)
10230 #ifdef CONFIG_RT_GROUP_SCHED
10231 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10234 /* We don't support RT-tasks being in separate groups */
10235 if (tsk->sched_class != &fair_sched_class)
10243 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10244 struct cgroup *old_cont, struct task_struct *tsk)
10246 sched_move_task(tsk);
10249 #ifdef CONFIG_FAIR_GROUP_SCHED
10250 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10253 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10256 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10258 struct task_group *tg = cgroup_tg(cgrp);
10260 return (u64) tg->shares;
10262 #endif /* CONFIG_FAIR_GROUP_SCHED */
10264 #ifdef CONFIG_RT_GROUP_SCHED
10265 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10268 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10271 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10273 return sched_group_rt_runtime(cgroup_tg(cgrp));
10276 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10279 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10282 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10284 return sched_group_rt_period(cgroup_tg(cgrp));
10286 #endif /* CONFIG_RT_GROUP_SCHED */
10288 static struct cftype cpu_files[] = {
10289 #ifdef CONFIG_FAIR_GROUP_SCHED
10292 .read_u64 = cpu_shares_read_u64,
10293 .write_u64 = cpu_shares_write_u64,
10296 #ifdef CONFIG_RT_GROUP_SCHED
10298 .name = "rt_runtime_us",
10299 .read_s64 = cpu_rt_runtime_read,
10300 .write_s64 = cpu_rt_runtime_write,
10303 .name = "rt_period_us",
10304 .read_u64 = cpu_rt_period_read_uint,
10305 .write_u64 = cpu_rt_period_write_uint,
10310 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10312 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10315 struct cgroup_subsys cpu_cgroup_subsys = {
10317 .create = cpu_cgroup_create,
10318 .destroy = cpu_cgroup_destroy,
10319 .can_attach = cpu_cgroup_can_attach,
10320 .attach = cpu_cgroup_attach,
10321 .populate = cpu_cgroup_populate,
10322 .subsys_id = cpu_cgroup_subsys_id,
10326 #endif /* CONFIG_CGROUP_SCHED */
10328 #ifdef CONFIG_CGROUP_CPUACCT
10331 * CPU accounting code for task groups.
10333 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10334 * (balbir@in.ibm.com).
10337 /* track cpu usage of a group of tasks and its child groups */
10339 struct cgroup_subsys_state css;
10340 /* cpuusage holds pointer to a u64-type object on every cpu */
10342 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10343 struct cpuacct *parent;
10346 struct cgroup_subsys cpuacct_subsys;
10348 /* return cpu accounting group corresponding to this container */
10349 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10351 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10352 struct cpuacct, css);
10355 /* return cpu accounting group to which this task belongs */
10356 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10358 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10359 struct cpuacct, css);
10362 /* create a new cpu accounting group */
10363 static struct cgroup_subsys_state *cpuacct_create(
10364 struct cgroup_subsys *ss, struct cgroup *cgrp)
10366 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10372 ca->cpuusage = alloc_percpu(u64);
10376 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10377 if (percpu_counter_init(&ca->cpustat[i], 0))
10378 goto out_free_counters;
10381 ca->parent = cgroup_ca(cgrp->parent);
10387 percpu_counter_destroy(&ca->cpustat[i]);
10388 free_percpu(ca->cpuusage);
10392 return ERR_PTR(-ENOMEM);
10395 /* destroy an existing cpu accounting group */
10397 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10399 struct cpuacct *ca = cgroup_ca(cgrp);
10402 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10403 percpu_counter_destroy(&ca->cpustat[i]);
10404 free_percpu(ca->cpuusage);
10408 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10410 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10413 #ifndef CONFIG_64BIT
10415 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10417 spin_lock_irq(&cpu_rq(cpu)->lock);
10419 spin_unlock_irq(&cpu_rq(cpu)->lock);
10427 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10429 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10431 #ifndef CONFIG_64BIT
10433 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10435 spin_lock_irq(&cpu_rq(cpu)->lock);
10437 spin_unlock_irq(&cpu_rq(cpu)->lock);
10443 /* return total cpu usage (in nanoseconds) of a group */
10444 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10446 struct cpuacct *ca = cgroup_ca(cgrp);
10447 u64 totalcpuusage = 0;
10450 for_each_present_cpu(i)
10451 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10453 return totalcpuusage;
10456 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10459 struct cpuacct *ca = cgroup_ca(cgrp);
10468 for_each_present_cpu(i)
10469 cpuacct_cpuusage_write(ca, i, 0);
10475 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10476 struct seq_file *m)
10478 struct cpuacct *ca = cgroup_ca(cgroup);
10482 for_each_present_cpu(i) {
10483 percpu = cpuacct_cpuusage_read(ca, i);
10484 seq_printf(m, "%llu ", (unsigned long long) percpu);
10486 seq_printf(m, "\n");
10490 static const char *cpuacct_stat_desc[] = {
10491 [CPUACCT_STAT_USER] = "user",
10492 [CPUACCT_STAT_SYSTEM] = "system",
10495 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10496 struct cgroup_map_cb *cb)
10498 struct cpuacct *ca = cgroup_ca(cgrp);
10501 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10502 s64 val = percpu_counter_read(&ca->cpustat[i]);
10503 val = cputime64_to_clock_t(val);
10504 cb->fill(cb, cpuacct_stat_desc[i], val);
10509 static struct cftype files[] = {
10512 .read_u64 = cpuusage_read,
10513 .write_u64 = cpuusage_write,
10516 .name = "usage_percpu",
10517 .read_seq_string = cpuacct_percpu_seq_read,
10521 .read_map = cpuacct_stats_show,
10525 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10527 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10531 * charge this task's execution time to its accounting group.
10533 * called with rq->lock held.
10535 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10537 struct cpuacct *ca;
10540 if (unlikely(!cpuacct_subsys.active))
10543 cpu = task_cpu(tsk);
10549 for (; ca; ca = ca->parent) {
10550 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10551 *cpuusage += cputime;
10558 * Charge the system/user time to the task's accounting group.
10560 static void cpuacct_update_stats(struct task_struct *tsk,
10561 enum cpuacct_stat_index idx, cputime_t val)
10563 struct cpuacct *ca;
10565 if (unlikely(!cpuacct_subsys.active))
10572 percpu_counter_add(&ca->cpustat[idx], val);
10578 struct cgroup_subsys cpuacct_subsys = {
10580 .create = cpuacct_create,
10581 .destroy = cpuacct_destroy,
10582 .populate = cpuacct_populate,
10583 .subsys_id = cpuacct_subsys_id,
10585 #endif /* CONFIG_CGROUP_CPUACCT */